+// SPDX-License-Identifier: GPL-2.0
/*
* Moving/copying garbage collector
*
*/
#include "bcachefs.h"
+#include "alloc_background.h"
+#include "alloc_foreground.h"
#include "btree_iter.h"
+#include "btree_update.h"
#include "buckets.h"
#include "clock.h"
+#include "disk_groups.h"
+#include "errcode.h"
+#include "error.h"
#include "extents.h"
#include "eytzinger.h"
#include "io.h"
#include <linux/freezer.h>
#include <linux/kthread.h>
#include <linux/math64.h>
+#include <linux/sched/task.h>
#include <linux/sort.h>
#include <linux/wait.h>
-/* Moving GC - IO loop */
-
-static int bucket_idx_cmp(const void *_l, const void *_r, size_t size)
+static inline int fragmentation_cmp(copygc_heap *heap,
+ struct copygc_heap_entry l,
+ struct copygc_heap_entry r)
{
- const struct bucket_heap_entry *l = _l;
- const struct bucket_heap_entry *r = _r;
-
- if (l->bucket < r->bucket)
- return -1;
- if (l->bucket > r->bucket)
- return 1;
- return 0;
+ return cmp_int(l.fragmentation, r.fragmentation);
}
-static const struct bch_extent_ptr *moving_pred(struct bch_dev *ca,
- struct bkey_s_c k)
+static int find_buckets_to_copygc(struct bch_fs *c)
{
- 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)
- };
+ copygc_heap *h = &c->copygc_heap;
+ struct btree_trans trans;
+ struct btree_iter iter;
+ struct bkey_s_c k;
+ int ret;
- size_t i = eytzinger0_find(h->data, h->used,
- sizeof(h->data[0]),
- bucket_idx_cmp, &search);
+ bch2_trans_init(&trans, c, 0, 0);
- if (i < h->used)
- return ptr;
- }
+ /*
+ * 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.
+ */
+ h->used = 0;
- return NULL;
-}
+ for_each_btree_key(&trans, iter, BTREE_ID_alloc, POS_MIN,
+ BTREE_ITER_PREFETCH, k, ret) {
+ struct bch_dev *ca = bch_dev_bkey_exists(c, iter.pos.inode);
+ struct copygc_heap_entry e;
+ struct bch_alloc_v4 a_convert;
+ const struct bch_alloc_v4 *a;
-static int issue_moving_gc_move(struct bch_dev *ca,
- struct moving_context *ctxt,
- struct bkey_s_c k)
-{
- struct bch_fs *c = ca->fs;
- const struct bch_extent_ptr *ptr;
- int ret;
+ a = bch2_alloc_to_v4(k, &a_convert);
- ptr = moving_pred(ca, k);
- if (!ptr) /* We raced - bucket's been reused */
- return 0;
+ if ((a->data_type != BCH_DATA_btree &&
+ a->data_type != BCH_DATA_user) ||
+ a->dirty_sectors >= ca->mi.bucket_size ||
+ bch2_bucket_is_open(c, iter.pos.inode, iter.pos.offset))
+ continue;
+
+ e = (struct copygc_heap_entry) {
+ .dev = iter.pos.inode,
+ .gen = a->gen,
+ .replicas = 1 + a->stripe_redundancy,
+ .fragmentation = div_u64((u64) a->dirty_sectors * (1ULL << 31),
+ ca->mi.bucket_size),
+ .sectors = a->dirty_sectors,
+ .bucket = iter.pos.offset,
+ };
+ heap_add_or_replace(h, e, -fragmentation_cmp, NULL);
+
+ }
+ bch2_trans_iter_exit(&trans, &iter);
- ret = bch2_data_move(c, ctxt, &ca->self, k, ptr);
- if (!ret)
- trace_gc_copy(k.k);
- else
- trace_moving_gc_alloc_fail(c, k.k->size);
+ bch2_trans_exit(&trans);
return ret;
}
-static void read_moving(struct bch_dev *ca, size_t buckets_to_move,
- u64 sectors_to_move)
+static int bch2_copygc(struct bch_fs *c)
{
- struct bch_fs *c = ca->fs;
- bucket_heap *h = &ca->copygc_heap;
+ copygc_heap *h = &c->copygc_heap;
+ struct copygc_heap_entry e;
+ struct bch_move_stats move_stats;
+ struct bch_dev *ca;
