Linux Audio

Check our new training course

In-person Linux kernel drivers training

Jun 16-20, 2025
Register
Loading...
Note: File does not exist in v3.1.
   1// SPDX-License-Identifier: GPL-2.0
   2/*
   3 * Copyright (C) 2010 Kent Overstreet <kent.overstreet@gmail.com>
   4 *
   5 * Uses a block device as cache for other block devices; optimized for SSDs.
   6 * All allocation is done in buckets, which should match the erase block size
   7 * of the device.
   8 *
   9 * Buckets containing cached data are kept on a heap sorted by priority;
  10 * bucket priority is increased on cache hit, and periodically all the buckets
  11 * on the heap have their priority scaled down. This currently is just used as
  12 * an LRU but in the future should allow for more intelligent heuristics.
  13 *
  14 * Buckets have an 8 bit counter; freeing is accomplished by incrementing the
  15 * counter. Garbage collection is used to remove stale pointers.
  16 *
  17 * Indexing is done via a btree; nodes are not necessarily fully sorted, rather
  18 * as keys are inserted we only sort the pages that have not yet been written.
  19 * When garbage collection is run, we resort the entire node.
  20 *
  21 * All configuration is done via sysfs; see Documentation/admin-guide/bcache.rst.
  22 */
  23
  24#include "bcache.h"
  25#include "btree.h"
  26#include "debug.h"
  27#include "extents.h"
  28
  29#include <linux/slab.h>
  30#include <linux/bitops.h>
  31#include <linux/hash.h>
  32#include <linux/kthread.h>
  33#include <linux/prefetch.h>
  34#include <linux/random.h>
  35#include <linux/rcupdate.h>
  36#include <linux/sched/clock.h>
  37#include <linux/rculist.h>
  38#include <linux/delay.h>
  39#include <trace/events/bcache.h>
  40
  41/*
  42 * Todo:
  43 * register_bcache: Return errors out to userspace correctly
  44 *
  45 * Writeback: don't undirty key until after a cache flush
  46 *
  47 * Create an iterator for key pointers
  48 *
  49 * On btree write error, mark bucket such that it won't be freed from the cache
  50 *
  51 * Journalling:
  52 *   Check for bad keys in replay
  53 *   Propagate barriers
  54 *   Refcount journal entries in journal_replay
  55 *
  56 * Garbage collection:
  57 *   Finish incremental gc
  58 *   Gc should free old UUIDs, data for invalid UUIDs
  59 *
  60 * Provide a way to list backing device UUIDs we have data cached for, and
  61 * probably how long it's been since we've seen them, and a way to invalidate
  62 * dirty data for devices that will never be attached again
  63 *
  64 * Keep 1 min/5 min/15 min statistics of how busy a block device has been, so
  65 * that based on that and how much dirty data we have we can keep writeback
  66 * from being starved
  67 *
  68 * Add a tracepoint or somesuch to watch for writeback starvation
  69 *
  70 * When btree depth > 1 and splitting an interior node, we have to make sure
  71 * alloc_bucket() cannot fail. This should be true but is not completely
  72 * obvious.
  73 *
  74 * Plugging?
  75 *
  76 * If data write is less than hard sector size of ssd, round up offset in open
  77 * bucket to the next whole sector
  78 *
  79 * Superblock needs to be fleshed out for multiple cache devices
  80 *
  81 * Add a sysfs tunable for the number of writeback IOs in flight
  82 *
  83 * Add a sysfs tunable for the number of open data buckets
  84 *
  85 * IO tracking: Can we track when one process is doing io on behalf of another?
  86 * IO tracking: Don't use just an average, weigh more recent stuff higher
  87 *
  88 * Test module load/unload
  89 */
  90
  91#define MAX_NEED_GC		64
  92#define MAX_SAVE_PRIO		72
  93#define MAX_GC_TIMES		100
  94#define MIN_GC_NODES		100
  95#define GC_SLEEP_MS		100
  96
  97#define PTR_DIRTY_BIT		(((uint64_t) 1 << 36))
  98
  99#define PTR_HASH(c, k)							\
 100	(((k)->ptr[0] >> c->bucket_bits) | PTR_GEN(k, 0))
 101
 102#define insert_lock(s, b)	((b)->level <= (s)->lock)
 103
 104/*
 105 * These macros are for recursing down the btree - they handle the details of
 106 * locking and looking up nodes in the cache for you. They're best treated as
 107 * mere syntax when reading code that uses them.
 108 *
 109 * op->lock determines whether we take a read or a write lock at a given depth.
 110 * If you've got a read lock and find that you need a write lock (i.e. you're
 111 * going to have to split), set op->lock and return -EINTR; btree_root() will
 112 * call you again and you'll have the correct lock.
 113 */
 114
 115/**
 116 * btree - recurse down the btree on a specified key
 117 * @fn:		function to call, which will be passed the child node
 118 * @key:	key to recurse on
 119 * @b:		parent btree node
 120 * @op:		pointer to struct btree_op
 121 */
 122#define btree(fn, key, b, op, ...)					\
 123({									\
 124	int _r, l = (b)->level - 1;					\
 125	bool _w = l <= (op)->lock;					\
 126	struct btree *_child = bch_btree_node_get((b)->c, op, key, l,	\
 127						  _w, b);		\
 128	if (!IS_ERR(_child)) {						\
 129		_r = bch_btree_ ## fn(_child, op, ##__VA_ARGS__);	\
 130		rw_unlock(_w, _child);					\
 131	} else								\
 132		_r = PTR_ERR(_child);					\
 133	_r;								\
 134})
 135
 136/**
 137 * btree_root - call a function on the root of the btree
 138 * @fn:		function to call, which will be passed the child node
 139 * @c:		cache set
 140 * @op:		pointer to struct btree_op
 141 */
 142#define btree_root(fn, c, op, ...)					\
 143({									\
 144	int _r = -EINTR;						\
 145	do {								\
 146		struct btree *_b = (c)->root;				\
 147		bool _w = insert_lock(op, _b);				\
 148		rw_lock(_w, _b, _b->level);				\
 149		if (_b == (c)->root &&					\
 150		    _w == insert_lock(op, _b)) {			\
 151			_r = bch_btree_ ## fn(_b, op, ##__VA_ARGS__);	\
 152		}							\
 153		rw_unlock(_w, _b);					\
 154		bch_cannibalize_unlock(c);				\
 155		if (_r == -EINTR)					\
 156			schedule();					\
 157	} while (_r == -EINTR);						\
 158									\
 159	finish_wait(&(c)->btree_cache_wait, &(op)->wait);		\
 160	_r;								\
 161})
 162
 163static inline struct bset *write_block(struct btree *b)
 164{
 165	return ((void *) btree_bset_first(b)) + b->written * block_bytes(b->c);
 166}
 167
 168static void bch_btree_init_next(struct btree *b)
 169{
 170	/* If not a leaf node, always sort */
 171	if (b->level && b->keys.nsets)
 172		bch_btree_sort(&b->keys, &b->c->sort);
 173	else
 174		bch_btree_sort_lazy(&b->keys, &b->c->sort);
 175
 176	if (b->written < btree_blocks(b))
 177		bch_bset_init_next(&b->keys, write_block(b),
 178				   bset_magic(&b->c->sb));
 179
 180}
 181
 182/* Btree key manipulation */
 183
 184void bkey_put(struct cache_set *c, struct bkey *k)
 185{
 186	unsigned int i;
 187
 188	for (i = 0; i < KEY_PTRS(k); i++)
 189		if (ptr_available(c, k, i))
 190			atomic_dec_bug(&PTR_BUCKET(c, k, i)->pin);
 191}
 192
 193/* Btree IO */
 194
 195static uint64_t btree_csum_set(struct btree *b, struct bset *i)
 196{
 197	uint64_t crc = b->key.ptr[0];
 198	void *data = (void *) i + 8, *end = bset_bkey_last(i);
 199
 200	crc = bch_crc64_update(crc, data, end - data);
 201	return crc ^ 0xffffffffffffffffULL;
 202}
 203
 204void bch_btree_node_read_done(struct btree *b)
 205{
 206	const char *err = "bad btree header";
 207	struct bset *i = btree_bset_first(b);
 208	struct btree_iter *iter;
 209
 210	/*
 211	 * c->fill_iter can allocate an iterator with more memory space
 212	 * than static MAX_BSETS.
 213	 * See the comment arount cache_set->fill_iter.
