HAMMER 49/Many: Enhance pruning code
[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 1.3 2008/05/31 18:37:57 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 check_prune(struct hammer_ioc_prune *prune, hammer_btree_elm_t elm,
48                         int *realign_cre, int *realign_del);
49 static int realign_prune(struct hammer_ioc_prune *prune, hammer_cursor_t cursor,
50                         int realign_cre, int realign_del);
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_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 realign_cre;
63         int realign_del;
64         int elm_array_size;
65
66         if (prune->nelms < 0 || prune->nelms > HAMMER_MAX_PRUNE_ELMS)
67                 return(EINVAL);
68         if (prune->beg_obj_id >= prune->end_obj_id)
69                 return(EINVAL);
70         if ((prune->head.flags & HAMMER_IOC_PRUNE_ALL) && prune->nelms)
71                 return(EINVAL);
72
73         prune->cur_localization = prune->end_localization;
74         prune->cur_obj_id = prune->end_obj_id;
75         prune->cur_key = HAMMER_MAX_KEY;
76
77         /*
78          * Copy element array from userland
79          */
80         elm_array_size = sizeof(*copy_elms) * prune->nelms;
81         user_elms = prune->elms;
82         copy_elms = kmalloc(elm_array_size, M_TEMP, M_WAITOK);
83         if ((error = copyin(user_elms, copy_elms, elm_array_size)) != 0)
84                 goto failed;
85         prune->elms = copy_elms;
86
87         /*
88          * Scan backwards.  Retries typically occur if a deadlock is detected.
89          */
90 retry:
91         error = hammer_init_cursor(trans, &cursor, NULL, NULL);
92         if (error) {
93                 hammer_done_cursor(&cursor);
94                 goto failed;
95         }
96         cursor.key_beg.localization = prune->beg_localization;
97         cursor.key_beg.obj_id = prune->beg_obj_id;
98         cursor.key_beg.key = HAMMER_MIN_KEY;
99         cursor.key_beg.create_tid = 1;
100         cursor.key_beg.delete_tid = 0;
101         cursor.key_beg.rec_type = HAMMER_MIN_RECTYPE;
102         cursor.key_beg.obj_type = 0;
103
104         cursor.key_end.localization = prune->cur_localization;
105         cursor.key_end.obj_id = prune->cur_obj_id;
106         cursor.key_end.key = prune->cur_key;
107         cursor.key_end.create_tid = HAMMER_MAX_TID - 1;
108         cursor.key_end.delete_tid = 0;
109         cursor.key_end.rec_type = HAMMER_MAX_RECTYPE;
110         cursor.key_end.obj_type = 0;
111
112         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE;
113         cursor.flags |= HAMMER_CURSOR_BACKEND;
114
115         /*
116          * This flag allows the B-Tree code to clean up loose ends.
117          */
118         cursor.flags |= HAMMER_CURSOR_PRUNING;
119
120         hammer_sync_lock_sh(trans);
121         error = hammer_btree_last(&cursor);
122         while (error == 0) {
123                 /*
124                  * Yield to more important tasks
125                  */
126                 if (trans->hmp->sync_lock.wanted) {
127                         hammer_sync_unlock(trans);
128                         tsleep(trans, 0, "hmrslo", hz / 10);
129                         hammer_sync_lock_sh(trans);
130                 }
131
132                 /*
133                  * Check for work
134                  */
135                 elm = &cursor.node->ondisk->elms[cursor.index];
136                 prune->cur_localization = elm->base.localization;
137                 prune->cur_obj_id = elm->base.obj_id;
138                 prune->cur_key = elm->base.key;
139
140                 if (prune->stat_oldest_tid > elm->leaf.base.create_tid)
141                         prune->stat_oldest_tid = elm->leaf.base.create_tid;
142
143                 if (check_prune(prune, elm, &realign_cre, &realign_del) == 0) {
144                         if (hammer_debug_general & 0x0200) {
145                                 kprintf("check %016llx %016llx: DELETE\n",
146                                         elm->base.obj_id, elm->base.key);
147                         }
148
149                         /*
150                          * NOTE: This can return EDEADLK
151                          *
152                          * Acquiring the sync lock guarantees that the
153                          * operation will not cross a synchronization
154                          * boundary (see the flusher).
