1 // SPDX-License-Identifier: GPL-2.0
3 * Primary bucket allocation code
5 * Copyright 2012 Google, Inc.
7 * Allocation in bcache is done in terms of buckets:
9 * Each bucket has associated an 8 bit gen; this gen corresponds to the gen in
10 * btree pointers - they must match for the pointer to be considered valid.
12 * Thus (assuming a bucket has no dirty data or metadata in it) we can reuse a
13 * bucket simply by incrementing its gen.
15 * The gens (along with the priorities; it's really the gens are important but
16 * the code is named as if it's the priorities) are written in an arbitrary list
17 * of buckets on disk, with a pointer to them in the journal header.
19 * When we invalidate a bucket, we have to write its new gen to disk and wait
20 * for that write to complete before we use it - otherwise after a crash we
21 * could have pointers that appeared to be good but pointed to data that had
24 * Since the gens and priorities are all stored contiguously on disk, we can
25 * batch this up: We fill up the free_inc list with freshly invalidated buckets,
26 * call prio_write(), and when prio_write() finishes we pull buckets off the
27 * free_inc list and optionally discard them.
29 * free_inc isn't the only freelist - if it was, we'd often have to sleep while
30 * priorities and gens were being written before we could allocate. c->free is a
31 * smaller freelist, and buckets on that list are always ready to be used.
33 * If we've got discards enabled, that happens when a bucket moves from the
34 * free_inc list to the free list.
36 * It's important to ensure that gens don't wrap around - with respect to
37 * either the oldest gen in the btree or the gen on disk. This is quite
38 * difficult to do in practice, but we explicitly guard against it anyways - if
39 * a bucket is in danger of wrapping around we simply skip invalidating it that
40 * time around, and we garbage collect or rewrite the priorities sooner than we
41 * would have otherwise.
43 * bch2_bucket_alloc() allocates a single bucket from a specific device.
45 * bch2_bucket_alloc_set() allocates one or more buckets from different devices
46 * in a given filesystem.
48 * invalidate_buckets() drives all the processes described above. It's called
49 * from bch2_bucket_alloc() and a few other places that need to make sure free
52 * invalidate_buckets_(lru|fifo)() find buckets that are available to be
53 * invalidated, and then invalidate them and stick them on the free_inc list -
54 * in either lru or fifo order.
58 #include "alloc_background.h"
59 #include "alloc_foreground.h"
64 #include "disk_groups.h"
69 #include <linux/math64.h>
70 #include <linux/rculist.h>
71 #include <linux/rcupdate.h>
74 * Open buckets represent a bucket that's currently being allocated from. They
77 * - They track buckets that have been partially allocated, allowing for
78 * sub-bucket sized allocations - they're used by the sector allocator below
80 * - They provide a reference to the buckets they own that mark and sweep GC
81 * can find, until the new allocation has a pointer to it inserted into the
84 * When allocating some space with the sector allocator, the allocation comes
85 * with a reference to an open bucket - the caller is required to put that
86 * reference _after_ doing the index update that makes its allocation reachable.
