Commit | Line | Data |
---|---|---|
1da177e4 | 1 | /* |
eedb5530 | 2 | * Copyright (c) 2000-2005 Silicon Graphics, Inc. All Rights Reserved. |
1da177e4 LT |
3 | * |
4 | * This program is free software; you can redistribute it and/or modify it | |
5 | * under the terms of version 2 of the GNU General Public License as | |
6 | * published by the Free Software Foundation. | |
7 | * | |
8 | * This program is distributed in the hope that it would be useful, but | |
9 | * WITHOUT ANY WARRANTY; without even the implied warranty of | |
10 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. | |
11 | * | |
12 | * Further, this software is distributed without any warranty that it is | |
13 | * free of the rightful claim of any third person regarding infringement | |
14 | * or the like. Any license provided herein, whether implied or | |
15 | * otherwise, applies only to this software file. Patent licenses, if | |
16 | * any, provided herein do not apply to combinations of this program with | |
17 | * other software, or any other product whatsoever. | |
18 | * | |
19 | * You should have received a copy of the GNU General Public License along | |
20 | * with this program; if not, write the Free Software Foundation, Inc., 59 | |
21 | * Temple Place - Suite 330, Boston MA 02111-1307, USA. | |
22 | * | |
23 | * Contact information: Silicon Graphics, Inc., 1600 Amphitheatre Pkwy, | |
24 | * Mountain View, CA 94043, or: | |
25 | * | |
26 | * http://www.sgi.com | |
27 | * | |
28 | * For further information regarding this notice, see: | |
29 | * | |
30 | * http://oss.sgi.com/projects/GenInfo/SGIGPLNoticeExplan/ | |
31 | */ | |
32 | ||
33 | /* | |
34 | * The xfs_buf.c code provides an abstract buffer cache model on top | |
35 | * of the Linux page cache. Cached metadata blocks for a file system | |
36 | * are hashed to the inode for the block device. xfs_buf.c assembles | |
37 | * buffers (xfs_buf_t) on demand to aggregate such cached pages for I/O. | |
38 | * | |
39 | * Written by Steve Lord, Jim Mostek, Russell Cattelan | |
40 | * and Rajagopal Ananthanarayanan ("ananth") at SGI. | |
41 | * | |
42 | */ | |
43 | ||
44 | #include <linux/stddef.h> | |
45 | #include <linux/errno.h> | |
46 | #include <linux/slab.h> | |
47 | #include <linux/pagemap.h> | |
48 | #include <linux/init.h> | |
49 | #include <linux/vmalloc.h> | |
50 | #include <linux/bio.h> | |
51 | #include <linux/sysctl.h> | |
52 | #include <linux/proc_fs.h> | |
53 | #include <linux/workqueue.h> | |
54 | #include <linux/percpu.h> | |
55 | #include <linux/blkdev.h> | |
56 | #include <linux/hash.h> | |
4df08c52 | 57 | #include <linux/kthread.h> |
1da177e4 LT |
58 | |
59 | #include "xfs_linux.h" | |
60 | ||
61 | /* | |
62 | * File wide globals | |
63 | */ | |
64 | ||
23ea4032 | 65 | STATIC kmem_cache_t *pagebuf_zone; |
1da177e4 | 66 | STATIC kmem_shaker_t pagebuf_shake; |
27496a8c | 67 | STATIC int xfsbufd_wakeup(int, gfp_t); |
1da177e4 | 68 | STATIC void pagebuf_delwri_queue(xfs_buf_t *, int); |
23ea4032 CH |
69 | |
70 | STATIC struct workqueue_struct *xfslogd_workqueue; | |
0829c360 | 71 | struct workqueue_struct *xfsdatad_workqueue; |
1da177e4 LT |
72 | |
73 | /* | |
74 | * Pagebuf debugging | |
75 | */ | |
76 | ||
77 | #ifdef PAGEBUF_TRACE | |
78 | void | |
79 | pagebuf_trace( | |
80 | xfs_buf_t *pb, | |
81 | char *id, | |
82 | void *data, | |
83 | void *ra) | |
84 | { | |
85 | ktrace_enter(pagebuf_trace_buf, | |
86 | pb, id, | |
87 | (void *)(unsigned long)pb->pb_flags, | |
88 | (void *)(unsigned long)pb->pb_hold.counter, | |
89 | (void *)(unsigned long)pb->pb_sema.count.counter, | |
90 | (void *)current, | |
91 | data, ra, | |
92 | (void *)(unsigned long)((pb->pb_file_offset>>32) & 0xffffffff), | |
93 | (void *)(unsigned long)(pb->pb_file_offset & 0xffffffff), | |
94 | (void *)(unsigned long)pb->pb_buffer_length, | |
95 | NULL, NULL, NULL, NULL, NULL); | |
96 | } | |
97 | ktrace_t *pagebuf_trace_buf; | |
98 | #define PAGEBUF_TRACE_SIZE 4096 | |
99 | #define PB_TRACE(pb, id, data) \ | |
100 | pagebuf_trace(pb, id, (void *)data, (void *)__builtin_return_address(0)) | |
101 | #else | |
102 | #define PB_TRACE(pb, id, data) do { } while (0) | |
103 | #endif | |
104 | ||
105 | #ifdef PAGEBUF_LOCK_TRACKING | |
106 | # define PB_SET_OWNER(pb) ((pb)->pb_last_holder = current->pid) | |
107 | # define PB_CLEAR_OWNER(pb) ((pb)->pb_last_holder = -1) | |
108 | # define PB_GET_OWNER(pb) ((pb)->pb_last_holder) | |
109 | #else | |
110 | # define PB_SET_OWNER(pb) do { } while (0) | |
111 | # define PB_CLEAR_OWNER(pb) do { } while (0) | |
112 | # define PB_GET_OWNER(pb) do { } while (0) | |
113 | #endif | |
114 | ||
115 | /* | |
116 | * Pagebuf allocation / freeing. | |
117 | */ | |
118 | ||
119 | #define pb_to_gfp(flags) \ | |
120 | ((((flags) & PBF_READ_AHEAD) ? __GFP_NORETRY : \ | |
121 | ((flags) & PBF_DONT_BLOCK) ? GFP_NOFS : GFP_KERNEL) | __GFP_NOWARN) | |
122 | ||
123 | #define pb_to_km(flags) \ | |
124 | (((flags) & PBF_DONT_BLOCK) ? KM_NOFS : KM_SLEEP) | |
125 | ||
126 | ||
127 | #define pagebuf_allocate(flags) \ | |
23ea4032 | 128 | kmem_zone_alloc(pagebuf_zone, pb_to_km(flags)) |
1da177e4 | 129 | #define pagebuf_deallocate(pb) \ |
23ea4032 | 130 | kmem_zone_free(pagebuf_zone, (pb)); |
1da177e4 LT |
131 | |
132 | /* | |
133 | * Page Region interfaces. | |
134 | * | |
135 | * For pages in filesystems where the blocksize is smaller than the | |
136 | * pagesize, we use the page->private field (long) to hold a bitmap | |
137 | * of uptodate regions within the page. | |
138 | * | |
139 | * Each such region is "bytes per page / bits per long" bytes long. | |
140 | * | |
141 | * NBPPR == number-of-bytes-per-page-region | |
142 | * BTOPR == bytes-to-page-region (rounded up) | |
143 | * BTOPRT == bytes-to-page-region-truncated (rounded down) | |
144 | */ | |
145 | #if (BITS_PER_LONG == 32) | |
146 | #define PRSHIFT (PAGE_CACHE_SHIFT - 5) /* (32 == 1<<5) */ | |
147 | #elif (BITS_PER_LONG == 64) | |
148 | #define PRSHIFT (PAGE_CACHE_SHIFT - 6) /* (64 == 1<<6) */ | |
149 | #else | |
150 | #error BITS_PER_LONG must be 32 or 64 | |
151 | #endif | |
152 | #define NBPPR (PAGE_CACHE_SIZE/BITS_PER_LONG) | |
153 | #define BTOPR(b) (((unsigned int)(b) + (NBPPR - 1)) >> PRSHIFT) | |
154 | #define BTOPRT(b) (((unsigned int)(b) >> PRSHIFT)) | |
155 | ||
156 | STATIC unsigned long | |
157 | page_region_mask( | |
158 | size_t offset, | |
159 | size_t length) | |
160 | { | |
161 | unsigned long mask; | |
162 | int first, final; | |
163 | ||
164 | first = BTOPR(offset); | |
165 | final = BTOPRT(offset + length - 1); | |
166 | first = min(first, final); | |
167 | ||
168 | mask = ~0UL; | |
169 | mask <<= BITS_PER_LONG - (final - first); | |
