// 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"
struct bio_vec bi_inline_vecs[0];
};
-static void move_free(struct closure *cl)
+static void move_free(struct moving_io *io)
{
- struct moving_io *io = container_of(cl, struct moving_io, cl);
struct moving_context *ctxt = io->write.ctxt;
struct bch_fs *c = ctxt->c;
bch2_data_update_exit(&io->write);
wake_up(&ctxt->wait);
- percpu_ref_put(&c->writes);
+ bch2_write_ref_put(c, BCH_WRITE_REF_move);
kfree(io);
}
-static void move_write_done(struct closure *cl)
+static void move_write_done(struct bch_write_op *op)
{
- struct moving_io *io = container_of(cl, struct moving_io, cl);
+ 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);
- closure_return_with_destructor(cl, move_free);
+ atomic_dec(&io->write.ctxt->write_ios);
+ move_free(io);
+ closure_put(&ctxt->cl);
}
-static void move_write(struct closure *cl)
+static void move_write(struct moving_io *io)
{
- struct moving_io *io = container_of(cl, struct moving_io, cl);
-
if (unlikely(io->rbio.bio.bi_status || io->rbio.hole)) {
- closure_return_with_destructor(cl, move_free);
+ move_free(io);
return;
}
+ closure_get(&io->write.ctxt->cl);
atomic_add(io->write_sectors, &io->write.ctxt->write_sectors);
+ atomic_inc(&io->write.ctxt->write_ios);
- bch2_data_update_read_done(&io->write, io->rbio.pick.crc, cl);
- continue_at(cl, move_write_done, NULL);
+ bch2_data_update_read_done(&io->write, io->rbio.pick.crc);
}
-static inline struct moving_io *next_pending_write(struct moving_context *ctxt)
+struct moving_io *bch2_moving_ctxt_next_pending_write(struct moving_context *ctxt)
{
struct moving_io *io =
list_first_entry_or_null(&ctxt->reads, struct moving_io, list);
struct moving_context *ctxt = io->write.ctxt;
atomic_sub(io->read_sectors, &ctxt->read_sectors);
+ atomic_dec(&ctxt->read_ios);
io->read_completed = true;
wake_up(&ctxt->wait);
closure_put(&ctxt->cl);
}
-static void do_pending_writes(struct moving_context *ctxt, struct btree_trans *trans)
+void bch2_moving_ctxt_do_pending_writes(struct moving_context *ctxt,
+ struct btree_trans *trans)
{
struct moving_io *io;
if (trans)
bch2_trans_unlock(trans);
- while ((io = next_pending_write(ctxt))) {
+ while ((io = bch2_moving_ctxt_next_pending_write(ctxt))) {
list_del(&io->list);
- closure_call(&io->cl, move_write, NULL, &ctxt->cl);
+ move_write(io);
}
}
-#define move_ctxt_wait_event(_ctxt, _trans, _cond) \
-do { \
- do_pending_writes(_ctxt, _trans); \
- \
- if (_cond) \
- break; \
- __wait_event((_ctxt)->wait, \
- next_pending_write(_ctxt) || (_cond)); \
-} while (1)
-
static void bch2_move_ctxt_wait_for_io(struct moving_context *ctxt,
struct btree_trans *trans)
{
{
move_ctxt_wait_event(ctxt, NULL, list_empty(&ctxt->reads));
closure_sync(&ctxt->cl);
+
EBUG_ON(atomic_read(&ctxt->write_sectors));
+ EBUG_ON(atomic_read(&ctxt->write_ios));
+ EBUG_ON(atomic_read(&ctxt->read_sectors));
+ EBUG_ON(atomic_read(&ctxt->read_ios));
if (ctxt->stats) {
progress_list_del(ctxt->c, ctxt->stats);
}
}
-void bch_move_stats_init(struct bch_move_stats *stats, char *name)
+void bch2_move_stats_init(struct bch_move_stats *stats, char *name)
{
memset(stats, 0, sizeof(*stats));
scnprintf(stats->name, sizeof(stats->name), "%s", name);
}
+static int bch2_extent_drop_ptrs(struct btree_trans *trans,
+ struct btree_iter *iter,
+ struct bkey_s_c k,
+ struct data_update_opts data_opts)
+{
+ struct bch_fs *c = trans->c;
+ struct bkey_i *n;
+ int ret;
+
+ n = bch2_bkey_make_mut(trans, k);
+ ret = PTR_ERR_OR_ZERO(n);
+ if (ret)
+ return ret;
+
+ while (data_opts.kill_ptrs) {
+ unsigned i = 0, drop = __fls(data_opts.kill_ptrs);
+ struct bch_extent_ptr *ptr;
+
+ bch2_bkey_drop_ptrs(bkey_i_to_s(n), ptr, i++ == drop);
+ data_opts.kill_ptrs ^= 1U << drop;
+ }
+
+ /*
+ * If the new extent no longer has any pointers, bch2_extent_normalize()
+ * will do the appropriate thing with it (turning it into a
