1 // SPDX-License-Identifier: GPL-2.0-only
3 * Copyright (C) 2017 Western Digital Corporation or its affiliates.
5 * This file is released under the GPL.
10 #include <linux/module.h>
12 #define DM_MSG_PREFIX "zoned reclaim"
15 struct dmz_metadata *metadata;
18 struct delayed_work work;
19 struct workqueue_struct *wq;
21 struct dm_kcopyd_client *kc;
22 struct dm_kcopyd_throttle kc_throttle;
27 /* Last target access time */
32 * Reclaim state flags.
39 * Number of seconds of target BIO inactivity to consider the target idle.
41 #define DMZ_IDLE_PERIOD (10UL * HZ)
44 * Percentage of unmapped (free) random zones below which reclaim starts
45 * even if the target is busy.
47 #define DMZ_RECLAIM_LOW_UNMAP_RND 30
50 * Percentage of unmapped (free) random zones above which reclaim will
51 * stop if the target is busy.
53 #define DMZ_RECLAIM_HIGH_UNMAP_RND 50
56 * Align a sequential zone write pointer to chunk_block.
58 static int dmz_reclaim_align_wp(struct dmz_reclaim *zrc, struct dm_zone *zone,
61 struct dmz_metadata *zmd = zrc->metadata;
62 sector_t wp_block = zone->wp_block;
63 unsigned int nr_blocks;
66 if (wp_block == block)
73 * Zeroout the space between the write
74 * pointer and the requested position.
76 nr_blocks = block - wp_block;
77 ret = blkdev_issue_zeroout(zrc->dev->bdev,
78 dmz_start_sect(zmd, zone) + dmz_blk2sect(wp_block),
79 dmz_blk2sect(nr_blocks), GFP_NOIO, 0);
82 "Align zone %u wp %llu to %llu (wp+%u) blocks failed %d",
83 dmz_id(zmd, zone), (unsigned long long)wp_block,
84 (unsigned long long)block, nr_blocks, ret);
88 zone->wp_block = block;
94 * dm_kcopyd_copy end notification.
96 static void dmz_reclaim_kcopy_end(int read_err, unsigned long write_err,
99 struct dmz_reclaim *zrc = context;
101 if (read_err || write_err)
106 clear_bit_unlock(DMZ_RECLAIM_KCOPY, &zrc->flags);
107 smp_mb__after_atomic();
108 wake_up_bit(&zrc->flags, DMZ_RECLAIM_KCOPY);
112 * Copy valid blocks of src_zone into dst_zone.
114 static int dmz_reclaim_copy(struct dmz_reclaim *zrc,
115 struct dm_zone *src_zone, struct dm_zone *dst_zone)
117 struct dmz_metadata *zmd = zrc->metadata;
118 struct dmz_dev *dev = zrc->dev;
119 struct dm_io_region src, dst;
120 sector_t block = 0, end_block;
122 sector_t src_zone_block;
123 sector_t dst_zone_block;
124 unsigned long flags = 0;
127 if (dmz_is_seq(src_zone))
128 end_block = src_zone->wp_block;
130 end_block = dev->zone_nr_blocks;
131 src_zone_block = dmz_start_block(zmd, src_zone);
132 dst_zone_block = dmz_start_block(zmd, dst_zone);
134 if (dmz_is_seq(dst_zone))
135 set_bit(DM_KCOPYD_WRITE_SEQ, &flags);
137 while (block < end_block) {
138 if (dev->flags & DMZ_BDEV_DYING)
141 /* Get a valid region from the source zone */
142 ret = dmz_first_valid_block(zmd, src_zone, &block);
148 * If we are writing in a sequential zone, we must make sure
149 * that writes are sequential. So Zeroout any eventual hole
152 if (dmz_is_seq(dst_zone)) {
153 ret = dmz_reclaim_align_wp(zrc, dst_zone, block);
158 src.bdev = dev->bdev;
159 src.sector = dmz_blk2sect(src_zone_block + block);
160 src.count = dmz_blk2sect(nr_blocks);
162 dst.bdev = dev->bdev;
163 dst.sector = dmz_blk2sect(dst_zone_block + block);
164 dst.count = src.count;
166 /* Copy the valid region */
167 set_bit(DMZ_RECLAIM_KCOPY, &zrc->flags);
168 dm_kcopyd_copy(zrc->kc, &src, 1, &dst, flags,
169 dmz_reclaim_kcopy_end, zrc);
171 /* Wait for copy to complete */
172 wait_on_bit_io(&zrc->flags, DMZ_RECLAIM_KCOPY,
173 TASK_UNINTERRUPTIBLE);
178 if (dmz_is_seq(dst_zone))
179 dst_zone->wp_block = block;
186 * Move valid blocks of dzone buffer zone into dzone (after its write pointer)
187 * and free the buffer zone.
