1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * Copyright (C) 2017 Western Digital Corporation or its affiliates.
4 *
5 * This file is released under the GPL.
6 */
7
8#include "dm-zoned.h"
9
10#include <linux/module.h>
11
12#define DM_MSG_PREFIX "zoned reclaim"
13
14struct dmz_reclaim {
15 struct dmz_metadata *metadata;
16
17 struct delayed_work work;
18 struct workqueue_struct *wq;
19
20 struct dm_kcopyd_client *kc;
21 struct dm_kcopyd_throttle kc_throttle;
22 int kc_err;
23
24 int dev_idx;
25
26 unsigned long flags;
27
28 /* Last target access time */
29 unsigned long atime;
30};
31
32/*
33 * Reclaim state flags.
34 */
35enum {
36 DMZ_RECLAIM_KCOPY,
37};
38
39/*
40 * Number of seconds of target BIO inactivity to consider the target idle.
41 */
42#define DMZ_IDLE_PERIOD (10UL * HZ)
43
44/*
45 * Percentage of unmapped (free) random zones below which reclaim starts
46 * even if the target is busy.
47 */
48#define DMZ_RECLAIM_LOW_UNMAP_ZONES 30
49
50/*
51 * Percentage of unmapped (free) random zones above which reclaim will
52 * stop if the target is busy.
53 */
54#define DMZ_RECLAIM_HIGH_UNMAP_ZONES 50
55
56/*
57 * Align a sequential zone write pointer to chunk_block.
58 */
59static int dmz_reclaim_align_wp(struct dmz_reclaim *zrc, struct dm_zone *zone,
60 sector_t block)
61{
62 struct dmz_metadata *zmd = zrc->metadata;
63 struct dmz_dev *dev = zone->dev;
64 sector_t wp_block = zone->wp_block;
65 unsigned int nr_blocks;
66 int ret;
67
68 if (wp_block == block)
69 return 0;
70
71 if (wp_block > block)
72 return -EIO;
73
74 /*
75 * Zeroout the space between the write
76 * pointer and the requested position.
77 */
78 nr_blocks = block - wp_block;
79 ret = blkdev_issue_zeroout(bdev: dev->bdev,
80 sector: dmz_start_sect(zmd, zone) + dmz_blk2sect(wp_block),
81 dmz_blk2sect(nr_blocks), GFP_NOIO, flags: 0);
82 if (ret) {
83 dmz_dev_err(dev,
84 "Align zone %u wp %llu to %llu (wp+%u) blocks failed %d",
85 zone->id, (unsigned long long)wp_block,
86 (unsigned long long)block, nr_blocks, ret);
87 dmz_check_bdev(dmz_dev: dev);
88 return ret;
89 }
90
91 zone->wp_block = block;
92
93 return 0;
94}
95
96/*
97 * dm_kcopyd_copy end notification.
98 */
99static void dmz_reclaim_kcopy_end(int read_err, unsigned long write_err,
100 void *context)
101{
102 struct dmz_reclaim *zrc = context;
103
104 if (read_err || write_err)
105 zrc->kc_err = -EIO;
106 else
107 zrc->kc_err = 0;
108
109 clear_bit_unlock(nr: DMZ_RECLAIM_KCOPY, addr: &zrc->flags);
110 smp_mb__after_atomic();
111 wake_up_bit(word: &zrc->flags, bit: DMZ_RECLAIM_KCOPY);
112}
113
114/*
115 * Copy valid blocks of src_zone into dst_zone.
