u8 replicas;
u32 fragmentation;
u32 sectors;
- u64 offset;
+ u64 bucket;
};
typedef HEAP(struct copygc_heap_entry) copygc_heap;
// SPDX-License-Identifier: GPL-2.0
#include "bcachefs.h"
+#include "alloc_background.h"
#include "alloc_foreground.h"
+#include "backpointers.h"
#include "bkey_buf.h"
#include "btree_gc.h"
#include "btree_update.h"
#include "btree_update_interior.h"
+#include "btree_write_buffer.h"
#include "disk_groups.h"
#include "ec.h"
#include "errcode.h"
+#include "error.h"
#include "inode.h"
#include "io.h"
#include "journal_reclaim.h"
struct moving_io *io = container_of(op, struct moving_io, write.op);
struct moving_context *ctxt = io->write.ctxt;
+ if (io->write.op.error)
+ ctxt->write_error = true;
+
atomic_sub(io->write_sectors, &io->write.ctxt->write_sectors);
move_free(io);
closure_put(&ctxt->cl);
return 0;
}
+static int move_get_io_opts(struct btree_trans *trans,
+ struct bch_io_opts *io_opts,
+ struct bkey_s_c k, u64 *cur_inum)
+{
+ struct bch_inode_unpacked inode;
+ int ret;
+
+ if (*cur_inum == k.k->p.inode)
+ return 0;
+
+ ret = lookup_inode(trans,
+ SPOS(0, k.k->p.inode, k.k->p.snapshot),
+ &inode);
+ if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
+ return ret;
+
+ if (!ret)
+ bch2_inode_opts_get(io_opts, trans->c, &inode);
+ else
+ *io_opts = bch2_opts_to_inode_opts(trans->c->opts);
+ *cur_inum = k.k->p.inode;
+ return 0;
+}
+
static int __bch2_move_data(struct moving_context *ctxt,
struct bpos start,
struct bpos end,
if (!bkey_extent_is_direct_data(k.k))
goto next_nondata;
- if (btree_id == BTREE_ID_extents &&
- cur_inum != k.k->p.inode) {
- struct bch_inode_unpacked inode;
-
- io_opts = bch2_opts_to_inode_opts(c->opts);
-
- ret = lookup_inode(&trans,
- SPOS(0, k.k->p.inode, k.k->p.snapshot),
- &inode);
- if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
- continue;
-
- if (!ret)
- bch2_inode_opts_get(&io_opts, c, &inode);
-
- cur_inum = k.k->p.inode;
- }
+ ret = move_get_io_opts(&trans, &io_opts, k, &cur_inum);
+ if (ret)
+ continue;
memset(&data_opts, 0, sizeof(data_opts));
if (!pred(c, arg, k, &io_opts, &data_opts))
return ret;
}
+static int verify_bucket_evacuated(struct btree_trans *trans, struct bpos bucket, int gen)
+{
+ struct bch_fs *c = trans->c;
+ struct btree_iter iter;
+ struct bkey_s_c k;
+ struct printbuf buf = PRINTBUF;
+ struct bch_backpointer bp;
+ u64 bp_offset = 0;
+ int ret;
+
+ bch2_trans_iter_init(trans, &iter, BTREE_ID_alloc,
+ bucket, BTREE_ITER_CACHED);
+again:
+ k = bch2_btree_iter_peek_slot(&iter);
+ ret = bkey_err(k);
+
+ if (!ret && k.k->type == KEY_TYPE_alloc_v4) {
+ struct bkey_s_c_alloc_v4 a = bkey_s_c_to_alloc_v4(k);
+
+ if (a.v->gen == gen &&
+ a.v->dirty_sectors) {
+ if (a.v->data_type == BCH_DATA_btree) {
+ bch2_trans_unlock(trans);
+ if (bch2_btree_interior_updates_flush(c))
+ goto again;
+ goto failed_to_evacuate;
+ }
+ }
+ }
+
+ bch2_trans_iter_exit(trans, &iter);
+ return ret;
+failed_to_evacuate:
+ bch2_trans_iter_exit(trans, &iter);
+
+ prt_printf(&buf, bch2_log_msg(c, "failed to evacuate bucket "));
+ bch2_bkey_val_to_text(&buf, c, k);
+
+ while (1) {
+ bch2_trans_begin(trans);
+
+ ret = bch2_get_next_backpointer(trans, bucket, gen,
+ &bp_offset, &bp);