170 | mask >>= BITS_PER_LONG - (final); | |
171 | ||
172 | ASSERT(offset + length <= PAGE_CACHE_SIZE); | |
173 | ASSERT((final - first) < BITS_PER_LONG && (final - first) >= 0); | |
174 | ||
175 | return mask; | |
176 | } | |
177 | ||
178 | STATIC inline void | |
179 | set_page_region( | |
180 | struct page *page, | |
181 | size_t offset, | |
182 | size_t length) | |
183 | { | |
4c21e2f2 HD |
184 | set_page_private(page, |
185 | page_private(page) | page_region_mask(offset, length)); | |
186 | if (page_private(page) == ~0UL) | |
1da177e4 LT |
187 | SetPageUptodate(page); |
188 | } | |
189 | ||
190 | STATIC inline int | |
191 | test_page_region( | |
192 | struct page *page, | |
193 | size_t offset, | |
194 | size_t length) | |
195 | { | |
196 | unsigned long mask = page_region_mask(offset, length); | |
197 | ||
4c21e2f2 | 198 | return (mask && (page_private(page) & mask) == mask); |
1da177e4 LT |
199 | } |
200 | ||
201 | /* | |
202 | * Mapping of multi-page buffers into contiguous virtual space | |
203 | */ | |
204 | ||
205 | typedef struct a_list { | |
206 | void *vm_addr; | |
207 | struct a_list *next; | |
208 | } a_list_t; | |
209 | ||
210 | STATIC a_list_t *as_free_head; | |
211 | STATIC int as_list_len; | |
212 | STATIC DEFINE_SPINLOCK(as_lock); | |
213 | ||
214 | /* | |
215 | * Try to batch vunmaps because they are costly. | |
216 | */ | |
217 | STATIC void | |
218 | free_address( | |
219 | void *addr) | |
220 | { | |
221 | a_list_t *aentry; | |
222 | ||
223 | aentry = kmalloc(sizeof(a_list_t), GFP_ATOMIC & ~__GFP_HIGH); | |
224 | if (likely(aentry)) { | |
225 | spin_lock(&as_lock); | |
226 | aentry->next = as_free_head; | |
227 | aentry->vm_addr = addr; | |
228 | as_free_head = aentry; | |
229 | as_list_len++; | |
230 | spin_unlock(&as_lock); | |
231 | } else { | |
232 | vunmap(addr); | |
233 | } | |
234 | } | |
235 | ||
236 | STATIC void | |
237 | purge_addresses(void) | |
238 | { | |
239 | a_list_t *aentry, *old; | |
240 | ||
241 | if (as_free_head == NULL) | |
242 | return; | |
243 | ||
244 | spin_lock(&as_lock); | |
245 | aentry = as_free_head; | |
246 | as_free_head = NULL; | |
247 | as_list_len = 0; | |
248 | spin_unlock(&as_lock); | |
249 | ||
250 | while ((old = aentry) != NULL) { | |
251 | vunmap(aentry->vm_addr); | |
252 | aentry = aentry->next; | |
253 | kfree(old); | |
254 | } | |
255 | } | |
256 | ||
257 | /* | |
258 | * Internal pagebuf object manipulation | |
259 | */ | |
260 | ||
261 | STATIC void | |
262 | _pagebuf_initialize( | |
263 | xfs_buf_t *pb, | |
264 | xfs_buftarg_t *target, | |
265 | loff_t range_base, | |
266 | size_t range_length, | |
267 | page_buf_flags_t flags) | |
268 | { | |
269 | /* | |
270 | * We don't want certain flags to appear in pb->pb_flags. | |
271 | */ | |
272 | flags &= ~(PBF_LOCK|PBF_MAPPED|PBF_DONT_BLOCK|PBF_READ_AHEAD); | |
273 | ||
274 | memset(pb, 0, sizeof(xfs_buf_t)); | |
275 | atomic_set(&pb->pb_hold, 1); | |
276 | init_MUTEX_LOCKED(&pb->pb_iodonesema); | |
277 | INIT_LIST_HEAD(&pb->pb_list); | |
278 | INIT_LIST_HEAD(&pb->pb_hash_list); | |
279 | init_MUTEX_LOCKED(&pb->pb_sema); /* held, no waiters */ | |
280 | PB_SET_OWNER(pb); | |
281 | pb->pb_target = target; | |
282 | pb->pb_file_offset = range_base; | |
283 | /* | |
284 | * Set buffer_length and count_desired to the same value initially. | |
285 | * I/O routines should use count_desired, which will be the same in | |
286 | * most cases but may be reset (e.g. XFS recovery). | |
287 | */ | |
288 | pb->pb_buffer_length = pb->pb_count_desired = range_length; | |
289 | pb->pb_flags = flags | PBF_NONE; | |
290 | pb->pb_bn = XFS_BUF_DADDR_NULL; | |
291 | atomic_set(&pb->pb_pin_count, 0); | |
292 | init_waitqueue_head(&pb->pb_waiters); | |
293 | ||
294 | XFS_STATS_INC(pb_create); | |
295 | PB_TRACE(pb, "initialize", target); | |
296 | } | |
297 | ||
298 | /* | |
299 | * Allocate a page array capable of holding a specified number | |
300 | * of pages, and point the page buf at it. | |
301 | */ | |
302 | STATIC int | |
303 | _pagebuf_get_pages( | |
304 | xfs_buf_t *pb, | |
305 | int page_count, | |
306 | page_buf_flags_t flags) | |
307 | { | |
308 | /* Make sure that we have a page list */ | |
309 | if (pb->pb_pages == NULL) { | |
310 | pb->pb_offset = page_buf_poff(pb->pb_file_offset); | |
311 | pb->pb_page_count = page_count; | |
312 | if (page_count <= PB_PAGES) { | |
313 | pb->pb_pages = pb->pb_page_array; | |
314 | } else { | |
315 | pb->pb_pages = kmem_alloc(sizeof(struct page *) * | |
316 | page_count, pb_to_km(flags)); | |
317 | if (pb->pb_pages == NULL) | |
318 | return -ENOMEM; | |
319 | } | |
320 | memset(pb->pb_pages, 0, sizeof(struct page *) * page_count); | |
321 | } | |
322 | return 0; | |
323 | } | |
324 | ||
325 | /* | |
326 | * Frees pb_pages if it was malloced. | |
327 | */ | |
328 | STATIC void | |
329 | _pagebuf_free_pages( | |
330 | xfs_buf_t *bp) | |
331 | { | |
332 | if (bp->pb_pages != bp->pb_page_array) { | |
333 | kmem_free(bp->pb_pages, | |
334 | bp->pb_page_count * sizeof(struct page *)); | |
335 | } | |
336 | } | |
337 | ||
338 | /* | |
339 | * Releases the specified buffer. | |
340 | * | |
341 | * The modification state of any associated pages is left unchanged. | |
342 | * The buffer most not be on any hash - use pagebuf_rele instead for | |
343 | * hashed and refcounted buffers | |
344 | */ | |
345 | void | |
346 | pagebuf_free( | |
347 | xfs_buf_t *bp) | |
348 | { | |
349 | PB_TRACE(bp, "free", 0); | |
350 | ||
351 | ASSERT(list_empty(&bp->pb_hash_list)); | |
352 | ||
353 | if (bp->pb_flags & _PBF_PAGE_CACHE) { | |
354 | uint i; | |
355 | ||
356 | if ((bp->pb_flags & PBF_MAPPED) && (bp->pb_page_count > 1)) | |
357 | free_address(bp->pb_addr - bp->pb_offset); | |
358 | ||
359 | for (i = 0; i < bp->pb_page_count; i++) | |
360 | page_cache_release(bp->pb_pages[i]); | |
361 | _pagebuf_free_pages(bp); | |
362 | } else if (bp->pb_flags & _PBF_KMEM_ALLOC) { | |
363 | /* | |
364 | * XXX(hch): bp->pb_count_desired might be incorrect (see | |
365 | * pagebuf_associate_memory for details), but fortunately | |
366 | * the Linux version of kmem_free ignores the len argument.. | |
367 | */ | |
368 | kmem_free(bp->pb_addr, bp->pb_count_desired); | |
369 | _pagebuf_free_pages(bp); | |
370 | } | |
371 | ||
372 | pagebuf_deallocate(bp); | |
373 | } | |
374 | ||
375 | /* | |
376 | * Finds all pages for buffer in question and builds it's page list. | |
377 | */ | |
378 | STATIC int | |
379 | _pagebuf_lookup_pages( | |
380 | xfs_buf_t *bp, | |
381 | uint flags) | |
382 | { | |
383 | struct address_space *mapping = bp->pb_target->pbr_mapping; | |
384 | size_t blocksize = bp->pb_target->pbr_bsize; | |
