1cde58dbcSMatthew Ahrens /*
2cde58dbcSMatthew Ahrens * CDDL HEADER START
3cde58dbcSMatthew Ahrens *
4cde58dbcSMatthew Ahrens * The contents of this file are subject to the terms of the
5cde58dbcSMatthew Ahrens * Common Development and Distribution License (the "License").
6cde58dbcSMatthew Ahrens * You may not use this file except in compliance with the License.
7cde58dbcSMatthew Ahrens *
8cde58dbcSMatthew Ahrens * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9cde58dbcSMatthew Ahrens * or http://www.opensolaris.org/os/licensing.
10cde58dbcSMatthew Ahrens * See the License for the specific language governing permissions
11cde58dbcSMatthew Ahrens * and limitations under the License.
12cde58dbcSMatthew Ahrens *
13cde58dbcSMatthew Ahrens * When distributing Covered Code, include this CDDL HEADER in each
14cde58dbcSMatthew Ahrens * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15cde58dbcSMatthew Ahrens * If applicable, add the following below this CDDL HEADER, with the
16cde58dbcSMatthew Ahrens * fields enclosed by brackets "[]" replaced with your own identifying
17cde58dbcSMatthew Ahrens * information: Portions Copyright [yyyy] [name of copyright owner]
18cde58dbcSMatthew Ahrens *
19cde58dbcSMatthew Ahrens * CDDL HEADER END
20cde58dbcSMatthew Ahrens */
21cde58dbcSMatthew Ahrens /*
22cde58dbcSMatthew Ahrens * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23a3905a45SSerapheim Dimitropoulos * Copyright (c) 2011, 2016 by Delphix. All rights reserved.
24c3d26abcSMatthew Ahrens * Copyright (c) 2014 Integros [integros.com]
251702cce7SAlek Pinchuk * Copyright (c) 2017 Datto Inc.
26cde58dbcSMatthew Ahrens */
27cde58dbcSMatthew Ahrens
28cde58dbcSMatthew Ahrens #include <sys/bpobj.h>
29cde58dbcSMatthew Ahrens #include <sys/zfs_context.h>
30cde58dbcSMatthew Ahrens #include <sys/refcount.h>
3119b94df9SMatthew Ahrens #include <sys/dsl_pool.h>
32f1745736SMatthew Ahrens #include <sys/zfeature.h>
33f1745736SMatthew Ahrens #include <sys/zap.h>
34f1745736SMatthew Ahrens
35f1745736SMatthew Ahrens /*
36f1745736SMatthew Ahrens * Return an empty bpobj, preferably the empty dummy one (dp_empty_bpobj).
37f1745736SMatthew Ahrens */
38f1745736SMatthew Ahrens uint64_t
bpobj_alloc_empty(objset_t * os,int blocksize,dmu_tx_t * tx)39f1745736SMatthew Ahrens bpobj_alloc_empty(objset_t *os, int blocksize, dmu_tx_t *tx)
40f1745736SMatthew Ahrens {
41f1745736SMatthew Ahrens spa_t *spa = dmu_objset_spa(os);
42f1745736SMatthew Ahrens dsl_pool_t *dp = dmu_objset_pool(os);
43f1745736SMatthew Ahrens
442acef22dSMatthew Ahrens if (spa_feature_is_enabled(spa, SPA_FEATURE_EMPTY_BPOBJ)) {
452acef22dSMatthew Ahrens if (!spa_feature_is_active(spa, SPA_FEATURE_EMPTY_BPOBJ)) {
46fb09f5aaSMadhav Suresh ASSERT0(dp->dp_empty_bpobj);
47f1745736SMatthew Ahrens dp->dp_empty_bpobj =
48b5152584SMatthew Ahrens bpobj_alloc(os, SPA_OLD_MAXBLOCKSIZE, tx);
49f1745736SMatthew Ahrens VERIFY(zap_add(os,
50f1745736SMatthew Ahrens DMU_POOL_DIRECTORY_OBJECT,
51f1745736SMatthew Ahrens DMU_POOL_EMPTY_BPOBJ, sizeof (uint64_t), 1,
52f1745736SMatthew Ahrens &dp->dp_empty_bpobj, tx) == 0);
53f1745736SMatthew Ahrens }
542acef22dSMatthew Ahrens spa_feature_incr(spa, SPA_FEATURE_EMPTY_BPOBJ, tx);
55f1745736SMatthew Ahrens ASSERT(dp->dp_empty_bpobj != 0);
56f1745736SMatthew Ahrens return (dp->dp_empty_bpobj);
57f1745736SMatthew Ahrens } else {
58f1745736SMatthew Ahrens return (bpobj_alloc(os, blocksize, tx));
59f1745736SMatthew Ahrens }
60f1745736SMatthew Ahrens }
61f1745736SMatthew Ahrens
62f1745736SMatthew Ahrens void
bpobj_decr_empty(objset_t * os,dmu_tx_t * tx)63f1745736SMatthew Ahrens bpobj_decr_empty(objset_t *os, dmu_tx_t *tx)
64f1745736SMatthew Ahrens {
65f1745736SMatthew Ahrens dsl_pool_t *dp = dmu_objset_pool(os);
66f1745736SMatthew Ahrens
672acef22dSMatthew Ahrens spa_feature_decr(dmu_objset_spa(os), SPA_FEATURE_EMPTY_BPOBJ, tx);
682acef22dSMatthew Ahrens if (!spa_feature_is_active(dmu_objset_spa(os),
