xref: /illumos-gate/usr/src/uts/common/fs/zfs/bptree.c (revision b89e420ae1290e425c29db875ec0c0546006eec7)
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) 2013 by Delphix. All rights reserved.
24  */
25 
26 #include <sys/arc.h>
27 #include <sys/bptree.h>
28 #include <sys/dmu.h>
29 #include <sys/dmu_objset.h>
30 #include <sys/dmu_tx.h>
31 #include <sys/dmu_traverse.h>
32 #include <sys/dsl_dataset.h>
33 #include <sys/dsl_dir.h>
34 #include <sys/dsl_pool.h>
35 #include <sys/dnode.h>
36 #include <sys/refcount.h>
37 #include <sys/spa.h>
38 
39 /*
40  * A bptree is a queue of root block pointers from destroyed datasets. When a
41  * dataset is destroyed its root block pointer is put on the end of the pool's
42  * bptree queue so the dataset's blocks can be freed asynchronously by
43  * dsl_scan_sync. This allows the delete operation to finish without traversing
44  * all the dataset's blocks.
45  *
46  * Note that while bt_begin and bt_end are only ever incremented in this code,
47  * they are effectively reset to 0 every time the entire bptree is freed because
48  * the bptree's object is destroyed and re-created.
49  */
50 
51 struct bptree_args {
52 	bptree_phys_t *ba_phys;	/* data in bonus buffer, dirtied if freeing */
53 	boolean_t ba_free;	/* true if freeing during traversal */
54 
55 	bptree_itor_t *ba_func;	/* function to call for each blockpointer */
56 	void *ba_arg;		/* caller supplied argument to ba_func */
57 	dmu_tx_t *ba_tx;	/* caller supplied tx, NULL if not freeing */
58 } bptree_args_t;
59 
60 uint64_t
61 bptree_alloc(objset_t *os, dmu_tx_t *tx)
62 {
63 	uint64_t obj;
64 	dmu_buf_t *db;
65 	bptree_phys_t *bt;
66 
67 	obj = dmu_object_alloc(os, DMU_OTN_UINT64_METADATA,
68 	    SPA_MAXBLOCKSIZE, DMU_OTN_UINT64_METADATA,
69 	    sizeof (bptree_phys_t), tx);
70 
71 	/*
72 	 * Bonus buffer contents are already initialized to 0, but for
73 	 * readability we make it explicit.
74 	 */
75 	VERIFY3U(0, ==, dmu_bonus_hold(os, obj, FTAG, &db));
76 	dmu_buf_will_dirty(db, tx);
77 	bt = db->db_data;
78 	bt->bt_begin = 0;
79 	bt->bt_end = 0;
80 	bt->bt_bytes = 0;
81 	bt->bt_comp = 0;
82 	bt->bt_uncomp = 0;
83 	dmu_buf_rele(db, FTAG);
84 
85 	return (obj);
86 }
87 
88 int
89 bptree_free(objset_t *os, uint64_t obj, dmu_tx_t *tx)
90 {
91 	dmu_buf_t *db;
92 	bptree_phys_t *bt;
93 
94 	VERIFY3U(0, ==, dmu_bonus_hold(os, obj, FTAG, &db));
95 	bt = db->db_data;
96 	ASSERT3U(bt->bt_begin, ==, bt->bt_end);
97 	ASSERT0(bt->bt_bytes);
98 	ASSERT0(bt->bt_comp);
99 	ASSERT0(bt->bt_uncomp);
100 	dmu_buf_rele(db, FTAG);
101 
102 	return (dmu_object_free(os, obj, tx));
103 }
104 
105 boolean_t
106 bptree_is_empty(objset_t *os, uint64_t obj)
107 {
108 	dmu_buf_t *db;
109 	bptree_phys_t *bt;
110 	boolean_t rv;
111 
112 	VERIFY0(dmu_bonus_hold(os, obj, FTAG, &db));
113 	bt = db->db_data;
114 	rv = (bt->bt_begin == bt->bt_end);
115 	dmu_buf_rele(db, FTAG);
116 	return (rv);
117 }
118 
119 void
120 bptree_add(objset_t *os, uint64_t obj, blkptr_t *bp, uint64_t birth_txg,
121     uint64_t bytes, uint64_t comp, uint64_t uncomp, dmu_tx_t *tx)
122 {
123 	dmu_buf_t *db;
124 	bptree_phys_t *bt;
125 	bptree_entry_phys_t bte = { 0 };
126 
127 	/*
128 	 * bptree objects are in the pool mos, therefore they can only be
129 	 * modified in syncing context. Furthermore, this is only modified
130 	 * by the sync thread, so no locking is necessary.
