xref: /dragonfly/sys/vfs/hammer/hammer_prune.c (revision 0ca59c34)
1 /*
2  * Copyright (c) 2008 The DragonFly Project.  All rights reserved.
3  *
4  * This code is derived from software contributed to The DragonFly Project
5  * by Matthew Dillon <dillon@backplane.com>
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  *
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in
15  *    the documentation and/or other materials provided with the
16  *    distribution.
17  * 3. Neither the name of The DragonFly Project nor the names of its
18  *    contributors may be used to endorse or promote products derived
19  *    from this software without specific, prior written permission.
20  *
21  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
24  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
25  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
26  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
27  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
28  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
29  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
30  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
31  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32  * SUCH DAMAGE.
33  *
34  * $DragonFly: src/sys/vfs/hammer/hammer_prune.c,v 1.19 2008/09/23 21:03:52 dillon Exp $
35  */
36 
37 #include "hammer.h"
38 
39 /*
40  * Iterate through the specified range of object ids and remove any
41  * deleted records that fall entirely within a prune modulo.
42  *
43  * A reverse iteration is used to prevent overlapping records from being
44  * created during the iteration due to alignments.  This also allows us
45  * to adjust alignments without blowing up the B-Tree.
46  */
47 static int prune_should_delete(struct hammer_ioc_prune *prune,
48 			       hammer_btree_leaf_elm_t elm);
49 static void prune_check_nlinks(hammer_cursor_t cursor,
50 			       hammer_btree_leaf_elm_t elm);
51 
52 int
53 hammer_ioc_prune(hammer_transaction_t trans, hammer_inode_t ip,
54 		 struct hammer_ioc_prune *prune)
55 {
56 	struct hammer_cursor cursor;
57 	hammer_btree_leaf_elm_t elm;
58 	struct hammer_ioc_prune_elm *copy_elms;
59 	struct hammer_ioc_prune_elm *user_elms;
60 	int error;
61 	int isdir;
62 	int elm_array_size;
63 	int seq;
64 	int64_t bytes;
65 	u_int32_t key_beg_localization;
66 
67 	if (prune->nelms < 0 || prune->nelms > HAMMER_MAX_PRUNE_ELMS)
68 		return(EINVAL);
69 	if ((prune->key_beg.localization | prune->key_end.localization) &
70 	    HAMMER_LOCALIZE_PSEUDOFS_MASK) {
71 		return(EINVAL);
72 	}
73 	if (prune->key_beg.localization > prune->key_end.localization)
74 		return(EINVAL);
75 	if (prune->key_beg.localization == prune->key_end.localization) {
76 		if (prune->key_beg.obj_id > prune->key_end.obj_id)
77 			return(EINVAL);
78 		/* key-space limitations - no check needed */
79 	}
80 	if ((prune->head.flags & HAMMER_IOC_PRUNE_ALL) && prune->nelms)
81 		return(EINVAL);
82 
83 	/*
84 	 * Ioctl caller has only set localization type to prune.
85 	 * Initialize cursor key localization with ip localization.
86 	 */
87 	key_beg_localization = prune->key_beg.localization;
88 	key_beg_localization &= HAMMER_LOCALIZE_MASK;
89 	key_beg_localization += ip->obj_localization;
90 
91 	prune->key_cur.localization = prune->key_end.localization;
92 	prune->key_cur.localization &= HAMMER_LOCALIZE_MASK;
93 	prune->key_cur.localization += ip->obj_localization;
94 
95 	prune->key_cur.obj_id = prune->key_end.obj_id;
96 	prune->key_cur.key = HAMMER_MAX_KEY;
97 
98 	/*
99 	 * Copy element array from userland
100 	 */
101 	elm_array_size = sizeof(*copy_elms) * prune->nelms;
102 	user_elms = prune->elms;
103 	copy_elms = kmalloc(elm_array_size, M_TEMP, M_WAITOK);
104 	if ((error = copyin(user_elms, copy_elms, elm_array_size)) != 0)
105 		goto failed;
106 	prune->elms = copy_elms;
107 
108 	seq = trans->hmp->flusher.done;
109 
110 	/*
111 	 * Scan backwards.  Retries typically occur if a deadlock is detected.
