xref: /netbsd-src/external/cddl/osnet/dist/uts/common/fs/zfs/dnode.c (revision 65d6baaaebacadd0dfed73fc006dfce41a386bfa)
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, 2016 by Delphix. All rights reserved.
24  * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved.
25  * Copyright (c) 2014 Integros [integros.com]
26  */
27 
28 #include <sys/zfs_context.h>
29 #include <sys/dbuf.h>
30 #include <sys/dnode.h>
31 #include <sys/dmu.h>
32 #include <sys/dmu_impl.h>
33 #include <sys/dmu_tx.h>
34 #include <sys/dmu_objset.h>
35 #include <sys/dsl_dir.h>
36 #include <sys/dsl_dataset.h>
37 #include <sys/spa.h>
38 #include <sys/zio.h>
39 #include <sys/dmu_zfetch.h>
40 #include <sys/range_tree.h>
41 
42 static kmem_cache_t *dnode_cache;
43 /*
44  * Define DNODE_STATS to turn on statistic gathering. By default, it is only
45  * turned on when DEBUG is also defined.
46  */
47 #ifdef	DEBUG
48 #define	DNODE_STATS
49 #endif	/* DEBUG */
50 
51 #ifdef	DNODE_STATS
52 #define	DNODE_STAT_ADD(stat)			((stat)++)
53 #else
54 #define	DNODE_STAT_ADD(stat)			/* nothing */
55 #endif	/* DNODE_STATS */
56 
57 static dnode_phys_t dnode_phys_zero;
58 
59 int zfs_default_bs = SPA_MINBLOCKSHIFT;
60 int zfs_default_ibs = DN_MAX_INDBLKSHIFT;
61 
62 #ifdef illumos
63 static kmem_cbrc_t dnode_move(void *, void *, size_t, void *);
64 #endif
65 
66 static int
dbuf_compare(const void * x1,const void * x2)67 dbuf_compare(const void *x1, const void *x2)
68 {
69 	const dmu_buf_impl_t *d1 = x1;
70 	const dmu_buf_impl_t *d2 = x2;
71 
72 	if (d1->db_level < d2->db_level) {
73 		return (-1);
74 	}
75 	if (d1->db_level > d2->db_level) {
76 		return (1);
77 	}
78 
79 	if (d1->db_blkid < d2->db_blkid) {
80 		return (-1);
81 	}
82 	if (d1->db_blkid > d2->db_blkid) {
83 		return (1);
84 	}
85 
86 	if (d1->db_state == DB_SEARCH) {
87 		ASSERT3S(d2->db_state, !=, DB_SEARCH);
88 		return (-1);
89 	} else if (d2->db_state == DB_SEARCH) {
90 		ASSERT3S(d1->db_state, !=, DB_SEARCH);
91 		return (1);
92 	}
93 
94 	if ((uintptr_t)d1 < (uintptr_t)d2) {
95 		return (-1);
96 	}
97 	if ((uintptr_t)d1 > (uintptr_t)d2) {
98 		return (1);
99 	}
100 	return (0);
101 }
102 
103 /* ARGSUSED */
104 static int
dnode_cons(void * arg,void * unused,int kmflag)105 dnode_cons(void *arg, void *unused, int kmflag)
106 {
107 	dnode_t *dn = arg;
108 	int i;
109 
110 	rw_init(&dn->dn_struct_rwlock, NULL, RW_DEFAULT, NULL);
111 	mutex_init(&dn->dn_mtx, NULL, MUTEX_DEFAULT, NULL);
112 	mutex_init(&dn->dn_dbufs_mtx, NULL, MUTEX_DEFAULT, NULL);
113 	cv_init(&dn->dn_notxholds, NULL, CV_DEFAULT, NULL);
114 
115 	/*
116 	 * Every dbuf has a reference, and dropping a tracked reference is
117 	 * O(number of references), so don't track dn_holds.
118 	 */
119 	refcount_create_untracked(&dn->dn_holds);
120 	refcount_create(&dn->dn_tx_holds);
121 	list_link_init(&dn->dn_link);
122 
123 	bzero(&dn->dn_next_nblkptr[0], sizeof (dn->dn_next_nblkptr));
124 	bzero(&dn->dn_next_nlevels[0], sizeof (dn->dn_next_nlevels));
125 	bzero(&dn->dn_next_indblkshift[0], sizeof (dn->dn_next_indblkshift));
126 	bzero(&dn->dn_next_bonustype[0], sizeof (dn->dn_next_bonustype));
127 	bzero(&dn->dn_rm_spillblk[0], sizeof (dn->dn_rm_spillblk));
128 	bzero(&dn->dn_next_bonuslen[0], sizeof (dn->dn_next_bonuslen));
129 	bzero(&dn->dn_next_blksz[0], sizeof (dn->dn_next_blksz));
130 
131 	for (i = 0; i < TXG_SIZE; i++) {
132 		list_link_init(&dn->dn_dirty_link[i]);
133 		dn->dn_free_ranges[i] = NULL;
134 		list_create(&dn->dn_dirty_records[i],
135 		    sizeof (dbuf_dirty_record_t),
136 		    offsetof(dbuf_dirty_record_t, dr_dirty_node));
137 	}
138 
139 	dn->dn_allocated_txg = 0;
140 	dn->dn_free_txg = 0;
141 	dn->dn_assigned_txg = 0;
142 	dn->dn_dirtyctx = 0;
143 	dn->dn_dirtyctx_firstset = NULL;
144 	dn->dn_bonus = NULL;
145 	dn->dn_have_spill = B_FALSE;
146 	dn->dn_zio = NULL;
147 	dn->dn_oldused = 0;
148 	dn->dn_oldflags = 0;
149 	dn->dn_olduid = 0;
150 	dn->dn_oldgid = 0;
151 	dn->dn_newuid = 0;
152 	dn->dn_newgid = 0;
153 	dn->dn_id_flags = 0;
154 
155 	dn->dn_dbufs_count = 0;
156 	avl_create(&dn->dn_dbufs, dbuf_compare, sizeof (dmu_buf_impl_t),
157 	    offsetof(dmu_buf_impl_t, db_link));
158 
159 	dn->dn_moved = 0;
160 	POINTER_INVALIDATE(&dn->dn_objset);
161 	return (0);
162 }
163 
164 /* ARGSUSED */
165 static void
dnode_dest(void * arg,void * unused)166 dnode_dest(void *arg, void *unused)
167 {
168 	int i;
169 	dnode_t *dn = arg;
170 
171 	rw_destroy(&dn->dn_struct_rwlock);
172 	mutex_destroy(&dn->dn_mtx);
173 	mutex_destroy(&dn->dn_dbufs_mtx);
174 	cv_destroy(&dn->dn_notxholds);
175 	refcount_destroy(&dn->dn_holds);
176 	refcount_destroy(&dn->dn_tx_holds);
177 	ASSERT(!list_link_active(&dn->dn_link));
178 
179 	for (i = 0; i < TXG_SIZE; i++) {
180 		ASSERT(!list_link_active(&dn->dn_dirty_link[i]));
181 		ASSERT3P(dn->dn_free_ranges[i], ==, NULL);
182 		list_destroy(&dn->dn_dirty_records[i]);
183 		ASSERT0(dn->dn_next_nblkptr[i]);
184 		ASSERT0(dn->dn_next_nlevels[i]);
185 		ASSERT0(dn->dn_next_indblkshift[i]);
186 		ASSERT0(dn->dn_next_bonustype[i]);
187 		ASSERT0(dn->dn_rm_spillblk[i]);
188 		ASSERT0(dn->dn_next_bonuslen[i]);
189 		ASSERT0(dn->dn_next_blksz[i]);
190 	}
191 
192 	ASSERT0(dn->dn_allocated_txg);
193 	ASSERT0(dn->dn_free_txg);
194 	ASSERT0(dn->dn_assigned_txg);
195 	ASSERT0(dn->dn_dirtyctx);
196 	ASSERT3P(dn->dn_dirtyctx_firstset, ==, NULL);
197 	ASSERT3P(dn->dn_bonus, ==, NULL);
198 	ASSERT(!dn->dn_have_spill);
199 	ASSERT3P(dn->dn_zio, ==, NULL);
200 	ASSERT0(dn->dn_oldused);
201 	ASSERT0(dn->dn_oldflags);
202 	ASSERT0(dn->dn_olduid);
203 	ASSERT0(dn->dn_oldgid);
204 	ASSERT0(dn->dn_newuid);
205 	ASSERT0(dn->dn_newgid);
206 	ASSERT0(dn->dn_id_flags);
207 
208 	ASSERT0(dn->dn_dbufs_count);
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;
409 
410 	dn = kmem_cache_alloc(dnode_cache, KM_SLEEP);
411 	ASSERT(!POINTER_IS_VALID(dn->dn_objset));
412 	dn->dn_moved = 0;
413 
414 	/*
415 	 * Defer setting dn_objset until the dnode is ready to be a candidate
416 	 * for the dnode_move() callback.
417 	 */
418 	dn->dn_object = object;
419 	dn->dn_dbuf = db;
420 	dn->dn_handle = dnh;
421 	dn->dn_phys = dnp;
422 
423 	if (dnp->dn_datablkszsec) {
424 		dnode_setdblksz(dn, dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT);
425 	} else {
426 		dn->dn_datablksz = 0;
427 		dn->dn_datablkszsec = 0;
428 		dn->dn_datablkshift = 0;
429 	}
430 	dn->dn_indblkshift = dnp->dn_indblkshift;
431 	dn->dn_nlevels = dnp->dn_nlevels;
432 	dn->dn_type = dnp->dn_type;
433 	dn->dn_nblkptr = dnp->dn_nblkptr;
434 	dn->dn_checksum = dnp->dn_checksum;
435 	dn->dn_compress = dnp->dn_compress;
436 	dn->dn_bonustype = dnp->dn_bonustype;
437 	dn->dn_bonuslen = dnp->dn_bonuslen;
438 	dn->dn_maxblkid = dnp->dn_maxblkid;
439 	dn->dn_have_spill = ((dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) != 0);
440 	dn->dn_id_flags = 0;
441 
442 	dmu_zfetch_init(&dn->dn_zfetch, dn);
443 
444 	ASSERT(DMU_OT_IS_VALID(dn->dn_phys->dn_type));
445 
446 	mutex_enter(&os->os_lock);
447 	if (dnh->dnh_dnode != NULL) {
448 		/* Lost the allocation race. */
449 		mutex_exit(&os->os_lock);
450 #ifdef __NetBSD__
451 		dmu_zfetch_fini(&dn->dn_zfetch);
452 #endif
453 		kmem_cache_free(dnode_cache, dn);
454 		return (dnh->dnh_dnode);
455 	}
456 
457 	/*
458 	 * Exclude special dnodes from os_dnodes so an empty os_dnodes
459 	 * signifies that the special dnodes have no references from
460 	 * their children (the entries in os_dnodes).  This allows
461 	 * dnode_destroy() to easily determine if the last child has
462 	 * been removed and then complete eviction of the objset.