 214	 */
 215	iter = mempool_alloc(&b->c->fill_iter, GFP_NOIO);
 216	iter->size = b->c->sb.bucket_size / b->c->sb.block_size;
 217	iter->used = 0;
 218
 219#ifdef CONFIG_BCACHE_DEBUG
 220	iter->b = &b->keys;
 221#endif
 222
 223	if (!i->seq)
 224		goto err;
 225
 226	for (;
 227	     b->written < btree_blocks(b) && i->seq == b->keys.set[0].data->seq;
 228	     i = write_block(b)) {
 229		err = "unsupported bset version";
 230		if (i->version > BCACHE_BSET_VERSION)
 231			goto err;
 232
 233		err = "bad btree header";
 234		if (b->written + set_blocks(i, block_bytes(b->c)) >
 235		    btree_blocks(b))
 236			goto err;
 237
 238		err = "bad magic";
 239		if (i->magic != bset_magic(&b->c->sb))
 240			goto err;
 241
 242		err = "bad checksum";
 243		switch (i->version) {
 244		case 0:
 245			if (i->csum != csum_set(i))
 246				goto err;
 247			break;
 248		case BCACHE_BSET_VERSION:
 249			if (i->csum != btree_csum_set(b, i))
 250				goto err;
 251			break;
 252		}
 253
 254		err = "empty set";
 255		if (i != b->keys.set[0].data && !i->keys)
 256			goto err;
 257
 258		bch_btree_iter_push(iter, i->start, bset_bkey_last(i));
 259
 260		b->written += set_blocks(i, block_bytes(b->c));
 261	}
 262
 263	err = "corrupted btree";
 264	for (i = write_block(b);
 265	     bset_sector_offset(&b->keys, i) < KEY_SIZE(&b->key);
 266	     i = ((void *) i) + block_bytes(b->c))
 267		if (i->seq == b->keys.set[0].data->seq)
 268			goto err;
 269
 270	bch_btree_sort_and_fix_extents(&b->keys, iter, &b->c->sort);
 271
 272	i = b->keys.set[0].data;
 273	err = "short btree key";
 274	if (b->keys.set[0].size &&
 275	    bkey_cmp(&b->key, &b->keys.set[0].end) < 0)
 276		goto err;
 277
 278	if (b->written < btree_blocks(b))
 279		bch_bset_init_next(&b->keys, write_block(b),
 280				   bset_magic(&b->c->sb));
 281out:
 282	mempool_free(iter, &b->c->fill_iter);
 283	return;
 284err:
 285	set_btree_node_io_error(b);
 286	bch_cache_set_error(b->c, "%s at bucket %zu, block %u, %u keys",
 287			    err, PTR_BUCKET_NR(b->c, &b->key, 0),
 288			    bset_block_offset(b, i), i->keys);
 289	goto out;
 290}
 291
 292static void btree_node_read_endio(struct bio *bio)
 293{
 294	struct closure *cl = bio->bi_private;
 295
 296	closure_put(cl);
 297}
 298
 299static void bch_btree_node_read(struct btree *b)
 300{
 301	uint64_t start_time = local_clock();
 302	struct closure cl;
 303	struct bio *bio;
 304
 305	trace_bcache_btree_read(b);
 306
 307	closure_init_stack(&cl);
 308
 309	bio = bch_bbio_alloc(b->c);
 310	bio->bi_iter.bi_size = KEY_SIZE(&b->key) << 9;
 311	bio->bi_end_io	= btree_node_read_endio;
 312	bio->bi_private	= &cl;
 313	bio->bi_opf = REQ_OP_READ | REQ_META;
 314
 315	bch_bio_map(bio, b->keys.set[0].data);
 316
 317	bch_submit_bbio(bio, b->c, &b->key, 0);
 318	closure_sync(&cl);
 319
 320	if (bio->bi_status)
 321		set_btree_node_io_error(b);
 322
 323	bch_bbio_free(bio, b->c);
 324
 325	if (btree_node_io_error(b))
 326		goto err;
 327
 328	bch_btree_node_read_done(b);
 329	bch_time_stats_update(&b->c->btree_read_time, start_time);
 330
 331	return;
 332err:
 333	bch_cache_set_error(b->c, "io error reading bucket %zu",
 334			    PTR_BUCKET_NR(b->c, &b->key, 0));
 335}
 336
 337static void btree_complete_write(struct btree *b, struct btree_write *w)
 338{
 339	if (w->prio_blocked &&
 340	    !atomic_sub_return(w->prio_blocked, &b->c->prio_blocked))
 341		wake_up_allocators(b->c);
 342
 343	if (w->journal) {
 344		atomic_dec_bug(w->journal);
 345		__closure_wake_up(&b->c->journal.wait);
 346	}
 347
 348	w->prio_blocked	= 0;
 349	w->journal	= NULL;
 350}
 351
 352static void btree_node_write_unlock(struct closure *cl)
 353{
 354	struct btree *b = container_of(cl, struct btree, io);
 355
 356	up(&b->io_mutex);
 357}
 358
 359static void __btree_node_write_done(struct closure *cl)
 360{
 361	struct btree *b = container_of(cl, struct btree, io);
 362	struct btree_write *w = btree_prev_write(b);
 363
 364	bch_bbio_free(b->bio, b->c);
 365	b->bio = NULL;
 366	btree_complete_write(b, w);
 367
 368	if (btree_node_dirty(b))
 369		schedule_delayed_work(&b->work, 30 * HZ);
 370
 371	closure_return_with_destructor(cl, btree_node_write_unlock);
 372}
 373
 374static void btree_node_write_done(struct closure *cl)
 375{
 376	struct btree *b = container_of(cl, struct btree, io);
 377
 378	bio_free_pages(b->bio);
 379	__btree_node_write_done(cl);
 380}
 381
 382static void btree_node_write_endio(struct bio *bio)
 383{
 384	struct closure *cl = bio->bi_private;
 385	struct btree *b = container_of(cl, struct btree, io);
 386
 387	if (bio->bi_status)
 388		set_btree_node_io_error(b);
 389
 390	bch_bbio_count_io_errors(b->c, bio, bio->bi_status, "writing btree");
 391	closure_put(cl);
 392}
 393
 394static void do_btree_node_write(struct btree *b)
 395{
 396	struct closure *cl = &b->io;
 397	struct bset *i = btree_bset_last(b);
 398	BKEY_PADDED(key) k;
 399
 400	i->version	= BCACHE_BSET_VERSION;
 401	i->csum		= btree_csum_set(b, i);
 402
 403	BUG_ON(b->bio);
 404	b->bio = bch_bbio_alloc(b->c);
 405
 406	b->bio->bi_end_io	= btree_node_write_endio;
 407	b->bio->bi_private	= cl;
 408	b->bio->bi_iter.bi_size	= roundup(set_bytes(i), block_bytes(b->c));
 409	b->bio->bi_opf		= REQ_OP_WRITE | REQ_META | REQ_FUA;
 410	bch_bio_map(b->bio, i);
 411
 412	/*
 413	 * If we're appending to a leaf node, we don't technically need FUA -
 414	 * this write just needs to be persisted before the next journal write,
 415	 * which will be marked FLUSH|FUA.
 416	 *
 417	 * Similarly if we're writing a new btree root - the pointer is going to
 418	 * be in the next journal entry.
 419	 *
 420	 * But if we're writing a new btree node (that isn't a root) or
 421	 * appending to a non leaf btree node, we need either FUA or a flush
 422	 * when we write the parent with the new pointer. FUA is cheaper than a
 423	 * flush, and writes appending to leaf nodes aren't blocking anything so
 424	 * just make all btree node writes FUA to keep things sane.
 425	 */
 426
 427	bkey_copy(&k.key, &b->key);
 428	SET_PTR_OFFSET(&k.key, 0, PTR_OFFSET(&k.key, 0) +
 429		       bset_sector_offset(&b->keys, i));
 430
 431	if (!bch_bio_alloc_pages(b->bio, __GFP_NOWARN|GFP_NOWAIT)) {
 432		struct bio_vec *bv;
 433		void *addr = (void *) ((unsigned long) i & ~(PAGE_SIZE - 1));
 434		struct bvec_iter_all iter_all;
 435
 436		bio_for_each_segment_all(bv, b->bio, iter_all) {
 437			memcpy(page_address(bv->bv_page), addr, PAGE_SIZE);
 438			addr += PAGE_SIZE;
 439		}
 440
 441		bch_submit_bbio(b->bio, b->c, &k.key, 0);
 442
 443		continue_at(cl, btree_node_write_done, NULL);
 444	} else {
 445		/*
 446		 * No problem for multipage bvec since the bio is
 447		 * just allocated
 448		 */
 449		b->bio->bi_vcnt = 0;
 450		bch_bio_map(b->bio, i);
 451
 452		bch_submit_bbio(b->bio, b->c, &k.key, 0);
 453
 454		closure_sync(cl);
 455		continue_at_nobarrier(cl, __btree_node_write_done, NULL);
 456	}
 457}
 458
 459void __bch_btree_node_write(struct btree *b, struct closure *parent)
 460{
 461	struct bset *i = btree_bset_last(b);
 462
 463	lockdep_assert_held(&b->write_lock);
 464
 465	trace_bcache_btree_write(b);
 466
 467	BUG_ON(current->bio_list);
 468	BUG_ON(b->written >= btree_blocks(b));
 469	BUG_ON(b->written && !i->keys);
 470	BUG_ON(btree_bset_first(b)->seq != i->seq);
 471	bch_check_keys(&b->keys, "writing");
 472
 473	cancel_delayed_work(&b->work);
 474
 475	/* If caller isn't waiting for write, parent refcount is cache set */
 476	down(&b->io_mutex);
 477	closure_init(&b->io, parent ?: &b->c->cl);
 478
 479	clear_bit(BTREE_NODE_dirty,	 &b->flags);
 480	change_bit(BTREE_NODE_write_idx, &b->flags);
 481
 482	do_btree_node_write(b);
 483
 484	atomic_long_add(set_blocks(i, block_bytes(b->c)) * b->c->sb.block_size,
 485			&PTR_CACHE(b->c, &b->key, 0)->btree_sectors_written);
 486
 487	b->written += set_blocks(i, block_bytes(b->c));
 488}
 489
 490void bch_btree_node_write(struct btree *b, struct closure *parent)
 491{
 492	unsigned int nsets = b->keys.nsets;
 493
 494	lockdep_assert_held(&b->lock);
 495
 496	__bch_btree_node_write(b, parent);
 497
 498	/*
 499	 * do verify if there was more than one set initially (i.e. we did a
 500	 * sort) and we sorted down to a single set:
 501	 */
 502	if (nsets && !b->keys.nsets)
 503		bch_btree_verify(b);
 504
 505	bch_btree_init_next(b);
 506}
 507
 508static void bch_btree_node_write_sync(struct btree *b)
 509{
 510	struct closure cl;
 511
 512	closure_init_stack(&cl);
 513
 514	mutex_lock(&b->write_lock);
 515	bch_btree_node_write(b, &cl);
 516	mutex_unlock(&b->write_lock);
 517
 518	closure_sync(&cl);
 519}
 520
 521static void btree_node_write_work(struct work_struct *w)
 522{
 523	struct btree *b = container_of(to_delayed_work(w), struct btree, work);
 524
 525	mutex_lock(&b->write_lock);
 526	if (btree_node_dirty(b))
 527		__bch_btree_node_write(b, NULL);
 528	mutex_unlock(&b->write_lock);
 529}
 530
 531static void bch_btree_leaf_dirty(struct btree *b, atomic_t *journal_ref)
 532{
 533	struct bset *i = btree_bset_last(b);
 534	struct btree_write *w = btree_current_write(b);
 535
 536	lockdep_assert_held(&b->write_lock);
 537
 538	BUG_ON(!b->written);
 539	BUG_ON(!i->keys);
 540
 541	if (!btree_node_dirty(b))
 542		schedule_delayed_work(&b->work, 30 * HZ);
 543
 544	set_btree_node_dirty(b);
 545
 546	if (journal_ref) {
 547		if (w->journal &&
 548		    journal_pin_cmp(b->c, w->journal, journal_ref)) {
 549			atomic_dec_bug(w->journal);
 550			w->journal = NULL;
 551		}
 552
 553		if (!w->journal) {
 554			w->journal = journal_ref;
 555			atomic_inc(w->journal);
 556		}
 557	}
 558
 559	/* Force write if set is too big */
 560	if (set_bytes(i) > PAGE_SIZE - 48 &&
 561	    !current->bio_list)
 562		bch_btree_node_write(b, NULL);
 563}
 564
 565/*
 566 * Btree in memory cache - allocation/freeing
 567 * mca -> memory cache
 568 */
 569
 570#define mca_reserve(c)	(((c->root && c->root->level)		\
 571			  ? c->root->level : 1) * 8 + 16)
 572#define mca_can_free(c)						\
 573	max_t(int, 0, c->btree_cache_used - mca_reserve(c))
 574
 575static void mca_data_free(struct btree *b)
 576{
 577	BUG_ON(b->io_mutex.count != 1);
 578
 579	bch_btree_keys_free(&b->keys);
 580
 581	b->c->btree_cache_used--;
 582	list_move(&b->list, &b->c->btree_cache_freed);
 583}
 584
 585static void mca_bucket_free(struct btree *b)
 586{
 587	BUG_ON(btree_node_dirty(b));
 588
 589	b->key.ptr[0] = 0;
 590	hlist_del_init_rcu(&b->hash);
 591	list_move(&b->list, &b->c->btree_cache_freeable);
 592}
 593
 594static unsigned int btree_order(struct bkey *k)
 595{
 596	return ilog2(KEY_SIZE(k) / PAGE_SECTORS ?: 1);
 597}
 598
 599static void mca_data_alloc(struct btree *b, struct bkey *k, gfp_t gfp)
 600{
 601	if (!bch_btree_keys_alloc(&b->keys,
 602				  max_t(unsigned int,
 603					ilog2(b->c->btree_pages),
 604					btree_order(k)),
 605				  gfp)) {
 606		b->c->btree_cache_used++;
 607		list_move(&b->list, &b->c->btree_cache);
 608	} else {
 609		list_move(&b->list, &b->c->btree_cache_freed);
 610	}
 611}
 612
 613static struct btree *mca_bucket_alloc(struct cache_set *c,
 614				      struct bkey *k, gfp_t gfp)
 615{
 616	/*
 617	 * kzalloc() is necessary here for initialization,
 618	 * see code comments in bch_btree_keys_init().
