xref: /openbmc/linux/drivers/md/raid5-cache.c (revision fd5e9fccbd504c5179ab57ff695c610bca8809d6)
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 
49 static 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 
82 struct 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  */
r5c_tree_index(struct r5conf * conf,sector_t sect)195 static 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  */
209 struct 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 */
242 enum 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 
r5c_is_writeback(struct r5l_log * log)250 bool r5c_is_writeback(struct r5l_log *log)
251 {
252 	return (log != NULL &&
253 		log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_BACK);
254 }
255 
r5l_ring_add(struct r5l_log * log,sector_t start,sector_t inc)256 static 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 
r5l_ring_distance(struct r5l_log * log,sector_t start,sector_t end)264 static 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 
r5l_has_free_space(struct r5l_log * log,sector_t size)273 static 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 
__r5l_set_io_unit_state(struct r5l_io_unit * io,enum r5l_io_unit_state state)283 static 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 
291 static void
r5c_return_dev_pending_writes(struct r5conf * conf,struct r5dev * dev)292 r5c_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 
r5c_handle_cached_data_endio(struct r5conf * conf,struct stripe_head * sh,int disks)307 void 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 
320 void r5l_wake_reclaim(struct r5l_log *log, sector_t space);
321 
322 /* Check whether we should flush some stripes to free up stripe cache */
r5c_check_stripe_cache_usage(struct r5conf * conf)323 void 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  */
r5c_check_cached_full_stripe(struct r5conf * conf)351 void 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  */
r5c_log_required_to_flush_cache(struct r5conf * conf)396 static 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  */
r5c_update_log_state(struct r5l_log * log)415 static 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  */
r5c_make_stripe_write_out(struct stripe_head * sh)448 void 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 
r5c_handle_data_cached(struct stripe_head * sh)462 static 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  */
r5c_handle_parity_cached(struct stripe_head * sh)478 static 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  */
r5c_finish_cache_stripe(struct stripe_head * sh)491 static 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 
r5l_io_run_stripes(struct r5l_io_unit * io)512 static 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 
r5l_log_run_stripes(struct r5l_log * log)526 static 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 
r5l_move_to_end_ios(struct r5l_log * log)542 static 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 
556 static void __r5l_stripe_write_finished(struct r5l_io_unit *io);
r5l_log_endio(struct bio * bio)557 static 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 
r5l_do_submit_io(struct r5l_log * log,struct r5l_io_unit * io)623 static 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 */
r5l_submit_io_async(struct work_struct * work)658 static 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 
r5c_disable_writeback_async(struct work_struct * work)679 static 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 	int locked = 0;
686 
687 	if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH)
688 		return;
689 	pr_info("md/raid:%s: Disabling writeback cache for degraded array.\n",
690 		mdname(mddev));
691 
692 	/* wait superblock change before suspend */
693 	wait_event(mddev->sb_wait,
694 		   !READ_ONCE(conf->log) ||
695 		   (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags) &&
696 		    (locked = mddev_trylock(mddev))));
697 	if (locked) {
698 		mddev_suspend(mddev);
699 		log->r5c_journal_mode = R5C_JOURNAL_MODE_WRITE_THROUGH;
700 		mddev_resume(mddev);
701 		mddev_unlock(mddev);
702 	}
703 }
704 
r5l_submit_current_io(struct r5l_log * log)705 static void r5l_submit_current_io(struct r5l_log *log)
706 {
707 	struct r5l_io_unit *io = log->current_io;
708 	struct r5l_meta_block *block;
709 	unsigned long flags;
710 	u32 crc;
711 	bool do_submit = true;
712 
713 	if (!io)
714 		return;
715 
716 	block = page_address(io->meta_page);
717 	block->meta_size = cpu_to_le32(io->meta_offset);
718 	crc = crc32c_le(log->uuid_checksum, block, PAGE_SIZE);
719 	block->checksum = cpu_to_le32(crc);
720 
721 	log->current_io = NULL;
722 	spin_lock_irqsave(&log->io_list_lock, flags);
723 	if (io->has_flush || io->has_fua) {
724 		if (io != list_first_entry(&log->running_ios,
725 					   struct r5l_io_unit, log_sibling)) {
726 			io->io_deferred = 1;
727 			do_submit = false;
728 		}
729 	}
730 	spin_unlock_irqrestore(&log->io_list_lock, flags);
731 	if (do_submit)
732 		r5l_do_submit_io(log, io);
733 }
734 
r5l_bio_alloc(struct r5l_log * log)735 static struct bio *r5l_bio_alloc(struct r5l_log *log)
736 {
737 	struct bio *bio = bio_alloc_bioset(log->rdev->bdev, BIO_MAX_VECS,
738 					   REQ_OP_WRITE, GFP_NOIO, &log->bs);
739 
740 	bio->bi_iter.bi_sector = log->rdev->data_offset + log->log_start;
741 
742 	return bio;
743 }
744 
r5_reserve_log_entry(struct r5l_log * log,struct r5l_io_unit * io)745 static void r5_reserve_log_entry(struct r5l_log *log, struct r5l_io_unit *io)
746 {
747 	log->log_start = r5l_ring_add(log, log->log_start, BLOCK_SECTORS);
748 
749 	r5c_update_log_state(log);
750 	/*
751 	 * If we filled up the log device start from the beginning again,
752 	 * which will require a new bio.
753 	 *
754 	 * Note: for this to work properly the log size needs to me a multiple
755 	 * of BLOCK_SECTORS.
756 	 */
757 	if (log->log_start == 0)
758 		io->need_split_bio = true;
759 
760 	io->log_end = log->log_start;
761 }
762 
r5l_new_meta(struct r5l_log * log)763 static struct r5l_io_unit *r5l_new_meta(struct r5l_log *log)
764 {
765 	struct r5l_io_unit *io;
766 	struct r5l_meta_block *block;
767 
768 	io = mempool_alloc(&log->io_pool, GFP_ATOMIC);
769 	if (!io)
770 		return NULL;
771 	memset(io, 0, sizeof(*io));
772 
773 	io->log = log;
774 	INIT_LIST_HEAD(&io->log_sibling);
775 	INIT_LIST_HEAD(&io->stripe_list);
776 	bio_list_init(&io->flush_barriers);
777 	io->state = IO_UNIT_RUNNING;
778 
779 	io->meta_page = mempool_alloc(&log->meta_pool, GFP_NOIO);
780 	block = page_address(io->meta_page);
781 	clear_page(block);
782 	block->magic = cpu_to_le32(R5LOG_MAGIC);
783 	block->version = R5LOG_VERSION;
784 	block->seq = cpu_to_le64(log->seq);
785 	block->position = cpu_to_le64(log->log_start);
786 
787 	io->log_start = log->log_start;
788 	io->meta_offset = sizeof(struct r5l_meta_block);
789 	io->seq = log->seq++;
790 
791 	io->current_bio = r5l_bio_alloc(log);
792 	io->current_bio->bi_end_io = r5l_log_endio;
793 	io->current_bio->bi_private = io;
794 	__bio_add_page(io->current_bio, io->meta_page, PAGE_SIZE, 0);
795 
796 	r5_reserve_log_entry(log, io);
797 
798 	spin_lock_irq(&log->io_list_lock);
799 	list_add_tail(&io->log_sibling, &log->running_ios);
800 	spin_unlock_irq(&log->io_list_lock);
801 
802 	return io;
803 }
804 
r5l_get_meta(struct r5l_log * log,unsigned int payload_size)805 static int r5l_get_meta(struct r5l_log *log, unsigned int payload_size)
806 {
807 	if (log->current_io &&
808 	    log->current_io->meta_offset + payload_size > PAGE_SIZE)
809 		r5l_submit_current_io(log);
810 
811 	if (!log->current_io) {
812 		log->current_io = r5l_new_meta(log);
813 		if (!log->current_io)
814 			return -ENOMEM;
815 	}
816 
817 	return 0;
818 }
819 
r5l_append_payload_meta(struct r5l_log * log,u16 type,sector_t location,u32 checksum1,u32 checksum2,bool checksum2_valid)820 static void r5l_append_payload_meta(struct r5l_log *log, u16 type,
821 				    sector_t location,
822 				    u32 checksum1, u32 checksum2,
823 				    bool checksum2_valid)
824 {
825 	struct r5l_io_unit *io = log->current_io;
826 	struct r5l_payload_data_parity *payload;
827 
828 	payload = page_address(io->meta_page) + io->meta_offset;
829 	payload->header.type = cpu_to_le16(type);
830 	payload->header.flags = cpu_to_le16(0);
831 	payload->size = cpu_to_le32((1 + !!checksum2_valid) <<
832 				    (PAGE_SHIFT - 9));
833 	payload->location = cpu_to_le64(location);
834 	payload->checksum[0] = cpu_to_le32(checksum1);
835 	if (checksum2_valid)
836 		payload->checksum[1] = cpu_to_le32(checksum2);
837 
838 	io->meta_offset += sizeof(struct r5l_payload_data_parity) +
839 		sizeof(__le32) * (1 + !!checksum2_valid);
840 }
841 
r5l_append_payload_page(struct r5l_log * log,struct page * page)842 static void r5l_append_payload_page(struct r5l_log *log, struct page *page)
843 {
844 	struct r5l_io_unit *io = log->current_io;
845 
846 	if (io->need_split_bio) {
847 		BUG_ON(io->split_bio);
848 		io->split_bio = io->current_bio;
849 		io->current_bio = r5l_bio_alloc(log);
850 		bio_chain(io->current_bio, io->split_bio);
851 		io->need_split_bio = false;
852 	}
853 
854 	if (!bio_add_page(io->current_bio, page, PAGE_SIZE, 0))
855 		BUG();
856 
857 	r5_reserve_log_entry(log, io);
858 }
859 
r5l_append_flush_payload(struct r5l_log * log,sector_t sect)860 static void r5l_append_flush_payload(struct r5l_log *log, sector_t sect)
861 {
862 	struct mddev *mddev = log->rdev->mddev;
863 	struct r5conf *conf = mddev->private;
864 	struct r5l_io_unit *io;
865 	struct r5l_payload_flush *payload;
866 	int meta_size;
867 
868 	/*
869 	 * payload_flush requires extra writes to the journal.
870 	 * To avoid handling the extra IO in quiesce, just skip
871 	 * flush_payload
872 	 */
873 	if (conf->quiesce)
874 		return;
875 
876 	mutex_lock(&log->io_mutex);
877 	meta_size = sizeof(struct r5l_payload_flush) + sizeof(__le64);
878 
879 	if (r5l_get_meta(log, meta_size)) {
880 		mutex_unlock(&log->io_mutex);
881 		return;
882 	}
883 
884 	/* current implementation is one stripe per flush payload */
885 	io = log->current_io;
886 	payload = page_address(io->meta_page) + io->meta_offset;
887 	payload->header.type = cpu_to_le16(R5LOG_PAYLOAD_FLUSH);
888 	payload->header.flags = cpu_to_le16(0);
889 	payload->size = cpu_to_le32(sizeof(__le64));
890 	payload->flush_stripes[0] = cpu_to_le64(sect);
891 	io->meta_offset += meta_size;
892 	/* multiple flush payloads count as one pending_stripe */
893 	if (!io->has_flush_payload) {
894 		io->has_flush_payload = 1;
895 		atomic_inc(&io->pending_stripe);
896 	}
897 	mutex_unlock(&log->io_mutex);
898 }
899 
r5l_log_stripe(struct r5l_log * log,struct stripe_head * sh,int data_pages,int parity_pages)900 static int r5l_log_stripe(struct r5l_log *log, struct stripe_head *sh,
901 			   int data_pages, int parity_pages)
902 {
903 	int i;
904 	int meta_size;
905 	int ret;
906 	struct r5l_io_unit *io;
907 
908 	meta_size =
909 		((sizeof(struct r5l_payload_data_parity) + sizeof(__le32))
910 		 * data_pages) +
911 		sizeof(struct r5l_payload_data_parity) +
912 		sizeof(__le32) * parity_pages;
913 
914 	ret = r5l_get_meta(log, meta_size);
915 	if (ret)
916 		return ret;
917 
918 	io = log->current_io;
919 
920 	if (test_and_clear_bit(STRIPE_R5C_PREFLUSH, &sh->state))
921 		io->has_flush = 1;
922 
923 	for (i = 0; i < sh->disks; i++) {
924 		if (!test_bit(R5_Wantwrite, &sh->dev[i].flags) ||
925 		    test_bit(R5_InJournal, &sh->dev[i].flags))
926 			continue;
927 		if (i == sh->pd_idx || i == sh->qd_idx)
928 			continue;
929 		if (test_bit(R5_WantFUA, &sh->dev[i].flags) &&
930 		    log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_BACK) {
931 			io->has_fua = 1;
932 			/*
933 			 * we need to flush journal to make sure recovery can
934 			 * reach the data with fua flag
935 			 */
936 			io->has_flush = 1;
937 		}
938 		r5l_append_payload_meta(log, R5LOG_PAYLOAD_DATA,
939 					raid5_compute_blocknr(sh, i, 0),
940 					sh->dev[i].log_checksum, 0, false);
941 		r5l_append_payload_page(log, sh->dev[i].page);
942 	}
943 
944 	if (parity_pages == 2) {
945 		r5l_append_payload_meta(log, R5LOG_PAYLOAD_PARITY,
946 					sh->sector, sh->dev[sh->pd_idx].log_checksum,
947 					sh->dev[sh->qd_idx].log_checksum, true);
948 		r5l_append_payload_page(log, sh->dev[sh->pd_idx].page);
949 		r5l_append_payload_page(log, sh->dev[sh->qd_idx].page);
950 	} else if (parity_pages == 1) {
951 		r5l_append_payload_meta(log, R5LOG_PAYLOAD_PARITY,
952 					sh->sector, sh->dev[sh->pd_idx].log_checksum,
953 					0, false);
954 		r5l_append_payload_page(log, sh->dev[sh->pd_idx].page);
955 	} else  /* Just writing data, not parity, in caching phase */
956 		BUG_ON(parity_pages != 0);
957 
958 	list_add_tail(&sh->log_list, &io->stripe_list);
959 	atomic_inc(&io->pending_stripe);
960 	sh->log_io = io;
961 
962 	if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH)
963 		return 0;
964 
965 	if (sh->log_start == MaxSector) {
966 		BUG_ON(!list_empty(&sh->r5c));
967 		sh->log_start = io->log_start;
968 		spin_lock_irq(&log->stripe_in_journal_lock);
969 		list_add_tail(&sh->r5c,
970 			      &log->stripe_in_journal_list);
971 		spin_unlock_irq(&log->stripe_in_journal_lock);
972 		atomic_inc(&log->stripe_in_journal_count);
973 	}
974 	return 0;
975 }
976 
977 /* add stripe to no_space_stripes, and then wake up reclaim */
r5l_add_no_space_stripe(struct r5l_log * log,struct stripe_head * sh)978 static inline void r5l_add_no_space_stripe(struct r5l_log *log,
979 					   struct stripe_head *sh)
980 {
981 	spin_lock(&log->no_space_stripes_lock);
982 	list_add_tail(&sh->log_list, &log->no_space_stripes);
983 	spin_unlock(&log->no_space_stripes_lock);
984 }
985 
986 /*
987  * running in raid5d, where reclaim could wait for raid5d too (when it flushes
988  * data from log to raid disks), so we shouldn't wait for reclaim here
989  */
r5l_write_stripe(struct r5l_log * log,struct stripe_head * sh)990 int r5l_write_stripe(struct r5l_log *log, struct stripe_head *sh)
991 {
992 	struct r5conf *conf = sh->raid_conf;
993 	int write_disks = 0;
994 	int data_pages, parity_pages;
995 	int reserve;
996 	int i;
997 	int ret = 0;
998 	bool wake_reclaim = false;
999 
1000 	if (!log)
1001 		return -EAGAIN;
1002 	/* Don't support stripe batch */
1003 	if (sh->log_io || !test_bit(R5_Wantwrite, &sh->dev[sh->pd_idx].flags) ||
1004 	    test_bit(STRIPE_SYNCING, &sh->state)) {
1005 		/* the stripe is written to log, we start writing it to raid */
1006 		clear_bit(STRIPE_LOG_TRAPPED, &sh->state);
1007 		return -EAGAIN;
1008 	}
1009 
1010 	WARN_ON(test_bit(STRIPE_R5C_CACHING, &sh->state));
1011 
1012 	for (i = 0; i < sh->disks; i++) {
1013 		void *addr;
1014 
1015 		if (!test_bit(R5_Wantwrite, &sh->dev[i].flags) ||
1016 		    test_bit(R5_InJournal, &sh->dev[i].flags))
1017 			continue;
1018 
1019 		write_disks++;
1020 		/* checksum is already calculated in last run */
1021 		if (test_bit(STRIPE_LOG_TRAPPED, &sh->state))
1022 			continue;
1023 		addr = kmap_atomic(sh->dev[i].page);
1024 		sh->dev[i].log_checksum = crc32c_le(log->uuid_checksum,
1025 						    addr, PAGE_SIZE);
1026 		kunmap_atomic(addr);
1027 	}
1028 	parity_pages = 1 + !!(sh->qd_idx >= 0);
1029 	data_pages = write_disks - parity_pages;
1030 
1031 	set_bit(STRIPE_LOG_TRAPPED, &sh->state);
1032 	/*
1033 	 * The stripe must enter state machine again to finish the write, so
1034 	 * don't delay.
