cec9a56dda146712ef06a917afaf1c326f67385e
[platform/kernel/linux-starfive.git] / drivers / lightnvm / pblk-gc.c
1 /*
2  * Copyright (C) 2016 CNEX Labs
3  * Initial release: Javier Gonzalez <javier@cnexlabs.com>
4  *                  Matias Bjorling <matias@cnexlabs.com>
5  *
6  * This program is free software; you can redistribute it and/or
7  * modify it under the terms of the GNU General Public License version
8  * 2 as published by the Free Software Foundation.
9  *
10  * This program is distributed in the hope that it will be useful, but
11  * WITHOUT ANY WARRANTY; without even the implied warranty of
12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
13  * General Public License for more details.
14  *
15  * pblk-gc.c - pblk's garbage collector
16  */
17
18 #include "pblk.h"
19 #include <linux/delay.h>
20
21 static void pblk_gc_free_gc_rq(struct pblk_gc_rq *gc_rq)
22 {
23         if (gc_rq->data)
24                 vfree(gc_rq->data);
25         kfree(gc_rq);
26 }
27
28 static int pblk_gc_write(struct pblk *pblk)
29 {
30         struct pblk_gc *gc = &pblk->gc;
31         struct pblk_gc_rq *gc_rq, *tgc_rq;
32         LIST_HEAD(w_list);
33
34         spin_lock(&gc->w_lock);
35         if (list_empty(&gc->w_list)) {
36                 spin_unlock(&gc->w_lock);
37                 return 1;
38         }
39
40         list_cut_position(&w_list, &gc->w_list, gc->w_list.prev);
41         gc->w_entries = 0;
42         spin_unlock(&gc->w_lock);
43
44         list_for_each_entry_safe(gc_rq, tgc_rq, &w_list, list) {
45                 pblk_write_gc_to_cache(pblk, gc_rq);
46                 list_del(&gc_rq->list);
47                 kref_put(&gc_rq->line->ref, pblk_line_put);
48                 pblk_gc_free_gc_rq(gc_rq);
49         }
50
51         return 0;
52 }
53
54 static void pblk_gc_writer_kick(struct pblk_gc *gc)
55 {
56         wake_up_process(gc->gc_writer_ts);
57 }
58
59 static void pblk_put_line_back(struct pblk *pblk, struct pblk_line *line)
60 {
61         struct pblk_line_mgmt *l_mg = &pblk->l_mg;
62         struct list_head *move_list;
63
64         spin_lock(&line->lock);
65         WARN_ON(line->state != PBLK_LINESTATE_GC);
66         line->state = PBLK_LINESTATE_CLOSED;
67         move_list = pblk_line_gc_list(pblk, line);
68         spin_unlock(&line->lock);
69
70         if (move_list) {
71                 spin_lock(&l_mg->gc_lock);
72                 list_add_tail(&line->list, move_list);
73                 spin_unlock(&l_mg->gc_lock);
74         }
75 }
76
77 static void pblk_gc_line_ws(struct work_struct *work)
78 {
79         struct pblk_line_ws *gc_rq_ws = container_of(work,
80                                                 struct pblk_line_ws, ws);
81         struct pblk *pblk = gc_rq_ws->pblk;
82         struct nvm_tgt_dev *dev = pblk->dev;
83         struct nvm_geo *geo = &dev->geo;
84         struct pblk_gc *gc = &pblk->gc;
85         struct pblk_line *line = gc_rq_ws->line;
86         struct pblk_gc_rq *gc_rq = gc_rq_ws->priv;
87         int ret;
88
89         up(&gc->gc_sem);
90
91         gc_rq->data = vmalloc(gc_rq->nr_secs * geo->sec_size);
92         if (!gc_rq->data) {
93                 pr_err("pblk: could not GC line:%d (%d/%d)\n",
94                                         line->id, *line->vsc, gc_rq->nr_secs);
95                 goto out;
96         }
97
98         /* Read from GC victim block */
99         ret = pblk_submit_read_gc(pblk, gc_rq);
100         if (ret) {
101                 pr_err("pblk: failed GC read in line:%d (err:%d)\n",
102                                                                 line->id, ret);
103                 goto out;
104         }
105
106         if (!gc_rq->secs_to_gc)
107                 goto out;
108
109 retry:
110         spin_lock(&gc->w_lock);
111         if (gc->w_entries >= PBLK_GC_RQ_QD) {
112                 spin_unlock(&gc->w_lock);
113                 pblk_gc_writer_kick(&pblk->gc);
114                 usleep_range(128, 256);
115                 goto retry;
116         }
117         gc->w_entries++;
118         list_add_tail(&gc_rq->list, &gc->w_list);
119         spin_unlock(&gc->w_lock);
