[mirror_zfs-debian.git] / module / zfs / bptree.c

/*
 * CDDL HEADER START
 *
 * The contents of this file are subject to the terms of the
 * Common Development and Distribution License (the "License").
 * You may not use this file except in compliance with the License.
 *
 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
 * or http://www.opensolaris.org/os/licensing.
 * See the License for the specific language governing permissions
 * and limitations under the License.
 *
 * When distributing Covered Code, include this CDDL HEADER in each
 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
 * If applicable, add the following below this CDDL HEADER, with the
 * fields enclosed by brackets "[]" replaced with your own identifying
 * information: Portions Copyright [yyyy] [name of copyright owner]
 *
 * CDDL HEADER END
 */

/*
 * Copyright (c) 2011, 2015 by Delphix. All rights reserved.
 */

#include <sys/arc.h>
#include <sys/bptree.h>
#include <sys/dmu.h>
#include <sys/dmu_objset.h>
#include <sys/dmu_tx.h>
#include <sys/dmu_traverse.h>
#include <sys/dsl_dataset.h>
#include <sys/dsl_dir.h>
#include <sys/dsl_pool.h>
#include <sys/dnode.h>
#include <sys/refcount.h>
#include <sys/spa.h>

/*
 * A bptree is a queue of root block pointers from destroyed datasets. When a
 * dataset is destroyed its root block pointer is put on the end of the pool's
 * bptree queue so the dataset's blocks can be freed asynchronously by
 * dsl_scan_sync. This allows the delete operation to finish without traversing
 * all the dataset's blocks.
 *
 * Note that while bt_begin and bt_end are only ever incremented in this code,
 * they are effectively reset to 0 every time the entire bptree is freed because
 * the bptree's object is destroyed and re-created.
 */

struct bptree_args {
	bptree_phys_t *ba_phys;	/* data in bonus buffer, dirtied if freeing */
	boolean_t ba_free;	/* true if freeing during traversal */

	bptree_itor_t *ba_func;	/* function to call for each blockpointer */
	void *ba_arg;		/* caller supplied argument to ba_func */
	dmu_tx_t *ba_tx;	/* caller supplied tx, NULL if not freeing */
} bptree_args_t;

uint64_t
bptree_alloc(objset_t *os, dmu_tx_t *tx)
{
	uint64_t obj;
	dmu_buf_t *db;
	bptree_phys_t *bt;

	obj = dmu_object_alloc(os, DMU_OTN_UINT64_METADATA,
	    SPA_OLD_MAXBLOCKSIZE, DMU_OTN_UINT64_METADATA,
	    sizeof (bptree_phys_t), tx);

	/*
	 * Bonus buffer contents are already initialized to 0, but for
	 * readability we make it explicit.
	 */
	VERIFY3U(0, ==, dmu_bonus_hold(os, obj, FTAG, &db));
	dmu_buf_will_dirty(db, tx);
	bt = db->db_data;
	bt->bt_begin = 0;
	bt->bt_end = 0;
	bt->bt_bytes = 0;
	bt->bt_comp = 0;
	bt->bt_uncomp = 0;
	dmu_buf_rele(db, FTAG);

	return (obj);
}

int
bptree_free(objset_t *os, uint64_t obj, dmu_tx_t *tx)
{
	dmu_buf_t *db;
	bptree_phys_t *bt;

	VERIFY3U(0, ==, dmu_bonus_hold(os, obj, FTAG, &db));
	bt = db->db_data;
	ASSERT3U(bt->bt_begin, ==, bt->bt_end);
	ASSERT0(bt->bt_bytes);
	ASSERT0(bt->bt_comp);
	ASSERT0(bt->bt_uncomp);
	dmu_buf_rele(db, FTAG);

	return (dmu_object_free(os, obj, tx));
}

boolean_t
bptree_is_empty(objset_t *os, uint64_t obj)
{
	dmu_buf_t *db;
	bptree_phys_t *bt;
	boolean_t rv;

	VERIFY0(dmu_bonus_hold(os, obj, FTAG, &db));
	bt = db->db_data;
	rv = (bt->bt_begin == bt->bt_end);
	dmu_buf_rele(db, FTAG);
	return (rv);
}

void
bptree_add(objset_t *os, uint64_t obj, blkptr_t *bp, uint64_t birth_txg,
    uint64_t bytes, uint64_t comp, uint64_t uncomp, dmu_tx_t *tx)
{
	dmu_buf_t *db;
	bptree_phys_t *bt;
	bptree_entry_phys_t *bte;