692acef22dSMatthew Ahrens SPA_FEATURE_EMPTY_BPOBJ)) {
70f1745736SMatthew Ahrens VERIFY3U(0, ==, zap_remove(dp->dp_meta_objset,
71f1745736SMatthew Ahrens DMU_POOL_DIRECTORY_OBJECT,
72f1745736SMatthew Ahrens DMU_POOL_EMPTY_BPOBJ, tx));
73f1745736SMatthew Ahrens VERIFY3U(0, ==, dmu_object_free(os, dp->dp_empty_bpobj, tx));
74f1745736SMatthew Ahrens dp->dp_empty_bpobj = 0;
75f1745736SMatthew Ahrens }
76f1745736SMatthew Ahrens }
77cde58dbcSMatthew Ahrens
78cde58dbcSMatthew Ahrens uint64_t
bpobj_alloc(objset_t * os,int blocksize,dmu_tx_t * tx)79cde58dbcSMatthew Ahrens bpobj_alloc(objset_t *os, int blocksize, dmu_tx_t *tx)
80cde58dbcSMatthew Ahrens {
81cde58dbcSMatthew Ahrens int size;
82cde58dbcSMatthew Ahrens
83cde58dbcSMatthew Ahrens if (spa_version(dmu_objset_spa(os)) < SPA_VERSION_BPOBJ_ACCOUNT)
84cde58dbcSMatthew Ahrens size = BPOBJ_SIZE_V0;
85cde58dbcSMatthew Ahrens else if (spa_version(dmu_objset_spa(os)) < SPA_VERSION_DEADLISTS)
86cde58dbcSMatthew Ahrens size = BPOBJ_SIZE_V1;
87cde58dbcSMatthew Ahrens else
88cde58dbcSMatthew Ahrens size = sizeof (bpobj_phys_t);
89cde58dbcSMatthew Ahrens
90cde58dbcSMatthew Ahrens return (dmu_object_alloc(os, DMU_OT_BPOBJ, blocksize,
91cde58dbcSMatthew Ahrens DMU_OT_BPOBJ_HDR, size, tx));
92cde58dbcSMatthew Ahrens }
93cde58dbcSMatthew Ahrens
94cde58dbcSMatthew Ahrens void
bpobj_free(objset_t * os,uint64_t obj,dmu_tx_t * tx)95cde58dbcSMatthew Ahrens bpobj_free(objset_t *os, uint64_t obj, dmu_tx_t *tx)
96cde58dbcSMatthew Ahrens {
97cde58dbcSMatthew Ahrens int64_t i;
98cde58dbcSMatthew Ahrens bpobj_t bpo;
99cde58dbcSMatthew Ahrens dmu_object_info_t doi;
100cde58dbcSMatthew Ahrens int epb;
101cde58dbcSMatthew Ahrens dmu_buf_t *dbuf = NULL;
102cde58dbcSMatthew Ahrens
103f1745736SMatthew Ahrens ASSERT(obj != dmu_objset_pool(os)->dp_empty_bpobj);
104b420f3adSRichard Lowe VERIFY3U(0, ==, bpobj_open(&bpo, os, obj));
105cde58dbcSMatthew Ahrens
106cde58dbcSMatthew Ahrens mutex_enter(&bpo.bpo_lock);
107cde58dbcSMatthew Ahrens
108cde58dbcSMatthew Ahrens if (!bpo.bpo_havesubobj || bpo.bpo_phys->bpo_subobjs == 0)
109cde58dbcSMatthew Ahrens goto out;
110cde58dbcSMatthew Ahrens
111b420f3adSRichard Lowe VERIFY3U(0, ==, dmu_object_info(os, bpo.bpo_phys->bpo_subobjs, &doi));
112cde58dbcSMatthew Ahrens epb = doi.doi_data_block_size / sizeof (uint64_t);
113cde58dbcSMatthew Ahrens
114cde58dbcSMatthew Ahrens for (i = bpo.bpo_phys->bpo_num_subobjs - 1; i >= 0; i--) {
115cde58dbcSMatthew Ahrens uint64_t *objarray;
116cde58dbcSMatthew Ahrens uint64_t offset, blkoff;
117cde58dbcSMatthew Ahrens
118cde58dbcSMatthew Ahrens offset = i * sizeof (uint64_t);
119cde58dbcSMatthew Ahrens blkoff = P2PHASE(i, epb);
120cde58dbcSMatthew Ahrens
121cde58dbcSMatthew Ahrens if (dbuf == NULL || dbuf->db_offset > offset) {
122cde58dbcSMatthew Ahrens if (dbuf)
123cde58dbcSMatthew Ahrens dmu_buf_rele(dbuf, FTAG);
124b420f3adSRichard Lowe VERIFY3U(0, ==, dmu_buf_hold(os,
125cde58dbcSMatthew Ahrens bpo.bpo_phys->bpo_subobjs, offset, FTAG, &dbuf, 0));
126cde58dbcSMatthew Ahrens }
127cde58dbcSMatthew Ahrens
128cde58dbcSMatthew Ahrens ASSERT3U(offset, >=, dbuf->db_offset);
129cde58dbcSMatthew Ahrens ASSERT3U(offset, <, dbuf->db_offset + dbuf->db_size);
130cde58dbcSMatthew Ahrens
131cde58dbcSMatthew Ahrens objarray = dbuf->db_data;
132cde58dbcSMatthew Ahrens bpobj_free(os, objarray[blkoff], tx);
133cde58dbcSMatthew Ahrens }
134cde58dbcSMatthew Ahrens if (dbuf) {
135cde58dbcSMatthew Ahrens dmu_buf_rele(dbuf, FTAG);
136cde58dbcSMatthew Ahrens dbuf = NULL;
137cde58dbcSMatthew Ahrens }
138b420f3adSRichard Lowe VERIFY3U(0, ==, dmu_object_free(os, bpo.bpo_phys->bpo_subobjs, tx));
139cde58dbcSMatthew Ahrens
140cde58dbcSMatthew Ahrens out:
141cde58dbcSMatthew Ahrens mutex_exit(&bpo.bpo_lock);
142cde58dbcSMatthew Ahrens bpobj_close(&bpo);
143cde58dbcSMatthew Ahrens
144b420f3adSRichard Lowe VERIFY3U(0, ==, dmu_object_free(os, obj, tx));
