xref: /titanic_41/usr/src/uts/common/fs/zfs/dmu_traverse.c (revision a31148363f598def767ac48c5d82e1572e44b935)
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 2010 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 
26 #include <sys/zfs_context.h>
27 #include <sys/dmu_objset.h>
28 #include <sys/dmu_traverse.h>
29 #include <sys/dsl_dataset.h>
30 #include <sys/dsl_dir.h>
31 #include <sys/dsl_pool.h>
32 #include <sys/dnode.h>
33 #include <sys/spa.h>
34 #include <sys/zio.h>
35 #include <sys/dmu_impl.h>
36 #include <sys/sa.h>
37 #include <sys/sa_impl.h>
38 #include <sys/callb.h>
39 
40 struct prefetch_data {
41 	kmutex_t pd_mtx;
42 	kcondvar_t pd_cv;
43 	int pd_blks_max;
44 	int pd_blks_fetched;
45 	int pd_flags;
46 	boolean_t pd_cancel;
47 	boolean_t pd_exited;
48 };
49 
50 struct traverse_data {
51 	spa_t *td_spa;
52 	uint64_t td_objset;
53 	blkptr_t *td_rootbp;
54 	uint64_t td_min_txg;
55 	int td_flags;
56 	struct prefetch_data *td_pfd;
57 	blkptr_cb_t *td_func;
58 	void *td_arg;
59 };
60 
61 static int traverse_dnode(struct traverse_data *td, const dnode_phys_t *dnp,
62     arc_buf_t *buf, uint64_t objset, uint64_t object);
63 
64 /* ARGSUSED */
65 static int
66 traverse_zil_block(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
67 {
68 	struct traverse_data *td = arg;
69 	zbookmark_t zb;
70 
71 	if (bp->blk_birth == 0)
72 		return (0);
73 
74 	if (claim_txg == 0 && bp->blk_birth >= spa_first_txg(td->td_spa))
75 		return (0);
76 
77 	SET_BOOKMARK(&zb, td->td_objset, ZB_ZIL_OBJECT, ZB_ZIL_LEVEL,
78 	    bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
79 
80 	(void) td->td_func(td->td_spa, zilog, bp, &zb, NULL, td->td_arg);
81 
82 	return (0);
83 }
84 
85 /* ARGSUSED */
86 static int
87 traverse_zil_record(zilog_t *zilog, lr_t *lrc, void *arg, uint64_t claim_txg)
88 {
89 	struct traverse_data *td = arg;
90 
91 	if (lrc->lrc_txtype == TX_WRITE) {
92 		lr_write_t *lr = (lr_write_t *)lrc;
93 		blkptr_t *bp = &lr->lr_blkptr;
94 		zbookmark_t zb;
95 
96 		if (bp->blk_birth == 0)
97 			return (0);
98 
99 		if (claim_txg == 0 || bp->blk_birth < claim_txg)
100 			return (0);
101 
102 		SET_BOOKMARK(&zb, td->td_objset, lr->lr_foid, ZB_ZIL_LEVEL,
103 		    lr->lr_offset / BP_GET_LSIZE(bp));
104 
105 		(void) td->td_func(td->td_spa, zilog, bp, &zb, NULL,
106 		    td->td_arg);
107 	}
108 	return (0);
109 }
110 
111 static void
112 traverse_zil(struct traverse_data *td, zil_header_t *zh)
113 {
114 	uint64_t claim_txg = zh->zh_claim_txg;
115 	zilog_t *zilog;
116 
117 	/*
118 	 * We only want to visit blocks that have been claimed but not yet
119 	 * replayed; plus, in read-only mode, blocks that are already stable.
