1// SPDX-License-Identifier: GPL-2.0
  2/*
  3 * Code for moving data off a device.
  4 */
  5
  6#include "bcachefs.h"
  7#include "bkey_buf.h"
  8#include "btree_update.h"
  9#include "btree_update_interior.h"
 10#include "buckets.h"
 11#include "errcode.h"
 12#include "extents.h"
 13#include "io_write.h"
 14#include "journal.h"
 15#include "keylist.h"
 16#include "migrate.h"
 17#include "move.h"
 18#include "replicas.h"
 19#include "super-io.h"
 20
 21static int drop_dev_ptrs(struct bch_fs *c, struct bkey_s k,
 22			 unsigned dev_idx, int flags, bool metadata)
 23{
 24	unsigned replicas = metadata ? c->opts.metadata_replicas : c->opts.data_replicas;
 25	unsigned lost = metadata ? BCH_FORCE_IF_METADATA_LOST : BCH_FORCE_IF_DATA_LOST;
 26	unsigned degraded = metadata ? BCH_FORCE_IF_METADATA_DEGRADED : BCH_FORCE_IF_DATA_DEGRADED;
 27	unsigned nr_good;
 28
 29	bch2_bkey_drop_device(k, dev_idx);
 30
 31	nr_good = bch2_bkey_durability(c, k.s_c);
 32	if ((!nr_good && !(flags & lost)) ||
 33	    (nr_good < replicas && !(flags & degraded)))
 34		return -EINVAL;
 35
 36	return 0;
 37}
 38
 39static int bch2_dev_usrdata_drop_key(struct btree_trans *trans,
 40				     struct btree_iter *iter,
 41				     struct bkey_s_c k,
 42				     unsigned dev_idx,
 43				     int flags)
 44{
 45	struct bch_fs *c = trans->c;
 46	struct bkey_i *n;
 47	int ret;
 48
 49	if (!bch2_bkey_has_device_c(k, dev_idx))
 50		return 0;
 51
 52	n = bch2_bkey_make_mut(trans, iter, &k, BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE);
 53	ret = PTR_ERR_OR_ZERO(n);
 54	if (ret)
 55		return ret;
 56
 57	ret = drop_dev_ptrs(c, bkey_i_to_s(n), dev_idx, flags, false);
 58	if (ret)
 59		return ret;
 60
 61	/*
 62	 * If the new extent no longer has any pointers, bch2_extent_normalize()
 63	 * will do the appropriate thing with it (turning it into a
 64	 * KEY_TYPE_error key, or just a discard if it was a cached extent)
 65	 */
 66	bch2_extent_normalize(c, bkey_i_to_s(n));
 67
 68	/*
 69	 * Since we're not inserting through an extent iterator
 70	 * (BTREE_ITER_ALL_SNAPSHOTS iterators aren't extent iterators),
 71	 * we aren't using the extent overwrite path to delete, we're
 72	 * just using the normal key deletion path:
 73	 */
 74	if (bkey_deleted(&n->k))
 75		n->k.size = 0;
 76	return 0;
 77}
 78
 79static int bch2_dev_usrdata_drop(struct bch_fs *c, unsigned dev_idx, int flags)
 80{
 81	struct btree_trans *trans = bch2_trans_get(c);
 82	enum btree_id id;
 83	int ret = 0;
 84
 85	for (id = 0; id < BTREE_ID_NR; id++) {
 86		if (!btree_type_has_ptrs(id))
 87			continue;
 88
 89		ret = for_each_btree_key_commit(trans, iter, id, POS_MIN,
 90				BTREE_ITER_PREFETCH|BTREE_ITER_ALL_SNAPSHOTS, k,
 91				NULL, NULL, BCH_TRANS_COMMIT_no_enospc,
 92			bch2_dev_usrdata_drop_key(trans, &iter, k, dev_idx, flags));
 93		if (ret)
 94			break;
 95	}
 96
 97	bch2_trans_put(trans);
 98
 99	return ret;
100}
101
102static int bch2_dev_metadata_drop(struct bch_fs *c, unsigned dev_idx, int flags)
103{
104	struct btree_trans *trans;
105	struct btree_iter iter;
106	struct closure cl;
107	struct btree *b;
108	struct bkey_buf k;
109	unsigned id;
110	int ret;
111
112	/* don't handle this yet: */
113	if (flags & BCH_FORCE_IF_METADATA_LOST)
114		return -EINVAL;
115
116	trans = bch2_trans_get(c);
117	bch2_bkey_buf_init(&k);
118	closure_init_stack(&cl);
119
120	for (id = 0; id < BTREE_ID_NR; id++) {
121		bch2_trans_node_iter_init(trans, &iter, id, POS_MIN, 0, 0,
122					  BTREE_ITER_PREFETCH);
123retry:
124		ret = 0;
125		while (bch2_trans_begin(trans),
126		       (b = bch2_btree_iter_peek_node(&iter)) &&
127		       !(ret = PTR_ERR_OR_ZERO(b))) {
128			if (!bch2_bkey_has_device_c(bkey_i_to_s_c(&b->key), dev_idx))
129				goto next;
130
131			bch2_bkey_buf_copy(&k, c, &b->key);
132
133			ret = drop_dev_ptrs(c, bkey_i_to_s(k.k),
134					    dev_idx, flags, true);
135			if (ret) {
136				bch_err(c, "Cannot drop device without losing data");
137				break;
138			}
139
140			ret = bch2_btree_node_update_key(trans, &iter, b, k.k, 0, false);
141			if (bch2_err_matches(ret, BCH_ERR_transaction_restart)) {
142				ret = 0;
143				continue;
144			}
145
146			bch_err_msg(c, ret, "updating btree node key");
147			if (ret)
148				break;
149next:
150			bch2_btree_iter_next_node(&iter);
151		}
152		if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
153			goto retry;
154
155		bch2_trans_iter_exit(trans, &iter);
156
157		if (ret)
158			goto err;
159	}
160
161	bch2_btree_interior_updates_flush(c);
162	ret = 0;
163err:
164	bch2_bkey_buf_exit(&k, c);
165	bch2_trans_put(trans);
166
167	BUG_ON(bch2_err_matches(ret, BCH_ERR_transaction_restart));
168
169	return ret;
170}
171
172int bch2_dev_data_drop(struct bch_fs *c, unsigned dev_idx, int flags)
173{
174	return bch2_dev_usrdata_drop(c, dev_idx, flags) ?:
175		bch2_dev_metadata_drop(c, dev_idx, flags);
176}