	/*
	 * bptree objects are in the pool mos, therefore they can only be
	 * modified in syncing context. Furthermore, this is only modified
	 * by the sync thread, so no locking is necessary.
	 */
	ASSERT(dmu_tx_is_syncing(tx));

	VERIFY3U(0, ==, dmu_bonus_hold(os, obj, FTAG, &db));
	bt = db->db_data;

	bte = kmem_zalloc(sizeof (*bte), KM_SLEEP);
	bte->be_birth_txg = birth_txg;
	bte->be_bp = *bp;
	dmu_write(os, obj, bt->bt_end * sizeof (*bte), sizeof (*bte), bte, tx);
	kmem_free(bte, sizeof (*bte));

	dmu_buf_will_dirty(db, tx);
	bt->bt_end++;
	bt->bt_bytes += bytes;
	bt->bt_comp += comp;
	bt->bt_uncomp += uncomp;
	dmu_buf_rele(db, FTAG);
}

/* ARGSUSED */
static int
bptree_visit_cb(spa_t *spa, zilog_t *zilog, const blkptr_t *bp,
    const zbookmark_phys_t *zb, const dnode_phys_t *dnp, void *arg)
{
	int err;
	struct bptree_args *ba = arg;

	if (bp == NULL || BP_IS_HOLE(bp))
		return (0);

	err = ba->ba_func(ba->ba_arg, bp, ba->ba_tx);
	if (err == 0 && ba->ba_free) {
		ba->ba_phys->bt_bytes -= bp_get_dsize_sync(spa, bp);
		ba->ba_phys->bt_comp -= BP_GET_PSIZE(bp);
		ba->ba_phys->bt_uncomp -= BP_GET_UCSIZE(bp);
	}
	return (err);
}

/*
 * If "free" is set:
 *  - It is assumed that "func" will be freeing the block pointers.
 *  - If "func" returns nonzero, the bookmark will be remembered and
 *    iteration will be restarted from this point on next invocation.
 *  - If an i/o error is encountered (e.g. "func" returns EIO or ECKSUM),
 *    bptree_iterate will remember the bookmark, continue traversing
 *    any additional entries, and return 0.
 *
 * If "free" is not set, traversal will stop and return an error if
 * an i/o error is encountered.
 *
 * In either case, if zfs_free_leak_on_eio is set, i/o errors will be
 * ignored and traversal will continue (i.e. TRAVERSE_HARD will be passed to
 * traverse_dataset_destroyed()).
 */
int
bptree_iterate(objset_t *os, uint64_t obj, boolean_t free, bptree_itor_t func,
    void *arg, dmu_tx_t *tx)
{
	boolean_t ioerr = B_FALSE;
	int err;
	uint64_t i;
	dmu_buf_t *db;
	struct bptree_args ba;

	ASSERT(!free || dmu_tx_is_syncing(tx));

	err = dmu_bonus_hold(os, obj, FTAG, &db);
	if (err != 0)
		return (err);

	if (free)
		dmu_buf_will_dirty(db, tx);

	ba.ba_phys = db->db_data;
	ba.ba_free = free;
	ba.ba_func = func;
	ba.ba_arg = arg;
	ba.ba_tx = tx;

	err = 0;
	for (i = ba.ba_phys->bt_begin; i < ba.ba_phys->bt_end; i++) {
		bptree_entry_phys_t bte;
		int flags = TRAVERSE_PREFETCH_METADATA | TRAVERSE_POST;

		err = dmu_read(os, obj, i * sizeof (bte), sizeof (bte),
		    &bte, DMU_READ_NO_PREFETCH);
		if (err != 0)
			break;

