summaryrefslogtreecommitdiff
path: root/libbcachefs/movinggc.c
blob: 2aa58b55779415763f8c8208ab83e4d873cd9a97 (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
/*
 * Moving/copying garbage collector
 *
 * Copyright 2012 Google, Inc.
 */

#include "bcachefs.h"
#include "btree_iter.h"
#include "btree_update.h"
#include "buckets.h"
#include "clock.h"
#include "extents.h"
#include "eytzinger.h"
#include "io.h"
#include "keylist.h"
#include "move.h"
#include "movinggc.h"
#include "super-io.h"

#include <trace/events/bcachefs.h>
#include <linux/freezer.h>
#include <linux/kthread.h>
#include <linux/math64.h>
#include <linux/sched/task.h>
#include <linux/sort.h>
#include <linux/wait.h>

/*
 * We can't use the entire copygc reserve in one iteration of copygc: we may
 * need the buckets we're freeing up to go back into the copygc reserve to make
 * forward progress, but if the copygc reserve is full they'll be available for
 * any allocation - and it's possible that in a given iteration, we free up most
 * of the buckets we're going to free before we allocate most of the buckets
 * we're going to allocate.
 *
 * If we only use half of the reserve per iteration, then in steady state we'll
 * always have room in the reserve for the buckets we're going to need in the
 * next iteration:
 */
#define COPYGC_BUCKETS_PER_ITER(ca)					\
	((ca)->free[RESERVE_MOVINGGC].size / 2)

/*
 * Max sectors to move per iteration: Have to take into account internal
 * fragmentation from the multiple write points for each generation:
 */
#define COPYGC_SECTORS_PER_ITER(ca)					\
	((ca)->mi.bucket_size *	COPYGC_BUCKETS_PER_ITER(ca))

static inline int sectors_used_cmp(copygc_heap *heap,
				   struct copygc_heap_entry l,
				   struct copygc_heap_entry r)
{
	return bucket_sectors_used(l.mark) - bucket_sectors_used(r.mark);
}

static int bucket_offset_cmp(const void *_l, const void *_r, size_t size)
{
	const struct copygc_heap_entry *l = _l;
	const struct copygc_heap_entry *r = _r;

	return (l->offset > r->offset) - (l->offset < r->offset);
}

static bool __copygc_pred(struct bch_dev *ca,
			  struct bkey_s_c_extent e)
{
	copygc_heap *h = &ca->copygc_heap;
	const struct bch_extent_ptr *ptr =
		bch2_extent_has_device(e, ca->dev_idx);

	if (ptr) {
		struct copygc_heap_entry search = { .offset = ptr->offset };

		size_t i = eytzinger0_find_le(h->data, h->used,
					      sizeof(h->data[0]),
					      bucket_offset_cmp, &search);

		return (i >= 0 &&
			ptr->offset < h->data[i].offset + ca->mi.bucket_size &&
			ptr->gen == h->data[i].mark.gen);
	}

	return false;
}

static enum data_cmd copygc_pred(struct bch_fs *c, void *arg,
				 enum bkey_type type,
				 struct bkey_s_c_extent e,
				 struct bch_io_opts *io_opts,
				 struct data_opts *data_opts)
{
	struct bch_dev *ca = arg;

	if (!__copygc_pred(ca, e))
		return DATA_SKIP;

	data_opts->target		= dev_to_target(ca->dev_idx);
	data_opts->btree_insert_flags	= BTREE_INSERT_USE_RESERVE;
	data_opts->rewrite_dev		= ca->dev_idx;
	return DATA_REWRITE;
}

static bool have_copygc_reserve(struct bch_dev *ca)
{
	bool ret;

	spin_lock(&ca->freelist_lock);
	ret = fifo_full(&ca->free[RESERVE_MOVINGGC]) ||
		ca->allocator_blocked;
	spin_unlock(&ca->freelist_lock);

	return ret;
}

static void bch2_copygc(struct bch_fs *c, struct bch_dev *ca)
{
	copygc_heap *h = &ca->copygc_heap;
	struct copygc_heap_entry e, *i;
	struct bucket_array *buckets;
	struct bch_move_stats move_stats;
	u64 sectors_to_move = 0, sectors_not_moved = 0;
	u64 buckets_to_move, buckets_not_moved = 0;
	size_t b;
	int ret;

	memset(&move_stats, 0, sizeof(move_stats));
	closure_wait_event(&c->freelist_wait, have_copygc_reserve(ca));

	/*
	 * Find buckets with lowest sector counts, skipping completely
	 * empty buckets, by building a maxheap sorted by sector count,
	 * and repeatedly replacing the maximum element until all
	 * buckets have been visited.
	 */
	h->used = 0;

