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 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2012, 2017 by Delphix. All rights reserved.
24 * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved.
25 * Copyright (c) 2014 Integros [integros.com]
26 * Copyright 2017 RackTop Systems.
27 */
28
29#include <sys/zfs_context.h>
30#include <sys/dbuf.h>
31#include <sys/dnode.h>
32#include <sys/dmu.h>
33#include <sys/dmu_impl.h>
34#include <sys/dmu_tx.h>
35#include <sys/dmu_objset.h>
36#include <sys/dsl_dir.h>
37#include <sys/dsl_dataset.h>
38#include <sys/spa.h>
39#include <sys/zio.h>
40#include <sys/dmu_zfetch.h>
41#include <sys/range_tree.h>
42#include <sys/zfs_project.h>
43
44dnode_stats_t dnode_stats = {
45	{ "dnode_hold_dbuf_hold",		KSTAT_DATA_UINT64 },
46	{ "dnode_hold_dbuf_read",		KSTAT_DATA_UINT64 },
47	{ "dnode_hold_alloc_hits",		KSTAT_DATA_UINT64 },
48	{ "dnode_hold_alloc_misses",		KSTAT_DATA_UINT64 },
49	{ "dnode_hold_alloc_interior",		KSTAT_DATA_UINT64 },
50	{ "dnode_hold_alloc_lock_retry",	KSTAT_DATA_UINT64 },
51	{ "dnode_hold_alloc_lock_misses",	KSTAT_DATA_UINT64 },
52	{ "dnode_hold_alloc_type_none",		KSTAT_DATA_UINT64 },
53	{ "dnode_hold_free_hits",		KSTAT_DATA_UINT64 },
54	{ "dnode_hold_free_misses",		KSTAT_DATA_UINT64 },
55	{ "dnode_hold_free_lock_misses",	KSTAT_DATA_UINT64 },
56	{ "dnode_hold_free_lock_retry",		KSTAT_DATA_UINT64 },
57	{ "dnode_hold_free_overflow",		KSTAT_DATA_UINT64 },
58	{ "dnode_hold_free_refcount",		KSTAT_DATA_UINT64 },
59	{ "dnode_free_interior_lock_retry",	KSTAT_DATA_UINT64 },
60	{ "dnode_allocate",			KSTAT_DATA_UINT64 },
61	{ "dnode_reallocate",			KSTAT_DATA_UINT64 },
62	{ "dnode_buf_evict",			KSTAT_DATA_UINT64 },
63	{ "dnode_alloc_next_chunk",		KSTAT_DATA_UINT64 },
64	{ "dnode_alloc_race",			KSTAT_DATA_UINT64 },
65	{ "dnode_alloc_next_block",		KSTAT_DATA_UINT64 },
66	{ "dnode_move_invalid",			KSTAT_DATA_UINT64 },
67	{ "dnode_move_recheck1",		KSTAT_DATA_UINT64 },
68	{ "dnode_move_recheck2",		KSTAT_DATA_UINT64 },
69	{ "dnode_move_special",			KSTAT_DATA_UINT64 },
70	{ "dnode_move_handle",			KSTAT_DATA_UINT64 },
71	{ "dnode_move_rwlock",			KSTAT_DATA_UINT64 },
72	{ "dnode_move_active",			KSTAT_DATA_UINT64 },
73};
74
75static kstat_t *dnode_ksp;
76static kmem_cache_t *dnode_cache;
77
78static dnode_phys_t dnode_phys_zero;
79
80int zfs_default_bs = SPA_MINBLOCKSHIFT;
81int zfs_default_ibs = DN_MAX_INDBLKSHIFT;
82
83#ifdef	_KERNEL
84static kmem_cbrc_t dnode_move(void *, void *, size_t, void *);
85#endif	/* _KERNEL */
86
87static int
88dbuf_compare(const void *x1, const void *x2)
89{
90	const dmu_buf_impl_t *d1 = x1;
91	const dmu_buf_impl_t *d2 = x2;
92
93	int cmp = AVL_CMP(d1->db_level, d2->db_level);
94	if (likely(cmp))
95		return (cmp);
96
97	cmp = AVL_CMP(d1->db_blkid, d2->db_blkid);
98	if (likely(cmp))
99		return (cmp);
100
101	if (d1->db_state == DB_SEARCH) {
102		ASSERT3S(d2->db_state, !=, DB_SEARCH);
103		return (-1);
104	} else if (d2->db_state == DB_SEARCH) {
105		ASSERT3S(d1->db_state, !=, DB_SEARCH);
106		return (1);
107	}
108
109	return (AVL_PCMP(d1, d2));
110}
111
112/* ARGSUSED */
113static int
114dnode_cons(void *arg, void *unused, int kmflag)
115{
116	dnode_t *dn = arg;
117	int i;
118
119	rw_init(&dn->dn_struct_rwlock, NULL, RW_DEFAULT, NULL);
120	mutex_init(&dn->dn_mtx, NULL, MUTEX_DEFAULT, NULL);
121	mutex_init(&dn->dn_dbufs_mtx, NULL, MUTEX_DEFAULT, NULL);
122	cv_init(&dn->dn_notxholds, NULL, CV_DEFAULT, NULL);
123
124	/*
125	 * Every dbuf has a reference, and dropping a tracked reference is
126	 * O(number of references), so don't track dn_holds.
127	 */
128	zfs_refcount_create_untracked(&dn->dn_holds);
129	zfs_refcount_create(&dn->dn_tx_holds);
130	list_link_init(&dn->dn_link);
131
132	bzero(&dn->dn_next_nblkptr[0], sizeof (dn->dn_next_nblkptr));
133	bzero(&dn->dn_next_nlevels[0], sizeof (dn->dn_next_nlevels));
134	bzero(&dn->dn_next_indblkshift[0], sizeof (dn->dn_next_indblkshift));
135	bzero(&dn->dn_next_bonustype[0], sizeof (dn->dn_next_bonustype));
136	bzero(&dn->dn_rm_spillblk[0], sizeof (dn->dn_rm_spillblk));
137	bzero(&dn->dn_next_bonuslen[0], sizeof (dn->dn_next_bonuslen));
138	bzero(&dn->dn_next_blksz[0], sizeof (dn->dn_next_blksz));
139	bzero(&dn->dn_next_maxblkid[0], sizeof (dn->dn_next_maxblkid));
140
141	for (i = 0; i < TXG_SIZE; i++) {
142		multilist_link_init(&dn->dn_dirty_link[i]);
143		dn->dn_free_ranges[i] = NULL;
144		list_create(&dn->dn_dirty_records[i],
145		    sizeof (dbuf_dirty_record_t),
146		    offsetof(dbuf_dirty_record_t, dr_dirty_node));
147	}
148
149	dn->dn_allocated_txg = 0;
150	dn->dn_free_txg = 0;
151	dn->dn_assigned_txg = 0;
152	dn->dn_dirty_txg = 0;
153	dn->dn_dirtyctx = 0;
154	dn->dn_dirtyctx_firstset = NULL;
155	dn->dn_bonus = NULL;
156	dn->dn_have_spill = B_FALSE;
157	dn->dn_zio = NULL;
158	dn->dn_oldused = 0;
159	dn->dn_oldflags = 0;
160	dn->dn_olduid = 0;
161	dn->dn_oldgid = 0;
162	dn->dn_oldprojid = ZFS_DEFAULT_PROJID;
163	dn->dn_newuid = 0;
164	dn->dn_newgid = 0;
165	dn->dn_newprojid = ZFS_DEFAULT_PROJID;
166	dn->dn_id_flags = 0;
167
168	dn->dn_dbufs_count = 0;
169	avl_create(&dn->dn_dbufs, dbuf_compare, sizeof (dmu_buf_impl_t),
170	    offsetof(dmu_buf_impl_t, db_link));
171
172	dn->dn_moved = 0;
173	return (0);
174}
175
176/* ARGSUSED */
177static void
178dnode_dest(void *arg, void *unused)
179{
180	int i;
181	dnode_t *dn = arg;
182
183	rw_destroy(&dn->dn_struct_rwlock);
184	mutex_destroy(&dn->dn_mtx);
185	mutex_destroy(&dn->dn_dbufs_mtx);
186	cv_destroy(&dn->dn_notxholds);
187	zfs_refcount_destroy(&dn->dn_holds);
188	zfs_refcount_destroy(&dn->dn_tx_holds);
189	ASSERT(!list_link_active(&dn->dn_link));
190
191	for (i = 0; i < TXG_SIZE; i++) {
192		ASSERT(!multilist_link_active(&dn->dn_dirty_link[i]));
193		ASSERT3P(dn->dn_free_ranges[i], ==, NULL);
194		list_destroy(&dn->dn_dirty_records[i]);
195		ASSERT0(dn->dn_next_nblkptr[i]);
196		ASSERT0(dn->dn_next_nlevels[i]);
197		ASSERT0(dn->dn_next_indblkshift[i]);
198		ASSERT0(dn->dn_next_bonustype[i]);
199		ASSERT0(dn->dn_rm_spillblk[i]);
200		ASSERT0(dn->dn_next_bonuslen[i]);
201		ASSERT0(dn->dn_next_blksz[i]);
202		ASSERT0(dn->dn_next_maxblkid[i]);
203	}
204
205	ASSERT0(dn->dn_allocated_txg);
206	ASSERT0(dn->dn_free_txg);
207	ASSERT0(dn->dn_assigned_txg);
208	ASSERT0(dn->dn_dirty_txg);
209	ASSERT0(dn->dn_dirtyctx);
210	ASSERT3P(dn->dn_dirtyctx_firstset, ==, NULL);
211	ASSERT3P(dn->dn_bonus, ==, NULL);
212	ASSERT(!dn->dn_have_spill);
213	ASSERT3P(dn->dn_zio, ==, NULL);
214	ASSERT0(dn->dn_oldused);
215	ASSERT0(dn->dn_oldflags);
216	ASSERT0(dn->dn_olduid);
217	ASSERT0(dn->dn_oldgid);
218	ASSERT0(dn->dn_oldprojid);
219	ASSERT0(dn->dn_newuid);
220	ASSERT0(dn->dn_newgid);
221	ASSERT0(dn->dn_newprojid);
222	ASSERT0(dn->dn_id_flags);
223
224	ASSERT0(dn->dn_dbufs_count);
225	avl_destroy(&dn->dn_dbufs);
226}
227
228void
229dnode_init(void)
230{
231	ASSERT(dnode_cache == NULL);
232	dnode_cache = kmem_cache_create("dnode_t",
233	    sizeof (dnode_t),
234	    0, dnode_cons, dnode_dest, NULL, NULL, NULL, 0);
235#ifdef	_KERNEL
236	kmem_cache_set_move(dnode_cache, dnode_move);
237
238	dnode_ksp = kstat_create("zfs", 0, "dnodestats", "misc",
239	    KSTAT_TYPE_NAMED, sizeof (dnode_stats) / sizeof (kstat_named_t),
240	    KSTAT_FLAG_VIRTUAL);
241	if (dnode_ksp != NULL) {
242		dnode_ksp->ks_data = &dnode_stats;
243		kstat_install(dnode_ksp);
244	}
245#endif	/* _KERNEL */
246}
247
248void
249dnode_fini(void)
250{
251	if (dnode_ksp != NULL) {
252		kstat_delete(dnode_ksp);
253		dnode_ksp = NULL;
254	}
255
256	kmem_cache_destroy(dnode_cache);
257	dnode_cache = NULL;
258}
259
260
261#ifdef ZFS_DEBUG
262void
263dnode_verify(dnode_t *dn)
264{
265	int drop_struct_lock = FALSE;
266
267	ASSERT(dn->dn_phys);
268	ASSERT(dn->dn_objset);
269	ASSERT(dn->dn_handle->dnh_dnode == dn);
270
271	ASSERT(DMU_OT_IS_VALID(dn->dn_phys->dn_type));
272
273	if (!(zfs_flags & ZFS_DEBUG_DNODE_VERIFY))
274		return;
275
276	if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
277		rw_enter(&dn->dn_struct_rwlock, RW_READER);
278		drop_struct_lock = TRUE;
279	}
280	if (dn->dn_phys->dn_type != DMU_OT_NONE || dn->dn_allocated_txg != 0) {
281		int i;
282		int max_bonuslen = DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots);
283		ASSERT3U(dn->dn_indblkshift, >=, 0);
284		ASSERT3U(dn->dn_indblkshift, <=, SPA_MAXBLOCKSHIFT);
285		if (dn->dn_datablkshift) {
286			ASSERT3U(dn->dn_datablkshift, >=, SPA_MINBLOCKSHIFT);
287			ASSERT3U(dn->dn_datablkshift, <=, SPA_MAXBLOCKSHIFT);
288			ASSERT3U(1<<dn->dn_datablkshift, ==, dn->dn_datablksz);
289		}
290		ASSERT3U(dn->dn_nlevels, <=, 30);
291		ASSERT(DMU_OT_IS_VALID(dn->dn_type));
292		ASSERT3U(dn->dn_nblkptr, >=, 1);
293		ASSERT3U(dn->dn_nblkptr, <=, DN_MAX_NBLKPTR);
294		ASSERT3U(dn->dn_bonuslen, <=, max_bonuslen);
295		ASSERT3U(dn->dn_datablksz, ==,
296		    dn->dn_datablkszsec << SPA_MINBLOCKSHIFT);
297		ASSERT3U(ISP2(dn->dn_datablksz), ==, dn->dn_datablkshift != 0);
298		ASSERT3U((dn->dn_nblkptr - 1) * sizeof (blkptr_t) +
299		    dn->dn_bonuslen, <=, max_bonuslen);
300		for (i = 0; i < TXG_SIZE; i++) {
301			ASSERT3U(dn->dn_next_nlevels[i], <=, dn->dn_nlevels);
302		}
303	}
304	if (dn->dn_phys->dn_type != DMU_OT_NONE)
305		ASSERT3U(dn->dn_phys->dn_nlevels, <=, dn->dn_nlevels);
306	ASSERT(DMU_OBJECT_IS_SPECIAL(dn->dn_object) || dn->dn_dbuf != NULL);
307	if (dn->dn_dbuf != NULL) {
308		ASSERT3P(dn->dn_phys, ==,
309		    (dnode_phys_t *)dn->dn_dbuf->db.db_data +
310		    (dn->dn_object % (dn->dn_dbuf->db.db_size >> DNODE_SHIFT)));
311	}
312	if (drop_struct_lock)
313		rw_exit(&dn->dn_struct_rwlock);
314}
315#endif
316
317void
318dnode_byteswap(dnode_phys_t *dnp)
319{
320	uint64_t *buf64 = (void*)&dnp->dn_blkptr;
321	int i;
322
323	if (dnp->dn_type == DMU_OT_NONE) {
324		bzero(dnp, sizeof (dnode_phys_t));
325		return;
326	}
327
328	dnp->dn_datablkszsec = BSWAP_16(dnp->dn_datablkszsec);
329	dnp->dn_bonuslen = BSWAP_16(dnp->dn_bonuslen);
330	dnp->dn_extra_slots = BSWAP_8(dnp->dn_extra_slots);
331	dnp->dn_maxblkid = BSWAP_64(dnp->dn_maxblkid);
332	dnp->dn_used = BSWAP_64(dnp->dn_used);
333
334	/*
335	 * dn_nblkptr is only one byte, so it's OK to read it in either
336	 * byte order.  We can't read dn_bouslen.
