+// SPDX-License-Identifier: GPL-2.0
/*
* Moving/copying garbage collector
*
*/
#include "bcachefs.h"
+#include "alloc_foreground.h"
#include "btree_iter.h"
+#include "btree_update.h"
#include "buckets.h"
#include "clock.h"
+#include "disk_groups.h"
+#include "error.h"
#include "extents.h"
+#include "eytzinger.h"
#include "io.h"
#include "keylist.h"
#include "move.h"
#include "movinggc.h"
+#include "super-io.h"
#include <trace/events/bcachefs.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 */
+/*
+ * We can't use the entire copygc reserve in one iteration of copygc: we may
+ * need the buckets we're freeing up to go back into the copygc reserve to make
+ * forward progress, but if the copygc reserve is full they'll be available for
+ * any allocation - and it's possible that in a given iteration, we free up most
+ * of the buckets we're going to free before we allocate most of the buckets
+ * we're going to allocate.
+ *
+ * If we only use half of the reserve per iteration, then in steady state we'll
+ * always have room in the reserve for the buckets we're going to need in the
+ * next iteration:
+ */
+#define COPYGC_BUCKETS_PER_ITER(ca) \
+ ((ca)->free[RESERVE_MOVINGGC].size / 2)
-static const struct bch_extent_ptr *moving_pred(struct bch_dev *ca,
- struct bkey_s_c k)
+static int bucket_offset_cmp(const void *_l, const void *_r, size_t size)
{
- const struct bch_extent_ptr *ptr;
+ const struct copygc_heap_entry *l = _l;
+ const struct copygc_heap_entry *r = _r;
- if (bkey_extent_is_data(k.k) &&
- (ptr = bch2_extent_has_device(bkey_s_c_to_extent(k),
- ca->dev_idx)) &&
- PTR_BUCKET(ca, ptr)->mark.copygc)
- return ptr;
-
- return NULL;
+ return cmp_int(l->dev, r->dev) ?:
+ cmp_int(l->offset, r->offset);
}
-static int issue_moving_gc_move(struct bch_dev *ca,
- struct moving_context *ctxt,
- struct bkey_s_c k)
+static enum data_cmd copygc_pred(struct bch_fs *c, void *arg,
+ struct bkey_s_c k,
+ struct bch_io_opts *io_opts,
+ struct data_opts *data_opts)
{
- struct bch_fs *c = ca->fs;
- const struct bch_extent_ptr *ptr;
- int ret;
-
- ptr = moving_pred(ca, k);
- if (!ptr) /* We raced - bucket's been reused */
- 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);
- return ret;
-}
-
-static void read_moving(struct bch_dev *ca, size_t buckets_to_move,
- u64 sectors_to_move)
-{
- struct bch_fs *c = ca->fs;
- struct bucket *g;
- struct moving_context ctxt;
- struct btree_iter iter;
- struct bkey_s_c k;
- u64 sectors_not_moved = 0;
- size_t buckets_not_moved = 0;
-
- 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);
-
- 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;
-
- if (!moving_pred(ca, k))
- goto next;
-
- if (issue_moving_gc_move(ca, &ctxt, k)) {
- bch2_btree_iter_unlock(&iter);
-
- /* memory allocation failure, wait for some IO to finish */
- bch2_move_ctxt_wait_for_io(&ctxt);
- continue;
+ copygc_heap *h = &c->copygc_heap;
+ struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
+ const union bch_extent_entry *entry;
+ struct extent_ptr_decoded p = { 0 };
+
+ bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
+ struct bch_dev *ca = bch_dev_bkey_exists(c, p.ptr.dev);
+ struct copygc_heap_entry search = {
+ .dev = p.ptr.dev,
+ .offset = p.ptr.offset,
+ };
+
+ ssize_t i = eytzinger0_find_le(h->data, h->used,
+ sizeof(h->data[0]),
+ bucket_offset_cmp, &search);
+#if 0
+ /* eytzinger search verify code: */
+ ssize_t j = -1, k;
+
+ for (k = 0; k < h->used; k++)
+ if (h->data[k].offset <= ptr->offset &&
+ (j < 0 || h->data[k].offset > h->data[j].offset))
+ j = k;
+
+ BUG_ON(i != j);
+#endif
+ if (i >= 0 &&
+ p.ptr.dev == h->data[i].dev &&
+ p.ptr.offset < h->data[i].offset + ca->mi.bucket_size &&
+ p.ptr.gen == h->data[i].gen) {
+ /*
+ * We need to use the journal reserve here, because
