fs/bcachefs/btree_gc.c

// SPDX-License-Identifier: GPL-2.0
/*
 * Copyright (C) 2010 Kent Overstreet <kent.overstreet@gmail.com>
 * Copyright (C) 2014 Datera Inc.
 */

#include "bcachefs.h"
#include "alloc_background.h"
#include "alloc_foreground.h"
#include "bkey_methods.h"
#include "btree_locking.h"
#include "btree_update_interior.h"
#include "btree_io.h"
#include "btree_gc.h"
#include "buckets.h"
#include "clock.h"
#include "debug.h"
#include "ec.h"
#include "error.h"
#include "extents.h"
#include "journal.h"
#include "journal_io.h"
#include "keylist.h"
#include "move.h"
#include "replicas.h"
#include "super-io.h"
#include "trace.h"

#include <linux/slab.h>
#include <linux/bitops.h>
#include <linux/freezer.h>
#include <linux/kthread.h>
#include <linux/preempt.h>
#include <linux/rcupdate.h>
#include <linux/sched/task.h>

static inline void __gc_pos_set(struct bch_fs *c, struct gc_pos new_pos)
{
	preempt_disable();
	write_seqcount_begin(&c->gc_pos_lock);
	c->gc_pos = new_pos;
	write_seqcount_end(&c->gc_pos_lock);
	preempt_enable();
}

static inline void gc_pos_set(struct bch_fs *c, struct gc_pos new_pos)
{
	BUG_ON(gc_pos_cmp(new_pos, c->gc_pos) <= 0);
	__gc_pos_set(c, new_pos);
}

/* range_checks - for validating min/max pos of each btree node: */

struct range_checks {
	struct range_level {
		struct bpos	min;
		struct bpos	max;
	}			l[BTREE_MAX_DEPTH];
	unsigned		depth;
};

static void btree_node_range_checks_init(struct range_checks *r, unsigned depth)
{
	unsigned i;

	for (i = 0; i < BTREE_MAX_DEPTH; i++)
		r->l[i].min = r->l[i].max = POS_MIN;
	r->depth = depth;
}

static void btree_node_range_checks(struct bch_fs *c, struct btree *b,
				    struct range_checks *r)
{
	struct range_level *l = &r->l[b->level];

	struct bpos expected_min = bkey_cmp(l->min, l->max)
		? btree_type_successor(b->btree_id, l->max)
		: l->max;

	bch2_fs_inconsistent_on(bkey_cmp(b->data->min_key, expected_min), c,
		"btree node has incorrect min key: %llu:%llu != %llu:%llu",
		b->data->min_key.inode,
		b->data->min_key.offset,
		expected_min.inode,
		expected_min.offset);

	l->max = b->data->max_key;

	if (b->level > r->depth) {
		l = &r->l[b->level - 1];

		bch2_fs_inconsistent_on(bkey_cmp(b->data->min_key, l->min), c,
			"btree node min doesn't match min of child nodes: %llu:%llu != %llu:%llu",
			b->data->min_key.inode,
			b->data->min_key.offset,
			l->min.inode,
			l->min.offset);

		bch2_fs_inconsistent_on(bkey_cmp(b->data->max_key, l->max), c,
			"btree node max doesn't match max of child nodes: %llu:%llu != %llu:%llu",
			b->data->max_key.inode,
			b->data->max_key.offset,
			l->max.inode,
			l->max.offset);

		if (bkey_cmp(b->data->max_key, POS_MAX))
			l->min = l->max =
				btree_type_successor(b->btree_id,
						     b->data->max_key);
	}
}

/* marking of btree keys/nodes: */

static int bch2_gc_mark_key(struct bch_fs *c, struct bkey_s_c k,
			    u8 *max_stale, bool initial)
{
	struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
	const struct bch_extent_ptr *ptr;
	struct gc_pos pos = { 0 };
	unsigned flags =
		BCH_BUCKET_MARK_GC|
		(initial ? BCH_BUCKET_MARK_NOATOMIC : 0);
	int ret = 0;

	if (initial) {
		BUG_ON(journal_seq_verify(c) &&
		       k.k->version.lo > journal_cur_seq(&c->journal));

		if (k.k->version.lo > atomic64_read(&c->key_version))
			atomic64_set(&c->key_version, k.k->version.lo);

		if (test_bit(BCH_FS_REBUILD_REPLICAS, &c->flags) ||
		    fsck_err_on(!bch2_bkey_replicas_marked(c, k, false), c,
				"superblock not marked as containing replicas (type %u)",
				k.k->type)) {
			ret = bch2_mark_bkey_replicas(c, k);
			if (ret)
				return ret;
		}

		bkey_for_each_ptr(ptrs, ptr) {
			struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
			size_t b = PTR_BUCKET_NR(ca, ptr);
			struct bucket *g = PTR_BUCKET(ca, ptr);

			if (mustfix_fsck_err_on(!g->gen_valid, c,
					"found ptr with missing gen in alloc btree,\n"
					"type %u gen %u",
					k.k->type, ptr->gen)) {
				g->_mark.gen = ptr->gen;
				g->gen_valid = 1;
				set_bit(b, ca->buckets_dirty);
			}

			if (mustfix_fsck_err_on(gen_cmp(ptr->gen, g->mark.gen) > 0, c,
					"%u ptr gen in the future: %u > %u",
					k.k->type, ptr->gen, g->mark.gen)) {
				g->_mark.gen = ptr->gen;
				g->gen_valid = 1;
				set_bit(b, ca->buckets_dirty);
				set_bit(BCH_FS_FIXED_GENS, &c->flags);
			}
		}
	}

	bkey_for_each_ptr(ptrs, ptr) {
		struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
		size_t b = PTR_BUCKET_NR(ca, ptr);

		if (gen_after(ca->oldest_gens[b], ptr->gen))
			ca->oldest_gens[b] = ptr->gen;