89 void __bch2_open_bucket_put(struct bch_fs *c, struct open_bucket *ob)
91 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->ptr.dev);
94 bch2_ec_bucket_written(c, ob);
98 percpu_down_read(&c->mark_lock);
101 bch2_mark_alloc_bucket(c, ca, PTR_BUCKET_NR(ca, &ob->ptr),
102 false, gc_pos_alloc(c, ob), 0);
106 spin_unlock(&ob->lock);
107 percpu_up_read(&c->mark_lock);
109 spin_lock(&c->freelist_lock);
110 ob->freelist = c->open_buckets_freelist;
111 c->open_buckets_freelist = ob - c->open_buckets;
112 c->open_buckets_nr_free++;
113 spin_unlock(&c->freelist_lock);
115 closure_wake_up(&c->open_buckets_wait);
118 void bch2_open_bucket_write_error(struct bch_fs *c,
119 struct open_buckets *obs,
122 struct open_bucket *ob;
125 open_bucket_for_each(c, obs, ob, i)
126 if (ob->ptr.dev == dev &&
128 bch2_ec_bucket_cancel(c, ob);
131 static struct open_bucket *bch2_open_bucket_alloc(struct bch_fs *c)
133 struct open_bucket *ob;
135 BUG_ON(!c->open_buckets_freelist || !c->open_buckets_nr_free);
137 ob = c->open_buckets + c->open_buckets_freelist;
138 c->open_buckets_freelist = ob->freelist;
139 atomic_set(&ob->pin, 1);
142 c->open_buckets_nr_free--;
146 static void open_bucket_free_unused(struct bch_fs *c,
147 struct write_point *wp,
148 struct open_bucket *ob)
150 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->ptr.dev);
151 bool may_realloc = wp->type == BCH_DATA_user;
153 BUG_ON(ca->open_buckets_partial_nr >
154 ARRAY_SIZE(ca->open_buckets_partial));
156 if (ca->open_buckets_partial_nr <
157 ARRAY_SIZE(ca->open_buckets_partial) &&
159 spin_lock(&c->freelist_lock);
160 ob->on_partial_list = true;
161 ca->open_buckets_partial[ca->open_buckets_partial_nr++] =
162 ob - c->open_buckets;
163 spin_unlock(&c->freelist_lock);
165 closure_wake_up(&c->open_buckets_wait);
166 closure_wake_up(&c->freelist_wait);
168 bch2_open_bucket_put(c, ob);
172 static void verify_not_stale(struct bch_fs *c, const struct open_buckets *obs)
174 #ifdef CONFIG_BCACHEFS_DEBUG
175 struct open_bucket *ob;
178 open_bucket_for_each(c, obs, ob, i) {
179 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->ptr.dev);
181 BUG_ON(ptr_stale(ca, &ob->ptr));
186 /* _only_ for allocating the journal on a new device: */
187 long bch2_bucket_alloc_new_fs(struct bch_dev *ca)
189 struct bucket_array *buckets;
193 buckets = bucket_array(ca);
195 for (b = ca->mi.first_bucket; b < ca->mi.nbuckets; b++)
196 if (is_available_bucket(buckets->b[b].mark))
204 static inline unsigned open_buckets_reserved(enum alloc_reserve reserve)
208 case RESERVE_BTREE_MOVINGGC:
210 case RESERVE_MOVINGGC:
211 return OPEN_BUCKETS_COUNT / 4;
213 return OPEN_BUCKETS_COUNT / 2;
218 * bch_bucket_alloc - allocate a single bucket from a specific device
220 * Returns index of bucket on success, 0 on failure
222 struct open_bucket *bch2_bucket_alloc(struct bch_fs *c, struct bch_dev *ca,
223 enum alloc_reserve reserve,
224 bool may_alloc_partial,
227 struct bucket_array *buckets;
228 struct open_bucket *ob;
231 spin_lock(&c->freelist_lock);
233 if (may_alloc_partial) {
236 for (i = ca->open_buckets_partial_nr - 1; i >= 0; --i) {
237 ob = c->open_buckets + ca->open_buckets_partial[i];
239 if (reserve <= ob->alloc_reserve) {