131 	 */
132 	ASSERT(dmu_tx_is_syncing(tx));
133 
134 	VERIFY3U(0, ==, dmu_bonus_hold(os, obj, FTAG, &db));
135 	bt = db->db_data;
136 
137 	bte.be_birth_txg = birth_txg;
138 	bte.be_bp = *bp;
139 	dmu_write(os, obj, bt->bt_end * sizeof (bte), sizeof (bte), &bte, tx);
140 
141 	dmu_buf_will_dirty(db, tx);
142 	bt->bt_end++;
143 	bt->bt_bytes += bytes;
144 	bt->bt_comp += comp;
145 	bt->bt_uncomp += uncomp;
146 	dmu_buf_rele(db, FTAG);
147 }
148 
149 /* ARGSUSED */
150 static int
151 bptree_visit_cb(spa_t *spa, zilog_t *zilog, const blkptr_t *bp,
152     const zbookmark_t *zb, const dnode_phys_t *dnp, void *arg)
153 {
154 	int err;
155 	struct bptree_args *ba = arg;
156 
157 	if (BP_IS_HOLE(bp))
158 		return (0);
159 
160 	err = ba->ba_func(ba->ba_arg, bp, ba->ba_tx);
161 	if (err == 0 && ba->ba_free) {
162 		ba->ba_phys->bt_bytes -= bp_get_dsize_sync(spa, bp);
163 		ba->ba_phys->bt_comp -= BP_GET_PSIZE(bp);
164 		ba->ba_phys->bt_uncomp -= BP_GET_UCSIZE(bp);
165 	}
166 	return (err);
167 }
168 
169 /*
170  * If "free" is set:
171  *  - It is assumed that "func" will be freeing the block pointers.
172  *  - If "func" returns nonzero, the bookmark will be remembered and
173  *    iteration will be restarted from this point on next invocation.
174  *  - If an i/o error is encountered (e.g. "func" returns EIO or ECKSUM),
175  *    bptree_iterate will remember the bookmark, continue traversing
176  *    any additional entries, and return 0.
177  *
178  * If "free" is not set, traversal will stop and return an error if
179  * an i/o error is encountered.
180  *
181  * In either case, if zfs_free_leak_on_eio is set, i/o errors will be
182  * ignored and traversal will continue (i.e. TRAVERSE_HARD will be passed to
183  * traverse_dataset_destroyed()).