337	 */
338	ASSERT(dnp->dn_indblkshift <= SPA_MAXBLOCKSHIFT);
339	ASSERT(dnp->dn_nblkptr <= DN_MAX_NBLKPTR);
340	for (i = 0; i < dnp->dn_nblkptr * sizeof (blkptr_t)/8; i++)
341		buf64[i] = BSWAP_64(buf64[i]);
342
343	/*
344	 * OK to check dn_bonuslen for zero, because it won't matter if
345	 * we have the wrong byte order.  This is necessary because the
346	 * dnode dnode is smaller than a regular dnode.
347	 */
348	if (dnp->dn_bonuslen != 0) {
349		/*
350		 * Note that the bonus length calculated here may be
351		 * longer than the actual bonus buffer.  This is because
352		 * we always put the bonus buffer after the last block
353		 * pointer (instead of packing it against the end of the
354		 * dnode buffer).
355		 */
356		int off = (dnp->dn_nblkptr-1) * sizeof (blkptr_t);
357		int slots = dnp->dn_extra_slots + 1;
358		size_t len = DN_SLOTS_TO_BONUSLEN(slots) - off;
359		ASSERT(DMU_OT_IS_VALID(dnp->dn_bonustype));
360		dmu_object_byteswap_t byteswap =
361		    DMU_OT_BYTESWAP(dnp->dn_bonustype);
362		dmu_ot_byteswap[byteswap].ob_func(dnp->dn_bonus + off, len);
363	}
364
365	/* Swap SPILL block if we have one */
366	if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR)
367		byteswap_uint64_array(DN_SPILL_BLKPTR(dnp), sizeof (blkptr_t));
368
369}
370
371void
372dnode_buf_byteswap(void *vbuf, size_t size)
373{
374	int i = 0;
375
376	ASSERT3U(sizeof (dnode_phys_t), ==, (1<<DNODE_SHIFT));
377	ASSERT((size & (sizeof (dnode_phys_t)-1)) == 0);
378
379	while (i < size) {
380		dnode_phys_t *dnp = (void *)(((char *)vbuf) + i);
381		dnode_byteswap(dnp);
382
383		i += DNODE_MIN_SIZE;
384		if (dnp->dn_type != DMU_OT_NONE)
385			i += dnp->dn_extra_slots * DNODE_MIN_SIZE;
386	}
387}
388
389void
390dnode_setbonuslen(dnode_t *dn, int newsize, dmu_tx_t *tx)
391{
392	ASSERT3U(zfs_refcount_count(&dn->dn_holds), >=, 1);
393
394	dnode_setdirty(dn, tx);
395	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
396	ASSERT3U(newsize, <=, DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots) -
397	    (dn->dn_nblkptr-1) * sizeof (blkptr_t));
398	dn->dn_bonuslen = newsize;
399	if (newsize == 0)
400		dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = DN_ZERO_BONUSLEN;
401	else
402		dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = dn->dn_bonuslen;
403	rw_exit(&dn->dn_struct_rwlock);
404}
405
406void
407dnode_setbonus_type(dnode_t *dn, dmu_object_type_t newtype, dmu_tx_t *tx)
408{
409	ASSERT3U(zfs_refcount_count(&dn->dn_holds), >=, 1);
410	dnode_setdirty(dn, tx);
411	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
412	dn->dn_bonustype = newtype;
413	dn->dn_next_bonustype[tx->tx_txg & TXG_MASK] = dn->dn_bonustype;
414	rw_exit(&dn->dn_struct_rwlock);
415}
416
417void
418dnode_rm_spill(dnode_t *dn, dmu_tx_t *tx)
419{
420	ASSERT3U(zfs_refcount_count(&dn->dn_holds), >=, 1);
421	ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
422	dnode_setdirty(dn, tx);
423	dn->dn_rm_spillblk[tx->tx_txg&TXG_MASK] = DN_KILL_SPILLBLK;
424	dn->dn_have_spill = B_FALSE;
425}
426
427static void
428dnode_setdblksz(dnode_t *dn, int size)
429{
430	ASSERT0(P2PHASE(size, SPA_MINBLOCKSIZE));
431	ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
432	ASSERT3U(size, >=, SPA_MINBLOCKSIZE);
433	ASSERT3U(size >> SPA_MINBLOCKSHIFT, <,
434	    1<<(sizeof (dn->dn_phys->dn_datablkszsec) * 8));
435	dn->dn_datablksz = size;
436	dn->dn_datablkszsec = size >> SPA_MINBLOCKSHIFT;
437	dn->dn_datablkshift = ISP2(size) ? highbit64(size - 1) : 0;
438}
439
440static dnode_t *
441dnode_create(objset_t *os, dnode_phys_t *dnp, dmu_buf_impl_t *db,
442    uint64_t object, dnode_handle_t *dnh)
443{
444	dnode_t *dn;
445
446	dn = kmem_cache_alloc(dnode_cache, KM_SLEEP);
447#ifdef _KERNEL
448	ASSERT(!POINTER_IS_VALID(dn->dn_objset));
449#endif /* _KERNEL */
450	dn->dn_moved = 0;
451
452	/*
453	 * Defer setting dn_objset until the dnode is ready to be a candidate
454	 * for the dnode_move() callback.
455	 */
456	dn->dn_object = object;
457	dn->dn_dbuf = db;
458	dn->dn_handle = dnh;
459	dn->dn_phys = dnp;
460
461	if (dnp->dn_datablkszsec) {
462		dnode_setdblksz(dn, dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT);
463	} else {
464		dn->dn_datablksz = 0;
465		dn->dn_datablkszsec = 0;
466		dn->dn_datablkshift = 0;
467	}
468	dn->dn_indblkshift = dnp->dn_indblkshift;
469	dn->dn_nlevels = dnp->dn_nlevels;
470	dn->dn_type = dnp->dn_type;
471	dn->dn_nblkptr = dnp->dn_nblkptr;
472	dn->dn_checksum = dnp->dn_checksum;
473	dn->dn_compress = dnp->dn_compress;
474	dn->dn_bonustype = dnp->dn_bonustype;
475	dn->dn_bonuslen = dnp->dn_bonuslen;
476	dn->dn_num_slots = dnp->dn_extra_slots + 1;
477	dn->dn_maxblkid = dnp->dn_maxblkid;
478	dn->dn_have_spill = ((dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) != 0);
479	dn->dn_id_flags = 0;
480
481	dmu_zfetch_init(&dn->dn_zfetch, dn);
482
483	ASSERT(DMU_OT_IS_VALID(dn->dn_phys->dn_type));
484	ASSERT(zrl_is_locked(&dnh->dnh_zrlock));
485	ASSERT(!DN_SLOT_IS_PTR(dnh->dnh_dnode));
486
487	mutex_enter(&os->os_lock);
488
489	/*
490	 * Exclude special dnodes from os_dnodes so an empty os_dnodes
491	 * signifies that the special dnodes have no references from
492	 * their children (the entries in os_dnodes).  This allows
493	 * dnode_destroy() to easily determine if the last child has
494	 * been removed and then complete eviction of the objset.
495	 */
496	if (!DMU_OBJECT_IS_SPECIAL(object))
497		list_insert_head(&os->os_dnodes, dn);
498	membar_producer();
499
500	/*
501	 * Everything else must be valid before assigning dn_objset
502	 * makes the dnode eligible for dnode_move().
503	 */
504	dn->dn_objset = os;
505
506	dnh->dnh_dnode = dn;
507	mutex_exit(&os->os_lock);
508
509	arc_space_consume(sizeof (dnode_t), ARC_SPACE_OTHER);
510
511	return (dn);
512}
513
514/*
515 * Caller must be holding the dnode handle, which is released upon return.
516 */
517static void
518dnode_destroy(dnode_t *dn)
519{
520	objset_t *os = dn->dn_objset;
521	boolean_t complete_os_eviction = B_FALSE;
522
523	ASSERT((dn->dn_id_flags & DN_ID_NEW_EXIST) == 0);
524
525	mutex_enter(&os->os_lock);
526	POINTER_INVALIDATE(&dn->dn_objset);
527	if (!DMU_OBJECT_IS_SPECIAL(dn->dn_object)) {
528		list_remove(&os->os_dnodes, dn);
529		complete_os_eviction =
530		    list_is_empty(&os->os_dnodes) &&
531		    list_link_active(&os->os_evicting_node);
532	}
533	mutex_exit(&os->os_lock);
534
535	/* the dnode can no longer move, so we can release the handle */
536	if (!zrl_is_locked(&dn->dn_handle->dnh_zrlock))
537		zrl_remove(&dn->dn_handle->dnh_zrlock);
538
539	dn->dn_allocated_txg = 0;
540	dn->dn_free_txg = 0;
541	dn->dn_assigned_txg = 0;
542	dn->dn_dirty_txg = 0;
543
544	dn->dn_dirtyctx = 0;
545	if (dn->dn_dirtyctx_firstset != NULL) {
546		kmem_free(dn->dn_dirtyctx_firstset, 1);
547		dn->dn_dirtyctx_firstset = NULL;
548	}
549	if (dn->dn_bonus != NULL) {
550		mutex_enter(&dn->dn_bonus->db_mtx);
551		dbuf_destroy(dn->dn_bonus);
552		dn->dn_bonus = NULL;
553	}
554	dn->dn_zio = NULL;
555
556	dn->dn_have_spill = B_FALSE;
557	dn->dn_oldused = 0;
558	dn->dn_oldflags = 0;
559	dn->dn_olduid = 0;
560	dn->dn_oldgid = 0;
561	dn->dn_oldprojid = ZFS_DEFAULT_PROJID;
562	dn->dn_newuid = 0;
563	dn->dn_newgid = 0;
564	dn->dn_newprojid = ZFS_DEFAULT_PROJID;
565	dn->dn_id_flags = 0;
566
567	dmu_zfetch_fini(&dn->dn_zfetch);
568	kmem_cache_free(dnode_cache, dn);
569	arc_space_return(sizeof (dnode_t), ARC_SPACE_OTHER);
570
571	if (complete_os_eviction)
572		dmu_objset_evict_done(os);
573}
574
575void
576dnode_allocate(dnode_t *dn, dmu_object_type_t ot, int blocksize, int ibs,
577    dmu_object_type_t bonustype, int bonuslen, int dn_slots, dmu_tx_t *tx)
578{
579	int i;
580
581	ASSERT3U(dn_slots, >, 0);
582	ASSERT3U(dn_slots << DNODE_SHIFT, <=,
583	    spa_maxdnodesize(dmu_objset_spa(dn->dn_objset)));
584	ASSERT3U(blocksize, <=,
585	    spa_maxblocksize(dmu_objset_spa(dn->dn_objset)));
586	if (blocksize == 0)
587		blocksize = 1 << zfs_default_bs;
588	else
589		blocksize = P2ROUNDUP(blocksize, SPA_MINBLOCKSIZE);
590
591	if (ibs == 0)
592		ibs = zfs_default_ibs;
593
594	ibs = MIN(MAX(ibs, DN_MIN_INDBLKSHIFT), DN_MAX_INDBLKSHIFT);
595
596	dprintf("os=%p obj=%" PRIu64 " txg=%" PRIu64
597	    " blocksize=%d ibs=%d dn_slots=%d\n",
598	    dn->dn_objset, dn->dn_object, tx->tx_txg, blocksize, ibs, dn_slots);
599	DNODE_STAT_BUMP(dnode_allocate);
600
601	ASSERT(dn->dn_type == DMU_OT_NONE);
602	ASSERT(bcmp(dn->dn_phys, &dnode_phys_zero, sizeof (dnode_phys_t)) == 0);
603	ASSERT(dn->dn_phys->dn_type == DMU_OT_NONE);
604	ASSERT(ot != DMU_OT_NONE);
605	ASSERT(DMU_OT_IS_VALID(ot));
606	ASSERT((bonustype == DMU_OT_NONE && bonuslen == 0) ||
607	    (bonustype == DMU_OT_SA && bonuslen == 0) ||
608	    (bonustype != DMU_OT_NONE && bonuslen != 0));