+ if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
+ continue;
+ if (ret)
+ break;
+ if (bp_offset == U64_MAX)
+ break;
+
+ k = bch2_backpointer_get_key(trans, &iter,
+ bucket, bp_offset, bp);
+ ret = bkey_err(k);
+ if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
+ continue;
+ if (ret)
+ break;
+ if (!k.k)
+ continue;
+ prt_newline(&buf);
+ bch2_bkey_val_to_text(&buf, c, k);
+ bch2_trans_iter_exit(trans, &iter);
+ }
+
+ bch2_print_string_as_lines(KERN_ERR, buf.buf);
+ printbuf_exit(&buf);
+ return 0;
+}
+
+int __bch2_evacuate_bucket(struct moving_context *ctxt,
+ struct bpos bucket, int gen,
+ struct data_update_opts _data_opts)
+{
+ struct bch_fs *c = ctxt->c;
+ struct bch_io_opts io_opts = bch2_opts_to_inode_opts(c->opts);
+ struct btree_trans trans;
+ struct btree_iter iter;
+ struct bkey_buf sk;
+ struct bch_backpointer bp;
+ struct bch_alloc_v4 a_convert;
+ const struct bch_alloc_v4 *a;
+ struct bkey_s_c k;
+ struct data_update_opts data_opts;
+ unsigned dirty_sectors, bucket_size;
+ u64 bp_offset = 0, cur_inum = U64_MAX;
+ int ret = 0;
+
+ bch2_bkey_buf_init(&sk);
+ bch2_trans_init(&trans, c, 0, 0);
+
+ bch2_trans_iter_init(&trans, &iter, BTREE_ID_alloc,
+ bucket, BTREE_ITER_CACHED);
+ ret = lockrestart_do(&trans,
+ bkey_err(k = bch2_btree_iter_peek_slot(&iter)));
+ bch2_trans_iter_exit(&trans, &iter);
+
+ if (ret) {
+ bch_err(c, "%s: error looking up alloc key: %s", __func__, bch2_err_str(ret));
+ goto err;
+ }
+
+ a = bch2_alloc_to_v4(k, &a_convert);
+ dirty_sectors = a->dirty_sectors;
+ bucket_size = bch_dev_bkey_exists(c, bucket.inode)->mi.bucket_size;
+
+ ret = bch2_btree_write_buffer_flush(&trans);
+ if (ret) {
+ bch_err(c, "%s: error flushing btree write buffer: %s", __func__, bch2_err_str(ret));
+ goto err;
+ }
+
+ while (!(ret = move_ratelimit(&trans, ctxt))) {
+ bch2_trans_begin(&trans);
+
+ ret = bch2_get_next_backpointer(&trans, bucket, gen,
+ &bp_offset, &bp);
+ if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
+ continue;
+ if (ret)
+ goto err;
+ if (bp_offset == U64_MAX)
+ break;
+
+ if (!bp.level) {
+ const struct bch_extent_ptr *ptr;
+ struct bkey_s_c k;
+ unsigned i = 0;
+
+ k = bch2_backpointer_get_key(&trans, &iter,
+ bucket, bp_offset, bp);
+ ret = bkey_err(k);
+ if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
+ continue;
+ if (ret)
+ goto err;
+ if (!k.k)
+ goto next;
+
+ bch2_bkey_buf_reassemble(&sk, c, k);
+ k = bkey_i_to_s_c(sk.k);
+
+ ret = move_get_io_opts(&trans, &io_opts, k, &cur_inum);
+ if (ret) {
+ bch2_trans_iter_exit(&trans, &iter);
+ continue;
+ }
+
+ data_opts = _data_opts;
+ data_opts.target = io_opts.background_target;
+ data_opts.rewrite_ptrs = 0;
+
+ bkey_for_each_ptr(bch2_bkey_ptrs_c(k), ptr) {
+ if (ptr->dev == bucket.inode)
+ data_opts.rewrite_ptrs |= 1U << i;
+ i++;
+ }
+
+ ret = bch2_move_extent(&trans, &iter, ctxt, io_opts,
+ bp.btree_id, k, data_opts);
+ bch2_trans_iter_exit(&trans, &iter);
+
+ if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
+ continue;
+ if (ret == -ENOMEM) {
+ /* memory allocation failure, wait for some IO to finish */
+ bch2_move_ctxt_wait_for_io(ctxt, &trans);
+ continue;
+ }
+ if (ret)
+ goto err;
+
+ if (ctxt->rate)
