xref: /titanic_41/usr/src/uts/common/fs/zfs/dnode.c (revision c4cb95767ae6931f185dc5c02edb165c2ff75b10)
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, 2015 by Delphix. All rights reserved.
24  * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved.
25  */
26 
27 #include <sys/zfs_context.h>
28 #include <sys/dbuf.h>
29 #include <sys/dnode.h>
30 #include <sys/dmu.h>
31 #include <sys/dmu_impl.h>
32 #include <sys/dmu_tx.h>
33 #include <sys/dmu_objset.h>
34 #include <sys/dsl_dir.h>
35 #include <sys/dsl_dataset.h>
36 #include <sys/spa.h>
37 #include <sys/zio.h>
38 #include <sys/dmu_zfetch.h>
39 #include <sys/range_tree.h>
40 
41 static kmem_cache_t *dnode_cache;
42 /*
43  * Define DNODE_STATS to turn on statistic gathering. By default, it is only
44  * turned on when DEBUG is also defined.
45  */
46 #ifdef	DEBUG
47 #define	DNODE_STATS
48 #endif	/* DEBUG */
49 
50 #ifdef	DNODE_STATS
51 #define	DNODE_STAT_ADD(stat)			((stat)++)
52 #else
53 #define	DNODE_STAT_ADD(stat)			/* nothing */
54 #endif	/* DNODE_STATS */
55 
56 static dnode_phys_t dnode_phys_zero;
57 
58 int zfs_default_bs = SPA_MINBLOCKSHIFT;
59 int zfs_default_ibs = DN_MAX_INDBLKSHIFT;
60 
61 static kmem_cbrc_t dnode_move(void *, void *, size_t, void *);
62 
63 static int
dbuf_compare(const void * x1,const void * x2)64 dbuf_compare(const void *x1, const void *x2)
65 {
66 	const dmu_buf_impl_t *d1 = x1;
67 	const dmu_buf_impl_t *d2 = x2;
68 
69 	if (d1->db_level < d2->db_level) {
70 		return (-1);
71 	}
72 	if (d1->db_level > d2->db_level) {
73 		return (1);
74 	}
75 
76 	if (d1->db_blkid < d2->db_blkid) {
77 		return (-1);
78 	}
79 	if (d1->db_blkid > d2->db_blkid) {
80 		return (1);
81 	}
82 
83 	if (d1->db_state == DB_SEARCH) {
84 		ASSERT3S(d2->db_state, !=, DB_SEARCH);
85 		return (-1);
86 	} else if (d2->db_state == DB_SEARCH) {
87 		ASSERT3S(d1->db_state, !=, DB_SEARCH);
88 		return (1);
89 	}
90 
91 	if ((uintptr_t)d1 < (uintptr_t)d2) {
92 		return (-1);
93 	}
94 	if ((uintptr_t)d1 > (uintptr_t)d2) {
95 		return (1);
96 	}
97 	return (0);
98 }
99 
100 /* ARGSUSED */
101 static int
dnode_cons(void * arg,void * unused,int kmflag)102 dnode_cons(void *arg, void *unused, int kmflag)
103 {
104 	dnode_t *dn = arg;
105 	int i;
106 
107 	rw_init(&dn->dn_struct_rwlock, NULL, RW_DEFAULT, NULL);
108 	mutex_init(&dn->dn_mtx, NULL, MUTEX_DEFAULT, NULL);
109 	mutex_init(&dn->dn_dbufs_mtx, NULL, MUTEX_DEFAULT, NULL);
110 	cv_init(&dn->dn_notxholds, NULL, CV_DEFAULT, NULL);
111 
112 	/*
113 	 * Every dbuf has a reference, and dropping a tracked reference is
114 	 * O(number of references), so don't track dn_holds.
115 	 */
116 	refcount_create_untracked(&dn->dn_holds);
117 	refcount_create(&dn->dn_tx_holds);
118 	list_link_init(&dn->dn_link);
119 
120 	bzero(&dn->dn_next_nblkptr[0], sizeof (dn->dn_next_nblkptr));
121 	bzero(&dn->dn_next_nlevels[0], sizeof (dn->dn_next_nlevels));
122 	bzero(&dn->dn_next_indblkshift[0], sizeof (dn->dn_next_indblkshift));
123 	bzero(&dn->dn_next_bonustype[0], sizeof (dn->dn_next_bonustype));
124 	bzero(&dn->dn_rm_spillblk[0], sizeof (dn->dn_rm_spillblk));
125 	bzero(&dn->dn_next_bonuslen[0], sizeof (dn->dn_next_bonuslen));
126 	bzero(&dn->dn_next_blksz[0], sizeof (dn->dn_next_blksz));
127 
128 	for (i = 0; i < TXG_SIZE; i++) {
129 		list_link_init(&dn->dn_dirty_link[i]);
130 		dn->dn_free_ranges[i] = NULL;
131 		list_create(&dn->dn_dirty_records[i],
132 		    sizeof (dbuf_dirty_record_t),
133 		    offsetof(dbuf_dirty_record_t, dr_dirty_node));
134 	}
135 
136 	dn->dn_allocated_txg = 0;
137 	dn->dn_free_txg = 0;
138 	dn->dn_assigned_txg = 0;
139 	dn->dn_dirtyctx = 0;
140 	dn->dn_dirtyctx_firstset = NULL;
141 	dn->dn_bonus = NULL;
142 	dn->dn_have_spill = B_FALSE;
143 	dn->dn_zio = NULL;
144 	dn->dn_oldused = 0;
145 	dn->dn_oldflags = 0;
146 	dn->dn_olduid = 0;
147 	dn->dn_oldgid = 0;
148 	dn->dn_newuid = 0;
149 	dn->dn_newgid = 0;
150 	dn->dn_id_flags = 0;
151 
152 	dn->dn_dbufs_count = 0;
153 	dn->dn_unlisted_l0_blkid = 0;
154 	avl_create(&dn->dn_dbufs, dbuf_compare, sizeof (dmu_buf_impl_t),
155 	    offsetof(dmu_buf_impl_t, db_link));
156 
157 	dn->dn_moved = 0;
158 	return (0);
159 }
160 
161 /* ARGSUSED */
162 static void
dnode_dest(void * arg,void * unused)163 dnode_dest(void *arg, void *unused)
164 {
165 	int i;
166 	dnode_t *dn = arg;
167 
168 	rw_destroy(&dn->dn_struct_rwlock);
169 	mutex_destroy(&dn->dn_mtx);
170 	mutex_destroy(&dn->dn_dbufs_mtx);
171 	cv_destroy(&dn->dn_notxholds);
172 	refcount_destroy(&dn->dn_holds);
173 	refcount_destroy(&dn->dn_tx_holds);
174 	ASSERT(!list_link_active(&dn->dn_link));
175 
176 	for (i = 0; i < TXG_SIZE; i++) {
177 		ASSERT(!list_link_active(&dn->dn_dirty_link[i]));
178 		ASSERT3P(dn->dn_free_ranges[i], ==, NULL);
179 		list_destroy(&dn->dn_dirty_records[i]);
180 		ASSERT0(dn->dn_next_nblkptr[i]);
181 		ASSERT0(dn->dn_next_nlevels[i]);
182 		ASSERT0(dn->dn_next_indblkshift[i]);
183 		ASSERT0(dn->dn_next_bonustype[i]);
184 		ASSERT0(dn->dn_rm_spillblk[i]);
185 		ASSERT0(dn->dn_next_bonuslen[i]);
186 		ASSERT0(dn->dn_next_blksz[i]);
187 	}
188 
189 	ASSERT0(dn->dn_allocated_txg);
190 	ASSERT0(dn->dn_free_txg);
191 	ASSERT0(dn->dn_assigned_txg);
192 	ASSERT0(dn->dn_dirtyctx);
193 	ASSERT3P(dn->dn_dirtyctx_firstset, ==, NULL);
194 	ASSERT3P(dn->dn_bonus, ==, NULL);
195 	ASSERT(!dn->dn_have_spill);
196 	ASSERT3P(dn->dn_zio, ==, NULL);
197 	ASSERT0(dn->dn_oldused);
198 	ASSERT0(dn->dn_oldflags);
199 	ASSERT0(dn->dn_olduid);
200 	ASSERT0(dn->dn_oldgid);
201 	ASSERT0(dn->dn_newuid);
202 	ASSERT0(dn->dn_newgid);
203 	ASSERT0(dn->dn_id_flags);
204 
205 	ASSERT0(dn->dn_dbufs_count);
206 	ASSERT0(dn->dn_unlisted_l0_blkid);
207 	avl_destroy(&dn->dn_dbufs);
208 }
209 
210 void
dnode_init(void)211 dnode_init(void)
212 {
213 	ASSERT(dnode_cache == NULL);
214 	dnode_cache = kmem_cache_create("dnode_t",
215 	    sizeof (dnode_t),
216 	    0, dnode_cons, dnode_dest, NULL, NULL, NULL, 0);
217 	kmem_cache_set_move(dnode_cache, dnode_move);
218 }
219 
220 void
dnode_fini(void)221 dnode_fini(void)
222 {
223 	kmem_cache_destroy(dnode_cache);
224 	dnode_cache = NULL;
225 }
226 
227 
228 #ifdef ZFS_DEBUG
229 void
dnode_verify(dnode_t * dn)230 dnode_verify(dnode_t *dn)
231 {
232 	int drop_struct_lock = FALSE;
233 
234 	ASSERT(dn->dn_phys);
235 	ASSERT(dn->dn_objset);
236 	ASSERT(dn->dn_handle->dnh_dnode == dn);
237 
238 	ASSERT(DMU_OT_IS_VALID(dn->dn_phys->dn_type));
239 
240 	if (!(zfs_flags & ZFS_DEBUG_DNODE_VERIFY))
241 		return;
242 
243 	if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
244 		rw_enter(&dn->dn_struct_rwlock, RW_READER);
245 		drop_struct_lock = TRUE;
246 	}
247 	if (dn->dn_phys->dn_type != DMU_OT_NONE || dn->dn_allocated_txg != 0) {
248 		int i;
249 		ASSERT3U(dn->dn_indblkshift, >=, 0);
250 		ASSERT3U(dn->dn_indblkshift, <=, SPA_MAXBLOCKSHIFT);
251 		if (dn->dn_datablkshift) {
252 			ASSERT3U(dn->dn_datablkshift, >=, SPA_MINBLOCKSHIFT);
253 			ASSERT3U(dn->dn_datablkshift, <=, SPA_MAXBLOCKSHIFT);
254 			ASSERT3U(1<<dn->dn_datablkshift, ==, dn->dn_datablksz);
255 		}
256 		ASSERT3U(dn->dn_nlevels, <=, 30);
257 		ASSERT(DMU_OT_IS_VALID(dn->dn_type));
258 		ASSERT3U(dn->dn_nblkptr, >=, 1);
259 		ASSERT3U(dn->dn_nblkptr, <=, DN_MAX_NBLKPTR);
260 		ASSERT3U(dn->dn_bonuslen, <=, DN_MAX_BONUSLEN);
261 		ASSERT3U(dn->dn_datablksz, ==,
262 		    dn->dn_datablkszsec << SPA_MINBLOCKSHIFT);
263 		ASSERT3U(ISP2(dn->dn_datablksz), ==, dn->dn_datablkshift != 0);
264 		ASSERT3U((dn->dn_nblkptr - 1) * sizeof (blkptr_t) +
265 		    dn->dn_bonuslen, <=, DN_MAX_BONUSLEN);
266 		for (i = 0; i < TXG_SIZE; i++) {
267 			ASSERT3U(dn->dn_next_nlevels[i], <=, dn->dn_nlevels);
268 		}
269 	}
270 	if (dn->dn_phys->dn_type != DMU_OT_NONE)
271 		ASSERT3U(dn->dn_phys->dn_nlevels, <=, dn->dn_nlevels);
272 	ASSERT(DMU_OBJECT_IS_SPECIAL(dn->dn_object) || dn->dn_dbuf != NULL);
273 	if (dn->dn_dbuf != NULL) {
274 		ASSERT3P(dn->dn_phys, ==,
275 		    (dnode_phys_t *)dn->dn_dbuf->db.db_data +
276 		    (dn->dn_object % (dn->dn_dbuf->db.db_size >> DNODE_SHIFT)));
277 	}
278 	if (drop_struct_lock)
279 		rw_exit(&dn->dn_struct_rwlock);
280 }
281 #endif
282 
283 void
dnode_byteswap(dnode_phys_t * dnp)284 dnode_byteswap(dnode_phys_t *dnp)
285 {
286 	uint64_t *buf64 = (void*)&dnp->dn_blkptr;
287 	int i;
288 
289 	if (dnp->dn_type == DMU_OT_NONE) {
290 		bzero(dnp, sizeof (dnode_phys_t));
291 		return;
292 	}
293 
294 	dnp->dn_datablkszsec = BSWAP_16(dnp->dn_datablkszsec);
295 	dnp->dn_bonuslen = BSWAP_16(dnp->dn_bonuslen);
296 	dnp->dn_maxblkid = BSWAP_64(dnp->dn_maxblkid);
297 	dnp->dn_used = BSWAP_64(dnp->dn_used);
298 
299 	/*
300 	 * dn_nblkptr is only one byte, so it's OK to read it in either
301 	 * byte order.  We can't read dn_bouslen.