155                          */
156                         isdir = (elm->base.rec_type == HAMMER_RECTYPE_DIRENTRY);
157
158                         error = hammer_delete_at_cursor(&cursor,
159                                                         &prune->stat_bytes);
160                         if (error)
161                                 break;
162
163                         if (isdir)
164                                 ++prune->stat_dirrecords;
165                         else
166                                 ++prune->stat_rawrecords;
167
168                         /*
169                          * The current record might now be the one after
170                          * the one we deleted, set ATEDISK to force us
171                          * to skip it (since we are iterating backwards).
172                          */
173                         cursor.flags |= HAMMER_CURSOR_ATEDISK;
174                 } else if (realign_cre >= 0 || realign_del >= 0) {
175                         error = realign_prune(prune, &cursor,
176                                               realign_cre, realign_del);
177                         if (error == 0) {
178                                 cursor.flags |= HAMMER_CURSOR_ATEDISK;
179                                 if (hammer_debug_general & 0x0200) {
180                                         kprintf("check %016llx %016llx: "
181                                                 "REALIGN\n",
182                                                 elm->base.obj_id,
183                                                 elm->base.key);
184                                 }
185                         }
186                 } else {
187                         cursor.flags |= HAMMER_CURSOR_ATEDISK;
188                         if (hammer_debug_general & 0x0100) {
189                                 kprintf("check %016llx %016llx: SKIP\n",
190                                         elm->base.obj_id, elm->base.key);
191                         }
192                 }
193                 ++prune->stat_scanrecords;
194
195                 /*
196                  * Bad hack for now, don't blow out the kernel's buffer
197                  * cache.  NOTE: We still hold locks on the cursor, we
198                  * cannot call the flusher synchronously.
199                  */
200                 if (trans->hmp->locked_dirty_count +
201                     trans->hmp->io_running_count > hammer_limit_dirtybufs) {
202                         hammer_flusher_async(trans->hmp);
203                         tsleep(trans, 0, "hmrslo", hz / 10);
204                 }
205                 error = hammer_signal_check(trans->hmp);
206                 if (error == 0)
207                         error = hammer_btree_iterate_reverse(&cursor);
208         }
209         hammer_sync_unlock(trans);
210         if (error == ENOENT)
211                 error = 0;
212         hammer_done_cursor(&cursor);
213         if (error == EDEADLK)
214                 goto retry;
215         if (error == EINTR) {
216                 prune->head.flags |= HAMMER_IOC_HEAD_INTR;
217                 error = 0;
218         }
219 failed:
220         prune->elms = user_elms;
221         kfree(copy_elms, M_TEMP);
222         return(error);
223 }
224
225 /*
226  * Check pruning list.  The list must be sorted in descending order.
227  */
228 static int
229 check_prune(struct hammer_ioc_prune *prune, hammer_btree_elm_t elm,
230             int *realign_cre, int *realign_del)
231 {
232         struct hammer_ioc_prune_elm *scan;
233         int i;
234
235         *realign_cre = -1;
236         *realign_del = -1;
237
238         /*
239          * If pruning everything remove all records with a non-zero
240          * delete_tid.
241          */
242         if (prune->head.flags & HAMMER_IOC_PRUNE_ALL) {
243                 if (elm->base.delete_tid != 0)
244                         return(0);
245                 return(-1);
246         }
247
248         for (i = 0; i < prune->nelms; ++i) {
249                 scan = &prune->elms[i];
250
251                 /*
252                  * Locate the scan index covering the create and delete TIDs.
253                  */
254                 if (*realign_cre < 0 &&
255                     elm->base.create_tid >= scan->beg_tid &&
256                     elm->base.create_tid < scan->end_tid) {
257                         *realign_cre = i;
258                 }
259                 if (*realign_del < 0 && elm->base.delete_tid &&
260                     elm->base.delete_tid > scan->beg_tid &&
261                     elm->base.delete_tid <= scan->end_tid) {
262                         *realign_del = i;
263                 }
264
265                 /*
266                  * Now check for loop termination.
267                  */
268                 if (elm->base.create_tid >= scan->end_tid ||
269                     elm->base.delete_tid > scan->end_tid) {
270                         break;
271                 }
272
273                 /*
274                  * Now determine if we can delete the record.
