xref: /trueos/sys/cddl/contrib/opensolaris/uts/common/fs/zfs/dnode.c (revision b878ae0412195c252dc69da54b764deab388e210)
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, 2014 by Delphix. All rights reserved.
24  */
25 
26 #include <sys/zfs_context.h>
27 #include <sys/dbuf.h>
28 #include <sys/dnode.h>
29 #include <sys/dmu.h>
30 #include <sys/dmu_impl.h>
31 #include <sys/dmu_tx.h>
32 #include <sys/dmu_objset.h>
33 #include <sys/dsl_dir.h>
34 #include <sys/dsl_dataset.h>
35 #include <sys/spa.h>
36 #include <sys/zio.h>
37 #include <sys/dmu_zfetch.h>
38 #include <sys/range_tree.h>
39 
40 static kmem_cache_t *dnode_cache;
41 /*
42  * Define DNODE_STATS to turn on statistic gathering. By default, it is only
43  * turned on when DEBUG is also defined.
44  */
45 #ifdef	DEBUG
46 #define	DNODE_STATS
47 #endif	/* DEBUG */
48 
49 #ifdef	DNODE_STATS
50 #define	DNODE_STAT_ADD(stat)			((stat)++)
51 #else
52 #define	DNODE_STAT_ADD(stat)			/* nothing */
53 #endif	/* DNODE_STATS */
54 
55 static dnode_phys_t dnode_phys_zero;
56 
57 int zfs_default_bs = SPA_MINBLOCKSHIFT;
58 int zfs_default_ibs = DN_MAX_INDBLKSHIFT;
59 
60 #ifdef sun
61 static kmem_cbrc_t dnode_move(void *, void *, size_t, void *);
62 #endif
63 
64 static int
dbuf_compare(const void * x1,const void * x2)65 dbuf_compare(const void *x1, const void *x2)
66 {
67 	const dmu_buf_impl_t *d1 = x1;
68 	const dmu_buf_impl_t *d2 = x2;
69 
70 	if (d1->db_level < d2->db_level) {
71 		return (-1);
72 	}
73 	if (d1->db_level > d2->db_level) {
74 		return (1);
75 	}
76 
77 	if (d1->db_blkid < d2->db_blkid) {
78 		return (-1);
79 	}
80 	if (d1->db_blkid > d2->db_blkid) {
81 		return (1);
82 	}
83 
84 	if (d1->db_state == DB_SEARCH) {
85 		ASSERT3S(d2->db_state, !=, DB_SEARCH);
86 		return (-1);
87 	} else if (d2->db_state == DB_SEARCH) {
88 		ASSERT3S(d1->db_state, !=, DB_SEARCH);
89 		return (1);
90 	}
91 
92 	if ((uintptr_t)d1 < (uintptr_t)d2) {
93 		return (-1);
94 	}
95 	if ((uintptr_t)d1 > (uintptr_t)d2) {
96 		return (1);
97 	}
98 	return (0);
99 }
100 
101 /* ARGSUSED */
102 static int
dnode_cons(void * arg,void * unused,int kmflag)103 dnode_cons(void *arg, void *unused, int kmflag)
104 {
105 	dnode_t *dn = arg;
106 	int i;
107 
108 	rw_init(&dn->dn_struct_rwlock, NULL, RW_DEFAULT, NULL);
109 	mutex_init(&dn->dn_mtx, NULL, MUTEX_DEFAULT, NULL);
110 	mutex_init(&dn->dn_dbufs_mtx, NULL, MUTEX_DEFAULT, NULL);
111 	cv_init(&dn->dn_notxholds, NULL, CV_DEFAULT, NULL);
112 
113 	/*
114 	 * Every dbuf has a reference, and dropping a tracked reference is
115 	 * O(number of references), so don't track dn_holds.
116 	 */
117 	refcount_create_untracked(&dn->dn_holds);
118 	refcount_create(&dn->dn_tx_holds);
119 	list_link_init(&dn->dn_link);
120 
121 	bzero(&dn->dn_next_nblkptr[0], sizeof (dn->dn_next_nblkptr));
122 	bzero(&dn->dn_next_nlevels[0], sizeof (dn->dn_next_nlevels));
123 	bzero(&dn->dn_next_indblkshift[0], sizeof (dn->dn_next_indblkshift));
124 	bzero(&dn->dn_next_bonustype[0], sizeof (dn->dn_next_bonustype));
125 	bzero(&dn->dn_rm_spillblk[0], sizeof (dn->dn_rm_spillblk));
126 	bzero(&dn->dn_next_bonuslen[0], sizeof (dn->dn_next_bonuslen));
127 	bzero(&dn->dn_next_blksz[0], sizeof (dn->dn_next_blksz));
128 
129 	for (i = 0; i < TXG_SIZE; i++) {
130 		list_link_init(&dn->dn_dirty_link[i]);
131 		dn->dn_free_ranges[i] = NULL;
132 		list_create(&dn->dn_dirty_records[i],
133 		    sizeof (dbuf_dirty_record_t),
134 		    offsetof(dbuf_dirty_record_t, dr_dirty_node));
135 	}
136 
137 	dn->dn_allocated_txg = 0;
138 	dn->dn_free_txg = 0;
139 	dn->dn_assigned_txg = 0;
140 	dn->dn_dirtyctx = 0;
141 	dn->dn_dirtyctx_firstset = NULL;
142 	dn->dn_bonus = NULL;
143 	dn->dn_have_spill = B_FALSE;
144 	dn->dn_zio = NULL;
145 	dn->dn_oldused = 0;
146 	dn->dn_oldflags = 0;
147 	dn->dn_olduid = 0;
148 	dn->dn_oldgid = 0;
149 	dn->dn_newuid = 0;
150 	dn->dn_newgid = 0;
151 	dn->dn_id_flags = 0;
152 
153 	dn->dn_dbufs_count = 0;
154 	dn->dn_unlisted_l0_blkid = 0;
155 	avl_create(&dn->dn_dbufs, dbuf_compare, sizeof (dmu_buf_impl_t),
156 	    offsetof(dmu_buf_impl_t, db_link));
157 
158 	dn->dn_moved = 0;
159 	POINTER_INVALIDATE(&dn->dn_objset);
160 	return (0);
161 }
162 
163 /* ARGSUSED */
164 static void
dnode_dest(void * arg,void * unused)165 dnode_dest(void *arg, void *unused)
166 {
167 	int i;
168 	dnode_t *dn = arg;
169 
170 	rw_destroy(&dn->dn_struct_rwlock);
171 	mutex_destroy(&dn->dn_mtx);
172 	mutex_destroy(&dn->dn_dbufs_mtx);
173 	cv_destroy(&dn->dn_notxholds);
174 	refcount_destroy(&dn->dn_holds);
175 	refcount_destroy(&dn->dn_tx_holds);
176 	ASSERT(!list_link_active(&dn->dn_link));
177 
178 	for (i = 0; i < TXG_SIZE; i++) {
179 		ASSERT(!list_link_active(&dn->dn_dirty_link[i]));
180 		ASSERT3P(dn->dn_free_ranges[i], ==, NULL);
181 		list_destroy(&dn->dn_dirty_records[i]);
182 		ASSERT0(dn->dn_next_nblkptr[i]);
183 		ASSERT0(dn->dn_next_nlevels[i]);
184 		ASSERT0(dn->dn_next_indblkshift[i]);
185 		ASSERT0(dn->dn_next_bonustype[i]);
186 		ASSERT0(dn->dn_rm_spillblk[i]);
187 		ASSERT0(dn->dn_next_bonuslen[i]);
188 		ASSERT0(dn->dn_next_blksz[i]);
189 	}
190 
191 	ASSERT0(dn->dn_allocated_txg);
192 	ASSERT0(dn->dn_free_txg);
193 	ASSERT0(dn->dn_assigned_txg);
194 	ASSERT0(dn->dn_dirtyctx);
195 	ASSERT3P(dn->dn_dirtyctx_firstset, ==, NULL);
196 	ASSERT3P(dn->dn_bonus, ==, NULL);
197 	ASSERT(!dn->dn_have_spill);
198 	ASSERT3P(dn->dn_zio, ==, NULL);
199 	ASSERT0(dn->dn_oldused);
200 	ASSERT0(dn->dn_oldflags);
201 	ASSERT0(dn->dn_olduid);
202 	ASSERT0(dn->dn_oldgid);
203 	ASSERT0(dn->dn_newuid);
204 	ASSERT0(dn->dn_newgid);
205 	ASSERT0(dn->dn_id_flags);
206 
207 	ASSERT0(dn->dn_dbufs_count);
208 	ASSERT0(dn->dn_unlisted_l0_blkid);
209 	avl_destroy(&dn->dn_dbufs);
210 }
211 
212 void
dnode_init(void)213 dnode_init(void)
214 {
215 	ASSERT(dnode_cache == NULL);
216 	dnode_cache = kmem_cache_create("dnode_t",
217 	    sizeof (dnode_t),
218 	    0, dnode_cons, dnode_dest, NULL, NULL, NULL, 0);
219 	kmem_cache_set_move(dnode_cache, dnode_move);
220 }
221 
222 void
dnode_fini(void)223 dnode_fini(void)
224 {
225 	kmem_cache_destroy(dnode_cache);
226 	dnode_cache = NULL;
227 }
228 
229 
230 #ifdef ZFS_DEBUG
231 void
dnode_verify(dnode_t * dn)232 dnode_verify(dnode_t *dn)
233 {
234 	int drop_struct_lock = FALSE;
235 
236 	ASSERT(dn->dn_phys);
237 	ASSERT(dn->dn_objset);
238 	ASSERT(dn->dn_handle->dnh_dnode == dn);
239 
240 	ASSERT(DMU_OT_IS_VALID(dn->dn_phys->dn_type));
241 
242 	if (!(zfs_flags & ZFS_DEBUG_DNODE_VERIFY))
243 		return;
244 
245 	if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
246 		rw_enter(&dn->dn_struct_rwlock, RW_READER);
247 		drop_struct_lock = TRUE;
248 	}
249 	if (dn->dn_phys->dn_type != DMU_OT_NONE || dn->dn_allocated_txg != 0) {
250 		int i;
251 		ASSERT3U(dn->dn_indblkshift, >=, 0);
252 		ASSERT3U(dn->dn_indblkshift, <=, SPA_MAXBLOCKSHIFT);
253 		if (dn->dn_datablkshift) {
254 			ASSERT3U(dn->dn_datablkshift, >=, SPA_MINBLOCKSHIFT);
255 			ASSERT3U(dn->dn_datablkshift, <=, SPA_MAXBLOCKSHIFT);
256 			ASSERT3U(1<<dn->dn_datablkshift, ==, dn->dn_datablksz);
257 		}
258 		ASSERT3U(dn->dn_nlevels, <=, 30);
259 		ASSERT(DMU_OT_IS_VALID(dn->dn_type));
260 		ASSERT3U(dn->dn_nblkptr, >=, 1);
261 		ASSERT3U(dn->dn_nblkptr, <=, DN_MAX_NBLKPTR);
262 		ASSERT3U(dn->dn_bonuslen, <=, DN_MAX_BONUSLEN);
263 		ASSERT3U(dn->dn_datablksz, ==,
264 		    dn->dn_datablkszsec << SPA_MINBLOCKSHIFT);
265 		ASSERT3U(ISP2(dn->dn_datablksz), ==, dn->dn_datablkshift != 0);
266 		ASSERT3U((dn->dn_nblkptr - 1) * sizeof (blkptr_t) +
267 		    dn->dn_bonuslen, <=, DN_MAX_BONUSLEN);
268 		for (i = 0; i < TXG_SIZE; i++) {
269 			ASSERT3U(dn->dn_next_nlevels[i], <=, dn->dn_nlevels);
270 		}
271 	}
272 	if (dn->dn_phys->dn_type != DMU_OT_NONE)
273 		ASSERT3U(dn->dn_phys->dn_nlevels, <=, dn->dn_nlevels);
274 	ASSERT(DMU_OBJECT_IS_SPECIAL(dn->dn_object) || dn->dn_dbuf != NULL);
275 	if (dn->dn_dbuf != NULL) {
276 		ASSERT3P(dn->dn_phys, ==,
277 		    (dnode_phys_t *)dn->dn_dbuf->db.db_data +
278 		    (dn->dn_object % (dn->dn_dbuf->db.db_size >> DNODE_SHIFT)));
279 	}
280 	if (drop_struct_lock)
281 		rw_exit(&dn->dn_struct_rwlock);
282 }
283 #endif
284 
285 void
dnode_byteswap(dnode_phys_t * dnp)286 dnode_byteswap(dnode_phys_t *dnp)
287 {
288 	uint64_t *buf64 = (void*)&dnp->dn_blkptr;
289 	int i;
290 
291 	if (dnp->dn_type == DMU_OT_NONE) {
292 		bzero(dnp, sizeof (dnode_phys_t));
293 		return;
294 	}
295 
296 	dnp->dn_datablkszsec = BSWAP_16(dnp->dn_datablkszsec);
297 	dnp->dn_bonuslen = BSWAP_16(dnp->dn_bonuslen);
298 	dnp->dn_maxblkid = BSWAP_64(dnp->dn_maxblkid);
299 	dnp->dn_used = BSWAP_64(dnp->dn_used);
300 
301 	/*
302 	 * dn_nblkptr is only one byte, so it's OK to read it in either
303 	 * byte order.  We can't read dn_bouslen.
