xref: /linux/fs/xfs/xfs_zone_gc.c (revision 7d4e49a77d9930c69751b9192448fda6ff9100f1)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Copyright (c) 2023-2025 Christoph Hellwig.
4  * Copyright (c) 2024-2025, Western Digital Corporation or its affiliates.
5  */
6 #include "xfs.h"
7 #include "xfs_shared.h"
8 #include "xfs_format.h"
9 #include "xfs_log_format.h"
10 #include "xfs_trans_resv.h"
11 #include "xfs_mount.h"
12 #include "xfs_inode.h"
13 #include "xfs_btree.h"
14 #include "xfs_trans.h"
15 #include "xfs_icache.h"
16 #include "xfs_rmap.h"
17 #include "xfs_rtbitmap.h"
18 #include "xfs_rtrmap_btree.h"
19 #include "xfs_zone_alloc.h"
20 #include "xfs_zone_priv.h"
21 #include "xfs_zones.h"
22 #include "xfs_trace.h"
23 
24 /*
25  * Implement Garbage Collection (GC) of partially used zoned.
26  *
27  * To support the purely sequential writes in each zone, zoned XFS needs to be
28  * able to move data remaining in a zone out of it to reset the zone to prepare
29  * for writing to it again.
30  *
31  * This is done by the GC thread implemented in this file.  To support that a
32  * number of zones (XFS_GC_ZONES) is reserved from the user visible capacity to
33  * write the garbage collected data into.
34  *
35  * Whenever the available space is below the chosen threshold, the GC thread
36  * looks for potential non-empty but not fully used zones that are worth
37  * reclaiming.  Once found the rmap for the victim zone is queried, and after
38  * a bit of sorting to reduce fragmentation, the still live extents are read
39  * into memory and written to the GC target zone, and the bmap btree of the
40  * files is updated to point to the new location.  To avoid taking the IOLOCK
41  * and MMAPLOCK for the entire GC process and thus affecting the latency of
42  * user reads and writes to the files, the GC writes are speculative and the
43  * I/O completion checks that no other writes happened for the affected regions
44  * before remapping.
45  *
46  * Once a zone does not contain any valid data, be that through GC or user
47  * block removal, it is queued for for a zone reset.  The reset operation
48  * carefully ensures that the RT device cache is flushed and all transactions
49  * referencing the rmap have been committed to disk.
50  */
51 
52 /*
53  * Size of each GC scratch pad.  This is also the upper bound for each
54  * GC I/O, which helps to keep latency down.
55  */
56 #define XFS_GC_CHUNK_SIZE	SZ_1M
57 
58 /*
59  * Scratchpad data to read GCed data into.
60  *
61  * The offset member tracks where the next allocation starts, and freed tracks
62  * the amount of space that is not used anymore.
63  */
64 #define XFS_ZONE_GC_NR_SCRATCH	2
65 struct xfs_zone_scratch {
66 	struct folio			*folio;
67 	unsigned int			offset;
68 	unsigned int			freed;
69 };
70 
71 /*
72  * Chunk that is read and written for each GC operation.
73  *
74  * Note that for writes to actual zoned devices, the chunk can be split when
75  * reaching the hardware limit.
76  */
77 struct xfs_gc_bio {
78 	struct xfs_zone_gc_data		*data;
79 
80 	/*
81 	 * Entry into the reading/writing/resetting list.  Only accessed from
82 	 * the GC thread, so no locking needed.
83 	 */
84 	struct list_head		entry;
85 
86 	/*
87 	 * State of this gc_bio.  Done means the current I/O completed.
88 	 * Set from the bio end I/O handler, read from the GC thread.
89 	 */
90 	enum {
91 		XFS_GC_BIO_NEW,
92 		XFS_GC_BIO_DONE,
93 	} state;
94 
95 	/*
96 	 * Pointer to the inode and byte range in the inode that this
97 	 * GC chunk is operating on.
98 	 */
99 	struct xfs_inode		*ip;
100 	loff_t				offset;
101 	unsigned int			len;
102 
103 	/*
104 	 * Existing startblock (in the zone to be freed) and newly assigned
105 	 * daddr in the zone GCed into.
106 	 */
107 	xfs_fsblock_t			old_startblock;
108 	xfs_daddr_t			new_daddr;
109 	struct xfs_zone_scratch		*scratch;
110 
111 	/* Are we writing to a sequential write required zone? */
112 	bool				is_seq;
113 
114 	/* Open Zone being written to */
115 	struct xfs_open_zone		*oz;
116 
117 	/* Bio used for reads and writes, including the bvec used by it */
118 	struct bio_vec			bv;
119 	struct bio			bio;	/* must be last */
120 };
121 
122 #define XFS_ZONE_GC_RECS		1024
123 
124 /* iterator, needs to be reinitialized for each victim zone */
125 struct xfs_zone_gc_iter {
126 	struct xfs_rtgroup		*victim_rtg;
127 	unsigned int			rec_count;
128 	unsigned int			rec_idx;
129 	xfs_agblock_t			next_startblock;
130 	struct xfs_rmap_irec		*recs;
131 };
132 
133 /*
134  * Per-mount GC state.
135  */
136 struct xfs_zone_gc_data {
137 	struct xfs_mount		*mp;
138 
139 	/* bioset used to allocate the gc_bios */
140 	struct bio_set			bio_set;
141 
142 	/*
143 	 * Scratchpad used, and index to indicated which one is used.
144 	 */
145 	struct xfs_zone_scratch		scratch[XFS_ZONE_GC_NR_SCRATCH];
146 	unsigned int			scratch_idx;
147 
148 	/*
149 	 * List of bios currently being read, written and reset.
150 	 * These lists are only accessed by the GC thread itself, and must only
151 	 * be processed in order.
152 	 */
153 	struct list_head		reading;
154 	struct list_head		writing;
155 	struct list_head		resetting;
156 
157 	/*
158 	 * Iterator for the victim zone.
159 	 */
160 	struct xfs_zone_gc_iter		iter;
161 };
162 
163 /*
164  * We aim to keep enough zones free in stock to fully use the open zone limit
165  * for data placement purposes. Additionally, the m_zonegc_low_space tunable
166  * can be set to make sure a fraction of the unused blocks are available for
167  * writing.
