2 * Copyright (C) 2016 CNEX Labs
3 * Initial release: Javier Gonzalez <javier@cnexlabs.com>
4 * Matias Bjorling <matias@cnexlabs.com>
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->sec_size);
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);
150 if (!invalid_bitmap) {
151 pr_err("pblk: could not allocate GC invalid bitmap\n");
155 emeta_buf = pblk_malloc(lm->emeta_len[0], l_mg->emeta_alloc_type,
158 pr_err("pblk: cannot use GC emeta\n");
159 goto fail_free_bitmap;
162 ret = pblk_line_read_emeta(pblk, line, emeta_buf);
164 pr_err("pblk: line %d read emeta failed (%d)\n", line->id, ret);
165 goto fail_free_emeta;
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.
173 ret = pblk_recov_check_emeta(pblk, emeta_buf);
175 pr_err("pblk: inconsistent emeta (line %d)\n", line->id);
176 goto fail_free_emeta;
179 lba_list = emeta_to_lbas(pblk, emeta_buf);
181 pr_err("pblk: could not interpret emeta (line %d)\n", line->id);
182 goto fail_free_emeta;
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);
191 pr_err("pblk: corrupted GC line (%d)\n", line->id);
192 goto fail_free_emeta;
197 gc_rq = kmalloc(sizeof(struct pblk_gc_rq), GFP_KERNEL);
199 goto fail_free_emeta;
203 bit = find_next_zero_bit(invalid_bitmap, lm->sec_per_line,
205 if (bit > line->emeta_ssec)
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);
212 if (unlikely(!nr_secs)) {
217 gc_rq->nr_secs = nr_secs;
220 gc_rq_ws = kmalloc(sizeof(struct pblk_line_ws), GFP_KERNEL);
222 goto fail_free_gc_rq;
224 gc_rq_ws->pblk = pblk;
225 gc_rq_ws->line = line;
226 gc_rq_ws->priv = gc_rq;
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.
232 while (down_timeout(&gc->gc_sem, msecs_to_jiffies(30000)))
235 kref_get(&line->ref);
237 INIT_WORK(&gc_rq_ws->ws, pblk_gc_line_ws);
238 queue_work(gc->gc_line_reader_wq, &gc_rq_ws->ws);
245 pblk_mfree(emeta_buf, l_mg->emeta_alloc_type);
247 kfree(invalid_bitmap);
249 kref_put(&line->ref, pblk_line_put);
250 atomic_dec(&gc->read_inflight_gc);
257 pblk_mfree(emeta_buf, l_mg->emeta_alloc_type);
259 kfree(invalid_bitmap);
263 pblk_put_line_back(pblk, line);
264 kref_put(&line->ref, pblk_line_put);
265 atomic_dec(&gc->read_inflight_gc);
267 pr_err("pblk: Failed to GC line %d\n", line->id);
270 static int pblk_gc_line(struct pblk *pblk, struct pblk_line *line)
272 struct pblk_gc *gc = &pblk->gc;
273 struct pblk_line_ws *line_ws;
275 pr_debug("pblk: line '%d' being reclaimed for GC\n", line->id);
277 line_ws = kmalloc(sizeof(struct pblk_line_ws), GFP_KERNEL);
281 line_ws->pblk = pblk;
282 line_ws->line = line;
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);
291 static void pblk_gc_reader_kick(struct pblk_gc *gc)
293 wake_up_process(gc->gc_reader_ts);
296 static void pblk_gc_kick(struct pblk *pblk)
298 struct pblk_gc *gc = &pblk->gc;
300 pblk_gc_writer_kick(gc);
301 pblk_gc_reader_kick(gc);
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));
311 static int pblk_gc_read(struct pblk *pblk)
313 struct pblk_gc *gc = &pblk->gc;
314 struct pblk_line *line;
