Linux Audio

Check our new training course

Loading...
v6.13.7
   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		}
 317	}
 318}
 319
 320void r5l_wake_reclaim(struct r5l_log *log, sector_t space);
 321
 322/* Check whether we should flush some stripes to free up stripe cache */
 323void r5c_check_stripe_cache_usage(struct r5conf *conf)
 324{
 325	int total_cached;
 326	struct r5l_log *log = READ_ONCE(conf->log);
 327
 328	if (!r5c_is_writeback(log))
 329		return;
 330
 331	total_cached = atomic_read(&conf->r5c_cached_partial_stripes) +
 332		atomic_read(&conf->r5c_cached_full_stripes);
 333
 334	/*
 335	 * The following condition is true for either of the following:
 336	 *   - stripe cache pressure high:
 337	 *          total_cached > 3/4 min_nr_stripes ||
 338	 *          empty_inactive_list_nr > 0
 339	 *   - stripe cache pressure moderate:
 340	 *          total_cached > 1/2 min_nr_stripes
 341	 */
 342	if (total_cached > conf->min_nr_stripes * 1 / 2 ||
 343	    atomic_read(&conf->empty_inactive_list_nr) > 0)
 344		r5l_wake_reclaim(log, 0);
 345}
 346
 347/*
 348 * flush cache when there are R5C_FULL_STRIPE_FLUSH_BATCH or more full
 349 * stripes in the cache
 350 */
 351void r5c_check_cached_full_stripe(struct r5conf *conf)
 352{
 353	struct r5l_log *log = READ_ONCE(conf->log);
 354
 355	if (!r5c_is_writeback(log))
 356		return;
 357
 358	/*
 359	 * wake up reclaim for R5C_FULL_STRIPE_FLUSH_BATCH cached stripes
 360	 * or a full stripe (chunk size / 4k stripes).
 361	 */
 362	if (atomic_read(&conf->r5c_cached_full_stripes) >=
 363	    min(R5C_FULL_STRIPE_FLUSH_BATCH(conf),
 364		conf->chunk_sectors >> RAID5_STRIPE_SHIFT(conf)))
 365		r5l_wake_reclaim(log, 0);
 366}
 367
 368/*
 369 * Total log space (in sectors) needed to flush all data in cache
 370 *
 371 * To avoid deadlock due to log space, it is necessary to reserve log
 372 * space to flush critical stripes (stripes that occupying log space near
 373 * last_checkpoint). This function helps check how much log space is
 374 * required to flush all cached stripes.
 375 *
 376 * To reduce log space requirements, two mechanisms are used to give cache
 377 * flush higher priorities:
 378 *    1. In handle_stripe_dirtying() and schedule_reconstruction(),
 379 *       stripes ALREADY in journal can be flushed w/o pending writes;
 380 *    2. In r5l_write_stripe() and r5c_cache_data(), stripes NOT in journal
 381 *       can be delayed (r5l_add_no_space_stripe).
 382 *
 383 * In cache flush, the stripe goes through 1 and then 2. For a stripe that
 384 * already passed 1, flushing it requires at most (conf->max_degraded + 1)
 385 * pages of journal space. For stripes that has not passed 1, flushing it
 386 * requires (conf->raid_disks + 1) pages of journal space. There are at
 387 * most (conf->group_cnt + 1) stripe that passed 1. So total journal space
 388 * required to flush all cached stripes (in pages) is:
 389 *
 390 *     (stripe_in_journal_count - group_cnt - 1) * (max_degraded + 1) +
 391 *     (group_cnt + 1) * (raid_disks + 1)
 392 * or
 393 *     (stripe_in_journal_count) * (max_degraded + 1) +
 394 *     (group_cnt + 1) * (raid_disks - max_degraded)
 395 */
 396static sector_t r5c_log_required_to_flush_cache(struct r5conf *conf)
 397{
 398	struct r5l_log *log = READ_ONCE(conf->log);
 399
 400	if (!r5c_is_writeback(log))
 401		return 0;
 402
 403	return BLOCK_SECTORS *
 404		((conf->max_degraded + 1) * atomic_read(&log->stripe_in_journal_count) +
 405		 (conf->raid_disks - conf->max_degraded) * (conf->group_cnt + 1));
 406}
 407
 408/*
 409 * evaluate log space usage and update R5C_LOG_TIGHT and R5C_LOG_CRITICAL
 410 *
 411 * R5C_LOG_TIGHT is set when free space on the log device is less than 3x of
 412 * reclaim_required_space. R5C_LOG_CRITICAL is set when free space on the log
 413 * device is less than 2x of reclaim_required_space.
 414 */
 415static inline void r5c_update_log_state(struct r5l_log *log)
 416{
 417	struct r5conf *conf = log->rdev->mddev->private;
 418	sector_t free_space;
 419	sector_t reclaim_space;
 420	bool wake_reclaim = false;
 421
 422	if (!r5c_is_writeback(log))
 423		return;
 424
 425	free_space = r5l_ring_distance(log, log->log_start,
 426				       log->last_checkpoint);
 427	reclaim_space = r5c_log_required_to_flush_cache(conf);
 428	if (free_space < 2 * reclaim_space)
 429		set_bit(R5C_LOG_CRITICAL, &conf->cache_state);
 430	else {
 431		if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state))
 432			wake_reclaim = true;
 433		clear_bit(R5C_LOG_CRITICAL, &conf->cache_state);
 434	}
 435	if (free_space < 3 * reclaim_space)
 436		set_bit(R5C_LOG_TIGHT, &conf->cache_state);
 437	else
 438		clear_bit(R5C_LOG_TIGHT, &conf->cache_state);
 439
 440	if (wake_reclaim)
 441		r5l_wake_reclaim(log, 0);
 442}
 443
 444/*
 445 * Put the stripe into writing-out phase by clearing STRIPE_R5C_CACHING.
 446 * This function should only be called in write-back mode.
 447 */
 448void r5c_make_stripe_write_out(struct stripe_head *sh)
 449{
 450	struct r5conf *conf = sh->raid_conf;
 451	struct r5l_log *log = READ_ONCE(conf->log);
 452
 453	BUG_ON(!r5c_is_writeback(log));
 454
 455	WARN_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state));
 456	clear_bit(STRIPE_R5C_CACHING, &sh->state);
 457
 458	if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
 459		atomic_inc(&conf->preread_active_stripes);
 460}
 461
 462static void r5c_handle_data_cached(struct stripe_head *sh)
 463{
 464	int i;
 465
 466	for (i = sh->disks; i--; )
 467		if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
 468			set_bit(R5_InJournal, &sh->dev[i].flags);
 469			clear_bit(R5_LOCKED, &sh->dev[i].flags);
 470		}
 471	clear_bit(STRIPE_LOG_TRAPPED, &sh->state);
 472}
 473
 474/*
 475 * this journal write must contain full parity,
 476 * it may also contain some data pages
 477 */
 478static void r5c_handle_parity_cached(struct stripe_head *sh)
 479{
 480	int i;
 481
 482	for (i = sh->disks; i--; )
 483		if (test_bit(R5_InJournal, &sh->dev[i].flags))
 484			set_bit(R5_Wantwrite, &sh->dev[i].flags);
 485}
 486
 487/*
 488 * Setting proper flags after writing (or flushing) data and/or parity to the
 489 * log device. This is called from r5l_log_endio() or r5l_log_flush_endio().
 490 */
 491static void r5c_finish_cache_stripe(struct stripe_head *sh)
 492{
 493	struct r5l_log *log = READ_ONCE(sh->raid_conf->log);
 494
 495	if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) {
 496		BUG_ON(test_bit(STRIPE_R5C_CACHING, &sh->state));
 497		/*
 498		 * Set R5_InJournal for parity dev[pd_idx]. This means
 499		 * all data AND parity in the journal. For RAID 6, it is
 500		 * NOT necessary to set the flag for dev[qd_idx], as the
 501		 * two parities are written out together.
 502		 */
 503		set_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags);
 504	} else if (test_bit(STRIPE_R5C_CACHING, &sh->state)) {
 505		r5c_handle_data_cached(sh);
 506	} else {
 507		r5c_handle_parity_cached(sh);
 508		set_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags);
 509	}
 510}
 511
 512static void r5l_io_run_stripes(struct r5l_io_unit *io)
 513{
 514	struct stripe_head *sh, *next;
 515
 516	list_for_each_entry_safe(sh, next, &io->stripe_list, log_list) {
 517		list_del_init(&sh->log_list);
 518
 519		r5c_finish_cache_stripe(sh);
 520
 521		set_bit(STRIPE_HANDLE, &sh->state);
 522		raid5_release_stripe(sh);
 523	}
 524}
 525
 526static void r5l_log_run_stripes(struct r5l_log *log)
 527{
 528	struct r5l_io_unit *io, *next;
 529
 530	lockdep_assert_held(&log->io_list_lock);
 531
 532	list_for_each_entry_safe(io, next, &log->running_ios, log_sibling) {
 533		/* don't change list order */
 534		if (io->state < IO_UNIT_IO_END)
 535			break;
 536
 537		list_move_tail(&io->log_sibling, &log->finished_ios);
 538		r5l_io_run_stripes(io);
 539	}
 540}
 541
 542static void r5l_move_to_end_ios(struct r5l_log *log)
 543{
 544	struct r5l_io_unit *io, *next;
 545
 546	lockdep_assert_held(&log->io_list_lock);
 547
 548	list_for_each_entry_safe(io, next, &log->running_ios, log_sibling) {
 549		/* don't change list order */
 550		if (io->state < IO_UNIT_IO_END)
 551			break;
 552		list_move_tail(&io->log_sibling, &log->io_end_ios);
 553	}
 554}
 555
 556static void __r5l_stripe_write_finished(struct r5l_io_unit *io);
 557static void r5l_log_endio(struct bio *bio)
 558{
 559	struct r5l_io_unit *io = bio->bi_private;
 560	struct r5l_io_unit *io_deferred;
 561	struct r5l_log *log = io->log;
 562	unsigned long flags;
 563	bool has_null_flush;
 564	bool has_flush_payload;
 565
 566	if (bio->bi_status)
 567		md_error(log->rdev->mddev, log->rdev);
 568
 569	bio_put(bio);
 570	mempool_free(io->meta_page, &log->meta_pool);
 571
 572	spin_lock_irqsave(&log->io_list_lock, flags);
 573	__r5l_set_io_unit_state(io, IO_UNIT_IO_END);
 574
 575	/*
 576	 * if the io doesn't not have null_flush or flush payload,
 577	 * it is not safe to access it after releasing io_list_lock.
 578	 * Therefore, it is necessary to check the condition with
 579	 * the lock held.
 580	 */
 581	has_null_flush = io->has_null_flush;
 582	has_flush_payload = io->has_flush_payload;
 583
 584	if (log->need_cache_flush && !list_empty(&io->stripe_list))
 585		r5l_move_to_end_ios(log);
 586	else
 587		r5l_log_run_stripes(log);
 588	if (!list_empty(&log->running_ios)) {
 589		/*
 590		 * FLUSH/FUA io_unit is deferred because of ordering, now we
 591		 * can dispatch it
 592		 */
 593		io_deferred = list_first_entry(&log->running_ios,
 594					       struct r5l_io_unit, log_sibling);
 595		if (io_deferred->io_deferred)
 596			schedule_work(&log->deferred_io_work);
 597	}
 598
 599	spin_unlock_irqrestore(&log->io_list_lock, flags);
 600
 601	if (log->need_cache_flush)
 602		md_wakeup_thread(log->rdev->mddev->thread);
 603
 604	/* finish flush only io_unit and PAYLOAD_FLUSH only io_unit */
 605	if (has_null_flush) {
 606		struct bio *bi;
 607
 608		WARN_ON(bio_list_empty(&io->flush_barriers));
 609		while ((bi = bio_list_pop(&io->flush_barriers)) != NULL) {
 610			bio_endio(bi);
 611			if (atomic_dec_and_test(&io->pending_stripe)) {
 612				__r5l_stripe_write_finished(io);
 613				return;
 614			}
 615		}
 616	}
 617	/* decrease pending_stripe for flush payload */
 618	if (has_flush_payload)
 619		if (atomic_dec_and_test(&io->pending_stripe))
 620			__r5l_stripe_write_finished(io);
 621}
 622
 623static void r5l_do_submit_io(struct r5l_log *log, struct r5l_io_unit *io)
 624{
 625	unsigned long flags;
 626
 627	spin_lock_irqsave(&log->io_list_lock, flags);
 628	__r5l_set_io_unit_state(io, IO_UNIT_IO_START);
 629	spin_unlock_irqrestore(&log->io_list_lock, flags);
 630
 631	/*
 632	 * In case of journal device failures, submit_bio will get error
 633	 * and calls endio, then active stripes will continue write
 634	 * process. Therefore, it is not necessary to check Faulty bit
 635	 * of journal device here.
 636	 *
 637	 * We can't check split_bio after current_bio is submitted. If
 638	 * io->split_bio is null, after current_bio is submitted, current_bio
 639	 * might already be completed and the io_unit is freed. We submit
 640	 * split_bio first to avoid the issue.
 641	 */
 642	if (io->split_bio) {
 643		if (io->has_flush)
 644			io->split_bio->bi_opf |= REQ_PREFLUSH;
 645		if (io->has_fua)
 646			io->split_bio->bi_opf |= REQ_FUA;
 647		submit_bio(io->split_bio);
 648	}
 649
 650	if (io->has_flush)
 651		io->current_bio->bi_opf |= REQ_PREFLUSH;
 652	if (io->has_fua)
 653		io->current_bio->bi_opf |= REQ_FUA;
 654	submit_bio(io->current_bio);
 655}
 656
 657/* deferred io_unit will be dispatched here */
 658static void r5l_submit_io_async(struct work_struct *work)
 659{
 660	struct r5l_log *log = container_of(work, struct r5l_log,
 661					   deferred_io_work);
 662	struct r5l_io_unit *io = NULL;
 663	unsigned long flags;
 664
 665	spin_lock_irqsave(&log->io_list_lock, flags);
 666	if (!list_empty(&log->running_ios)) {
 667		io = list_first_entry(&log->running_ios, struct r5l_io_unit,
 668				      log_sibling);
 669		if (!io->io_deferred)
 670			io = NULL;
 671		else
 672			io->io_deferred = 0;
 673	}
 674	spin_unlock_irqrestore(&log->io_list_lock, flags);
 675	if (io)
 676		r5l_do_submit_io(log, io);
 677}
 678
 679static void r5c_disable_writeback_async(struct work_struct *work)
 680{
 681	struct r5l_log *log = container_of(work, struct r5l_log,
 682					   disable_writeback_work);
 683	struct mddev *mddev = log->rdev->mddev;
 684	struct r5conf *conf = mddev->private;
 
