xref: /onnv-gate/usr/src/uts/common/fs/zfs/dsl_deadlist.c (revision 12470:54258108784b)
1*12470SMatthew.Ahrens@Sun.COM /*
2*12470SMatthew.Ahrens@Sun.COM  * CDDL HEADER START
3*12470SMatthew.Ahrens@Sun.COM  *
4*12470SMatthew.Ahrens@Sun.COM  * The contents of this file are subject to the terms of the
5*12470SMatthew.Ahrens@Sun.COM  * Common Development and Distribution License (the "License").
6*12470SMatthew.Ahrens@Sun.COM  * You may not use this file except in compliance with the License.
7*12470SMatthew.Ahrens@Sun.COM  *
8*12470SMatthew.Ahrens@Sun.COM  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9*12470SMatthew.Ahrens@Sun.COM  * or http://www.opensolaris.org/os/licensing.
10*12470SMatthew.Ahrens@Sun.COM  * See the License for the specific language governing permissions
11*12470SMatthew.Ahrens@Sun.COM  * and limitations under the License.
12*12470SMatthew.Ahrens@Sun.COM  *
13*12470SMatthew.Ahrens@Sun.COM  * When distributing Covered Code, include this CDDL HEADER in each
14*12470SMatthew.Ahrens@Sun.COM  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15*12470SMatthew.Ahrens@Sun.COM  * If applicable, add the following below this CDDL HEADER, with the
16*12470SMatthew.Ahrens@Sun.COM  * fields enclosed by brackets "[]" replaced with your own identifying
17*12470SMatthew.Ahrens@Sun.COM  * information: Portions Copyright [yyyy] [name of copyright owner]
18*12470SMatthew.Ahrens@Sun.COM  *
19*12470SMatthew.Ahrens@Sun.COM  * CDDL HEADER END
20*12470SMatthew.Ahrens@Sun.COM  */
21*12470SMatthew.Ahrens@Sun.COM /*
22*12470SMatthew.Ahrens@Sun.COM  * Copyright (c) 2010, Oracle and/or its affiliates. All rights reserved.
23*12470SMatthew.Ahrens@Sun.COM  */
24*12470SMatthew.Ahrens@Sun.COM 
25*12470SMatthew.Ahrens@Sun.COM #include <sys/dsl_dataset.h>
26*12470SMatthew.Ahrens@Sun.COM #include <sys/dmu.h>
27*12470SMatthew.Ahrens@Sun.COM #include <sys/refcount.h>
28*12470SMatthew.Ahrens@Sun.COM #include <sys/zap.h>
29*12470SMatthew.Ahrens@Sun.COM #include <sys/zfs_context.h>
30*12470SMatthew.Ahrens@Sun.COM #include <sys/dsl_pool.h>
31*12470SMatthew.Ahrens@Sun.COM 
32*12470SMatthew.Ahrens@Sun.COM static int
dsl_deadlist_compare(const void * arg1,const void * arg2)33*12470SMatthew.Ahrens@Sun.COM dsl_deadlist_compare(const void *arg1, const void *arg2)
34*12470SMatthew.Ahrens@Sun.COM {
35*12470SMatthew.Ahrens@Sun.COM 	const dsl_deadlist_entry_t *dle1 = arg1;
36*12470SMatthew.Ahrens@Sun.COM 	const dsl_deadlist_entry_t *dle2 = arg2;
37*12470SMatthew.Ahrens@Sun.COM 
38*12470SMatthew.Ahrens@Sun.COM 	if (dle1->dle_mintxg < dle2->dle_mintxg)
39*12470SMatthew.Ahrens@Sun.COM 		return (-1);
40*12470SMatthew.Ahrens@Sun.COM 	else if (dle1->dle_mintxg > dle2->dle_mintxg)
41*12470SMatthew.Ahrens@Sun.COM 		return (+1);
42*12470SMatthew.Ahrens@Sun.COM 	else
43*12470SMatthew.Ahrens@Sun.COM 		return (0);
44*12470SMatthew.Ahrens@Sun.COM }
45*12470SMatthew.Ahrens@Sun.COM 
46*12470SMatthew.Ahrens@Sun.COM static void
dsl_deadlist_load_tree(dsl_deadlist_t * dl)47*12470SMatthew.Ahrens@Sun.COM dsl_deadlist_load_tree(dsl_deadlist_t *dl)
48*12470SMatthew.Ahrens@Sun.COM {
49*12470SMatthew.Ahrens@Sun.COM 	zap_cursor_t zc;
50*12470SMatthew.Ahrens@Sun.COM 	zap_attribute_t za;
51*12470SMatthew.Ahrens@Sun.COM 
52*12470SMatthew.Ahrens@Sun.COM 	ASSERT(!dl->dl_oldfmt);
53*12470SMatthew.Ahrens@Sun.COM 	if (dl->dl_havetree)
54*12470SMatthew.Ahrens@Sun.COM 		return;
55*12470SMatthew.Ahrens@Sun.COM 
56*12470SMatthew.Ahrens@Sun.COM 	avl_create(&dl->dl_tree, dsl_deadlist_compare,
57*12470SMatthew.Ahrens@Sun.COM 	    sizeof (dsl_deadlist_entry_t),
58*12470SMatthew.Ahrens@Sun.COM 	    offsetof(dsl_deadlist_entry_t, dle_node));
59*12470SMatthew.Ahrens@Sun.COM 	for (zap_cursor_init(&zc, dl->dl_os, dl->dl_object);
