HAMMER 60C/many: Mirroring
[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 1.12 2008/07/04 07:25:36 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
64         if (prune->nelms < 0 || prune->nelms > HAMMER_MAX_PRUNE_ELMS)
65                 return(EINVAL);
66         if ((prune->key_beg.localization | prune->key_end.localization) &
67             HAMMER_LOCALIZE_PSEUDOFS_MASK) {
68                 return(EINVAL);
69         }
70         if (prune->key_beg.localization > prune->key_end.localization)
71                 return(EINVAL);
72         if (prune->key_beg.localization == prune->key_end.localization) {
73                 if (prune->key_beg.obj_id > prune->key_end.obj_id)
74                         return(EINVAL);
75                 /* key-space limitations - no check needed */
76         }
77         if ((prune->head.flags & HAMMER_IOC_PRUNE_ALL) && prune->nelms)
78                 return(EINVAL);
79
80         prune->key_cur.localization = prune->key_end.localization +
81                                       ip->obj_localization;
82         prune->key_cur.obj_id = prune->key_end.obj_id;
83         prune->key_cur.key = HAMMER_MAX_KEY;
84
85         /*
86          * Copy element array from userland
87          */
88         elm_array_size = sizeof(*copy_elms) * prune->nelms;
89         user_elms = prune->elms;
90         copy_elms = kmalloc(elm_array_size, M_TEMP, M_WAITOK);
91         if ((error = copyin(user_elms, copy_elms, elm_array_size)) != 0)
92                 goto failed;
93         prune->elms = copy_elms;
94
95         /*
96          * Scan backwards.  Retries typically occur if a deadlock is detected.
97          */
98 retry:
99         error = hammer_init_cursor(trans, &cursor, NULL, NULL);
100         if (error) {
101                 hammer_done_cursor(&cursor);
102                 goto failed;
103         }
104         cursor.key_beg.localization = prune->key_beg.localization +
105                                       ip->obj_localization;
106         cursor.key_beg.obj_id = prune->key_beg.obj_id;
107         cursor.key_beg.key = HAMMER_MIN_KEY;
108         cursor.key_beg.create_tid = 1;
109         cursor.key_beg.delete_tid = 0;
110         cursor.key_beg.rec_type = HAMMER_MIN_RECTYPE;
111         cursor.key_beg.obj_type = 0;
112
113         cursor.key_end.localization = prune->key_cur.localization;
114         cursor.key_end.obj_id = prune->key_cur.obj_id;
115         cursor.key_end.key = prune->key_cur.key;
116         cursor.key_end.create_tid = HAMMER_MAX_TID - 1;
117         cursor.key_end.delete_tid = 0;
118         cursor.key_end.rec_type = HAMMER_MAX_RECTYPE;
119         cursor.key_end.obj_type = 0;
120
121         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE;
122         cursor.flags |= HAMMER_CURSOR_BACKEND;
123
124         /*
125          * This flag allows the B-Tree code to clean up loose ends.
126          */
127         cursor.flags |= HAMMER_CURSOR_PRUNING;
128
129         hammer_sync_lock_sh(trans);
130         error = hammer_btree_last(&cursor);
131
132         while (error == 0) {
133                 /*
134                  * Check for work
135                  */
136                 elm = &cursor.node->ondisk->elms[cursor.index].leaf;
137                 prune->key_cur = elm->base;
138
139                 /*
140                  * Yield to more important tasks
141                  */
142                 if ((error = hammer_signal_check(trans->hmp)) != 0)
143                         break;
144                 if (trans->hmp->sync_lock.wanted) {
145                         hammer_sync_unlock(trans);
146                         tsleep(trans, 0, "hmrslo", hz / 10);
147                         hammer_sync_lock_sh(trans);
148                 }
149                 if (hammer_flusher_meta_limit(trans->hmp) ||
150                     hammer_flusher_undo_exhausted(trans, 2)) {
151                         error = EWOULDBLOCK;
152                         break;
153                 }
154
155                 if (prune->stat_oldest_tid > elm->base.create_tid)
156                         prune->stat_oldest_tid = elm->base.create_tid;
157
158                 if (hammer_debug_general & 0x0200) {
159                         kprintf("check %016llx %016llx cre=%016llx del=%016llx\n",
160                                         elm->base.obj_id,
161                                         elm->base.key,
162                                         elm->base.create_tid,
163                                         elm->base.delete_tid);
164                 }
165                                 
166                 if (prune_should_delete(prune, elm)) {
167                         if (hammer_debug_general & 0x0200) {
168                                 kprintf("check %016llx %016llx: DELETE\n",
169                                         elm->base.obj_id, elm->base.key);
170                         }
171
172                         /*
173                          * NOTE: This can return EDEADLK
174                          *
175                          * Acquiring the sync lock guarantees that the
176                          * operation will not cross a synchronization
177                          * boundary (see the flusher).