145cde58dbcSMatthew Ahrens }
146cde58dbcSMatthew Ahrens
147cde58dbcSMatthew Ahrens int
bpobj_open(bpobj_t * bpo,objset_t * os,uint64_t object)148cde58dbcSMatthew Ahrens bpobj_open(bpobj_t *bpo, objset_t *os, uint64_t object)
149cde58dbcSMatthew Ahrens {
150cde58dbcSMatthew Ahrens dmu_object_info_t doi;
151cde58dbcSMatthew Ahrens int err;
152cde58dbcSMatthew Ahrens
153cde58dbcSMatthew Ahrens err = dmu_object_info(os, object, &doi);
154cde58dbcSMatthew Ahrens if (err)
155cde58dbcSMatthew Ahrens return (err);
156cde58dbcSMatthew Ahrens
157cde58dbcSMatthew Ahrens bzero(bpo, sizeof (*bpo));
158cde58dbcSMatthew Ahrens mutex_init(&bpo->bpo_lock, NULL, MUTEX_DEFAULT, NULL);
159cde58dbcSMatthew Ahrens
160cde58dbcSMatthew Ahrens ASSERT(bpo->bpo_dbuf == NULL);
161cde58dbcSMatthew Ahrens ASSERT(bpo->bpo_phys == NULL);
162cde58dbcSMatthew Ahrens ASSERT(object != 0);
163cde58dbcSMatthew Ahrens ASSERT3U(doi.doi_type, ==, DMU_OT_BPOBJ);
164cde58dbcSMatthew Ahrens ASSERT3U(doi.doi_bonus_type, ==, DMU_OT_BPOBJ_HDR);
165cde58dbcSMatthew Ahrens
166837b568bSGeorge Wilson err = dmu_bonus_hold(os, object, bpo, &bpo->bpo_dbuf);
167837b568bSGeorge Wilson if (err)
168837b568bSGeorge Wilson return (err);
169837b568bSGeorge Wilson
170cde58dbcSMatthew Ahrens bpo->bpo_os = os;
171cde58dbcSMatthew Ahrens bpo->bpo_object = object;
172cde58dbcSMatthew Ahrens bpo->bpo_epb = doi.doi_data_block_size >> SPA_BLKPTRSHIFT;
173cde58dbcSMatthew Ahrens bpo->bpo_havecomp = (doi.doi_bonus_size > BPOBJ_SIZE_V0);
174cde58dbcSMatthew Ahrens bpo->bpo_havesubobj = (doi.doi_bonus_size > BPOBJ_SIZE_V1);
175cde58dbcSMatthew Ahrens bpo->bpo_phys = bpo->bpo_dbuf->db_data;
176cde58dbcSMatthew Ahrens return (0);
177cde58dbcSMatthew Ahrens }
178cde58dbcSMatthew Ahrens
1795cabbc6bSPrashanth Sreenivasa boolean_t
bpobj_is_open(const bpobj_t * bpo)1805cabbc6bSPrashanth Sreenivasa bpobj_is_open(const bpobj_t *bpo)
1815cabbc6bSPrashanth Sreenivasa {
1825cabbc6bSPrashanth Sreenivasa return (bpo->bpo_object != 0);
1835cabbc6bSPrashanth Sreenivasa }
1845cabbc6bSPrashanth Sreenivasa
185cde58dbcSMatthew Ahrens void
bpobj_close(bpobj_t * bpo)186cde58dbcSMatthew Ahrens bpobj_close(bpobj_t *bpo)
187cde58dbcSMatthew Ahrens {
188cde58dbcSMatthew Ahrens /* Lame workaround for closing a bpobj that was never opened. */
189cde58dbcSMatthew Ahrens if (bpo->bpo_object == 0)
190cde58dbcSMatthew Ahrens return;
191cde58dbcSMatthew Ahrens
192cde58dbcSMatthew Ahrens dmu_buf_rele(bpo->bpo_dbuf, bpo);
193cde58dbcSMatthew Ahrens if (bpo->bpo_cached_dbuf != NULL)
194cde58dbcSMatthew Ahrens dmu_buf_rele(bpo->bpo_cached_dbuf, bpo);
195cde58dbcSMatthew Ahrens bpo->bpo_dbuf = NULL;
196cde58dbcSMatthew Ahrens bpo->bpo_phys = NULL;
197cde58dbcSMatthew Ahrens bpo->bpo_cached_dbuf = NULL;
198837b568bSGeorge Wilson bpo->bpo_object = 0;
199cde58dbcSMatthew Ahrens
200cde58dbcSMatthew Ahrens mutex_destroy(&bpo->bpo_lock);
201cde58dbcSMatthew Ahrens }
202cde58dbcSMatthew Ahrens
2035cabbc6bSPrashanth Sreenivasa boolean_t
bpobj_is_empty(bpobj_t * bpo)2045cabbc6bSPrashanth Sreenivasa bpobj_is_empty(bpobj_t *bpo)
2055d7b4d43SMatthew Ahrens {
2065cabbc6bSPrashanth Sreenivasa return (bpo->bpo_phys->bpo_num_blkptrs == 0 &&
2075cabbc6bSPrashanth Sreenivasa (!bpo->bpo_havesubobj || bpo->bpo_phys->bpo_num_subobjs == 0));
2085d7b4d43SMatthew Ahrens }
2095d7b4d43SMatthew Ahrens
210cde58dbcSMatthew Ahrens static int
bpobj_iterate_impl(bpobj_t * bpo,bpobj_itor_t func,void * arg,dmu_tx_t * tx,boolean_t free)211cde58dbcSMatthew Ahrens bpobj_iterate_impl(bpobj_t *bpo, bpobj_itor_t func, void *arg, dmu_tx_t *tx,
212cde58dbcSMatthew Ahrens boolean_t free)
213cde58dbcSMatthew Ahrens {
214cde58dbcSMatthew Ahrens dmu_object_info_t doi;
215cde58dbcSMatthew Ahrens int epb;
216cde58dbcSMatthew Ahrens int64_t i;
217cde58dbcSMatthew Ahrens int err = 0;
218cde58dbcSMatthew Ahrens dmu_buf_t *dbuf = NULL;
219cde58dbcSMatthew Ahrens
2205cabbc6bSPrashanth Sreenivasa ASSERT(bpobj_is_open(bpo));