		if (zfs_free_leak_on_eio)
			flags |= TRAVERSE_HARD;
		zfs_dbgmsg("bptree index %lld: traversing from min_txg=%lld "
		    "bookmark %lld/%lld/%lld/%lld",
		    (longlong_t)i,
		    (longlong_t)bte.be_birth_txg,
		    (longlong_t)bte.be_zb.zb_objset,
		    (longlong_t)bte.be_zb.zb_object,
		    (longlong_t)bte.be_zb.zb_level,
		    (longlong_t)bte.be_zb.zb_blkid);
		err = traverse_dataset_destroyed(os->os_spa, &bte.be_bp,
		    bte.be_birth_txg, &bte.be_zb, flags,
		    bptree_visit_cb, &ba);
		if (free) {
			/*
			 * The callback has freed the visited block pointers.
			 * Record our traversal progress on disk, either by
			 * updating this record's bookmark, or by logically
			 * removing this record by advancing bt_begin.
			 */
			if (err != 0) {
				/* save bookmark for future resume */
				ASSERT3U(bte.be_zb.zb_objset, ==,
				    ZB_DESTROYED_OBJSET);
				ASSERT0(bte.be_zb.zb_level);
				dmu_write(os, obj, i * sizeof (bte),
				    sizeof (bte), &bte, tx);
				if (err == EIO || err == ECKSUM ||
				    err == ENXIO) {
					/*
					 * Skip the rest of this tree and
					 * continue on to the next entry.
					 */
					err = 0;
					ioerr = B_TRUE;
				} else {
					break;
				}
			} else if (ioerr) {
				/*
				 * This entry is finished, but there were
				 * i/o errors on previous entries, so we
				 * can't adjust bt_begin.  Set this entry's
				 * be_birth_txg such that it will be
				 * treated as a no-op in future traversals.
				 */
				bte.be_birth_txg = UINT64_MAX;
				dmu_write(os, obj, i * sizeof (bte),
				    sizeof (bte), &bte, tx);
			}

			if (!ioerr) {
				ba.ba_phys->bt_begin++;
				(void) dmu_free_range(os, obj,
				    i * sizeof (bte), sizeof (bte), tx);
			}
		} else if (err != 0) {
			break;
		}
	}

	ASSERT(!free || err != 0 || ioerr ||
	    ba.ba_phys->bt_begin == ba.ba_phys->bt_end);

	/* if all blocks are free there should be no used space */
	if (ba.ba_phys->bt_begin == ba.ba_phys->bt_end) {
		if (zfs_free_leak_on_eio) {
			ba.ba_phys->bt_bytes = 0;
			ba.ba_phys->bt_comp = 0;
			ba.ba_phys->bt_uncomp = 0;
		}

		ASSERT0(ba.ba_phys->bt_bytes);
		ASSERT0(ba.ba_phys->bt_comp);
		ASSERT0(ba.ba_phys->bt_uncomp);
	}

	dmu_buf_rele(db, FTAG);