60*12470SMatthew.Ahrens@Sun.COM 	    zap_cursor_retrieve(&zc, &za) == 0;
61*12470SMatthew.Ahrens@Sun.COM 	    zap_cursor_advance(&zc)) {
62*12470SMatthew.Ahrens@Sun.COM 		dsl_deadlist_entry_t *dle = kmem_alloc(sizeof (*dle), KM_SLEEP);
63*12470SMatthew.Ahrens@Sun.COM 		dle->dle_mintxg = strtonum(za.za_name, NULL);
64*12470SMatthew.Ahrens@Sun.COM 		VERIFY3U(0, ==, bpobj_open(&dle->dle_bpobj, dl->dl_os,
65*12470SMatthew.Ahrens@Sun.COM 		    za.za_first_integer));
66*12470SMatthew.Ahrens@Sun.COM 		avl_add(&dl->dl_tree, dle);
67*12470SMatthew.Ahrens@Sun.COM 	}
68*12470SMatthew.Ahrens@Sun.COM 	zap_cursor_fini(&zc);
69*12470SMatthew.Ahrens@Sun.COM 	dl->dl_havetree = B_TRUE;
70*12470SMatthew.Ahrens@Sun.COM }
71*12470SMatthew.Ahrens@Sun.COM 
72*12470SMatthew.Ahrens@Sun.COM void
dsl_deadlist_open(dsl_deadlist_t * dl,objset_t * os,uint64_t object)73*12470SMatthew.Ahrens@Sun.COM dsl_deadlist_open(dsl_deadlist_t *dl, objset_t *os, uint64_t object)
74*12470SMatthew.Ahrens@Sun.COM {
75*12470SMatthew.Ahrens@Sun.COM 	dmu_object_info_t doi;
76*12470SMatthew.Ahrens@Sun.COM 
77*12470SMatthew.Ahrens@Sun.COM 	mutex_init(&dl->dl_lock, NULL, MUTEX_DEFAULT, NULL);
78*12470SMatthew.Ahrens@Sun.COM 	dl->dl_os = os;
79*12470SMatthew.Ahrens@Sun.COM 	dl->dl_object = object;
80*12470SMatthew.Ahrens@Sun.COM 	VERIFY3U(0, ==, dmu_bonus_hold(os, object, dl, &dl->dl_dbuf));
81*12470SMatthew.Ahrens@Sun.COM 	dmu_object_info_from_db(dl->dl_dbuf, &doi);
82*12470SMatthew.Ahrens@Sun.COM 	if (doi.doi_type == DMU_OT_BPOBJ) {
83*12470SMatthew.Ahrens@Sun.COM 		dmu_buf_rele(dl->dl_dbuf, dl);
84*12470SMatthew.Ahrens@Sun.COM 		dl->dl_dbuf = NULL;
85*12470SMatthew.Ahrens@Sun.COM 		dl->dl_oldfmt = B_TRUE;
86*12470SMatthew.Ahrens@Sun.COM 		VERIFY3U(0, ==, bpobj_open(&dl->dl_bpobj, os, object));
87*12470SMatthew.Ahrens@Sun.COM 		return;
88*12470SMatthew.Ahrens@Sun.COM 	}
89*12470SMatthew.Ahrens@Sun.COM 
90*12470SMatthew.Ahrens@Sun.COM 	dl->dl_oldfmt = B_FALSE;
91*12470SMatthew.Ahrens@Sun.COM 	dl->dl_phys = dl->dl_dbuf->db_data;
92*12470SMatthew.Ahrens@Sun.COM 	dl->dl_havetree = B_FALSE;
93*12470SMatthew.Ahrens@Sun.COM }
94*12470SMatthew.Ahrens@Sun.COM 
95*12470SMatthew.Ahrens@Sun.COM void
dsl_deadlist_close(dsl_deadlist_t * dl)96*12470SMatthew.Ahrens@Sun.COM dsl_deadlist_close(dsl_deadlist_t *dl)
97*12470SMatthew.Ahrens@Sun.COM {
98*12470SMatthew.Ahrens@Sun.COM 	void *cookie = NULL;
99*12470SMatthew.Ahrens@Sun.COM 	dsl_deadlist_entry_t *dle;
100*12470SMatthew.Ahrens@Sun.COM 
101*12470SMatthew.Ahrens@Sun.COM 	if (dl->dl_oldfmt) {
102*12470SMatthew.Ahrens@Sun.COM 		dl->dl_oldfmt = B_FALSE;
103*12470SMatthew.Ahrens@Sun.COM 		bpobj_close(&dl->dl_bpobj);
104*12470SMatthew.Ahrens@Sun.COM 		return;
105*12470SMatthew.Ahrens@Sun.COM 	}
106*12470SMatthew.Ahrens@Sun.COM 
107*12470SMatthew.Ahrens@Sun.COM 	if (dl->dl_havetree) {
108*12470SMatthew.Ahrens@Sun.COM 		while ((dle = avl_destroy_nodes(&dl->dl_tree, &cookie))
109*12470SMatthew.Ahrens@Sun.COM 		    != NULL) {
110*12470SMatthew.Ahrens@Sun.COM 			bpobj_close(&dle->dle_bpobj);
111*12470SMatthew.Ahrens@Sun.COM 			kmem_free(dle, sizeof (*dle));
112*12470SMatthew.Ahrens@Sun.COM 		}
113*12470SMatthew.Ahrens@Sun.COM 		avl_destroy(&dl->dl_tree);
114*12470SMatthew.Ahrens@Sun.COM 	}
115*12470SMatthew.Ahrens@Sun.COM 	dmu_buf_rele(dl->dl_dbuf, dl);
116*12470SMatthew.Ahrens@Sun.COM 	mutex_destroy(&dl->dl_lock);
117*12470SMatthew.Ahrens@Sun.COM 	dl->dl_dbuf = NULL;
118*12470SMatthew.Ahrens@Sun.COM 	dl->dl_phys = NULL;
119*12470SMatthew.Ahrens@Sun.COM }
120*12470SMatthew.Ahrens@Sun.COM 
121*12470SMatthew.Ahrens@Sun.COM uint64_t
dsl_deadlist_alloc(objset_t * os,dmu_tx_t * tx)122*12470SMatthew.Ahrens@Sun.COM dsl_deadlist_alloc(objset_t *os, dmu_tx_t *tx)