 685
 686	if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH)
 687		return;
 688	pr_info("md/raid:%s: Disabling writeback cache for degraded array.\n",
 689		mdname(mddev));
 690
 691	/* wait superblock change before suspend */
 692	wait_event(mddev->sb_wait,
 693		   !READ_ONCE(conf->log) ||
 694		   !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags));
 695
 696	log = READ_ONCE(conf->log);
 697	if (log) {
 698		mddev_suspend(mddev, false);
 699		log->r5c_journal_mode = R5C_JOURNAL_MODE_WRITE_THROUGH;
 700		mddev_resume(mddev);
 
 701	}
 702}
 703
 704static void r5l_submit_current_io(struct r5l_log *log)
 705{
 706	struct r5l_io_unit *io = log->current_io;
 
 707	struct r5l_meta_block *block;
 708	unsigned long flags;
 709	u32 crc;
 710	bool do_submit = true;
 711
 712	if (!io)
 713		return;
 714
 715	block = page_address(io->meta_page);
 716	block->meta_size = cpu_to_le32(io->meta_offset);
 717	crc = crc32c_le(log->uuid_checksum, block, PAGE_SIZE);
 718	block->checksum = cpu_to_le32(crc);
 
 719
 720	log->current_io = NULL;
 721	spin_lock_irqsave(&log->io_list_lock, flags);
 722	if (io->has_flush || io->has_fua) {
 723		if (io != list_first_entry(&log->running_ios,
 724					   struct r5l_io_unit, log_sibling)) {
 725			io->io_deferred = 1;
 726			do_submit = false;
 727		}
 728	}
 729	spin_unlock_irqrestore(&log->io_list_lock, flags);
 730	if (do_submit)
 731		r5l_do_submit_io(log, io);
 732}
 733
 734static struct bio *r5l_bio_alloc(struct r5l_log *log)
 735{
 736	struct bio *bio = bio_alloc_bioset(log->rdev->bdev, BIO_MAX_VECS,
 737					   REQ_OP_WRITE, GFP_NOIO, &log->bs);
 738
 
