1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Moving/copying garbage collector
4  *
5  * Copyright 2012 Google, Inc.
6  */
7 
8 #include "bcachefs.h"
9 #include "alloc_background.h"
10 #include "alloc_foreground.h"
11 #include "btree_iter.h"
12 #include "btree_update.h"
13 #include "btree_write_buffer.h"
14 #include "buckets.h"
15 #include "clock.h"
16 #include "errcode.h"
17 #include "error.h"
18 #include "lru.h"
19 #include "move.h"
20 #include "movinggc.h"
21 #include "trace.h"
22 
23 #include <linux/freezer.h>
24 #include <linux/kthread.h>
25 #include <linux/math64.h>
26 #include <linux/sched/task.h>
27 #include <linux/wait.h>
28 
29 struct buckets_in_flight {
30 	struct rhashtable		table;
31 	struct move_bucket_in_flight	*first;
32 	struct move_bucket_in_flight	*last;
33 	size_t				nr;
34 	size_t				sectors;
35 };
36 
37 static const struct rhashtable_params bch_move_bucket_params = {
38 	.head_offset		= offsetof(struct move_bucket_in_flight, hash),
39 	.key_offset		= offsetof(struct move_bucket_in_flight, bucket.k),
40 	.key_len		= sizeof(struct move_bucket_key),
41 	.automatic_shrinking	= true,
42 };
43 
44 static struct move_bucket_in_flight *
move_bucket_in_flight_add(struct buckets_in_flight * list,struct move_bucket b)45 move_bucket_in_flight_add(struct buckets_in_flight *list, struct move_bucket b)
46 {
47 	struct move_bucket_in_flight *new = kzalloc(sizeof(*new), GFP_KERNEL);
48 	int ret;
49 
50 	if (!new)
51 		return ERR_PTR(-ENOMEM);
52 
53 	new->bucket = b;
54 
55 	ret = rhashtable_lookup_insert_fast(&list->table, &new->hash,
56 					    bch_move_bucket_params);
57 	if (ret) {
58 		kfree(new);
59 		return ERR_PTR(ret);
60 	}
61 
62 	if (!list->first)
63 		list->first = new;
64 	else
65 		list->last->next = new;
66 
67 	list->last = new;
68 	list->nr++;
69 	list->sectors += b.sectors;
70 	return new;
71 }
72 
bch2_bucket_is_movable(struct btree_trans * trans,struct move_bucket * b,u64 time)73 static int bch2_bucket_is_movable(struct btree_trans *trans,
74 				  struct move_bucket *b, u64 time)
75 {
76 	struct bch_fs *c = trans->c;
77 
78 	if (bch2_bucket_is_open(c, b->k.bucket.inode, b->k.bucket.offset))
79 		return 0;
80 
81 	struct btree_iter iter;
82 	struct bkey_s_c k = bch2_bkey_get_iter(trans, &iter, BTREE_ID_alloc,
83 				       b->k.bucket, BTREE_ITER_cached);
84 	int ret = bkey_err(k);
85 	if (ret)
86 		return ret;
87 
88 	struct bch_dev *ca = bch2_dev_tryget(c, k.k->p.inode);
89 	if (!ca)
90 		goto out;
91 
92 	if (ca->mi.state != BCH_MEMBER_STATE_rw ||
93 	    !bch2_dev_is_online(ca))
94 		goto out_put;
95 
96 	struct bch_alloc_v4 _a;
97 	const struct bch_alloc_v4 *a = bch2_alloc_to_v4(k, &_a);
98 	b->k.gen	= a->gen;
99 	b->sectors	= bch2_bucket_sectors_dirty(*a);
100 	u64 lru_idx	= alloc_lru_idx_fragmentation(*a, ca);
101 
102 	ret = lru_idx && lru_idx <= time;
103 out_put:
104 	bch2_dev_put(ca);
105 out:
106 	bch2_trans_iter_exit(trans, &iter);
107 	return ret;
108 }
109 
move_buckets_wait(struct moving_context * ctxt,struct buckets_in_flight * list,bool flush)110 static void move_buckets_wait(struct moving_context *ctxt,
111 			      struct buckets_in_flight *list,
112 			      bool flush)
113 {
114 	struct move_bucket_in_flight *i;
115 	int ret;
116 
117 	while ((i = list->first)) {
118 		if (flush)
119 			move_ctxt_wait_event(ctxt, !atomic_read(&i->count));
120 
121 		if (atomic_read(&i->count))
122 			break;
123 
124 		list->first = i->next;
125 		if (!list->first)
126 			list->last = NULL;
127 
128 		list->nr--;
129 		list->sectors -= i->bucket.sectors;
130 
131 		ret = rhashtable_remove_fast(&list->table, &i->hash,
132 					     bch_move_bucket_params);
133 		BUG_ON(ret);
134 		kfree(i);
135 	}
136 
137 	bch2_trans_unlock_long(ctxt->trans);
138 }
139 
bucket_in_flight(struct buckets_in_flight * list,struct move_bucket_key k)140 static bool bucket_in_flight(struct buckets_in_flight *list,
141 			     struct move_bucket_key k)
142 {
143 	return rhashtable_lookup_fast(&list->table, &k, bch_move_bucket_params);
144 }
145 
146 typedef DARRAY(struct move_bucket) move_buckets;
147 