		*max_stale = max(*max_stale, ptr_stale(ca, ptr));
	}

	bch2_mark_key(c, k, true, k.k->size, pos, NULL, 0, flags);
fsck_err:
	return ret;
}

static int btree_gc_mark_node(struct bch_fs *c, struct btree *b,
			      u8 *max_stale, bool initial)
{
	struct btree_node_iter iter;
	struct bkey unpacked;
	struct bkey_s_c k;
	int ret = 0;

	*max_stale = 0;

	if (!btree_node_type_needs_gc(btree_node_type(b)))
		return 0;

	for_each_btree_node_key_unpack(b, k, &iter,
				       &unpacked) {
		bch2_bkey_debugcheck(c, b, k);

		ret = bch2_gc_mark_key(c, k, max_stale, initial);
		if (ret)
			break;
	}

	return ret;
}

static int bch2_gc_btree(struct bch_fs *c, enum btree_id btree_id,
			 bool initial)
{
	struct btree_iter iter;
	struct btree *b;
	struct range_checks r;
	unsigned depth = btree_node_type_needs_gc(btree_id) ? 0 : 1;
	u8 max_stale;
	int ret = 0;

	gc_pos_set(c, gc_pos_btree(btree_id, POS_MIN, 0));

	/*
	 * if expensive_debug_checks is on, run range_checks on all leaf nodes:
	 *
	 * and on startup, we have to read every btree node (XXX: only if it was
	 * an unclean shutdown)
	 */
	if (initial || expensive_debug_checks(c))
		depth = 0;

	btree_node_range_checks_init(&r, depth);

	__for_each_btree_node(&iter, c, btree_id, POS_MIN,
			      0, depth, BTREE_ITER_PREFETCH, b) {
		btree_node_range_checks(c, b, &r);

		bch2_verify_btree_nr_keys(b);

		ret = btree_gc_mark_node(c, b, &max_stale, initial);
		if (ret)
			break;

		gc_pos_set(c, gc_pos_btree_node(b));

		if (!initial) {
			if (max_stale > 64)
				bch2_btree_node_rewrite(c, &iter,
						b->data->keys.seq,
						BTREE_INSERT_USE_RESERVE|
						BTREE_INSERT_NOWAIT|
						BTREE_INSERT_GC_LOCK_HELD);
			else if (!btree_gc_rewrite_disabled(c) &&
				 (btree_gc_always_rewrite(c) || max_stale > 16))
				bch2_btree_node_rewrite(c, &iter,
						b->data->keys.seq,
						BTREE_INSERT_NOWAIT|
						BTREE_INSERT_GC_LOCK_HELD);
		}

		bch2_btree_iter_cond_resched(&iter);
	}
	ret = bch2_btree_iter_unlock(&iter) ?: ret;
	if (ret)
		return ret;

	mutex_lock(&c->btree_root_lock);

	b = c->btree_roots[btree_id].b;
	if (!btree_node_fake(b))
		bch2_gc_mark_key(c, bkey_i_to_s_c(&b->key),
				 &max_stale, initial);
	gc_pos_set(c, gc_pos_btree_root(b->btree_id));

	mutex_unlock(&c->btree_root_lock);
	return 0;
}

static inline int btree_id_gc_phase_cmp(enum btree_id l, enum btree_id r)
{
	return  (int) btree_id_to_gc_phase(l) -
		(int) btree_id_to_gc_phase(r);
}

static int bch2_gc_btrees(struct bch_fs *c, struct list_head *journal,
			  bool initial)
{
	enum btree_id ids[BTREE_ID_NR];
	u8 max_stale;
	unsigned i;

	for (i = 0; i < BTREE_ID_NR; i++)
		ids[i] = i;
	bubble_sort(ids, BTREE_ID_NR, btree_id_gc_phase_cmp);

	for (i = 0; i < BTREE_ID_NR; i++) {
		enum btree_id id = ids[i];
		enum btree_node_type type = __btree_node_type(0, id);

		int ret = bch2_gc_btree(c, id, initial);
		if (ret)
			return ret;

		if (journal && btree_node_type_needs_gc(type)) {
			struct bkey_i *k, *n;
			struct jset_entry *j;
			struct journal_replay *r;
			int ret;

			list_for_each_entry(r, journal, list)
				for_each_jset_key(k, n, j, &r->j) {
					if (type == __btree_node_type(j->level, j->btree_id)) {
						ret = bch2_gc_mark_key(c,
							bkey_i_to_s_c(k),
							&max_stale, initial);
						if (ret)
							return ret;
					}
				}
		}
	}

	return 0;
}

static void mark_metadata_sectors(struct bch_fs *c, struct bch_dev *ca,
				  u64 start, u64 end,
				  enum bch_data_type type,
				  unsigned flags)
{
	u64 b = sector_to_bucket(ca, start);

	do {
		unsigned sectors =
			min_t(u64, bucket_to_sector(ca, b + 1), end) - start;

		bch2_mark_metadata_bucket(c, ca, b, type, sectors,
					  gc_phase(GC_PHASE_SB), flags);
		b++;
		start += sectors;
	} while (start < end);
}

void bch2_mark_dev_superblock(struct bch_fs *c, struct bch_dev *ca,
			      unsigned flags)
{
	struct bch_sb_layout *layout = &ca->disk_sb.sb->layout;
	unsigned i;
	u64 b;