609	ASSERT(DMU_OT_IS_VALID(bonustype));
610	ASSERT3U(bonuslen, <=, DN_SLOTS_TO_BONUSLEN(dn_slots));
611	ASSERT(dn->dn_type == DMU_OT_NONE);
612	ASSERT0(dn->dn_maxblkid);
613	ASSERT0(dn->dn_allocated_txg);
614	ASSERT0(dn->dn_dirty_txg);
615	ASSERT0(dn->dn_assigned_txg);
616	ASSERT(zfs_refcount_is_zero(&dn->dn_tx_holds));
617	ASSERT3U(zfs_refcount_count(&dn->dn_holds), <=, 1);
618	ASSERT(avl_is_empty(&dn->dn_dbufs));
619
620	for (i = 0; i < TXG_SIZE; i++) {
621		ASSERT0(dn->dn_next_nblkptr[i]);
622		ASSERT0(dn->dn_next_nlevels[i]);
623		ASSERT0(dn->dn_next_indblkshift[i]);
624		ASSERT0(dn->dn_next_bonuslen[i]);
625		ASSERT0(dn->dn_next_bonustype[i]);
626		ASSERT0(dn->dn_rm_spillblk[i]);
627		ASSERT0(dn->dn_next_blksz[i]);
628		ASSERT0(dn->dn_next_maxblkid[i]);
629		ASSERT(!multilist_link_active(&dn->dn_dirty_link[i]));
630		ASSERT3P(list_head(&dn->dn_dirty_records[i]), ==, NULL);
631		ASSERT3P(dn->dn_free_ranges[i], ==, NULL);
632	}
633
634	dn->dn_type = ot;
635	dnode_setdblksz(dn, blocksize);
636	dn->dn_indblkshift = ibs;
637	dn->dn_nlevels = 1;
638	dn->dn_num_slots = dn_slots;
639	if (bonustype == DMU_OT_SA) /* Maximize bonus space for SA */
640		dn->dn_nblkptr = 1;
641	else {
642		dn->dn_nblkptr = MIN(DN_MAX_NBLKPTR,
643		    1 + ((DN_SLOTS_TO_BONUSLEN(dn_slots) - bonuslen) >>
644		    SPA_BLKPTRSHIFT));
645	}
646
647	dn->dn_bonustype = bonustype;
648	dn->dn_bonuslen = bonuslen;
649	dn->dn_checksum = ZIO_CHECKSUM_INHERIT;
650	dn->dn_compress = ZIO_COMPRESS_INHERIT;
651	dn->dn_dirtyctx = 0;
652
653	dn->dn_free_txg = 0;
654	if (dn->dn_dirtyctx_firstset) {
655		kmem_free(dn->dn_dirtyctx_firstset, 1);
656		dn->dn_dirtyctx_firstset = NULL;
657	}
658
659	dn->dn_allocated_txg = tx->tx_txg;
660	dn->dn_id_flags = 0;
661
662	dnode_setdirty(dn, tx);
663	dn->dn_next_indblkshift[tx->tx_txg & TXG_MASK] = ibs;
664	dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = dn->dn_bonuslen;
665	dn->dn_next_bonustype[tx->tx_txg & TXG_MASK] = dn->dn_bonustype;
666	dn->dn_next_blksz[tx->tx_txg & TXG_MASK] = dn->dn_datablksz;
667}
668
669void
670dnode_reallocate(dnode_t *dn, dmu_object_type_t ot, int blocksize,
671    dmu_object_type_t bonustype, int bonuslen, int dn_slots,
672    boolean_t keep_spill, dmu_tx_t *tx)
673{
674	int nblkptr;
675
676	ASSERT3U(blocksize, >=, SPA_MINBLOCKSIZE);
677	ASSERT3U(blocksize, <=,
678	    spa_maxblocksize(dmu_objset_spa(dn->dn_objset)));
679	ASSERT0(blocksize % SPA_MINBLOCKSIZE);
680	ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT || dmu_tx_private_ok(tx));
681	ASSERT(tx->tx_txg != 0);
682	ASSERT((bonustype == DMU_OT_NONE && bonuslen == 0) ||
683	    (bonustype != DMU_OT_NONE && bonuslen != 0) ||
684	    (bonustype == DMU_OT_SA && bonuslen == 0));
685	ASSERT(DMU_OT_IS_VALID(bonustype));
686	ASSERT3U(bonuslen, <=,
687	    DN_BONUS_SIZE(spa_maxdnodesize(dmu_objset_spa(dn->dn_objset))));
688	ASSERT3U(bonuslen, <=, DN_BONUS_SIZE(dn_slots << DNODE_SHIFT));
689
690	dnode_free_interior_slots(dn);
691	DNODE_STAT_BUMP(dnode_reallocate);
692
693	/* clean up any unreferenced dbufs */
694	dnode_evict_dbufs(dn);
695
696	dn->dn_id_flags = 0;
697
698	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
699	dnode_setdirty(dn, tx);
700	if (dn->dn_datablksz != blocksize) {
701		/* change blocksize */
702		ASSERT(dn->dn_maxblkid == 0 &&
703		    (BP_IS_HOLE(&dn->dn_phys->dn_blkptr[0]) ||
704		    dnode_block_freed(dn, 0)));
705		dnode_setdblksz(dn, blocksize);
706		dn->dn_next_blksz[tx->tx_txg&TXG_MASK] = blocksize;
707	}
708	if (dn->dn_bonuslen != bonuslen)
709		dn->dn_next_bonuslen[tx->tx_txg&TXG_MASK] = bonuslen;
710
711	if (bonustype == DMU_OT_SA) /* Maximize bonus space for SA */
712		nblkptr = 1;
713	else
714		nblkptr = MIN(DN_MAX_NBLKPTR,
715		    1 + ((DN_SLOTS_TO_BONUSLEN(dn_slots) - bonuslen) >>
716		    SPA_BLKPTRSHIFT));
717	if (dn->dn_bonustype != bonustype)
718		dn->dn_next_bonustype[tx->tx_txg&TXG_MASK] = bonustype;
719	if (dn->dn_nblkptr != nblkptr)
720		dn->dn_next_nblkptr[tx->tx_txg&TXG_MASK] = nblkptr;
721	if (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR && !keep_spill) {
722		dbuf_rm_spill(dn, tx);
723		dnode_rm_spill(dn, tx);
724	}
725	rw_exit(&dn->dn_struct_rwlock);
726
727	/* change type */
728	dn->dn_type = ot;
729
730	/* change bonus size and type */
731	mutex_enter(&dn->dn_mtx);
732	dn->dn_bonustype = bonustype;
733	dn->dn_bonuslen = bonuslen;
734	dn->dn_num_slots = dn_slots;
735	dn->dn_nblkptr = nblkptr;
736	dn->dn_checksum = ZIO_CHECKSUM_INHERIT;
737	dn->dn_compress = ZIO_COMPRESS_INHERIT;
738	ASSERT3U(dn->dn_nblkptr, <=, DN_MAX_NBLKPTR);
739
740	/* fix up the bonus db_size */
741	if (dn->dn_bonus) {
742		dn->dn_bonus->db.db_size =
743		    DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots) -
744		    (dn->dn_nblkptr - 1) * sizeof (blkptr_t);
745		ASSERT(dn->dn_bonuslen <= dn->dn_bonus->db.db_size);
746	}
747
748	dn->dn_allocated_txg = tx->tx_txg;
749	mutex_exit(&dn->dn_mtx);
750}
751
752#ifdef	_KERNEL
753static void
754dnode_move_impl(dnode_t *odn, dnode_t *ndn)
755{
756	int i;
757
758	ASSERT(!RW_LOCK_HELD(&odn->dn_struct_rwlock));
759	ASSERT(MUTEX_NOT_HELD(&odn->dn_mtx));
760	ASSERT(MUTEX_NOT_HELD(&odn->dn_dbufs_mtx));
761	ASSERT(!RW_LOCK_HELD(&odn->dn_zfetch.zf_rwlock));
762
763	/* Copy fields. */
764	ndn->dn_objset = odn->dn_objset;
765	ndn->dn_object = odn->dn_object;
766	ndn->dn_dbuf = odn->dn_dbuf;
767	ndn->dn_handle = odn->dn_handle;
768	ndn->dn_phys = odn->dn_phys;
769	ndn->dn_type = odn->dn_type;
770	ndn->dn_bonuslen = odn->dn_bonuslen;
771	ndn->dn_bonustype = odn->dn_bonustype;
772	ndn->dn_nblkptr = odn->dn_nblkptr;
773	ndn->dn_checksum = odn->dn_checksum;
774	ndn->dn_compress = odn->dn_compress;
775	ndn->dn_nlevels = odn->dn_nlevels;
776	ndn->dn_indblkshift = odn->dn_indblkshift;
777	ndn->dn_datablkshift = odn->dn_datablkshift;
778	ndn->dn_datablkszsec = odn->dn_datablkszsec;
779	ndn->dn_datablksz = odn->dn_datablksz;
780	ndn->dn_maxblkid = odn->dn_maxblkid;
781	ndn->dn_num_slots = odn->dn_num_slots;
782	bcopy(&odn->dn_next_type[0], &ndn->dn_next_type[0],
783	    sizeof (odn->dn_next_type));
784	bcopy(&odn->dn_next_nblkptr[0], &ndn->dn_next_nblkptr[0],
785	    sizeof (odn->dn_next_nblkptr));
786	bcopy(&odn->dn_next_nlevels[0], &ndn->dn_next_nlevels[0],
787	    sizeof (odn->dn_next_nlevels));
788	bcopy(&odn->dn_next_indblkshift[0], &ndn->dn_next_indblkshift[0],
789	    sizeof (odn->dn_next_indblkshift));
790	bcopy(&odn->dn_next_bonustype[0], &ndn->dn_next_bonustype[0],
791	    sizeof (odn->dn_next_bonustype));
792	bcopy(&odn->dn_rm_spillblk[0], &ndn->dn_rm_spillblk[0],
793	    sizeof (odn->dn_rm_spillblk));
794	bcopy(&odn->dn_next_bonuslen[0], &ndn->dn_next_bonuslen[0],
795	    sizeof (odn->dn_next_bonuslen));
796	bcopy(&odn->dn_next_blksz[0], &ndn->dn_next_blksz[0],
797	    sizeof (odn->dn_next_blksz));
798	bcopy(&odn->dn_next_maxblkid[0], &ndn->dn_next_maxblkid[0],
799	    sizeof (odn->dn_next_maxblkid));
800	for (i = 0; i < TXG_SIZE; i++) {
801		list_move_tail(&ndn->dn_dirty_records[i],
802		    &odn->dn_dirty_records[i]);
803	}
804	bcopy(&odn->dn_free_ranges[0], &ndn->dn_free_ranges[0],
805	    sizeof (odn->dn_free_ranges));
806	ndn->dn_allocated_txg = odn->dn_allocated_txg;
807	ndn->dn_free_txg = odn->dn_free_txg;
808	ndn->dn_assigned_txg = odn->dn_assigned_txg;
809	ndn->dn_dirty_txg = odn->dn_dirty_txg;
810	ndn->dn_dirtyctx = odn->dn_dirtyctx;
811	ndn->dn_dirtyctx_firstset = odn->dn_dirtyctx_firstset;
812	ASSERT(zfs_refcount_count(&odn->dn_tx_holds) == 0);
813	zfs_refcount_transfer(&ndn->dn_holds, &odn->dn_holds);
814	ASSERT(avl_is_empty(&ndn->dn_dbufs));
815	avl_swap(&ndn->dn_dbufs, &odn->dn_dbufs);
816	ndn->dn_dbufs_count = odn->dn_dbufs_count;
817	ndn->dn_bonus = odn->dn_bonus;
818	ndn->dn_have_spill = odn->dn_have_spill;
819	ndn->dn_zio = odn->dn_zio;
820	ndn->dn_oldused = odn->dn_oldused;
821	ndn->dn_oldflags = odn->dn_oldflags;
822	ndn->dn_olduid = odn->dn_olduid;
823	ndn->dn_oldgid = odn->dn_oldgid;
824	ndn->dn_oldprojid = odn->dn_oldprojid;
825	ndn->dn_newuid = odn->dn_newuid;
826	ndn->dn_newgid = odn->dn_newgid;
827	ndn->dn_newprojid = odn->dn_newprojid;
828	ndn->dn_id_flags = odn->dn_id_flags;
829	dmu_zfetch_init(&ndn->dn_zfetch, NULL);
830	list_move_tail(&ndn->dn_zfetch.zf_stream, &odn->dn_zfetch.zf_stream);
831	ndn->dn_zfetch.zf_dnode = odn->dn_zfetch.zf_dnode;
832
833	/*
834	 * Update back pointers. Updating the handle fixes the back pointer of
835	 * every descendant dbuf as well as the bonus dbuf.
836	 */
837	ASSERT(ndn->dn_handle->dnh_dnode == odn);
838	ndn->dn_handle->dnh_dnode = ndn;
839	if (ndn->dn_zfetch.zf_dnode == odn) {
840		ndn->dn_zfetch.zf_dnode = ndn;
841	}
842
843	/*
844	 * Invalidate the original dnode by clearing all of its back pointers.