385 | size_t size = bp->pb_count_desired; | |
386 | size_t nbytes, offset; | |
27496a8c | 387 | gfp_t gfp_mask = pb_to_gfp(flags); |
1da177e4 LT |
388 | unsigned short page_count, i; |
389 | pgoff_t first; | |
390 | loff_t end; | |
391 | int error; | |
392 | ||
393 | end = bp->pb_file_offset + bp->pb_buffer_length; | |
394 | page_count = page_buf_btoc(end) - page_buf_btoct(bp->pb_file_offset); | |
395 | ||
396 | error = _pagebuf_get_pages(bp, page_count, flags); | |
397 | if (unlikely(error)) | |
398 | return error; | |
399 | bp->pb_flags |= _PBF_PAGE_CACHE; | |
400 | ||
401 | offset = bp->pb_offset; | |
402 | first = bp->pb_file_offset >> PAGE_CACHE_SHIFT; | |
403 | ||
404 | for (i = 0; i < bp->pb_page_count; i++) { | |
405 | struct page *page; | |
406 | uint retries = 0; | |
407 | ||
408 | retry: | |
409 | page = find_or_create_page(mapping, first + i, gfp_mask); | |
410 | if (unlikely(page == NULL)) { | |
411 | if (flags & PBF_READ_AHEAD) { | |
412 | bp->pb_page_count = i; | |
413 | for (i = 0; i < bp->pb_page_count; i++) | |
414 | unlock_page(bp->pb_pages[i]); | |
415 | return -ENOMEM; | |
416 | } | |
417 | ||
418 | /* | |
419 | * This could deadlock. | |
420 | * | |
421 | * But until all the XFS lowlevel code is revamped to | |
422 | * handle buffer allocation failures we can't do much. | |
423 | */ | |
424 | if (!(++retries % 100)) | |
425 | printk(KERN_ERR | |
426 | "XFS: possible memory allocation " | |
427 | "deadlock in %s (mode:0x%x)\n", | |
428 | __FUNCTION__, gfp_mask); | |
429 | ||
430 | XFS_STATS_INC(pb_page_retries); | |
23ea4032 | 431 | xfsbufd_wakeup(0, gfp_mask); |
1da177e4 LT |
432 | blk_congestion_wait(WRITE, HZ/50); |
433 | goto retry; | |
434 | } | |
435 | ||
436 | XFS_STATS_INC(pb_page_found); | |
437 | ||
438 | nbytes = min_t(size_t, size, PAGE_CACHE_SIZE - offset); | |
439 | size -= nbytes; | |
440 | ||
441 | if (!PageUptodate(page)) { | |
442 | page_count--; | |
443 | if (blocksize >= PAGE_CACHE_SIZE) { | |
444 | if (flags & PBF_READ) | |
445 | bp->pb_locked = 1; | |
446 | } else if (!PagePrivate(page)) { | |
447 | if (test_page_region(page, offset, nbytes)) | |
448 | page_count++; | |
449 | } | |
450 | } | |
451 | ||
452 | bp->pb_pages[i] = page; | |
453 | offset = 0; | |
454 | } | |
455 | ||
456 | if (!bp->pb_locked) { | |
457 | for (i = 0; i < bp->pb_page_count; i++) | |
458 | unlock_page(bp->pb_pages[i]); | |
459 | } | |
460 | ||
461 | if (page_count) { | |
462 | /* if we have any uptodate pages, mark that in the buffer */ | |
463 | bp->pb_flags &= ~PBF_NONE; | |
464 | ||
465 | /* if some pages aren't uptodate, mark that in the buffer */ | |
466 | if (page_count != bp->pb_page_count) | |
467 | bp->pb_flags |= PBF_PARTIAL; | |
468 | } | |
469 | ||
470 | PB_TRACE(bp, "lookup_pages", (long)page_count); | |
471 | return error; | |
472 | } | |
473 | ||
474 | /* | |
475 | * Map buffer into kernel address-space if nessecary. | |
476 | */ | |
477 | STATIC int | |
478 | _pagebuf_map_pages( | |
479 | xfs_buf_t *bp, | |
480 | uint flags) | |
481 | { | |
482 | /* A single page buffer is always mappable */ | |
483 | if (bp->pb_page_count == 1) { | |
484 | bp->pb_addr = page_address(bp->pb_pages[0]) + bp->pb_offset; | |
485 | bp->pb_flags |= PBF_MAPPED; | |
486 | } else if (flags & PBF_MAPPED) { | |
487 | if (as_list_len > 64) | |
488 | purge_addresses(); | |
489 | bp->pb_addr = vmap(bp->pb_pages, bp->pb_page_count, | |
490 | VM_MAP, PAGE_KERNEL); | |
491 | if (unlikely(bp->pb_addr == NULL)) | |
492 | return -ENOMEM; | |
493 | bp->pb_addr += bp->pb_offset; | |
494 | bp->pb_flags |= PBF_MAPPED; | |
495 | } | |
496 | ||
497 | return 0; | |
498 | } | |
499 | ||
500 | /* | |
501 | * Finding and Reading Buffers | |
502 | */ | |
503 | ||
504 | /* | |
505 | * _pagebuf_find | |
506 | * | |
507 | * Looks up, and creates if absent, a lockable buffer for | |
508 | * a given range of an inode. The buffer is returned | |
509 | * locked. If other overlapping buffers exist, they are | |
510 | * released before the new buffer is created and locked, | |
511 | * which may imply that this call will block until those buffers | |
512 | * are unlocked. No I/O is implied by this call. | |
513 | */ | |
514 | xfs_buf_t * | |
515 | _pagebuf_find( | |
516 | xfs_buftarg_t *btp, /* block device target */ | |
517 | loff_t ioff, /* starting offset of range */ | |
518 | size_t isize, /* length of range */ | |
519 | page_buf_flags_t flags, /* PBF_TRYLOCK */ | |
520 | xfs_buf_t *new_pb)/* newly allocated buffer */ | |
521 | { | |
522 | loff_t range_base; | |
523 | size_t range_length; | |
524 | xfs_bufhash_t *hash; | |
525 | xfs_buf_t *pb, *n; | |
526 | ||
527 | range_base = (ioff << BBSHIFT); | |
528 | range_length = (isize << BBSHIFT); | |
529 | ||
530 | /* Check for IOs smaller than the sector size / not sector aligned */ | |
531 | ASSERT(!(range_length < (1 << btp->pbr_sshift))); | |
532 | ASSERT(!(range_base & (loff_t)btp->pbr_smask)); | |
533 | ||
534 | hash = &btp->bt_hash[hash_long((unsigned long)ioff, btp->bt_hashshift)]; | |
535 | ||
536 | spin_lock(&hash->bh_lock); | |
537 | ||
538 | list_for_each_entry_safe(pb, n, &hash->bh_list, pb_hash_list) { | |
539 | ASSERT(btp == pb->pb_target); | |
540 | if (pb->pb_file_offset == range_base && | |
541 | pb->pb_buffer_length == range_length) { | |
542 | /* | |
543 | * If we look at something bring it to the | |
544 | * front of the list for next time. | |
545 | */ | |
546 | atomic_inc(&pb->pb_hold); | |
547 | list_move(&pb->pb_hash_list, &hash->bh_list); | |
548 | goto found; | |
549 | } | |
550 | } | |
551 | ||
552 | /* No match found */ | |
553 | if (new_pb) { | |
554 | _pagebuf_initialize(new_pb, btp, range_base, | |
555 | range_length, flags); | |
556 | new_pb->pb_hash = hash; | |
557 | list_add(&new_pb->pb_hash_list, &hash->bh_list); | |
558 | } else { | |
559 | XFS_STATS_INC(pb_miss_locked); | |
560 | } | |
561 | ||
562 | spin_unlock(&hash->bh_lock); | |
563 | return new_pb; | |
564 | ||
565 | found: | |
566 | spin_unlock(&hash->bh_lock); | |
567 | ||
568 | /* Attempt to get the semaphore without sleeping, | |
569 | * if this does not work then we need to drop the | |
570 | * spinlock and do a hard attempt on the semaphore. | |
571 | */ | |
572 | if (down_trylock(&pb->pb_sema)) { | |
573 | if (!(flags & PBF_TRYLOCK)) { | |
574 | /* wait for buffer ownership */ | |
575 | PB_TRACE(pb, "get_lock", 0); | |
576 | pagebuf_lock(pb); | |
577 | XFS_STATS_INC(pb_get_locked_waited); | |
578 | } else { | |
579 | /* We asked for a trylock and failed, no need | |
580 | * to look at file offset and length here, we | |
581 | * know that this pagebuf at least overlaps our | |