116 */
117static int dmz_reclaim_copy(struct dmz_reclaim *zrc,
118 struct dm_zone *src_zone, struct dm_zone *dst_zone)
119{
120 struct dmz_metadata *zmd = zrc->metadata;
121 struct dm_io_region src, dst;
122 sector_t block = 0, end_block;
123 sector_t nr_blocks;
124 sector_t src_zone_block;
125 sector_t dst_zone_block;
126 unsigned long flags = 0;
127 int ret;
128
129 if (dmz_is_seq(src_zone))
130 end_block = src_zone->wp_block;
131 else
132 end_block = dmz_zone_nr_blocks(zmd);
133 src_zone_block = dmz_start_block(zmd, zone: src_zone);
134 dst_zone_block = dmz_start_block(zmd, zone: dst_zone);
135
136 if (dmz_is_seq(dst_zone))
137 flags |= BIT(DM_KCOPYD_WRITE_SEQ);
138
139 while (block < end_block) {
140 if (src_zone->dev->flags & DMZ_BDEV_DYING)
141 return -EIO;
142 if (dst_zone->dev->flags & DMZ_BDEV_DYING)
143 return -EIO;
144
145 if (dmz_reclaim_should_terminate(src_zone))
146 return -EINTR;
147
148 /* Get a valid region from the source zone */
149 ret = dmz_first_valid_block(zmd, zone: src_zone, chunk_block: &block);
150 if (ret <= 0)
151 return ret;
152 nr_blocks = ret;
153
154 /*
155 * If we are writing in a sequential zone, we must make sure
156 * that writes are sequential. So Zeroout any eventual hole
157 * between writes.
158 */
159 if (dmz_is_seq(dst_zone)) {
160 ret = dmz_reclaim_align_wp(zrc, zone: dst_zone, block);
161 if (ret)
162 return ret;
163 }
164
165 src.bdev = src_zone->dev->bdev;
166 src.sector = dmz_blk2sect(src_zone_block + block);
167 src.count = dmz_blk2sect(nr_blocks);
168
169 dst.bdev = dst_zone->dev->bdev;
170 dst.sector = dmz_blk2sect(dst_zone_block + block);
171 dst.count = src.count;
172
173 /* Copy the valid region */
174 set_bit(nr: DMZ_RECLAIM_KCOPY, addr: &zrc->flags);
175 dm_kcopyd_copy(kc: zrc->kc, from: &src, num_dests: 1, dests: &dst, flags,
176 fn: dmz_reclaim_kcopy_end, context: zrc);
177
178 /* Wait for copy to complete */
179 wait_on_bit_io(word: &zrc->flags, bit: DMZ_RECLAIM_KCOPY,
180 TASK_UNINTERRUPTIBLE);
181 if (zrc->kc_err)
182 return zrc->kc_err;
183
184 block += nr_blocks;
185 if (dmz_is_seq(dst_zone))
186 dst_zone->wp_block = block;
187 }
188
189 return 0;
190}
191
192/*
193 * Move valid blocks of dzone buffer zone into dzone (after its write pointer)
194 * and free the buffer zone.
195 */
196static int dmz_reclaim_buf(struct dmz_reclaim *zrc, struct dm_zone *dzone)
197{
198 struct dm_zone *bzone = dzone->bzone;
199 sector_t chunk_block = dzone->wp_block;
200 struct dmz_metadata *zmd = zrc->metadata;
201 int ret;
202
203 DMDEBUG("(%s/%u): Chunk %u, move buf zone %u (weight %u) to data zone %u (weight %u)",
204 dmz_metadata_label(zmd), zrc->dev_idx,
205 dzone->chunk, bzone->id, dmz_weight(bzone),
206 dzone->id, dmz_weight(dzone));
207
208 /* Flush data zone into the buffer zone */
209 ret = dmz_reclaim_copy(zrc, src_zone: bzone, dst_zone: dzone);
210 if (ret < 0)
211 return ret;
212
213 dmz_lock_flush(zmd);
214
215 /* Validate copied blocks */
216 ret = dmz_merge_valid_blocks(zmd, from_zone: bzone, to_zone: dzone, chunk_block);
217 if (ret == 0) {
218 /* Free the buffer zone */
219 dmz_invalidate_blocks(zmd, zone: bzone, chunk_block: 0, nr_blocks: dmz_zone_nr_blocks(zmd));
220 dmz_lock_map(zmd);
221 dmz_unmap_zone(zmd, zone: bzone);
222 dmz_unlock_zone_reclaim(zone: dzone);
223 dmz_free_zone(zmd, zone: bzone);
224 dmz_unlock_map(zmd);
225 }
226
227 dmz_unlock_flush(zmd);
228
229 return ret;
230}
231
232/*
233 * Merge valid blocks of dzone into its buffer zone and free dzone.