1035 	 */
1036 	clear_bit(STRIPE_DELAYED, &sh->state);
1037 	atomic_inc(&sh->count);
1038 
1039 	mutex_lock(&log->io_mutex);
1040 	/* meta + data */
1041 	reserve = (1 + write_disks) << (PAGE_SHIFT - 9);
1042 
1043 	if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) {
1044 		if (!r5l_has_free_space(log, reserve)) {
1045 			r5l_add_no_space_stripe(log, sh);
1046 			wake_reclaim = true;
1047 		} else {
1048 			ret = r5l_log_stripe(log, sh, data_pages, parity_pages);
1049 			if (ret) {
1050 				spin_lock_irq(&log->io_list_lock);
1051 				list_add_tail(&sh->log_list,
1052 					      &log->no_mem_stripes);
1053 				spin_unlock_irq(&log->io_list_lock);
1054 			}
1055 		}
1056 	} else {  /* R5C_JOURNAL_MODE_WRITE_BACK */
1057 		/*
1058 		 * log space critical, do not process stripes that are
1059 		 * not in cache yet (sh->log_start == MaxSector).
1060 		 */
1061 		if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state) &&
1062 		    sh->log_start == MaxSector) {
1063 			r5l_add_no_space_stripe(log, sh);
1064 			wake_reclaim = true;
1065 			reserve = 0;
1066 		} else if (!r5l_has_free_space(log, reserve)) {
1067 			if (sh->log_start == log->last_checkpoint)
1068 				BUG();
1069 			else
1070 				r5l_add_no_space_stripe(log, sh);
1071 		} else {
1072 			ret = r5l_log_stripe(log, sh, data_pages, parity_pages);
1073 			if (ret) {
1074 				spin_lock_irq(&log->io_list_lock);
1075 				list_add_tail(&sh->log_list,
1076 					      &log->no_mem_stripes);
1077 				spin_unlock_irq(&log->io_list_lock);
1078 			}
1079 		}
1080 	}
1081 
1082 	mutex_unlock(&log->io_mutex);
1083 	if (wake_reclaim)
1084 		r5l_wake_reclaim(log, reserve);
1085 	return 0;
1086 }
1087 
r5l_write_stripe_run(struct r5l_log * log)1088 void r5l_write_stripe_run(struct r5l_log *log)
1089 {
1090 	if (!log)
1091 		return;
1092 	mutex_lock(&log->io_mutex);
1093 	r5l_submit_current_io(log);
1094 	mutex_unlock(&log->io_mutex);
1095 }
1096 
r5l_handle_flush_request(struct r5l_log * log,struct bio * bio)1097 int r5l_handle_flush_request(struct r5l_log *log, struct bio *bio)
1098 {
1099 	if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) {
1100 		/*
1101 		 * in write through (journal only)
1102 		 * we flush log disk cache first, then write stripe data to
1103 		 * raid disks. So if bio is finished, the log disk cache is
1104 		 * flushed already. The recovery guarantees we can recovery
1105 		 * the bio from log disk, so we don't need to flush again
1106 		 */
1107 		if (bio->bi_iter.bi_size == 0) {
1108 			bio_endio(bio);
1109 			return 0;
1110 		}
1111 		bio->bi_opf &= ~REQ_PREFLUSH;
1112 	} else {
1113 		/* write back (with cache) */
1114 		if (bio->bi_iter.bi_size == 0) {
1115 			mutex_lock(&log->io_mutex);
1116 			r5l_get_meta(log, 0);
1117 			bio_list_add(&log->current_io->flush_barriers, bio);
1118 			log->current_io->has_flush = 1;
1119 			log->current_io->has_null_flush = 1;
1120 			atomic_inc(&log->current_io->pending_stripe);
1121 			r5l_submit_current_io(log);
1122 			mutex_unlock(&log->io_mutex);
1123 			return 0;
1124 		}
1125 	}
1126 	return -EAGAIN;
1127 }
1128 
1129 /* This will run after log space is reclaimed */
r5l_run_no_space_stripes(struct r5l_log * log)1130 static void r5l_run_no_space_stripes(struct r5l_log *log)
1131 {
1132 	struct stripe_head *sh;
1133 
1134 	spin_lock(&log->no_space_stripes_lock);
1135 	while (!list_empty(&log->no_space_stripes)) {
1136 		sh = list_first_entry(&log->no_space_stripes,
1137 				      struct stripe_head, log_list);
1138 		list_del_init(&sh->log_list);
1139 		set_bit(STRIPE_HANDLE, &sh->state);
1140 		raid5_release_stripe(sh);
1141 	}
1142 	spin_unlock(&log->no_space_stripes_lock);
1143 }
1144 
1145 /*
1146  * calculate new last_checkpoint
1147  * for write through mode, returns log->next_checkpoint
1148  * for write back, returns log_start of first sh in stripe_in_journal_list
1149  */
r5c_calculate_new_cp(struct r5conf * conf)1150 static sector_t r5c_calculate_new_cp(struct r5conf *conf)
1151 {
1152 	struct stripe_head *sh;
1153 	struct r5l_log *log = READ_ONCE(conf->log);
1154 	sector_t new_cp;
1155 	unsigned long flags;
1156 
1157 	if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH)
1158 		return log->next_checkpoint;
1159 
1160 	spin_lock_irqsave(&log->stripe_in_journal_lock, flags);
1161 	if (list_empty(&log->stripe_in_journal_list)) {
1162 		/* all stripes flushed */
1163 		spin_unlock_irqrestore(&log->stripe_in_journal_lock, flags);
1164 		return log->next_checkpoint;
1165 	}
1166 	sh = list_first_entry(&log->stripe_in_journal_list,
1167 			      struct stripe_head, r5c);
1168 	new_cp = sh->log_start;
1169 	spin_unlock_irqrestore(&log->stripe_in_journal_lock, flags);
1170 	return new_cp;
1171 }
1172 
r5l_reclaimable_space(struct r5l_log * log)1173 static sector_t r5l_reclaimable_space(struct r5l_log *log)
1174 {
1175 	struct r5conf *conf = log->rdev->mddev->private;
1176 
1177 	return r5l_ring_distance(log, log->last_checkpoint,
1178 				 r5c_calculate_new_cp(conf));
1179 }
1180 
r5l_run_no_mem_stripe(struct r5l_log * log)1181 static void r5l_run_no_mem_stripe(struct r5l_log *log)
1182 {
1183 	struct stripe_head *sh;
1184 
1185 	lockdep_assert_held(&log->io_list_lock);
1186 
1187 	if (!list_empty(&log->no_mem_stripes)) {
1188 		sh = list_first_entry(&log->no_mem_stripes,
1189 				      struct stripe_head, log_list);
1190 		list_del_init(&sh->log_list);
1191 		set_bit(STRIPE_HANDLE, &sh->state);
1192 		raid5_release_stripe(sh);
1193 	}
1194 }
1195 
r5l_complete_finished_ios(struct r5l_log * log)1196 static bool r5l_complete_finished_ios(struct r5l_log *log)
1197 {
1198 	struct r5l_io_unit *io, *next;
1199 	bool found = false;
1200 
1201 	lockdep_assert_held(&log->io_list_lock);
1202 
1203 	list_for_each_entry_safe(io, next, &log->finished_ios, log_sibling) {
1204 		/* don't change list order */
1205 		if (io->state < IO_UNIT_STRIPE_END)
1206 			break;
1207 
1208 		log->next_checkpoint = io->log_start;
1209 
1210 		list_del(&io->log_sibling);
1211 		mempool_free(io, &log->io_pool);
1212 		r5l_run_no_mem_stripe(log);
1213 
1214 		found = true;
1215 	}
1216 
1217 	return found;
1218 }
1219 
__r5l_stripe_write_finished(struct r5l_io_unit * io)1220 static void __r5l_stripe_write_finished(struct r5l_io_unit *io)
1221 {
1222 	struct r5l_log *log = io->log;
1223 	struct r5conf *conf = log->rdev->mddev->private;
1224 	unsigned long flags;
1225 
1226 	spin_lock_irqsave(&log->io_list_lock, flags);
1227 	__r5l_set_io_unit_state(io, IO_UNIT_STRIPE_END);
1228 
1229 	if (!r5l_complete_finished_ios(log)) {
1230 		spin_unlock_irqrestore(&log->io_list_lock, flags);
1231 		return;
1232 	}
1233 
1234 	if (r5l_reclaimable_space(log) > log->max_free_space ||
1235 	    test_bit(R5C_LOG_TIGHT, &conf->cache_state))
1236 		r5l_wake_reclaim(log, 0);
1237 
1238 	spin_unlock_irqrestore(&log->io_list_lock, flags);
1239 	wake_up(&log->iounit_wait);
1240 }
1241 
r5l_stripe_write_finished(struct stripe_head * sh)1242 void r5l_stripe_write_finished(struct stripe_head *sh)
1243 {
1244 	struct r5l_io_unit *io;
1245 
1246 	io = sh->log_io;
1247 	sh->log_io = NULL;
1248 
1249 	if (io && atomic_dec_and_test(&io->pending_stripe))
1250 		__r5l_stripe_write_finished(io);
1251 }
1252 
r5l_log_flush_endio(struct bio * bio)1253 static void r5l_log_flush_endio(struct bio *bio)
1254 {
1255 	struct r5l_log *log = container_of(bio, struct r5l_log,
1256 		flush_bio);
1257 	unsigned long flags;
1258 	struct r5l_io_unit *io;
1259 
1260 	if (bio->bi_status)
1261 		md_error(log->rdev->mddev, log->rdev);
1262 	bio_uninit(bio);
1263 
1264 	spin_lock_irqsave(&log->io_list_lock, flags);
1265 	list_for_each_entry(io, &log->flushing_ios, log_sibling)
1266 		r5l_io_run_stripes(io);
1267 	list_splice_tail_init(&log->flushing_ios, &log->finished_ios);
1268 	spin_unlock_irqrestore(&log->io_list_lock, flags);
1269 }
1270 
1271 /*
1272  * Starting dispatch IO to raid.