 619	 */
 620	struct btree *b = kzalloc(sizeof(struct btree), gfp);
 621
 622	if (!b)
 623		return NULL;
 624
 625	init_rwsem(&b->lock);
 626	lockdep_set_novalidate_class(&b->lock);
 627	mutex_init(&b->write_lock);
 628	lockdep_set_novalidate_class(&b->write_lock);
 629	INIT_LIST_HEAD(&b->list);
 630	INIT_DELAYED_WORK(&b->work, btree_node_write_work);
 631	b->c = c;
 632	sema_init(&b->io_mutex, 1);
 633
 634	mca_data_alloc(b, k, gfp);
 635	return b;
 636}
 637
 638static int mca_reap(struct btree *b, unsigned int min_order, bool flush)
 639{
 640	struct closure cl;
 641
 642	closure_init_stack(&cl);
 643	lockdep_assert_held(&b->c->bucket_lock);
 644
 645	if (!down_write_trylock(&b->lock))
 646		return -ENOMEM;
 647
 648	BUG_ON(btree_node_dirty(b) && !b->keys.set[0].data);
 649
 650	if (b->keys.page_order < min_order)
 651		goto out_unlock;
 652
 653	if (!flush) {
 654		if (btree_node_dirty(b))
 655			goto out_unlock;
 656
 657		if (down_trylock(&b->io_mutex))
 658			goto out_unlock;
 659		up(&b->io_mutex);
 660	}
 661
 662retry:
 663	/*
 664	 * BTREE_NODE_dirty might be cleared in btree_flush_btree() by
 665	 * __bch_btree_node_write(). To avoid an extra flush, acquire
 666	 * b->write_lock before checking BTREE_NODE_dirty bit.
 667	 */
 668	mutex_lock(&b->write_lock);
 669	/*
 670	 * If this btree node is selected in btree_flush_write() by journal
 671	 * code, delay and retry until the node is flushed by journal code
 672	 * and BTREE_NODE_journal_flush bit cleared by btree_flush_write().
 673	 */
 674	if (btree_node_journal_flush(b)) {
 675		pr_debug("bnode %p is flushing by journal, retry", b);
 676		mutex_unlock(&b->write_lock);
 677		udelay(1);
 678		goto retry;
 679	}
 680
 681	if (btree_node_dirty(b))
 682		__bch_btree_node_write(b, &cl);
 683	mutex_unlock(&b->write_lock);
 684
 685	closure_sync(&cl);
 686
 687	/* wait for any in flight btree write */
 688	down(&b->io_mutex);
 689	up(&b->io_mutex);
 690
 691	return 0;
 692out_unlock:
 693	rw_unlock(true, b);
 694	return -ENOMEM;
 695}
 696
 697static unsigned long bch_mca_scan(struct shrinker *shrink,
 698				  struct shrink_control *sc)
 699{
 700	struct cache_set *c = container_of(shrink, struct cache_set, shrink);
 701	struct btree *b, *t;
 702	unsigned long i, nr = sc->nr_to_scan;
 703	unsigned long freed = 0;
 704	unsigned int btree_cache_used;
 705
 706	if (c->shrinker_disabled)
 707		return SHRINK_STOP;
 708
 709	if (c->btree_cache_alloc_lock)
 710		return SHRINK_STOP;
 711
 712	/* Return -1 if we can't do anything right now */
 713	if (sc->gfp_mask & __GFP_IO)
 714		mutex_lock(&c->bucket_lock);
 715	else if (!mutex_trylock(&c->bucket_lock))
 716		return -1;
 717
 718	/*
 719	 * It's _really_ critical that we don't free too many btree nodes - we
 720	 * have to always leave ourselves a reserve. The reserve is how we
 721	 * guarantee that allocating memory for a new btree node can always
 722	 * succeed, so that inserting keys into the btree can always succeed and
 723	 * IO can always make forward progress:
 724	 */
 725	nr /= c->btree_pages;
 726	nr = min_t(unsigned long, nr, mca_can_free(c));
 727
 728	i = 0;
 729	btree_cache_used = c->btree_cache_used;
 730	list_for_each_entry_safe(b, t, &c->btree_cache_freeable, list) {
 731		if (nr <= 0)
 732			goto out;
 733
 734		if (++i > 3 &&
 735		    !mca_reap(b, 0, false)) {
 736			mca_data_free(b);
 737			rw_unlock(true, b);
 738			freed++;
 739		}
 740		nr--;
 741	}
 742
 743	for (;  (nr--) && i < btree_cache_used; i++) {
 744		if (list_empty(&c->btree_cache))
 745			goto out;
 746
 747		b = list_first_entry(&c->btree_cache, struct btree, list);
 748		list_rotate_left(&c->btree_cache);
 749
 750		if (!b->accessed &&
 751		    !mca_reap(b, 0, false)) {
 752			mca_bucket_free(b);
 753			mca_data_free(b);
 754			rw_unlock(true, b);
 755			freed++;
 756		} else
 757			b->accessed = 0;
 758	}
 759out:
 760	mutex_unlock(&c->bucket_lock);
 761	return freed * c->btree_pages;
 762}
 763
 764static unsigned long bch_mca_count(struct shrinker *shrink,
 765				   struct shrink_control *sc)
 766{
 767	struct cache_set *c = container_of(shrink, struct cache_set, shrink);
 768
 769	if (c->shrinker_disabled)
 770		return 0;
 771
 772	if (c->btree_cache_alloc_lock)
 773		return 0;
 774
 775	return mca_can_free(c) * c->btree_pages;
 776}
 777
 778void bch_btree_cache_free(struct cache_set *c)
 779{
 780	struct btree *b;
 781	struct closure cl;
 782
 783	closure_init_stack(&cl);
 784
 785	if (c->shrink.list.next)
 786		unregister_shrinker(&c->shrink);
 787
 788	mutex_lock(&c->bucket_lock);
 789
 790#ifdef CONFIG_BCACHE_DEBUG
 791	if (c->verify_data)
 792		list_move(&c->verify_data->list, &c->btree_cache);
 793
 794	free_pages((unsigned long) c->verify_ondisk, ilog2(bucket_pages(c)));
 795#endif
 796
 797	list_splice(&c->btree_cache_freeable,
 798		    &c->btree_cache);
 799
 800	while (!list_empty(&c->btree_cache)) {
 801		b = list_first_entry(&c->btree_cache, struct btree, list);
 802
 803		/*
 804		 * This function is called by cache_set_free(), no I/O
 805		 * request on cache now, it is unnecessary to acquire
 806		 * b->write_lock before clearing BTREE_NODE_dirty anymore.
 807		 */
 808		if (btree_node_dirty(b)) {
 809			btree_complete_write(b, btree_current_write(b));
 810			clear_bit(BTREE_NODE_dirty, &b->flags);
 811		}
 812		mca_data_free(b);
 813	}
 814
 815	while (!list_empty(&c->btree_cache_freed)) {
 816		b = list_first_entry(&c->btree_cache_freed,
 817				     struct btree, list);
 818		list_del(&b->list);
 819		cancel_delayed_work_sync(&b->work);
 820		kfree(b);
 821	}
 822
 823	mutex_unlock(&c->bucket_lock);
 824}
 825
 826int bch_btree_cache_alloc(struct cache_set *c)
 827{
 828	unsigned int i;
 829
 830	for (i = 0; i < mca_reserve(c); i++)
 831		if (!mca_bucket_alloc(c, &ZERO_KEY, GFP_KERNEL))
 832			return -ENOMEM;
 833
 834	list_splice_init(&c->btree_cache,
 835			 &c->btree_cache_freeable);
 836
 837#ifdef CONFIG_BCACHE_DEBUG
 838	mutex_init(&c->verify_lock);
 839
 840	c->verify_ondisk = (void *)
 841		__get_free_pages(GFP_KERNEL, ilog2(bucket_pages(c)));
 842
 843	c->verify_data = mca_bucket_alloc(c, &ZERO_KEY, GFP_KERNEL);
 844
 845	if (c->verify_data &&
 846	    c->verify_data->keys.set->data)
 847		list_del_init(&c->verify_data->list);
 848	else
 849		c->verify_data = NULL;
 850#endif
 851
 852	c->shrink.count_objects = bch_mca_count;
 853	c->shrink.scan_objects = bch_mca_scan;
 854	c->shrink.seeks = 4;
 855	c->shrink.batch = c->btree_pages * 2;
 856
 857	if (register_shrinker(&c->shrink))
 858		pr_warn("bcache: %s: could not register shrinker",
 859				__func__);
 860
 861	return 0;
 862}
 863
 864/* Btree in memory cache - hash table */
 865
 866static struct hlist_head *mca_hash(struct cache_set *c, struct bkey *k)
 867{
 868	return &c->bucket_hash[hash_32(PTR_HASH(c, k), BUCKET_HASH_BITS)];
 869}
 870
 871static struct btree *mca_find(struct cache_set *c, struct bkey *k)
 872{
 873	struct btree *b;
 874
 875	rcu_read_lock();
 876	hlist_for_each_entry_rcu(b, mca_hash(c, k), hash)
 877		if (PTR_HASH(c, &b->key) == PTR_HASH(c, k))
 878			goto out;
 879	b = NULL;
 880out:
 881	rcu_read_unlock();
 882	return b;
 883}
 884
 885static int mca_cannibalize_lock(struct cache_set *c, struct btree_op *op)
 886{
 887	struct task_struct *old;
 888
 889	old = cmpxchg(&c->btree_cache_alloc_lock, NULL, current);
 890	if (old && old != current) {
 891		if (op)
 892			prepare_to_wait(&c->btree_cache_wait, &op->wait,
 893					TASK_UNINTERRUPTIBLE);
 894		return -EINTR;
 895	}
 896
 897	return 0;
 898}
 899
 900static struct btree *mca_cannibalize(struct cache_set *c, struct btree_op *op,
 901				     struct bkey *k)
 902{
 903	struct btree *b;
 904
 905	trace_bcache_btree_cache_cannibalize(c);
 906
 907	if (mca_cannibalize_lock(c, op))
 908		return ERR_PTR(-EINTR);
 909
 910	list_for_each_entry_reverse(b, &c->btree_cache, list)
 911		if (!mca_reap(b, btree_order(k), false))
 912			return b;
 913
 914	list_for_each_entry_reverse(b, &c->btree_cache, list)
 915		if (!mca_reap(b, btree_order(k), true))
 916			return b;
 917
 918	WARN(1, "btree cache cannibalize failed\n");
 919	return ERR_PTR(-ENOMEM);
 920}
 921
 922/*
 923 * We can only have one thread cannibalizing other cached btree nodes at a time,
 924 * or we'll deadlock. We use an open coded mutex to ensure that, which a
 925 * cannibalize_bucket() will take. This means every time we unlock the root of
 926 * the btree, we need to release this lock if we have it held.
 927 */
 928static void bch_cannibalize_unlock(struct cache_set *c)
 929{
 930	if (c->btree_cache_alloc_lock == current) {
 931		c->btree_cache_alloc_lock = NULL;
 932		wake_up(&c->btree_cache_wait);
 933	}
 934}
 935
 936static struct btree *mca_alloc(struct cache_set *c, struct btree_op *op,
 937			       struct bkey *k, int level)
 938{
 939	struct btree *b;
 940
 941	BUG_ON(current->bio_list);
 942
 943	lockdep_assert_held(&c->bucket_lock);
 944
 945	if (mca_find(c, k))
 946		return NULL;
 947
 948	/* btree_free() doesn't free memory; it sticks the node on the end of
 949	 * the list. Check if there's any freed nodes there:
 950	 */
 951	list_for_each_entry(b, &c->btree_cache_freeable, list)
 952		if (!mca_reap(b, btree_order(k), false))
 953			goto out;
 954
 955	/* We never free struct btree itself, just the memory that holds the on
 956	 * disk node. Check the freed list before allocating a new one:
 957	 */
 958	list_for_each_entry(b, &c->btree_cache_freed, list)
 959		if (!mca_reap(b, 0, false)) {
 960			mca_data_alloc(b, k, __GFP_NOWARN|GFP_NOIO);
 961			if (!b->keys.set[0].data)
 962				goto err;
 963			else
 964				goto out;
 965		}
 966
 967	b = mca_bucket_alloc(c, k, __GFP_NOWARN|GFP_NOIO);
 968	if (!b)
 969		goto err;
 970
 971	BUG_ON(!down_write_trylock(&b->lock));
 972	if (!b->keys.set->data)
 973		goto err;
 974out:
 975	BUG_ON(b->io_mutex.count != 1);
 976
 977	bkey_copy(&b->key, k);
 978	list_move(&b->list, &c->btree_cache);
 979	hlist_del_init_rcu(&b->hash);
 980	hlist_add_head_rcu(&b->hash, mca_hash(c, k));
 981
 982	lock_set_subclass(&b->lock.dep_map, level + 1, _THIS_IP_);
 983	b->parent	= (void *) ~0UL;
 984	b->flags	= 0;
 985	b->written	= 0;
 986	b->level	= level;
 987
 988	if (!b->level)
 989		bch_btree_keys_init(&b->keys, &bch_extent_keys_ops,
 990				    &b->c->expensive_debug_checks);
 991	else
 992		bch_btree_keys_init(&b->keys, &bch_btree_keys_ops,
 993				    &b->c->expensive_debug_checks);
 994
 995	return b;
 996err:
 997	if (b)
 998		rw_unlock(true, b);
 999
1000	b = mca_cannibalize(c, op, k);
1001	if (!IS_ERR(b))
1002		goto out;
1003
1004	return b;
1005}
1006
1007/*
1008 * bch_btree_node_get - find a btree node in the cache and lock it, reading it
1009 * in from disk if necessary.