234 */
235static int dmz_reclaim_seq_data(struct dmz_reclaim *zrc, struct dm_zone *dzone)
236{
237 unsigned int chunk = dzone->chunk;
238 struct dm_zone *bzone = dzone->bzone;
239 struct dmz_metadata *zmd = zrc->metadata;
240 int ret = 0;
241
242 DMDEBUG("(%s/%u): Chunk %u, move data zone %u (weight %u) to buf zone %u (weight %u)",
243 dmz_metadata_label(zmd), zrc->dev_idx,
244 chunk, dzone->id, dmz_weight(dzone),
245 bzone->id, dmz_weight(bzone));
246
247 /* Flush data zone into the buffer zone */
248 ret = dmz_reclaim_copy(zrc, src_zone: dzone, dst_zone: bzone);
249 if (ret < 0)
250 return ret;
251
252 dmz_lock_flush(zmd);
253
254 /* Validate copied blocks */
255 ret = dmz_merge_valid_blocks(zmd, from_zone: dzone, to_zone: bzone, chunk_block: 0);
256 if (ret == 0) {
257 /*
258 * Free the data zone and remap the chunk to
259 * the buffer zone.
260 */
261 dmz_invalidate_blocks(zmd, zone: dzone, chunk_block: 0, nr_blocks: dmz_zone_nr_blocks(zmd));
262 dmz_lock_map(zmd);
263 dmz_unmap_zone(zmd, zone: bzone);
264 dmz_unmap_zone(zmd, zone: dzone);
265 dmz_unlock_zone_reclaim(zone: dzone);
266 dmz_free_zone(zmd, zone: dzone);
267 dmz_map_zone(zmd, zone: bzone, chunk);
268 dmz_unlock_map(zmd);
269 }
270
271 dmz_unlock_flush(zmd);
272
273 return ret;
274}
275
276/*
277 * Move valid blocks of the random data zone dzone into a free sequential zone.
278 * Once blocks are moved, remap the zone chunk to the sequential zone.
279 */
280static int dmz_reclaim_rnd_data(struct dmz_reclaim *zrc, struct dm_zone *dzone)
281{
282 unsigned int chunk = dzone->chunk;
283 struct dm_zone *szone = NULL;
284 struct dmz_metadata *zmd = zrc->metadata;
285 int ret;
286 int alloc_flags = DMZ_ALLOC_SEQ;
287
288 /* Get a free random or sequential zone */
289 dmz_lock_map(zmd);
290again:
291 szone = dmz_alloc_zone(zmd, dev_idx: zrc->dev_idx,
292 flags: alloc_flags | DMZ_ALLOC_RECLAIM);
293 if (!szone && alloc_flags == DMZ_ALLOC_SEQ && dmz_nr_cache_zones(zmd)) {
294 alloc_flags = DMZ_ALLOC_RND;
295 goto again;
296 }
297 dmz_unlock_map(zmd);
298 if (!szone)
299 return -ENOSPC;
300
301 DMDEBUG("(%s/%u): Chunk %u, move %s zone %u (weight %u) to %s zone %u",
302 dmz_metadata_label(zmd), zrc->dev_idx, chunk,
303 dmz_is_cache(dzone) ? "cache" : "rnd",
304 dzone->id, dmz_weight(dzone),
305 dmz_is_rnd(szone) ? "rnd" : "seq", szone->id);
306
307 /* Flush the random data zone into the sequential zone */
308 ret = dmz_reclaim_copy(zrc, src_zone: dzone, dst_zone: szone);
309
310 dmz_lock_flush(zmd);
311
312 if (ret == 0) {
313 /* Validate copied blocks */
314 ret = dmz_copy_valid_blocks(zmd, from_zone: dzone, to_zone: szone);
315 }
316 if (ret) {
317 /* Free the sequential zone */
318 dmz_lock_map(zmd);
319 dmz_free_zone(zmd, zone: szone);
320 dmz_unlock_map(zmd);
321 } else {
322 /* Free the data zone and remap the chunk */
323 dmz_invalidate_blocks(zmd, zone: dzone, chunk_block: 0, nr_blocks: dmz_zone_nr_blocks(zmd));
324 dmz_lock_map(zmd);
325 dmz_unmap_zone(zmd, zone: dzone);
326 dmz_unlock_zone_reclaim(zone: dzone);
327 dmz_free_zone(zmd, zone: dzone);
328 dmz_map_zone(zmd, zone: szone, chunk);
329 dmz_unlock_map(zmd);
330 }
331
332 dmz_unlock_flush(zmd);
333
334 return ret;
335}
336
337/*
338 * Reclaim an empty zone.