221cde58dbcSMatthew Ahrens mutex_enter(&bpo->bpo_lock);
222cde58dbcSMatthew Ahrens
223cde58dbcSMatthew Ahrens if (free)
224cde58dbcSMatthew Ahrens dmu_buf_will_dirty(bpo->bpo_dbuf, tx);
225cde58dbcSMatthew Ahrens
226cde58dbcSMatthew Ahrens for (i = bpo->bpo_phys->bpo_num_blkptrs - 1; i >= 0; i--) {
227cde58dbcSMatthew Ahrens blkptr_t *bparray;
228cde58dbcSMatthew Ahrens blkptr_t *bp;
229cde58dbcSMatthew Ahrens uint64_t offset, blkoff;
230cde58dbcSMatthew Ahrens
231cde58dbcSMatthew Ahrens offset = i * sizeof (blkptr_t);
232cde58dbcSMatthew Ahrens blkoff = P2PHASE(i, bpo->bpo_epb);
233cde58dbcSMatthew Ahrens
234cde58dbcSMatthew Ahrens if (dbuf == NULL || dbuf->db_offset > offset) {
235cde58dbcSMatthew Ahrens if (dbuf)
236cde58dbcSMatthew Ahrens dmu_buf_rele(dbuf, FTAG);
237cde58dbcSMatthew Ahrens err = dmu_buf_hold(bpo->bpo_os, bpo->bpo_object, offset,
238cde58dbcSMatthew Ahrens FTAG, &dbuf, 0);
239cde58dbcSMatthew Ahrens if (err)
240cde58dbcSMatthew Ahrens break;
241cde58dbcSMatthew Ahrens }
242cde58dbcSMatthew Ahrens
243cde58dbcSMatthew Ahrens ASSERT3U(offset, >=, dbuf->db_offset);
244cde58dbcSMatthew Ahrens ASSERT3U(offset, <, dbuf->db_offset + dbuf->db_size);
245cde58dbcSMatthew Ahrens
246cde58dbcSMatthew Ahrens bparray = dbuf->db_data;
247cde58dbcSMatthew Ahrens bp = &bparray[blkoff];
248cde58dbcSMatthew Ahrens err = func(arg, bp, tx);
249cde58dbcSMatthew Ahrens if (err)
250cde58dbcSMatthew Ahrens break;
251cde58dbcSMatthew Ahrens if (free) {
252cde58dbcSMatthew Ahrens bpo->bpo_phys->bpo_bytes -=
253cde58dbcSMatthew Ahrens bp_get_dsize_sync(dmu_objset_spa(bpo->bpo_os), bp);
254cde58dbcSMatthew Ahrens ASSERT3S(bpo->bpo_phys->bpo_bytes, >=, 0);
255cde58dbcSMatthew Ahrens if (bpo->bpo_havecomp) {
256cde58dbcSMatthew Ahrens bpo->bpo_phys->bpo_comp -= BP_GET_PSIZE(bp);
257cde58dbcSMatthew Ahrens bpo->bpo_phys->bpo_uncomp -= BP_GET_UCSIZE(bp);
258cde58dbcSMatthew Ahrens }
259cde58dbcSMatthew Ahrens bpo->bpo_phys->bpo_num_blkptrs--;
260cde58dbcSMatthew Ahrens ASSERT3S(bpo->bpo_phys->bpo_num_blkptrs, >=, 0);
261cde58dbcSMatthew Ahrens }
262cde58dbcSMatthew Ahrens }
263cde58dbcSMatthew Ahrens if (dbuf) {
264cde58dbcSMatthew Ahrens dmu_buf_rele(dbuf, FTAG);
265cde58dbcSMatthew Ahrens dbuf = NULL;
266cde58dbcSMatthew Ahrens }
267cde58dbcSMatthew Ahrens if (free) {
268b420f3adSRichard Lowe VERIFY3U(0, ==, dmu_free_range(bpo->bpo_os, bpo->bpo_object,
269*eb633035STom Caputi (i + 1) * sizeof (blkptr_t), DMU_OBJECT_END, tx));
270cde58dbcSMatthew Ahrens }
271cde58dbcSMatthew Ahrens if (err || !bpo->bpo_havesubobj || bpo->bpo_phys->bpo_subobjs == 0)
272cde58dbcSMatthew Ahrens goto out;
273cde58dbcSMatthew Ahrens
274cde58dbcSMatthew Ahrens ASSERT(bpo->bpo_havecomp);
275cde58dbcSMatthew Ahrens err = dmu_object_info(bpo->bpo_os, bpo->bpo_phys->bpo_subobjs, &doi);
276a0e4757dSLin Ling if (err) {
277a0e4757dSLin Ling mutex_exit(&bpo->bpo_lock);
278cde58dbcSMatthew Ahrens return (err);
279a0e4757dSLin Ling }
2802acef22dSMatthew Ahrens ASSERT3U(doi.doi_type, ==, DMU_OT_BPOBJ_SUBOBJ);
281cde58dbcSMatthew Ahrens epb = doi.doi_data_block_size / sizeof (uint64_t);
282cde58dbcSMatthew Ahrens
283cde58dbcSMatthew Ahrens for (i = bpo->bpo_phys->bpo_num_subobjs - 1; i >= 0; i--) {
284cde58dbcSMatthew Ahrens uint64_t *objarray;
285cde58dbcSMatthew Ahrens uint64_t offset, blkoff;
286cde58dbcSMatthew Ahrens bpobj_t sublist;
287cde58dbcSMatthew Ahrens uint64_t used_before, comp_before, uncomp_before;
288cde58dbcSMatthew Ahrens uint64_t used_after, comp_after, uncomp_after;
289cde58dbcSMatthew Ahrens
290cde58dbcSMatthew Ahrens offset = i * sizeof (uint64_t);
291cde58dbcSMatthew Ahrens blkoff = P2PHASE(i, epb);
292cde58dbcSMatthew Ahrens
293cde58dbcSMatthew Ahrens if (dbuf == NULL || dbuf->db_offset > offset) {
294cde58dbcSMatthew Ahrens if (dbuf)
295cde58dbcSMatthew Ahrens dmu_buf_rele(dbuf, FTAG);
296cde58dbcSMatthew Ahrens err = dmu_buf_hold(bpo->bpo_os,