463 	 */
464 	if (!DMU_OBJECT_IS_SPECIAL(object))
465 		list_insert_head(&os->os_dnodes, dn);
466 	membar_producer();
467 
468 	/*
469 	 * Everything else must be valid before assigning dn_objset
470 	 * makes the dnode eligible for dnode_move().
471 	 */
472 	dn->dn_objset = os;
473 
474 	dnh->dnh_dnode = dn;
475 	mutex_exit(&os->os_lock);
476 
477 	arc_space_consume(sizeof (dnode_t), ARC_SPACE_OTHER);
478 	return (dn);
479 }
480 
481 /*
482  * Caller must be holding the dnode handle, which is released upon return.
483  */
484 static void
dnode_destroy(dnode_t * dn)485 dnode_destroy(dnode_t *dn)
486 {
487 	objset_t *os = dn->dn_objset;
488 	boolean_t complete_os_eviction = B_FALSE;
489 
490 	ASSERT((dn->dn_id_flags & DN_ID_NEW_EXIST) == 0);
491 
492 	mutex_enter(&os->os_lock);
493 	POINTER_INVALIDATE(&dn->dn_objset);
494 	if (!DMU_OBJECT_IS_SPECIAL(dn->dn_object)) {
495 		list_remove(&os->os_dnodes, dn);
496 		complete_os_eviction =
497 		    list_is_empty(&os->os_dnodes) &&
498 		    list_link_active(&os->os_evicting_node);
499 	}
500 	mutex_exit(&os->os_lock);
501 
502 	/* the dnode can no longer move, so we can release the handle */
503 	zrl_remove(&dn->dn_handle->dnh_zrlock);
504 
505 	dn->dn_allocated_txg = 0;
506 	dn->dn_free_txg = 0;
507 	dn->dn_assigned_txg = 0;
508 
509 	dn->dn_dirtyctx = 0;
510 	if (dn->dn_dirtyctx_firstset != NULL) {
511 		kmem_free(dn->dn_dirtyctx_firstset, 1);
512 		dn->dn_dirtyctx_firstset = NULL;
513 	}
514 	if (dn->dn_bonus != NULL) {
515 		mutex_enter(&dn->dn_bonus->db_mtx);
516 		dbuf_destroy(dn->dn_bonus);
517 		dn->dn_bonus = NULL;
518 	}
519 	dn->dn_zio = NULL;
520 
521 	dn->dn_have_spill = B_FALSE;
522 	dn->dn_oldused = 0;
523 	dn->dn_oldflags = 0;
524 	dn->dn_olduid = 0;
525 	dn->dn_oldgid = 0;
526 	dn->dn_newuid = 0;
527 	dn->dn_newgid = 0;
528 	dn->dn_id_flags = 0;
529 
530 	dmu_zfetch_fini(&dn->dn_zfetch);
531 	kmem_cache_free(dnode_cache, dn);
532 	arc_space_return(sizeof (dnode_t), ARC_SPACE_OTHER);
533 
534 	if (complete_os_eviction)
535 		dmu_objset_evict_done(os);
536 }
537 
538 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)539 dnode_allocate(dnode_t *dn, dmu_object_type_t ot, int blocksize, int ibs,
540     dmu_object_type_t bonustype, int bonuslen, dmu_tx_t *tx)
541 {
542 	int i;
543 
544 	ASSERT3U(blocksize, <=,
545 	    spa_maxblocksize(dmu_objset_spa(dn->dn_objset)));
546 	if (blocksize == 0)
547 		blocksize = 1 << zfs_default_bs;
548 	else
549 		blocksize = P2ROUNDUP(blocksize, SPA_MINBLOCKSIZE);
550 
551 	if (ibs == 0)
552 		ibs = zfs_default_ibs;
553 
554 	ibs = MIN(MAX(ibs, DN_MIN_INDBLKSHIFT), DN_MAX_INDBLKSHIFT);
555 
556 	dprintf("os=%p obj=%llu txg=%llu blocksize=%d ibs=%d\n", dn->dn_objset,
557 	    dn->dn_object, tx->tx_txg, blocksize, ibs);
558 
559 	ASSERT(dn->dn_type == DMU_OT_NONE);
560 	ASSERT(bcmp(dn->dn_phys, &dnode_phys_zero, sizeof (dnode_phys_t)) == 0);
561 	ASSERT(dn->dn_phys->dn_type == DMU_OT_NONE);
562 	ASSERT(ot != DMU_OT_NONE);
563 	ASSERT(DMU_OT_IS_VALID(ot));
564 	ASSERT((bonustype == DMU_OT_NONE && bonuslen == 0) ||
565 	    (bonustype == DMU_OT_SA && bonuslen == 0) ||
566 	    (bonustype != DMU_OT_NONE && bonuslen != 0));
567 	ASSERT(DMU_OT_IS_VALID(bonustype));
568 	ASSERT3U(bonuslen, <=, DN_MAX_BONUSLEN);
569 	ASSERT(dn->dn_type == DMU_OT_NONE);
570 	ASSERT0(dn->dn_maxblkid);
571 	ASSERT0(dn->dn_allocated_txg);
572 	ASSERT0(dn->dn_assigned_txg);
573 	ASSERT(refcount_is_zero(&dn->dn_tx_holds));
574 	ASSERT3U(refcount_count(&dn->dn_holds), <=, 1);
575 	ASSERT(avl_is_empty(&dn->dn_dbufs));
576 
577 	for (i = 0; i < TXG_SIZE; i++) {
578 		ASSERT0(dn->dn_next_nblkptr[i]);
579 		ASSERT0(dn->dn_next_nlevels[i]);
580 		ASSERT0(dn->dn_next_indblkshift[i]);
581 		ASSERT0(dn->dn_next_bonuslen[i]);
582 		ASSERT0(dn->dn_next_bonustype[i]);
583 		ASSERT0(dn->dn_rm_spillblk[i]);
584 		ASSERT0(dn->dn_next_blksz[i]);
585 		ASSERT(!list_link_active(&dn->dn_dirty_link[i]));
586 		ASSERT3P(list_head(&dn->dn_dirty_records[i]), ==, NULL);
587 		ASSERT3P(dn->dn_free_ranges[i], ==, NULL);
588 	}
589 
590 	dn->dn_type = ot;
591 	dnode_setdblksz(dn, blocksize);
592 	dn->dn_indblkshift = ibs;
593 	dn->dn_nlevels = 1;
594 	if (bonustype == DMU_OT_SA) /* Maximize bonus space for SA */
595 		dn->dn_nblkptr = 1;
596 	else
597 		dn->dn_nblkptr = 1 +
598 		    ((DN_MAX_BONUSLEN - bonuslen) >> SPA_BLKPTRSHIFT);
599 	dn->dn_bonustype = bonustype;
600 	dn->dn_bonuslen = bonuslen;
601 	dn->dn_checksum = ZIO_CHECKSUM_INHERIT;
602 	dn->dn_compress = ZIO_COMPRESS_INHERIT;
603 	dn->dn_dirtyctx = 0;
604 
605 	dn->dn_free_txg = 0;
606 	if (dn->dn_dirtyctx_firstset) {
607 		kmem_free(dn->dn_dirtyctx_firstset, 1);
608 		dn->dn_dirtyctx_firstset = NULL;
609 	}
610 
611 	dn->dn_allocated_txg = tx->tx_txg;
612 	dn->dn_id_flags = 0;
613 
614 	dnode_setdirty(dn, tx);
615 	dn->dn_next_indblkshift[tx->tx_txg & TXG_MASK] = ibs;
616 	dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = dn->dn_bonuslen;
617 	dn->dn_next_bonustype[tx->tx_txg & TXG_MASK] = dn->dn_bonustype;
618 	dn->dn_next_blksz[tx->tx_txg & TXG_MASK] = dn->dn_datablksz;
619 }
620 
621 void
dnode_reallocate(dnode_t * dn,dmu_object_type_t ot,int blocksize,dmu_object_type_t bonustype,int bonuslen,dmu_tx_t * tx)622 dnode_reallocate(dnode_t *dn, dmu_object_type_t ot, int blocksize,
623     dmu_object_type_t bonustype, int bonuslen, dmu_tx_t *tx)
624 {
625 	int nblkptr;
626 
627 	ASSERT3U(blocksize, >=, SPA_MINBLOCKSIZE);
628 	ASSERT3U(blocksize, <=,
629 	    spa_maxblocksize(dmu_objset_spa(dn->dn_objset)));
630 	ASSERT0(blocksize % SPA_MINBLOCKSIZE);
631 	ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT || dmu_tx_private_ok(tx));
632 	ASSERT(tx->tx_txg != 0);
633 	ASSERT((bonustype == DMU_OT_NONE && bonuslen == 0) ||
634 	    (bonustype != DMU_OT_NONE && bonuslen != 0) ||
635 	    (bonustype == DMU_OT_SA && bonuslen == 0));
636 	ASSERT(DMU_OT_IS_VALID(bonustype));
637 	ASSERT3U(bonuslen, <=, DN_MAX_BONUSLEN);
638 
639 	/* clean up any unreferenced dbufs */
640 	dnode_evict_dbufs(dn);
641 
642 	dn->dn_id_flags = 0;
643 
644 	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
645 	dnode_setdirty(dn, tx);
646 	if (dn->dn_datablksz != blocksize) {
647 		/* change blocksize */
648 		ASSERT(dn->dn_maxblkid == 0 &&
649 		    (BP_IS_HOLE(&dn->dn_phys->dn_blkptr[0]) ||
650 		    dnode_block_freed(dn, 0)));
651 		dnode_setdblksz(dn, blocksize);
652 		dn->dn_next_blksz[tx->tx_txg&TXG_MASK] = blocksize;
653 	}
654 	if (dn->dn_bonuslen != bonuslen)
655 		dn->dn_next_bonuslen[tx->tx_txg&TXG_MASK] = bonuslen;
656 
657 	if (bonustype == DMU_OT_SA) /* Maximize bonus space for SA */
658 		nblkptr = 1;
659 	else
660 		nblkptr = 1 + ((DN_MAX_BONUSLEN - bonuslen) >> SPA_BLKPTRSHIFT);
661 	if (dn->dn_bonustype != bonustype)
662 		dn->dn_next_bonustype[tx->tx_txg&TXG_MASK] = bonustype;
663 	if (dn->dn_nblkptr != nblkptr)
664 		dn->dn_next_nblkptr[tx->tx_txg&TXG_MASK] = nblkptr;