302 	 */
303 	ASSERT(dnp->dn_indblkshift <= SPA_MAXBLOCKSHIFT);
304 	ASSERT(dnp->dn_nblkptr <= DN_MAX_NBLKPTR);
305 	for (i = 0; i < dnp->dn_nblkptr * sizeof (blkptr_t)/8; i++)
306 		buf64[i] = BSWAP_64(buf64[i]);
307 
308 	/*
309 	 * OK to check dn_bonuslen for zero, because it won't matter if
310 	 * we have the wrong byte order.  This is necessary because the
311 	 * dnode dnode is smaller than a regular dnode.
312 	 */
313 	if (dnp->dn_bonuslen != 0) {
314 		/*
315 		 * Note that the bonus length calculated here may be
316 		 * longer than the actual bonus buffer.  This is because
317 		 * we always put the bonus buffer after the last block
318 		 * pointer (instead of packing it against the end of the
319 		 * dnode buffer).
320 		 */
321 		int off = (dnp->dn_nblkptr-1) * sizeof (blkptr_t);
322 		size_t len = DN_MAX_BONUSLEN - off;
323 		ASSERT(DMU_OT_IS_VALID(dnp->dn_bonustype));
324 		dmu_object_byteswap_t byteswap =
325 		    DMU_OT_BYTESWAP(dnp->dn_bonustype);
326 		dmu_ot_byteswap[byteswap].ob_func(dnp->dn_bonus + off, len);
327 	}
328 
329 	/* Swap SPILL block if we have one */
330 	if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR)
331 		byteswap_uint64_array(&dnp->dn_spill, sizeof (blkptr_t));
332 
333 }
334 
335 void
dnode_buf_byteswap(void * vbuf,size_t size)336 dnode_buf_byteswap(void *vbuf, size_t size)
337 {
338 	dnode_phys_t *buf = vbuf;
339 	int i;
340 
341 	ASSERT3U(sizeof (dnode_phys_t), ==, (1<<DNODE_SHIFT));
342 	ASSERT((size & (sizeof (dnode_phys_t)-1)) == 0);
343 
344 	size >>= DNODE_SHIFT;
345 	for (i = 0; i < size; i++) {
346 		dnode_byteswap(buf);
347 		buf++;
348 	}
349 }
350 
351 void
dnode_setbonuslen(dnode_t * dn,int newsize,dmu_tx_t * tx)352 dnode_setbonuslen(dnode_t *dn, int newsize, dmu_tx_t *tx)
353 {
354 	ASSERT3U(refcount_count(&dn->dn_holds), >=, 1);
355 
356 	dnode_setdirty(dn, tx);
357 	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
358 	ASSERT3U(newsize, <=, DN_MAX_BONUSLEN -
359 	    (dn->dn_nblkptr-1) * sizeof (blkptr_t));
360 	dn->dn_bonuslen = newsize;
361 	if (newsize == 0)
362 		dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = DN_ZERO_BONUSLEN;
363 	else
364 		dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = dn->dn_bonuslen;
365 	rw_exit(&dn->dn_struct_rwlock);
366 }
367 
368 void
dnode_setbonus_type(dnode_t * dn,dmu_object_type_t newtype,dmu_tx_t * tx)369 dnode_setbonus_type(dnode_t *dn, dmu_object_type_t newtype, dmu_tx_t *tx)
370 {
371 	ASSERT3U(refcount_count(&dn->dn_holds), >=, 1);
372 	dnode_setdirty(dn, tx);
373 	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
374 	dn->dn_bonustype = newtype;
375 	dn->dn_next_bonustype[tx->tx_txg & TXG_MASK] = dn->dn_bonustype;
376 	rw_exit(&dn->dn_struct_rwlock);
377 }
378 
379 void
dnode_rm_spill(dnode_t * dn,dmu_tx_t * tx)380 dnode_rm_spill(dnode_t *dn, dmu_tx_t *tx)
381 {
382 	ASSERT3U(refcount_count(&dn->dn_holds), >=, 1);
383 	ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
384 	dnode_setdirty(dn, tx);
385 	dn->dn_rm_spillblk[tx->tx_txg&TXG_MASK] = DN_KILL_SPILLBLK;
386 	dn->dn_have_spill = B_FALSE;
387 }
388 
389 static void
dnode_setdblksz(dnode_t * dn,int size)390 dnode_setdblksz(dnode_t *dn, int size)
391 {
392 	ASSERT0(P2PHASE(size, SPA_MINBLOCKSIZE));
393 	ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
394 	ASSERT3U(size, >=, SPA_MINBLOCKSIZE);
395 	ASSERT3U(size >> SPA_MINBLOCKSHIFT, <,
396 	    1<<(sizeof (dn->dn_phys->dn_datablkszsec) * 8));
397 	dn->dn_datablksz = size;
398 	dn->dn_datablkszsec = size >> SPA_MINBLOCKSHIFT;
399 	dn->dn_datablkshift = ISP2(size) ? highbit64(size - 1) : 0;
400 }
401 
402 static dnode_t *
dnode_create(objset_t * os,dnode_phys_t * dnp,dmu_buf_impl_t * db,uint64_t object,dnode_handle_t * dnh)403 dnode_create(objset_t *os, dnode_phys_t *dnp, dmu_buf_impl_t *db,
404     uint64_t object, dnode_handle_t *dnh)
405 {
406 	dnode_t *dn;
407 
408 	dn = kmem_cache_alloc(dnode_cache, KM_SLEEP);
409 	ASSERT(!POINTER_IS_VALID(dn->dn_objset));
410 	dn->dn_moved = 0;
411 
412 	/*
413 	 * Defer setting dn_objset until the dnode is ready to be a candidate
414 	 * for the dnode_move() callback.
415 	 */
416 	dn->dn_object = object;
417 	dn->dn_dbuf = db;
418 	dn->dn_handle = dnh;
419 	dn->dn_phys = dnp;
420 
421 	if (dnp->dn_datablkszsec) {
422 		dnode_setdblksz(dn, dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT);
423 	} else {
424 		dn->dn_datablksz = 0;
425 		dn->dn_datablkszsec = 0;
426 		dn->dn_datablkshift = 0;
427 	}
428 	dn->dn_indblkshift = dnp->dn_indblkshift;
429 	dn->dn_nlevels = dnp->dn_nlevels;
430 	dn->dn_type = dnp->dn_type;
431 	dn->dn_nblkptr = dnp->dn_nblkptr;
432 	dn->dn_checksum = dnp->dn_checksum;
433 	dn->dn_compress = dnp->dn_compress;
434 	dn->dn_bonustype = dnp->dn_bonustype;
435 	dn->dn_bonuslen = dnp->dn_bonuslen;
436 	dn->dn_maxblkid = dnp->dn_maxblkid;
437 	dn->dn_have_spill = ((dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) != 0);
438 	dn->dn_id_flags = 0;
439 
440 	dmu_zfetch_init(&dn->dn_zfetch, dn);
441 
442 	ASSERT(DMU_OT_IS_VALID(dn->dn_phys->dn_type));
443 
444 	mutex_enter(&os->os_lock);
445 	if (dnh->dnh_dnode != NULL) {
446 		/* Lost the allocation race. */
447 		mutex_exit(&os->os_lock);
448 		kmem_cache_free(dnode_cache, dn);
449 		return (dnh->dnh_dnode);
450 	}
451 
452 	/*
453 	 * Exclude special dnodes from os_dnodes so an empty os_dnodes
454 	 * signifies that the special dnodes have no references from
455 	 * their children (the entries in os_dnodes).  This allows
456 	 * dnode_destroy() to easily determine if the last child has
457 	 * been removed and then complete eviction of the objset.
458 	 */
459 	if (!DMU_OBJECT_IS_SPECIAL(object))
460 		list_insert_head(&os->os_dnodes, dn);
461 	membar_producer();
462 
463 	/*
464 	 * Everything else must be valid before assigning dn_objset
465 	 * makes the dnode eligible for dnode_move().
466 	 */
467 	dn->dn_objset = os;
468 
469 	dnh->dnh_dnode = dn;
470 	mutex_exit(&os->os_lock);
471 
472 	arc_space_consume(sizeof (dnode_t), ARC_SPACE_OTHER);
473 	return (dn);
474 }
475 
476 /*
477  * Caller must be holding the dnode handle, which is released upon return.