845	 */
846	odn->dn_dbuf = NULL;
847	odn->dn_handle = NULL;
848	avl_create(&odn->dn_dbufs, dbuf_compare, sizeof (dmu_buf_impl_t),
849	    offsetof(dmu_buf_impl_t, db_link));
850	odn->dn_dbufs_count = 0;
851	odn->dn_bonus = NULL;
852	odn->dn_zfetch.zf_dnode = NULL;
853
854	/*
855	 * Set the low bit of the objset pointer to ensure that dnode_move()
856	 * recognizes the dnode as invalid in any subsequent callback.
857	 */
858	POINTER_INVALIDATE(&odn->dn_objset);
859
860	/*
861	 * Satisfy the destructor.
862	 */
863	for (i = 0; i < TXG_SIZE; i++) {
864		list_create(&odn->dn_dirty_records[i],
865		    sizeof (dbuf_dirty_record_t),
866		    offsetof(dbuf_dirty_record_t, dr_dirty_node));
867		odn->dn_free_ranges[i] = NULL;
868		odn->dn_next_nlevels[i] = 0;
869		odn->dn_next_indblkshift[i] = 0;
870		odn->dn_next_bonustype[i] = 0;
871		odn->dn_rm_spillblk[i] = 0;
872		odn->dn_next_bonuslen[i] = 0;
873		odn->dn_next_blksz[i] = 0;
874	}
875	odn->dn_allocated_txg = 0;
876	odn->dn_free_txg = 0;
877	odn->dn_assigned_txg = 0;
878	odn->dn_dirty_txg = 0;
879	odn->dn_dirtyctx = 0;
880	odn->dn_dirtyctx_firstset = NULL;
881	odn->dn_have_spill = B_FALSE;
882	odn->dn_zio = NULL;
883	odn->dn_oldused = 0;
884	odn->dn_oldflags = 0;
885	odn->dn_olduid = 0;
886	odn->dn_oldgid = 0;
887	odn->dn_oldprojid = ZFS_DEFAULT_PROJID;
888	odn->dn_newuid = 0;
889	odn->dn_newgid = 0;
890	odn->dn_newprojid = ZFS_DEFAULT_PROJID;
891	odn->dn_id_flags = 0;
892
893	/*
894	 * Mark the dnode.
895	 */
896	ndn->dn_moved = 1;
897	odn->dn_moved = (uint8_t)-1;
898}
899
900/*ARGSUSED*/
901static kmem_cbrc_t
902dnode_move(void *buf, void *newbuf, size_t size, void *arg)
903{
904	dnode_t *odn = buf, *ndn = newbuf;
905	objset_t *os;
906	int64_t refcount;
907	uint32_t dbufs;
908
909	/*
910	 * The dnode is on the objset's list of known dnodes if the objset
911	 * pointer is valid. We set the low bit of the objset pointer when
912	 * freeing the dnode to invalidate it, and the memory patterns written
913	 * by kmem (baddcafe and deadbeef) set at least one of the two low bits.
914	 * A newly created dnode sets the objset pointer last of all to indicate
915	 * that the dnode is known and in a valid state to be moved by this
916	 * function.
917	 */
918	os = odn->dn_objset;
919	if (!POINTER_IS_VALID(os)) {
920		DNODE_STAT_BUMP(dnode_move_invalid);
921		return (KMEM_CBRC_DONT_KNOW);
922	}
923
924	/*
925	 * Ensure that the objset does not go away during the move.
926	 */
927	rw_enter(&os_lock, RW_WRITER);
928	if (os != odn->dn_objset) {
929		rw_exit(&os_lock);
930		DNODE_STAT_BUMP(dnode_move_recheck1);
931		return (KMEM_CBRC_DONT_KNOW);
932	}
933
934	/*
935	 * If the dnode is still valid, then so is the objset. We know that no
936	 * valid objset can be freed while we hold os_lock, so we can safely
937	 * ensure that the objset remains in use.
938	 */
939	mutex_enter(&os->os_lock);
940
941	/*
942	 * Recheck the objset pointer in case the dnode was removed just before
943	 * acquiring the lock.
944	 */
945	if (os != odn->dn_objset) {
946		mutex_exit(&os->os_lock);
947		rw_exit(&os_lock);
948		DNODE_STAT_BUMP(dnode_move_recheck2);
949		return (KMEM_CBRC_DONT_KNOW);
950	}
951
952	/*
953	 * At this point we know that as long as we hold os->os_lock, the dnode
954	 * cannot be freed and fields within the dnode can be safely accessed.
955	 * The objset listing this dnode cannot go away as long as this dnode is
956	 * on its list.
957	 */
958	rw_exit(&os_lock);
959	if (DMU_OBJECT_IS_SPECIAL(odn->dn_object)) {
960		mutex_exit(&os->os_lock);
961		DNODE_STAT_BUMP(dnode_move_special);
962		return (KMEM_CBRC_NO);
963	}
964	ASSERT(odn->dn_dbuf != NULL); /* only "special" dnodes have no parent */
965
966	/*
967	 * Lock the dnode handle to prevent the dnode from obtaining any new
968	 * holds. This also prevents the descendant dbufs and the bonus dbuf
969	 * from accessing the dnode, so that we can discount their holds. The
970	 * handle is safe to access because we know that while the dnode cannot
971	 * go away, neither can its handle. Once we hold dnh_zrlock, we can
972	 * safely move any dnode referenced only by dbufs.
973	 */
974	if (!zrl_tryenter(&odn->dn_handle->dnh_zrlock)) {
975		mutex_exit(&os->os_lock);
976		DNODE_STAT_BUMP(dnode_move_handle);
977		return (KMEM_CBRC_LATER);
978	}
979
980	/*
981	 * Ensure a consistent view of the dnode's holds and the dnode's dbufs.
982	 * We need to guarantee that there is a hold for every dbuf in order to
983	 * determine whether the dnode is actively referenced. Falsely matching
984	 * a dbuf to an active hold would lead to an unsafe move. It's possible
985	 * that a thread already having an active dnode hold is about to add a
986	 * dbuf, and we can't compare hold and dbuf counts while the add is in
987	 * progress.
988	 */
989	if (!rw_tryenter(&odn->dn_struct_rwlock, RW_WRITER)) {
990		zrl_exit(&odn->dn_handle->dnh_zrlock);
991		mutex_exit(&os->os_lock);
992		DNODE_STAT_BUMP(dnode_move_rwlock);
993		return (KMEM_CBRC_LATER);
994	}
995
996	/*
997	 * A dbuf may be removed (evicted) without an active dnode hold. In that
998	 * case, the dbuf count is decremented under the handle lock before the
999	 * dbuf's hold is released. This order ensures that if we count the hold
1000	 * after the dbuf is removed but before its hold is released, we will
1001	 * treat the unmatched hold as active and exit safely. If we count the
1002	 * hold before the dbuf is removed, the hold is discounted, and the
1003	 * removal is blocked until the move completes.
1004	 */
1005	refcount = zfs_refcount_count(&odn->dn_holds);
1006	ASSERT(refcount >= 0);
1007	dbufs = odn->dn_dbufs_count;
1008
1009	/* We can't have more dbufs than dnode holds. */
1010	ASSERT3U(dbufs, <=, refcount);
1011	DTRACE_PROBE3(dnode__move, dnode_t *, odn, int64_t, refcount,
1012	    uint32_t, dbufs);
1013
1014	if (refcount > dbufs) {
1015		rw_exit(&odn->dn_struct_rwlock);
1016		zrl_exit(&odn->dn_handle->dnh_zrlock);
1017		mutex_exit(&os->os_lock);
1018		DNODE_STAT_BUMP(dnode_move_active);
1019		return (KMEM_CBRC_LATER);
1020	}
1021
1022	rw_exit(&odn->dn_struct_rwlock);
1023
1024	/*
1025	 * At this point we know that anyone with a hold on the dnode is not
1026	 * actively referencing it. The dnode is known and in a valid state to
1027	 * move. We're holding the locks needed to execute the critical section.
1028	 */
1029	dnode_move_impl(odn, ndn);
1030
1031	list_link_replace(&odn->dn_link, &ndn->dn_link);
1032	/* If the dnode was safe to move, the refcount cannot have changed. */
1033	ASSERT(refcount == zfs_refcount_count(&ndn->dn_holds));
1034	ASSERT(dbufs == ndn->dn_dbufs_count);
1035	zrl_exit(&ndn->dn_handle->dnh_zrlock); /* handle has moved */
1036	mutex_exit(&os->os_lock);
1037
1038	return (KMEM_CBRC_YES);
1039}
1040#endif	/* _KERNEL */
1041
1042static void
1043dnode_slots_hold(dnode_children_t *children, int idx, int slots)
1044{
1045	ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK);
1046
1047	for (int i = idx; i < idx + slots; i++) {
1048		dnode_handle_t *dnh = &children->dnc_children[i];
1049		zrl_add(&dnh->dnh_zrlock);
1050	}
1051}
1052
1053static void
1054dnode_slots_rele(dnode_children_t *children, int idx, int slots)
1055{
1056	ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK);
1057
1058	for (int i = idx; i < idx + slots; i++) {
1059		dnode_handle_t *dnh = &children->dnc_children[i];
1060
1061		if (zrl_is_locked(&dnh->dnh_zrlock))
1062			zrl_exit(&dnh->dnh_zrlock);
1063		else
1064			zrl_remove(&dnh->dnh_zrlock);
1065	}
1066}
1067
1068static int
1069dnode_slots_tryenter(dnode_children_t *children, int idx, int slots)
1070{
1071	ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK);
1072
1073	for (int i = idx; i < idx + slots; i++) {
1074		dnode_handle_t *dnh = &children->dnc_children[i];
1075
1076		if (!zrl_tryenter(&dnh->dnh_zrlock)) {
1077			for (int j = idx; j < i; j++) {
1078				dnh = &children->dnc_children[j];
1079				zrl_exit(&dnh->dnh_zrlock);
1080			}
1081
1082			return (0);
1083		}
1084	}
1085
1086	return (1);
1087}
1088
1089static void
1090dnode_set_slots(dnode_children_t *children, int idx, int slots, void *ptr)
1091{
1092	ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK);
1093
1094	for (int i = idx; i < idx + slots; i++) {
1095		dnode_handle_t *dnh = &children->dnc_children[i];
1096		dnh->dnh_dnode = ptr;
1097	}
1098}
1099
1100static boolean_t
1101dnode_check_slots_free(dnode_children_t *children, int idx, int slots)
1102{
1103	ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK);
1104
1105	/*
1106	 * If all dnode slots are either already free or
1107	 * evictable return B_TRUE.
1108	 */
1109	for (int i = idx; i < idx + slots; i++) {
1110		dnode_handle_t *dnh = &children->dnc_children[i];
1111		dnode_t *dn = dnh->dnh_dnode;
1112
1113		if (dn == DN_SLOT_FREE) {
1114			continue;
1115		} else if (DN_SLOT_IS_PTR(dn)) {
1116			mutex_enter(&dn->dn_mtx);
1117			boolean_t can_free = (dn->dn_type == DMU_OT_NONE &&
1118			    zfs_refcount_is_zero(&dn->dn_holds) &&
1119			    !DNODE_IS_DIRTY(dn));
1120			mutex_exit(&dn->dn_mtx);
1121
1122			if (!can_free)
1123				return (B_FALSE);
1124			else
1125				continue;
1126		} else {
1127			return (B_FALSE);
1128		}
1129	}
1130
1131	return (B_TRUE);
1132}
1133
1134static void
1135dnode_reclaim_slots(dnode_children_t *children, int idx, int slots)
1136{
1137	ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK);
1138
1139	for (int i = idx; i < idx + slots; i++) {
1140		dnode_handle_t *dnh = &children->dnc_children[i];
1141
1142		ASSERT(zrl_is_locked(&dnh->dnh_zrlock));
1143
1144		if (DN_SLOT_IS_PTR(dnh->dnh_dnode)) {
1145			ASSERT3S(dnh->dnh_dnode->dn_type, ==, DMU_OT_NONE);
1146			dnode_destroy(dnh->dnh_dnode);
1147			dnh->dnh_dnode = DN_SLOT_FREE;
1148		}
1149	}
1150}
1151
1152void
1153dnode_free_interior_slots(dnode_t *dn)
1154{
1155	dnode_children_t *children = dmu_buf_get_user(&dn->dn_dbuf->db);
1156	int epb = dn->dn_dbuf->db.db_size >> DNODE_SHIFT;
1157	int idx = (dn->dn_object & (epb - 1)) + 1;
1158	int slots = dn->dn_num_slots - 1;
1159
1160	if (slots == 0)
1161		return;
1162
1163	ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK);
1164
1165	while (!dnode_slots_tryenter(children, idx, slots))
1166		DNODE_STAT_BUMP(dnode_free_interior_lock_retry);
1167
1168	dnode_set_slots(children, idx, slots, DN_SLOT_FREE);
1169	dnode_slots_rele(children, idx, slots);
1170}
1171
1172void
1173dnode_special_close(dnode_handle_t *dnh)
1174{
1175	dnode_t *dn = dnh->dnh_dnode;
1176
1177	/*
1178	 * Wait for final references to the dnode to clear.  This can
1179	 * only happen if the arc is asynchronously evicting state that
1180	 * has a hold on this dnode while we are trying to evict this
1181	 * dnode.