297cde58dbcSMatthew Ahrens bpo->bpo_phys->bpo_subobjs, offset, FTAG, &dbuf, 0);
298cde58dbcSMatthew Ahrens if (err)
299cde58dbcSMatthew Ahrens break;
300cde58dbcSMatthew Ahrens }
301cde58dbcSMatthew Ahrens
302cde58dbcSMatthew Ahrens ASSERT3U(offset, >=, dbuf->db_offset);
303cde58dbcSMatthew Ahrens ASSERT3U(offset, <, dbuf->db_offset + dbuf->db_size);
304cde58dbcSMatthew Ahrens
305cde58dbcSMatthew Ahrens objarray = dbuf->db_data;
306cde58dbcSMatthew Ahrens err = bpobj_open(&sublist, bpo->bpo_os, objarray[blkoff]);
307cde58dbcSMatthew Ahrens if (err)
308cde58dbcSMatthew Ahrens break;
309cde58dbcSMatthew Ahrens if (free) {
310cde58dbcSMatthew Ahrens err = bpobj_space(&sublist,
311cde58dbcSMatthew Ahrens &used_before, &comp_before, &uncomp_before);
312b67dde11SWill Andrews if (err != 0) {
313b67dde11SWill Andrews bpobj_close(&sublist);
314cde58dbcSMatthew Ahrens break;
315cde58dbcSMatthew Ahrens }
316b67dde11SWill Andrews }
317cde58dbcSMatthew Ahrens err = bpobj_iterate_impl(&sublist, func, arg, tx, free);
318cde58dbcSMatthew Ahrens if (free) {
319b420f3adSRichard Lowe VERIFY3U(0, ==, bpobj_space(&sublist,
320cde58dbcSMatthew Ahrens &used_after, &comp_after, &uncomp_after));
321cde58dbcSMatthew Ahrens bpo->bpo_phys->bpo_bytes -= used_before - used_after;
322cde58dbcSMatthew Ahrens ASSERT3S(bpo->bpo_phys->bpo_bytes, >=, 0);
3234bc4cb79SMatthew Ahrens bpo->bpo_phys->bpo_comp -= comp_before - comp_after;
324cde58dbcSMatthew Ahrens bpo->bpo_phys->bpo_uncomp -=
325cde58dbcSMatthew Ahrens uncomp_before - uncomp_after;
326cde58dbcSMatthew Ahrens }
327cde58dbcSMatthew Ahrens
328cde58dbcSMatthew Ahrens bpobj_close(&sublist);
329cde58dbcSMatthew Ahrens if (err)
330cde58dbcSMatthew Ahrens break;
331cde58dbcSMatthew Ahrens if (free) {
332cde58dbcSMatthew Ahrens err = dmu_object_free(bpo->bpo_os,
333cde58dbcSMatthew Ahrens objarray[blkoff], tx);
334cde58dbcSMatthew Ahrens if (err)
335cde58dbcSMatthew Ahrens break;
336cde58dbcSMatthew Ahrens bpo->bpo_phys->bpo_num_subobjs--;
337cde58dbcSMatthew Ahrens ASSERT3S(bpo->bpo_phys->bpo_num_subobjs, >=, 0);
338cde58dbcSMatthew Ahrens }
339cde58dbcSMatthew Ahrens }
340cde58dbcSMatthew Ahrens if (dbuf) {
341cde58dbcSMatthew Ahrens dmu_buf_rele(dbuf, FTAG);
342cde58dbcSMatthew Ahrens dbuf = NULL;
343cde58dbcSMatthew Ahrens }
344cde58dbcSMatthew Ahrens if (free) {
345b420f3adSRichard Lowe VERIFY3U(0, ==, dmu_free_range(bpo->bpo_os,
346cde58dbcSMatthew Ahrens bpo->bpo_phys->bpo_subobjs,
347*eb633035STom Caputi (i + 1) * sizeof (uint64_t), DMU_OBJECT_END, tx));
348cde58dbcSMatthew Ahrens }
349cde58dbcSMatthew Ahrens
350cde58dbcSMatthew Ahrens out:
351cde58dbcSMatthew Ahrens /* If there are no entries, there should be no bytes. */
3525cabbc6bSPrashanth Sreenivasa if (bpobj_is_empty(bpo)) {
3535d7b4d43SMatthew Ahrens ASSERT0(bpo->bpo_phys->bpo_bytes);
3545d7b4d43SMatthew Ahrens ASSERT0(bpo->bpo_phys->bpo_comp);
3555d7b4d43SMatthew Ahrens ASSERT0(bpo->bpo_phys->bpo_uncomp);
3565d7b4d43SMatthew Ahrens }
357cde58dbcSMatthew Ahrens
358cde58dbcSMatthew Ahrens mutex_exit(&bpo->bpo_lock);
359cde58dbcSMatthew Ahrens return (err);
360cde58dbcSMatthew Ahrens }
361cde58dbcSMatthew Ahrens
362cde58dbcSMatthew Ahrens /*
363cde58dbcSMatthew Ahrens * Iterate and remove the entries. If func returns nonzero, iteration
364cde58dbcSMatthew Ahrens * will stop and that entry will not be removed.
365cde58dbcSMatthew Ahrens */
366cde58dbcSMatthew Ahrens int
bpobj_iterate(bpobj_t * bpo,bpobj_itor_t func,void * arg,dmu_tx_t * tx)367cde58dbcSMatthew Ahrens bpobj_iterate(bpobj_t *bpo, bpobj_itor_t func, void *arg, dmu_tx_t *tx)
368cde58dbcSMatthew Ahrens {
369cde58dbcSMatthew Ahrens return (bpobj_iterate_impl(bpo, func, arg, tx, B_TRUE));
370cde58dbcSMatthew Ahrens }
371cde58dbcSMatthew Ahrens
372cde58dbcSMatthew Ahrens /*
373cde58dbcSMatthew Ahrens * Iterate the entries. If func returns nonzero, iteration will stop.