582 | * pagebuf and is locked, therefore our buffer | |
583 | * either does not exist, or is this buffer | |
584 | */ | |
585 | ||
586 | pagebuf_rele(pb); | |
587 | XFS_STATS_INC(pb_busy_locked); | |
588 | return (NULL); | |
589 | } | |
590 | } else { | |
591 | /* trylock worked */ | |
592 | PB_SET_OWNER(pb); | |
593 | } | |
594 | ||
2f926587 DC |
595 | if (pb->pb_flags & PBF_STALE) { |
596 | ASSERT((pb->pb_flags & _PBF_DELWRI_Q) == 0); | |
1da177e4 | 597 | pb->pb_flags &= PBF_MAPPED; |
2f926587 | 598 | } |
1da177e4 LT |
599 | PB_TRACE(pb, "got_lock", 0); |
600 | XFS_STATS_INC(pb_get_locked); | |
601 | return (pb); | |
602 | } | |
603 | ||
604 | /* | |
605 | * xfs_buf_get_flags assembles a buffer covering the specified range. | |
606 | * | |
607 | * Storage in memory for all portions of the buffer will be allocated, | |
608 | * although backing storage may not be. | |
609 | */ | |
610 | xfs_buf_t * | |
611 | xfs_buf_get_flags( /* allocate a buffer */ | |
612 | xfs_buftarg_t *target,/* target for buffer */ | |
613 | loff_t ioff, /* starting offset of range */ | |
614 | size_t isize, /* length of range */ | |
615 | page_buf_flags_t flags) /* PBF_TRYLOCK */ | |
616 | { | |
617 | xfs_buf_t *pb, *new_pb; | |
618 | int error = 0, i; | |
619 | ||
620 | new_pb = pagebuf_allocate(flags); | |
621 | if (unlikely(!new_pb)) | |
622 | return NULL; | |
623 | ||
624 | pb = _pagebuf_find(target, ioff, isize, flags, new_pb); | |
625 | if (pb == new_pb) { | |
626 | error = _pagebuf_lookup_pages(pb, flags); | |
627 | if (error) | |
628 | goto no_buffer; | |
629 | } else { | |
630 | pagebuf_deallocate(new_pb); | |
631 | if (unlikely(pb == NULL)) | |
632 | return NULL; | |
633 | } | |
634 | ||
635 | for (i = 0; i < pb->pb_page_count; i++) | |
636 | mark_page_accessed(pb->pb_pages[i]); | |
637 | ||
638 | if (!(pb->pb_flags & PBF_MAPPED)) { | |
639 | error = _pagebuf_map_pages(pb, flags); | |
640 | if (unlikely(error)) { | |
641 | printk(KERN_WARNING "%s: failed to map pages\n", | |
642 | __FUNCTION__); | |
643 | goto no_buffer; | |
644 | } | |
645 | } | |
646 | ||
647 | XFS_STATS_INC(pb_get); | |
648 | ||
649 | /* | |
650 | * Always fill in the block number now, the mapped cases can do | |
651 | * their own overlay of this later. | |
652 | */ | |
653 | pb->pb_bn = ioff; | |
654 | pb->pb_count_desired = pb->pb_buffer_length; | |
655 | ||
656 | PB_TRACE(pb, "get", (unsigned long)flags); | |
657 | return pb; | |
658 | ||
659 | no_buffer: | |
660 | if (flags & (PBF_LOCK | PBF_TRYLOCK)) | |
661 | pagebuf_unlock(pb); | |
662 | pagebuf_rele(pb); | |
663 | return NULL; | |
664 | } | |
665 | ||
666 | xfs_buf_t * | |
667 | xfs_buf_read_flags( | |
668 | xfs_buftarg_t *target, | |
669 | loff_t ioff, | |
670 | size_t isize, | |
671 | page_buf_flags_t flags) | |
672 | { | |
673 | xfs_buf_t *pb; | |
674 | ||
675 | flags |= PBF_READ; | |
676 | ||
677 | pb = xfs_buf_get_flags(target, ioff, isize, flags); | |
678 | if (pb) { | |
679 | if (PBF_NOT_DONE(pb)) { | |
680 | PB_TRACE(pb, "read", (unsigned long)flags); | |
681 | XFS_STATS_INC(pb_get_read); | |
682 | pagebuf_iostart(pb, flags); | |
683 | } else if (flags & PBF_ASYNC) { | |
684 | PB_TRACE(pb, "read_async", (unsigned long)flags); | |
685 | /* | |
686 | * Read ahead call which is already satisfied, | |
687 | * drop the buffer | |
688 | */ | |
689 | goto no_buffer; | |
690 | } else { | |
691 | PB_TRACE(pb, "read_done", (unsigned long)flags); | |
692 | /* We do not want read in the flags */ | |
693 | pb->pb_flags &= ~PBF_READ; | |
694 | } | |
695 | } | |
696 | ||
697 | return pb; | |
698 | ||
699 | no_buffer: | |
700 | if (flags & (PBF_LOCK | PBF_TRYLOCK)) | |
701 | pagebuf_unlock(pb); | |
702 | pagebuf_rele(pb); | |
703 | return NULL; | |
704 | } | |
705 | ||
1da177e4 LT |
706 | /* |
707 | * If we are not low on memory then do the readahead in a deadlock | |
708 | * safe manner. | |
709 | */ | |
710 | void | |
711 | pagebuf_readahead( | |
712 | xfs_buftarg_t *target, | |
713 | loff_t ioff, | |
714 | size_t isize, | |
715 | page_buf_flags_t flags) | |
716 | { | |
717 | struct backing_dev_info *bdi; | |
718 | ||
719 | bdi = target->pbr_mapping->backing_dev_info; | |
720 | if (bdi_read_congested(bdi)) | |
721 | return; | |
722 | ||
723 | flags |= (PBF_TRYLOCK|PBF_ASYNC|PBF_READ_AHEAD); | |
724 | xfs_buf_read_flags(target, ioff, isize, flags); | |
725 | } | |
726 | ||
727 | xfs_buf_t * | |
728 | pagebuf_get_empty( | |
729 | size_t len, | |
730 | xfs_buftarg_t *target) | |
731 | { | |
732 | xfs_buf_t *pb; | |
733 | ||
734 | pb = pagebuf_allocate(0); | |
735 | if (pb) | |
736 | _pagebuf_initialize(pb, target, 0, len, 0); | |
737 | return pb; | |
738 | } | |
739 | ||
740 | static inline struct page * | |
741 | mem_to_page( | |
742 | void *addr) | |
743 | { | |
744 | if (((unsigned long)addr < VMALLOC_START) || | |
745 | ((unsigned long)addr >= VMALLOC_END)) { | |
746 | return virt_to_page(addr); | |
747 | } else { | |
748 | return vmalloc_to_page(addr); | |
749 | } | |
750 | } | |
751 | ||
752 | int | |
753 | pagebuf_associate_memory( | |
754 | xfs_buf_t *pb, | |
755 | void *mem, | |
756 | size_t len) | |
757 | { | |
758 | int rval; | |
759 | int i = 0; | |
760 | size_t ptr; | |
761 | size_t end, end_cur; | |
762 | off_t offset; | |
763 | int page_count; | |
764 | ||
765 | page_count = PAGE_CACHE_ALIGN(len) >> PAGE_CACHE_SHIFT; | |
766 | offset = (off_t) mem - ((off_t)mem & PAGE_CACHE_MASK); | |
767 | if (offset && (len > PAGE_CACHE_SIZE)) | |
768 | page_count++; | |
769 | ||
770 | /* Free any previous set of page pointers */ | |
771 | if (pb->pb_pages) | |
772 | _pagebuf_free_pages(pb); | |
773 | ||
774 | pb->pb_pages = NULL; | |
775 | pb->pb_addr = mem; | |
776 | ||
777 | rval = _pagebuf_get_pages(pb, page_count, 0); | |
778 | if (rval) | |
779 | return rval; | |
780 | ||
781 | pb->pb_offset = offset; | |
782 | ptr = (size_t) mem & PAGE_CACHE_MASK; | |
783 | end = PAGE_CACHE_ALIGN((size_t) mem + len); | |
784 | end_cur = end; | |
785 | /* set up first page */ | |
786 | pb->pb_pages[0] = mem_to_page(mem); | |
787 | ||
788 | ptr += PAGE_CACHE_SIZE; | |
789 | pb->pb_page_count = ++i; | |
790 | while (ptr < end) { | |
791 | pb->pb_pages[i] = mem_to_page((void *)ptr); | |
792 | pb->pb_page_count = ++i; | |
793 | ptr += PAGE_CACHE_SIZE; | |
794 | } | |
795 | pb->pb_locked = 0; | |
796 | ||
797 | pb->pb_count_desired = pb->pb_buffer_length = len; | |
798 | pb->pb_flags |= PBF_MAPPED; | |
799 | ||
800 | return 0; | |
801 | } | |
802 | ||
803 | xfs_buf_t * | |
804 | pagebuf_get_no_daddr( | |
805 | size_t len, | |
806 | xfs_buftarg_t *target) | |
807 | { | |
808 | size_t malloc_len = len; | |
809 | xfs_buf_t *bp; | |
810 | void *data; | |
811 | int error; | |
