kernel - Incidental mplock removal - NTP support
[dragonfly.git] / sys / vfs / hammer / hammer_prune.c
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         uint32_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                  * Filesystem went read-only during rebalancing
158                  */
159                 if (trans->hmp->ronly) {
160                         error = EROFS;
161                         break;
162                 }
163
164                 /*
165                  * Yield to more important tasks
166                  */
167                 if ((error = hammer_signal_check(trans->hmp)) != 0)
168                         break;
169
170                 if (prune->stat_oldest_tid > elm->base.create_tid)
171                         prune->stat_oldest_tid = elm->base.create_tid;
172
173                 if (hammer_debug_general & 0x0200) {
174                         hdkprintf("check %016jx %016jx cre=%016jx del=%016jx\n",
175                                         (intmax_t)elm->base.obj_id,
176                                         (intmax_t)elm->base.key,
177                                         (intmax_t)elm->base.create_tid,
178                                         (intmax_t)elm->base.delete_tid);
179                 }
180
181                 if (prune_should_delete(prune, elm)) {
182                         if (hammer_debug_general & 0x0200) {
183                                 hdkprintf("check %016jx %016jx: DELETE\n",
184                                         (intmax_t)elm->base.obj_id,
185                                         (intmax_t)elm->base.key);
186                         }
187
188                         /*
189                          * NOTE: This can return EDEADLK
190                          *
191                          * Acquiring the sync lock guarantees that the
192                          * operation will not cross a synchronization
193                          * boundary (see the flusher).
194                          *
195                          * We dont need to track inodes or next_tid when
196                          * we are destroying deleted records.
197                          */
198                         isdir = (elm->base.rec_type == HAMMER_RECTYPE_DIRENTRY);
199
200                         hammer_sync_lock_sh(trans);
201                         error = hammer_delete_at_cursor(&cursor,
202                                                         HAMMER_DELETE_DESTROY,
203                                                         cursor.trans->tid,
204                                                         cursor.trans->time32,
205                                                         0, &bytes);
206                         hammer_sync_unlock(trans);
207                         if (error)
208                                 break;
209
210                         if (isdir)
211                                 ++prune->stat_dirrecords;
212                         else
213                                 ++prune->stat_rawrecords;
214                         prune->stat_bytes += bytes;
215
216                         /*
217                          * The current record might now be the one after
218                          * the one we deleted, set ATEDISK to force us
219                          * to skip it (since we are iterating backwards).
220                          */
221                         cursor.flags |= HAMMER_CURSOR_ATEDISK;
222                 } else {
223                         /*
224                          * Nothing to delete, but we may have to check other
225                          * things.
226                          */
227                         prune_check_nlinks(&cursor, elm);
228                         cursor.flags |= HAMMER_CURSOR_ATEDISK;
229                         if (hammer_debug_general & 0x0100) {
230                                 hdkprintf("check %016jx %016jx: SKIP\n",
231                                         (intmax_t)elm->base.obj_id,
232                                         (intmax_t)elm->base.key);
233                         }
234                 }
235                 ++prune->stat_scanrecords;
236
237                 /*
238                  * WARNING: See warnings in hammer_unlock_cursor() function.
239                  */
240                 while (hammer_flusher_meta_halflimit(trans->hmp) ||
241                        hammer_flusher_undo_exhausted(trans, 2)) {
242                         hammer_unlock_cursor(&cursor);
243                         hammer_flusher_wait(trans->hmp, seq);
244                         hammer_lock_cursor(&cursor);
245                         seq = hammer_flusher_async_one(trans->hmp);
246                 }
247                 hammer_sync_lock_sh(trans);
248                 error = hammer_btree_iterate_reverse(&cursor);
249                 hammer_sync_unlock(trans);
250         }
251         if (error == ENOENT)
252                 error = 0;
253         hammer_done_cursor(&cursor);
254         if (error == EDEADLK)
255                 goto retry;
256         if (error == EINTR) {
257                 prune->head.flags |= HAMMER_IOC_HEAD_INTR;
258                 error = 0;
259         }
260 failed:
261         prune->key_cur.localization &= HAMMER_LOCALIZE_MASK;
262         prune->elms = user_elms;
263         kfree(copy_elms, M_TEMP);
264         return(error);
265 }
266
267 /*
268  * Check pruning list.  The list must be sorted in descending order.