120
121         pblk_gc_writer_kick(&pblk->gc);
122
123         kfree(gc_rq_ws);
124         return;
125
126 out:
127         pblk_gc_free_gc_rq(gc_rq);
128         kref_put(&line->ref, pblk_line_put);
129         kfree(gc_rq_ws);
130 }
131
132 static void pblk_gc_line_prepare_ws(struct work_struct *work)
133 {
134         struct pblk_line_ws *line_ws = container_of(work, struct pblk_line_ws,
135                                                                         ws);
136         struct pblk *pblk = line_ws->pblk;
137         struct pblk_line *line = line_ws->line;
138         struct pblk_line_mgmt *l_mg = &pblk->l_mg;
139         struct pblk_line_meta *lm = &pblk->lm;
140         struct pblk_gc *gc = &pblk->gc;
141         struct line_emeta *emeta_buf;
142         struct pblk_line_ws *gc_rq_ws;
143         struct pblk_gc_rq *gc_rq;
144         __le64 *lba_list;
145         unsigned long *invalid_bitmap;
146         int sec_left, nr_secs, bit;
147         int ret;
148
149         invalid_bitmap = kmalloc(lm->sec_bitmap_len, GFP_KERNEL);
150         if (!invalid_bitmap) {
151                 pr_err("pblk: could not allocate GC invalid bitmap\n");
152                 goto fail_free_ws;
153         }
154
155         emeta_buf = pblk_malloc(lm->emeta_len[0], l_mg->emeta_alloc_type,
156                                                                 GFP_KERNEL);
157         if (!emeta_buf) {
158                 pr_err("pblk: cannot use GC emeta\n");
159                 goto fail_free_bitmap;
160         }
161
162         ret = pblk_line_read_emeta(pblk, line, emeta_buf);
163         if (ret) {
164                 pr_err("pblk: line %d read emeta failed (%d)\n", line->id, ret);
165                 goto fail_free_emeta;
166         }
167
168         /* If this read fails, it means that emeta is corrupted. For now, leave
169          * the line untouched. TODO: Implement a recovery routine that scans and
170          * moves all sectors on the line.
171          */
172
173         ret = pblk_recov_check_emeta(pblk, emeta_buf);
174         if (ret) {
175                 pr_err("pblk: inconsistent emeta (line %d)\n", line->id);
176                 goto fail_free_emeta;
177         }
178
179         lba_list = emeta_to_lbas(pblk, emeta_buf);
180         if (!lba_list) {
181                 pr_err("pblk: could not interpret emeta (line %d)\n", line->id);
182                 goto fail_free_emeta;
183         }
184
185         spin_lock(&line->lock);
186         bitmap_copy(invalid_bitmap, line->invalid_bitmap, lm->sec_per_line);
187         sec_left = pblk_line_vsc(line);
188         spin_unlock(&line->lock);
189
190         if (sec_left < 0) {
191                 pr_err("pblk: corrupted GC line (%d)\n", line->id);
192                 goto fail_free_emeta;
193         }
194
195         bit = -1;
196 next_rq:
197         gc_rq = kmalloc(sizeof(struct pblk_gc_rq), GFP_KERNEL);
198         if (!gc_rq)
199                 goto fail_free_emeta;
200
201         nr_secs = 0;
202         do {
203                 bit = find_next_zero_bit(invalid_bitmap, lm->sec_per_line,
204                                                                 bit + 1);
205                 if (bit > line->emeta_ssec)
206                         break;
207
208                 gc_rq->paddr_list[nr_secs] = bit;
209                 gc_rq->lba_list[nr_secs++] = le64_to_cpu(lba_list[bit]);
210         } while (nr_secs < pblk->max_write_pgs);
211
212         if (unlikely(!nr_secs)) {
213                 kfree(gc_rq);
214                 goto out;
215         }
216
217         gc_rq->nr_secs = nr_secs;
218         gc_rq->line = line;
219
220         gc_rq_ws = kmalloc(sizeof(struct pblk_line_ws), GFP_KERNEL);
221         if (!gc_rq_ws)
222                 goto fail_free_gc_rq;
223
224         gc_rq_ws->pblk = pblk;
225         gc_rq_ws->line = line;
226         gc_rq_ws->priv = gc_rq;
227
228         /* The write GC path can be much slower than the read GC one due to
229          * the budget imposed by the rate-limiter. Balance in case that we get
230          * back pressure from the write GC path.