	/*
	 * This conditional is kind of gross, but we may be called from the
	 * device add path, before the new device has actually been added to the
	 * running filesystem:
	 */
	if (c) {
		lockdep_assert_held(&c->sb_lock);
		percpu_down_read(&c->usage_lock);
	} else {
		preempt_disable();
	}

	for (i = 0; i < layout->nr_superblocks; i++) {
		u64 offset = le64_to_cpu(layout->sb_offset[i]);

		if (offset == BCH_SB_SECTOR)
			mark_metadata_sectors(c, ca, 0, BCH_SB_SECTOR,
					      BCH_DATA_SB, flags);

		mark_metadata_sectors(c, ca, offset,
				      offset + (1 << layout->sb_max_size_bits),
				      BCH_DATA_SB, flags);
	}

	for (i = 0; i < ca->journal.nr; i++) {
		b = ca->journal.buckets[i];
		bch2_mark_metadata_bucket(c, ca, b, BCH_DATA_JOURNAL,
					  ca->mi.bucket_size,
					  gc_phase(GC_PHASE_SB), flags);
	}

	if (c) {
		percpu_up_read(&c->usage_lock);
	} else {
		preempt_enable();
	}
}

static void bch2_mark_superblocks(struct bch_fs *c)
{
	struct bch_dev *ca;
	unsigned i;

	mutex_lock(&c->sb_lock);
	gc_pos_set(c, gc_phase(GC_PHASE_SB));

	for_each_online_member(ca, c, i)
		bch2_mark_dev_superblock(c, ca, BCH_BUCKET_MARK_GC);
	mutex_unlock(&c->sb_lock);
}

/* Also see bch2_pending_btree_node_free_insert_done() */
static void bch2_mark_pending_btree_node_frees(struct bch_fs *c)
{
	struct gc_pos pos = { 0 };
	struct btree_update *as;
	struct pending_btree_node_free *d;

	mutex_lock(&c->btree_interior_update_lock);
	gc_pos_set(c, gc_phase(GC_PHASE_PENDING_DELETE));

	for_each_pending_btree_node_free(c, as, d)
		if (d->index_update_done)
			bch2_mark_key(c, bkey_i_to_s_c(&d->key),
				      true, 0,
				      pos, NULL, 0,
				      BCH_BUCKET_MARK_GC);

	mutex_unlock(&c->btree_interior_update_lock);
}

static void bch2_mark_allocator_buckets(struct bch_fs *c)
{
	struct bch_dev *ca;
	struct open_bucket *ob;
	size_t i, j, iter;
	unsigned ci;

	percpu_down_read(&c->usage_lock);

	spin_lock(&c->freelist_lock);
	gc_pos_set(c, gc_pos_alloc(c, NULL));

	for_each_member_device(ca, c, ci) {
		fifo_for_each_entry(i, &ca->free_inc, iter)
			bch2_mark_alloc_bucket(c, ca, i, true,
					       gc_pos_alloc(c, NULL),
					       BCH_BUCKET_MARK_GC);



		for (j = 0; j < RESERVE_NR; j++)
			fifo_for_each_entry(i, &ca->free[j], iter)
				bch2_mark_alloc_bucket(c, ca, i, true,
						       gc_pos_alloc(c, NULL),
						       BCH_BUCKET_MARK_GC);
	}

	spin_unlock(&c->freelist_lock);

	for (ob = c->open_buckets;
	     ob < c->open_buckets + ARRAY_SIZE(c->open_buckets);
	     ob++) {
		spin_lock(&ob->lock);
		if (ob->valid) {
			gc_pos_set(c, gc_pos_alloc(c, ob));
			ca = bch_dev_bkey_exists(c, ob->ptr.dev);
			bch2_mark_alloc_bucket(c, ca, PTR_BUCKET_NR(ca, &ob->ptr), true,
					       gc_pos_alloc(c, ob),
					       BCH_BUCKET_MARK_GC);
		}
		spin_unlock(&ob->lock);
	}

	percpu_up_read(&c->usage_lock);
}

static void bch2_gc_free(struct bch_fs *c)
{
	struct bch_dev *ca;
	unsigned i;

	genradix_free(&c->stripes[1]);

	for_each_member_device(ca, c, i) {
		kvpfree(rcu_dereference_protected(ca->buckets[1], 1),
			sizeof(struct bucket_array) +
			ca->mi.nbuckets * sizeof(struct bucket));
		ca->buckets[1] = NULL;

		free_percpu(ca->usage[1]);
		ca->usage[1] = NULL;
	}

	free_percpu(c->usage[1]);
	c->usage[1] = NULL;
}

static void bch2_gc_done_nocheck(struct bch_fs *c)
{
	struct bch_dev *ca;
	unsigned i;
	int cpu;