1010 *
1011 * If IO is necessary and running under generic_make_request, returns -EAGAIN.
1012 *
1013 * The btree node will have either a read or a write lock held, depending on
1014 * level and op->lock.
1015 */
1016struct btree *bch_btree_node_get(struct cache_set *c, struct btree_op *op,
1017				 struct bkey *k, int level, bool write,
1018				 struct btree *parent)
1019{
1020	int i = 0;
1021	struct btree *b;
1022
1023	BUG_ON(level < 0);
1024retry:
1025	b = mca_find(c, k);
1026
1027	if (!b) {
1028		if (current->bio_list)
1029			return ERR_PTR(-EAGAIN);
1030
1031		mutex_lock(&c->bucket_lock);
1032		b = mca_alloc(c, op, k, level);
1033		mutex_unlock(&c->bucket_lock);
1034
1035		if (!b)
1036			goto retry;
1037		if (IS_ERR(b))
1038			return b;
1039
1040		bch_btree_node_read(b);
1041
1042		if (!write)
1043			downgrade_write(&b->lock);
1044	} else {
1045		rw_lock(write, b, level);
1046		if (PTR_HASH(c, &b->key) != PTR_HASH(c, k)) {
1047			rw_unlock(write, b);
1048			goto retry;
1049		}
1050		BUG_ON(b->level != level);
1051	}
1052
1053	if (btree_node_io_error(b)) {
1054		rw_unlock(write, b);
1055		return ERR_PTR(-EIO);
1056	}
1057
1058	BUG_ON(!b->written);
1059
1060	b->parent = parent;
1061	b->accessed = 1;
1062
1063	for (; i <= b->keys.nsets && b->keys.set[i].size; i++) {
1064		prefetch(b->keys.set[i].tree);
1065		prefetch(b->keys.set[i].data);
1066	}
1067
1068	for (; i <= b->keys.nsets; i++)
1069		prefetch(b->keys.set[i].data);
1070
1071	return b;
1072}
1073
1074static void btree_node_prefetch(struct btree *parent, struct bkey *k)
1075{
1076	struct btree *b;
1077
1078	mutex_lock(&parent->c->bucket_lock);
1079	b = mca_alloc(parent->c, NULL, k, parent->level - 1);
1080	mutex_unlock(&parent->c->bucket_lock);
1081
1082	if (!IS_ERR_OR_NULL(b)) {
1083		b->parent = parent;
1084		bch_btree_node_read(b);
1085		rw_unlock(true, b);
1086	}
1087}
1088
1089/* Btree alloc */
1090
1091static void btree_node_free(struct btree *b)
1092{
1093	trace_bcache_btree_node_free(b);
1094
1095	BUG_ON(b == b->c->root);
1096
1097retry:
1098	mutex_lock(&b->write_lock);
1099	/*
1100	 * If the btree node is selected and flushing in btree_flush_write(),
1101	 * delay and retry until the BTREE_NODE_journal_flush bit cleared,
1102	 * then it is safe to free the btree node here. Otherwise this btree
1103	 * node will be in race condition.
1104	 */
1105	if (btree_node_journal_flush(b)) {
1106		mutex_unlock(&b->write_lock);
1107		pr_debug("bnode %p journal_flush set, retry", b);
1108		udelay(1);
1109		goto retry;
1110	}
1111
1112	if (btree_node_dirty(b)) {
1113		btree_complete_write(b, btree_current_write(b));
1114		clear_bit(BTREE_NODE_dirty, &b->flags);
1115	}
1116
1117	mutex_unlock(&b->write_lock);
1118
1119	cancel_delayed_work(&b->work);
1120
1121	mutex_lock(&b->c->bucket_lock);
1122	bch_bucket_free(b->c, &b->key);
1123	mca_bucket_free(b);
1124	mutex_unlock(&b->c->bucket_lock);
1125}
1126
1127struct btree *__bch_btree_node_alloc(struct cache_set *c, struct btree_op *op,
1128				     int level, bool wait,
1129				     struct btree *parent)
1130{
1131	BKEY_PADDED(key) k;
1132	struct btree *b = ERR_PTR(-EAGAIN);
1133
1134	mutex_lock(&c->bucket_lock);
1135retry:
1136	if (__bch_bucket_alloc_set(c, RESERVE_BTREE, &k.key, 1, wait))
1137		goto err;
1138
1139	bkey_put(c, &k.key);
1140	SET_KEY_SIZE(&k.key, c->btree_pages * PAGE_SECTORS);
1141
1142	b = mca_alloc(c, op, &k.key, level);
1143	if (IS_ERR(b))
1144		goto err_free;
1145
1146	if (!b) {
1147		cache_bug(c,
1148			"Tried to allocate bucket that was in btree cache");
1149		goto retry;
1150	}
1151
1152	b->accessed = 1;
1153	b->parent = parent;
1154	bch_bset_init_next(&b->keys, b->keys.set->data, bset_magic(&b->c->sb));
1155
1156	mutex_unlock(&c->bucket_lock);
1157
1158	trace_bcache_btree_node_alloc(b);
1159	return b;
1160err_free:
1161	bch_bucket_free(c, &k.key);
1162err:
1163	mutex_unlock(&c->bucket_lock);
1164
1165	trace_bcache_btree_node_alloc_fail(c);
1166	return b;
1167}
1168
1169static struct btree *bch_btree_node_alloc(struct cache_set *c,
1170					  struct btree_op *op, int level,
1171					  struct btree *parent)
1172{
1173	return __bch_btree_node_alloc(c, op, level, op != NULL, parent);
1174}
1175
1176static struct btree *btree_node_alloc_replacement(struct btree *b,
1177						  struct btree_op *op)
1178{
1179	struct btree *n = bch_btree_node_alloc(b->c, op, b->level, b->parent);
1180
1181	if (!IS_ERR_OR_NULL(n)) {
1182		mutex_lock(&n->write_lock);
1183		bch_btree_sort_into(&b->keys, &n->keys, &b->c->sort);
1184		bkey_copy_key(&n->key, &b->key);
1185		mutex_unlock(&n->write_lock);
1186	}
1187
1188	return n;
1189}
1190
1191static void make_btree_freeing_key(struct btree *b, struct bkey *k)
1192{
1193	unsigned int i;
1194
1195	mutex_lock(&b->c->bucket_lock);
1196
1197	atomic_inc(&b->c->prio_blocked);
1198
1199	bkey_copy(k, &b->key);
1200	bkey_copy_key(k, &ZERO_KEY);
1201
1202	for (i = 0; i < KEY_PTRS(k); i++)
1203		SET_PTR_GEN(k, i,
1204			    bch_inc_gen(PTR_CACHE(b->c, &b->key, i),
1205					PTR_BUCKET(b->c, &b->key, i)));
1206
1207	mutex_unlock(&b->c->bucket_lock);
1208}
1209
1210static int btree_check_reserve(struct btree *b, struct btree_op *op)
1211{
1212	struct cache_set *c = b->c;
1213	struct cache *ca;
1214	unsigned int i, reserve = (c->root->level - b->level) * 2 + 1;
1215
1216	mutex_lock(&c->bucket_lock);
1217
1218	for_each_cache(ca, c, i)
1219		if (fifo_used(&ca->free[RESERVE_BTREE]) < reserve) {
1220			if (op)
1221				prepare_to_wait(&c->btree_cache_wait, &op->wait,
1222						TASK_UNINTERRUPTIBLE);
1223			mutex_unlock(&c->bucket_lock);
1224			return -EINTR;
1225		}
1226
1227	mutex_unlock(&c->bucket_lock);
1228
1229	return mca_cannibalize_lock(b->c, op);
1230}
1231
1232/* Garbage collection */
1233
1234static uint8_t __bch_btree_mark_key(struct cache_set *c, int level,
1235				    struct bkey *k)
1236{
1237	uint8_t stale = 0;
1238	unsigned int i;
1239	struct bucket *g;
1240
1241	/*
1242	 * ptr_invalid() can't return true for the keys that mark btree nodes as
1243	 * freed, but since ptr_bad() returns true we'll never actually use them
1244	 * for anything and thus we don't want mark their pointers here
1245	 */
1246	if (!bkey_cmp(k, &ZERO_KEY))
1247		return stale;
1248
1249	for (i = 0; i < KEY_PTRS(k); i++) {
1250		if (!ptr_available(c, k, i))
1251			continue;
1252
1253		g = PTR_BUCKET(c, k, i);
1254
1255		if (gen_after(g->last_gc, PTR_GEN(k, i)))
1256			g->last_gc = PTR_GEN(k, i);
1257
1258		if (ptr_stale(c, k, i)) {
1259			stale = max(stale, ptr_stale(c, k, i));
1260			continue;
1261		}
1262
1263		cache_bug_on(GC_MARK(g) &&
1264			     (GC_MARK(g) == GC_MARK_METADATA) != (level != 0),
1265			     c, "inconsistent ptrs: mark = %llu, level = %i",
1266			     GC_MARK(g), level);
1267
1268		if (level)
1269			SET_GC_MARK(g, GC_MARK_METADATA);
1270		else if (KEY_DIRTY(k))
1271			SET_GC_MARK(g, GC_MARK_DIRTY);
1272		else if (!GC_MARK(g))
1273			SET_GC_MARK(g, GC_MARK_RECLAIMABLE);
1274
1275		/* guard against overflow */
1276		SET_GC_SECTORS_USED(g, min_t(unsigned int,
1277					     GC_SECTORS_USED(g) + KEY_SIZE(k),
1278					     MAX_GC_SECTORS_USED));
1279
1280		BUG_ON(!GC_SECTORS_USED(g));
1281	}
1282
1283	return stale;
1284}
1285
1286#define btree_mark_key(b, k)	__bch_btree_mark_key(b->c, b->level, k)
1287
1288void bch_initial_mark_key(struct cache_set *c, int level, struct bkey *k)
1289{
1290	unsigned int i;
1291
1292	for (i = 0; i < KEY_PTRS(k); i++)
1293		if (ptr_available(c, k, i) &&
1294		    !ptr_stale(c, k, i)) {
1295			struct bucket *b = PTR_BUCKET(c, k, i);
1296
1297			b->gen = PTR_GEN(k, i);
1298
1299			if (level && bkey_cmp(k, &ZERO_KEY))
1300				b->prio = BTREE_PRIO;
1301			else if (!level && b->prio == BTREE_PRIO)
1302				b->prio = INITIAL_PRIO;
1303		}
1304
1305	__bch_btree_mark_key(c, level, k);
1306}
1307
