X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=libbcachefs%2Fmovinggc.c;h=d079ba7aa65349bd8c6fbe73f65dee2872a249a0;hb=1dc5842111fa587eb58112f4ae3c2048b1c1a006;hp=72cbb9d50f69162ce1b525789efc0a34508424ca;hpb=c35fbbc025c6099969befb4dfaf065215cf40cf3;p=bcachefs-tools-debian diff --git a/libbcachefs/movinggc.c b/libbcachefs/movinggc.c index 72cbb9d..d079ba7 100644 --- a/libbcachefs/movinggc.c +++ b/libbcachefs/movinggc.c @@ -1,3 +1,4 @@ +// SPDX-License-Identifier: GPL-2.0 /* * Moving/copying garbage collector * @@ -5,323 +6,436 @@ */ #include "bcachefs.h" +#include "alloc_background.h" +#include "alloc_foreground.h" #include "btree_iter.h" +#include "btree_update.h" +#include "btree_write_buffer.h" #include "buckets.h" #include "clock.h" -#include "extents.h" -#include "eytzinger.h" -#include "io.h" -#include "keylist.h" +#include "errcode.h" +#include "error.h" +#include "lru.h" #include "move.h" #include "movinggc.h" +#include "trace.h" -#include #include #include #include -#include +#include #include -/* Moving GC - IO loop */ - -static int bucket_idx_cmp(const void *_l, const void *_r, size_t size) +struct buckets_in_flight { + struct rhashtable table; + struct move_bucket_in_flight *first; + struct move_bucket_in_flight *last; + size_t nr; + size_t sectors; +}; + +static const struct rhashtable_params bch_move_bucket_params = { + .head_offset = offsetof(struct move_bucket_in_flight, hash), + .key_offset = offsetof(struct move_bucket_in_flight, bucket.k), + .key_len = sizeof(struct move_bucket_key), +}; + +static struct move_bucket_in_flight * +move_bucket_in_flight_add(struct buckets_in_flight *list, struct move_bucket b) { - const struct bucket_heap_entry *l = _l; - const struct bucket_heap_entry *r = _r; + struct move_bucket_in_flight *new = kzalloc(sizeof(*new), GFP_KERNEL); + int ret; - if (l->bucket < r->bucket) - return -1; - if (l->bucket > r->bucket) - return 1; - return 0; -} + if (!new) + return ERR_PTR(-ENOMEM); -static const struct bch_extent_ptr *moving_pred(struct bch_dev *ca, - struct bkey_s_c k) -{ - bucket_heap *h = &ca->copygc_heap; - const struct bch_extent_ptr *ptr; - - if (bkey_extent_is_data(k.k) && - (ptr = bch2_extent_has_device(bkey_s_c_to_extent(k), - ca->dev_idx))) { - struct bucket_heap_entry search = { - .bucket = PTR_BUCKET_NR(ca, ptr) - }; - - size_t i = eytzinger0_find(h->data, h->used, - sizeof(h->data[0]), - bucket_idx_cmp, &search); - - if (i < h->used) - return ptr; + new->bucket = b; + + ret = rhashtable_lookup_insert_fast(&list->table, &new->hash, + bch_move_bucket_params); + if (ret) { + kfree(new); + return ERR_PTR(ret); } - return NULL; + if (!list->first) + list->first = new; + else + list->last->next = new; + + list->last = new; + list->nr++; + list->sectors += b.sectors; + return new; } -static int issue_moving_gc_move(struct bch_dev *ca, - struct moving_context *ctxt, - struct bkey_s_c k) +static int bch2_bucket_is_movable(struct btree_trans *trans, + struct move_bucket *b, u64 time) { - struct bch_fs *c = ca->fs; - const struct bch_extent_ptr *ptr; + struct btree_iter iter; + struct bkey_s_c k; + struct bch_alloc_v4 _a; + const struct bch_alloc_v4 *a; int ret; - ptr = moving_pred(ca, k); - if (!ptr) /* We raced - bucket's been reused */ + if (bch2_bucket_is_open(trans->c, + b->k.bucket.inode, + b->k.bucket.offset)) return 0; - ret = bch2_data_move(c, ctxt, &ca->copygc_write_point, k, ptr); - if (!ret) - trace_gc_copy(k.k); - else - trace_moving_gc_alloc_fail(c, k.k->size); + k = bch2_bkey_get_iter(trans, &iter, BTREE_ID_alloc, + b->k.bucket, BTREE_ITER_CACHED); + ret = bkey_err(k); + if (ret) + return ret; + + a = bch2_alloc_to_v4(k, &_a); + b->k.gen = a->gen; + b->sectors = bch2_bucket_sectors_dirty(*a); + + ret = data_type_movable(a->data_type) && + a->fragmentation_lru && + a->fragmentation_lru <= time; + + bch2_trans_iter_exit(trans, &iter); return ret; } -static void read_moving(struct bch_dev *ca, size_t buckets_to_move, - u64 sectors_to_move) +static void move_buckets_wait(struct moving_context *ctxt, + struct buckets_in_flight *list, + bool flush) { - struct bch_fs *c = ca->fs; - bucket_heap *h = &ca->copygc_heap; - struct moving_context ctxt; + struct move_bucket_in_flight *i; + int ret; + + while ((i = list->first)) { + if (flush) + move_ctxt_wait_event(ctxt, !atomic_read(&i->count)); + + if (atomic_read(&i->count)) + break; + + list->first = i->next; + if (!list->first) + list->last = NULL; + + list->nr--; + list->sectors -= i->bucket.sectors; + + ret = rhashtable_remove_fast(&list->table, &i->hash, + bch_move_bucket_params); + BUG_ON(ret); + kfree(i); + } + + bch2_trans_unlock_long(ctxt->trans); +} + +static bool bucket_in_flight(struct buckets_in_flight *list, + struct move_bucket_key k) +{ + return rhashtable_lookup_fast(&list->table, &k, bch_move_bucket_params); +} + +typedef DARRAY(struct move_bucket) move_buckets; + +static int bch2_copygc_get_buckets(struct moving_context *ctxt, + struct buckets_in_flight *buckets_in_flight, + move_buckets *buckets) +{ + struct btree_trans *trans = ctxt->trans; + struct bch_fs *c = trans->c; struct btree_iter iter; struct bkey_s_c k; - u64 sectors_not_moved = 0; - size_t buckets_not_moved = 0; - struct bucket_heap_entry *i; - - bch2_ratelimit_reset(&ca->moving_gc_pd.rate); - bch2_move_ctxt_init(&ctxt, &ca->moving_gc_pd.rate, - SECTORS_IN_FLIGHT_PER_DEVICE); - bch2_btree_iter_init(&iter, c, BTREE_ID_EXTENTS, POS_MIN, - BTREE_ITER_PREFETCH); - - while (1) { - if (kthread_should_stop()) - goto out; - if (bch2_move_ctxt_wait(&ctxt)) - goto out; - k = bch2_btree_iter_peek(&iter); - if (!k.k) - break; - if (btree_iter_err(k)) - goto out; + size_t nr_to_get = max_t(size_t, 16U, buckets_in_flight->nr / 4); + size_t saw = 0, in_flight = 0, not_movable = 0, sectors = 0; + struct bpos last_flushed_pos = POS_MIN; + int ret; + + move_buckets_wait(ctxt, buckets_in_flight, false); + + ret = bch2_btree_write_buffer_tryflush(trans); + if (bch2_err_matches(ret, EROFS)) + return ret; + + if (bch2_fs_fatal_err_on(ret, c, "%s: error %s from bch2_btree_write_buffer_tryflush()", + __func__, bch2_err_str(ret))) + return ret; - if (!moving_pred(ca, k)) + ret = for_each_btree_key2_upto(trans, iter, BTREE_ID_lru, + lru_pos(BCH_LRU_FRAGMENTATION_START, 0, 0), + lru_pos(BCH_LRU_FRAGMENTATION_START, U64_MAX, LRU_TIME_MAX), + 0, k, ({ + int ret2 = bch2_check_lru_key(trans, &iter, k, &last_flushed_pos); + if (ret2) { + ret2 = ret2 < 0 ? ret2 : 0; goto next; + } - if (issue_moving_gc_move(ca, &ctxt, k)) { - bch2_btree_iter_unlock(&iter); + saw++; - /* memory allocation failure, wait for some IO to finish */ - bch2_move_ctxt_wait_for_io(&ctxt); - continue; + struct move_bucket b = { .k.bucket = u64_to_bucket(k.k->p.offset) }; + + if (!bch2_bucket_is_movable(trans, &b, lru_pos_time(k.k->p))) + not_movable++; + else if (bucket_in_flight(buckets_in_flight, b.k)) + in_flight++; + else { + ret2 = darray_push(buckets, b) ?: buckets->nr >= nr_to_get; + if (ret2 >= 0) + sectors += b.sectors; } next: - bch2_btree_iter_advance_pos(&iter); - //bch2_btree_iter_cond_resched(&iter); + ret2; + })); - /* unlock before calling moving_context_wait() */ - bch2_btree_iter_unlock(&iter); - cond_resched(); - } + pr_debug("have: %zu (%zu) saw %zu in flight %zu not movable %zu got %zu (%zu)/%zu buckets ret %i", + buckets_in_flight->nr, buckets_in_flight->sectors, + saw, in_flight, not_movable, buckets->nr, sectors, nr_to_get, ret); - bch2_btree_iter_unlock(&iter); - bch2_move_ctxt_exit(&ctxt); - trace_moving_gc_end(ca, ctxt.sectors_moved, ctxt.keys_moved, - buckets_to_move); + return ret < 0 ? ret : 0; +} - /* don't check this if we bailed out early: */ - for (i = h->data; i < h->data + h->used; i++) { - struct bucket_mark m = READ_ONCE(ca->buckets[i->bucket].mark); +noinline +static int bch2_copygc(struct moving_context *ctxt, + struct buckets_in_flight *buckets_in_flight, + bool *did_work) +{ + struct btree_trans *trans = ctxt->trans; + struct bch_fs *c = trans->c; + struct data_update_opts data_opts = { + .btree_insert_flags = BCH_WATERMARK_copygc, + }; + move_buckets buckets = { 0 }; + struct move_bucket_in_flight *f; + struct move_bucket *i; + u64 moved = atomic64_read(&ctxt->stats->sectors_moved); + int ret = 0; + + ret = bch2_copygc_get_buckets(ctxt, buckets_in_flight, &buckets); + if (ret) + goto err; + + darray_for_each(buckets, i) { + if (unlikely(freezing(current))) + break; - if (i->mark.gen == m.gen && bucket_sectors_used(m)) { - sectors_not_moved += bucket_sectors_used(m); - buckets_not_moved++; + f = move_bucket_in_flight_add(buckets_in_flight, *i); + ret = PTR_ERR_OR_ZERO(f); + if (ret == -EEXIST) { /* rare race: copygc_get_buckets returned same bucket more than once */ + ret = 0; + continue; } + if (ret == -ENOMEM) { /* flush IO, continue later */ + ret = 0; + break; + } + + ret = __bch2_evacuate_bucket(ctxt, f, f->bucket.k.bucket, + f->bucket.k.gen, data_opts); + if (ret) + goto err; + + *did_work = true; } +err: + darray_exit(&buckets); + + /* no entries in LRU btree found, or got to end: */ + if (bch2_err_matches(ret, ENOENT)) + ret = 0; + + if (ret < 0 && !bch2_err_matches(ret, EROFS)) + bch_err_msg(c, ret, "from bch2_move_data()"); - if (sectors_not_moved) - bch_warn(c, "copygc finished but %llu/%llu sectors, %zu/%zu buckets not moved", - sectors_not_moved, sectors_to_move, - buckets_not_moved, buckets_to_move); - return; -out: - bch2_btree_iter_unlock(&iter); - bch2_move_ctxt_exit(&ctxt); - trace_moving_gc_end(ca, ctxt.sectors_moved, ctxt.keys_moved, - buckets_to_move); + moved = atomic64_read(&ctxt->stats->sectors_moved) - moved; + trace_and_count(c, copygc, c, moved, 0, 0, 0); + return ret; } -static bool have_copygc_reserve(struct bch_dev *ca) +/* + * Copygc runs when the amount of fragmented data is above some arbitrary + * threshold: + * + * The threshold at the limit - when the device is full - is the amount of space + * we reserved in bch2_recalc_capacity; we can't have more than that amount of + * disk space stranded due to fragmentation and store everything we have + * promised to store. + * + * But we don't want to be running copygc unnecessarily when the device still + * has plenty of free space - rather, we want copygc to smoothly run every so + * often and continually reduce the amount of fragmented space as the device + * fills up. So, we increase the threshold by half the current free space. + */ +unsigned long bch2_copygc_wait_amount(struct bch_fs *c) { - bool ret; + struct bch_dev *ca; + unsigned dev_idx; + s64 wait = S64_MAX, fragmented_allowed, fragmented; + unsigned i; - spin_lock(&ca->freelist_lock); - ret = fifo_used(&ca->free[RESERVE_MOVINGGC]) >= - COPYGC_BUCKETS_PER_ITER(ca); - spin_unlock(&ca->freelist_lock); + for_each_rw_member(ca, c, dev_idx) { + struct bch_dev_usage usage = bch2_dev_usage_read(ca); - return