231          */
232         while (down_timeout(&gc->gc_sem, msecs_to_jiffies(30000)))
233                 io_schedule();
234
235         kref_get(&line->ref);
236
237         INIT_WORK(&gc_rq_ws->ws, pblk_gc_line_ws);
238         queue_work(gc->gc_line_reader_wq, &gc_rq_ws->ws);
239
240         sec_left -= nr_secs;
241         if (sec_left > 0)
242                 goto next_rq;
243
244 out:
245         pblk_mfree(emeta_buf, l_mg->emeta_alloc_type);
246         kfree(line_ws);
247         kfree(invalid_bitmap);
248
249         kref_put(&line->ref, pblk_line_put);
250         atomic_dec(&gc->read_inflight_gc);
251
252         return;
253
254 fail_free_gc_rq:
255         kfree(gc_rq);
256 fail_free_emeta:
257         pblk_mfree(emeta_buf, l_mg->emeta_alloc_type);
258 fail_free_bitmap:
259         kfree(invalid_bitmap);
260 fail_free_ws:
261         kfree(line_ws);
262
263         pblk_put_line_back(pblk, line);
264         kref_put(&line->ref, pblk_line_put);
265         atomic_dec(&gc->read_inflight_gc);
266
267         pr_err("pblk: Failed to GC line %d\n", line->id);
268 }
269
270 static int pblk_gc_line(struct pblk *pblk, struct pblk_line *line)
271 {
272         struct pblk_gc *gc = &pblk->gc;
273         struct pblk_line_ws *line_ws;
274
275         pr_debug("pblk: line '%d' being reclaimed for GC\n", line->id);
276
277         line_ws = kmalloc(sizeof(struct pblk_line_ws), GFP_KERNEL);
278         if (!line_ws)
279                 return -ENOMEM;
280
281         line_ws->pblk = pblk;
282         line_ws->line = line;
283
284         atomic_inc(&gc->pipeline_gc);
285         INIT_WORK(&line_ws->ws, pblk_gc_line_prepare_ws);
286         queue_work(gc->gc_reader_wq, &line_ws->ws);
287
288         return 0;
289 }
290
291 static void pblk_gc_reader_kick(struct pblk_gc *gc)
292 {
293         wake_up_process(gc->gc_reader_ts);
294 }
295
296 static void pblk_gc_kick(struct pblk *pblk)
297 {
298         struct pblk_gc *gc = &pblk->gc;
299
300         pblk_gc_writer_kick(gc);
301         pblk_gc_reader_kick(gc);
302
303         /* If we're shutting down GC, let's not start it up again */
304         if (gc->gc_enabled) {
305                 wake_up_process(gc->gc_ts);
306                 mod_timer(&gc->gc_timer,
307                           jiffies + msecs_to_jiffies(GC_TIME_MSECS));
308         }
309 }
310
311 static int pblk_gc_read(struct pblk *pblk)
312 {
313         struct pblk_gc *gc = &pblk->gc;
314         struct pblk_line *line;
315
316         spin_lock(&gc->r_lock);
317         if (list_empty(&gc->r_list)) {
318                 spin_unlock(&gc->r_lock);
319                 return 1;
320         }
321
322         line = list_first_entry(&gc->r_list, struct pblk_line, list);
323         list_del(&line->list);
324         spin_unlock(&gc->r_lock);
325
326         pblk_gc_kick(pblk);
327
328         if (pblk_gc_line(pblk, line))
329                 pr_err("pblk: failed to GC line %d\n", line->id);
330
331         return 0;
332 }
333
334 static struct pblk_line *pblk_gc_get_victim_line(struct pblk *pblk,
335                                                  struct list_head *group_list)
336 {
337         struct pblk_line *line, *victim;
338         int line_vsc, victim_vsc;
339
340         victim = list_first_entry(group_list, struct pblk_line, list);
341         list_for_each_entry(line, group_list, list) {
342                 line_vsc = le32_to_cpu(*line->vsc);
343                 victim_vsc = le32_to_cpu(*victim->vsc);
344                 if (line_vsc < victim_vsc)
345                         victim = line;
346         }
347
348         return victim;
349 }
350
351 static bool pblk_gc_should_run(struct pblk_gc *gc, struct pblk_rl *rl)