1308void bch_update_bucket_in_use(struct cache_set *c, struct gc_stat *stats)
1309{
1310	stats->in_use = (c->nbuckets - c->avail_nbuckets) * 100 / c->nbuckets;
1311}
1312
1313static bool btree_gc_mark_node(struct btree *b, struct gc_stat *gc)
1314{
1315	uint8_t stale = 0;
1316	unsigned int keys = 0, good_keys = 0;
1317	struct bkey *k;
1318	struct btree_iter iter;
1319	struct bset_tree *t;
1320
1321	gc->nodes++;
1322
1323	for_each_key_filter(&b->keys, k, &iter, bch_ptr_invalid) {
1324		stale = max(stale, btree_mark_key(b, k));
1325		keys++;
1326
1327		if (bch_ptr_bad(&b->keys, k))
1328			continue;
1329
1330		gc->key_bytes += bkey_u64s(k);
1331		gc->nkeys++;
1332		good_keys++;
1333
1334		gc->data += KEY_SIZE(k);
1335	}
1336
1337	for (t = b->keys.set; t <= &b->keys.set[b->keys.nsets]; t++)
1338		btree_bug_on(t->size &&
1339			     bset_written(&b->keys, t) &&
1340			     bkey_cmp(&b->key, &t->end) < 0,
1341			     b, "found short btree key in gc");
1342
1343	if (b->c->gc_always_rewrite)
1344		return true;
1345
1346	if (stale > 10)
1347		return true;
1348
1349	if ((keys - good_keys) * 2 > keys)
1350		return true;
1351
1352	return false;
1353}
1354
1355#define GC_MERGE_NODES	4U
1356
1357struct gc_merge_info {
1358	struct btree	*b;
1359	unsigned int	keys;
1360};
1361
1362static int bch_btree_insert_node(struct btree *b, struct btree_op *op,
1363				 struct keylist *insert_keys,
1364				 atomic_t *journal_ref,
1365				 struct bkey *replace_key);
1366
1367static int btree_gc_coalesce(struct btree *b, struct btree_op *op,
1368			     struct gc_stat *gc, struct gc_merge_info *r)
1369{
1370	unsigned int i, nodes = 0, keys = 0, blocks;
1371	struct btree *new_nodes[GC_MERGE_NODES];
1372	struct keylist keylist;
1373	struct closure cl;
1374	struct bkey *k;
1375
1376	bch_keylist_init(&keylist);
1377
1378	if (btree_check_reserve(b, NULL))
1379		return 0;
1380
1381	memset(new_nodes, 0, sizeof(new_nodes));
1382	closure_init_stack(&cl);
1383
1384	while (nodes < GC_MERGE_NODES && !IS_ERR_OR_NULL(r[nodes].b))
1385		keys += r[nodes++].keys;
1386
1387	blocks = btree_default_blocks(b->c) * 2 / 3;
1388
1389	if (nodes < 2 ||
1390	    __set_blocks(b->keys.set[0].data, keys,
1391			 block_bytes(b->c)) > blocks * (nodes - 1))
1392		return 0;
1393
1394	for (i = 0; i < nodes; i++) {
1395		new_nodes[i] = btree_node_alloc_replacement(r[i].b, NULL);
1396		if (IS_ERR_OR_NULL(new_nodes[i]))
1397			goto out_nocoalesce;
1398	}
1399
1400	/*
1401	 * We have to check the reserve here, after we've allocated our new
1402	 * nodes, to make sure the insert below will succeed - we also check
1403	 * before as an optimization to potentially avoid a bunch of expensive
1404	 * allocs/sorts
1405	 */
1406	if (btree_check_reserve(b, NULL))
1407		goto out_nocoalesce;
1408
1409	for (i = 0; i < nodes; i++)
1410		mutex_lock(&new_nodes[i]->write_lock);
1411
1412	for (i = nodes - 1; i > 0; --i) {
1413		struct bset *n1 = btree_bset_first(new_nodes[i]);
1414		struct bset *n2 = btree_bset_first(new_nodes[i - 1]);
1415		struct bkey *k, *last = NULL;
1416
1417		keys = 0;
1418
1419		if (i > 1) {
1420			for (k = n2->start;
1421			     k < bset_bkey_last(n2);
1422			     k = bkey_next(k)) {
1423				if (__set_blocks(n1, n1->keys + keys +
1424						 bkey_u64s(k),
1425						 block_bytes(b->c)) > blocks)
1426					break;
1427
1428				last = k;
1429				keys += bkey_u64s(k);
1430			}
1431		} else {
1432			/*
1433			 * Last node we're not getting rid of - we're getting
1434			 * rid of the node at r[0]. Have to try and fit all of
1435			 * the remaining keys into this node; we can't ensure
1436			 * they will always fit due to rounding and variable
1437			 * length keys (shouldn't be possible in practice,
1438			 * though)
1439			 */
1440			if (__set_blocks(n1, n1->keys + n2->keys,
1441					 block_bytes(b->c)) >
1442			    btree_blocks(new_nodes[i]))
1443				goto out_nocoalesce;
1444
1445			keys = n2->keys;
1446			/* Take the key of the node we're getting rid of */
1447			last = &r->b->key;
1448		}
1449
1450		BUG_ON(__set_blocks(n1, n1->keys + keys, block_bytes(b->c)) >
1451		       btree_blocks(new_nodes[i]));
1452
1453		if (last)
1454			bkey_copy_key(&new_nodes[i]->key, last);
1455
1456		memcpy(bset_bkey_last(n1),
1457		       n2->start,
1458		       (void *) bset_bkey_idx(n2, keys) - (void *) n2->start);
1459
1460		n1->keys += keys;
1461		r[i].keys = n1->keys;
1462
1463		memmove(n2->start,
1464			bset_bkey_idx(n2, keys),
1465			(void *) bset_bkey_last(n2) -
1466			(void *) bset_bkey_idx(n2, keys));
1467
1468		n2->keys -= keys;
1469
1470		if (__bch_keylist_realloc(&keylist,
1471					  bkey_u64s(&new_nodes[i]->key)))
1472			goto out_nocoalesce;
1473
1474		bch_btree_node_write(new_nodes[i], &cl);
1475		bch_keylist_add(&keylist, &new_nodes[i]->key);
1476	}
1477
1478	for (i = 0; i < nodes; i++)
1479		mutex_unlock(&new_nodes[i]->write_lock);
1480
1481	closure_sync(&cl);
1482
1483	/* We emptied out this node */
1484	BUG_ON(btree_bset_first(new_nodes[0])->keys);
1485	btree_node_free(new_nodes[0]);
1486	rw_unlock(true, new_nodes[0]);
1487	new_nodes[0] = NULL;
1488
1489	for (i = 0; i < nodes; i++) {
1490		if (__bch_keylist_realloc(&keylist, bkey_u64s(&r[i].b->key)))
1491			goto out_nocoalesce;
1492
1493		make_btree_freeing_key(r[i].b, keylist.top);
1494		bch_keylist_push(&keylist);
1495	}
1496
1497	bch_btree_insert_node(b, op, &keylist, NULL, NULL);
1498	BUG_ON(!bch_keylist_empty(&keylist));
1499
1500	for (i = 0; i < nodes; i++) {
1501		btree_node_free(r[i].b);
1502		rw_unlock(true, r[i].b);
1503
1504		r[i].b = new_nodes[i];
1505	}
1506
1507	memmove(r, r + 1, sizeof(r[0]) * (nodes - 1));
1508	r[nodes - 1].b = ERR_PTR(-EINTR);
1509
1510	trace_bcache_btree_gc_coalesce(nodes);
1511	gc->nodes--;
1512
1513	bch_keylist_free(&keylist);
1514
1515	/* Invalidated our iterator */
1516	return -EINTR;
1517
1518out_nocoalesce:
1519	closure_sync(&cl);
1520
1521	while ((k = bch_keylist_pop(&keylist)))
1522		if (!bkey_cmp(k, &ZERO_KEY))
1523			atomic_dec(&b->c->prio_blocked);
1524	bch_keylist_free(&keylist);
1525
1526	for (i = 0; i < nodes; i++)
1527		if (!IS_ERR_OR_NULL(new_nodes[i])) {
1528			btree_node_free(new_nodes[i]);
1529			rw_unlock(true, new_nodes[i]);
1530		}
1531	return 0;
1532}
1533
1534static int btree_gc_rewrite_node(struct btree *b, struct btree_op *op,
1535				 struct btree *replace)
1536{
1537	struct keylist keys;
1538	struct btree *n;
1539
1540	if (btree_check_reserve(b, NULL))
1541		return 0;
1542
1543	n = btree_node_alloc_replacement(replace, NULL);
1544
1545	/* recheck reserve after allocating replacement node */
1546	if (btree_check_reserve(b, NULL)) {
1547		btree_node_free(n);
1548		rw_unlock(true, n);
1549		return 0;
1550	}
1551
1552	bch_btree_node_write_sync(n);
1553
1554	bch_keylist_init(&keys);
1555	bch_keylist_add(&keys, &n->key);
1556
1557	make_btree_freeing_key(replace, keys.top);
1558	bch_keylist_push(&keys);
1559
1560	bch_btree_insert_node(b, op, &keys, NULL, NULL);
1561	BUG_ON(!bch_keylist_empty(&keys));
1562
1563	btree_node_free(replace);
1564	rw_unlock(true, n);
1565
1566	/* Invalidated our iterator */
1567	return -EINTR;
1568}
1569
1570static unsigned int btree_gc_count_keys(struct btree *b)
1571{
1572	struct bkey *k;
1573	struct btree_iter iter;
1574	unsigned int ret = 0;
1575
1576	for_each_key_filter(&b->keys, k, &iter, bch_ptr_bad)
1577		ret += bkey_u64s(k);
1578
1579	return ret;
1580}
1581
1582static size_t btree_gc_min_nodes(struct cache_set *c)
1583{
1584	size_t min_nodes;
1585
1586	/*
1587	 * Since incremental GC would stop 100ms when front
1588	 * side I/O comes, so when there are many btree nodes,
1589	 * if GC only processes constant (100) nodes each time,
1590	 * GC would last a long time, and the front side I/Os
1591	 * would run out of the buckets (since no new bucket
1592	 * can be allocated during GC), and be blocked again.