ret; + fragmented_allowed = ((__dev_buckets_available(ca, usage, BCH_WATERMARK_stripe) * + ca->mi.bucket_size) >> 1); + fragmented = 0; + + for (i = 0; i < BCH_DATA_NR; i++) + if (data_type_movable(i)) + fragmented += usage.d[i].fragmented; + + wait = min(wait, max(0LL, fragmented_allowed - fragmented)); + } + + return wait; } -static inline int sectors_used_cmp(bucket_heap *heap, - struct bucket_heap_entry l, - struct bucket_heap_entry r) +void bch2_copygc_wait_to_text(struct printbuf *out, struct bch_fs *c) { - return bucket_sectors_used(l.mark) - bucket_sectors_used(r.mark); + prt_printf(out, "Currently waiting for: "); + prt_human_readable_u64(out, max(0LL, c->copygc_wait - + atomic64_read(&c->io_clock[WRITE].now)) << 9); + prt_newline(out); + + prt_printf(out, "Currently waiting since: "); + prt_human_readable_u64(out, max(0LL, + atomic64_read(&c->io_clock[WRITE].now) - + c->copygc_wait_at) << 9); + prt_newline(out); + + prt_printf(out, "Currently calculated wait: "); + prt_human_readable_u64(out, bch2_copygc_wait_amount(c)); + prt_newline(out); } -static void bch2_moving_gc(struct bch_dev *ca) +static int bch2_copygc_thread(void *arg) { - struct bch_fs *c = ca->fs; - struct bucket *g; - u64 sectors_to_move = 0; - size_t buckets_to_move, buckets_unused = 0; - struct bucket_heap_entry e, *i; - int reserve_sectors; - - if (!have_copygc_reserve(ca)) { - struct closure cl; - - closure_init_stack(&cl); - while (1) { - closure_wait(&c->freelist_wait, &cl); - if (have_copygc_reserve(ca)) - break; - closure_sync(&cl); - } - closure_wake_up(&c->freelist_wait); + struct bch_fs *c = arg; + struct moving_context ctxt; + struct bch_move_stats move_stats; + struct io_clock *clock = &c->io_clock[WRITE]; + struct buckets_in_flight *buckets; + u64 last, wait; + int ret = 0; + + buckets = kzalloc(sizeof(struct buckets_in_flight), GFP_KERNEL); + if (!buckets) + return -ENOMEM; + ret = rhashtable_init(&buckets->table, &bch_move_bucket_params); + if (ret) { + kfree(buckets); + bch_err_msg(c, ret, "allocating copygc buckets in flight"); + return ret; } - reserve_sectors = COPYGC_SECTORS_PER_ITER(ca); - - trace_moving_gc_start(ca); - - /* - * Find buckets with lowest sector counts, skipping completely - * empty buckets, by building a maxheap sorted by sector count, - * and repeatedly replacing the maximum element until all - * buckets have been visited. - */ - - /* - * We need bucket marks to be up to date - gc can't be recalculating - * them: - */ - down_read(&c->gc_lock); - ca->copygc_heap.used = 0; - for_each_bucket(g, ca) { - struct bucket_mark m = READ_ONCE(g->mark); - struct bucket_heap_entry e = { g - ca->buckets, m }; - - if (bucket_unused(m)) { - buckets_unused++; - continue; - } + set_freezable(); - if (m.owned_by_allocator || - m.data_type != BUCKET_DATA) - continue; + bch2_move_stats_init(&move_stats, "copygc"); + bch2_moving_ctxt_init(&ctxt, c, NULL, &move_stats, + writepoint_ptr(&c->copygc_write_point), + false); - if (bucket_sectors_used(m) >= ca->mi.bucket_size) - continue; + while (!ret && !kthread_should_stop()) { + bool did_work = false; - heap_add_or_replace(&ca->copygc_heap, e, -sectors_used_cmp); - } - up_read(&c->gc_lock); + bch2_trans_unlock_long(ctxt.trans); + cond_resched(); - for (i = ca->copygc_heap.data; - i < ca->copygc_heap.data + ca->copygc_heap.used; - i++) - sectors_to_move += bucket_sectors_used(i->mark); + if (!c->copy_gc_enabled) { + move_buckets_wait(&ctxt, buckets, true); + kthread_wait_freezable(c->copy_gc_enabled); + } - while (sectors_to_move > COPYGC_SECTORS_PER_ITER(ca)) { - BUG_ON(!heap_pop(&ca->copygc_heap, e, -sectors_used_cmp)); - sectors_to_move -= bucket_sectors_used(e.mark); - } + if (unlikely(freezing(current))) { + move_buckets_wait(&ctxt, buckets, true); + __refrigerator(false); + continue; + } - buckets_to_move = ca->copygc_heap.used; + last = atomic64_read(&clock->now); + wait = bch2_copygc_wait_amount(c); - eytzinger0_sort(ca->copygc_heap.data, - ca->copygc_heap.used, - sizeof(ca->copygc_heap.data[0]), - bucket_idx_cmp, NULL); + if (wait > clock->max_slop) { + c->copygc_wait_at = last; + c->copygc_wait = last + wait; + move_buckets_wait(&ctxt, buckets, true); + trace_and_count(c, copygc_wait, c, wait, last + wait); + bch2_kthread_io_clock_wait(clock, last + wait, + MAX_SCHEDULE_TIMEOUT); + continue; + } - read_moving(ca, buckets_to_move, sectors_to_move); -} + c->copygc_wait = 0; -static int bch2_moving_gc_thread(void *arg) -{ - struct bch_dev *ca = arg; - struct bch_fs *c = ca->fs; - struct io_clock *clock = &c->io_clock[WRITE]; - unsigned long last; - u64 available, want, next; + c->copygc_running = true; + ret = bch2_copygc(&ctxt, buckets, &did_work); + c->copygc_running = false; - set_freezable(); + wake_up(&c->copygc_running_wq); - while (!kthread_should_stop()) { - if (kthread_wait_freezable(c->copy_gc_enabled)) - break; + if (!wait && !did_work) { + u64 min_member_capacity = bch2_min_rw_member_capacity(c); - last = atomic_long_read(&clock->now); - /* - * don't start copygc until less than half the gc reserve is - * available: - */ - available = dev_buckets_available(ca); - want = div64_u64((ca->mi.nbuckets - ca->mi.first_bucket) * - c->opts.gc_reserve_percent, 200); - if (available > want) { - next = last + (available - want) * - ca->mi.bucket_size; - bch2_kthread_io_clock_wait(clock, next); - continue; - } + if (min_member_capacity == U64_MAX) + min_member_capacity = 128 * 2048; - bch2_moving_gc(ca); + bch2_trans_unlock_long(ctxt.trans); + bch2_kthread_io_clock_wait(clock, last + (min_member_capacity >> 6), + MAX_SCHEDULE_TIMEOUT); + } } + move_buckets_wait(&ctxt, buckets, true); + + rhashtable_destroy(&buckets->table); + kfree(buckets); + bch2_moving_ctxt_exit(&ctxt); + bch2_move_stats_exit(&move_stats, c); + return 0; } -void bch2_moving_gc_stop(struct bch_dev *ca) +void bch2_copygc_stop(struct bch_fs *c) { - ca->moving_gc_pd.rate.rate = UINT_MAX; - bch2_ratelimit_reset(&ca->moving_gc_pd.rate); - - if (ca->moving_gc_read) - kthread_stop(ca->moving_gc_read); - ca->moving_gc_read = NULL; + if (c->copygc_thread) { + kthread_stop(c->copygc_thread); + put_task_struct(c->copygc_thread); + } + c->copygc_thread = NULL; } -int bch2_moving_gc_start(struct bch_dev *ca) +int bch2_copygc_start(struct bch_fs *c) { struct task_struct *t; + int ret; - BUG_ON(ca->moving_gc_read); + if (c->copygc_thread) + return 0; - if (ca->fs->opts.nochanges) + if (c->opts.nochanges) return 0; - if (bch2_fs_init_fault("moving_gc_start")) + if (bch2_fs_init_fault("copygc_start")) return -ENOMEM; - t = kthread_create(bch2_moving_gc_thread, ca, "bch_copygc_read"); - if (IS_ERR(t)) - return PTR_ERR(t); + t = kthread_create(bch2_copygc_thread, c, "bch-copygc/%s", c->name); + ret = PTR_ERR_OR_ZERO(t); + if (ret) { + bch_err_msg(c, ret, "creating copygc thread"); + return ret; + } + + get_task_struct(t); - ca->moving_gc_read = t; - wake_up_process(ca->moving_gc_read); + c->copygc_thread = t; + wake_up_process(c->copygc_thread); return 0; } -void bch2_dev_moving_gc_init(struct bch_dev *ca) +void bch2_fs_copygc_init(struct bch_fs *c) { - bch2_pd_controller_init(&ca->moving_gc_pd); - ca->moving_gc_pd.d_term = 0; + init_waitqueue_head(&c->copygc_running_wq); + c->copygc_running = false; }