123*12470SMatthew.Ahrens@Sun.COM {
124*12470SMatthew.Ahrens@Sun.COM 	if (spa_version(dmu_objset_spa(os)) < SPA_VERSION_DEADLISTS)
125*12470SMatthew.Ahrens@Sun.COM 		return (bpobj_alloc(os, SPA_MAXBLOCKSIZE, tx));
126*12470SMatthew.Ahrens@Sun.COM 	return (zap_create(os, DMU_OT_DEADLIST, DMU_OT_DEADLIST_HDR,
127*12470SMatthew.Ahrens@Sun.COM 	    sizeof (dsl_deadlist_phys_t), tx));
128*12470SMatthew.Ahrens@Sun.COM }
129*12470SMatthew.Ahrens@Sun.COM 
130*12470SMatthew.Ahrens@Sun.COM void
dsl_deadlist_free(objset_t * os,uint64_t dlobj,dmu_tx_t * tx)131*12470SMatthew.Ahrens@Sun.COM dsl_deadlist_free(objset_t *os, uint64_t dlobj, dmu_tx_t *tx)
132*12470SMatthew.Ahrens@Sun.COM {
133*12470SMatthew.Ahrens@Sun.COM 	dmu_object_info_t doi;
134*12470SMatthew.Ahrens@Sun.COM 	zap_cursor_t zc;
135*12470SMatthew.Ahrens@Sun.COM 	zap_attribute_t za;
136*12470SMatthew.Ahrens@Sun.COM 
137*12470SMatthew.Ahrens@Sun.COM 	VERIFY3U(0, ==, dmu_object_info(os, dlobj, &doi));
138*12470SMatthew.Ahrens@Sun.COM 	if (doi.doi_type == DMU_OT_BPOBJ) {
139*12470SMatthew.Ahrens@Sun.COM 		bpobj_free(os, dlobj, tx);
140*12470SMatthew.Ahrens@Sun.COM 		return;
141*12470SMatthew.Ahrens@Sun.COM 	}
142*12470SMatthew.Ahrens@Sun.COM 
143*12470SMatthew.Ahrens@Sun.COM 	for (zap_cursor_init(&zc, os, dlobj);
144*12470SMatthew.Ahrens@Sun.COM 	    zap_cursor_retrieve(&zc, &za) == 0;
145*12470SMatthew.Ahrens@Sun.COM 	    zap_cursor_advance(&zc))
146*12470SMatthew.Ahrens@Sun.COM 		bpobj_free(os, za.za_first_integer, tx);
147*12470SMatthew.Ahrens@Sun.COM 	zap_cursor_fini(&zc);
148*12470SMatthew.Ahrens@Sun.COM 	VERIFY3U(0, ==, dmu_object_free(os, dlobj, tx));
149*12470SMatthew.Ahrens@Sun.COM }
150*12470SMatthew.Ahrens@Sun.COM 
151*12470SMatthew.Ahrens@Sun.COM void
dsl_deadlist_insert(dsl_deadlist_t * dl,const blkptr_t * bp,dmu_tx_t * tx)152*12470SMatthew.Ahrens@Sun.COM dsl_deadlist_insert(dsl_deadlist_t *dl, const blkptr_t *bp, dmu_tx_t *tx)
153*12470SMatthew.Ahrens@Sun.COM {
154*12470SMatthew.Ahrens@Sun.COM 	dsl_deadlist_entry_t dle_tofind;
155*12470SMatthew.Ahrens@Sun.COM 	dsl_deadlist_entry_t *dle;
156*12470SMatthew.Ahrens@Sun.COM 	avl_index_t where;
157*12470SMatthew.Ahrens@Sun.COM 
158*12470SMatthew.Ahrens@Sun.COM 	if (dl->dl_oldfmt) {
159*12470SMatthew.Ahrens@Sun.COM 		bpobj_enqueue(&dl->dl_bpobj, bp, tx);
160*12470SMatthew.Ahrens@Sun.COM 		return;
161*12470SMatthew.Ahrens@Sun.COM 	}
162*12470SMatthew.Ahrens@Sun.COM 
163*12470SMatthew.Ahrens@Sun.COM 	dsl_deadlist_load_tree(dl);
164*12470SMatthew.Ahrens@Sun.COM 
165*12470SMatthew.Ahrens@Sun.COM 	dmu_buf_will_dirty(dl->dl_dbuf, tx);
166*12470SMatthew.Ahrens@Sun.COM 	mutex_enter(&dl->dl_lock);
167*12470SMatthew.Ahrens@Sun.COM 	dl->dl_phys->dl_used +=
168*12470SMatthew.Ahrens@Sun.COM 	    bp_get_dsize_sync(dmu_objset_spa(dl->dl_os), bp);
169*12470SMatthew.Ahrens@Sun.COM 	dl->dl_phys->dl_comp += BP_GET_PSIZE(bp);
170*12470SMatthew.Ahrens@Sun.COM 	dl->dl_phys->dl_uncomp += BP_GET_UCSIZE(bp);
171*12470SMatthew.Ahrens@Sun.COM 	mutex_exit(&dl->dl_lock);
172*12470SMatthew.Ahrens@Sun.COM 
173*12470SMatthew.Ahrens@Sun.COM 	dle_tofind.dle_mintxg = bp->blk_birth;
174*12470SMatthew.Ahrens@Sun.COM 	dle = avl_find(&dl->dl_tree, &dle_tofind, &where);
175*12470SMatthew.Ahrens@Sun.COM 	if (dle == NULL)
176*12470SMatthew.Ahrens@Sun.COM 		dle = avl_nearest(&dl->dl_tree, where, AVL_BEFORE);
177*12470SMatthew.Ahrens@Sun.COM 	else
178*12470SMatthew.Ahrens@Sun.COM 		dle = AVL_PREV(&dl->dl_tree, dle);
179*12470SMatthew.Ahrens@Sun.COM 	bpobj_enqueue(&dle->dle_bpobj, bp, tx);
180*12470SMatthew.Ahrens@Sun.COM }
181*12470SMatthew.Ahrens@Sun.COM 
182*12470SMatthew.Ahrens@Sun.COM /*
183*12470SMatthew.Ahrens@Sun.COM  * Insert new key in deadlist, which must be > all current entries.