269  *
270  * Return non-zero if the record should be deleted.
271  */
272 static int
273 prune_should_delete(struct hammer_ioc_prune *prune, hammer_btree_leaf_elm_t elm)
274 {
275         struct hammer_ioc_prune_elm *scan;
276         int i;
277
278         /*
279          * If pruning everything remove all records with a non-zero
280          * delete_tid.
281          */
282         if (prune->head.flags & HAMMER_IOC_PRUNE_ALL) {
283                 if (elm->base.delete_tid != 0)
284                         return(1);
285                 return(0);
286         }
287
288         for (i = 0; i < prune->nelms; ++i) {
289                 scan = &prune->elms[i];
290
291                 /*
292                  * Check for loop termination.
293                  */
294                 if (elm->base.create_tid >= scan->end_tid ||
295                     elm->base.delete_tid > scan->end_tid) {
296                         break;
297                 }
298
299                 /*
300                  * Determine if we can delete the record.
301                  */
302                 if (elm->base.delete_tid &&
303                     elm->base.create_tid >= scan->beg_tid &&
304                     elm->base.delete_tid <= scan->end_tid &&
305                     (elm->base.create_tid - scan->beg_tid) / scan->mod_tid ==
306                     (elm->base.delete_tid - scan->beg_tid) / scan->mod_tid) {
307                         return(1);
308                 }
309         }
310         return(0);
311 }
312
313 /*
314  * Dangling inodes can occur if processes are holding open descriptors on
315  * deleted files as-of when a machine crashes.  When we find one simply
316  * acquire the inode and release it.  The inode handling code will then
317  * do the right thing.
318  */
319 static
320 void
321 prune_check_nlinks(hammer_cursor_t cursor, hammer_btree_leaf_elm_t elm)
322 {
323         hammer_inode_t ip;
324         int error;
325
326         if (elm->base.rec_type != HAMMER_RECTYPE_INODE)
327                 return;
328         if (elm->base.delete_tid != 0)
329                 return;
330         if (hammer_btree_extract_data(cursor))
331                 return;
332         if (cursor->data->inode.nlinks)
333                 return;
334         hammer_cursor_downgrade(cursor);
335         ip = hammer_get_inode(cursor->trans, NULL, elm->base.obj_id,
336                       HAMMER_MAX_TID,
337                       elm->base.localization & HAMMER_LOCALIZE_PSEUDOFS_MASK,
338                       0, &error);
339         if (ip) {
340                 if (hammer_debug_general & 0x0001) {
341                         hdkprintf("pruning disconnected inode %016jx\n",
342                                 (intmax_t)elm->base.obj_id);
343                 }
344                 hammer_rel_inode(ip, 0);
345                 hammer_inode_waitreclaims(cursor->trans);
346         } else {
347                 hkprintf("unable to prune disconnected inode %016jx\n",
348                         (intmax_t)elm->base.obj_id);
349         }
350 }
351
352 #if 0
353
354 /*
355  * NOTE: THIS CODE HAS BEEN REMOVED!  Pruning no longer attempts to realign
356  *       adjacent records because it seriously interferes with every
357  *       mirroring algorithm I could come up with.
358  *
359  *       This means that historical accesses beyond the first snapshot
360  *       softlink should be on snapshot boundaries only.  Historical
361  *       accesses from "now" to the first snapshot softlink continue to
362  *       be fine-grained.