665 	if (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
666 		dbuf_rm_spill(dn, tx);
667 		dnode_rm_spill(dn, tx);
668 	}
669 	rw_exit(&dn->dn_struct_rwlock);
670 
671 	/* change type */
672 	dn->dn_type = ot;
673 
674 	/* change bonus size and type */
675 	mutex_enter(&dn->dn_mtx);
676 	dn->dn_bonustype = bonustype;
677 	dn->dn_bonuslen = bonuslen;
678 	dn->dn_nblkptr = nblkptr;
679 	dn->dn_checksum = ZIO_CHECKSUM_INHERIT;
680 	dn->dn_compress = ZIO_COMPRESS_INHERIT;
681 	ASSERT3U(dn->dn_nblkptr, <=, DN_MAX_NBLKPTR);
682 
683 	/* fix up the bonus db_size */
684 	if (dn->dn_bonus) {
685 		dn->dn_bonus->db.db_size =
686 		    DN_MAX_BONUSLEN - (dn->dn_nblkptr-1) * sizeof (blkptr_t);
687 		ASSERT(dn->dn_bonuslen <= dn->dn_bonus->db.db_size);
688 	}
689 
690 	dn->dn_allocated_txg = tx->tx_txg;
691 	mutex_exit(&dn->dn_mtx);
692 }
693 
694 #ifdef	DNODE_STATS
695 static struct {
696 	uint64_t dms_dnode_invalid;
697 	uint64_t dms_dnode_recheck1;
698 	uint64_t dms_dnode_recheck2;
699 	uint64_t dms_dnode_special;
700 	uint64_t dms_dnode_handle;
701 	uint64_t dms_dnode_rwlock;
702 	uint64_t dms_dnode_active;
703 } dnode_move_stats;
704 #endif	/* DNODE_STATS */
705 
706 static void
dnode_move_impl(dnode_t * odn,dnode_t * ndn)707 dnode_move_impl(dnode_t *odn, dnode_t *ndn)
708 {
709 	int i;
710 
711 	ASSERT(!RW_LOCK_HELD(&odn->dn_struct_rwlock));
712 	ASSERT(MUTEX_NOT_HELD(&odn->dn_mtx));
713 	ASSERT(MUTEX_NOT_HELD(&odn->dn_dbufs_mtx));
714 	ASSERT(!RW_LOCK_HELD(&odn->dn_zfetch.zf_rwlock));
715 
716 	/* Copy fields. */
717 	ndn->dn_objset = odn->dn_objset;
718 	ndn->dn_object = odn->dn_object;
719 	ndn->dn_dbuf = odn->dn_dbuf;
720 	ndn->dn_handle = odn->dn_handle;
721 	ndn->dn_phys = odn->dn_phys;
722 	ndn->dn_type = odn->dn_type;
723 	ndn->dn_bonuslen = odn->dn_bonuslen;
724 	ndn->dn_bonustype = odn->dn_bonustype;
725 	ndn->dn_nblkptr = odn->dn_nblkptr;
726 	ndn->dn_checksum = odn->dn_checksum;
727 	ndn->dn_compress = odn->dn_compress;
728 	ndn->dn_nlevels = odn->dn_nlevels;
729 	ndn->dn_indblkshift = odn->dn_indblkshift;
730 	ndn->dn_datablkshift = odn->dn_datablkshift;
731 	ndn->dn_datablkszsec = odn->dn_datablkszsec;
732 	ndn->dn_datablksz = odn->dn_datablksz;
733 	ndn->dn_maxblkid = odn->dn_maxblkid;
734 	bcopy(&odn->dn_next_nblkptr[0], &ndn->dn_next_nblkptr[0],
735 	    sizeof (odn->dn_next_nblkptr));
736 	bcopy(&odn->dn_next_nlevels[0], &ndn->dn_next_nlevels[0],
737 	    sizeof (odn->dn_next_nlevels));
738 	bcopy(&odn->dn_next_indblkshift[0], &ndn->dn_next_indblkshift[0],
739 	    sizeof (odn->dn_next_indblkshift));
740 	bcopy(&odn->dn_next_bonustype[0], &ndn->dn_next_bonustype[0],
741 	    sizeof (odn->dn_next_bonustype));
742 	bcopy(&odn->dn_rm_spillblk[0], &ndn->dn_rm_spillblk[0],
743 	    sizeof (odn->dn_rm_spillblk));
744 	bcopy(&odn->dn_next_bonuslen[0], &ndn->dn_next_bonuslen[0],
745 	    sizeof (odn->dn_next_bonuslen));
746 	bcopy(&odn->dn_next_blksz[0], &ndn->dn_next_blksz[0],
747 	    sizeof (odn->dn_next_blksz));
748 	for (i = 0; i < TXG_SIZE; i++) {
749 		list_move_tail(&ndn->dn_dirty_records[i],
750 		    &odn->dn_dirty_records[i]);
751 	}
752 	bcopy(&odn->dn_free_ranges[0], &ndn->dn_free_ranges[0],
753 	    sizeof (odn->dn_free_ranges));
754 	ndn->dn_allocated_txg = odn->dn_allocated_txg;
755 	ndn->dn_free_txg = odn->dn_free_txg;
756 	ndn->dn_assigned_txg = odn->dn_assigned_txg;
757 	ndn->dn_dirtyctx = odn->dn_dirtyctx;
758 	ndn->dn_dirtyctx_firstset = odn->dn_dirtyctx_firstset;
759 	ASSERT(refcount_count(&odn->dn_tx_holds) == 0);
760 	refcount_transfer(&ndn->dn_holds, &odn->dn_holds);
761 	ASSERT(avl_is_empty(&ndn->dn_dbufs));
762 	avl_swap(&ndn->dn_dbufs, &odn->dn_dbufs);
763 	ndn->dn_dbufs_count = odn->dn_dbufs_count;
764 	ndn->dn_bonus = odn->dn_bonus;
765 	ndn->dn_have_spill = odn->dn_have_spill;
766 	ndn->dn_zio = odn->dn_zio;
767 	ndn->dn_oldused = odn->dn_oldused;
768 	ndn->dn_oldflags = odn->dn_oldflags;
769 	ndn->dn_olduid = odn->dn_olduid;
770 	ndn->dn_oldgid = odn->dn_oldgid;
771 	ndn->dn_newuid = odn->dn_newuid;
772 	ndn->dn_newgid = odn->dn_newgid;
773 	ndn->dn_id_flags = odn->dn_id_flags;
774 	dmu_zfetch_init(&ndn->dn_zfetch, NULL);
775 	list_move_tail(&ndn->dn_zfetch.zf_stream, &odn->dn_zfetch.zf_stream);
776 	ndn->dn_zfetch.zf_dnode = odn->dn_zfetch.zf_dnode;
777 
778 	/*
779 	 * Update back pointers. Updating the handle fixes the back pointer of
780 	 * every descendant dbuf as well as the bonus dbuf.
781 	 */
782 	ASSERT(ndn->dn_handle->dnh_dnode == odn);
783 	ndn->dn_handle->dnh_dnode = ndn;
784 	if (ndn->dn_zfetch.zf_dnode == odn) {
785 		ndn->dn_zfetch.zf_dnode = ndn;
786 	}
787 
788 	/*
789 	 * Invalidate the original dnode by clearing all of its back pointers.
790 	 */
791 	odn->dn_dbuf = NULL;
792 	odn->dn_handle = NULL;
793 	avl_create(&odn->dn_dbufs, dbuf_compare, sizeof (dmu_buf_impl_t),
794 	    offsetof(dmu_buf_impl_t, db_link));
795 	odn->dn_dbufs_count = 0;
796 	odn->dn_bonus = NULL;
797 	odn->dn_zfetch.zf_dnode = NULL;
798 
799 	/*
800 	 * Set the low bit of the objset pointer to ensure that dnode_move()
801 	 * recognizes the dnode as invalid in any subsequent callback.
802 	 */
803 	POINTER_INVALIDATE(&odn->dn_objset);
804 
805 	/*
806 	 * Satisfy the destructor.
807 	 */
808 	for (i = 0; i < TXG_SIZE; i++) {
809 		list_create(&odn->dn_dirty_records[i],
810 		    sizeof (dbuf_dirty_record_t),
811 		    offsetof(dbuf_dirty_record_t, dr_dirty_node));
812 		odn->dn_free_ranges[i] = NULL;
813 		odn->dn_next_nlevels[i] = 0;
814 		odn->dn_next_indblkshift[i] = 0;
815 		odn->dn_next_bonustype[i] = 0;
816 		odn->dn_rm_spillblk[i] = 0;
817 		odn->dn_next_bonuslen[i] = 0;
818 		odn->dn_next_blksz[i] = 0;
819 	}
820 	odn->dn_allocated_txg = 0;
821 	odn->dn_free_txg = 0;
822 	odn->dn_assigned_txg = 0;
823 	odn->dn_dirtyctx = 0;
824 	odn->dn_dirtyctx_firstset = NULL;
825 	odn->dn_have_spill = B_FALSE;
826 	odn->dn_zio = NULL;
827 	odn->dn_oldused = 0;
828 	odn->dn_oldflags = 0;
829 	odn->dn_olduid = 0;
830 	odn->dn_oldgid = 0;
831 	odn->dn_newuid = 0;
832 	odn->dn_newgid = 0;
833 	odn->dn_id_flags = 0;
834 
835 	/*
836 	 * Mark the dnode.
837 	 */
838 	ndn->dn_moved = 1;
839 	odn->dn_moved = (uint8_t)-1;
840 }
841 
842 #ifdef illumos
843 #ifdef	_KERNEL
844 /*ARGSUSED*/
845 static kmem_cbrc_t
dnode_move(void * buf,void * newbuf,size_t size,void * arg)846 dnode_move(void *buf, void *newbuf, size_t size, void *arg)
847 {
848 	dnode_t *odn = buf, *ndn = newbuf;
849 	objset_t *os;
850 	int64_t refcount;
851 	uint32_t dbufs;
852 
853 	/*
854 	 * The dnode is on the objset's list of known dnodes if the objset
855 	 * pointer is valid. We set the low bit of the objset pointer when
856 	 * freeing the dnode to invalidate it, and the memory patterns written
857 	 * by kmem (baddcafe and deadbeef) set at least one of the two low bits.
858 	 * A newly created dnode sets the objset pointer last of all to indicate
859 	 * that the dnode is known and in a valid state to be moved by this
860 	 * function.