bch2_copygc_get_buckets(struct moving_context * ctxt,struct buckets_in_flight * buckets_in_flight,move_buckets * buckets)148 static int bch2_copygc_get_buckets(struct moving_context *ctxt,
149 			struct buckets_in_flight *buckets_in_flight,
150 			move_buckets *buckets)
151 {
152 	struct btree_trans *trans = ctxt->trans;
153 	struct bch_fs *c = trans->c;
154 	size_t nr_to_get = max_t(size_t, 16U, buckets_in_flight->nr / 4);
155 	size_t saw = 0, in_flight = 0, not_movable = 0, sectors = 0;
156 	int ret;
157 
158 	move_buckets_wait(ctxt, buckets_in_flight, false);
159 
160 	ret = bch2_btree_write_buffer_tryflush(trans);
161 	if (bch2_err_matches(ret, EROFS))
162 		return ret;
163 
164 	if (bch2_fs_fatal_err_on(ret, c, "%s: from bch2_btree_write_buffer_tryflush()", bch2_err_str(ret)))
165 		return ret;
166 
167 	bch2_trans_begin(trans);
168 
169 	ret = for_each_btree_key_max(trans, iter, BTREE_ID_lru,
170 				  lru_pos(BCH_LRU_BUCKET_FRAGMENTATION, 0, 0),
171 				  lru_pos(BCH_LRU_BUCKET_FRAGMENTATION, U64_MAX, LRU_TIME_MAX),
172 				  0, k, ({
173 		struct move_bucket b = { .k.bucket = u64_to_bucket(k.k->p.offset) };
174 		int ret2 = 0;
175 
176 		saw++;
177 
178 		ret2 = bch2_bucket_is_movable(trans, &b, lru_pos_time(k.k->p));
179 		if (ret2 < 0)
180 			goto err;
181 
182 		if (!ret2)
183 			not_movable++;
184 		else if (bucket_in_flight(buckets_in_flight, b.k))
185 			in_flight++;
186 		else {
187 			ret2 = darray_push(buckets, b);
188 			if (ret2)
189 				goto err;
190 			sectors += b.sectors;
191 		}
192 
193 		ret2 = buckets->nr >= nr_to_get;
194 err:
195 		ret2;
196 	}));
197 
198 	pr_debug("have: %zu (%zu) saw %zu in flight %zu not movable %zu got %zu (%zu)/%zu buckets ret %i",
199 		 buckets_in_flight->nr, buckets_in_flight->sectors,
200 		 saw, in_flight, not_movable, buckets->nr, sectors, nr_to_get, ret);
201 
202 	return ret < 0 ? ret : 0;
203 }
204 
205 noinline
bch2_copygc(struct moving_context * ctxt,struct buckets_in_flight * buckets_in_flight,bool * did_work)206 static int bch2_copygc(struct moving_context *ctxt,
207 		       struct buckets_in_flight *buckets_in_flight,
208 		       bool *did_work)
209 {
210 	struct btree_trans *trans = ctxt->trans;
211 	struct bch_fs *c = trans->c;
212 	struct data_update_opts data_opts = {
213 		.btree_insert_flags = BCH_WATERMARK_copygc,
214 	};
215 	move_buckets buckets = { 0 };
216 	struct move_bucket_in_flight *f;
217 	u64 sectors_seen	= atomic64_read(&ctxt->stats->sectors_seen);
218 	u64 sectors_moved	= atomic64_read(&ctxt->stats->sectors_moved);
219 	int ret = 0;
220 
221 	ret = bch2_copygc_get_buckets(ctxt, buckets_in_flight, &buckets);
222 	if (ret)
223 		goto err;
224 
225 	darray_for_each(buckets, i) {
226 		if (kthread_should_stop() || freezing(current))
227 			break;
228 
229 		f = move_bucket_in_flight_add(buckets_in_flight, *i);
230 		ret = PTR_ERR_OR_ZERO(f);
231 		if (ret == -EEXIST) { /* rare race: copygc_get_buckets returned same bucket more than once */
232 			ret = 0;
233 			continue;
234 		}
235 		if (ret == -ENOMEM) { /* flush IO, continue later */
236 			ret = 0;
237 			break;
238 		}
239 
240 		ret = bch2_evacuate_bucket(ctxt, f, f->bucket.k.bucket,
241 					     f->bucket.k.gen, data_opts);
242 		if (ret)
243 			goto err;
244 
245 		*did_work = true;
246 	}
247 err:
248 
249 	/* no entries in LRU btree found, or got to end: */
250 	if (bch2_err_matches(ret, ENOENT))
251 		ret = 0;
252 
253 	if (ret < 0 && !bch2_err_matches(ret, EROFS))
254 		bch_err_msg(c, ret, "from bch2_move_data()");
255 
256 	sectors_seen	= atomic64_read(&ctxt->stats->sectors_seen) - sectors_seen;
257 	sectors_moved	= atomic64_read(&ctxt->stats->sectors_moved) - sectors_moved;
258 	trace_and_count(c, copygc, c, buckets.nr, sectors_seen, sectors_moved);
259 
260 	darray_exit(&buckets);
261 	return ret;
262 }
263 
264 /*
265  * Copygc runs when the amount of fragmented data is above some arbitrary
266  * threshold:
267  *
268  * The threshold at the limit - when the device is full - is the amount of space
269  * we reserved in bch2_recalc_capacity; we can't have more than that amount of
270  * disk space stranded due to fragmentation and store everything we have
271  * promised to store.
