xref: /titanic_41/usr/src/uts/common/fs/zfs/dmu_traverse.c (revision fb9f9b975cb9214fec5dab37d461199adab9b964)
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, Version 1.0 only
6  * (the "License").  You may not use this file except in compliance
7  * with the License.
8  *
9  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
10  * or http://www.opensolaris.org/os/licensing.
11  * See the License for the specific language governing permissions
12  * and limitations under the License.
13  *
14  * When distributing Covered Code, include this CDDL HEADER in each
15  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
16  * If applicable, add the following below this CDDL HEADER, with the
17  * fields enclosed by brackets "[]" replaced with your own identifying
18  * information: Portions Copyright [yyyy] [name of copyright owner]
19  *
20  * CDDL HEADER END
21  */
22 /*
23  * Copyright 2005 Sun Microsystems, Inc.  All rights reserved.
24  * Use is subject to license terms.
25  */
26 
27 #pragma ident	"%Z%%M%	%I%	%E% SMI"
28 
29 #include <sys/zfs_context.h>
30 #include <sys/dmu_objset.h>
31 #include <sys/dmu_traverse.h>
32 #include <sys/dsl_dataset.h>
33 #include <sys/dsl_dir.h>
34 #include <sys/dsl_pool.h>
35 #include <sys/dnode.h>
36 #include <sys/spa.h>
37 #include <sys/zio.h>
38 #include <sys/dmu_impl.h>
39 
40 #define	BP_SPAN_SHIFT(level, width)	((level) * (width))
41 
42 #define	BP_EQUAL(b1, b2)				\
43 	(DVA_EQUAL(BP_IDENTITY(b1), BP_IDENTITY(b2)) &&	\
44 	(b1)->blk_birth == (b2)->blk_birth)
45 
46 /*
47  * Compare two bookmarks.
48  *
49  * For ADVANCE_PRE, the visitation order is:
50  *
51  *	objset 0, 1, 2, ..., ZB_MAXOBJSET.
52  *	object 0, 1, 2, ..., ZB_MAXOBJECT.
53  *	blkoff 0, 1, 2, ...
54  *	level ZB_MAXLEVEL, ..., 2, 1, 0.
55  *
56  * where blkoff = blkid << BP_SPAN_SHIFT(level, width), and thus a valid
57  * ordering vector is:
58  *
59  *	< objset, object, blkoff, -level >
60  *
61  * For ADVANCE_POST, the starting offsets aren't sequential but ending
62  * offsets [blkoff = (blkid + 1) << BP_SPAN_SHIFT(level, width)] are.
63  * The visitation order is:
64  *
65  *	objset 1, 2, ..., ZB_MAXOBJSET, 0.
66  *	object 1, 2, ..., ZB_MAXOBJECT, 0.
67  *	blkoff 1, 2, ...
68  *	level 0, 1, 2, ..., ZB_MAXLEVEL.
69  *
70  * and thus a valid ordering vector is:
71  *
72  *	< objset - 1, object - 1, blkoff, level >
73  *
74  * Both orderings can be expressed as:
75  *
76  *	< objset + bias, object + bias, blkoff, level ^ bias >
77  *
78  * where 'bias' is either 0 or -1 (for ADVANCE_PRE or ADVANCE_POST)
79  * and 'blkoff' is (blkid - bias) << BP_SPAN_SHIFT(level, wshift).
80  *
81  * Special case: an objset's osphys is represented as level -1 of object 0.
82  * It is always either the very first or very last block we visit in an objset.
83  * Therefore, if either bookmark's level is -1, level alone determines order.