478  */
479 static void
dnode_destroy(dnode_t * dn)480 dnode_destroy(dnode_t *dn)
481 {
482 	objset_t *os = dn->dn_objset;
483 	boolean_t complete_os_eviction = B_FALSE;
484 
485 	ASSERT((dn->dn_id_flags & DN_ID_NEW_EXIST) == 0);
486 
487 	mutex_enter(&os->os_lock);
488 	POINTER_INVALIDATE(&dn->dn_objset);
489 	if (!DMU_OBJECT_IS_SPECIAL(dn->dn_object)) {
490 		list_remove(&os->os_dnodes, dn);
491 		complete_os_eviction =
492 		    list_is_empty(&os->os_dnodes) &&
493 		    list_link_active(&os->os_evicting_node);
494 	}
495 	mutex_exit(&os->os_lock);
496 
497 	/* the dnode can no longer move, so we can release the handle */
498 	zrl_remove(&dn->dn_handle->dnh_zrlock);
499 
500 	dn->dn_allocated_txg = 0;
501 	dn->dn_free_txg = 0;
502 	dn->dn_assigned_txg = 0;
503 
504 	dn->dn_dirtyctx = 0;
505 	if (dn->dn_dirtyctx_firstset != NULL) {
506 		kmem_free(dn->dn_dirtyctx_firstset, 1);
507 		dn->dn_dirtyctx_firstset = NULL;
508 	}
509 	if (dn->dn_bonus != NULL) {
510 		mutex_enter(&dn->dn_bonus->db_mtx);
511 		dbuf_evict(dn->dn_bonus);
512 		dn->dn_bonus = NULL;
513 	}
514 	dn->dn_zio = NULL;
515 
516 	dn->dn_have_spill = B_FALSE;
517 	dn->dn_oldused = 0;
518 	dn->dn_oldflags = 0;
519 	dn->dn_olduid = 0;
520 	dn->dn_oldgid = 0;
521 	dn->dn_newuid = 0;
522 	dn->dn_newgid = 0;
523 	dn->dn_id_flags = 0;
524 	dn->dn_unlisted_l0_blkid = 0;
525 
526 	dmu_zfetch_rele(&dn->dn_zfetch);
527 	kmem_cache_free(dnode_cache, dn);
528 	arc_space_return(sizeof (dnode_t), ARC_SPACE_OTHER);
529 
530 	if (complete_os_eviction)
531 		dmu_objset_evict_done(os);
532 }
533 
534 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)535 dnode_allocate(dnode_t *dn, dmu_object_type_t ot, int blocksize, int ibs,
536     dmu_object_type_t bonustype, int bonuslen, dmu_tx_t *tx)
537 {
538 	int i;
539 
540 	ASSERT3U(blocksize, <=,
541 	    spa_maxblocksize(dmu_objset_spa(dn->dn_objset)));
542 	if (blocksize == 0)
543 		blocksize = 1 << zfs_default_bs;
544 	else
545 		blocksize = P2ROUNDUP(blocksize, SPA_MINBLOCKSIZE);
546 
547 	if (ibs == 0)
548 		ibs = zfs_default_ibs;
549 
550 	ibs = MIN(MAX(ibs, DN_MIN_INDBLKSHIFT), DN_MAX_INDBLKSHIFT);
551 
552 	dprintf("os=%p obj=%llu txg=%llu blocksize=%d ibs=%d\n", dn->dn_objset,
553 	    dn->dn_object, tx->tx_txg, blocksize, ibs);
554 
555 	ASSERT(dn->dn_type == DMU_OT_NONE);
556 	ASSERT(bcmp(dn->dn_phys, &dnode_phys_zero, sizeof (dnode_phys_t)) == 0);
557 	ASSERT(dn->dn_phys->dn_type == DMU_OT_NONE);
558 	ASSERT(ot != DMU_OT_NONE);
559 	ASSERT(DMU_OT_IS_VALID(ot));
560 	ASSERT((bonustype == DMU_OT_NONE && bonuslen == 0) ||
561 	    (bonustype == DMU_OT_SA && bonuslen == 0) ||
562 	    (bonustype != DMU_OT_NONE && bonuslen != 0));
563 	ASSERT(DMU_OT_IS_VALID(bonustype));
564 	ASSERT3U(bonuslen, <=, DN_MAX_BONUSLEN);
565 	ASSERT(dn->dn_type == DMU_OT_NONE);
566 	ASSERT0(dn->dn_maxblkid);
567 	ASSERT0(dn->dn_allocated_txg);
568 	ASSERT0(dn->dn_assigned_txg);
569 	ASSERT(refcount_is_zero(&dn->dn_tx_holds));
570 	ASSERT3U(refcount_count(&dn->dn_holds), <=, 1);
571 	ASSERT(avl_is_empty(&dn->dn_dbufs));
572 
573 	for (i = 0; i < TXG_SIZE; i++) {
574 		ASSERT0(dn->dn_next_nblkptr[i]);
575 		ASSERT0(dn->dn_next_nlevels[i]);
576 		ASSERT0(dn->dn_next_indblkshift[i]);
577 		ASSERT0(dn->dn_next_bonuslen[i]);
578 		ASSERT0(dn->dn_next_bonustype[i]);
579 		ASSERT0(dn->dn_rm_spillblk[i]);
580 		ASSERT0(dn->dn_next_blksz[i]);
581 		ASSERT(!list_link_active(&dn->dn_dirty_link[i]));
582 		ASSERT3P(list_head(&dn->dn_dirty_records[i]), ==, NULL);
583 		ASSERT3P(dn->dn_free_ranges[i], ==, NULL);
584 	}
585 
586 	dn->dn_type = ot;
587 	dnode_setdblksz(dn, blocksize);
588 	dn->dn_indblkshift = ibs;
589 	dn->dn_nlevels = 1;
590 	if (bonustype == DMU_OT_SA) /* Maximize bonus space for SA */
591 		dn->dn_nblkptr = 1;
592 	else
593 		dn->dn_nblkptr = 1 +
594 		    ((DN_MAX_BONUSLEN - bonuslen) >> SPA_BLKPTRSHIFT);
595 	dn->dn_bonustype = bonustype;
596 	dn->dn_bonuslen = bonuslen;
597 	dn->dn_checksum = ZIO_CHECKSUM_INHERIT;
598 	dn->dn_compress = ZIO_COMPRESS_INHERIT;
599 	dn->dn_dirtyctx = 0;
600 
601 	dn->dn_free_txg = 0;
602 	if (dn->dn_dirtyctx_firstset) {
603 		kmem_free(dn->dn_dirtyctx_firstset, 1);
604 		dn->dn_dirtyctx_firstset = NULL;
605 	}
606 
607 	dn->dn_allocated_txg = tx->tx_txg;
608 	dn->dn_id_flags = 0;
609 
610 	dnode_setdirty(dn, tx);
611 	dn->dn_next_indblkshift[tx->tx_txg & TXG_MASK] = ibs;
612 	dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = dn->dn_bonuslen;
613 	dn->dn_next_bonustype[tx->tx_txg & TXG_MASK] = dn->dn_bonustype;
614 	dn->dn_next_blksz[tx->tx_txg & TXG_MASK] = dn->dn_datablksz;
615 }
616 
617 void
dnode_reallocate(dnode_t * dn,dmu_object_type_t ot,int blocksize,dmu_object_type_t bonustype,int bonuslen,dmu_tx_t * tx)618 dnode_reallocate(dnode_t *dn, dmu_object_type_t ot, int blocksize,
619     dmu_object_type_t bonustype, int bonuslen, dmu_tx_t *tx)
620 {
621 	int nblkptr;
622 
623 	ASSERT3U(blocksize, >=, SPA_MINBLOCKSIZE);
624 	ASSERT3U(blocksize, <=,
625 	    spa_maxblocksize(dmu_objset_spa(dn->dn_objset)));
626 	ASSERT0(blocksize % SPA_MINBLOCKSIZE);
627 	ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT || dmu_tx_private_ok(tx));
628 	ASSERT(tx->tx_txg != 0);
629 	ASSERT((bonustype == DMU_OT_NONE && bonuslen == 0) ||
630 	    (bonustype != DMU_OT_NONE && bonuslen != 0) ||
631 	    (bonustype == DMU_OT_SA && bonuslen == 0));
632 	ASSERT(DMU_OT_IS_VALID(bonustype));
633 	ASSERT3U(bonuslen, <=, DN_MAX_BONUSLEN);
634 
635 	/* clean up any unreferenced dbufs */
636 	dnode_evict_dbufs(dn);
637 
638 	dn->dn_id_flags = 0;
639 
640 	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
641 	dnode_setdirty(dn, tx);
642 	if (dn->dn_datablksz != blocksize) {
643 		/* change blocksize */
644 		ASSERT(dn->dn_maxblkid == 0 &&
645 		    (BP_IS_HOLE(&dn->dn_phys->dn_blkptr[0]) ||
646 		    dnode_block_freed(dn, 0)));
647 		dnode_setdblksz(dn, blocksize);
648 		dn->dn_next_blksz[tx->tx_txg&TXG_MASK] = blocksize;
649 	}
650 	if (dn->dn_bonuslen != bonuslen)
651 		dn->dn_next_bonuslen[tx->tx_txg&TXG_MASK] = bonuslen;
652 
653 	if (bonustype == DMU_OT_SA) /* Maximize bonus space for SA */
654 		nblkptr = 1;
655 	else
656 		nblkptr = 1 + ((DN_MAX_BONUSLEN - bonuslen) >> SPA_BLKPTRSHIFT);
657 	if (dn->dn_bonustype != bonustype)
658 		dn->dn_next_bonustype[tx->tx_txg&TXG_MASK] = bonustype;
659 	if (dn->dn_nblkptr != nblkptr)
660 		dn->dn_next_nblkptr[tx->tx_txg&TXG_MASK] = nblkptr;
661 	if (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
662 		dbuf_rm_spill(dn, tx);
663 		dnode_rm_spill(dn, tx);
664 	}
665 	rw_exit(&dn->dn_struct_rwlock);
666 
667 	/* change type */
668 	dn->dn_type = ot;
669 
670 	/* change bonus size and type */
671 	mutex_enter(&dn->dn_mtx);
672 	dn->dn_bonustype = bonustype;
673 	dn->dn_bonuslen = bonuslen;
674 	dn->dn_nblkptr = nblkptr;
675 	dn->dn_checksum = ZIO_CHECKSUM_INHERIT;
676 	dn->dn_compress = ZIO_COMPRESS_INHERIT;
677 	ASSERT3U(dn->dn_nblkptr, <=, DN_MAX_NBLKPTR);
678 
679 	/* fix up the bonus db_size */
680 	if (dn->dn_bonus) {
681 		dn->dn_bonus->db.db_size =
682 		    DN_MAX_BONUSLEN - (dn->dn_nblkptr-1) * sizeof (blkptr_t);
683 		ASSERT(dn->dn_bonuslen <= dn->dn_bonus->db.db_size);
684 	}
685 
686 	dn->dn_allocated_txg = tx->tx_txg;
687 	mutex_exit(&dn->dn_mtx);
688 }
689 
690 #ifdef	DNODE_STATS
691 static struct {
692 	uint64_t dms_dnode_invalid;
693 	uint64_t dms_dnode_recheck1;
694 	uint64_t dms_dnode_recheck2;
695 	uint64_t dms_dnode_special;
696 	uint64_t dms_dnode_handle;
697 	uint64_t dms_dnode_rwlock;
698 	uint64_t dms_dnode_active;
699 } dnode_move_stats;
700 #endif	/* DNODE_STATS */
701 
702 static void
dnode_move_impl(dnode_t * odn,dnode_t * ndn)703 dnode_move_impl(dnode_t *odn, dnode_t *ndn)
704 {
705 	int i;
706 
707 	ASSERT(!RW_LOCK_HELD(&odn->dn_struct_rwlock));
708 	ASSERT(MUTEX_NOT_HELD(&odn->dn_mtx));
709 	ASSERT(MUTEX_NOT_HELD(&odn->dn_dbufs_mtx));
710 	ASSERT(!RW_LOCK_HELD(&odn->dn_zfetch.zf_rwlock));
711 
712 	/* Copy fields. */
713 	ndn->dn_objset = odn->dn_objset;
714 	ndn->dn_object = odn->dn_object;
715 	ndn->dn_dbuf = odn->dn_dbuf;
716 	ndn->dn_handle = odn->dn_handle;
717 	ndn->dn_phys = odn->dn_phys;
718 	ndn->dn_type = odn->dn_type;
719 	ndn->dn_bonuslen = odn->dn_bonuslen;
720 	ndn->dn_bonustype = odn->dn_bonustype;
721 	ndn->dn_nblkptr = odn->dn_nblkptr;
722 	ndn->dn_checksum = odn->dn_checksum;
723 	ndn->dn_compress = odn->dn_compress;
724 	ndn->dn_nlevels = odn->dn_nlevels;
725 	ndn->dn_indblkshift = odn->dn_indblkshift;
726 	ndn->dn_datablkshift = odn->dn_datablkshift;
727 	ndn->dn_datablkszsec = odn->dn_datablkszsec;
728 	ndn->dn_datablksz = odn->dn_datablksz;
729 	ndn->dn_maxblkid = odn->dn_maxblkid;
730 	bcopy(&odn->dn_next_nblkptr[0], &ndn->dn_next_nblkptr[0],
731 	    sizeof (odn->dn_next_nblkptr));
732 	bcopy(&odn->dn_next_nlevels[0], &ndn->dn_next_nlevels[0],
733 	    sizeof (odn->dn_next_nlevels));
734 	bcopy(&odn->dn_next_indblkshift[0], &ndn->dn_next_indblkshift[0],
735 	    sizeof (odn->dn_next_indblkshift));
736 	bcopy(&odn->dn_next_bonustype[0], &ndn->dn_next_bonustype[0],
737 	    sizeof (odn->dn_next_bonustype));
738 	bcopy(&odn->dn_rm_spillblk[0], &ndn->dn_rm_spillblk[0],
739 	    sizeof (odn->dn_rm_spillblk));
740 	bcopy(&odn->dn_next_bonuslen[0], &ndn->dn_next_bonuslen[0],
741 	    sizeof (odn->dn_next_bonuslen));
742 	bcopy(&odn->dn_next_blksz[0], &ndn->dn_next_blksz[0],
743 	    sizeof (odn->dn_next_blksz));
744 	for (i = 0; i < TXG_SIZE; i++) {
745 		list_move_tail(&ndn->dn_dirty_records[i],
746 		    &odn->dn_dirty_records[i]);
747 	}
748 	bcopy(&odn->dn_free_ranges[0], &ndn->dn_free_ranges[0],
749 	    sizeof (odn->dn_free_ranges));
750 	ndn->dn_allocated_txg = odn->dn_allocated_txg;
751 	ndn->dn_free_txg = odn->dn_free_txg;
752 	ndn->dn_assigned_txg = odn->dn_assigned_txg;
753 	ndn->dn_dirtyctx = odn->dn_dirtyctx;
754 	ndn->dn_dirtyctx_firstset = odn->dn_dirtyctx_firstset;
755 	ASSERT(refcount_count(&odn->dn_tx_holds) == 0);
756 	refcount_transfer(&ndn->dn_holds, &odn->dn_holds);
757 	ASSERT(avl_is_empty(&ndn->dn_dbufs));
758 	avl_swap(&ndn->dn_dbufs, &odn->dn_dbufs);
759 	ndn->dn_dbufs_count = odn->dn_dbufs_count;
760 	ndn->dn_unlisted_l0_blkid = odn->dn_unlisted_l0_blkid;
761 	ndn->dn_bonus = odn->dn_bonus;
762 	ndn->dn_have_spill = odn->dn_have_spill;
763 	ndn->dn_zio = odn->dn_zio;
764 	ndn->dn_oldused = odn->dn_oldused;
765 	ndn->dn_oldflags = odn->dn_oldflags;
766 	ndn->dn_olduid = odn->dn_olduid;
767 	ndn->dn_oldgid = odn->dn_oldgid;
768 	ndn->dn_newuid = odn->dn_newuid;
769 	ndn->dn_newgid = odn->dn_newgid;
770 	ndn->dn_id_flags = odn->dn_id_flags;
771 	dmu_zfetch_init(&ndn->dn_zfetch, NULL);
772 	list_move_tail(&ndn->dn_zfetch.zf_stream, &odn->dn_zfetch.zf_stream);
773 	ndn->dn_zfetch.zf_dnode = odn->dn_zfetch.zf_dnode;
774 	ndn->dn_zfetch.zf_stream_cnt = odn->dn_zfetch.zf_stream_cnt;
775 	ndn->dn_zfetch.zf_alloc_fail = odn->dn_zfetch.zf_alloc_fail;
776 
777 	/*
778 	 * Update back pointers. Updating the handle fixes the back pointer of
779 	 * every descendant dbuf as well as the bonus dbuf.