304 	 */
305 	ASSERT(dnp->dn_indblkshift <= SPA_MAXBLOCKSHIFT);
306 	ASSERT(dnp->dn_nblkptr <= DN_MAX_NBLKPTR);
307 	for (i = 0; i < dnp->dn_nblkptr * sizeof (blkptr_t)/8; i++)
308 		buf64[i] = BSWAP_64(buf64[i]);
309 
310 	/*
311 	 * OK to check dn_bonuslen for zero, because it won't matter if
312 	 * we have the wrong byte order.  This is necessary because the
313 	 * dnode dnode is smaller than a regular dnode.
314 	 */
315 	if (dnp->dn_bonuslen != 0) {
316 		/*
317 		 * Note that the bonus length calculated here may be
318 		 * longer than the actual bonus buffer.  This is because
319 		 * we always put the bonus buffer after the last block
320 		 * pointer (instead of packing it against the end of the
321 		 * dnode buffer).
322 		 */
323 		int off = (dnp->dn_nblkptr-1) * sizeof (blkptr_t);
324 		size_t len = DN_MAX_BONUSLEN - off;
325 		ASSERT(DMU_OT_IS_VALID(dnp->dn_bonustype));
326 		dmu_object_byteswap_t byteswap =
327 		    DMU_OT_BYTESWAP(dnp->dn_bonustype);
328 		dmu_ot_byteswap[byteswap].ob_func(dnp->dn_bonus + off, len);
329 	}
330 
331 	/* Swap SPILL block if we have one */
332 	if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR)
333 		byteswap_uint64_array(&dnp->dn_spill, sizeof (blkptr_t));
334 
335 }
336 
337 void
dnode_buf_byteswap(void * vbuf,size_t size)338 dnode_buf_byteswap(void *vbuf, size_t size)
339 {
340 	dnode_phys_t *buf = vbuf;
341 	int i;
342 
343 	ASSERT3U(sizeof (dnode_phys_t), ==, (1<<DNODE_SHIFT));
344 	ASSERT((size & (sizeof (dnode_phys_t)-1)) == 0);
345 
346 	size >>= DNODE_SHIFT;
347 	for (i = 0; i < size; i++) {
348 		dnode_byteswap(buf);
349 		buf++;
350 	}
351 }
352 
353 void
dnode_setbonuslen(dnode_t * dn,int newsize,dmu_tx_t * tx)354 dnode_setbonuslen(dnode_t *dn, int newsize, dmu_tx_t *tx)
355 {
356 	ASSERT3U(refcount_count(&dn->dn_holds), >=, 1);
357 
358 	dnode_setdirty(dn, tx);
359 	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
360 	ASSERT3U(newsize, <=, DN_MAX_BONUSLEN -
361 	    (dn->dn_nblkptr-1) * sizeof (blkptr_t));
362 	dn->dn_bonuslen = newsize;
363 	if (newsize == 0)
364 		dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = DN_ZERO_BONUSLEN;
365 	else
366 		dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = dn->dn_bonuslen;
367 	rw_exit(&dn->dn_struct_rwlock);
368 }
369 
370 void
dnode_setbonus_type(dnode_t * dn,dmu_object_type_t newtype,dmu_tx_t * tx)371 dnode_setbonus_type(dnode_t *dn, dmu_object_type_t newtype, dmu_tx_t *tx)
372 {
373 	ASSERT3U(refcount_count(&dn->dn_holds), >=, 1);
374 	dnode_setdirty(dn, tx);
375 	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
376 	dn->dn_bonustype = newtype;
377 	dn->dn_next_bonustype[tx->tx_txg & TXG_MASK] = dn->dn_bonustype;
378 	rw_exit(&dn->dn_struct_rwlock);
379 }
380 
381 void
dnode_rm_spill(dnode_t * dn,dmu_tx_t * tx)382 dnode_rm_spill(dnode_t *dn, dmu_tx_t *tx)
383 {
384 	ASSERT3U(refcount_count(&dn->dn_holds), >=, 1);
385 	ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
386 	dnode_setdirty(dn, tx);
387 	dn->dn_rm_spillblk[tx->tx_txg&TXG_MASK] = DN_KILL_SPILLBLK;
388 	dn->dn_have_spill = B_FALSE;
389 }
390 
391 static void
dnode_setdblksz(dnode_t * dn,int size)392 dnode_setdblksz(dnode_t *dn, int size)
393 {
394 	ASSERT0(P2PHASE(size, SPA_MINBLOCKSIZE));
395 	ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
396 	ASSERT3U(size, >=, SPA_MINBLOCKSIZE);
397 	ASSERT3U(size >> SPA_MINBLOCKSHIFT, <,
398 	    1<<(sizeof (dn->dn_phys->dn_datablkszsec) * 8));
399 	dn->dn_datablksz = size;
400 	dn->dn_datablkszsec = size >> SPA_MINBLOCKSHIFT;
401 	dn->dn_datablkshift = ISP2(size) ? highbit64(size - 1) : 0;
402 }
403 
404 static dnode_t *
dnode_create(objset_t * os,dnode_phys_t * dnp,dmu_buf_impl_t * db,uint64_t object,dnode_handle_t * dnh)405 dnode_create(objset_t *os, dnode_phys_t *dnp, dmu_buf_impl_t *db,
406     uint64_t object, dnode_handle_t *dnh)
407 {
408 	dnode_t *dn = kmem_cache_alloc(dnode_cache, KM_SLEEP);
409 
410 	ASSERT(!POINTER_IS_VALID(dn->dn_objset));
411 	dn->dn_moved = 0;
412 
413 	/*
414 	 * Defer setting dn_objset until the dnode is ready to be a candidate
415 	 * for the dnode_move() callback.
416 	 */
417 	dn->dn_object = object;
418 	dn->dn_dbuf = db;
419 	dn->dn_handle = dnh;
420 	dn->dn_phys = dnp;
421 
422 	if (dnp->dn_datablkszsec) {
423 		dnode_setdblksz(dn, dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT);
424 	} else {
425 		dn->dn_datablksz = 0;
426 		dn->dn_datablkszsec = 0;
427 		dn->dn_datablkshift = 0;
428 	}
429 	dn->dn_indblkshift = dnp->dn_indblkshift;
430 	dn->dn_nlevels = dnp->dn_nlevels;
431 	dn->dn_type = dnp->dn_type;
432 	dn->dn_nblkptr = dnp->dn_nblkptr;
433 	dn->dn_checksum = dnp->dn_checksum;
434 	dn->dn_compress = dnp->dn_compress;
435 	dn->dn_bonustype = dnp->dn_bonustype;
436 	dn->dn_bonuslen = dnp->dn_bonuslen;
437 	dn->dn_maxblkid = dnp->dn_maxblkid;
438 	dn->dn_have_spill = ((dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) != 0);
439 	dn->dn_id_flags = 0;
440 
441 	dmu_zfetch_init(&dn->dn_zfetch, dn);
442 
443 	ASSERT(DMU_OT_IS_VALID(dn->dn_phys->dn_type));
444 
445 	mutex_enter(&os->os_lock);
446 	list_insert_head(&os->os_dnodes, dn);
447 	membar_producer();
448 	/*
449 	 * Everything else must be valid before assigning dn_objset makes the
450 	 * dnode eligible for dnode_move().
451 	 */
452 	dn->dn_objset = os;
453 	mutex_exit(&os->os_lock);
454 
455 	arc_space_consume(sizeof (dnode_t), ARC_SPACE_OTHER);
456 	return (dn);
457 }
458 
459 /*
460  * Caller must be holding the dnode handle, which is released upon return.