861 	 */
862 	os = odn->dn_objset;
863 	if (!POINTER_IS_VALID(os)) {
864 		DNODE_STAT_ADD(dnode_move_stats.dms_dnode_invalid);
865 		return (KMEM_CBRC_DONT_KNOW);
866 	}
867 
868 	/*
869 	 * Ensure that the objset does not go away during the move.
870 	 */
871 	rw_enter(&os_lock, RW_WRITER);
872 	if (os != odn->dn_objset) {
873 		rw_exit(&os_lock);
874 		DNODE_STAT_ADD(dnode_move_stats.dms_dnode_recheck1);
875 		return (KMEM_CBRC_DONT_KNOW);
876 	}
877 
878 	/*
879 	 * If the dnode is still valid, then so is the objset. We know that no
880 	 * valid objset can be freed while we hold os_lock, so we can safely
881 	 * ensure that the objset remains in use.
882 	 */
883 	mutex_enter(&os->os_lock);
884 
885 	/*
886 	 * Recheck the objset pointer in case the dnode was removed just before
887 	 * acquiring the lock.
888 	 */
889 	if (os != odn->dn_objset) {
890 		mutex_exit(&os->os_lock);
891 		rw_exit(&os_lock);
892 		DNODE_STAT_ADD(dnode_move_stats.dms_dnode_recheck2);
893 		return (KMEM_CBRC_DONT_KNOW);
894 	}
895 
896 	/*
897 	 * At this point we know that as long as we hold os->os_lock, the dnode
898 	 * cannot be freed and fields within the dnode can be safely accessed.
899 	 * The objset listing this dnode cannot go away as long as this dnode is
900 	 * on its list.
901 	 */
902 	rw_exit(&os_lock);
903 	if (DMU_OBJECT_IS_SPECIAL(odn->dn_object)) {
904 		mutex_exit(&os->os_lock);
905 		DNODE_STAT_ADD(dnode_move_stats.dms_dnode_special);
906 		return (KMEM_CBRC_NO);
907 	}
908 	ASSERT(odn->dn_dbuf != NULL); /* only "special" dnodes have no parent */
909 
910 	/*
911 	 * Lock the dnode handle to prevent the dnode from obtaining any new
912 	 * holds. This also prevents the descendant dbufs and the bonus dbuf
913 	 * from accessing the dnode, so that we can discount their holds. The
914 	 * handle is safe to access because we know that while the dnode cannot
915 	 * go away, neither can its handle. Once we hold dnh_zrlock, we can
916 	 * safely move any dnode referenced only by dbufs.
917 	 */
918 	if (!zrl_tryenter(&odn->dn_handle->dnh_zrlock)) {
919 		mutex_exit(&os->os_lock);
920 		DNODE_STAT_ADD(dnode_move_stats.dms_dnode_handle);
921 		return (KMEM_CBRC_LATER);
922 	}
923 
924 	/*
925 	 * Ensure a consistent view of the dnode's holds and the dnode's dbufs.
926 	 * We need to guarantee that there is a hold for every dbuf in order to
927 	 * determine whether the dnode is actively referenced. Falsely matching
928 	 * a dbuf to an active hold would lead to an unsafe move. It's possible
929 	 * that a thread already having an active dnode hold is about to add a
930 	 * dbuf, and we can't compare hold and dbuf counts while the add is in
931 	 * progress.
932 	 */
933 	if (!rw_tryenter(&odn->dn_struct_rwlock, RW_WRITER)) {
934 		zrl_exit(&odn->dn_handle->dnh_zrlock);
935 		mutex_exit(&os->os_lock);
936 		DNODE_STAT_ADD(dnode_move_stats.dms_dnode_rwlock);
937 		return (KMEM_CBRC_LATER);
938 	}
939 
940 	/*
941 	 * A dbuf may be removed (evicted) without an active dnode hold. In that
942 	 * case, the dbuf count is decremented under the handle lock before the
943 	 * dbuf's hold is released. This order ensures that if we count the hold
944 	 * after the dbuf is removed but before its hold is released, we will
945 	 * treat the unmatched hold as active and exit safely. If we count the
946 	 * hold before the dbuf is removed, the hold is discounted, and the
947 	 * removal is blocked until the move completes.
948 	 */
949 	refcount = refcount_count(&odn->dn_holds);
950 	ASSERT(refcount >= 0);
951 	dbufs = odn->dn_dbufs_count;
952 
953 	/* We can't have more dbufs than dnode holds. */
954 	ASSERT3U(dbufs, <=, refcount);
955 	DTRACE_PROBE3(dnode__move, dnode_t *, odn, int64_t, refcount,
956 	    uint32_t, dbufs);
957 
958 	if (refcount > dbufs) {
959 		rw_exit(&odn->dn_struct_rwlock);
960 		zrl_exit(&odn->dn_handle->dnh_zrlock);
961 		mutex_exit(&os->os_lock);
962 		DNODE_STAT_ADD(dnode_move_stats.dms_dnode_active);
963 		return (KMEM_CBRC_LATER);
964 	}
965 
966 	rw_exit(&odn->dn_struct_rwlock);
967 
968 	/*
969 	 * At this point we know that anyone with a hold on the dnode is not
970 	 * actively referencing it. The dnode is known and in a valid state to
971 	 * move. We're holding the locks needed to execute the critical section.
972 	 */
973 	dnode_move_impl(odn, ndn);
974 
975 	list_link_replace(&odn->dn_link, &ndn->dn_link);
976 	/* If the dnode was safe to move, the refcount cannot have changed. */
977 	ASSERT(refcount == refcount_count(&ndn->dn_holds));
978 	ASSERT(dbufs == ndn->dn_dbufs_count);
979 	zrl_exit(&ndn->dn_handle->dnh_zrlock); /* handle has moved */
980 	mutex_exit(&os->os_lock);
981 
982 	return (KMEM_CBRC_YES);
983 }
984 #endif	/* _KERNEL */
985 #endif	/* illumos */
986 
987 void
dnode_special_close(dnode_handle_t * dnh)988 dnode_special_close(dnode_handle_t *dnh)
989 {
990 	dnode_t *dn = dnh->dnh_dnode;
991 
992 	/*
993 	 * Wait for final references to the dnode to clear.  This can
994 	 * only happen if the arc is asyncronously evicting state that
995 	 * has a hold on this dnode while we are trying to evict this
996 	 * dnode.
997 	 */
998 	while (refcount_count(&dn->dn_holds) > 0)
999 		delay(1);
1000 	ASSERT(dn->dn_dbuf == NULL ||
1001 	    dmu_buf_get_user(&dn->dn_dbuf->db) == NULL);
1002 	zrl_add(&dnh->dnh_zrlock);
1003 	dnode_destroy(dn); /* implicit zrl_remove() */
1004 	zrl_destroy(&dnh->dnh_zrlock);
1005 	dnh->dnh_dnode = NULL;
1006 }
1007 
1008 void
dnode_special_open(objset_t * os,dnode_phys_t * dnp,uint64_t object,dnode_handle_t * dnh)1009 dnode_special_open(objset_t *os, dnode_phys_t *dnp, uint64_t object,
1010     dnode_handle_t *dnh)
1011 {
1012 	dnode_t *dn;
1013 
1014 	dn = dnode_create(os, dnp, NULL, object, dnh);
1015 	zrl_init(&dnh->dnh_zrlock);
1016 	DNODE_VERIFY(dn);
1017 }
1018 
1019 static void
dnode_buf_evict_async(void * dbu)1020 dnode_buf_evict_async(void *dbu)
1021 {
1022 	dnode_children_t *children_dnodes = dbu;
1023 	int i;
1024 
1025 	for (i = 0; i < children_dnodes->dnc_count; i++) {
1026 		dnode_handle_t *dnh = &children_dnodes->dnc_children[i];
1027 		dnode_t *dn;
1028 
1029 		/*
1030 		 * The dnode handle lock guards against the dnode moving to
1031 		 * another valid address, so there is no need here to guard
1032 		 * against changes to or from NULL.
1033 		 */
1034 		if (dnh->dnh_dnode == NULL) {
1035 			zrl_destroy(&dnh->dnh_zrlock);
1036 			continue;
1037 		}
1038 
1039 		zrl_add(&dnh->dnh_zrlock);
1040 		dn = dnh->dnh_dnode;
1041 		/*
1042 		 * If there are holds on this dnode, then there should
1043 		 * be holds on the dnode's containing dbuf as well; thus
1044 		 * it wouldn't be eligible for eviction and this function
1045 		 * would not have been called.
1046 		 */
1047 		ASSERT(refcount_is_zero(&dn->dn_holds));
1048 		ASSERT(refcount_is_zero(&dn->dn_tx_holds));
1049 
1050 		dnode_destroy(dn); /* implicit zrl_remove() */
1051 		zrl_destroy(&dnh->dnh_zrlock);
1052 		dnh->dnh_dnode = NULL;
1053 	}
1054 	kmem_free(children_dnodes, sizeof (dnode_children_t) +
1055 	    children_dnodes->dnc_count * sizeof (dnode_handle_t));
1056 }
1057 
1058 /*
1059  * errors:
1060  * EINVAL - invalid object number.
1061  * EIO - i/o error.
1062  * succeeds even for free dnodes.
1063  */
1064 int
dnode_hold_impl(objset_t * os,uint64_t object,int flag,void * tag,dnode_t ** dnp)1065 dnode_hold_impl(objset_t *os, uint64_t object, int flag,
1066     void *tag, dnode_t **dnp)
1067 {
1068 	int epb, idx, err;
1069 	int drop_struct_lock = FALSE;
1070 	int type;
1071 	uint64_t blk;
1072 	dnode_t *mdn, *dn;
1073 	dmu_buf_impl_t *db;
1074 	dnode_children_t *children_dnodes;
1075 	dnode_handle_t *dnh;
1076 
1077 	/*
1078 	 * If you are holding the spa config lock as writer, you shouldn't
1079 	 * be asking the DMU to do *anything* unless it's the root pool
1080 	 * which may require us to read from the root filesystem while
1081 	 * holding some (not all) of the locks as writer.
1082 	 */
1083 	ASSERT(spa_config_held(os->os_spa, SCL_ALL, RW_WRITER) == 0 ||
1084 	    (spa_is_root(os->os_spa) &&
1085 	    spa_config_held(os->os_spa, SCL_STATE, RW_WRITER)));
1086 
1087 	if (object == DMU_USERUSED_OBJECT || object == DMU_GROUPUSED_OBJECT) {
1088 		dn = (object == DMU_USERUSED_OBJECT) ?