84  */
85 static int
86 compare_bookmark(zbookmark_t *szb, zbookmark_t *ezb, dnode_phys_t *dnp,
87     int advance)
88 {
89 	int bias = (advance & ADVANCE_PRE) ? 0 : -1;
90 	uint64_t sblkoff, eblkoff;
91 	int slevel, elevel, wshift;
92 
93 	if (szb->zb_objset + bias < ezb->zb_objset + bias)
94 		return (-1);
95 
96 	if (szb->zb_objset + bias > ezb->zb_objset + bias)
97 		return (1);
98 
99 	slevel = szb->zb_level;
100 	elevel = ezb->zb_level;
101 
102 	if ((slevel | elevel) < 0)
103 		return ((slevel ^ bias) - (elevel ^ bias));
104 
105 	if (szb->zb_object + bias < ezb->zb_object + bias)
106 		return (-1);
107 
108 	if (szb->zb_object + bias > ezb->zb_object + bias)
109 		return (1);
110 
111 	if (dnp == NULL)
112 		return (0);
113 
114 	wshift = dnp->dn_indblkshift - SPA_BLKPTRSHIFT;
115 
116 	sblkoff = (szb->zb_blkid - bias) << BP_SPAN_SHIFT(slevel, wshift);
117 	eblkoff = (ezb->zb_blkid - bias) << BP_SPAN_SHIFT(elevel, wshift);
118 
119 	if (sblkoff < eblkoff)
120 		return (-1);
121 
122 	if (sblkoff > eblkoff)
123 		return (1);
124 
125 	return ((elevel ^ bias) - (slevel ^ bias));
126 }
127 
128 #define	SET_BOOKMARK(zb, objset, object, level, blkid)	\
129 {							\
130 	(zb)->zb_objset = objset;			\
131 	(zb)->zb_object = object;			\
132 	(zb)->zb_level = level;				\
133 	(zb)->zb_blkid = blkid;				\
134 }
135 
136 #define	SET_BOOKMARK_LB(zb, level, blkid)		\
137 {							\
138 	(zb)->zb_level = level;				\
139 	(zb)->zb_blkid = blkid;				\
140 }
141 
142 static int
143 advance_objset(zseg_t *zseg, uint64_t objset, int advance)
144 {
145 	zbookmark_t *zb = &zseg->seg_start;
146 
147 	if (advance & ADVANCE_PRE) {
148 		if (objset >= ZB_MAXOBJSET)
149 			return (ERANGE);
150 		SET_BOOKMARK(zb, objset, 0, -1, 0);
151 	} else {
152 		if (objset >= ZB_MAXOBJSET)
153 			objset = 0;
154 		SET_BOOKMARK(zb, objset, 1, 0, 0);
155 	}
156 
157 	if (compare_bookmark(zb, &zseg->seg_end, NULL, advance) > 0)
158 		return (ERANGE);
159 
160 	return (EAGAIN);
161 }
162 
163 static int
164 advance_object(zseg_t *zseg, uint64_t object, int advance)
165 {
166 	zbookmark_t *zb = &zseg->seg_start;
167 
168 	if (advance & ADVANCE_PRE) {
169 		if (object >= ZB_MAXOBJECT) {
170 			SET_BOOKMARK(zb, zb->zb_objset + 1, 0, -1, 0);
171 		} else {
172 			SET_BOOKMARK(zb, zb->zb_objset, object, ZB_MAXLEVEL, 0);
173 		}
174 	} else {
175 		if (zb->zb_object == 0) {
176 			SET_BOOKMARK(zb, zb->zb_objset, 0, -1, 0);
177 		} else {
178 			if (object >= ZB_MAXOBJECT)
179 				object = 0;
180 			SET_BOOKMARK(zb, zb->zb_objset, object, 0, 0);
181 		}
182 	}
183 
184 	if (compare_bookmark(zb, &zseg->seg_end, NULL, advance) > 0)
185 		return (ERANGE);
186 
187 	return (EAGAIN);
188 }
189 
190 static int
191 advance_from_osphys(zseg_t *zseg, int advance)
192 {
193 	zbookmark_t *zb = &zseg->seg_start;
194 
195 	ASSERT(zb->zb_object == 0);
196 	ASSERT(zb->zb_level == -1);
197 	ASSERT(zb->zb_blkid == 0);
198 
199 	if (advance & ADVANCE_PRE) {
200 		SET_BOOKMARK_LB(zb, ZB_MAXLEVEL, 0);