1273  * io_unit(meta) consists of a log. There is one situation we want to avoid. A
1274  * broken meta in the middle of a log causes recovery can't find meta at the
1275  * head of log. If operations require meta at the head persistent in log, we
1276  * must make sure meta before it persistent in log too. A case is:
1277  *
1278  * stripe data/parity is in log, we start write stripe to raid disks. stripe
1279  * data/parity must be persistent in log before we do the write to raid disks.
1280  *
1281  * The solution is we restrictly maintain io_unit list order. In this case, we
1282  * only write stripes of an io_unit to raid disks till the io_unit is the first
1283  * one whose data/parity is in log.
1284  */
r5l_flush_stripe_to_raid(struct r5l_log * log)1285 void r5l_flush_stripe_to_raid(struct r5l_log *log)
1286 {
1287 	bool do_flush;
1288 
1289 	if (!log || !log->need_cache_flush)
1290 		return;
1291 
1292 	spin_lock_irq(&log->io_list_lock);
1293 	/* flush bio is running */
1294 	if (!list_empty(&log->flushing_ios)) {
1295 		spin_unlock_irq(&log->io_list_lock);
1296 		return;
1297 	}
1298 	list_splice_tail_init(&log->io_end_ios, &log->flushing_ios);
1299 	do_flush = !list_empty(&log->flushing_ios);
1300 	spin_unlock_irq(&log->io_list_lock);
1301 
1302 	if (!do_flush)
1303 		return;
1304 	bio_init(&log->flush_bio, log->rdev->bdev, NULL, 0,
1305 		  REQ_OP_WRITE | REQ_PREFLUSH);
1306 	log->flush_bio.bi_end_io = r5l_log_flush_endio;
1307 	submit_bio(&log->flush_bio);
1308 }
1309 
1310 static void r5l_write_super(struct r5l_log *log, sector_t cp);
r5l_write_super_and_discard_space(struct r5l_log * log,sector_t end)1311 static void r5l_write_super_and_discard_space(struct r5l_log *log,
1312 	sector_t end)
1313 {
1314 	struct block_device *bdev = log->rdev->bdev;
1315 	struct mddev *mddev;
1316 
1317 	r5l_write_super(log, end);
1318 
1319 	if (!bdev_max_discard_sectors(bdev))
1320 		return;
1321 
1322 	mddev = log->rdev->mddev;
1323 	/*
1324 	 * Discard could zero data, so before discard we must make sure
1325 	 * superblock is updated to new log tail. Updating superblock (either
1326 	 * directly call md_update_sb() or depend on md thread) must hold
1327 	 * reconfig mutex. On the other hand, raid5_quiesce is called with
1328 	 * reconfig_mutex hold. The first step of raid5_quiesce() is waiting
1329 	 * for all IO finish, hence waiting for reclaim thread, while reclaim
1330 	 * thread is calling this function and waiting for reconfig mutex. So
1331 	 * there is a deadlock. We workaround this issue with a trylock.
1332 	 * FIXME: we could miss discard if we can't take reconfig mutex
1333 	 */
1334 	set_mask_bits(&mddev->sb_flags, 0,
1335 		BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1336 	if (!mddev_trylock(mddev))
1337 		return;
1338 	md_update_sb(mddev, 1);
1339 	mddev_unlock(mddev);
1340 
1341 	/* discard IO error really doesn't matter, ignore it */
1342 	if (log->last_checkpoint < end) {
1343 		blkdev_issue_discard(bdev,
1344 				log->last_checkpoint + log->rdev->data_offset,
1345 				end - log->last_checkpoint, GFP_NOIO);
1346 	} else {
1347 		blkdev_issue_discard(bdev,
1348 				log->last_checkpoint + log->rdev->data_offset,
1349 				log->device_size - log->last_checkpoint,
1350 				GFP_NOIO);
1351 		blkdev_issue_discard(bdev, log->rdev->data_offset, end,
1352 				GFP_NOIO);
1353 	}
1354 }
1355 
1356 /*
1357  * r5c_flush_stripe moves stripe from cached list to handle_list. When called,
1358  * the stripe must be on r5c_cached_full_stripes or r5c_cached_partial_stripes.
1359  *
1360  * must hold conf->device_lock
1361  */
r5c_flush_stripe(struct r5conf * conf,struct stripe_head * sh)1362 static void r5c_flush_stripe(struct r5conf *conf, struct stripe_head *sh)
1363 {
1364 	BUG_ON(list_empty(&sh->lru));
1365 	BUG_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state));
1366 	BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
1367 
1368 	/*
1369 	 * The stripe is not ON_RELEASE_LIST, so it is safe to call
1370 	 * raid5_release_stripe() while holding conf->device_lock
1371 	 */
1372 	BUG_ON(test_bit(STRIPE_ON_RELEASE_LIST, &sh->state));
1373 	lockdep_assert_held(&conf->device_lock);
1374 
1375 	list_del_init(&sh->lru);
1376 	atomic_inc(&sh->count);
1377 
1378 	set_bit(STRIPE_HANDLE, &sh->state);
1379 	atomic_inc(&conf->active_stripes);
1380 	r5c_make_stripe_write_out(sh);
1381 
1382 	if (test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state))
1383 		atomic_inc(&conf->r5c_flushing_partial_stripes);
1384 	else
1385 		atomic_inc(&conf->r5c_flushing_full_stripes);
1386 	raid5_release_stripe(sh);
1387 }
1388 
1389 /*
1390  * if num == 0, flush all full stripes
1391  * if num > 0, flush all full stripes. If less than num full stripes are
1392  *             flushed, flush some partial stripes until totally num stripes are
1393  *             flushed or there is no more cached stripes.
1394  */
r5c_flush_cache(struct r5conf * conf,int num)1395 void r5c_flush_cache(struct r5conf *conf, int num)
1396 {
1397 	int count;
1398 	struct stripe_head *sh, *next;
1399 
1400 	lockdep_assert_held(&conf->device_lock);
1401 	if (!READ_ONCE(conf->log))
1402 		return;
1403 
1404 	count = 0;
1405 	list_for_each_entry_safe(sh, next, &conf->r5c_full_stripe_list, lru) {
1406 		r5c_flush_stripe(conf, sh);
1407 		count++;
1408 	}
1409 
1410 	if (count >= num)
1411 		return;
1412 	list_for_each_entry_safe(sh, next,
1413 				 &conf->r5c_partial_stripe_list, lru) {
1414 		r5c_flush_stripe(conf, sh);
1415 		if (++count >= num)
1416 			break;
1417 	}
1418 }
1419 
r5c_do_reclaim(struct r5conf * conf)1420 static void r5c_do_reclaim(struct r5conf *conf)
1421 {
1422 	struct r5l_log *log = READ_ONCE(conf->log);
1423 	struct stripe_head *sh;
1424 	int count = 0;
1425 	unsigned long flags;
1426 	int total_cached;
1427 	int stripes_to_flush;
1428 	int flushing_partial, flushing_full;
1429 
1430 	if (!r5c_is_writeback(log))
1431 		return;
1432 
1433 	flushing_partial = atomic_read(&conf->r5c_flushing_partial_stripes);
1434 	flushing_full = atomic_read(&conf->r5c_flushing_full_stripes);
1435 	total_cached = atomic_read(&conf->r5c_cached_partial_stripes) +
1436 		atomic_read(&conf->r5c_cached_full_stripes) -
1437 		flushing_full - flushing_partial;
1438 
1439 	if (total_cached > conf->min_nr_stripes * 3 / 4 ||
1440 	    atomic_read(&conf->empty_inactive_list_nr) > 0)
1441 		/*
1442 		 * if stripe cache pressure high, flush all full stripes and
1443 		 * some partial stripes
1444 		 */
1445 		stripes_to_flush = R5C_RECLAIM_STRIPE_GROUP;
1446 	else if (total_cached > conf->min_nr_stripes * 1 / 2 ||
1447 		 atomic_read(&conf->r5c_cached_full_stripes) - flushing_full >
1448 		 R5C_FULL_STRIPE_FLUSH_BATCH(conf))
1449 		/*
1450 		 * if stripe cache pressure moderate, or if there is many full
1451 		 * stripes,flush all full stripes
1452 		 */
1453 		stripes_to_flush = 0;
1454 	else
1455 		/* no need to flush */
1456 		stripes_to_flush = -1;
1457 
1458 	if (stripes_to_flush >= 0) {
1459 		spin_lock_irqsave(&conf->device_lock, flags);
1460 		r5c_flush_cache(conf, stripes_to_flush);
1461 		spin_unlock_irqrestore(&conf->device_lock, flags);
1462 	}
1463 
1464 	/* if log space is tight, flush stripes on stripe_in_journal_list */
1465 	if (test_bit(R5C_LOG_TIGHT, &conf->cache_state)) {
1466 		spin_lock_irqsave(&log->stripe_in_journal_lock, flags);
1467 		spin_lock(&conf->device_lock);
1468 		list_for_each_entry(sh, &log->stripe_in_journal_list, r5c) {
1469 			/*
1470 			 * stripes on stripe_in_journal_list could be in any
1471 			 * state of the stripe_cache state machine. In this
1472 			 * case, we only want to flush stripe on
1473 			 * r5c_cached_full/partial_stripes. The following
1474 			 * condition makes sure the stripe is on one of the
1475 			 * two lists.
1476 			 */
1477 			if (!list_empty(&sh->lru) &&
1478 			    !test_bit(STRIPE_HANDLE, &sh->state) &&
1479 			    atomic_read(&sh->count) == 0) {
1480 				r5c_flush_stripe(conf, sh);
1481 				if (count++ >= R5C_RECLAIM_STRIPE_GROUP)
1482 					break;
1483 			}
1484 		}
1485 		spin_unlock(&conf->device_lock);
1486 		spin_unlock_irqrestore(&log->stripe_in_journal_lock, flags);
1487 	}
1488 
1489 	if (!test_bit(R5C_LOG_CRITICAL, &conf->cache_state))
1490 		r5l_run_no_space_stripes(log);
1491 
1492 	md_wakeup_thread(conf->mddev->thread);
1493 }
1494 
r5l_do_reclaim(struct r5l_log * log)1495 static void r5l_do_reclaim(struct r5l_log *log)
1496 {
1497 	struct r5conf *conf = log->rdev->mddev->private;
1498 	sector_t reclaim_target = xchg(&log->reclaim_target, 0);
1499 	sector_t reclaimable;
1500 	sector_t next_checkpoint;
1501 	bool write_super;
1502 
1503 	spin_lock_irq(&log->io_list_lock);
1504 	write_super = r5l_reclaimable_space(log) > log->max_free_space ||
1505 		reclaim_target != 0 || !list_empty(&log->no_space_stripes);
1506 	/*
1507 	 * move proper io_unit to reclaim list. We should not change the order.
1508 	 * reclaimable/unreclaimable io_unit can be mixed in the list, we
1509 	 * shouldn't reuse space of an unreclaimable io_unit
1510 	 */
1511 	while (1) {
1512 		reclaimable = r5l_reclaimable_space(log);
1513 		if (reclaimable >= reclaim_target ||
1514 		    (list_empty(&log->running_ios) &&
1515 		     list_empty(&log->io_end_ios) &&
1516 		     list_empty(&log->flushing_ios) &&
1517 		     list_empty(&log->finished_ios)))
1518 			break;
1519 
1520 		md_wakeup_thread(log->rdev->mddev->thread);
1521 		wait_event_lock_irq(log->iounit_wait,
1522 				    r5l_reclaimable_space(log) > reclaimable,
1523 				    log->io_list_lock);
1524 	}
1525 
1526 	next_checkpoint = r5c_calculate_new_cp(conf);
1527 	spin_unlock_irq(&log->io_list_lock);
1528 
1529 	if (reclaimable == 0 || !write_super)
1530 		return;
1531 
1532 	/*
1533 	 * write_super will flush cache of each raid disk. We must write super
1534 	 * here, because the log area might be reused soon and we don't want to
1535 	 * confuse recovery
1536 	 */
1537 	r5l_write_super_and_discard_space(log, next_checkpoint);
1538 
1539 	mutex_lock(&log->io_mutex);
1540 	log->last_checkpoint = next_checkpoint;
1541 	r5c_update_log_state(log);
1542 	mutex_unlock(&log->io_mutex);
1543 
1544 	r5l_run_no_space_stripes(log);
1545 }
1546 
r5l_reclaim_thread(struct md_thread * thread)1547 static void r5l_reclaim_thread(struct md_thread *thread)
1548 {
1549 	struct mddev *mddev = thread->mddev;
1550 	struct r5conf *conf = mddev->private;
1551 	struct r5l_log *log = READ_ONCE(conf->log);
1552 
1553 	if (!log)
1554 		return;
1555 	r5c_do_reclaim(conf);
1556 	r5l_do_reclaim(log);
1557 }
1558 
r5l_wake_reclaim(struct r5l_log * log,sector_t space)1559 void r5l_wake_reclaim(struct r5l_log *log, sector_t space)
1560 {
1561 	unsigned long target;
1562 	unsigned long new = (unsigned long)space; /* overflow in theory */
1563 
1564 	if (!log)
1565 		return;
1566 
1567 	target = READ_ONCE(log->reclaim_target);
1568 	do {
1569 		if (new < target)
1570 			return;
1571 	} while (!try_cmpxchg(&log->reclaim_target, &target, new));
1572 	md_wakeup_thread(log->reclaim_thread);
1573 }
1574 
r5l_quiesce(struct r5l_log * log,int quiesce)1575 void r5l_quiesce(struct r5l_log *log, int quiesce)
1576 {
1577 	struct mddev *mddev = log->rdev->mddev;
1578 	struct md_thread *thread = rcu_dereference_protected(
1579 		log->reclaim_thread, lockdep_is_held(&mddev->reconfig_mutex));
1580 
1581 	if (quiesce) {
1582 		/* make sure r5l_write_super_and_discard_space exits */
1583 		wake_up(&mddev->sb_wait);
1584 		kthread_park(thread->tsk);
1585 		r5l_wake_reclaim(log, MaxSector);
1586 		r5l_do_reclaim(log);
1587 	} else
1588 		kthread_unpark(thread->tsk);
1589 }
1590 
r5l_log_disk_error(struct r5conf * conf)1591 bool r5l_log_disk_error(struct r5conf *conf)
1592 {
1593 	struct r5l_log *log = READ_ONCE(conf->log);
1594 
1595 	/* don't allow write if journal disk is missing */
1596 	if (!log)
1597 		return test_bit(MD_HAS_JOURNAL, &conf->mddev->flags);
1598 	else
1599 		return test_bit(Faulty, &log->rdev->flags);
1600 }
1601 
1602 #define R5L_RECOVERY_PAGE_POOL_SIZE 256
1603 
1604 struct r5l_recovery_ctx {
1605 	struct page *meta_page;		/* current meta */
1606 	sector_t meta_total_blocks;	/* total size of current meta and data */
1607 	sector_t pos;			/* recovery position */
1608 	u64 seq;			/* recovery position seq */
1609 	int data_parity_stripes;	/* number of data_parity stripes */
1610 	int data_only_stripes;		/* number of data_only stripes */
1611 	struct list_head cached_list;
1612 
1613 	/*
1614 	 * read ahead page pool (ra_pool)
1615 	 * in recovery, log is read sequentially. It is not efficient to
1616 	 * read every page with sync_page_io(). The read ahead page pool
1617 	 * reads multiple pages with one IO, so further log read can
1618 	 * just copy data from the pool.