780 	 */
781 	ASSERT(ndn->dn_handle->dnh_dnode == odn);
782 	ndn->dn_handle->dnh_dnode = ndn;
783 	if (ndn->dn_zfetch.zf_dnode == odn) {
784 		ndn->dn_zfetch.zf_dnode = ndn;
785 	}
786 
787 	/*
788 	 * Invalidate the original dnode by clearing all of its back pointers.
789 	 */
790 	odn->dn_dbuf = NULL;
791 	odn->dn_handle = NULL;
792 	avl_create(&odn->dn_dbufs, dbuf_compare, sizeof (dmu_buf_impl_t),
793 	    offsetof(dmu_buf_impl_t, db_link));
794 	odn->dn_dbufs_count = 0;
795 	odn->dn_unlisted_l0_blkid = 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	_KERNEL
843 /*ARGSUSED*/
844 static kmem_cbrc_t
dnode_move(void * buf,void * newbuf,size_t size,void * arg)845 dnode_move(void *buf, void *newbuf, size_t size, void *arg)
846 {
847 	dnode_t *odn = buf, *ndn = newbuf;
848 	objset_t *os;
849 	int64_t refcount;
850 	uint32_t dbufs;
851 
852 	/*
853 	 * The dnode is on the objset's list of known dnodes if the objset
854 	 * pointer is valid. We set the low bit of the objset pointer when
855 	 * freeing the dnode to invalidate it, and the memory patterns written
856 	 * by kmem (baddcafe and deadbeef) set at least one of the two low bits.
857 	 * A newly created dnode sets the objset pointer last of all to indicate
858 	 * that the dnode is known and in a valid state to be moved by this
859 	 * function.
860 	 */
861 	os = odn->dn_objset;
862 	if (!POINTER_IS_VALID(os)) {
863 		DNODE_STAT_ADD(dnode_move_stats.dms_dnode_invalid);
864 		return (KMEM_CBRC_DONT_KNOW);
865 	}
866 
867 	/*
868 	 * Ensure that the objset does not go away during the move.
869 	 */
870 	rw_enter(&os_lock, RW_WRITER);
871 	if (os != odn->dn_objset) {
872 		rw_exit(&os_lock);
873 		DNODE_STAT_ADD(dnode_move_stats.dms_dnode_recheck1);
874 		return (KMEM_CBRC_DONT_KNOW);
875 	}
876 
877 	/*
878 	 * If the dnode is still valid, then so is the objset. We know that no
879 	 * valid objset can be freed while we hold os_lock, so we can safely
880 	 * ensure that the objset remains in use.
881 	 */
882 	mutex_enter(&os->os_lock);
883 
884 	/*
885 	 * Recheck the objset pointer in case the dnode was removed just before
886 	 * acquiring the lock.
887 	 */
888 	if (os != odn->dn_objset) {
889 		mutex_exit(&os->os_lock);
890 		rw_exit(&os_lock);
891 		DNODE_STAT_ADD(dnode_move_stats.dms_dnode_recheck2);
892 		return (KMEM_CBRC_DONT_KNOW);
893 	}
894 
895 	/*
896 	 * At this point we know that as long as we hold os->os_lock, the dnode
897 	 * cannot be freed and fields within the dnode can be safely accessed.
898 	 * The objset listing this dnode cannot go away as long as this dnode is
899 	 * on its list.
900 	 */
901 	rw_exit(&os_lock);
902 	if (DMU_OBJECT_IS_SPECIAL(odn->dn_object)) {
903 		mutex_exit(&os->os_lock);
904 		DNODE_STAT_ADD(dnode_move_stats.dms_dnode_special);
905 		return (KMEM_CBRC_NO);
906 	}
907 	ASSERT(odn->dn_dbuf != NULL); /* only "special" dnodes have no parent */
908 
909 	/*
910 	 * Lock the dnode handle to prevent the dnode from obtaining any new
911 	 * holds. This also prevents the descendant dbufs and the bonus dbuf
912 	 * from accessing the dnode, so that we can discount their holds. The
913 	 * handle is safe to access because we know that while the dnode cannot
914 	 * go away, neither can its handle. Once we hold dnh_zrlock, we can
915 	 * safely move any dnode referenced only by dbufs.
916 	 */
917 	if (!zrl_tryenter(&odn->dn_handle->dnh_zrlock)) {
918 		mutex_exit(&os->os_lock);
919 		DNODE_STAT_ADD(dnode_move_stats.dms_dnode_handle);
920 		return (KMEM_CBRC_LATER);
921 	}
922 
923 	/*
924 	 * Ensure a consistent view of the dnode's holds and the dnode's dbufs.
925 	 * We need to guarantee that there is a hold for every dbuf in order to
926 	 * determine whether the dnode is actively referenced. Falsely matching
927 	 * a dbuf to an active hold would lead to an unsafe move. It's possible
928 	 * that a thread already having an active dnode hold is about to add a
929 	 * dbuf, and we can't compare hold and dbuf counts while the add is in
930 	 * progress.
931 	 */
932 	if (!rw_tryenter(&odn->dn_struct_rwlock, RW_WRITER)) {
933 		zrl_exit(&odn->dn_handle->dnh_zrlock);
934 		mutex_exit(&os->os_lock);
935 		DNODE_STAT_ADD(dnode_move_stats.dms_dnode_rwlock);
936 		return (KMEM_CBRC_LATER);
937 	}
938 
939 	/*
940 	 * A dbuf may be removed (evicted) without an active dnode hold. In that
941 	 * case, the dbuf count is decremented under the handle lock before the
942 	 * dbuf's hold is released. This order ensures that if we count the hold
943 	 * after the dbuf is removed but before its hold is released, we will
944 	 * treat the unmatched hold as active and exit safely. If we count the
945 	 * hold before the dbuf is removed, the hold is discounted, and the
946 	 * removal is blocked until the move completes.
947 	 */
948 	refcount = refcount_count(&odn->dn_holds);
949 	ASSERT(refcount >= 0);
950 	dbufs = odn->dn_dbufs_count;
951 
952 	/* We can't have more dbufs than dnode holds. */
953 	ASSERT3U(dbufs, <=, refcount);
954 	DTRACE_PROBE3(dnode__move, dnode_t *, odn, int64_t, refcount,
955 	    uint32_t, dbufs);
956 
957 	if (refcount > dbufs) {
958 		rw_exit(&odn->dn_struct_rwlock);
959 		zrl_exit(&odn->dn_handle->dnh_zrlock);
960 		mutex_exit(&os->os_lock);
961 		DNODE_STAT_ADD(dnode_move_stats.dms_dnode_active);
962 		return (KMEM_CBRC_LATER);
963 	}
964 
965 	rw_exit(&odn->dn_struct_rwlock);
966 
967 	/*
968 	 * At this point we know that anyone with a hold on the dnode is not
969 	 * actively referencing it. The dnode is known and in a valid state to
970 	 * move. We're holding the locks needed to execute the critical section.
971 	 */
972 	dnode_move_impl(odn, ndn);
973 
974 	list_link_replace(&odn->dn_link, &ndn->dn_link);
975 	/* If the dnode was safe to move, the refcount cannot have changed. */
976 	ASSERT(refcount == refcount_count(&ndn->dn_holds));
977 	ASSERT(dbufs == ndn->dn_dbufs_count);
978 	zrl_exit(&ndn->dn_handle->dnh_zrlock); /* handle has moved */
979 	mutex_exit(&os->os_lock);
980 
981 	return (KMEM_CBRC_YES);
982 }
983 #endif	/* _KERNEL */
984 
985 void
dnode_special_close(dnode_handle_t * dnh)986 dnode_special_close(dnode_handle_t *dnh)
987 {
988 	dnode_t *dn = dnh->dnh_dnode;
989 
990 	/*
991 	 * Wait for final references to the dnode to clear.  This can
992 	 * only happen if the arc is asyncronously evicting state that
993 	 * has a hold on this dnode while we are trying to evict this
994 	 * dnode.