201 	} else {
202 		if (zb->zb_objset == 0)
203 			return (ERANGE);
204 		SET_BOOKMARK(zb, zb->zb_objset + 1, 1, 0, 0);
205 	}
206 
207 	if (compare_bookmark(zb, &zseg->seg_end, NULL, advance) > 0)
208 		return (ERANGE);
209 
210 	return (EAGAIN);
211 }
212 
213 static int
214 advance_block(zseg_t *zseg, dnode_phys_t *dnp, int rc, int advance)
215 {
216 	zbookmark_t *zb = &zseg->seg_start;
217 	int wshift = dnp->dn_indblkshift - SPA_BLKPTRSHIFT;
218 	int maxlevel = dnp->dn_nlevels - 1;
219 	int level = zb->zb_level;
220 	uint64_t blkid = zb->zb_blkid;
221 
222 	if (advance & ADVANCE_PRE) {
223 		if (level > 0 && rc == 0) {
224 			level--;
225 			blkid <<= wshift;
226 		} else {
227 			blkid++;
228 
229 			if ((blkid << BP_SPAN_SHIFT(level, wshift)) >
230 			    dnp->dn_maxblkid)
231 				return (ERANGE);
232 
233 			while (level < maxlevel) {
234 				if (P2PHASE(blkid, 1ULL << wshift))
235 					break;
236 				blkid >>= wshift;
237 				level++;
238 			}
239 		}
240 	} else {
241 		if (level >= maxlevel || P2PHASE(blkid + 1, 1ULL << wshift)) {
242 			blkid = (blkid + 1) << BP_SPAN_SHIFT(level, wshift);
243 			level = 0;
244 		} else {
245 			blkid >>= wshift;
246 			level++;
247 		}
248 
249 		while ((blkid << BP_SPAN_SHIFT(level, wshift)) >
250 		    dnp->dn_maxblkid) {
251 			if (level == maxlevel)
252 				return (ERANGE);
253 			blkid >>= wshift;
254 			level++;
255 		}
256 	}
257 	SET_BOOKMARK_LB(zb, level, blkid);
258 
259 	if (compare_bookmark(zb, &zseg->seg_end, dnp, advance) > 0)
260 		return (ERANGE);
261 
262 	return (EAGAIN);
263 }
264 
265 static int
266 traverse_callback(traverse_handle_t *th, zseg_t *zseg, traverse_blk_cache_t *bc)
267 {
268 	/*
269 	 * Before we issue the callback, prune against maxtxg.
270 	 *
271 	 * We prune against mintxg before we get here because it's a big win.
272 	 * If a given block was born in txg 37, then we know that the entire
273 	 * subtree below that block must have been born in txg 37 or earlier.
274 	 * We can therefore lop off huge branches of the tree as we go.
275 	 *
276 	 * There's no corresponding optimization for maxtxg because knowing
277 	 * that bp->blk_birth >= maxtxg doesn't imply anything about the bp's
278 	 * children.  In fact, the copy-on-write design of ZFS ensures that
279 	 * top-level blocks will pretty much always be new.
280 	 *
281 	 * Therefore, in the name of simplicity we don't prune against
282 	 * maxtxg until the last possible moment -- that being right now.
283 	 */
284 	if (bc->bc_errno == 0 && bc->bc_blkptr.blk_birth >= zseg->seg_maxtxg)
285 		return (0);
286 
287 	if (bc->bc_errno == 0) {
288 		zbookmark_t *zb = &bc->bc_bookmark;
289 		zbookmark_t *szb = &zseg->seg_start;
290 		zbookmark_t *ezb = &zseg->seg_end;
291 		zbookmark_t *lzb = &th->th_lastcb;
292 		dnode_phys_t *dnp = bc->bc_dnode;
293 
294 		/*
295 		 * Debugging: verify that the order we visit things
296 		 * agrees with the order defined by compare_bookmark().