1089 		    DMU_USERUSED_DNODE(os) : DMU_GROUPUSED_DNODE(os);
1090 		if (dn == NULL)
1091 			return (SET_ERROR(ENOENT));
1092 		type = dn->dn_type;
1093 		if ((flag & DNODE_MUST_BE_ALLOCATED) && type == DMU_OT_NONE)
1094 			return (SET_ERROR(ENOENT));
1095 		if ((flag & DNODE_MUST_BE_FREE) && type != DMU_OT_NONE)
1096 			return (SET_ERROR(EEXIST));
1097 		DNODE_VERIFY(dn);
1098 		(void) refcount_add(&dn->dn_holds, tag);
1099 		*dnp = dn;
1100 		return (0);
1101 	}
1102 
1103 	if (object == 0 || object >= DN_MAX_OBJECT)
1104 		return (SET_ERROR(EINVAL));
1105 
1106 	mdn = DMU_META_DNODE(os);
1107 	ASSERT(mdn->dn_object == DMU_META_DNODE_OBJECT);
1108 
1109 	DNODE_VERIFY(mdn);
1110 
1111 	if (!RW_WRITE_HELD(&mdn->dn_struct_rwlock)) {
1112 		rw_enter(&mdn->dn_struct_rwlock, RW_READER);
1113 		drop_struct_lock = TRUE;
1114 	}
1115 
1116 	blk = dbuf_whichblock(mdn, 0, object * sizeof (dnode_phys_t));
1117 
1118 	db = dbuf_hold(mdn, blk, FTAG);
1119 	if (drop_struct_lock)
1120 		rw_exit(&mdn->dn_struct_rwlock);
1121 	if (db == NULL)
1122 		return (SET_ERROR(EIO));
1123 	err = dbuf_read(db, NULL, DB_RF_CANFAIL);
1124 	if (err) {
1125 		dbuf_rele(db, FTAG);
1126 		return (err);
1127 	}
1128 
1129 	ASSERT3U(db->db.db_size, >=, 1<<DNODE_SHIFT);
1130 	epb = db->db.db_size >> DNODE_SHIFT;
1131 
1132 	idx = object & (epb-1);
1133 
1134 	ASSERT(DB_DNODE(db)->dn_type == DMU_OT_DNODE);
1135 	children_dnodes = dmu_buf_get_user(&db->db);
1136 	if (children_dnodes == NULL) {
1137 		int i;
1138 		dnode_children_t *winner;
1139 		children_dnodes = kmem_zalloc(sizeof (dnode_children_t) +
1140 		    epb * sizeof (dnode_handle_t), KM_SLEEP);
1141 		children_dnodes->dnc_count = epb;
1142 		dnh = &children_dnodes->dnc_children[0];
1143 		for (i = 0; i < epb; i++) {
1144 			zrl_init(&dnh[i].dnh_zrlock);
1145 		}
1146 		dmu_buf_init_user(&children_dnodes->dnc_dbu, NULL,
1147 		    dnode_buf_evict_async, NULL);
1148 		winner = dmu_buf_set_user(&db->db, &children_dnodes->dnc_dbu);
1149 		if (winner != NULL) {
1150 
1151 			for (i = 0; i < epb; i++) {
1152 				zrl_destroy(&dnh[i].dnh_zrlock);
1153 			}
1154 
1155 			kmem_free(children_dnodes, sizeof (dnode_children_t) +
1156 			    epb * sizeof (dnode_handle_t));
1157 			children_dnodes = winner;
1158 		}
1159 	}
1160 	ASSERT(children_dnodes->dnc_count == epb);
1161 
1162 	dnh = &children_dnodes->dnc_children[idx];
1163 	zrl_add(&dnh->dnh_zrlock);
1164 	dn = dnh->dnh_dnode;
1165 	if (dn == NULL) {
1166 		dnode_phys_t *phys = (dnode_phys_t *)db->db.db_data+idx;
1167 
1168 		dn = dnode_create(os, phys, db, object, dnh);
1169 	}
1170 
1171 	mutex_enter(&dn->dn_mtx);
1172 	type = dn->dn_type;
1173 	if (dn->dn_free_txg ||
1174 	    ((flag & DNODE_MUST_BE_ALLOCATED) && type == DMU_OT_NONE) ||
1175 	    ((flag & DNODE_MUST_BE_FREE) &&
1176 	    (type != DMU_OT_NONE || !refcount_is_zero(&dn->dn_holds)))) {
1177 		mutex_exit(&dn->dn_mtx);
1178 		zrl_remove(&dnh->dnh_zrlock);
1179 		dbuf_rele(db, FTAG);
1180 		return (type == DMU_OT_NONE ? ENOENT : EEXIST);
1181 	}
1182 	if (refcount_add(&dn->dn_holds, tag) == 1)
1183 		dbuf_add_ref(db, dnh);
1184 	mutex_exit(&dn->dn_mtx);
1185 
1186 	/* Now we can rely on the hold to prevent the dnode from moving. */
1187 	zrl_remove(&dnh->dnh_zrlock);
1188 
1189 	DNODE_VERIFY(dn);
1190 	ASSERT3P(dn->dn_dbuf, ==, db);
1191 	ASSERT3U(dn->dn_object, ==, object);
1192 	dbuf_rele(db, FTAG);
1193 
1194 	*dnp = dn;
1195 	return (0);
1196 }
1197 
1198 /*
1199  * Return held dnode if the object is allocated, NULL if not.
1200  */
1201 int
dnode_hold(objset_t * os,uint64_t object,void * tag,dnode_t ** dnp)1202 dnode_hold(objset_t *os, uint64_t object, void *tag, dnode_t **dnp)
1203 {
1204 	return (dnode_hold_impl(os, object, DNODE_MUST_BE_ALLOCATED, tag, dnp));
1205 }
1206 
1207 /*
1208  * Can only add a reference if there is already at least one
1209  * reference on the dnode.  Returns FALSE if unable to add a
1210  * new reference.
1211  */
1212 boolean_t
dnode_add_ref(dnode_t * dn,void * tag)1213 dnode_add_ref(dnode_t *dn, void *tag)
1214 {
1215 	mutex_enter(&dn->dn_mtx);
1216 	if (refcount_is_zero(&dn->dn_holds)) {
1217 		mutex_exit(&dn->dn_mtx);
1218 		return (FALSE);
1219 	}
1220 	VERIFY(1 < refcount_add(&dn->dn_holds, tag));
1221 	mutex_exit(&dn->dn_mtx);
1222 	return (TRUE);
1223 }
1224 
1225 void
dnode_rele(dnode_t * dn,void * tag)1226 dnode_rele(dnode_t *dn, void *tag)
1227 {
1228 	mutex_enter(&dn->dn_mtx);
1229 	dnode_rele_and_unlock(dn, tag);
1230 }
1231 
1232 void
dnode_rele_and_unlock(dnode_t * dn,void * tag)1233 dnode_rele_and_unlock(dnode_t *dn, void *tag)
1234 {
1235 	uint64_t refs;
1236 	/* Get while the hold prevents the dnode from moving. */
1237 	dmu_buf_impl_t *db = dn->dn_dbuf;
1238 	dnode_handle_t *dnh = dn->dn_handle;
1239 
1240 	refs = refcount_remove(&dn->dn_holds, tag);
1241 	mutex_exit(&dn->dn_mtx);
1242 
1243 	/*
1244 	 * It's unsafe to release the last hold on a dnode by dnode_rele() or
1245 	 * indirectly by dbuf_rele() while relying on the dnode handle to
1246 	 * prevent the dnode from moving, since releasing the last hold could
1247 	 * result in the dnode's parent dbuf evicting its dnode handles. For
1248 	 * that reason anyone calling dnode_rele() or dbuf_rele() without some
1249 	 * other direct or indirect hold on the dnode must first drop the dnode
1250 	 * handle.
1251 	 */
1252 	ASSERT(refs > 0 || dnh->dnh_zrlock.zr_owner != curthread);
1253 
1254 	/* NOTE: the DNODE_DNODE does not have a dn_dbuf */
1255 	if (refs == 0 && db != NULL) {
1256 		/*
1257 		 * Another thread could add a hold to the dnode handle in
1258 		 * dnode_hold_impl() while holding the parent dbuf. Since the
1259 		 * hold on the parent dbuf prevents the handle from being
1260 		 * destroyed, the hold on the handle is OK. We can't yet assert
1261 		 * that the handle has zero references, but that will be
1262 		 * asserted anyway when the handle gets destroyed.
1263 		 */
1264 		dbuf_rele(db, dnh);
1265 	}
1266 }
1267 
1268 void
dnode_setdirty(dnode_t * dn,dmu_tx_t * tx)1269 dnode_setdirty(dnode_t *dn, dmu_tx_t *tx)
1270 {
1271 	objset_t *os = dn->dn_objset;
1272 	uint64_t txg = tx->tx_txg;
1273 
1274 	if (DMU_OBJECT_IS_SPECIAL(dn->dn_object)) {
1275 		dsl_dataset_dirty(os->os_dsl_dataset, tx);
1276 		return;
1277 	}
1278 
1279 	DNODE_VERIFY(dn);
1280 
1281 #ifdef ZFS_DEBUG
1282 	mutex_enter(&dn->dn_mtx);
1283 	ASSERT(dn->dn_phys->dn_type || dn->dn_allocated_txg);
1284 	ASSERT(dn->dn_free_txg == 0 || dn->dn_free_txg >= txg);
1285 	mutex_exit(&dn->dn_mtx);
1286 #endif
1287 
1288 	/*
1289 	 * Determine old uid/gid when necessary
1290 	 */
1291 	dmu_objset_userquota_get_ids(dn, B_TRUE, tx);
1292 
1293 	mutex_enter(&os->os_lock);
1294 
1295 	/*
1296 	 * If we are already marked dirty, we're done.
1297 	 */
1298 	if (list_link_active(&dn->dn_dirty_link[txg & TXG_MASK])) {
1299 		mutex_exit(&os->os_lock);
1300 		return;
1301 	}
1302 
1303 	ASSERT(!refcount_is_zero(&dn->dn_holds) ||
1304 	    !avl_is_empty(&dn->dn_dbufs));
1305 	ASSERT(dn->dn_datablksz != 0);
1306 	ASSERT0(dn->dn_next_bonuslen[txg&TXG_MASK]);
1307 	ASSERT0(dn->dn_next_blksz[txg&TXG_MASK]);
1308 	ASSERT0(dn->dn_next_bonustype[txg&TXG_MASK]);
1309 
1310 	dprintf_ds(os->os_dsl_dataset, "obj=%llu txg=%llu\n",
1311 	    dn->dn_object, txg);
1312 
1313 	if (dn->dn_free_txg > 0 && dn->dn_free_txg <= txg) {
1314 		list_insert_tail(&os->os_free_dnodes[txg&TXG_MASK], dn);
1315 	} else {
1316 		list_insert_tail(&os->os_dirty_dnodes[txg&TXG_MASK], dn);
1317 	}
1318 
1319 	mutex_exit(&os->os_lock);
1320 
1321 	/*
1322 	 * The dnode maintains a hold on its containing dbuf as
1323 	 * long as there are holds on it.  Each instantiated child
1324 	 * dbuf maintains a hold on the dnode.  When the last child
1325 	 * drops its hold, the dnode will drop its hold on the
1326 	 * containing dbuf. We add a "dirty hold" here so that the
1327 	 * dnode will hang around after we finish processing its
1328 	 * children.