995 	 */
996 	while (refcount_count(&dn->dn_holds) > 0)
997 		delay(1);
998 	ASSERT(dn->dn_dbuf == NULL ||
999 	    dmu_buf_get_user(&dn->dn_dbuf->db) == NULL);
1000 	zrl_add(&dnh->dnh_zrlock);
1001 	dnode_destroy(dn); /* implicit zrl_remove() */
1002 	zrl_destroy(&dnh->dnh_zrlock);
1003 	dnh->dnh_dnode = NULL;
1004 }
1005 
1006 void
dnode_special_open(objset_t * os,dnode_phys_t * dnp,uint64_t object,dnode_handle_t * dnh)1007 dnode_special_open(objset_t *os, dnode_phys_t *dnp, uint64_t object,
1008     dnode_handle_t *dnh)
1009 {
1010 	dnode_t *dn;
1011 
1012 	dn = dnode_create(os, dnp, NULL, object, dnh);
1013 	zrl_init(&dnh->dnh_zrlock);
1014 	DNODE_VERIFY(dn);
1015 }
1016 
1017 static void
dnode_buf_evict_async(void * dbu)1018 dnode_buf_evict_async(void *dbu)
1019 {
1020 	dnode_children_t *children_dnodes = dbu;
1021 	int i;
1022 
1023 	for (i = 0; i < children_dnodes->dnc_count; i++) {
1024 		dnode_handle_t *dnh = &children_dnodes->dnc_children[i];
1025 		dnode_t *dn;
1026 
1027 		/*
1028 		 * The dnode handle lock guards against the dnode moving to
1029 		 * another valid address, so there is no need here to guard
1030 		 * against changes to or from NULL.
1031 		 */
1032 		if (dnh->dnh_dnode == NULL) {
1033 			zrl_destroy(&dnh->dnh_zrlock);
1034 			continue;
1035 		}
1036 
1037 		zrl_add(&dnh->dnh_zrlock);
1038 		dn = dnh->dnh_dnode;
1039 		/*
1040 		 * If there are holds on this dnode, then there should
1041 		 * be holds on the dnode's containing dbuf as well; thus
1042 		 * it wouldn't be eligible for eviction and this function
1043 		 * would not have been called.
1044 		 */
1045 		ASSERT(refcount_is_zero(&dn->dn_holds));
1046 		ASSERT(refcount_is_zero(&dn->dn_tx_holds));
1047 
1048 		dnode_destroy(dn); /* implicit zrl_remove() */
1049 		zrl_destroy(&dnh->dnh_zrlock);
1050 		dnh->dnh_dnode = NULL;
1051 	}
1052 	kmem_free(children_dnodes, sizeof (dnode_children_t) +
1053 	    children_dnodes->dnc_count * sizeof (dnode_handle_t));
1054 }
1055 
1056 /*
1057  * errors:
1058  * EINVAL - invalid object number.
1059  * EIO - i/o error.
1060  * succeeds even for free dnodes.
1061  */
1062 int
dnode_hold_impl(objset_t * os,uint64_t object,int flag,void * tag,dnode_t ** dnp)1063 dnode_hold_impl(objset_t *os, uint64_t object, int flag,
1064     void *tag, dnode_t **dnp)
1065 {
1066 	int epb, idx, err;
1067 	int drop_struct_lock = FALSE;
1068 	int type;
1069 	uint64_t blk;
1070 	dnode_t *mdn, *dn;
1071 	dmu_buf_impl_t *db;
1072 	dnode_children_t *children_dnodes;
1073 	dnode_handle_t *dnh;
1074 
1075 	/*
1076 	 * If you are holding the spa config lock as writer, you shouldn't
1077 	 * be asking the DMU to do *anything* unless it's the root pool
1078 	 * which may require us to read from the root filesystem while
1079 	 * holding some (not all) of the locks as writer.
1080 	 */
1081 	ASSERT(spa_config_held(os->os_spa, SCL_ALL, RW_WRITER) == 0 ||
1082 	    (spa_is_root(os->os_spa) &&
1083 	    spa_config_held(os->os_spa, SCL_STATE, RW_WRITER)));
1084 
1085 	if (object == DMU_USERUSED_OBJECT || object == DMU_GROUPUSED_OBJECT) {
1086 		dn = (object == DMU_USERUSED_OBJECT) ?
1087 		    DMU_USERUSED_DNODE(os) : DMU_GROUPUSED_DNODE(os);
1088 		if (dn == NULL)
1089 			return (SET_ERROR(ENOENT));
1090 		type = dn->dn_type;
1091 		if ((flag & DNODE_MUST_BE_ALLOCATED) && type == DMU_OT_NONE)
1092 			return (SET_ERROR(ENOENT));
1093 		if ((flag & DNODE_MUST_BE_FREE) && type != DMU_OT_NONE)
1094 			return (SET_ERROR(EEXIST));
1095 		DNODE_VERIFY(dn);
1096 		(void) refcount_add(&dn->dn_holds, tag);
1097 		*dnp = dn;
1098 		return (0);
1099 	}
1100 
1101 	if (object == 0 || object >= DN_MAX_OBJECT)
1102 		return (SET_ERROR(EINVAL));
1103 
1104 	mdn = DMU_META_DNODE(os);
1105 	ASSERT(mdn->dn_object == DMU_META_DNODE_OBJECT);
1106 
1107 	DNODE_VERIFY(mdn);
1108 
1109 	if (!RW_WRITE_HELD(&mdn->dn_struct_rwlock)) {
1110 		rw_enter(&mdn->dn_struct_rwlock, RW_READER);
1111 		drop_struct_lock = TRUE;
1112 	}
1113 
1114 	blk = dbuf_whichblock(mdn, 0, object * sizeof (dnode_phys_t));
1115 
1116 	db = dbuf_hold(mdn, blk, FTAG);
1117 	if (drop_struct_lock)
1118 		rw_exit(&mdn->dn_struct_rwlock);
1119 	if (db == NULL)
1120 		return (SET_ERROR(EIO));
1121 	err = dbuf_read(db, NULL, DB_RF_CANFAIL);
1122 	if (err) {
1123 		dbuf_rele(db, FTAG);
1124 		return (err);
1125 	}
1126 
1127 	ASSERT3U(db->db.db_size, >=, 1<<DNODE_SHIFT);
1128 	epb = db->db.db_size >> DNODE_SHIFT;
1129 
1130 	idx = object & (epb-1);
1131 
1132 	ASSERT(DB_DNODE(db)->dn_type == DMU_OT_DNODE);
1133 	children_dnodes = dmu_buf_get_user(&db->db);
1134 	if (children_dnodes == NULL) {
1135 		int i;
1136 		dnode_children_t *winner;
1137 		children_dnodes = kmem_zalloc(sizeof (dnode_children_t) +
1138 		    epb * sizeof (dnode_handle_t), KM_SLEEP);
1139 		children_dnodes->dnc_count = epb;
1140 		dnh = &children_dnodes->dnc_children[0];
1141 		for (i = 0; i < epb; i++) {
1142 			zrl_init(&dnh[i].dnh_zrlock);
1143 		}
1144 		dmu_buf_init_user(&children_dnodes->dnc_dbu, NULL,
1145 		    dnode_buf_evict_async, NULL);
1146 		winner = dmu_buf_set_user(&db->db, &children_dnodes->dnc_dbu);
1147 		if (winner != NULL) {
1148 
1149 			for (i = 0; i < epb; i++) {
1150 				zrl_destroy(&dnh[i].dnh_zrlock);
1151 			}
1152 
1153 			kmem_free(children_dnodes, sizeof (dnode_children_t) +
1154 			    epb * sizeof (dnode_handle_t));
1155 			children_dnodes = winner;
1156 		}
1157 	}
1158 	ASSERT(children_dnodes->dnc_count == epb);
1159 
1160 	dnh = &children_dnodes->dnc_children[idx];
1161 	zrl_add(&dnh->dnh_zrlock);
1162 	dn = dnh->dnh_dnode;
1163 	if (dn == NULL) {
1164 		dnode_phys_t *phys = (dnode_phys_t *)db->db.db_data+idx;
1165 
1166 		dn = dnode_create(os, phys, db, object, dnh);
1167 	}
1168 
1169 	mutex_enter(&dn->dn_mtx);
1170 	type = dn->dn_type;
1171 	if (dn->dn_free_txg ||
1172 	    ((flag & DNODE_MUST_BE_ALLOCATED) && type == DMU_OT_NONE) ||
1173 	    ((flag & DNODE_MUST_BE_FREE) &&
1174 	    (type != DMU_OT_NONE || !refcount_is_zero(&dn->dn_holds)))) {
1175 		mutex_exit(&dn->dn_mtx);
1176 		zrl_remove(&dnh->dnh_zrlock);
1177 		dbuf_rele(db, FTAG);
1178 		return (type == DMU_OT_NONE ? ENOENT : EEXIST);
1179 	}
1180 	if (refcount_add(&dn->dn_holds, tag) == 1)
1181 		dbuf_add_ref(db, dnh);
1182 	mutex_exit(&dn->dn_mtx);
1183 
1184 	/* Now we can rely on the hold to prevent the dnode from moving. */
1185 	zrl_remove(&dnh->dnh_zrlock);
1186 
1187 	DNODE_VERIFY(dn);
1188 	ASSERT3P(dn->dn_dbuf, ==, db);
1189 	ASSERT3U(dn->dn_object, ==, object);
1190 	dbuf_rele(db, FTAG);
1191 
1192 	*dnp = dn;
1193 	return (0);
1194 }
1195 
1196 /*
1197  * Return held dnode if the object is allocated, NULL if not.
1198  */
1199 int
dnode_hold(objset_t * os,uint64_t object,void * tag,dnode_t ** dnp)1200 dnode_hold(objset_t *os, uint64_t object, void *tag, dnode_t **dnp)
1201 {
1202 	return (dnode_hold_impl(os, object, DNODE_MUST_BE_ALLOCATED, tag, dnp));
1203 }
1204 
1205 /*
1206  * Can only add a reference if there is already at least one
1207  * reference on the dnode.  Returns FALSE if unable to add a
1208  * new reference.
1209  */
1210 boolean_t
dnode_add_ref(dnode_t * dn,void * tag)1211 dnode_add_ref(dnode_t *dn, void *tag)
1212 {
1213 	mutex_enter(&dn->dn_mtx);
1214 	if (refcount_is_zero(&dn->dn_holds)) {
1215 		mutex_exit(&dn->dn_mtx);
1216 		return (FALSE);
1217 	}
1218 	VERIFY(1 < refcount_add(&dn->dn_holds, tag));
1219 	mutex_exit(&dn->dn_mtx);
1220 	return (TRUE);
1221 }
1222 
1223 void
dnode_rele(dnode_t * dn,void * tag)1224 dnode_rele(dnode_t *dn, void *tag)
1225 {
1226 	mutex_enter(&dn->dn_mtx);
1227 	dnode_rele_and_unlock(dn, tag);
1228 }
1229 
1230 void
dnode_rele_and_unlock(dnode_t * dn,void * tag)1231 dnode_rele_and_unlock(dnode_t *dn, void *tag)
1232 {
1233 	uint64_t refs;
1234 	/* Get while the hold prevents the dnode from moving. */
1235 	dmu_buf_impl_t *db = dn->dn_dbuf;
1236 	dnode_handle_t *dnh = dn->dn_handle;
1237 
1238 	refs = refcount_remove(&dn->dn_holds, tag);
1239 	mutex_exit(&dn->dn_mtx);
1240 
1241 	/*
1242 	 * It's unsafe to release the last hold on a dnode by dnode_rele() or
1243 	 * indirectly by dbuf_rele() while relying on the dnode handle to
1244 	 * prevent the dnode from moving, since releasing the last hold could
1245 	 * result in the dnode's parent dbuf evicting its dnode handles. For
1246 	 * that reason anyone calling dnode_rele() or dbuf_rele() without some
1247 	 * other direct or indirect hold on the dnode must first drop the dnode
1248 	 * handle.