184*12470SMatthew.Ahrens@Sun.COM  * mintxg is not inclusive.
185*12470SMatthew.Ahrens@Sun.COM  */
186*12470SMatthew.Ahrens@Sun.COM void
dsl_deadlist_add_key(dsl_deadlist_t * dl,uint64_t mintxg,dmu_tx_t * tx)187*12470SMatthew.Ahrens@Sun.COM dsl_deadlist_add_key(dsl_deadlist_t *dl, uint64_t mintxg, dmu_tx_t *tx)
188*12470SMatthew.Ahrens@Sun.COM {
189*12470SMatthew.Ahrens@Sun.COM 	uint64_t obj;
190*12470SMatthew.Ahrens@Sun.COM 	dsl_deadlist_entry_t *dle;
191*12470SMatthew.Ahrens@Sun.COM 
192*12470SMatthew.Ahrens@Sun.COM 	if (dl->dl_oldfmt)
193*12470SMatthew.Ahrens@Sun.COM 		return;
194*12470SMatthew.Ahrens@Sun.COM 
195*12470SMatthew.Ahrens@Sun.COM 	dsl_deadlist_load_tree(dl);
196*12470SMatthew.Ahrens@Sun.COM 
197*12470SMatthew.Ahrens@Sun.COM 	dle = kmem_alloc(sizeof (*dle), KM_SLEEP);
198*12470SMatthew.Ahrens@Sun.COM 	dle->dle_mintxg = mintxg;
199*12470SMatthew.Ahrens@Sun.COM 	obj = bpobj_alloc(dl->dl_os, SPA_MAXBLOCKSIZE, tx);
200*12470SMatthew.Ahrens@Sun.COM 	VERIFY3U(0, ==, bpobj_open(&dle->dle_bpobj, dl->dl_os, obj));
201*12470SMatthew.Ahrens@Sun.COM 	avl_add(&dl->dl_tree, dle);
202*12470SMatthew.Ahrens@Sun.COM 
203*12470SMatthew.Ahrens@Sun.COM 	VERIFY3U(0, ==, zap_add_int_key(dl->dl_os, dl->dl_object,
204*12470SMatthew.Ahrens@Sun.COM 	    mintxg, obj, tx));
205*12470SMatthew.Ahrens@Sun.COM }
206*12470SMatthew.Ahrens@Sun.COM 
207*12470SMatthew.Ahrens@Sun.COM /*
208*12470SMatthew.Ahrens@Sun.COM  * Remove this key, merging its entries into the previous key.
209*12470SMatthew.Ahrens@Sun.COM  */
210*12470SMatthew.Ahrens@Sun.COM void
dsl_deadlist_remove_key(dsl_deadlist_t * dl,uint64_t mintxg,dmu_tx_t * tx)211*12470SMatthew.Ahrens@Sun.COM dsl_deadlist_remove_key(dsl_deadlist_t *dl, uint64_t mintxg, dmu_tx_t *tx)
212*12470SMatthew.Ahrens@Sun.COM {
213*12470SMatthew.Ahrens@Sun.COM 	dsl_deadlist_entry_t dle_tofind;
214*12470SMatthew.Ahrens@Sun.COM 	dsl_deadlist_entry_t *dle, *dle_prev;
215*12470SMatthew.Ahrens@Sun.COM 
216*12470SMatthew.Ahrens@Sun.COM 	if (dl->dl_oldfmt)
217*12470SMatthew.Ahrens@Sun.COM 		return;
218*12470SMatthew.Ahrens@Sun.COM 
219*12470SMatthew.Ahrens@Sun.COM 	dsl_deadlist_load_tree(dl);
220*12470SMatthew.Ahrens@Sun.COM 
221*12470SMatthew.Ahrens@Sun.COM 	dle_tofind.dle_mintxg = mintxg;
222*12470SMatthew.Ahrens@Sun.COM 	dle = avl_find(&dl->dl_tree, &dle_tofind, NULL);
223*12470SMatthew.Ahrens@Sun.COM 	dle_prev = AVL_PREV(&dl->dl_tree, dle);
224*12470SMatthew.Ahrens@Sun.COM 
225*12470SMatthew.Ahrens@Sun.COM 	bpobj_enqueue_subobj(&dle_prev->dle_bpobj,
226*12470SMatthew.Ahrens@Sun.COM 	    dle->dle_bpobj.bpo_object, tx);
227*12470SMatthew.Ahrens@Sun.COM 
228*12470SMatthew.Ahrens@Sun.COM 	avl_remove(&dl->dl_tree, dle);
229*12470SMatthew.Ahrens@Sun.COM 	bpobj_close(&dle->dle_bpobj);
230*12470SMatthew.Ahrens@Sun.COM 	kmem_free(dle, sizeof (*dle));
231*12470SMatthew.Ahrens@Sun.COM 
232*12470SMatthew.Ahrens@Sun.COM 	VERIFY3U(0, ==, zap_remove_int(dl->dl_os, dl->dl_object, mintxg, tx));
233*12470SMatthew.Ahrens@Sun.COM }
234*12470SMatthew.Ahrens@Sun.COM 
235*12470SMatthew.Ahrens@Sun.COM /*
236*12470SMatthew.Ahrens@Sun.COM  * Walk ds's snapshots to regenerate generate ZAP & AVL.