461  */
462 static void
dnode_destroy(dnode_t * dn)463 dnode_destroy(dnode_t *dn)
464 {
465 	objset_t *os = dn->dn_objset;
466 
467 	ASSERT((dn->dn_id_flags & DN_ID_NEW_EXIST) == 0);
468 
469 	mutex_enter(&os->os_lock);
470 	POINTER_INVALIDATE(&dn->dn_objset);
471 	list_remove(&os->os_dnodes, dn);
472 	mutex_exit(&os->os_lock);
473 
474 	/* the dnode can no longer move, so we can release the handle */
475 	zrl_remove(&dn->dn_handle->dnh_zrlock);
476 
477 	dn->dn_allocated_txg = 0;
478 	dn->dn_free_txg = 0;
479 	dn->dn_assigned_txg = 0;
480 
481 	dn->dn_dirtyctx = 0;
482 	if (dn->dn_dirtyctx_firstset != NULL) {
483 		kmem_free(dn->dn_dirtyctx_firstset, 1);
484 		dn->dn_dirtyctx_firstset = NULL;
485 	}
486 	if (dn->dn_bonus != NULL) {
487 		mutex_enter(&dn->dn_bonus->db_mtx);
488 		dbuf_evict(dn->dn_bonus);
489 		dn->dn_bonus = NULL;
490 	}
491 	dn->dn_zio = NULL;
492 
493 	dn->dn_have_spill = B_FALSE;
494 	dn->dn_oldused = 0;
495 	dn->dn_oldflags = 0;
496 	dn->dn_olduid = 0;
497 	dn->dn_oldgid = 0;
498 	dn->dn_newuid = 0;
499 	dn->dn_newgid = 0;
500 	dn->dn_id_flags = 0;
501 	dn->dn_unlisted_l0_blkid = 0;
502 
503 	dmu_zfetch_rele(&dn->dn_zfetch);
504 	kmem_cache_free(dnode_cache, dn);
505 	arc_space_return(sizeof (dnode_t), ARC_SPACE_OTHER);
506 }
507 
508 void
dnode_allocate(dnode_t * dn,dmu_object_type_t ot,int blocksize,int ibs,dmu_object_type_t bonustype,int bonuslen,dmu_tx_t * tx)509 dnode_allocate(dnode_t *dn, dmu_object_type_t ot, int blocksize, int ibs,
510     dmu_object_type_t bonustype, int bonuslen, dmu_tx_t *tx)
511 {
512 	int i;
513 
514 	ASSERT3U(blocksize, <=,
515 	    spa_maxblocksize(dmu_objset_spa(dn->dn_objset)));
516 	if (blocksize == 0)
517 		blocksize = 1 << zfs_default_bs;
518 	else
519 		blocksize = P2ROUNDUP(blocksize, SPA_MINBLOCKSIZE);
520 
521 	if (ibs == 0)
522 		ibs = zfs_default_ibs;
523 
524 	ibs = MIN(MAX(ibs, DN_MIN_INDBLKSHIFT), DN_MAX_INDBLKSHIFT);
525 
526 	dprintf("os=%p obj=%llu txg=%llu blocksize=%d ibs=%d\n", dn->dn_objset,
527 	    dn->dn_object, tx->tx_txg, blocksize, ibs);
528 
529 	ASSERT(dn->dn_type == DMU_OT_NONE);
530 	ASSERT(bcmp(dn->dn_phys, &dnode_phys_zero, sizeof (dnode_phys_t)) == 0);
531 	ASSERT(dn->dn_phys->dn_type == DMU_OT_NONE);
532 	ASSERT(ot != DMU_OT_NONE);
533 	ASSERT(DMU_OT_IS_VALID(ot));
534 	ASSERT((bonustype == DMU_OT_NONE && bonuslen == 0) ||
535 	    (bonustype == DMU_OT_SA && bonuslen == 0) ||
536 	    (bonustype != DMU_OT_NONE && bonuslen != 0));
537 	ASSERT(DMU_OT_IS_VALID(bonustype));
538 	ASSERT3U(bonuslen, <=, DN_MAX_BONUSLEN);
539 	ASSERT(dn->dn_type == DMU_OT_NONE);
540 	ASSERT0(dn->dn_maxblkid);
541 	ASSERT0(dn->dn_allocated_txg);
542 	ASSERT0(dn->dn_assigned_txg);
543 	ASSERT(refcount_is_zero(&dn->dn_tx_holds));
544 	ASSERT3U(refcount_count(&dn->dn_holds), <=, 1);
545 	ASSERT(avl_is_empty(&dn->dn_dbufs));
546 
547 	for (i = 0; i < TXG_SIZE; i++) {
548 		ASSERT0(dn->dn_next_nblkptr[i]);
549 		ASSERT0(dn->dn_next_nlevels[i]);
550 		ASSERT0(dn->dn_next_indblkshift[i]);
551 		ASSERT0(dn->dn_next_bonuslen[i]);
552 		ASSERT0(dn->dn_next_bonustype[i]);
553 		ASSERT0(dn->dn_rm_spillblk[i]);
554 		ASSERT0(dn->dn_next_blksz[i]);
555 		ASSERT(!list_link_active(&dn->dn_dirty_link[i]));
556 		ASSERT3P(list_head(&dn->dn_dirty_records[i]), ==, NULL);
557 		ASSERT3P(dn->dn_free_ranges[i], ==, NULL);
558 	}
559 
560 	dn->dn_type = ot;
561 	dnode_setdblksz(dn, blocksize);
562 	dn->dn_indblkshift = ibs;
563 	dn->dn_nlevels = 1;
564 	if (bonustype == DMU_OT_SA) /* Maximize bonus space for SA */
565 		dn->dn_nblkptr = 1;
566 	else
567 		dn->dn_nblkptr = 1 +
568 		    ((DN_MAX_BONUSLEN - bonuslen) >> SPA_BLKPTRSHIFT);
569 	dn->dn_bonustype = bonustype;
570 	dn->dn_bonuslen = bonuslen;
571 	dn->dn_checksum = ZIO_CHECKSUM_INHERIT;
572 	dn->dn_compress = ZIO_COMPRESS_INHERIT;
573 	dn->dn_dirtyctx = 0;
574 
575 	dn->dn_free_txg = 0;
576 	if (dn->dn_dirtyctx_firstset) {
577 		kmem_free(dn->dn_dirtyctx_firstset, 1);
578 		dn->dn_dirtyctx_firstset = NULL;
579 	}
580 
581 	dn->dn_allocated_txg = tx->tx_txg;
582 	dn->dn_id_flags = 0;
583 
584 	dnode_setdirty(dn, tx);
585 	dn->dn_next_indblkshift[tx->tx_txg & TXG_MASK] = ibs;
586 	dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = dn->dn_bonuslen;
587 	dn->dn_next_bonustype[tx->tx_txg & TXG_MASK] = dn->dn_bonustype;
588 	dn->dn_next_blksz[tx->tx_txg & TXG_MASK] = dn->dn_datablksz;
589 }
590 
591 void
dnode_reallocate(dnode_t * dn,dmu_object_type_t ot,int blocksize,dmu_object_type_t bonustype,int bonuslen,dmu_tx_t * tx)592 dnode_reallocate(dnode_t *dn, dmu_object_type_t ot, int blocksize,
593     dmu_object_type_t bonustype, int bonuslen, dmu_tx_t *tx)
594 {
595 	int nblkptr;
596 
597 	ASSERT3U(blocksize, >=, SPA_MINBLOCKSIZE);
598 	ASSERT3U(blocksize, <=,
599 	    spa_maxblocksize(dmu_objset_spa(dn->dn_objset)));
600 	ASSERT0(blocksize % SPA_MINBLOCKSIZE);
601 	ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT || dmu_tx_private_ok(tx));
602 	ASSERT(tx->tx_txg != 0);
603 	ASSERT((bonustype == DMU_OT_NONE && bonuslen == 0) ||
604 	    (bonustype != DMU_OT_NONE && bonuslen != 0) ||
605 	    (bonustype == DMU_OT_SA && bonuslen == 0));
606 	ASSERT(DMU_OT_IS_VALID(bonustype));
607 	ASSERT3U(bonuslen, <=, DN_MAX_BONUSLEN);
608 
609 	/* clean up any unreferenced dbufs */
610 	dnode_evict_dbufs(dn);
611 
612 	dn->dn_id_flags = 0;
613 
614 	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
615 	dnode_setdirty(dn, tx);
616 	if (dn->dn_datablksz != blocksize) {
617 		/* change blocksize */
618 		ASSERT(dn->dn_maxblkid == 0 &&
619 		    (BP_IS_HOLE(&dn->dn_phys->dn_blkptr[0]) ||
620 		    dnode_block_freed(dn, 0)));
621 		dnode_setdblksz(dn, blocksize);
622 		dn->dn_next_blksz[tx->tx_txg&TXG_MASK] = blocksize;
623 	}
624 	if (dn->dn_bonuslen != bonuslen)
625 		dn->dn_next_bonuslen[tx->tx_txg&TXG_MASK] = bonuslen;
626 
627 	if (bonustype == DMU_OT_SA) /* Maximize bonus space for SA */
628 		nblkptr = 1;
629 	else
630 		nblkptr = 1 + ((DN_MAX_BONUSLEN - bonuslen) >> SPA_BLKPTRSHIFT);
631 	if (dn->dn_bonustype != bonustype)
632 		dn->dn_next_bonustype[tx->tx_txg&TXG_MASK] = bonustype;
633 	if (dn->dn_nblkptr != nblkptr)
634 		dn->dn_next_nblkptr[tx->tx_txg&TXG_MASK] = nblkptr;
635 	if (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
636 		dbuf_rm_spill(dn, tx);
637 		dnode_rm_spill(dn, tx);
638 	}
639 	rw_exit(&dn->dn_struct_rwlock);
640 
641 	/* change type */
642 	dn->dn_type = ot;
643 
644 	/* change bonus size and type */
645 	mutex_enter(&dn->dn_mtx);
646 	dn->dn_bonustype = bonustype;
647 	dn->dn_bonuslen = bonuslen;
648 	dn->dn_nblkptr = nblkptr;
649 	dn->dn_checksum = ZIO_CHECKSUM_INHERIT;
650 	dn->dn_compress = ZIO_COMPRESS_INHERIT;
651 	ASSERT3U(dn->dn_nblkptr, <=, DN_MAX_NBLKPTR);
652 
653 	/* fix up the bonus db_size */
654 	if (dn->dn_bonus) {
655 		dn->dn_bonus->db.db_size =
656 		    DN_MAX_BONUSLEN - (dn->dn_nblkptr-1) * sizeof (blkptr_t);
657 		ASSERT(dn->dn_bonuslen <= dn->dn_bonus->db.db_size);
658 	}
659 
660 	dn->dn_allocated_txg = tx->tx_txg;
661 	mutex_exit(&dn->dn_mtx);
662 }
663 
664 #ifdef	DNODE_STATS
665 static struct {
666 	uint64_t dms_dnode_invalid;
667 	uint64_t dms_dnode_recheck1;
668 	uint64_t dms_dnode_recheck2;
669 	uint64_t dms_dnode_special;
670 	uint64_t dms_dnode_handle;
671 	uint64_t dms_dnode_rwlock;
672 	uint64_t dms_dnode_active;
673 } dnode_move_stats;
674 #endif	/* DNODE_STATS */
675 
676 static void
dnode_move_impl(dnode_t * odn,dnode_t * ndn)677 dnode_move_impl(dnode_t *odn, dnode_t *ndn)
678 {
679 	int i;
680 
681 	ASSERT(!RW_LOCK_HELD(&odn->dn_struct_rwlock));
682 	ASSERT(MUTEX_NOT_HELD(&odn->dn_mtx));
683 	ASSERT(MUTEX_NOT_HELD(&odn->dn_dbufs_mtx));
684 	ASSERT(!RW_LOCK_HELD(&odn->dn_zfetch.zf_rwlock));
685 
686 	/* Copy fields. */
687 	ndn->dn_objset = odn->dn_objset;
688 	ndn->dn_object = odn->dn_object;
689 	ndn->dn_dbuf = odn->dn_dbuf;
690 	ndn->dn_handle = odn->dn_handle;
691 	ndn->dn_phys = odn->dn_phys;
692 	ndn->dn_type = odn->dn_type;
693 	ndn->dn_bonuslen = odn->dn_bonuslen;
694 	ndn->dn_bonustype = odn->dn_bonustype;
695 	ndn->dn_nblkptr = odn->dn_nblkptr;
696 	ndn->dn_checksum = odn->dn_checksum;
697 	ndn->dn_compress = odn->dn_compress;
698 	ndn->dn_nlevels = odn->dn_nlevels;
699 	ndn->dn_indblkshift = odn->dn_indblkshift;
700 	ndn->dn_datablkshift = odn->dn_datablkshift;
701 	ndn->dn_datablkszsec = odn->dn_datablkszsec;
702 	ndn->dn_datablksz = odn->dn_datablksz;
703 	ndn->dn_maxblkid = odn->dn_maxblkid;
704 	bcopy(&odn->dn_next_nblkptr[0], &ndn->dn_next_nblkptr[0],
705 	    sizeof (odn->dn_next_nblkptr));
706 	bcopy(&odn->dn_next_nlevels[0], &ndn->dn_next_nlevels[0],
707 	    sizeof (odn->dn_next_nlevels));
708 	bcopy(&odn->dn_next_indblkshift[0], &ndn->dn_next_indblkshift[0],
709 	    sizeof (odn->dn_next_indblkshift));
710 	bcopy(&odn->dn_next_bonustype[0], &ndn->dn_next_bonustype[0],
711 	    sizeof (odn->dn_next_bonustype));
712 	bcopy(&odn->dn_rm_spillblk[0], &ndn->dn_rm_spillblk[0],
713 	    sizeof (odn->dn_rm_spillblk));
714 	bcopy(&odn->dn_next_bonuslen[0], &ndn->dn_next_bonuslen[0],
715 	    sizeof (odn->dn_next_bonuslen));
716 	bcopy(&odn->dn_next_blksz[0], &ndn->dn_next_blksz[0],
717 	    sizeof (odn->dn_next_blksz));
718 	for (i = 0; i < TXG_SIZE; i++) {
719 		list_move_tail(&ndn->dn_dirty_records[i],
720 		    &odn->dn_dirty_records[i]);
721 	}
722 	bcopy(&odn->dn_free_ranges[0], &ndn->dn_free_ranges[0],
723 	    sizeof (odn->dn_free_ranges));
724 	ndn->dn_allocated_txg = odn->dn_allocated_txg;
725 	ndn->dn_free_txg = odn->dn_free_txg;
726 	ndn->dn_assigned_txg = odn->dn_assigned_txg;
727 	ndn->dn_dirtyctx = odn->dn_dirtyctx;
728 	ndn->dn_dirtyctx_firstset = odn->dn_dirtyctx_firstset;
729 	ASSERT(refcount_count(&odn->dn_tx_holds) == 0);
730 	refcount_transfer(&ndn->dn_holds, &odn->dn_holds);
731 	ASSERT(avl_is_empty(&ndn->dn_dbufs));
732 	avl_swap(&ndn->dn_dbufs, &odn->dn_dbufs);
733 	ndn->dn_dbufs_count = odn->dn_dbufs_count;
734 	ndn->dn_unlisted_l0_blkid = odn->dn_unlisted_l0_blkid;
735 	ndn->dn_bonus = odn->dn_bonus;
736 	ndn->dn_have_spill = odn->dn_have_spill;
737 	ndn->dn_zio = odn->dn_zio;
738 	ndn->dn_oldused = odn->dn_oldused;
739 	ndn->dn_oldflags = odn->dn_oldflags;
740 	ndn->dn_olduid = odn->dn_olduid;
741 	ndn->dn_oldgid = odn->dn_oldgid;
742 	ndn->dn_newuid = odn->dn_newuid;
743 	ndn->dn_newgid = odn->dn_newgid;
744 	ndn->dn_id_flags = odn->dn_id_flags;
745 	dmu_zfetch_init(&ndn->dn_zfetch, NULL);
746 	list_move_tail(&ndn->dn_zfetch.zf_stream, &odn->dn_zfetch.zf_stream);
747 	ndn->dn_zfetch.zf_dnode = odn->dn_zfetch.zf_dnode;
748 	ndn->dn_zfetch.zf_stream_cnt = odn->dn_zfetch.zf_stream_cnt;
749 	ndn->dn_zfetch.zf_alloc_fail = odn->dn_zfetch.zf_alloc_fail;
750 
751 	/*
752 	 * Update back pointers. Updating the handle fixes the back pointer of
753 	 * every descendant dbuf as well as the bonus dbuf.
