2 * Copyright (C) 2016 CNEX Labs
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.
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.
15 * pblk-gc.c - pblk's garbage collector
19 #include <linux/delay.h>
21 static void pblk_gc_free_gc_rq(struct pblk_gc_rq *gc_rq)
28 static int pblk_gc_write(struct pblk *pblk)
30 struct pblk_gc *gc = &pblk->gc;
31 struct pblk_gc_rq *gc_rq, *tgc_rq;
34 spin_lock(&gc->w_lock);
35 if (list_empty(&gc->w_list)) {
36 spin_unlock(&gc->w_lock);
40 list_cut_position(&w_list, &gc->w_list, gc->w_list.prev);
42 spin_unlock(&gc->w_lock);
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);
54 static void pblk_gc_writer_kick(struct pblk_gc *gc)
56 wake_up_process(gc->gc_writer_ts);
59 static void pblk_put_line_back(struct pblk *pblk, struct pblk_line *line)
61 struct pblk_line_mgmt *l_mg = &pblk->l_mg;
62 struct list_head *move_list;
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);
71 spin_lock(&l_mg->gc_lock);
72 list_add_tail(&line->list, move_list);
73 spin_unlock(&l_mg->gc_lock);
77 static void pblk_gc_line_ws(struct work_struct *work)
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;
91 gc_rq->data = vmalloc(gc_rq->nr_secs * geo->csecs);
93 pr_err("pblk: could not GC line:%d (%d/%d)\n",
94 line->id, *line->vsc, gc_rq->nr_secs);
98 /* Read from GC victim block */
99 ret = pblk_submit_read_gc(pblk, gc_rq);
101 pr_err("pblk: failed GC read in line:%d (err:%d)\n",
106 if (!gc_rq->secs_to_gc)
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);
118 list_add_tail(&gc_rq->list, &gc->w_list);
119 spin_unlock(&gc->w_lock);
121 pblk_gc_writer_kick(&pblk->gc);
127 pblk_gc_free_gc_rq(gc_rq);
128 kref_put(&line->ref, pblk_line_put);
132 static void pblk_gc_line_prepare_ws(struct work_struct *work)
134 struct pblk_line_ws *line_ws = container_of(work, struct pblk_line_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;
145 unsigned long *invalid_bitmap;
146 int sec_left, nr_secs, bit;
149 invalid_bitmap = kmalloc(lm->sec_bitmap_len, GFP_KERNEL);
153 emeta_buf = pblk_malloc(lm->emeta_len[0], l_mg->emeta_alloc_type,
156 pr_err("pblk: cannot use GC emeta\n");
157 goto fail_free_bitmap;
160 ret = pblk_line_read_emeta(pblk, line, emeta_buf);
162 pr_err("pblk: line %d read emeta failed (%d)\n", line->id, ret);
163 goto fail_free_emeta;
166 /* If this read fails, it means that emeta is corrupted. For now, leave
167 * the line untouched. TODO: Implement a recovery routine that scans and
168 * moves all sectors on the line.
171 ret = pblk_recov_check_emeta(pblk, emeta_buf);
173 pr_err("pblk: inconsistent emeta (line %d)\n", line->id);
174 goto fail_free_emeta;
177 lba_list = emeta_to_lbas(pblk, emeta_buf);
179 pr_err("pblk: could not interpret emeta (line %d)\n", line->id);
180 goto fail_free_emeta;
183 spin_lock(&line->lock);
184 bitmap_copy(invalid_bitmap, line->invalid_bitmap, lm->sec_per_line);
185 sec_left = pblk_line_vsc(line);
186 spin_unlock(&line->lock);
189 pr_err("pblk: corrupted GC line (%d)\n", line->id);
190 goto fail_free_emeta;
195 gc_rq = kmalloc(sizeof(struct pblk_gc_rq), GFP_KERNEL);
197 goto fail_free_emeta;
201 bit = find_next_zero_bit(invalid_bitmap, lm->sec_per_line,
203 if (bit > line->emeta_ssec)
206 gc_rq->paddr_list[nr_secs] = bit;
207 gc_rq->lba_list[nr_secs++] = le64_to_cpu(lba_list[bit]);
208 } while (nr_secs < pblk->max_write_pgs);
210 if (unlikely(!nr_secs)) {
215 gc_rq->nr_secs = nr_secs;
218 gc_rq_ws = kmalloc(sizeof(struct pblk_line_ws), GFP_KERNEL);
220 goto fail_free_gc_rq;
222 gc_rq_ws->pblk = pblk;
223 gc_rq_ws->line = line;
224 gc_rq_ws->priv = gc_rq;
226 /* The write GC path can be much slower than the read GC one due to
227 * the budget imposed by the rate-limiter. Balance in case that we get
228 * back pressure from the write GC path.