 
 739	bio->bi_iter.bi_sector = log->rdev->data_offset + log->log_start;
 740
 741	return bio;
 742}
 743
 744static void r5_reserve_log_entry(struct r5l_log *log, struct r5l_io_unit *io)
 745{
 746	log->log_start = r5l_ring_add(log, log->log_start, BLOCK_SECTORS);
 747
 748	r5c_update_log_state(log);
 749	/*
 750	 * If we filled up the log device start from the beginning again,
 751	 * which will require a new bio.
 752	 *
 753	 * Note: for this to work properly the log size needs to me a multiple
 754	 * of BLOCK_SECTORS.
 755	 */
 756	if (log->log_start == 0)
 757		io->need_split_bio = true;
 758
 759	io->log_end = log->log_start;
 760}
 761
 762static struct r5l_io_unit *r5l_new_meta(struct r5l_log *log)
 763{
 764	struct r5l_io_unit *io;
 765	struct r5l_meta_block *block;
 766
 767	io = mempool_alloc(&log->io_pool, GFP_ATOMIC);
 768	if (!io)
 769		return NULL;
 770	memset(io, 0, sizeof(*io));
 771
 772	io->log = log;
 773	INIT_LIST_HEAD(&io->log_sibling);
 774	INIT_LIST_HEAD(&io->stripe_list);
 775	bio_list_init(&io->flush_barriers);
 776	io->state = IO_UNIT_RUNNING;
 777
 778	io->meta_page = mempool_alloc(&log->meta_pool, GFP_NOIO);
 779	block = page_address(io->meta_page);
 780	clear_page(block);
 781	block->magic = cpu_to_le32(R5LOG_MAGIC);
 782	block->version = R5LOG_VERSION;
 783	block->seq = cpu_to_le64(log->seq);
 784	block->position = cpu_to_le64(log->log_start);
 785
 786	io->log_start = log->log_start;
 787	io->meta_offset = sizeof(struct r5l_meta_block);
 788	io->seq = log->seq++;
 789
 790	io->current_bio = r5l_bio_alloc(log);
 791	io->current_bio->bi_end_io = r5l_log_endio;
 792	io->current_bio->bi_private = io;
 793	__bio_add_page(io->current_bio, io->meta_page, PAGE_SIZE, 0);
 794
 795	r5_reserve_log_entry(log, io);
 796
 797	spin_lock_irq(&log->io_list_lock);
 798	list_add_tail(&io->log_sibling, &log->running_ios);
 799	spin_unlock_irq(&log->io_list_lock);
 800
 801	return io;
 802}
 803
 804static int r5l_get_meta(struct r5l_log *log, unsigned int payload_size)
 805{
 806	if (log->current_io &&
 807	    log->current_io->meta_offset + payload_size > PAGE_SIZE)
 808		r5l_submit_current_io(log);
 809
 810	if (!log->current_io) {
 811		log->current_io = r5l_new_meta(log);
 812		if (!log->current_io)
 813			return -ENOMEM;
 814	}
 815
 816	return 0;
 817}
 818
 819static void r5l_append_payload_meta(struct r5l_log *log, u16 type,
 820				    sector_t location,
 821				    u32 checksum1, u32 checksum2,
 822				    bool checksum2_valid)
 823{
 824	struct r5l_io_unit *io = log->current_io;
 825	struct r5l_payload_data_parity *payload;
 826
 827	payload = page_address(io->meta_page) + io->meta_offset;
 828	payload->header.type = cpu_to_le16(type);
 829	payload->header.flags = cpu_to_le16(0);
 830	payload->size = cpu_to_le32((1 + !!checksum2_valid) <<
 831				    (PAGE_SHIFT - 9));
 832	payload->location = cpu_to_le64(location);
 833	payload->checksum[0] = cpu_to_le32(checksum1);
 834	if (checksum2_valid)
 835		payload->checksum[1] = cpu_to_le32(checksum2);
 836
 837	io->meta_offset += sizeof(struct r5l_payload_data_parity) +
 838		sizeof(__le32) * (1 + !!checksum2_valid);
 839}
 840
 841static void r5l_append_payload_page(struct r5l_log *log, struct page *page)
 842{
 843	struct r5l_io_unit *io = log->current_io;
 844
 845	if (io->need_split_bio) {
 846		BUG_ON(io->split_bio);
 847		io->split_bio = io->current_bio;
 848		io->current_bio = r5l_bio_alloc(log);
 849		bio_chain(io->current_bio, io->split_bio);
 850		io->need_split_bio = false;
 851	}
 852
 853	if (!bio_add_page(io->current_bio, page, PAGE_SIZE, 0))
 854		BUG();
 855
 856	r5_reserve_log_entry(log, io);
 857}
 858
 859static void r5l_append_flush_payload(struct r5l_log *log, sector_t sect)
 860{
 861	struct mddev *mddev = log->rdev->mddev;
 862	struct r5conf *conf = mddev->private;
 863	struct r5l_io_unit *io;
 864	struct r5l_payload_flush *payload;
 865	int meta_size;
 866
 867	/*
 868	 * payload_flush requires extra writes to the journal.
 869	 * To avoid handling the extra IO in quiesce, just skip
 870	 * flush_payload
 871	 */
 872	if (conf->quiesce)
 873		return;
 874
 875	mutex_lock(&log->io_mutex);
 876	meta_size = sizeof(struct r5l_payload_flush) + sizeof(__le64);
 877
 878	if (r5l_get_meta(log, meta_size)) {
 879		mutex_unlock(&log->io_mutex);
 880		return;
 881	}
 882
 883	/* current implementation is one stripe per flush payload */
 884	io = log->current_io;
 885	payload = page_address(io->meta_page) + io->meta_offset;
 886	payload->header.type = cpu_to_le16(R5LOG_PAYLOAD_FLUSH);
 887	payload->header.flags = cpu_to_le16(0);
 888	payload->size = cpu_to_le32(sizeof(__le64));
 889	payload->flush_stripes[0] = cpu_to_le64(sect);
 890	io->meta_offset += meta_size;
 891	/* multiple flush payloads count as one pending_stripe */
 892	if (!io->has_flush_payload) {
 893		io->has_flush_payload = 1;
 894		atomic_inc(&io->pending_stripe);
 895	}
 896	mutex_unlock(&log->io_mutex);
 897}
 898
 899static int r5l_log_stripe(struct r5l_log *log, struct stripe_head *sh,
 900			   int data_pages, int parity_pages)
 901{
 902	int i;
 903	int meta_size;
 904	int ret;
 905	struct r5l_io_unit *io;
 906
 907	meta_size =
 908		((sizeof(struct r5l_payload_data_parity) + sizeof(__le32))
 909		 * data_pages) +
 910		sizeof(struct r5l_payload_data_parity) +
 911		sizeof(__le32) * parity_pages;
 912
 913	ret = r5l_get_meta(log, meta_size);
 914	if (ret)
 915		return ret;
 916
 917	io = log->current_io;
 918
 919	if (test_and_clear_bit(STRIPE_R5C_PREFLUSH, &sh->state))
 920		io->has_flush = 1;
 921
 922	for (i = 0; i < sh->disks; i++) {
 923		if (!test_bit(R5_Wantwrite, &sh->dev[i].flags) ||
 924		    test_bit(R5_InJournal, &sh->dev[i].flags))
 925			continue;
 926		if (i == sh->pd_idx || i == sh->qd_idx)
 927			continue;
 928		if (test_bit(R5_WantFUA, &sh->dev[i].flags) &&
 929		    log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_BACK) {
 930			io->has_fua = 1;
 931			/*
 932			 * we need to flush journal to make sure recovery can
 933			 * reach the data with fua flag
 934			 */
 935			io->has_flush = 1;
 936		}
 937		r5l_append_payload_meta(log, R5LOG_PAYLOAD_DATA,
 938					raid5_compute_blocknr(sh, i, 0),
 939					sh->dev[i].log_checksum, 0, false);
 940		r5l_append_payload_page(log, sh->dev[i].page);
 941	}
 942
 943	if (parity_pages == 2) {
 944		r5l_append_payload_meta(log, R5LOG_PAYLOAD_PARITY,
 945					sh->sector, sh->dev[sh->pd_idx].log_checksum,
 946					sh->dev[sh->qd_idx].log_checksum, true);
 947		r5l_append_payload_page(log, sh->dev[sh->pd_idx].page);
 948		r5l_append_payload_page(log, sh->dev[sh->qd_idx].page);
 949	} else if (parity_pages == 1) {
 950		r5l_append_payload_meta(log, R5LOG_PAYLOAD_PARITY,
 951					sh->sector, sh->dev[sh->pd_idx].log_checksum,
 952					0, false);
 953		r5l_append_payload_page(log, sh->dev[sh->pd_idx].page);
 954	} else  /* Just writing data, not parity, in caching phase */
 955		BUG_ON(parity_pages != 0);
 956
 957	list_add_tail(&sh->log_list, &io->stripe_list);
 958	atomic_inc(&io->pending_stripe);
 959	sh->log_io = io;
 960
 961	if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH)
 962		return 0;
 963
 964	if (sh->log_start == MaxSector) {
 965		BUG_ON(!list_empty(&sh->r5c));
 966		sh->log_start = io->log_start;
 967		spin_lock_irq(&log->stripe_in_journal_lock);
 968		list_add_tail(&sh->r5c,
 969			      &log->stripe_in_journal_list);
 970		spin_unlock_irq(&log->stripe_in_journal_lock);
 971		atomic_inc(&log->stripe_in_journal_count);
 972	}
 973	return 0;
 974}
 975
 976/* add stripe to no_space_stripes, and then wake up reclaim */
 977static inline void r5l_add_no_space_stripe(struct r5l_log *log,
 978					   struct stripe_head *sh)
 979{
 980	spin_lock(&log->no_space_stripes_lock);
 981	list_add_tail(&sh->log_list, &log->no_space_stripes);
 982	spin_unlock(&log->no_space_stripes_lock);
 983}
 984
 985/*
 986 * running in raid5d, where reclaim could wait for raid5d too (when it flushes
 987 * data from log to raid disks), so we shouldn't wait for reclaim here
 988 */
 989int r5l_write_stripe(struct r5l_log *log, struct stripe_head *sh)
 990{
 991	struct r5conf *conf = sh->raid_conf;
 992	int write_disks = 0;
 993	int data_pages, parity_pages;
 994	int reserve;
 995	int i;
 996	int ret = 0;
 997	bool wake_reclaim = false;
 998
 999	if (!log)
1000		return -EAGAIN;
1001	/* Don't support stripe batch */
1002	if (sh->log_io || !test_bit(R5_Wantwrite, &sh->dev[sh->pd_idx].flags) ||
1003	    test_bit(STRIPE_SYNCING, &sh->state)) {
1004		/* the stripe is written to log, we start writing it to raid */
1005		clear_bit(STRIPE_LOG_TRAPPED, &sh->state);
1006		return -EAGAIN;
1007	}
1008
1009	WARN_ON(test_bit(STRIPE_R5C_CACHING, &sh->state));
1010
1011	for (i = 0; i < sh->disks; i++) {
1012		void *addr;
1013
1014		if (!test_bit(R5_Wantwrite, &sh->dev[i].flags) ||
1015		    test_bit(R5_InJournal, &sh->dev[i].flags))
1016			continue;
1017
1018		write_disks++;
1019		/* checksum is already calculated in last run */
1020		if (test_bit(STRIPE_LOG_TRAPPED, &sh->state))
1021			continue;
1022		addr = kmap_atomic(sh->dev[i].page);
1023		sh->dev[i].log_checksum = crc32c_le(log->uuid_checksum,
1024						    addr, PAGE_SIZE);
1025		kunmap_atomic(addr);
1026	}
1027	parity_pages = 1 + !!(sh->qd_idx >= 0);
1028	data_pages = write_disks - parity_pages;
1029
1030	set_bit(STRIPE_LOG_TRAPPED, &sh->state);
1031	/*
1032	 * The stripe must enter state machine again to finish the write, so
1033	 * don't delay.
1034	 */
1035	clear_bit(STRIPE_DELAYED, &sh->state);
1036	atomic_inc(&sh->count);
1037
1038	mutex_lock(&log->io_mutex);
1039	/* meta + data */
1040	reserve = (1 + write_disks) << (PAGE_SHIFT - 9);
1041
1042	if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) {
1043		if (!r5l_has_free_space(log, reserve)) {
1044			r5l_add_no_space_stripe(log, sh);
1045			wake_reclaim = true;
1046		} else {
1047			ret = r5l_log_stripe(log, sh, data_pages, parity_pages);
1048			if (ret) {
1049				spin_lock_irq(&log->io_list_lock);
1050				list_add_tail(&sh->log_list,
1051					      &log->no_mem_stripes);
1052				spin_unlock_irq(&log->io_list_lock);
1053			}
1054		}
1055	} else {  /* R5C_JOURNAL_MODE_WRITE_BACK */
1056		/*
1057		 * log space critical, do not process stripes that are
1058		 * not in cache yet (sh->log_start == MaxSector).
1059		 */
1060		if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state) &&
1061		    sh->log_start == MaxSector) {
1062			r5l_add_no_space_stripe(log, sh);
1063			wake_reclaim = true;
1064			reserve = 0;
1065		} else if (!r5l_has_free_space(log, reserve)) {
1066			if (sh->log_start == log->last_checkpoint)
1067				BUG();
1068			else
1069				r5l_add_no_space_stripe(log, sh);
1070		} else {
1071			ret = r5l_log_stripe(log, sh, data_pages, parity_pages);
1072			if (ret) {
1073				spin_lock_irq(&log->io_list_lock);
1074				list_add_tail(&sh->log_list,
1075					      &log->no_mem_stripes);
1076				spin_unlock_irq(&log->io_list_lock);
1077			}
1078		}
1079	}
1080
1081	mutex_unlock(&log->io_mutex);
1082	if (wake_reclaim)
1083		r5l_wake_reclaim(log, reserve);
1084	return 0;
1085}
1086
1087void r5l_write_stripe_run(struct r5l_log *log)
1088{
1089	if (!log)
1090		return;
1091	mutex_lock(&log->io_mutex);
1092	r5l_submit_current_io(log);
1093	mutex_unlock(&log->io_mutex);
1094}
1095
1096int r5l_handle_flush_request(struct r5l_log *log, struct bio *bio)
1097{
1098	if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) {
1099		/*
1100		 * in write through (journal only)
1101		 * we flush log disk cache first, then write stripe data to
1102		 * raid disks. So if bio is finished, the log disk cache is
1103		 * flushed already. The recovery guarantees we can recovery
1104		 * the bio from log disk, so we don't need to flush again
1105		 */
1106		if (bio->bi_iter.bi_size == 0) {
1107			bio_endio(bio);
1108			return 0;
1109		}
1110		bio->bi_opf &= ~REQ_PREFLUSH;
1111	} else {
1112		/* write back (with cache) */
1113		if (bio->bi_iter.bi_size == 0) {
1114			mutex_lock(&log->io_mutex);
1115			r5l_get_meta(log, 0);
1116			bio_list_add(&log->current_io->flush_barriers, bio);
1117			log->current_io->has_flush = 1;
1118			log->current_io->has_null_flush = 1;
1119			atomic_inc(&log->current_io->pending_stripe);
1120			r5l_submit_current_io(log);
1121			mutex_unlock(&log->io_mutex);
1122			return 0;
1123		}
1124	}
1125	return -EAGAIN;
1126}
1127
1128/* This will run after log space is reclaimed */
1129static void r5l_run_no_space_stripes(struct r5l_log *log)
1130{
1131	struct stripe_head *sh;
1132
1133	spin_lock(&log->no_space_stripes_lock);
1134	while (!list_empty(&log->no_space_stripes)) {
1135		sh = list_first_entry(&log->no_space_stripes,
1136				      struct stripe_head, log_list);
1137		list_del_init(&sh->log_list);
1138		set_bit(STRIPE_HANDLE, &sh->state);
1139		raid5_release_stripe(sh);
1140	}
1141	spin_unlock(&log->no_space_stripes_lock);
1142}
1143
1144/*
1145 * calculate new last_checkpoint
1146 * for write through mode, returns log->next_checkpoint
1147 * for write back, returns log_start of first sh in stripe_in_journal_list
1148 */
1149static sector_t r5c_calculate_new_cp(struct r5conf *conf)
1150{
1151	struct stripe_head *sh;
1152	struct r5l_log *log = READ_ONCE(conf->log);
1153	sector_t new_cp;
1154	unsigned long flags;
1155
1156	if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH)
1157		return log->next_checkpoint;
1158
1159	spin_lock_irqsave(&log->stripe_in_journal_lock, flags);
1160	if (list_empty(&log->stripe_in_journal_list)) {
1161		/* all stripes flushed */
1162		spin_unlock_irqrestore(&log->stripe_in_journal_lock, flags);
1163		return log->next_checkpoint;
1164	}
1165	sh = list_first_entry(&log->stripe_in_journal_list,
1166			      struct stripe_head, r5c);
1167	new_cp = sh->log_start;
1168	spin_unlock_irqrestore(&log->stripe_in_journal_lock, flags);
1169	return new_cp;
1170}
1171
1172static sector_t r5l_reclaimable_space(struct r5l_log *log)
1173{
1174	struct r5conf *conf = log->rdev->mddev->private;
1175
1176	return r5l_ring_distance(log, log->last_checkpoint,
1177				 r5c_calculate_new_cp(conf));
1178}
1179
1180static void r5l_run_no_mem_stripe(struct r5l_log *log)
1181{
1182	struct stripe_head *sh;
1183
1184	lockdep_assert_held(&log->io_list_lock);
1185
1186	if (!list_empty(&log->no_mem_stripes)) {
1187		sh = list_first_entry(&log->no_mem_stripes,
1188				      struct stripe_head, log_list);
1189		list_del_init(&sh->log_list);
1190		set_bit(STRIPE_HANDLE, &sh->state);
1191		raid5_release_stripe(sh);
1192	}
1193}
1194
1195static bool r5l_complete_finished_ios(struct r5l_log *log)
1196{
1197	struct r5l_io_unit *io, *next;
1198	bool found = false;
1199
1200	lockdep_assert_held(&log->io_list_lock);
1201
1202	list_for_each_entry_safe(io, next, &log->finished_ios, log_sibling) {
1203		/* don't change list order */
1204		if (io->state < IO_UNIT_STRIPE_END)
1205			break;
1206
1207		log->next_checkpoint = io->log_start;
1208
1209		list_del(&io->log_sibling);
1210		mempool_free(io, &log->io_pool);
1211		r5l_run_no_mem_stripe(log);
1212
1213		found = true;
1214	}
1215
1216	return found;
1217}
1218
1219static void __r5l_stripe_write_finished(struct r5l_io_unit *io)
1220{
1221	struct r5l_log *log = io->log;
1222	struct r5conf *conf = log->rdev->mddev->private;
1223	unsigned long flags;
1224
1225	spin_lock_irqsave(&log->io_list_lock, flags);
1226	__r5l_set_io_unit_state(io, IO_UNIT_STRIPE_END);
1227
1228	if (!r5l_complete_finished_ios(log)) {
1229		spin_unlock_irqrestore(&log->io_list_lock, flags);
1230		return;
1231	}
1232
1233	if (r5l_reclaimable_space(log) > log->max_free_space ||
1234	    test_bit(R5C_LOG_TIGHT, &conf->cache_state))
1235		r5l_wake_reclaim(log, 0);
1236
1237	spin_unlock_irqrestore(&log->io_list_lock, flags);
1238	wake_up(&log->iounit_wait);
1239}
1240
1241void r5l_stripe_write_finished(struct stripe_head *sh)
1242{
1243	struct r5l_io_unit *io;
1244
1245	io = sh->log_io;
1246	sh->log_io = NULL;
1247
1248	if (io && atomic_dec_and_test(&io->pending_stripe))
1249		__r5l_stripe_write_finished(io);
1250}
1251
1252static void r5l_log_flush_endio(struct bio *bio)
1253{
1254	struct r5l_log *log = container_of(bio, struct r5l_log,
1255		flush_bio);
1256	unsigned long flags;
1257	struct r5l_io_unit *io;
1258
1259	if (bio->bi_status)
1260		md_error(log->rdev->mddev, log->rdev);
1261	bio_uninit(bio);
1262
1263	spin_lock_irqsave(&log->io_list_lock, flags);
1264	list_for_each_entry(io, &log->flushing_ios, log_sibling)
1265		r5l_io_run_stripes(io);
1266	list_splice_tail_init(&log->flushing_ios, &log->finished_ios);
1267	spin_unlock_irqrestore(&log->io_list_lock, flags);
1268}
1269
1270/*
1271 * Starting dispatch IO to raid.
1272 * io_unit(meta) consists of a log. There is one situation we want to avoid. A
1273 * broken meta in the middle of a log causes recovery can't find meta at the
1274 * head of log. If operations require meta at the head persistent in log, we
1275 * must make sure meta before it persistent in log too. A case is:
1276 *
1277 * stripe data/parity is in log, we start write stripe to raid disks. stripe
1278 * data/parity must be persistent in log before we do the write to raid disks.
1279 *
1280 * The solution is we restrictly maintain io_unit list order. In this case, we
1281 * only write stripes of an io_unit to raid disks till the io_unit is the first
1282 * one whose data/parity is in log.
1283 */
1284void r5l_flush_stripe_to_raid(struct r5l_log *log)
1285{
1286	bool do_flush;
1287
1288	if (!log || !log->need_cache_flush)
1289		return;
1290
1291	spin_lock_irq(&log->io_list_lock);
1292	/* flush bio is running */
1293	if (!list_empty(&log->flushing_ios)) {
1294		spin_unlock_irq(&log->io_list_lock);
1295		return;
1296	}
1297	list_splice_tail_init(&log->io_end_ios, &log->flushing_ios);
1298	do_flush = !list_empty(&log->flushing_ios);
1299	spin_unlock_irq(&log->io_list_lock);
1300
1301	if (!do_flush)
1302		return;
1303	bio_init(&log->flush_bio, log->rdev->bdev, NULL, 0,
1304		  REQ_OP_WRITE | REQ_PREFLUSH);
1305	log->flush_bio.bi_end_io = r5l_log_flush_endio;
 