+ unsigned dev_idx;
+ size_t heap_size = 0;
struct moving_context ctxt;
- 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;
+ struct data_update_opts data_opts = {
+ .btree_insert_flags = BTREE_INSERT_USE_RESERVE|JOURNAL_WATERMARK_copygc,
+ };
+ int ret = 0;
- if (!moving_pred(ca, k))
- goto next;
+ bch2_move_stats_init(&move_stats, "copygc");
- if (issue_moving_gc_move(ca, &ctxt, k)) {
- bch2_btree_iter_unlock(&iter);
+ for_each_rw_member(ca, c, dev_idx)
+ heap_size += ca->mi.nbuckets >> 7;
- /* memory allocation failure, wait for some IO to finish */
- bch2_move_ctxt_wait_for_io(&ctxt);
- continue;
+ if (h->size < heap_size) {
+ free_heap(&c->copygc_heap);
+ if (!init_heap(&c->copygc_heap, heap_size, GFP_KERNEL)) {
+ bch_err(c, "error allocating copygc heap");
+ return 0;
}
-next:
- bch2_btree_iter_advance_pos(&iter);
- //bch2_btree_iter_cond_resched(&iter);
-
- /* unlock before calling moving_context_wait() */
- bch2_btree_iter_unlock(&iter);
- cond_resched();
}
- bch2_btree_iter_unlock(&iter);
- bch2_move_ctxt_exit(&ctxt);
- trace_moving_gc_end(ca, ctxt.sectors_moved, ctxt.keys_moved,
- buckets_to_move);
-
- /* 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);
-
- if (i->mark.gen == m.gen && bucket_sectors_used(m)) {
- sectors_not_moved += bucket_sectors_used(m);
- buckets_not_moved++;
- }
+ ret = find_buckets_to_copygc(c);
+ if (ret) {
+ bch2_fs_fatal_error(c, "error walking buckets to copygc!");
+ return ret;
}
- 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);
-}
+ if (!h->used) {
+ s64 wait = S64_MAX, dev_wait;
+ u64 dev_min_wait_fragmented = 0;
+ u64 dev_min_wait_allowed = 0;
+ int dev_min_wait = -1;
+
+ for_each_rw_member(ca, c, dev_idx) {
+ struct bch_dev_usage usage = bch2_dev_usage_read(ca);
+ s64 allowed = ((__dev_buckets_available(ca, usage, RESERVE_none) *
+ ca->mi.bucket_size) >> 1);
+ s64 fragmented = usage.d[BCH_DATA_user].fragmented;
+
+ dev_wait = max(0LL, allowed - fragmented);
+
+ if (dev_min_wait < 0 || dev_wait < wait) {
+ dev_min_wait = dev_idx;
+ dev_min_wait_fragmented = fragmented;
+ dev_min_wait_allowed = allowed;
+ }
+ }
-static bool have_copygc_reserve(struct bch_dev *ca)
-{
- bool ret;
+ bch_err_ratelimited(c, "copygc requested to run but found no buckets to move! dev %u fragmented %llu allowed %llu",
+ dev_min_wait, dev_min_wait_fragmented, dev_min_wait_allowed);
+ return 0;
+ }
- spin_lock(&ca->freelist_lock);
- ret = fifo_used(&ca->free[RESERVE_MOVINGGC]) >=
- COPYGC_BUCKETS_PER_ITER(ca);
- spin_unlock(&ca->freelist_lock);
+ heap_resort(h, fragmentation_cmp, NULL);
- return ret;
-}
+ bch2_moving_ctxt_init(&ctxt, c, NULL, &move_stats,
+ writepoint_ptr(&c->copygc_write_point),
+ false);
-static inline int sectors_used_cmp(bucket_heap *heap,
- struct bucket_heap_entry l,
- struct bucket_heap_entry r)
-{
- return bucket_sectors_used(l.mark) - bucket_sectors_used(r.mark);
-}
-
-static void bch2_moving_gc(struct bch_dev *ca)
-{
- 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);
+ /* not correct w.r.t. device removal */
+ while (h->used && !ret) {
+ BUG_ON(!heap_pop(h, e, -fragmentation_cmp, NULL));
+ ret = __bch2_evacuate_bucket(&ctxt, POS(e.dev, e.bucket), e.gen,
+ data_opts);
}
- 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.