374cde58dbcSMatthew Ahrens */
375cde58dbcSMatthew Ahrens int
bpobj_iterate_nofree(bpobj_t * bpo,bpobj_itor_t func,void * arg,dmu_tx_t * tx)376cde58dbcSMatthew Ahrens bpobj_iterate_nofree(bpobj_t *bpo, bpobj_itor_t func, void *arg, dmu_tx_t *tx)
377cde58dbcSMatthew Ahrens {
378cde58dbcSMatthew Ahrens return (bpobj_iterate_impl(bpo, func, arg, tx, B_FALSE));
379cde58dbcSMatthew Ahrens }
380cde58dbcSMatthew Ahrens
381cde58dbcSMatthew Ahrens void
bpobj_enqueue_subobj(bpobj_t * bpo,uint64_t subobj,dmu_tx_t * tx)382cde58dbcSMatthew Ahrens bpobj_enqueue_subobj(bpobj_t *bpo, uint64_t subobj, dmu_tx_t *tx)
383cde58dbcSMatthew Ahrens {
384cde58dbcSMatthew Ahrens bpobj_t subbpo;
3854bc4cb79SMatthew Ahrens uint64_t used, comp, uncomp, subsubobjs;
386cde58dbcSMatthew Ahrens
3875cabbc6bSPrashanth Sreenivasa ASSERT(bpobj_is_open(bpo));
3885cabbc6bSPrashanth Sreenivasa ASSERT(subobj != 0);
389cde58dbcSMatthew Ahrens ASSERT(bpo->bpo_havesubobj);
390cde58dbcSMatthew Ahrens ASSERT(bpo->bpo_havecomp);
391f1745736SMatthew Ahrens ASSERT(bpo->bpo_object != dmu_objset_pool(bpo->bpo_os)->dp_empty_bpobj);
392f1745736SMatthew Ahrens
393f1745736SMatthew Ahrens if (subobj == dmu_objset_pool(bpo->bpo_os)->dp_empty_bpobj) {
394f1745736SMatthew Ahrens bpobj_decr_empty(bpo->bpo_os, tx);
395f1745736SMatthew Ahrens return;
396f1745736SMatthew Ahrens }
397cde58dbcSMatthew Ahrens
398b420f3adSRichard Lowe VERIFY3U(0, ==, bpobj_open(&subbpo, bpo->bpo_os, subobj));
399b420f3adSRichard Lowe VERIFY3U(0, ==, bpobj_space(&subbpo, &used, &comp, &uncomp));
400cde58dbcSMatthew Ahrens
4015cabbc6bSPrashanth Sreenivasa if (bpobj_is_empty(&subbpo)) {
402cde58dbcSMatthew Ahrens /* No point in having an empty subobj. */
4034bc4cb79SMatthew Ahrens bpobj_close(&subbpo);
404cde58dbcSMatthew Ahrens bpobj_free(bpo->bpo_os, subobj, tx);
405cde58dbcSMatthew Ahrens return;
406cde58dbcSMatthew Ahrens }
407cde58dbcSMatthew Ahrens
408a3905a45SSerapheim Dimitropoulos mutex_enter(&bpo->bpo_lock);
409cde58dbcSMatthew Ahrens dmu_buf_will_dirty(bpo->bpo_dbuf, tx);
410cde58dbcSMatthew Ahrens if (bpo->bpo_phys->bpo_subobjs == 0) {
411cde58dbcSMatthew Ahrens bpo->bpo_phys->bpo_subobjs = dmu_object_alloc(bpo->bpo_os,
412b5152584SMatthew Ahrens DMU_OT_BPOBJ_SUBOBJ, SPA_OLD_MAXBLOCKSIZE,
413b5152584SMatthew Ahrens DMU_OT_NONE, 0, tx);
414cde58dbcSMatthew Ahrens }
415cde58dbcSMatthew Ahrens
4163b2aab18SMatthew Ahrens dmu_object_info_t doi;
4173b2aab18SMatthew Ahrens ASSERT0(dmu_object_info(bpo->bpo_os, bpo->bpo_phys->bpo_subobjs, &doi));
4183b2aab18SMatthew Ahrens ASSERT3U(doi.doi_type, ==, DMU_OT_BPOBJ_SUBOBJ);
4193b2aab18SMatthew Ahrens
420cde58dbcSMatthew Ahrens dmu_write(bpo->bpo_os, bpo->bpo_phys->bpo_subobjs,
421cde58dbcSMatthew Ahrens bpo->bpo_phys->bpo_num_subobjs * sizeof (subobj),
422cde58dbcSMatthew Ahrens sizeof (subobj), &subobj, tx);
423cde58dbcSMatthew Ahrens bpo->bpo_phys->bpo_num_subobjs++;
4244bc4cb79SMatthew Ahrens
4254bc4cb79SMatthew Ahrens /*
4264bc4cb79SMatthew Ahrens * If subobj has only one block of subobjs, then move subobj's
4274bc4cb79SMatthew Ahrens * subobjs to bpo's subobj list directly. This reduces
4284bc4cb79SMatthew Ahrens * recursion in bpobj_iterate due to nested subobjs.
4294bc4cb79SMatthew Ahrens */
4304bc4cb79SMatthew Ahrens subsubobjs = subbpo.bpo_phys->bpo_subobjs;
4314bc4cb79SMatthew Ahrens if (subsubobjs != 0) {
4324bc4cb79SMatthew Ahrens dmu_object_info_t doi;
4334bc4cb79SMatthew Ahrens
434b420f3adSRichard Lowe VERIFY3U(0, ==, dmu_object_info(bpo->bpo_os, subsubobjs, &doi));
4354bc4cb79SMatthew Ahrens if (doi.doi_max_offset == doi.doi_data_block_size) {
4364bc4cb79SMatthew Ahrens dmu_buf_t *subdb;
4374bc4cb79SMatthew Ahrens uint64_t numsubsub = subbpo.bpo_phys->bpo_num_subobjs;
4384bc4cb79SMatthew Ahrens
439b420f3adSRichard Lowe VERIFY3U(0, ==, dmu_buf_hold(bpo->bpo_os, subsubobjs,
4404bc4cb79SMatthew Ahrens 0, FTAG, &subdb, 0));
441d0475637SMatthew Ahrens /*
442d0475637SMatthew Ahrens * Make sure that we are not asking dmu_write()
443d0475637SMatthew Ahrens * to write more data than we have in our buffer.