754 	 */
755 	ASSERT(ndn->dn_handle->dnh_dnode == odn);
756 	ndn->dn_handle->dnh_dnode = ndn;
757 	if (ndn->dn_zfetch.zf_dnode == odn) {
758 		ndn->dn_zfetch.zf_dnode = ndn;
759 	}
760 
761 	/*
762 	 * Invalidate the original dnode by clearing all of its back pointers.
763 	 */
764 	odn->dn_dbuf = NULL;
765 	odn->dn_handle = NULL;
766 	avl_create(&odn->dn_dbufs, dbuf_compare, sizeof (dmu_buf_impl_t),
767 	    offsetof(dmu_buf_impl_t, db_link));
768 	odn->dn_dbufs_count = 0;
769 	odn->dn_unlisted_l0_blkid = 0;
770 	odn->dn_bonus = NULL;
771 	odn->dn_zfetch.zf_dnode = NULL;
772 
773 	/*
774 	 * Set the low bit of the objset pointer to ensure that dnode_move()
775 	 * recognizes the dnode as invalid in any subsequent callback.
776 	 */
777 	POINTER_INVALIDATE(&odn->dn_objset);
778 
779 	/*
780 	 * Satisfy the destructor.
781 	 */
782 	for (i = 0; i < TXG_SIZE; i++) {
783 		list_create(&odn->dn_dirty_records[i],
784 		    sizeof (dbuf_dirty_record_t),
785 		    offsetof(dbuf_dirty_record_t, dr_dirty_node));
786 		odn->dn_free_ranges[i] = NULL;
787 		odn->dn_next_nlevels[i] = 0;
788 		odn->dn_next_indblkshift[i] = 0;
789 		odn->dn_next_bonustype[i] = 0;
790 		odn->dn_rm_spillblk[i] = 0;
791 		odn->dn_next_bonuslen[i] = 0;
792 		odn->dn_next_blksz[i] = 0;
793 	}
794 	odn->dn_allocated_txg = 0;
795 	odn->dn_free_txg = 0;
796 	odn->dn_assigned_txg = 0;
797 	odn->dn_dirtyctx = 0;
798 	odn->dn_dirtyctx_firstset = NULL;
799 	odn->dn_have_spill = B_FALSE;
800 	odn->dn_zio = NULL;
801 	odn->dn_oldused = 0;
802 	odn->dn_oldflags = 0;
803 	odn->dn_olduid = 0;
804 	odn->dn_oldgid = 0;
805 	odn->dn_newuid = 0;
806 	odn->dn_newgid = 0;
807 	odn->dn_id_flags = 0;
808 
809 	/*
810 	 * Mark the dnode.
811 	 */
812 	ndn->dn_moved = 1;
813 	odn->dn_moved = (uint8_t)-1;
814 }
815 
816 #ifdef sun
817 #ifdef	_KERNEL
818 /*ARGSUSED*/
819 static kmem_cbrc_t
dnode_move(void * buf,void * newbuf,size_t size,void * arg)820 dnode_move(void *buf, void *newbuf, size_t size, void *arg)
821 {
822 	dnode_t *odn = buf, *ndn = newbuf;
823 	objset_t *os;
824 	int64_t refcount;
825 	uint32_t dbufs;
826 
827 	/*
828 	 * The dnode is on the objset's list of known dnodes if the objset
829 	 * pointer is valid. We set the low bit of the objset pointer when
830 	 * freeing the dnode to invalidate it, and the memory patterns written
831 	 * by kmem (baddcafe and deadbeef) set at least one of the two low bits.
832 	 * A newly created dnode sets the objset pointer last of all to indicate
833 	 * that the dnode is known and in a valid state to be moved by this
834 	 * function.
835 	 */
836 	os = odn->dn_objset;
837 	if (!POINTER_IS_VALID(os)) {
838 		DNODE_STAT_ADD(dnode_move_stats.dms_dnode_invalid);
839 		return (KMEM_CBRC_DONT_KNOW);
840 	}
841 
842 	/*
843 	 * Ensure that the objset does not go away during the move.
844 	 */
845 	rw_enter(&os_lock, RW_WRITER);
846 	if (os != odn->dn_objset) {
847 		rw_exit(&os_lock);
848 		DNODE_STAT_ADD(dnode_move_stats.dms_dnode_recheck1);
849 		return (KMEM_CBRC_DONT_KNOW);
850 	}
851 
852 	/*
853 	 * If the dnode is still valid, then so is the objset. We know that no
854 	 * valid objset can be freed while we hold os_lock, so we can safely
855 	 * ensure that the objset remains in use.
856 	 */
857 	mutex_enter(&os->os_lock);
858 
859 	/*
860 	 * Recheck the objset pointer in case the dnode was removed just before
861 	 * acquiring the lock.
862 	 */
863 	if (os != odn->dn_objset) {
864 		mutex_exit(&os->os_lock);
865 		rw_exit(&os_lock);
866 		DNODE_STAT_ADD(dnode_move_stats.dms_dnode_recheck2);
867 		return (KMEM_CBRC_DONT_KNOW);
868 	}
869 
870 	/*
871 	 * At this point we know that as long as we hold os->os_lock, the dnode
872 	 * cannot be freed and fields within the dnode can be safely accessed.
873 	 * The objset listing this dnode cannot go away as long as this dnode is
874 	 * on its list.
875 	 */
876 	rw_exit(&os_lock);
877 	if (DMU_OBJECT_IS_SPECIAL(odn->dn_object)) {
878 		mutex_exit(&os->os_lock);
879 		DNODE_STAT_ADD(dnode_move_stats.dms_dnode_special);
880 		return (KMEM_CBRC_NO);
881 	}
882 	ASSERT(odn->dn_dbuf != NULL); /* only "special" dnodes have no parent */
883 
884 	/*
885 	 * Lock the dnode handle to prevent the dnode from obtaining any new
886 	 * holds. This also prevents the descendant dbufs and the bonus dbuf
887 	 * from accessing the dnode, so that we can discount their holds. The
888 	 * handle is safe to access because we know that while the dnode cannot
889 	 * go away, neither can its handle. Once we hold dnh_zrlock, we can
890 	 * safely move any dnode referenced only by dbufs.
891 	 */
892 	if (!zrl_tryenter(&odn->dn_handle->dnh_zrlock)) {
893 		mutex_exit(&os->os_lock);
894 		DNODE_STAT_ADD(dnode_move_stats.dms_dnode_handle);
895 		return (KMEM_CBRC_LATER);
896 	}
897 
898 	/*
899 	 * Ensure a consistent view of the dnode's holds and the dnode's dbufs.
900 	 * We need to guarantee that there is a hold for every dbuf in order to
901 	 * determine whether the dnode is actively referenced. Falsely matching
902 	 * a dbuf to an active hold would lead to an unsafe move. It's possible
903 	 * that a thread already having an active dnode hold is about to add a
904 	 * dbuf, and we can't compare hold and dbuf counts while the add is in
905 	 * progress.
906 	 */
907 	if (!rw_tryenter(&odn->dn_struct_rwlock, RW_WRITER)) {
908 		zrl_exit(&odn->dn_handle->dnh_zrlock);
909 		mutex_exit(&os->os_lock);
910 		DNODE_STAT_ADD(dnode_move_stats.dms_dnode_rwlock);
911 		return (KMEM_CBRC_LATER);
912 	}
913 
914 	/*
915 	 * A dbuf may be removed (evicted) without an active dnode hold. In that
916 	 * case, the dbuf count is decremented under the handle lock before the
917 	 * dbuf's hold is released. This order ensures that if we count the hold
918 	 * after the dbuf is removed but before its hold is released, we will
919 	 * treat the unmatched hold as active and exit safely. If we count the
920 	 * hold before the dbuf is removed, the hold is discounted, and the
921 	 * removal is blocked until the move completes.
922 	 */
923 	refcount = refcount_count(&odn->dn_holds);
924 	ASSERT(refcount >= 0);
925 	dbufs = odn->dn_dbufs_count;
926 
927 	/* We can't have more dbufs than dnode holds. */
928 	ASSERT3U(dbufs, <=, refcount);
929 	DTRACE_PROBE3(dnode__move, dnode_t *, odn, int64_t, refcount,
930 	    uint32_t, dbufs);
931 
932 	if (refcount > dbufs) {
933 		rw_exit(&odn->dn_struct_rwlock);
934 		zrl_exit(&odn->dn_handle->dnh_zrlock);
935 		mutex_exit(&os->os_lock);
936 		DNODE_STAT_ADD(dnode_move_stats.dms_dnode_active);
937 		return (KMEM_CBRC_LATER);
938 	}
939 
940 	rw_exit(&odn->dn_struct_rwlock);
941 
942 	/*
943 	 * At this point we know that anyone with a hold on the dnode is not
944 	 * actively referencing it. The dnode is known and in a valid state to
945 	 * move. We're holding the locks needed to execute the critical section.