178                          */
179                         isdir = (elm->base.rec_type == HAMMER_RECTYPE_DIRENTRY);
180
181                         error = hammer_delete_at_cursor(&cursor,
182                                                         HAMMER_DELETE_DESTROY,
183                                                         &prune->stat_bytes);
184                         if (error)
185                                 break;
186
187                         if (isdir)
188                                 ++prune->stat_dirrecords;
189                         else
190                                 ++prune->stat_rawrecords;
191
192                         /*
193                          * The current record might now be the one after
194                          * the one we deleted, set ATEDISK to force us
195                          * to skip it (since we are iterating backwards).
196                          */
197                         cursor.flags |= HAMMER_CURSOR_ATEDISK;
198                 } else {
199                         /*
200                          * Nothing to delete, but we may have to check other
201                          * things.
202                          */
203                         prune_check_nlinks(&cursor, elm);
204                         cursor.flags |= HAMMER_CURSOR_ATEDISK;
205                         if (hammer_debug_general & 0x0100) {
206                                 kprintf("check %016llx %016llx: SKIP\n",
207                                         elm->base.obj_id, elm->base.key);
208                         }
209                 }
210                 ++prune->stat_scanrecords;
211                 error = hammer_btree_iterate_reverse(&cursor);
212         }
213         hammer_sync_unlock(trans);
214         if (error == ENOENT)
215                 error = 0;
216         hammer_done_cursor(&cursor);
217         if (error == EWOULDBLOCK) {
218                 hammer_flusher_sync(trans->hmp);
219                 goto retry;
220         }
221         if (error == EDEADLK)
222                 goto retry;
223         if (error == EINTR) {
224                 prune->head.flags |= HAMMER_IOC_HEAD_INTR;
225                 error = 0;
226         }
227 failed:
228         prune->key_cur.localization &= HAMMER_LOCALIZE_MASK;
229         prune->elms = user_elms;
230         kfree(copy_elms, M_TEMP);
231         return(error);
232 }
233
234 /*
235  * Check pruning list.  The list must be sorted in descending order.
236  *
237  * Return non-zero if the record should be deleted.
238  */
239 static int
240 prune_should_delete(struct hammer_ioc_prune *prune, hammer_btree_leaf_elm_t elm)
241 {
242         struct hammer_ioc_prune_elm *scan;
243         int i;
244
245         /*
246          * If pruning everything remove all records with a non-zero
247          * delete_tid.
248          */
249         if (prune->head.flags & HAMMER_IOC_PRUNE_ALL) {
250                 if (elm->base.delete_tid != 0)
251                         return(1);
252                 return(0);
253         }
254
255         for (i = 0; i < prune->nelms; ++i) {
256                 scan = &prune->elms[i];
257
258                 /*
259                  * Check for loop termination.
260                  */
261                 if (elm->base.create_tid >= scan->end_tid ||
262                     elm->base.delete_tid > scan->end_tid) {
263                         break;
264                 }
265
266                 /*
267                  * Determine if we can delete the record.
268                  */
269                 if (elm->base.delete_tid &&
270                     elm->base.create_tid >= scan->beg_tid &&
271                     elm->base.delete_tid <= scan->end_tid &&
272                     (elm->base.create_tid - scan->beg_tid) / scan->mod_tid ==
273                     (elm->base.delete_tid - scan->beg_tid) / scan->mod_tid) {
274                         return(1);
275                 }
276         }
277         return(0);
278 }
279
280 /*
281  * Dangling inodes can occur if processes are holding open descriptors on
282  * deleted files as-of when a machine crashes.  When we find one simply
283  * acquire the inode and release it.  The inode handling code will then
284  * do the right thing.
285  */
286 static
287 void
288 prune_check_nlinks(hammer_cursor_t cursor, hammer_btree_leaf_elm_t elm)
289 {
290         hammer_inode_t ip;
291         int error;
292
293         if (elm->base.rec_type != HAMMER_RECTYPE_INODE)
294                 return;
295         if (elm->base.delete_tid != 0)
296                 return;
297         if (hammer_btree_extract(cursor, HAMMER_CURSOR_GET_DATA))
298                 return;
299         if (cursor->data->inode.nlinks)
300                 return;
301         hammer_cursor_downgrade(cursor);
302         hammer_sync_unlock(cursor->trans);
303         ip = hammer_get_inode(cursor->trans, NULL, elm->base.obj_id,
304                       HAMMER_MAX_TID,
305                       elm->base.localization & HAMMER_LOCALIZE_PSEUDOFS_MASK,
306                       0, &error);
307         if (ip) {
308                 kprintf("pruning disconnected inode %016llx\n",
309                         elm->base.obj_id);
310                 hammer_rel_inode(ip, 0);
311         } else {
312                 kprintf("unable to prune disconnected inode %016llx\n",
313                         elm->base.obj_id);
314         }
315         hammer_sync_lock_sh(cursor->trans);
316 }
317
318 #if 0
319
320 /*
321  * NOTE: THIS CODE HAS BEEN REMOVED!  Pruning no longer attempts to realign
322  *       adjacent records because it seriously interferes with every 
323  *       mirroring algorithm I could come up with.