1593	 * So GC should not process constant nodes, but varied
1594	 * nodes according to the number of btree nodes, which
1595	 * realized by dividing GC into constant(100) times,
1596	 * so when there are many btree nodes, GC can process
1597	 * more nodes each time, otherwise, GC will process less
1598	 * nodes each time (but no less than MIN_GC_NODES)
1599	 */
1600	min_nodes = c->gc_stats.nodes / MAX_GC_TIMES;
1601	if (min_nodes < MIN_GC_NODES)
1602		min_nodes = MIN_GC_NODES;
1603
1604	return min_nodes;
1605}
1606
1607
1608static int btree_gc_recurse(struct btree *b, struct btree_op *op,
1609			    struct closure *writes, struct gc_stat *gc)
1610{
1611	int ret = 0;
1612	bool should_rewrite;
1613	struct bkey *k;
1614	struct btree_iter iter;
1615	struct gc_merge_info r[GC_MERGE_NODES];
1616	struct gc_merge_info *i, *last = r + ARRAY_SIZE(r) - 1;
1617
1618	bch_btree_iter_init(&b->keys, &iter, &b->c->gc_done);
1619
1620	for (i = r; i < r + ARRAY_SIZE(r); i++)
1621		i->b = ERR_PTR(-EINTR);
1622
1623	while (1) {
1624		k = bch_btree_iter_next_filter(&iter, &b->keys, bch_ptr_bad);
1625		if (k) {
1626			r->b = bch_btree_node_get(b->c, op, k, b->level - 1,
1627						  true, b);
1628			if (IS_ERR(r->b)) {
1629				ret = PTR_ERR(r->b);
1630				break;
1631			}
1632
1633			r->keys = btree_gc_count_keys(r->b);
1634
1635			ret = btree_gc_coalesce(b, op, gc, r);
1636			if (ret)
1637				break;
1638		}
1639
1640		if (!last->b)
1641			break;
1642
1643		if (!IS_ERR(last->b)) {
1644			should_rewrite = btree_gc_mark_node(last->b, gc);
1645			if (should_rewrite) {
1646				ret = btree_gc_rewrite_node(b, op, last->b);
1647				if (ret)
1648					break;
1649			}
1650
1651			if (last->b->level) {
1652				ret = btree_gc_recurse(last->b, op, writes, gc);
1653				if (ret)
1654					break;
1655			}
1656
1657			bkey_copy_key(&b->c->gc_done, &last->b->key);
1658
1659			/*
1660			 * Must flush leaf nodes before gc ends, since replace
1661			 * operations aren't journalled
1662			 */
1663			mutex_lock(&last->b->write_lock);
1664			if (btree_node_dirty(last->b))
1665				bch_btree_node_write(last->b, writes);
1666			mutex_unlock(&last->b->write_lock);
1667			rw_unlock(true, last->b);
1668		}
1669
1670		memmove(r + 1, r, sizeof(r[0]) * (GC_MERGE_NODES - 1));
1671		r->b = NULL;
1672
1673		if (atomic_read(&b->c->search_inflight) &&
1674		    gc->nodes >= gc->nodes_pre + btree_gc_min_nodes(b->c)) {
1675			gc->nodes_pre =  gc->nodes;
1676			ret = -EAGAIN;
1677			break;
1678		}
1679
1680		if (need_resched()) {
1681			ret = -EAGAIN;
1682			break;
1683		}
1684	}
1685
1686	for (i = r; i < r + ARRAY_SIZE(r); i++)
1687		if (!IS_ERR_OR_NULL(i->b)) {
1688			mutex_lock(&i->b->write_lock);
1689			if (btree_node_dirty(i->b))
1690				bch_btree_node_write(i->b, writes);
1691			mutex_unlock(&i->b->write_lock);
1692			rw_unlock(true, i->b);
1693		}
1694
1695	return ret;
1696}
1697
1698static int bch_btree_gc_root(struct btree *b, struct btree_op *op,
1699			     struct closure *writes, struct gc_stat *gc)
1700{
1701	struct btree *n = NULL;
1702	int ret = 0;
1703	bool should_rewrite;
1704
1705	should_rewrite = btree_gc_mark_node(b, gc);
1706	if (should_rewrite) {
1707		n = btree_node_alloc_replacement(b, NULL);
1708
1709		if (!IS_ERR_OR_NULL(n)) {
1710			bch_btree_node_write_sync(n);
1711
1712			bch_btree_set_root(n);
1713			btree_node_free(b);
1714			rw_unlock(true, n);
1715
1716			return -EINTR;
1717		}
1718	}
1719
1720	__bch_btree_mark_key(b->c, b->level + 1, &b->key);
1721
1722	if (b->level) {
1723		ret = btree_gc_recurse(b, op, writes, gc);
1724		if (ret)
1725			return ret;
1726	}
1727
1728	bkey_copy_key(&b->c->gc_done, &b->key);
1729
1730	return ret;
1731}
1732
1733static void btree_gc_start(struct cache_set *c)
1734{
1735	struct cache *ca;
1736	struct bucket *b;
1737	unsigned int i;
1738
1739	if (!c->gc_mark_valid)
1740		return;
1741
1742	mutex_lock(&c->bucket_lock);
1743
1744	c->gc_mark_valid = 0;
1745	c->gc_done = ZERO_KEY;
1746
1747	for_each_cache(ca, c, i)
1748		for_each_bucket(b, ca) {
1749			b->last_gc = b->gen;
1750			if (!atomic_read(&b->pin)) {
1751				SET_GC_MARK(b, 0);
1752				SET_GC_SECTORS_USED(b, 0);
1753			}
1754		}
1755
1756	mutex_unlock(&c->bucket_lock);
1757}
1758
1759static void bch_btree_gc_finish(struct cache_set *c)
1760{
1761	struct bucket *b;
1762	struct cache *ca;
1763	unsigned int i;
1764
1765	mutex_lock(&c->bucket_lock);
1766
1767	set_gc_sectors(c);
1768	c->gc_mark_valid = 1;
1769	c->need_gc	= 0;
1770
1771	for (i = 0; i < KEY_PTRS(&c->uuid_bucket); i++)
1772		SET_GC_MARK(PTR_BUCKET(c, &c->uuid_bucket, i),
1773			    GC_MARK_METADATA);
1774
1775	/* don't reclaim buckets to which writeback keys point */
1776	rcu_read_lock();
1777	for (i = 0; i < c->devices_max_used; i++) {
1778		struct bcache_device *d = c->devices[i];
1779		struct cached_dev *dc;
1780		struct keybuf_key *w, *n;
1781		unsigned int j;
1782
1783		if (!d || UUID_FLASH_ONLY(&c->uuids[i]))
1784			continue;
1785		dc = container_of(d, struct cached_dev, disk);
1786
1787		spin_lock(&dc->writeback_keys.lock);
1788		rbtree_postorder_for_each_entry_safe(w, n,
1789					&dc->writeback_keys.keys, node)
1790			for (j = 0; j < KEY_PTRS(&w->key); j++)
1791				SET_GC_MARK(PTR_BUCKET(c, &w->key, j),
1792					    GC_MARK_DIRTY);
1793		spin_unlock(&dc->writeback_keys.lock);
1794	}
1795	rcu_read_unlock();
1796
1797	c->avail_nbuckets = 0;
1798	for_each_cache(ca, c, i) {
1799		uint64_t *i;
1800
1801		ca->invalidate_needs_gc = 0;
1802
1803		for (i = ca->sb.d; i < ca->sb.d + ca->sb.keys; i++)
1804			SET_GC_MARK(ca->buckets + *i, GC_MARK_METADATA);
1805
1806		for (i = ca->prio_buckets;
1807		     i < ca->prio_buckets + prio_buckets(ca) * 2; i++)
1808			SET_GC_MARK(ca->buckets + *i, GC_MARK_METADATA);
1809
1810		for_each_bucket(b, ca) {
1811			c->need_gc	= max(c->need_gc, bucket_gc_gen(b));
1812
1813			if (atomic_read(&b->pin))
1814				continue;
1815
1816			BUG_ON(!GC_MARK(b) && GC_SECTORS_USED(b));
1817
1818			if (!GC_MARK(b) || GC_MARK(b) == GC_MARK_RECLAIMABLE)
1819				c->avail_nbuckets++;
1820		}
1821	}
1822
1823	mutex_unlock(&c->bucket_lock);
1824}
1825
1826static void bch_btree_gc(struct cache_set *c)
1827{
1828	int ret;
1829	struct gc_stat stats;
1830	struct closure writes;
1831	struct btree_op op;
1832	uint64_t start_time = local_clock();
1833
1834	trace_bcache_gc_start(c);
1835
1836	memset(&stats, 0, sizeof(struct gc_stat));
1837	closure_init_stack(&writes);
1838	bch_btree_op_init(&op, SHRT_MAX);
1839
1840	btree_gc_start(c);
1841
1842	/* if CACHE_SET_IO_DISABLE set, gc thread should stop too */
1843	do {
1844		ret = btree_root(gc_root, c, &op, &writes, &stats);
1845		closure_sync(&writes);
1846		cond_resched();
1847
1848		if (ret == -EAGAIN)
1849			schedule_timeout_interruptible(msecs_to_jiffies
1850						       (GC_SLEEP_MS));
1851		else if (ret)
1852			pr_warn("gc failed!");
1853	} while (ret && !test_bit(CACHE_SET_IO_DISABLE, &c->flags));
1854
1855	bch_btree_gc_finish(c);
1856	wake_up_allocators(c);
1857
1858	bch_time_stats_update(&c->btree_gc_time, start_time);
1859
1860	stats.key_bytes *= sizeof(uint64_t);
1861	stats.data	<<= 9;
1862	bch_update_bucket_in_use(c, &stats);
1863	memcpy(&c->gc_stats, &stats, sizeof(struct gc_stat));
1864
1865	trace_bcache_gc_end(c);
1866
1867	bch_moving_gc(c);
1868}
1869
1870static bool gc_should_run(struct cache_set *c)
1871{
1872	struct cache *ca;
1873	unsigned int i;
1874
1875	for_each_cache(ca, c, i)
1876		if (ca->invalidate_needs_gc)
1877			return true;
1878
1879	if (atomic_read(&c->sectors_to_gc) < 0)
1880		return true;
1881
1882	return false;
1883}
1884
1885static int bch_gc_thread(void *arg)
1886{
1887	struct cache_set *c = arg;
1888
1889	while (1) {
1890		wait_event_interruptible(c->gc_wait,
1891			   kthread_should_stop() ||
1892			   test_bit(CACHE_SET_IO_DISABLE, &c->flags) ||
1893			   gc_should_run(c));
1894
1895		if (kthread_should_stop() ||
1896		    test_bit(CACHE_SET_IO_DISABLE, &c->flags))
1897			break;
1898
1899		set_gc_sectors(c);
1900		bch_btree_gc(c);
1901	}
1902
1903	wait_for_kthread_stop();
1904	return 0;
1905}
1906
1907int bch_gc_thread_start(struct cache_set *c)
1908{
1909	c->gc_thread = kthread_run(bch_gc_thread, c, "bcache_gc");
1910	return PTR_ERR_OR_ZERO(c->gc_thread);
1911}
1912
1913/* Initial partial gc */
1914
1915static int bch_btree_check_recurse(struct btree *b, struct btree_op *op)
1916{
1917	int ret = 0;
1918	struct bkey *k, *p = NULL;
1919	struct btree_iter iter;
1920
1921	for_each_key_filter(&b->keys, k, &iter, bch_ptr_invalid)
1922		bch_initial_mark_key(b->c, b->level, k);
1923
1924	bch_initial_mark_key(b->c, b->level + 1, &b->key);
1925
1926	if (b->level) {
1927		bch_btree_iter_init(&b->keys, &iter, NULL);
1928
1929		do {
1930			k = bch_btree_iter_next_filter(&iter, &b->keys,
1931						       bch_ptr_bad);
1932			if (k) {
1933				btree_node_prefetch(b, k);
1934				/*
1935				 * initiallize c->gc_stats.nodes
1936				 * for incremental GC
1937				 */
1938				b->c->gc_stats.nodes++;
1939			}
1940
1941			if (p)
1942				ret = btree(check_recurse, p, b, op);
1943
1944			p = k;
1945		} while (p && !ret);
1946	}
1947
1948	return ret;
1949}
1950
1951int bch_btree_check(struct cache_set *c)
1952{
1953	struct btree_op op;
1954
1955	bch_btree_op_init(&op, SHRT_MAX);
1956
1957	return btree_root(check_recurse, c, &op);
1958}
1959
1960void bch_initial_gc_finish(struct cache_set *c)
1961{
1962	struct cache *ca;
1963	struct bucket *b;
1964	unsigned int i;
1965
1966	bch_btree_gc_finish(c);
1967
1968	mutex_lock(&c->bucket_lock);
1969
1970	/*
1971	 * We need to put some unused buckets directly on the prio freelist in
1972	 * order to get the allocator thread started - it needs freed buckets in
1973	 * order to rewrite the prios and gens, and it needs to rewrite prios
1974	 * and gens in order to free buckets.