	/*
	 * We need bucket marks to be up to date - gc can't be recalculating
	 * them:
	 */
	down_read(&c->gc_lock);
	down_read(&ca->bucket_lock);
	buckets = bucket_array(ca);

	for (b = buckets->first_bucket; b < buckets->nbuckets; b++) {
		struct bucket_mark m = READ_ONCE(buckets->b[b].mark);
		struct copygc_heap_entry e;

		if (m.owned_by_allocator ||
		    m.data_type != BCH_DATA_USER ||
		    !bucket_sectors_used(m) ||
		    bucket_sectors_used(m) >= ca->mi.bucket_size)
			continue;

		e = (struct copygc_heap_entry) {
			.offset = bucket_to_sector(ca, b),
			.mark	= m
		};
		heap_add_or_replace(h, e, -sectors_used_cmp);
	}
	up_read(&ca->bucket_lock);
	up_read(&c->gc_lock);

	for (i = h->data; i < h->data + h->used; i++)
		sectors_to_move += bucket_sectors_used(i->mark);

	while (sectors_to_move > COPYGC_SECTORS_PER_ITER(ca)) {
		BUG_ON(!heap_pop(h, e, -sectors_used_cmp));
		sectors_to_move -= bucket_sectors_used(e.mark);
	}

	buckets_to_move = h->used;

	if (!buckets_to_move)
		return;

	eytzinger0_sort(h->data, h->used,
			sizeof(h->data[0]),
			bucket_offset_cmp, NULL);

	ret = bch2_move_data(c, &ca->copygc_pd.rate,
			     writepoint_ptr(&ca->copygc_write_point),
			     POS_MIN, POS_MAX,
			     copygc_pred, ca,
			     &move_stats);

	down_read(&ca->bucket_lock);
	buckets = bucket_array(ca);
	for (i = h->data; i < h->data + h->used; i++) {
		size_t b = sector_to_bucket(ca, i->offset);
		struct bucket_mark m = READ_ONCE(buckets->b[b].mark);

		if (i->mark.gen == m.gen && bucket_sectors_used(m)) {
			sectors_not_moved += bucket_sectors_used(m);
			buckets_not_moved++;
		}
	}
	up_read(&ca->bucket_lock);

	if (sectors_not_moved && !ret)
		bch_warn(c, "copygc finished but %llu/%llu sectors, %llu/%llu buckets not moved",
			 sectors_not_moved, sectors_to_move,
			 buckets_not_moved, buckets_to_move);

	trace_copygc(ca,
		     atomic64_read(&move_stats.sectors_moved), sectors_not_moved,
		     buckets_to_move, buckets_not_moved);
}

static int bch2_copygc_thread(void *arg)
{
	struct bch_dev *ca = arg;
	struct bch_fs *c = ca->fs;
	struct io_clock *clock = &c->io_clock[WRITE];
	unsigned long last;
	u64 available, want, next;

	set_freezable();

	while (!kthread_should_stop()) {
		if (kthread_wait_freezable(c->copy_gc_enabled))
			break;

		last = atomic_long_read(&clock->now);
		/*
		 * don't start copygc until less than half the gc reserve is
		 * available:
		 */
		available = dev_buckets_available(c, ca);
		want = div64_u64((ca->mi.nbuckets - ca->mi.first_bucket) *
				 c->opts.gc_reserve_percent, 200);
		if (available > want) {
			next = last + (available - want) *
				ca->mi.bucket_size;
			bch2_kthread_io_clock_wait(clock, next);
			continue;
		}

		bch2_copygc(c, ca);
	}

	return 0;
}

void bch2_copygc_stop(struct bch_dev *ca)
{
	ca->copygc_pd.rate.rate = UINT_MAX;
	bch2_ratelimit_reset(&ca->copygc_pd.rate);

	if (ca->copygc_thread) {
		kthread_stop(ca->copygc_thread);
		put_task_struct(ca->copygc_thread);
	}
	ca->copygc_thread = NULL;
}

int bch2_copygc_start(struct bch_fs *c, struct bch_dev *ca)
{
	struct task_struct *t;

	BUG_ON(ca->copygc_thread);

	if (c->opts.nochanges)
		return 0;

	if (bch2_fs_init_fault("copygc_start"))
		return -ENOMEM;

	t = kthread_create(bch2_copygc_thread, ca, "bch_copygc");
	if (IS_ERR(t))
		return PTR_ERR(t);

	get_task_struct(t);

	ca->copygc_thread = t;
	wake_up_process(ca->copygc_thread);

	return 0;
}

void bch2_dev_copygc_init(struct bch_dev *ca)
{
	bch2_pd_controller_init(&ca->copygc_pd);
	ca->copygc_pd.d_term = 0;
}