189 static int dmz_reclaim_buf(struct dmz_reclaim *zrc, struct dm_zone *dzone)
191 struct dm_zone *bzone = dzone->bzone;
192 sector_t chunk_block = dzone->wp_block;
193 struct dmz_metadata *zmd = zrc->metadata;
196 dmz_dev_debug(zrc->dev,
197 "Chunk %u, move buf zone %u (weight %u) to data zone %u (weight %u)",
198 dzone->chunk, dmz_id(zmd, bzone), dmz_weight(bzone),
199 dmz_id(zmd, dzone), dmz_weight(dzone));
201 /* Flush data zone into the buffer zone */
202 ret = dmz_reclaim_copy(zrc, bzone, dzone);
208 /* Validate copied blocks */
209 ret = dmz_merge_valid_blocks(zmd, bzone, dzone, chunk_block);
211 /* Free the buffer zone */
212 dmz_invalidate_blocks(zmd, bzone, 0, zrc->dev->zone_nr_blocks);
214 dmz_unmap_zone(zmd, bzone);
215 dmz_unlock_zone_reclaim(dzone);
216 dmz_free_zone(zmd, bzone);
220 dmz_unlock_flush(zmd);
226 * Merge valid blocks of dzone into its buffer zone and free dzone.
228 static int dmz_reclaim_seq_data(struct dmz_reclaim *zrc, struct dm_zone *dzone)
230 unsigned int chunk = dzone->chunk;
231 struct dm_zone *bzone = dzone->bzone;
232 struct dmz_metadata *zmd = zrc->metadata;
235 dmz_dev_debug(zrc->dev,
236 "Chunk %u, move data zone %u (weight %u) to buf zone %u (weight %u)",
237 chunk, dmz_id(zmd, dzone), dmz_weight(dzone),
238 dmz_id(zmd, bzone), dmz_weight(bzone));
240 /* Flush data zone into the buffer zone */
241 ret = dmz_reclaim_copy(zrc, dzone, bzone);
247 /* Validate copied blocks */
248 ret = dmz_merge_valid_blocks(zmd, dzone, bzone, 0);
251 * Free the data zone and remap the chunk to
254 dmz_invalidate_blocks(zmd, dzone, 0, zrc->dev->zone_nr_blocks);
256 dmz_unmap_zone(zmd, bzone);
257 dmz_unmap_zone(zmd, dzone);
258 dmz_unlock_zone_reclaim(dzone);
259 dmz_free_zone(zmd, dzone);
260 dmz_map_zone(zmd, bzone, chunk);
264 dmz_unlock_flush(zmd);
270 * Move valid blocks of the random data zone dzone into a free sequential zone.
271 * Once blocks are moved, remap the zone chunk to the sequential zone.