352 {
353         unsigned int nr_blocks_free, nr_blocks_need;
354
355         nr_blocks_need = pblk_rl_high_thrs(rl);
356         nr_blocks_free = pblk_rl_nr_free_blks(rl);
357
358         /* This is not critical, no need to take lock here */
359         return ((gc->gc_active) && (nr_blocks_need > nr_blocks_free));
360 }
361
362 void pblk_gc_free_full_lines(struct pblk *pblk)
363 {
364         struct pblk_line_mgmt *l_mg = &pblk->l_mg;
365         struct pblk_gc *gc = &pblk->gc;
366         struct pblk_line *line;
367
368         do {
369                 spin_lock(&l_mg->gc_lock);
370                 if (list_empty(&l_mg->gc_full_list)) {
371                         spin_unlock(&l_mg->gc_lock);
372                         return;
373                 }
374
375                 line = list_first_entry(&l_mg->gc_full_list,
376                                                         struct pblk_line, list);
377
378                 spin_lock(&line->lock);
379                 WARN_ON(line->state != PBLK_LINESTATE_CLOSED);
380                 line->state = PBLK_LINESTATE_GC;
381                 spin_unlock(&line->lock);
382
383                 list_del(&line->list);
384                 spin_unlock(&l_mg->gc_lock);
385
386                 atomic_inc(&gc->pipeline_gc);
387                 kref_put(&line->ref, pblk_line_put);
388         } while (1);
389 }
390
391 /*
392  * Lines with no valid sectors will be returned to the free list immediately. If
393  * GC is activated - either because the free block count is under the determined
394  * threshold, or because it is being forced from user space - only lines with a
395  * high count of invalid sectors will be recycled.
396  */
397 static void pblk_gc_run(struct pblk *pblk)
398 {
399         struct pblk_line_mgmt *l_mg = &pblk->l_mg;
400         struct pblk_gc *gc = &pblk->gc;
401         struct pblk_line *line;
402         struct list_head *group_list;
403         bool run_gc;
404         int read_inflight_gc, gc_group = 0, prev_group = 0;
405
406         pblk_gc_free_full_lines(pblk);
407
408         run_gc = pblk_gc_should_run(&pblk->gc, &pblk->rl);
409         if (!run_gc || (atomic_read(&gc->read_inflight_gc) >= PBLK_GC_L_QD))
410                 return;
411
412 next_gc_group:
413         group_list = l_mg->gc_lists[gc_group++];
414
415         do {
416                 spin_lock(&l_mg->gc_lock);
417                 if (list_empty(group_list)) {
418                         spin_unlock(&l_mg->gc_lock);
419                         break;
420                 }
421
422                 line = pblk_gc_get_victim_line(pblk, group_list);
423
424                 spin_lock(&line->lock);
425                 WARN_ON(line->state != PBLK_LINESTATE_CLOSED);
426                 line->state = PBLK_LINESTATE_GC;
427                 spin_unlock(&line->lock);
428
429                 list_del(&line->list);
430                 spin_unlock(&l_mg->gc_lock);
431
432                 spin_lock(&gc->r_lock);
433                 list_add_tail(&line->list, &gc->r_list);
434                 spin_unlock(&gc->r_lock);
435
436                 read_inflight_gc = atomic_inc_return(&gc->read_inflight_gc);
437                 pblk_gc_reader_kick(gc);
438
439                 prev_group = 1;
440
441                 /* No need to queue up more GC lines than we can handle */
442                 run_gc = pblk_gc_should_run(&pblk->gc, &pblk->rl);
443                 if (!run_gc || read_inflight_gc >= PBLK_GC_L_QD)
444                         break;
445         } while (1);
446
447         if (!prev_group && pblk->rl.rb_state > gc_group &&
448                                                 gc_group < PBLK_GC_NR_LISTS)