339 */
340static void dmz_reclaim_empty(struct dmz_reclaim *zrc, struct dm_zone *dzone)
341{
342 struct dmz_metadata *zmd = zrc->metadata;
343
344 dmz_lock_flush(zmd);
345 dmz_lock_map(zmd);
346 dmz_unmap_zone(zmd, zone: dzone);
347 dmz_unlock_zone_reclaim(zone: dzone);
348 dmz_free_zone(zmd, zone: dzone);
349 dmz_unlock_map(zmd);
350 dmz_unlock_flush(zmd);
351}
352
353/*
354 * Test if the target device is idle.
355 */
356static inline int dmz_target_idle(struct dmz_reclaim *zrc)
357{
358 return time_is_before_jiffies(zrc->atime + DMZ_IDLE_PERIOD);
359}
360
361/*
362 * Find a candidate zone for reclaim and process it.
363 */
364static int dmz_do_reclaim(struct dmz_reclaim *zrc)
365{
366 struct dmz_metadata *zmd = zrc->metadata;
367 struct dm_zone *dzone;
368 struct dm_zone *rzone;
369 unsigned long start;
370 int ret;
371
372 /* Get a data zone */
373 dzone = dmz_get_zone_for_reclaim(zmd, dev_idx: zrc->dev_idx,
374 idle: dmz_target_idle(zrc));
375 if (!dzone) {
376 DMDEBUG("(%s/%u): No zone found to reclaim",
377 dmz_metadata_label(zmd), zrc->dev_idx);
378 return -EBUSY;
379 }
380 rzone = dzone;
381
382 start = jiffies;
383 if (dmz_is_cache(dzone) || dmz_is_rnd(dzone)) {
384 if (!dmz_weight(dzone)) {
385 /* Empty zone */
386 dmz_reclaim_empty(zrc, dzone);
387 ret = 0;
388 } else {
389 /*
390 * Reclaim the random data zone by moving its
391 * valid data blocks to a free sequential zone.
392 */
393 ret = dmz_reclaim_rnd_data(zrc, dzone);
394 }
395 } else {
396 struct dm_zone *bzone = dzone->bzone;
397 sector_t chunk_block = 0;
398
399 ret = dmz_first_valid_block(zmd, zone: bzone, chunk_block: &chunk_block);
400 if (ret < 0)
401 goto out;
402
403 if (ret == 0 || chunk_block >= dzone->wp_block) {
404 /*
405 * The buffer zone is empty or its valid blocks are
406 * after the data zone write pointer.
407 */
408 ret = dmz_reclaim_buf(zrc, dzone);
409 rzone = bzone;
410 } else {
411 /*
412 * Reclaim the data zone by merging it into the
413 * buffer zone so that the buffer zone itself can
414 * be later reclaimed.