120 	 */
121 	if (claim_txg == 0 && spa_writeable(td->td_spa))
122 		return;
123 
124 	zilog = zil_alloc(spa_get_dsl(td->td_spa)->dp_meta_objset, zh);
125 
126 	(void) zil_parse(zilog, traverse_zil_block, traverse_zil_record, td,
127 	    claim_txg);
128 
129 	zil_free(zilog);
130 }
131 
132 static int
133 traverse_visitbp(struct traverse_data *td, const dnode_phys_t *dnp,
134     arc_buf_t *pbuf, blkptr_t *bp, const zbookmark_t *zb)
135 {
136 	zbookmark_t czb;
137 	int err = 0, lasterr = 0;
138 	arc_buf_t *buf = NULL;
139 	struct prefetch_data *pd = td->td_pfd;
140 	boolean_t hard = td->td_flags & TRAVERSE_HARD;
141 
142 	if (bp->blk_birth == 0) {
143 		err = td->td_func(td->td_spa, NULL, NULL, zb, dnp, td->td_arg);
144 		return (err);
145 	}
146 
147 	if (bp->blk_birth <= td->td_min_txg)
148 		return (0);
149 
150 	if (pd && !pd->pd_exited &&
151 	    ((pd->pd_flags & TRAVERSE_PREFETCH_DATA) ||
152 	    BP_GET_TYPE(bp) == DMU_OT_DNODE || BP_GET_LEVEL(bp) > 0)) {
153 		mutex_enter(&pd->pd_mtx);
154 		ASSERT(pd->pd_blks_fetched >= 0);
155 		while (pd->pd_blks_fetched == 0 && !pd->pd_exited)
156 			cv_wait(&pd->pd_cv, &pd->pd_mtx);
157 		pd->pd_blks_fetched--;
158 		cv_broadcast(&pd->pd_cv);
159 		mutex_exit(&pd->pd_mtx);
160 	}
161 
162 	if (td->td_flags & TRAVERSE_PRE) {
163 		err = td->td_func(td->td_spa, NULL, bp, zb, dnp, td->td_arg);
164 		if (err)
165 			return (err);
166 	}
167 
168 	if (BP_GET_LEVEL(bp) > 0) {
169 		uint32_t flags = ARC_WAIT;
170 		int i;
171 		blkptr_t *cbp;
172 		int epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT;
173 
174 		err = arc_read(NULL, td->td_spa, bp, pbuf,
175 		    arc_getbuf_func, &buf,
176 		    ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb);
177 		if (err)
178 			return (err);
179 
180 		/* recursively visitbp() blocks below this */
181 		cbp = buf->b_data;
182 		for (i = 0; i < epb; i++, cbp++) {
183 			SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object,
184 			    zb->zb_level - 1,
185 			    zb->zb_blkid * epb + i);
186 			err = traverse_visitbp(td, dnp, buf, cbp, &czb);
187 			if (err) {
188 				if (!hard)
189 					break;
190 				lasterr = err;
191 			}
192 		}
193 	} else if (BP_GET_TYPE(bp) == DMU_OT_DNODE) {
194 		uint32_t flags = ARC_WAIT;
195 		int i;
196 		int epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT;
197 
198 		err = arc_read(NULL, td->td_spa, bp, pbuf,
199 		    arc_getbuf_func, &buf,
200 		    ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb);
201 		if (err)
202 			return (err);
203 
204 		/* recursively visitbp() blocks below this */
205 		dnp = buf->b_data;
206 		for (i = 0; i < epb; i++, dnp++) {
207 			err = traverse_dnode(td, dnp, buf, zb->zb_objset,
208 			    zb->zb_blkid * epb + i);
209 			if (err) {
210 				if (!hard)
211 					break;
212 				lasterr = err;
213 			}
214 		}
215 	} else if (BP_GET_TYPE(bp) == DMU_OT_OBJSET) {
216 		uint32_t flags = ARC_WAIT;
217 		objset_phys_t *osp;
218 		dnode_phys_t *dnp;
219 
220 		err = arc_read_nolock(NULL, td->td_spa, bp,
221 		    arc_getbuf_func, &buf,
222 		    ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb);
223 		if (err)
224 			return (err);
225 
226 		osp = buf->b_data;
227 		traverse_zil(td, &osp->os_zil_header);
228 
229 		dnp = &osp->os_meta_dnode;
230 		err = traverse_dnode(td, dnp, buf, zb->zb_objset,
231 		    DMU_META_DNODE_OBJECT);
232 		if (err && hard) {
233 			lasterr = err;
234 			err = 0;
235 		}