1249 	 */
1250 	ASSERT(refs > 0 || dnh->dnh_zrlock.zr_owner != curthread);
1251 
1252 	/* NOTE: the DNODE_DNODE does not have a dn_dbuf */
1253 	if (refs == 0 && db != NULL) {
1254 		/*
1255 		 * Another thread could add a hold to the dnode handle in
1256 		 * dnode_hold_impl() while holding the parent dbuf. Since the
1257 		 * hold on the parent dbuf prevents the handle from being
1258 		 * destroyed, the hold on the handle is OK. We can't yet assert
1259 		 * that the handle has zero references, but that will be
1260 		 * asserted anyway when the handle gets destroyed.
1261 		 */
1262 		dbuf_rele(db, dnh);
1263 	}
1264 }
1265 
1266 void
dnode_setdirty(dnode_t * dn,dmu_tx_t * tx)1267 dnode_setdirty(dnode_t *dn, dmu_tx_t *tx)
1268 {
1269 	objset_t *os = dn->dn_objset;
1270 	uint64_t txg = tx->tx_txg;
1271 
1272 	if (DMU_OBJECT_IS_SPECIAL(dn->dn_object)) {
1273 		dsl_dataset_dirty(os->os_dsl_dataset, tx);
1274 		return;
1275 	}
1276 
1277 	DNODE_VERIFY(dn);
1278 
1279 #ifdef ZFS_DEBUG
1280 	mutex_enter(&dn->dn_mtx);
1281 	ASSERT(dn->dn_phys->dn_type || dn->dn_allocated_txg);
1282 	ASSERT(dn->dn_free_txg == 0 || dn->dn_free_txg >= txg);
1283 	mutex_exit(&dn->dn_mtx);
1284 #endif
1285 
1286 	/*
1287 	 * Determine old uid/gid when necessary
1288 	 */
1289 	dmu_objset_userquota_get_ids(dn, B_TRUE, tx);
1290 
1291 	mutex_enter(&os->os_lock);
1292 
1293 	/*
1294 	 * If we are already marked dirty, we're done.
1295 	 */
1296 	if (list_link_active(&dn->dn_dirty_link[txg & TXG_MASK])) {
1297 		mutex_exit(&os->os_lock);
1298 		return;
1299 	}
1300 
1301 	ASSERT(!refcount_is_zero(&dn->dn_holds) ||
1302 	    !avl_is_empty(&dn->dn_dbufs));
1303 	ASSERT(dn->dn_datablksz != 0);
1304 	ASSERT0(dn->dn_next_bonuslen[txg&TXG_MASK]);
1305 	ASSERT0(dn->dn_next_blksz[txg&TXG_MASK]);
1306 	ASSERT0(dn->dn_next_bonustype[txg&TXG_MASK]);
1307 
1308 	dprintf_ds(os->os_dsl_dataset, "obj=%llu txg=%llu\n",
1309 	    dn->dn_object, txg);
1310 
1311 	if (dn->dn_free_txg > 0 && dn->dn_free_txg <= txg) {
1312 		list_insert_tail(&os->os_free_dnodes[txg&TXG_MASK], dn);
1313 	} else {
1314 		list_insert_tail(&os->os_dirty_dnodes[txg&TXG_MASK], dn);
1315 	}
1316 
1317 	mutex_exit(&os->os_lock);
1318 
1319 	/*
1320 	 * The dnode maintains a hold on its containing dbuf as
1321 	 * long as there are holds on it.  Each instantiated child
1322 	 * dbuf maintains a hold on the dnode.  When the last child
1323 	 * drops its hold, the dnode will drop its hold on the
1324 	 * containing dbuf. We add a "dirty hold" here so that the
1325 	 * dnode will hang around after we finish processing its
1326 	 * children.
1327 	 */
1328 	VERIFY(dnode_add_ref(dn, (void *)(uintptr_t)tx->tx_txg));
1329 
1330 	(void) dbuf_dirty(dn->dn_dbuf, tx);
1331 
1332 	dsl_dataset_dirty(os->os_dsl_dataset, tx);
1333 }
1334 
1335 void
dnode_free(dnode_t * dn,dmu_tx_t * tx)1336 dnode_free(dnode_t *dn, dmu_tx_t *tx)
1337 {
1338 	int txgoff = tx->tx_txg & TXG_MASK;
1339 
1340 	dprintf("dn=%p txg=%llu\n", dn, tx->tx_txg);
1341 
1342 	/* we should be the only holder... hopefully */
1343 	/* ASSERT3U(refcount_count(&dn->dn_holds), ==, 1); */
1344 
1345 	mutex_enter(&dn->dn_mtx);
1346 	if (dn->dn_type == DMU_OT_NONE || dn->dn_free_txg) {
1347 		mutex_exit(&dn->dn_mtx);
1348 		return;
1349 	}
1350 	dn->dn_free_txg = tx->tx_txg;
1351 	mutex_exit(&dn->dn_mtx);
1352 
1353 	/*
1354 	 * If the dnode is already dirty, it needs to be moved from
1355 	 * the dirty list to the free list.
1356 	 */
1357 	mutex_enter(&dn->dn_objset->os_lock);
1358 	if (list_link_active(&dn->dn_dirty_link[txgoff])) {
1359 		list_remove(&dn->dn_objset->os_dirty_dnodes[txgoff], dn);
1360 		list_insert_tail(&dn->dn_objset->os_free_dnodes[txgoff], dn);
1361 		mutex_exit(&dn->dn_objset->os_lock);
1362 	} else {
1363 		mutex_exit(&dn->dn_objset->os_lock);
1364 		dnode_setdirty(dn, tx);
1365 	}
1366 }
1367 
1368 /*
1369  * Try to change the block size for the indicated dnode.  This can only
1370  * succeed if there are no blocks allocated or dirty beyond first block
1371  */
1372 int
dnode_set_blksz(dnode_t * dn,uint64_t size,int ibs,dmu_tx_t * tx)1373 dnode_set_blksz(dnode_t *dn, uint64_t size, int ibs, dmu_tx_t *tx)
1374 {
1375 	dmu_buf_impl_t *db;
1376 	int err;
1377 
1378 	ASSERT3U(size, <=, spa_maxblocksize(dmu_objset_spa(dn->dn_objset)));
1379 	if (size == 0)
1380 		size = SPA_MINBLOCKSIZE;
1381 	else
1382 		size = P2ROUNDUP(size, SPA_MINBLOCKSIZE);
1383 
1384 	if (ibs == dn->dn_indblkshift)
1385 		ibs = 0;
1386 
1387 	if (size >> SPA_MINBLOCKSHIFT == dn->dn_datablkszsec && ibs == 0)
1388 		return (0);
1389 
1390 	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1391 
1392 	/* Check for any allocated blocks beyond the first */
1393 	if (dn->dn_maxblkid != 0)
1394 		goto fail;
1395 
1396 	mutex_enter(&dn->dn_dbufs_mtx);
1397 	for (db = avl_first(&dn->dn_dbufs); db != NULL;
1398 	    db = AVL_NEXT(&dn->dn_dbufs, db)) {
1399 		if (db->db_blkid != 0 && db->db_blkid != DMU_BONUS_BLKID &&
1400 		    db->db_blkid != DMU_SPILL_BLKID) {
1401 			mutex_exit(&dn->dn_dbufs_mtx);
1402 			goto fail;
1403 		}
1404 	}
1405 	mutex_exit(&dn->dn_dbufs_mtx);
1406 
1407 	if (ibs && dn->dn_nlevels != 1)
1408 		goto fail;
1409 
1410 	/* resize the old block */
1411 	err = dbuf_hold_impl(dn, 0, 0, TRUE, FALSE, FTAG, &db);
1412 	if (err == 0)
1413 		dbuf_new_size(db, size, tx);
1414 	else if (err != ENOENT)
1415 		goto fail;
1416 
1417 	dnode_setdblksz(dn, size);
1418 	dnode_setdirty(dn, tx);
1419 	dn->dn_next_blksz[tx->tx_txg&TXG_MASK] = size;
1420 	if (ibs) {
1421 		dn->dn_indblkshift = ibs;
1422 		dn->dn_next_indblkshift[tx->tx_txg&TXG_MASK] = ibs;
1423 	}
1424 	/* rele after we have fixed the blocksize in the dnode */
1425 	if (db)
1426 		dbuf_rele(db, FTAG);
1427 
1428 	rw_exit(&dn->dn_struct_rwlock);
1429 	return (0);
1430 
1431 fail:
1432 	rw_exit(&dn->dn_struct_rwlock);
1433 	return (SET_ERROR(ENOTSUP));
1434 }
1435 
1436 /* read-holding callers must not rely on the lock being continuously held */
1437 void
dnode_new_blkid(dnode_t * dn,uint64_t blkid,dmu_tx_t * tx,boolean_t have_read)1438 dnode_new_blkid(dnode_t *dn, uint64_t blkid, dmu_tx_t *tx, boolean_t have_read)
1439 {
1440 	uint64_t txgoff = tx->tx_txg & TXG_MASK;
1441 	int epbs, new_nlevels;
1442 	uint64_t sz;
1443 
1444 	ASSERT(blkid != DMU_BONUS_BLKID);
1445 
1446 	ASSERT(have_read ?
1447 	    RW_READ_HELD(&dn->dn_struct_rwlock) :
1448 	    RW_WRITE_HELD(&dn->dn_struct_rwlock));
1449 
1450 	/*
1451 	 * if we have a read-lock, check to see if we need to do any work
1452 	 * before upgrading to a write-lock.
1453 	 */
1454 	if (have_read) {
1455 		if (blkid <= dn->dn_maxblkid)
1456 			return;
1457 
1458 		if (!rw_tryupgrade(&dn->dn_struct_rwlock)) {
1459 			rw_exit(&dn->dn_struct_rwlock);
1460 			rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1461 		}
1462 	}
1463 
1464 	if (blkid <= dn->dn_maxblkid)
1465 		goto out;
1466 
1467 	dn->dn_maxblkid = blkid;
1468 
1469 	/*
1470 	 * Compute the number of levels necessary to support the new maxblkid.