1975	 *
1976	 * This is only safe for buckets that have no live data in them, which
1977	 * there should always be some of.
1978	 */
1979	for_each_cache(ca, c, i) {
1980		for_each_bucket(b, ca) {
1981			if (fifo_full(&ca->free[RESERVE_PRIO]) &&
1982			    fifo_full(&ca->free[RESERVE_BTREE]))
1983				break;
1984
1985			if (bch_can_invalidate_bucket(ca, b) &&
1986			    !GC_MARK(b)) {
1987				__bch_invalidate_one_bucket(ca, b);
1988				if (!fifo_push(&ca->free[RESERVE_PRIO],
1989				   b - ca->buckets))
1990					fifo_push(&ca->free[RESERVE_BTREE],
1991						  b - ca->buckets);
1992			}
1993		}
1994	}
1995
1996	mutex_unlock(&c->bucket_lock);
1997}
1998
1999/* Btree insertion */
2000
2001static bool btree_insert_key(struct btree *b, struct bkey *k,
2002			     struct bkey *replace_key)
2003{
2004	unsigned int status;
2005
2006	BUG_ON(bkey_cmp(k, &b->key) > 0);
2007
2008	status = bch_btree_insert_key(&b->keys, k, replace_key);
2009	if (status != BTREE_INSERT_STATUS_NO_INSERT) {
2010		bch_check_keys(&b->keys, "%u for %s", status,
2011			       replace_key ? "replace" : "insert");
2012
2013		trace_bcache_btree_insert_key(b, k, replace_key != NULL,
2014					      status);
2015		return true;
2016	} else
2017		return false;
2018}
2019
2020static size_t insert_u64s_remaining(struct btree *b)
2021{
2022	long ret = bch_btree_keys_u64s_remaining(&b->keys);
2023
2024	/*
2025	 * Might land in the middle of an existing extent and have to split it
2026	 */
2027	if (b->keys.ops->is_extents)
2028		ret -= KEY_MAX_U64S;
2029
2030	return max(ret, 0L);
2031}
2032
2033static bool bch_btree_insert_keys(struct btree *b, struct btree_op *op,
2034				  struct keylist *insert_keys,
2035				  struct bkey *replace_key)
2036{
2037	bool ret = false;
2038	int oldsize = bch_count_data(&b->keys);
2039
2040	while (!bch_keylist_empty(insert_keys)) {
2041		struct bkey *k = insert_keys->keys;
2042
2043		if (bkey_u64s(k) > insert_u64s_remaining(b))
2044			break;
2045
2046		if (bkey_cmp(k, &b->key) <= 0) {
2047			if (!b->level)
2048				bkey_put(b->c, k);
2049
2050			ret |= btree_insert_key(b, k, replace_key);
2051			bch_keylist_pop_front(insert_keys);
2052		} else if (bkey_cmp(&START_KEY(k), &b->key) < 0) {
2053			BKEY_PADDED(key) temp;
2054			bkey_copy(&temp.key, insert_keys->keys);
2055
2056			bch_cut_back(&b->key, &temp.key);
2057			bch_cut_front(&b->key, insert_keys->keys);
2058
2059			ret |= btree_insert_key(b, &temp.key, replace_key);
2060			break;
2061		} else {
2062			break;
2063		}
2064	}
2065
2066	if (!ret)
2067		op->insert_collision = true;
2068
2069	BUG_ON(!bch_keylist_empty(insert_keys) && b->level);
2070
2071	BUG_ON(bch_count_data(&b->keys) < oldsize);
2072	return ret;
2073}
2074
2075static int btree_split(struct btree *b, struct btree_op *op,
2076		       struct keylist *insert_keys,
2077		       struct bkey *replace_key)
2078{
2079	bool split;
2080	struct btree *n1, *n2 = NULL, *n3 = NULL;
2081	uint64_t start_time = local_clock();
2082	struct closure cl;
2083	struct keylist parent_keys;
2084
2085	closure_init_stack(&cl);
2086	bch_keylist_init(&parent_keys);
2087
2088	if (btree_check_reserve(b, op)) {
2089		if (!b->level)
2090			return -EINTR;
2091		else
2092			WARN(1, "insufficient reserve for split\n");
2093	}
2094
2095	n1 = btree_node_alloc_replacement(b, op);
2096	if (IS_ERR(n1))
2097		goto err;
2098
2099	split = set_blocks(btree_bset_first(n1),
2100			   block_bytes(n1->c)) > (btree_blocks(b) * 4) / 5;
2101
2102	if (split) {
2103		unsigned int keys = 0;
2104
2105		trace_bcache_btree_node_split(b, btree_bset_first(n1)->keys);
2106
2107		n2 = bch_btree_node_alloc(b->c, op, b->level, b->parent);
2108		if (IS_ERR(n2))
2109			goto err_free1;
2110
2111		if (!b->parent) {
2112			n3 = bch_btree_node_alloc(b->c, op, b->level + 1, NULL);
2113			if (IS_ERR(n3))
2114				goto err_free2;
2115		}
2116
2117		mutex_lock(&n1->write_lock);
2118		mutex_lock(&n2->write_lock);
2119
2120		bch_btree_insert_keys(n1, op, insert_keys, replace_key);
2121
2122		/*
2123		 * Has to be a linear search because we don't have an auxiliary
2124		 * search tree yet
2125		 */
2126
2127		while (keys < (btree_bset_first(n1)->keys * 3) / 5)
2128			keys += bkey_u64s(bset_bkey_idx(btree_bset_first(n1),
2129							keys));
2130
2131		bkey_copy_key(&n1->key,
2132			      bset_bkey_idx(btree_bset_first(n1), keys));
2133		keys += bkey_u64s(bset_bkey_idx(btree_bset_first(n1), keys));
2134
2135		btree_bset_first(n2)->keys = btree_bset_first(n1)->keys - keys;
2136		btree_bset_first(n1)->keys = keys;
2137
2138		memcpy(btree_bset_first(n2)->start,
2139		       bset_bkey_last(btree_bset_first(n1)),
2140		       btree_bset_first(n2)->keys * sizeof(uint64_t));
2141
2142		bkey_copy_key(&n2->key, &b->key);
2143
2144		bch_keylist_add(&parent_keys, &n2->key);
2145		bch_btree_node_write(n2, &cl);
2146		mutex_unlock(&n2->write_lock);
2147		rw_unlock(true, n2);
2148	} else {
2149		trace_bcache_btree_node_compact(b, btree_bset_first(n1)->keys);
2150
2151		mutex_lock(&n1->write_lock);
2152		bch_btree_insert_keys(n1, op, insert_keys, replace_key);
2153	}
2154
2155	bch_keylist_add(&parent_keys, &n1->key);
2156	bch_btree_node_write(n1, &cl);
2157	mutex_unlock(&n1->write_lock);
2158
2159	if (n3) {
2160		/* Depth increases, make a new root */
2161		mutex_lock(&n3->write_lock);
2162		bkey_copy_key(&n3->key, &MAX_KEY);
2163		bch_btree_insert_keys(n3, op, &parent_keys, NULL);
2164		bch_btree_node_write(n3, &cl);
2165		mutex_unlock(&n3->write_lock);
2166
2167		closure_sync(&cl);
2168		bch_btree_set_root(n3);
2169		rw_unlock(true, n3);
2170	} else if (!b->parent) {
2171		/* Root filled up but didn't need to be split */
2172		closure_sync(&cl);
2173		bch_btree_set_root(n1);
2174	} else {
2175		/* Split a non root node */
2176		closure_sync(&cl);
2177		make_btree_freeing_key(b, parent_keys.top);
2178		bch_keylist_push(&parent_keys);
2179
2180		bch_btree_insert_node(b->parent, op, &parent_keys, NULL, NULL);
2181		BUG_ON(!bch_keylist_empty(&parent_keys));
2182	}
2183
2184	btree_node_free(b);
2185	rw_unlock(true, n1);
2186
2187	bch_time_stats_update(&b->c->btree_split_time, start_time);
2188
2189	return 0;
2190err_free2:
2191	bkey_put(b->c, &n2->key);
2192	btree_node_free(n2);
2193	rw_unlock(true, n2);
2194err_free1:
2195	bkey_put(b->c, &n1->key);
2196	btree_node_free(n1);
2197	rw_unlock(true, n1);
2198err:
2199	WARN(1, "bcache: btree split failed (level %u)", b->level);
2200
2201	if (n3 == ERR_PTR(-EAGAIN) ||
2202	    n2 == ERR_PTR(-EAGAIN) ||
2203	    n1 == ERR_PTR(-EAGAIN))
2204		return -EAGAIN;
2205
2206	return -ENOMEM;
2207}
2208
2209static int bch_btree_insert_node(struct btree *b, struct btree_op *op,
2210				 struct keylist *insert_keys,
2211				 atomic_t *journal_ref,
2212				 struct bkey *replace_key)
2213{
2214	struct closure cl;
2215
2216	BUG_ON(b->level && replace_key);
2217
2218	closure_init_stack(&cl);
2219
2220	mutex_lock(&b->write_lock);
2221
2222	if (write_block(b) != btree_bset_last(b) &&
2223	    b->keys.last_set_unwritten)
2224		bch_btree_init_next(b); /* just wrote a set */
2225
2226	if (bch_keylist_nkeys(insert_keys) > insert_u64s_remaining(b)) {
2227		mutex_unlock(&b->write_lock);
2228		goto split;
2229	}
2230
2231	BUG_ON(write_block(b) != btree_bset_last(b));
2232
2233	if (bch_btree_insert_keys(b, op, insert_keys, replace_key)) {
2234		if (!b->level)
2235			bch_btree_leaf_dirty(b, journal_ref);
2236		else
2237			bch_btree_node_write(b, &cl);
2238	}
2239
2240	mutex_unlock(&b->write_lock);
2241
2242	/* wait for btree node write if necessary, after unlock */
2243	closure_sync(&cl);
2244
2245	return 0;
2246split:
2247	if (current->bio_list) {
2248		op->lock = b->c->root->level + 1;
2249		return -EAGAIN;
2250	} else if (op->lock <= b->c->root->level) {
2251		op->lock = b->c->root->level + 1;
2252		return -EINTR;
2253	} else {
2254		/* Invalidated all iterators */
2255		int ret = btree_split(b, op, insert_keys, replace_key);
2256
2257		if (bch_keylist_empty(insert_keys))
2258			return 0;
2259		else if (!ret)
2260			return -EINTR;
2261		return ret;
2262	}
2263}
2264
2265int bch_btree_insert_check_key(struct btree *b, struct btree_op *op,
2266			       struct bkey *check_key)
2267{
2268	int ret = -EINTR;
2269	uint64_t btree_ptr = b->key.ptr[0];
2270	unsigned long seq = b->seq;
2271	struct keylist insert;
2272	bool upgrade = op->lock == -1;
2273
2274	bch_keylist_init(&insert);
2275
2276	if (upgrade) {
2277		rw_unlock(false, b);
2278		rw_lock(true, b, b->level);
2279
2280		if (b->key.ptr[0] != btree_ptr ||
2281		    b->seq != seq + 1) {
2282			op->lock = b->level;
2283			goto out;
2284		}
2285	}
2286
2287	SET_KEY_PTRS(check_key, 1);
2288	get_random_bytes(&check_key->ptr[0], sizeof(uint64_t));
2289
2290	SET_PTR_DEV(check_key, 0, PTR_CHECK_DEV);
2291
2292	bch_keylist_add(&insert, check_key);
2293