184  */
185 int
186 bptree_iterate(objset_t *os, uint64_t obj, boolean_t free, bptree_itor_t func,
187     void *arg, dmu_tx_t *tx)
188 {
189 	boolean_t ioerr = B_FALSE;
190 	int err;
191 	uint64_t i;
192 	dmu_buf_t *db;
193 	struct bptree_args ba;
194 
195 	ASSERT(!free || dmu_tx_is_syncing(tx));
196 
197 	err = dmu_bonus_hold(os, obj, FTAG, &db);
198 	if (err != 0)
199 		return (err);
200 
201 	if (free)
202 		dmu_buf_will_dirty(db, tx);
203 
204 	ba.ba_phys = db->db_data;
205 	ba.ba_free = free;
206 	ba.ba_func = func;
207 	ba.ba_arg = arg;
208 	ba.ba_tx = tx;
209 
210 	err = 0;
211 	for (i = ba.ba_phys->bt_begin; i < ba.ba_phys->bt_end; i++) {
212 		bptree_entry_phys_t bte;
213 		int flags = TRAVERSE_PREFETCH_METADATA | TRAVERSE_POST;
214 
215 		err = dmu_read(os, obj, i * sizeof (bte), sizeof (bte),
216 		    &bte, DMU_READ_NO_PREFETCH);
217 		if (err != 0)
218 			break;
219 
220 		if (zfs_free_leak_on_eio)
221 			flags |= TRAVERSE_HARD;
222 		zfs_dbgmsg("bptree index %d: traversing from min_txg=%lld "
223 		    "bookmark %lld/%lld/%lld/%lld",
224 		    i, (longlong_t)bte.be_birth_txg,
225 		    (longlong_t)bte.be_zb.zb_objset,
226 		    (longlong_t)bte.be_zb.zb_object,
227 		    (longlong_t)bte.be_zb.zb_level,
228 		    (longlong_t)bte.be_zb.zb_blkid);
229 		err = traverse_dataset_destroyed(os->os_spa, &bte.be_bp,
230 		    bte.be_birth_txg, &bte.be_zb, flags,
231 		    bptree_visit_cb, &ba);
232 		if (free) {
233 			/*
234 			 * The callback has freed the visited block pointers.
235 			 * Record our traversal progress on disk, either by
236 			 * updating this record's bookmark, or by logically
237 			 * removing this record by advancing bt_begin.
238 			 */
239 			if (err != 0) {
240 				/* save bookmark for future resume */
241 				ASSERT3U(bte.be_zb.zb_objset, ==,
242 				    ZB_DESTROYED_OBJSET);
243 				ASSERT0(bte.be_zb.zb_level);
244 				dmu_write(os, obj, i * sizeof (bte),
245 				    sizeof (bte), &bte, tx);
246 				if (err == EIO || err == ECKSUM ||
247 				    err == ENXIO) {
248 					/*
249 					 * Skip the rest of this tree and
250 					 * continue on to the next entry.
251 					 */
252 					err = 0;
253 					ioerr = B_TRUE;
254 				} else {
255 					break;
256 				}
257 			} else if (ioerr) {
258 				/*
259 				 * This entry is finished, but there were
260 				 * i/o errors on previous entries, so we
261 				 * can't adjust bt_begin.  Set this entry's
262 				 * be_birth_txg such that it will be
263 				 * treated as a no-op in future traversals.
264 				 */
265 				bte.be_birth_txg = UINT64_MAX;
266 				dmu_write(os, obj, i * sizeof (bte),
267 				    sizeof (bte), &bte, tx);
268 			}
269 
270 			if (!ioerr) {
271 				ba.ba_phys->bt_begin++;
272 				(void) dmu_free_range(os, obj,
273 				    i * sizeof (bte), sizeof (bte), tx);
274 			}
275 		} else if (err != 0) {
276 			break;
277 		}
278 	}
279 
280 	ASSERT(!free || err != 0 || ioerr ||
281 	    ba.ba_phys->bt_begin == ba.ba_phys->bt_end);
282 
283 	/* if all blocks are free there should be no used space */
284 	if (ba.ba_phys->bt_begin == ba.ba_phys->bt_end) {
285 		if (zfs_free_leak_on_eio) {
286 			ba.ba_phys->bt_bytes = 0;
287 			ba.ba_phys->bt_comp = 0;
288 			ba.ba_phys->bt_uncomp = 0;
289 		}
290 
291 		ASSERT0(ba.ba_phys->bt_bytes);
292 		ASSERT0(ba.ba_phys->bt_comp);
293 		ASSERT0(ba.ba_phys->bt_uncomp);
294 	}
295 
296 	dmu_buf_rele(db, FTAG);
297 
298 	return (err);
299 }
300