+ bch2_ratelimit_increment(ctxt->rate, k.k->size);
+ atomic64_add(k.k->size, &ctxt->stats->sectors_seen);
+ } else {
+ struct btree *b;
+
+ b = bch2_backpointer_get_node(&trans, &iter,
+ bucket, bp_offset, bp);
+ ret = PTR_ERR_OR_ZERO(b);
+ if (ret == -BCH_ERR_backpointer_to_overwritten_btree_node)
+ continue;
+ if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
+ continue;
+ if (ret)
+ goto err;
+ if (!b)
+ goto next;
+
+ ret = bch2_btree_node_rewrite(&trans, &iter, b, 0);
+ bch2_trans_iter_exit(&trans, &iter);
+
+ if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
+ continue;
+ if (ret)
+ goto err;
+
+ if (ctxt->rate)
+ bch2_ratelimit_increment(ctxt->rate,
+ c->opts.btree_node_size >> 9);
+ atomic64_add(c->opts.btree_node_size >> 9, &ctxt->stats->sectors_seen);
+ atomic64_add(c->opts.btree_node_size >> 9, &ctxt->stats->sectors_moved);
+ }
+next:
+ bp_offset++;
+ }
+
+ trace_evacuate_bucket(c, &bucket, dirty_sectors, bucket_size, ret);
+
+ if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG) && gen >= 0) {
+ bch2_trans_unlock(&trans);
+ move_ctxt_wait_event(ctxt, NULL, list_empty(&ctxt->reads));
+ closure_sync(&ctxt->cl);
+ if (!ctxt->write_error)
+ lockrestart_do(&trans, verify_bucket_evacuated(&trans, bucket, gen));
+ }
+err:
+ bch2_trans_exit(&trans);
+ bch2_bkey_buf_exit(&sk, c);
+ return ret;
+}
+
+int bch2_evacuate_bucket(struct bch_fs *c,
+ struct bpos bucket, int gen,
+ struct data_update_opts data_opts,
+ struct bch_ratelimit *rate,
+ struct bch_move_stats *stats,
+ struct write_point_specifier wp,
+ bool wait_on_copygc)
+{
+ struct moving_context ctxt;
+ int ret;
+
+ bch2_moving_ctxt_init(&ctxt, c, rate, stats, wp, wait_on_copygc);
+ ret = __bch2_evacuate_bucket(&ctxt, bucket, gen, data_opts);
+ bch2_moving_ctxt_exit(&ctxt);
+
+ return ret;
+}
+
typedef bool (*move_btree_pred)(struct bch_fs *, void *,
struct btree *, struct bch_io_opts *,
struct data_update_opts *);
struct bch_move_stats *stats;
struct write_point_specifier wp;
bool wait_on_copygc;
+ bool write_error;
/* For waiting on outstanding reads and writes: */
struct closure cl;
bool,
move_pred_fn, void *);
+int __bch2_evacuate_bucket(struct moving_context *,
+ struct bpos, int,
+ struct data_update_opts);
+int bch2_evacuate_bucket(struct bch_fs *, struct bpos, int,
+ struct data_update_opts,
+ struct bch_ratelimit *,
+ struct bch_move_stats *,
+ struct write_point_specifier,
+ bool);
int bch2_data_job(struct bch_fs *,
struct bch_move_stats *,
struct bch_ioctl_data);
#include <linux/sort.h>
#include <linux/wait.h>
-static int bucket_offset_cmp(const void *_l, const void *_r, size_t size)
-{
- const struct copygc_heap_entry *l = _l;
- const struct copygc_heap_entry *r = _r;
-
- return cmp_int(l->dev, r->dev) ?:
- cmp_int(l->offset, r->offset);
-}
-
-static bool copygc_pred(struct bch_fs *c, void *arg,
- struct bkey_s_c k,
- struct bch_io_opts *io_opts,
- struct data_update_opts *data_opts)
-{
- 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 };
- unsigned i = 0;
-
- /*
- * 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->rewrite_ptrs = 0;
- data_opts->target = io_opts->background_target;
- data_opts->extra_replicas = 0;
- data_opts->btree_insert_flags = BTREE_INSERT_USE_RESERVE|