273 static int dmz_reclaim_rnd_data(struct dmz_reclaim *zrc, struct dm_zone *dzone)
275 unsigned int chunk = dzone->chunk;
276 struct dm_zone *szone = NULL;
277 struct dmz_metadata *zmd = zrc->metadata;
280 /* Get a free sequential zone */
282 szone = dmz_alloc_zone(zmd, DMZ_ALLOC_RECLAIM);
287 dmz_dev_debug(zrc->dev,
288 "Chunk %u, move rnd zone %u (weight %u) to seq zone %u",
289 chunk, dmz_id(zmd, dzone), dmz_weight(dzone),
292 /* Flush the random data zone into the sequential zone */
293 ret = dmz_reclaim_copy(zrc, dzone, szone);
298 /* Validate copied blocks */
299 ret = dmz_copy_valid_blocks(zmd, dzone, szone);
302 /* Free the sequential zone */
304 dmz_free_zone(zmd, szone);
307 /* Free the data zone and remap the chunk */
308 dmz_invalidate_blocks(zmd, dzone, 0, zrc->dev->zone_nr_blocks);
310 dmz_unmap_zone(zmd, dzone);
311 dmz_unlock_zone_reclaim(dzone);
312 dmz_free_zone(zmd, dzone);
313 dmz_map_zone(zmd, szone, chunk);
317 dmz_unlock_flush(zmd);
323 * Reclaim an empty zone.
325 static void dmz_reclaim_empty(struct dmz_reclaim *zrc, struct dm_zone *dzone)
327 struct dmz_metadata *zmd = zrc->metadata;
331 dmz_unmap_zone(zmd, dzone);
332 dmz_unlock_zone_reclaim(dzone);
333 dmz_free_zone(zmd, dzone);
335 dmz_unlock_flush(zmd);
339 * Find a candidate zone for reclaim and process it.
341 static int dmz_do_reclaim(struct dmz_reclaim *zrc)
343 struct dmz_metadata *zmd = zrc->metadata;
344 struct dm_zone *dzone;
345 struct dm_zone *rzone;
349 /* Get a data zone */
350 dzone = dmz_get_zone_for_reclaim(zmd);
352 return PTR_ERR(dzone);
356 if (dmz_is_rnd(dzone)) {
357 if (!dmz_weight(dzone)) {
359 dmz_reclaim_empty(zrc, dzone);
363 * Reclaim the random data zone by moving its
364 * valid data blocks to a free sequential zone.
366 ret = dmz_reclaim_rnd_data(zrc, dzone);
371 struct dm_zone *bzone = dzone->bzone;
372 sector_t chunk_block = 0;
374 ret = dmz_first_valid_block(zmd, bzone, &chunk_block);
378 if (ret == 0 || chunk_block >= dzone->wp_block) {
380 * The buffer zone is empty or its valid blocks are
381 * after the data zone write pointer.
383 ret = dmz_reclaim_buf(zrc, dzone);
387 * Reclaim the data zone by merging it into the
388 * buffer zone so that the buffer zone itself can
389 * be later reclaimed.
391 ret = dmz_reclaim_seq_data(zrc, dzone);
397 dmz_unlock_zone_reclaim(dzone);
401 ret = dmz_flush_metadata(zrc->metadata);
403 dmz_dev_debug(zrc->dev,
404 "Metadata flush for zone %u failed, err %d\n",
405 dmz_id(zmd, rzone), ret);
409 dmz_dev_debug(zrc->dev, "Reclaimed zone %u in %u ms",
410 dmz_id(zmd, rzone), jiffies_to_msecs(jiffies - start));
415 * Test if the target device is idle.
417 static inline int dmz_target_idle(struct dmz_reclaim *zrc)
419 return time_is_before_jiffies(zrc->atime + DMZ_IDLE_PERIOD);
423 * Test if reclaim is necessary.
425 static bool dmz_should_reclaim(struct dmz_reclaim *zrc)
427 struct dmz_metadata *zmd = zrc->metadata;
428 unsigned int nr_rnd = dmz_nr_rnd_zones(zmd);
429 unsigned int nr_unmap_rnd = dmz_nr_unmap_rnd_zones(zmd);
430 unsigned int p_unmap_rnd = nr_unmap_rnd * 100 / nr_rnd;
432 /* Reclaim when idle */
433 if (dmz_target_idle(zrc) && nr_unmap_rnd < nr_rnd)
436 /* If there are still plenty of random zones, do not reclaim */
437 if (p_unmap_rnd >= DMZ_RECLAIM_HIGH_UNMAP_RND)
441 * If the percentage of unmapped random zones is low,
442 * reclaim even if the target is busy.
