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