946 	 */
947 	dnode_move_impl(odn, ndn);
948 
949 	list_link_replace(&odn->dn_link, &ndn->dn_link);
950 	/* If the dnode was safe to move, the refcount cannot have changed. */
951 	ASSERT(refcount == refcount_count(&ndn->dn_holds));
952 	ASSERT(dbufs == ndn->dn_dbufs_count);
953 	zrl_exit(&ndn->dn_handle->dnh_zrlock); /* handle has moved */
954 	mutex_exit(&os->os_lock);
955 
956 	return (KMEM_CBRC_YES);
957 }
958 #endif	/* _KERNEL */
959 #endif	/* sun */
960 
961 void
dnode_special_close(dnode_handle_t * dnh)962 dnode_special_close(dnode_handle_t *dnh)
963 {
964 	dnode_t *dn = dnh->dnh_dnode;
965 
966 	/*
967 	 * Wait for final references to the dnode to clear.  This can
968 	 * only happen if the arc is asyncronously evicting state that
969 	 * has a hold on this dnode while we are trying to evict this
970 	 * dnode.
971 	 */
972 	while (refcount_count(&dn->dn_holds) > 0)
973 		delay(1);
974 	zrl_add(&dnh->dnh_zrlock);
975 	dnode_destroy(dn); /* implicit zrl_remove() */
976 	zrl_destroy(&dnh->dnh_zrlock);
977 	dnh->dnh_dnode = NULL;
978 }
979 
980 dnode_t *
dnode_special_open(objset_t * os,dnode_phys_t * dnp,uint64_t object,dnode_handle_t * dnh)981 dnode_special_open(objset_t *os, dnode_phys_t *dnp, uint64_t object,
982     dnode_handle_t *dnh)
983 {
984 	dnode_t *dn = dnode_create(os, dnp, NULL, object, dnh);
985 	dnh->dnh_dnode = dn;
986 	zrl_init(&dnh->dnh_zrlock);
987 	DNODE_VERIFY(dn);
988 	return (dn);
989 }
990 
991 static void
dnode_buf_pageout(dmu_buf_t * db,void * arg)992 dnode_buf_pageout(dmu_buf_t *db, void *arg)
993 {
994 	dnode_children_t *children_dnodes = arg;
995 	int i;
996 	int epb = db->db_size >> DNODE_SHIFT;
997 
998 	ASSERT(epb == children_dnodes->dnc_count);
999 
1000 	for (i = 0; i < epb; i++) {
1001 		dnode_handle_t *dnh = &children_dnodes->dnc_children[i];
1002 		dnode_t *dn;
1003 
1004 		/*
1005 		 * The dnode handle lock guards against the dnode moving to
1006 		 * another valid address, so there is no need here to guard
1007 		 * against changes to or from NULL.
1008 		 */
1009 		if (dnh->dnh_dnode == NULL) {
1010 			zrl_destroy(&dnh->dnh_zrlock);
1011 			continue;
1012 		}
1013 
1014 		zrl_add(&dnh->dnh_zrlock);
1015 		dn = dnh->dnh_dnode;
1016 		/*
1017 		 * If there are holds on this dnode, then there should
1018 		 * be holds on the dnode's containing dbuf as well; thus
1019 		 * it wouldn't be eligible for eviction and this function
1020 		 * would not have been called.
1021 		 */
1022 		ASSERT(refcount_is_zero(&dn->dn_holds));
1023 		ASSERT(refcount_is_zero(&dn->dn_tx_holds));
1024 
1025 		dnode_destroy(dn); /* implicit zrl_remove() */
1026 		zrl_destroy(&dnh->dnh_zrlock);
1027 		dnh->dnh_dnode = NULL;
1028 	}
1029 	kmem_free(children_dnodes, sizeof (dnode_children_t) +
1030 	    epb * sizeof (dnode_handle_t));
1031 }
1032 
1033 /*
1034  * errors:
1035  * EINVAL - invalid object number.
1036  * EIO - i/o error.
1037  * succeeds even for free dnodes.
1038  */
1039 int
dnode_hold_impl(objset_t * os,uint64_t object,int flag,void * tag,dnode_t ** dnp)1040 dnode_hold_impl(objset_t *os, uint64_t object, int flag,
1041     void *tag, dnode_t **dnp)
1042 {
1043 	int epb, idx, err;
1044 	int drop_struct_lock = FALSE;
1045 	int type;
1046 	uint64_t blk;
1047 	dnode_t *mdn, *dn;
1048 	dmu_buf_impl_t *db;
1049 	dnode_children_t *children_dnodes;
1050 	dnode_handle_t *dnh;
1051 
1052 	/*
1053 	 * If you are holding the spa config lock as writer, you shouldn't
1054 	 * be asking the DMU to do *anything* unless it's the root pool
1055 	 * which may require us to read from the root filesystem while
1056 	 * holding some (not all) of the locks as writer.
1057 	 */
1058 	ASSERT(spa_config_held(os->os_spa, SCL_ALL, RW_WRITER) == 0 ||
1059 	    (spa_is_root(os->os_spa) &&
1060 	    spa_config_held(os->os_spa, SCL_STATE, RW_WRITER)));
1061 
1062 	if (object == DMU_USERUSED_OBJECT || object == DMU_GROUPUSED_OBJECT) {
1063 		dn = (object == DMU_USERUSED_OBJECT) ?
1064 		    DMU_USERUSED_DNODE(os) : DMU_GROUPUSED_DNODE(os);
1065 		if (dn == NULL)
1066 			return (SET_ERROR(ENOENT));
1067 		type = dn->dn_type;
1068 		if ((flag & DNODE_MUST_BE_ALLOCATED) && type == DMU_OT_NONE)
1069 			return (SET_ERROR(ENOENT));
1070 		if ((flag & DNODE_MUST_BE_FREE) && type != DMU_OT_NONE)
1071 			return (SET_ERROR(EEXIST));
1072 		DNODE_VERIFY(dn);
1073 		(void) refcount_add(&dn->dn_holds, tag);
1074 		*dnp = dn;
1075 		return (0);
1076 	}
1077 
1078 	if (object == 0 || object >= DN_MAX_OBJECT)
1079 		return (SET_ERROR(EINVAL));
1080 
1081 	mdn = DMU_META_DNODE(os);
1082 	ASSERT(mdn->dn_object == DMU_META_DNODE_OBJECT);
1083 
1084 	DNODE_VERIFY(mdn);
1085 
1086 	if (!RW_WRITE_HELD(&mdn->dn_struct_rwlock)) {
1087 		rw_enter(&mdn->dn_struct_rwlock, RW_READER);
1088 		drop_struct_lock = TRUE;
1089 	}
1090 
1091 	blk = dbuf_whichblock(mdn, object * sizeof (dnode_phys_t));
1092 
1093 	db = dbuf_hold(mdn, blk, FTAG);
1094 	if (drop_struct_lock)
1095 		rw_exit(&mdn->dn_struct_rwlock);
1096 	if (db == NULL)
1097 		return (SET_ERROR(EIO));
1098 	err = dbuf_read(db, NULL, DB_RF_CANFAIL);
1099 	if (err) {
1100 		dbuf_rele(db, FTAG);
1101 		return (err);
1102 	}
1103 
1104 	ASSERT3U(db->db.db_size, >=, 1<<DNODE_SHIFT);
1105 	epb = db->db.db_size >> DNODE_SHIFT;
1106 
1107 	idx = object & (epb-1);
1108 
1109 	ASSERT(DB_DNODE(db)->dn_type == DMU_OT_DNODE);
1110 	children_dnodes = dmu_buf_get_user(&db->db);
1111 	if (children_dnodes == NULL) {
1112 		int i;
1113 		dnode_children_t *winner;
1114 		children_dnodes = kmem_zalloc(sizeof (dnode_children_t) +
1115 		    epb * sizeof (dnode_handle_t), KM_SLEEP);
1116 		children_dnodes->dnc_count = epb;
1117 		dnh = &children_dnodes->dnc_children[0];
1118 		for (i = 0; i < epb; i++) {
1119 			zrl_init(&dnh[i].dnh_zrlock);
1120 			dnh[i].dnh_dnode = NULL;
1121 		}
1122 		if (winner = dmu_buf_set_user(&db->db, children_dnodes,
1123 		    dnode_buf_pageout)) {
1124 
1125 			for (i = 0; i < epb; i++) {
1126 				zrl_destroy(&dnh[i].dnh_zrlock);
1127 			}
1128 
1129 			kmem_free(children_dnodes, sizeof (dnode_children_t) +
1130 			    epb * sizeof (dnode_handle_t));
1131 			children_dnodes = winner;
1132 		}
1133 	}
1134 	ASSERT(children_dnodes->dnc_count == epb);
1135 
1136 	dnh = &children_dnodes->dnc_children[idx];
1137 	zrl_add(&dnh->dnh_zrlock);
1138 	if ((dn = dnh->dnh_dnode) == NULL) {
1139 		dnode_phys_t *phys = (dnode_phys_t *)db->db.db_data+idx;
1140 		dnode_t *winner;
1141 
1142 		dn = dnode_create(os, phys, db, object, dnh);
1143 		winner = atomic_cas_ptr(&dnh->dnh_dnode, NULL, dn);
1144 		if (winner != NULL) {
1145 			zrl_add(&dnh->dnh_zrlock);
1146 			dnode_destroy(dn); /* implicit zrl_remove() */
1147 			dn = winner;
1148 		}
1149 	}
1150 
1151 	mutex_enter(&dn->dn_mtx);
1152 	type = dn->dn_type;
1153 	if (dn->dn_free_txg ||
1154 	    ((flag & DNODE_MUST_BE_ALLOCATED) && type == DMU_OT_NONE) ||
1155 	    ((flag & DNODE_MUST_BE_FREE) &&
1156 	    (type != DMU_OT_NONE || !refcount_is_zero(&dn->dn_holds)))) {
1157 		mutex_exit(&dn->dn_mtx);
1158 		zrl_remove(&dnh->dnh_zrlock);
1159 		dbuf_rele(db, FTAG);
1160 		return (type == DMU_OT_NONE ? ENOENT : EEXIST);
1161 	}
1162 	mutex_exit(&dn->dn_mtx);
1163 
1164 	if (refcount_add(&dn->dn_holds, tag) == 1)
1165 		dbuf_add_ref(db, dnh);
1166 	/* Now we can rely on the hold to prevent the dnode from moving. */
1167 	zrl_remove(&dnh->dnh_zrlock);
1168 
1169 	DNODE_VERIFY(dn);
1170 	ASSERT3P(dn->dn_dbuf, ==, db);
1171 	ASSERT3U(dn->dn_object, ==, object);
1172 	dbuf_rele(db, FTAG);
1173 
1174 	*dnp = dn;
1175 	return (0);
1176 }
1177 
1178 /*
1179  * Return held dnode if the object is allocated, NULL if not.
1180  */
1181 int
dnode_hold(objset_t * os,uint64_t object,void * tag,dnode_t ** dnp)1182 dnode_hold(objset_t *os, uint64_t object, void *tag, dnode_t **dnp)
1183 {
1184 	return (dnode_hold_impl(os, object, DNODE_MUST_BE_ALLOCATED, tag, dnp));
1185 }
1186 
1187 /*
1188  * Can only add a reference if there is already at least one
1189  * reference on the dnode.  Returns FALSE if unable to add a
1190  * new reference.