240 array_remove_item(ca->open_buckets_partial,
241 ca->open_buckets_partial_nr,
243 ob->on_partial_list = false;
244 ob->alloc_reserve = reserve;
245 spin_unlock(&c->freelist_lock);
251 if (unlikely(c->open_buckets_nr_free <= open_buckets_reserved(reserve))) {
253 closure_wait(&c->open_buckets_wait, cl);
255 if (!c->blocked_allocate_open_bucket)
256 c->blocked_allocate_open_bucket = local_clock();
258 spin_unlock(&c->freelist_lock);
259 trace_open_bucket_alloc_fail(ca, reserve);
260 return ERR_PTR(-OPEN_BUCKETS_EMPTY);
263 if (likely(fifo_pop(&ca->free[RESERVE_NONE], bucket)))
267 case RESERVE_BTREE_MOVINGGC:
268 case RESERVE_MOVINGGC:
269 if (fifo_pop(&ca->free[RESERVE_MOVINGGC], bucket))
277 closure_wait(&c->freelist_wait, cl);
279 if (!c->blocked_allocate)
280 c->blocked_allocate = local_clock();
282 spin_unlock(&c->freelist_lock);
284 trace_bucket_alloc_fail(ca, reserve);
285 return ERR_PTR(-FREELIST_EMPTY);
287 verify_not_on_freelist(c, ca, bucket);
289 ob = bch2_open_bucket_alloc(c);
291 spin_lock(&ob->lock);
292 buckets = bucket_array(ca);
295 ob->sectors_free = ca->mi.bucket_size;
296 ob->alloc_reserve = reserve;
297 ob->ptr = (struct bch_extent_ptr) {
298 .type = 1 << BCH_EXTENT_ENTRY_ptr,
299 .gen = buckets->b[bucket].mark.gen,
300 .offset = bucket_to_sector(ca, bucket),
304 spin_unlock(&ob->lock);
306 if (c->blocked_allocate_open_bucket) {
307 bch2_time_stats_update(
308 &c->times[BCH_TIME_blocked_allocate_open_bucket],
309 c->blocked_allocate_open_bucket);
310 c->blocked_allocate_open_bucket = 0;
313 if (c->blocked_allocate) {
314 bch2_time_stats_update(
315 &c->times[BCH_TIME_blocked_allocate],
316 c->blocked_allocate);
317 c->blocked_allocate = 0;
320 spin_unlock(&c->freelist_lock);
322 bch2_wake_allocator(ca);
324 trace_bucket_alloc(ca, reserve);
328 static int __dev_stripe_cmp(struct dev_stripe_state *stripe,
329 unsigned l, unsigned r)
331 return ((stripe->next_alloc[l] > stripe->next_alloc[r]) -
332 (stripe->next_alloc[l] < stripe->next_alloc[r]));
335 #define dev_stripe_cmp(l, r) __dev_stripe_cmp(stripe, l, r)
337 struct dev_alloc_list bch2_dev_alloc_list(struct bch_fs *c,
338 struct dev_stripe_state *stripe,
339 struct bch_devs_mask *devs)
341 struct dev_alloc_list ret = { .nr = 0 };
344 for_each_set_bit(i, devs->d, BCH_SB_MEMBERS_MAX)
345 ret.devs[ret.nr++] = i;
347 bubble_sort(ret.devs, ret.nr, dev_stripe_cmp);
351 void bch2_dev_stripe_increment(struct bch_dev *ca,
352 struct dev_stripe_state *stripe)
354 u64 *v = stripe->next_alloc + ca->dev_idx;
355 u64 free_space = dev_buckets_free(ca);
356 u64 free_space_inv = free_space
357 ? div64_u64(1ULL << 48, free_space)
361 if (*v + free_space_inv >= *v)
362 *v += free_space_inv;
366 for (v = stripe->next_alloc;
367 v < stripe->next_alloc + ARRAY_SIZE(stripe->next_alloc); v++)
368 *v = *v < scale ? 0 : *v - scale;
371 #define BUCKET_MAY_ALLOC_PARTIAL (1 << 0)
372 #define BUCKET_ALLOC_USE_DURABILITY (1 << 1)
374 static void add_new_bucket(struct bch_fs *c,
375 struct open_buckets *ptrs,
376 struct bch_devs_mask *devs_may_alloc,
377 unsigned *nr_effective,
380 struct open_bucket *ob)
382 unsigned durability =
383 bch_dev_bkey_exists(c, ob->ptr.dev)->mi.durability;
385 __clear_bit(ob->ptr.dev, devs_may_alloc->d);