415 */
416 ret = dmz_reclaim_seq_data(zrc, dzone);
417 }
418 }
419out:
420 if (ret) {
421 if (ret == -EINTR)
422 DMDEBUG("(%s/%u): reclaim zone %u interrupted",
423 dmz_metadata_label(zmd), zrc->dev_idx,
424 rzone->id);
425 else
426 DMDEBUG("(%s/%u): Failed to reclaim zone %u, err %d",
427 dmz_metadata_label(zmd), zrc->dev_idx,
428 rzone->id, ret);
429 dmz_unlock_zone_reclaim(zone: dzone);
430 return ret;
431 }
432
433 ret = dmz_flush_metadata(zmd: zrc->metadata);
434 if (ret) {
435 DMDEBUG("(%s/%u): Metadata flush for zone %u failed, err %d",
436 dmz_metadata_label(zmd), zrc->dev_idx, rzone->id, ret);
437 return ret;
438 }
439
440 DMDEBUG("(%s/%u): Reclaimed zone %u in %u ms",
441 dmz_metadata_label(zmd), zrc->dev_idx,
442 rzone->id, jiffies_to_msecs(jiffies - start));
443 return 0;
444}
445
446static unsigned int dmz_reclaim_percentage(struct dmz_reclaim *zrc)
447{
448 struct dmz_metadata *zmd = zrc->metadata;
449 unsigned int nr_cache = dmz_nr_cache_zones(zmd);
450 unsigned int nr_unmap, nr_zones;
451
452 if (nr_cache) {
453 nr_zones = nr_cache;
454 nr_unmap = dmz_nr_unmap_cache_zones(zmd);
455 } else {
456 nr_zones = dmz_nr_rnd_zones(zmd, idx: zrc->dev_idx);
457 nr_unmap = dmz_nr_unmap_rnd_zones(zmd, idx: zrc->dev_idx);
458 }
459 if (nr_unmap <= 1)
460 return 0;
461 return nr_unmap * 100 / nr_zones;
462}
463
464/*
465 * Test if reclaim is necessary.
466 */
467static bool dmz_should_reclaim(struct dmz_reclaim *zrc, unsigned int p_unmap)
468{
469 unsigned int nr_reclaim;
470
471 nr_reclaim = dmz_nr_rnd_zones(zmd: zrc->metadata, idx: zrc->dev_idx);
472
473 if (dmz_nr_cache_zones(zmd: zrc->metadata)) {
474 /*
475 * The first device in a multi-device
476 * setup only contains cache zones, so
477 * never start reclaim there.
478 */
479 if (zrc->dev_idx == 0)
480 return false;
481 nr_reclaim += dmz_nr_cache_zones(zmd: zrc->metadata);
482 }
483
484 /* Reclaim when idle */
485 if (dmz_target_idle(zrc) && nr_reclaim)
486 return true;
487
488 /* If there are still plenty of cache zones, do not reclaim */
489 if (p_unmap >= DMZ_RECLAIM_HIGH_UNMAP_ZONES)
490 return false;
491
492 /*
493 * If the percentage of unmapped cache zones is low,
494 * reclaim even if the target is busy.
495 */
496 return p_unmap <= DMZ_RECLAIM_LOW_UNMAP_ZONES;
497}
498
499/*
500 * Reclaim work function.
501 */
502static void dmz_reclaim_work(struct work_struct *work)
503{
504 struct dmz_reclaim *zrc = container_of(work, struct dmz_reclaim, work.work);
505 struct dmz_metadata *zmd = zrc->metadata;
506 unsigned int p_unmap;
507 int ret;
508
509 if (dmz_dev_is_dying(zmd))
510 return;
511
512 p_unmap = dmz_reclaim_percentage(zrc);
513 if (!dmz_should_reclaim(zrc, p_unmap)) {
514 mod_delayed_work(wq: zrc->wq, dwork: &zrc->work, DMZ_IDLE_PERIOD);
515 return;
516 }
517
518 /*
519 * We need to start reclaiming random zones: set up zone copy
520 * throttling to either go fast if we are very low on random zones
521 * and slower if there are still some free random zones to avoid
522 * as much as possible to negatively impact the user workload.