+ * - journal reclaim depends on btree key cache
+ * flushing to make forward progress,
+ * - which has to make forward progress when the
+ * journal is pre-reservation full,
+ * - and depends on allocation - meaning allocator and
+ * copygc
+ */
+
+ data_opts->target = io_opts->background_target;
+ data_opts->nr_replicas = 1;
+ data_opts->btree_insert_flags = BTREE_INSERT_USE_RESERVE|
+ BTREE_INSERT_JOURNAL_RESERVED;
+ data_opts->rewrite_dev = p.ptr.dev;
+
+ if (p.has_ec)
+ data_opts->nr_replicas += p.ec.redundancy;
+
+ return DATA_REWRITE;
}
-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_each_bucket(g, ca)
- if (g->mark.copygc && bucket_sectors_used(g)) {
- sectors_not_moved += bucket_sectors_used(g);
- buckets_not_moved++;
- }
-
- 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);
+ return DATA_SKIP;
}
static bool have_copygc_reserve(struct bch_dev *ca)
{
bool ret;
- spin_lock(&ca->freelist_lock);
- ret = fifo_used(&ca->free[RESERVE_MOVINGGC]) >=
- COPYGC_BUCKETS_PER_ITER(ca);
- spin_unlock(&ca->freelist_lock);
+ spin_lock(&ca->fs->freelist_lock);
+ ret = fifo_full(&ca->free[RESERVE_MOVINGGC]) ||
+ ca->allocator_state != ALLOCATOR_running;
+ spin_unlock(&ca->fs->freelist_lock);
return ret;
}
-static void bch2_moving_gc(struct bch_dev *ca)
+static inline int fragmentation_cmp(copygc_heap *heap,
+ struct copygc_heap_entry l,
+ struct copygc_heap_entry r)
{
- struct bch_fs *c = ca->fs;
- struct bucket *g;
- struct bucket_mark new;
- u64 sectors_to_move;
- size_t buckets_to_move, buckets_unused = 0;
- struct bucket_heap_entry e;
- unsigned sectors_used, 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);
- }
+ return cmp_int(l.fragmentation, r.fragmentation);
+}
- reserve_sectors = COPYGC_SECTORS_PER_ITER(ca);
+static int bch2_copygc(struct bch_fs *c)
+{
+ copygc_heap *h = &c->copygc_heap;
+ struct copygc_heap_entry e, *i;
+ struct bucket_array *buckets;
+ struct bch_move_stats move_stats;
+ u64 sectors_to_move = 0, sectors_to_write = 0, sectors_not_moved = 0;
+ u64 sectors_reserved = 0;
+ u64 buckets_to_move, buckets_not_moved = 0;
+ struct bch_dev *ca;
+ unsigned dev_idx;
+ size_t b, heap_size = 0;
+ int ret;
- trace_moving_gc_start(ca);
+ bch_move_stats_init(&move_stats, "copygc");
/*
* Find buckets with lowest sector counts, skipping completely
* and repeatedly replacing the maximum element until all
* buckets have been visited.
*/
+ h->used = 0;
- /*
- * We need bucket marks to be up to date, so gc can't be recalculating
- * them, and we don't want the allocator invalidating a bucket after
- * we've decided to evacuate it but before we set copygc:
- */
- down_read(&c->gc_lock);
- mutex_lock(&ca->heap_lock);
- mutex_lock(&ca->fs->bucket_lock);
-
- ca->heap.used = 0;
- for_each_bucket(g, ca) {
- bucket_cmpxchg(g, new, new.copygc = 0);
+ for_each_rw_member(ca, c, dev_idx)
+ heap_size += ca->mi.nbuckets >> 7;
- if (bucket_unused(g)) {
- buckets_unused++;
- 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;
}
+ }
- if (g->mark.owned_by_allocator ||
- g->mark.data_type != BUCKET_DATA)
- continue;
+ for_each_rw_member(ca, c, dev_idx) {
+ closure_wait_event(&c->freelist_wait, have_copygc_reserve(ca));
+
+ spin_lock(&ca->fs->freelist_lock);
+ sectors_reserved += fifo_used(&ca->free[RESERVE_MOVINGGC]) * ca->mi.bucket_size;
+ spin_unlock(&ca->fs->freelist_lock);
+
+ down_read(&ca->bucket_lock);
+ buckets = bucket_array(ca);
+
+ for (b = buckets->first_bucket; b < buckets->nbuckets; b++) {
+ struct bucket *g = buckets->b + b;
+ struct bucket_mark m = READ_ONCE(g->mark);
+ struct copygc_heap_entry e;
+
+ if (m.owned_by_allocator ||
+ m.data_type != BCH_DATA_user ||
+ !bucket_sectors_used(m) ||
+ bucket_sectors_used(m) >= ca->mi.bucket_size)
+ continue;
+