- JOURNAL_WATERMARK_copygc;
-
- 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 eytz;
-
- if (p.ptr.cached)
- continue;
-
- eytz = 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 (eytz >= 0 &&
- p.ptr.dev == h->data[eytz].dev &&
- p.ptr.offset < h->data[eytz].offset + ca->mi.bucket_size &&
- p.ptr.gen == h->data[eytz].gen)
- data_opts->rewrite_ptrs |= 1U << i;
-
- i++;
- }
-
- return data_opts->rewrite_ptrs != 0;
-}
-
static inline int fragmentation_cmp(copygc_heap *heap,
struct copygc_heap_entry l,
struct copygc_heap_entry r)
return cmp_int(l.fragmentation, r.fragmentation);
}
-static int walk_buckets_to_copygc(struct bch_fs *c)
+static int find_buckets_to_copygc(struct bch_fs *c)
{
copygc_heap *h = &c->copygc_heap;
struct btree_trans trans;
bch2_trans_init(&trans, c, 0, 0);
+ /*
+ * 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;
+
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);
a = bch2_alloc_to_v4(k, &a_convert);
- if (a->data_type != BCH_DATA_user ||
+ 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;
.fragmentation = div_u64((u64) a->dirty_sectors * (1ULL << 31),
ca->mi.bucket_size),
.sectors = a->dirty_sectors,
- .offset = bucket_to_sector(ca, iter.pos.offset),
+ .bucket = iter.pos.offset,
};
heap_add_or_replace(h, e, -fragmentation_cmp, NULL);
return ret;
}
-static int bucket_inorder_cmp(const void *_l, const void *_r)
-{
- const struct copygc_heap_entry *l = _l;
- const struct copygc_heap_entry *r = _r;
-
- return cmp_int(l->dev, r->dev) ?: cmp_int(l->offset, r->offset);
-}
-
-static int check_copygc_was_done(struct bch_fs *c,
- u64 *sectors_not_moved,
- u64 *buckets_not_moved)
-{
- copygc_heap *h = &c->copygc_heap;
- struct btree_trans trans;
- struct btree_iter iter;
- struct bkey_s_c k;
- struct bch_alloc_v4 a;
- struct copygc_heap_entry *i;
- int ret = 0;
-
- sort(h->data, h->used, sizeof(h->data[0]), bucket_inorder_cmp, NULL);
-
- bch2_trans_init(&trans, c, 0, 0);
- bch2_trans_iter_init(&trans, &iter, BTREE_ID_alloc, POS_MIN, 0);
-
- for (i = h->data; i < h->data + h->used; i++) {
- struct bch_dev *ca = bch_dev_bkey_exists(c, i->dev);
-
- bch2_btree_iter_set_pos(&iter, POS(i->dev, sector_to_bucket(ca, i->offset)));
-
- ret = lockrestart_do(&trans,
- bkey_err(k = bch2_btree_iter_peek_slot(&iter)));
- if (ret)
- break;
-
- bch2_alloc_to_v4(k, &a);
-
- if (a.gen == i->gen && a.dirty_sectors) {
- *sectors_not_moved += a.dirty_sectors;
- *buckets_not_moved += 1;
- }
- }
- bch2_trans_iter_exit(&trans, &iter);
-
- bch2_trans_exit(&trans);
- return ret;
-}
-
static int bch2_copygc(struct bch_fs *c)
{
copygc_heap *h = &c->copygc_heap;
- struct copygc_heap_entry e, *i;
+ struct copygc_heap_entry e;
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 heap_size = 0;
- int ret;
+ struct moving_context ctxt;
+ struct data_update_opts data_opts = {
+ .btree_insert_flags = BTREE_INSERT_USE_RESERVE|JOURNAL_WATERMARK_copygc,
+ };
+ int ret = 0;
bch2_move_stats_init(&move_stats, "copygc");
- /*
- * 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;
-
for_each_rw_member(ca, c, dev_idx)
heap_size += ca->mi.nbuckets >> 7;
}
}
- for_each_rw_member(ca, c, dev_idx) {