237*12470SMatthew.Ahrens@Sun.COM  */
238*12470SMatthew.Ahrens@Sun.COM static void
dsl_deadlist_regenerate(objset_t * os,uint64_t dlobj,uint64_t mrs_obj,dmu_tx_t * tx)239*12470SMatthew.Ahrens@Sun.COM dsl_deadlist_regenerate(objset_t *os, uint64_t dlobj,
240*12470SMatthew.Ahrens@Sun.COM     uint64_t mrs_obj, dmu_tx_t *tx)
241*12470SMatthew.Ahrens@Sun.COM {
242*12470SMatthew.Ahrens@Sun.COM 	dsl_deadlist_t dl;
243*12470SMatthew.Ahrens@Sun.COM 	dsl_pool_t *dp = dmu_objset_pool(os);
244*12470SMatthew.Ahrens@Sun.COM 
245*12470SMatthew.Ahrens@Sun.COM 	dsl_deadlist_open(&dl, os, dlobj);
246*12470SMatthew.Ahrens@Sun.COM 	if (dl.dl_oldfmt) {
247*12470SMatthew.Ahrens@Sun.COM 		dsl_deadlist_close(&dl);
248*12470SMatthew.Ahrens@Sun.COM 		return;
249*12470SMatthew.Ahrens@Sun.COM 	}
250*12470SMatthew.Ahrens@Sun.COM 
251*12470SMatthew.Ahrens@Sun.COM 	while (mrs_obj != 0) {
252*12470SMatthew.Ahrens@Sun.COM 		dsl_dataset_t *ds;
253*12470SMatthew.Ahrens@Sun.COM 		VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, mrs_obj, FTAG, &ds));
254*12470SMatthew.Ahrens@Sun.COM 		dsl_deadlist_add_key(&dl, ds->ds_phys->ds_prev_snap_txg, tx);
255*12470SMatthew.Ahrens@Sun.COM 		mrs_obj = ds->ds_phys->ds_prev_snap_obj;
256*12470SMatthew.Ahrens@Sun.COM 		dsl_dataset_rele(ds, FTAG);
257*12470SMatthew.Ahrens@Sun.COM 	}
258*12470SMatthew.Ahrens@Sun.COM 	dsl_deadlist_close(&dl);
259*12470SMatthew.Ahrens@Sun.COM }
260*12470SMatthew.Ahrens@Sun.COM 
261*12470SMatthew.Ahrens@Sun.COM uint64_t
dsl_deadlist_clone(dsl_deadlist_t * dl,uint64_t maxtxg,uint64_t mrs_obj,dmu_tx_t * tx)262*12470SMatthew.Ahrens@Sun.COM dsl_deadlist_clone(dsl_deadlist_t *dl, uint64_t maxtxg,
263*12470SMatthew.Ahrens@Sun.COM     uint64_t mrs_obj, dmu_tx_t *tx)
264*12470SMatthew.Ahrens@Sun.COM {
265*12470SMatthew.Ahrens@Sun.COM 	dsl_deadlist_entry_t *dle;
266*12470SMatthew.Ahrens@Sun.COM 	uint64_t newobj;
267*12470SMatthew.Ahrens@Sun.COM 
268*12470SMatthew.Ahrens@Sun.COM 	newobj = dsl_deadlist_alloc(dl->dl_os, tx);
269*12470SMatthew.Ahrens@Sun.COM 
270*12470SMatthew.Ahrens@Sun.COM 	if (dl->dl_oldfmt) {
271*12470SMatthew.Ahrens@Sun.COM 		dsl_deadlist_regenerate(dl->dl_os, newobj, mrs_obj, tx);
272*12470SMatthew.Ahrens@Sun.COM 		return (newobj);
273*12470SMatthew.Ahrens@Sun.COM 	}
274*12470SMatthew.Ahrens@Sun.COM 
275*12470SMatthew.Ahrens@Sun.COM 	dsl_deadlist_load_tree(dl);
276*12470SMatthew.Ahrens@Sun.COM 
277*12470SMatthew.Ahrens@Sun.COM 	for (dle = avl_first(&dl->dl_tree); dle;
278*12470SMatthew.Ahrens@Sun.COM 	    dle = AVL_NEXT(&dl->dl_tree, dle)) {
279*12470SMatthew.Ahrens@Sun.COM 		uint64_t obj;
280*12470SMatthew.Ahrens@Sun.COM 
281*12470SMatthew.Ahrens@Sun.COM 		if (dle->dle_mintxg >= maxtxg)
282*12470SMatthew.Ahrens@Sun.COM 			break;
283*12470SMatthew.Ahrens@Sun.COM 
284*12470SMatthew.Ahrens@Sun.COM 		obj = bpobj_alloc(dl->dl_os, SPA_MAXBLOCKSIZE, tx);
285*12470SMatthew.Ahrens@Sun.COM 		VERIFY3U(0, ==, zap_add_int_key(dl->dl_os, newobj,
286*12470SMatthew.Ahrens@Sun.COM 		    dle->dle_mintxg, obj, tx));
287*12470SMatthew.Ahrens@Sun.COM 	}
288*12470SMatthew.Ahrens@Sun.COM 	return (newobj);
289*12470SMatthew.Ahrens@Sun.COM }
290*12470SMatthew.Ahrens@Sun.COM 
291*12470SMatthew.Ahrens@Sun.COM void
dsl_deadlist_space(dsl_deadlist_t * dl,uint64_t * usedp,uint64_t * compp,uint64_t * uncompp)292*12470SMatthew.Ahrens@Sun.COM dsl_deadlist_space(dsl_deadlist_t *dl,
293*12470SMatthew.Ahrens@Sun.COM     uint64_t *usedp, uint64_t *compp, uint64_t *uncompp)
294*12470SMatthew.Ahrens@Sun.COM {
295*12470SMatthew.Ahrens@Sun.COM 	if (dl->dl_oldfmt) {
296*12470SMatthew.Ahrens@Sun.COM 		VERIFY3U(0, ==, bpobj_space(&dl->dl_bpobj,
297*12470SMatthew.Ahrens@Sun.COM 		    usedp, compp, uncompp));
298*12470SMatthew.Ahrens@Sun.COM 		return;
299*12470SMatthew.Ahrens@Sun.COM 	}
300*12470SMatthew.Ahrens@Sun.COM 
301*12470SMatthew.Ahrens@Sun.COM 	mutex_enter(&dl->dl_lock);
302*12470SMatthew.Ahrens@Sun.COM 	*usedp = dl->dl_phys->dl_used;
303*12470SMatthew.Ahrens@Sun.COM 	*compp = dl->dl_phys->dl_comp;
304*12470SMatthew.Ahrens@Sun.COM 	*uncompp = dl->dl_phys->dl_uncomp;
305*12470SMatthew.Ahrens@Sun.COM 	mutex_exit(&dl->dl_lock);
306*12470SMatthew.Ahrens@Sun.COM }
307*12470SMatthew.Ahrens@Sun.COM 
308*12470SMatthew.Ahrens@Sun.COM /*
309*12470SMatthew.Ahrens@Sun.COM  * return space used in the range (mintxg, maxtxg].