112 	 */
113 retry:
114 	error = hammer_init_cursor(trans, &cursor, NULL, NULL);
115 	if (error) {
116 		hammer_done_cursor(&cursor);
117 		goto failed;
118 	}
119 	cursor.key_beg.localization = key_beg_localization;
120 	cursor.key_beg.obj_id = prune->key_beg.obj_id;
121 	cursor.key_beg.key = HAMMER_MIN_KEY;
122 	cursor.key_beg.create_tid = 1;
123 	cursor.key_beg.delete_tid = 0;
124 	cursor.key_beg.rec_type = HAMMER_MIN_RECTYPE;
125 	cursor.key_beg.obj_type = 0;
126 
127 	cursor.key_end.localization = prune->key_cur.localization;
128 	cursor.key_end.obj_id = prune->key_cur.obj_id;
129 	cursor.key_end.key = prune->key_cur.key;
130 	cursor.key_end.create_tid = HAMMER_MAX_TID - 1;
131 	cursor.key_end.delete_tid = 0;
132 	cursor.key_end.rec_type = HAMMER_MAX_RECTYPE;
133 	cursor.key_end.obj_type = 0;
134 
135 	cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE;
136 	cursor.flags |= HAMMER_CURSOR_BACKEND;
137 
138 	/*
139 	 * This flag allows the B-Tree code to clean up loose ends.  At
140 	 * the moment (XXX) it also means we have to hold the sync lock
141 	 * through the iteration.
142 	 */
143 	cursor.flags |= HAMMER_CURSOR_PRUNING;
144 
145 	hammer_sync_lock_sh(trans);
146 	error = hammer_btree_last(&cursor);
147 	hammer_sync_unlock(trans);
148 
149 	while (error == 0) {
150 		/*
151 		 * Check for work
152 		 */
153 		elm = &cursor.node->ondisk->elms[cursor.index].leaf;
154 		prune->key_cur = elm->base;
155 
156 		/*
157 		 * Yield to more important tasks
158 		 */
159 		if ((error = hammer_signal_check(trans->hmp)) != 0)
160 			break;
161 
162 		if (prune->stat_oldest_tid > elm->base.create_tid)
163 			prune->stat_oldest_tid = elm->base.create_tid;
164 
165 		if (hammer_debug_general & 0x0200) {
166 			hdkprintf("check %016llx %016llx cre=%016llx del=%016llx\n",
167 					(long long)elm->base.obj_id,
168 					(long long)elm->base.key,
169 					(long long)elm->base.create_tid,
170 					(long long)elm->base.delete_tid);
171 		}
172 
173 		if (prune_should_delete(prune, elm)) {
174 			if (hammer_debug_general & 0x0200) {
175 				hdkprintf("check %016llx %016llx: DELETE\n",
176 					(long long)elm->base.obj_id,
177 					(long long)elm->base.key);
178 			}
179 
180 			/*
181 			 * NOTE: This can return EDEADLK
182 			 *
183 			 * Acquiring the sync lock guarantees that the
184 			 * operation will not cross a synchronization
185 			 * boundary (see the flusher).
186 			 *
187 			 * We dont need to track inodes or next_tid when
188 			 * we are destroying deleted records.
189 			 */
190 			isdir = (elm->base.rec_type == HAMMER_RECTYPE_DIRENTRY);
191 
192 			hammer_sync_lock_sh(trans);
193 			error = hammer_delete_at_cursor(&cursor,
194 							HAMMER_DELETE_DESTROY,
195 							cursor.trans->tid,
196 							cursor.trans->time32,
197 							0, &bytes);
198 			hammer_sync_unlock(trans);
199 			if (error)
200 				break;
201 
202 			if (isdir)
203 				++prune->stat_dirrecords;
204 			else
205 				++prune->stat_rawrecords;
206 			prune->stat_bytes += bytes;
207 
208 			/*
209 			 * The current record might now be the one after
210 			 * the one we deleted, set ATEDISK to force us
211 			 * to skip it (since we are iterating backwards).
212 			 */
213 			cursor.flags |= HAMMER_CURSOR_ATEDISK;
214 		} else {
215 			/*
216 			 * Nothing to delete, but we may have to check other
217 			 * things.
218 			 */
219 			prune_check_nlinks(&cursor, elm);
220 			cursor.flags |= HAMMER_CURSOR_ATEDISK;
221 			if (hammer_debug_general & 0x0100) {
222 				hdkprintf("check %016llx %016llx: SKIP\n",
223 					(long long)elm->base.obj_id,
224 					(long long)elm->base.key);
225 			}
226 		}
227 		++prune->stat_scanrecords;
228 
229 		/*
230 		 * WARNING: See warnings in hammer_unlock_cursor() function.