1619 	 */
1620 	struct page *ra_pool[R5L_RECOVERY_PAGE_POOL_SIZE];
1621 	struct bio_vec ra_bvec[R5L_RECOVERY_PAGE_POOL_SIZE];
1622 	sector_t pool_offset;	/* offset of first page in the pool */
1623 	int total_pages;	/* total allocated pages */
1624 	int valid_pages;	/* pages with valid data */
1625 };
1626 
r5l_recovery_allocate_ra_pool(struct r5l_log * log,struct r5l_recovery_ctx * ctx)1627 static int r5l_recovery_allocate_ra_pool(struct r5l_log *log,
1628 					    struct r5l_recovery_ctx *ctx)
1629 {
1630 	struct page *page;
1631 
1632 	ctx->valid_pages = 0;
1633 	ctx->total_pages = 0;
1634 	while (ctx->total_pages < R5L_RECOVERY_PAGE_POOL_SIZE) {
1635 		page = alloc_page(GFP_KERNEL);
1636 
1637 		if (!page)
1638 			break;
1639 		ctx->ra_pool[ctx->total_pages] = page;
1640 		ctx->total_pages += 1;
1641 	}
1642 
1643 	if (ctx->total_pages == 0)
1644 		return -ENOMEM;
1645 
1646 	ctx->pool_offset = 0;
1647 	return 0;
1648 }
1649 
r5l_recovery_free_ra_pool(struct r5l_log * log,struct r5l_recovery_ctx * ctx)1650 static void r5l_recovery_free_ra_pool(struct r5l_log *log,
1651 					struct r5l_recovery_ctx *ctx)
1652 {
1653 	int i;
1654 
1655 	for (i = 0; i < ctx->total_pages; ++i)
1656 		put_page(ctx->ra_pool[i]);
1657 }
1658 
1659 /*
1660  * fetch ctx->valid_pages pages from offset
1661  * In normal cases, ctx->valid_pages == ctx->total_pages after the call.
1662  * However, if the offset is close to the end of the journal device,
1663  * ctx->valid_pages could be smaller than ctx->total_pages
1664  */
r5l_recovery_fetch_ra_pool(struct r5l_log * log,struct r5l_recovery_ctx * ctx,sector_t offset)1665 static int r5l_recovery_fetch_ra_pool(struct r5l_log *log,
1666 				      struct r5l_recovery_ctx *ctx,
1667 				      sector_t offset)
1668 {
1669 	struct bio bio;
1670 	int ret;
1671 
1672 	bio_init(&bio, log->rdev->bdev, ctx->ra_bvec,
1673 		 R5L_RECOVERY_PAGE_POOL_SIZE, REQ_OP_READ);
1674 	bio.bi_iter.bi_sector = log->rdev->data_offset + offset;
1675 
1676 	ctx->valid_pages = 0;
1677 	ctx->pool_offset = offset;
1678 
1679 	while (ctx->valid_pages < ctx->total_pages) {
1680 		__bio_add_page(&bio, ctx->ra_pool[ctx->valid_pages], PAGE_SIZE,
1681 			       0);
1682 		ctx->valid_pages += 1;
1683 
1684 		offset = r5l_ring_add(log, offset, BLOCK_SECTORS);
1685 
1686 		if (offset == 0)  /* reached end of the device */
1687 			break;
1688 	}
1689 
1690 	ret = submit_bio_wait(&bio);
1691 	bio_uninit(&bio);
1692 	return ret;
1693 }
1694 
1695 /*
1696  * try read a page from the read ahead page pool, if the page is not in the
1697  * pool, call r5l_recovery_fetch_ra_pool
1698  */
r5l_recovery_read_page(struct r5l_log * log,struct r5l_recovery_ctx * ctx,struct page * page,sector_t offset)1699 static int r5l_recovery_read_page(struct r5l_log *log,
1700 				  struct r5l_recovery_ctx *ctx,
1701 				  struct page *page,
1702 				  sector_t offset)
1703 {
1704 	int ret;
1705 
1706 	if (offset < ctx->pool_offset ||
1707 	    offset >= ctx->pool_offset + ctx->valid_pages * BLOCK_SECTORS) {
1708 		ret = r5l_recovery_fetch_ra_pool(log, ctx, offset);
1709 		if (ret)
1710 			return ret;
1711 	}
1712 
1713 	BUG_ON(offset < ctx->pool_offset ||
1714 	       offset >= ctx->pool_offset + ctx->valid_pages * BLOCK_SECTORS);
1715 
1716 	memcpy(page_address(page),
1717 	       page_address(ctx->ra_pool[(offset - ctx->pool_offset) >>
1718 					 BLOCK_SECTOR_SHIFT]),
1719 	       PAGE_SIZE);
1720 	return 0;
1721 }
1722 
r5l_recovery_read_meta_block(struct r5l_log * log,struct r5l_recovery_ctx * ctx)1723 static int r5l_recovery_read_meta_block(struct r5l_log *log,
1724 					struct r5l_recovery_ctx *ctx)
1725 {
1726 	struct page *page = ctx->meta_page;
1727 	struct r5l_meta_block *mb;
1728 	u32 crc, stored_crc;
1729 	int ret;
1730 
1731 	ret = r5l_recovery_read_page(log, ctx, page, ctx->pos);
1732 	if (ret != 0)
1733 		return ret;
1734 
1735 	mb = page_address(page);
1736 	stored_crc = le32_to_cpu(mb->checksum);
1737 	mb->checksum = 0;
1738 
1739 	if (le32_to_cpu(mb->magic) != R5LOG_MAGIC ||
1740 	    le64_to_cpu(mb->seq) != ctx->seq ||
1741 	    mb->version != R5LOG_VERSION ||
1742 	    le64_to_cpu(mb->position) != ctx->pos)
1743 		return -EINVAL;
1744 
1745 	crc = crc32c_le(log->uuid_checksum, mb, PAGE_SIZE);
1746 	if (stored_crc != crc)
1747 		return -EINVAL;
1748 
1749 	if (le32_to_cpu(mb->meta_size) > PAGE_SIZE)
1750 		return -EINVAL;
1751 
1752 	ctx->meta_total_blocks = BLOCK_SECTORS;
1753 
1754 	return 0;
1755 }
1756 
1757 static void
r5l_recovery_create_empty_meta_block(struct r5l_log * log,struct page * page,sector_t pos,u64 seq)1758 r5l_recovery_create_empty_meta_block(struct r5l_log *log,
1759 				     struct page *page,
1760 				     sector_t pos, u64 seq)
1761 {
1762 	struct r5l_meta_block *mb;
1763 
1764 	mb = page_address(page);
1765 	clear_page(mb);
1766 	mb->magic = cpu_to_le32(R5LOG_MAGIC);
1767 	mb->version = R5LOG_VERSION;
1768 	mb->meta_size = cpu_to_le32(sizeof(struct r5l_meta_block));
1769 	mb->seq = cpu_to_le64(seq);
1770 	mb->position = cpu_to_le64(pos);
1771 }
1772 
r5l_log_write_empty_meta_block(struct r5l_log * log,sector_t pos,u64 seq)1773 static int r5l_log_write_empty_meta_block(struct r5l_log *log, sector_t pos,
1774 					  u64 seq)
1775 {
1776 	struct page *page;
1777 	struct r5l_meta_block *mb;
1778 
1779 	page = alloc_page(GFP_KERNEL);
1780 	if (!page)
1781 		return -ENOMEM;
1782 	r5l_recovery_create_empty_meta_block(log, page, pos, seq);
1783 	mb = page_address(page);
1784 	mb->checksum = cpu_to_le32(crc32c_le(log->uuid_checksum,
1785 					     mb, PAGE_SIZE));
1786 	if (!sync_page_io(log->rdev, pos, PAGE_SIZE, page, REQ_OP_WRITE |
1787 			  REQ_SYNC | REQ_FUA, false)) {
1788 		__free_page(page);
1789 		return -EIO;
1790 	}
1791 	__free_page(page);
1792 	return 0;
1793 }
1794 
1795 /*
1796  * r5l_recovery_load_data and r5l_recovery_load_parity uses flag R5_Wantwrite
1797  * to mark valid (potentially not flushed) data in the journal.
1798  *
1799  * We already verified checksum in r5l_recovery_verify_data_checksum_for_mb,
1800  * so there should not be any mismatch here.
1801  */
r5l_recovery_load_data(struct r5l_log * log,struct stripe_head * sh,struct r5l_recovery_ctx * ctx,struct r5l_payload_data_parity * payload,sector_t log_offset)1802 static void r5l_recovery_load_data(struct r5l_log *log,
1803 				   struct stripe_head *sh,
1804 				   struct r5l_recovery_ctx *ctx,
1805 				   struct r5l_payload_data_parity *payload,
1806 				   sector_t log_offset)
1807 {
1808 	struct mddev *mddev = log->rdev->mddev;
1809 	struct r5conf *conf = mddev->private;
1810 	int dd_idx;
1811 
1812 	raid5_compute_sector(conf,
1813 			     le64_to_cpu(payload->location), 0,
1814 			     &dd_idx, sh);
1815 	r5l_recovery_read_page(log, ctx, sh->dev[dd_idx].page, log_offset);
1816 	sh->dev[dd_idx].log_checksum =
1817 		le32_to_cpu(payload->checksum[0]);
1818 	ctx->meta_total_blocks += BLOCK_SECTORS;
1819 
1820 	set_bit(R5_Wantwrite, &sh->dev[dd_idx].flags);
1821 	set_bit(STRIPE_R5C_CACHING, &sh->state);
1822 }
1823 
r5l_recovery_load_parity(struct r5l_log * log,struct stripe_head * sh,struct r5l_recovery_ctx * ctx,struct r5l_payload_data_parity * payload,sector_t log_offset)1824 static void r5l_recovery_load_parity(struct r5l_log *log,
1825 				     struct stripe_head *sh,
1826 				     struct r5l_recovery_ctx *ctx,
1827 				     struct r5l_payload_data_parity *payload,
1828 				     sector_t log_offset)
1829 {
1830 	struct mddev *mddev = log->rdev->mddev;
1831 	struct r5conf *conf = mddev->private;
1832 
1833 	ctx->meta_total_blocks += BLOCK_SECTORS * conf->max_degraded;
1834 	r5l_recovery_read_page(log, ctx, sh->dev[sh->pd_idx].page, log_offset);
1835 	sh->dev[sh->pd_idx].log_checksum =
1836 		le32_to_cpu(payload->checksum[0]);
1837 	set_bit(R5_Wantwrite, &sh->dev[sh->pd_idx].flags);
1838 
1839 	if (sh->qd_idx >= 0) {
1840 		r5l_recovery_read_page(
1841 			log, ctx, sh->dev[sh->qd_idx].page,
1842 			r5l_ring_add(log, log_offset, BLOCK_SECTORS));
1843 		sh->dev[sh->qd_idx].log_checksum =
1844 			le32_to_cpu(payload->checksum[1]);
1845 		set_bit(R5_Wantwrite, &sh->dev[sh->qd_idx].flags);
1846 	}
1847 	clear_bit(STRIPE_R5C_CACHING, &sh->state);
1848 }
1849 
r5l_recovery_reset_stripe(struct stripe_head * sh)1850 static void r5l_recovery_reset_stripe(struct stripe_head *sh)
1851 {
1852 	int i;
1853 
1854 	sh->state = 0;
1855 	sh->log_start = MaxSector;
1856 	for (i = sh->disks; i--; )
1857 		sh->dev[i].flags = 0;
1858 }
1859 
1860 static void
r5l_recovery_replay_one_stripe(struct r5conf * conf,struct stripe_head * sh,struct r5l_recovery_ctx * ctx)1861 r5l_recovery_replay_one_stripe(struct r5conf *conf,
1862 			       struct stripe_head *sh,
1863 			       struct r5l_recovery_ctx *ctx)
1864 {
1865 	struct md_rdev *rdev, *rrdev;
1866 	int disk_index;
1867 	int data_count = 0;
1868 
1869 	for (disk_index = 0; disk_index < sh->disks; disk_index++) {
1870 		if (!test_bit(R5_Wantwrite, &sh->dev[disk_index].flags))
1871 			continue;
1872 		if (disk_index == sh->qd_idx || disk_index == sh->pd_idx)
1873 			continue;
1874 		data_count++;
1875 	}
1876 
1877 	/*
1878 	 * stripes that only have parity must have been flushed
1879 	 * before the crash that we are now recovering from, so
1880 	 * there is nothing more to recovery.