+ * KEY_TYPE_error key, or just a discard if it was a cached extent)
+ */
+ bch2_extent_normalize(c, bkey_i_to_s(n));
+
+ /*
+ * Since we're not inserting through an extent iterator
+ * (BTREE_ITER_ALL_SNAPSHOTS iterators aren't extent iterators),
+ * we aren't using the extent overwrite path to delete, we're
+ * just using the normal key deletion path:
+ */
+ if (bkey_deleted(&n->k))
+ n->k.size = 0;
+
+ return bch2_trans_update(trans, iter, n, BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE) ?:
+ bch2_trans_commit(trans, NULL, NULL, BTREE_INSERT_NOFAIL);
+}
+
static int bch2_move_extent(struct btree_trans *trans,
+ struct btree_iter *iter,
struct moving_context *ctxt,
struct bch_io_opts io_opts,
enum btree_id btree_id,
unsigned sectors = k.k->size, pages;
int ret = -ENOMEM;
- if (!percpu_ref_tryget_live(&c->writes))
- return -EROFS;
+ bch2_data_update_opts_normalize(k, &data_opts);
+
+ if (!data_opts.rewrite_ptrs &&
+ !data_opts.extra_replicas) {
+ if (data_opts.kill_ptrs)
+ return bch2_extent_drop_ptrs(trans, iter, k, data_opts);
+ return 0;
+ }
+
+ if (!bch2_write_ref_tryget(c, BCH_WRITE_REF_move))
+ return -BCH_ERR_erofs_no_writes;
+
+ /*
+ * Before memory allocations & taking nocow locks in
+ * bch2_data_update_init():
+ */
+ bch2_trans_unlock(trans);
/* write path might have to decompress data: */
bkey_for_each_ptr_decode(k.k, ptrs, p, entry)
io->rbio.bio.bi_iter.bi_sector = bkey_start_offset(k.k);
io->rbio.bio.bi_end_io = move_read_endio;
- ret = bch2_data_update_init(c, &io->write, ctxt->wp, io_opts,
- data_opts, btree_id, k);
- if (ret)
+ ret = bch2_data_update_init(trans, ctxt, &io->write, ctxt->wp,
+ io_opts, data_opts, btree_id, k);
+ if (ret && ret != -BCH_ERR_unwritten_extent_update)
goto err_free_pages;
io->write.ctxt = ctxt;
+ io->write.op.end_io = move_write_done;
atomic64_inc(&ctxt->stats->keys_moved);
atomic64_add(k.k->size, &ctxt->stats->sectors_moved);
- this_cpu_add(c->counters[BCH_COUNTER_io_move], k.k->size);
- trace_move_extent(k.k);
+ if (ret == -BCH_ERR_unwritten_extent_update) {
+ bch2_update_unwritten_extent(trans, &io->write);
+ move_free(io);
+ return 0;
+ }
+
+ BUG_ON(ret);
+
+ this_cpu_add(c->counters[BCH_COUNTER_io_move], k.k->size);
+ this_cpu_add(c->counters[BCH_COUNTER_move_extent_read], k.k->size);
+ trace_move_extent_read(k.k);
atomic_add(io->read_sectors, &ctxt->read_sectors);
+ atomic_inc(&ctxt->read_ios);
list_add_tail(&io->list, &ctxt->reads);
/*
err_free:
kfree(io);
err:
- percpu_ref_put(&c->writes);
- trace_move_alloc_mem_fail(k.k);
+ bch2_write_ref_put(c, BCH_WRITE_REF_move);
+ trace_and_count(c, move_extent_alloc_mem_fail, k.k);
return ret;
}
if (ret)
goto err;
- if (!k.k || bkey_cmp(k.k->p, pos)) {
+ if (!k.k || !bkey_eq(k.k->p, pos)) {
ret = -ENOENT;
goto err;
}
}
} while (delay);
+ /*
+ * XXX: these limits really ought to be per device, SSDs and hard drives
+ * will want different limits
+ */
move_ctxt_wait_event(ctxt, trans,
- atomic_read(&ctxt->write_sectors) <
- c->opts.move_bytes_in_flight >> 9);
-
- move_ctxt_wait_event(ctxt, trans,
- atomic_read(&ctxt->read_sectors) <
- c->opts.move_bytes_in_flight >> 9);
+ atomic_read(&ctxt->write_sectors) < c->opts.move_bytes_in_flight >> 9 &&
+ atomic_read(&ctxt->read_sectors) < c->opts.move_bytes_in_flight >> 9 &&
+ atomic_read(&ctxt->write_ios) < c->opts.move_ios_in_flight &&
+ atomic_read(&ctxt->read_ios) < c->opts.move_ios_in_flight);
return 0;
}
if (*cur_inum == k.k->p.inode)
return 0;
- *io_opts = bch2_opts_to_inode_opts(trans->c->opts);
-
ret = lookup_inode(trans,
SPOS(0, k.k->p.inode, k.k->p.snapshot),
&inode);
- if (ret == -EINTR)
+ if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
return ret;
if (!ret)
- bch2_io_opts_apply(io_opts, bch2_inode_opts_get(&inode));
-
+ 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;
}
break;
ret = bkey_err(k);
- if (ret == -EINTR)
+ if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