231 		 */
232 		while (hammer_flusher_meta_halflimit(trans->hmp) ||
233 		       hammer_flusher_undo_exhausted(trans, 2)) {
234 			hammer_unlock_cursor(&cursor);
235 			hammer_flusher_wait(trans->hmp, seq);
236 			hammer_lock_cursor(&cursor);
237 			seq = hammer_flusher_async_one(trans->hmp);
238 		}
239 		hammer_sync_lock_sh(trans);
240 		error = hammer_btree_iterate_reverse(&cursor);
241 		hammer_sync_unlock(trans);
242 	}
243 	if (error == ENOENT)
244 		error = 0;
245 	hammer_done_cursor(&cursor);
246 	if (error == EDEADLK)
247 		goto retry;
248 	if (error == EINTR) {
249 		prune->head.flags |= HAMMER_IOC_HEAD_INTR;
250 		error = 0;
251 	}
252 failed:
253 	prune->key_cur.localization &= HAMMER_LOCALIZE_MASK;
254 	prune->elms = user_elms;
255 	kfree(copy_elms, M_TEMP);
256 	return(error);
257 }
258 
259 /*
260  * Check pruning list.  The list must be sorted in descending order.
261  *
262  * Return non-zero if the record should be deleted.
263  */
264 static int
265 prune_should_delete(struct hammer_ioc_prune *prune, hammer_btree_leaf_elm_t elm)
266 {
267 	struct hammer_ioc_prune_elm *scan;
268 	int i;
269 
270 	/*
271 	 * If pruning everything remove all records with a non-zero
272 	 * delete_tid.
273 	 */
274 	if (prune->head.flags & HAMMER_IOC_PRUNE_ALL) {
275 		if (elm->base.delete_tid != 0)
276 			return(1);
277 		return(0);
278 	}
279 
280 	for (i = 0; i < prune->nelms; ++i) {
281 		scan = &prune->elms[i];
282 
283 		/*
284 		 * Check for loop termination.
285 		 */
286 		if (elm->base.create_tid >= scan->end_tid ||
287 		    elm->base.delete_tid > scan->end_tid) {
288 			break;
289 		}
290 
291 		/*
292 		 * Determine if we can delete the record.
293 		 */
294 		if (elm->base.delete_tid &&
295 		    elm->base.create_tid >= scan->beg_tid &&
296 		    elm->base.delete_tid <= scan->end_tid &&
297 		    (elm->base.create_tid - scan->beg_tid) / scan->mod_tid ==
298 		    (elm->base.delete_tid - scan->beg_tid) / scan->mod_tid) {
299 			return(1);
300 		}
301 	}
302 	return(0);
303 }
304 
305 /*
306  * Dangling inodes can occur if processes are holding open descriptors on
307  * deleted files as-of when a machine crashes.  When we find one simply
308  * acquire the inode and release it.  The inode handling code will then
309  * do the right thing.
310  */
311 static
312 void
313 prune_check_nlinks(hammer_cursor_t cursor, hammer_btree_leaf_elm_t elm)
314 {
315 	hammer_inode_t ip;
316 	int error;
317 
318 	if (elm->base.rec_type != HAMMER_RECTYPE_INODE)
319 		return;
320 	if (elm->base.delete_tid != 0)
321 		return;
322 	if (hammer_btree_extract(cursor, HAMMER_CURSOR_GET_DATA))
323 		return;
324 	if (cursor->data->inode.nlinks)
325 		return;
326 	hammer_cursor_downgrade(cursor);
327 	ip = hammer_get_inode(cursor->trans, NULL, elm->base.obj_id,
328 		      HAMMER_MAX_TID,
329 		      elm->base.localization & HAMMER_LOCALIZE_PSEUDOFS_MASK,
330 		      0, &error);
331 	if (ip) {
332 		if (hammer_debug_general & 0x0001) {
333 			hdkprintf("pruning disconnected inode %016llx\n",
334 				(long long)elm->base.obj_id);
335 		}
336 		hammer_rel_inode(ip, 0);
337 		hammer_inode_waitreclaims(cursor->trans);
338 	} else {
339 		hkprintf("unable to prune disconnected inode %016llx\n",
340 			(long long)elm->base.obj_id);
341 	}
342 }
343 
344 #if 0
345 
346 /*
347  * NOTE: THIS CODE HAS BEEN REMOVED!  Pruning no longer attempts to realign
348  *	 adjacent records because it seriously interferes with every
349  *	 mirroring algorithm I could come up with.