386 *nr_effective += (flags & BUCKET_ALLOC_USE_DURABILITY)
388 *have_cache |= !durability;
390 ob_push(c, ptrs, ob);
393 enum bucket_alloc_ret
394 bch2_bucket_alloc_set(struct bch_fs *c,
395 struct open_buckets *ptrs,
396 struct dev_stripe_state *stripe,
397 struct bch_devs_mask *devs_may_alloc,
398 unsigned nr_replicas,
399 unsigned *nr_effective,
401 enum alloc_reserve reserve,
405 struct dev_alloc_list devs_sorted =
406 bch2_dev_alloc_list(c, stripe, devs_may_alloc);
408 enum bucket_alloc_ret ret = INSUFFICIENT_DEVICES;
411 BUG_ON(*nr_effective >= nr_replicas);
413 for (i = 0; i < devs_sorted.nr; i++) {
414 struct open_bucket *ob;
416 ca = rcu_dereference(c->devs[devs_sorted.devs[i]]);
420 if (!ca->mi.durability && *have_cache)
423 ob = bch2_bucket_alloc(c, ca, reserve,
424 flags & BUCKET_MAY_ALLOC_PARTIAL, cl);
433 add_new_bucket(c, ptrs, devs_may_alloc,
434 nr_effective, have_cache, flags, ob);
436 bch2_dev_stripe_increment(ca, stripe);
438 if (*nr_effective >= nr_replicas)
439 return ALLOC_SUCCESS;
445 /* Allocate from stripes: */
448 * if we can't allocate a new stripe because there are already too many
449 * partially filled stripes, force allocating from an existing stripe even when
450 * it's to a device we don't want:
453 static enum bucket_alloc_ret
454 bucket_alloc_from_stripe(struct bch_fs *c,
455 struct open_buckets *ptrs,
456 struct write_point *wp,
457 struct bch_devs_mask *devs_may_alloc,
459 unsigned erasure_code,
460 unsigned nr_replicas,
461 unsigned *nr_effective,
466 struct dev_alloc_list devs_sorted;
467 struct ec_stripe_head *h;
468 struct open_bucket *ob;
478 if (ec_open_bucket(c, ptrs))
481 h = bch2_ec_stripe_head_get(c, target, 0, nr_replicas - 1,
482 wp == &c->copygc_write_point,
489 devs_sorted = bch2_dev_alloc_list(c, &wp->stripe, devs_may_alloc);
491 for (i = 0; i < devs_sorted.nr; i++)
492 open_bucket_for_each(c, &h->s->blocks, ob, ec_idx)
493 if (ob->ptr.dev == devs_sorted.devs[i] &&
494 !test_and_set_bit(h->s->data_block_idx[ec_idx],
495 h->s->blocks_allocated))
499 ca = bch_dev_bkey_exists(c, ob->ptr.dev);
501 ob->ec_idx = h->s->data_block_idx[ec_idx];
504 add_new_bucket(c, ptrs, devs_may_alloc,
505 nr_effective, have_cache, flags, ob);
506 atomic_inc(&h->s->pin);
508 bch2_ec_stripe_head_put(c, h);
512 /* Sector allocator */
514 static void get_buckets_from_writepoint(struct bch_fs *c,
515 struct open_buckets *ptrs,
516 struct write_point *wp,
517 struct bch_devs_mask *devs_may_alloc,
518 unsigned nr_replicas,
519 unsigned *nr_effective,
524 struct open_buckets ptrs_skip = { .nr = 0 };
525 struct open_bucket *ob;
528 open_bucket_for_each(c, &wp->ptrs, ob, i) {
529 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->ptr.dev);
531 if (*nr_effective < nr_replicas &&
532 test_bit(ob->ptr.dev, devs_may_alloc->d) &&
533 (ca->mi.durability ||
534 (wp->type == BCH_DATA_user && !*have_cache)) &&
535 (ob->ec || !need_ec)) {
536 add_new_bucket(c, ptrs, devs_may_alloc,
537 nr_effective, have_cache,
540 ob_push(c, &ptrs_skip, ob);
543 wp->ptrs = ptrs_skip;
546 static enum bucket_alloc_ret
547 open_bucket_add_buckets(struct bch_fs *c,
548 struct open_buckets *ptrs,
549 struct write_point *wp,
550 struct bch_devs_list *devs_have,