168  */
169 bool
xfs_zoned_need_gc(struct xfs_mount * mp)170 xfs_zoned_need_gc(
171 	struct xfs_mount	*mp)
172 {
173 	s64			available, free, threshold;
174 	s32			remainder;
175 
176 	if (!xfs_group_marked(mp, XG_TYPE_RTG, XFS_RTG_RECLAIMABLE))
177 		return false;
178 
179 	available = xfs_estimate_freecounter(mp, XC_FREE_RTAVAILABLE);
180 
181 	if (available <
182 	    mp->m_groups[XG_TYPE_RTG].blocks *
183 	    (mp->m_max_open_zones - XFS_OPEN_GC_ZONES))
184 		return true;
185 
186 	free = xfs_estimate_freecounter(mp, XC_FREE_RTEXTENTS);
187 
188 	threshold = div_s64_rem(free, 100, &remainder);
189 	threshold = threshold * mp->m_zonegc_low_space +
190 		    remainder * div_s64(mp->m_zonegc_low_space, 100);
191 
192 	if (available < threshold)
193 		return true;
194 
195 	return false;
196 }
197 
198 static struct xfs_zone_gc_data *
xfs_zone_gc_data_alloc(struct xfs_mount * mp)199 xfs_zone_gc_data_alloc(
200 	struct xfs_mount	*mp)
201 {
202 	struct xfs_zone_gc_data	*data;
203 	int			i;
204 
205 	data = kzalloc(sizeof(*data), GFP_KERNEL);
206 	if (!data)
207 		return NULL;
208 	data->iter.recs = kcalloc(XFS_ZONE_GC_RECS, sizeof(*data->iter.recs),
209 			GFP_KERNEL);
210 	if (!data->iter.recs)
211 		goto out_free_data;
212 
213 	/*
214 	 * We actually only need a single bio_vec.  It would be nice to have
215 	 * a flag that only allocates the inline bvecs and not the separate
216 	 * bvec pool.
217 	 */
218 	if (bioset_init(&data->bio_set, 16, offsetof(struct xfs_gc_bio, bio),
219 			BIOSET_NEED_BVECS))
220 		goto out_free_recs;
221 	for (i = 0; i < XFS_ZONE_GC_NR_SCRATCH; i++) {
222 		data->scratch[i].folio =
223 			folio_alloc(GFP_KERNEL, get_order(XFS_GC_CHUNK_SIZE));
224 		if (!data->scratch[i].folio)
225 			goto out_free_scratch;
226 	}
227 	INIT_LIST_HEAD(&data->reading);
228 	INIT_LIST_HEAD(&data->writing);
229 	INIT_LIST_HEAD(&data->resetting);
230 	data->mp = mp;
231 	return data;
232 
233 out_free_scratch:
234 	while (--i >= 0)
235 		folio_put(data->scratch[i].folio);
236 	bioset_exit(&data->bio_set);
237 out_free_recs:
238 	kfree(data->iter.recs);
239 out_free_data:
240 	kfree(data);
241 	return NULL;
242 }
243 
244 static void
xfs_zone_gc_data_free(struct xfs_zone_gc_data * data)245 xfs_zone_gc_data_free(
246 	struct xfs_zone_gc_data	*data)
247 {
248 	int			i;
249 
250 	for (i = 0; i < XFS_ZONE_GC_NR_SCRATCH; i++)
251 		folio_put(data->scratch[i].folio);
252 	bioset_exit(&data->bio_set);
253 	kfree(data->iter.recs);
254 	kfree(data);
255 }
256 
257 static void
xfs_zone_gc_iter_init(struct xfs_zone_gc_iter * iter,struct xfs_rtgroup * victim_rtg)258 xfs_zone_gc_iter_init(
259 	struct xfs_zone_gc_iter	*iter,
260 	struct xfs_rtgroup	*victim_rtg)
261 
262 {
263 	iter->next_startblock = 0;
264 	iter->rec_count = 0;
265 	iter->rec_idx = 0;
266 	iter->victim_rtg = victim_rtg;
267 }
268 
269 /*
270  * Query the rmap of the victim zone to gather the records to evacuate.
271  */
272 static int
xfs_zone_gc_query_cb(struct xfs_btree_cur * cur,const struct xfs_rmap_irec * irec,void * private)273 xfs_zone_gc_query_cb(
274 	struct xfs_btree_cur	*cur,
275 	const struct xfs_rmap_irec *irec,
276 	void			*private)
277 {
278 	struct xfs_zone_gc_iter	*iter = private;
279 
280 	ASSERT(!XFS_RMAP_NON_INODE_OWNER(irec->rm_owner));
281 	ASSERT(!xfs_is_sb_inum(cur->bc_mp, irec->rm_owner));
282 	ASSERT(!(irec->rm_flags & (XFS_RMAP_ATTR_FORK | XFS_RMAP_BMBT_BLOCK)));
283 
284 	iter->recs[iter->rec_count] = *irec;
285 	if (++iter->rec_count == XFS_ZONE_GC_RECS) {
286 		iter->next_startblock =
287 			irec->rm_startblock + irec->rm_blockcount;
288 		return 1;
289 	}
290 	return 0;
291 }
292 
293 static int
xfs_zone_gc_rmap_rec_cmp(const void * a,const void * b)294 xfs_zone_gc_rmap_rec_cmp(
295 	const void			*a,
296 	const void			*b)
297 {
298 	const struct xfs_rmap_irec	*reca = a;
299 	const struct xfs_rmap_irec	*recb = b;
300 	int				diff;
301 
302 	diff = cmp_int(reca->rm_owner, recb->rm_owner);
303 	if (diff)
304 		return diff;
305 	return cmp_int(reca->rm_offset, recb->rm_offset);
306 }
307 
308 static int
xfs_zone_gc_query(struct xfs_mount * mp,struct xfs_zone_gc_iter * iter)309 xfs_zone_gc_query(
310 	struct xfs_mount	*mp,
311 	struct xfs_zone_gc_iter	*iter)
312 {
313 	struct xfs_rtgroup	*rtg = iter->victim_rtg;
314 	struct xfs_rmap_irec	ri_low = { };
315 	struct xfs_rmap_irec	ri_high;
316 	struct xfs_btree_cur	*cur;
317 	struct xfs_trans	*tp;
318 	int			error;
319 
320 	ASSERT(iter->next_startblock <= rtg_blocks(rtg));
321 	if (iter->next_startblock == rtg_blocks(rtg))
322 		goto done;
323 
324 	ASSERT(iter->next_startblock < rtg_blocks(rtg));
325 	ri_low.rm_startblock = iter->next_startblock;
326 	memset(&ri_high, 0xFF, sizeof(ri_high));
327 
328 	iter->rec_idx = 0;
329 	iter->rec_count = 0;
330 
331 	error = xfs_trans_alloc_empty(mp, &tp);
332 	if (error)
333 		return error;
334 
335 	xfs_rtgroup_lock(rtg, XFS_RTGLOCK_RMAP);
336 	cur = xfs_rtrmapbt_init_cursor(tp, rtg);
337 	error = xfs_rmap_query_range(cur, &ri_low, &ri_high,
338 			xfs_zone_gc_query_cb, iter);
339 	xfs_rtgroup_unlock(rtg, XFS_RTGLOCK_RMAP);
340 	xfs_btree_del_cursor(cur, error < 0 ? error : 0);
341 	xfs_trans_cancel(tp);
342 
343 	if (error < 0)
344 		return error;
345 
346 	/*
347 	 * Sort the rmap records by inode number and increasing offset to
348 	 * defragment the mappings.