275                  */
276                 if (elm->base.delete_tid &&
277                     elm->base.create_tid >= scan->beg_tid &&
278                     elm->base.delete_tid <= scan->end_tid &&
279                     (elm->base.create_tid - scan->beg_tid) / scan->mod_tid ==
280                     (elm->base.delete_tid - scan->beg_tid) / scan->mod_tid) {
281                         return(0);
282                 }
283         }
284         return(-1);
285 }
286
287 /*
288  * Align the record to cover any gaps created through the deletion of
289  * records within the pruning space.  If we were to just delete the records
290  * there would be gaps which in turn would cause a snapshot that is NOT on
291  * a pruning boundary to appear corrupt to the user.  Forcing alignment
292  * of the create_tid and delete_tid for retained records 'reconnects'
293  * the previously contiguous space, making it contiguous again after the
294  * deletions.
295  *
296  * The use of a reverse iteration allows us to safely align the records and
297  * related elements without creating temporary overlaps.  XXX we should
298  * add ordering dependancies for record buffers to guarantee consistency
299  * during recovery.
300  */
301 static int
302 realign_prune(struct hammer_ioc_prune *prune,
303               hammer_cursor_t cursor, int realign_cre, int realign_del)
304 {
305         struct hammer_ioc_prune_elm *scan;
306         hammer_btree_elm_t elm;
307         hammer_tid_t delta;
308         hammer_tid_t tid;
309         int error;
310
311         hammer_cursor_downgrade(cursor);
312
313         elm = &cursor->node->ondisk->elms[cursor->index];
314         ++prune->stat_realignments;
315
316         /*
317          * Align the create_tid.  By doing a reverse iteration we guarantee
318          * that all records after our current record have already been
319          * aligned, allowing us to safely correct the right-hand-boundary
320          * (because no record to our right is otherwise exactly matching
321          * will have a create_tid to the left of our aligned create_tid).
322          */
323         error = 0;
324         if (realign_cre >= 0) {
325                 scan = &prune->elms[realign_cre];
326
327                 delta = (elm->leaf.base.create_tid - scan->beg_tid) % 
328                         scan->mod_tid;
329                 if (delta) {
330                         tid = elm->leaf.base.create_tid - delta + scan->mod_tid;
331
332                         /* can EDEADLK */
333                         error = hammer_btree_correct_rhb(cursor, tid + 1);
334                         if (error == 0) {
335                                 error = hammer_btree_extract(cursor,
336                                                      HAMMER_CURSOR_GET_LEAF);
337                         }
338                         if (error == 0) {
339                                 /* can EDEADLK */
340                                 error = hammer_cursor_upgrade(cursor);
341                         }
342                         if (error == 0) {
343                                 hammer_modify_node(cursor->trans, cursor->node,
344                                             &elm->leaf.base.create_tid,
345                                             sizeof(elm->leaf.base.create_tid));
346                                 elm->leaf.base.create_tid = tid;
347                                 hammer_modify_node_done(cursor->node);
348                         }
349                 }
350         }
351
352         /*
353          * Align the delete_tid.  This only occurs if the record is historical
354          * was deleted at some point.  Realigning the delete_tid does not
355          * move the record within the B-Tree but may cause it to temporarily
356          * overlap a record that has not yet been pruned.
357          */
358         if (error == 0 && realign_del >= 0) {
359                 scan = &prune->elms[realign_del];
360
361                 delta = (elm->leaf.base.delete_tid - scan->beg_tid) % 
362                         scan->mod_tid;
363                 if (delta) {
364                         error = hammer_btree_extract(cursor,
365                                                      HAMMER_CURSOR_GET_LEAF);
366                         if (error == 0) {
367                                 hammer_modify_node(cursor->trans, cursor->node,
368                                             &elm->leaf.base.delete_tid,
369                                             sizeof(elm->leaf.base.delete_tid));
370                                 elm->leaf.base.delete_tid =
371                                             elm->leaf.base.delete_tid -
372                                             delta + scan->mod_tid;
373                                 hammer_modify_node_done(cursor->node);
374                         }
375                 }
376         }
377         return (error);
378 }
379