1329 	 */
1330 	VERIFY(dnode_add_ref(dn, (void *)(uintptr_t)tx->tx_txg));
1331 
1332 	(void) dbuf_dirty(dn->dn_dbuf, tx);
1333 
1334 	dsl_dataset_dirty(os->os_dsl_dataset, tx);
1335 }
1336 
1337 void
dnode_free(dnode_t * dn,dmu_tx_t * tx)1338 dnode_free(dnode_t *dn, dmu_tx_t *tx)
1339 {
1340 	int txgoff = tx->tx_txg & TXG_MASK;
1341 
1342 	dprintf("dn=%p txg=%llu\n", dn, tx->tx_txg);
1343 
1344 	/* we should be the only holder... hopefully */
1345 	/* ASSERT3U(refcount_count(&dn->dn_holds), ==, 1); */
1346 
1347 	mutex_enter(&dn->dn_mtx);
1348 	if (dn->dn_type == DMU_OT_NONE || dn->dn_free_txg) {
1349 		mutex_exit(&dn->dn_mtx);
1350 		return;
1351 	}
1352 	dn->dn_free_txg = tx->tx_txg;
1353 	mutex_exit(&dn->dn_mtx);
1354 
1355 	/*
1356 	 * If the dnode is already dirty, it needs to be moved from
1357 	 * the dirty list to the free list.
1358 	 */
1359 	mutex_enter(&dn->dn_objset->os_lock);
1360 	if (list_link_active(&dn->dn_dirty_link[txgoff])) {
1361 		list_remove(&dn->dn_objset->os_dirty_dnodes[txgoff], dn);
1362 		list_insert_tail(&dn->dn_objset->os_free_dnodes[txgoff], dn);
1363 		mutex_exit(&dn->dn_objset->os_lock);
1364 	} else {
1365 		mutex_exit(&dn->dn_objset->os_lock);
1366 		dnode_setdirty(dn, tx);
1367 	}
1368 }
1369 
1370 /*
1371  * Try to change the block size for the indicated dnode.  This can only
1372  * succeed if there are no blocks allocated or dirty beyond first block
1373  */
1374 int
dnode_set_blksz(dnode_t * dn,uint64_t size,int ibs,dmu_tx_t * tx)1375 dnode_set_blksz(dnode_t *dn, uint64_t size, int ibs, dmu_tx_t *tx)
1376 {
1377 	dmu_buf_impl_t *db;
1378 	int err;
1379 
1380 	ASSERT3U(size, <=, spa_maxblocksize(dmu_objset_spa(dn->dn_objset)));
1381 	if (size == 0)
1382 		size = SPA_MINBLOCKSIZE;
1383 	else
1384 		size = P2ROUNDUP(size, SPA_MINBLOCKSIZE);
1385 
1386 	if (ibs == dn->dn_indblkshift)
1387 		ibs = 0;
1388 
1389 	if (size >> SPA_MINBLOCKSHIFT == dn->dn_datablkszsec && ibs == 0)
1390 		return (0);
1391 
1392 	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1393 
1394 	/* Check for any allocated blocks beyond the first */
1395 	if (dn->dn_maxblkid != 0)
1396 		goto fail;
1397 
1398 	mutex_enter(&dn->dn_dbufs_mtx);
1399 	for (db = avl_first(&dn->dn_dbufs); db != NULL;
1400 	    db = AVL_NEXT(&dn->dn_dbufs, db)) {
1401 		if (db->db_blkid != 0 && db->db_blkid != DMU_BONUS_BLKID &&
1402 		    db->db_blkid != DMU_SPILL_BLKID) {
1403 			mutex_exit(&dn->dn_dbufs_mtx);
1404 			goto fail;
1405 		}
1406 	}
1407 	mutex_exit(&dn->dn_dbufs_mtx);
1408 
1409 	if (ibs && dn->dn_nlevels != 1)
1410 		goto fail;
1411 
1412 	/* resize the old block */
1413 	err = dbuf_hold_impl(dn, 0, 0, TRUE, FALSE, FTAG, &db);
1414 	if (err == 0)
1415 		dbuf_new_size(db, size, tx);
1416 	else if (err != ENOENT)
1417 		goto fail;
1418 
1419 	dnode_setdblksz(dn, size);
1420 	dnode_setdirty(dn, tx);
1421 	dn->dn_next_blksz[tx->tx_txg&TXG_MASK] = size;
1422 	if (ibs) {
1423 		dn->dn_indblkshift = ibs;
1424 		dn->dn_next_indblkshift[tx->tx_txg&TXG_MASK] = ibs;
1425 	}
1426 	/* rele after we have fixed the blocksize in the dnode */
1427 	if (db)
1428 		dbuf_rele(db, FTAG);
1429 
1430 	rw_exit(&dn->dn_struct_rwlock);
1431 	return (0);
1432 
1433 fail:
1434 	rw_exit(&dn->dn_struct_rwlock);
1435 	return (SET_ERROR(ENOTSUP));
1436 }
1437 
1438 /* read-holding callers must not rely on the lock being continuously held */
1439 void
dnode_new_blkid(dnode_t * dn,uint64_t blkid,dmu_tx_t * tx,boolean_t have_read)1440 dnode_new_blkid(dnode_t *dn, uint64_t blkid, dmu_tx_t *tx, boolean_t have_read)
1441 {
1442 	uint64_t txgoff = tx->tx_txg & TXG_MASK;
1443 	int epbs, new_nlevels;
1444 	uint64_t sz;
1445 
1446 	ASSERT(blkid != DMU_BONUS_BLKID);
1447 
1448 	ASSERT(have_read ?
1449 	    RW_READ_HELD(&dn->dn_struct_rwlock) :
1450 	    RW_WRITE_HELD(&dn->dn_struct_rwlock));
1451 
1452 	/*
1453 	 * if we have a read-lock, check to see if we need to do any work
1454 	 * before upgrading to a write-lock.
1455 	 */
1456 	if (have_read) {
1457 		if (blkid <= dn->dn_maxblkid)
1458 			return;
1459 
1460 		if (!rw_tryupgrade(&dn->dn_struct_rwlock)) {
1461 			rw_exit(&dn->dn_struct_rwlock);
1462 			rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1463 		}
1464 	}
1465 
1466 	if (blkid <= dn->dn_maxblkid)
1467 		goto out;
1468 
1469 	dn->dn_maxblkid = blkid;
1470 
1471 	/*
1472 	 * Compute the number of levels necessary to support the new maxblkid.
1473 	 */
1474 	new_nlevels = 1;
1475 	epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1476 	for (sz = dn->dn_nblkptr;
1477 	    sz <= blkid && sz >= dn->dn_nblkptr; sz <<= epbs)
1478 		new_nlevels++;
1479 
1480 	if (new_nlevels > dn->dn_nlevels) {
1481 		int old_nlevels = dn->dn_nlevels;
1482 		dmu_buf_impl_t *db;
1483 		list_t *list;
1484 		dbuf_dirty_record_t *new, *dr, *dr_next;
1485 
1486 		dn->dn_nlevels = new_nlevels;
1487 
1488 		ASSERT3U(new_nlevels, >, dn->dn_next_nlevels[txgoff]);
1489 		dn->dn_next_nlevels[txgoff] = new_nlevels;
1490 
1491 		/* dirty the left indirects */
1492 		db = dbuf_hold_level(dn, old_nlevels, 0, FTAG);
1493 		ASSERT(db != NULL);
1494 		new = dbuf_dirty(db, tx);
1495 		dbuf_rele(db, FTAG);
1496 
1497 		/* transfer the dirty records to the new indirect */
1498 		mutex_enter(&dn->dn_mtx);
1499 		mutex_enter(&new->dt.di.dr_mtx);
1500 		list = &dn->dn_dirty_records[txgoff];
1501 		for (dr = list_head(list); dr; dr = dr_next) {
1502 			dr_next = list_next(&dn->dn_dirty_records[txgoff], dr);
1503 			if (dr->dr_dbuf->db_level != new_nlevels-1 &&
1504 			    dr->dr_dbuf->db_blkid != DMU_BONUS_BLKID &&
1505 			    dr->dr_dbuf->db_blkid != DMU_SPILL_BLKID) {
1506 				ASSERT(dr->dr_dbuf->db_level == old_nlevels-1);
1507 				list_remove(&dn->dn_dirty_records[txgoff], dr);
1508 				list_insert_tail(&new->dt.di.dr_children, dr);
1509 				dr->dr_parent = new;
1510 			}
1511 		}
1512 		mutex_exit(&new->dt.di.dr_mtx);
1513 		mutex_exit(&dn->dn_mtx);
1514 	}
1515 
1516 out:
1517 	if (have_read)
1518 		rw_downgrade(&dn->dn_struct_rwlock);
1519 }
1520 
1521 static void
dnode_dirty_l1(dnode_t * dn,uint64_t l1blkid,dmu_tx_t * tx)1522 dnode_dirty_l1(dnode_t *dn, uint64_t l1blkid, dmu_tx_t *tx)
1523 {
1524 	dmu_buf_impl_t *db = dbuf_hold_level(dn, 1, l1blkid, FTAG);
1525 	if (db != NULL) {
1526 		dmu_buf_will_dirty(&db->db, tx);
1527 		dbuf_rele(db, FTAG);
1528 	}
1529 }
1530 
1531 void
dnode_free_range(dnode_t * dn,uint64_t off,uint64_t len,dmu_tx_t * tx)1532 dnode_free_range(dnode_t *dn, uint64_t off, uint64_t len, dmu_tx_t *tx)
1533 {
1534 	dmu_buf_impl_t *db;
1535 	uint64_t blkoff, blkid, nblks;
1536 	int blksz, blkshift, head, tail;
1537 	int trunc = FALSE;
1538 	int epbs;
1539 
1540 	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1541 	blksz = dn->dn_datablksz;
1542 	blkshift = dn->dn_datablkshift;
1543 	epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1544 
1545 	if (len == DMU_OBJECT_END) {
1546 		len = UINT64_MAX - off;
1547 		trunc = TRUE;
1548 	}
1549 
1550 	/*
1551 	 * First, block align the region to free:
1552 	 */
1553 	if (ISP2(blksz)) {
1554 		head = P2NPHASE(off, blksz);
1555 		blkoff = P2PHASE(off, blksz);
1556 		if ((off >> blkshift) > dn->dn_maxblkid)
1557 			goto out;
1558 	} else {
1559 		ASSERT(dn->dn_maxblkid == 0);
1560 		if (off == 0 && len >= blksz) {
1561 			/*
1562 			 * Freeing the whole block; fast-track this request.