349 	 *
350 	 * This could be further enhanced by an even bigger look ahead window,
351 	 * but that's better left until we have better detection of changes to
352 	 * inode mapping to avoid the potential of GCing already dead data.
353 	 */
354 	sort(iter->recs, iter->rec_count, sizeof(iter->recs[0]),
355 			xfs_zone_gc_rmap_rec_cmp, NULL);
356 
357 	if (error == 0) {
358 		/*
359 		 * We finished iterating through the zone.
360 		 */
361 		iter->next_startblock = rtg_blocks(rtg);
362 		if (iter->rec_count == 0)
363 			goto done;
364 	}
365 
366 	return 0;
367 done:
368 	xfs_rtgroup_rele(iter->victim_rtg);
369 	iter->victim_rtg = NULL;
370 	return 0;
371 }
372 
373 static bool
xfs_zone_gc_iter_next(struct xfs_mount * mp,struct xfs_zone_gc_iter * iter,struct xfs_rmap_irec * chunk_rec,struct xfs_inode ** ipp)374 xfs_zone_gc_iter_next(
375 	struct xfs_mount	*mp,
376 	struct xfs_zone_gc_iter	*iter,
377 	struct xfs_rmap_irec	*chunk_rec,
378 	struct xfs_inode	**ipp)
379 {
380 	struct xfs_rmap_irec	*irec;
381 	int			error;
382 
383 	if (!iter->victim_rtg)
384 		return false;
385 
386 retry:
387 	if (iter->rec_idx == iter->rec_count) {
388 		error = xfs_zone_gc_query(mp, iter);
389 		if (error)
390 			goto fail;
391 		if (!iter->victim_rtg)
392 			return false;
393 	}
394 
395 	irec = &iter->recs[iter->rec_idx];
396 	error = xfs_iget(mp, NULL, irec->rm_owner,
397 			XFS_IGET_UNTRUSTED | XFS_IGET_DONTCACHE, 0, ipp);
398 	if (error) {
399 		/*
400 		 * If the inode was already deleted, skip over it.
401 		 */
402 		if (error == -ENOENT) {
403 			iter->rec_idx++;
404 			goto retry;
405 		}
406 		goto fail;
407 	}
408 
409 	if (!S_ISREG(VFS_I(*ipp)->i_mode) || !XFS_IS_REALTIME_INODE(*ipp)) {
410 		iter->rec_idx++;
411 		xfs_irele(*ipp);
412 		goto retry;
413 	}
414 
415 	*chunk_rec = *irec;
416 	return true;
417 
418 fail:
419 	xfs_force_shutdown(mp, SHUTDOWN_META_IO_ERROR);
420 	return false;
421 }
422 
423 static void
xfs_zone_gc_iter_advance(struct xfs_zone_gc_iter * iter,xfs_extlen_t count_fsb)424 xfs_zone_gc_iter_advance(
425 	struct xfs_zone_gc_iter	*iter,
426 	xfs_extlen_t		count_fsb)
427 {
428 	struct xfs_rmap_irec	*irec = &iter->recs[iter->rec_idx];
429 
430 	irec->rm_offset += count_fsb;
431 	irec->rm_startblock += count_fsb;
432 	irec->rm_blockcount -= count_fsb;
433 	if (!irec->rm_blockcount)
434 		iter->rec_idx++;
435 }
436 
437 static struct xfs_rtgroup *
xfs_zone_gc_pick_victim_from(struct xfs_mount * mp,uint32_t bucket)438 xfs_zone_gc_pick_victim_from(
439 	struct xfs_mount	*mp,
440 	uint32_t		bucket)
441 {
442 	struct xfs_zone_info	*zi = mp->m_zone_info;
443 	uint32_t		victim_used = U32_MAX;
444 	struct xfs_rtgroup	*victim_rtg = NULL;
445 	uint32_t		bit;
446 
447 	if (!zi->zi_used_bucket_entries[bucket])
448 		return NULL;
449 
450 	for_each_set_bit(bit, zi->zi_used_bucket_bitmap[bucket],
451 			mp->m_sb.sb_rgcount) {
452 		struct xfs_rtgroup *rtg = xfs_rtgroup_grab(mp, bit);
453 
454 		if (!rtg)
455 			continue;
456 
457 		/* skip zones that are just waiting for a reset */
458 		if (rtg_rmap(rtg)->i_used_blocks == 0 ||
459 		    rtg_rmap(rtg)->i_used_blocks >= victim_used) {
460 			xfs_rtgroup_rele(rtg);
461 			continue;
462 		}
463 
464 		if (victim_rtg)
465 			xfs_rtgroup_rele(victim_rtg);
466 		victim_rtg = rtg;
467 		victim_used = rtg_rmap(rtg)->i_used_blocks;
468 
469 		/*
470 		 * Any zone that is less than 1 percent used is fair game for
471 		 * instant reclaim. All of these zones are in the last
472 		 * bucket, so avoid the expensive division for the zones
473 		 * in the other buckets.