812 | ||
813 | bp = pagebuf_allocate(0); | |
814 | if (unlikely(bp == NULL)) | |
815 | goto fail; | |
816 | _pagebuf_initialize(bp, target, 0, len, PBF_FORCEIO); | |
817 | ||
818 | try_again: | |
819 | data = kmem_alloc(malloc_len, KM_SLEEP | KM_MAYFAIL); | |
820 | if (unlikely(data == NULL)) | |
821 | goto fail_free_buf; | |
822 | ||
823 | /* check whether alignment matches.. */ | |
824 | if ((__psunsigned_t)data != | |
825 | ((__psunsigned_t)data & ~target->pbr_smask)) { | |
826 | /* .. else double the size and try again */ | |
827 | kmem_free(data, malloc_len); | |
828 | malloc_len <<= 1; | |
829 | goto try_again; | |
830 | } | |
831 | ||
832 | error = pagebuf_associate_memory(bp, data, len); | |
833 | if (error) | |
834 | goto fail_free_mem; | |
835 | bp->pb_flags |= _PBF_KMEM_ALLOC; | |
836 | ||
837 | pagebuf_unlock(bp); | |
838 | ||
839 | PB_TRACE(bp, "no_daddr", data); | |
840 | return bp; | |
841 | fail_free_mem: | |
842 | kmem_free(data, malloc_len); | |
843 | fail_free_buf: | |
844 | pagebuf_free(bp); | |
845 | fail: | |
846 | return NULL; | |
847 | } | |
848 | ||
849 | /* | |
850 | * pagebuf_hold | |
851 | * | |
852 | * Increment reference count on buffer, to hold the buffer concurrently | |
853 | * with another thread which may release (free) the buffer asynchronously. | |
854 | * | |
855 | * Must hold the buffer already to call this function. | |
856 | */ | |
857 | void | |
858 | pagebuf_hold( | |
859 | xfs_buf_t *pb) | |
860 | { | |
861 | atomic_inc(&pb->pb_hold); | |
862 | PB_TRACE(pb, "hold", 0); | |
863 | } | |
864 | ||
865 | /* | |
866 | * pagebuf_rele | |
867 | * | |
868 | * pagebuf_rele releases a hold on the specified buffer. If the | |
869 | * the hold count is 1, pagebuf_rele calls pagebuf_free. | |
870 | */ | |
871 | void | |
872 | pagebuf_rele( | |
873 | xfs_buf_t *pb) | |
874 | { | |
875 | xfs_bufhash_t *hash = pb->pb_hash; | |
876 | ||
877 | PB_TRACE(pb, "rele", pb->pb_relse); | |
878 | ||
2f926587 DC |
879 | /* |
880 | * pagebuf_lookup buffers are not hashed, not delayed write, | |
881 | * and don't have their own release routines. Special case. | |
882 | */ | |
883 | if (unlikely(!hash)) { | |
884 | ASSERT(!pb->pb_relse); | |
885 | if (atomic_dec_and_test(&pb->pb_hold)) | |
886 | xfs_buf_free(pb); | |
887 | return; | |
888 | } | |
889 | ||
1da177e4 LT |
890 | if (atomic_dec_and_lock(&pb->pb_hold, &hash->bh_lock)) { |
891 | int do_free = 1; | |
892 | ||
893 | if (pb->pb_relse) { | |
894 | atomic_inc(&pb->pb_hold); | |
895 | spin_unlock(&hash->bh_lock); | |
896 | (*(pb->pb_relse)) (pb); | |
897 | spin_lock(&hash->bh_lock); | |
898 | do_free = 0; | |
899 | } | |
900 | ||
2f926587 | 901 | if (pb->pb_flags & PBF_FS_MANAGED) { |
1da177e4 LT |
902 | do_free = 0; |
903 | } | |
904 | ||
905 | if (do_free) { | |
2f926587 | 906 | ASSERT((pb->pb_flags & (PBF_DELWRI|_PBF_DELWRI_Q)) == 0); |
1da177e4 LT |
907 | list_del_init(&pb->pb_hash_list); |
908 | spin_unlock(&hash->bh_lock); | |
909 | pagebuf_free(pb); | |
910 | } else { | |
911 | spin_unlock(&hash->bh_lock); | |
912 | } | |
2f926587 DC |
913 | } else { |
914 | /* | |
915 | * Catch reference count leaks | |
916 | */ | |
917 | ASSERT(atomic_read(&pb->pb_hold) >= 0); | |
1da177e4 LT |
918 | } |
919 | } | |
920 | ||
921 | ||
922 | /* | |
923 | * Mutual exclusion on buffers. Locking model: | |
924 | * | |
925 | * Buffers associated with inodes for which buffer locking | |
926 | * is not enabled are not protected by semaphores, and are | |
927 | * assumed to be exclusively owned by the caller. There is a | |
928 | * spinlock in the buffer, used by the caller when concurrent | |
929 | * access is possible. | |
930 | */ | |
931 | ||
932 | /* | |
933 | * pagebuf_cond_lock | |
934 | * | |
935 | * pagebuf_cond_lock locks a buffer object, if it is not already locked. | |
936 | * Note that this in no way | |
937 | * locks the underlying pages, so it is only useful for synchronizing | |
938 | * concurrent use of page buffer objects, not for synchronizing independent | |
939 | * access to the underlying pages. | |
940 | */ | |
941 | int | |
942 | pagebuf_cond_lock( /* lock buffer, if not locked */ | |
943 | /* returns -EBUSY if locked) */ | |
944 | xfs_buf_t *pb) | |
945 | { | |
946 | int locked; | |
947 | ||
948 | locked = down_trylock(&pb->pb_sema) == 0; | |
949 | if (locked) { | |
950 | PB_SET_OWNER(pb); | |
951 | } | |
952 | PB_TRACE(pb, "cond_lock", (long)locked); | |
953 | return(locked ? 0 : -EBUSY); | |
954 | } | |
955 | ||
956 | #if defined(DEBUG) || defined(XFS_BLI_TRACE) | |
957 | /* | |
958 | * pagebuf_lock_value | |
959 | * | |
960 | * Return lock value for a pagebuf | |
961 | */ | |
962 | int | |
963 | pagebuf_lock_value( | |
964 | xfs_buf_t *pb) | |
965 | { | |
966 | return(atomic_read(&pb->pb_sema.count)); | |
967 | } | |
968 | #endif | |
969 | ||
970 | /* | |
971 | * pagebuf_lock | |
972 | * | |
973 | * pagebuf_lock locks a buffer object. Note that this in no way | |
974 | * locks the underlying pages, so it is only useful for synchronizing | |
975 | * concurrent use of page buffer objects, not for synchronizing independent | |
976 | * access to the underlying pages. | |
977 | */ | |
978 | int | |
979 | pagebuf_lock( | |
980 | xfs_buf_t *pb) | |
981 | { | |
982 | PB_TRACE(pb, "lock", 0); | |
983 | if (atomic_read(&pb->pb_io_remaining)) | |
984 | blk_run_address_space(pb->pb_target->pbr_mapping); | |
985 | down(&pb->pb_sema); | |
986 | PB_SET_OWNER(pb); | |
987 | PB_TRACE(pb, "locked", 0); | |
988 | return 0; | |
989 | } | |
990 | ||
991 | /* | |
992 | * pagebuf_unlock | |
993 | * | |
994 | * pagebuf_unlock releases the lock on the buffer object created by | |
2f926587 DC |
995 | * pagebuf_lock or pagebuf_cond_lock (not any pinning of underlying pages |
996 | * created by pagebuf_pin). | |
997 | * | |
998 | * If the buffer is marked delwri but is not queued, do so before we | |
999 | * unlock the buffer as we need to set flags correctly. We also need to | |
1000 | * take a reference for the delwri queue because the unlocker is going to | |
1001 | * drop their's and they don't know we just queued it. | |
1da177e4 LT |
1002 | */ |
1003 | void | |
1004 | pagebuf_unlock( /* unlock buffer */ | |
1005 | xfs_buf_t *pb) /* buffer to unlock */ | |
1006 | { | |
2f926587 DC |
1007 | if ((pb->pb_flags & (PBF_DELWRI|_PBF_DELWRI_Q)) == PBF_DELWRI) { |
1008 | atomic_inc(&pb->pb_hold); | |
1009 | pb->pb_flags |= PBF_ASYNC; | |
1010 | pagebuf_delwri_queue(pb, 0); | |
1011 | } | |
1012 | ||
1da177e4 LT |
1013 | PB_CLEAR_OWNER(pb); |
1014 | up(&pb->pb_sema); | |
1015 | PB_TRACE(pb, "unlock", 0); | |
1016 | } | |
1017 | ||
1018 | ||
1019 | /* | |
1020 | * Pinning Buffer Storage in Memory | |