552 unsigned erasure_code,
553 unsigned nr_replicas,
554 unsigned *nr_effective,
556 enum alloc_reserve reserve,
560 struct bch_devs_mask devs;
561 struct open_bucket *ob;
562 struct closure *cl = NULL;
563 enum bucket_alloc_ret ret;
567 devs = target_rw_devs(c, wp->type, target);
570 /* Don't allocate from devices we already have pointers to: */
571 for (i = 0; i < devs_have->nr; i++)
572 __clear_bit(devs_have->devs[i], devs.d);
574 open_bucket_for_each(c, ptrs, ob, i)
575 __clear_bit(ob->ptr.dev, devs.d);
578 if (!ec_open_bucket(c, ptrs)) {
579 get_buckets_from_writepoint(c, ptrs, wp, &devs,
580 nr_replicas, nr_effective,
581 have_cache, flags, true);
582 if (*nr_effective >= nr_replicas)
586 if (!ec_open_bucket(c, ptrs)) {
587 ret = bucket_alloc_from_stripe(c, ptrs, wp, &devs,
588 target, erasure_code,
589 nr_replicas, nr_effective,
590 have_cache, flags, _cl);
591 if (ret == FREELIST_EMPTY ||
592 ret == OPEN_BUCKETS_EMPTY)
594 if (*nr_effective >= nr_replicas)
599 get_buckets_from_writepoint(c, ptrs, wp, &devs,
600 nr_replicas, nr_effective,
601 have_cache, flags, false);
602 if (*nr_effective >= nr_replicas)
605 percpu_down_read(&c->mark_lock);
610 * Try nonblocking first, so that if one device is full we'll try from
613 ret = bch2_bucket_alloc_set(c, ptrs, &wp->stripe, &devs,
614 nr_replicas, nr_effective, have_cache,
616 if (ret && ret != INSUFFICIENT_DEVICES && !cl && _cl) {
622 percpu_up_read(&c->mark_lock);
627 void bch2_open_buckets_stop_dev(struct bch_fs *c, struct bch_dev *ca,
628 struct open_buckets *obs)
630 struct open_buckets ptrs = { .nr = 0 };
631 struct open_bucket *ob, *ob2;
634 open_bucket_for_each(c, obs, ob, i) {
635 bool drop = !ca || ob->ptr.dev == ca->dev_idx;
637 if (!drop && ob->ec) {
638 mutex_lock(&ob->ec->lock);
639 open_bucket_for_each(c, &ob->ec->blocks, ob2, j)
640 drop |= ob2->ptr.dev == ca->dev_idx;
641 open_bucket_for_each(c, &ob->ec->parity, ob2, j)
642 drop |= ob2->ptr.dev == ca->dev_idx;
643 mutex_unlock(&ob->ec->lock);
647 bch2_open_bucket_put(c, ob);
649 ob_push(c, &ptrs, ob);
655 void bch2_writepoint_stop(struct bch_fs *c, struct bch_dev *ca,
656 struct write_point *wp)
658 mutex_lock(&wp->lock);
659 bch2_open_buckets_stop_dev(c, ca, &wp->ptrs);
660 mutex_unlock(&wp->lock);
663 static inline struct hlist_head *writepoint_hash(struct bch_fs *c,
664 unsigned long write_point)
667 hash_long(write_point, ilog2(ARRAY_SIZE(c->write_points_hash)));
669 return &c->write_points_hash[hash];
672 static struct write_point *__writepoint_find(struct hlist_head *head,
673 unsigned long write_point)
675 struct write_point *wp;
677 hlist_for_each_entry_rcu(wp, head, node)
678 if (wp->write_point == write_point)
684 static inline bool too_many_writepoints(struct bch_fs *c, unsigned factor)
686 u64 stranded = c->write_points_nr * c->bucket_size_max;
687 u64 free = bch2_fs_usage_read_short(c).free;
689 return stranded * factor > free;
692 static bool try_increase_writepoints(struct bch_fs *c)
694 struct write_point *wp;
696 if (c->write_points_nr == ARRAY_SIZE(c->write_points) ||
697 too_many_writepoints(c, 32))
700 wp = c->write_points + c->write_points_nr++;
701 hlist_add_head_rcu(&wp->node, writepoint_hash(c, wp->write_point));