449                 goto next_gc_group;
450 }
451
452 static void pblk_gc_timer(struct timer_list *t)
453 {
454         struct pblk *pblk = from_timer(pblk, t, gc.gc_timer);
455
456         pblk_gc_kick(pblk);
457 }
458
459 static int pblk_gc_ts(void *data)
460 {
461         struct pblk *pblk = data;
462
463         while (!kthread_should_stop()) {
464                 pblk_gc_run(pblk);
465                 set_current_state(TASK_INTERRUPTIBLE);
466                 io_schedule();
467         }
468
469         return 0;
470 }
471
472 static int pblk_gc_writer_ts(void *data)
473 {
474         struct pblk *pblk = data;
475
476         while (!kthread_should_stop()) {
477                 if (!pblk_gc_write(pblk))
478                         continue;
479                 set_current_state(TASK_INTERRUPTIBLE);
480                 io_schedule();
481         }
482
483         return 0;
484 }
485
486 static int pblk_gc_reader_ts(void *data)
487 {
488         struct pblk *pblk = data;
489         struct pblk_gc *gc = &pblk->gc;
490
491         while (!kthread_should_stop()) {
492                 if (!pblk_gc_read(pblk))
493                         continue;
494                 set_current_state(TASK_INTERRUPTIBLE);
495                 io_schedule();
496         }
497
498 #ifdef CONFIG_NVM_DEBUG
499         pr_info("pblk: flushing gc pipeline, %d lines left\n",
500                 atomic_read(&gc->pipeline_gc));
501 #endif
502
503         do {
504                 if (!atomic_read(&gc->pipeline_gc))
505                         break;
506
507                 schedule();
508         } while (1);
509
510         return 0;
511 }
512
513 static void pblk_gc_start(struct pblk *pblk)
514 {
515         pblk->gc.gc_active = 1;
516         pr_debug("pblk: gc start\n");
517 }
518
519 void pblk_gc_should_start(struct pblk *pblk)
520 {
521         struct pblk_gc *gc = &pblk->gc;
522
523         if (gc->gc_enabled && !gc->gc_active) {
524                 pblk_gc_start(pblk);
525                 pblk_gc_kick(pblk);
526         }
527 }
528
529 /*
530  * If flush_wq == 1 then no lock should be held by the caller since
531  * flush_workqueue can sleep
532  */
533 static void pblk_gc_stop(struct pblk *pblk, int flush_wq)
534 {
535         pblk->gc.gc_active = 0;
536         pr_debug("pblk: gc stop\n");
537 }
538
539 void pblk_gc_should_stop(struct pblk *pblk)
540 {
541         struct pblk_gc *gc = &pblk->gc;
542
543         if (gc->gc_active && !gc->gc_forced)
544                 pblk_gc_stop(pblk, 0);
545 }
546
547 void pblk_gc_should_kick(struct pblk *pblk)
548 {
549         pblk_rl_update_rates(&pblk->rl);
550 }
551
552 void pblk_gc_sysfs_state_show(struct pblk *pblk, int *gc_enabled,
553                               int *gc_active)
554 {
555         struct pblk_gc *gc = &pblk->gc;
556
557         spin_lock(&gc->lock);
558         *gc_enabled = gc->gc_enabled;
559         *gc_active = gc->gc_active;
560         spin_unlock(&gc->lock);
561 }
562
563 int pblk_gc_sysfs_force(struct pblk *pblk, int force)
564 {
565         struct pblk_gc *gc = &pblk->gc;
566
567         if (force < 0 || force > 1)
568                 return -EINVAL;
569
570         spin_lock(&gc->lock);
571         gc->gc_forced = force;
572
573         if (force)
574                 gc->gc_enabled = 1;
575         else
576                 gc->gc_enabled = 0;
577         spin_unlock(&gc->lock);
578
579         pblk_gc_should_start(pblk);
580
581         return 0;
582 }
583
584 int pblk_gc_init(struct pblk *pblk)
585 {
586         struct pblk_gc *gc = &pblk->gc;
587         int ret;
588
589         gc->gc_ts = kthread_create(pblk_gc_ts, pblk, "pblk-gc-ts");