1021 | */ | |
1022 | ||
1023 | /* | |
1024 | * pagebuf_pin | |
1025 | * | |
1026 | * pagebuf_pin locks all of the memory represented by a buffer in | |
1027 | * memory. Multiple calls to pagebuf_pin and pagebuf_unpin, for | |
1028 | * the same or different buffers affecting a given page, will | |
1029 | * properly count the number of outstanding "pin" requests. The | |
1030 | * buffer may be released after the pagebuf_pin and a different | |
1031 | * buffer used when calling pagebuf_unpin, if desired. | |
1032 | * pagebuf_pin should be used by the file system when it wants be | |
1033 | * assured that no attempt will be made to force the affected | |
1034 | * memory to disk. It does not assure that a given logical page | |
1035 | * will not be moved to a different physical page. | |
1036 | */ | |
1037 | void | |
1038 | pagebuf_pin( | |
1039 | xfs_buf_t *pb) | |
1040 | { | |
1041 | atomic_inc(&pb->pb_pin_count); | |
1042 | PB_TRACE(pb, "pin", (long)pb->pb_pin_count.counter); | |
1043 | } | |
1044 | ||
1045 | /* | |
1046 | * pagebuf_unpin | |
1047 | * | |
1048 | * pagebuf_unpin reverses the locking of memory performed by | |
1049 | * pagebuf_pin. Note that both functions affected the logical | |
1050 | * pages associated with the buffer, not the buffer itself. | |
1051 | */ | |
1052 | void | |
1053 | pagebuf_unpin( | |
1054 | xfs_buf_t *pb) | |
1055 | { | |
1056 | if (atomic_dec_and_test(&pb->pb_pin_count)) { | |
1057 | wake_up_all(&pb->pb_waiters); | |
1058 | } | |
1059 | PB_TRACE(pb, "unpin", (long)pb->pb_pin_count.counter); | |
1060 | } | |
1061 | ||
1062 | int | |
1063 | pagebuf_ispin( | |
1064 | xfs_buf_t *pb) | |
1065 | { | |
1066 | return atomic_read(&pb->pb_pin_count); | |
1067 | } | |
1068 | ||
1069 | /* | |
1070 | * pagebuf_wait_unpin | |
1071 | * | |
1072 | * pagebuf_wait_unpin waits until all of the memory associated | |
1073 | * with the buffer is not longer locked in memory. It returns | |
1074 | * immediately if none of the affected pages are locked. | |
1075 | */ | |
1076 | static inline void | |
1077 | _pagebuf_wait_unpin( | |
1078 | xfs_buf_t *pb) | |
1079 | { | |
1080 | DECLARE_WAITQUEUE (wait, current); | |
1081 | ||
1082 | if (atomic_read(&pb->pb_pin_count) == 0) | |
1083 | return; | |
1084 | ||
1085 | add_wait_queue(&pb->pb_waiters, &wait); | |
1086 | for (;;) { | |
1087 | set_current_state(TASK_UNINTERRUPTIBLE); | |
1088 | if (atomic_read(&pb->pb_pin_count) == 0) | |
1089 | break; | |
1090 | if (atomic_read(&pb->pb_io_remaining)) | |
1091 | blk_run_address_space(pb->pb_target->pbr_mapping); | |
1092 | schedule(); | |
1093 | } | |
1094 | remove_wait_queue(&pb->pb_waiters, &wait); | |
1095 | set_current_state(TASK_RUNNING); | |
1096 | } | |
1097 | ||
1098 | /* | |
1099 | * Buffer Utility Routines | |
1100 | */ | |
1101 | ||
1102 | /* | |
1103 | * pagebuf_iodone | |
1104 | * | |
1105 | * pagebuf_iodone marks a buffer for which I/O is in progress | |
1106 | * done with respect to that I/O. The pb_iodone routine, if | |
1107 | * present, will be called as a side-effect. | |
1108 | */ | |
1109 | STATIC void | |
1110 | pagebuf_iodone_work( | |
1111 | void *v) | |
1112 | { | |
1113 | xfs_buf_t *bp = (xfs_buf_t *)v; | |
1114 | ||
1115 | if (bp->pb_iodone) | |
1116 | (*(bp->pb_iodone))(bp); | |
1117 | else if (bp->pb_flags & PBF_ASYNC) | |
1118 | xfs_buf_relse(bp); | |
1119 | } | |
1120 | ||
1121 | void | |
1122 | pagebuf_iodone( | |
1123 | xfs_buf_t *pb, | |
1124 | int dataio, | |
1125 | int schedule) | |
1126 | { | |
1127 | pb->pb_flags &= ~(PBF_READ | PBF_WRITE); | |
1128 | if (pb->pb_error == 0) { | |
1129 | pb->pb_flags &= ~(PBF_PARTIAL | PBF_NONE); | |
1130 | } | |
1131 | ||
1132 | PB_TRACE(pb, "iodone", pb->pb_iodone); | |
1133 | ||
1134 | if ((pb->pb_iodone) || (pb->pb_flags & PBF_ASYNC)) { | |
1135 | if (schedule) { | |
1136 | INIT_WORK(&pb->pb_iodone_work, pagebuf_iodone_work, pb); | |
23ea4032 CH |
1137 | queue_work(dataio ? xfsdatad_workqueue : |
1138 | xfslogd_workqueue, &pb->pb_iodone_work); | |
1da177e4 LT |
1139 | } else { |
1140 | pagebuf_iodone_work(pb); | |
1141 | } | |
1142 | } else { | |
1143 | up(&pb->pb_iodonesema); | |
1144 | } | |
1145 | } | |
1146 | ||
1147 | /* | |
1148 | * pagebuf_ioerror | |
1149 | * | |
1150 | * pagebuf_ioerror sets the error code for a buffer. | |
1151 | */ | |
1152 | void | |
1153 | pagebuf_ioerror( /* mark/clear buffer error flag */ | |
1154 | xfs_buf_t *pb, /* buffer to mark */ | |
1155 | int error) /* error to store (0 if none) */ | |
1156 | { | |
1157 | ASSERT(error >= 0 && error <= 0xffff); | |
1158 | pb->pb_error = (unsigned short)error; | |
1159 | PB_TRACE(pb, "ioerror", (unsigned long)error); | |
1160 | } | |
1161 | ||
1162 | /* | |
1163 | * pagebuf_iostart | |
1164 | * | |
1165 | * pagebuf_iostart initiates I/O on a buffer, based on the flags supplied. | |
1166 | * If necessary, it will arrange for any disk space allocation required, | |
1167 | * and it will break up the request if the block mappings require it. | |
1168 | * The pb_iodone routine in the buffer supplied will only be called | |
1169 | * when all of the subsidiary I/O requests, if any, have been completed. | |
1170 | * pagebuf_iostart calls the pagebuf_ioinitiate routine or | |
1171 | * pagebuf_iorequest, if the former routine is not defined, to start | |
1172 | * the I/O on a given low-level request. | |
1173 | */ | |
1174 | int | |
1175 | pagebuf_iostart( /* start I/O on a buffer */ | |
1176 | xfs_buf_t *pb, /* buffer to start */ | |
1177 | page_buf_flags_t flags) /* PBF_LOCK, PBF_ASYNC, PBF_READ, */ | |
1178 | /* PBF_WRITE, PBF_DELWRI, */ | |
1179 | /* PBF_DONT_BLOCK */ | |
1180 | { | |
1181 | int status = 0; | |
1182 | ||
1183 | PB_TRACE(pb, "iostart", (unsigned long)flags); | |
1184 | ||
1185 | if (flags & PBF_DELWRI) { | |
1186 | pb->pb_flags &= ~(PBF_READ | PBF_WRITE | PBF_ASYNC); | |
1187 | pb->pb_flags |= flags & (PBF_DELWRI | PBF_ASYNC); | |
1188 | pagebuf_delwri_queue(pb, 1); | |
1189 | return status; | |
1190 | } | |
1191 | ||
1192 | pb->pb_flags &= ~(PBF_READ | PBF_WRITE | PBF_ASYNC | PBF_DELWRI | \ | |
1193 | PBF_READ_AHEAD | _PBF_RUN_QUEUES); | |
1194 | pb->pb_flags |= flags & (PBF_READ | PBF_WRITE | PBF_ASYNC | \ | |
1195 | PBF_READ_AHEAD | _PBF_RUN_QUEUES); | |
1196 | ||
1197 | BUG_ON(pb->pb_bn == XFS_BUF_DADDR_NULL); | |
1198 | ||
1199 | /* For writes allow an alternate strategy routine to precede | |
1200 | * the actual I/O request (which may not be issued at all in | |
1201 | * a shutdown situation, for example). | |
1202 | */ | |
1203 | status = (flags & PBF_WRITE) ? | |
1204 | pagebuf_iostrategy(pb) : pagebuf_iorequest(pb); | |
1205 | ||