474 		 */
475 		if (bucket == 0 &&
476 		    rtg_rmap(rtg)->i_used_blocks < rtg_blocks(rtg) / 100)
477 			break;
478 	}
479 
480 	return victim_rtg;
481 }
482 
483 /*
484  * Iterate through all zones marked as reclaimable and find a candidate to
485  * reclaim.
486  */
487 static bool
xfs_zone_gc_select_victim(struct xfs_zone_gc_data * data)488 xfs_zone_gc_select_victim(
489 	struct xfs_zone_gc_data	*data)
490 {
491 	struct xfs_zone_gc_iter	*iter = &data->iter;
492 	struct xfs_mount	*mp = data->mp;
493 	struct xfs_zone_info	*zi = mp->m_zone_info;
494 	struct xfs_rtgroup	*victim_rtg = NULL;
495 	unsigned int		bucket;
496 
497 	if (xfs_is_shutdown(mp))
498 		return false;
499 
500 	if (iter->victim_rtg)
501 		return true;
502 
503 	/*
504 	 * Don't start new work if we are asked to stop or park.
505 	 */
506 	if (kthread_should_stop() || kthread_should_park())
507 		return false;
508 
509 	if (!xfs_zoned_need_gc(mp))
510 		return false;
511 
512 	spin_lock(&zi->zi_used_buckets_lock);
513 	for (bucket = 0; bucket < XFS_ZONE_USED_BUCKETS; bucket++) {
514 		victim_rtg = xfs_zone_gc_pick_victim_from(mp, bucket);
515 		if (victim_rtg)
516 			break;
517 	}
518 	spin_unlock(&zi->zi_used_buckets_lock);
519 
520 	if (!victim_rtg)
521 		return false;
522 
523 	trace_xfs_zone_gc_select_victim(victim_rtg, bucket);
524 	xfs_zone_gc_iter_init(iter, victim_rtg);
525 	return true;
526 }
527 
528 static struct xfs_open_zone *
xfs_zone_gc_steal_open(struct xfs_zone_info * zi)529 xfs_zone_gc_steal_open(
530 	struct xfs_zone_info	*zi)
531 {
532 	struct xfs_open_zone	*oz, *found = NULL;
533 
534 	spin_lock(&zi->zi_open_zones_lock);
535 	list_for_each_entry(oz, &zi->zi_open_zones, oz_entry) {
536 		if (!found ||
537 		    oz->oz_write_pointer < found->oz_write_pointer)
538 			found = oz;
539 	}
540 
541 	if (found) {
542 		found->oz_is_gc = true;
543 		list_del_init(&found->oz_entry);
544 		zi->zi_nr_open_zones--;
545 	}
546 
547 	spin_unlock(&zi->zi_open_zones_lock);
548 	return found;
549 }
550 
551 static struct xfs_open_zone *
xfs_zone_gc_select_target(struct xfs_mount * mp)552 xfs_zone_gc_select_target(
553 	struct xfs_mount	*mp)
554 {
555 	struct xfs_zone_info	*zi = mp->m_zone_info;
556 	struct xfs_open_zone	*oz = zi->zi_open_gc_zone;
557 
558 	/*
559 	 * We need to wait for pending writes to finish.
560 	 */
561 	if (oz && oz->oz_written < rtg_blocks(oz->oz_rtg))
562 		return NULL;
563 
564 	ASSERT(zi->zi_nr_open_zones <=
565 		mp->m_max_open_zones - XFS_OPEN_GC_ZONES);
566 	oz = xfs_open_zone(mp, WRITE_LIFE_NOT_SET, true);
567 	if (oz)
568 		trace_xfs_zone_gc_target_opened(oz->oz_rtg);
569 	spin_lock(&zi->zi_open_zones_lock);
570 	zi->zi_open_gc_zone = oz;
571 	spin_unlock(&zi->zi_open_zones_lock);
572 	return oz;
573 }
574 
575 /*
576  * Ensure we have a valid open zone to write the GC data to.
577  *
578  * If the current target zone has space keep writing to it, else first wait for
579  * all pending writes and then pick a new one.
580  */
581 static struct xfs_open_zone *
xfs_zone_gc_ensure_target(struct xfs_mount * mp)582 xfs_zone_gc_ensure_target(
583 	struct xfs_mount	*mp)
584 {
585 	struct xfs_open_zone	*oz = mp->m_zone_info->zi_open_gc_zone;
586 
587 	if (!oz || oz->oz_write_pointer == rtg_blocks(oz->oz_rtg))
588 		return xfs_zone_gc_select_target(mp);
589 	return oz;
590 }
591 
592 static unsigned int
xfs_zone_gc_scratch_available(struct xfs_zone_gc_data * data)593 xfs_zone_gc_scratch_available(
594 	struct xfs_zone_gc_data	*data)
595 {
596 	return XFS_GC_CHUNK_SIZE - data->scratch[data->scratch_idx].offset;
597 }
598 
599 static bool
xfs_zone_gc_space_available(struct xfs_zone_gc_data * data)600 xfs_zone_gc_space_available(
601 	struct xfs_zone_gc_data	*data)
602 {
603 	struct xfs_open_zone	*oz;
604 
605 	oz = xfs_zone_gc_ensure_target(data->mp);
606 	if (!oz)
607 		return false;
608 	return oz->oz_write_pointer < rtg_blocks(oz->oz_rtg) &&
609 		xfs_zone_gc_scratch_available(data);
610 }
611 
612 static void
xfs_zone_gc_end_io(struct bio * bio)613 xfs_zone_gc_end_io(
614 	struct bio		*bio)
615 {
616 	struct xfs_gc_bio	*chunk =
617 		container_of(bio, struct xfs_gc_bio, bio);
618 	struct xfs_zone_gc_data	*data = chunk->data;
619 
620 	WRITE_ONCE(chunk->state, XFS_GC_BIO_DONE);
621 	wake_up_process(data->mp->m_zone_info->zi_gc_thread);
622 }
623 
624 static struct xfs_open_zone *
xfs_zone_gc_alloc_blocks(struct xfs_zone_gc_data * data,xfs_extlen_t * count_fsb,xfs_daddr_t * daddr,bool * is_seq)625 xfs_zone_gc_alloc_blocks(
626 	struct xfs_zone_gc_data	*data,
627 	xfs_extlen_t		*count_fsb,
628 	xfs_daddr_t		*daddr,
629 	bool			*is_seq)
630 {
631 	struct xfs_mount	*mp = data->mp;
632 	struct xfs_open_zone	*oz;
633 
634 	oz = xfs_zone_gc_ensure_target(mp);
635 	if (!oz)
636 		return NULL;
637 
638 	*count_fsb = min(*count_fsb,
639 		XFS_B_TO_FSB(mp, xfs_zone_gc_scratch_available(data)));
640 
641 	/*
642 	 * Directly allocate GC blocks from the reserved pool.