1306	submit_bio(&log->flush_bio);
1307}
1308
1309static void r5l_write_super(struct r5l_log *log, sector_t cp);
1310static void r5l_write_super_and_discard_space(struct r5l_log *log,
1311	sector_t end)
1312{
1313	struct block_device *bdev = log->rdev->bdev;
1314	struct mddev *mddev;
1315
1316	r5l_write_super(log, end);
1317
1318	if (!bdev_max_discard_sectors(bdev))
1319		return;
1320
1321	mddev = log->rdev->mddev;
1322	/*
1323	 * Discard could zero data, so before discard we must make sure
1324	 * superblock is updated to new log tail. Updating superblock (either
1325	 * directly call md_update_sb() or depend on md thread) must hold
1326	 * reconfig mutex. On the other hand, raid5_quiesce is called with
1327	 * reconfig_mutex hold. The first step of raid5_quiesce() is waiting
1328	 * for all IO finish, hence waiting for reclaim thread, while reclaim
1329	 * thread is calling this function and waiting for reconfig mutex. So
1330	 * there is a deadlock. We workaround this issue with a trylock.
1331	 * FIXME: we could miss discard if we can't take reconfig mutex
1332	 */
1333	set_mask_bits(&mddev->sb_flags, 0,
1334		BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1335	if (!mddev_trylock(mddev))
1336		return;
1337	md_update_sb(mddev, 1);
1338	mddev_unlock(mddev);
1339
1340	/* discard IO error really doesn't matter, ignore it */
1341	if (log->last_checkpoint < end) {
1342		blkdev_issue_discard(bdev,
1343				log->last_checkpoint + log->rdev->data_offset,
1344				end - log->last_checkpoint, GFP_NOIO);
1345	} else {
1346		blkdev_issue_discard(bdev,
1347				log->last_checkpoint + log->rdev->data_offset,
1348				log->device_size - log->last_checkpoint,
1349				GFP_NOIO);
1350		blkdev_issue_discard(bdev, log->rdev->data_offset, end,
1351				GFP_NOIO);
1352	}
1353}
1354
1355/*
1356 * r5c_flush_stripe moves stripe from cached list to handle_list. When called,
1357 * the stripe must be on r5c_cached_full_stripes or r5c_cached_partial_stripes.
1358 *
1359 * must hold conf->device_lock
1360 */
1361static void r5c_flush_stripe(struct r5conf *conf, struct stripe_head *sh)
1362{
1363	BUG_ON(list_empty(&sh->lru));
1364	BUG_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state));
1365	BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
1366
1367	/*
1368	 * The stripe is not ON_RELEASE_LIST, so it is safe to call
1369	 * raid5_release_stripe() while holding conf->device_lock
1370	 */
1371	BUG_ON(test_bit(STRIPE_ON_RELEASE_LIST, &sh->state));
1372	lockdep_assert_held(&conf->device_lock);
1373
1374	list_del_init(&sh->lru);
1375	atomic_inc(&sh->count);
1376
1377	set_bit(STRIPE_HANDLE, &sh->state);
1378	atomic_inc(&conf->active_stripes);
1379	r5c_make_stripe_write_out(sh);
1380
1381	if (test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state))
1382		atomic_inc(&conf->r5c_flushing_partial_stripes);
1383	else
1384		atomic_inc(&conf->r5c_flushing_full_stripes);
1385	raid5_release_stripe(sh);
1386}
1387
1388/*
1389 * if num == 0, flush all full stripes
1390 * if num > 0, flush all full stripes. If less than num full stripes are
1391 *             flushed, flush some partial stripes until totally num stripes are
1392 *             flushed or there is no more cached stripes.
1393 */
1394void r5c_flush_cache(struct r5conf *conf, int num)
1395{
1396	int count;
1397	struct stripe_head *sh, *next;
1398
1399	lockdep_assert_held(&conf->device_lock);
1400	if (!READ_ONCE(conf->log))
1401		return;
1402
1403	count = 0;
1404	list_for_each_entry_safe(sh, next, &conf->r5c_full_stripe_list, lru) {
1405		r5c_flush_stripe(conf, sh);
1406		count++;
1407	}
1408
1409	if (count >= num)
1410		return;
1411	list_for_each_entry_safe(sh, next,
1412				 &conf->r5c_partial_stripe_list, lru) {
1413		r5c_flush_stripe(conf, sh);
1414		if (++count >= num)
1415			break;
1416	}
1417}
1418
1419static void r5c_do_reclaim(struct r5conf *conf)
1420{
1421	struct r5l_log *log = READ_ONCE(conf->log);
1422	struct stripe_head *sh;
1423	int count = 0;
1424	unsigned long flags;
1425	int total_cached;
1426	int stripes_to_flush;
1427	int flushing_partial, flushing_full;
1428
1429	if (!r5c_is_writeback(log))
1430		return;
1431
1432	flushing_partial = atomic_read(&conf->r5c_flushing_partial_stripes);
1433	flushing_full = atomic_read(&conf->r5c_flushing_full_stripes);
1434	total_cached = atomic_read(&conf->r5c_cached_partial_stripes) +
1435		atomic_read(&conf->r5c_cached_full_stripes) -
1436		flushing_full - flushing_partial;
1437
1438	if (total_cached > conf->min_nr_stripes * 3 / 4 ||
1439	    atomic_read(&conf->empty_inactive_list_nr) > 0)
1440		/*
1441		 * if stripe cache pressure high, flush all full stripes and
1442		 * some partial stripes
1443		 */
1444		stripes_to_flush = R5C_RECLAIM_STRIPE_GROUP;
1445	else if (total_cached > conf->min_nr_stripes * 1 / 2 ||
1446		 atomic_read(&conf->r5c_cached_full_stripes) - flushing_full >
1447		 R5C_FULL_STRIPE_FLUSH_BATCH(conf))
1448		/*
1449		 * if stripe cache pressure moderate, or if there is many full
1450		 * stripes,flush all full stripes
1451		 */
1452		stripes_to_flush = 0;
1453	else
1454		/* no need to flush */
1455		stripes_to_flush = -1;
1456
1457	if (stripes_to_flush >= 0) {
1458		spin_lock_irqsave(&conf->device_lock, flags);
1459		r5c_flush_cache(conf, stripes_to_flush);
1460		spin_unlock_irqrestore(&conf->device_lock, flags);
1461	}
1462
1463	/* if log space is tight, flush stripes on stripe_in_journal_list */
1464	if (test_bit(R5C_LOG_TIGHT, &conf->cache_state)) {
1465		spin_lock_irqsave(&log->stripe_in_journal_lock, flags);
1466		spin_lock(&conf->device_lock);
1467		list_for_each_entry(sh, &log->stripe_in_journal_list, r5c) {
1468			/*
1469			 * stripes on stripe_in_journal_list could be in any
1470			 * state of the stripe_cache state machine. In this
1471			 * case, we only want to flush stripe on
1472			 * r5c_cached_full/partial_stripes. The following
1473			 * condition makes sure the stripe is on one of the
1474			 * two lists.
1475			 */
1476			if (!list_empty(&sh->lru) &&
1477			    !test_bit(STRIPE_HANDLE, &sh->state) &&
1478			    atomic_read(&sh->count) == 0) {
1479				r5c_flush_stripe(conf, sh);
1480				if (count++ >= R5C_RECLAIM_STRIPE_GROUP)
1481					break;
1482			}
1483		}
1484		spin_unlock(&conf->device_lock);
1485		spin_unlock_irqrestore(&log->stripe_in_journal_lock, flags);
1486	}
1487
1488	if (!test_bit(R5C_LOG_CRITICAL, &conf->cache_state))
1489		r5l_run_no_space_stripes(log);
1490
1491	md_wakeup_thread(conf->mddev->thread);
1492}
1493
1494static void r5l_do_reclaim(struct r5l_log *log)
1495{
1496	struct r5conf *conf = log->rdev->mddev->private;
1497	sector_t reclaim_target = xchg(&log->reclaim_target, 0);
1498	sector_t reclaimable;
1499	sector_t next_checkpoint;
1500	bool write_super;
1501
1502	spin_lock_irq(&log->io_list_lock);
1503	write_super = r5l_reclaimable_space(log) > log->max_free_space ||
1504		reclaim_target != 0 || !list_empty(&log->no_space_stripes);
1505	/*
1506	 * move proper io_unit to reclaim list. We should not change the order.
1507	 * reclaimable/unreclaimable io_unit can be mixed in the list, we
1508	 * shouldn't reuse space of an unreclaimable io_unit
1509	 */
1510	while (1) {
1511		reclaimable = r5l_reclaimable_space(log);
1512		if (reclaimable >= reclaim_target ||
1513		    (list_empty(&log->running_ios) &&
1514		     list_empty(&log->io_end_ios) &&
1515		     list_empty(&log->flushing_ios) &&
1516		     list_empty(&log->finished_ios)))
1517			break;
1518
1519		md_wakeup_thread(log->rdev->mddev->thread);
1520		wait_event_lock_irq(log->iounit_wait,
1521				    r5l_reclaimable_space(log) > reclaimable,
1522				    log->io_list_lock);
1523	}
1524
1525	next_checkpoint = r5c_calculate_new_cp(conf);
1526	spin_unlock_irq(&log->io_list_lock);
1527
1528	if (reclaimable == 0 || !write_super)
1529		return;
1530
1531	/*
1532	 * write_super will flush cache of each raid disk. We must write super
1533	 * here, because the log area might be reused soon and we don't want to
1534	 * confuse recovery
1535	 */
1536	r5l_write_super_and_discard_space(log, next_checkpoint);
1537
1538	mutex_lock(&log->io_mutex);
1539	log->last_checkpoint = next_checkpoint;
1540	r5c_update_log_state(log);
1541	mutex_unlock(&log->io_mutex);
1542
1543	r5l_run_no_space_stripes(log);
1544}
1545
1546static void r5l_reclaim_thread(struct md_thread *thread)
1547{
1548	struct mddev *mddev = thread->mddev;
1549	struct r5conf *conf = mddev->private;
1550	struct r5l_log *log = READ_ONCE(conf->log);
1551
1552	if (!log)
1553		return;
1554	r5c_do_reclaim(conf);
1555	r5l_do_reclaim(log);
1556}
1557
1558void r5l_wake_reclaim(struct r5l_log *log, sector_t space)
1559{
1560	unsigned long target;
1561	unsigned long new = (unsigned long)space; /* overflow in theory */
1562
1563	if (!log)
1564		return;
1565
1566	target = READ_ONCE(log->reclaim_target);
1567	do {
 
1568		if (new < target)
1569			return;
1570	} while (!try_cmpxchg(&log->reclaim_target, &target, new));
1571	md_wakeup_thread(log->reclaim_thread);
1572}
1573
1574void r5l_quiesce(struct r5l_log *log, int quiesce)
1575{
1576	struct mddev *mddev = log->rdev->mddev;
1577	struct md_thread *thread = rcu_dereference_protected(
1578		log->reclaim_thread, lockdep_is_held(&mddev->reconfig_mutex));
1579
1580	if (quiesce) {
1581		/* make sure r5l_write_super_and_discard_space exits */
 
1582		wake_up(&mddev->sb_wait);
1583		kthread_park(thread->tsk);
1584		r5l_wake_reclaim(log, MaxSector);
1585		r5l_do_reclaim(log);
1586	} else
1587		kthread_unpark(thread->tsk);
1588}
1589
1590bool r5l_log_disk_error(struct r5conf *conf)
1591{
1592	struct r5l_log *log = READ_ONCE(conf->log);
1593
1594	/* don't allow write if journal disk is missing */
 
 
 
1595	if (!log)
1596		return test_bit(MD_HAS_JOURNAL, &conf->mddev->flags);
1597	else
1598		return test_bit(Faulty, &log->rdev->flags);
 
 
1599}
1600
1601#define R5L_RECOVERY_PAGE_POOL_SIZE 256
1602
1603struct r5l_recovery_ctx {
1604	struct page *meta_page;		/* current meta */
1605	sector_t meta_total_blocks;	/* total size of current meta and data */
1606	sector_t pos;			/* recovery position */
1607	u64 seq;			/* recovery position seq */
1608	int data_parity_stripes;	/* number of data_parity stripes */
1609	int data_only_stripes;		/* number of data_only stripes */
1610	struct list_head cached_list;
1611
1612	/*
1613	 * read ahead page pool (ra_pool)
1614	 * in recovery, log is read sequentially. It is not efficient to
1615	 * read every page with sync_page_io(). The read ahead page pool
1616	 * reads multiple pages with one IO, so further log read can
1617	 * just copy data from the pool.
1618	 */
1619	struct page *ra_pool[R5L_RECOVERY_PAGE_POOL_SIZE];
1620	struct bio_vec ra_bvec[R5L_RECOVERY_PAGE_POOL_SIZE];
1621	sector_t pool_offset;	/* offset of first page in the pool */
1622	int total_pages;	/* total allocated pages */
1623	int valid_pages;	/* pages with valid data */
 
1624};
1625
1626static int r5l_recovery_allocate_ra_pool(struct r5l_log *log,
1627					    struct r5l_recovery_ctx *ctx)
1628{
1629	struct page *page;
1630
 
 
 
 
1631	ctx->valid_pages = 0;
1632	ctx->total_pages = 0;
1633	while (ctx->total_pages < R5L_RECOVERY_PAGE_POOL_SIZE) {
1634		page = alloc_page(GFP_KERNEL);
1635
1636		if (!page)
1637			break;
1638		ctx->ra_pool[ctx->total_pages] = page;
1639		ctx->total_pages += 1;
1640	}
1641
1642	if (ctx->total_pages == 0)
 
1643		return -ENOMEM;
 
1644
1645	ctx->pool_offset = 0;
1646	return 0;
1647}
1648
1649static void r5l_recovery_free_ra_pool(struct r5l_log *log,
1650					struct r5l_recovery_ctx *ctx)
1651{
1652	int i;
1653
1654	for (i = 0; i < ctx->total_pages; ++i)
1655		put_page(ctx->ra_pool[i]);
 