297 		 */
298 		ASSERT(compare_bookmark(zb, ezb, dnp, th->th_advance) <= 0);
299 		ASSERT(compare_bookmark(zb, szb, dnp, th->th_advance) == 0);
300 		ASSERT(compare_bookmark(lzb, zb, dnp, th->th_advance) < 0 ||
301 		    lzb->zb_level == ZB_NO_LEVEL);
302 		*lzb = *zb;
303 	}
304 
305 	th->th_callbacks++;
306 	return (th->th_func(bc, th->th_spa, th->th_arg));
307 }
308 
309 static int
310 traverse_read(traverse_handle_t *th, traverse_blk_cache_t *bc, blkptr_t *bp,
311 	dnode_phys_t *dnp)
312 {
313 	zbookmark_t *zb = &bc->bc_bookmark;
314 	int error;
315 
316 	th->th_hits++;
317 
318 	bc->bc_dnode = dnp;
319 	bc->bc_errno = 0;
320 
321 	if (BP_EQUAL(&bc->bc_blkptr, bp))
322 		return (0);
323 
324 	bc->bc_blkptr = *bp;
325 
326 	if (bc->bc_data == NULL)
327 		return (0);
328 
329 	if (BP_IS_HOLE(bp)) {
330 		ASSERT(th->th_advance & ADVANCE_HOLES);
331 		return (0);
332 	}
333 
334 	if (compare_bookmark(zb, &th->th_noread, dnp, 0) == 0) {
335 		error = EIO;
336 	} else if (arc_tryread(th->th_spa, bp, bc->bc_data) == 0) {
337 		error = 0;
338 		th->th_arc_hits++;
339 	} else {
340 		error = zio_wait(zio_read(NULL, th->th_spa, bp, bc->bc_data,
341 		    BP_GET_LSIZE(bp), NULL, NULL, ZIO_PRIORITY_SYNC_READ,
342 		    th->th_zio_flags | ZIO_FLAG_DONT_CACHE));
343 
344 		if (BP_SHOULD_BYTESWAP(bp) && error == 0)
345 			(zb->zb_level > 0 ? byteswap_uint64_array :
346 			    dmu_ot[BP_GET_TYPE(bp)].ot_byteswap)(bc->bc_data,
347 			    BP_GET_LSIZE(bp));
348 		th->th_reads++;
349 	}
350 
351 	if (error) {
352 		bc->bc_errno = error;
353 		error = traverse_callback(th, NULL, bc);
354 		ASSERT(error == EAGAIN || error == EINTR || error == ERESTART);
355 		bc->bc_blkptr.blk_birth = -1ULL;
356 	}
357 
358 	dprintf("cache %02x error %d <%llu, %llu, %d, %llx>\n",
359 	    bc - &th->th_cache[0][0], error,
360 	    zb->zb_objset, zb->zb_object, zb->zb_level, zb->zb_blkid);
361 
362 	return (error);
363 }
364 
365 static int
366 find_block(traverse_handle_t *th, zseg_t *zseg, dnode_phys_t *dnp, int depth)
367 {
368 	zbookmark_t *zb = &zseg->seg_start;
369 	traverse_blk_cache_t *bc;
370 	blkptr_t *bp = dnp->dn_blkptr;
371 	int i, first, level;
372 	int nbp = dnp->dn_nblkptr;
373 	int minlevel = zb->zb_level;
374 	int maxlevel = dnp->dn_nlevels - 1;
375 	int wshift = dnp->dn_indblkshift - SPA_BLKPTRSHIFT;
376 	int bp_shift = BP_SPAN_SHIFT(maxlevel - minlevel, wshift);
377 	uint64_t blkid = zb->zb_blkid >> bp_shift;
378 	int do_holes = (th->th_advance & ADVANCE_HOLES) && depth == ZB_DN_CACHE;
379 	int rc;
380 
381 	if (minlevel > maxlevel || blkid >= nbp)
382 		return (ERANGE);
383 
384 	for (level = maxlevel; level >= minlevel; level--) {