350  *
351  *	 This means that historical accesses beyond the first snapshot
352  *	 softlink should be on snapshot boundaries only.  Historical
353  *	 accesses from "now" to the first snapshot softlink continue to
354  *	 be fine-grained.
355  *
356  * NOTE: It also looks like there's a bug in the removed code.  It is believed
357  *	 that create_tid can sometimes get set to 0xffffffffffffffff.  Just as
358  *	 well we no longer try to do this fancy shit.  Probably the attempt to
359  *	 correct the rhb is blowing up the cursor's indexing or addressing mapping.
360  *
361  * Align the record to cover any gaps created through the deletion of
362  * records within the pruning space.  If we were to just delete the records
363  * there would be gaps which in turn would cause a snapshot that is NOT on
364  * a pruning boundary to appear corrupt to the user.  Forcing alignment
365  * of the create_tid and delete_tid for retained records 'reconnects'
366  * the previously contiguous space, making it contiguous again after the
367  * deletions.
368  *
369  * The use of a reverse iteration allows us to safely align the records and
370  * related elements without creating temporary overlaps.  XXX we should
371  * add ordering dependancies for record buffers to guarantee consistency
372  * during recovery.
373  */
374 static int
375 realign_prune(struct hammer_ioc_prune *prune,
376 	      hammer_cursor_t cursor, int realign_cre, int realign_del)
377 {
378 	struct hammer_ioc_prune_elm *scan;
379 	hammer_btree_elm_t elm;
380 	hammer_tid_t delta;
381 	hammer_tid_t tid;
382 	int error;
383 
384 	hammer_cursor_downgrade(cursor);
385 
386 	elm = &cursor->node->ondisk->elms[cursor->index];
387 	++prune->stat_realignments;
388 
389 	/*
390 	 * Align the create_tid.  By doing a reverse iteration we guarantee
391 	 * that all records after our current record have already been
392 	 * aligned, allowing us to safely correct the right-hand-boundary
393 	 * (because no record to our right is otherwise exactly matching
394 	 * will have a create_tid to the left of our aligned create_tid).
395 	 */
396 	error = 0;
397 	if (realign_cre >= 0) {
398 		scan = &prune->elms[realign_cre];
399 
400 		delta = (elm->leaf.base.create_tid - scan->beg_tid) %
401 			scan->mod_tid;
402 		if (delta) {
403 			tid = elm->leaf.base.create_tid - delta + scan->mod_tid;
404 
405 			/* can EDEADLK */
406 			error = hammer_btree_correct_rhb(cursor, tid + 1);
407 			if (error == 0) {
408 				error = hammer_btree_extract(cursor,
409 						     HAMMER_CURSOR_GET_LEAF);
410 			}
411 			if (error == 0) {
412 				/* can EDEADLK */
413 				error = hammer_cursor_upgrade(cursor);
414 			}
415 			if (error == 0) {
416 				hammer_modify_node(cursor->trans, cursor->node,
417 					    &elm->leaf.base.create_tid,
418 					    sizeof(elm->leaf.base.create_tid));
419 				elm->leaf.base.create_tid = tid;
420 				hammer_modify_node_done(cursor->node);
421 			}
422 		}
423 	}
424 
425 	/*
426 	 * Align the delete_tid.  This only occurs if the record is historical
427 	 * was deleted at some point.  Realigning the delete_tid does not
428 	 * move the record within the B-Tree but may cause it to temporarily
429 	 * overlap a record that has not yet been pruned.
430 	 */
431 	if (error == 0 && realign_del >= 0) {
432 		scan = &prune->elms[realign_del];
433 
434 		delta = (elm->leaf.base.delete_tid - scan->beg_tid) %
435 			scan->mod_tid;
436 		if (delta) {
437 			error = hammer_btree_extract(cursor,
438 						     HAMMER_CURSOR_GET_LEAF);
439 			if (error == 0) {
440 				hammer_modify_node(cursor->trans, cursor->node,
441 					    &elm->leaf.base.delete_tid,
442 					    sizeof(elm->leaf.base.delete_tid));
443 				elm->leaf.base.delete_tid =
444 					    elm->leaf.base.delete_tid -
445 					    delta + scan->mod_tid;
446 				hammer_modify_node_done(cursor->node);
447 			}
448 		}
449 	}
450 	return (error);
451 }
452 
453 #endif
454