324  *
325  *       This means that historical accesses beyond the first snapshot
326  *       softlink should be on snapshot boundaries only.  Historical
327  *       accesses from "now" to the first snapshot softlink continue to
328  *       be fine-grained.
329  *
330  * NOTE: It also looks like there's a bug in the removed code.  It is believed
331  *       that create_tid can sometimes get set to 0xffffffffffffffff.  Just as
332  *       well we no longer try to do this fancy shit.  Probably the attempt to
333  *       correct the rhb is blowing up the cursor's indexing or addressing mapping.
334  *
335  * Align the record to cover any gaps created through the deletion of
336  * records within the pruning space.  If we were to just delete the records
337  * there would be gaps which in turn would cause a snapshot that is NOT on
338  * a pruning boundary to appear corrupt to the user.  Forcing alignment
339  * of the create_tid and delete_tid for retained records 'reconnects'
340  * the previously contiguous space, making it contiguous again after the
341  * deletions.
342  *
343  * The use of a reverse iteration allows us to safely align the records and
344  * related elements without creating temporary overlaps.  XXX we should
345  * add ordering dependancies for record buffers to guarantee consistency
346  * during recovery.
347  */
348 static int
349 realign_prune(struct hammer_ioc_prune *prune,
350               hammer_cursor_t cursor, int realign_cre, int realign_del)
351 {
352         struct hammer_ioc_prune_elm *scan;
353         hammer_btree_elm_t elm;
354         hammer_tid_t delta;
355         hammer_tid_t tid;
356         int error;
357
358         hammer_cursor_downgrade(cursor);
359
360         elm = &cursor->node->ondisk->elms[cursor->index];
361         ++prune->stat_realignments;
362
363         /*
364          * Align the create_tid.  By doing a reverse iteration we guarantee
365          * that all records after our current record have already been
366          * aligned, allowing us to safely correct the right-hand-boundary
367          * (because no record to our right is otherwise exactly matching
368          * will have a create_tid to the left of our aligned create_tid).
369          */
370         error = 0;
371         if (realign_cre >= 0) {
372                 scan = &prune->elms[realign_cre];
373
374                 delta = (elm->leaf.base.create_tid - scan->beg_tid) % 
375                         scan->mod_tid;
376                 if (delta) {
377                         tid = elm->leaf.base.create_tid - delta + scan->mod_tid;
378
379                         /* can EDEADLK */
380                         error = hammer_btree_correct_rhb(cursor, tid + 1);
381                         if (error == 0) {
382                                 error = hammer_btree_extract(cursor,
383                                                      HAMMER_CURSOR_GET_LEAF);
384                         }
385                         if (error == 0) {
386                                 /* can EDEADLK */
387                                 error = hammer_cursor_upgrade(cursor);
388                         }
389                         if (error == 0) {
390                                 hammer_modify_node(cursor->trans, cursor->node,
391                                             &elm->leaf.base.create_tid,
392                                             sizeof(elm->leaf.base.create_tid));
393                                 elm->leaf.base.create_tid = tid;
394                                 hammer_modify_node_done(cursor->node);
395                         }
396                 }
397         }
398
399         /*
400          * Align the delete_tid.  This only occurs if the record is historical
401          * was deleted at some point.  Realigning the delete_tid does not
402          * move the record within the B-Tree but may cause it to temporarily
403          * overlap a record that has not yet been pruned.
404          */
405         if (error == 0 && realign_del >= 0) {
406                 scan = &prune->elms[realign_del];
407
408                 delta = (elm->leaf.base.delete_tid - scan->beg_tid) % 
409                         scan->mod_tid;
410                 if (delta) {
411                         error = hammer_btree_extract(cursor,
412                                                      HAMMER_CURSOR_GET_LEAF);
413                         if (error == 0) {
414                                 hammer_modify_node(cursor->trans, cursor->node,
415                                             &elm->leaf.base.delete_tid,
416                                             sizeof(elm->leaf.base.delete_tid));
417                                 elm->leaf.base.delete_tid =
418                                             elm->leaf.base.delete_tid -
419                                             delta + scan->mod_tid;
420                                 hammer_modify_node_done(cursor->node);
421                         }
422                 }
423         }
424         return (error);
425 }
426
427 #endif