1881 	 */
1882 	if (data_count == 0)
1883 		goto out;
1884 
1885 	for (disk_index = 0; disk_index < sh->disks; disk_index++) {
1886 		if (!test_bit(R5_Wantwrite, &sh->dev[disk_index].flags))
1887 			continue;
1888 
1889 		/* in case device is broken */
1890 		rcu_read_lock();
1891 		rdev = rcu_dereference(conf->disks[disk_index].rdev);
1892 		if (rdev) {
1893 			atomic_inc(&rdev->nr_pending);
1894 			rcu_read_unlock();
1895 			sync_page_io(rdev, sh->sector, PAGE_SIZE,
1896 				     sh->dev[disk_index].page, REQ_OP_WRITE,
1897 				     false);
1898 			rdev_dec_pending(rdev, rdev->mddev);
1899 			rcu_read_lock();
1900 		}
1901 		rrdev = rcu_dereference(conf->disks[disk_index].replacement);
1902 		if (rrdev) {
1903 			atomic_inc(&rrdev->nr_pending);
1904 			rcu_read_unlock();
1905 			sync_page_io(rrdev, sh->sector, PAGE_SIZE,
1906 				     sh->dev[disk_index].page, REQ_OP_WRITE,
1907 				     false);
1908 			rdev_dec_pending(rrdev, rrdev->mddev);
1909 			rcu_read_lock();
1910 		}
1911 		rcu_read_unlock();
1912 	}
1913 	ctx->data_parity_stripes++;
1914 out:
1915 	r5l_recovery_reset_stripe(sh);
1916 }
1917 
1918 static struct stripe_head *
r5c_recovery_alloc_stripe(struct r5conf * conf,sector_t stripe_sect,int noblock)1919 r5c_recovery_alloc_stripe(
1920 		struct r5conf *conf,
1921 		sector_t stripe_sect,
1922 		int noblock)
1923 {
1924 	struct stripe_head *sh;
1925 
1926 	sh = raid5_get_active_stripe(conf, NULL, stripe_sect,
1927 				     noblock ? R5_GAS_NOBLOCK : 0);
1928 	if (!sh)
1929 		return NULL;  /* no more stripe available */
1930 
1931 	r5l_recovery_reset_stripe(sh);
1932 
1933 	return sh;
1934 }
1935 
1936 static struct stripe_head *
r5c_recovery_lookup_stripe(struct list_head * list,sector_t sect)1937 r5c_recovery_lookup_stripe(struct list_head *list, sector_t sect)
1938 {
1939 	struct stripe_head *sh;
1940 
1941 	list_for_each_entry(sh, list, lru)
1942 		if (sh->sector == sect)
1943 			return sh;
1944 	return NULL;
1945 }
1946 
1947 static void
r5c_recovery_drop_stripes(struct list_head * cached_stripe_list,struct r5l_recovery_ctx * ctx)1948 r5c_recovery_drop_stripes(struct list_head *cached_stripe_list,
1949 			  struct r5l_recovery_ctx *ctx)
1950 {
1951 	struct stripe_head *sh, *next;
1952 
1953 	list_for_each_entry_safe(sh, next, cached_stripe_list, lru) {
1954 		r5l_recovery_reset_stripe(sh);
1955 		list_del_init(&sh->lru);
1956 		raid5_release_stripe(sh);
1957 	}
1958 }
1959 
1960 static void
r5c_recovery_replay_stripes(struct list_head * cached_stripe_list,struct r5l_recovery_ctx * ctx)1961 r5c_recovery_replay_stripes(struct list_head *cached_stripe_list,
1962 			    struct r5l_recovery_ctx *ctx)
1963 {
1964 	struct stripe_head *sh, *next;
1965 
1966 	list_for_each_entry_safe(sh, next, cached_stripe_list, lru)
1967 		if (!test_bit(STRIPE_R5C_CACHING, &sh->state)) {
1968 			r5l_recovery_replay_one_stripe(sh->raid_conf, sh, ctx);
1969 			list_del_init(&sh->lru);
1970 			raid5_release_stripe(sh);
1971 		}
1972 }
1973 
1974 /* if matches return 0; otherwise return -EINVAL */
1975 static int
r5l_recovery_verify_data_checksum(struct r5l_log * log,struct r5l_recovery_ctx * ctx,struct page * page,sector_t log_offset,__le32 log_checksum)1976 r5l_recovery_verify_data_checksum(struct r5l_log *log,
1977 				  struct r5l_recovery_ctx *ctx,
1978 				  struct page *page,
1979 				  sector_t log_offset, __le32 log_checksum)
1980 {
1981 	void *addr;
1982 	u32 checksum;
1983 
1984 	r5l_recovery_read_page(log, ctx, page, log_offset);
1985 	addr = kmap_atomic(page);
1986 	checksum = crc32c_le(log->uuid_checksum, addr, PAGE_SIZE);
1987 	kunmap_atomic(addr);
1988 	return (le32_to_cpu(log_checksum) == checksum) ? 0 : -EINVAL;
1989 }
1990 
1991 /*
1992  * before loading data to stripe cache, we need verify checksum for all data,
1993  * if there is mismatch for any data page, we drop all data in the mata block
1994  */
1995 static int
r5l_recovery_verify_data_checksum_for_mb(struct r5l_log * log,struct r5l_recovery_ctx * ctx)1996 r5l_recovery_verify_data_checksum_for_mb(struct r5l_log *log,
1997 					 struct r5l_recovery_ctx *ctx)
1998 {
1999 	struct mddev *mddev = log->rdev->mddev;
2000 	struct r5conf *conf = mddev->private;
2001 	struct r5l_meta_block *mb = page_address(ctx->meta_page);
2002 	sector_t mb_offset = sizeof(struct r5l_meta_block);
2003 	sector_t log_offset = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS);
2004 	struct page *page;
2005 	struct r5l_payload_data_parity *payload;
2006 	struct r5l_payload_flush *payload_flush;
2007 
2008 	page = alloc_page(GFP_KERNEL);
2009 	if (!page)
2010 		return -ENOMEM;
2011 
2012 	while (mb_offset < le32_to_cpu(mb->meta_size)) {
2013 		payload = (void *)mb + mb_offset;
2014 		payload_flush = (void *)mb + mb_offset;
2015 
2016 		if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_DATA) {
2017 			if (r5l_recovery_verify_data_checksum(
2018 				    log, ctx, page, log_offset,
2019 				    payload->checksum[0]) < 0)
2020 				goto mismatch;
2021 		} else if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_PARITY) {
2022 			if (r5l_recovery_verify_data_checksum(
2023 				    log, ctx, page, log_offset,
2024 				    payload->checksum[0]) < 0)
2025 				goto mismatch;
2026 			if (conf->max_degraded == 2 && /* q for RAID 6 */
2027 			    r5l_recovery_verify_data_checksum(
2028 				    log, ctx, page,
2029 				    r5l_ring_add(log, log_offset,
2030 						 BLOCK_SECTORS),
2031 				    payload->checksum[1]) < 0)
2032 				goto mismatch;
2033 		} else if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_FLUSH) {
2034 			/* nothing to do for R5LOG_PAYLOAD_FLUSH here */
2035 		} else /* not R5LOG_PAYLOAD_DATA/PARITY/FLUSH */
2036 			goto mismatch;
2037 
2038 		if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_FLUSH) {
2039 			mb_offset += sizeof(struct r5l_payload_flush) +
2040 				le32_to_cpu(payload_flush->size);
2041 		} else {
2042 			/* DATA or PARITY payload */
2043 			log_offset = r5l_ring_add(log, log_offset,
2044 						  le32_to_cpu(payload->size));
2045 			mb_offset += sizeof(struct r5l_payload_data_parity) +
2046 				sizeof(__le32) *
2047 				(le32_to_cpu(payload->size) >> (PAGE_SHIFT - 9));
2048 		}
2049 
2050 	}
2051 
2052 	put_page(page);
2053 	return 0;
2054 
2055 mismatch:
2056 	put_page(page);
2057 	return -EINVAL;
2058 }
2059 
2060 /*
2061  * Analyze all data/parity pages in one meta block
2062  * Returns:
2063  * 0 for success
2064  * -EINVAL for unknown playload type
2065  * -EAGAIN for checksum mismatch of data page
2066  * -ENOMEM for run out of memory (alloc_page failed or run out of stripes)
2067  */
2068 static int
r5c_recovery_analyze_meta_block(struct r5l_log * log,struct r5l_recovery_ctx * ctx,struct list_head * cached_stripe_list)2069 r5c_recovery_analyze_meta_block(struct r5l_log *log,
2070 				struct r5l_recovery_ctx *ctx,
2071 				struct list_head *cached_stripe_list)
2072 {
2073 	struct mddev *mddev = log->rdev->mddev;
2074 	struct r5conf *conf = mddev->private;
2075 	struct r5l_meta_block *mb;
2076 	struct r5l_payload_data_parity *payload;
2077 	struct r5l_payload_flush *payload_flush;
2078 	int mb_offset;
2079 	sector_t log_offset;
2080 	sector_t stripe_sect;
2081 	struct stripe_head *sh;
2082 	int ret;
2083 
2084 	/*
2085 	 * for mismatch in data blocks, we will drop all data in this mb, but
2086 	 * we will still read next mb for other data with FLUSH flag, as
2087 	 * io_unit could finish out of order.
2088 	 */
2089 	ret = r5l_recovery_verify_data_checksum_for_mb(log, ctx);
2090 	if (ret == -EINVAL)
2091 		return -EAGAIN;
2092 	else if (ret)
2093 		return ret;   /* -ENOMEM duo to alloc_page() failed */
2094 
2095 	mb = page_address(ctx->meta_page);
2096 	mb_offset = sizeof(struct r5l_meta_block);
2097 	log_offset = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS);
2098 
2099 	while (mb_offset < le32_to_cpu(mb->meta_size)) {
2100 		int dd;
2101 
2102 		payload = (void *)mb + mb_offset;
2103 		payload_flush = (void *)mb + mb_offset;
2104 
2105 		if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_FLUSH) {
2106 			int i, count;
2107 
2108 			count = le32_to_cpu(payload_flush->size) / sizeof(__le64);
2109 			for (i = 0; i < count; ++i) {
2110 				stripe_sect = le64_to_cpu(payload_flush->flush_stripes[i]);
2111 				sh = r5c_recovery_lookup_stripe(cached_stripe_list,
2112 								stripe_sect);
2113 				if (sh) {
2114 					WARN_ON(test_bit(STRIPE_R5C_CACHING, &sh->state));
2115 					r5l_recovery_reset_stripe(sh);
2116 					list_del_init(&sh->lru);
2117 					raid5_release_stripe(sh);
2118 				}
2119 			}
2120 
2121 			mb_offset += sizeof(struct r5l_payload_flush) +
2122 				le32_to_cpu(payload_flush->size);
2123 			continue;
2124 		}
2125 
2126 		/* DATA or PARITY payload */
2127 		stripe_sect = (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_DATA) ?
2128 			raid5_compute_sector(
2129 				conf, le64_to_cpu(payload->location), 0, &dd,
2130 				NULL)
2131 			: le64_to_cpu(payload->location);
2132 
2133 		sh = r5c_recovery_lookup_stripe(cached_stripe_list,
2134 						stripe_sect);
2135 
2136 		if (!sh) {
2137 			sh = r5c_recovery_alloc_stripe(conf, stripe_sect, 1);
2138 			/*
2139 			 * cannot get stripe from raid5_get_active_stripe
2140 			 * try replay some stripes
2141 			 */
2142 			if (!sh) {
2143 				r5c_recovery_replay_stripes(
2144 					cached_stripe_list, ctx);
2145 				sh = r5c_recovery_alloc_stripe(
2146 					conf, stripe_sect, 1);
2147 			}
2148 			if (!sh) {
2149 				int new_size = conf->min_nr_stripes * 2;
2150 				pr_debug("md/raid:%s: Increasing stripe cache size to %d to recovery data on journal.\n",
2151 					mdname(mddev),
2152 					new_size);
2153 				ret = raid5_set_cache_size(mddev, new_size);
2154 				if (conf->min_nr_stripes <= new_size / 2) {
2155 					pr_err("md/raid:%s: Cannot increase cache size, ret=%d, new_size=%d, min_nr_stripes=%d, max_nr_stripes=%d\n",
2156 						mdname(mddev),
2157 						ret,
2158 						new_size,
2159 						conf->min_nr_stripes,
2160 						conf->max_nr_stripes);
2161 					return -ENOMEM;
2162 				}
2163 				sh = r5c_recovery_alloc_stripe(
2164 					conf, stripe_sect, 0);
2165 			}
2166 			if (!sh) {
2167 				pr_err("md/raid:%s: Cannot get enough stripes due to memory pressure. Recovery failed.\n",
2168 					mdname(mddev));
2169 				return -ENOMEM;
2170 			}
2171 			list_add_tail(&sh->lru, cached_stripe_list);
2172 		}
2173 
2174 		if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_DATA) {
2175 			if (!test_bit(STRIPE_R5C_CACHING, &sh->state) &&
2176 			    test_bit(R5_Wantwrite, &sh->dev[sh->pd_idx].flags)) {
2177 				r5l_recovery_replay_one_stripe(conf, sh, ctx);
2178 				list_move_tail(&sh->lru, cached_stripe_list);
2179 			}
2180 			r5l_recovery_load_data(log, sh, ctx, payload,
2181 					       log_offset);
2182 		} else if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_PARITY)
2183 			r5l_recovery_load_parity(log, sh, ctx, payload,
2184 						 log_offset);
2185 		else
2186 			return -EINVAL;
2187 
2188 		log_offset = r5l_ring_add(log, log_offset,
2189 					  le32_to_cpu(payload->size));
2190 
2191 		mb_offset += sizeof(struct r5l_payload_data_parity) +
2192 			sizeof(__le32) *
2193 			(le32_to_cpu(payload->size) >> (PAGE_SHIFT - 9));
2194 	}
2195 
2196 	return 0;
2197 }
2198 
2199 /*
2200  * Load the stripe into cache. The stripe will be written out later by
2201  * the stripe cache state machine.