1182	 */
1183	while (zfs_refcount_count(&dn->dn_holds) > 0)
1184		delay(1);
1185	ASSERT(dn->dn_dbuf == NULL ||
1186	    dmu_buf_get_user(&dn->dn_dbuf->db) == NULL);
1187	zrl_add(&dnh->dnh_zrlock);
1188	dnode_destroy(dn); /* implicit zrl_remove() */
1189	zrl_destroy(&dnh->dnh_zrlock);
1190	dnh->dnh_dnode = NULL;
1191}
1192
1193void
1194dnode_special_open(objset_t *os, dnode_phys_t *dnp, uint64_t object,
1195    dnode_handle_t *dnh)
1196{
1197	dnode_t *dn;
1198
1199	zrl_init(&dnh->dnh_zrlock);
1200	zrl_tryenter(&dnh->dnh_zrlock);
1201
1202	dn = dnode_create(os, dnp, NULL, object, dnh);
1203	DNODE_VERIFY(dn);
1204
1205	zrl_exit(&dnh->dnh_zrlock);
1206}
1207
1208static void
1209dnode_buf_evict_async(void *dbu)
1210{
1211	dnode_children_t *dnc = dbu;
1212
1213	DNODE_STAT_BUMP(dnode_buf_evict);
1214
1215	for (int i = 0; i < dnc->dnc_count; i++) {
1216		dnode_handle_t *dnh = &dnc->dnc_children[i];
1217		dnode_t *dn;
1218
1219		/*
1220		 * The dnode handle lock guards against the dnode moving to
1221		 * another valid address, so there is no need here to guard
1222		 * against changes to or from NULL.
1223		 */
1224		if (!DN_SLOT_IS_PTR(dnh->dnh_dnode)) {
1225			zrl_destroy(&dnh->dnh_zrlock);
1226			dnh->dnh_dnode = DN_SLOT_UNINIT;
1227			continue;
1228		}
1229
1230		zrl_add(&dnh->dnh_zrlock);
1231		dn = dnh->dnh_dnode;
1232		/*
1233		 * If there are holds on this dnode, then there should
1234		 * be holds on the dnode's containing dbuf as well; thus
1235		 * it wouldn't be eligible for eviction and this function
1236		 * would not have been called.
1237		 */
1238		ASSERT(zfs_refcount_is_zero(&dn->dn_holds));
1239		ASSERT(zfs_refcount_is_zero(&dn->dn_tx_holds));
1240
1241		dnode_destroy(dn); /* implicit zrl_remove() for first slot */
1242		zrl_destroy(&dnh->dnh_zrlock);
1243		dnh->dnh_dnode = DN_SLOT_UNINIT;
1244	}
1245	kmem_free(dnc, sizeof (dnode_children_t) +
1246	    dnc->dnc_count * sizeof (dnode_handle_t));
1247}
1248
1249/*
1250 * When the DNODE_MUST_BE_FREE flag is set, the "slots" parameter is used
1251 * to ensure the hole at the specified object offset is large enough to
1252 * hold the dnode being created. The slots parameter is also used to ensure
1253 * a dnode does not span multiple dnode blocks. In both of these cases, if
1254 * a failure occurs, ENOSPC is returned. Keep in mind, these failure cases
1255 * are only possible when using DNODE_MUST_BE_FREE.
1256 *
1257 * If the DNODE_MUST_BE_ALLOCATED flag is set, "slots" must be 0.
1258 * dnode_hold_impl() will check if the requested dnode is already consumed
1259 * as an extra dnode slot by an large dnode, in which case it returns
1260 * ENOENT.
1261 *
1262 * If the DNODE_DRY_RUN flag is set, we don't actually hold the dnode, just
1263 * return whether the hold would succeed or not. tag and dnp should set to
1264 * NULL in this case.
1265 *
1266 * errors:
1267 * EINVAL - invalid object number or flags.
1268 * ENOSPC - hole too small to fulfill "slots" request (DNODE_MUST_BE_FREE)
1269 * EEXIST - Refers to an allocated dnode (DNODE_MUST_BE_FREE)
1270 *        - Refers to a freeing dnode (DNODE_MUST_BE_FREE)
1271 *        - Refers to an interior dnode slot (DNODE_MUST_BE_ALLOCATED)
1272 * ENOENT - The requested dnode is not allocated (DNODE_MUST_BE_ALLOCATED)
1273 *        - The requested dnode is being freed (DNODE_MUST_BE_ALLOCATED)
1274 * EIO    - i/o error error when reading the meta dnode dbuf.
1275 * succeeds even for free dnodes.
1276 */
1277int
1278dnode_hold_impl(objset_t *os, uint64_t object, int flag, int slots,
1279    void *tag, dnode_t **dnp)
1280{
1281	int epb, idx, err;
1282	int drop_struct_lock = FALSE;
1283	int type;
1284	uint64_t blk;
1285	dnode_t *mdn, *dn;
1286	dmu_buf_impl_t *db;
1287	dnode_children_t *dnc;
1288	dnode_phys_t *dn_block;
1289	dnode_handle_t *dnh;
1290
1291	ASSERT(!(flag & DNODE_MUST_BE_ALLOCATED) || (slots == 0));
1292	ASSERT(!(flag & DNODE_MUST_BE_FREE) || (slots > 0));
1293	IMPLY(flag & DNODE_DRY_RUN, (tag == NULL) && (dnp == NULL));
1294
1295	/*
1296	 * If you are holding the spa config lock as writer, you shouldn't
1297	 * be asking the DMU to do *anything* unless it's the root pool
1298	 * which may require us to read from the root filesystem while
1299	 * holding some (not all) of the locks as writer.
1300	 */
1301	ASSERT(spa_config_held(os->os_spa, SCL_ALL, RW_WRITER) == 0 ||
1302	    (spa_is_root(os->os_spa) &&
1303	    spa_config_held(os->os_spa, SCL_STATE, RW_WRITER)));
1304
1305	ASSERT((flag & DNODE_MUST_BE_ALLOCATED) || (flag & DNODE_MUST_BE_FREE));
1306
1307	if (object == DMU_USERUSED_OBJECT || object == DMU_GROUPUSED_OBJECT ||
1308	    object == DMU_PROJECTUSED_OBJECT) {
1309		if (object == DMU_USERUSED_OBJECT)
1310			dn = DMU_USERUSED_DNODE(os);
1311		else if (object == DMU_GROUPUSED_OBJECT)
1312			dn = DMU_GROUPUSED_DNODE(os);
1313		else
1314			dn = DMU_PROJECTUSED_DNODE(os);
1315		if (dn == NULL)
1316			return (SET_ERROR(ENOENT));
1317		type = dn->dn_type;
1318		if ((flag & DNODE_MUST_BE_ALLOCATED) && type == DMU_OT_NONE)
1319			return (SET_ERROR(ENOENT));
1320		if ((flag & DNODE_MUST_BE_FREE) && type != DMU_OT_NONE)
1321			return (SET_ERROR(EEXIST));
1322		DNODE_VERIFY(dn);
1323		/* Don't actually hold if dry run, just return 0 */
1324		if (!(flag & DNODE_DRY_RUN)) {
1325			(void) zfs_refcount_add(&dn->dn_holds, tag);
1326			*dnp = dn;
1327		}
1328		return (0);
1329	}
1330
1331	if (object == 0 || object >= DN_MAX_OBJECT)
1332		return (SET_ERROR(EINVAL));
1333
1334	mdn = DMU_META_DNODE(os);
1335	ASSERT(mdn->dn_object == DMU_META_DNODE_OBJECT);
1336
1337	DNODE_VERIFY(mdn);
1338
1339	if (!RW_WRITE_HELD(&mdn->dn_struct_rwlock)) {
1340		rw_enter(&mdn->dn_struct_rwlock, RW_READER);
1341		drop_struct_lock = TRUE;
1342	}
1343
1344	blk = dbuf_whichblock(mdn, 0, object * sizeof (dnode_phys_t));
1345
1346	db = dbuf_hold(mdn, blk, FTAG);
1347	if (drop_struct_lock)
1348		rw_exit(&mdn->dn_struct_rwlock);
1349	if (db == NULL) {
1350		DNODE_STAT_BUMP(dnode_hold_dbuf_hold);
1351		return (SET_ERROR(EIO));
1352	}
1353	/*
1354	 * We do not need to decrypt to read the dnode so it doesn't matter
1355	 * if we get the encrypted or decrypted version.
1356	 */
1357	err = dbuf_read(db, NULL, DB_RF_CANFAIL | DB_RF_NO_DECRYPT);
1358	if (err) {
1359		DNODE_STAT_BUMP(dnode_hold_dbuf_read);
1360		dbuf_rele(db, FTAG);
1361		return (err);
1362	}
1363
1364	ASSERT3U(db->db.db_size, >=, 1<<DNODE_SHIFT);
1365	epb = db->db.db_size >> DNODE_SHIFT;
1366
1367	idx = object & (epb - 1);
1368	dn_block = (dnode_phys_t *)db->db.db_data;
1369
1370	ASSERT(DB_DNODE(db)->dn_type == DMU_OT_DNODE);
1371	dnc = dmu_buf_get_user(&db->db);
1372	dnh = NULL;
1373	if (dnc == NULL) {
1374		dnode_children_t *winner;
1375		int skip = 0;
1376
1377		dnc = kmem_zalloc(sizeof (dnode_children_t) +
1378		    epb * sizeof (dnode_handle_t), KM_SLEEP);
1379		dnc->dnc_count = epb;
1380		dnh = &dnc->dnc_children[0];
1381
1382		/* Initialize dnode slot status from dnode_phys_t */
1383		for (int i = 0; i < epb; i++) {
1384			zrl_init(&dnh[i].dnh_zrlock);
1385
1386			if (skip) {
1387				skip--;
1388				continue;
1389			}
1390
1391			if (dn_block[i].dn_type != DMU_OT_NONE) {
1392				int interior = dn_block[i].dn_extra_slots;
1393
1394				dnode_set_slots(dnc, i, 1, DN_SLOT_ALLOCATED);
1395				dnode_set_slots(dnc, i + 1, interior,
1396				    DN_SLOT_INTERIOR);
1397				skip = interior;
1398			} else {
1399				dnh[i].dnh_dnode = DN_SLOT_FREE;
1400				skip = 0;
1401			}
1402		}
1403
1404		dmu_buf_init_user(&dnc->dnc_dbu, NULL,
1405		    dnode_buf_evict_async, NULL);
1406		winner = dmu_buf_set_user(&db->db, &dnc->dnc_dbu);
1407		if (winner != NULL) {
1408
1409			for (int i = 0; i < epb; i++)
1410				zrl_destroy(&dnh[i].dnh_zrlock);
1411
1412			kmem_free(dnc, sizeof (dnode_children_t) +
1413			    epb * sizeof (dnode_handle_t));
1414			dnc = winner;
1415		}
1416	}
1417
1418	ASSERT(dnc->dnc_count == epb);
1419	dn = DN_SLOT_UNINIT;
1420
1421	if (flag & DNODE_MUST_BE_ALLOCATED) {
1422		slots = 1;
1423
1424		while (dn == DN_SLOT_UNINIT) {
1425			dnode_slots_hold(dnc, idx, slots);
1426			dnh = &dnc->dnc_children[idx];
1427
1428			if (DN_SLOT_IS_PTR(dnh->dnh_dnode)) {
1429				dn = dnh->dnh_dnode;
1430				break;
1431			} else if (dnh->dnh_dnode == DN_SLOT_INTERIOR) {
1432				DNODE_STAT_BUMP(dnode_hold_alloc_interior);
1433				dnode_slots_rele(dnc, idx, slots);
1434				dbuf_rele(db, FTAG);
1435				return (SET_ERROR(EEXIST));
1436			} else if (dnh->dnh_dnode != DN_SLOT_ALLOCATED) {
1437				DNODE_STAT_BUMP(dnode_hold_alloc_misses);
1438				dnode_slots_rele(dnc, idx, slots);
1439				dbuf_rele(db, FTAG);
1440				return (SET_ERROR(ENOENT));
1441			}
1442
1443			dnode_slots_rele(dnc, idx, slots);
1444			if (!dnode_slots_tryenter(dnc, idx, slots)) {
1445				DNODE_STAT_BUMP(dnode_hold_alloc_lock_retry);
1446				continue;
1447			}
1448
1449			/*
1450			 * Someone else won the race and called dnode_create()
1451			 * after we checked DN_SLOT_IS_PTR() above but before
1452			 * we acquired the lock.
1453			 */
1454			if (DN_SLOT_IS_PTR(dnh->dnh_dnode)) {
1455				DNODE_STAT_BUMP(dnode_hold_alloc_lock_misses);
1456				dn = dnh->dnh_dnode;
1457			} else {
1458				dn = dnode_create(os, dn_block + idx, db,
1459				    object, dnh);
1460			}
1461		}
1462
1463		mutex_enter(&dn->dn_mtx);
1464		if (dn->dn_type == DMU_OT_NONE || dn->dn_free_txg != 0) {
1465			DNODE_STAT_BUMP(dnode_hold_alloc_type_none);
1466			mutex_exit(&dn->dn_mtx);
1467			dnode_slots_rele(dnc, idx, slots);
1468			dbuf_rele(db, FTAG);
1469			return (SET_ERROR(ENOENT));
1470		}
1471
1472		/* Don't actually hold if dry run, just return 0 */
1473		if (flag & DNODE_DRY_RUN) {
1474			mutex_exit(&dn->dn_mtx);
1475			dnode_slots_rele(dnc, idx, slots);
1476			dbuf_rele(db, FTAG);
1477			return (0);
1478		}
1479
1480		DNODE_STAT_BUMP(dnode_hold_alloc_hits);
1481	} else if (flag & DNODE_MUST_BE_FREE) {
1482
1483		if (idx + slots - 1 >= DNODES_PER_BLOCK) {
1484			DNODE_STAT_BUMP(dnode_hold_free_overflow);
1485			dbuf_rele(db, FTAG);
1486			return (SET_ERROR(ENOSPC));
1487		}
1488
1489		while (dn == DN_SLOT_UNINIT) {
1490			dnode_slots_hold(dnc, idx, slots);
1491
1492			if (!dnode_check_slots_free(dnc, idx, slots)) {
1493				DNODE_STAT_BUMP(dnode_hold_free_misses);
1494				dnode_slots_rele(dnc, idx, slots);
1495				dbuf_rele(db, FTAG);
1496				return (SET_ERROR(ENOSPC));
1497			}
1498
1499			dnode_slots_rele(dnc, idx, slots);
1500			if (!dnode_slots_tryenter(dnc, idx, slots)) {
1501				DNODE_STAT_BUMP(dnode_hold_free_lock_retry);
1502				continue;
1503			}
1504
1505			if (!dnode_check_slots_free(dnc, idx, slots)) {
1506				DNODE_STAT_BUMP(dnode_hold_free_lock_misses);
1507				dnode_slots_rele(dnc, idx, slots);
1508				dbuf_rele(db, FTAG);
1509				return (SET_ERROR(ENOSPC));
1510			}
1511
1512			/*
1513			 * Allocated but otherwise free dnodes which would
1514			 * be in the interior of a multi-slot dnodes need
1515			 * to be freed.  Single slot dnodes can be safely
1516			 * re-purposed as a performance optimization.