385 		first = P2PHASE(blkid, 1ULL << wshift);
386 
387 		for (i = first; i < nbp; i++)
388 			if (bp[i].blk_birth > zseg->seg_mintxg ||
389 			    BP_IS_HOLE(&bp[i]) && do_holes)
390 				break;
391 
392 		if (i != first) {
393 			i--;
394 			SET_BOOKMARK_LB(zb, level, blkid + (i - first));
395 			return (ENOTBLK);
396 		}
397 
398 		bc = &th->th_cache[depth][level];
399 
400 		SET_BOOKMARK(&bc->bc_bookmark, zb->zb_objset, zb->zb_object,
401 		    level, blkid);
402 
403 		if (rc = traverse_read(th, bc, bp + i, dnp)) {
404 			if (rc != EAGAIN) {
405 				SET_BOOKMARK_LB(zb, level, blkid);
406 			}
407 			return (rc);
408 		}
409 
410 		if (BP_IS_HOLE(&bp[i])) {
411 			SET_BOOKMARK_LB(zb, level, blkid);
412 			th->th_lastcb.zb_level = ZB_NO_LEVEL;
413 			return (0);
414 		}
415 
416 		nbp = 1 << wshift;
417 		bp = bc->bc_data;
418 		bp_shift -= wshift;
419 		blkid = zb->zb_blkid >> bp_shift;
420 	}
421 
422 	return (0);
423 }
424 
425 static int
426 get_dnode(traverse_handle_t *th, uint64_t objset, dnode_phys_t *mdn,
427     uint64_t *objectp, dnode_phys_t **dnpp, uint64_t txg, int type, int depth)
428 {
429 	zseg_t zseg;
430 	zbookmark_t *zb = &zseg.seg_start;
431 	uint64_t object = *objectp;
432 	int i, rc;
433 
434 	SET_BOOKMARK(zb, objset, 0, 0, object / DNODES_PER_BLOCK);
435 	SET_BOOKMARK(&zseg.seg_end, objset, 0, 0, ZB_MAXBLKID);
436 
437 	zseg.seg_mintxg = txg;
438 	zseg.seg_maxtxg = -1ULL;
439 
440 	for (;;) {
441 		rc = find_block(th, &zseg, mdn, depth);
442 
443 		if (rc == EAGAIN || rc == EINTR || rc == ERANGE)
444 			break;
445 
446 		if (rc == 0 && zb->zb_level == 0) {
447 			dnode_phys_t *dnp = th->th_cache[depth][0].bc_data;
448 			for (i = 0; i < DNODES_PER_BLOCK; i++) {
449 				object = (zb->zb_blkid * DNODES_PER_BLOCK) + i;
450 				if (object >= *objectp &&
451 				    dnp[i].dn_type != DMU_OT_NONE &&
452 				    (type == -1 || dnp[i].dn_type == type)) {
453 					*objectp = object;
454 					*dnpp = &dnp[i];
455 					return (0);
456 				}
457 			}
458 		}
459 
460 		rc = advance_block(&zseg, mdn, rc, ADVANCE_PRE);
461 
462 		if (rc == ERANGE)
463 			break;
464 	}
465 
466 	if (rc == ERANGE)
467 		*objectp = ZB_MAXOBJECT;
468 
469 	return (rc);
470 }
471 
472 static int
473 traverse_segment(traverse_handle_t *th, zseg_t *zseg, blkptr_t *mosbp)
474 {
475 	zbookmark_t *zb = &zseg->seg_start;
476 	traverse_blk_cache_t *bc;
477 	dnode_phys_t *dn, *dn_tmp;
478 	int worklimit = 1000;
479 	int rc;
480 
481 	dprintf("<%llu, %llu, %d, %llx>\n",
482 	    zb->zb_objset, zb->zb_object, zb->zb_level, zb->zb_blkid);
483 
484 	bc = &th->th_cache[ZB_MOS_CACHE][ZB_MAXLEVEL - 1];
485 	dn = &((objset_phys_t *)bc->bc_data)->os_meta_dnode;
486 