643 	 *
644 	 * If we'd take them from the normal pool we could be stealing blocks
645 	 * from a regular writer, which would then have to wait for GC and
646 	 * deadlock.
647 	 */
648 	spin_lock(&mp->m_sb_lock);
649 	*count_fsb = min(*count_fsb,
650 			rtg_blocks(oz->oz_rtg) - oz->oz_write_pointer);
651 	*count_fsb = min3(*count_fsb,
652 			mp->m_free[XC_FREE_RTEXTENTS].res_avail,
653 			mp->m_free[XC_FREE_RTAVAILABLE].res_avail);
654 	mp->m_free[XC_FREE_RTEXTENTS].res_avail -= *count_fsb;
655 	mp->m_free[XC_FREE_RTAVAILABLE].res_avail -= *count_fsb;
656 	spin_unlock(&mp->m_sb_lock);
657 
658 	if (!*count_fsb)
659 		return NULL;
660 
661 	*daddr = xfs_gbno_to_daddr(&oz->oz_rtg->rtg_group, 0);
662 	*is_seq = bdev_zone_is_seq(mp->m_rtdev_targp->bt_bdev, *daddr);
663 	if (!*is_seq)
664 		*daddr += XFS_FSB_TO_BB(mp, oz->oz_write_pointer);
665 	oz->oz_write_pointer += *count_fsb;
666 	atomic_inc(&oz->oz_ref);
667 	return oz;
668 }
669 
670 static bool
xfs_zone_gc_start_chunk(struct xfs_zone_gc_data * data)671 xfs_zone_gc_start_chunk(
672 	struct xfs_zone_gc_data	*data)
673 {
674 	struct xfs_zone_gc_iter	*iter = &data->iter;
675 	struct xfs_mount	*mp = data->mp;
676 	struct block_device	*bdev = mp->m_rtdev_targp->bt_bdev;
677 	struct xfs_open_zone	*oz;
678 	struct xfs_rmap_irec	irec;
679 	struct xfs_gc_bio	*chunk;
680 	struct xfs_inode	*ip;
681 	struct bio		*bio;
682 	xfs_daddr_t		daddr;
683 	bool			is_seq;
684 
685 	if (xfs_is_shutdown(mp))
686 		return false;
687 
688 	if (!xfs_zone_gc_iter_next(mp, iter, &irec, &ip))
689 		return false;
690 	oz = xfs_zone_gc_alloc_blocks(data, &irec.rm_blockcount, &daddr,
691 			&is_seq);
692 	if (!oz) {
693 		xfs_irele(ip);
694 		return false;
695 	}
696 
697 	bio = bio_alloc_bioset(bdev, 1, REQ_OP_READ, GFP_NOFS, &data->bio_set);
698 
699 	chunk = container_of(bio, struct xfs_gc_bio, bio);
700 	chunk->ip = ip;
701 	chunk->offset = XFS_FSB_TO_B(mp, irec.rm_offset);
702 	chunk->len = XFS_FSB_TO_B(mp, irec.rm_blockcount);
703 	chunk->old_startblock =
704 		xfs_rgbno_to_rtb(iter->victim_rtg, irec.rm_startblock);
705 	chunk->new_daddr = daddr;
706 	chunk->is_seq = is_seq;
707 	chunk->scratch = &data->scratch[data->scratch_idx];
708 	chunk->data = data;
709 	chunk->oz = oz;
710 
711 	bio->bi_iter.bi_sector = xfs_rtb_to_daddr(mp, chunk->old_startblock);
712 	bio->bi_end_io = xfs_zone_gc_end_io;
713 	bio_add_folio_nofail(bio, chunk->scratch->folio, chunk->len,
714 			chunk->scratch->offset);
715 	chunk->scratch->offset += chunk->len;
716 	if (chunk->scratch->offset == XFS_GC_CHUNK_SIZE) {
717 		data->scratch_idx =
718 			(data->scratch_idx + 1) % XFS_ZONE_GC_NR_SCRATCH;
719 	}
720 	WRITE_ONCE(chunk->state, XFS_GC_BIO_NEW);
721 	list_add_tail(&chunk->entry, &data->reading);
722 	xfs_zone_gc_iter_advance(iter, irec.rm_blockcount);
723 
724 	submit_bio(bio);
725 	return true;
726 }
727 
728 static void
xfs_zone_gc_free_chunk(struct xfs_gc_bio * chunk)729 xfs_zone_gc_free_chunk(
730 	struct xfs_gc_bio	*chunk)
731 {
732 	list_del(&chunk->entry);
733 	xfs_open_zone_put(chunk->oz);
734 	xfs_irele(chunk->ip);
735 	bio_put(&chunk->bio);
736 }
737 
738 static void
xfs_zone_gc_submit_write(struct xfs_zone_gc_data * data,struct xfs_gc_bio * chunk)739 xfs_zone_gc_submit_write(
740 	struct xfs_zone_gc_data	*data,
741 	struct xfs_gc_bio	*chunk)
742 {
743 	if (chunk->is_seq) {
744 		chunk->bio.bi_opf &= ~REQ_OP_WRITE;
745 		chunk->bio.bi_opf |= REQ_OP_ZONE_APPEND;
746 	}
747 	chunk->bio.bi_iter.bi_sector = chunk->new_daddr;
748 	chunk->bio.bi_end_io = xfs_zone_gc_end_io;
749 	submit_bio(&chunk->bio);
750 }
751 
752 static struct xfs_gc_bio *
xfs_zone_gc_split_write(struct xfs_zone_gc_data * data,struct xfs_gc_bio * chunk)753 xfs_zone_gc_split_write(