1563 			 * Note that we won't dirty any indirect blocks,
1564 			 * which is fine because we will be freeing the entire
1565 			 * file and thus all indirect blocks will be freed
1566 			 * by free_children().
1567 			 */
1568 			blkid = 0;
1569 			nblks = 1;
1570 			goto done;
1571 		} else if (off >= blksz) {
1572 			/* Freeing past end-of-data */
1573 			goto out;
1574 		} else {
1575 			/* Freeing part of the block. */
1576 			head = blksz - off;
1577 			ASSERT3U(head, >, 0);
1578 		}
1579 		blkoff = off;
1580 	}
1581 	/* zero out any partial block data at the start of the range */
1582 	if (head) {
1583 		ASSERT3U(blkoff + head, ==, blksz);
1584 		if (len < head)
1585 			head = len;
1586 		if (dbuf_hold_impl(dn, 0, dbuf_whichblock(dn, 0, off),
1587 		    TRUE, FALSE, FTAG, &db) == 0) {
1588 			caddr_t data;
1589 
1590 			/* don't dirty if it isn't on disk and isn't 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 				data = db->db.db_data;
1597 				bzero(data + blkoff, head);
1598 			}
1599 			dbuf_rele(db, FTAG);
1600 		}
1601 		off += head;
1602 		len -= head;
1603 	}
1604 
1605 	/* If the range was less than one block, we're done */
1606 	if (len == 0)
1607 		goto out;
1608 
1609 	/* If the remaining range is past end of file, we're done */
1610 	if ((off >> blkshift) > dn->dn_maxblkid)
1611 		goto out;
1612 
1613 	ASSERT(ISP2(blksz));
1614 	if (trunc)
1615 		tail = 0;
1616 	else
1617 		tail = P2PHASE(len, blksz);
1618 
1619 	ASSERT0(P2PHASE(off, blksz));
1620 	/* zero out any partial block data at the end of the range */
1621 	if (tail) {
1622 		if (len < tail)
1623 			tail = len;
1624 		if (dbuf_hold_impl(dn, 0, dbuf_whichblock(dn, 0, off+len),
1625 		    TRUE, FALSE, FTAG, &db) == 0) {
1626 			/* don't dirty if not on disk and not dirty */
1627 			if (db->db_last_dirty ||
1628 			    (db->db_blkptr && !BP_IS_HOLE(db->db_blkptr))) {
1629 				rw_exit(&dn->dn_struct_rwlock);
1630 				dmu_buf_will_dirty(&db->db, tx);
1631 				rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1632 				bzero(db->db.db_data, tail);
1633 			}
1634 			dbuf_rele(db, FTAG);
1635 		}
1636 		len -= tail;
1637 	}
1638 
1639 	/* If the range did not include a full block, we are done */
1640 	if (len == 0)
1641 		goto out;
1642 
1643 	ASSERT(IS_P2ALIGNED(off, blksz));
1644 	ASSERT(trunc || IS_P2ALIGNED(len, blksz));
1645 	blkid = off >> blkshift;
1646 	nblks = len >> blkshift;
1647 	if (trunc)
1648 		nblks += 1;
1649 
1650 	/*
1651 	 * Dirty all the indirect blocks in this range.  Note that only
1652 	 * the first and last indirect blocks can actually be written
1653 	 * (if they were partially freed) -- they must be dirtied, even if
1654 	 * they do not exist on disk yet.  The interior blocks will
1655 	 * be freed by free_children(), so they will not actually be written.
1656 	 * Even though these interior blocks will not be written, we
1657 	 * dirty them for two reasons:
1658 	 *
1659 	 *  - It ensures that the indirect blocks remain in memory until
1660 	 *    syncing context.  (They have already been prefetched by
1661 	 *    dmu_tx_hold_free(), so we don't have to worry about reading
1662 	 *    them serially here.)
1663 	 *
1664 	 *  - The dirty space accounting will put pressure on the txg sync
1665 	 *    mechanism to begin syncing, and to delay transactions if there
1666 	 *    is a large amount of freeing.  Even though these indirect
1667 	 *    blocks will not be written, we could need to write the same
1668 	 *    amount of space if we copy the freed BPs into deadlists.
1669 	 */
1670 	if (dn->dn_nlevels > 1) {
1671 		uint64_t first, last;
1672 
1673 		first = blkid >> epbs;
1674 		dnode_dirty_l1(dn, first, tx);
1675 		if (trunc)
1676 			last = dn->dn_maxblkid >> epbs;
1677 		else
1678 			last = (blkid + nblks - 1) >> epbs;
1679 		if (last != first)
1680 			dnode_dirty_l1(dn, last, tx);
1681 
1682 		int shift = dn->dn_datablkshift + dn->dn_indblkshift -
1683 		    SPA_BLKPTRSHIFT;
1684 		for (uint64_t i = first + 1; i < last; i++) {
1685 			/*
1686 			 * Set i to the blockid of the next non-hole
1687 			 * level-1 indirect block at or after i.  Note
1688 			 * that dnode_next_offset() operates in terms of
1689 			 * level-0-equivalent bytes.
1690 			 */
1691 			uint64_t ibyte = i << shift;
1692 			int err = dnode_next_offset(dn, DNODE_FIND_HAVELOCK,
1693 			    &ibyte, 2, 1, 0);
1694 			i = ibyte >> shift;
1695 			if (i >= last)
1696 				break;
1697 
1698 			/*
1699 			 * Normally we should not see an error, either
1700 			 * from dnode_next_offset() or dbuf_hold_level()
1701 			 * (except for ESRCH from dnode_next_offset).
1702 			 * If there is an i/o error, then when we read
1703 			 * this block in syncing context, it will use
1704 			 * ZIO_FLAG_MUSTSUCCEED, and thus hang/panic according
1705 			 * to the "failmode" property.  dnode_next_offset()
1706 			 * doesn't have a flag to indicate MUSTSUCCEED.
1707 			 */
1708 			if (err != 0)
1709 				break;
1710 
1711 			dnode_dirty_l1(dn, i, tx);
1712 		}
1713 	}
1714 
1715 done:
1716 	/*
1717 	 * Add this range to the dnode range list.
1718 	 * We will finish up this free operation in the syncing phase.
1719 	 */
1720 	mutex_enter(&dn->dn_mtx);
1721 	int txgoff = tx->tx_txg & TXG_MASK;
1722 	if (dn->dn_free_ranges[txgoff] == NULL) {
1723 		dn->dn_free_ranges[txgoff] =
1724 		    range_tree_create(NULL, NULL, &dn->dn_mtx);
1725 	}
1726 	range_tree_clear(dn->dn_free_ranges[txgoff], blkid, nblks);
1727 	range_tree_add(dn->dn_free_ranges[txgoff], blkid, nblks);
1728 	dprintf_dnode(dn, "blkid=%llu nblks=%llu txg=%llu\n",
1729 	    blkid, nblks, tx->tx_txg);
1730 	mutex_exit(&dn->dn_mtx);
1731 
1732 	dbuf_free_range(dn, blkid, blkid + nblks - 1, tx);
1733 	dnode_setdirty(dn, tx);
1734 out:
1735 
1736 	rw_exit(&dn->dn_struct_rwlock);
1737 }
1738 
1739 static boolean_t
dnode_spill_freed(dnode_t * dn)1740 dnode_spill_freed(dnode_t *dn)
1741 {
1742 	int i;
1743 
1744 	mutex_enter(&dn->dn_mtx);
1745 	for (i = 0; i < TXG_SIZE; i++) {
1746 		if (dn->dn_rm_spillblk[i] == DN_KILL_SPILLBLK)
1747 			break;
1748 	}
1749 	mutex_exit(&dn->dn_mtx);
1750 	return (i < TXG_SIZE);
1751 }
1752 
1753 /* return TRUE if this blkid was freed in a recent txg, or FALSE if it wasn't */
1754 uint64_t
dnode_block_freed(dnode_t * dn,uint64_t blkid)1755 dnode_block_freed(dnode_t *dn, uint64_t blkid)
1756 {
1757 	void *dp = spa_get_dsl(dn->dn_objset->os_spa);
1758 	int i;
1759 
1760 	if (blkid == DMU_BONUS_BLKID)
1761 		return (FALSE);
1762 
1763 	/*
1764 	 * If we're in the process of opening the pool, dp will not be
1765 	 * set yet, but there shouldn't be anything dirty.
1766 	 */
1767 	if (dp == NULL)
1768 		return (FALSE);
1769 
1770 	if (dn->dn_free_txg)
1771 		return (TRUE);
1772 
1773 	if (blkid == DMU_SPILL_BLKID)
1774 		return (dnode_spill_freed(dn));
1775 
1776 	mutex_enter(&dn->dn_mtx);
1777 	for (i = 0; i < TXG_SIZE; i++) {
1778 		if (dn->dn_free_ranges[i] != NULL &&
1779 		    range_tree_contains(dn->dn_free_ranges[i], blkid, 1))
1780 			break;
1781 	}
1782 	mutex_exit(&dn->dn_mtx);
1783 	return (i < TXG_SIZE);
1784 }
1785 
1786 /* call from syncing context when we actually write/free space for this dnode */
1787 void
dnode_diduse_space(dnode_t * dn,int64_t delta)1788 dnode_diduse_space(dnode_t *dn, int64_t delta)
1789 {
1790 	uint64_t space;
1791 	dprintf_dnode(dn, "dn=%p dnp=%p used=%llu delta=%lld\n",
1792 	    dn, dn->dn_phys,
1793 	    (u_longlong_t)dn->dn_phys->dn_used,
1794 	    (longlong_t)delta);
1795 
1796 	mutex_enter(&dn->dn_mtx);
1797 	space = DN_USED_BYTES(dn->dn_phys);
1798 	if (delta > 0) {
1799 		ASSERT3U(space + delta, >=, space); /* no overflow */
1800 	} else {
1801 		ASSERT3U(space, >=, -delta); /* no underflow */
1802 	}
1803 	space += delta;
1804 	if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_DNODE_BYTES) {
1805 		ASSERT((dn->dn_phys->dn_flags & DNODE_FLAG_USED_BYTES) == 0);
1806 		ASSERT0(P2PHASE(space, 1<<DEV_BSHIFT));
1807 		dn->dn_phys->dn_used = space >> DEV_BSHIFT;
1808 	} else {
1809 		dn->dn_phys->dn_used = space;
1810 		dn->dn_phys->dn_flags |= DNODE_FLAG_USED_BYTES;
1811 	}
1812 	mutex_exit(&dn->dn_mtx);
1813 }
1814 
1815 /*
1816  * Call when we think we're going to write/free space in open context to track
1817  * the amount of memory in use by the currently open txg.