316 spin_lock(&gc->r_lock);
317 if (list_empty(&gc->r_list)) {
318 spin_unlock(&gc->r_lock);
322 line = list_first_entry(&gc->r_list, struct pblk_line, list);
323 list_del(&line->list);
324 spin_unlock(&gc->r_lock);
328 if (pblk_gc_line(pblk, line))
329 pr_err("pblk: failed to GC line %d\n", line->id);
334 static struct pblk_line *pblk_gc_get_victim_line(struct pblk *pblk,
335 struct list_head *group_list)
337 struct pblk_line *line, *victim;
338 int line_vsc, victim_vsc;
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)
351 static bool pblk_gc_should_run(struct pblk_gc *gc, struct pblk_rl *rl)
353 unsigned int nr_blocks_free, nr_blocks_need;
355 nr_blocks_need = pblk_rl_high_thrs(rl);
356 nr_blocks_free = pblk_rl_nr_free_blks(rl);
358 /* This is not critical, no need to take lock here */
359 return ((gc->gc_active) && (nr_blocks_need > nr_blocks_free));
362 void pblk_gc_free_full_lines(struct pblk *pblk)
364 struct pblk_line_mgmt *l_mg = &pblk->l_mg;
365 struct pblk_gc *gc = &pblk->gc;
366 struct pblk_line *line;
369 spin_lock(&l_mg->gc_lock);
370 if (list_empty(&l_mg->gc_full_list)) {
371 spin_unlock(&l_mg->gc_lock);
375 line = list_first_entry(&l_mg->gc_full_list,
376 struct pblk_line, list);
378 spin_lock(&line->lock);
379 WARN_ON(line->state != PBLK_LINESTATE_CLOSED);
380 line->state = PBLK_LINESTATE_GC;
381 spin_unlock(&line->lock);
383 list_del(&line->list);
384 spin_unlock(&l_mg->gc_lock);
386 atomic_inc(&gc->pipeline_gc);
387 kref_put(&line->ref, pblk_line_put);
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.
397 static void pblk_gc_run(struct pblk *pblk)
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;
404 int read_inflight_gc, gc_group = 0, prev_group = 0;
406 pblk_gc_free_full_lines(pblk);
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))
413 group_list = l_mg->gc_lists[gc_group++];
416 spin_lock(&l_mg->gc_lock);
417 if (list_empty(group_list)) {
418 spin_unlock(&l_mg->gc_lock);
422 line = pblk_gc_get_victim_line(pblk, group_list);
424 spin_lock(&line->lock);
425 WARN_ON(line->state != PBLK_LINESTATE_CLOSED);
426 line->state = PBLK_LINESTATE_GC;
427 spin_unlock(&line->lock);
429 list_del(&line->list);
430 spin_unlock(&l_mg->gc_lock);
432 spin_lock(&gc->r_lock);
433 list_add_tail(&line->list, &gc->r_list);
434 spin_unlock(&gc->r_lock);
436 read_inflight_gc = atomic_inc_return(&gc->read_inflight_gc);
437 pblk_gc_reader_kick(gc);
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)
447 if (!prev_group && pblk->rl.rb_state > gc_group &&
448 gc_group < PBLK_GC_NR_LISTS)
452 static void pblk_gc_timer(struct timer_list *t)
454 struct pblk *pblk = from_timer(pblk, t, gc.gc_timer);
459 static int pblk_gc_ts(void *data)
461 struct pblk *pblk = data;
463 while (!kthread_should_stop()) {
465 set_current_state(TASK_INTERRUPTIBLE);
472 static int pblk_gc_writer_ts(void *data)
474 struct pblk *pblk = data;
476 while (!kthread_should_stop()) {
477 if (!pblk_gc_write(pblk))
479 set_current_state(TASK_INTERRUPTIBLE);
486 static int pblk_gc_reader_ts(void *data)