2294	ret = bch_btree_insert_node(b, op, &insert, NULL, NULL);
2295
2296	BUG_ON(!ret && !bch_keylist_empty(&insert));
2297out:
2298	if (upgrade)
2299		downgrade_write(&b->lock);
2300	return ret;
2301}
2302
2303struct btree_insert_op {
2304	struct btree_op	op;
2305	struct keylist	*keys;
2306	atomic_t	*journal_ref;
2307	struct bkey	*replace_key;
2308};
2309
2310static int btree_insert_fn(struct btree_op *b_op, struct btree *b)
2311{
2312	struct btree_insert_op *op = container_of(b_op,
2313					struct btree_insert_op, op);
2314
2315	int ret = bch_btree_insert_node(b, &op->op, op->keys,
2316					op->journal_ref, op->replace_key);
2317	if (ret && !bch_keylist_empty(op->keys))
2318		return ret;
2319	else
2320		return MAP_DONE;
2321}
2322
2323int bch_btree_insert(struct cache_set *c, struct keylist *keys,
2324		     atomic_t *journal_ref, struct bkey *replace_key)
2325{
2326	struct btree_insert_op op;
2327	int ret = 0;
2328
2329	BUG_ON(current->bio_list);
2330	BUG_ON(bch_keylist_empty(keys));
2331
2332	bch_btree_op_init(&op.op, 0);
2333	op.keys		= keys;
2334	op.journal_ref	= journal_ref;
2335	op.replace_key	= replace_key;
2336
2337	while (!ret && !bch_keylist_empty(keys)) {
2338		op.op.lock = 0;
2339		ret = bch_btree_map_leaf_nodes(&op.op, c,
2340					       &START_KEY(keys->keys),
2341					       btree_insert_fn);
2342	}
2343
2344	if (ret) {
2345		struct bkey *k;
2346
2347		pr_err("error %i", ret);
2348
2349		while ((k = bch_keylist_pop(keys)))
2350			bkey_put(c, k);
2351	} else if (op.op.insert_collision)
2352		ret = -ESRCH;
2353
2354	return ret;
2355}
2356
2357void bch_btree_set_root(struct btree *b)
2358{
2359	unsigned int i;
2360	struct closure cl;
2361
2362	closure_init_stack(&cl);
2363
2364	trace_bcache_btree_set_root(b);
2365
2366	BUG_ON(!b->written);
2367
2368	for (i = 0; i < KEY_PTRS(&b->key); i++)
2369		BUG_ON(PTR_BUCKET(b->c, &b->key, i)->prio != BTREE_PRIO);
2370
2371	mutex_lock(&b->c->bucket_lock);
2372	list_del_init(&b->list);
2373	mutex_unlock(&b->c->bucket_lock);
2374
2375	b->c->root = b;
2376
2377	bch_journal_meta(b->c, &cl);
2378	closure_sync(&cl);
2379}
2380
2381/* Map across nodes or keys */
2382
2383static int bch_btree_map_nodes_recurse(struct btree *b, struct btree_op *op,
2384				       struct bkey *from,
2385				       btree_map_nodes_fn *fn, int flags)
2386{
2387	int ret = MAP_CONTINUE;
2388
2389	if (b->level) {
2390		struct bkey *k;
2391		struct btree_iter iter;
2392
2393		bch_btree_iter_init(&b->keys, &iter, from);
2394
2395		while ((k = bch_btree_iter_next_filter(&iter, &b->keys,
2396						       bch_ptr_bad))) {
2397			ret = btree(map_nodes_recurse, k, b,
2398				    op, from, fn, flags);
2399			from = NULL;
2400
2401			if (ret != MAP_CONTINUE)
2402				return ret;
2403		}
2404	}
2405
2406	if (!b->level || flags == MAP_ALL_NODES)
2407		ret = fn(op, b);
2408
2409	return ret;
2410}
2411
2412int __bch_btree_map_nodes(struct btree_op *op, struct cache_set *c,
2413			  struct bkey *from, btree_map_nodes_fn *fn, int flags)
2414{
2415	return btree_root(map_nodes_recurse, c, op, from, fn, flags);
2416}
2417
2418static int bch_btree_map_keys_recurse(struct btree *b, struct btree_op *op,
2419				      struct bkey *from, btree_map_keys_fn *fn,
2420				      int flags)
2421{
2422	int ret = MAP_CONTINUE;
2423	struct bkey *k;
2424	struct btree_iter iter;
2425
2426	bch_btree_iter_init(&b->keys, &iter, from);
2427
2428	while ((k = bch_btree_iter_next_filter(&iter, &b->keys, bch_ptr_bad))) {
2429		ret = !b->level
2430			? fn(op, b, k)
2431			: btree(map_keys_recurse, k, b, op, from, fn, flags);
2432		from = NULL;
2433
2434		if (ret != MAP_CONTINUE)
2435			return ret;
2436	}
2437
2438	if (!b->level && (flags & MAP_END_KEY))
2439		ret = fn(op, b, &KEY(KEY_INODE(&b->key),
2440				     KEY_OFFSET(&b->key), 0));
2441
2442	return ret;
2443}
2444
2445int bch_btree_map_keys(struct btree_op *op, struct cache_set *c,
2446		       struct bkey *from, btree_map_keys_fn *fn, int flags)
2447{
2448	return btree_root(map_keys_recurse, c, op, from, fn, flags);
2449}
2450
2451/* Keybuf code */
2452
2453static inline int keybuf_cmp(struct keybuf_key *l, struct keybuf_key *r)
2454{
2455	/* Overlapping keys compare equal */
2456	if (bkey_cmp(&l->key, &START_KEY(&r->key)) <= 0)
2457		return -1;
2458	if (bkey_cmp(&START_KEY(&l->key), &r->key) >= 0)
2459		return 1;
2460	return 0;
2461}
2462
2463static inline int keybuf_nonoverlapping_cmp(struct keybuf_key *l,
2464					    struct keybuf_key *r)
2465{
2466	return clamp_t(int64_t, bkey_cmp(&l->key, &r->key), -1, 1);
2467}
2468
2469struct refill {
2470	struct btree_op	op;
2471	unsigned int	nr_found;
2472	struct keybuf	*buf;
2473	struct bkey	*end;
2474	keybuf_pred_fn	*pred;
2475};
2476
2477static int refill_keybuf_fn(struct btree_op *op, struct btree *b,
2478			    struct bkey *k)
2479{
2480	struct refill *refill = container_of(op, struct refill, op);
2481	struct keybuf *buf = refill->buf;
2482	int ret = MAP_CONTINUE;
2483
2484	if (bkey_cmp(k, refill->end) > 0) {
2485		ret = MAP_DONE;
2486		goto out;
2487	}
2488
2489	if (!KEY_SIZE(k)) /* end key */
2490		goto out;
2491
2492	if (refill->pred(buf, k)) {
2493		struct keybuf_key *w;
2494
2495		spin_lock(&buf->lock);
2496
2497		w = array_alloc(&buf->freelist);
2498		if (!w) {
2499			spin_unlock(&buf->lock);
2500			return MAP_DONE;
2501		}
2502
2503		w->private = NULL;
2504		bkey_copy(&w->key, k);
2505
2506		if (RB_INSERT(&buf->keys, w, node, keybuf_cmp))
2507			array_free(&buf->freelist, w);
2508		else
2509			refill->nr_found++;
2510
2511		if (array_freelist_empty(&buf->freelist))
2512			ret = MAP_DONE;
2513
2514		spin_unlock(&buf->lock);
2515	}
2516out:
2517	buf->last_scanned = *k;
2518	return ret;
2519}
2520
2521void bch_refill_keybuf(struct cache_set *c, struct keybuf *buf,
2522		       struct bkey *end, keybuf_pred_fn *pred)
2523{
2524	struct bkey start = buf->last_scanned;
2525	struct refill refill;
2526
2527	cond_resched();
2528
2529	bch_btree_op_init(&refill.op, -1);
2530	refill.nr_found	= 0;
2531	refill.buf	= buf;
2532	refill.end	= end;
2533	refill.pred	= pred;
2534
2535	bch_btree_map_keys(&refill.op, c, &buf->last_scanned,
2536			   refill_keybuf_fn, MAP_END_KEY);
2537
2538	trace_bcache_keyscan(refill.nr_found,
2539			     KEY_INODE(&start), KEY_OFFSET(&start),
2540			     KEY_INODE(&buf->last_scanned),
2541			     KEY_OFFSET(&buf->last_scanned));
2542
2543	spin_lock(&buf->lock);
2544
2545	if (!RB_EMPTY_ROOT(&buf->keys)) {
2546		struct keybuf_key *w;
2547
2548		w = RB_FIRST(&buf->keys, struct keybuf_key, node);
2549		buf->start	= START_KEY(&w->key);
2550
2551		w = RB_LAST(&buf->keys, struct keybuf_key, node);
2552		buf->end	= w->key;
2553	} else {
2554		buf->start	= MAX_KEY;
2555		buf->end	= MAX_KEY;
2556	}
2557
2558	spin_unlock(&buf->lock);
2559}
2560
2561static void __bch_keybuf_del(struct keybuf *buf, struct keybuf_key *w)
2562{
2563	rb_erase(&w->node, &buf->keys);
2564	array_free(&buf->freelist, w);
2565}
2566
2567void bch_keybuf_del(struct keybuf *buf, struct keybuf_key *w)
2568{
2569	spin_lock(&buf->lock);
2570	__bch_keybuf_del(buf, w);
2571	spin_unlock(&buf->lock);
2572}
2573
2574bool bch_keybuf_check_overlapping(struct keybuf *buf, struct bkey *start,
2575				  struct bkey *end)
2576{
2577	bool ret = false;
2578	struct keybuf_key *p, *w, s;
2579
2580	s.key = *start;
2581
2582	if (bkey_cmp(end, &buf->start) <= 0 ||
2583	    bkey_cmp(start, &buf->end) >= 0)
2584		return false;
2585
2586	spin_lock(&buf->lock);
2587	w = RB_GREATER(&buf->keys, s, node, keybuf_nonoverlapping_cmp);
2588
2589	while (w && bkey_cmp(&START_KEY(&w->key), end) < 0) {
2590		p = w;
2591		w = RB_NEXT(w, node);
2592
2593		if (p->private)
2594			ret = true;
2595		else
2596			__bch_keybuf_del(buf, p);
2597	}
2598
2599	spin_unlock(&buf->lock);
2600	return ret;
2601}
2602
2603struct keybuf_key *bch_keybuf_next(struct keybuf *buf)
2604{
2605	struct keybuf_key *w;
2606
2607	spin_lock(&buf->lock);
2608
2609	w = RB_FIRST(&buf->keys, struct keybuf_key, node);
2610
2611	while (w && w->private)
2612		w = RB_NEXT(w, node);
2613
2614	if (w)
2615		w->private = ERR_PTR(-EINTR);
2616
2617	spin_unlock(&buf->lock);
2618	return w;
2619}
2620
2621struct keybuf_key *bch_keybuf_next_rescan(struct cache_set *c,
2622					  struct keybuf *buf,
2623					  struct bkey *end,
2624					  keybuf_pred_fn *pred)
2625{
2626	struct keybuf_key *ret;
2627
2628	while (1) {
2629		ret = bch_keybuf_next(buf);
2630		if (ret)
2631			break;
2632
2633		if (bkey_cmp(&buf->last_scanned, end) >= 0) {
2634			pr_debug("scan finished");
2635			break;
2636		}
2637
2638		bch_refill_keybuf(c, buf, end, pred);
2639	}
2640
2641	return ret;
2642}
2643
2644void bch_keybuf_init(struct keybuf *buf)
2645{
2646	buf->last_scanned	= MAX_KEY;
2647	buf->keys		= RB_ROOT;
2648
2649	spin_lock_init(&buf->lock);
2650	array_allocator_init(&buf->freelist);
2651}