1517			 */
1518			if (slots > 1)
1519				dnode_reclaim_slots(dnc, idx + 1, slots - 1);
1520
1521			dnh = &dnc->dnc_children[idx];
1522			if (DN_SLOT_IS_PTR(dnh->dnh_dnode)) {
1523				dn = dnh->dnh_dnode;
1524			} else {
1525				dn = dnode_create(os, dn_block + idx, db,
1526				    object, dnh);
1527			}
1528		}
1529
1530		mutex_enter(&dn->dn_mtx);
1531		if (!zfs_refcount_is_zero(&dn->dn_holds) || dn->dn_free_txg) {
1532			DNODE_STAT_BUMP(dnode_hold_free_refcount);
1533			mutex_exit(&dn->dn_mtx);
1534			dnode_slots_rele(dnc, idx, slots);
1535			dbuf_rele(db, FTAG);
1536			return (SET_ERROR(EEXIST));
1537		}
1538
1539		/* Don't actually hold if dry run, just return 0 */
1540		if (flag & DNODE_DRY_RUN) {
1541			mutex_exit(&dn->dn_mtx);
1542			dnode_slots_rele(dnc, idx, slots);
1543			dbuf_rele(db, FTAG);
1544			return (0);
1545		}
1546
1547		dnode_set_slots(dnc, idx + 1, slots - 1, DN_SLOT_INTERIOR);
1548		DNODE_STAT_BUMP(dnode_hold_free_hits);
1549	} else {
1550		dbuf_rele(db, FTAG);
1551		return (SET_ERROR(EINVAL));
1552	}
1553
1554	ASSERT0(dn->dn_free_txg);
1555
1556	if (zfs_refcount_add(&dn->dn_holds, tag) == 1)
1557		dbuf_add_ref(db, dnh);
1558
1559	mutex_exit(&dn->dn_mtx);
1560
1561	/* Now we can rely on the hold to prevent the dnode from moving. */
1562	dnode_slots_rele(dnc, idx, slots);
1563
1564	DNODE_VERIFY(dn);
1565	ASSERT3P(dn->dn_dbuf, ==, db);
1566	ASSERT3U(dn->dn_object, ==, object);
1567	dbuf_rele(db, FTAG);
1568
1569	*dnp = dn;
1570	return (0);
1571}
1572
1573/*
1574 * Return held dnode if the object is allocated, NULL if not.
1575 */
1576int
1577dnode_hold(objset_t *os, uint64_t object, void *tag, dnode_t **dnp)
1578{
1579	return (dnode_hold_impl(os, object, DNODE_MUST_BE_ALLOCATED, 0, tag,
1580	    dnp));
1581}
1582
1583/*
1584 * Can only add a reference if there is already at least one
1585 * reference on the dnode.  Returns FALSE if unable to add a
1586 * new reference.
1587 */
1588boolean_t
1589dnode_add_ref(dnode_t *dn, void *tag)
1590{
1591	mutex_enter(&dn->dn_mtx);
1592	if (zfs_refcount_is_zero(&dn->dn_holds)) {
1593		mutex_exit(&dn->dn_mtx);
1594		return (FALSE);
1595	}
1596	VERIFY(1 < zfs_refcount_add(&dn->dn_holds, tag));
1597	mutex_exit(&dn->dn_mtx);
1598	return (TRUE);
1599}
1600
1601void
1602dnode_rele(dnode_t *dn, void *tag)
1603{
1604	mutex_enter(&dn->dn_mtx);
1605	dnode_rele_and_unlock(dn, tag, B_FALSE);
1606}
1607
1608void
1609dnode_rele_and_unlock(dnode_t *dn, void *tag, boolean_t evicting)
1610{
1611	uint64_t refs;
1612	/* Get while the hold prevents the dnode from moving. */
1613	dmu_buf_impl_t *db = dn->dn_dbuf;
1614	dnode_handle_t *dnh = dn->dn_handle;
1615
1616	refs = zfs_refcount_remove(&dn->dn_holds, tag);
1617	mutex_exit(&dn->dn_mtx);
1618
1619	/*
1620	 * It's unsafe to release the last hold on a dnode by dnode_rele() or
1621	 * indirectly by dbuf_rele() while relying on the dnode handle to
1622	 * prevent the dnode from moving, since releasing the last hold could
1623	 * result in the dnode's parent dbuf evicting its dnode handles. For
1624	 * that reason anyone calling dnode_rele() or dbuf_rele() without some
1625	 * other direct or indirect hold on the dnode must first drop the dnode
1626	 * handle.
1627	 */
1628	ASSERT(refs > 0 || dnh->dnh_zrlock.zr_owner != curthread);
1629
1630	/* NOTE: the DNODE_DNODE does not have a dn_dbuf */
1631	if (refs == 0 && db != NULL) {
1632		/*
1633		 * Another thread could add a hold to the dnode handle in
1634		 * dnode_hold_impl() while holding the parent dbuf. Since the
1635		 * hold on the parent dbuf prevents the handle from being
1636		 * destroyed, the hold on the handle is OK. We can't yet assert
1637		 * that the handle has zero references, but that will be
1638		 * asserted anyway when the handle gets destroyed.
1639		 */
1640		mutex_enter(&db->db_mtx);
1641		dbuf_rele_and_unlock(db, dnh, evicting);
1642	}
1643}
1644
1645/*
1646 * Test whether we can create a dnode at the specified location.
1647 */
1648int
1649dnode_try_claim(objset_t *os, uint64_t object, int slots)
1650{
1651	return (dnode_hold_impl(os, object, DNODE_MUST_BE_FREE | DNODE_DRY_RUN,
1652	    slots, NULL, NULL));
1653}
1654
1655void
1656dnode_setdirty(dnode_t *dn, dmu_tx_t *tx)
1657{
1658	objset_t *os = dn->dn_objset;
1659	uint64_t txg = tx->tx_txg;
1660
1661	if (DMU_OBJECT_IS_SPECIAL(dn->dn_object)) {
1662		dsl_dataset_dirty(os->os_dsl_dataset, tx);
1663		return;
1664	}
1665
1666	DNODE_VERIFY(dn);
1667
1668#ifdef ZFS_DEBUG
1669	mutex_enter(&dn->dn_mtx);
1670	ASSERT(dn->dn_phys->dn_type || dn->dn_allocated_txg);
1671	ASSERT(dn->dn_free_txg == 0 || dn->dn_free_txg >= txg);
1672	mutex_exit(&dn->dn_mtx);
1673#endif
1674
1675	/*
1676	 * Determine old uid/gid when necessary
1677	 */
1678	dmu_objset_userquota_get_ids(dn, B_TRUE, tx);
1679
1680	multilist_t *dirtylist = os->os_dirty_dnodes[txg & TXG_MASK];
1681	multilist_sublist_t *mls = multilist_sublist_lock_obj(dirtylist, dn);
1682
1683	/*
1684	 * If we are already marked dirty, we're done.
1685	 */
1686	if (multilist_link_active(&dn->dn_dirty_link[txg & TXG_MASK])) {
1687		multilist_sublist_unlock(mls);
1688		return;
1689	}
1690
1691	ASSERT(!zfs_refcount_is_zero(&dn->dn_holds) ||
1692	    !avl_is_empty(&dn->dn_dbufs));
1693	ASSERT(dn->dn_datablksz != 0);
1694	ASSERT0(dn->dn_next_bonuslen[txg&TXG_MASK]);
1695	ASSERT0(dn->dn_next_blksz[txg&TXG_MASK]);
1696	ASSERT0(dn->dn_next_bonustype[txg&TXG_MASK]);
1697
1698	dprintf_ds(os->os_dsl_dataset, "obj=%llu txg=%llu\n",
1699	    dn->dn_object, txg);
1700
1701	multilist_sublist_insert_head(mls, dn);
1702
1703	multilist_sublist_unlock(mls);
1704
1705	/*
1706	 * The dnode maintains a hold on its containing dbuf as
1707	 * long as there are holds on it.  Each instantiated child
1708	 * dbuf maintains a hold on the dnode.  When the last child
1709	 * drops its hold, the dnode will drop its hold on the
1710	 * containing dbuf. We add a "dirty hold" here so that the
1711	 * dnode will hang around after we finish processing its
1712	 * children.
1713	 */
1714	VERIFY(dnode_add_ref(dn, (void *)(uintptr_t)tx->tx_txg));
1715
1716	(void) dbuf_dirty(dn->dn_dbuf, tx);
1717
1718	dsl_dataset_dirty(os->os_dsl_dataset, tx);
1719}
1720
1721void
1722dnode_free(dnode_t *dn, dmu_tx_t *tx)
1723{
1724	mutex_enter(&dn->dn_mtx);
1725	if (dn->dn_type == DMU_OT_NONE || dn->dn_free_txg) {
1726		mutex_exit(&dn->dn_mtx);
1727		return;
1728	}
1729	dn->dn_free_txg = tx->tx_txg;
1730	mutex_exit(&dn->dn_mtx);
1731
1732	dnode_setdirty(dn, tx);
1733}
1734
1735/*
1736 * Try to change the block size for the indicated dnode.  This can only
1737 * succeed if there are no blocks allocated or dirty beyond first block
1738 */
1739int
1740dnode_set_blksz(dnode_t *dn, uint64_t size, int ibs, dmu_tx_t *tx)
1741{
1742	dmu_buf_impl_t *db;
1743	int err;
1744
1745	ASSERT3U(size, <=, spa_maxblocksize(dmu_objset_spa(dn->dn_objset)));
1746	if (size == 0)
1747		size = SPA_MINBLOCKSIZE;
1748	else
1749		size = P2ROUNDUP(size, SPA_MINBLOCKSIZE);
1750
1751	if (ibs == dn->dn_indblkshift)
1752		ibs = 0;
1753
1754	if (size >> SPA_MINBLOCKSHIFT == dn->dn_datablkszsec && ibs == 0)
1755		return (0);
1756
1757	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1758
1759	/* Check for any allocated blocks beyond the first */
1760	if (dn->dn_maxblkid != 0)
1761		goto fail;
1762
1763	mutex_enter(&dn->dn_dbufs_mtx);
1764	for (db = avl_first(&dn->dn_dbufs); db != NULL;
1765	    db = AVL_NEXT(&dn->dn_dbufs, db)) {
1766		if (db->db_blkid != 0 && db->db_blkid != DMU_BONUS_BLKID &&
1767		    db->db_blkid != DMU_SPILL_BLKID) {
1768			mutex_exit(&dn->dn_dbufs_mtx);
1769			goto fail;
1770		}
1771	}
1772	mutex_exit(&dn->dn_dbufs_mtx);
1773
1774	if (ibs && dn->dn_nlevels != 1)
1775		goto fail;
1776
1777	/* resize the old block */
1778	err = dbuf_hold_impl(dn, 0, 0, TRUE, FALSE, FTAG, &db);
1779	if (err == 0)
1780		dbuf_new_size(db, size, tx);
1781	else if (err != ENOENT)
1782		goto fail;
1783
1784	dnode_setdblksz(dn, size);
1785	dnode_setdirty(dn, tx);
1786	dn->dn_next_blksz[tx->tx_txg&TXG_MASK] = size;
1787	if (ibs) {
1788		dn->dn_indblkshift = ibs;
1789		dn->dn_next_indblkshift[tx->tx_txg&TXG_MASK] = ibs;
1790	}
1791	/* rele after we have fixed the blocksize in the dnode */
1792	if (db)
1793		dbuf_rele(db, FTAG);
1794
1795	rw_exit(&dn->dn_struct_rwlock);
1796	return (0);
1797
1798fail:
1799	rw_exit(&dn->dn_struct_rwlock);
1800	return (SET_ERROR(ENOTSUP));
1801}
1802
1803static void
1804dnode_set_nlevels_impl(dnode_t *dn, int new_nlevels, dmu_tx_t *tx)
1805{
1806	uint64_t txgoff = tx->tx_txg & TXG_MASK;
1807	int old_nlevels = dn->dn_nlevels;
1808	dmu_buf_impl_t *db;
1809	list_t *list;
1810	dbuf_dirty_record_t *new, *dr, *dr_next;
1811
1812	ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
1813
1814	dn->dn_nlevels = new_nlevels;
1815
1816	ASSERT3U(new_nlevels, >, dn->dn_next_nlevels[txgoff]);
1817	dn->dn_next_nlevels[txgoff] = new_nlevels;
1818
1819	/* dirty the left indirects */
1820	db = dbuf_hold_level(dn, old_nlevels, 0, FTAG);
1821	ASSERT(db != NULL);
1822	new = dbuf_dirty(db, tx);
1823	dbuf_rele(db, FTAG);
1824
1825	/* transfer the dirty records to the new indirect */
1826	mutex_enter(&dn->dn_mtx);
1827	mutex_enter(&new->dt.di.dr_mtx);
1828	list = &dn->dn_dirty_records[txgoff];
1829	for (dr = list_head(list); dr; dr = dr_next) {
1830		dr_next = list_next(&dn->dn_dirty_records[txgoff], dr);
1831		if (dr->dr_dbuf->db_level != new_nlevels-1 &&
1832		    dr->dr_dbuf->db_blkid != DMU_BONUS_BLKID &&
1833		    dr->dr_dbuf->db_blkid != DMU_SPILL_BLKID) {
1834			ASSERT(dr->dr_dbuf->db_level == old_nlevels-1);
1835			list_remove(&dn->dn_dirty_records[txgoff], dr);
1836			list_insert_tail(&new->dt.di.dr_children, dr);
1837			dr->dr_parent = new;
1838		}
1839	}
1840	mutex_exit(&new->dt.di.dr_mtx);
1841	mutex_exit(&dn->dn_mtx);
1842}
1843
1844int
1845dnode_set_nlevels(dnode_t *dn, int nlevels, dmu_tx_t *tx)
1846{
1847	int ret = 0;
1848
1849	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1850
1851	if (dn->dn_nlevels == nlevels) {
1852		ret = 0;
1853		goto out;
1854	} else if (nlevels < dn->dn_nlevels) {
1855		ret = SET_ERROR(EINVAL);
1856		goto out;
1857	}
1858
1859	dnode_set_nlevels_impl(dn, nlevels, tx);
1860
1861out:
1862	rw_exit(&dn->dn_struct_rwlock);
1863	return (ret);
1864}
1865
1866/* read-holding callers must not rely on the lock being continuously held */
1867void
1868dnode_new_blkid(dnode_t *dn, uint64_t blkid, dmu_tx_t *tx, boolean_t have_read,
1869    boolean_t force)
1870{
1871	int epbs, new_nlevels;
1872	uint64_t sz;