487 	SET_BOOKMARK(&bc->bc_bookmark, 0, 0, -1, 0);
488 
489 	rc = traverse_read(th, bc, mosbp, dn);
490 
491 	if (rc)		/* If we get ERESTART, we've got nowhere left to go */
492 		return (rc == ERESTART ? EINTR : rc);
493 
494 	ASSERT(dn->dn_nlevels < ZB_MAXLEVEL);
495 
496 	if (zb->zb_objset != 0) {
497 		uint64_t objset = zb->zb_objset;
498 		dsl_dataset_phys_t *dsp;
499 
500 		rc = get_dnode(th, 0, dn, &objset, &dn_tmp, 0,
501 		    DMU_OT_DSL_DATASET, ZB_MOS_CACHE);
502 
503 		if (objset != zb->zb_objset)
504 			rc = advance_objset(zseg, objset, th->th_advance);
505 
506 		if (rc != 0)
507 			return (rc);
508 
509 		dsp = DN_BONUS(dn_tmp);
510 
511 		bc = &th->th_cache[ZB_MDN_CACHE][ZB_MAXLEVEL - 1];
512 		dn = &((objset_phys_t *)bc->bc_data)->os_meta_dnode;
513 
514 		SET_BOOKMARK(&bc->bc_bookmark, objset, 0, -1, 0);
515 
516 		rc = traverse_read(th, bc, &dsp->ds_bp, dn);
517 
518 		if (rc != 0) {
519 			if (rc == ERESTART)
520 				rc = advance_objset(zseg, zb->zb_objset + 1,
521 				    th->th_advance);
522 			return (rc);
523 		}
524 
525 		if (th->th_advance & ADVANCE_PRUNE)
526 			zseg->seg_mintxg =
527 			    MAX(zseg->seg_mintxg, dsp->ds_prev_snap_txg);
528 	}
529 
530 	if (zb->zb_level == -1) {
531 		ASSERT(zb->zb_object == 0);
532 
533 		if (bc->bc_blkptr.blk_birth > zseg->seg_mintxg) {
534 			rc = traverse_callback(th, zseg, bc);
535 			if (rc) {
536 				ASSERT(rc == EINTR);
537 				return (rc);
538 			}
539 		}
540 
541 		return (advance_from_osphys(zseg, th->th_advance));
542 	}
543 
544 	if (zb->zb_object != 0) {
545 		uint64_t object = zb->zb_object;
546 
547 		rc = get_dnode(th, zb->zb_objset, dn, &object, &dn_tmp,
548 		    zseg->seg_mintxg, -1, ZB_MDN_CACHE);
549 
550 		if (object != zb->zb_object)
551 			rc = advance_object(zseg, object, th->th_advance);
552 
553 		if (rc != 0)
554 			return (rc);
555 
556 		dn = dn_tmp;
557 	}
558 
559 	if (zb->zb_level == ZB_MAXLEVEL)
560 		zb->zb_level = dn->dn_nlevels - 1;
561 
562 	for (;;) {
563 		rc = find_block(th, zseg, dn, ZB_DN_CACHE);
564 
565 		if (rc == EAGAIN || rc == EINTR || rc == ERANGE)
566 			break;
567 
568 		if (rc == 0) {
569 			bc = &th->th_cache[ZB_DN_CACHE][zb->zb_level];
570 			ASSERT(bc->bc_dnode == dn);
571 			ASSERT(bc->bc_blkptr.blk_birth <= mosbp->blk_birth);
572 			rc = traverse_callback(th, zseg, bc);
573 			if (rc) {
574 				ASSERT(rc == EINTR);
575 				return (rc);
576 			}
577 			if (BP_IS_HOLE(&bc->bc_blkptr)) {
578 				ASSERT(th->th_advance & ADVANCE_HOLES);
579 				rc = ENOTBLK;
580 			}
581 		}
582 
583 		rc = advance_block(zseg, dn, rc, th->th_advance);
584 
585 		if (rc == ERANGE)
586 			break;
587 
588 		/*
589 		 * Give spa_sync() a chance to run.