	{
		struct genradix_iter dst_iter = genradix_iter_init(&c->stripes[0], 0);
		struct genradix_iter src_iter = genradix_iter_init(&c->stripes[1], 0);
		struct stripe *dst, *src;

		c->ec_stripes_heap.used = 0;

		while ((dst = genradix_iter_peek(&dst_iter, &c->stripes[0])) &&
		       (src = genradix_iter_peek(&src_iter, &c->stripes[1]))) {
			*dst = *src;

			if (dst->alive)
				bch2_stripes_heap_insert(c, dst, dst_iter.pos);

			genradix_iter_advance(&dst_iter, &c->stripes[0]);
			genradix_iter_advance(&src_iter, &c->stripes[1]);
		}
	}

	for_each_member_device(ca, c, i) {
		struct bucket_array *src = __bucket_array(ca, 1);

		memcpy(__bucket_array(ca, 0), src,
		       sizeof(struct bucket_array) +
		       sizeof(struct bucket) * src->nbuckets);
	};

	for_each_member_device(ca, c, i) {
		struct bch_dev_usage *p;

		for_each_possible_cpu(cpu) {
			p = per_cpu_ptr(ca->usage[0], cpu);
			memset(p, 0, sizeof(*p));
		}

		preempt_disable();
		*this_cpu_ptr(ca->usage[0]) = __bch2_dev_usage_read(ca, 1);
		preempt_enable();
	}

	{
		struct bch_fs_usage src = __bch2_fs_usage_read(c, 1);
		struct bch_fs_usage *p;

		for_each_possible_cpu(cpu) {
			p = per_cpu_ptr(c->usage[0], cpu);
			memset(p, 0, offsetof(typeof(*p), online_reserved));
		}

		preempt_disable();
		memcpy(this_cpu_ptr(c->usage[0]),
		       &src,
		       offsetof(typeof(*p), online_reserved));
		preempt_enable();
	}

}

static void bch2_gc_done(struct bch_fs *c, bool initial)
{
	struct bch_dev *ca;
	unsigned i;
	int cpu;

#define copy_field(_f, _msg, ...)					\
	if (dst._f != src._f) {						\
		bch_err(c, _msg ": got %llu, should be %llu, fixing"\
			, ##__VA_ARGS__, dst._f, src._f);		\
		dst._f = src._f;					\
	}
#define copy_stripe_field(_f, _msg, ...)				\
	if (dst->_f != src->_f) {					\
		bch_err_ratelimited(c, "stripe %zu has wrong "_msg	\
			": got %u, should be %u, fixing",		\
			dst_iter.pos, ##__VA_ARGS__,			\
			dst->_f, src->_f);				\
		dst->_f = src->_f;					\
	}
#define copy_bucket_field(_f)						\
	if (dst->b[b].mark._f != src->b[b].mark._f) {			\
		bch_err_ratelimited(c, "dev %u bucket %zu has wrong " #_f\
			": got %u, should be %u, fixing",		\
			i, b, dst->b[b].mark._f, src->b[b].mark._f);	\
		dst->b[b]._mark._f = src->b[b].mark._f;			\
	}
#define copy_dev_field(_f, _msg, ...)					\
	copy_field(_f, "dev %u has wrong " _msg, i, ##__VA_ARGS__)
#define copy_fs_field(_f, _msg, ...)					\
	copy_field(_f, "fs has wrong " _msg, ##__VA_ARGS__)

	percpu_down_write(&c->usage_lock);

	if (initial) {
		bch2_gc_done_nocheck(c);
		goto out;
	}

	{
		struct genradix_iter dst_iter = genradix_iter_init(&c->stripes[0], 0);
		struct genradix_iter src_iter = genradix_iter_init(&c->stripes[1], 0);
		struct stripe *dst, *src;
		unsigned i;

		c->ec_stripes_heap.used = 0;

		while ((dst = genradix_iter_peek(&dst_iter, &c->stripes[0])) &&
		       (src = genradix_iter_peek(&src_iter, &c->stripes[1]))) {
			copy_stripe_field(alive,	"alive");
			copy_stripe_field(sectors,	"sectors");
			copy_stripe_field(algorithm,	"algorithm");
			copy_stripe_field(nr_blocks,	"nr_blocks");
			copy_stripe_field(nr_redundant,	"nr_redundant");
			copy_stripe_field(blocks_nonempty.counter,
					  "blocks_nonempty");

			for (i = 0; i < ARRAY_SIZE(dst->block_sectors); i++)
				copy_stripe_field(block_sectors[i].counter,
						  "block_sectors[%u]", i);

			if (dst->alive)
				bch2_stripes_heap_insert(c, dst, dst_iter.pos);

			genradix_iter_advance(&dst_iter, &c->stripes[0]);
			genradix_iter_advance(&src_iter, &c->stripes[1]);
		}
	}

	for_each_member_device(ca, c, i) {
		struct bucket_array *dst = __bucket_array(ca, 0);
		struct bucket_array *src = __bucket_array(ca, 1);
		size_t b;

		if (initial) {
			memcpy(dst, src,
			       sizeof(struct bucket_array) +
			       sizeof(struct bucket) * dst->nbuckets);
		}