	return (err);
}
Commit	Line	Data
9ae529ec CS	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	/*
cae5b340	23	* Copyright (c) 2011, 2015 by Delphix. All rights reserved.
9ae529ec CS	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	*
a08ee875	46	* Note that while bt_begin and bt_end are only ever incremented in this code,
9ae529ec CS	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,
e10b0808	68	SPA_OLD_MAXBLOCKSIZE, DMU_OTN_UINT64_METADATA,
9ae529ec CS	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);
c06d4368 AX	97	ASSERT0(bt->bt_bytes);
	98	ASSERT0(bt->bt_comp);
	99	ASSERT0(bt->bt_uncomp);
9ae529ec CS	100	dmu_buf_rele(db, FTAG);
	101
	102	return (dmu_object_free(os, obj, tx));
	103	}
	104
ea04106b AX	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
9ae529ec CS	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;
ea04106b	125	bptree_entry_phys_t *bte;
9ae529ec CS	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
ea04106b AX	137	bte = kmem_zalloc(sizeof (*bte), KM_SLEEP);
	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	kmem_free(bte, sizeof (*bte));
9ae529ec CS	142
	143	dmu_buf_will_dirty(db, tx);
	144	bt->bt_end++;
	145	bt->bt_bytes += bytes;
	146	bt->bt_comp += comp;
	147	bt->bt_uncomp += uncomp;
	148	dmu_buf_rele(db, FTAG);
	149	}
	150
	151	/* ARGSUSED */
	152	static int
c06d4368	153	bptree_visit_cb(spa_t spa, zilog_t zilog, const blkptr_t *bp,
ea04106b	154	const zbookmark_phys_t zb, const dnode_phys_t dnp, void *arg)
9ae529ec CS	155	{
	156	int err;
	157	struct bptree_args *ba = arg;
	158
cae5b340	159	if (bp == NULL \|\| BP_IS_HOLE(bp))
9ae529ec CS	160	return (0);
	161
	162	err = ba->ba_func(ba->ba_arg, bp, ba->ba_tx);
	163	if (err == 0 && ba->ba_free) {
	164	ba->ba_phys->bt_bytes -= bp_get_dsize_sync(spa, bp);
	165	ba->ba_phys->bt_comp -= BP_GET_PSIZE(bp);
	166	ba->ba_phys->bt_uncomp -= BP_GET_UCSIZE(bp);
	167	}
	168	return (err);
	169	}
	170
ea04106b AX	171	/*
	172	* If "free" is set:
	173	* - It is assumed that "func" will be freeing the block pointers.
	174	* - If "func" returns nonzero, the bookmark will be remembered and
	175	* iteration will be restarted from this point on next invocation.
	176	* - If an i/o error is encountered (e.g. "func" returns EIO or ECKSUM),
	177	* bptree_iterate will remember the bookmark, continue traversing
	178	* any additional entries, and return 0.
	179	*
	180	* If "free" is not set, traversal will stop and return an error if
	181	* an i/o error is encountered.
	182	*
	183	* In either case, if zfs_free_leak_on_eio is set, i/o errors will be
	184	* ignored and traversal will continue (i.e. TRAVERSE_HARD will be passed to
	185	* traverse_dataset_destroyed()).
	186	*/
9ae529ec CS	187	int
	188	bptree_iterate(objset_t *os, uint64_t obj, boolean_t free, bptree_itor_t func,
	189	void arg, dmu_tx_t tx)
	190	{
ea04106b	191	boolean_t ioerr = B_FALSE;
9ae529ec CS	192	int err;
	193	uint64_t i;
	194	dmu_buf_t *db;
	195	struct bptree_args ba;
	196
	197	ASSERT(!free \|\| dmu_tx_is_syncing(tx));
	198
	199	err = dmu_bonus_hold(os, obj, FTAG, &db);
	200	if (err != 0)
	201	return (err);
	202
	203	if (free)
	204	dmu_buf_will_dirty(db, tx);
	205
	206	ba.ba_phys = db->db_data;
	207	ba.ba_free = free;
	208	ba.ba_func = func;
	209	ba.ba_arg = arg;
	210	ba.ba_tx = tx;
	211
	212	err = 0;
	213	for (i = ba.ba_phys->bt_begin; i < ba.ba_phys->bt_end; i++) {
	214	bptree_entry_phys_t bte;
a08ee875	215	int flags = TRAVERSE_PREFETCH_METADATA \| TRAVERSE_POST;
9ae529ec	216
9ae529ec CS	217	err = dmu_read(os, obj, i * sizeof (bte), sizeof (bte),
	218	&bte, DMU_READ_NO_PREFETCH);
	219	if (err != 0)
	220	break;
	221
ea04106b	222	if (zfs_free_leak_on_eio)
a08ee875	223	flags \|= TRAVERSE_HARD;
ea04106b AX	224	zfs_dbgmsg("bptree index %lld: traversing from min_txg=%lld "
ea04106b AX	225	"bookmark %lld/%lld/%lld/%lld",
cae5b340 AX	226	(longlong_t)i,
cae5b340 AX	227	(longlong_t)bte.be_birth_txg,
ea04106b AX	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);
9ae529ec	232	err = traverse_dataset_destroyed(os->os_spa, &bte.be_bp,
a08ee875	233	bte.be_birth_txg, &bte.be_zb, flags,
9ae529ec CS	234	bptree_visit_cb, &ba);
9ae529ec CS	235	if (free) {
ea04106b AX	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) {
9ae529ec CS	243	/* save bookmark for future resume */
	244	ASSERT3U(bte.be_zb.zb_objset, ==,
	245	ZB_DESTROYED_OBJSET);
c06d4368	246	ASSERT0(bte.be_zb.zb_level);
9ae529ec CS	247	dmu_write(os, obj, i * sizeof (bte),
9ae529ec CS	248	sizeof (bte), &bte, tx);
ea04106b AX	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) {
a08ee875	261	/*
ea04106b AX	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.
a08ee875	267	*/
ea04106b AX	268	bte.be_birth_txg = UINT64_MAX;
	269	dmu_write(os, obj, i * sizeof (bte),
	270	sizeof (bte), &bte, tx);
a08ee875 LG	271	}
a08ee875 LG	272
ea04106b AX	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;
9ae529ec CS	280	}
	281	}
	282
ea04106b AX	283	ASSERT(!free \|\| err != 0 \|\| ioerr \|\|
ea04106b AX	284	ba.ba_phys->bt_begin == ba.ba_phys->bt_end);
9ae529ec CS	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) {
ea04106b AX	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
c06d4368 AX	294	ASSERT0(ba.ba_phys->bt_bytes);
	295	ASSERT0(ba.ba_phys->bt_comp);
	296	ASSERT0(ba.ba_phys->bt_uncomp);
9ae529ec CS	297	}
	298
	299	dmu_buf_rele(db, FTAG);
	300
	301	return (err);
	302	}