2202  */
r5c_recovery_load_one_stripe(struct r5l_log * log,struct stripe_head * sh)2203 static void r5c_recovery_load_one_stripe(struct r5l_log *log,
2204 					 struct stripe_head *sh)
2205 {
2206 	struct r5dev *dev;
2207 	int i;
2208 
2209 	for (i = sh->disks; i--; ) {
2210 		dev = sh->dev + i;
2211 		if (test_and_clear_bit(R5_Wantwrite, &dev->flags)) {
2212 			set_bit(R5_InJournal, &dev->flags);
2213 			set_bit(R5_UPTODATE, &dev->flags);
2214 		}
2215 	}
2216 }
2217 
2218 /*
2219  * Scan through the log for all to-be-flushed data
2220  *
2221  * For stripes with data and parity, namely Data-Parity stripe
2222  * (STRIPE_R5C_CACHING == 0), we simply replay all the writes.
2223  *
2224  * For stripes with only data, namely Data-Only stripe
2225  * (STRIPE_R5C_CACHING == 1), we load them to stripe cache state machine.
2226  *
2227  * For a stripe, if we see data after parity, we should discard all previous
2228  * data and parity for this stripe, as these data are already flushed to
2229  * the array.
2230  *
2231  * At the end of the scan, we return the new journal_tail, which points to
2232  * first data-only stripe on the journal device, or next invalid meta block.
2233  */
r5c_recovery_flush_log(struct r5l_log * log,struct r5l_recovery_ctx * ctx)2234 static int r5c_recovery_flush_log(struct r5l_log *log,
2235 				  struct r5l_recovery_ctx *ctx)
2236 {
2237 	struct stripe_head *sh;
2238 	int ret = 0;
2239 
2240 	/* scan through the log */
2241 	while (1) {
2242 		if (r5l_recovery_read_meta_block(log, ctx))
2243 			break;
2244 
2245 		ret = r5c_recovery_analyze_meta_block(log, ctx,
2246 						      &ctx->cached_list);
2247 		/*
2248 		 * -EAGAIN means mismatch in data block, in this case, we still
2249 		 * try scan the next metablock
2250 		 */
2251 		if (ret && ret != -EAGAIN)
2252 			break;   /* ret == -EINVAL or -ENOMEM */
2253 		ctx->seq++;
2254 		ctx->pos = r5l_ring_add(log, ctx->pos, ctx->meta_total_blocks);
2255 	}
2256 
2257 	if (ret == -ENOMEM) {
2258 		r5c_recovery_drop_stripes(&ctx->cached_list, ctx);
2259 		return ret;
2260 	}
2261 
2262 	/* replay data-parity stripes */
2263 	r5c_recovery_replay_stripes(&ctx->cached_list, ctx);
2264 
2265 	/* load data-only stripes to stripe cache */
2266 	list_for_each_entry(sh, &ctx->cached_list, lru) {
2267 		WARN_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state));
2268 		r5c_recovery_load_one_stripe(log, sh);
2269 		ctx->data_only_stripes++;
2270 	}
2271 
2272 	return 0;
2273 }
2274 
2275 /*
2276  * we did a recovery. Now ctx.pos points to an invalid meta block. New
2277  * log will start here. but we can't let superblock point to last valid
2278  * meta block. The log might looks like:
2279  * | meta 1| meta 2| meta 3|
2280  * meta 1 is valid, meta 2 is invalid. meta 3 could be valid. If
2281  * superblock points to meta 1, we write a new valid meta 2n.  if crash
2282  * happens again, new recovery will start from meta 1. Since meta 2n is
2283  * valid now, recovery will think meta 3 is valid, which is wrong.
2284  * The solution is we create a new meta in meta2 with its seq == meta
2285  * 1's seq + 10000 and let superblock points to meta2. The same recovery
2286  * will not think meta 3 is a valid meta, because its seq doesn't match
2287  */
2288 
2289 /*
2290  * Before recovery, the log looks like the following
2291  *
2292  *   ---------------------------------------------
2293  *   |           valid log        | invalid log  |
2294  *   ---------------------------------------------
2295  *   ^
2296  *   |- log->last_checkpoint
2297  *   |- log->last_cp_seq
2298  *
2299  * Now we scan through the log until we see invalid entry
2300  *
2301  *   ---------------------------------------------
2302  *   |           valid log        | invalid log  |
2303  *   ---------------------------------------------
2304  *   ^                            ^
2305  *   |- log->last_checkpoint      |- ctx->pos
2306  *   |- log->last_cp_seq          |- ctx->seq
2307  *
2308  * From this point, we need to increase seq number by 10 to avoid
2309  * confusing next recovery.
2310  *
2311  *   ---------------------------------------------
2312  *   |           valid log        | invalid log  |
2313  *   ---------------------------------------------
2314  *   ^                              ^
2315  *   |- log->last_checkpoint        |- ctx->pos+1
2316  *   |- log->last_cp_seq            |- ctx->seq+10001
2317  *
2318  * However, it is not safe to start the state machine yet, because data only
2319  * parities are not yet secured in RAID. To save these data only parities, we
2320  * rewrite them from seq+11.
2321  *
2322  *   -----------------------------------------------------------------
2323  *   |           valid log        | data only stripes | invalid log  |
2324  *   -----------------------------------------------------------------
2325  *   ^                                                ^
2326  *   |- log->last_checkpoint                          |- ctx->pos+n
2327  *   |- log->last_cp_seq                              |- ctx->seq+10000+n
2328  *
2329  * If failure happens again during this process, the recovery can safe start
2330  * again from log->last_checkpoint.
2331  *
2332  * Once data only stripes are rewritten to journal, we move log_tail
2333  *
2334  *   -----------------------------------------------------------------
2335  *   |     old log        |    data only stripes    | invalid log  |
2336  *   -----------------------------------------------------------------
2337  *                        ^                         ^
2338  *                        |- log->last_checkpoint   |- ctx->pos+n
2339  *                        |- log->last_cp_seq       |- ctx->seq+10000+n
2340  *
2341  * Then we can safely start the state machine. If failure happens from this
2342  * point on, the recovery will start from new log->last_checkpoint.
2343  */
2344 static int
r5c_recovery_rewrite_data_only_stripes(struct r5l_log * log,struct r5l_recovery_ctx * ctx)2345 r5c_recovery_rewrite_data_only_stripes(struct r5l_log *log,
2346 				       struct r5l_recovery_ctx *ctx)
2347 {
2348 	struct stripe_head *sh;
2349 	struct mddev *mddev = log->rdev->mddev;
2350 	struct page *page;
2351 	sector_t next_checkpoint = MaxSector;
2352 
2353 	page = alloc_page(GFP_KERNEL);
2354 	if (!page) {
2355 		pr_err("md/raid:%s: cannot allocate memory to rewrite data only stripes\n",
2356 		       mdname(mddev));
2357 		return -ENOMEM;
2358 	}
2359 
2360 	WARN_ON(list_empty(&ctx->cached_list));
2361 
2362 	list_for_each_entry(sh, &ctx->cached_list, lru) {
2363 		struct r5l_meta_block *mb;
2364 		int i;
2365 		int offset;
2366 		sector_t write_pos;
2367 
2368 		WARN_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state));
2369 		r5l_recovery_create_empty_meta_block(log, page,
2370 						     ctx->pos, ctx->seq);
2371 		mb = page_address(page);
2372 		offset = le32_to_cpu(mb->meta_size);
2373 		write_pos = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS);
2374 
2375 		for (i = sh->disks; i--; ) {
2376 			struct r5dev *dev = &sh->dev[i];
2377 			struct r5l_payload_data_parity *payload;
2378 			void *addr;
2379 
2380 			if (test_bit(R5_InJournal, &dev->flags)) {
2381 				payload = (void *)mb + offset;
2382 				payload->header.type = cpu_to_le16(
2383 					R5LOG_PAYLOAD_DATA);
2384 				payload->size = cpu_to_le32(BLOCK_SECTORS);
2385 				payload->location = cpu_to_le64(
2386 					raid5_compute_blocknr(sh, i, 0));
2387 				addr = kmap_atomic(dev->page);
2388 				payload->checksum[0] = cpu_to_le32(
2389 					crc32c_le(log->uuid_checksum, addr,
2390 						  PAGE_SIZE));
2391 				kunmap_atomic(addr);
2392 				sync_page_io(log->rdev, write_pos, PAGE_SIZE,
2393 					     dev->page, REQ_OP_WRITE, false);
2394 				write_pos = r5l_ring_add(log, write_pos,
2395 							 BLOCK_SECTORS);
2396 				offset += sizeof(__le32) +
2397 					sizeof(struct r5l_payload_data_parity);
2398 
2399 			}
2400 		}
2401 		mb->meta_size = cpu_to_le32(offset);
2402 		mb->checksum = cpu_to_le32(crc32c_le(log->uuid_checksum,
2403 						     mb, PAGE_SIZE));
2404 		sync_page_io(log->rdev, ctx->pos, PAGE_SIZE, page,
2405 			     REQ_OP_WRITE | REQ_SYNC | REQ_FUA, false);
2406 		sh->log_start = ctx->pos;
2407 		list_add_tail(&sh->r5c, &log->stripe_in_journal_list);
2408 		atomic_inc(&log->stripe_in_journal_count);
2409 		ctx->pos = write_pos;
2410 		ctx->seq += 1;
2411 		next_checkpoint = sh->log_start;
2412 	}
2413 	log->next_checkpoint = next_checkpoint;
2414 	__free_page(page);
2415 	return 0;
2416 }
2417 
r5c_recovery_flush_data_only_stripes(struct r5l_log * log,struct r5l_recovery_ctx * ctx)2418 static void r5c_recovery_flush_data_only_stripes(struct r5l_log *log,
2419 						 struct r5l_recovery_ctx *ctx)
2420 {
2421 	struct mddev *mddev = log->rdev->mddev;
2422 	struct r5conf *conf = mddev->private;
2423 	struct stripe_head *sh, *next;
2424 	bool cleared_pending = false;
2425 
2426 	if (ctx->data_only_stripes == 0)
2427 		return;
2428 
2429 	if (test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2430 		cleared_pending = true;
2431 		clear_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags);
2432 	}
2433 	log->r5c_journal_mode = R5C_JOURNAL_MODE_WRITE_BACK;
2434 
2435 	list_for_each_entry_safe(sh, next, &ctx->cached_list, lru) {
2436 		r5c_make_stripe_write_out(sh);
2437 		set_bit(STRIPE_HANDLE, &sh->state);
2438 		list_del_init(&sh->lru);
2439 		raid5_release_stripe(sh);
2440 	}
2441 
2442 	/* reuse conf->wait_for_quiescent in recovery */
2443 	wait_event(conf->wait_for_quiescent,
2444 		   atomic_read(&conf->active_stripes) == 0);
2445 
2446 	log->r5c_journal_mode = R5C_JOURNAL_MODE_WRITE_THROUGH;
2447 	if (cleared_pending)
2448 		set_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags);
2449 }
2450 
r5l_recovery_log(struct r5l_log * log)2451 static int r5l_recovery_log(struct r5l_log *log)
2452 {
2453 	struct mddev *mddev = log->rdev->mddev;
2454 	struct r5l_recovery_ctx *ctx;
2455 	int ret;
2456 	sector_t pos;
2457 
2458 	ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
2459 	if (!ctx)
2460 		return -ENOMEM;
2461 
2462 	ctx->pos = log->last_checkpoint;
2463 	ctx->seq = log->last_cp_seq;
2464 	INIT_LIST_HEAD(&ctx->cached_list);
2465 	ctx->meta_page = alloc_page(GFP_KERNEL);
2466 
2467 	if (!ctx->meta_page) {
2468 		ret =  -ENOMEM;
2469 		goto meta_page;
2470 	}
2471 
2472 	if (r5l_recovery_allocate_ra_pool(log, ctx) != 0) {
2473 		ret = -ENOMEM;
2474 		goto ra_pool;
2475 	}
2476 
2477 	ret = r5c_recovery_flush_log(log, ctx);
2478 
2479 	if (ret)
2480 		goto error;
2481 
2482 	pos = ctx->pos;
2483 	ctx->seq += 10000;
2484 
2485 	if ((ctx->data_only_stripes == 0) && (ctx->data_parity_stripes == 0))
2486 		pr_info("md/raid:%s: starting from clean shutdown\n",
2487 			 mdname(mddev));
2488 	else
2489 		pr_info("md/raid:%s: recovering %d data-only stripes and %d data-parity stripes\n",
2490 			 mdname(mddev), ctx->data_only_stripes,
2491 			 ctx->data_parity_stripes);
2492 
2493 	if (ctx->data_only_stripes == 0) {
2494 		log->next_checkpoint = ctx->pos;
2495 		r5l_log_write_empty_meta_block(log, ctx->pos, ctx->seq++);
2496 		ctx->pos = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS);
2497 	} else if (r5c_recovery_rewrite_data_only_stripes(log, ctx)) {
2498 		pr_err("md/raid:%s: failed to rewrite stripes to journal\n",
2499 		       mdname(mddev));
2500 		ret =  -EIO;
2501 		goto error;
2502 	}
2503 
2504 	log->log_start = ctx->pos;
2505 	log->seq = ctx->seq;
2506 	log->last_checkpoint = pos;
2507 	r5l_write_super(log, pos);
2508 
2509 	r5c_recovery_flush_data_only_stripes(log, ctx);
2510 	ret = 0;
2511 error:
2512 	r5l_recovery_free_ra_pool(log, ctx);
2513 ra_pool:
2514 	__free_page(ctx->meta_page);
2515 meta_page:
2516 	kfree(ctx);
2517 	return ret;
2518 }
2519 
r5l_write_super(struct r5l_log * log,sector_t cp)2520 static void r5l_write_super(struct r5l_log *log, sector_t cp)
2521 {
2522 	struct mddev *mddev = log->rdev->mddev;
2523 
2524 	log->rdev->journal_tail = cp;
2525 	set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
2526 }
2527 
r5c_journal_mode_show(struct mddev * mddev,char * page)2528 static ssize_t r5c_journal_mode_show(struct mddev *mddev, char *page)
2529 {
2530 	struct r5conf *conf;
2531 	int ret;
2532 
2533 	ret = mddev_lock(mddev);
2534 	if (ret)
2535 		return ret;
2536 
2537 	conf = mddev->private;
2538 	if (!conf || !conf->log)
2539 		goto out_unlock;
2540 
2541 	switch (conf->log->r5c_journal_mode) {
2542 	case R5C_JOURNAL_MODE_WRITE_THROUGH:
2543 		ret = snprintf(
2544 			page, PAGE_SIZE, "[%s] %s\n",
2545 			r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_THROUGH],
2546 			r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_BACK]);
2547 		break;
2548 	case R5C_JOURNAL_MODE_WRITE_BACK:
2549 		ret = snprintf(
2550 			page, PAGE_SIZE, "%s [%s]\n",
2551 			r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_THROUGH],
2552 			r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_BACK]);
2553 		break;
2554 	default:
2555 		ret = 0;
2556 	}
2557 
2558 out_unlock:
2559 	mddev_unlock(mddev);
2560 	return ret;
2561 }
2562 
2563 /*
2564  * Set journal cache mode on @mddev (external API initially needed by dm-raid).