		for (b = 0; b < src->nbuckets; b++) {
			copy_bucket_field(gen);
			copy_bucket_field(data_type);
			copy_bucket_field(owned_by_allocator);
			copy_bucket_field(stripe);
			copy_bucket_field(dirty_sectors);
			copy_bucket_field(cached_sectors);
		}
	};

	for_each_member_device(ca, c, i) {
		struct bch_dev_usage dst = __bch2_dev_usage_read(ca, 0);
		struct bch_dev_usage src = __bch2_dev_usage_read(ca, 1);
		struct bch_dev_usage *p;
		unsigned b;

		for (b = 0; b < BCH_DATA_NR; b++)
			copy_dev_field(buckets[b],
				       "buckets[%s]", bch2_data_types[b]);
		copy_dev_field(buckets_alloc, "buckets_alloc");
		copy_dev_field(buckets_ec, "buckets_ec");

		for (b = 0; b < BCH_DATA_NR; b++)
			copy_dev_field(sectors[b],
				       "sectors[%s]", bch2_data_types[b]);
		copy_dev_field(sectors_fragmented,
			       "sectors_fragmented");

		for_each_possible_cpu(cpu) {
			p = per_cpu_ptr(ca->usage[0], cpu);
			memset(p, 0, sizeof(*p));
		}

		preempt_disable();
		p = this_cpu_ptr(ca->usage[0]);
		*p = dst;
		preempt_enable();
	}

	{
		struct bch_fs_usage dst = __bch2_fs_usage_read(c, 0);
		struct bch_fs_usage src = __bch2_fs_usage_read(c, 1);
		struct bch_fs_usage *p;
		unsigned r, b;

		for (r = 0; r < BCH_REPLICAS_MAX; r++) {
			for (b = 0; b < BCH_DATA_NR; b++)
				copy_fs_field(replicas[r].data[b],
					      "replicas[%i].data[%s]",
					      r, bch2_data_types[b]);
			copy_fs_field(replicas[r].ec_data,
				      "replicas[%i].ec_data", r);
			copy_fs_field(replicas[r].persistent_reserved,
				      "replicas[%i].persistent_reserved", r);
		}

		for (b = 0; b < BCH_DATA_NR; b++)
			copy_fs_field(buckets[b],
				      "buckets[%s]", bch2_data_types[b]);
		copy_fs_field(nr_inodes, "nr_inodes");

		for_each_possible_cpu(cpu) {
			p = per_cpu_ptr(c->usage[0], cpu);
			memset(p, 0, offsetof(typeof(*p), online_reserved));
		}

		preempt_disable();
		p = this_cpu_ptr(c->usage[0]);
		memcpy(p, &dst, offsetof(typeof(*p), online_reserved));
		preempt_enable();
	}
out:
	percpu_up_write(&c->usage_lock);

#undef copy_fs_field
#undef copy_dev_field
#undef copy_bucket_field
#undef copy_stripe_field
#undef copy_field
}

static int bch2_gc_start(struct bch_fs *c)
{
	struct bch_dev *ca;
	unsigned i;

	/*
	 * indicate to stripe code that we need to allocate for the gc stripes
	 * radix tree, too
	 */
	gc_pos_set(c, gc_phase(GC_PHASE_START));

	BUG_ON(c->usage[1]);

	c->usage[1] = alloc_percpu(struct bch_fs_usage);
	if (!c->usage[1])
		return -ENOMEM;

	for_each_member_device(ca, c, i) {
		BUG_ON(ca->buckets[1]);
		BUG_ON(ca->usage[1]);

		ca->buckets[1] = kvpmalloc(sizeof(struct bucket_array) +
				ca->mi.nbuckets * sizeof(struct bucket),
				GFP_KERNEL|__GFP_ZERO);
		if (!ca->buckets[1]) {
			percpu_ref_put(&ca->ref);
			return -ENOMEM;
		}

		ca->usage[1] = alloc_percpu(struct bch_dev_usage);
		if (!ca->usage[1]) {
			percpu_ref_put(&ca->ref);
			return -ENOMEM;
		}
	}

	percpu_down_write(&c->usage_lock);

	for_each_member_device(ca, c, i) {
		struct bucket_array *dst = __bucket_array(ca, 1);
		struct bucket_array *src = __bucket_array(ca, 0);
		size_t b;

		dst->first_bucket	= src->first_bucket;
		dst->nbuckets		= src->nbuckets;

		for (b = 0; b < src->nbuckets; b++)
			dst->b[b]._mark.gen = src->b[b].mark.gen;
	};

	percpu_up_write(&c->usage_lock);

	return bch2_ec_mem_alloc(c, true);
}

/**
 * bch2_gc - walk _all_ references to buckets, and recompute them:
 *
 * Order matters here:
 *  - Concurrent GC relies on the fact that we have a total ordering for
 *    everything that GC walks - see  gc_will_visit_node(),
 *    gc_will_visit_root()
 *
 *  - also, references move around in the course of index updates and
 *    various other crap: everything needs to agree on the ordering
 *    references are allowed to move around in - e.g., we're allowed to
 *    start with a reference owned by an open_bucket (the allocator) and
 *    move it to the btree, but not the reverse.
 *
 *    This is necessary to ensure that gc doesn't miss references that
 *    move around - if references move backwards in the ordering GC
 *    uses, GC could skip past them
 */
int bch2_gc(struct bch_fs *c, struct list_head *journal, bool initial)
{
	struct bch_dev *ca;
	u64 start_time = local_clock();
	unsigned i, iter = 0;
	int ret;