1656}
1657
1658/*
1659 * fetch ctx->valid_pages pages from offset
1660 * In normal cases, ctx->valid_pages == ctx->total_pages after the call.
1661 * However, if the offset is close to the end of the journal device,
1662 * ctx->valid_pages could be smaller than ctx->total_pages
1663 */
1664static int r5l_recovery_fetch_ra_pool(struct r5l_log *log,
1665				      struct r5l_recovery_ctx *ctx,
1666				      sector_t offset)
1667{
1668	struct bio bio;
1669	int ret;
1670
1671	bio_init(&bio, log->rdev->bdev, ctx->ra_bvec,
1672		 R5L_RECOVERY_PAGE_POOL_SIZE, REQ_OP_READ);
1673	bio.bi_iter.bi_sector = log->rdev->data_offset + offset;
1674
1675	ctx->valid_pages = 0;
1676	ctx->pool_offset = offset;
1677
1678	while (ctx->valid_pages < ctx->total_pages) {
1679		__bio_add_page(&bio, ctx->ra_pool[ctx->valid_pages], PAGE_SIZE,
1680			       0);
1681		ctx->valid_pages += 1;
1682
1683		offset = r5l_ring_add(log, offset, BLOCK_SECTORS);
1684
1685		if (offset == 0)  /* reached end of the device */
1686			break;
1687	}
1688
1689	ret = submit_bio_wait(&bio);
1690	bio_uninit(&bio);
1691	return ret;
1692}
1693
1694/*
1695 * try read a page from the read ahead page pool, if the page is not in the
1696 * pool, call r5l_recovery_fetch_ra_pool
1697 */
1698static int r5l_recovery_read_page(struct r5l_log *log,
1699				  struct r5l_recovery_ctx *ctx,
1700				  struct page *page,
1701				  sector_t offset)
1702{
1703	int ret;
1704
1705	if (offset < ctx->pool_offset ||
1706	    offset >= ctx->pool_offset + ctx->valid_pages * BLOCK_SECTORS) {
1707		ret = r5l_recovery_fetch_ra_pool(log, ctx, offset);
1708		if (ret)
1709			return ret;
1710	}
1711
1712	BUG_ON(offset < ctx->pool_offset ||
1713	       offset >= ctx->pool_offset + ctx->valid_pages * BLOCK_SECTORS);
1714
1715	memcpy(page_address(page),
1716	       page_address(ctx->ra_pool[(offset - ctx->pool_offset) >>
1717					 BLOCK_SECTOR_SHIFT]),
1718	       PAGE_SIZE);
1719	return 0;
1720}
1721
1722static int r5l_recovery_read_meta_block(struct r5l_log *log,
1723					struct r5l_recovery_ctx *ctx)
1724{
1725	struct page *page = ctx->meta_page;
1726	struct r5l_meta_block *mb;
1727	u32 crc, stored_crc;
1728	int ret;
1729
1730	ret = r5l_recovery_read_page(log, ctx, page, ctx->pos);
1731	if (ret != 0)
1732		return ret;
1733
1734	mb = page_address(page);
1735	stored_crc = le32_to_cpu(mb->checksum);
1736	mb->checksum = 0;
1737
1738	if (le32_to_cpu(mb->magic) != R5LOG_MAGIC ||
1739	    le64_to_cpu(mb->seq) != ctx->seq ||
1740	    mb->version != R5LOG_VERSION ||
1741	    le64_to_cpu(mb->position) != ctx->pos)
1742		return -EINVAL;
1743
1744	crc = crc32c_le(log->uuid_checksum, mb, PAGE_SIZE);
1745	if (stored_crc != crc)
1746		return -EINVAL;
1747
1748	if (le32_to_cpu(mb->meta_size) > PAGE_SIZE)
1749		return -EINVAL;
1750
1751	ctx->meta_total_blocks = BLOCK_SECTORS;
1752
1753	return 0;
1754}
1755
1756static void
1757r5l_recovery_create_empty_meta_block(struct r5l_log *log,
1758				     struct page *page,
1759				     sector_t pos, u64 seq)
1760{
1761	struct r5l_meta_block *mb;
1762
1763	mb = page_address(page);
1764	clear_page(mb);
1765	mb->magic = cpu_to_le32(R5LOG_MAGIC);
1766	mb->version = R5LOG_VERSION;
1767	mb->meta_size = cpu_to_le32(sizeof(struct r5l_meta_block));
1768	mb->seq = cpu_to_le64(seq);
1769	mb->position = cpu_to_le64(pos);
1770}
1771
1772static int r5l_log_write_empty_meta_block(struct r5l_log *log, sector_t pos,
1773					  u64 seq)
1774{
1775	struct page *page;
1776	struct r5l_meta_block *mb;
1777
1778	page = alloc_page(GFP_KERNEL);
1779	if (!page)
1780		return -ENOMEM;
1781	r5l_recovery_create_empty_meta_block(log, page, pos, seq);
1782	mb = page_address(page);
1783	mb->checksum = cpu_to_le32(crc32c_le(log->uuid_checksum,
1784					     mb, PAGE_SIZE));
1785	if (!sync_page_io(log->rdev, pos, PAGE_SIZE, page, REQ_OP_WRITE |
1786			  REQ_SYNC | REQ_FUA, false)) {
1787		__free_page(page);
1788		return -EIO;
1789	}
1790	__free_page(page);
1791	return 0;
1792}
1793
1794/*
1795 * r5l_recovery_load_data and r5l_recovery_load_parity uses flag R5_Wantwrite
1796 * to mark valid (potentially not flushed) data in the journal.
1797 *
1798 * We already verified checksum in r5l_recovery_verify_data_checksum_for_mb,
1799 * so there should not be any mismatch here.
1800 */
1801static void r5l_recovery_load_data(struct r5l_log *log,
1802				   struct stripe_head *sh,
1803				   struct r5l_recovery_ctx *ctx,
1804				   struct r5l_payload_data_parity *payload,
1805				   sector_t log_offset)
1806{
1807	struct mddev *mddev = log->rdev->mddev;
1808	struct r5conf *conf = mddev->private;
1809	int dd_idx;
1810
1811	raid5_compute_sector(conf,
1812			     le64_to_cpu(payload->location), 0,
1813			     &dd_idx, sh);
1814	r5l_recovery_read_page(log, ctx, sh->dev[dd_idx].page, log_offset);
1815	sh->dev[dd_idx].log_checksum =
1816		le32_to_cpu(payload->checksum[0]);
1817	ctx->meta_total_blocks += BLOCK_SECTORS;
1818
1819	set_bit(R5_Wantwrite, &sh->dev[dd_idx].flags);
1820	set_bit(STRIPE_R5C_CACHING, &sh->state);
1821}
1822
1823static void r5l_recovery_load_parity(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
1832	ctx->meta_total_blocks += BLOCK_SECTORS * conf->max_degraded;
1833	r5l_recovery_read_page(log, ctx, sh->dev[sh->pd_idx].page, log_offset);
1834	sh->dev[sh->pd_idx].log_checksum =
1835		le32_to_cpu(payload->checksum[0]);
1836	set_bit(R5_Wantwrite, &sh->dev[sh->pd_idx].flags);
1837
1838	if (sh->qd_idx >= 0) {
1839		r5l_recovery_read_page(
1840			log, ctx, sh->dev[sh->qd_idx].page,
1841			r5l_ring_add(log, log_offset, BLOCK_SECTORS));
1842		sh->dev[sh->qd_idx].log_checksum =
1843			le32_to_cpu(payload->checksum[1]);
1844		set_bit(R5_Wantwrite, &sh->dev[sh->qd_idx].flags);
1845	}
1846	clear_bit(STRIPE_R5C_CACHING, &sh->state);
1847}
1848
1849static void r5l_recovery_reset_stripe(struct stripe_head *sh)
1850{
1851	int i;
1852
1853	sh->state = 0;
1854	sh->log_start = MaxSector;
1855	for (i = sh->disks; i--; )
1856		sh->dev[i].flags = 0;
1857}
1858
1859static void
1860r5l_recovery_replay_one_stripe(struct r5conf *conf,
1861			       struct stripe_head *sh,
1862			       struct r5l_recovery_ctx *ctx)
1863{
1864	struct md_rdev *rdev, *rrdev;
1865	int disk_index;
1866	int data_count = 0;
1867
1868	for (disk_index = 0; disk_index < sh->disks; disk_index++) {
1869		if (!test_bit(R5_Wantwrite, &sh->dev[disk_index].flags))
1870			continue;
1871		if (disk_index == sh->qd_idx || disk_index == sh->pd_idx)
1872			continue;
1873		data_count++;
1874	}
1875
1876	/*
1877	 * stripes that only have parity must have been flushed
1878	 * before the crash that we are now recovering from, so
1879	 * there is nothing more to recovery.
1880	 */
1881	if (data_count == 0)
1882		goto out;
1883
1884	for (disk_index = 0; disk_index < sh->disks; disk_index++) {
1885		if (!test_bit(R5_Wantwrite, &sh->dev[disk_index].flags))
1886			continue;
1887
1888		/* in case device is broken */
1889		rdev = conf->disks[disk_index].rdev;
 
1890		if (rdev) {
1891			atomic_inc(&rdev->nr_pending);
 
1892			sync_page_io(rdev, sh->sector, PAGE_SIZE,
1893				     sh->dev[disk_index].page, REQ_OP_WRITE,
1894				     false);
1895			rdev_dec_pending(rdev, rdev->mddev);
 
1896		}
1897		rrdev = conf->disks[disk_index].replacement;
1898		if (rrdev) {
1899			atomic_inc(&rrdev->nr_pending);
 
1900			sync_page_io(rrdev, sh->sector, PAGE_SIZE,
1901				     sh->dev[disk_index].page, REQ_OP_WRITE,
1902				     false);
1903			rdev_dec_pending(rrdev, rrdev->mddev);
 