272  *
273  * But we don't want to be running copygc unnecessarily when the device still
274  * has plenty of free space - rather, we want copygc to smoothly run every so
275  * often and continually reduce the amount of fragmented space as the device
276  * fills up. So, we increase the threshold by half the current free space.
277  */
bch2_copygc_wait_amount(struct bch_fs * c)278 unsigned long bch2_copygc_wait_amount(struct bch_fs *c)
279 {
280 	s64 wait = S64_MAX, fragmented_allowed, fragmented;
281 
282 	for_each_rw_member(c, ca) {
283 		struct bch_dev_usage_full usage_full = bch2_dev_usage_full_read(ca);
284 		struct bch_dev_usage usage;
285 
286 		for (unsigned i = 0; i < BCH_DATA_NR; i++)
287 			usage.buckets[i] = usage_full.d[i].buckets;
288 
289 		fragmented_allowed = ((__dev_buckets_available(ca, usage, BCH_WATERMARK_stripe) *
290 				       ca->mi.bucket_size) >> 1);
291 		fragmented = 0;
292 
293 		for (unsigned i = 0; i < BCH_DATA_NR; i++)
294 			if (data_type_movable(i))
295 				fragmented += usage_full.d[i].fragmented;
296 
297 		wait = min(wait, max(0LL, fragmented_allowed - fragmented));
298 	}
299 
300 	return wait;
301 }
302 
bch2_copygc_wait_to_text(struct printbuf * out,struct bch_fs * c)303 void bch2_copygc_wait_to_text(struct printbuf *out, struct bch_fs *c)
304 {
305 	printbuf_tabstop_push(out, 32);
306 	prt_printf(out, "running:\t%u\n",		c->copygc_running);
307 	prt_printf(out, "copygc_wait:\t%llu\n",		c->copygc_wait);
308 	prt_printf(out, "copygc_wait_at:\t%llu\n",	c->copygc_wait_at);
309 
310 	prt_printf(out, "Currently waiting for:\t");
311 	prt_human_readable_u64(out, max(0LL, c->copygc_wait -
312 					atomic64_read(&c->io_clock[WRITE].now)) << 9);
313 	prt_newline(out);
314 
315 	prt_printf(out, "Currently waiting since:\t");
316 	prt_human_readable_u64(out, max(0LL,
317 					atomic64_read(&c->io_clock[WRITE].now) -
318 					c->copygc_wait_at) << 9);
319 	prt_newline(out);
320 
321 	prt_printf(out, "Currently calculated wait:\t");
322 	prt_human_readable_u64(out, bch2_copygc_wait_amount(c));
323 	prt_newline(out);
324 
325 	rcu_read_lock();
326 	struct task_struct *t = rcu_dereference(c->copygc_thread);
327 	if (t)
328 		get_task_struct(t);
329 	rcu_read_unlock();
330 
331 	if (t) {
332 		bch2_prt_task_backtrace(out, t, 0, GFP_KERNEL);
333 		put_task_struct(t);
334 	}
335 }
336 
bch2_copygc_thread(void * arg)337 static int bch2_copygc_thread(void *arg)
338 {
339 	struct bch_fs *c = arg;
340 	struct moving_context ctxt;
341 	struct bch_move_stats move_stats;
342 	struct io_clock *clock = &c->io_clock[WRITE];
343 	struct buckets_in_flight *buckets;
344 	u64 last, wait;
345 	int ret = 0;
346 
347 	buckets = kzalloc(sizeof(struct buckets_in_flight), GFP_KERNEL);
348 	if (!buckets)
349 		return -ENOMEM;
350 	ret = rhashtable_init(&buckets->table, &bch_move_bucket_params);
351 	bch_err_msg(c, ret, "allocating copygc buckets in flight");
352 	if (ret) {
353 		kfree(buckets);
354 		return ret;
355 	}
356 
357 	set_freezable();
358 
359 	/*
360 	 * Data move operations can't run until after check_snapshots has
361 	 * completed, and bch2_snapshot_is_ancestor() is available.