236 		if (err == 0 && arc_buf_size(buf) >= sizeof (objset_phys_t)) {
237 			dnp = &osp->os_userused_dnode;
238 			err = traverse_dnode(td, dnp, buf, zb->zb_objset,
239 			    DMU_USERUSED_OBJECT);
240 		}
241 		if (err && hard) {
242 			lasterr = err;
243 			err = 0;
244 		}
245 		if (err == 0 && arc_buf_size(buf) >= sizeof (objset_phys_t)) {
246 			dnp = &osp->os_groupused_dnode;
247 			err = traverse_dnode(td, dnp, buf, zb->zb_objset,
248 			    DMU_GROUPUSED_OBJECT);
249 		}
250 	}
251 
252 	if (buf)
253 		(void) arc_buf_remove_ref(buf, &buf);
254 
255 	if (err == 0 && lasterr == 0 && (td->td_flags & TRAVERSE_POST))
256 		err = td->td_func(td->td_spa, NULL, bp, zb, dnp, td->td_arg);
257 
258 	return (err != 0 ? err : lasterr);
259 }
260 
261 static int
262 traverse_dnode(struct traverse_data *td, const dnode_phys_t *dnp,
263     arc_buf_t *buf, uint64_t objset, uint64_t object)
264 {
265 	int j, err = 0, lasterr = 0;
266 	zbookmark_t czb;
267 	boolean_t hard = (td->td_flags & TRAVERSE_HARD);
268 
269 	for (j = 0; j < dnp->dn_nblkptr; j++) {
270 		SET_BOOKMARK(&czb, objset, object, dnp->dn_nlevels - 1, j);
271 		err = traverse_visitbp(td, dnp, buf,
272 		    (blkptr_t *)&dnp->dn_blkptr[j], &czb);
273 		if (err) {
274 			if (!hard)
275 				break;
276 			lasterr = err;
277 		}
278 		if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
279 			SET_BOOKMARK(&czb, objset,
280 			    object, 0, DMU_SPILL_BLKID);
281 			err = traverse_visitbp(td, dnp, buf,
282 			    (blkptr_t *)&dnp->dn_spill, &czb);
283 			if (err) {
284 				if (!hard)
285 					break;
286 				lasterr = err;
287 			}
288 		}
289 	}
290 	return (err != 0 ? err : lasterr);
291 }
292 
293 /* ARGSUSED */
294 static int
295 traverse_prefetcher(spa_t *spa, zilog_t *zilog, const blkptr_t *bp,
296     const zbookmark_t *zb, const dnode_phys_t *dnp, void *arg)
297 {
298 	struct prefetch_data *pfd = arg;
299 	uint32_t aflags = ARC_NOWAIT | ARC_PREFETCH;
300 
301 	ASSERT(pfd->pd_blks_fetched >= 0);
302 	if (pfd->pd_cancel)
303 		return (EINTR);
304 
305 	if (bp == NULL || !((pfd->pd_flags & TRAVERSE_PREFETCH_DATA) ||
306 	    BP_GET_TYPE(bp) == DMU_OT_DNODE || BP_GET_LEVEL(bp) > 0) ||
307 	    BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG)
308 		return (0);
309 
310 	mutex_enter(&pfd->pd_mtx);
311 	while (!pfd->pd_cancel && pfd->pd_blks_fetched >= pfd->pd_blks_max)
312 		cv_wait(&pfd->pd_cv, &pfd->pd_mtx);
313 	pfd->pd_blks_fetched++;
314 	cv_broadcast(&pfd->pd_cv);
315 	mutex_exit(&pfd->pd_mtx);
316 
317 	(void) arc_read_nolock(NULL, spa, bp, NULL, NULL,
318 	    ZIO_PRIORITY_ASYNC_READ,
319 	    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
320 	    &aflags, zb);
321 
322 	return (0);
323 }
324 
325 static void
326 traverse_prefetch_thread(void *arg)
327 {
328 	struct traverse_data *td_main = arg;
329 	struct traverse_data td = *td_main;
330 	zbookmark_t czb;
331 
332 	td.td_func = traverse_prefetcher;
333 	td.td_arg = td_main->td_pfd;
334 	td.td_pfd = NULL;
335 
336 	SET_BOOKMARK(&czb, td.td_objset,
337 	    ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
338 	(void) traverse_visitbp(&td, NULL, NULL, td.td_rootbp, &czb);
339 
340 	mutex_enter(&td_main->td_pfd->pd_mtx);
341 	td_main->td_pfd->pd_exited = B_TRUE;
342 	cv_broadcast(&td_main->td_pfd->pd_cv);
343 	mutex_exit(&td_main->td_pfd->pd_mtx);
344 }
345 
346 /*
347  * NB: dataset must not be changing on-disk (eg, is a snapshot or we are
348  * in syncing context).