1471 	 */
1472 	new_nlevels = 1;
1473 	epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1474 	for (sz = dn->dn_nblkptr;
1475 	    sz <= blkid && sz >= dn->dn_nblkptr; sz <<= epbs)
1476 		new_nlevels++;
1477 
1478 	if (new_nlevels > dn->dn_nlevels) {
1479 		int old_nlevels = dn->dn_nlevels;
1480 		dmu_buf_impl_t *db;
1481 		list_t *list;
1482 		dbuf_dirty_record_t *new, *dr, *dr_next;
1483 
1484 		dn->dn_nlevels = new_nlevels;
1485 
1486 		ASSERT3U(new_nlevels, >, dn->dn_next_nlevels[txgoff]);
1487 		dn->dn_next_nlevels[txgoff] = new_nlevels;
1488 
1489 		/* dirty the left indirects */
1490 		db = dbuf_hold_level(dn, old_nlevels, 0, FTAG);
1491 		ASSERT(db != NULL);
1492 		new = dbuf_dirty(db, tx);
1493 		dbuf_rele(db, FTAG);
1494 
1495 		/* transfer the dirty records to the new indirect */
1496 		mutex_enter(&dn->dn_mtx);
1497 		mutex_enter(&new->dt.di.dr_mtx);
1498 		list = &dn->dn_dirty_records[txgoff];
1499 		for (dr = list_head(list); dr; dr = dr_next) {
1500 			dr_next = list_next(&dn->dn_dirty_records[txgoff], dr);
1501 			if (dr->dr_dbuf->db_level != new_nlevels-1 &&
1502 			    dr->dr_dbuf->db_blkid != DMU_BONUS_BLKID &&
1503 			    dr->dr_dbuf->db_blkid != DMU_SPILL_BLKID) {
1504 				ASSERT(dr->dr_dbuf->db_level == old_nlevels-1);
1505 				list_remove(&dn->dn_dirty_records[txgoff], dr);
1506 				list_insert_tail(&new->dt.di.dr_children, dr);
1507 				dr->dr_parent = new;
1508 			}
1509 		}
1510 		mutex_exit(&new->dt.di.dr_mtx);
1511 		mutex_exit(&dn->dn_mtx);
1512 	}
1513 
1514 out:
1515 	if (have_read)
1516 		rw_downgrade(&dn->dn_struct_rwlock);
1517 }
1518 
1519 static void
dnode_dirty_l1(dnode_t * dn,uint64_t l1blkid,dmu_tx_t * tx)1520 dnode_dirty_l1(dnode_t *dn, uint64_t l1blkid, dmu_tx_t *tx)
1521 {
1522 	dmu_buf_impl_t *db = dbuf_hold_level(dn, 1, l1blkid, FTAG);
1523 	if (db != NULL) {
1524 		dmu_buf_will_dirty(&db->db, tx);
1525 		dbuf_rele(db, FTAG);
1526 	}
1527 }
1528 
1529 void
dnode_free_range(dnode_t * dn,uint64_t off,uint64_t len,dmu_tx_t * tx)1530 dnode_free_range(dnode_t *dn, uint64_t off, uint64_t len, dmu_tx_t *tx)
1531 {
1532 	dmu_buf_impl_t *db;
1533 	uint64_t blkoff, blkid, nblks;
1534 	int blksz, blkshift, head, tail;
1535 	int trunc = FALSE;
1536 	int epbs;
1537 
1538 	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1539 	blksz = dn->dn_datablksz;
1540 	blkshift = dn->dn_datablkshift;
1541 	epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1542 
1543 	if (len == DMU_OBJECT_END) {
1544 		len = UINT64_MAX - off;
1545 		trunc = TRUE;
1546 	}
1547 
1548 	/*
1549 	 * First, block align the region to free:
1550 	 */
1551 	if (ISP2(blksz)) {
1552 		head = P2NPHASE(off, blksz);
1553 		blkoff = P2PHASE(off, blksz);
1554 		if ((off >> blkshift) > dn->dn_maxblkid)
1555 			goto out;
1556 	} else {
1557 		ASSERT(dn->dn_maxblkid == 0);
1558 		if (off == 0 && len >= blksz) {
1559 			/*
1560 			 * Freeing the whole block; fast-track this request.
1561 			 * Note that we won't dirty any indirect blocks,
1562 			 * which is fine because we will be freeing the entire
1563 			 * file and thus all indirect blocks will be freed
1564 			 * by free_children().
1565 			 */
1566 			blkid = 0;
1567 			nblks = 1;
1568 			goto done;
1569 		} else if (off >= blksz) {
1570 			/* Freeing past end-of-data */
1571 			goto out;
1572 		} else {
1573 			/* Freeing part of the block. */
1574 			head = blksz - off;
1575 			ASSERT3U(head, >, 0);
1576 		}
1577 		blkoff = off;
1578 	}
1579 	/* zero out any partial block data at the start of the range */
1580 	if (head) {
1581 		ASSERT3U(blkoff + head, ==, blksz);
1582 		if (len < head)
1583 			head = len;
1584 		if (dbuf_hold_impl(dn, 0, dbuf_whichblock(dn, 0, off),
1585 		    TRUE, FALSE, FTAG, &db) == 0) {
1586 			caddr_t data;
1587 
1588 			/* don't dirty if it isn't on disk and isn't dirty */
1589 			if (db->db_last_dirty ||
1590 			    (db->db_blkptr && !BP_IS_HOLE(db->db_blkptr))) {
1591 				rw_exit(&dn->dn_struct_rwlock);
1592 				dmu_buf_will_dirty(&db->db, tx);
1593 				rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1594 				data = db->db.db_data;
1595 				bzero(data + blkoff, head);
1596 			}
1597 			dbuf_rele(db, FTAG);
1598 		}
1599 		off += head;
1600 		len -= head;
1601 	}
1602 
1603 	/* If the range was less than one block, we're done */
1604 	if (len == 0)
1605 		goto out;
1606 
1607 	/* If the remaining range is past end of file, we're done */
1608 	if ((off >> blkshift) > dn->dn_maxblkid)
1609 		goto out;
1610 
1611 	ASSERT(ISP2(blksz));
1612 	if (trunc)
1613 		tail = 0;
1614 	else
1615 		tail = P2PHASE(len, blksz);
1616 
1617 	ASSERT0(P2PHASE(off, blksz));
1618 	/* zero out any partial block data at the end of the range */
1619 	if (tail) {
1620 		if (len < tail)
1621 			tail = len;
1622 		if (dbuf_hold_impl(dn, 0, dbuf_whichblock(dn, 0, off+len),
1623 		    TRUE, FALSE, FTAG, &db) == 0) {
1624 			/* don't dirty if not on disk and not dirty */
1625 			if (db->db_last_dirty ||
1626 			    (db->db_blkptr && !BP_IS_HOLE(db->db_blkptr))) {
1627 				rw_exit(&dn->dn_struct_rwlock);
1628 				dmu_buf_will_dirty(&db->db, tx);
1629 				rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1630 				bzero(db->db.db_data, tail);
1631 			}
1632 			dbuf_rele(db, FTAG);
1633 		}
1634 		len -= tail;
1635 	}
1636 
1637 	/* If the range did not include a full block, we are done */
1638 	if (len == 0)
1639 		goto out;
1640 
1641 	ASSERT(IS_P2ALIGNED(off, blksz));
1642 	ASSERT(trunc || IS_P2ALIGNED(len, blksz));
1643 	blkid = off >> blkshift;
1644 	nblks = len >> blkshift;
1645 	if (trunc)
1646 		nblks += 1;
1647 
1648 	/*
1649 	 * Dirty all the indirect blocks in this range.  Note that only
1650 	 * the first and last indirect blocks can actually be written
1651 	 * (if they were partially freed) -- they must be dirtied, even if
1652 	 * they do not exist on disk yet.  The interior blocks will
1653 	 * be freed by free_children(), so they will not actually be written.
1654 	 * Even though these interior blocks will not be written, we
1655 	 * dirty them for two reasons:
1656 	 *
1657 	 *  - It ensures that the indirect blocks remain in memory until
1658 	 *    syncing context.  (They have already been prefetched by
1659 	 *    dmu_tx_hold_free(), so we don't have to worry about reading
1660 	 *    them serially here.)
1661 	 *
1662 	 *  - The dirty space accounting will put pressure on the txg sync
1663 	 *    mechanism to begin syncing, and to delay transactions if there
1664 	 *    is a large amount of freeing.  Even though these indirect
1665 	 *    blocks will not be written, we could need to write the same
1666 	 *    amount of space if we copy the freed BPs into deadlists.
1667 	 */
1668 	if (dn->dn_nlevels > 1) {
1669 		uint64_t first, last;
1670 
1671 		first = blkid >> epbs;
1672 		dnode_dirty_l1(dn, first, tx);
1673 		if (trunc)
1674 			last = dn->dn_maxblkid >> epbs;
1675 		else
1676 			last = (blkid + nblks - 1) >> epbs;
1677 		if (last != first)
1678 			dnode_dirty_l1(dn, last, tx);
1679 
1680 		int shift = dn->dn_datablkshift + dn->dn_indblkshift -
1681 		    SPA_BLKPTRSHIFT;
1682 		for (uint64_t i = first + 1; i < last; i++) {
1683 			/*
1684 			 * Set i to the blockid of the next non-hole
1685 			 * level-1 indirect block at or after i.  Note
1686 			 * that dnode_next_offset() operates in terms of
1687 			 * level-0-equivalent bytes.
1688 			 */
1689 			uint64_t ibyte = i << shift;
1690 			int err = dnode_next_offset(dn, DNODE_FIND_HAVELOCK,
1691 			    &ibyte, 2, 1, 0);
1692 			i = ibyte >> shift;
1693 			if (i >= last)
1694 				break;
1695 
1696 			/*
1697 			 * Normally we should not see an error, either
1698 			 * from dnode_next_offset() or dbuf_hold_level()
1699 			 * (except for ESRCH from dnode_next_offset).
1700 			 * If there is an i/o error, then when we read
1701 			 * this block in syncing context, it will use
1702 			 * ZIO_FLAG_MUSTSUCCEED, and thus hang/panic according
1703 			 * to the "failmode" property.  dnode_next_offset()
1704 			 * doesn't have a flag to indicate MUSTSUCCEED.
1705 			 */
1706 			if (err != 0)
1707 				break;
1708 
1709 			dnode_dirty_l1(dn, i, tx);
1710 		}
1711 	}
1712 
1713 done:
1714 	/*
1715 	 * Add this range to the dnode range list.
1716 	 * We will finish up this free operation in the syncing phase.
1717 	 */
1718 	mutex_enter(&dn->dn_mtx);
1719 	int txgoff = tx->tx_txg & TXG_MASK;
1720 	if (dn->dn_free_ranges[txgoff] == NULL) {
1721 		dn->dn_free_ranges[txgoff] =
1722 		    range_tree_create(NULL, NULL, &dn->dn_mtx);
1723 	}
1724 	range_tree_clear(dn->dn_free_ranges[txgoff], blkid, nblks);
1725 	range_tree_add(dn->dn_free_ranges[txgoff], blkid, nblks);
1726 	dprintf_dnode(dn, "blkid=%llu nblks=%llu txg=%llu\n",
1727 	    blkid, nblks, tx->tx_txg);
1728 	mutex_exit(&dn->dn_mtx);
1729 
1730 	dbuf_free_range(dn, blkid, blkid + nblks - 1, tx);
1731 	dnode_setdirty(dn, tx);
1732 out:
1733 
1734 	rw_exit(&dn->dn_struct_rwlock);
1735 }
1736 
1737 static boolean_t
dnode_spill_freed(dnode_t * dn)1738 dnode_spill_freed(dnode_t *dn)
1739 {
1740 	int i;
1741 
1742 	mutex_enter(&dn->dn_mtx);
1743 	for (i = 0; i < TXG_SIZE; i++) {
1744 		if (dn->dn_rm_spillblk[i] == DN_KILL_SPILLBLK)
1745 			break;
1746 	}
1747 	mutex_exit(&dn->dn_mtx);
1748 	return (i < TXG_SIZE);
1749 }
1750 
1751 /* return TRUE if this blkid was freed in a recent txg, or FALSE if it wasn't */
1752 uint64_t
dnode_block_freed(dnode_t * dn,uint64_t blkid)1753 dnode_block_freed(dnode_t *dn, uint64_t blkid)
1754 {
1755 	void *dp = spa_get_dsl(dn->dn_objset->os_spa);
1756 	int i;
1757 
1758 	if (blkid == DMU_BONUS_BLKID)
1759 		return (FALSE);
1760 
1761 	/*
1762 	 * If we're in the process of opening the pool, dp will not be
1763 	 * set yet, but there shouldn't be anything dirty.
