Merge branch 'vendor/GCC50'
[dragonfly.git] / sys / vfs / hammer / hammer_reblock.c
1 /*
2  * Copyright (c) 2008-2012 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 /*
35  * HAMMER reblocker - This code frees up fragmented physical space
36  *
37  * HAMMER only keeps track of free space on a big-block basis.  A big-block
38  * containing holes can only be freed by migrating the remaining data in
39  * that big-block into a new big-block, then freeing the big-block.
40  *
41  * This function is called from an ioctl or via the hammer support thread.
42  */
43
44 #include "hammer.h"
45
46 static int hammer_reblock_helper(struct hammer_ioc_reblock *reblock,
47                                  hammer_cursor_t cursor,
48                                  hammer_btree_elm_t elm);
49 static int hammer_reblock_data(struct hammer_ioc_reblock *reblock,
50                                 hammer_cursor_t cursor, hammer_btree_elm_t elm);
51 static int hammer_reblock_leaf_node(struct hammer_ioc_reblock *reblock,
52                                 hammer_cursor_t cursor, hammer_btree_elm_t elm);
53 static int hammer_reblock_int_node(struct hammer_ioc_reblock *reblock,
54                                 hammer_cursor_t cursor, hammer_btree_elm_t elm);
55
56 int
57 hammer_ioc_reblock(hammer_transaction_t trans, hammer_inode_t ip,
58                    struct hammer_ioc_reblock *reblock)
59 {
60         struct hammer_cursor cursor;
61         hammer_btree_elm_t elm;
62         int checkspace_count;
63         int error;
64         int seq;
65         int slop;
66         u_int32_t key_end_localization;
67
68         if ((reblock->key_beg.localization | reblock->key_end.localization) &
69             HAMMER_LOCALIZE_PSEUDOFS_MASK) {
70                 return(EINVAL);
71         }
72         if (reblock->key_beg.obj_id >= reblock->key_end.obj_id)
73                 return(EINVAL);
74         if (reblock->free_level < 0 ||
75             reblock->free_level > HAMMER_BIGBLOCK_SIZE)
76                 return(EINVAL);
77
78         /*
79          * A fill_percentage <= 20% is considered an emergency.  free_level is
80          * inverted from fill_percentage.
81          */
82         if (reblock->free_level >= HAMMER_BIGBLOCK_SIZE * 8 / 10)
83                 slop = HAMMER_CHKSPC_EMERGENCY;
84         else
85                 slop = HAMMER_CHKSPC_REBLOCK;
86
87         /*
88          * Ioctl caller has only set localization type to reblock.
89          * Initialize cursor key localization with ip localization.
90          */
91         reblock->key_cur = reblock->key_beg;
92         reblock->key_cur.localization &= HAMMER_LOCALIZE_MASK;
93         reblock->key_cur.localization += ip->obj_localization;
94
95         key_end_localization = reblock->key_end.localization;
96         key_end_localization &= HAMMER_LOCALIZE_MASK;
97         key_end_localization += ip->obj_localization;
98
99         checkspace_count = 0;
100         seq = trans->hmp->flusher.done;
101 retry:
102         error = hammer_init_cursor(trans, &cursor, NULL, NULL);
103         if (error) {
104                 hammer_done_cursor(&cursor);
105                 goto failed;
106         }
107         cursor.key_beg.localization = reblock->key_cur.localization;
108         cursor.key_beg.obj_id = reblock->key_cur.obj_id;
109         cursor.key_beg.key = HAMMER_MIN_KEY;
110         cursor.key_beg.create_tid = 1;
111         cursor.key_beg.delete_tid = 0;
112         cursor.key_beg.rec_type = HAMMER_MIN_RECTYPE;
113         cursor.key_beg.obj_type = 0;
114
115         cursor.key_end.localization = key_end_localization;
116         cursor.key_end.obj_id = reblock->key_end.obj_id;
117         cursor.key_end.key = HAMMER_MAX_KEY;
118         cursor.key_end.create_tid = HAMMER_MAX_TID - 1;
119         cursor.key_end.delete_tid = 0;
120         cursor.key_end.rec_type = HAMMER_MAX_RECTYPE;
121         cursor.key_end.obj_type = 0;
122
123         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE;
124         cursor.flags |= HAMMER_CURSOR_BACKEND;
125         cursor.flags |= HAMMER_CURSOR_NOSWAPCACHE;
126
127         /*
128          * This flag allows the btree scan code to return internal nodes,
129          * so we can reblock them in addition to the leafs.  Only specify it
130          * if we intend to reblock B-Tree nodes.