230 while (down_timeout(&gc->gc_sem, msecs_to_jiffies(30000)))
233 kref_get(&line->ref);
235 INIT_WORK(&gc_rq_ws->ws, pblk_gc_line_ws);
236 queue_work(gc->gc_line_reader_wq, &gc_rq_ws->ws);
243 pblk_mfree(emeta_buf, l_mg->emeta_alloc_type);
245 kfree(invalid_bitmap);
247 kref_put(&line->ref, pblk_line_put);
248 atomic_dec(&gc->read_inflight_gc);
255 pblk_mfree(emeta_buf, l_mg->emeta_alloc_type);
257 kfree(invalid_bitmap);
261 pblk_put_line_back(pblk, line);
262 kref_put(&line->ref, pblk_line_put);
263 atomic_dec(&gc->read_inflight_gc);
265 pr_err("pblk: Failed to GC line %d\n", line->id);
268 static int pblk_gc_line(struct pblk *pblk, struct pblk_line *line)
270 struct pblk_gc *gc = &pblk->gc;
271 struct pblk_line_ws *line_ws;
273 pr_debug("pblk: line '%d' being reclaimed for GC\n", line->id);
275 line_ws = kmalloc(sizeof(struct pblk_line_ws), GFP_KERNEL);
279 line_ws->pblk = pblk;
280 line_ws->line = line;
282 atomic_inc(&gc->pipeline_gc);
283 INIT_WORK(&line_ws->ws, pblk_gc_line_prepare_ws);
284 queue_work(gc->gc_reader_wq, &line_ws->ws);
289 static void pblk_gc_reader_kick(struct pblk_gc *gc)
291 wake_up_process(gc->gc_reader_ts);
294 static void pblk_gc_kick(struct pblk *pblk)
296 struct pblk_gc *gc = &pblk->gc;
298 pblk_gc_writer_kick(gc);
299 pblk_gc_reader_kick(gc);
301 /* If we're shutting down GC, let's not start it up again */
302 if (gc->gc_enabled) {
303 wake_up_process(gc->gc_ts);
304 mod_timer(&gc->gc_timer,
305 jiffies + msecs_to_jiffies(GC_TIME_MSECS));
309 static int pblk_gc_read(struct pblk *pblk)
311 struct pblk_gc *gc = &pblk->gc;
312 struct pblk_line *line;
314 spin_lock(&gc->r_lock);
315 if (list_empty(&gc->r_list)) {
316 spin_unlock(&gc->r_lock);
320 line = list_first_entry(&gc->r_list, struct pblk_line, list);
321 list_del(&line->list);
322 spin_unlock(&gc->r_lock);
326 if (pblk_gc_line(pblk, line))
327 pr_err("pblk: failed to GC line %d\n", line->id);
332 static struct pblk_line *pblk_gc_get_victim_line(struct pblk *pblk,
333 struct list_head *group_list)
335 struct pblk_line *line, *victim;
336 int line_vsc, victim_vsc;
338 victim = list_first_entry(group_list, struct pblk_line, list);
339 list_for_each_entry(line, group_list, list) {
340 line_vsc = le32_to_cpu(*line->vsc);
341 victim_vsc = le32_to_cpu(*victim->vsc);
342 if (line_vsc < victim_vsc)
349 static bool pblk_gc_should_run(struct pblk_gc *gc, struct pblk_rl *rl)
351 unsigned int nr_blocks_free, nr_blocks_need;
353 nr_blocks_need = pblk_rl_high_thrs(rl);
354 nr_blocks_free = pblk_rl_nr_free_blks(rl);
356 /* This is not critical, no need to take lock here */
357 return ((gc->gc_active) && (nr_blocks_need > nr_blocks_free));
360 void pblk_gc_free_full_lines(struct pblk *pblk)
362 struct pblk_line_mgmt *l_mg = &pblk->l_mg;
363 struct pblk_gc *gc = &pblk->gc;
364 struct pblk_line *line;
367 spin_lock(&l_mg->gc_lock);
368 if (list_empty(&l_mg->gc_full_list)) {
369 spin_unlock(&l_mg->gc_lock);
373 line = list_first_entry(&l_mg->gc_full_list,
374 struct pblk_line, list);
376 spin_lock(&line->lock);
377 WARN_ON(line->state != PBLK_LINESTATE_CLOSED);
378 line->state = PBLK_LINESTATE_GC;
379 spin_unlock(&line->lock);
381 list_del(&line->list);
382 spin_unlock(&l_mg->gc_lock);
384 atomic_inc(&gc->pipeline_gc);
385 kref_put(&line->ref, pblk_line_put);
390 * Lines with no valid sectors will be returned to the free list immediately. If
391 * GC is activated - either because the free block count is under the determined
392 * threshold, or because it is being forced from user space - only lines with a
393 * high count of invalid sectors will be recycled.
