1/*
2 * CDDL HEADER START
3 *
4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
7 *
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
12 *
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 *
19 * CDDL HEADER END
20 */
21
22/*
23 * Copyright (c) 2011, 2015 by Delphix. All rights reserved.
24 * Copyright (c) 2014 Integros [integros.com]
25 */
26
27#include <sys/arc.h>
28#include <sys/bptree.h>
29#include <sys/dmu.h>
30#include <sys/dmu_objset.h>
31#include <sys/dmu_tx.h>
32#include <sys/dmu_traverse.h>
33#include <sys/dsl_dataset.h>
34#include <sys/dsl_dir.h>
35#include <sys/dsl_pool.h>
36#include <sys/dnode.h>
37#include <sys/refcount.h>
38#include <sys/spa.h>
39
40/*
41 * A bptree is a queue of root block pointers from destroyed datasets. When a
42 * dataset is destroyed its root block pointer is put on the end of the pool's
43 * bptree queue so the dataset's blocks can be freed asynchronously by
44 * dsl_scan_sync. This allows the delete operation to finish without traversing
45 * all the dataset's blocks.
46 *
47 * Note that while bt_begin and bt_end are only ever incremented in this code,
48 * they are effectively reset to 0 every time the entire bptree is freed because
49 * the bptree's object is destroyed and re-created.
50 */
51
52struct bptree_args {
53	bptree_phys_t *ba_phys;	/* data in bonus buffer, dirtied if freeing */
54	boolean_t ba_free;	/* true if freeing during traversal */
55
56	bptree_itor_t *ba_func;	/* function to call for each blockpointer */
57	void *ba_arg;		/* caller supplied argument to ba_func */
58	dmu_tx_t *ba_tx;	/* caller supplied tx, NULL if not freeing */
59} bptree_args_t;
60
61uint64_t
62bptree_alloc(objset_t *os, dmu_tx_t *tx)
63{
64	uint64_t obj;
65	dmu_buf_t *db;
66	bptree_phys_t *bt;
67
68	obj = dmu_object_alloc(os, DMU_OTN_UINT64_METADATA,
69	    SPA_OLD_MAXBLOCKSIZE, DMU_OTN_UINT64_METADATA,
70	    sizeof (bptree_phys_t), tx);
71
72	/*
73	 * Bonus buffer contents are already initialized to 0, but for
74	 * readability we make it explicit.
75	 */
76	VERIFY3U(0, ==, dmu_bonus_hold(os, obj, FTAG, &db));
77	dmu_buf_will_dirty(db, tx);
78	bt = db->db_data;
79	bt->bt_begin = 0;
80	bt->bt_end = 0;
81	bt->bt_bytes = 0;
82	bt->bt_comp = 0;
83	bt->bt_uncomp = 0;
84	dmu_buf_rele(db, FTAG);
85
86	return (obj);
87}
88
89int
90bptree_free(objset_t *os, uint64_t obj, dmu_tx_t *tx)
91{
92	dmu_buf_t *db;
93	bptree_phys_t *bt;
94
95	VERIFY3U(0, ==, dmu_bonus_hold(os, obj, FTAG, &db));
96	bt = db->db_data;
97	ASSERT3U(bt->bt_begin, ==, bt->bt_end);
98	ASSERT0(bt->bt_bytes);
99	ASSERT0(bt->bt_comp);
100	ASSERT0(bt->bt_uncomp);
101	dmu_buf_rele(db, FTAG);
102
103	return (dmu_object_free(os, obj, tx));
104}
105
106boolean_t
107bptree_is_empty(objset_t *os, uint64_t obj)
108{
109	dmu_buf_t *db;
110	bptree_phys_t *bt;
111	boolean_t rv;
112
113	VERIFY0(dmu_bonus_hold(os, obj, FTAG, &db));
114	bt = db->db_data;
115	rv = (bt->bt_begin == bt->bt_end);
116	dmu_buf_rele(db, FTAG);
117	return (rv);
118}
119
120void
121bptree_add(objset_t *os, uint64_t obj, blkptr_t *bp, uint64_t birth_txg,
122    uint64_t bytes, uint64_t comp, uint64_t uncomp, dmu_tx_t *tx)
123{
124	dmu_buf_t *db;
125	bptree_phys_t *bt;
126	bptree_entry_phys_t bte = { 0 };
127
128	/*
129	 * bptree objects are in the pool mos, therefore they can only be
130	 * modified in syncing context. Furthermore, this is only modified
131	 * by the sync thread, so no locking is necessary.
132	 */
133	ASSERT(dmu_tx_is_syncing(tx));
134
135	VERIFY3U(0, ==, dmu_bonus_hold(os, obj, FTAG, &db));
136	bt = db->db_data;
137
138	bte.be_birth_txg = birth_txg;
139	bte.be_bp = *bp;
140	dmu_write(os, obj, bt->bt_end * sizeof (bte), sizeof (bte), &bte, tx);
141
142	dmu_buf_will_dirty(db, tx);
143	bt->bt_end++;
144	bt->bt_bytes += bytes;
145	bt->bt_comp += comp;
146	bt->bt_uncomp += uncomp;
147	dmu_buf_rele(db, FTAG);
148}
149
150/* ARGSUSED */
151static int
152bptree_visit_cb(spa_t *spa, zilog_t *zilog, const blkptr_t *bp,
153    const zbookmark_phys_t *zb, const dnode_phys_t *dnp, void *arg)
154{
155	int err;
156	struct bptree_args *ba = arg;
157
158	if (bp == NULL || BP_IS_HOLE(bp))
159		return (0);
160
161	err = ba->ba_func(ba->ba_arg, bp, ba->ba_tx);
162	if (err == 0 && ba->ba_free) {
163		ba->ba_phys->bt_bytes -= bp_get_dsize_sync(spa, bp);
164		ba->ba_phys->bt_comp -= BP_GET_PSIZE(bp);
165		ba->ba_phys->bt_uncomp -= BP_GET_UCSIZE(bp);
166	}
167	return (err);
168}
169
170/*
171 * If "free" is set:
172 *  - It is assumed that "func" will be freeing the block pointers.