754 	struct xfs_zone_gc_data	*data,
755 	struct xfs_gc_bio	*chunk)
756 {
757 	struct queue_limits	*lim =
758 		&bdev_get_queue(chunk->bio.bi_bdev)->limits;
759 	struct xfs_gc_bio	*split_chunk;
760 	int			split_sectors;
761 	unsigned int		split_len;
762 	struct bio		*split;
763 	unsigned int		nsegs;
764 
765 	if (!chunk->is_seq)
766 		return NULL;
767 
768 	split_sectors = bio_split_rw_at(&chunk->bio, lim, &nsegs,
769 			lim->max_zone_append_sectors << SECTOR_SHIFT);
770 	if (!split_sectors)
771 		return NULL;
772 
773 	/* ensure the split chunk is still block size aligned */
774 	split_sectors = ALIGN_DOWN(split_sectors << SECTOR_SHIFT,
775 			data->mp->m_sb.sb_blocksize) >> SECTOR_SHIFT;
776 	split_len = split_sectors << SECTOR_SHIFT;
777 
778 	split = bio_split(&chunk->bio, split_sectors, GFP_NOFS, &data->bio_set);
779 	split_chunk = container_of(split, struct xfs_gc_bio, bio);
780 	split_chunk->data = data;
781 	ihold(VFS_I(chunk->ip));
782 	split_chunk->ip = chunk->ip;
783 	split_chunk->is_seq = chunk->is_seq;
784 	split_chunk->scratch = chunk->scratch;
785 	split_chunk->offset = chunk->offset;
786 	split_chunk->len = split_len;
787 	split_chunk->old_startblock = chunk->old_startblock;
788 	split_chunk->new_daddr = chunk->new_daddr;
789 	split_chunk->oz = chunk->oz;
790 	atomic_inc(&chunk->oz->oz_ref);
791 
792 	chunk->offset += split_len;
793 	chunk->len -= split_len;
794 	chunk->old_startblock += XFS_B_TO_FSB(data->mp, split_len);
795 
796 	/* add right before the original chunk */
797 	WRITE_ONCE(split_chunk->state, XFS_GC_BIO_NEW);
798 	list_add_tail(&split_chunk->entry, &chunk->entry);
799 	return split_chunk;
800 }
801 
802 static void
xfs_zone_gc_write_chunk(struct xfs_gc_bio * chunk)803 xfs_zone_gc_write_chunk(
804 	struct xfs_gc_bio	*chunk)
805 {
806 	struct xfs_zone_gc_data	*data = chunk->data;
807 	struct xfs_mount	*mp = chunk->ip->i_mount;
808 	phys_addr_t		bvec_paddr =
809 		bvec_phys(bio_first_bvec_all(&chunk->bio));
810 	struct xfs_gc_bio	*split_chunk;
811 
812 	if (chunk->bio.bi_status)
813 		xfs_force_shutdown(mp, SHUTDOWN_META_IO_ERROR);
814 	if (xfs_is_shutdown(mp)) {
815 		xfs_zone_gc_free_chunk(chunk);
816 		return;
817 	}
818 
819 	WRITE_ONCE(chunk->state, XFS_GC_BIO_NEW);
820 	list_move_tail(&chunk->entry, &data->writing);
821 
822 	bio_reset(&chunk->bio, mp->m_rtdev_targp->bt_bdev, REQ_OP_WRITE);
823 	bio_add_folio_nofail(&chunk->bio, chunk->scratch->folio, chunk->len,
824 			offset_in_folio(chunk->scratch->folio, bvec_paddr));
825 
826 	while ((split_chunk = xfs_zone_gc_split_write(data, chunk)))
827 		xfs_zone_gc_submit_write(data, split_chunk);
828 	xfs_zone_gc_submit_write(data, chunk);
829 }
830 
831 static void
xfs_zone_gc_finish_chunk(struct xfs_gc_bio * chunk)832 xfs_zone_gc_finish_chunk(
833 	struct xfs_gc_bio	*chunk)
834 {
835 	uint			iolock = XFS_IOLOCK_EXCL | XFS_MMAPLOCK_EXCL;
836 	struct xfs_inode	*ip = chunk->ip;
837 	struct xfs_mount	*mp = ip->i_mount;
838 	int			error;
839 
840 	if (chunk->bio.bi_status)
841 		xfs_force_shutdown(mp, SHUTDOWN_META_IO_ERROR);
842 	if (xfs_is_shutdown(mp)) {
843 		xfs_zone_gc_free_chunk(chunk);
844 		return;
845 	}
846 
847 	chunk->scratch->freed += chunk->len;
848 	if (chunk->scratch->freed == chunk->scratch->offset) {
849 		chunk->scratch->offset = 0;
850 		chunk->scratch->freed = 0;
851 	}
852 
853 	/*
854 	 * Cycle through the iolock and wait for direct I/O and layouts to
855 	 * ensure no one is reading from the old mapping before it goes away.
856 	 *
857 	 * Note that xfs_zoned_end_io() below checks that no other writer raced
858 	 * with us to update the mapping by checking that the old startblock
859 	 * didn't change.
860 	 */
861 	xfs_ilock(ip, iolock);
862 	error = xfs_break_layouts(VFS_I(ip), &iolock, BREAK_UNMAP);
863 	if (!error)
864 		inode_dio_wait(VFS_I(ip));
865 	xfs_iunlock(ip, iolock);
866 	if (error)
867 		goto free;
868 
869 	if (chunk->is_seq)
870 		chunk->new_daddr = chunk->bio.bi_iter.bi_sector;
871 	error = xfs_zoned_end_io(ip, chunk->offset, chunk->len,
872 			chunk->new_daddr, chunk->oz, chunk->old_startblock);
873 free:
874 	if (error)
875 		xfs_force_shutdown(mp, SHUTDOWN_META_IO_ERROR);
876 	xfs_zone_gc_free_chunk(chunk);
877 }
878 
879 static void
xfs_zone_gc_finish_reset(struct xfs_gc_bio * chunk)880 xfs_zone_gc_finish_reset(
881 	struct xfs_gc_bio	*chunk)
882 {
883 	struct xfs_rtgroup	*rtg = chunk->bio.bi_private;
884 	struct xfs_mount	*mp = rtg_mount(rtg);
885 	struct xfs_zone_info	*zi = mp->m_zone_info;
886 
887 	if (chunk->bio.bi_status) {
888 		xfs_force_shutdown(mp, SHUTDOWN_META_IO_ERROR);
889 		goto out;
890 	}
891 
892 	xfs_group_set_mark(&rtg->rtg_group, XFS_RTG_FREE);
893 	atomic_inc(&zi->zi_nr_free_zones);
894 
895 	xfs_zoned_add_available(mp, rtg_blocks(rtg));
896 
897 	wake_up_all(&zi->zi_zone_wait);
898 out:
899 	list_del(&chunk->entry);
900 	bio_put(&chunk->bio);
901 }
902 
903 static bool
xfs_zone_gc_prepare_reset(struct bio * bio,struct xfs_rtgroup * rtg)904 xfs_zone_gc_prepare_reset(
905 	struct bio		*bio,
906 	struct xfs_rtgroup	*rtg)
907 {
908 	trace_xfs_zone_reset(rtg);
909 
910 	ASSERT(rtg_rmap(rtg)->i_used_blocks == 0);
911 	bio->bi_iter.bi_sector = xfs_gbno_to_daddr(&rtg->rtg_group, 0);
912 	if (!bdev_zone_is_seq(bio->bi_bdev, bio->bi_iter.bi_sector)) {
913 		if (!bdev_max_discard_sectors(bio->bi_bdev))
914 			return false;
915 		bio->bi_opf = REQ_OP_DISCARD | REQ_SYNC;
916 		bio->bi_iter.bi_size =
917 			XFS_FSB_TO_B(rtg_mount(rtg), rtg_blocks(rtg));
918 	}
919 
920 	return true;
921 }
922 
923 int
xfs_zone_gc_reset_sync(struct xfs_rtgroup * rtg)924 xfs_zone_gc_reset_sync(
925 	struct xfs_rtgroup	*rtg)
926 {
927 	int			error = 0;
928 	struct bio		bio;
929 
930 	bio_init(&bio, rtg_mount(rtg)->m_rtdev_targp->bt_bdev, NULL, 0,
931 			REQ_OP_ZONE_RESET);
932 	if (xfs_zone_gc_prepare_reset(&bio, rtg))
933 		error = submit_bio_wait(&bio);
934 	bio_uninit(&bio);
935 
936 	return error;
937 }
938 
939 static void
xfs_zone_gc_reset_zones(struct xfs_zone_gc_data * data,struct xfs_group * reset_list)940 xfs_zone_gc_reset_zones(
941 	struct xfs_zone_gc_data	*data,
942 	struct xfs_group	*reset_list)
943 {
944 	struct xfs_group	*next = reset_list;
945 
946 	if (blkdev_issue_flush(data->mp->m_rtdev_targp->bt_bdev) < 0) {
947 		xfs_force_shutdown(data->mp, SHUTDOWN_META_IO_ERROR);
948 		return;
949 	}
950 
951 	do {
952 		struct xfs_rtgroup	*rtg = to_rtg(next);
953 		struct xfs_gc_bio	*chunk;
954 		struct bio		*bio;
955 
956 		xfs_log_force_inode(rtg_rmap(rtg));
957 
958 		next = rtg_group(rtg)->xg_next_reset;
959 		rtg_group(rtg)->xg_next_reset = NULL;
960 
961 		bio = bio_alloc_bioset(rtg_mount(rtg)->m_rtdev_targp->bt_bdev,
962 				0, REQ_OP_ZONE_RESET, GFP_NOFS, &data->bio_set);
963 		bio->bi_private = rtg;
964 		bio->bi_end_io = xfs_zone_gc_end_io;
965 
966 		chunk = container_of(bio, struct xfs_gc_bio, bio);
967 		chunk->data = data;
968 		WRITE_ONCE(chunk->state, XFS_GC_BIO_NEW);
969 		list_add_tail(&chunk->entry, &data->resetting);
970 
971 		/*
972 		 * Also use the bio to drive the state machine when neither
973 		 * zone reset nor discard is supported to keep things simple.
974 		 */
975 		if (xfs_zone_gc_prepare_reset(bio, rtg))
976 			submit_bio(bio);
977 		else
978 			bio_endio(bio);
979 	} while (next);
980 }
981 
982 /*
983  * Handle the work to read and write data for GC and to reset the zones,
984  * including handling all completions.
985  *
986  * Note that the order of the chunks is preserved so that we don't undo the
987  * optimal order established by xfs_zone_gc_query().
988  */
989 static bool
xfs_zone_gc_handle_work(struct xfs_zone_gc_data * data)990 xfs_zone_gc_handle_work(
991 	struct xfs_zone_gc_data	*data)
992 {
993 	struct xfs_zone_info	*zi = data->mp->m_zone_info;
994 	struct xfs_gc_bio	*chunk, *next;
995 	struct xfs_group	*reset_list;
996 	struct blk_plug		plug;
997 
998 	spin_lock(&zi->zi_reset_list_lock);
999 	reset_list = zi->zi_reset_list;
1000 	zi->zi_reset_list = NULL;
1001 	spin_unlock(&zi->zi_reset_list_lock);
1002 
1003 	if (!xfs_zone_gc_select_victim(data) ||
1004 	    !xfs_zone_gc_space_available(data)) {
1005 		if (list_empty(&data->reading) &&
1006 		    list_empty(&data->writing) &&
1007 		    list_empty(&data->resetting) &&
1008 		    !reset_list)
1009 			return false;
1010 	}
1011 
1012 	__set_current_state(TASK_RUNNING);
1013 	try_to_freeze();
1014 
1015 	if (reset_list)
1016 		xfs_zone_gc_reset_zones(data, reset_list);
1017 
1018 	list_for_each_entry_safe(chunk, next, &data->resetting, entry) {
1019 		if (READ_ONCE(chunk->state) != XFS_GC_BIO_DONE)
1020 			break;
1021 		xfs_zone_gc_finish_reset(chunk);
1022 	}
1023 
1024 	list_for_each_entry_safe(chunk, next, &data->writing, entry) {
1025 		if (READ_ONCE(chunk->state) != XFS_GC_BIO_DONE)
1026 			break;
1027 		xfs_zone_gc_finish_chunk(chunk);
1028 	}
1029 
1030 	blk_start_plug(&plug);
1031 	list_for_each_entry_safe(chunk, next, &data->reading, entry) {
1032 		if (READ_ONCE(chunk->state) != XFS_GC_BIO_DONE)
1033 			break;
1034 		xfs_zone_gc_write_chunk(chunk);
1035 	}
1036 	blk_finish_plug(&plug);
1037 
1038 	blk_start_plug(&plug);
1039 	while (xfs_zone_gc_start_chunk(data))
1040 		;
1041 	blk_finish_plug(&plug);
1042 	return true;
1043 }
1044 
1045 /*
1046  * Note that the current GC algorithm would break reflinks and thus duplicate
1047  * data that was shared by multiple owners before.  Because of that reflinks
1048  * are currently not supported on zoned file systems and can't be created or
1049  * mounted.