1191  */
1192 boolean_t
dnode_add_ref(dnode_t * dn,void * tag)1193 dnode_add_ref(dnode_t *dn, void *tag)
1194 {
1195 	mutex_enter(&dn->dn_mtx);
1196 	if (refcount_is_zero(&dn->dn_holds)) {
1197 		mutex_exit(&dn->dn_mtx);
1198 		return (FALSE);
1199 	}
1200 	VERIFY(1 < refcount_add(&dn->dn_holds, tag));
1201 	mutex_exit(&dn->dn_mtx);
1202 	return (TRUE);
1203 }
1204 
1205 void
dnode_rele(dnode_t * dn,void * tag)1206 dnode_rele(dnode_t *dn, void *tag)
1207 {
1208 	uint64_t refs;
1209 	/* Get while the hold prevents the dnode from moving. */
1210 	dmu_buf_impl_t *db = dn->dn_dbuf;
1211 	dnode_handle_t *dnh = dn->dn_handle;
1212 
1213 	mutex_enter(&dn->dn_mtx);
1214 	refs = refcount_remove(&dn->dn_holds, tag);
1215 	mutex_exit(&dn->dn_mtx);
1216 
1217 	/*
1218 	 * It's unsafe to release the last hold on a dnode by dnode_rele() or
1219 	 * indirectly by dbuf_rele() while relying on the dnode handle to
1220 	 * prevent the dnode from moving, since releasing the last hold could
1221 	 * result in the dnode's parent dbuf evicting its dnode handles. For
1222 	 * that reason anyone calling dnode_rele() or dbuf_rele() without some
1223 	 * other direct or indirect hold on the dnode must first drop the dnode
1224 	 * handle.
1225 	 */
1226 	ASSERT(refs > 0 || dnh->dnh_zrlock.zr_owner != curthread);
1227 
1228 	/* NOTE: the DNODE_DNODE does not have a dn_dbuf */
1229 	if (refs == 0 && db != NULL) {
1230 		/*
1231 		 * Another thread could add a hold to the dnode handle in
1232 		 * dnode_hold_impl() while holding the parent dbuf. Since the
1233 		 * hold on the parent dbuf prevents the handle from being
1234 		 * destroyed, the hold on the handle is OK. We can't yet assert
1235 		 * that the handle has zero references, but that will be
1236 		 * asserted anyway when the handle gets destroyed.
1237 		 */
1238 		dbuf_rele(db, dnh);
1239 	}
1240 }
1241 
1242 void
dnode_setdirty(dnode_t * dn,dmu_tx_t * tx)1243 dnode_setdirty(dnode_t *dn, dmu_tx_t *tx)
1244 {
1245 	objset_t *os = dn->dn_objset;
1246 	uint64_t txg = tx->tx_txg;
1247 
1248 	if (DMU_OBJECT_IS_SPECIAL(dn->dn_object)) {
1249 		dsl_dataset_dirty(os->os_dsl_dataset, tx);
1250 		return;
1251 	}
1252 
1253 	DNODE_VERIFY(dn);
1254 
1255 #ifdef ZFS_DEBUG
1256 	mutex_enter(&dn->dn_mtx);
1257 	ASSERT(dn->dn_phys->dn_type || dn->dn_allocated_txg);
1258 	ASSERT(dn->dn_free_txg == 0 || dn->dn_free_txg >= txg);
1259 	mutex_exit(&dn->dn_mtx);
1260 #endif
1261 
1262 	/*
1263 	 * Determine old uid/gid when necessary
1264 	 */
1265 	dmu_objset_userquota_get_ids(dn, B_TRUE, tx);
1266 
1267 	mutex_enter(&os->os_lock);
1268 
1269 	/*
1270 	 * If we are already marked dirty, we're done.
1271 	 */
1272 	if (list_link_active(&dn->dn_dirty_link[txg & TXG_MASK])) {
1273 		mutex_exit(&os->os_lock);
1274 		return;
1275 	}
1276 
1277 	ASSERT(!refcount_is_zero(&dn->dn_holds) ||
1278 	    !avl_is_empty(&dn->dn_dbufs));
1279 	ASSERT(dn->dn_datablksz != 0);
1280 	ASSERT0(dn->dn_next_bonuslen[txg&TXG_MASK]);
1281 	ASSERT0(dn->dn_next_blksz[txg&TXG_MASK]);
1282 	ASSERT0(dn->dn_next_bonustype[txg&TXG_MASK]);
1283 
1284 	dprintf_ds(os->os_dsl_dataset, "obj=%llu txg=%llu\n",
1285 	    dn->dn_object, txg);
1286 
1287 	if (dn->dn_free_txg > 0 && dn->dn_free_txg <= txg) {
1288 		list_insert_tail(&os->os_free_dnodes[txg&TXG_MASK], dn);
1289 	} else {
1290 		list_insert_tail(&os->os_dirty_dnodes[txg&TXG_MASK], dn);
1291 	}
1292 
1293 	mutex_exit(&os->os_lock);
1294 
1295 	/*
1296 	 * The dnode maintains a hold on its containing dbuf as
1297 	 * long as there are holds on it.  Each instantiated child
1298 	 * dbuf maintains a hold on the dnode.  When the last child
1299 	 * drops its hold, the dnode will drop its hold on the
1300 	 * containing dbuf. We add a "dirty hold" here so that the
1301 	 * dnode will hang around after we finish processing its
1302 	 * children.
1303 	 */
1304 	VERIFY(dnode_add_ref(dn, (void *)(uintptr_t)tx->tx_txg));
1305 
1306 	(void) dbuf_dirty(dn->dn_dbuf, tx);
1307 
1308 	dsl_dataset_dirty(os->os_dsl_dataset, tx);
1309 }
1310 
1311 void
dnode_free(dnode_t * dn,dmu_tx_t * tx)1312 dnode_free(dnode_t *dn, dmu_tx_t *tx)
1313 {
1314 	int txgoff = tx->tx_txg & TXG_MASK;
1315 
1316 	dprintf("dn=%p txg=%llu\n", dn, tx->tx_txg);
1317 
1318 	/* we should be the only holder... hopefully */
1319 	/* ASSERT3U(refcount_count(&dn->dn_holds), ==, 1); */
1320 
1321 	mutex_enter(&dn->dn_mtx);
1322 	if (dn->dn_type == DMU_OT_NONE || dn->dn_free_txg) {
1323 		mutex_exit(&dn->dn_mtx);
1324 		return;
1325 	}
1326 	dn->dn_free_txg = tx->tx_txg;
1327 	mutex_exit(&dn->dn_mtx);
1328 
1329 	/*
1330 	 * If the dnode is already dirty, it needs to be moved from
1331 	 * the dirty list to the free list.
1332 	 */
1333 	mutex_enter(&dn->dn_objset->os_lock);
1334 	if (list_link_active(&dn->dn_dirty_link[txgoff])) {
1335 		list_remove(&dn->dn_objset->os_dirty_dnodes[txgoff], dn);
1336 		list_insert_tail(&dn->dn_objset->os_free_dnodes[txgoff], dn);
1337 		mutex_exit(&dn->dn_objset->os_lock);
1338 	} else {
1339 		mutex_exit(&dn->dn_objset->os_lock);
1340 		dnode_setdirty(dn, tx);
1341 	}
1342 }
1343 
1344 /*
1345  * Try to change the block size for the indicated dnode.  This can only
1346  * succeed if there are no blocks allocated or dirty beyond first block
1347  */
1348 int
dnode_set_blksz(dnode_t * dn,uint64_t size,int ibs,dmu_tx_t * tx)1349 dnode_set_blksz(dnode_t *dn, uint64_t size, int ibs, dmu_tx_t *tx)
1350 {
1351 	dmu_buf_impl_t *db;
1352 	int err;
1353 
1354 	ASSERT3U(size, <=, spa_maxblocksize(dmu_objset_spa(dn->dn_objset)));
1355 	if (size == 0)
1356 		size = SPA_MINBLOCKSIZE;
1357 	else
1358 		size = P2ROUNDUP(size, SPA_MINBLOCKSIZE);
1359 
1360 	if (ibs == dn->dn_indblkshift)
1361 		ibs = 0;
1362 
1363 	if (size >> SPA_MINBLOCKSHIFT == dn->dn_datablkszsec && ibs == 0)
1364 		return (0);
1365 
1366 	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1367 
1368 	/* Check for any allocated blocks beyond the first */
1369 	if (dn->dn_maxblkid != 0)
1370 		goto fail;
1371 
1372 	mutex_enter(&dn->dn_dbufs_mtx);
1373 	for (db = avl_first(&dn->dn_dbufs); db != NULL;
1374 	    db = AVL_NEXT(&dn->dn_dbufs, db)) {
1375 		if (db->db_blkid != 0 && db->db_blkid != DMU_BONUS_BLKID &&
1376 		    db->db_blkid != DMU_SPILL_BLKID) {
1377 			mutex_exit(&dn->dn_dbufs_mtx);
1378 			goto fail;
1379 		}
1380 	}
1381 	mutex_exit(&dn->dn_dbufs_mtx);
1382 
1383 	if (ibs && dn->dn_nlevels != 1)
1384 		goto fail;
1385 
1386 	/* resize the old block */
1387 	err = dbuf_hold_impl(dn, 0, 0, TRUE, FTAG, &db);
1388 	if (err == 0)
1389 		dbuf_new_size(db, size, tx);
1390 	else if (err != ENOENT)
1391 		goto fail;
1392 
1393 	dnode_setdblksz(dn, size);
1394 	dnode_setdirty(dn, tx);
1395 	dn->dn_next_blksz[tx->tx_txg&TXG_MASK] = size;
1396 	if (ibs) {
1397 		dn->dn_indblkshift = ibs;
1398 		dn->dn_next_indblkshift[tx->tx_txg&TXG_MASK] = ibs;
1399 	}
1400 	/* rele after we have fixed the blocksize in the dnode */
1401 	if (db)
1402 		dbuf_rele(db, FTAG);
1403 
1404 	rw_exit(&dn->dn_struct_rwlock);
1405 	return (0);
1406 
1407 fail:
1408 	rw_exit(&dn->dn_struct_rwlock);
1409 	return (SET_ERROR(ENOTSUP));
1410 }
1411 
1412 /* read-holding callers must not rely on the lock being continuously held */
1413 void
dnode_new_blkid(dnode_t * dn,uint64_t blkid,dmu_tx_t * tx,boolean_t have_read)1414 dnode_new_blkid(dnode_t *dn, uint64_t blkid, dmu_tx_t *tx, boolean_t have_read)
1415 {
1416 	uint64_t txgoff = tx->tx_txg & TXG_MASK;
1417 	int epbs, new_nlevels;
1418 	uint64_t sz;
1419 
1420 	ASSERT(blkid != DMU_BONUS_BLKID);
1421 
1422 	ASSERT(have_read ?
1423 	    RW_READ_HELD(&dn->dn_struct_rwlock) :
1424 	    RW_WRITE_HELD(&dn->dn_struct_rwlock));
1425 
1426 	/*
1427 	 * if we have a read-lock, check to see if we need to do any work
1428 	 * before upgrading to a write-lock.