444d0475637SMatthew Ahrens */
445d0475637SMatthew Ahrens VERIFY3U(subdb->db_size, >=,
446d0475637SMatthew Ahrens numsubsub * sizeof (subobj));
4474bc4cb79SMatthew Ahrens dmu_write(bpo->bpo_os, bpo->bpo_phys->bpo_subobjs,
4484bc4cb79SMatthew Ahrens bpo->bpo_phys->bpo_num_subobjs * sizeof (subobj),
4494bc4cb79SMatthew Ahrens numsubsub * sizeof (subobj), subdb->db_data, tx);
4504bc4cb79SMatthew Ahrens dmu_buf_rele(subdb, FTAG);
4514bc4cb79SMatthew Ahrens bpo->bpo_phys->bpo_num_subobjs += numsubsub;
4524bc4cb79SMatthew Ahrens
4534bc4cb79SMatthew Ahrens dmu_buf_will_dirty(subbpo.bpo_dbuf, tx);
4544bc4cb79SMatthew Ahrens subbpo.bpo_phys->bpo_subobjs = 0;
455b420f3adSRichard Lowe VERIFY3U(0, ==, dmu_object_free(bpo->bpo_os,
4564bc4cb79SMatthew Ahrens subsubobjs, tx));
4574bc4cb79SMatthew Ahrens }
4584bc4cb79SMatthew Ahrens }
459cde58dbcSMatthew Ahrens bpo->bpo_phys->bpo_bytes += used;
460cde58dbcSMatthew Ahrens bpo->bpo_phys->bpo_comp += comp;
461cde58dbcSMatthew Ahrens bpo->bpo_phys->bpo_uncomp += uncomp;
462cde58dbcSMatthew Ahrens mutex_exit(&bpo->bpo_lock);
4634bc4cb79SMatthew Ahrens
4644bc4cb79SMatthew Ahrens bpobj_close(&subbpo);
465cde58dbcSMatthew Ahrens }
466cde58dbcSMatthew Ahrens
467cde58dbcSMatthew Ahrens void
bpobj_enqueue(bpobj_t * bpo,const blkptr_t * bp,dmu_tx_t * tx)468cde58dbcSMatthew Ahrens bpobj_enqueue(bpobj_t *bpo, const blkptr_t *bp, dmu_tx_t *tx)
469cde58dbcSMatthew Ahrens {
470cde58dbcSMatthew Ahrens blkptr_t stored_bp = *bp;
471cde58dbcSMatthew Ahrens uint64_t offset;
472cde58dbcSMatthew Ahrens int blkoff;
473cde58dbcSMatthew Ahrens blkptr_t *bparray;
474cde58dbcSMatthew Ahrens
4755cabbc6bSPrashanth Sreenivasa ASSERT(bpobj_is_open(bpo));
476cde58dbcSMatthew Ahrens ASSERT(!BP_IS_HOLE(bp));
477f1745736SMatthew Ahrens ASSERT(bpo->bpo_object != dmu_objset_pool(bpo->bpo_os)->dp_empty_bpobj);
478cde58dbcSMatthew Ahrens
4795d7b4d43SMatthew Ahrens if (BP_IS_EMBEDDED(bp)) {
4805d7b4d43SMatthew Ahrens /*
4815d7b4d43SMatthew Ahrens * The bpobj will compress better without the payload.
4825d7b4d43SMatthew Ahrens *
4835d7b4d43SMatthew Ahrens * Note that we store EMBEDDED bp's because they have an
4845d7b4d43SMatthew Ahrens * uncompressed size, which must be accounted for. An
4855d7b4d43SMatthew Ahrens * alternative would be to add their size to bpo_uncomp
4865d7b4d43SMatthew Ahrens * without storing the bp, but that would create additional
4875d7b4d43SMatthew Ahrens * complications: bpo_uncomp would be inconsistent with the
4885d7b4d43SMatthew Ahrens * set of BP's stored, and bpobj_iterate() wouldn't visit
4895d7b4d43SMatthew Ahrens * all the space accounted for in the bpobj.
4905d7b4d43SMatthew Ahrens */
4915d7b4d43SMatthew Ahrens bzero(&stored_bp, sizeof (stored_bp));
4925d7b4d43SMatthew Ahrens stored_bp.blk_prop = bp->blk_prop;
4935d7b4d43SMatthew Ahrens stored_bp.blk_birth = bp->blk_birth;
4945d7b4d43SMatthew Ahrens } else if (!BP_GET_DEDUP(bp)) {
4955d7b4d43SMatthew Ahrens /* The bpobj will compress better without the checksum */
4965d7b4d43SMatthew Ahrens bzero(&stored_bp.blk_cksum, sizeof (stored_bp.blk_cksum));
4975d7b4d43SMatthew Ahrens }
4985d7b4d43SMatthew Ahrens
499cde58dbcSMatthew Ahrens /* We never need the fill count. */
500cde58dbcSMatthew Ahrens stored_bp.blk_fill = 0;
501cde58dbcSMatthew Ahrens
502cde58dbcSMatthew Ahrens mutex_enter(&bpo->bpo_lock);
503cde58dbcSMatthew Ahrens
504cde58dbcSMatthew Ahrens offset = bpo->bpo_phys->bpo_num_blkptrs * sizeof (stored_bp);
505cde58dbcSMatthew Ahrens blkoff = P2PHASE(bpo->bpo_phys->bpo_num_blkptrs, bpo->bpo_epb);
506cde58dbcSMatthew Ahrens
507cde58dbcSMatthew Ahrens if (bpo->bpo_cached_dbuf == NULL ||
508cde58dbcSMatthew Ahrens offset < bpo->bpo_cached_dbuf->db_offset ||
509cde58dbcSMatthew Ahrens offset >= bpo->bpo_cached_dbuf->db_offset +
510cde58dbcSMatthew Ahrens bpo->bpo_cached_dbuf->db_size) {
511cde58dbcSMatthew Ahrens if (bpo->bpo_cached_dbuf)
512cde58dbcSMatthew Ahrens dmu_buf_rele(bpo->bpo_cached_dbuf, bpo);
513b420f3adSRichard Lowe VERIFY3U(0, ==, dmu_buf_hold(bpo->bpo_os, bpo->bpo_object,
514cde58dbcSMatthew Ahrens offset, bpo, &bpo->bpo_cached_dbuf, 0));
515cde58dbcSMatthew Ahrens }
516cde58dbcSMatthew Ahrens
517cde58dbcSMatthew Ahrens dmu_buf_will_dirty(bpo->bpo_cached_dbuf, tx);
518cde58dbcSMatthew Ahrens bparray = bpo->bpo_cached_dbuf->db_data;
519cde58dbcSMatthew Ahrens bparray[blkoff] = stored_bp;
520cde58dbcSMatthew Ahrens
521cde58dbcSMatthew Ahrens dmu_buf_will_dirty(bpo->bpo_dbuf, tx);
522cde58dbcSMatthew Ahrens bpo->bpo_phys->bpo_num_blkptrs++;