131          */
132         if (reblock->head.flags & HAMMER_IOC_DO_BTREE)
133                 cursor.flags |= HAMMER_CURSOR_REBLOCKING;
134
135         error = hammer_btree_first(&cursor);
136         while (error == 0) {
137                 /*
138                  * Internal or Leaf node
139                  */
140                 KKASSERT(cursor.index < cursor.node->ondisk->count);
141                 elm = &cursor.node->ondisk->elms[cursor.index];
142                 reblock->key_cur.obj_id = elm->base.obj_id;
143                 reblock->key_cur.localization = elm->base.localization;
144
145                 /*
146                  * Yield to more important tasks
147                  */
148                 if ((error = hammer_signal_check(trans->hmp)) != 0)
149                         break;
150
151                 /*
152                  * If there is insufficient free space it may be due to
153                  * reserved bigblocks, which flushing might fix.
154                  *
155                  * We must force a retest in case the unlocked cursor is
156                  * moved to the end of the leaf, or moved to an internal
157                  * node.
158                  *
159                  * WARNING: See warnings in hammer_unlock_cursor() function.
160                  */
161                 if (hammer_checkspace(trans->hmp, slop)) {
162                         if (++checkspace_count == 10) {
163                                 error = ENOSPC;
164                                 break;
165                         }
166                         hammer_unlock_cursor(&cursor);
167                         cursor.flags |= HAMMER_CURSOR_RETEST;
168                         hammer_flusher_wait(trans->hmp, seq);
169                         hammer_lock_cursor(&cursor);
170                         seq = hammer_flusher_async(trans->hmp, NULL);
171                         goto skip;
172                 }
173
174                 /*
175                  * Acquiring the sync_lock prevents the operation from
176                  * crossing a synchronization boundary.
177                  *
178                  * NOTE: cursor.node may have changed on return.
179                  *
180                  * WARNING: See warnings in hammer_unlock_cursor() function.
181                  */
182                 hammer_sync_lock_sh(trans);
183                 error = hammer_reblock_helper(reblock, &cursor, elm);
184                 hammer_sync_unlock(trans);
185
186                 while (hammer_flusher_meta_halflimit(trans->hmp) ||
187                        hammer_flusher_undo_exhausted(trans, 2)) {
188                         hammer_unlock_cursor(&cursor);
189                         hammer_flusher_wait(trans->hmp, seq);
190                         hammer_lock_cursor(&cursor);
191                         seq = hammer_flusher_async_one(trans->hmp);
192                 }
193
194                 /*
195                  * Setup for iteration, our cursor flags may be modified by
196                  * other threads while we are unlocked.
197                  */
198                 cursor.flags |= HAMMER_CURSOR_ATEDISK;
199
200                 /*
201                  * We allocate data buffers, which atm we don't track
202                  * dirty levels for because we allow the kernel to write
203                  * them.  But if we allocate too many we can still deadlock
204                  * the buffer cache.
205                  *
206                  * WARNING: See warnings in hammer_unlock_cursor() function.
207                  *          (The cursor's node and element may change!)