395 static void pblk_gc_run(struct pblk *pblk)
397 struct pblk_line_mgmt *l_mg = &pblk->l_mg;
398 struct pblk_gc *gc = &pblk->gc;
399 struct pblk_line *line;
400 struct list_head *group_list;
402 int read_inflight_gc, gc_group = 0, prev_group = 0;
404 pblk_gc_free_full_lines(pblk);
406 run_gc = pblk_gc_should_run(&pblk->gc, &pblk->rl);
407 if (!run_gc || (atomic_read(&gc->read_inflight_gc) >= PBLK_GC_L_QD))
411 group_list = l_mg->gc_lists[gc_group++];
414 spin_lock(&l_mg->gc_lock);
415 if (list_empty(group_list)) {
416 spin_unlock(&l_mg->gc_lock);
420 line = pblk_gc_get_victim_line(pblk, group_list);
422 spin_lock(&line->lock);
423 WARN_ON(line->state != PBLK_LINESTATE_CLOSED);
424 line->state = PBLK_LINESTATE_GC;
425 spin_unlock(&line->lock);
427 list_del(&line->list);
428 spin_unlock(&l_mg->gc_lock);
430 spin_lock(&gc->r_lock);
431 list_add_tail(&line->list, &gc->r_list);
432 spin_unlock(&gc->r_lock);
434 read_inflight_gc = atomic_inc_return(&gc->read_inflight_gc);
435 pblk_gc_reader_kick(gc);
439 /* No need to queue up more GC lines than we can handle */
440 run_gc = pblk_gc_should_run(&pblk->gc, &pblk->rl);
441 if (!run_gc || read_inflight_gc >= PBLK_GC_L_QD)
445 if (!prev_group && pblk->rl.rb_state > gc_group &&
446 gc_group < PBLK_GC_NR_LISTS)
450 static void pblk_gc_timer(struct timer_list *t)
452 struct pblk *pblk = from_timer(pblk, t, gc.gc_timer);
457 static int pblk_gc_ts(void *data)
459 struct pblk *pblk = data;
461 while (!kthread_should_stop()) {
463 set_current_state(TASK_INTERRUPTIBLE);
470 static int pblk_gc_writer_ts(void *data)
472 struct pblk *pblk = data;
474 while (!kthread_should_stop()) {
475 if (!pblk_gc_write(pblk))
477 set_current_state(TASK_INTERRUPTIBLE);
484 static int pblk_gc_reader_ts(void *data)
486 struct pblk *pblk = data;
487 struct pblk_gc *gc = &pblk->gc;
489 while (!kthread_should_stop()) {
490 if (!pblk_gc_read(pblk))
492 set_current_state(TASK_INTERRUPTIBLE);
496 #ifdef CONFIG_NVM_DEBUG
497 pr_info("pblk: flushing gc pipeline, %d lines left\n",
498 atomic_read(&gc->pipeline_gc));
502 if (!atomic_read(&gc->pipeline_gc))
511 static void pblk_gc_start(struct pblk *pblk)
513 pblk->gc.gc_active = 1;
514 pr_debug("pblk: gc start\n");
517 void pblk_gc_should_start(struct pblk *pblk)
519 struct pblk_gc *gc = &pblk->gc;
521 if (gc->gc_enabled && !gc->gc_active) {
527 void pblk_gc_should_stop(struct pblk *pblk)
529 struct pblk_gc *gc = &pblk->gc;
531 if (gc->gc_active && !gc->gc_forced)
535 void pblk_gc_should_kick(struct pblk *pblk)
537 pblk_rl_update_rates(&pblk->rl);
540 void pblk_gc_sysfs_state_show(struct pblk *pblk, int *gc_enabled,
543 struct pblk_gc *gc = &pblk->gc;
545 spin_lock(&gc->lock);
546 *gc_enabled = gc->gc_enabled;
547 *gc_active = gc->gc_active;
548 spin_unlock(&gc->lock);