- */
+ bch2_moving_ctxt_exit(&ctxt);
- /*
- * 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;
- }
+ if (ret < 0 && !bch2_err_matches(ret, EROFS))
+ bch_err(c, "error from bch2_move_data() in copygc: %s", bch2_err_str(ret));
- if (m.owned_by_allocator ||
- m.data_type != BUCKET_DATA)
- continue;
+ trace_and_count(c, copygc, c, atomic64_read(&move_stats.sectors_moved), 0, 0, 0);
+ return ret;
+}
- if (bucket_sectors_used(m) >= ca->mi.bucket_size)
- continue;
+/*
+ * 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)
+{
+ struct bch_dev *ca;
+ unsigned dev_idx;
+ s64 wait = S64_MAX, fragmented_allowed, fragmented;
- heap_add_or_replace(&ca->copygc_heap, e, -sectors_used_cmp);
- }
- up_read(&c->gc_lock);
+ for_each_rw_member(ca, c, dev_idx) {
+ struct bch_dev_usage usage = bch2_dev_usage_read(ca);
- for (i = ca->copygc_heap.data;
- i < ca->copygc_heap.data + ca->copygc_heap.used;
- i++)
- sectors_to_move += bucket_sectors_used(i->mark);
+ fragmented_allowed = ((__dev_buckets_available(ca, usage, RESERVE_none) *
+ ca->mi.bucket_size) >> 1);
+ fragmented = usage.d[BCH_DATA_user].fragmented;
- 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);
+ wait = min(wait, max(0LL, fragmented_allowed - fragmented));
}
- buckets_to_move = ca->copygc_heap.used;
-
- eytzinger0_sort(ca->copygc_heap.data,
- ca->copygc_heap.used,
- sizeof(ca->copygc_heap.data[0]),
- bucket_idx_cmp, NULL);
-
- read_moving(ca, buckets_to_move, sectors_to_move);
+ return wait;
}
-static int bch2_moving_gc_thread(void *arg)
+static int bch2_copygc_thread(void *arg)
{
- struct bch_dev *ca = arg;
- struct bch_fs *c = ca->fs;
+ struct bch_fs *c = arg;
struct io_clock *clock = &c->io_clock[WRITE];
- unsigned long last;
- u64 available, want, next;
+ u64 last, wait;
+ int ret = 0;
set_freezable();
- while (!kthread_should_stop()) {
+ while (!ret && !kthread_should_stop()) {
+ cond_resched();
+
if (kthread_wait_freezable(c->copy_gc_enabled))
break;
- 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);
+ last = atomic64_read(&clock->now);
+ wait = bch2_copygc_wait_amount(c);
+
+ if (wait > clock->max_slop) {
+ trace_and_count(c, copygc_wait, c, wait, last + wait);
+ c->copygc_wait = last + wait;
+ bch2_kthread_io_clock_wait(clock, last + wait,
+ MAX_SCHEDULE_TIMEOUT);
continue;
}
- bch2_moving_gc(ca);
+ c->copygc_wait = 0;
+
+ c->copygc_running = true;
+ ret = bch2_copygc(c);
+ c->copygc_running = false;
+
+ wake_up(&c->copygc_running_wq);
}
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(c, "error creating copygc thread: %s", bch2_err_str(ret));
+ 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;
}