523 */
524 if (dmz_target_idle(zrc) || p_unmap < DMZ_RECLAIM_LOW_UNMAP_ZONES / 2) {
525 /* Idle or very low percentage: go fast */
526 zrc->kc_throttle.throttle = 100;
527 } else {
528 /* Busy but we still have some random zone: throttle */
529 zrc->kc_throttle.throttle = min(75U, 100U - p_unmap / 2);
530 }
531
532 DMDEBUG("(%s/%u): Reclaim (%u): %s, %u%% free zones (%u/%u cache %u/%u random)",
533 dmz_metadata_label(zmd), zrc->dev_idx,
534 zrc->kc_throttle.throttle,
535 (dmz_target_idle(zrc) ? "Idle" : "Busy"),
536 p_unmap, dmz_nr_unmap_cache_zones(zmd),
537 dmz_nr_cache_zones(zmd),
538 dmz_nr_unmap_rnd_zones(zmd, zrc->dev_idx),
539 dmz_nr_rnd_zones(zmd, zrc->dev_idx));
540
541 ret = dmz_do_reclaim(zrc);
542 if (ret && ret != -EINTR) {
543 if (!dmz_check_dev(zmd))
544 return;
545 }
546
547 dmz_schedule_reclaim(zrc);
548}
549
550/*
551 * Initialize reclaim.
552 */
553int dmz_ctr_reclaim(struct dmz_metadata *zmd,
554 struct dmz_reclaim **reclaim, int idx)
555{
556 struct dmz_reclaim *zrc;
557 int ret;
558
559 zrc = kzalloc(size: sizeof(struct dmz_reclaim), GFP_KERNEL);
560 if (!zrc)
561 return -ENOMEM;
562
563 zrc->metadata = zmd;
564 zrc->atime = jiffies;
565 zrc->dev_idx = idx;
566
567 /* Reclaim kcopyd client */
568 zrc->kc = dm_kcopyd_client_create(throttle: &zrc->kc_throttle);
569 if (IS_ERR(ptr: zrc->kc)) {
570 ret = PTR_ERR(ptr: zrc->kc);
571 zrc->kc = NULL;
572 goto err;
573 }
574
575 /* Reclaim work */
576 INIT_DELAYED_WORK(&zrc->work, dmz_reclaim_work);
577 zrc->wq = alloc_ordered_workqueue("dmz_rwq_%s_%d", WQ_MEM_RECLAIM,
578 dmz_metadata_label(zmd), idx);
579 if (!zrc->wq) {
580 ret = -ENOMEM;
581 goto err;
582 }
583
584 *reclaim = zrc;
585 queue_delayed_work(wq: zrc->wq, dwork: &zrc->work, delay: 0);
586
587 return 0;
588err:
589 if (zrc->kc)
590 dm_kcopyd_client_destroy(kc: zrc->kc);
591 kfree(objp: zrc);
592
593 return ret;
594}
595
596/*
597 * Terminate reclaim.
598 */
599void dmz_dtr_reclaim(struct dmz_reclaim *zrc)
600{
601 cancel_delayed_work_sync(dwork: &zrc->work);
602 destroy_workqueue(wq: zrc->wq);
603 dm_kcopyd_client_destroy(kc: zrc->kc);
604 kfree(objp: zrc);
605}
606
607/*
608 * Suspend reclaim.
609 */
610void dmz_suspend_reclaim(struct dmz_reclaim *zrc)
611{
612 cancel_delayed_work_sync(dwork: &zrc->work);
613}
614
615/*
616 * Resume reclaim.
617 */
618void dmz_resume_reclaim(struct dmz_reclaim *zrc)
619{
620 queue_delayed_work(wq: zrc->wq, dwork: &zrc->work, DMZ_IDLE_PERIOD);
621}
622
623/*
624 * BIO accounting.
625 */
626void dmz_reclaim_bio_acc(struct dmz_reclaim *zrc)
627{
628 zrc->atime = jiffies;
629}
630
631/*
632 * Start reclaim if necessary.
633 */
634void dmz_schedule_reclaim(struct dmz_reclaim *zrc)
635{
636 unsigned int p_unmap = dmz_reclaim_percentage(zrc);
637
638 if (dmz_should_reclaim(zrc, p_unmap))
639 mod_delayed_work(wq: zrc->wq, dwork: &zrc->work, delay: 0);
640}
641

source code of linux/drivers/md/dm-zoned-reclaim.c