362 	 */
363 	kthread_wait_freezable(c->recovery_pass_done > BCH_RECOVERY_PASS_check_snapshots ||
364 			       kthread_should_stop());
365 
366 	bch2_move_stats_init(&move_stats, "copygc");
367 	bch2_moving_ctxt_init(&ctxt, c, NULL, &move_stats,
368 			      writepoint_ptr(&c->copygc_write_point),
369 			      false);
370 
371 	while (!ret && !kthread_should_stop()) {
372 		bool did_work = false;
373 
374 		bch2_trans_unlock_long(ctxt.trans);
375 		cond_resched();
376 
377 		if (!c->opts.copygc_enabled) {
378 			move_buckets_wait(&ctxt, buckets, true);
379 			kthread_wait_freezable(c->opts.copygc_enabled ||
380 					       kthread_should_stop());
381 		}
382 
383 		if (unlikely(freezing(current))) {
384 			move_buckets_wait(&ctxt, buckets, true);
385 			__refrigerator(false);
386 			continue;
387 		}
388 
389 		last = atomic64_read(&clock->now);
390 		wait = bch2_copygc_wait_amount(c);
391 
392 		if (wait > clock->max_slop) {
393 			c->copygc_wait_at = last;
394 			c->copygc_wait = last + wait;
395 			move_buckets_wait(&ctxt, buckets, true);
396 			trace_and_count(c, copygc_wait, c, wait, last + wait);
397 			bch2_kthread_io_clock_wait(clock, last + wait,
398 					MAX_SCHEDULE_TIMEOUT);
399 			continue;
400 		}
401 
402 		c->copygc_wait = 0;
403 
404 		c->copygc_running = true;
405 		ret = bch2_copygc(&ctxt, buckets, &did_work);
406 		c->copygc_running = false;
407 
408 		wake_up(&c->copygc_running_wq);
409 
410 		if (!wait && !did_work) {
411 			u64 min_member_capacity = bch2_min_rw_member_capacity(c);
412 
413 			if (min_member_capacity == U64_MAX)
414 				min_member_capacity = 128 * 2048;
415 
416 			move_buckets_wait(&ctxt, buckets, true);
417 			bch2_kthread_io_clock_wait(clock, last + (min_member_capacity >> 6),
418 					MAX_SCHEDULE_TIMEOUT);
419 		}
420 	}
421 
422 	move_buckets_wait(&ctxt, buckets, true);
423 
424 	rhashtable_destroy(&buckets->table);
425 	kfree(buckets);
426 	bch2_moving_ctxt_exit(&ctxt);
427 	bch2_move_stats_exit(&move_stats, c);
428 
429 	return 0;
430 }
431 
bch2_copygc_stop(struct bch_fs * c)432 void bch2_copygc_stop(struct bch_fs *c)
433 {
434 	if (c->copygc_thread) {
435 		kthread_stop(c->copygc_thread);
436 		put_task_struct(c->copygc_thread);
437 	}
438 	c->copygc_thread = NULL;
439 }
440 
bch2_copygc_start(struct bch_fs * c)441 int bch2_copygc_start(struct bch_fs *c)
442 {
443 	struct task_struct *t;
444 	int ret;
445 
446 	if (c->copygc_thread)
447 		return 0;
448 
449 	if (c->opts.nochanges)
450 		return 0;
451 
452 	if (bch2_fs_init_fault("copygc_start"))
453 		return -ENOMEM;
454 
455 	t = kthread_create(bch2_copygc_thread, c, "bch-copygc/%s", c->name);
456 	ret = PTR_ERR_OR_ZERO(t);
457 	bch_err_msg(c, ret, "creating copygc thread");
458 	if (ret)
459 		return ret;
460 
461 	get_task_struct(t);
462 
463 	c->copygc_thread = t;
464 	wake_up_process(c->copygc_thread);
465 
466 	return 0;
467 }
468 
bch2_fs_copygc_init(struct bch_fs * c)469 void bch2_fs_copygc_init(struct bch_fs *c)
470 {
471 	init_waitqueue_head(&c->copygc_running_wq);
472 	c->copygc_running = false;
473 }
474