	trace_gc_start(c);

	down_write(&c->gc_lock);
again:
	ret = bch2_gc_start(c);
	if (ret)
		goto out;

	bch2_mark_superblocks(c);

	ret = bch2_gc_btrees(c, journal, initial);
	if (ret)
		goto out;

	bch2_mark_pending_btree_node_frees(c);
	bch2_mark_allocator_buckets(c);

	c->gc_count++;
out:
	if (!ret && test_bit(BCH_FS_FIXED_GENS, &c->flags)) {
		/*
		 * XXX: make sure gens we fixed got saved
		 */
		if (iter++ <= 2) {
			bch_info(c, "Fixed gens, restarting mark and sweep:");
			clear_bit(BCH_FS_FIXED_GENS, &c->flags);
			goto again;
		}

		bch_info(c, "Unable to fix bucket gens, looping");
		ret = -EINVAL;
	}

	if (!ret)
		bch2_gc_done(c, initial);

	/* Indicates that gc is no longer in progress: */
	__gc_pos_set(c, gc_phase(GC_PHASE_NOT_RUNNING));

	bch2_gc_free(c);
	up_write(&c->gc_lock);

	if (!ret && initial)
		set_bit(BCH_FS_INITIAL_GC_DONE, &c->flags);

	trace_gc_end(c);
	bch2_time_stats_update(&c->times[BCH_TIME_btree_gc], start_time);

	/*
	 * Wake up allocator in case it was waiting for buckets
	 * because of not being able to inc gens
	 */
	for_each_member_device(ca, c, i)
		bch2_wake_allocator(ca);

	/*
	 * At startup, allocations can happen directly instead of via the
	 * allocator thread - issue wakeup in case they blocked on gc_lock:
	 */
	closure_wake_up(&c->freelist_wait);
	return ret;
}

/* Btree coalescing */

static void recalc_packed_keys(struct btree *b)
{
	struct bset *i = btree_bset_first(b);
	struct bkey_packed *k;

	memset(&b->nr, 0, sizeof(b->nr));

	BUG_ON(b->nsets != 1);

	vstruct_for_each(i, k)
		btree_keys_account_key_add(&b->nr, 0, k);
}

static void bch2_coalesce_nodes(struct bch_fs *c, struct btree_iter *iter,
				struct btree *old_nodes[GC_MERGE_NODES])
{
	struct btree *parent = btree_node_parent(iter, old_nodes[0]);
	unsigned i, nr_old_nodes, nr_new_nodes, u64s = 0;
	unsigned blocks = btree_blocks(c) * 2 / 3;
	struct btree *new_nodes[GC_MERGE_NODES];
	struct btree_update *as;
	struct keylist keylist;
	struct bkey_format_state format_state;
	struct bkey_format new_format;

	memset(new_nodes, 0, sizeof(new_nodes));
	bch2_keylist_init(&keylist, NULL);

	/* Count keys that are not deleted */
	for (i = 0; i < GC_MERGE_NODES && old_nodes[i]; i++)
		u64s += old_nodes[i]->nr.live_u64s;

	nr_old_nodes = nr_new_nodes = i;

	/* Check if all keys in @old_nodes could fit in one fewer node */
	if (nr_old_nodes <= 1 ||
	    __vstruct_blocks(struct btree_node, c->block_bits,
			     DIV_ROUND_UP(u64s, nr_old_nodes - 1)) > blocks)
		return;

	/* Find a format that all keys in @old_nodes can pack into */
	bch2_bkey_format_init(&format_state);

	for (i = 0; i < nr_old_nodes; i++)
		__bch2_btree_calc_format(&format_state, old_nodes[i]);

	new_format = bch2_bkey_format_done(&format_state);

	/* Check if repacking would make any nodes too big to fit */
	for (i = 0; i < nr_old_nodes; i++)
		if (!bch2_btree_node_format_fits(c, old_nodes[i], &new_format)) {
			trace_btree_gc_coalesce_fail(c,
					BTREE_GC_COALESCE_FAIL_FORMAT_FITS);
			return;
		}

	if (bch2_keylist_realloc(&keylist, NULL, 0,
			(BKEY_U64s + BKEY_EXTENT_U64s_MAX) * nr_old_nodes)) {
		trace_btree_gc_coalesce_fail(c,
				BTREE_GC_COALESCE_FAIL_KEYLIST_REALLOC);
		return;
	}

	as = bch2_btree_update_start(c, iter->btree_id,
			btree_update_reserve_required(c, parent) + nr_old_nodes,
			BTREE_INSERT_NOFAIL|
			BTREE_INSERT_USE_RESERVE,
			NULL);
	if (IS_ERR(as)) {
		trace_btree_gc_coalesce_fail(c,
				BTREE_GC_COALESCE_FAIL_RESERVE_GET);
		bch2_keylist_free(&keylist, NULL);
		return;
	}

	trace_btree_gc_coalesce(c, old_nodes[0]);

	for (i = 0; i < nr_old_nodes; i++)
		bch2_btree_interior_update_will_free_node(as, old_nodes[i]);

	/* Repack everything with @new_format and sort down to one bset */
	for (i = 0; i < nr_old_nodes; i++)
		new_nodes[i] =
			__bch2_btree_node_alloc_replacement(as, old_nodes[i],
							    new_format);