1206 | /* Wait for I/O if we are not an async request. | |
1207 | * Note: async I/O request completion will release the buffer, | |
1208 | * and that can already be done by this point. So using the | |
1209 | * buffer pointer from here on, after async I/O, is invalid. | |
1210 | */ | |
1211 | if (!status && !(flags & PBF_ASYNC)) | |
1212 | status = pagebuf_iowait(pb); | |
1213 | ||
1214 | return status; | |
1215 | } | |
1216 | ||
1217 | /* | |
1218 | * Helper routine for pagebuf_iorequest | |
1219 | */ | |
1220 | ||
1221 | STATIC __inline__ int | |
1222 | _pagebuf_iolocked( | |
1223 | xfs_buf_t *pb) | |
1224 | { | |
1225 | ASSERT(pb->pb_flags & (PBF_READ|PBF_WRITE)); | |
1226 | if (pb->pb_flags & PBF_READ) | |
1227 | return pb->pb_locked; | |
1228 | return 0; | |
1229 | } | |
1230 | ||
1231 | STATIC __inline__ void | |
1232 | _pagebuf_iodone( | |
1233 | xfs_buf_t *pb, | |
1234 | int schedule) | |
1235 | { | |
1236 | if (atomic_dec_and_test(&pb->pb_io_remaining) == 1) { | |
1237 | pb->pb_locked = 0; | |
1238 | pagebuf_iodone(pb, (pb->pb_flags & PBF_FS_DATAIOD), schedule); | |
1239 | } | |
1240 | } | |
1241 | ||
1242 | STATIC int | |
1243 | bio_end_io_pagebuf( | |
1244 | struct bio *bio, | |
1245 | unsigned int bytes_done, | |
1246 | int error) | |
1247 | { | |
1248 | xfs_buf_t *pb = (xfs_buf_t *)bio->bi_private; | |
eedb5530 NS |
1249 | unsigned int blocksize = pb->pb_target->pbr_bsize; |
1250 | struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1; | |
1da177e4 LT |
1251 | |
1252 | if (bio->bi_size) | |
1253 | return 1; | |
1254 | ||
1255 | if (!test_bit(BIO_UPTODATE, &bio->bi_flags)) | |
1256 | pb->pb_error = EIO; | |
1257 | ||
eedb5530 | 1258 | do { |
1da177e4 LT |
1259 | struct page *page = bvec->bv_page; |
1260 | ||
eedb5530 NS |
1261 | if (unlikely(pb->pb_error)) { |
1262 | if (pb->pb_flags & PBF_READ) | |
1263 | ClearPageUptodate(page); | |
1da177e4 LT |
1264 | SetPageError(page); |
1265 | } else if (blocksize == PAGE_CACHE_SIZE) { | |
1266 | SetPageUptodate(page); | |
1267 | } else if (!PagePrivate(page) && | |
1268 | (pb->pb_flags & _PBF_PAGE_CACHE)) { | |
1269 | set_page_region(page, bvec->bv_offset, bvec->bv_len); | |
1270 | } | |
1271 | ||
eedb5530 NS |
1272 | if (--bvec >= bio->bi_io_vec) |
1273 | prefetchw(&bvec->bv_page->flags); | |
1274 | ||
1da177e4 LT |
1275 | if (_pagebuf_iolocked(pb)) { |
1276 | unlock_page(page); | |
1277 | } | |
eedb5530 | 1278 | } while (bvec >= bio->bi_io_vec); |
1da177e4 LT |
1279 | |
1280 | _pagebuf_iodone(pb, 1); | |
1281 | bio_put(bio); | |
1282 | return 0; | |
1283 | } | |
1284 | ||
1285 | STATIC void | |
1286 | _pagebuf_ioapply( | |
1287 | xfs_buf_t *pb) | |
1288 | { | |
1289 | int i, rw, map_i, total_nr_pages, nr_pages; | |
1290 | struct bio *bio; | |
1291 | int offset = pb->pb_offset; | |
1292 | int size = pb->pb_count_desired; | |
1293 | sector_t sector = pb->pb_bn; | |
1294 | unsigned int blocksize = pb->pb_target->pbr_bsize; | |
1295 | int locking = _pagebuf_iolocked(pb); | |
1296 | ||
1297 | total_nr_pages = pb->pb_page_count; | |
1298 | map_i = 0; | |
1299 | ||
1300 | if (pb->pb_flags & _PBF_RUN_QUEUES) { | |
1301 | pb->pb_flags &= ~_PBF_RUN_QUEUES; | |
1302 | rw = (pb->pb_flags & PBF_READ) ? READ_SYNC : WRITE_SYNC; | |
1303 | } else { | |
1304 | rw = (pb->pb_flags & PBF_READ) ? READ : WRITE; | |
1305 | } | |
1306 | ||
1307 | /* Special code path for reading a sub page size pagebuf in -- | |
1308 | * we populate up the whole page, and hence the other metadata | |
1309 | * in the same page. This optimization is only valid when the | |
1310 | * filesystem block size and the page size are equal. | |
1311 | */ | |
1312 | if ((pb->pb_buffer_length < PAGE_CACHE_SIZE) && | |
1313 | (pb->pb_flags & PBF_READ) && locking && | |
1314 | (blocksize == PAGE_CACHE_SIZE)) { | |
1315 | bio = bio_alloc(GFP_NOIO, 1); | |
1316 | ||
1317 | bio->bi_bdev = pb->pb_target->pbr_bdev; | |
1318 | bio->bi_sector = sector - (offset >> BBSHIFT); | |
1319 | bio->bi_end_io = bio_end_io_pagebuf; | |
1320 | bio->bi_private = pb; | |
1321 | ||
1322 | bio_add_page(bio, pb->pb_pages[0], PAGE_CACHE_SIZE, 0); | |
1323 | size = 0; | |
1324 | ||
1325 | atomic_inc(&pb->pb_io_remaining); | |
1326 | ||
1327 | goto submit_io; | |
1328 | } | |
1329 | ||
1330 | /* Lock down the pages which we need to for the request */ | |
1331 | if (locking && (pb->pb_flags & PBF_WRITE) && (pb->pb_locked == 0)) { | |
1332 | for (i = 0; size; i++) { | |
1333 | int nbytes = PAGE_CACHE_SIZE - offset; | |
1334 | struct page *page = pb->pb_pages[i]; | |
1335 | ||
1336 | if (nbytes > size) | |
1337 | nbytes = size; | |
1338 | ||
1339 | lock_page(page); | |
1340 | ||
1341 | size -= nbytes; | |
1342 | offset = 0; | |
1343 | } | |
1344 | offset = pb->pb_offset; | |
1345 | size = pb->pb_count_desired; | |
1346 | } | |
1347 | ||
1348 | next_chunk: | |
1349 | atomic_inc(&pb->pb_io_remaining); | |
1350 | nr_pages = BIO_MAX_SECTORS >> (PAGE_SHIFT - BBSHIFT); | |
1351 | if (nr_pages > total_nr_pages) | |
1352 | nr_pages = total_nr_pages; | |
1353 | ||
1354 | bio = bio_alloc(GFP_NOIO, nr_pages); | |
1355 | bio->bi_bdev = pb->pb_target->pbr_bdev; | |
1356 | bio->bi_sector = sector; | |
1357 | bio->bi_end_io = bio_end_io_pagebuf; | |
1358 | bio->bi_private = pb; | |
1359 | ||
1360 | for (; size && nr_pages; nr_pages--, map_i++) { | |
1361 | int nbytes = PAGE_CACHE_SIZE - offset; | |
1362 | ||
1363 | if (nbytes > size) | |
1364 | nbytes = size; | |
1365 | ||
1366 | if (bio_add_page(bio, pb->pb_pages[map_i], | |
1367 | nbytes, offset) < nbytes) | |
1368 | break; | |
1369 | ||
1370 | offset = 0; | |
1371 | sector += nbytes >> BBSHIFT; | |
1372 | size -= nbytes; | |
1373 | total_nr_pages--; | |
1374 | } | |
1375 | ||
1376 | submit_io: | |
1377 | if (likely(bio->bi_size)) { | |
1378 | submit_bio(rw, bio); | |
1379 | if (size) | |
1380 | goto next_chunk; | |
1381 | } else { | |
1382 | bio_put(bio); | |
1383 | pagebuf_ioerror(pb, EIO); | |
1384 | } | |
1385 | } | |
1386 | ||
1387 | /* | |
1388 | * pagebuf_iorequest -- the core I/O request routine. | |
1389 | */ | |
1390 | int | |
1391 | pagebuf_iorequest( /* start real I/O */ | |
1392 | xfs_buf_t *pb) /* buffer to convey to device */ | |
1393 | { | |
1394 | PB_TRACE(pb, "iorequest", 0); | |
1395 | ||
1396 | if (pb->pb_flags & PBF_DELWRI) { | |
1397 | pagebuf_delwri_queue(pb, 1); | |
1398 | return 0; | |
1399 | } | |
1400 | ||
1401 | if (pb->pb_flags & PBF_WRITE) { | |
1402 | _pagebuf_wait_unpin(pb); | |
1403 | } | |
1404 | ||
1405 | pagebuf_hold(pb); | |
1406 | ||
1407 | /* Set the count to 1 initially, this will stop an I/O | |
1408 | * completion callout which happens before we have started | |
1409 | * all the I/O from calling pagebuf_iodone too early. | |