1818  */
1819 void
dnode_willuse_space(dnode_t * dn,int64_t space,dmu_tx_t * tx)1820 dnode_willuse_space(dnode_t *dn, int64_t space, dmu_tx_t *tx)
1821 {
1822 	objset_t *os = dn->dn_objset;
1823 	dsl_dataset_t *ds = os->os_dsl_dataset;
1824 	int64_t aspace = spa_get_asize(os->os_spa, space);
1825 
1826 	if (ds != NULL) {
1827 		dsl_dir_willuse_space(ds->ds_dir, aspace, tx);
1828 		dsl_pool_dirty_space(dmu_tx_pool(tx), space, tx);
1829 	}
1830 
1831 	dmu_tx_willuse_space(tx, aspace);
1832 }
1833 
1834 /*
1835  * Scans a block at the indicated "level" looking for a hole or data,
1836  * depending on 'flags'.
1837  *
1838  * If level > 0, then we are scanning an indirect block looking at its
1839  * pointers.  If level == 0, then we are looking at a block of dnodes.
1840  *
1841  * If we don't find what we are looking for in the block, we return ESRCH.
1842  * Otherwise, return with *offset pointing to the beginning (if searching
1843  * forwards) or end (if searching backwards) of the range covered by the
1844  * block pointer we matched on (or dnode).
1845  *
1846  * The basic search algorithm used below by dnode_next_offset() is to
1847  * use this function to search up the block tree (widen the search) until
1848  * we find something (i.e., we don't return ESRCH) and then search back
1849  * down the tree (narrow the search) until we reach our original search
1850  * level.
1851  */
1852 static int
dnode_next_offset_level(dnode_t * dn,int flags,uint64_t * offset,int lvl,uint64_t blkfill,uint64_t txg)1853 dnode_next_offset_level(dnode_t *dn, int flags, uint64_t *offset,
1854     int lvl, uint64_t blkfill, uint64_t txg)
1855 {
1856 	dmu_buf_impl_t *db = NULL;
1857 	void *data = NULL;
1858 	uint64_t epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
1859 	uint64_t epb = 1ULL << epbs;
1860 	uint64_t minfill, maxfill;
1861 	boolean_t hole;
1862 	int i, inc, error, span;
1863 
1864 	dprintf("probing object %llu offset %llx level %d of %u\n",
1865 	    dn->dn_object, *offset, lvl, dn->dn_phys->dn_nlevels);
1866 
1867 	hole = ((flags & DNODE_FIND_HOLE) != 0);
1868 	inc = (flags & DNODE_FIND_BACKWARDS) ? -1 : 1;
1869 	ASSERT(txg == 0 || !hole);
1870 
1871 	if (lvl == dn->dn_phys->dn_nlevels) {
1872 		error = 0;
1873 		epb = dn->dn_phys->dn_nblkptr;
1874 		data = dn->dn_phys->dn_blkptr;
1875 	} else {
1876 		uint64_t blkid = dbuf_whichblock(dn, lvl, *offset);
1877 		error = dbuf_hold_impl(dn, lvl, blkid, TRUE, FALSE, FTAG, &db);
1878 		if (error) {
1879 			if (error != ENOENT)
1880 				return (error);
1881 			if (hole)
1882 				return (0);
1883 			/*
1884 			 * This can only happen when we are searching up
1885 			 * the block tree for data.  We don't really need to
1886 			 * adjust the offset, as we will just end up looking
1887 			 * at the pointer to this block in its parent, and its
1888 			 * going to be unallocated, so we will skip over it.
1889 			 */
1890 			return (SET_ERROR(ESRCH));
1891 		}
1892 		error = dbuf_read(db, NULL, DB_RF_CANFAIL | DB_RF_HAVESTRUCT);
1893 		if (error) {
1894 			dbuf_rele(db, FTAG);
1895 			return (error);
1896 		}
1897 		data = db->db.db_data;
1898 	}
1899 
1900 
1901 	if (db != NULL && txg != 0 && (db->db_blkptr == NULL ||
1902 	    db->db_blkptr->blk_birth <= txg ||
1903 	    BP_IS_HOLE(db->db_blkptr))) {
1904 		/*
1905 		 * This can only happen when we are searching up the tree
1906 		 * and these conditions mean that we need to keep climbing.
1907 		 */
1908 		error = SET_ERROR(ESRCH);
1909 	} else if (lvl == 0) {
1910 		dnode_phys_t *dnp = data;
1911 		span = DNODE_SHIFT;
1912 		ASSERT(dn->dn_type == DMU_OT_DNODE);
1913 
1914 		for (i = (*offset >> span) & (blkfill - 1);
1915 		    i >= 0 && i < blkfill; i += inc) {
1916 			if ((dnp[i].dn_type == DMU_OT_NONE) == hole)
1917 				break;
1918 			*offset += (1ULL << span) * inc;
1919 		}
1920 		if (i < 0 || i == blkfill)
1921 			error = SET_ERROR(ESRCH);
1922 	} else {
1923 		blkptr_t *bp = data;
1924 		uint64_t start = *offset;
1925 		span = (lvl - 1) * epbs + dn->dn_datablkshift;
1926 		minfill = 0;
1927 		maxfill = blkfill << ((lvl - 1) * epbs);
1928 
1929 		if (hole)
1930 			maxfill--;
1931 		else
1932 			minfill++;
1933 
1934 		*offset = *offset >> span;
1935 		for (i = BF64_GET(*offset, 0, epbs);
1936 		    i >= 0 && i < epb; i += inc) {
1937 			if (BP_GET_FILL(&bp[i]) >= minfill &&
1938 			    BP_GET_FILL(&bp[i]) <= maxfill &&
1939 			    (hole || bp[i].blk_birth > txg))
1940 				break;
1941 			if (inc > 0 || *offset > 0)
1942 				*offset += inc;
1943 		}
1944 		*offset = *offset << span;
1945 		if (inc < 0) {
1946 			/* traversing backwards; position offset at the end */
1947 			ASSERT3U(*offset, <=, start);
1948 			*offset = MIN(*offset + (1ULL << span) - 1, start);
1949 		} else if (*offset < start) {
1950 			*offset = start;
1951 		}
1952 		if (i < 0 || i >= epb)
1953 			error = SET_ERROR(ESRCH);
1954 	}
1955 
1956 	if (db)
1957 		dbuf_rele(db, FTAG);
1958 
1959 	return (error);
1960 }
1961 
1962 /*
1963  * Find the next hole, data, or sparse region at or after *offset.
1964  * The value 'blkfill' tells us how many items we expect to find
1965  * in an L0 data block; this value is 1 for normal objects,
1966  * DNODES_PER_BLOCK for the meta dnode, and some fraction of
1967  * DNODES_PER_BLOCK when searching for sparse regions thereof.
1968  *
1969  * Examples:
1970  *
1971  * dnode_next_offset(dn, flags, offset, 1, 1, 0);
1972  *	Finds the next/previous hole/data in a file.
1973  *	Used in dmu_offset_next().
1974  *
1975  * dnode_next_offset(mdn, flags, offset, 0, DNODES_PER_BLOCK, txg);
1976  *	Finds the next free/allocated dnode an objset's meta-dnode.
1977  *	Only finds objects that have new contents since txg (ie.
1978  *	bonus buffer changes and content removal are ignored).
1979  *	Used in dmu_object_next().
1980  *
1981  * dnode_next_offset(mdn, DNODE_FIND_HOLE, offset, 2, DNODES_PER_BLOCK >> 2, 0);
1982  *	Finds the next L2 meta-dnode bp that's at most 1/4 full.
1983  *	Used in dmu_object_alloc().
1984  */
1985 int
dnode_next_offset(dnode_t * dn,int flags,uint64_t * offset,int minlvl,uint64_t blkfill,uint64_t txg)1986 dnode_next_offset(dnode_t *dn, int flags, uint64_t *offset,
1987     int minlvl, uint64_t blkfill, uint64_t txg)
1988 {
1989 	uint64_t initial_offset = *offset;
1990 	int lvl, maxlvl;
1991 	int error = 0;
1992 
1993 	if (!(flags & DNODE_FIND_HAVELOCK))
1994 		rw_enter(&dn->dn_struct_rwlock, RW_READER);
1995 
1996 	if (dn->dn_phys->dn_nlevels == 0) {
1997 		error = SET_ERROR(ESRCH);
1998 		goto out;
1999 	}
2000 
2001 	if (dn->dn_datablkshift == 0) {
2002 		if (*offset < dn->dn_datablksz) {
2003 			if (flags & DNODE_FIND_HOLE)
2004 				*offset = dn->dn_datablksz;
2005 		} else {
2006 			error = SET_ERROR(ESRCH);
2007 		}
2008 		goto out;
2009 	}
2010 
2011 	maxlvl = dn->dn_phys->dn_nlevels;
2012 
2013 	for (lvl = minlvl; lvl <= maxlvl; lvl++) {
2014 		error = dnode_next_offset_level(dn,
2015 		    flags, offset, lvl, blkfill, txg);
2016 		if (error != ESRCH)
2017 			break;
2018 	}
2019 
2020 	while (error == 0 && --lvl >= minlvl) {
2021 		error = dnode_next_offset_level(dn,
2022 		    flags, offset, lvl, blkfill, txg);
2023 	}
2024 
2025 	/*
2026 	 * There's always a "virtual hole" at the end of the object, even
2027 	 * if all BP's which physically exist are non-holes.
2028 	 */
2029 	if ((flags & DNODE_FIND_HOLE) && error == ESRCH && txg == 0 &&
2030 	    minlvl == 1 && blkfill == 1 && !(flags & DNODE_FIND_BACKWARDS)) {
2031 		error = 0;
2032 	}
2033 
2034 	if (error == 0 && (flags & DNODE_FIND_BACKWARDS ?
2035 	    initial_offset < *offset : initial_offset > *offset))
2036 		error = SET_ERROR(ESRCH);
2037 out:
2038 	if (!(flags & DNODE_FIND_HAVELOCK))
2039 		rw_exit(&dn->dn_struct_rwlock);
2040 
2041 	return (error);
2042 }
2043