1050  */
1051 static int
xfs_zoned_gcd(void * private)1052 xfs_zoned_gcd(
1053 	void			*private)
1054 {
1055 	struct xfs_zone_gc_data	*data = private;
1056 	struct xfs_mount	*mp = data->mp;
1057 	struct xfs_zone_info	*zi = mp->m_zone_info;
1058 	unsigned int		nofs_flag;
1059 
1060 	nofs_flag = memalloc_nofs_save();
1061 	set_freezable();
1062 
1063 	for (;;) {
1064 		set_current_state(TASK_INTERRUPTIBLE | TASK_FREEZABLE);
1065 		xfs_set_zonegc_running(mp);
1066 		if (xfs_zone_gc_handle_work(data))
1067 			continue;
1068 
1069 		if (list_empty(&data->reading) &&
1070 		    list_empty(&data->writing) &&
1071 		    list_empty(&data->resetting) &&
1072 		    !zi->zi_reset_list) {
1073 			xfs_clear_zonegc_running(mp);
1074 			xfs_zoned_resv_wake_all(mp);
1075 
1076 			if (kthread_should_stop()) {
1077 				__set_current_state(TASK_RUNNING);
1078 				break;
1079 			}
1080 
1081 			if (kthread_should_park()) {
1082 				__set_current_state(TASK_RUNNING);
1083 				kthread_parkme();
1084 				continue;
1085 			}
1086 		}
1087 
1088 		schedule();
1089 	}
1090 	xfs_clear_zonegc_running(mp);
1091 
1092 	if (data->iter.victim_rtg)
1093 		xfs_rtgroup_rele(data->iter.victim_rtg);
1094 
1095 	memalloc_nofs_restore(nofs_flag);
1096 	xfs_zone_gc_data_free(data);
1097 	return 0;
1098 }
1099 
1100 void
xfs_zone_gc_start(struct xfs_mount * mp)1101 xfs_zone_gc_start(
1102 	struct xfs_mount	*mp)
1103 {
1104 	if (xfs_has_zoned(mp))
1105 		kthread_unpark(mp->m_zone_info->zi_gc_thread);
1106 }
1107 
1108 void
xfs_zone_gc_stop(struct xfs_mount * mp)1109 xfs_zone_gc_stop(
1110 	struct xfs_mount	*mp)
1111 {
1112 	if (xfs_has_zoned(mp))
1113 		kthread_park(mp->m_zone_info->zi_gc_thread);
1114 }
1115 
1116 int
xfs_zone_gc_mount(struct xfs_mount * mp)1117 xfs_zone_gc_mount(
1118 	struct xfs_mount	*mp)
1119 {
1120 	struct xfs_zone_info	*zi = mp->m_zone_info;
1121 	struct xfs_zone_gc_data	*data;
1122 	struct xfs_open_zone	*oz;
1123 	int			error;
1124 
1125 	/*
1126 	 * If there are no free zones available for GC, pick the open zone with
1127 	 * the least used space to GC into.  This should only happen after an
1128 	 * unclean shutdown near ENOSPC while GC was ongoing.
1129 	 *
1130 	 * We also need to do this for the first gc zone allocation if we
1131 	 * unmounted while at the open limit.
1132 	 */
1133 	if (!xfs_group_marked(mp, XG_TYPE_RTG, XFS_RTG_FREE) ||
1134 	    zi->zi_nr_open_zones == mp->m_max_open_zones)
1135 		oz = xfs_zone_gc_steal_open(zi);
1136 	else
1137 		oz = xfs_open_zone(mp, WRITE_LIFE_NOT_SET, true);
1138 	if (!oz) {
1139 		xfs_warn(mp, "unable to allocate a zone for gc");
1140 		error = -EIO;
1141 		goto out;
1142 	}
1143 
1144 	trace_xfs_zone_gc_target_opened(oz->oz_rtg);
1145 	zi->zi_open_gc_zone = oz;
1146 
1147 	data = xfs_zone_gc_data_alloc(mp);
1148 	if (!data) {
1149 		error = -ENOMEM;
1150 		goto out_put_gc_zone;
1151 	}
1152 
1153 	mp->m_zone_info->zi_gc_thread = kthread_create(xfs_zoned_gcd, data,
1154 			"xfs-zone-gc/%s", mp->m_super->s_id);
1155 	if (IS_ERR(mp->m_zone_info->zi_gc_thread)) {
1156 		xfs_warn(mp, "unable to create zone gc thread");
1157 		error = PTR_ERR(mp->m_zone_info->zi_gc_thread);
1158 		goto out_free_gc_data;
1159 	}
1160 
1161 	/* xfs_zone_gc_start will unpark for rw mounts */
1162 	kthread_park(mp->m_zone_info->zi_gc_thread);
1163 	return 0;
1164 
1165 out_free_gc_data:
1166 	kfree(data);
1167 out_put_gc_zone:
1168 	xfs_open_zone_put(zi->zi_open_gc_zone);
1169 out:
1170 	return error;
1171 }
1172 
1173 void
xfs_zone_gc_unmount(struct xfs_mount * mp)1174 xfs_zone_gc_unmount(
1175 	struct xfs_mount	*mp)
1176 {
1177 	struct xfs_zone_info	*zi = mp->m_zone_info;
1178 
1179 	kthread_stop(zi->zi_gc_thread);
1180 	if (zi->zi_open_gc_zone)
1181 		xfs_open_zone_put(zi->zi_open_gc_zone);
1182 }
1183