310*12470SMatthew.Ahrens@Sun.COM  * Includes maxtxg, does not include mintxg.
311*12470SMatthew.Ahrens@Sun.COM  * mintxg and maxtxg must both be keys in the deadlist (unless maxtxg is
312*12470SMatthew.Ahrens@Sun.COM  * UINT64_MAX).
313*12470SMatthew.Ahrens@Sun.COM  */
314*12470SMatthew.Ahrens@Sun.COM void
dsl_deadlist_space_range(dsl_deadlist_t * dl,uint64_t mintxg,uint64_t maxtxg,uint64_t * usedp,uint64_t * compp,uint64_t * uncompp)315*12470SMatthew.Ahrens@Sun.COM dsl_deadlist_space_range(dsl_deadlist_t *dl, uint64_t mintxg, uint64_t maxtxg,
316*12470SMatthew.Ahrens@Sun.COM     uint64_t *usedp, uint64_t *compp, uint64_t *uncompp)
317*12470SMatthew.Ahrens@Sun.COM {
318*12470SMatthew.Ahrens@Sun.COM 	dsl_deadlist_entry_t dle_tofind;
319*12470SMatthew.Ahrens@Sun.COM 	dsl_deadlist_entry_t *dle;
320*12470SMatthew.Ahrens@Sun.COM 	avl_index_t where;
321*12470SMatthew.Ahrens@Sun.COM 
322*12470SMatthew.Ahrens@Sun.COM 	if (dl->dl_oldfmt) {
323*12470SMatthew.Ahrens@Sun.COM 		VERIFY3U(0, ==, bpobj_space_range(&dl->dl_bpobj,
324*12470SMatthew.Ahrens@Sun.COM 		    mintxg, maxtxg, usedp, compp, uncompp));
325*12470SMatthew.Ahrens@Sun.COM 		return;
326*12470SMatthew.Ahrens@Sun.COM 	}
327*12470SMatthew.Ahrens@Sun.COM 
328*12470SMatthew.Ahrens@Sun.COM 	dsl_deadlist_load_tree(dl);
329*12470SMatthew.Ahrens@Sun.COM 	*usedp = *compp = *uncompp = 0;
330*12470SMatthew.Ahrens@Sun.COM 
331*12470SMatthew.Ahrens@Sun.COM 	dle_tofind.dle_mintxg = mintxg;
332*12470SMatthew.Ahrens@Sun.COM 	dle = avl_find(&dl->dl_tree, &dle_tofind, &where);
333*12470SMatthew.Ahrens@Sun.COM 	/*
334*12470SMatthew.Ahrens@Sun.COM 	 * If we don't find this mintxg, there shouldn't be anything
335*12470SMatthew.Ahrens@Sun.COM 	 * after it either.
336*12470SMatthew.Ahrens@Sun.COM 	 */
337*12470SMatthew.Ahrens@Sun.COM 	ASSERT(dle != NULL ||
338*12470SMatthew.Ahrens@Sun.COM 	    avl_nearest(&dl->dl_tree, where, AVL_AFTER) == NULL);
339*12470SMatthew.Ahrens@Sun.COM 	for (; dle && dle->dle_mintxg < maxtxg;
340*12470SMatthew.Ahrens@Sun.COM 	    dle = AVL_NEXT(&dl->dl_tree, dle)) {
341*12470SMatthew.Ahrens@Sun.COM 		uint64_t used, comp, uncomp;
342*12470SMatthew.Ahrens@Sun.COM 
343*12470SMatthew.Ahrens@Sun.COM 		VERIFY3U(0, ==, bpobj_space(&dle->dle_bpobj,
344*12470SMatthew.Ahrens@Sun.COM 		    &used, &comp, &uncomp));
345*12470SMatthew.Ahrens@Sun.COM 
346*12470SMatthew.Ahrens@Sun.COM 		*usedp += used;
347*12470SMatthew.Ahrens@Sun.COM 		*compp += comp;
348*12470SMatthew.Ahrens@Sun.COM 		*uncompp += uncomp;
349*12470SMatthew.Ahrens@Sun.COM 	}
350*12470SMatthew.Ahrens@Sun.COM }
351*12470SMatthew.Ahrens@Sun.COM 
352*12470SMatthew.Ahrens@Sun.COM static void
dsl_deadlist_insert_bpobj(dsl_deadlist_t * dl,uint64_t obj,uint64_t birth,dmu_tx_t * tx)353*12470SMatthew.Ahrens@Sun.COM dsl_deadlist_insert_bpobj(dsl_deadlist_t *dl, uint64_t obj, uint64_t birth,
354*12470SMatthew.Ahrens@Sun.COM     dmu_tx_t *tx)
355*12470SMatthew.Ahrens@Sun.COM {
356*12470SMatthew.Ahrens@Sun.COM 	dsl_deadlist_entry_t dle_tofind;
357*12470SMatthew.Ahrens@Sun.COM 	dsl_deadlist_entry_t *dle;
358*12470SMatthew.Ahrens@Sun.COM 	avl_index_t where;
359*12470SMatthew.Ahrens@Sun.COM 	uint64_t used, comp, uncomp;
360*12470SMatthew.Ahrens@Sun.COM 	bpobj_t bpo;
361*12470SMatthew.Ahrens@Sun.COM 
362*12470SMatthew.Ahrens@Sun.COM 	VERIFY3U(0, ==, bpobj_open(&bpo, dl->dl_os, obj));
363*12470SMatthew.Ahrens@Sun.COM 	VERIFY3U(0, ==, bpobj_space(&bpo, &used, &comp, &uncomp));
364*12470SMatthew.Ahrens@Sun.COM 	bpobj_close(&bpo);
365*12470SMatthew.Ahrens@Sun.COM 
366*12470SMatthew.Ahrens@Sun.COM 	dsl_deadlist_load_tree(dl);
367*12470SMatthew.Ahrens@Sun.COM 
368*12470SMatthew.Ahrens@Sun.COM 	dmu_buf_will_dirty(dl->dl_dbuf, tx);
369*12470SMatthew.Ahrens@Sun.COM 	mutex_enter(&dl->dl_lock);