1429 	 */
1430 	if (have_read) {
1431 		if (blkid <= dn->dn_maxblkid)
1432 			return;
1433 
1434 		if (!rw_tryupgrade(&dn->dn_struct_rwlock)) {
1435 			rw_exit(&dn->dn_struct_rwlock);
1436 			rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1437 		}
1438 	}
1439 
1440 	if (blkid <= dn->dn_maxblkid)
1441 		goto out;
1442 
1443 	dn->dn_maxblkid = blkid;
1444 
1445 	/*
1446 	 * Compute the number of levels necessary to support the new maxblkid.
1447 	 */
1448 	new_nlevels = 1;
1449 	epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1450 	for (sz = dn->dn_nblkptr;
1451 	    sz <= blkid && sz >= dn->dn_nblkptr; sz <<= epbs)
1452 		new_nlevels++;
1453 
1454 	if (new_nlevels > dn->dn_nlevels) {
1455 		int old_nlevels = dn->dn_nlevels;
1456 		dmu_buf_impl_t *db;
1457 		list_t *list;
1458 		dbuf_dirty_record_t *new, *dr, *dr_next;
1459 
1460 		dn->dn_nlevels = new_nlevels;
1461 
1462 		ASSERT3U(new_nlevels, >, dn->dn_next_nlevels[txgoff]);
1463 		dn->dn_next_nlevels[txgoff] = new_nlevels;
1464 
1465 		/* dirty the left indirects */
1466 		db = dbuf_hold_level(dn, old_nlevels, 0, FTAG);
1467 		ASSERT(db != NULL);
1468 		new = dbuf_dirty(db, tx);
1469 		dbuf_rele(db, FTAG);
1470 
1471 		/* transfer the dirty records to the new indirect */
1472 		mutex_enter(&dn->dn_mtx);
1473 		mutex_enter(&new->dt.di.dr_mtx);
1474 		list = &dn->dn_dirty_records[txgoff];
1475 		for (dr = list_head(list); dr; dr = dr_next) {
1476 			dr_next = list_next(&dn->dn_dirty_records[txgoff], dr);
1477 			if (dr->dr_dbuf->db_level != new_nlevels-1 &&
1478 			    dr->dr_dbuf->db_blkid != DMU_BONUS_BLKID &&
1479 			    dr->dr_dbuf->db_blkid != DMU_SPILL_BLKID) {
1480 				ASSERT(dr->dr_dbuf->db_level == old_nlevels-1);
1481 				list_remove(&dn->dn_dirty_records[txgoff], dr);
1482 				list_insert_tail(&new->dt.di.dr_children, dr);
1483 				dr->dr_parent = new;
1484 			}
1485 		}
1486 		mutex_exit(&new->dt.di.dr_mtx);
1487 		mutex_exit(&dn->dn_mtx);
1488 	}
1489 
1490 out:
1491 	if (have_read)
1492 		rw_downgrade(&dn->dn_struct_rwlock);
1493 }
1494 
1495 void
dnode_free_range(dnode_t * dn,uint64_t off,uint64_t len,dmu_tx_t * tx)1496 dnode_free_range(dnode_t *dn, uint64_t off, uint64_t len, dmu_tx_t *tx)
1497 {
1498 	dmu_buf_impl_t *db;
1499 	uint64_t blkoff, blkid, nblks;
1500 	int blksz, blkshift, head, tail;
1501 	int trunc = FALSE;
1502 	int epbs;
1503 
1504 	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1505 	blksz = dn->dn_datablksz;
1506 	blkshift = dn->dn_datablkshift;
1507 	epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1508 
1509 	if (len == DMU_OBJECT_END) {
1510 		len = UINT64_MAX - off;
1511 		trunc = TRUE;
1512 	}
1513 
1514 	/*
1515 	 * First, block align the region to free:
1516 	 */
1517 	if (ISP2(blksz)) {
1518 		head = P2NPHASE(off, blksz);
1519 		blkoff = P2PHASE(off, blksz);
1520 		if ((off >> blkshift) > dn->dn_maxblkid)
1521 			goto out;
1522 	} else {
1523 		ASSERT(dn->dn_maxblkid == 0);
1524 		if (off == 0 && len >= blksz) {
1525 			/*
1526 			 * Freeing the whole block; fast-track this request.
1527 			 * Note that we won't dirty any indirect blocks,
1528 			 * which is fine because we will be freeing the entire
1529 			 * file and thus all indirect blocks will be freed
1530 			 * by free_children().
1531 			 */
1532 			blkid = 0;
1533 			nblks = 1;
1534 			goto done;
1535 		} else if (off >= blksz) {
1536 			/* Freeing past end-of-data */
1537 			goto out;
1538 		} else {
1539 			/* Freeing part of the block. */
1540 			head = blksz - off;
1541 			ASSERT3U(head, >, 0);
1542 		}
1543 		blkoff = off;
1544 	}
1545 	/* zero out any partial block data at the start of the range */
1546 	if (head) {
1547 		ASSERT3U(blkoff + head, ==, blksz);
1548 		if (len < head)
1549 			head = len;
1550 		if (dbuf_hold_impl(dn, 0, dbuf_whichblock(dn, off), TRUE,
1551 		    FTAG, &db) == 0) {
1552 			caddr_t data;
1553 
1554 			/* don't dirty if it isn't on disk and isn't dirty */
1555 			if (db->db_last_dirty ||
1556 			    (db->db_blkptr && !BP_IS_HOLE(db->db_blkptr))) {
1557 				rw_exit(&dn->dn_struct_rwlock);
1558 				dmu_buf_will_dirty(&db->db, tx);
1559 				rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1560 				data = db->db.db_data;
1561 				bzero(data + blkoff, head);
1562 			}
1563 			dbuf_rele(db, FTAG);
1564 		}
1565 		off += head;
1566 		len -= head;
1567 	}
1568 
1569 	/* If the range was less than one block, we're done */
1570 	if (len == 0)
1571 		goto out;
1572 
1573 	/* If the remaining range is past end of file, we're done */
1574 	if ((off >> blkshift) > dn->dn_maxblkid)
1575 		goto out;
1576 
1577 	ASSERT(ISP2(blksz));
1578 	if (trunc)
1579 		tail = 0;
1580 	else
1581 		tail = P2PHASE(len, blksz);
1582 
1583 	ASSERT0(P2PHASE(off, blksz));
1584 	/* zero out any partial block data at the end of the range */
1585 	if (tail) {
1586 		if (len < tail)
1587 			tail = len;
1588 		if (dbuf_hold_impl(dn, 0, dbuf_whichblock(dn, off+len),
1589 		    TRUE, FTAG, &db) == 0) {
1590 			/* don't dirty if not on disk and not dirty */
1591 			if (db->db_last_dirty ||
1592 			    (db->db_blkptr && !BP_IS_HOLE(db->db_blkptr))) {
1593 				rw_exit(&dn->dn_struct_rwlock);
1594 				dmu_buf_will_dirty(&db->db, tx);
1595 				rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1596 				bzero(db->db.db_data, tail);
1597 			}
1598 			dbuf_rele(db, FTAG);
1599 		}
1600 		len -= tail;
1601 	}
1602 
1603 	/* If the range did not include a full block, we are done */
1604 	if (len == 0)
1605 		goto out;
1606 
1607 	ASSERT(IS_P2ALIGNED(off, blksz));
1608 	ASSERT(trunc || IS_P2ALIGNED(len, blksz));
1609 	blkid = off >> blkshift;
1610 	nblks = len >> blkshift;
1611 	if (trunc)
1612 		nblks += 1;
1613 
1614 	/*
1615 	 * Dirty the first and last indirect blocks, as they (and/or their
1616 	 * parents) will need to be written out if they were only
1617 	 * partially freed.  Interior indirect blocks will be themselves freed,
1618 	 * by free_children(), so they need not be dirtied.  Note that these
1619 	 * interior blocks have already been prefetched by dmu_tx_hold_free().
1620 	 */
1621 	if (dn->dn_nlevels > 1) {
1622 		uint64_t first, last;
1623 
1624 		first = blkid >> epbs;
1625 		if (db = dbuf_hold_level(dn, 1, first, FTAG)) {
1626 			dmu_buf_will_dirty(&db->db, tx);
1627 			dbuf_rele(db, FTAG);
1628 		}
1629 		if (trunc)
1630 			last = dn->dn_maxblkid >> epbs;
1631 		else
1632 			last = (blkid + nblks - 1) >> epbs;
1633 		if (last > first && (db = dbuf_hold_level(dn, 1, last, FTAG))) {
1634 			dmu_buf_will_dirty(&db->db, tx);
1635 			dbuf_rele(db, FTAG);
1636 		}
1637 	}
1638 
1639 done:
1640 	/*
1641 	 * Add this range to the dnode range list.
1642 	 * We will finish up this free operation in the syncing phase.
1643 	 */
1644 	mutex_enter(&dn->dn_mtx);
1645 	int txgoff = tx->tx_txg & TXG_MASK;
1646 	if (dn->dn_free_ranges[txgoff] == NULL) {
1647 		dn->dn_free_ranges[txgoff] =
1648 		    range_tree_create(NULL, NULL, &dn->dn_mtx);
1649 	}
1650 	range_tree_clear(dn->dn_free_ranges[txgoff], blkid, nblks);
1651 	range_tree_add(dn->dn_free_ranges[txgoff], blkid, nblks);
1652 	dprintf_dnode(dn, "blkid=%llu nblks=%llu txg=%llu\n",
1653 	    blkid, nblks, tx->tx_txg);
1654 	mutex_exit(&dn->dn_mtx);
1655 
1656 	dbuf_free_range(dn, blkid, blkid + nblks - 1, tx);
1657 	dnode_setdirty(dn, tx);
1658 out:
1659 
1660 	rw_exit(&dn->dn_struct_rwlock);
1661 }
1662 
1663 static boolean_t
dnode_spill_freed(dnode_t * dn)1664 dnode_spill_freed(dnode_t *dn)
1665 {
1666 	int i;
1667 
1668 	mutex_enter(&dn->dn_mtx);
1669 	for (i = 0; i < TXG_SIZE; i++) {
1670 		if (dn->dn_rm_spillblk[i] == DN_KILL_SPILLBLK)
1671 			break;
1672 	}
1673 	mutex_exit(&dn->dn_mtx);
1674 	return (i < TXG_SIZE);
1675 }
1676 
1677 /* return TRUE if this blkid was freed in a recent txg, or FALSE if it wasn't */
1678 uint64_t
dnode_block_freed(dnode_t * dn,uint64_t blkid)1679 dnode_block_freed(dnode_t *dn, uint64_t blkid)
1680 {
1681 	void *dp = spa_get_dsl(dn->dn_objset->os_spa);
1682 	int i;
1683 
1684 	if (blkid == DMU_BONUS_BLKID)
1685 		return (FALSE);
1686 
1687 	/*
1688 	 * If we're in the process of opening the pool, dp will not be
1689 	 * set yet, but there shouldn't be anything dirty.