705 static bool try_decrease_writepoints(struct bch_fs *c,
708 struct write_point *wp;
710 mutex_lock(&c->write_points_hash_lock);
711 if (c->write_points_nr < old_nr) {
712 mutex_unlock(&c->write_points_hash_lock);
716 if (c->write_points_nr == 1 ||
717 !too_many_writepoints(c, 8)) {
718 mutex_unlock(&c->write_points_hash_lock);
722 wp = c->write_points + --c->write_points_nr;
724 hlist_del_rcu(&wp->node);
725 mutex_unlock(&c->write_points_hash_lock);
727 bch2_writepoint_stop(c, NULL, wp);
731 static struct write_point *writepoint_find(struct bch_fs *c,
732 unsigned long write_point)
734 struct write_point *wp, *oldest;
735 struct hlist_head *head;
737 if (!(write_point & 1UL)) {
738 wp = (struct write_point *) write_point;
739 mutex_lock(&wp->lock);
743 head = writepoint_hash(c, write_point);
745 wp = __writepoint_find(head, write_point);
748 mutex_lock(&wp->lock);
749 if (wp->write_point == write_point)
751 mutex_unlock(&wp->lock);
756 for (wp = c->write_points;
757 wp < c->write_points + c->write_points_nr; wp++)
758 if (!oldest || time_before64(wp->last_used, oldest->last_used))
761 mutex_lock(&oldest->lock);
762 mutex_lock(&c->write_points_hash_lock);
763 if (oldest >= c->write_points + c->write_points_nr ||
764 try_increase_writepoints(c)) {
765 mutex_unlock(&c->write_points_hash_lock);
766 mutex_unlock(&oldest->lock);
767 goto restart_find_oldest;
770 wp = __writepoint_find(head, write_point);
771 if (wp && wp != oldest) {
772 mutex_unlock(&c->write_points_hash_lock);
773 mutex_unlock(&oldest->lock);
778 hlist_del_rcu(&wp->node);
779 wp->write_point = write_point;
780 hlist_add_head_rcu(&wp->node, head);
781 mutex_unlock(&c->write_points_hash_lock);
783 wp->last_used = sched_clock();
788 * Get us an open_bucket we can allocate from, return with it locked:
790 struct write_point *bch2_alloc_sectors_start(struct bch_fs *c,
792 unsigned erasure_code,
793 struct write_point_specifier write_point,
794 struct bch_devs_list *devs_have,
795 unsigned nr_replicas,
796 unsigned nr_replicas_required,
797 enum alloc_reserve reserve,
801 struct write_point *wp;
802 struct open_bucket *ob;
803 struct open_buckets ptrs;
804 unsigned nr_effective, write_points_nr;
805 unsigned ob_flags = 0;
807 enum bucket_alloc_ret ret;
810 if (!(flags & BCH_WRITE_ONLY_SPECIFIED_DEVS))
811 ob_flags |= BUCKET_ALLOC_USE_DURABILITY;
813 BUG_ON(!nr_replicas || !nr_replicas_required);
817 write_points_nr = c->write_points_nr;
820 wp = writepoint_find(c, write_point.v);
822 if (wp->type == BCH_DATA_user)
823 ob_flags |= BUCKET_MAY_ALLOC_PARTIAL;
825 /* metadata may not allocate on cache devices: */
826 if (wp->type != BCH_DATA_user)
829 if (!target || (flags & BCH_WRITE_ONLY_SPECIFIED_DEVS)) {
830 ret = open_bucket_add_buckets(c, &ptrs, wp, devs_have,
831 target, erasure_code,
832 nr_replicas, &nr_effective,
833 &have_cache, reserve,
836 ret = open_bucket_add_buckets(c, &ptrs, wp, devs_have,
837 target, erasure_code,
838 nr_replicas, &nr_effective,
839 &have_cache, reserve,
844 ret = open_bucket_add_buckets(c, &ptrs, wp, devs_have,
846 nr_replicas, &nr_effective,
847 &have_cache, reserve,
851 BUG_ON(!ret && nr_effective < nr_replicas);
853 if (erasure_code && !ec_open_bucket(c, &ptrs))