208                  */
209                 if (bd_heatup()) {
210                         hammer_unlock_cursor(&cursor);
211                         bwillwrite(HAMMER_XBUFSIZE);
212                         hammer_lock_cursor(&cursor);
213                 }
214                 vm_wait_nominal();
215 skip:
216                 if (error == 0) {
217                         error = hammer_btree_iterate(&cursor);
218                 }
219         }
220         if (error == ENOENT)
221                 error = 0;
222         hammer_done_cursor(&cursor);
223         if (error == EWOULDBLOCK) {
224                 hammer_flusher_sync(trans->hmp);
225                 goto retry;
226         }
227         if (error == EDEADLK)
228                 goto retry;
229         if (error == EINTR) {
230                 reblock->head.flags |= HAMMER_IOC_HEAD_INTR;
231                 error = 0;
232         }
233 failed:
234         reblock->key_cur.localization &= HAMMER_LOCALIZE_MASK;
235         return(error);
236 }
237
238 /*
239  * Reblock the B-Tree (leaf) node, record, and/or data if necessary.
240  *
241  * XXX We have no visibility into internal B-Tree nodes at the moment,
242  * only leaf nodes.
243  */
244 static int
245 hammer_reblock_helper(struct hammer_ioc_reblock *reblock,
246                       hammer_cursor_t cursor, hammer_btree_elm_t elm)
247 {
248         hammer_mount_t hmp;
249         hammer_off_t tmp_offset;
250         hammer_node_ondisk_t ondisk;
251         struct hammer_btree_leaf_elm leaf;
252         int error;
253         int bytes;
254         int cur;
255         int iocflags;
256
257         error = 0;
258         hmp = cursor->trans->hmp;
259
260         /*
261          * Reblock data.  Note that data embedded in a record is reblocked
262          * by the record reblock code.  Data processing only occurs at leaf
263          * nodes and for RECORD element types.
264          */
265         if (cursor->node->ondisk->type != HAMMER_BTREE_TYPE_LEAF)
266                 goto skip;
267         if (elm->leaf.base.btype != HAMMER_BTREE_TYPE_RECORD)
268                 return(0);
269         tmp_offset = elm->leaf.data_offset;
270         if (tmp_offset == 0)
271                 goto skip;
272         if (error)
273                 goto skip;
274
275         /*
276          * NOTE: Localization restrictions may also have been set-up, we can't
277          *       just set the match flags willy-nilly here.
278          */
279         switch(elm->leaf.base.rec_type) {
280         case HAMMER_RECTYPE_INODE:
281         case HAMMER_RECTYPE_SNAPSHOT:
282         case HAMMER_RECTYPE_CONFIG:
283                 iocflags = HAMMER_IOC_DO_INODES;
284                 break;
285         case HAMMER_RECTYPE_EXT:
286         case HAMMER_RECTYPE_FIX:
287         case HAMMER_RECTYPE_PFS:
288         case HAMMER_RECTYPE_DIRENTRY:
289                 iocflags = HAMMER_IOC_DO_DIRS;
290                 break;
291         case HAMMER_RECTYPE_DATA:
292         case HAMMER_RECTYPE_DB:
293                 iocflags = HAMMER_IOC_DO_DATA;
294                 break;
295         default:
296                 iocflags = 0;
297                 break;
298         }
299         if (reblock->head.flags & iocflags) {
300                 ++reblock->data_count;
301                 reblock->data_byte_count += elm->leaf.data_len;
302                 bytes = hammer_blockmap_getfree(hmp, tmp_offset, &cur, &error);
303                 if (hammer_debug_general & 0x4000)
304                         kprintf("D %6d/%d\n", bytes, reblock->free_level);
305                 if (error == 0 && (cur == 0 || reblock->free_level == 0) &&
306                     bytes >= reblock->free_level) {
307                         /*
308                          * This is nasty, the uncache code may have to get
309                          * vnode locks and because of that we can't hold
310                          * the cursor locked.
311                          *
312                          * WARNING: See warnings in hammer_unlock_cursor()
313                          *          function.