444 return p_unmap_rnd <= DMZ_RECLAIM_LOW_UNMAP_RND;
448 * Reclaim work function.
450 static void dmz_reclaim_work(struct work_struct *work)
452 struct dmz_reclaim *zrc = container_of(work, struct dmz_reclaim, work.work);
453 struct dmz_metadata *zmd = zrc->metadata;
454 unsigned int nr_rnd, nr_unmap_rnd;
455 unsigned int p_unmap_rnd;
458 if (dmz_bdev_is_dying(zrc->dev))
461 if (!dmz_should_reclaim(zrc)) {
462 mod_delayed_work(zrc->wq, &zrc->work, DMZ_IDLE_PERIOD);
467 * We need to start reclaiming random zones: set up zone copy
468 * throttling to either go fast if we are very low on random zones
469 * and slower if there are still some free random zones to avoid
470 * as much as possible to negatively impact the user workload.
472 nr_rnd = dmz_nr_rnd_zones(zmd);
473 nr_unmap_rnd = dmz_nr_unmap_rnd_zones(zmd);
474 p_unmap_rnd = nr_unmap_rnd * 100 / nr_rnd;
475 if (dmz_target_idle(zrc) || p_unmap_rnd < DMZ_RECLAIM_LOW_UNMAP_RND / 2) {
476 /* Idle or very low percentage: go fast */
477 zrc->kc_throttle.throttle = 100;
479 /* Busy but we still have some random zone: throttle */
480 zrc->kc_throttle.throttle = min(75U, 100U - p_unmap_rnd / 2);
483 dmz_dev_debug(zrc->dev,
484 "Reclaim (%u): %s, %u%% free rnd zones (%u/%u)",
485 zrc->kc_throttle.throttle,
486 (dmz_target_idle(zrc) ? "Idle" : "Busy"),
487 p_unmap_rnd, nr_unmap_rnd, nr_rnd);
489 ret = dmz_do_reclaim(zrc);
491 dmz_dev_debug(zrc->dev, "Reclaim error %d\n", ret);
494 * LLD might be performing some error handling sequence
495 * at the underlying device. To not interfere, do not
496 * attempt to schedule the next reclaim run immediately.
501 dmz_schedule_reclaim(zrc);
505 * Initialize reclaim.
507 int dmz_ctr_reclaim(struct dmz_dev *dev, struct dmz_metadata *zmd,
508 struct dmz_reclaim **reclaim)
510 struct dmz_reclaim *zrc;
513 zrc = kzalloc(sizeof(struct dmz_reclaim), GFP_KERNEL);
519 zrc->atime = jiffies;
521 /* Reclaim kcopyd client */
522 zrc->kc = dm_kcopyd_client_create(&zrc->kc_throttle);
523 if (IS_ERR(zrc->kc)) {
524 ret = PTR_ERR(zrc->kc);
530 INIT_DELAYED_WORK(&zrc->work, dmz_reclaim_work);
531 zrc->wq = alloc_ordered_workqueue("dmz_rwq_%s", WQ_MEM_RECLAIM,
539 queue_delayed_work(zrc->wq, &zrc->work, 0);
544 dm_kcopyd_client_destroy(zrc->kc);
553 void dmz_dtr_reclaim(struct dmz_reclaim *zrc)
555 cancel_delayed_work_sync(&zrc->work);
556 destroy_workqueue(zrc->wq);
557 dm_kcopyd_client_destroy(zrc->kc);
564 void dmz_suspend_reclaim(struct dmz_reclaim *zrc)
566 cancel_delayed_work_sync(&zrc->work);
572 void dmz_resume_reclaim(struct dmz_reclaim *zrc)
574 queue_delayed_work(zrc->wq, &zrc->work, DMZ_IDLE_PERIOD);
580 void dmz_reclaim_bio_acc(struct dmz_reclaim *zrc)
582 zrc->atime = jiffies;
586 * Start reclaim if necessary.
588 void dmz_schedule_reclaim(struct dmz_reclaim *zrc)
590 if (dmz_should_reclaim(zrc))
591 mod_delayed_work(zrc->wq, &zrc->work, 0);