551 int pblk_gc_sysfs_force(struct pblk *pblk, int force)
553 struct pblk_gc *gc = &pblk->gc;
555 if (force < 0 || force > 1)
558 spin_lock(&gc->lock);
559 gc->gc_forced = force;
565 spin_unlock(&gc->lock);
567 pblk_gc_should_start(pblk);
572 int pblk_gc_init(struct pblk *pblk)
574 struct pblk_gc *gc = &pblk->gc;
577 gc->gc_ts = kthread_create(pblk_gc_ts, pblk, "pblk-gc-ts");
578 if (IS_ERR(gc->gc_ts)) {
579 pr_err("pblk: could not allocate GC main kthread\n");
580 return PTR_ERR(gc->gc_ts);
583 gc->gc_writer_ts = kthread_create(pblk_gc_writer_ts, pblk,
584 "pblk-gc-writer-ts");
585 if (IS_ERR(gc->gc_writer_ts)) {
586 pr_err("pblk: could not allocate GC writer kthread\n");
587 ret = PTR_ERR(gc->gc_writer_ts);
588 goto fail_free_main_kthread;
591 gc->gc_reader_ts = kthread_create(pblk_gc_reader_ts, pblk,
592 "pblk-gc-reader-ts");
593 if (IS_ERR(gc->gc_reader_ts)) {
594 pr_err("pblk: could not allocate GC reader kthread\n");
595 ret = PTR_ERR(gc->gc_reader_ts);
596 goto fail_free_writer_kthread;
599 timer_setup(&gc->gc_timer, pblk_gc_timer, 0);
600 mod_timer(&gc->gc_timer, jiffies + msecs_to_jiffies(GC_TIME_MSECS));
606 atomic_set(&gc->read_inflight_gc, 0);
607 atomic_set(&gc->pipeline_gc, 0);
609 /* Workqueue that reads valid sectors from a line and submit them to the
610 * GC writer to be recycled.
612 gc->gc_line_reader_wq = alloc_workqueue("pblk-gc-line-reader-wq",
613 WQ_MEM_RECLAIM | WQ_UNBOUND, PBLK_GC_MAX_READERS);
614 if (!gc->gc_line_reader_wq) {
615 pr_err("pblk: could not allocate GC line reader workqueue\n");
617 goto fail_free_reader_kthread;
620 /* Workqueue that prepare lines for GC */
621 gc->gc_reader_wq = alloc_workqueue("pblk-gc-line_wq",
622 WQ_MEM_RECLAIM | WQ_UNBOUND, 1);
623 if (!gc->gc_reader_wq) {
624 pr_err("pblk: could not allocate GC reader workqueue\n");
626 goto fail_free_reader_line_wq;
629 spin_lock_init(&gc->lock);
630 spin_lock_init(&gc->w_lock);
631 spin_lock_init(&gc->r_lock);
633 sema_init(&gc->gc_sem, PBLK_GC_RQ_QD);
635 INIT_LIST_HEAD(&gc->w_list);
636 INIT_LIST_HEAD(&gc->r_list);
640 fail_free_reader_line_wq:
641 destroy_workqueue(gc->gc_line_reader_wq);
642 fail_free_reader_kthread:
643 kthread_stop(gc->gc_reader_ts);
644 fail_free_writer_kthread:
645 kthread_stop(gc->gc_writer_ts);
646 fail_free_main_kthread:
647 kthread_stop(gc->gc_ts);
652 void pblk_gc_exit(struct pblk *pblk)
654 struct pblk_gc *gc = &pblk->gc;
657 del_timer_sync(&gc->gc_timer);
661 kthread_stop(gc->gc_ts);
663 if (gc->gc_reader_ts)
664 kthread_stop(gc->gc_reader_ts);
666 flush_workqueue(gc->gc_reader_wq);
667 destroy_workqueue(gc->gc_reader_wq);
669 flush_workqueue(gc->gc_line_reader_wq);
670 destroy_workqueue(gc->gc_line_reader_wq);
672 if (gc->gc_writer_ts)
673 kthread_stop(gc->gc_writer_ts);