173 *  - If "func" returns nonzero, the bookmark will be remembered and
174 *    iteration will be restarted from this point on next invocation.
175 *  - If an i/o error is encountered (e.g. "func" returns EIO or ECKSUM),
176 *    bptree_iterate will remember the bookmark, continue traversing
177 *    any additional entries, and return 0.
178 *
179 * If "free" is not set, traversal will stop and return an error if
180 * an i/o error is encountered.
181 *
182 * In either case, if zfs_free_leak_on_eio is set, i/o errors will be
183 * ignored and traversal will continue (i.e. TRAVERSE_HARD will be passed to
184 * traverse_dataset_destroyed()).
185 */
186int
187bptree_iterate(objset_t *os, uint64_t obj, boolean_t free, bptree_itor_t func,
188    void *arg, dmu_tx_t *tx)
189{
190	boolean_t ioerr = B_FALSE;
191	int err;
192	uint64_t i;
193	dmu_buf_t *db;
194	struct bptree_args ba;
195
196	ASSERT(!free || dmu_tx_is_syncing(tx));
197
198	err = dmu_bonus_hold(os, obj, FTAG, &db);
199	if (err != 0)
200		return (err);
201
202	if (free)
203		dmu_buf_will_dirty(db, tx);
204
205	ba.ba_phys = db->db_data;
206	ba.ba_free = free;
207	ba.ba_func = func;
208	ba.ba_arg = arg;
209	ba.ba_tx = tx;
210
211	err = 0;
212	for (i = ba.ba_phys->bt_begin; i < ba.ba_phys->bt_end; i++) {
213		bptree_entry_phys_t bte;
214		int flags = TRAVERSE_PREFETCH_METADATA | TRAVERSE_POST
215		    | TRAVERSE_NO_DECRYPT;
216
217		err = dmu_read(os, obj, i * sizeof (bte), sizeof (bte),
218		    &bte, DMU_READ_NO_PREFETCH);
219		if (err != 0)
220			break;
221
222		if (zfs_free_leak_on_eio)
223			flags |= TRAVERSE_HARD;
224		zfs_dbgmsg("bptree index %lld: traversing from min_txg=%lld "
225		    "bookmark %lld/%lld/%lld/%lld",
226		    (longlong_t)i,
227		    (longlong_t)bte.be_birth_txg,
228		    (longlong_t)bte.be_zb.zb_objset,
229		    (longlong_t)bte.be_zb.zb_object,
230		    (longlong_t)bte.be_zb.zb_level,
231		    (longlong_t)bte.be_zb.zb_blkid);
232		err = traverse_dataset_destroyed(os->os_spa, &bte.be_bp,
233		    bte.be_birth_txg, &bte.be_zb, flags,
234		    bptree_visit_cb, &ba);
235		if (free) {
236			/*
237			 * The callback has freed the visited block pointers.
238			 * Record our traversal progress on disk, either by
239			 * updating this record's bookmark, or by logically
240			 * removing this record by advancing bt_begin.
241			 */
242			if (err != 0) {
243				/* save bookmark for future resume */
244				ASSERT3U(bte.be_zb.zb_objset, ==,
245				    ZB_DESTROYED_OBJSET);
246				ASSERT0(bte.be_zb.zb_level);
247				dmu_write(os, obj, i * sizeof (bte),
248				    sizeof (bte), &bte, tx);
249				if (err == EIO || err == ECKSUM ||
250				    err == ENXIO) {
251					/*
252					 * Skip the rest of this tree and
253					 * continue on to the next entry.
254					 */
255					err = 0;
256					ioerr = B_TRUE;
257				} else {
258					break;
259				}
260			} else if (ioerr) {
261				/*
262				 * This entry is finished, but there were
263				 * i/o errors on previous entries, so we
264				 * can't adjust bt_begin.  Set this entry's
265				 * be_birth_txg such that it will be
266				 * treated as a no-op in future traversals.
267				 */
268				bte.be_birth_txg = UINT64_MAX;
269				dmu_write(os, obj, i * sizeof (bte),
270				    sizeof (bte), &bte, tx);
271			}
272
273			if (!ioerr) {
274				ba.ba_phys->bt_begin++;
275				(void) dmu_free_range(os, obj,
276				    i * sizeof (bte), sizeof (bte), tx);
277			}
278		} else if (err != 0) {
279			break;
280		}
281	}
282
283	ASSERT(!free || err != 0 || ioerr ||
284	    ba.ba_phys->bt_begin == ba.ba_phys->bt_end);
285
286	/* if all blocks are free there should be no used space */
287	if (ba.ba_phys->bt_begin == ba.ba_phys->bt_end) {
288		if (zfs_free_leak_on_eio) {
289			ba.ba_phys->bt_bytes = 0;
290			ba.ba_phys->bt_comp = 0;
291			ba.ba_phys->bt_uncomp = 0;
292		}
293
294		ASSERT0(ba.ba_phys->bt_bytes);
295		ASSERT0(ba.ba_phys->bt_comp);
296		ASSERT0(ba.ba_phys->bt_uncomp);
297	}
298
299	dmu_buf_rele(db, FTAG);
300
301	return (err);
302}
303