370*12470SMatthew.Ahrens@Sun.COM 	dl->dl_phys->dl_used += used;
371*12470SMatthew.Ahrens@Sun.COM 	dl->dl_phys->dl_comp += comp;
372*12470SMatthew.Ahrens@Sun.COM 	dl->dl_phys->dl_uncomp += uncomp;
373*12470SMatthew.Ahrens@Sun.COM 	mutex_exit(&dl->dl_lock);
374*12470SMatthew.Ahrens@Sun.COM 
375*12470SMatthew.Ahrens@Sun.COM 	dle_tofind.dle_mintxg = birth;
376*12470SMatthew.Ahrens@Sun.COM 	dle = avl_find(&dl->dl_tree, &dle_tofind, &where);
377*12470SMatthew.Ahrens@Sun.COM 	if (dle == NULL)
378*12470SMatthew.Ahrens@Sun.COM 		dle = avl_nearest(&dl->dl_tree, where, AVL_BEFORE);
379*12470SMatthew.Ahrens@Sun.COM 	bpobj_enqueue_subobj(&dle->dle_bpobj, obj, tx);
380*12470SMatthew.Ahrens@Sun.COM }
381*12470SMatthew.Ahrens@Sun.COM 
382*12470SMatthew.Ahrens@Sun.COM static int
dsl_deadlist_insert_cb(void * arg,const blkptr_t * bp,dmu_tx_t * tx)383*12470SMatthew.Ahrens@Sun.COM dsl_deadlist_insert_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
384*12470SMatthew.Ahrens@Sun.COM {
385*12470SMatthew.Ahrens@Sun.COM 	dsl_deadlist_t *dl = arg;
386*12470SMatthew.Ahrens@Sun.COM 	dsl_deadlist_insert(dl, bp, tx);
387*12470SMatthew.Ahrens@Sun.COM 	return (0);
388*12470SMatthew.Ahrens@Sun.COM }
389*12470SMatthew.Ahrens@Sun.COM 
390*12470SMatthew.Ahrens@Sun.COM /*
391*12470SMatthew.Ahrens@Sun.COM  * Merge the deadlist pointed to by 'obj' into dl.  obj will be left as
392*12470SMatthew.Ahrens@Sun.COM  * an empty deadlist.
393*12470SMatthew.Ahrens@Sun.COM  */
394*12470SMatthew.Ahrens@Sun.COM void
dsl_deadlist_merge(dsl_deadlist_t * dl,uint64_t obj,dmu_tx_t * tx)395*12470SMatthew.Ahrens@Sun.COM dsl_deadlist_merge(dsl_deadlist_t *dl, uint64_t obj, dmu_tx_t *tx)
396*12470SMatthew.Ahrens@Sun.COM {
397*12470SMatthew.Ahrens@Sun.COM 	zap_cursor_t zc;
398*12470SMatthew.Ahrens@Sun.COM 	zap_attribute_t za;
399*12470SMatthew.Ahrens@Sun.COM 	dmu_buf_t *bonus;
400*12470SMatthew.Ahrens@Sun.COM 	dsl_deadlist_phys_t *dlp;
401*12470SMatthew.Ahrens@Sun.COM 	dmu_object_info_t doi;
402*12470SMatthew.Ahrens@Sun.COM 
403*12470SMatthew.Ahrens@Sun.COM 	VERIFY3U(0, ==, dmu_object_info(dl->dl_os, obj, &doi));
404*12470SMatthew.Ahrens@Sun.COM 	if (doi.doi_type == DMU_OT_BPOBJ) {
405*12470SMatthew.Ahrens@Sun.COM 		bpobj_t bpo;
406*12470SMatthew.Ahrens@Sun.COM 		VERIFY3U(0, ==, bpobj_open(&bpo, dl->dl_os, obj));
407*12470SMatthew.Ahrens@Sun.COM 		VERIFY3U(0, ==, bpobj_iterate(&bpo,
408*12470SMatthew.Ahrens@Sun.COM 		    dsl_deadlist_insert_cb, dl, tx));
409*12470SMatthew.Ahrens@Sun.COM 		bpobj_close(&bpo);
410*12470SMatthew.Ahrens@Sun.COM 		return;
411*12470SMatthew.Ahrens@Sun.COM 	}
412*12470SMatthew.Ahrens@Sun.COM 
413*12470SMatthew.Ahrens@Sun.COM 	for (zap_cursor_init(&zc, dl->dl_os, obj);
414*12470SMatthew.Ahrens@Sun.COM 	    zap_cursor_retrieve(&zc, &za) == 0;
415*12470SMatthew.Ahrens@Sun.COM 	    zap_cursor_advance(&zc)) {
416*12470SMatthew.Ahrens@Sun.COM 		uint64_t mintxg = strtonum(za.za_name, NULL);
417*12470SMatthew.Ahrens@Sun.COM 		dsl_deadlist_insert_bpobj(dl, za.za_first_integer, mintxg, tx);
418*12470SMatthew.Ahrens@Sun.COM 		VERIFY3U(0, ==, zap_remove_int(dl->dl_os, obj, mintxg, tx));
419*12470SMatthew.Ahrens@Sun.COM 	}
420*12470SMatthew.Ahrens@Sun.COM 	zap_cursor_fini(&zc);
421*12470SMatthew.Ahrens@Sun.COM 
422*12470SMatthew.Ahrens@Sun.COM 	VERIFY3U(0, ==, dmu_bonus_hold(dl->dl_os, obj, FTAG, &bonus));
423*12470SMatthew.Ahrens@Sun.COM 	dlp = bonus->db_data;
424*12470SMatthew.Ahrens@Sun.COM 	dmu_buf_will_dirty(bonus, tx);
425*12470SMatthew.Ahrens@Sun.COM 	bzero(dlp, sizeof (*dlp));
426*12470SMatthew.Ahrens@Sun.COM 	dmu_buf_rele(bonus, FTAG);
427*12470SMatthew.Ahrens@Sun.COM }
428*12470SMatthew.Ahrens@Sun.COM 
429*12470SMatthew.Ahrens@Sun.COM /*
430*12470SMatthew.Ahrens@Sun.COM  * Remove entries on dl that are >= mintxg, and put them on the bpobj.