590 		 */
591 		if (spa_traverse_wanted(th->th_spa)) {
592 			th->th_syncs++;
593 			return (EAGAIN);
594 		}
595 
596 		if (--worklimit == 0)
597 			return (EAGAIN);
598 	}
599 
600 	if (rc == ERANGE)
601 		rc = advance_object(zseg, zb->zb_object + 1, th->th_advance);
602 
603 	return (rc);
604 }
605 
606 /*
607  * It is the caller's responsibility to ensure that the dsl_dataset_t
608  * doesn't go away during traversal.
609  */
610 int
611 traverse_dsl_dataset(dsl_dataset_t *ds, uint64_t txg_start, int advance,
612     blkptr_cb_t func, void *arg)
613 {
614 	spa_t *spa = ds->ds_dir->dd_pool->dp_spa;
615 	traverse_handle_t *th;
616 	int err;
617 
618 	th = traverse_init(spa, func, arg, advance, ZIO_FLAG_MUSTSUCCEED);
619 
620 	traverse_add_objset(th, txg_start, -1ULL, ds->ds_object);
621 
622 	while ((err = traverse_more(th)) == EAGAIN)
623 		continue;
624 
625 	traverse_fini(th);
626 	return (err);
627 }
628 
629 int
630 traverse_more(traverse_handle_t *th)
631 {
632 	zseg_t *zseg = list_head(&th->th_seglist);
633 	uint64_t save_txg;	/* XXX won't be necessary with real itinerary */
634 	krwlock_t *rw = spa_traverse_rwlock(th->th_spa);
635 	blkptr_t *mosbp = spa_get_rootblkptr(th->th_spa);
636 	int rc;
637 
638 	if (zseg == NULL)
639 		return (0);
640 
641 	th->th_restarts++;
642 
643 	save_txg = zseg->seg_mintxg;
644 
645 	if (!(th->th_advance & ADVANCE_NOLOCK))
646 		rw_enter(rw, RW_READER);
647 
648 	rc = traverse_segment(th, zseg, mosbp);
649 	ASSERT(rc == ERANGE || rc == EAGAIN || rc == EINTR);
650 
651 	if (!(th->th_advance & ADVANCE_NOLOCK))
652 		rw_exit(rw);
653 
654 	zseg->seg_mintxg = save_txg;
655 
656 	if (rc == ERANGE) {
657 		list_remove(&th->th_seglist, zseg);
658 		kmem_free(zseg, sizeof (*zseg));
659 		return (EAGAIN);
660 	}
661 
662 	return (rc);
663 }
664 
665 /*
666  * Note: (mintxg, maxtxg) is an open interval; mintxg and maxtxg themselves
667  * are not included.  The blocks covered by this segment will all have
668  * mintxg < birth < maxtxg.
669  */
670 static void
671 traverse_add_segment(traverse_handle_t *th, uint64_t mintxg, uint64_t maxtxg,
672     uint64_t sobjset, uint64_t sobject, int slevel, uint64_t sblkid,
673     uint64_t eobjset, uint64_t eobject, int elevel, uint64_t eblkid)
674 {
675 	zseg_t *zseg;
676 
677 	zseg = kmem_alloc(sizeof (zseg_t), KM_SLEEP);
678 
679 	zseg->seg_mintxg = mintxg;
680 	zseg->seg_maxtxg = maxtxg;
681 
682 	zseg->seg_start.zb_objset = sobjset;
683 	zseg->seg_start.zb_object = sobject;
684 	zseg->seg_start.zb_level = slevel;
685 	zseg->seg_start.zb_blkid = sblkid;
686 
687 	zseg->seg_end.zb_objset = eobjset;
688 	zseg->seg_end.zb_object = eobject;
689 	zseg->seg_end.zb_level = elevel;
690 	zseg->seg_end.zb_blkid = eblkid;
691 
692 	list_insert_tail(&th->th_seglist, zseg);
693 }
694 