1690 	 */
1691 	if (dp == NULL)
1692 		return (FALSE);
1693 
1694 	if (dn->dn_free_txg)
1695 		return (TRUE);
1696 
1697 	if (blkid == DMU_SPILL_BLKID)
1698 		return (dnode_spill_freed(dn));
1699 
1700 	mutex_enter(&dn->dn_mtx);
1701 	for (i = 0; i < TXG_SIZE; i++) {
1702 		if (dn->dn_free_ranges[i] != NULL &&
1703 		    range_tree_contains(dn->dn_free_ranges[i], blkid, 1))
1704 			break;
1705 	}
1706 	mutex_exit(&dn->dn_mtx);
1707 	return (i < TXG_SIZE);
1708 }
1709 
1710 /* call from syncing context when we actually write/free space for this dnode */
1711 void
dnode_diduse_space(dnode_t * dn,int64_t delta)1712 dnode_diduse_space(dnode_t *dn, int64_t delta)
1713 {
1714 	uint64_t space;
1715 	dprintf_dnode(dn, "dn=%p dnp=%p used=%llu delta=%lld\n",
1716 	    dn, dn->dn_phys,
1717 	    (u_longlong_t)dn->dn_phys->dn_used,
1718 	    (longlong_t)delta);
1719 
1720 	mutex_enter(&dn->dn_mtx);
1721 	space = DN_USED_BYTES(dn->dn_phys);
1722 	if (delta > 0) {
1723 		ASSERT3U(space + delta, >=, space); /* no overflow */
1724 	} else {
1725 		ASSERT3U(space, >=, -delta); /* no underflow */
1726 	}
1727 	space += delta;
1728 	if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_DNODE_BYTES) {
1729 		ASSERT((dn->dn_phys->dn_flags & DNODE_FLAG_USED_BYTES) == 0);
1730 		ASSERT0(P2PHASE(space, 1<<DEV_BSHIFT));
1731 		dn->dn_phys->dn_used = space >> DEV_BSHIFT;
1732 	} else {
1733 		dn->dn_phys->dn_used = space;
1734 		dn->dn_phys->dn_flags |= DNODE_FLAG_USED_BYTES;
1735 	}
1736 	mutex_exit(&dn->dn_mtx);
1737 }
1738 
1739 /*
1740  * Call when we think we're going to write/free space in open context to track
1741  * the amount of memory in use by the currently open txg.
1742  */
1743 void
dnode_willuse_space(dnode_t * dn,int64_t space,dmu_tx_t * tx)1744 dnode_willuse_space(dnode_t *dn, int64_t space, dmu_tx_t *tx)
1745 {
1746 	objset_t *os = dn->dn_objset;
1747 	dsl_dataset_t *ds = os->os_dsl_dataset;
1748 	int64_t aspace = spa_get_asize(os->os_spa, space);
1749 
1750 	if (ds != NULL) {
1751 		dsl_dir_willuse_space(ds->ds_dir, aspace, tx);
1752 		dsl_pool_dirty_space(dmu_tx_pool(tx), space, tx);
1753 	}
1754 
1755 	dmu_tx_willuse_space(tx, aspace);
1756 }
1757 
1758 /*
1759  * Scans a block at the indicated "level" looking for a hole or data,
1760  * depending on 'flags'.
1761  *
1762  * If level > 0, then we are scanning an indirect block looking at its
1763  * pointers.  If level == 0, then we are looking at a block of dnodes.
1764  *
1765  * If we don't find what we are looking for in the block, we return ESRCH.
1766  * Otherwise, return with *offset pointing to the beginning (if searching
1767  * forwards) or end (if searching backwards) of the range covered by the
1768  * block pointer we matched on (or dnode).
1769  *
1770  * The basic search algorithm used below by dnode_next_offset() is to
1771  * use this function to search up the block tree (widen the search) until
1772  * we find something (i.e., we don't return ESRCH) and then search back
1773  * down the tree (narrow the search) until we reach our original search
1774  * level.
1775  */
1776 static int
dnode_next_offset_level(dnode_t * dn,int flags,uint64_t * offset,int lvl,uint64_t blkfill,uint64_t txg)1777 dnode_next_offset_level(dnode_t *dn, int flags, uint64_t *offset,
1778 	int lvl, uint64_t blkfill, uint64_t txg)
1779 {
1780 	dmu_buf_impl_t *db = NULL;
1781 	void *data = NULL;
1782 	uint64_t epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
1783 	uint64_t epb = 1ULL << epbs;
1784 	uint64_t minfill, maxfill;
1785 	boolean_t hole;
1786 	int i, inc, error, span;
1787 
1788 	dprintf("probing object %llu offset %llx level %d of %u\n",
1789 	    dn->dn_object, *offset, lvl, dn->dn_phys->dn_nlevels);
1790 
1791 	hole = ((flags & DNODE_FIND_HOLE) != 0);
1792 	inc = (flags & DNODE_FIND_BACKWARDS) ? -1 : 1;
1793 	ASSERT(txg == 0 || !hole);
1794 
1795 	if (lvl == dn->dn_phys->dn_nlevels) {
1796 		error = 0;
1797 		epb = dn->dn_phys->dn_nblkptr;
1798 		data = dn->dn_phys->dn_blkptr;
1799 	} else {
1800 		uint64_t blkid = dbuf_whichblock(dn, *offset) >> (epbs * lvl);
1801 		error = dbuf_hold_impl(dn, lvl, blkid, TRUE, FTAG, &db);
1802 		if (error) {
1803 			if (error != ENOENT)
1804 				return (error);
1805 			if (hole)
1806 				return (0);
1807 			/*
1808 			 * This can only happen when we are searching up
1809 			 * the block tree for data.  We don't really need to
1810 			 * adjust the offset, as we will just end up looking
1811 			 * at the pointer to this block in its parent, and its
1812 			 * going to be unallocated, so we will skip over it.
1813 			 */
1814 			return (SET_ERROR(ESRCH));
1815 		}
1816 		error = dbuf_read(db, NULL, DB_RF_CANFAIL | DB_RF_HAVESTRUCT);
1817 		if (error) {
1818 			dbuf_rele(db, FTAG);
1819 			return (error);
1820 		}
1821 		data = db->db.db_data;
1822 	}
1823 
1824 
1825 	if (db != NULL && txg != 0 && (db->db_blkptr == NULL ||
1826 	    db->db_blkptr->blk_birth <= txg ||
1827 	    BP_IS_HOLE(db->db_blkptr))) {
1828 		/*
1829 		 * This can only happen when we are searching up the tree
1830 		 * and these conditions mean that we need to keep climbing.
1831 		 */
1832 		error = SET_ERROR(ESRCH);
1833 	} else if (lvl == 0) {
1834 		dnode_phys_t *dnp = data;
1835 		span = DNODE_SHIFT;
1836 		ASSERT(dn->dn_type == DMU_OT_DNODE);
1837 
1838 		for (i = (*offset >> span) & (blkfill - 1);
1839 		    i >= 0 && i < blkfill; i += inc) {
1840 			if ((dnp[i].dn_type == DMU_OT_NONE) == hole)
1841 				break;
1842 			*offset += (1ULL << span) * inc;
1843 		}
1844 		if (i < 0 || i == blkfill)
1845 			error = SET_ERROR(ESRCH);
1846 	} else {
1847 		blkptr_t *bp = data;
1848 		uint64_t start = *offset;
1849 		span = (lvl - 1) * epbs + dn->dn_datablkshift;
1850 		minfill = 0;
1851 		maxfill = blkfill << ((lvl - 1) * epbs);
1852 
1853 		if (hole)
1854 			maxfill--;
1855 		else
1856 			minfill++;
1857 
1858 		*offset = *offset >> span;
1859 		for (i = BF64_GET(*offset, 0, epbs);
1860 		    i >= 0 && i < epb; i += inc) {
1861 			if (BP_GET_FILL(&bp[i]) >= minfill &&
1862 			    BP_GET_FILL(&bp[i]) <= maxfill &&
1863 			    (hole || bp[i].blk_birth > txg))
1864 				break;
1865 			if (inc > 0 || *offset > 0)
1866 				*offset += inc;
1867 		}
1868 		*offset = *offset << span;
1869 		if (inc < 0) {
1870 			/* traversing backwards; position offset at the end */
1871 			ASSERT3U(*offset, <=, start);
1872 			*offset = MIN(*offset + (1ULL << span) - 1, start);
1873 		} else if (*offset < start) {
1874 			*offset = start;
1875 		}
1876 		if (i < 0 || i >= epb)
1877 			error = SET_ERROR(ESRCH);
1878 	}
1879 
1880 	if (db)
1881 		dbuf_rele(db, FTAG);
1882 
1883 	return (error);
1884 }
1885 
1886 /*
1887  * Find the next hole, data, or sparse region at or after *offset.
1888  * The value 'blkfill' tells us how many items we expect to find
1889  * in an L0 data block; this value is 1 for normal objects,
1890  * DNODES_PER_BLOCK for the meta dnode, and some fraction of
1891  * DNODES_PER_BLOCK when searching for sparse regions thereof.
1892  *
1893  * Examples:
1894  *
1895  * dnode_next_offset(dn, flags, offset, 1, 1, 0);
1896  *	Finds the next/previous hole/data in a file.
1897  *	Used in dmu_offset_next().
1898  *
1899  * dnode_next_offset(mdn, flags, offset, 0, DNODES_PER_BLOCK, txg);
1900  *	Finds the next free/allocated dnode an objset's meta-dnode.
1901  *	Only finds objects that have new contents since txg (ie.
1902  *	bonus buffer changes and content removal are ignored).
1903  *	Used in dmu_object_next().
1904  *
1905  * dnode_next_offset(mdn, DNODE_FIND_HOLE, offset, 2, DNODES_PER_BLOCK >> 2, 0);
1906  *	Finds the next L2 meta-dnode bp that's at most 1/4 full.
1907  *	Used in dmu_object_alloc().
1908  */
1909 int
dnode_next_offset(dnode_t * dn,int flags,uint64_t * offset,int minlvl,uint64_t blkfill,uint64_t txg)1910 dnode_next_offset(dnode_t *dn, int flags, uint64_t *offset,
1911     int minlvl, uint64_t blkfill, uint64_t txg)
1912 {
1913 	uint64_t initial_offset = *offset;
1914 	int lvl, maxlvl;
1915 	int error = 0;
1916 
1917 	if (!(flags & DNODE_FIND_HAVELOCK))
1918 		rw_enter(&dn->dn_struct_rwlock, RW_READER);
1919 
1920 	if (dn->dn_phys->dn_nlevels == 0) {
1921 		error = SET_ERROR(ESRCH);
1922 		goto out;
1923 	}
1924 
1925 	if (dn->dn_datablkshift == 0) {
1926 		if (*offset < dn->dn_datablksz) {
1927 			if (flags & DNODE_FIND_HOLE)
1928 				*offset = dn->dn_datablksz;
1929 		} else {
1930 			error = SET_ERROR(ESRCH);
1931 		}
1932 		goto out;
1933 	}
1934 
1935 	maxlvl = dn->dn_phys->dn_nlevels;
1936 
1937 	for (lvl = minlvl; lvl <= maxlvl; lvl++) {
1938 		error = dnode_next_offset_level(dn,
1939 		    flags, offset, lvl, blkfill, txg);
1940 		if (error != ESRCH)
1941 			break;
1942 	}
1943 
1944 	while (error == 0 && --lvl >= minlvl) {
1945 		error = dnode_next_offset_level(dn,
1946 		    flags, offset, lvl, blkfill, txg);
1947 	}
1948 
1949 	/*
1950 	 * There's always a "virtual hole" at the end of the object, even
1951 	 * if all BP's which physically exist are non-holes.
1952 	 */
1953 	if ((flags & DNODE_FIND_HOLE) && error == ESRCH && txg == 0 &&
1954 	    minlvl == 1 && blkfill == 1 && !(flags & DNODE_FIND_BACKWARDS)) {
1955 		error = 0;
1956 	}
1957 
1958 	if (error == 0 && (flags & DNODE_FIND_BACKWARDS ?
1959 	    initial_offset < *offset : initial_offset > *offset))
1960 		error = SET_ERROR(ESRCH);
1961 out:
1962 	if (!(flags & DNODE_FIND_HAVELOCK))
1963 		rw_exit(&dn->dn_struct_rwlock);
1964 
1965 	return (error);
1966 }
1967