431*12470SMatthew.Ahrens@Sun.COM  */
432*12470SMatthew.Ahrens@Sun.COM void
dsl_deadlist_move_bpobj(dsl_deadlist_t * dl,bpobj_t * bpo,uint64_t mintxg,dmu_tx_t * tx)433*12470SMatthew.Ahrens@Sun.COM dsl_deadlist_move_bpobj(dsl_deadlist_t *dl, bpobj_t *bpo, uint64_t mintxg,
434*12470SMatthew.Ahrens@Sun.COM     dmu_tx_t *tx)
435*12470SMatthew.Ahrens@Sun.COM {
436*12470SMatthew.Ahrens@Sun.COM 	dsl_deadlist_entry_t dle_tofind;
437*12470SMatthew.Ahrens@Sun.COM 	dsl_deadlist_entry_t *dle;
438*12470SMatthew.Ahrens@Sun.COM 	avl_index_t where;
439*12470SMatthew.Ahrens@Sun.COM 
440*12470SMatthew.Ahrens@Sun.COM 	ASSERT(!dl->dl_oldfmt);
441*12470SMatthew.Ahrens@Sun.COM 	dmu_buf_will_dirty(dl->dl_dbuf, tx);
442*12470SMatthew.Ahrens@Sun.COM 	dsl_deadlist_load_tree(dl);
443*12470SMatthew.Ahrens@Sun.COM 
444*12470SMatthew.Ahrens@Sun.COM 	dle_tofind.dle_mintxg = mintxg;
445*12470SMatthew.Ahrens@Sun.COM 	dle = avl_find(&dl->dl_tree, &dle_tofind, &where);
446*12470SMatthew.Ahrens@Sun.COM 	if (dle == NULL)
447*12470SMatthew.Ahrens@Sun.COM 		dle = avl_nearest(&dl->dl_tree, where, AVL_AFTER);
448*12470SMatthew.Ahrens@Sun.COM 	while (dle) {
449*12470SMatthew.Ahrens@Sun.COM 		uint64_t used, comp, uncomp;
450*12470SMatthew.Ahrens@Sun.COM 		dsl_deadlist_entry_t *dle_next;
451*12470SMatthew.Ahrens@Sun.COM 
452*12470SMatthew.Ahrens@Sun.COM 		bpobj_enqueue_subobj(bpo, dle->dle_bpobj.bpo_object, tx);
453*12470SMatthew.Ahrens@Sun.COM 
454*12470SMatthew.Ahrens@Sun.COM 		VERIFY3U(0, ==, bpobj_space(&dle->dle_bpobj,
455*12470SMatthew.Ahrens@Sun.COM 		    &used, &comp, &uncomp));
456*12470SMatthew.Ahrens@Sun.COM 		mutex_enter(&dl->dl_lock);
457*12470SMatthew.Ahrens@Sun.COM 		ASSERT3U(dl->dl_phys->dl_used, >=, used);
458*12470SMatthew.Ahrens@Sun.COM 		ASSERT3U(dl->dl_phys->dl_comp, >=, comp);
459*12470SMatthew.Ahrens@Sun.COM 		ASSERT3U(dl->dl_phys->dl_uncomp, >=, uncomp);
460*12470SMatthew.Ahrens@Sun.COM 		dl->dl_phys->dl_used -= used;
461*12470SMatthew.Ahrens@Sun.COM 		dl->dl_phys->dl_comp -= comp;
462*12470SMatthew.Ahrens@Sun.COM 		dl->dl_phys->dl_uncomp -= uncomp;
463*12470SMatthew.Ahrens@Sun.COM 		mutex_exit(&dl->dl_lock);
464*12470SMatthew.Ahrens@Sun.COM 
465*12470SMatthew.Ahrens@Sun.COM 		VERIFY3U(0, ==, zap_remove_int(dl->dl_os, dl->dl_object,
466*12470SMatthew.Ahrens@Sun.COM 		    dle->dle_mintxg, tx));
467*12470SMatthew.Ahrens@Sun.COM 
468*12470SMatthew.Ahrens@Sun.COM 		dle_next = AVL_NEXT(&dl->dl_tree, dle);
469*12470SMatthew.Ahrens@Sun.COM 		avl_remove(&dl->dl_tree, dle);
470*12470SMatthew.Ahrens@Sun.COM 		bpobj_close(&dle->dle_bpobj);
471*12470SMatthew.Ahrens@Sun.COM 		kmem_free(dle, sizeof (*dle));
472*12470SMatthew.Ahrens@Sun.COM 		dle = dle_next;
473*12470SMatthew.Ahrens@Sun.COM 	}
474*12470SMatthew.Ahrens@Sun.COM }
475