2565  *
2566  * @mode as defined in 'enum r5c_journal_mode'.
2567  *
2568  */
r5c_journal_mode_set(struct mddev * mddev,int mode)2569 int r5c_journal_mode_set(struct mddev *mddev, int mode)
2570 {
2571 	struct r5conf *conf;
2572 
2573 	if (mode < R5C_JOURNAL_MODE_WRITE_THROUGH ||
2574 	    mode > R5C_JOURNAL_MODE_WRITE_BACK)
2575 		return -EINVAL;
2576 
2577 	conf = mddev->private;
2578 	if (!conf || !conf->log)
2579 		return -ENODEV;
2580 
2581 	if (raid5_calc_degraded(conf) > 0 &&
2582 	    mode == R5C_JOURNAL_MODE_WRITE_BACK)
2583 		return -EINVAL;
2584 
2585 	mddev_suspend(mddev);
2586 	conf->log->r5c_journal_mode = mode;
2587 	mddev_resume(mddev);
2588 
2589 	pr_debug("md/raid:%s: setting r5c cache mode to %d: %s\n",
2590 		 mdname(mddev), mode, r5c_journal_mode_str[mode]);
2591 	return 0;
2592 }
2593 EXPORT_SYMBOL(r5c_journal_mode_set);
2594 
r5c_journal_mode_store(struct mddev * mddev,const char * page,size_t length)2595 static ssize_t r5c_journal_mode_store(struct mddev *mddev,
2596 				      const char *page, size_t length)
2597 {
2598 	int mode = ARRAY_SIZE(r5c_journal_mode_str);
2599 	size_t len = length;
2600 	int ret;
2601 
2602 	if (len < 2)
2603 		return -EINVAL;
2604 
2605 	if (page[len - 1] == '\n')
2606 		len--;
2607 
2608 	while (mode--)
2609 		if (strlen(r5c_journal_mode_str[mode]) == len &&
2610 		    !strncmp(page, r5c_journal_mode_str[mode], len))
2611 			break;
2612 	ret = mddev_lock(mddev);
2613 	if (ret)
2614 		return ret;
2615 	ret = r5c_journal_mode_set(mddev, mode);
2616 	mddev_unlock(mddev);
2617 	return ret ?: length;
2618 }
2619 
2620 struct md_sysfs_entry
2621 r5c_journal_mode = __ATTR(journal_mode, 0644,
2622 			  r5c_journal_mode_show, r5c_journal_mode_store);
2623 
2624 /*
2625  * Try handle write operation in caching phase. This function should only
2626  * be called in write-back mode.
2627  *
2628  * If all outstanding writes can be handled in caching phase, returns 0
2629  * If writes requires write-out phase, call r5c_make_stripe_write_out()
2630  * and returns -EAGAIN
2631  */
r5c_try_caching_write(struct r5conf * conf,struct stripe_head * sh,struct stripe_head_state * s,int disks)2632 int r5c_try_caching_write(struct r5conf *conf,
2633 			  struct stripe_head *sh,
2634 			  struct stripe_head_state *s,
2635 			  int disks)
2636 {
2637 	struct r5l_log *log = READ_ONCE(conf->log);
2638 	int i;
2639 	struct r5dev *dev;
2640 	int to_cache = 0;
2641 	void __rcu **pslot;
2642 	sector_t tree_index;
2643 	int ret;
2644 	uintptr_t refcount;
2645 
2646 	BUG_ON(!r5c_is_writeback(log));
2647 
2648 	if (!test_bit(STRIPE_R5C_CACHING, &sh->state)) {
2649 		/*
2650 		 * There are two different scenarios here:
2651 		 *  1. The stripe has some data cached, and it is sent to
2652 		 *     write-out phase for reclaim
2653 		 *  2. The stripe is clean, and this is the first write
2654 		 *
2655 		 * For 1, return -EAGAIN, so we continue with
2656 		 * handle_stripe_dirtying().
2657 		 *
2658 		 * For 2, set STRIPE_R5C_CACHING and continue with caching
2659 		 * write.
2660 		 */
2661 
2662 		/* case 1: anything injournal or anything in written */
2663 		if (s->injournal > 0 || s->written > 0)
2664 			return -EAGAIN;
2665 		/* case 2 */
2666 		set_bit(STRIPE_R5C_CACHING, &sh->state);
2667 	}
2668 
2669 	/*
2670 	 * When run in degraded mode, array is set to write-through mode.
2671 	 * This check helps drain pending write safely in the transition to
2672 	 * write-through mode.
2673 	 *
2674 	 * When a stripe is syncing, the write is also handled in write
2675 	 * through mode.
2676 	 */
2677 	if (s->failed || test_bit(STRIPE_SYNCING, &sh->state)) {
2678 		r5c_make_stripe_write_out(sh);
2679 		return -EAGAIN;
2680 	}
2681 
2682 	for (i = disks; i--; ) {
2683 		dev = &sh->dev[i];
2684 		/* if non-overwrite, use writing-out phase */
2685 		if (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags) &&
2686 		    !test_bit(R5_InJournal, &dev->flags)) {
2687 			r5c_make_stripe_write_out(sh);
2688 			return -EAGAIN;
2689 		}
2690 	}
2691 
2692 	/* if the stripe is not counted in big_stripe_tree, add it now */
2693 	if (!test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state) &&
2694 	    !test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state)) {
2695 		tree_index = r5c_tree_index(conf, sh->sector);
2696 		spin_lock(&log->tree_lock);
2697 		pslot = radix_tree_lookup_slot(&log->big_stripe_tree,
2698 					       tree_index);
2699 		if (pslot) {
2700 			refcount = (uintptr_t)radix_tree_deref_slot_protected(
2701 				pslot, &log->tree_lock) >>
2702 				R5C_RADIX_COUNT_SHIFT;
2703 			radix_tree_replace_slot(
2704 				&log->big_stripe_tree, pslot,
2705 				(void *)((refcount + 1) << R5C_RADIX_COUNT_SHIFT));
2706 		} else {
2707 			/*
2708 			 * this radix_tree_insert can fail safely, so no
2709 			 * need to call radix_tree_preload()
2710 			 */
2711 			ret = radix_tree_insert(
2712 				&log->big_stripe_tree, tree_index,
2713 				(void *)(1 << R5C_RADIX_COUNT_SHIFT));
2714 			if (ret) {
2715 				spin_unlock(&log->tree_lock);
2716 				r5c_make_stripe_write_out(sh);
2717 				return -EAGAIN;
2718 			}
2719 		}
2720 		spin_unlock(&log->tree_lock);
2721 
2722 		/*
2723 		 * set STRIPE_R5C_PARTIAL_STRIPE, this shows the stripe is
2724 		 * counted in the radix tree
2725 		 */
2726 		set_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state);
2727 		atomic_inc(&conf->r5c_cached_partial_stripes);
2728 	}
2729 
2730 	for (i = disks; i--; ) {
2731 		dev = &sh->dev[i];
2732 		if (dev->towrite) {
2733 			set_bit(R5_Wantwrite, &dev->flags);
2734 			set_bit(R5_Wantdrain, &dev->flags);
2735 			set_bit(R5_LOCKED, &dev->flags);
2736 			to_cache++;
2737 		}
2738 	}
2739 
2740 	if (to_cache) {
2741 		set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2742 		/*
2743 		 * set STRIPE_LOG_TRAPPED, which triggers r5c_cache_data()
2744 		 * in ops_run_io(). STRIPE_LOG_TRAPPED will be cleared in
2745 		 * r5c_handle_data_cached()
2746 		 */
2747 		set_bit(STRIPE_LOG_TRAPPED, &sh->state);
2748 	}
2749 
2750 	return 0;
2751 }
2752 
2753 /*
2754  * free extra pages (orig_page) we allocated for prexor
2755  */
r5c_release_extra_page(struct stripe_head * sh)2756 void r5c_release_extra_page(struct stripe_head *sh)
2757 {
2758 	struct r5conf *conf = sh->raid_conf;
2759 	int i;
2760 	bool using_disk_info_extra_page;
2761 
2762 	using_disk_info_extra_page =
2763 		sh->dev[0].orig_page == conf->disks[0].extra_page;
2764 
2765 	for (i = sh->disks; i--; )
2766 		if (sh->dev[i].page != sh->dev[i].orig_page) {
2767 			struct page *p = sh->dev[i].orig_page;
2768 
2769 			sh->dev[i].orig_page = sh->dev[i].page;
2770 			clear_bit(R5_OrigPageUPTDODATE, &sh->dev[i].flags);
2771 
2772 			if (!using_disk_info_extra_page)
2773 				put_page(p);
2774 		}
2775 
2776 	if (using_disk_info_extra_page) {
2777 		clear_bit(R5C_EXTRA_PAGE_IN_USE, &conf->cache_state);
2778 		md_wakeup_thread(conf->mddev->thread);
2779 	}
2780 }
2781 
r5c_use_extra_page(struct stripe_head * sh)2782 void r5c_use_extra_page(struct stripe_head *sh)
2783 {
2784 	struct r5conf *conf = sh->raid_conf;
2785 	int i;
2786 	struct r5dev *dev;
2787 
2788 	for (i = sh->disks; i--; ) {
2789 		dev = &sh->dev[i];
2790 		if (dev->orig_page != dev->page)
2791 			put_page(dev->orig_page);
2792 		dev->orig_page = conf->disks[i].extra_page;
2793 	}
2794 }
2795 
2796 /*
2797  * clean up the stripe (clear R5_InJournal for dev[pd_idx] etc.) after the
2798  * stripe is committed to RAID disks.
2799  */
r5c_finish_stripe_write_out(struct r5conf * conf,struct stripe_head * sh,struct stripe_head_state * s)2800 void r5c_finish_stripe_write_out(struct r5conf *conf,
2801 				 struct stripe_head *sh,
2802 				 struct stripe_head_state *s)
2803 {
2804 	struct r5l_log *log = READ_ONCE(conf->log);
2805 	int i;
2806 	int do_wakeup = 0;
2807 	sector_t tree_index;
2808 	void __rcu **pslot;
2809 	uintptr_t refcount;
2810 
2811 	if (!log || !test_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags))
2812 		return;
2813 
2814 	WARN_ON(test_bit(STRIPE_R5C_CACHING, &sh->state));
2815 	clear_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags);
2816 
2817 	if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH)
2818 		return;
2819 
2820 	for (i = sh->disks; i--; ) {
2821 		clear_bit(R5_InJournal, &sh->dev[i].flags);
2822 		if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2823 			do_wakeup = 1;
2824 	}
2825 
2826 	/*
2827 	 * analyse_stripe() runs before r5c_finish_stripe_write_out(),
2828 	 * We updated R5_InJournal, so we also update s->injournal.