590         if (IS_ERR(gc->gc_ts)) {
591                 pr_err("pblk: could not allocate GC main kthread\n");
592                 return PTR_ERR(gc->gc_ts);
593         }
594
595         gc->gc_writer_ts = kthread_create(pblk_gc_writer_ts, pblk,
596                                                         "pblk-gc-writer-ts");
597         if (IS_ERR(gc->gc_writer_ts)) {
598                 pr_err("pblk: could not allocate GC writer kthread\n");
599                 ret = PTR_ERR(gc->gc_writer_ts);
600                 goto fail_free_main_kthread;
601         }
602
603         gc->gc_reader_ts = kthread_create(pblk_gc_reader_ts, pblk,
604                                                         "pblk-gc-reader-ts");
605         if (IS_ERR(gc->gc_reader_ts)) {
606                 pr_err("pblk: could not allocate GC reader kthread\n");
607                 ret = PTR_ERR(gc->gc_reader_ts);
608                 goto fail_free_writer_kthread;
609         }
610
611         timer_setup(&gc->gc_timer, pblk_gc_timer, 0);
612         mod_timer(&gc->gc_timer, jiffies + msecs_to_jiffies(GC_TIME_MSECS));
613
614         gc->gc_active = 0;
615         gc->gc_forced = 0;
616         gc->gc_enabled = 1;
617         gc->w_entries = 0;
618         atomic_set(&gc->read_inflight_gc, 0);
619         atomic_set(&gc->pipeline_gc, 0);
620
621         /* Workqueue that reads valid sectors from a line and submit them to the
622          * GC writer to be recycled.
623          */
624         gc->gc_line_reader_wq = alloc_workqueue("pblk-gc-line-reader-wq",
625                         WQ_MEM_RECLAIM | WQ_UNBOUND, PBLK_GC_MAX_READERS);
626         if (!gc->gc_line_reader_wq) {
627                 pr_err("pblk: could not allocate GC line reader workqueue\n");
628                 ret = -ENOMEM;
629                 goto fail_free_reader_kthread;
630         }
631
632         /* Workqueue that prepare lines for GC */
633         gc->gc_reader_wq = alloc_workqueue("pblk-gc-line_wq",
634                                         WQ_MEM_RECLAIM | WQ_UNBOUND, 1);
635         if (!gc->gc_reader_wq) {
636                 pr_err("pblk: could not allocate GC reader workqueue\n");
637                 ret = -ENOMEM;
638                 goto fail_free_reader_line_wq;
639         }
640
641         spin_lock_init(&gc->lock);
642         spin_lock_init(&gc->w_lock);
643         spin_lock_init(&gc->r_lock);
644
645         sema_init(&gc->gc_sem, PBLK_GC_RQ_QD);
646
647         INIT_LIST_HEAD(&gc->w_list);
648         INIT_LIST_HEAD(&gc->r_list);
649
650         return 0;
651
652 fail_free_reader_line_wq:
653         destroy_workqueue(gc->gc_line_reader_wq);
654 fail_free_reader_kthread:
655         kthread_stop(gc->gc_reader_ts);
656 fail_free_writer_kthread:
657         kthread_stop(gc->gc_writer_ts);
658 fail_free_main_kthread:
659         kthread_stop(gc->gc_ts);
660
661         return ret;
662 }
663
664 void pblk_gc_exit(struct pblk *pblk)
665 {
666         struct pblk_gc *gc = &pblk->gc;
667
668         gc->gc_enabled = 0;
669         del_timer_sync(&gc->gc_timer);
670         pblk_gc_stop(pblk, 1);
671
672         if (gc->gc_ts)
673                 kthread_stop(gc->gc_ts);
674
675         if (gc->gc_reader_ts)
676                 kthread_stop(gc->gc_reader_ts);
677
678         flush_workqueue(gc->gc_reader_wq);
679         if (gc->gc_reader_wq)
680                 destroy_workqueue(gc->gc_reader_wq);
681
682         flush_workqueue(gc->gc_line_reader_wq);
683         if (gc->gc_line_reader_wq)
684                 destroy_workqueue(gc->gc_line_reader_wq);
685
686         if (gc->gc_writer_ts)
687                 kthread_stop(gc->gc_writer_ts);
688 }