1873
1874	ASSERT(blkid != DMU_BONUS_BLKID);
1875
1876	ASSERT(have_read ?
1877	    RW_READ_HELD(&dn->dn_struct_rwlock) :
1878	    RW_WRITE_HELD(&dn->dn_struct_rwlock));
1879
1880	/*
1881	 * if we have a read-lock, check to see if we need to do any work
1882	 * before upgrading to a write-lock.
1883	 */
1884	if (have_read) {
1885		if (blkid <= dn->dn_maxblkid)
1886			return;
1887
1888		if (!rw_tryupgrade(&dn->dn_struct_rwlock)) {
1889			rw_exit(&dn->dn_struct_rwlock);
1890			rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1891		}
1892	}
1893
1894	/*
1895	 * Raw sends (indicated by the force flag) require that we take the
1896	 * given blkid even if the value is lower than the current value.
1897	 */
1898	if (!force && blkid <= dn->dn_maxblkid)
1899		goto out;
1900
1901	/*
1902	 * We use the (otherwise unused) top bit of dn_next_maxblkid[txgoff]
1903	 * to indicate that this field is set. This allows us to set the
1904	 * maxblkid to 0 on an existing object in dnode_sync().
1905	 */
1906	dn->dn_maxblkid = blkid;
1907	dn->dn_next_maxblkid[tx->tx_txg & TXG_MASK] =
1908	    blkid | DMU_NEXT_MAXBLKID_SET;
1909
1910	/*
1911	 * Compute the number of levels necessary to support the new maxblkid.
1912	 * Raw sends will ensure nlevels is set correctly for us.
1913	 */
1914	new_nlevels = 1;
1915	epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1916	for (sz = dn->dn_nblkptr;
1917	    sz <= blkid && sz >= dn->dn_nblkptr; sz <<= epbs)
1918		new_nlevels++;
1919
1920	if (!force) {
1921		if (new_nlevels > dn->dn_nlevels)
1922			dnode_set_nlevels_impl(dn, new_nlevels, tx);
1923	} else {
1924		ASSERT3U(dn->dn_nlevels, >=, new_nlevels);
1925	}
1926
1927out:
1928	if (have_read)
1929		rw_downgrade(&dn->dn_struct_rwlock);
1930}
1931
1932static void
1933dnode_dirty_l1(dnode_t *dn, uint64_t l1blkid, dmu_tx_t *tx)
1934{
1935	dmu_buf_impl_t *db = dbuf_hold_level(dn, 1, l1blkid, FTAG);
1936	if (db != NULL) {
1937		dmu_buf_will_dirty(&db->db, tx);
1938		dbuf_rele(db, FTAG);
1939	}
1940}
1941
1942/*
1943 * Dirty all the in-core level-1 dbufs in the range specified by start_blkid
1944 * and end_blkid.
1945 */
1946static void
1947dnode_dirty_l1range(dnode_t *dn, uint64_t start_blkid, uint64_t end_blkid,
1948    dmu_tx_t *tx)
1949{
1950	dmu_buf_impl_t db_search;
1951	dmu_buf_impl_t *db;
1952	avl_index_t where;
1953
1954	mutex_enter(&dn->dn_dbufs_mtx);
1955
1956	db_search.db_level = 1;
1957	db_search.db_blkid = start_blkid + 1;
1958	db_search.db_state = DB_SEARCH;
1959	for (;;) {
1960
1961		db = avl_find(&dn->dn_dbufs, &db_search, &where);
1962		if (db == NULL)
1963			db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER);
1964
1965		if (db == NULL || db->db_level != 1 ||
1966		    db->db_blkid >= end_blkid) {
1967			break;
1968		}
1969
1970		/*
1971		 * Setup the next blkid we want to search for.
1972		 */
1973		db_search.db_blkid = db->db_blkid + 1;
1974		ASSERT3U(db->db_blkid, >=, start_blkid);
1975
1976		/*
1977		 * If the dbuf transitions to DB_EVICTING while we're trying
1978		 * to dirty it, then we will be unable to discover it in
1979		 * the dbuf hash table. This will result in a call to
1980		 * dbuf_create() which needs to acquire the dn_dbufs_mtx
1981		 * lock. To avoid a deadlock, we drop the lock before
1982		 * dirtying the level-1 dbuf.
1983		 */
1984		mutex_exit(&dn->dn_dbufs_mtx);
1985		dnode_dirty_l1(dn, db->db_blkid, tx);
1986		mutex_enter(&dn->dn_dbufs_mtx);
1987	}
1988
1989#ifdef ZFS_DEBUG
1990	/*
1991	 * Walk all the in-core level-1 dbufs and verify they have been dirtied.
1992	 */
1993	db_search.db_level = 1;
1994	db_search.db_blkid = start_blkid + 1;
1995	db_search.db_state = DB_SEARCH;
1996	db = avl_find(&dn->dn_dbufs, &db_search, &where);
1997	if (db == NULL)
1998		db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER);
1999	for (; db != NULL; db = AVL_NEXT(&dn->dn_dbufs, db)) {
2000		if (db->db_level != 1 || db->db_blkid >= end_blkid)
2001			break;
2002		ASSERT(db->db_dirtycnt > 0);
2003	}
2004#endif
2005	mutex_exit(&dn->dn_dbufs_mtx);
2006}
2007
2008void
2009dnode_free_range(dnode_t *dn, uint64_t off, uint64_t len, dmu_tx_t *tx)
2010{
2011	dmu_buf_impl_t *db;
2012	uint64_t blkoff, blkid, nblks;
2013	int blksz, blkshift, head, tail;
2014	int trunc = FALSE;
2015	int epbs;
2016
2017	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
2018	blksz = dn->dn_datablksz;
2019	blkshift = dn->dn_datablkshift;
2020	epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
2021
2022	if (len == DMU_OBJECT_END) {
2023		len = UINT64_MAX - off;
2024		trunc = TRUE;
2025	}
2026
2027	/*
2028	 * First, block align the region to free:
2029	 */
2030	if (ISP2(blksz)) {
2031		head = P2NPHASE(off, blksz);
2032		blkoff = P2PHASE(off, blksz);
2033		if ((off >> blkshift) > dn->dn_maxblkid)
2034			goto out;
2035	} else {
2036		ASSERT(dn->dn_maxblkid == 0);
2037		if (off == 0 && len >= blksz) {
2038			/*
2039			 * Freeing the whole block; fast-track this request.
2040			 */
2041			blkid = 0;
2042			nblks = 1;
2043			if (dn->dn_nlevels > 1)
2044				dnode_dirty_l1(dn, 0, tx);
2045			goto done;
2046		} else if (off >= blksz) {
2047			/* Freeing past end-of-data */
2048			goto out;
2049		} else {
2050			/* Freeing part of the block. */
2051			head = blksz - off;
2052			ASSERT3U(head, >, 0);
2053		}
2054		blkoff = off;
2055	}
2056	/* zero out any partial block data at the start of the range */
2057	if (head) {
2058		ASSERT3U(blkoff + head, ==, blksz);
2059		if (len < head)
2060			head = len;
2061		if (dbuf_hold_impl(dn, 0, dbuf_whichblock(dn, 0, off),
2062		    TRUE, FALSE, FTAG, &db) == 0) {
2063			caddr_t data;
2064
2065			/* don't dirty if it isn't on disk and isn't dirty */
2066			if (db->db_last_dirty ||
2067			    (db->db_blkptr && !BP_IS_HOLE(db->db_blkptr))) {
2068				rw_exit(&dn->dn_struct_rwlock);
2069				dmu_buf_will_dirty(&db->db, tx);
2070				rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
2071				data = db->db.db_data;
2072				bzero(data + blkoff, head);
2073			}
2074			dbuf_rele(db, FTAG);
2075		}
2076		off += head;
2077		len -= head;
2078	}
2079
2080	/* If the range was less than one block, we're done */
2081	if (len == 0)
2082		goto out;
2083
2084	/* If the remaining range is past end of file, we're done */
2085	if ((off >> blkshift) > dn->dn_maxblkid)
2086		goto out;
2087
2088	ASSERT(ISP2(blksz));
2089	if (trunc)
2090		tail = 0;
2091	else
2092		tail = P2PHASE(len, blksz);
2093
2094	ASSERT0(P2PHASE(off, blksz));
2095	/* zero out any partial block data at the end of the range */
2096	if (tail) {
2097		if (len < tail)
2098			tail = len;
2099		if (dbuf_hold_impl(dn, 0, dbuf_whichblock(dn, 0, off+len),
2100		    TRUE, FALSE, FTAG, &db) == 0) {
2101			/* don't dirty if not on disk and not dirty */
2102			if (db->db_last_dirty ||
2103			    (db->db_blkptr && !BP_IS_HOLE(db->db_blkptr))) {
2104				rw_exit(&dn->dn_struct_rwlock);
2105				dmu_buf_will_dirty(&db->db, tx);
2106				rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
2107				bzero(db->db.db_data, tail);
2108			}
2109			dbuf_rele(db, FTAG);
2110		}
2111		len -= tail;
2112	}
2113
2114	/* If the range did not include a full block, we are done */
2115	if (len == 0)
2116		goto out;
2117
2118	ASSERT(IS_P2ALIGNED(off, blksz));
2119	ASSERT(trunc || IS_P2ALIGNED(len, blksz));
2120	blkid = off >> blkshift;
2121	nblks = len >> blkshift;
2122	if (trunc)
2123		nblks += 1;
2124
2125	/*
2126	 * Dirty all the indirect blocks in this range.  Note that only
2127	 * the first and last indirect blocks can actually be written
2128	 * (if they were partially freed) -- they must be dirtied, even if
2129	 * they do not exist on disk yet.  The interior blocks will
2130	 * be freed by free_children(), so they will not actually be written.
2131	 * Even though these interior blocks will not be written, we
2132	 * dirty them for two reasons:
2133	 *
2134	 *  - It ensures that the indirect blocks remain in memory until
2135	 *    syncing context.  (They have already been prefetched by
2136	 *    dmu_tx_hold_free(), so we don't have to worry about reading
2137	 *    them serially here.)
2138	 *
2139	 *  - The dirty space accounting will put pressure on the txg sync
2140	 *    mechanism to begin syncing, and to delay transactions if there
2141	 *    is a large amount of freeing.  Even though these indirect
2142	 *    blocks will not be written, we could need to write the same
2143	 *    amount of space if we copy the freed BPs into deadlists.
2144	 */
2145	if (dn->dn_nlevels > 1) {
2146		uint64_t first, last;
2147
2148		first = blkid >> epbs;
2149		dnode_dirty_l1(dn, first, tx);
2150		if (trunc)
2151			last = dn->dn_maxblkid >> epbs;
2152		else
2153			last = (blkid + nblks - 1) >> epbs;
2154		if (last != first)
2155			dnode_dirty_l1(dn, last, tx);
2156
2157		dnode_dirty_l1range(dn, first, last, tx);
2158
2159		int shift = dn->dn_datablkshift + dn->dn_indblkshift -
2160		    SPA_BLKPTRSHIFT;
2161		for (uint64_t i = first + 1; i < last; i++) {
2162			/*
2163			 * Set i to the blockid of the next non-hole
2164			 * level-1 indirect block at or after i.  Note
2165			 * that dnode_next_offset() operates in terms of
2166			 * level-0-equivalent bytes.