854 pr_debug("failed to get ec bucket: ret %u", ret);
856 if (ret == INSUFFICIENT_DEVICES &&
857 nr_effective >= nr_replicas_required)
863 /* Free buckets we didn't use: */
864 open_bucket_for_each(c, &wp->ptrs, ob, i)
865 open_bucket_free_unused(c, wp, ob);
869 wp->sectors_free = UINT_MAX;
871 open_bucket_for_each(c, &wp->ptrs, ob, i)
872 wp->sectors_free = min(wp->sectors_free, ob->sectors_free);
874 BUG_ON(!wp->sectors_free || wp->sectors_free == UINT_MAX);
876 verify_not_stale(c, &wp->ptrs);
880 open_bucket_for_each(c, &wp->ptrs, ob, i)
881 if (ptrs.nr < ARRAY_SIZE(ptrs.v))
882 ob_push(c, &ptrs, ob);
884 open_bucket_free_unused(c, wp, ob);
887 mutex_unlock(&wp->lock);
889 if (ret == FREELIST_EMPTY &&
890 try_decrease_writepoints(c, write_points_nr))
894 case OPEN_BUCKETS_EMPTY:
896 return cl ? ERR_PTR(-EAGAIN) : ERR_PTR(-ENOSPC);
897 case INSUFFICIENT_DEVICES:
898 return ERR_PTR(-EROFS);
905 * Append pointers to the space we just allocated to @k, and mark @sectors space
906 * as allocated out of @ob
908 void bch2_alloc_sectors_append_ptrs(struct bch_fs *c, struct write_point *wp,
909 struct bkey_i *k, unsigned sectors)
912 struct open_bucket *ob;
915 BUG_ON(sectors > wp->sectors_free);
916 wp->sectors_free -= sectors;
918 open_bucket_for_each(c, &wp->ptrs, ob, i) {
919 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->ptr.dev);
920 struct bch_extent_ptr tmp = ob->ptr;
922 tmp.cached = !ca->mi.durability &&
923 wp->type == BCH_DATA_user;
925 tmp.offset += ca->mi.bucket_size - ob->sectors_free;
926 bch2_bkey_append_ptr(k, tmp);
928 BUG_ON(sectors > ob->sectors_free);
929 ob->sectors_free -= sectors;
934 * Append pointers to the space we just allocated to @k, and mark @sectors space
935 * as allocated out of @ob
937 void bch2_alloc_sectors_done(struct bch_fs *c, struct write_point *wp)
939 struct open_buckets ptrs = { .nr = 0 }, keep = { .nr = 0 };
940 struct open_bucket *ob;
943 open_bucket_for_each(c, &wp->ptrs, ob, i)
944 ob_push(c, !ob->sectors_free ? &ptrs : &keep, ob);
947 mutex_unlock(&wp->lock);
949 bch2_open_buckets_put(c, &ptrs);
952 static inline void writepoint_init(struct write_point *wp,
953 enum bch_data_type type)
955 mutex_init(&wp->lock);
959 void bch2_fs_allocator_foreground_init(struct bch_fs *c)
961 struct open_bucket *ob;
962 struct write_point *wp;
964 mutex_init(&c->write_points_hash_lock);
965 c->write_points_nr = ARRAY_SIZE(c->write_points);
967 /* open bucket 0 is a sentinal NULL: */
968 spin_lock_init(&c->open_buckets[0].lock);
970 for (ob = c->open_buckets + 1;
971 ob < c->open_buckets + ARRAY_SIZE(c->open_buckets); ob++) {
972 spin_lock_init(&ob->lock);
973 c->open_buckets_nr_free++;
975 ob->freelist = c->open_buckets_freelist;
976 c->open_buckets_freelist = ob - c->open_buckets;
979 writepoint_init(&c->btree_write_point, BCH_DATA_btree);
980 writepoint_init(&c->rebalance_write_point, BCH_DATA_user);
981 writepoint_init(&c->copygc_write_point, BCH_DATA_user);
983 for (wp = c->write_points;
984 wp < c->write_points + c->write_points_nr; wp++) {
985 writepoint_init(wp, BCH_DATA_user);
987 wp->last_used = sched_clock();
988 wp->write_point = (unsigned long) wp;
989 hlist_add_head_rcu(&wp->node,
990 writepoint_hash(c, wp->write_point));