314                          */
315                         leaf = elm->leaf;
316                         hammer_unlock_cursor(cursor);
317                         hammer_io_direct_uncache(hmp, &leaf);
318                         hammer_lock_cursor(cursor);
319
320                         /*
321                          * elm may have become stale or invalid, reload it.
322                          * ondisk variable is temporary only.  Note that
323                          * cursor->node and thus cursor->node->ondisk may
324                          * also changed.
325                          */
326                         ondisk = cursor->node->ondisk;
327                         elm = &ondisk->elms[cursor->index];
328                         if (cursor->flags & HAMMER_CURSOR_RETEST) {
329                                 kprintf("hammer: debug: retest on "
330                                         "reblocker uncache\n");
331                                 error = EDEADLK;
332                         } else if (ondisk->type != HAMMER_BTREE_TYPE_LEAF ||
333                                    cursor->index >= ondisk->count) {
334                                 kprintf("hammer: debug: shifted on "
335                                         "reblocker uncache\n");
336                                 error = EDEADLK;
337                         } else if (bcmp(&elm->leaf, &leaf, sizeof(leaf))) {
338                                 kprintf("hammer: debug: changed on "
339                                         "reblocker uncache\n");
340                                 error = EDEADLK;
341                         }
342                         if (error == 0)
343                                 error = hammer_cursor_upgrade(cursor);
344                         if (error == 0) {
345                                 KKASSERT(cursor->index < ondisk->count);
346                                 error = hammer_reblock_data(reblock,
347                                                             cursor, elm);
348                         }
349                         if (error == 0) {
350                                 ++reblock->data_moves;
351                                 reblock->data_byte_moves += elm->leaf.data_len;
352                         }
353                 }
354         }
355
356 skip:
357         /*
358          * Reblock a B-Tree internal or leaf node.  A leaf node is reblocked
359          * on initial entry only (element 0).  An internal node is reblocked
360          * when entered upward from its first leaf node only (also element 0).
361          * Further revisits of the internal node (index > 0) are ignored.
362          */
363         tmp_offset = cursor->node->node_offset;
364         if (cursor->index == 0 &&
365             error == 0 && (reblock->head.flags & HAMMER_IOC_DO_BTREE)) {
366                 ++reblock->btree_count;
367                 bytes = hammer_blockmap_getfree(hmp, tmp_offset, &cur, &error);
368                 if (hammer_debug_general & 0x4000)
369                         kprintf("B %6d/%d\n", bytes, reblock->free_level);
370                 if (error == 0 && (cur == 0 || reblock->free_level == 0) &&
371                     bytes >= reblock->free_level) {
372                         error = hammer_cursor_upgrade(cursor);
373                         if (error == 0) {
374                                 if (cursor->parent) {
375                                         KKASSERT(cursor->parent_index <
376                                                  cursor->parent->ondisk->count);
377                                         elm = &cursor->parent->ondisk->elms[cursor->parent_index];
378                                 } else {
379                                         elm = NULL;
380                                 }
381                                 switch(cursor->node->ondisk->type) {
382                                 case HAMMER_BTREE_TYPE_LEAF:
383                                         error = hammer_reblock_leaf_node(
384                                                         reblock, cursor, elm);
385                                         break;
386                                 case HAMMER_BTREE_TYPE_INTERNAL:
387                                         error = hammer_reblock_int_node(
388                                                         reblock, cursor, elm);
389                                         break;
390                                 default:
391                                         panic("Illegal B-Tree node type");
392                                 }
393                         }
394                         if (error == 0) {
395                                 ++reblock->btree_moves;
396                         }
397                 }
398         }
399
400         hammer_cursor_downgrade(cursor);
401         return(error);
402 }
403
404 /*
405  * Reblock a record's data.  Both the B-Tree element and record pointers
406  * to the data must be adjusted.
407  */
408 static int
409 hammer_reblock_data(struct hammer_ioc_reblock *reblock,
410                     hammer_cursor_t cursor, hammer_btree_elm_t elm)
411 {
412         struct hammer_buffer *data_buffer = NULL;
413         hammer_off_t ndata_offset;
414         int error;
415         void *ndata;
416
417         error = hammer_btree_extract(cursor, HAMMER_CURSOR_GET_DATA |
418                                              HAMMER_CURSOR_GET_LEAF);
419         if (error)
420                 return (error);
421         ndata = hammer_alloc_data(cursor->trans, elm->leaf.data_len,
422                                   elm->leaf.base.rec_type,
423                                   &ndata_offset, &data_buffer,
424                                   0, &error);
425         if (error)
426                 goto done;
427         hammer_io_notmeta(data_buffer);
428
429         /*
430          * Move the data.  Note that we must invalidate any cached
431          * data buffer in the cursor before calling blockmap_free.
432          * The blockmap_free may free up the entire big-block and
433          * will not be able to invalidate it if the cursor is holding
434          * a data buffer cached in that big block.
435          */
436         hammer_modify_buffer(cursor->trans, data_buffer, NULL, 0);
437         bcopy(cursor->data, ndata, elm->leaf.data_len);
438         hammer_modify_buffer_done(data_buffer);
439         hammer_cursor_invalidate_cache(cursor);
440
441         hammer_blockmap_free(cursor->trans,
442                              elm->leaf.data_offset, elm->leaf.data_len);
443
444         hammer_modify_node(cursor->trans, cursor->node,
445                            &elm->leaf.data_offset, sizeof(hammer_off_t));
446         elm->leaf.data_offset = ndata_offset;
447         hammer_modify_node_done(cursor->node);
448
449 done:
450         if (data_buffer)
451                 hammer_rel_buffer(data_buffer, 0);
452         return (error);
453 }
454
455 /*
456  * Reblock a B-Tree leaf node.  The parent must be adjusted to point to
457  * the new copy of the leaf node.
458  *
459  * elm is a pointer to the parent element pointing at cursor.node.
460  */
461 static int
462 hammer_reblock_leaf_node(struct hammer_ioc_reblock *reblock,
463                          hammer_cursor_t cursor, hammer_btree_elm_t elm)
464 {
465         hammer_node_t onode;
466         hammer_node_t nnode;
467         int error;
468
469         /*
470          * Don't supply a hint when allocating the leaf.  Fills are done
471          * from the leaf upwards.
472          */
473         onode = cursor->node;
474         nnode = hammer_alloc_btree(cursor->trans, 0, &error);
475
476         if (nnode == NULL)
477                 return (error);
478
479         /*
480          * Move the node
481          */
482         hammer_lock_ex(&nnode->lock);
483         hammer_modify_node_noundo(cursor->trans, nnode);
484         bcopy(onode->ondisk, nnode->ondisk, sizeof(*nnode->ondisk));
485
486         if (elm) {
487                 /*
488                  * We are not the root of the B-Tree 
489                  */
490                 hammer_modify_node(cursor->trans, cursor->parent,
491                                    &elm->internal.subtree_offset,
492                                    sizeof(elm->internal.subtree_offset));
493                 elm->internal.subtree_offset = nnode->node_offset;
494                 hammer_modify_node_done(cursor->parent);
495         } else {
496                 /*
497                  * We are the root of the B-Tree
498                  */
499                 hammer_volume_t volume;
500                         
501                 volume = hammer_get_root_volume(cursor->trans->hmp, &error);
502                 KKASSERT(error == 0);
503
504                 hammer_modify_volume_field(cursor->trans, volume,
505                                            vol0_btree_root);
506                 volume->ondisk->vol0_btree_root = nnode->node_offset;
507                 hammer_modify_volume_done(volume);
508                 hammer_rel_volume(volume, 0);
509         }
510
511         hammer_cursor_replaced_node(onode, nnode);
512         hammer_delete_node(cursor->trans, onode);
513
514         if (hammer_debug_general & 0x4000) {
515                 kprintf("REBLOCK LNODE %016llx -> %016llx\n",
516                         (long long)onode->node_offset,
517                         (long long)nnode->node_offset);
518         }
519         hammer_modify_node_done(nnode);
520         cursor->node = nnode;
521
522         hammer_unlock(&onode->lock);
523         hammer_rel_node(onode);
524
525         return (error);
526 }
527
528 /*
529  * Reblock a B-Tree internal node.  The parent must be adjusted to point to
530  * the new copy of the internal node, and the node's children's parent
531  * pointers must also be adjusted to point to the new copy.
532  *
533  * elm is a pointer to the parent element pointing at cursor.node.
534  */
535 static int
536 hammer_reblock_int_node(struct hammer_ioc_reblock *reblock,
537                          hammer_cursor_t cursor, hammer_btree_elm_t elm)
538 {
539         struct hammer_node_lock lockroot;
540         hammer_node_t onode;
541         hammer_node_t nnode;
542         int error;
543         int i;
544
545         hammer_node_lock_init(&lockroot, cursor->node);
546         error = hammer_btree_lock_children(cursor, 1, &lockroot, NULL);
547         if (error)
548                 goto done;
549
550         onode = cursor->node;
551         nnode = hammer_alloc_btree(cursor->trans, 0, &error);
552
553         if (nnode == NULL)
554                 goto done;
555
556         /*
557          * Move the node.  Adjust the parent's pointer to us first.
558          */
559         hammer_lock_ex(&nnode->lock);
560         hammer_modify_node_noundo(cursor->trans, nnode);
561         bcopy(onode->ondisk, nnode->ondisk, sizeof(*nnode->ondisk));
562
563         if (elm) {
564                 /*
565                  * We are not the root of the B-Tree 
566                  */
567                 hammer_modify_node(cursor->trans, cursor->parent,
568                                    &elm->internal.subtree_offset,
569                                    sizeof(elm->internal.subtree_offset));
570                 elm->internal.subtree_offset = nnode->node_offset;
571                 hammer_modify_node_done(cursor->parent);
572         } else {
573                 /*
574                  * We are the root of the B-Tree
575                  */
576                 hammer_volume_t volume;
577                         
578                 volume = hammer_get_root_volume(cursor->trans->hmp, &error);
579                 KKASSERT(error == 0);
580
581                 hammer_modify_volume_field(cursor->trans, volume,
582                                            vol0_btree_root);
583                 volume->ondisk->vol0_btree_root = nnode->node_offset;
584                 hammer_modify_volume_done(volume);
585                 hammer_rel_volume(volume, 0);
586         }
587
588         /*
589          * Now adjust our children's pointers to us.
590          */
591         for (i = 0; i < nnode->ondisk->count; ++i) {
592                 elm = &nnode->ondisk->elms[i];
593                 error = btree_set_parent(cursor->trans, nnode, elm);
594                 if (error)
595                         panic("reblock internal node: fixup problem");
596         }
597
598         /*
599          * Clean up.
600          *
601          * The new node replaces the current node in the cursor.  The cursor
602          * expects it to be locked so leave it locked.  Discard onode.
603          */
604         hammer_cursor_replaced_node(onode, nnode);
605         hammer_delete_node(cursor->trans, onode);
606
607         if (hammer_debug_general & 0x4000) {
608                 kprintf("REBLOCK INODE %016llx -> %016llx\n",
609                         (long long)onode->node_offset,
610                         (long long)nnode->node_offset);
611         }
612         hammer_modify_node_done(nnode);
613         cursor->node = nnode;
614
615         hammer_unlock(&onode->lock);
616         hammer_rel_node(onode);
617
618 done:
619         hammer_btree_unlock_children(cursor->trans->hmp, &lockroot, NULL);
620         return (error);
621 }
622