1904		}
 
1905	}
1906	ctx->data_parity_stripes++;
1907out:
1908	r5l_recovery_reset_stripe(sh);
1909}
1910
1911static struct stripe_head *
1912r5c_recovery_alloc_stripe(
1913		struct r5conf *conf,
1914		sector_t stripe_sect,
1915		int noblock)
1916{
1917	struct stripe_head *sh;
1918
1919	sh = raid5_get_active_stripe(conf, NULL, stripe_sect,
1920				     noblock ? R5_GAS_NOBLOCK : 0);
1921	if (!sh)
1922		return NULL;  /* no more stripe available */
1923
1924	r5l_recovery_reset_stripe(sh);
1925
1926	return sh;
1927}
1928
1929static struct stripe_head *
1930r5c_recovery_lookup_stripe(struct list_head *list, sector_t sect)
1931{
1932	struct stripe_head *sh;
1933
1934	list_for_each_entry(sh, list, lru)
1935		if (sh->sector == sect)
1936			return sh;
1937	return NULL;
1938}
1939
1940static void
1941r5c_recovery_drop_stripes(struct list_head *cached_stripe_list,
1942			  struct r5l_recovery_ctx *ctx)
1943{
1944	struct stripe_head *sh, *next;
1945
1946	list_for_each_entry_safe(sh, next, cached_stripe_list, lru) {
1947		r5l_recovery_reset_stripe(sh);
1948		list_del_init(&sh->lru);
1949		raid5_release_stripe(sh);
1950	}
1951}
1952
1953static void
1954r5c_recovery_replay_stripes(struct list_head *cached_stripe_list,
1955			    struct r5l_recovery_ctx *ctx)
1956{
1957	struct stripe_head *sh, *next;
1958
1959	list_for_each_entry_safe(sh, next, cached_stripe_list, lru)
1960		if (!test_bit(STRIPE_R5C_CACHING, &sh->state)) {
1961			r5l_recovery_replay_one_stripe(sh->raid_conf, sh, ctx);
1962			list_del_init(&sh->lru);
1963			raid5_release_stripe(sh);
1964		}
1965}
1966
1967/* if matches return 0; otherwise return -EINVAL */
1968static int
1969r5l_recovery_verify_data_checksum(struct r5l_log *log,
1970				  struct r5l_recovery_ctx *ctx,
1971				  struct page *page,
1972				  sector_t log_offset, __le32 log_checksum)
1973{
1974	void *addr;
1975	u32 checksum;
1976
1977	r5l_recovery_read_page(log, ctx, page, log_offset);
1978	addr = kmap_atomic(page);
1979	checksum = crc32c_le(log->uuid_checksum, addr, PAGE_SIZE);
1980	kunmap_atomic(addr);
1981	return (le32_to_cpu(log_checksum) == checksum) ? 0 : -EINVAL;
1982}
1983
1984/*
1985 * before loading data to stripe cache, we need verify checksum for all data,
1986 * if there is mismatch for any data page, we drop all data in the mata block
1987 */
1988static int
1989r5l_recovery_verify_data_checksum_for_mb(struct r5l_log *log,
1990					 struct r5l_recovery_ctx *ctx)
1991{
1992	struct mddev *mddev = log->rdev->mddev;
1993	struct r5conf *conf = mddev->private;
1994	struct r5l_meta_block *mb = page_address(ctx->meta_page);
1995	sector_t mb_offset = sizeof(struct r5l_meta_block);
1996	sector_t log_offset = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS);
1997	struct page *page;
1998	struct r5l_payload_data_parity *payload;
1999	struct r5l_payload_flush *payload_flush;
2000
2001	page = alloc_page(GFP_KERNEL);
2002	if (!page)
2003		return -ENOMEM;
2004
2005	while (mb_offset < le32_to_cpu(mb->meta_size)) {
2006		payload = (void *)mb + mb_offset;
2007		payload_flush = (void *)mb + mb_offset;
2008
2009		if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_DATA) {
2010			if (r5l_recovery_verify_data_checksum(
2011				    log, ctx, page, log_offset,
2012				    payload->checksum[0]) < 0)
2013				goto mismatch;
2014		} else if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_PARITY) {
2015			if (r5l_recovery_verify_data_checksum(
2016				    log, ctx, page, log_offset,
2017				    payload->checksum[0]) < 0)
2018				goto mismatch;
2019			if (conf->max_degraded == 2 && /* q for RAID 6 */
2020			    r5l_recovery_verify_data_checksum(
2021				    log, ctx, page,
2022				    r5l_ring_add(log, log_offset,
2023						 BLOCK_SECTORS),
2024				    payload->checksum[1]) < 0)
2025				goto mismatch;
2026		} else if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_FLUSH) {
2027			/* nothing to do for R5LOG_PAYLOAD_FLUSH here */
2028		} else /* not R5LOG_PAYLOAD_DATA/PARITY/FLUSH */
2029			goto mismatch;
2030
2031		if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_FLUSH) {
2032			mb_offset += sizeof(struct r5l_payload_flush) +
2033				le32_to_cpu(payload_flush->size);
2034		} else {
2035			/* DATA or PARITY payload */
2036			log_offset = r5l_ring_add(log, log_offset,
2037						  le32_to_cpu(payload->size));
2038			mb_offset += sizeof(struct r5l_payload_data_parity) +
2039				sizeof(__le32) *
2040				(le32_to_cpu(payload->size) >> (PAGE_SHIFT - 9));
2041		}
2042
2043	}
2044
2045	put_page(page);
2046	return 0;
2047
2048mismatch:
2049	put_page(page);
2050	return -EINVAL;
2051}
2052
2053/*
2054 * Analyze all data/parity pages in one meta block
2055 * Returns:
2056 * 0 for success
2057 * -EINVAL for unknown playload type
2058 * -EAGAIN for checksum mismatch of data page
2059 * -ENOMEM for run out of memory (alloc_page failed or run out of stripes)
2060 */
2061static int
2062r5c_recovery_analyze_meta_block(struct r5l_log *log,
2063				struct r5l_recovery_ctx *ctx,
2064				struct list_head *cached_stripe_list)
2065{
2066	struct mddev *mddev = log->rdev->mddev;
2067	struct r5conf *conf = mddev->private;
2068	struct r5l_meta_block *mb;
2069	struct r5l_payload_data_parity *payload;
2070	struct r5l_payload_flush *payload_flush;
2071	int mb_offset;
2072	sector_t log_offset;
2073	sector_t stripe_sect;
2074	struct stripe_head *sh;
2075	int ret;
2076
2077	/*
2078	 * for mismatch in data blocks, we will drop all data in this mb, but
2079	 * we will still read next mb for other data with FLUSH flag, as
2080	 * io_unit could finish out of order.
2081	 */
2082	ret = r5l_recovery_verify_data_checksum_for_mb(log, ctx);
2083	if (ret == -EINVAL)
2084		return -EAGAIN;
2085	else if (ret)
2086		return ret;   /* -ENOMEM duo to alloc_page() failed */
2087
2088	mb = page_address(ctx->meta_page);
2089	mb_offset = sizeof(struct r5l_meta_block);
2090	log_offset = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS);
2091
2092	while (mb_offset < le32_to_cpu(mb->meta_size)) {
2093		int dd;
2094
2095		payload = (void *)mb + mb_offset;
2096		payload_flush = (void *)mb + mb_offset;
2097
2098		if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_FLUSH) {
2099			int i, count;
2100
2101			count = le32_to_cpu(payload_flush->size) / sizeof(__le64);
2102			for (i = 0; i < count; ++i) {
2103				stripe_sect = le64_to_cpu(payload_flush->flush_stripes[i]);
2104				sh = r5c_recovery_lookup_stripe(cached_stripe_list,
2105								stripe_sect);
2106				if (sh) {
2107					WARN_ON(test_bit(STRIPE_R5C_CACHING, &sh->state));
2108					r5l_recovery_reset_stripe(sh);
2109					list_del_init(&sh->lru);
2110					raid5_release_stripe(sh);
2111				}
2112			}
2113
2114			mb_offset += sizeof(struct r5l_payload_flush) +
2115				le32_to_cpu(payload_flush->size);
2116			continue;
2117		}
2118
2119		/* DATA or PARITY payload */
2120		stripe_sect = (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_DATA) ?
2121			raid5_compute_sector(
2122				conf, le64_to_cpu(payload->location), 0, &dd,
2123				NULL)
2124			: le64_to_cpu(payload->location);
2125
2126		sh = r5c_recovery_lookup_stripe(cached_stripe_list,
2127						stripe_sect);
2128
2129		if (!sh) {
2130			sh = r5c_recovery_alloc_stripe(conf, stripe_sect, 1);
2131			/*
2132			 * cannot get stripe from raid5_get_active_stripe
2133			 * try replay some stripes
2134			 */
2135			if (!sh) {
2136				r5c_recovery_replay_stripes(
2137					cached_stripe_list, ctx);
2138				sh = r5c_recovery_alloc_stripe(
2139					conf, stripe_sect, 1);
2140			}
2141			if (!sh) {
2142				int new_size = conf->min_nr_stripes * 2;
2143				pr_debug("md/raid:%s: Increasing stripe cache size to %d to recovery data on journal.\n",
2144					mdname(mddev),
2145					new_size);
2146				ret = raid5_set_cache_size(mddev, new_size);
2147				if (conf->min_nr_stripes <= new_size / 2) {
2148					pr_err("md/raid:%s: Cannot increase cache size, ret=%d, new_size=%d, min_nr_stripes=%d, max_nr_stripes=%d\n",
2149						mdname(mddev),
2150						ret,
2151						new_size,
2152						conf->min_nr_stripes,
2153						conf->max_nr_stripes);
2154					return -ENOMEM;
2155				}
2156				sh = r5c_recovery_alloc_stripe(
2157					conf, stripe_sect, 0);
2158			}
2159			if (!sh) {
2160				pr_err("md/raid:%s: Cannot get enough stripes due to memory pressure. Recovery failed.\n",
2161					mdname(mddev));
2162				return -ENOMEM;
2163			}
2164			list_add_tail(&sh->lru, cached_stripe_list);
2165		}
2166
2167		if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_DATA) {
2168			if (!test_bit(STRIPE_R5C_CACHING, &sh->state) &&
2169			    test_bit(R5_Wantwrite, &sh->dev[sh->pd_idx].flags)) {
2170				r5l_recovery_replay_one_stripe(conf, sh, ctx);
2171				list_move_tail(&sh->lru, cached_stripe_list);
2172			}
2173			r5l_recovery_load_data(log, sh, ctx, payload,
2174					       log_offset);
2175		} else if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_PARITY)
2176			r5l_recovery_load_parity(log, sh, ctx, payload,
2177						 log_offset);
2178		else
2179			return -EINVAL;
2180
2181		log_offset = r5l_ring_add(log, log_offset,
2182					  le32_to_cpu(payload->size));
2183
2184		mb_offset += sizeof(struct r5l_payload_data_parity) +
2185			sizeof(__le32) *
2186			(le32_to_cpu(payload->size) >> (PAGE_SHIFT - 9));
2187	}
2188
2189	return 0;
2190}
2191
2192/*
2193 * Load the stripe into cache. The stripe will be written out later by
2194 * the stripe cache state machine.
2195 */
2196static void r5c_recovery_load_one_stripe(struct r5l_log *log,
2197					 struct stripe_head *sh)
2198{
2199	struct r5dev *dev;
2200	int i;
2201
2202	for (i = sh->disks; i--; ) {
2203		dev = sh->dev + i;
2204		if (test_and_clear_bit(R5_Wantwrite, &dev->flags)) {
2205			set_bit(R5_InJournal, &dev->flags);
2206			set_bit(R5_UPTODATE, &dev->flags);
2207		}
2208	}
2209}
2210
2211/*
2212 * Scan through the log for all to-be-flushed data
2213 *
2214 * For stripes with data and parity, namely Data-Parity stripe
2215 * (STRIPE_R5C_CACHING == 0), we simply replay all the writes.
2216 *
2217 * For stripes with only data, namely Data-Only stripe
2218 * (STRIPE_R5C_CACHING == 1), we load them to stripe cache state machine.
2219 *
2220 * For a stripe, if we see data after parity, we should discard all previous
2221 * data and parity for this stripe, as these data are already flushed to
2222 * the array.
2223 *
2224 * At the end of the scan, we return the new journal_tail, which points to
2225 * first data-only stripe on the journal device, or next invalid meta block.
2226 */
2227static int r5c_recovery_flush_log(struct r5l_log *log,
2228				  struct r5l_recovery_ctx *ctx)
2229{
2230	struct stripe_head *sh;
2231	int ret = 0;
2232
2233	/* scan through the log */
2234	while (1) {
2235		if (r5l_recovery_read_meta_block(log, ctx))
2236			break;
2237
2238		ret = r5c_recovery_analyze_meta_block(log, ctx,
2239						      &ctx->cached_list);
2240		/*
2241		 * -EAGAIN means mismatch in data block, in this case, we still
2242		 * try scan the next metablock
2243		 */
2244		if (ret && ret != -EAGAIN)
2245			break;   /* ret == -EINVAL or -ENOMEM */
2246		ctx->seq++;
2247		ctx->pos = r5l_ring_add(log, ctx->pos, ctx->meta_total_blocks);
2248	}
2249
2250	if (ret == -ENOMEM) {
2251		r5c_recovery_drop_stripes(&ctx->cached_list, ctx);
2252		return ret;
2253	}
2254
2255	/* replay data-parity stripes */
2256	r5c_recovery_replay_stripes(&ctx->cached_list, ctx);
2257
2258	/* load data-only stripes to stripe cache */
2259	list_for_each_entry(sh, &ctx->cached_list, lru) {
2260		WARN_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state));
2261		r5c_recovery_load_one_stripe(log, sh);
2262		ctx->data_only_stripes++;
2263	}
2264
2265	return 0;
2266}
2267
2268/*
2269 * we did a recovery. Now ctx.pos points to an invalid meta block. New
2270 * log will start here. but we can't let superblock point to last valid
2271 * meta block. The log might looks like:
2272 * | meta 1| meta 2| meta 3|
2273 * meta 1 is valid, meta 2 is invalid. meta 3 could be valid. If
2274 * superblock points to meta 1, we write a new valid meta 2n.  if crash
2275 * happens again, new recovery will start from meta 1. Since meta 2n is
2276 * valid now, recovery will think meta 3 is valid, which is wrong.
2277 * The solution is we create a new meta in meta2 with its seq == meta
2278 * 1's seq + 10000 and let superblock points to meta2. The same recovery
2279 * will not think meta 3 is a valid meta, because its seq doesn't match
2280 */
2281
2282/*
2283 * Before recovery, the log looks like the following
2284 *
2285 *   ---------------------------------------------
2286 *   |           valid log        | invalid log  |
2287 *   ---------------------------------------------
2288 *   ^
2289 *   |- log->last_checkpoint
2290 *   |- log->last_cp_seq
2291 *
2292 * Now we scan through the log until we see invalid entry
2293 *
2294 *   ---------------------------------------------
2295 *   |           valid log        | invalid log  |
2296 *   ---------------------------------------------
2297 *   ^                            ^
2298 *   |- log->last_checkpoint      |- ctx->pos
2299 *   |- log->last_cp_seq          |- ctx->seq
2300 *
2301 * From this point, we need to increase seq number by 10 to avoid
2302 * confusing next recovery.
2303 *
2304 *   ---------------------------------------------
2305 *   |           valid log        | invalid log  |
2306 *   ---------------------------------------------
2307 *   ^                              ^
2308 *   |- log->last_checkpoint        |- ctx->pos+1
2309 *   |- log->last_cp_seq            |- ctx->seq+10001
2310 *
2311 * However, it is not safe to start the state machine yet, because data only
2312 * parities are not yet secured in RAID. To save these data only parities, we
2313 * rewrite them from seq+11.
2314 *
2315 *   -----------------------------------------------------------------
2316 *   |           valid log        | data only stripes | invalid log  |
2317 *   -----------------------------------------------------------------
2318 *   ^                                                ^
2319 *   |- log->last_checkpoint                          |- ctx->pos+n
2320 *   |- log->last_cp_seq                              |- ctx->seq+10000+n
2321 *
2322 * If failure happens again during this process, the recovery can safe start
2323 * again from log->last_checkpoint.
2324 *
2325 * Once data only stripes are rewritten to journal, we move log_tail
2326 *
2327 *   -----------------------------------------------------------------
2328 *   |     old log        |    data only stripes    | invalid log  |
2329 *   -----------------------------------------------------------------
2330 *                        ^                         ^
2331 *                        |- log->last_checkpoint   |- ctx->pos+n
2332 *                        |- log->last_cp_seq       |- ctx->seq+10000+n
2333 *
2334 * Then we can safely start the state machine. If failure happens from this
2335 * point on, the recovery will start from new log->last_checkpoint.
2336 */
2337static int
2338r5c_recovery_rewrite_data_only_stripes(struct r5l_log *log,
2339				       struct r5l_recovery_ctx *ctx)
2340{
2341	struct stripe_head *sh;
2342	struct mddev *mddev = log->rdev->mddev;
2343	struct page *page;
2344	sector_t next_checkpoint = MaxSector;
2345
2346	page = alloc_page(GFP_KERNEL);
2347	if (!page) {
2348		pr_err("md/raid:%s: cannot allocate memory to rewrite data only stripes\n",
2349		       mdname(mddev));
2350		return -ENOMEM;
2351	}
2352
2353	WARN_ON(list_empty(&ctx->cached_list));
2354
2355	list_for_each_entry(sh, &ctx->cached_list, lru) {
2356		struct r5l_meta_block *mb;
2357		int i;
2358		int offset;
2359		sector_t write_pos;
2360
2361		WARN_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state));
2362		r5l_recovery_create_empty_meta_block(log, page,
2363						     ctx->pos, ctx->seq);
2364		mb = page_address(page);
2365		offset = le32_to_cpu(mb->meta_size);
2366		write_pos = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS);
2367
2368		for (i = sh->disks; i--; ) {
2369			struct r5dev *dev = &sh->dev[i];
2370			struct r5l_payload_data_parity *payload;
2371			void *addr;
2372
2373			if (test_bit(R5_InJournal, &dev->flags)) {
2374				payload = (void *)mb + offset;
2375				payload->header.type = cpu_to_le16(
2376					R5LOG_PAYLOAD_DATA);
2377				payload->size = cpu_to_le32(BLOCK_SECTORS);
2378				payload->location = cpu_to_le64(
2379					raid5_compute_blocknr(sh, i, 0));
2380				addr = kmap_atomic(dev->page);
2381				payload->checksum[0] = cpu_to_le32(
2382					crc32c_le(log->uuid_checksum, addr,
2383						  PAGE_SIZE));
2384				kunmap_atomic(addr);
2385				sync_page_io(log->rdev, write_pos, PAGE_SIZE,
2386					     dev->page, REQ_OP_WRITE, false);
2387				write_pos = r5l_ring_add(log, write_pos,
2388							 BLOCK_SECTORS);
2389				offset += sizeof(__le32) +
2390					sizeof(struct r5l_payload_data_parity);
2391
2392			}
2393		}
2394		mb->meta_size = cpu_to_le32(offset);
2395		mb->checksum = cpu_to_le32(crc32c_le(log->uuid_checksum,
2396						     mb, PAGE_SIZE));
2397		sync_page_io(log->rdev, ctx->pos, PAGE_SIZE, page,
2398			     REQ_OP_WRITE | REQ_SYNC | REQ_FUA, false);
2399		sh->log_start = ctx->pos;
2400		list_add_tail(&sh->r5c, &log->stripe_in_journal_list);
2401		atomic_inc(&log->stripe_in_journal_count);
2402		ctx->pos = write_pos;
2403		ctx->seq += 1;
2404		next_checkpoint = sh->log_start;
2405	}
2406	log->next_checkpoint = next_checkpoint;
2407	__free_page(page);
2408	return 0;
2409}
2410
2411static void r5c_recovery_flush_data_only_stripes(struct r5l_log *log,
2412						 struct r5l_recovery_ctx *ctx)
2413{
2414	struct mddev *mddev = log->rdev->mddev;
2415	struct r5conf *conf = mddev->private;
2416	struct stripe_head *sh, *next;
2417	bool cleared_pending = false;
2418
2419	if (ctx->data_only_stripes == 0)
2420		return;
2421
2422	if (test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2423		cleared_pending = true;
2424		clear_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags);
2425	}
2426	log->r5c_journal_mode = R5C_JOURNAL_MODE_WRITE_BACK;
2427
2428	list_for_each_entry_safe(sh, next, &ctx->cached_list, lru) {
2429		r5c_make_stripe_write_out(sh);
2430		set_bit(STRIPE_HANDLE, &sh->state);
2431		list_del_init(&sh->lru);
2432		raid5_release_stripe(sh);
2433	}
2434
2435	/* reuse conf->wait_for_quiescent in recovery */
2436	wait_event(conf->wait_for_quiescent,
2437		   atomic_read(&conf->active_stripes) == 0);
2438
2439	log->r5c_journal_mode = R5C_JOURNAL_MODE_WRITE_THROUGH;
2440	if (cleared_pending)
2441		set_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags);
2442}
2443
2444static int r5l_recovery_log(struct r5l_log *log)
2445{
2446	struct mddev *mddev = log->rdev->mddev;
2447	struct r5l_recovery_ctx *ctx;
2448	int ret;
2449	sector_t pos;
2450
2451	ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
2452	if (!ctx)
2453		return -ENOMEM;
2454
2455	ctx->pos = log->last_checkpoint;
2456	ctx->seq = log->last_cp_seq;
2457	INIT_LIST_HEAD(&ctx->cached_list);
2458	ctx->meta_page = alloc_page(GFP_KERNEL);
2459
2460	if (!ctx->meta_page) {
2461		ret =  -ENOMEM;
2462		goto meta_page;
2463	}
2464
2465	if (r5l_recovery_allocate_ra_pool(log, ctx) != 0) {
2466		ret = -ENOMEM;
2467		goto ra_pool;
2468	}
2469
2470	ret = r5c_recovery_flush_log(log, ctx);
2471
2472	if (ret)
2473		goto error;
2474
2475	pos = ctx->pos;
2476	ctx->seq += 10000;
2477
2478	if ((ctx->data_only_stripes == 0) && (ctx->data_parity_stripes == 0))
2479		pr_info("md/raid:%s: starting from clean shutdown\n",
2480			 mdname(mddev));
2481	else
2482		pr_info("md/raid:%s: recovering %d data-only stripes and %d data-parity stripes\n",
2483			 mdname(mddev), ctx->data_only_stripes,
2484			 ctx->data_parity_stripes);
2485
2486	if (ctx->data_only_stripes == 0) {
2487		log->next_checkpoint = ctx->pos;
2488		r5l_log_write_empty_meta_block(log, ctx->pos, ctx->seq++);
2489		ctx->pos = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS);
2490	} else if (r5c_recovery_rewrite_data_only_stripes(log, ctx)) {
2491		pr_err("md/raid:%s: failed to rewrite stripes to journal\n",
2492		       mdname(mddev));
2493		ret =  -EIO;
2494		goto error;
2495	}
2496
2497	log->log_start = ctx->pos;
2498	log->seq = ctx->seq;
2499	log->last_checkpoint = pos;
2500	r5l_write_super(log, pos);
2501
2502	r5c_recovery_flush_data_only_stripes(log, ctx);
2503	ret = 0;
2504error:
2505	r5l_recovery_free_ra_pool(log, ctx);
2506ra_pool:
2507	__free_page(ctx->meta_page);
2508meta_page:
2509	kfree(ctx);
2510	return ret;
2511}
2512
2513static void r5l_write_super(struct r5l_log *log, sector_t cp)
2514{
2515	struct mddev *mddev = log->rdev->mddev;
2516
2517	log->rdev->journal_tail = cp;
2518	set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
2519}
2520
2521static ssize_t r5c_journal_mode_show(struct mddev *mddev, char *page)
2522{
2523	struct r5conf *conf;
2524	int ret;
2525
2526	ret = mddev_lock(mddev);
2527	if (ret)
2528		return ret;
2529
2530	conf = mddev->private;
2531	if (!conf || !conf->log)
2532		goto out_unlock;
 