continue;
if (ret)
break;
- if (bkey_cmp(bkey_start_pos(k.k), end) >= 0)
+ if (bkey_ge(bkey_start_pos(k.k), end))
break;
ctxt->stats->pos = iter.pos;
/*
* The iterator gets unlocked by __bch2_read_extent - need to
* save a copy of @k elsewhere:
- */
+ */
bch2_bkey_buf_reassemble(&sk, c, k);
k = bkey_i_to_s_c(sk.k);
+ bch2_trans_unlock(&trans);
- ret2 = bch2_move_extent(&trans, ctxt, io_opts,
+ ret2 = bch2_move_extent(&trans, &iter, ctxt, io_opts,
btree_id, k, data_opts);
if (ret2) {
- if (ret2 == -EINTR)
+ if (bch2_err_matches(ret2, BCH_ERR_transaction_restart))
continue;
if (ret2 == -ENOMEM) {
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,
if (a.v->gen == gen &&
a.v->dirty_sectors) {
- struct printbuf buf = PRINTBUF;
-
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;
}
-
- prt_str(&buf, "failed to evacuate bucket ");
- bch2_bkey_val_to_text(&buf, c, k);
-
- bch2_trans_inconsistent(trans, "%s", buf.buf);
- printbuf_exit(&buf);
}
}
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,
+ BTREE_ITER_CACHED);
+ 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 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 (ret == -EINTR)
+ &bp_offset, &bp,
+ BTREE_ITER_CACHED);
+ if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
continue;
if (ret)
goto err;
k = bch2_backpointer_get_key(&trans, &iter,
bucket, bp_offset, bp);
ret = bkey_err(k);
- if (ret == -EINTR)
+ if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
continue;
if (ret)
goto err;
if (!k.k)
- continue;
+ goto next;
bch2_bkey_buf_reassemble(&sk, c, k);
k = bkey_i_to_s_c(sk.k);
- bch2_trans_iter_exit(&trans, &iter);
ret = move_get_io_opts(&trans, &io_opts, k, &cur_inum);
- if (ret)
+ if (ret) {
+ bch2_trans_iter_exit(&trans, &iter);
continue;
+ }
data_opts = _data_opts;
data_opts.target = io_opts.background_target;
i++;
}
- ret = bch2_move_extent(&trans, ctxt, io_opts,
+ ret = bch2_move_extent(&trans, &iter, ctxt, io_opts,
bp.btree_id, k, data_opts);
- if (ret == -EINTR)
+ 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 */
b = bch2_backpointer_get_node(&trans, &iter,
bucket, bp_offset, bp);
ret = PTR_ERR_OR_ZERO(b);
- if (ret == -EINTR)
+ 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)
- continue;
+ goto next;
ret = bch2_btree_node_rewrite(&trans, &iter, b, 0);
bch2_trans_iter_exit(&trans, &iter);
- if (ret == -EINTR)
+ if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
continue;
if (ret)
goto err;
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));
goto next;
ret = bch2_btree_node_rewrite(&trans, &iter, b, 0) ?: ret;
- if (ret == -EINTR)
+ if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
continue;
if (ret)
break;
next:
bch2_btree_iter_next_node(&iter);
}
- if (ret == -EINTR)
+ if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
goto retry;
bch2_trans_iter_exit(&trans, &iter);
bch2_trans_exit(&trans);
if (ret)
- bch_err(c, "error %i in bch2_move_btree", ret);
+ bch_err(c, "error in %s(): %s", __func__, bch2_err_str(ret));
bch2_btree_interior_updates_flush(c);
i++;
}
- return data_opts->rewrite_ptrs != 0;;
+ return data_opts->rewrite_ptrs != 0;
}
static bool rereplicate_btree_pred(struct bch_fs *c, void *arg,
switch (op.op) {
case BCH_DATA_OP_REREPLICATE:
- bch_move_stats_init(stats, "rereplicate");
+ bch2_move_stats_init(stats, "rereplicate");
stats->data_type = BCH_DATA_journal;
ret = bch2_journal_flush_device_pins(&c->journal, -1);
if (op.migrate.dev >= c->sb.nr_devices)
return -EINVAL;
- bch_move_stats_init(stats, "migrate");
+ bch2_move_stats_init(stats, "migrate");
stats->data_type = BCH_DATA_journal;
ret = bch2_journal_flush_device_pins(&c->journal, op.migrate.dev);
ret = bch2_replicas_gc2(c) ?: ret;
break;
case BCH_DATA_OP_REWRITE_OLD_NODES:
- bch_move_stats_init(stats, "rewrite_old_nodes");
+ bch2_move_stats_init(stats, "rewrite_old_nodes");
ret = bch2_scan_old_btree_nodes(c, stats);
break;
default: