1 /* -*- mode: c; c-basic-offset: 8; -*-
2 * vim: noexpandtab sw=8 ts=8 sts=0:
3 *
4 * localalloc.c
5 *
6 * Node local data allocation
7 *
8 * Copyright (C) 2002, 2004 Oracle. All rights reserved.
9 *
10 * This program is free software; you can redistribute it and/or
11 * modify it under the terms of the GNU General Public
12 * License as published by the Free Software Foundation; either
13 * version 2 of the License, or (at your option) any later version.
14 *
15 * This program is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
18 * General Public License for more details.
19 *
20 * You should have received a copy of the GNU General Public
21 * License along with this program; if not, write to the
22 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
23 * Boston, MA 021110-1307, USA.
24 */
25
26 #include <linux/fs.h>
27 #include <linux/types.h>
28 #include <linux/slab.h>
29 #include <linux/highmem.h>
30 #include <linux/bitops.h>
31 #include <linux/debugfs.h>
32
33 #define MLOG_MASK_PREFIX ML_DISK_ALLOC
34 #include <cluster/masklog.h>
35
36 #include "ocfs2.h"
37
38 #include "alloc.h"
39 #include "blockcheck.h"
40 #include "dlmglue.h"
41 #include "inode.h"
42 #include "journal.h"
43 #include "localalloc.h"
44 #include "suballoc.h"
45 #include "super.h"
46 #include "sysfile.h"
47
48 #include "buffer_head_io.h"
49
50 #define OCFS2_LOCAL_ALLOC(dinode) (&((dinode)->id2.i_lab))
51
52 static u32 ocfs2_local_alloc_count_bits(struct ocfs2_dinode *alloc);
53
54 static int ocfs2_local_alloc_find_clear_bits(struct ocfs2_super *osb,
55 struct ocfs2_dinode *alloc,
56 u32 numbits);
57
58 static void ocfs2_clear_local_alloc(struct ocfs2_dinode *alloc);
59
60 static int ocfs2_sync_local_to_main(struct ocfs2_super *osb,
61 handle_t *handle,
62 struct ocfs2_dinode *alloc,
63 struct inode *main_bm_inode,
64 struct buffer_head *main_bm_bh);
65
66 static int ocfs2_local_alloc_reserve_for_window(struct ocfs2_super *osb,
67 struct ocfs2_alloc_context **ac,
68 struct inode **bitmap_inode,
69 struct buffer_head **bitmap_bh);
70
71 static int ocfs2_local_alloc_new_window(struct ocfs2_super *osb,
72 handle_t *handle,
73 struct ocfs2_alloc_context *ac);
74
75 static int ocfs2_local_alloc_slide_window(struct ocfs2_super *osb,
76 struct inode *local_alloc_inode);
77
78 #ifdef CONFIG_OCFS2_FS_STATS
79
ocfs2_la_debug_open(struct inode * inode,struct file * file)80 static int ocfs2_la_debug_open(struct inode *inode, struct file *file)
81 {
82 file->private_data = inode->i_private;
83 return 0;
84 }
85
86 #define LA_DEBUG_BUF_SZ PAGE_CACHE_SIZE
87 #define LA_DEBUG_VER 1
ocfs2_la_debug_read(struct file * file,char __user * userbuf,size_t count,loff_t * ppos)88 static ssize_t ocfs2_la_debug_read(struct file *file, char __user *userbuf,
89 size_t count, loff_t *ppos)
90 {
91 static DEFINE_MUTEX(la_debug_mutex);
92 struct ocfs2_super *osb = file->private_data;
93 int written, ret;
94 char *buf = osb->local_alloc_debug_buf;
95
96 mutex_lock(&la_debug_mutex);
97 memset(buf, 0, LA_DEBUG_BUF_SZ);
98
99 written = snprintf(buf, LA_DEBUG_BUF_SZ,
100 "0x%x\t0x%llx\t%u\t%u\t0x%x\n",
101 LA_DEBUG_VER,
102 (unsigned long long)osb->la_last_gd,
103 osb->local_alloc_default_bits,
104 osb->local_alloc_bits, osb->local_alloc_state);
105
106 ret = simple_read_from_buffer(userbuf, count, ppos, buf, written);
107
108 mutex_unlock(&la_debug_mutex);
109 return ret;
110 }
111
112 static const struct file_operations ocfs2_la_debug_fops = {
113 .open = ocfs2_la_debug_open,
114 .read = ocfs2_la_debug_read,
115 };
116
ocfs2_init_la_debug(struct ocfs2_super * osb)117 static void ocfs2_init_la_debug(struct ocfs2_super *osb)
118 {
119 osb->local_alloc_debug_buf = kmalloc(LA_DEBUG_BUF_SZ, GFP_NOFS);
120 if (!osb->local_alloc_debug_buf)
121 return;
122
123 osb->local_alloc_debug = debugfs_create_file("local_alloc_stats",
124 S_IFREG|S_IRUSR,
125 osb->osb_debug_root,
126 osb,
127 &ocfs2_la_debug_fops);
128 if (!osb->local_alloc_debug) {
129 kfree(osb->local_alloc_debug_buf);
130 osb->local_alloc_debug_buf = NULL;
131 }
132 }
133
ocfs2_shutdown_la_debug(struct ocfs2_super * osb)134 static void ocfs2_shutdown_la_debug(struct ocfs2_super *osb)
135 {
136 if (osb->local_alloc_debug)
137 debugfs_remove(osb->local_alloc_debug);
138
139 if (osb->local_alloc_debug_buf)
140 kfree(osb->local_alloc_debug_buf);
141
142 osb->local_alloc_debug_buf = NULL;
143 osb->local_alloc_debug = NULL;
144 }
145 #else /* CONFIG_OCFS2_FS_STATS */
ocfs2_init_la_debug(struct ocfs2_super * osb)146 static void ocfs2_init_la_debug(struct ocfs2_super *osb)
147 {
148 return;
149 }
ocfs2_shutdown_la_debug(struct ocfs2_super * osb)150 static void ocfs2_shutdown_la_debug(struct ocfs2_super *osb)
151 {
152 return;
153 }
154 #endif
155
ocfs2_la_state_enabled(struct ocfs2_super * osb)156 static inline int ocfs2_la_state_enabled(struct ocfs2_super *osb)
157 {
158 return (osb->local_alloc_state == OCFS2_LA_THROTTLED ||
159 osb->local_alloc_state == OCFS2_LA_ENABLED);
160 }
161
ocfs2_local_alloc_seen_free_bits(struct ocfs2_super * osb,unsigned int num_clusters)162 void ocfs2_local_alloc_seen_free_bits(struct ocfs2_super *osb,
163 unsigned int num_clusters)
164 {
165 spin_lock(&osb->osb_lock);
166 if (osb->local_alloc_state == OCFS2_LA_DISABLED ||
167 osb->local_alloc_state == OCFS2_LA_THROTTLED)
168 if (num_clusters >= osb->local_alloc_default_bits) {
169 cancel_delayed_work(&osb->la_enable_wq);
170 osb->local_alloc_state = OCFS2_LA_ENABLED;
171 }
172 spin_unlock(&osb->osb_lock);
173 }
174
ocfs2_la_enable_worker(struct work_struct * work)175 void ocfs2_la_enable_worker(struct work_struct *work)
176 {
177 struct ocfs2_super *osb =
178 container_of(work, struct ocfs2_super,
179 la_enable_wq.work);
180 spin_lock(&osb->osb_lock);
181 osb->local_alloc_state = OCFS2_LA_ENABLED;
182 spin_unlock(&osb->osb_lock);
183 }
184
185 /*
186 * Tell us whether a given allocation should use the local alloc
187 * file. Otherwise, it has to go to the main bitmap.
188 *
189 * This function does semi-dirty reads of local alloc size and state!
190 * This is ok however, as the values are re-checked once under mutex.
191 */
ocfs2_alloc_should_use_local(struct ocfs2_super * osb,u64 bits)192 int ocfs2_alloc_should_use_local(struct ocfs2_super *osb, u64 bits)
193 {
194 int ret = 0;
195 int la_bits;
196
197 spin_lock(&osb->osb_lock);
198 la_bits = osb->local_alloc_bits;
199
200 if (!ocfs2_la_state_enabled(osb))
201 goto bail;
202
203 /* la_bits should be at least twice the size (in clusters) of
204 * a new block group. We want to be sure block group
205 * allocations go through the local alloc, so allow an
206 * allocation to take up to half the bitmap. */
207 if (bits > (la_bits / 2))
208 goto bail;
209
210 ret = 1;
211 bail:
212 mlog(0, "state=%d, bits=%llu, la_bits=%d, ret=%d\n",
213 osb->local_alloc_state, (unsigned long long)bits, la_bits, ret);
214 spin_unlock(&osb->osb_lock);
215 return ret;
216 }
217
ocfs2_load_local_alloc(struct ocfs2_super * osb)218 int ocfs2_load_local_alloc(struct ocfs2_super *osb)
219 {
220 int status = 0;
221 struct ocfs2_dinode *alloc = NULL;
222 struct buffer_head *alloc_bh = NULL;
223 u32 num_used;
224 struct inode *inode = NULL;
225 struct ocfs2_local_alloc *la;
226
227 mlog_entry_void();
228
229 ocfs2_init_la_debug(osb);
230
231 if (osb->local_alloc_bits == 0)
232 goto bail;
233
234 if (osb->local_alloc_bits >= osb->bitmap_cpg) {
235 mlog(ML_NOTICE, "Requested local alloc window %d is larger "
236 "than max possible %u. Using defaults.\n",
237 osb->local_alloc_bits, (osb->bitmap_cpg - 1));
238 osb->local_alloc_bits =
239 ocfs2_megabytes_to_clusters(osb->sb,
240 OCFS2_DEFAULT_LOCAL_ALLOC_SIZE);
241 }
242
243 /* read the alloc off disk */
244 inode = ocfs2_get_system_file_inode(osb, LOCAL_ALLOC_SYSTEM_INODE,
245 osb->slot_num);
246 if (!inode) {
247 status = -EINVAL;
248 mlog_errno(status);
249 goto bail;
250 }
251
252 status = ocfs2_read_inode_block_full(inode, &alloc_bh,
253 OCFS2_BH_IGNORE_CACHE);
254 if (status < 0) {
255 mlog_errno(status);
256 goto bail;
257 }
258
259 alloc = (struct ocfs2_dinode *) alloc_bh->b_data;
260 la = OCFS2_LOCAL_ALLOC(alloc);
261
262 if (!(le32_to_cpu(alloc->i_flags) &
263 (OCFS2_LOCAL_ALLOC_FL|OCFS2_BITMAP_FL))) {
264 mlog(ML_ERROR, "Invalid local alloc inode, %llu\n",
265 (unsigned long long)OCFS2_I(inode)->ip_blkno);
266 status = -EINVAL;
267 goto bail;
268 }
269
270 if ((la->la_size == 0) ||
271 (le16_to_cpu(la->la_size) > ocfs2_local_alloc_size(inode->i_sb))) {
272 mlog(ML_ERROR, "Local alloc size is invalid (la_size = %u)\n",
273 le16_to_cpu(la->la_size));
274 status = -EINVAL;
275 goto bail;
276 }
277
278 /* do a little verification. */
279 num_used = ocfs2_local_alloc_count_bits(alloc);
280
281 /* hopefully the local alloc has always been recovered before
282 * we load it. */
283 if (num_used
284 || alloc->id1.bitmap1.i_used
285 || alloc->id1.bitmap1.i_total
286 || la->la_bm_off)
287 mlog(ML_ERROR, "Local alloc hasn't been recovered!\n"
288 "found = %u, set = %u, taken = %u, off = %u\n",
289 num_used, le32_to_cpu(alloc->id1.bitmap1.i_used),
290 le32_to_cpu(alloc->id1.bitmap1.i_total),
291 OCFS2_LOCAL_ALLOC(alloc)->la_bm_off);
292
293 osb->local_alloc_bh = alloc_bh;
294 osb->local_alloc_state = OCFS2_LA_ENABLED;
295
296 bail:
297 if (status < 0)
298 brelse(alloc_bh);
299 if (inode)
300 iput(inode);
301
302 if (status < 0)
303 ocfs2_shutdown_la_debug(osb);
304
305 mlog(0, "Local alloc window bits = %d\n", osb->local_alloc_bits);
306
307 mlog_exit(status);
308 return status;
309 }
310
311 /*
312 * return any unused bits to the bitmap and write out a clean
313 * local_alloc.
314 *
315 * local_alloc_bh is optional. If not passed, we will simply use the
316 * one off osb. If you do pass it however, be warned that it *will* be
317 * returned brelse'd and NULL'd out.*/
ocfs2_shutdown_local_alloc(struct ocfs2_super * osb)318 void ocfs2_shutdown_local_alloc(struct ocfs2_super *osb)
319 {
320 int status;
321 handle_t *handle;
322 struct inode *local_alloc_inode = NULL;
323 struct buffer_head *bh = NULL;
324 struct buffer_head *main_bm_bh = NULL;
325 struct inode *main_bm_inode = NULL;
326 struct ocfs2_dinode *alloc_copy = NULL;
327 struct ocfs2_dinode *alloc = NULL;
328
329 mlog_entry_void();
330
331 cancel_delayed_work(&osb->la_enable_wq);
332 flush_workqueue(ocfs2_wq);
333
334 ocfs2_shutdown_la_debug(osb);
335
336 if (osb->local_alloc_state == OCFS2_LA_UNUSED)
337 goto out;
338
339 local_alloc_inode =
340 ocfs2_get_system_file_inode(osb,
341 LOCAL_ALLOC_SYSTEM_INODE,
342 osb->slot_num);
343 if (!local_alloc_inode) {
344 status = -ENOENT;
345 mlog_errno(status);
346 goto out;
347 }
348
349 osb->local_alloc_state = OCFS2_LA_DISABLED;
350
351 main_bm_inode = ocfs2_get_system_file_inode(osb,
352 GLOBAL_BITMAP_SYSTEM_INODE,
353 OCFS2_INVALID_SLOT);
354 if (!main_bm_inode) {
355 status = -EINVAL;
356 mlog_errno(status);
357 goto out;
358 }
359
360 mutex_lock(&main_bm_inode->i_mutex);
361
362 status = ocfs2_inode_lock(main_bm_inode, &main_bm_bh, 1);
363 if (status < 0) {
364 mlog_errno(status);
365 goto out_mutex;
366 }
367
368 /* WINDOW_MOVE_CREDITS is a bit heavy... */
369 handle = ocfs2_start_trans(osb, OCFS2_WINDOW_MOVE_CREDITS);
370 if (IS_ERR(handle)) {
371 mlog_errno(PTR_ERR(handle));
372 handle = NULL;
373 goto out_unlock;
374 }
375
376 bh = osb->local_alloc_bh;
377 alloc = (struct ocfs2_dinode *) bh->b_data;
378
379 alloc_copy = kmalloc(bh->b_size, GFP_NOFS);
380 if (!alloc_copy) {
381 status = -ENOMEM;
382 goto out_commit;
383 }
384 memcpy(alloc_copy, alloc, bh->b_size);
385
386 status = ocfs2_journal_access_di(handle, local_alloc_inode, bh,
387 OCFS2_JOURNAL_ACCESS_WRITE);
388 if (status < 0) {
389 mlog_errno(status);
390 goto out_commit;
391 }
392
393 ocfs2_clear_local_alloc(alloc);
394
395 status = ocfs2_journal_dirty(handle, bh);
396 if (status < 0) {
397 mlog_errno(status);
398 goto out_commit;
399 }
400
401 brelse(bh);
402 osb->local_alloc_bh = NULL;
403 osb->local_alloc_state = OCFS2_LA_UNUSED;
404
405 status = ocfs2_sync_local_to_main(osb, handle, alloc_copy,
406 main_bm_inode, main_bm_bh);
407 if (status < 0)
408 mlog_errno(status);
409
410 out_commit:
411 ocfs2_commit_trans(osb, handle);
412
413 out_unlock:
414 brelse(main_bm_bh);
415
416 ocfs2_inode_unlock(main_bm_inode, 1);
417
418 out_mutex:
419 mutex_unlock(&main_bm_inode->i_mutex);
420 iput(main_bm_inode);
421
422 out:
423 if (local_alloc_inode)
424 iput(local_alloc_inode);
425
426 if (alloc_copy)
427 kfree(alloc_copy);
428
429 mlog_exit_void();
430 }
431
432 /*
433 * We want to free the bitmap bits outside of any recovery context as
434 * we'll need a cluster lock to do so, but we must clear the local
435 * alloc before giving up the recovered nodes journal. To solve this,
436 * we kmalloc a copy of the local alloc before it's change for the
437 * caller to process with ocfs2_complete_local_alloc_recovery
438 */
ocfs2_begin_local_alloc_recovery(struct ocfs2_super * osb,int slot_num,struct ocfs2_dinode ** alloc_copy)439 int ocfs2_begin_local_alloc_recovery(struct ocfs2_super *osb,
440 int slot_num,
441 struct ocfs2_dinode **alloc_copy)
442 {
443 int status = 0;
444 struct buffer_head *alloc_bh = NULL;
445 struct inode *inode = NULL;
446 struct ocfs2_dinode *alloc;
447
448 mlog_entry("(slot_num = %d)\n", slot_num);
449
450 *alloc_copy = NULL;
451
452 inode = ocfs2_get_system_file_inode(osb,
453 LOCAL_ALLOC_SYSTEM_INODE,
454 slot_num);
455 if (!inode) {
456 status = -EINVAL;
457 mlog_errno(status);
458 goto bail;
459 }
460
461 mutex_lock(&inode->i_mutex);
462
463 status = ocfs2_read_inode_block_full(inode, &alloc_bh,
464 OCFS2_BH_IGNORE_CACHE);
465 if (status < 0) {
466 mlog_errno(status);
467 goto bail;
468 }
469
470 *alloc_copy = kmalloc(alloc_bh->b_size, GFP_KERNEL);
471 if (!(*alloc_copy)) {
472 status = -ENOMEM;
473 goto bail;
474 }
475 memcpy((*alloc_copy), alloc_bh->b_data, alloc_bh->b_size);
476
477 alloc = (struct ocfs2_dinode *) alloc_bh->b_data;
478 ocfs2_clear_local_alloc(alloc);
479
480 ocfs2_compute_meta_ecc(osb->sb, alloc_bh->b_data, &alloc->i_check);
481 status = ocfs2_write_block(osb, alloc_bh, inode);
482 if (status < 0)
483 mlog_errno(status);
484
485 bail:
486 if ((status < 0) && (*alloc_copy)) {
487 kfree(*alloc_copy);
488 *alloc_copy = NULL;
489 }
490
491 brelse(alloc_bh);
492
493 if (inode) {
494 mutex_unlock(&inode->i_mutex);
495 iput(inode);
496 }
497
498 mlog_exit(status);
499 return status;
500 }
501
502 /*
503 * Step 2: By now, we've completed the journal recovery, we've stamped
504 * a clean local alloc on disk and dropped the node out of the
505 * recovery map. Dlm locks will no longer stall, so lets clear out the
506 * main bitmap.
507 */
ocfs2_complete_local_alloc_recovery(struct ocfs2_super * osb,struct ocfs2_dinode * alloc)508 int ocfs2_complete_local_alloc_recovery(struct ocfs2_super *osb,
509 struct ocfs2_dinode *alloc)
510 {
511 int status;
512 handle_t *handle;
513 struct buffer_head *main_bm_bh = NULL;
514 struct inode *main_bm_inode;
515
516 mlog_entry_void();
517
518 main_bm_inode = ocfs2_get_system_file_inode(osb,
519 GLOBAL_BITMAP_SYSTEM_INODE,
520 OCFS2_INVALID_SLOT);
521 if (!main_bm_inode) {
522 status = -EINVAL;
523 mlog_errno(status);
524 goto out;
525 }
526
527 mutex_lock(&main_bm_inode->i_mutex);
528
529 status = ocfs2_inode_lock(main_bm_inode, &main_bm_bh, 1);
530 if (status < 0) {
531 mlog_errno(status);
532 goto out_mutex;
533 }
534
535 handle = ocfs2_start_trans(osb, OCFS2_WINDOW_MOVE_CREDITS);
536 if (IS_ERR(handle)) {
537 status = PTR_ERR(handle);
538 handle = NULL;
539 mlog_errno(status);
540 goto out_unlock;
541 }
542
543 /* we want the bitmap change to be recorded on disk asap */
544 handle->h_sync = 1;
545
546 status = ocfs2_sync_local_to_main(osb, handle, alloc,
547 main_bm_inode, main_bm_bh);
548 if (status < 0)
549 mlog_errno(status);
550
551 ocfs2_commit_trans(osb, handle);
552
553 out_unlock:
554 ocfs2_inode_unlock(main_bm_inode, 1);
555
556 out_mutex:
557 mutex_unlock(&main_bm_inode->i_mutex);
558
559 brelse(main_bm_bh);
560
561 iput(main_bm_inode);
562
563 out:
564 if (!status)
565 ocfs2_init_inode_steal_slot(osb);
566 mlog_exit(status);
567 return status;
568 }
569
570 /* Check to see if the local alloc window is within ac->ac_max_block */
ocfs2_local_alloc_in_range(struct inode * inode,struct ocfs2_alloc_context * ac,u32 bits_wanted)571 static int ocfs2_local_alloc_in_range(struct inode *inode,
572 struct ocfs2_alloc_context *ac,
573 u32 bits_wanted)
574 {
575 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
576 struct ocfs2_dinode *alloc;
577 struct ocfs2_local_alloc *la;
578 int start;
579 u64 block_off;
580
581 if (!ac->ac_max_block)
582 return 1;
583
584 alloc = (struct ocfs2_dinode *) osb->local_alloc_bh->b_data;
585 la = OCFS2_LOCAL_ALLOC(alloc);
586
587 start = ocfs2_local_alloc_find_clear_bits(osb, alloc, bits_wanted);
588 if (start == -1) {
589 mlog_errno(-ENOSPC);
590 return 0;
591 }
592
593 /*
594 * Converting (bm_off + start + bits_wanted) to blocks gives us
595 * the blkno just past our actual allocation. This is perfect
596 * to compare with ac_max_block.
597 */
598 block_off = ocfs2_clusters_to_blocks(inode->i_sb,
599 le32_to_cpu(la->la_bm_off) +
600 start + bits_wanted);
601 mlog(0, "Checking %llu against %llu\n",
602 (unsigned long long)block_off,
603 (unsigned long long)ac->ac_max_block);
604 if (block_off > ac->ac_max_block)
605 return 0;
606
607 return 1;
608 }
609
610 /*
611 * make sure we've got at least bits_wanted contiguous bits in the
612 * local alloc. You lose them when you drop i_mutex.
613 *
614 * We will add ourselves to the transaction passed in, but may start
615 * our own in order to shift windows.
616 */
ocfs2_reserve_local_alloc_bits(struct ocfs2_super * osb,u32 bits_wanted,struct ocfs2_alloc_context * ac)617 int ocfs2_reserve_local_alloc_bits(struct ocfs2_super *osb,
618 u32 bits_wanted,
619 struct ocfs2_alloc_context *ac)
620 {
621 int status;
622 struct ocfs2_dinode *alloc;
623 struct inode *local_alloc_inode;
624 unsigned int free_bits;
625
626 mlog_entry_void();
627
628 BUG_ON(!ac);
629
630 local_alloc_inode =
631 ocfs2_get_system_file_inode(osb,
632 LOCAL_ALLOC_SYSTEM_INODE,
633 osb->slot_num);
634 if (!local_alloc_inode) {
635 status = -ENOENT;
636 mlog_errno(status);
637 goto bail;
638 }
639
640 mutex_lock(&local_alloc_inode->i_mutex);
641
642 /*
643 * We must double check state and allocator bits because
644 * another process may have changed them while holding i_mutex.
645 */
646 spin_lock(&osb->osb_lock);
647 if (!ocfs2_la_state_enabled(osb) ||
648 (bits_wanted > osb->local_alloc_bits)) {
649 spin_unlock(&osb->osb_lock);
650 status = -ENOSPC;
651 goto bail;
652 }
653 spin_unlock(&osb->osb_lock);
654
655 alloc = (struct ocfs2_dinode *) osb->local_alloc_bh->b_data;
656
657 #ifdef CONFIG_OCFS2_DEBUG_FS
658 if (le32_to_cpu(alloc->id1.bitmap1.i_used) !=
659 ocfs2_local_alloc_count_bits(alloc)) {
660 ocfs2_error(osb->sb, "local alloc inode %llu says it has "
661 "%u free bits, but a count shows %u",
662 (unsigned long long)le64_to_cpu(alloc->i_blkno),
663 le32_to_cpu(alloc->id1.bitmap1.i_used),
664 ocfs2_local_alloc_count_bits(alloc));
665 status = -EIO;
666 goto bail;
667 }
668 #endif
669
670 free_bits = le32_to_cpu(alloc->id1.bitmap1.i_total) -
671 le32_to_cpu(alloc->id1.bitmap1.i_used);
672 if (bits_wanted > free_bits) {
673 /* uhoh, window change time. */
674 status =
675 ocfs2_local_alloc_slide_window(osb, local_alloc_inode);
676 if (status < 0) {
677 if (status != -ENOSPC)
678 mlog_errno(status);
679 goto bail;
680 }
681
682 /*
683 * Under certain conditions, the window slide code
684 * might have reduced the number of bits available or
685 * disabled the the local alloc entirely. Re-check
686 * here and return -ENOSPC if necessary.
687 */
688 status = -ENOSPC;
689 if (!ocfs2_la_state_enabled(osb))
690 goto bail;
691
692 free_bits = le32_to_cpu(alloc->id1.bitmap1.i_total) -
693 le32_to_cpu(alloc->id1.bitmap1.i_used);
694 if (bits_wanted > free_bits)
695 goto bail;
696 }
697
698 if (ac->ac_max_block)
699 mlog(0, "Calling in_range for max block %llu\n",
700 (unsigned long long)ac->ac_max_block);
701
702 if (!ocfs2_local_alloc_in_range(local_alloc_inode, ac,
703 bits_wanted)) {
704 /*
705 * The window is outside ac->ac_max_block.
706 * This errno tells the caller to keep localalloc enabled
707 * but to get the allocation from the main bitmap.
708 */
709 status = -EFBIG;
710 goto bail;
711 }
712
713 ac->ac_inode = local_alloc_inode;
714 /* We should never use localalloc from another slot */
715 ac->ac_alloc_slot = osb->slot_num;
716 ac->ac_which = OCFS2_AC_USE_LOCAL;
717 get_bh(osb->local_alloc_bh);
718 ac->ac_bh = osb->local_alloc_bh;
719 status = 0;
720 bail:
721 if (status < 0 && local_alloc_inode) {
722 mutex_unlock(&local_alloc_inode->i_mutex);
723 iput(local_alloc_inode);
724 }
725
726 mlog(0, "bits=%d, slot=%d, ret=%d\n", bits_wanted, osb->slot_num,
727 status);
728
729 mlog_exit(status);
730 return status;
731 }
732
ocfs2_claim_local_alloc_bits(struct ocfs2_super * osb,handle_t * handle,struct ocfs2_alloc_context * ac,u32 bits_wanted,u32 * bit_off,u32 * num_bits)733 int ocfs2_claim_local_alloc_bits(struct ocfs2_super *osb,
734 handle_t *handle,
735 struct ocfs2_alloc_context *ac,
736 u32 bits_wanted,
737 u32 *bit_off,
738 u32 *num_bits)
739 {
740 int status, start;
741 struct inode *local_alloc_inode;
742 void *bitmap;
743 struct ocfs2_dinode *alloc;
744 struct ocfs2_local_alloc *la;
745
746 mlog_entry_void();
747 BUG_ON(ac->ac_which != OCFS2_AC_USE_LOCAL);
748
749 local_alloc_inode = ac->ac_inode;
750 alloc = (struct ocfs2_dinode *) osb->local_alloc_bh->b_data;
751 la = OCFS2_LOCAL_ALLOC(alloc);
752
753 start = ocfs2_local_alloc_find_clear_bits(osb, alloc, bits_wanted);
754 if (start == -1) {
755 /* TODO: Shouldn't we just BUG here? */
756 status = -ENOSPC;
757 mlog_errno(status);
758 goto bail;
759 }
760
761 bitmap = la->la_bitmap;
762 *bit_off = le32_to_cpu(la->la_bm_off) + start;
763 /* local alloc is always contiguous by nature -- we never
764 * delete bits from it! */
765 *num_bits = bits_wanted;
766
767 status = ocfs2_journal_access_di(handle, local_alloc_inode,
768 osb->local_alloc_bh,
769 OCFS2_JOURNAL_ACCESS_WRITE);
770 if (status < 0) {
771 mlog_errno(status);
772 goto bail;
773 }
774
775 while(bits_wanted--)
776 ocfs2_set_bit(start++, bitmap);
777
778 le32_add_cpu(&alloc->id1.bitmap1.i_used, *num_bits);
779
780 status = ocfs2_journal_dirty(handle, osb->local_alloc_bh);
781 if (status < 0) {
782 mlog_errno(status);
783 goto bail;
784 }
785
786 status = 0;
787 bail:
788 mlog_exit(status);
789 return status;
790 }
791
ocfs2_local_alloc_count_bits(struct ocfs2_dinode * alloc)792 static u32 ocfs2_local_alloc_count_bits(struct ocfs2_dinode *alloc)
793 {
794 int i;
795 u8 *buffer;
796 u32 count = 0;
797 struct ocfs2_local_alloc *la = OCFS2_LOCAL_ALLOC(alloc);
798
799 mlog_entry_void();
800
801 buffer = la->la_bitmap;
802 for (i = 0; i < le16_to_cpu(la->la_size); i++)
803 count += hweight8(buffer[i]);
804
805 mlog_exit(count);
806 return count;
807 }
808
ocfs2_local_alloc_find_clear_bits(struct ocfs2_super * osb,struct ocfs2_dinode * alloc,u32 numbits)809 static int ocfs2_local_alloc_find_clear_bits(struct ocfs2_super *osb,
810 struct ocfs2_dinode *alloc,
811 u32 numbits)
812 {
813 int numfound, bitoff, left, startoff, lastzero;
814 void *bitmap = NULL;
815
816 mlog_entry("(numbits wanted = %u)\n", numbits);
817
818 if (!alloc->id1.bitmap1.i_total) {
819 mlog(0, "No bits in my window!\n");
820 bitoff = -1;
821 goto bail;
822 }
823
824 bitmap = OCFS2_LOCAL_ALLOC(alloc)->la_bitmap;
825
826 numfound = bitoff = startoff = 0;
827 lastzero = -1;
828 left = le32_to_cpu(alloc->id1.bitmap1.i_total);
829 while ((bitoff = ocfs2_find_next_zero_bit(bitmap, left, startoff)) != -1) {
830 if (bitoff == left) {
831 /* mlog(0, "bitoff (%d) == left", bitoff); */
832 break;
833 }
834 /* mlog(0, "Found a zero: bitoff = %d, startoff = %d, "
835 "numfound = %d\n", bitoff, startoff, numfound);*/
836
837 /* Ok, we found a zero bit... is it contig. or do we
838 * start over?*/
839 if (bitoff == startoff) {
840 /* we found a zero */
841 numfound++;
842 startoff++;
843 } else {
844 /* got a zero after some ones */
845 numfound = 1;
846 startoff = bitoff+1;
847 }
848 /* we got everything we needed */
849 if (numfound == numbits) {
850 /* mlog(0, "Found it all!\n"); */
851 break;
852 }
853 }
854
855 mlog(0, "Exiting loop, bitoff = %d, numfound = %d\n", bitoff,
856 numfound);
857
858 if (numfound == numbits)
859 bitoff = startoff - numfound;
860 else
861 bitoff = -1;
862
863 bail:
864 mlog_exit(bitoff);
865 return bitoff;
866 }
867
ocfs2_clear_local_alloc(struct ocfs2_dinode * alloc)868 static void ocfs2_clear_local_alloc(struct ocfs2_dinode *alloc)
869 {
870 struct ocfs2_local_alloc *la = OCFS2_LOCAL_ALLOC(alloc);
871 int i;
872 mlog_entry_void();
873
874 alloc->id1.bitmap1.i_total = 0;
875 alloc->id1.bitmap1.i_used = 0;
876 la->la_bm_off = 0;
877 for(i = 0; i < le16_to_cpu(la->la_size); i++)
878 la->la_bitmap[i] = 0;
879
880 mlog_exit_void();
881 }
882
883 #if 0
884 /* turn this on and uncomment below to aid debugging window shifts. */
885 static void ocfs2_verify_zero_bits(unsigned long *bitmap,
886 unsigned int start,
887 unsigned int count)
888 {
889 unsigned int tmp = count;
890 while(tmp--) {
891 if (ocfs2_test_bit(start + tmp, bitmap)) {
892 printk("ocfs2_verify_zero_bits: start = %u, count = "
893 "%u\n", start, count);
894 printk("ocfs2_verify_zero_bits: bit %u is set!",
895 start + tmp);
896 BUG();
897 }
898 }
899 }
900 #endif
901
902 /*
903 * sync the local alloc to main bitmap.
904 *
905 * assumes you've already locked the main bitmap -- the bitmap inode
906 * passed is used for caching.
907 */
ocfs2_sync_local_to_main(struct ocfs2_super * osb,handle_t * handle,struct ocfs2_dinode * alloc,struct inode * main_bm_inode,struct buffer_head * main_bm_bh)908 static int ocfs2_sync_local_to_main(struct ocfs2_super *osb,
909 handle_t *handle,
910 struct ocfs2_dinode *alloc,
911 struct inode *main_bm_inode,
912 struct buffer_head *main_bm_bh)
913 {
914 int status = 0;
915 int bit_off, left, count, start;
916 u64 la_start_blk;
917 u64 blkno;
918 void *bitmap;
919 struct ocfs2_local_alloc *la = OCFS2_LOCAL_ALLOC(alloc);
920
921 mlog_entry("total = %u, used = %u\n",
922 le32_to_cpu(alloc->id1.bitmap1.i_total),
923 le32_to_cpu(alloc->id1.bitmap1.i_used));
924
925 if (!alloc->id1.bitmap1.i_total) {
926 mlog(0, "nothing to sync!\n");
927 goto bail;
928 }
929
930 if (le32_to_cpu(alloc->id1.bitmap1.i_used) ==
931 le32_to_cpu(alloc->id1.bitmap1.i_total)) {
932 mlog(0, "all bits were taken!\n");
933 goto bail;
934 }
935
936 la_start_blk = ocfs2_clusters_to_blocks(osb->sb,
937 le32_to_cpu(la->la_bm_off));
938 bitmap = la->la_bitmap;
939 start = count = bit_off = 0;
940 left = le32_to_cpu(alloc->id1.bitmap1.i_total);
941
942 while ((bit_off = ocfs2_find_next_zero_bit(bitmap, left, start))
943 != -1) {
944 if ((bit_off < left) && (bit_off == start)) {
945 count++;
946 start++;
947 continue;
948 }
949 if (count) {
950 blkno = la_start_blk +
951 ocfs2_clusters_to_blocks(osb->sb,
952 start - count);
953
954 mlog(0, "freeing %u bits starting at local alloc bit "
955 "%u (la_start_blk = %llu, blkno = %llu)\n",
956 count, start - count,
957 (unsigned long long)la_start_blk,
958 (unsigned long long)blkno);
959
960 status = ocfs2_free_clusters(handle, main_bm_inode,
961 main_bm_bh, blkno, count);
962 if (status < 0) {
963 mlog_errno(status);
964 goto bail;
965 }
966 }
967 if (bit_off >= left)
968 break;
969 count = 1;
970 start = bit_off + 1;
971 }
972
973 bail:
974 mlog_exit(status);
975 return status;
976 }
977
978 enum ocfs2_la_event {
979 OCFS2_LA_EVENT_SLIDE, /* Normal window slide. */
980 OCFS2_LA_EVENT_FRAGMENTED, /* The global bitmap has
981 * enough bits theoretically
982 * free, but a contiguous
983 * allocation could not be
984 * found. */
985 OCFS2_LA_EVENT_ENOSPC, /* Global bitmap doesn't have
986 * enough bits free to satisfy
987 * our request. */
988 };
989 #define OCFS2_LA_ENABLE_INTERVAL (30 * HZ)
990 /*
991 * Given an event, calculate the size of our next local alloc window.
992 *
993 * This should always be called under i_mutex of the local alloc inode
994 * so that local alloc disabling doesn't race with processes trying to
995 * use the allocator.
996 *
997 * Returns the state which the local alloc was left in. This value can
998 * be ignored by some paths.
999 */
ocfs2_recalc_la_window(struct ocfs2_super * osb,enum ocfs2_la_event event)1000 static int ocfs2_recalc_la_window(struct ocfs2_super *osb,
1001 enum ocfs2_la_event event)
1002 {
1003 unsigned int bits;
1004 int state;
1005
1006 spin_lock(&osb->osb_lock);
1007 if (osb->local_alloc_state == OCFS2_LA_DISABLED) {
1008 WARN_ON_ONCE(osb->local_alloc_state == OCFS2_LA_DISABLED);
1009 goto out_unlock;
1010 }
1011
1012 /*
1013 * ENOSPC and fragmentation are treated similarly for now.
1014 */
1015 if (event == OCFS2_LA_EVENT_ENOSPC ||
1016 event == OCFS2_LA_EVENT_FRAGMENTED) {
1017 /*
1018 * We ran out of contiguous space in the primary
1019 * bitmap. Drastically reduce the number of bits used
1020 * by local alloc until we have to disable it.
1021 */
1022 bits = osb->local_alloc_bits >> 1;
1023 if (bits > ocfs2_megabytes_to_clusters(osb->sb, 1)) {
1024 /*
1025 * By setting state to THROTTLED, we'll keep
1026 * the number of local alloc bits used down
1027 * until an event occurs which would give us
1028 * reason to assume the bitmap situation might
1029 * have changed.
1030 */
1031 osb->local_alloc_state = OCFS2_LA_THROTTLED;
1032 osb->local_alloc_bits = bits;
1033 } else {
1034 osb->local_alloc_state = OCFS2_LA_DISABLED;
1035 }
1036 queue_delayed_work(ocfs2_wq, &osb->la_enable_wq,
1037 OCFS2_LA_ENABLE_INTERVAL);
1038 goto out_unlock;
1039 }
1040
1041 /*
1042 * Don't increase the size of the local alloc window until we
1043 * know we might be able to fulfill the request. Otherwise, we
1044 * risk bouncing around the global bitmap during periods of
1045 * low space.
1046 */
1047 if (osb->local_alloc_state != OCFS2_LA_THROTTLED)
1048 osb->local_alloc_bits = osb->local_alloc_default_bits;
1049
1050 out_unlock:
1051 state = osb->local_alloc_state;
1052 spin_unlock(&osb->osb_lock);
1053
1054 return state;
1055 }
1056
ocfs2_local_alloc_reserve_for_window(struct ocfs2_super * osb,struct ocfs2_alloc_context ** ac,struct inode ** bitmap_inode,struct buffer_head ** bitmap_bh)1057 static int ocfs2_local_alloc_reserve_for_window(struct ocfs2_super *osb,
1058 struct ocfs2_alloc_context **ac,
1059 struct inode **bitmap_inode,
1060 struct buffer_head **bitmap_bh)
1061 {
1062 int status;
1063
1064 *ac = kzalloc(sizeof(struct ocfs2_alloc_context), GFP_KERNEL);
1065 if (!(*ac)) {
1066 status = -ENOMEM;
1067 mlog_errno(status);
1068 goto bail;
1069 }
1070
1071 retry_enospc:
1072 (*ac)->ac_bits_wanted = osb->local_alloc_bits;
1073
1074 status = ocfs2_reserve_cluster_bitmap_bits(osb, *ac);
1075 if (status == -ENOSPC) {
1076 if (ocfs2_recalc_la_window(osb, OCFS2_LA_EVENT_ENOSPC) ==
1077 OCFS2_LA_DISABLED)
1078 goto bail;
1079
1080 ocfs2_free_ac_resource(*ac);
1081 memset(*ac, 0, sizeof(struct ocfs2_alloc_context));
1082 goto retry_enospc;
1083 }
1084 if (status < 0) {
1085 mlog_errno(status);
1086 goto bail;
1087 }
1088
1089 *bitmap_inode = (*ac)->ac_inode;
1090 igrab(*bitmap_inode);
1091 *bitmap_bh = (*ac)->ac_bh;
1092 get_bh(*bitmap_bh);
1093 status = 0;
1094 bail:
1095 if ((status < 0) && *ac) {
1096 ocfs2_free_alloc_context(*ac);
1097 *ac = NULL;
1098 }
1099
1100 mlog_exit(status);
1101 return status;
1102 }
1103
1104 /*
1105 * pass it the bitmap lock in lock_bh if you have it.
1106 */
ocfs2_local_alloc_new_window(struct ocfs2_super * osb,handle_t * handle,struct ocfs2_alloc_context * ac)1107 static int ocfs2_local_alloc_new_window(struct ocfs2_super *osb,
1108 handle_t *handle,
1109 struct ocfs2_alloc_context *ac)
1110 {
1111 int status = 0;
1112 u32 cluster_off, cluster_count;
1113 struct ocfs2_dinode *alloc = NULL;
1114 struct ocfs2_local_alloc *la;
1115
1116 mlog_entry_void();
1117
1118 alloc = (struct ocfs2_dinode *) osb->local_alloc_bh->b_data;
1119 la = OCFS2_LOCAL_ALLOC(alloc);
1120
1121 if (alloc->id1.bitmap1.i_total)
1122 mlog(0, "asking me to alloc a new window over a non-empty "
1123 "one\n");
1124
1125 mlog(0, "Allocating %u clusters for a new window.\n",
1126 osb->local_alloc_bits);
1127
1128 /* Instruct the allocation code to try the most recently used
1129 * cluster group. We'll re-record the group used this pass
1130 * below. */
1131 ac->ac_last_group = osb->la_last_gd;
1132
1133 /* we used the generic suballoc reserve function, but we set
1134 * everything up nicely, so there's no reason why we can't use
1135 * the more specific cluster api to claim bits. */
1136 status = ocfs2_claim_clusters(osb, handle, ac, osb->local_alloc_bits,
1137 &cluster_off, &cluster_count);
1138 if (status == -ENOSPC) {
1139 retry_enospc:
1140 /*
1141 * Note: We could also try syncing the journal here to
1142 * allow use of any free bits which the current
1143 * transaction can't give us access to. --Mark
1144 */
1145 if (ocfs2_recalc_la_window(osb, OCFS2_LA_EVENT_FRAGMENTED) ==
1146 OCFS2_LA_DISABLED)
1147 goto bail;
1148
1149 status = ocfs2_claim_clusters(osb, handle, ac,
1150 osb->local_alloc_bits,
1151 &cluster_off,
1152 &cluster_count);
1153 if (status == -ENOSPC)
1154 goto retry_enospc;
1155 /*
1156 * We only shrunk the *minimum* number of in our
1157 * request - it's entirely possible that the allocator
1158 * might give us more than we asked for.
1159 */
1160 if (status == 0) {
1161 spin_lock(&osb->osb_lock);
1162 osb->local_alloc_bits = cluster_count;
1163 spin_unlock(&osb->osb_lock);
1164 }
1165 }
1166 if (status < 0) {
1167 if (status != -ENOSPC)
1168 mlog_errno(status);
1169 goto bail;
1170 }
1171
1172 osb->la_last_gd = ac->ac_last_group;
1173
1174 la->la_bm_off = cpu_to_le32(cluster_off);
1175 alloc->id1.bitmap1.i_total = cpu_to_le32(cluster_count);
1176 /* just in case... In the future when we find space ourselves,
1177 * we don't have to get all contiguous -- but we'll have to
1178 * set all previously used bits in bitmap and update
1179 * la_bits_set before setting the bits in the main bitmap. */
1180 alloc->id1.bitmap1.i_used = 0;
1181 memset(OCFS2_LOCAL_ALLOC(alloc)->la_bitmap, 0,
1182 le16_to_cpu(la->la_size));
1183
1184 mlog(0, "New window allocated:\n");
1185 mlog(0, "window la_bm_off = %u\n",
1186 OCFS2_LOCAL_ALLOC(alloc)->la_bm_off);
1187 mlog(0, "window bits = %u\n", le32_to_cpu(alloc->id1.bitmap1.i_total));
1188
1189 bail:
1190 mlog_exit(status);
1191 return status;
1192 }
1193
1194 /* Note that we do *NOT* lock the local alloc inode here as
1195 * it's been locked already for us. */
ocfs2_local_alloc_slide_window(struct ocfs2_super * osb,struct inode * local_alloc_inode)1196 static int ocfs2_local_alloc_slide_window(struct ocfs2_super *osb,
1197 struct inode *local_alloc_inode)
1198 {
1199 int status = 0;
1200 struct buffer_head *main_bm_bh = NULL;
1201 struct inode *main_bm_inode = NULL;
1202 handle_t *handle = NULL;
1203 struct ocfs2_dinode *alloc;
1204 struct ocfs2_dinode *alloc_copy = NULL;
1205 struct ocfs2_alloc_context *ac = NULL;
1206
1207 mlog_entry_void();
1208
1209 ocfs2_recalc_la_window(osb, OCFS2_LA_EVENT_SLIDE);
1210
1211 /* This will lock the main bitmap for us. */
1212 status = ocfs2_local_alloc_reserve_for_window(osb,
1213 &ac,
1214 &main_bm_inode,
1215 &main_bm_bh);
1216 if (status < 0) {
1217 if (status != -ENOSPC)
1218 mlog_errno(status);
1219 goto bail;
1220 }
1221
1222 handle = ocfs2_start_trans(osb, OCFS2_WINDOW_MOVE_CREDITS);
1223 if (IS_ERR(handle)) {
1224 status = PTR_ERR(handle);
1225 handle = NULL;
1226 mlog_errno(status);
1227 goto bail;
1228 }
1229
1230 alloc = (struct ocfs2_dinode *) osb->local_alloc_bh->b_data;
1231
1232 /* We want to clear the local alloc before doing anything
1233 * else, so that if we error later during this operation,
1234 * local alloc shutdown won't try to double free main bitmap
1235 * bits. Make a copy so the sync function knows which bits to
1236 * free. */
1237 alloc_copy = kmalloc(osb->local_alloc_bh->b_size, GFP_NOFS);
1238 if (!alloc_copy) {
1239 status = -ENOMEM;
1240 mlog_errno(status);
1241 goto bail;
1242 }
1243 memcpy(alloc_copy, alloc, osb->local_alloc_bh->b_size);
1244
1245 status = ocfs2_journal_access_di(handle, local_alloc_inode,
1246 osb->local_alloc_bh,
1247 OCFS2_JOURNAL_ACCESS_WRITE);
1248 if (status < 0) {
1249 mlog_errno(status);
1250 goto bail;
1251 }
1252
1253 ocfs2_clear_local_alloc(alloc);
1254
1255 status = ocfs2_journal_dirty(handle, osb->local_alloc_bh);
1256 if (status < 0) {
1257 mlog_errno(status);
1258 goto bail;
1259 }
1260
1261 status = ocfs2_sync_local_to_main(osb, handle, alloc_copy,
1262 main_bm_inode, main_bm_bh);
1263 if (status < 0) {
1264 mlog_errno(status);
1265 goto bail;
1266 }
1267
1268 status = ocfs2_local_alloc_new_window(osb, handle, ac);
1269 if (status < 0) {
1270 if (status != -ENOSPC)
1271 mlog_errno(status);
1272 goto bail;
1273 }
1274
1275 atomic_inc(&osb->alloc_stats.moves);
1276
1277 status = 0;
1278 bail:
1279 if (handle)
1280 ocfs2_commit_trans(osb, handle);
1281
1282 brelse(main_bm_bh);
1283
1284 if (main_bm_inode)
1285 iput(main_bm_inode);
1286
1287 if (alloc_copy)
1288 kfree(alloc_copy);
1289
1290 if (ac)
1291 ocfs2_free_alloc_context(ac);
1292
1293 mlog_exit(status);
1294 return status;
1295 }
1296
1297