	/*
	 * Conceptually we concatenate the nodes together and slice them
	 * up at different boundaries.
	 */
	for (i = nr_new_nodes - 1; i > 0; --i) {
		struct btree *n1 = new_nodes[i];
		struct btree *n2 = new_nodes[i - 1];

		struct bset *s1 = btree_bset_first(n1);
		struct bset *s2 = btree_bset_first(n2);
		struct bkey_packed *k, *last = NULL;

		/* Calculate how many keys from @n2 we could fit inside @n1 */
		u64s = 0;

		for (k = s2->start;
		     k < vstruct_last(s2) &&
		     vstruct_blocks_plus(n1->data, c->block_bits,
					 u64s + k->u64s) <= blocks;
		     k = bkey_next(k)) {
			last = k;
			u64s += k->u64s;
		}

		if (u64s == le16_to_cpu(s2->u64s)) {
			/* n2 fits entirely in n1 */
			n1->key.k.p = n1->data->max_key = n2->data->max_key;

			memcpy_u64s(vstruct_last(s1),
				    s2->start,
				    le16_to_cpu(s2->u64s));
			le16_add_cpu(&s1->u64s, le16_to_cpu(s2->u64s));

			set_btree_bset_end(n1, n1->set);

			six_unlock_write(&n2->lock);
			bch2_btree_node_free_never_inserted(c, n2);
			six_unlock_intent(&n2->lock);

			memmove(new_nodes + i - 1,
				new_nodes + i,
				sizeof(new_nodes[0]) * (nr_new_nodes - i));
			new_nodes[--nr_new_nodes] = NULL;
		} else if (u64s) {
			/* move part of n2 into n1 */
			n1->key.k.p = n1->data->max_key =
				bkey_unpack_pos(n1, last);

			n2->data->min_key =
				btree_type_successor(iter->btree_id,
						     n1->data->max_key);

			memcpy_u64s(vstruct_last(s1),
				    s2->start, u64s);
			le16_add_cpu(&s1->u64s, u64s);

			memmove(s2->start,
				vstruct_idx(s2, u64s),
				(le16_to_cpu(s2->u64s) - u64s) * sizeof(u64));
			s2->u64s = cpu_to_le16(le16_to_cpu(s2->u64s) - u64s);

			set_btree_bset_end(n1, n1->set);
			set_btree_bset_end(n2, n2->set);
		}
	}

	for (i = 0; i < nr_new_nodes; i++) {
		struct btree *n = new_nodes[i];

		recalc_packed_keys(n);
		btree_node_reset_sib_u64s(n);

		bch2_btree_build_aux_trees(n);
		six_unlock_write(&n->lock);

		bch2_btree_node_write(c, n, SIX_LOCK_intent);
	}

	/*
	 * The keys for the old nodes get deleted. We don't want to insert keys
	 * that compare equal to the keys for the new nodes we'll also be
	 * inserting - we can't because keys on a keylist must be strictly
	 * greater than the previous keys, and we also don't need to since the
	 * key for the new node will serve the same purpose (overwriting the key
	 * for the old node).
	 */
	for (i = 0; i < nr_old_nodes; i++) {
		struct bkey_i delete;
		unsigned j;

		for (j = 0; j < nr_new_nodes; j++)
			if (!bkey_cmp(old_nodes[i]->key.k.p,
				      new_nodes[j]->key.k.p))
				goto next;

		bkey_init(&delete.k);
		delete.k.p = old_nodes[i]->key.k.p;
		bch2_keylist_add_in_order(&keylist, &delete);
next:
		i = i;
	}

	/*
	 * Keys for the new nodes get inserted: bch2_btree_insert_keys() only
	 * does the lookup once and thus expects the keys to be in sorted order
	 * so we have to make sure the new keys are correctly ordered with
	 * respect to the deleted keys added in the previous loop
	 */
	for (i = 0; i < nr_new_nodes; i++)
		bch2_keylist_add_in_order(&keylist, &new_nodes[i]->key);

	/* Insert the newly coalesced nodes */
	bch2_btree_insert_node(as, parent, iter, &keylist, 0);

	BUG_ON(!bch2_keylist_empty(&keylist));

	BUG_ON(iter->l[old_nodes[0]->level].b != old_nodes[0]);

	bch2_btree_iter_node_replace(iter, new_nodes[0]);

	for (i = 0; i < nr_new_nodes; i++)
		bch2_open_buckets_put(c, &new_nodes[i]->ob);

	/* Free the old nodes and update our sliding window */
	for (i = 0; i < nr_old_nodes; i++) {
		bch2_btree_node_free_inmem(c, old_nodes[i], iter);

		/*
		 * the index update might have triggered a split, in which case
		 * the nodes we coalesced - the new nodes we just created -
		 * might not be sibling nodes anymore - don't add them to the
		 * sliding window (except the first):
		 */
		if (!i) {
			old_nodes[i] = new_nodes[i];
		} else {
			old_nodes[i] = NULL;
			if (new_nodes[i])
				six_unlock_intent(&new_nodes[i]->lock);
		}
	}

	bch2_btree_update_done(as);
	bch2_keylist_free(&keylist, NULL);
}

static int bch2_coalesce_btree(struct bch_fs *c, enum btree_id btree_id)
{
	struct btree_iter iter;
	struct btree *b;
	bool kthread = (current->flags & PF_KTHREAD) != 0;
	unsigned i;

	/* Sliding window of adjacent btree nodes */
	struct btree *merge[GC_MERGE_NODES];
	u32 lock_seq[GC_MERGE_NODES];

	/*
	 * XXX: We don't have a good way of positively matching on sibling nodes
	 * that have the same parent - this code works by handling the cases
	 * where they might not have the same parent, and is thus fragile. Ugh.
	 *
	 * Perhaps redo this to use multiple linked iterators?
	 */
	memset(merge, 0, sizeof(merge));

	__for_each_btree_node(&iter, c, btree_id, POS_MIN,
			      BTREE_MAX_DEPTH, 0,
			      BTREE_ITER_PREFETCH, b) {
		memmove(merge + 1, merge,
			sizeof(merge) - sizeof(merge[0]));
		memmove(lock_seq + 1, lock_seq,
			sizeof(lock_seq) - sizeof(lock_seq[0]));

		merge[0] = b;

		for (i = 1; i < GC_MERGE_NODES; i++) {
			if (!merge[i] ||
			    !six_relock_intent(&merge[i]->lock, lock_seq[i]))
				break;

			if (merge[i]->level != merge[0]->level) {
				six_unlock_intent(&merge[i]->lock);
				break;
			}
		}
		memset(merge + i, 0, (GC_MERGE_NODES - i) * sizeof(merge[0]));

		bch2_coalesce_nodes(c, &iter, merge);

		for (i = 1; i < GC_MERGE_NODES && merge[i]; i++) {
			lock_seq[i] = merge[i]->lock.state.seq;
			six_unlock_intent(&merge[i]->lock);
		}

		lock_seq[0] = merge[0]->lock.state.seq;

		if (kthread && kthread_should_stop()) {
			bch2_btree_iter_unlock(&iter);
			return -ESHUTDOWN;
		}

		bch2_btree_iter_cond_resched(&iter);

		/*
		 * If the parent node wasn't relocked, it might have been split
		 * and the nodes in our sliding window might not have the same
		 * parent anymore - blow away the sliding window:
		 */
		if (btree_iter_node(&iter, iter.level + 1) &&
		    !btree_node_intent_locked(&iter, iter.level + 1))
			memset(merge + 1, 0,
			       (GC_MERGE_NODES - 1) * sizeof(merge[0]));
	}
	return bch2_btree_iter_unlock(&iter);
}

/**
 * bch_coalesce - coalesce adjacent nodes with low occupancy
 */
void bch2_coalesce(struct bch_fs *c)
{
	enum btree_id id;

	down_read(&c->gc_lock);
	trace_gc_coalesce_start(c);

	for (id = 0; id < BTREE_ID_NR; id++) {
		int ret = c->btree_roots[id].b
			? bch2_coalesce_btree(c, id)
			: 0;

		if (ret) {
			if (ret != -ESHUTDOWN)
				bch_err(c, "btree coalescing failed: %d", ret);
			return;
		}
	}

	trace_gc_coalesce_end(c);
	up_read(&c->gc_lock);
}

static int bch2_gc_thread(void *arg)
{
	struct bch_fs *c = arg;
	struct io_clock *clock = &c->io_clock[WRITE];
	unsigned long last = atomic_long_read(&clock->now);
	unsigned last_kick = atomic_read(&c->kick_gc);
	int ret;

	set_freezable();

	while (1) {
		while (1) {
			set_current_state(TASK_INTERRUPTIBLE);

			if (kthread_should_stop()) {
				__set_current_state(TASK_RUNNING);
				return 0;
			}

			if (atomic_read(&c->kick_gc) != last_kick)
				break;

			if (c->btree_gc_periodic) {
				unsigned long next = last + c->capacity / 16;

				if (atomic_long_read(&clock->now) >= next)
					break;

				bch2_io_clock_schedule_timeout(clock, next);
			} else {
				schedule();
			}

			try_to_freeze();
		}
		__set_current_state(TASK_RUNNING);

		last = atomic_long_read(&clock->now);
		last_kick = atomic_read(&c->kick_gc);

		ret = bch2_gc(c, NULL, false);
		if (ret)
			bch_err(c, "btree gc failed: %i", ret);

		debug_check_no_locks_held();
	}

	return 0;
}

void bch2_gc_thread_stop(struct bch_fs *c)
{
	struct task_struct *p;

	p = c->gc_thread;
	c->gc_thread = NULL;

	if (p) {
		kthread_stop(p);
		put_task_struct(p);
	}
}

int bch2_gc_thread_start(struct bch_fs *c)
{
	struct task_struct *p;

	BUG_ON(c->gc_thread);

	p = kthread_create(bch2_gc_thread, c, "bch_gc");
	if (IS_ERR(p))
		return PTR_ERR(p);

	get_task_struct(p);
	c->gc_thread = p;
	wake_up_process(p);
	return 0;
}

/* Initial GC computes bucket marks during startup */

int bch2_initial_gc(struct bch_fs *c, struct list_head *journal)
{
	int ret = bch2_gc(c, journal, true);

	/*
	 * Skip past versions that might have possibly been used (as nonces),
	 * but hadn't had their pointers written:
	 */
	if (c->sb.encryption_type)
		atomic64_add(1 << 16, &c->key_version);

	return ret;
}