2829 	 */
2830 	s->injournal = 0;
2831 
2832 	if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2833 		if (atomic_dec_and_test(&conf->pending_full_writes))
2834 			md_wakeup_thread(conf->mddev->thread);
2835 
2836 	if (do_wakeup)
2837 		wake_up(&conf->wait_for_overlap);
2838 
2839 	spin_lock_irq(&log->stripe_in_journal_lock);
2840 	list_del_init(&sh->r5c);
2841 	spin_unlock_irq(&log->stripe_in_journal_lock);
2842 	sh->log_start = MaxSector;
2843 
2844 	atomic_dec(&log->stripe_in_journal_count);
2845 	r5c_update_log_state(log);
2846 
2847 	/* stop counting this stripe in big_stripe_tree */
2848 	if (test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state) ||
2849 	    test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state)) {
2850 		tree_index = r5c_tree_index(conf, sh->sector);
2851 		spin_lock(&log->tree_lock);
2852 		pslot = radix_tree_lookup_slot(&log->big_stripe_tree,
2853 					       tree_index);
2854 		BUG_ON(pslot == NULL);
2855 		refcount = (uintptr_t)radix_tree_deref_slot_protected(
2856 			pslot, &log->tree_lock) >>
2857 			R5C_RADIX_COUNT_SHIFT;
2858 		if (refcount == 1)
2859 			radix_tree_delete(&log->big_stripe_tree, tree_index);
2860 		else
2861 			radix_tree_replace_slot(
2862 				&log->big_stripe_tree, pslot,
2863 				(void *)((refcount - 1) << R5C_RADIX_COUNT_SHIFT));
2864 		spin_unlock(&log->tree_lock);
2865 	}
2866 
2867 	if (test_and_clear_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state)) {
2868 		BUG_ON(atomic_read(&conf->r5c_cached_partial_stripes) == 0);
2869 		atomic_dec(&conf->r5c_flushing_partial_stripes);
2870 		atomic_dec(&conf->r5c_cached_partial_stripes);
2871 	}
2872 
2873 	if (test_and_clear_bit(STRIPE_R5C_FULL_STRIPE, &sh->state)) {
2874 		BUG_ON(atomic_read(&conf->r5c_cached_full_stripes) == 0);
2875 		atomic_dec(&conf->r5c_flushing_full_stripes);
2876 		atomic_dec(&conf->r5c_cached_full_stripes);
2877 	}
2878 
2879 	r5l_append_flush_payload(log, sh->sector);
2880 	/* stripe is flused to raid disks, we can do resync now */
2881 	if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state))
2882 		set_bit(STRIPE_HANDLE, &sh->state);
2883 }
2884 
r5c_cache_data(struct r5l_log * log,struct stripe_head * sh)2885 int r5c_cache_data(struct r5l_log *log, struct stripe_head *sh)
2886 {
2887 	struct r5conf *conf = sh->raid_conf;
2888 	int pages = 0;
2889 	int reserve;
2890 	int i;
2891 	int ret = 0;
2892 
2893 	BUG_ON(!log);
2894 
2895 	for (i = 0; i < sh->disks; i++) {
2896 		void *addr;
2897 
2898 		if (!test_bit(R5_Wantwrite, &sh->dev[i].flags))
2899 			continue;
2900 		addr = kmap_atomic(sh->dev[i].page);
2901 		sh->dev[i].log_checksum = crc32c_le(log->uuid_checksum,
2902 						    addr, PAGE_SIZE);
2903 		kunmap_atomic(addr);
2904 		pages++;
2905 	}
2906 	WARN_ON(pages == 0);
2907 
2908 	/*
2909 	 * The stripe must enter state machine again to call endio, so
2910 	 * don't delay.
2911 	 */
2912 	clear_bit(STRIPE_DELAYED, &sh->state);
2913 	atomic_inc(&sh->count);
2914 
2915 	mutex_lock(&log->io_mutex);
2916 	/* meta + data */
2917 	reserve = (1 + pages) << (PAGE_SHIFT - 9);
2918 
2919 	if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state) &&
2920 	    sh->log_start == MaxSector)
2921 		r5l_add_no_space_stripe(log, sh);
2922 	else if (!r5l_has_free_space(log, reserve)) {
2923 		if (sh->log_start == log->last_checkpoint)
2924 			BUG();
2925 		else
2926 			r5l_add_no_space_stripe(log, sh);
2927 	} else {
2928 		ret = r5l_log_stripe(log, sh, pages, 0);
2929 		if (ret) {
2930 			spin_lock_irq(&log->io_list_lock);
2931 			list_add_tail(&sh->log_list, &log->no_mem_stripes);
2932 			spin_unlock_irq(&log->io_list_lock);
2933 		}
2934 	}
2935 
2936 	mutex_unlock(&log->io_mutex);
2937 	return 0;
2938 }
2939 
2940 /* check whether this big stripe is in write back cache. */
r5c_big_stripe_cached(struct r5conf * conf,sector_t sect)2941 bool r5c_big_stripe_cached(struct r5conf *conf, sector_t sect)
2942 {
2943 	struct r5l_log *log = READ_ONCE(conf->log);
2944 	sector_t tree_index;
2945 	void *slot;
2946 
2947 	if (!log)
2948 		return false;
2949 
2950 	WARN_ON_ONCE(!rcu_read_lock_held());
2951 	tree_index = r5c_tree_index(conf, sect);
2952 	slot = radix_tree_lookup(&log->big_stripe_tree, tree_index);
2953 	return slot != NULL;
2954 }
2955 
r5l_load_log(struct r5l_log * log)2956 static int r5l_load_log(struct r5l_log *log)
2957 {
2958 	struct md_rdev *rdev = log->rdev;
2959 	struct page *page;
2960 	struct r5l_meta_block *mb;
2961 	sector_t cp = log->rdev->journal_tail;
2962 	u32 stored_crc, expected_crc;
2963 	bool create_super = false;
2964 	int ret = 0;
2965 
2966 	/* Make sure it's valid */
2967 	if (cp >= rdev->sectors || round_down(cp, BLOCK_SECTORS) != cp)
2968 		cp = 0;
2969 	page = alloc_page(GFP_KERNEL);
2970 	if (!page)
2971 		return -ENOMEM;
2972 
2973 	if (!sync_page_io(rdev, cp, PAGE_SIZE, page, REQ_OP_READ, false)) {
2974 		ret = -EIO;
2975 		goto ioerr;
2976 	}
2977 	mb = page_address(page);
2978 
2979 	if (le32_to_cpu(mb->magic) != R5LOG_MAGIC ||
2980 	    mb->version != R5LOG_VERSION) {
2981 		create_super = true;
2982 		goto create;
2983 	}
2984 	stored_crc = le32_to_cpu(mb->checksum);
2985 	mb->checksum = 0;
2986 	expected_crc = crc32c_le(log->uuid_checksum, mb, PAGE_SIZE);
2987 	if (stored_crc != expected_crc) {
2988 		create_super = true;
2989 		goto create;
2990 	}
2991 	if (le64_to_cpu(mb->position) != cp) {
2992 		create_super = true;
2993 		goto create;
2994 	}
2995 create:
2996 	if (create_super) {
2997 		log->last_cp_seq = get_random_u32();
2998 		cp = 0;
2999 		r5l_log_write_empty_meta_block(log, cp, log->last_cp_seq);
3000 		/*
3001 		 * Make sure super points to correct address. Log might have
3002 		 * data very soon. If super hasn't correct log tail address,
3003 		 * recovery can't find the log
3004 		 */
3005 		r5l_write_super(log, cp);
3006 	} else
3007 		log->last_cp_seq = le64_to_cpu(mb->seq);
3008 
3009 	log->device_size = round_down(rdev->sectors, BLOCK_SECTORS);
3010 	log->max_free_space = log->device_size >> RECLAIM_MAX_FREE_SPACE_SHIFT;
3011 	if (log->max_free_space > RECLAIM_MAX_FREE_SPACE)
3012 		log->max_free_space = RECLAIM_MAX_FREE_SPACE;
3013 	log->last_checkpoint = cp;
3014 
3015 	__free_page(page);
3016 
3017 	if (create_super) {
3018 		log->log_start = r5l_ring_add(log, cp, BLOCK_SECTORS);
3019 		log->seq = log->last_cp_seq + 1;
3020 		log->next_checkpoint = cp;
3021 	} else
3022 		ret = r5l_recovery_log(log);
3023 
3024 	r5c_update_log_state(log);
3025 	return ret;
3026 ioerr:
3027 	__free_page(page);
3028 	return ret;
3029 }
3030 
r5l_start(struct r5l_log * log)3031 int r5l_start(struct r5l_log *log)
3032 {
3033 	int ret;
3034 
3035 	if (!log)
3036 		return 0;
3037 
3038 	ret = r5l_load_log(log);
3039 	if (ret) {
3040 		struct mddev *mddev = log->rdev->mddev;
3041 		struct r5conf *conf = mddev->private;
3042 
3043 		r5l_exit_log(conf);
3044 	}
3045 	return ret;
3046 }
3047 
r5c_update_on_rdev_error(struct mddev * mddev,struct md_rdev * rdev)3048 void r5c_update_on_rdev_error(struct mddev *mddev, struct md_rdev *rdev)
3049 {
3050 	struct r5conf *conf = mddev->private;
3051 	struct r5l_log *log = READ_ONCE(conf->log);
3052 
3053 	if (!log)
3054 		return;
3055 
3056 	if ((raid5_calc_degraded(conf) > 0 ||
3057 	     test_bit(Journal, &rdev->flags)) &&
3058 	    log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_BACK)
3059 		schedule_work(&log->disable_writeback_work);
3060 }
3061 
r5l_init_log(struct r5conf * conf,struct md_rdev * rdev)3062 int r5l_init_log(struct r5conf *conf, struct md_rdev *rdev)
3063 {
3064 	struct r5l_log *log;
3065 	struct md_thread *thread;
3066 	int ret;
3067 
3068 	pr_debug("md/raid:%s: using device %pg as journal\n",
3069 		 mdname(conf->mddev), rdev->bdev);
3070 
3071 	if (PAGE_SIZE != 4096)
3072 		return -EINVAL;
3073 
3074 	/*
3075 	 * The PAGE_SIZE must be big enough to hold 1 r5l_meta_block and
3076 	 * raid_disks r5l_payload_data_parity.
3077 	 *
3078 	 * Write journal and cache does not work for very big array
3079 	 * (raid_disks > 203)
3080 	 */
3081 	if (sizeof(struct r5l_meta_block) +
3082 	    ((sizeof(struct r5l_payload_data_parity) + sizeof(__le32)) *
3083 	     conf->raid_disks) > PAGE_SIZE) {
3084 		pr_err("md/raid:%s: write journal/cache doesn't work for array with %d disks\n",
3085 		       mdname(conf->mddev), conf->raid_disks);
3086 		return -EINVAL;
3087 	}
3088 
3089 	log = kzalloc(sizeof(*log), GFP_KERNEL);
3090 	if (!log)
3091 		return -ENOMEM;
3092 	log->rdev = rdev;
3093 	log->need_cache_flush = bdev_write_cache(rdev->bdev);
3094 	log->uuid_checksum = crc32c_le(~0, rdev->mddev->uuid,
3095 				       sizeof(rdev->mddev->uuid));
3096 
3097 	mutex_init(&log->io_mutex);
3098 
3099 	spin_lock_init(&log->io_list_lock);
3100 	INIT_LIST_HEAD(&log->running_ios);
3101 	INIT_LIST_HEAD(&log->io_end_ios);
3102 	INIT_LIST_HEAD(&log->flushing_ios);
3103 	INIT_LIST_HEAD(&log->finished_ios);
3104 
3105 	log->io_kc = KMEM_CACHE(r5l_io_unit, 0);
3106 	if (!log->io_kc)
3107 		goto io_kc;
3108 
3109 	ret = mempool_init_slab_pool(&log->io_pool, R5L_POOL_SIZE, log->io_kc);
3110 	if (ret)
3111 		goto io_pool;
3112 
3113 	ret = bioset_init(&log->bs, R5L_POOL_SIZE, 0, BIOSET_NEED_BVECS);
3114 	if (ret)
3115 		goto io_bs;
3116 
3117 	ret = mempool_init_page_pool(&log->meta_pool, R5L_POOL_SIZE, 0);
3118 	if (ret)
3119 		goto out_mempool;
3120 
3121 	spin_lock_init(&log->tree_lock);
3122 	INIT_RADIX_TREE(&log->big_stripe_tree, GFP_NOWAIT | __GFP_NOWARN);
3123 
3124 	thread = md_register_thread(r5l_reclaim_thread, log->rdev->mddev,
3125 				    "reclaim");
3126 	if (!thread)
3127 		goto reclaim_thread;
3128 
3129 	thread->timeout = R5C_RECLAIM_WAKEUP_INTERVAL;
3130 	rcu_assign_pointer(log->reclaim_thread, thread);
3131 
3132 	init_waitqueue_head(&log->iounit_wait);
3133 
3134 	INIT_LIST_HEAD(&log->no_mem_stripes);
3135 
3136 	INIT_LIST_HEAD(&log->no_space_stripes);
3137 	spin_lock_init(&log->no_space_stripes_lock);
3138 
3139 	INIT_WORK(&log->deferred_io_work, r5l_submit_io_async);
3140 	INIT_WORK(&log->disable_writeback_work, r5c_disable_writeback_async);
3141 
3142 	log->r5c_journal_mode = R5C_JOURNAL_MODE_WRITE_THROUGH;
3143 	INIT_LIST_HEAD(&log->stripe_in_journal_list);
3144 	spin_lock_init(&log->stripe_in_journal_lock);
3145 	atomic_set(&log->stripe_in_journal_count, 0);
3146 
3147 	WRITE_ONCE(conf->log, log);
3148 
3149 	set_bit(MD_HAS_JOURNAL, &conf->mddev->flags);
3150 	return 0;
3151 
3152 reclaim_thread:
3153 	mempool_exit(&log->meta_pool);
3154 out_mempool:
3155 	bioset_exit(&log->bs);
3156 io_bs:
3157 	mempool_exit(&log->io_pool);
3158 io_pool:
3159 	kmem_cache_destroy(log->io_kc);
3160 io_kc:
3161 	kfree(log);
3162 	return -EINVAL;
3163 }
3164 
r5l_exit_log(struct r5conf * conf)3165 void r5l_exit_log(struct r5conf *conf)
3166 {
3167 	struct r5l_log *log = conf->log;
3168 
3169 	md_unregister_thread(conf->mddev, &log->reclaim_thread);
3170 
3171 	/*
3172 	 * 'reconfig_mutex' is held by caller, set 'confg->log' to NULL to
3173 	 * ensure disable_writeback_work wakes up and exits.
3174 	 */
3175 	WRITE_ONCE(conf->log, NULL);
3176 	wake_up(&conf->mddev->sb_wait);
3177 	flush_work(&log->disable_writeback_work);
3178 
3179 	mempool_exit(&log->meta_pool);
3180 	bioset_exit(&log->bs);
3181 	mempool_exit(&log->io_pool);
3182 	kmem_cache_destroy(log->io_kc);
3183 	kfree(log);
3184 }
3185