363  *
364  * NOTE: It also looks like there's a bug in the removed code.  It is believed
365  *       that create_tid can sometimes get set to 0xffffffffffffffff.  Just as
366  *       well we no longer try to do this fancy shit.  Probably the attempt to
367  *       correct the rhb is blowing up the cursor's indexing or addressing mapping.
368  *
369  * Align the record to cover any gaps created through the deletion of
370  * records within the pruning space.  If we were to just delete the records
371  * there would be gaps which in turn would cause a snapshot that is NOT on
372  * a pruning boundary to appear corrupt to the user.  Forcing alignment
373  * of the create_tid and delete_tid for retained records 'reconnects'
374  * the previously contiguous space, making it contiguous again after the
375  * deletions.
376  *
377  * The use of a reverse iteration allows us to safely align the records and
378  * related elements without creating temporary overlaps.  XXX we should
379  * add ordering dependancies for record buffers to guarantee consistency
380  * during recovery.
381  */
382 static int
383 realign_prune(struct hammer_ioc_prune *prune,
384               hammer_cursor_t cursor, int realign_cre, int realign_del)
385 {
386         struct hammer_ioc_prune_elm *scan;
387         hammer_btree_elm_t elm;
388         hammer_tid_t delta;
389         hammer_tid_t tid;
390         int error;
391
392         hammer_cursor_downgrade(cursor);
393
394         elm = &cursor->node->ondisk->elms[cursor->index];
395         ++prune->stat_realignments;
396
397         /*
398          * Align the create_tid.  By doing a reverse iteration we guarantee
399          * that all records after our current record have already been
400          * aligned, allowing us to safely correct the right-hand-boundary
401          * (because no record to our right is otherwise exactly matching
402          * will have a create_tid to the left of our aligned create_tid).
403          */
404         error = 0;
405         if (realign_cre >= 0) {
406                 scan = &prune->elms[realign_cre];
407
408                 delta = (elm->leaf.base.create_tid - scan->beg_tid) %
409                         scan->mod_tid;
410                 if (delta) {
411                         tid = elm->leaf.base.create_tid - delta + scan->mod_tid;
412
413                         /* can EDEADLK */
414                         error = hammer_btree_correct_rhb(cursor, tid + 1);
415                         if (error == 0) {
416                                 error = hammer_btree_extract_leaf(cursor);
417                         }
418                         if (error == 0) {
419                                 /* can EDEADLK */
420                                 error = hammer_cursor_upgrade(cursor);
421                         }
422                         if (error == 0) {
423                                 hammer_modify_node(cursor->trans, cursor->node,
424                                             &elm->leaf.base.create_tid,
425                                             sizeof(elm->leaf.base.create_tid));
426                                 elm->leaf.base.create_tid = tid;
427                                 hammer_modify_node_done(cursor->node);
428                         }
429                 }
430         }
431
432         /*
433          * Align the delete_tid.  This only occurs if the record is historical
434          * was deleted at some point.  Realigning the delete_tid does not
435          * move the record within the B-Tree but may cause it to temporarily
436          * overlap a record that has not yet been pruned.
437          */
438         if (error == 0 && realign_del >= 0) {
439                 scan = &prune->elms[realign_del];
440
441                 delta = (elm->leaf.base.delete_tid - scan->beg_tid) %
442                         scan->mod_tid;
443                 if (delta) {
444                         error = hammer_btree_extract_leaf(cursor);
445                         if (error == 0) {
446                                 hammer_modify_node(cursor->trans, cursor->node,
447                                             &elm->leaf.base.delete_tid,
448                                             sizeof(elm->leaf.base.delete_tid));
449                                 elm->leaf.base.delete_tid =
450                                             elm->leaf.base.delete_tid -
451                                             delta + scan->mod_tid;
452                                 hammer_modify_node_done(cursor->node);
453                         }
454                 }
455         }
456         return (error);
457 }
458
459 #endif