523cde58dbcSMatthew Ahrens bpo->bpo_phys->bpo_bytes +=
524cde58dbcSMatthew Ahrens bp_get_dsize_sync(dmu_objset_spa(bpo->bpo_os), bp);
525cde58dbcSMatthew Ahrens if (bpo->bpo_havecomp) {
526cde58dbcSMatthew Ahrens bpo->bpo_phys->bpo_comp += BP_GET_PSIZE(bp);
527cde58dbcSMatthew Ahrens bpo->bpo_phys->bpo_uncomp += BP_GET_UCSIZE(bp);
528cde58dbcSMatthew Ahrens }
529cde58dbcSMatthew Ahrens mutex_exit(&bpo->bpo_lock);
530cde58dbcSMatthew Ahrens }
531cde58dbcSMatthew Ahrens
532cde58dbcSMatthew Ahrens struct space_range_arg {
533cde58dbcSMatthew Ahrens spa_t *spa;
534cde58dbcSMatthew Ahrens uint64_t mintxg;
535cde58dbcSMatthew Ahrens uint64_t maxtxg;
536cde58dbcSMatthew Ahrens uint64_t used;
537cde58dbcSMatthew Ahrens uint64_t comp;
538cde58dbcSMatthew Ahrens uint64_t uncomp;
539cde58dbcSMatthew Ahrens };
540cde58dbcSMatthew Ahrens
541cde58dbcSMatthew Ahrens /* ARGSUSED */
542cde58dbcSMatthew Ahrens static int
space_range_cb(void * arg,const blkptr_t * bp,dmu_tx_t * tx)543cde58dbcSMatthew Ahrens space_range_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
544cde58dbcSMatthew Ahrens {
545cde58dbcSMatthew Ahrens struct space_range_arg *sra = arg;
546cde58dbcSMatthew Ahrens
547cde58dbcSMatthew Ahrens if (bp->blk_birth > sra->mintxg && bp->blk_birth <= sra->maxtxg) {
54819b94df9SMatthew Ahrens if (dsl_pool_sync_context(spa_get_dsl(sra->spa)))
549cde58dbcSMatthew Ahrens sra->used += bp_get_dsize_sync(sra->spa, bp);
55019b94df9SMatthew Ahrens else
55119b94df9SMatthew Ahrens sra->used += bp_get_dsize(sra->spa, bp);
552cde58dbcSMatthew Ahrens sra->comp += BP_GET_PSIZE(bp);
553cde58dbcSMatthew Ahrens sra->uncomp += BP_GET_UCSIZE(bp);
554cde58dbcSMatthew Ahrens }
555cde58dbcSMatthew Ahrens return (0);
556cde58dbcSMatthew Ahrens }
557cde58dbcSMatthew Ahrens
558cde58dbcSMatthew Ahrens int
bpobj_space(bpobj_t * bpo,uint64_t * usedp,uint64_t * compp,uint64_t * uncompp)559cde58dbcSMatthew Ahrens bpobj_space(bpobj_t *bpo, uint64_t *usedp, uint64_t *compp, uint64_t *uncompp)
560cde58dbcSMatthew Ahrens {
5615cabbc6bSPrashanth Sreenivasa ASSERT(bpobj_is_open(bpo));
562cde58dbcSMatthew Ahrens mutex_enter(&bpo->bpo_lock);
563cde58dbcSMatthew Ahrens
564cde58dbcSMatthew Ahrens *usedp = bpo->bpo_phys->bpo_bytes;
565cde58dbcSMatthew Ahrens if (bpo->bpo_havecomp) {
566cde58dbcSMatthew Ahrens *compp = bpo->bpo_phys->bpo_comp;
567cde58dbcSMatthew Ahrens *uncompp = bpo->bpo_phys->bpo_uncomp;
568cde58dbcSMatthew Ahrens mutex_exit(&bpo->bpo_lock);
569cde58dbcSMatthew Ahrens return (0);
570cde58dbcSMatthew Ahrens } else {
571cde58dbcSMatthew Ahrens mutex_exit(&bpo->bpo_lock);
572cde58dbcSMatthew Ahrens return (bpobj_space_range(bpo, 0, UINT64_MAX,
573cde58dbcSMatthew Ahrens usedp, compp, uncompp));
574cde58dbcSMatthew Ahrens }
575cde58dbcSMatthew Ahrens }
576cde58dbcSMatthew Ahrens
577cde58dbcSMatthew Ahrens /*
578cde58dbcSMatthew Ahrens * Return the amount of space in the bpobj which is:
579cde58dbcSMatthew Ahrens * mintxg < blk_birth <= maxtxg
580cde58dbcSMatthew Ahrens */
581cde58dbcSMatthew Ahrens int
bpobj_space_range(bpobj_t * bpo,uint64_t mintxg,uint64_t maxtxg,uint64_t * usedp,uint64_t * compp,uint64_t * uncompp)582cde58dbcSMatthew Ahrens bpobj_space_range(bpobj_t *bpo, uint64_t mintxg, uint64_t maxtxg,
583cde58dbcSMatthew Ahrens uint64_t *usedp, uint64_t *compp, uint64_t *uncompp)
584cde58dbcSMatthew Ahrens {
585cde58dbcSMatthew Ahrens struct space_range_arg sra = { 0 };
586cde58dbcSMatthew Ahrens int err;
587cde58dbcSMatthew Ahrens
5885cabbc6bSPrashanth Sreenivasa ASSERT(bpobj_is_open(bpo));
5895cabbc6bSPrashanth Sreenivasa
590cde58dbcSMatthew Ahrens /*
591cde58dbcSMatthew Ahrens * As an optimization, if they want the whole txg range, just
592cde58dbcSMatthew Ahrens * get bpo_bytes rather than iterating over the bps.
593cde58dbcSMatthew Ahrens */
594cde58dbcSMatthew Ahrens if (mintxg < TXG_INITIAL && maxtxg == UINT64_MAX && bpo->bpo_havecomp)
595cde58dbcSMatthew Ahrens return (bpobj_space(bpo, usedp, compp, uncompp));
596cde58dbcSMatthew Ahrens
597cde58dbcSMatthew Ahrens sra.spa = dmu_objset_spa(bpo->bpo_os);
598cde58dbcSMatthew Ahrens sra.mintxg = mintxg;
599cde58dbcSMatthew Ahrens sra.maxtxg = maxtxg;
600cde58dbcSMatthew Ahrens
601cde58dbcSMatthew Ahrens err = bpobj_iterate_nofree(bpo, space_range_cb, &sra, NULL);
602cde58dbcSMatthew Ahrens *usedp = sra.used;
603cde58dbcSMatthew Ahrens *compp = sra.comp;
604cde58dbcSMatthew Ahrens *uncompp = sra.uncomp;
605cde58dbcSMatthew Ahrens return (err);
606cde58dbcSMatthew Ahrens }
607