 
2533
2534	switch (conf->log->r5c_journal_mode) {
2535	case R5C_JOURNAL_MODE_WRITE_THROUGH:
2536		ret = snprintf(
2537			page, PAGE_SIZE, "[%s] %s\n",
2538			r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_THROUGH],
2539			r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_BACK]);
2540		break;
2541	case R5C_JOURNAL_MODE_WRITE_BACK:
2542		ret = snprintf(
2543			page, PAGE_SIZE, "%s [%s]\n",
2544			r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_THROUGH],
2545			r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_BACK]);
2546		break;
2547	default:
2548		ret = 0;
2549	}
2550
2551out_unlock:
2552	mddev_unlock(mddev);
2553	return ret;
2554}
2555
2556/*
2557 * Set journal cache mode on @mddev (external API initially needed by dm-raid).
2558 *
2559 * @mode as defined in 'enum r5c_journal_mode'.
2560 *
2561 */
2562int r5c_journal_mode_set(struct mddev *mddev, int mode)
2563{
2564	struct r5conf *conf;
2565
2566	if (mode < R5C_JOURNAL_MODE_WRITE_THROUGH ||
2567	    mode > R5C_JOURNAL_MODE_WRITE_BACK)
2568		return -EINVAL;
2569
2570	conf = mddev->private;
2571	if (!conf || !conf->log)
2572		return -ENODEV;
2573
2574	if (raid5_calc_degraded(conf) > 0 &&
2575	    mode == R5C_JOURNAL_MODE_WRITE_BACK)
2576		return -EINVAL;
2577
 
2578	conf->log->r5c_journal_mode = mode;
 
2579
2580	pr_debug("md/raid:%s: setting r5c cache mode to %d: %s\n",
2581		 mdname(mddev), mode, r5c_journal_mode_str[mode]);
2582	return 0;
2583}
2584EXPORT_SYMBOL(r5c_journal_mode_set);
2585
2586static ssize_t r5c_journal_mode_store(struct mddev *mddev,
2587				      const char *page, size_t length)
2588{
2589	int mode = ARRAY_SIZE(r5c_journal_mode_str);
2590	size_t len = length;
2591	int ret;
2592
2593	if (len < 2)
2594		return -EINVAL;
2595
2596	if (page[len - 1] == '\n')
2597		len--;
2598
2599	while (mode--)
2600		if (strlen(r5c_journal_mode_str[mode]) == len &&
2601		    !strncmp(page, r5c_journal_mode_str[mode], len))
2602			break;
2603	ret = mddev_suspend_and_lock(mddev);
2604	if (ret)
2605		return ret;
2606	ret = r5c_journal_mode_set(mddev, mode);
2607	mddev_unlock_and_resume(mddev);
2608	return ret ?: length;
2609}
2610
2611struct md_sysfs_entry
2612r5c_journal_mode = __ATTR(journal_mode, 0644,
2613			  r5c_journal_mode_show, r5c_journal_mode_store);
2614
2615/*
2616 * Try handle write operation in caching phase. This function should only
2617 * be called in write-back mode.
2618 *
2619 * If all outstanding writes can be handled in caching phase, returns 0
2620 * If writes requires write-out phase, call r5c_make_stripe_write_out()
2621 * and returns -EAGAIN
2622 */
2623int r5c_try_caching_write(struct r5conf *conf,
2624			  struct stripe_head *sh,
2625			  struct stripe_head_state *s,
2626			  int disks)
2627{
2628	struct r5l_log *log = READ_ONCE(conf->log);
2629	int i;
2630	struct r5dev *dev;
2631	int to_cache = 0;
2632	void __rcu **pslot;
2633	sector_t tree_index;
2634	int ret;
2635	uintptr_t refcount;
2636
2637	BUG_ON(!r5c_is_writeback(log));
2638
2639	if (!test_bit(STRIPE_R5C_CACHING, &sh->state)) {
2640		/*
2641		 * There are two different scenarios here:
2642		 *  1. The stripe has some data cached, and it is sent to
2643		 *     write-out phase for reclaim
2644		 *  2. The stripe is clean, and this is the first write
2645		 *
2646		 * For 1, return -EAGAIN, so we continue with
2647		 * handle_stripe_dirtying().
2648		 *
2649		 * For 2, set STRIPE_R5C_CACHING and continue with caching
2650		 * write.
2651		 */
2652
2653		/* case 1: anything injournal or anything in written */
2654		if (s->injournal > 0 || s->written > 0)
2655			return -EAGAIN;
2656		/* case 2 */
2657		set_bit(STRIPE_R5C_CACHING, &sh->state);
2658	}
2659
2660	/*
2661	 * When run in degraded mode, array is set to write-through mode.
2662	 * This check helps drain pending write safely in the transition to
2663	 * write-through mode.
2664	 *
2665	 * When a stripe is syncing, the write is also handled in write
2666	 * through mode.
2667	 */
2668	if (s->failed || test_bit(STRIPE_SYNCING, &sh->state)) {
2669		r5c_make_stripe_write_out(sh);
2670		return -EAGAIN;
2671	}
2672
2673	for (i = disks; i--; ) {
2674		dev = &sh->dev[i];
2675		/* if non-overwrite, use writing-out phase */
2676		if (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags) &&
2677		    !test_bit(R5_InJournal, &dev->flags)) {
2678			r5c_make_stripe_write_out(sh);
2679			return -EAGAIN;
2680		}
2681	}
2682
2683	/* if the stripe is not counted in big_stripe_tree, add it now */
2684	if (!test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state) &&
2685	    !test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state)) {
2686		tree_index = r5c_tree_index(conf, sh->sector);
2687		spin_lock(&log->tree_lock);
2688		pslot = radix_tree_lookup_slot(&log->big_stripe_tree,
2689					       tree_index);
2690		if (pslot) {
2691			refcount = (uintptr_t)radix_tree_deref_slot_protected(
2692				pslot, &log->tree_lock) >>
2693				R5C_RADIX_COUNT_SHIFT;
2694			radix_tree_replace_slot(
2695				&log->big_stripe_tree, pslot,
2696				(void *)((refcount + 1) << R5C_RADIX_COUNT_SHIFT));
2697		} else {
2698			/*
2699			 * this radix_tree_insert can fail safely, so no
2700			 * need to call radix_tree_preload()
2701			 */
2702			ret = radix_tree_insert(
2703				&log->big_stripe_tree, tree_index,
2704				(void *)(1 << R5C_RADIX_COUNT_SHIFT));
2705			if (ret) {
2706				spin_unlock(&log->tree_lock);
2707				r5c_make_stripe_write_out(sh);
2708				return -EAGAIN;
2709			}
2710		}
2711		spin_unlock(&log->tree_lock);
2712
2713		/*
2714		 * set STRIPE_R5C_PARTIAL_STRIPE, this shows the stripe is
2715		 * counted in the radix tree
2716		 */
2717		set_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state);
2718		atomic_inc(&conf->r5c_cached_partial_stripes);
2719	}
2720
2721	for (i = disks; i--; ) {
2722		dev = &sh->dev[i];
2723		if (dev->towrite) {
2724			set_bit(R5_Wantwrite, &dev->flags);
2725			set_bit(R5_Wantdrain, &dev->flags);
2726			set_bit(R5_LOCKED, &dev->flags);
2727			to_cache++;
2728		}
2729	}
2730
2731	if (to_cache) {
2732		set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2733		/*
2734		 * set STRIPE_LOG_TRAPPED, which triggers r5c_cache_data()
2735		 * in ops_run_io(). STRIPE_LOG_TRAPPED will be cleared in
2736		 * r5c_handle_data_cached()
2737		 */
2738		set_bit(STRIPE_LOG_TRAPPED, &sh->state);
2739	}
2740
2741	return 0;
2742}
2743
2744/*
2745 * free extra pages (orig_page) we allocated for prexor
2746 */
2747void r5c_release_extra_page(struct stripe_head *sh)
2748{
2749	struct r5conf *conf = sh->raid_conf;
2750	int i;
2751	bool using_disk_info_extra_page;
2752
2753	using_disk_info_extra_page =
2754		sh->dev[0].orig_page == conf->disks[0].extra_page;
2755
2756	for (i = sh->disks; i--; )
2757		if (sh->dev[i].page != sh->dev[i].orig_page) {
2758			struct page *p = sh->dev[i].orig_page;
2759
2760			sh->dev[i].orig_page = sh->dev[i].page;
2761			clear_bit(R5_OrigPageUPTDODATE, &sh->dev[i].flags);
2762
2763			if (!using_disk_info_extra_page)
2764				put_page(p);
2765		}
2766
2767	if (using_disk_info_extra_page) {
2768		clear_bit(R5C_EXTRA_PAGE_IN_USE, &conf->cache_state);
2769		md_wakeup_thread(conf->mddev->thread);
2770	}
2771}
2772
2773void r5c_use_extra_page(struct stripe_head *sh)
2774{
2775	struct r5conf *conf = sh->raid_conf;
2776	int i;
2777	struct r5dev *dev;
2778
2779	for (i = sh->disks; i--; ) {
2780		dev = &sh->dev[i];
2781		if (dev->orig_page != dev->page)
2782			put_page(dev->orig_page);
2783		dev->orig_page = conf->disks[i].extra_page;
2784	}
2785}
2786
2787/*
2788 * clean up the stripe (clear R5_InJournal for dev[pd_idx] etc.) after the
2789 * stripe is committed to RAID disks.
2790 */
2791void r5c_finish_stripe_write_out(struct r5conf *conf,
2792				 struct stripe_head *sh,
2793				 struct stripe_head_state *s)
2794{
2795	struct r5l_log *log = READ_ONCE(conf->log);
2796	int i;
 
2797	sector_t tree_index;
2798	void __rcu **pslot;
2799	uintptr_t refcount;
2800
2801	if (!log || !test_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags))
2802		return;
2803
2804	WARN_ON(test_bit(STRIPE_R5C_CACHING, &sh->state));
2805	clear_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags);
2806
2807	if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH)
2808		return;
2809
2810	for (i = sh->disks; i--; ) {
2811		clear_bit(R5_InJournal, &sh->dev[i].flags);
2812		if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2813			wake_up_bit(&sh->dev[i].flags, R5_Overlap);
2814	}
2815
2816	/*
2817	 * analyse_stripe() runs before r5c_finish_stripe_write_out(),
2818	 * We updated R5_InJournal, so we also update s->injournal.
2819	 */
2820	s->injournal = 0;
2821
2822	if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2823		if (atomic_dec_and_test(&conf->pending_full_writes))
2824			md_wakeup_thread(conf->mddev->thread);
2825
 