- struct bch_dev_usage usage = bch2_dev_usage_read(ca);
-
- u64 avail = max_t(s64, 0,
- usage.d[BCH_DATA_free].buckets +
- usage.d[BCH_DATA_need_discard].buckets -
- ca->nr_open_buckets -
- bch2_dev_buckets_reserved(ca, RESERVE_movinggc));
-
- avail = min(avail, ca->mi.nbuckets >> 6);
-
- sectors_reserved += avail * ca->mi.bucket_size;
- }
-
- ret = walk_buckets_to_copygc(c);
+ ret = find_buckets_to_copygc(c);
if (ret) {
bch2_fs_fatal_error(c, "error walking buckets to copygc!");
return ret;
return 0;
}
- /*
- * Our btree node allocations also come out of RESERVE_movingc:
- */
- sectors_reserved = (sectors_reserved * 3) / 4;
- if (!sectors_reserved) {
- bch2_fs_fatal_error(c, "stuck, ran out of copygc reserve!");
- return -1;
- }
+ heap_resort(h, fragmentation_cmp, NULL);
- for (i = h->data; i < h->data + h->used; i++) {
- sectors_to_move += i->sectors;
- sectors_to_write += i->sectors * i->replicas;
- }
+ bch2_moving_ctxt_init(&ctxt, c, NULL, &move_stats,
+ writepoint_ptr(&c->copygc_write_point),
+ false);
- while (sectors_to_write > sectors_reserved) {
+ /* not correct w.r.t. device removal */
+ while (h->used && !ret) {
BUG_ON(!heap_pop(h, e, -fragmentation_cmp, NULL));
- sectors_to_write -= e.sectors * e.replicas;
+ ret = __bch2_evacuate_bucket(&ctxt, POS(e.dev, e.bucket), e.gen,
+ data_opts);
}
- buckets_to_move = h->used;
+ bch2_moving_ctxt_exit(&ctxt);
- if (!buckets_to_move) {
- bch_err_ratelimited(c, "copygc cannot run - sectors_reserved %llu!",
- sectors_reserved);
- 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,
- &move_stats,
- writepoint_ptr(&c->copygc_write_point),
- false,
- copygc_pred, NULL);
if (ret < 0 && !bch2_err_matches(ret, EROFS))
bch_err(c, "error from bch2_move_data() in copygc: %s", bch2_err_str(ret));
- if (ret)
- return ret;
-
- ret = check_copygc_was_done(c, §ors_not_moved, &buckets_not_moved);
- if (ret) {
- bch_err(c, "error %i from check_copygc_was_done()", ret);
- return ret;
- }
- if (sectors_not_moved)
- 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_and_count(c, copygc, c,
- atomic64_read(&move_stats.sectors_moved), sectors_not_moved,
- buckets_to_move, buckets_not_moved);
- return 0;
+ trace_and_count(c, copygc, c, atomic64_read(&move_stats.sectors_moved), 0, 0, 0);
+ return ret;
}
/*
__entry->sectors_moved, __entry->keys_moved)
);
+TRACE_EVENT(evacuate_bucket,
+ TP_PROTO(struct bch_fs *c, struct bpos *bucket,
+ unsigned sectors, unsigned bucket_size,
+ int ret),
+ TP_ARGS(c, bucket, sectors, bucket_size, ret),
+
+ TP_STRUCT__entry(
+ __field(dev_t, dev )
+ __field(u64, member )
+ __field(u64, bucket )
+ __field(u32, sectors )
+ __field(u32, bucket_size )
+ __field(int, ret )
+ ),
+
+ TP_fast_assign(
+ __entry->dev = c->dev;
+ __entry->member = bucket->inode;
+ __entry->bucket = bucket->offset;
+ __entry->sectors = sectors;
+ __entry->bucket_size = bucket_size;
+ __entry->ret = ret;
+ ),
+
+ TP_printk("%d,%d %llu:%llu sectors %u/%u ret %i",
+ MAJOR(__entry->dev), MINOR(__entry->dev),
+ __entry->member, __entry->bucket,
+ __entry->sectors, __entry->bucket_size,
+ __entry->ret)
+);
+
TRACE_EVENT(copygc,
TP_PROTO(struct bch_fs *c,
u64 sectors_moved, u64 sectors_not_moved,
projects / linux-block.git / commitdiff