349  */
350 static int
351 traverse_impl(spa_t *spa, uint64_t objset, blkptr_t *rootbp,
352     uint64_t txg_start, int flags, blkptr_cb_t func, void *arg)
353 {
354 	struct traverse_data td;
355 	struct prefetch_data pd = { 0 };
356 	zbookmark_t czb;
357 	int err;
358 
359 	td.td_spa = spa;
360 	td.td_objset = objset;
361 	td.td_rootbp = rootbp;
362 	td.td_min_txg = txg_start;
363 	td.td_func = func;
364 	td.td_arg = arg;
365 	td.td_pfd = &pd;
366 	td.td_flags = flags;
367 
368 	pd.pd_blks_max = 100;
369 	pd.pd_flags = flags;
370 	mutex_init(&pd.pd_mtx, NULL, MUTEX_DEFAULT, NULL);
371 	cv_init(&pd.pd_cv, NULL, CV_DEFAULT, NULL);
372 
373 	if (!(flags & TRAVERSE_PREFETCH) ||
374 	    0 == taskq_dispatch(system_taskq, traverse_prefetch_thread,
375 	    &td, TQ_NOQUEUE))
376 		pd.pd_exited = B_TRUE;
377 
378 	SET_BOOKMARK(&czb, objset,
379 	    ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
380 	err = traverse_visitbp(&td, NULL, NULL, rootbp, &czb);
381 
382 	mutex_enter(&pd.pd_mtx);
383 	pd.pd_cancel = B_TRUE;
384 	cv_broadcast(&pd.pd_cv);
385 	while (!pd.pd_exited)
386 		cv_wait(&pd.pd_cv, &pd.pd_mtx);
387 	mutex_exit(&pd.pd_mtx);
388 
389 	mutex_destroy(&pd.pd_mtx);
390 	cv_destroy(&pd.pd_cv);
391 
392 	return (err);
393 }
394 
395 /*
396  * NB: dataset must not be changing on-disk (eg, is a snapshot or we are
397  * in syncing context).
398  */
399 int
400 traverse_dataset(dsl_dataset_t *ds, uint64_t txg_start, int flags,
401     blkptr_cb_t func, void *arg)
402 {
403 	return (traverse_impl(ds->ds_dir->dd_pool->dp_spa, ds->ds_object,
404 	    &ds->ds_phys->ds_bp, txg_start, flags, func, arg));
405 }
406 
407 /*
408  * NB: pool must not be changing on-disk (eg, from zdb or sync context).
409  */
410 int
411 traverse_pool(spa_t *spa, uint64_t txg_start, int flags,
412     blkptr_cb_t func, void *arg)
413 {
414 	int err, lasterr = 0;
415 	uint64_t obj;
416 	dsl_pool_t *dp = spa_get_dsl(spa);
417 	objset_t *mos = dp->dp_meta_objset;
418 	boolean_t hard = (flags & TRAVERSE_HARD);
419 
420 	/* visit the MOS */
421 	err = traverse_impl(spa, 0, spa_get_rootblkptr(spa),
422 	    txg_start, flags, func, arg);
423 	if (err)
424 		return (err);
425 
426 	/* visit each dataset */
427 	for (obj = 1; err == 0 || (err != ESRCH && hard);
428 	    err = dmu_object_next(mos, &obj, FALSE, txg_start)) {
429 		dmu_object_info_t doi;
430 
431 		err = dmu_object_info(mos, obj, &doi);
432 		if (err) {
433 			if (!hard)
434 				return (err);
435 			lasterr = err;
436 			continue;
437 		}
438 
439 		if (doi.doi_type == DMU_OT_DSL_DATASET) {
440 			dsl_dataset_t *ds;
441 			uint64_t txg = txg_start;
442 
443 			rw_enter(&dp->dp_config_rwlock, RW_READER);
444 			err = dsl_dataset_hold_obj(dp, obj, FTAG, &ds);
445 			rw_exit(&dp->dp_config_rwlock);
446 			if (err) {
447 				if (!hard)
448 					return (err);
449 				lasterr = err;
450 				continue;
451 			}
452 			if (ds->ds_phys->ds_prev_snap_txg > txg)
453 				txg = ds->ds_phys->ds_prev_snap_txg;
454 			err = traverse_dataset(ds, txg, flags, func, arg);
455 			dsl_dataset_rele(ds, FTAG);
456 			if (err) {
457 				if (!hard)
458 					return (err);
459 				lasterr = err;
460 			}
461 		}
462 	}
463 	if (err == ESRCH)
464 		err = 0;
465 	return (err != 0 ? err : lasterr);
466 }
467