 
 
2826	spin_lock_irq(&log->stripe_in_journal_lock);
2827	list_del_init(&sh->r5c);
2828	spin_unlock_irq(&log->stripe_in_journal_lock);
2829	sh->log_start = MaxSector;
2830
2831	atomic_dec(&log->stripe_in_journal_count);
2832	r5c_update_log_state(log);
2833
2834	/* stop counting this stripe in big_stripe_tree */
2835	if (test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state) ||
2836	    test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state)) {
2837		tree_index = r5c_tree_index(conf, sh->sector);
2838		spin_lock(&log->tree_lock);
2839		pslot = radix_tree_lookup_slot(&log->big_stripe_tree,
2840					       tree_index);
2841		BUG_ON(pslot == NULL);
2842		refcount = (uintptr_t)radix_tree_deref_slot_protected(
2843			pslot, &log->tree_lock) >>
2844			R5C_RADIX_COUNT_SHIFT;
2845		if (refcount == 1)
2846			radix_tree_delete(&log->big_stripe_tree, tree_index);
2847		else
2848			radix_tree_replace_slot(
2849				&log->big_stripe_tree, pslot,
2850				(void *)((refcount - 1) << R5C_RADIX_COUNT_SHIFT));
2851		spin_unlock(&log->tree_lock);
2852	}
2853
2854	if (test_and_clear_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state)) {
2855		BUG_ON(atomic_read(&conf->r5c_cached_partial_stripes) == 0);
2856		atomic_dec(&conf->r5c_flushing_partial_stripes);
2857		atomic_dec(&conf->r5c_cached_partial_stripes);
2858	}
2859
2860	if (test_and_clear_bit(STRIPE_R5C_FULL_STRIPE, &sh->state)) {
2861		BUG_ON(atomic_read(&conf->r5c_cached_full_stripes) == 0);
2862		atomic_dec(&conf->r5c_flushing_full_stripes);
2863		atomic_dec(&conf->r5c_cached_full_stripes);
2864	}
2865
2866	r5l_append_flush_payload(log, sh->sector);
2867	/* stripe is flused to raid disks, we can do resync now */
2868	if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state))
2869		set_bit(STRIPE_HANDLE, &sh->state);
2870}
2871
2872int r5c_cache_data(struct r5l_log *log, struct stripe_head *sh)
2873{
2874	struct r5conf *conf = sh->raid_conf;
2875	int pages = 0;
2876	int reserve;
2877	int i;
2878	int ret = 0;
2879
2880	BUG_ON(!log);
2881
2882	for (i = 0; i < sh->disks; i++) {
2883		void *addr;
2884
2885		if (!test_bit(R5_Wantwrite, &sh->dev[i].flags))
2886			continue;
2887		addr = kmap_atomic(sh->dev[i].page);
2888		sh->dev[i].log_checksum = crc32c_le(log->uuid_checksum,
2889						    addr, PAGE_SIZE);
2890		kunmap_atomic(addr);
2891		pages++;
2892	}
2893	WARN_ON(pages == 0);
2894
2895	/*
2896	 * The stripe must enter state machine again to call endio, so
2897	 * don't delay.
2898	 */
2899	clear_bit(STRIPE_DELAYED, &sh->state);
2900	atomic_inc(&sh->count);
2901
2902	mutex_lock(&log->io_mutex);
2903	/* meta + data */
2904	reserve = (1 + pages) << (PAGE_SHIFT - 9);
2905
2906	if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state) &&
2907	    sh->log_start == MaxSector)
2908		r5l_add_no_space_stripe(log, sh);
2909	else if (!r5l_has_free_space(log, reserve)) {
2910		if (sh->log_start == log->last_checkpoint)
2911			BUG();
2912		else
2913			r5l_add_no_space_stripe(log, sh);
2914	} else {
2915		ret = r5l_log_stripe(log, sh, pages, 0);
2916		if (ret) {
2917			spin_lock_irq(&log->io_list_lock);
2918			list_add_tail(&sh->log_list, &log->no_mem_stripes);
2919			spin_unlock_irq(&log->io_list_lock);
2920		}
2921	}
2922
2923	mutex_unlock(&log->io_mutex);
2924	return 0;
2925}
2926
2927/* check whether this big stripe is in write back cache. */
2928bool r5c_big_stripe_cached(struct r5conf *conf, sector_t sect)
2929{
2930	struct r5l_log *log = READ_ONCE(conf->log);
2931	sector_t tree_index;
2932	void *slot;
2933
2934	if (!log)
2935		return false;
2936
 
2937	tree_index = r5c_tree_index(conf, sect);
2938	slot = radix_tree_lookup(&log->big_stripe_tree, tree_index);
2939	return slot != NULL;
2940}
2941
2942static int r5l_load_log(struct r5l_log *log)
2943{
2944	struct md_rdev *rdev = log->rdev;
2945	struct page *page;
2946	struct r5l_meta_block *mb;
2947	sector_t cp = log->rdev->journal_tail;
2948	u32 stored_crc, expected_crc;
2949	bool create_super = false;
2950	int ret = 0;
2951
2952	/* Make sure it's valid */
2953	if (cp >= rdev->sectors || round_down(cp, BLOCK_SECTORS) != cp)
2954		cp = 0;
2955	page = alloc_page(GFP_KERNEL);
2956	if (!page)
2957		return -ENOMEM;
2958
2959	if (!sync_page_io(rdev, cp, PAGE_SIZE, page, REQ_OP_READ, false)) {
2960		ret = -EIO;
2961		goto ioerr;
2962	}
2963	mb = page_address(page);
2964
2965	if (le32_to_cpu(mb->magic) != R5LOG_MAGIC ||
2966	    mb->version != R5LOG_VERSION) {
2967		create_super = true;
2968		goto create;
2969	}
2970	stored_crc = le32_to_cpu(mb->checksum);
2971	mb->checksum = 0;
2972	expected_crc = crc32c_le(log->uuid_checksum, mb, PAGE_SIZE);
2973	if (stored_crc != expected_crc) {
2974		create_super = true;
2975		goto create;
2976	}
2977	if (le64_to_cpu(mb->position) != cp) {
2978		create_super = true;
2979		goto create;
2980	}
2981create:
2982	if (create_super) {
2983		log->last_cp_seq = get_random_u32();
2984		cp = 0;
2985		r5l_log_write_empty_meta_block(log, cp, log->last_cp_seq);
2986		/*
2987		 * Make sure super points to correct address. Log might have
2988		 * data very soon. If super hasn't correct log tail address,
2989		 * recovery can't find the log
2990		 */
2991		r5l_write_super(log, cp);
2992	} else
2993		log->last_cp_seq = le64_to_cpu(mb->seq);
2994
2995	log->device_size = round_down(rdev->sectors, BLOCK_SECTORS);
2996	log->max_free_space = log->device_size >> RECLAIM_MAX_FREE_SPACE_SHIFT;
2997	if (log->max_free_space > RECLAIM_MAX_FREE_SPACE)
2998		log->max_free_space = RECLAIM_MAX_FREE_SPACE;
2999	log->last_checkpoint = cp;
3000
3001	__free_page(page);
3002
3003	if (create_super) {
3004		log->log_start = r5l_ring_add(log, cp, BLOCK_SECTORS);
3005		log->seq = log->last_cp_seq + 1;
3006		log->next_checkpoint = cp;
3007	} else
3008		ret = r5l_recovery_log(log);
3009
3010	r5c_update_log_state(log);
3011	return ret;
3012ioerr:
3013	__free_page(page);
3014	return ret;
3015}
3016
3017int r5l_start(struct r5l_log *log)
3018{
3019	int ret;
3020
3021	if (!log)
3022		return 0;
3023
3024	ret = r5l_load_log(log);
3025	if (ret) {
3026		struct mddev *mddev = log->rdev->mddev;
3027		struct r5conf *conf = mddev->private;
3028
3029		r5l_exit_log(conf);
3030	}
3031	return ret;
3032}
3033
3034void r5c_update_on_rdev_error(struct mddev *mddev, struct md_rdev *rdev)
3035{
3036	struct r5conf *conf = mddev->private;
3037	struct r5l_log *log = READ_ONCE(conf->log);
3038
3039	if (!log)
3040		return;
3041
3042	if ((raid5_calc_degraded(conf) > 0 ||
3043	     test_bit(Journal, &rdev->flags)) &&
3044	    log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_BACK)
3045		schedule_work(&log->disable_writeback_work);
3046}
3047
3048int r5l_init_log(struct r5conf *conf, struct md_rdev *rdev)
3049{
 
3050	struct r5l_log *log;
3051	struct md_thread *thread;
3052	int ret;
3053
3054	pr_debug("md/raid:%s: using device %pg as journal\n",
3055		 mdname(conf->mddev), rdev->bdev);
3056
3057	if (PAGE_SIZE != 4096)
3058		return -EINVAL;
3059
3060	/*
3061	 * The PAGE_SIZE must be big enough to hold 1 r5l_meta_block and
3062	 * raid_disks r5l_payload_data_parity.
3063	 *
3064	 * Write journal and cache does not work for very big array
3065	 * (raid_disks > 203)
3066	 */
3067	if (sizeof(struct r5l_meta_block) +
3068	    ((sizeof(struct r5l_payload_data_parity) + sizeof(__le32)) *
3069	     conf->raid_disks) > PAGE_SIZE) {
3070		pr_err("md/raid:%s: write journal/cache doesn't work for array with %d disks\n",
3071		       mdname(conf->mddev), conf->raid_disks);
3072		return -EINVAL;
3073	}
3074
3075	log = kzalloc(sizeof(*log), GFP_KERNEL);
3076	if (!log)
3077		return -ENOMEM;
3078	log->rdev = rdev;
3079	log->need_cache_flush = bdev_write_cache(rdev->bdev);
 
 
3080	log->uuid_checksum = crc32c_le(~0, rdev->mddev->uuid,
3081				       sizeof(rdev->mddev->uuid));
3082
3083	mutex_init(&log->io_mutex);
3084
3085	spin_lock_init(&log->io_list_lock);
3086	INIT_LIST_HEAD(&log->running_ios);
3087	INIT_LIST_HEAD(&log->io_end_ios);
3088	INIT_LIST_HEAD(&log->flushing_ios);
3089	INIT_LIST_HEAD(&log->finished_ios);
 
3090
3091	log->io_kc = KMEM_CACHE(r5l_io_unit, 0);
3092	if (!log->io_kc)
3093		goto io_kc;
3094
3095	ret = mempool_init_slab_pool(&log->io_pool, R5L_POOL_SIZE, log->io_kc);
3096	if (ret)
3097		goto io_pool;
3098
3099	ret = bioset_init(&log->bs, R5L_POOL_SIZE, 0, BIOSET_NEED_BVECS);
3100	if (ret)
3101		goto io_bs;
3102
3103	ret = mempool_init_page_pool(&log->meta_pool, R5L_POOL_SIZE, 0);
3104	if (ret)
3105		goto out_mempool;
3106
3107	spin_lock_init(&log->tree_lock);
3108	INIT_RADIX_TREE(&log->big_stripe_tree, GFP_NOWAIT | __GFP_NOWARN);
3109
3110	thread = md_register_thread(r5l_reclaim_thread, log->rdev->mddev,
3111				    "reclaim");
3112	if (!thread)
3113		goto reclaim_thread;
3114
3115	thread->timeout = R5C_RECLAIM_WAKEUP_INTERVAL;
3116	rcu_assign_pointer(log->reclaim_thread, thread);
3117
3118	init_waitqueue_head(&log->iounit_wait);
3119
3120	INIT_LIST_HEAD(&log->no_mem_stripes);
3121
3122	INIT_LIST_HEAD(&log->no_space_stripes);
3123	spin_lock_init(&log->no_space_stripes_lock);
3124
3125	INIT_WORK(&log->deferred_io_work, r5l_submit_io_async);
3126	INIT_WORK(&log->disable_writeback_work, r5c_disable_writeback_async);
3127
3128	log->r5c_journal_mode = R5C_JOURNAL_MODE_WRITE_THROUGH;
3129	INIT_LIST_HEAD(&log->stripe_in_journal_list);
3130	spin_lock_init(&log->stripe_in_journal_lock);
3131	atomic_set(&log->stripe_in_journal_count, 0);
3132
3133	WRITE_ONCE(conf->log, log);
3134
3135	set_bit(MD_HAS_JOURNAL, &conf->mddev->flags);
3136	return 0;
3137
 
 
3138reclaim_thread:
3139	mempool_exit(&log->meta_pool);
3140out_mempool:
3141	bioset_exit(&log->bs);
3142io_bs:
3143	mempool_exit(&log->io_pool);
3144io_pool:
3145	kmem_cache_destroy(log->io_kc);
3146io_kc:
3147	kfree(log);
3148	return -EINVAL;
3149}
3150
3151void r5l_exit_log(struct r5conf *conf)
3152{
3153	struct r5l_log *log = conf->log;
3154
3155	md_unregister_thread(conf->mddev, &log->reclaim_thread);
 
3156
3157	/*
3158	 * 'reconfig_mutex' is held by caller, set 'confg->log' to NULL to
3159	 * ensure disable_writeback_work wakes up and exits.
3160	 */
3161	WRITE_ONCE(conf->log, NULL);
3162	wake_up(&conf->mddev->sb_wait);
3163	flush_work(&log->disable_writeback_work);
3164
3165	mempool_exit(&log->meta_pool);
3166	bioset_exit(&log->bs);
3167	mempool_exit(&log->io_pool);
3168	kmem_cache_destroy(log->io_kc);
3169	kfree(log);
3170}
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}