488 struct pblk *pblk = data;
489 struct pblk_gc *gc = &pblk->gc;
491 while (!kthread_should_stop()) {
492 if (!pblk_gc_read(pblk))
494 set_current_state(TASK_INTERRUPTIBLE);
498 #ifdef CONFIG_NVM_DEBUG
499 pr_info("pblk: flushing gc pipeline, %d lines left\n",
500 atomic_read(&gc->pipeline_gc));
504 if (!atomic_read(&gc->pipeline_gc))
513 static void pblk_gc_start(struct pblk *pblk)
515 pblk->gc.gc_active = 1;
516 pr_debug("pblk: gc start\n");
519 void pblk_gc_should_start(struct pblk *pblk)
521 struct pblk_gc *gc = &pblk->gc;
523 if (gc->gc_enabled && !gc->gc_active) {
530 * If flush_wq == 1 then no lock should be held by the caller since
531 * flush_workqueue can sleep
533 static void pblk_gc_stop(struct pblk *pblk, int flush_wq)
535 pblk->gc.gc_active = 0;
536 pr_debug("pblk: gc stop\n");
539 void pblk_gc_should_stop(struct pblk *pblk)
541 struct pblk_gc *gc = &pblk->gc;
543 if (gc->gc_active && !gc->gc_forced)
544 pblk_gc_stop(pblk, 0);
547 void pblk_gc_should_kick(struct pblk *pblk)
549 pblk_rl_update_rates(&pblk->rl);
552 void pblk_gc_sysfs_state_show(struct pblk *pblk, int *gc_enabled,
555 struct pblk_gc *gc = &pblk->gc;
557 spin_lock(&gc->lock);
558 *gc_enabled = gc->gc_enabled;
559 *gc_active = gc->gc_active;
560 spin_unlock(&gc->lock);
563 int pblk_gc_sysfs_force(struct pblk *pblk, int force)
565 struct pblk_gc *gc = &pblk->gc;
567 if (force < 0 || force > 1)
570 spin_lock(&gc->lock);
571 gc->gc_forced = force;
577 spin_unlock(&gc->lock);
579 pblk_gc_should_start(pblk);
584 int pblk_gc_init(struct pblk *pblk)
586 struct pblk_gc *gc = &pblk->gc;
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);
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;
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;
611 timer_setup(&gc->gc_timer, pblk_gc_timer, 0);
612 mod_timer(&gc->gc_timer, jiffies + msecs_to_jiffies(GC_TIME_MSECS));
618 atomic_set(&gc->read_inflight_gc, 0);
619 atomic_set(&gc->pipeline_gc, 0);
621 /* Workqueue that reads valid sectors from a line and submit them to the
622 * GC writer to be recycled.
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");
629 goto fail_free_reader_kthread;
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");
638 goto fail_free_reader_line_wq;
641 spin_lock_init(&gc->lock);
642 spin_lock_init(&gc->w_lock);
643 spin_lock_init(&gc->r_lock);
645 sema_init(&gc->gc_sem, PBLK_GC_RQ_QD);
647 INIT_LIST_HEAD(&gc->w_list);
648 INIT_LIST_HEAD(&gc->r_list);
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);
664 void pblk_gc_exit(struct pblk *pblk)
666 struct pblk_gc *gc = &pblk->gc;
669 del_timer_sync(&gc->gc_timer);
670 pblk_gc_stop(pblk, 1);
673 kthread_stop(gc->gc_ts);
675 if (gc->gc_reader_ts)
676 kthread_stop(gc->gc_reader_ts);
678 flush_workqueue(gc->gc_reader_wq);
679 if (gc->gc_reader_wq)
680 destroy_workqueue(gc->gc_reader_wq);
682 flush_workqueue(gc->gc_line_reader_wq);
683 if (gc->gc_line_reader_wq)
684 destroy_workqueue(gc->gc_line_reader_wq);
686 if (gc->gc_writer_ts)
687 kthread_stop(gc->gc_writer_ts);