2167			 */
2168			uint64_t ibyte = i << shift;
2169			int err = dnode_next_offset(dn, DNODE_FIND_HAVELOCK,
2170			    &ibyte, 2, 1, 0);
2171			i = ibyte >> shift;
2172			if (i >= last)
2173				break;
2174
2175			/*
2176			 * Normally we should not see an error, either
2177			 * from dnode_next_offset() or dbuf_hold_level()
2178			 * (except for ESRCH from dnode_next_offset).
2179			 * If there is an i/o error, then when we read
2180			 * this block in syncing context, it will use
2181			 * ZIO_FLAG_MUSTSUCCEED, and thus hang/panic according
2182			 * to the "failmode" property.  dnode_next_offset()
2183			 * doesn't have a flag to indicate MUSTSUCCEED.
2184			 */
2185			if (err != 0)
2186				break;
2187
2188			dnode_dirty_l1(dn, i, tx);
2189		}
2190	}
2191
2192done:
2193	/*
2194	 * Add this range to the dnode range list.
2195	 * We will finish up this free operation in the syncing phase.
2196	 */
2197	mutex_enter(&dn->dn_mtx);
2198	int txgoff = tx->tx_txg & TXG_MASK;
2199	if (dn->dn_free_ranges[txgoff] == NULL) {
2200		dn->dn_free_ranges[txgoff] = range_tree_create(NULL, NULL);
2201	}
2202	range_tree_clear(dn->dn_free_ranges[txgoff], blkid, nblks);
2203	range_tree_add(dn->dn_free_ranges[txgoff], blkid, nblks);
2204	dprintf_dnode(dn, "blkid=%llu nblks=%llu txg=%llu\n",
2205	    blkid, nblks, tx->tx_txg);
2206	mutex_exit(&dn->dn_mtx);
2207
2208	dbuf_free_range(dn, blkid, blkid + nblks - 1, tx);
2209	dnode_setdirty(dn, tx);
2210out:
2211
2212	rw_exit(&dn->dn_struct_rwlock);
2213}
2214
2215static boolean_t
2216dnode_spill_freed(dnode_t *dn)
2217{
2218	int i;
2219
2220	mutex_enter(&dn->dn_mtx);
2221	for (i = 0; i < TXG_SIZE; i++) {
2222		if (dn->dn_rm_spillblk[i] == DN_KILL_SPILLBLK)
2223			break;
2224	}
2225	mutex_exit(&dn->dn_mtx);
2226	return (i < TXG_SIZE);
2227}
2228
2229/* return TRUE if this blkid was freed in a recent txg, or FALSE if it wasn't */
2230uint64_t
2231dnode_block_freed(dnode_t *dn, uint64_t blkid)
2232{
2233	void *dp = spa_get_dsl(dn->dn_objset->os_spa);
2234	int i;
2235
2236	if (blkid == DMU_BONUS_BLKID)
2237		return (FALSE);
2238
2239	/*
2240	 * If we're in the process of opening the pool, dp will not be
2241	 * set yet, but there shouldn't be anything dirty.
2242	 */
2243	if (dp == NULL)
2244		return (FALSE);
2245
2246	if (dn->dn_free_txg)
2247		return (TRUE);
2248
2249	if (blkid == DMU_SPILL_BLKID)
2250		return (dnode_spill_freed(dn));
2251
2252	mutex_enter(&dn->dn_mtx);
2253	for (i = 0; i < TXG_SIZE; i++) {
2254		if (dn->dn_free_ranges[i] != NULL &&
2255		    range_tree_contains(dn->dn_free_ranges[i], blkid, 1))
2256			break;
2257	}
2258	mutex_exit(&dn->dn_mtx);
2259	return (i < TXG_SIZE);
2260}
2261
2262/* call from syncing context when we actually write/free space for this dnode */
2263void
2264dnode_diduse_space(dnode_t *dn, int64_t delta)
2265{
2266	uint64_t space;
2267	dprintf_dnode(dn, "dn=%p dnp=%p used=%llu delta=%lld\n",
2268	    dn, dn->dn_phys,
2269	    (u_longlong_t)dn->dn_phys->dn_used,
2270	    (longlong_t)delta);
2271
2272	mutex_enter(&dn->dn_mtx);
2273	space = DN_USED_BYTES(dn->dn_phys);
2274	if (delta > 0) {
2275		ASSERT3U(space + delta, >=, space); /* no overflow */
2276	} else {
2277		ASSERT3U(space, >=, -delta); /* no underflow */
2278	}
2279	space += delta;
2280	if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_DNODE_BYTES) {
2281		ASSERT((dn->dn_phys->dn_flags & DNODE_FLAG_USED_BYTES) == 0);
2282		ASSERT0(P2PHASE(space, 1<<DEV_BSHIFT));
2283		dn->dn_phys->dn_used = space >> DEV_BSHIFT;
2284	} else {
2285		dn->dn_phys->dn_used = space;
2286		dn->dn_phys->dn_flags |= DNODE_FLAG_USED_BYTES;
2287	}
2288	mutex_exit(&dn->dn_mtx);
2289}
2290
2291/*
2292 * Scans a block at the indicated "level" looking for a hole or data,
2293 * depending on 'flags'.
2294 *
2295 * If level > 0, then we are scanning an indirect block looking at its
2296 * pointers.  If level == 0, then we are looking at a block of dnodes.
2297 *
2298 * If we don't find what we are looking for in the block, we return ESRCH.
2299 * Otherwise, return with *offset pointing to the beginning (if searching
2300 * forwards) or end (if searching backwards) of the range covered by the
2301 * block pointer we matched on (or dnode).
2302 *
2303 * The basic search algorithm used below by dnode_next_offset() is to
2304 * use this function to search up the block tree (widen the search) until
2305 * we find something (i.e., we don't return ESRCH) and then search back
2306 * down the tree (narrow the search) until we reach our original search
2307 * level.
2308 */
2309static int
2310dnode_next_offset_level(dnode_t *dn, int flags, uint64_t *offset,
2311    int lvl, uint64_t blkfill, uint64_t txg)
2312{
2313	dmu_buf_impl_t *db = NULL;
2314	void *data = NULL;
2315	uint64_t epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
2316	uint64_t epb = 1ULL << epbs;
2317	uint64_t minfill, maxfill;
2318	boolean_t hole;
2319	int i, inc, error, span;
2320
2321	dprintf("probing object %llu offset %llx level %d of %u\n",
2322	    dn->dn_object, *offset, lvl, dn->dn_phys->dn_nlevels);
2323
2324	hole = ((flags & DNODE_FIND_HOLE) != 0);
2325	inc = (flags & DNODE_FIND_BACKWARDS) ? -1 : 1;
2326	ASSERT(txg == 0 || !hole);
2327
2328	if (lvl == dn->dn_phys->dn_nlevels) {
2329		error = 0;
2330		epb = dn->dn_phys->dn_nblkptr;
2331		data = dn->dn_phys->dn_blkptr;
2332	} else {
2333		uint64_t blkid = dbuf_whichblock(dn, lvl, *offset);
2334		error = dbuf_hold_impl(dn, lvl, blkid, TRUE, FALSE, FTAG, &db);
2335		if (error) {
2336			if (error != ENOENT)
2337				return (error);
2338			if (hole)
2339				return (0);
2340			/*
2341			 * This can only happen when we are searching up
2342			 * the block tree for data.  We don't really need to
2343			 * adjust the offset, as we will just end up looking
2344			 * at the pointer to this block in its parent, and its
2345			 * going to be unallocated, so we will skip over it.
2346			 */
2347			return (SET_ERROR(ESRCH));
2348		}
2349		error = dbuf_read(db, NULL,
2350		    DB_RF_CANFAIL | DB_RF_HAVESTRUCT | DB_RF_NO_DECRYPT);
2351		if (error) {
2352			dbuf_rele(db, FTAG);
2353			return (error);
2354		}
2355		data = db->db.db_data;
2356	}
2357
2358
2359	if (db != NULL && txg != 0 && (db->db_blkptr == NULL ||
2360	    db->db_blkptr->blk_birth <= txg ||
2361	    BP_IS_HOLE(db->db_blkptr))) {
2362		/*
2363		 * This can only happen when we are searching up the tree
2364		 * and these conditions mean that we need to keep climbing.
2365		 */
2366		error = SET_ERROR(ESRCH);
2367	} else if (lvl == 0) {
2368		dnode_phys_t *dnp = data;
2369
2370		ASSERT(dn->dn_type == DMU_OT_DNODE);
2371		ASSERT(!(flags & DNODE_FIND_BACKWARDS));
2372
2373		for (i = (*offset >> DNODE_SHIFT) & (blkfill - 1);
2374		    i < blkfill; i += dnp[i].dn_extra_slots + 1) {
2375			if ((dnp[i].dn_type == DMU_OT_NONE) == hole)
2376				break;
2377		}
2378
2379		if (i == blkfill)
2380			error = SET_ERROR(ESRCH);
2381
2382		*offset = (*offset & ~(DNODE_BLOCK_SIZE - 1)) +
2383		    (i << DNODE_SHIFT);
2384	} else {
2385		blkptr_t *bp = data;
2386		uint64_t start = *offset;
2387		span = (lvl - 1) * epbs + dn->dn_datablkshift;
2388		minfill = 0;
2389		maxfill = blkfill << ((lvl - 1) * epbs);
2390
2391		if (hole)
2392			maxfill--;
2393		else
2394			minfill++;
2395
2396		*offset = *offset >> span;
2397		for (i = BF64_GET(*offset, 0, epbs);
2398		    i >= 0 && i < epb; i += inc) {
2399			if (BP_GET_FILL(&bp[i]) >= minfill &&
2400			    BP_GET_FILL(&bp[i]) <= maxfill &&
2401			    (hole || bp[i].blk_birth > txg))
2402				break;
2403			if (inc > 0 || *offset > 0)
2404				*offset += inc;
2405		}
2406		*offset = *offset << span;
2407		if (inc < 0) {
2408			/* traversing backwards; position offset at the end */
2409			ASSERT3U(*offset, <=, start);
2410			*offset = MIN(*offset + (1ULL << span) - 1, start);
2411		} else if (*offset < start) {
2412			*offset = start;
2413		}
2414		if (i < 0 || i >= epb)
2415			error = SET_ERROR(ESRCH);
2416	}
2417
2418	if (db)
2419		dbuf_rele(db, FTAG);
2420
2421	return (error);
2422}
2423
2424/*
2425 * Find the next hole, data, or sparse region at or after *offset.
2426 * The value 'blkfill' tells us how many items we expect to find
2427 * in an L0 data block; this value is 1 for normal objects,
2428 * DNODES_PER_BLOCK for the meta dnode, and some fraction of
2429 * DNODES_PER_BLOCK when searching for sparse regions thereof.
2430 *
2431 * Examples:
2432 *
2433 * dnode_next_offset(dn, flags, offset, 1, 1, 0);
2434 *	Finds the next/previous hole/data in a file.
2435 *	Used in dmu_offset_next().
2436 *
2437 * dnode_next_offset(mdn, flags, offset, 0, DNODES_PER_BLOCK, txg);
2438 *	Finds the next free/allocated dnode an objset's meta-dnode.
2439 *	Only finds objects that have new contents since txg (ie.
2440 *	bonus buffer changes and content removal are ignored).
2441 *	Used in dmu_object_next().
2442 *
2443 * dnode_next_offset(mdn, DNODE_FIND_HOLE, offset, 2, DNODES_PER_BLOCK >> 2, 0);
2444 *	Finds the next L2 meta-dnode bp that's at most 1/4 full.
2445 *	Used in dmu_object_alloc().
2446 */
2447int
2448dnode_next_offset(dnode_t *dn, int flags, uint64_t *offset,
2449    int minlvl, uint64_t blkfill, uint64_t txg)
2450{
2451	uint64_t initial_offset = *offset;
2452	int lvl, maxlvl;
2453	int error = 0;
2454
2455	if (!(flags & DNODE_FIND_HAVELOCK))
2456		rw_enter(&dn->dn_struct_rwlock, RW_READER);
2457
2458	if (dn->dn_phys->dn_nlevels == 0) {
2459		error = SET_ERROR(ESRCH);
2460		goto out;
2461	}
2462
2463	if (dn->dn_datablkshift == 0) {
2464		if (*offset < dn->dn_datablksz) {
2465			if (flags & DNODE_FIND_HOLE)
2466				*offset = dn->dn_datablksz;
2467		} else {
2468			error = SET_ERROR(ESRCH);
2469		}
2470		goto out;
2471	}
2472
2473	maxlvl = dn->dn_phys->dn_nlevels;
2474
2475	for (lvl = minlvl; lvl <= maxlvl; lvl++) {
2476		error = dnode_next_offset_level(dn,
2477		    flags, offset, lvl, blkfill, txg);
2478		if (error != ESRCH)
2479			break;
2480	}
2481
2482	while (error == 0 && --lvl >= minlvl) {
2483		error = dnode_next_offset_level(dn,
2484		    flags, offset, lvl, blkfill, txg);
2485	}
2486
2487	/*
2488	 * There's always a "virtual hole" at the end of the object, even
2489	 * if all BP's which physically exist are non-holes.
2490	 */
2491	if ((flags & DNODE_FIND_HOLE) && error == ESRCH && txg == 0 &&
2492	    minlvl == 1 && blkfill == 1 && !(flags & DNODE_FIND_BACKWARDS)) {
2493		error = 0;
2494	}
2495
2496	if (error == 0 && (flags & DNODE_FIND_BACKWARDS ?
2497	    initial_offset < *offset : initial_offset > *offset))
2498		error = SET_ERROR(ESRCH);
2499out:
2500	if (!(flags & DNODE_FIND_HAVELOCK))
2501		rw_exit(&dn->dn_struct_rwlock);
2502
2503	return (error);
2504}
2505