695 void
696 traverse_add_dnode(traverse_handle_t *th, uint64_t mintxg, uint64_t maxtxg,
697     uint64_t objset, uint64_t object)
698 {
699 	if (th->th_advance & ADVANCE_PRE)
700 		traverse_add_segment(th, mintxg, maxtxg,
701 		    objset, object, ZB_MAXLEVEL, 0,
702 		    objset, object, 0, ZB_MAXBLKID);
703 	else
704 		traverse_add_segment(th, mintxg, maxtxg,
705 		    objset, object, 0, 0,
706 		    objset, object, 0, ZB_MAXBLKID);
707 }
708 
709 void
710 traverse_add_objset(traverse_handle_t *th, uint64_t mintxg, uint64_t maxtxg,
711     uint64_t objset)
712 {
713 	if (th->th_advance & ADVANCE_PRE)
714 		traverse_add_segment(th, mintxg, maxtxg,
715 		    objset, 0, -1, 0,
716 		    objset, ZB_MAXOBJECT, 0, ZB_MAXBLKID);
717 	else
718 		traverse_add_segment(th, mintxg, maxtxg,
719 		    objset, 1, 0, 0,
720 		    objset, 0, -1, 0);
721 }
722 
723 void
724 traverse_add_pool(traverse_handle_t *th, uint64_t mintxg, uint64_t maxtxg)
725 {
726 	if (th->th_advance & ADVANCE_PRE)
727 		traverse_add_segment(th, mintxg, maxtxg,
728 		    0, 0, -1, 0,
729 		    ZB_MAXOBJSET, ZB_MAXOBJECT, 0, ZB_MAXBLKID);
730 	else
731 		traverse_add_segment(th, mintxg, maxtxg,
732 		    1, 1, 0, 0,
733 		    0, 0, -1, 0);
734 }
735 
736 traverse_handle_t *
737 traverse_init(spa_t *spa, blkptr_cb_t func, void *arg, int advance,
738     int zio_flags)
739 {
740 	traverse_handle_t *th;
741 	int d, l;
742 
743 	th = kmem_zalloc(sizeof (*th), KM_SLEEP);
744 
745 	th->th_spa = spa;
746 	th->th_func = func;
747 	th->th_arg = arg;
748 	th->th_advance = advance;
749 	th->th_lastcb.zb_level = ZB_NO_LEVEL;
750 	th->th_noread.zb_level = ZB_NO_LEVEL;
751 	th->th_zio_flags = zio_flags;
752 
753 	list_create(&th->th_seglist, sizeof (zseg_t),
754 	    offsetof(zseg_t, seg_node));
755 
756 	for (d = 0; d < ZB_DEPTH; d++) {
757 		for (l = 0; l < ZB_MAXLEVEL; l++) {
758 			if ((advance & ADVANCE_DATA) ||
759 			    l != 0 || d != ZB_DN_CACHE)
760 				th->th_cache[d][l].bc_data =
761 				    zio_buf_alloc(SPA_MAXBLOCKSIZE);
762 		}
763 	}
764 
765 	return (th);
766 }
767 
768 void
769 traverse_fini(traverse_handle_t *th)
770 {
771 	int d, l;
772 	zseg_t *zseg;
773 
774 	for (d = 0; d < ZB_DEPTH; d++)
775 		for (l = 0; l < ZB_MAXLEVEL; l++)
776 			if (th->th_cache[d][l].bc_data != NULL)
777 				zio_buf_free(th->th_cache[d][l].bc_data,
778 				    SPA_MAXBLOCKSIZE);
779 
780 	while ((zseg = list_head(&th->th_seglist)) != NULL) {
781 		list_remove(&th->th_seglist, zseg);
782 		kmem_free(zseg, sizeof (*zseg));
783 	}
784 
785 	list_destroy(&th->th_seglist);
786 
787 	dprintf("%llu hit, %llu ARC, %llu IO, %llu cb, %llu sync, %llu again\n",
788 	    th->th_hits, th->th_arc_hits, th->th_reads, th->th_callbacks,
789 	    th->th_syncs, th->th_restarts);
790 
791 	kmem_free(th, sizeof (*th));
792 }
793