1764 	 */
1765 	if (dp == NULL)
1766 		return (FALSE);
1767 
1768 	if (dn->dn_free_txg)
1769 		return (TRUE);
1770 
1771 	if (blkid == DMU_SPILL_BLKID)
1772 		return (dnode_spill_freed(dn));
1773 
1774 	mutex_enter(&dn->dn_mtx);
1775 	for (i = 0; i < TXG_SIZE; i++) {
1776 		if (dn->dn_free_ranges[i] != NULL &&
1777 		    range_tree_contains(dn->dn_free_ranges[i], blkid, 1))
1778 			break;
1779 	}
1780 	mutex_exit(&dn->dn_mtx);
1781 	return (i < TXG_SIZE);
1782 }
1783 
1784 /* call from syncing context when we actually write/free space for this dnode */
1785 void
dnode_diduse_space(dnode_t * dn,int64_t delta)1786 dnode_diduse_space(dnode_t *dn, int64_t delta)
1787 {
1788 	uint64_t space;
1789 	dprintf_dnode(dn, "dn=%p dnp=%p used=%llu delta=%lld\n",
1790 	    dn, dn->dn_phys,
1791 	    (u_longlong_t)dn->dn_phys->dn_used,
1792 	    (longlong_t)delta);
1793 
1794 	mutex_enter(&dn->dn_mtx);
1795 	space = DN_USED_BYTES(dn->dn_phys);
1796 	if (delta > 0) {
1797 		ASSERT3U(space + delta, >=, space); /* no overflow */
1798 	} else {
1799 		ASSERT3U(space, >=, -delta); /* no underflow */
1800 	}
1801 	space += delta;
1802 	if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_DNODE_BYTES) {
1803 		ASSERT((dn->dn_phys->dn_flags & DNODE_FLAG_USED_BYTES) == 0);
1804 		ASSERT0(P2PHASE(space, 1<<DEV_BSHIFT));
1805 		dn->dn_phys->dn_used = space >> DEV_BSHIFT;
1806 	} else {
1807 		dn->dn_phys->dn_used = space;
1808 		dn->dn_phys->dn_flags |= DNODE_FLAG_USED_BYTES;
1809 	}
1810 	mutex_exit(&dn->dn_mtx);
1811 }
1812 
1813 /*
1814  * Call when we think we're going to write/free space in open context to track
1815  * the amount of memory in use by the currently open txg.
1816  */
1817 void
dnode_willuse_space(dnode_t * dn,int64_t space,dmu_tx_t * tx)1818 dnode_willuse_space(dnode_t *dn, int64_t space, dmu_tx_t *tx)
1819 {
1820 	objset_t *os = dn->dn_objset;
1821 	dsl_dataset_t *ds = os->os_dsl_dataset;
1822 	int64_t aspace = spa_get_asize(os->os_spa, space);
1823 
1824 	if (ds != NULL) {
1825 		dsl_dir_willuse_space(ds->ds_dir, aspace, tx);
1826 		dsl_pool_dirty_space(dmu_tx_pool(tx), space, tx);
1827 	}
1828 
1829 	dmu_tx_willuse_space(tx, aspace);
1830 }
1831 
1832 /*
1833  * Scans a block at the indicated "level" looking for a hole or data,
1834  * depending on 'flags'.
1835  *
1836  * If level > 0, then we are scanning an indirect block looking at its
1837  * pointers.  If level == 0, then we are looking at a block of dnodes.
1838  *
1839  * If we don't find what we are looking for in the block, we return ESRCH.
1840  * Otherwise, return with *offset pointing to the beginning (if searching
1841  * forwards) or end (if searching backwards) of the range covered by the
1842  * block pointer we matched on (or dnode).
1843  *
1844  * The basic search algorithm used below by dnode_next_offset() is to
1845  * use this function to search up the block tree (widen the search) until
1846  * we find something (i.e., we don't return ESRCH) and then search back
1847  * down the tree (narrow the search) until we reach our original search
1848  * level.
1849  */
1850 static int
dnode_next_offset_level(dnode_t * dn,int flags,uint64_t * offset,int lvl,uint64_t blkfill,uint64_t txg)1851 dnode_next_offset_level(dnode_t *dn, int flags, uint64_t *offset,
1852     int lvl, uint64_t blkfill, uint64_t txg)
1853 {
1854 	dmu_buf_impl_t *db = NULL;
1855 	void *data = NULL;
1856 	uint64_t epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
1857 	uint64_t epb = 1ULL << epbs;
1858 	uint64_t minfill, maxfill;
1859 	boolean_t hole;
1860 	int i, inc, error, span;
1861 
1862 	dprintf("probing object %llu offset %llx level %d of %u\n",
1863 	    dn->dn_object, *offset, lvl, dn->dn_phys->dn_nlevels);
1864 
1865 	hole = ((flags & DNODE_FIND_HOLE) != 0);
1866 	inc = (flags & DNODE_FIND_BACKWARDS) ? -1 : 1;
1867 	ASSERT(txg == 0 || !hole);
1868 
1869 	if (lvl == dn->dn_phys->dn_nlevels) {
1870 		error = 0;
1871 		epb = dn->dn_phys->dn_nblkptr;
1872 		data = dn->dn_phys->dn_blkptr;
1873 	} else {
1874 		uint64_t blkid = dbuf_whichblock(dn, lvl, *offset);
1875 		error = dbuf_hold_impl(dn, lvl, blkid, TRUE, FALSE, FTAG, &db);
1876 		if (error) {
1877 			if (error != ENOENT)
1878 				return (error);
1879 			if (hole)
1880 				return (0);
1881 			/*
1882 			 * This can only happen when we are searching up
1883 			 * the block tree for data.  We don't really need to
1884 			 * adjust the offset, as we will just end up looking
1885 			 * at the pointer to this block in its parent, and its
1886 			 * going to be unallocated, so we will skip over it.
1887 			 */
1888 			return (SET_ERROR(ESRCH));
1889 		}
1890 		error = dbuf_read(db, NULL, DB_RF_CANFAIL | DB_RF_HAVESTRUCT);
1891 		if (error) {
1892 			dbuf_rele(db, FTAG);
1893 			return (error);
1894 		}
1895 		data = db->db.db_data;
1896 	}
1897 
1898 
1899 	if (db != NULL && txg != 0 && (db->db_blkptr == NULL ||
1900 	    db->db_blkptr->blk_birth <= txg ||
1901 	    BP_IS_HOLE(db->db_blkptr))) {
1902 		/*
1903 		 * This can only happen when we are searching up the tree
1904 		 * and these conditions mean that we need to keep climbing.
1905 		 */
1906 		error = SET_ERROR(ESRCH);
1907 	} else if (lvl == 0) {
1908 		dnode_phys_t *dnp = data;
1909 		span = DNODE_SHIFT;
1910 		ASSERT(dn->dn_type == DMU_OT_DNODE);
1911 
1912 		for (i = (*offset >> span) & (blkfill - 1);
1913 		    i >= 0 && i < blkfill; i += inc) {
1914 			if ((dnp[i].dn_type == DMU_OT_NONE) == hole)
1915 				break;
1916 			*offset += (1ULL << span) * inc;
1917 		}
1918 		if (i < 0 || i == blkfill)
1919 			error = SET_ERROR(ESRCH);
1920 	} else {
1921 		blkptr_t *bp = data;
1922 		uint64_t start = *offset;
1923 		span = (lvl - 1) * epbs + dn->dn_datablkshift;
1924 		minfill = 0;
1925 		maxfill = blkfill << ((lvl - 1) * epbs);
1926 
1927 		if (hole)
1928 			maxfill--;
1929 		else
1930 			minfill++;
1931 
1932 		*offset = *offset >> span;
1933 		for (i = BF64_GET(*offset, 0, epbs);
1934 		    i >= 0 && i < epb; i += inc) {
1935 			if (BP_GET_FILL(&bp[i]) >= minfill &&
1936 			    BP_GET_FILL(&bp[i]) <= maxfill &&
1937 			    (hole || bp[i].blk_birth > txg))
1938 				break;
1939 			if (inc > 0 || *offset > 0)
1940 				*offset += inc;
1941 		}
1942 		*offset = *offset << span;
1943 		if (inc < 0) {
1944 			/* traversing backwards; position offset at the end */
1945 			ASSERT3U(*offset, <=, start);
1946 			*offset = MIN(*offset + (1ULL << span) - 1, start);
1947 		} else if (*offset < start) {
1948 			*offset = start;
1949 		}
1950 		if (i < 0 || i >= epb)
1951 			error = SET_ERROR(ESRCH);
1952 	}
1953 
1954 	if (db)
1955 		dbuf_rele(db, FTAG);
1956 
1957 	return (error);
1958 }
1959 
1960 /*
1961  * Find the next hole, data, or sparse region at or after *offset.
1962  * The value 'blkfill' tells us how many items we expect to find
1963  * in an L0 data block; this value is 1 for normal objects,
1964  * DNODES_PER_BLOCK for the meta dnode, and some fraction of
1965  * DNODES_PER_BLOCK when searching for sparse regions thereof.
1966  *
1967  * Examples:
1968  *
1969  * dnode_next_offset(dn, flags, offset, 1, 1, 0);
1970  *	Finds the next/previous hole/data in a file.
1971  *	Used in dmu_offset_next().
1972  *
1973  * dnode_next_offset(mdn, flags, offset, 0, DNODES_PER_BLOCK, txg);
1974  *	Finds the next free/allocated dnode an objset's meta-dnode.
1975  *	Only finds objects that have new contents since txg (ie.
1976  *	bonus buffer changes and content removal are ignored).
1977  *	Used in dmu_object_next().
1978  *
1979  * dnode_next_offset(mdn, DNODE_FIND_HOLE, offset, 2, DNODES_PER_BLOCK >> 2, 0);
1980  *	Finds the next L2 meta-dnode bp that's at most 1/4 full.
1981  *	Used in dmu_object_alloc().
1982  */
1983 int
dnode_next_offset(dnode_t * dn,int flags,uint64_t * offset,int minlvl,uint64_t blkfill,uint64_t txg)1984 dnode_next_offset(dnode_t *dn, int flags, uint64_t *offset,
1985     int minlvl, uint64_t blkfill, uint64_t txg)
1986 {
1987 	uint64_t initial_offset = *offset;
1988 	int lvl, maxlvl;
1989 	int error = 0;
1990 
1991 	if (!(flags & DNODE_FIND_HAVELOCK))
1992 		rw_enter(&dn->dn_struct_rwlock, RW_READER);
1993 
1994 	if (dn->dn_phys->dn_nlevels == 0) {
1995 		error = SET_ERROR(ESRCH);
1996 		goto out;
1997 	}
1998 
1999 	if (dn->dn_datablkshift == 0) {
2000 		if (*offset < dn->dn_datablksz) {
2001 			if (flags & DNODE_FIND_HOLE)
2002 				*offset = dn->dn_datablksz;
2003 		} else {
2004 			error = SET_ERROR(ESRCH);
2005 		}
2006 		goto out;
2007 	}
2008 
2009 	maxlvl = dn->dn_phys->dn_nlevels;
2010 
2011 	for (lvl = minlvl; lvl <= maxlvl; lvl++) {
2012 		error = dnode_next_offset_level(dn,
2013 		    flags, offset, lvl, blkfill, txg);
2014 		if (error != ESRCH)
2015 			break;
2016 	}
2017 
2018 	while (error == 0 && --lvl >= minlvl) {
2019 		error = dnode_next_offset_level(dn,
2020 		    flags, offset, lvl, blkfill, txg);
2021 	}
2022 
2023 	/*
2024 	 * There's always a "virtual hole" at the end of the object, even
2025 	 * if all BP's which physically exist are non-holes.
2026 	 */
2027 	if ((flags & DNODE_FIND_HOLE) && error == ESRCH && txg == 0 &&
2028 	    minlvl == 1 && blkfill == 1 && !(flags & DNODE_FIND_BACKWARDS)) {
2029 		error = 0;
2030 	}
2031 
2032 	if (error == 0 && (flags & DNODE_FIND_BACKWARDS ?
2033 	    initial_offset < *offset : initial_offset > *offset))
2034 		error = SET_ERROR(ESRCH);
2035 out:
2036 	if (!(flags & DNODE_FIND_HAVELOCK))
2037 		rw_exit(&dn->dn_struct_rwlock);
2038 
2039 	return (error);
2040 }
2041