+ WARN_ON(m.stripe && !g->stripe_redundancy);
+
+ e = (struct copygc_heap_entry) {
+ .dev = dev_idx,
+ .gen = m.gen,
+ .replicas = 1 + g->stripe_redundancy,
+ .fragmentation = bucket_sectors_used(m) * (1U << 15)
+ / ca->mi.bucket_size,
+ .sectors = bucket_sectors_used(m),
+ .offset = bucket_to_sector(ca, b),
+ };
+ heap_add_or_replace(h, e, -fragmentation_cmp, NULL);
+ }
+ up_read(&ca->bucket_lock);
+ }
- sectors_used = bucket_sectors_used(g);
+ /*
+ * Our btree node allocations also come out of RESERVE_MOVINGGC:
+ */
+ sectors_reserved = (sectors_reserved * 3) / 4;
+ if (!sectors_reserved) {
+ bch2_fs_fatal_error(c, "stuck, ran out of copygc reserve!");
+ return -1;
+ }
- if (sectors_used >= ca->mi.bucket_size)
- continue;
+ for (i = h->data; i < h->data + h->used; i++) {
+ sectors_to_move += i->sectors;
+ sectors_to_write += i->sectors * i->replicas;
+ }
- bucket_heap_push(ca, g, sectors_used);
+ while (sectors_to_write > sectors_reserved) {
+ BUG_ON(!heap_pop(h, e, -fragmentation_cmp, NULL));
+ sectors_to_write -= e.sectors * e.replicas;
}
- sectors_to_move = 0;
- for (i = 0; i < ca->heap.used; i++)
- sectors_to_move += ca->heap.data[i].val;
+ buckets_to_move = h->used;
- while (sectors_to_move > COPYGC_SECTORS_PER_ITER(ca)) {
- BUG_ON(!heap_pop(&ca->heap, e, bucket_min_cmp));
- sectors_to_move -= e.val;
+ if (!buckets_to_move)
+ return 0;
+
+ eytzinger0_sort(h->data, h->used,
+ sizeof(h->data[0]),
+ bucket_offset_cmp, NULL);
+
+ ret = bch2_move_data(c,
+ 0, POS_MIN,
+ BTREE_ID_NR, POS_MAX,
+ NULL,
+ writepoint_ptr(&c->copygc_write_point),
+ copygc_pred, NULL,
+ &move_stats);
+
+ for_each_rw_member(ca, c, dev_idx) {
+ down_read(&ca->bucket_lock);
+ buckets = bucket_array(ca);
+ for (i = h->data; i < h->data + h->used; i++) {
+ struct bucket_mark m;
+ size_t b;
+
+ if (i->dev != dev_idx)
+ continue;
+
+ b = sector_to_bucket(ca, i->offset);
+ m = READ_ONCE(buckets->b[b].mark);
+
+ if (i->gen == m.gen &&
+ bucket_sectors_used(m)) {
+ sectors_not_moved += bucket_sectors_used(m);
+ buckets_not_moved++;
+ }
+ }
+ up_read(&ca->bucket_lock);
}
- for (i = 0; i < ca->heap.used; i++)
- bucket_cmpxchg(ca->heap.data[i].g, new, new.copygc = 1);
+ if (sectors_not_moved && !ret)
+ bch_warn_ratelimited(c,
+ "copygc finished but %llu/%llu sectors, %llu/%llu buckets not moved (move stats: moved %llu sectors, raced %llu keys, %llu sectors)",
+ sectors_not_moved, sectors_to_move,
+ buckets_not_moved, buckets_to_move,
+ atomic64_read(&move_stats.sectors_moved),
+ atomic64_read(&move_stats.keys_raced),
+ atomic64_read(&move_stats.sectors_raced));
+
+ trace_copygc(c,
+ atomic64_read(&move_stats.sectors_moved), sectors_not_moved,
+ buckets_to_move, buckets_not_moved);
+ return 0;
+}
+
+/*
+ * 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;
+
+ for_each_rw_member(ca, c, dev_idx) {
+ struct bch_dev_usage usage = bch2_dev_usage_read(ca);
- buckets_to_move = ca->heap.used;
+ fragmented_allowed = ((__dev_buckets_reclaimable(ca, usage) *
+ ca->mi.bucket_size) >> 1);
+ fragmented = usage.d[BCH_DATA_user].fragmented;
- mutex_unlock(&ca->fs->bucket_lock);
- mutex_unlock(&ca->heap_lock);
- up_read(&c->gc_lock);
+ wait = min(wait, max(0LL, fragmented_allowed - fragmented));
+ }
- 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;
set_freezable();
while (!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_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;
+
+ if (bch2_copygc(c))
+ break;
}
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;
- 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))
+ t = kthread_create(bch2_copygc_thread, c, "bch-copygc/%s", c->name);
+ if (IS_ERR(t)) {
+ bch_err(c, "error creating copygc thread: %li", PTR_ERR(t));
return PTR_ERR(t);
+ }
+
+ 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;
}