1410 | */ | |
1411 | atomic_set(&pb->pb_io_remaining, 1); | |
1412 | _pagebuf_ioapply(pb); | |
1413 | _pagebuf_iodone(pb, 0); | |
1414 | ||
1415 | pagebuf_rele(pb); | |
1416 | return 0; | |
1417 | } | |
1418 | ||
1419 | /* | |
1420 | * pagebuf_iowait | |
1421 | * | |
1422 | * pagebuf_iowait waits for I/O to complete on the buffer supplied. | |
1423 | * It returns immediately if no I/O is pending. In any case, it returns | |
1424 | * the error code, if any, or 0 if there is no error. | |
1425 | */ | |
1426 | int | |
1427 | pagebuf_iowait( | |
1428 | xfs_buf_t *pb) | |
1429 | { | |
1430 | PB_TRACE(pb, "iowait", 0); | |
1431 | if (atomic_read(&pb->pb_io_remaining)) | |
1432 | blk_run_address_space(pb->pb_target->pbr_mapping); | |
1433 | down(&pb->pb_iodonesema); | |
1434 | PB_TRACE(pb, "iowaited", (long)pb->pb_error); | |
1435 | return pb->pb_error; | |
1436 | } | |
1437 | ||
1438 | caddr_t | |
1439 | pagebuf_offset( | |
1440 | xfs_buf_t *pb, | |
1441 | size_t offset) | |
1442 | { | |
1443 | struct page *page; | |
1444 | ||
1445 | offset += pb->pb_offset; | |
1446 | ||
1447 | page = pb->pb_pages[offset >> PAGE_CACHE_SHIFT]; | |
1448 | return (caddr_t) page_address(page) + (offset & (PAGE_CACHE_SIZE - 1)); | |
1449 | } | |
1450 | ||
1451 | /* | |
1452 | * pagebuf_iomove | |
1453 | * | |
1454 | * Move data into or out of a buffer. | |
1455 | */ | |
1456 | void | |
1457 | pagebuf_iomove( | |
1458 | xfs_buf_t *pb, /* buffer to process */ | |
1459 | size_t boff, /* starting buffer offset */ | |
1460 | size_t bsize, /* length to copy */ | |
1461 | caddr_t data, /* data address */ | |
1462 | page_buf_rw_t mode) /* read/write flag */ | |
1463 | { | |
1464 | size_t bend, cpoff, csize; | |
1465 | struct page *page; | |
1466 | ||
1467 | bend = boff + bsize; | |
1468 | while (boff < bend) { | |
1469 | page = pb->pb_pages[page_buf_btoct(boff + pb->pb_offset)]; | |
1470 | cpoff = page_buf_poff(boff + pb->pb_offset); | |
1471 | csize = min_t(size_t, | |
1472 | PAGE_CACHE_SIZE-cpoff, pb->pb_count_desired-boff); | |
1473 | ||
1474 | ASSERT(((csize + cpoff) <= PAGE_CACHE_SIZE)); | |
1475 | ||
1476 | switch (mode) { | |
1477 | case PBRW_ZERO: | |
1478 | memset(page_address(page) + cpoff, 0, csize); | |
1479 | break; | |
1480 | case PBRW_READ: | |
1481 | memcpy(data, page_address(page) + cpoff, csize); | |
1482 | break; | |
1483 | case PBRW_WRITE: | |
1484 | memcpy(page_address(page) + cpoff, data, csize); | |
1485 | } | |
1486 | ||
1487 | boff += csize; | |
1488 | data += csize; | |
1489 | } | |
1490 | } | |
1491 | ||
1492 | /* | |
1493 | * Handling of buftargs. | |
1494 | */ | |
1495 | ||
1496 | /* | |
1497 | * Wait for any bufs with callbacks that have been submitted but | |
1498 | * have not yet returned... walk the hash list for the target. | |
1499 | */ | |
1500 | void | |
1501 | xfs_wait_buftarg( | |
1502 | xfs_buftarg_t *btp) | |
1503 | { | |
1504 | xfs_buf_t *bp, *n; | |
1505 | xfs_bufhash_t *hash; | |
1506 | uint i; | |
1507 | ||
1508 | for (i = 0; i < (1 << btp->bt_hashshift); i++) { | |
1509 | hash = &btp->bt_hash[i]; | |
1510 | again: | |
1511 | spin_lock(&hash->bh_lock); | |
1512 | list_for_each_entry_safe(bp, n, &hash->bh_list, pb_hash_list) { | |
1513 | ASSERT(btp == bp->pb_target); | |
1514 | if (!(bp->pb_flags & PBF_FS_MANAGED)) { | |
1515 | spin_unlock(&hash->bh_lock); | |
2f926587 DC |
1516 | /* |
1517 | * Catch superblock reference count leaks | |
1518 | * immediately | |
1519 | */ | |
1520 | BUG_ON(bp->pb_bn == 0); | |
1da177e4 LT |
1521 | delay(100); |
1522 | goto again; | |
1523 | } | |
1524 | } | |
1525 | spin_unlock(&hash->bh_lock); | |
1526 | } | |
1527 | } | |
1528 | ||
1529 | /* | |
1530 | * Allocate buffer hash table for a given target. | |
1531 | * For devices containing metadata (i.e. not the log/realtime devices) | |
1532 | * we need to allocate a much larger hash table. | |
1533 | */ | |
1534 | STATIC void | |
1535 | xfs_alloc_bufhash( | |
1536 | xfs_buftarg_t *btp, | |
1537 | int external) | |
1538 | { | |
1539 | unsigned int i; | |
1540 | ||
1541 | btp->bt_hashshift = external ? 3 : 8; /* 8 or 256 buckets */ | |
1542 | btp->bt_hashmask = (1 << btp->bt_hashshift) - 1; | |
1543 | btp->bt_hash = kmem_zalloc((1 << btp->bt_hashshift) * | |
1544 | sizeof(xfs_bufhash_t), KM_SLEEP); | |
1545 | for (i = 0; i < (1 << btp->bt_hashshift); i++) { | |
1546 | spin_lock_init(&btp->bt_hash[i].bh_lock); | |
1547 | INIT_LIST_HEAD(&btp->bt_hash[i].bh_list); | |
1548 | } | |
1549 | } | |
1550 | ||
1551 | STATIC void | |
1552 | xfs_free_bufhash( | |
1553 | xfs_buftarg_t *btp) | |
1554 | { | |
1555 | kmem_free(btp->bt_hash, | |
1556 | (1 << btp->bt_hashshift) * sizeof(xfs_bufhash_t)); | |
1557 | btp->bt_hash = NULL; | |
1558 | } | |
1559 | ||
1560 | void | |
1561 | xfs_free_buftarg( | |
1562 | xfs_buftarg_t *btp, | |
1563 | int external) | |
1564 | { | |
1565 | xfs_flush_buftarg(btp, 1); | |
1566 | if (external) | |
1567 | xfs_blkdev_put(btp->pbr_bdev); | |
1568 | xfs_free_bufhash(btp); | |
1569 | iput(btp->pbr_mapping->host); | |
1570 | kmem_free(btp, sizeof(*btp)); | |
1571 | } | |
1572 | ||
1da177e4 LT |
1573 | STATIC int |
1574 | xfs_setsize_buftarg_flags( | |
1575 | xfs_buftarg_t *btp, | |
1576 | unsigned int blocksize, | |
1577 | unsigned int sectorsize, | |
1578 | int verbose) | |
1579 | { | |
1580 | btp->pbr_bsize = blocksize; | |
1581 | btp->pbr_sshift = ffs(sectorsize) - 1; | |
1582 | btp->pbr_smask = sectorsize - 1; | |
1583 | ||
1584 | if (set_blocksize(btp->pbr_bdev, sectorsize)) { | |
1585 | printk(KERN_WARNING | |
1586 | "XFS: Cannot set_blocksize to %u on device %s\n", | |
1587 | sectorsize, XFS_BUFTARG_NAME(btp)); | |
1588 | return EINVAL; | |
1589 | } | |
1590 | ||
1591 | if (verbose && | |
1592 | (PAGE_CACHE_SIZE / BITS_PER_LONG) > sectorsize) { | |
1593 | printk(KERN_WARNING | |
1594 | "XFS: %u byte sectors in use on device %s. " | |
1595 | "This is suboptimal; %u or greater is ideal.\n", | |
1596 | sectorsize, XFS_BUFTARG_NAME(btp), | |
1597 | (unsigned int)PAGE_CACHE_SIZE / BITS_PER_LONG); | |
1598 | } | |
1599 | ||
1600 | return 0; | |
1601 | } | |
1602 | ||
1603 | /* | |
1604 | * When allocating the initial buffer target we have not yet | |
1605 | * read in the superblock, so don't know what sized sectors | |
1606 | * are being used is at this early stage. Play safe. | |
1607 | */ | |
1608 | STATIC int | |
1609 | xfs_setsize_buftarg_early( | |
1610 | xfs_buftarg_t *btp, | |
1611 | struct block_device *bdev) | |
1612 | { | |
1613 | return xfs_setsize_buftarg_flags(btp, | |
1614 | PAGE_CACHE_SIZE, bdev_hardsect_size(bdev), 0); | |
1615 | } | |
1616 | ||
1617 | int | |
1618 | xfs_setsize_buftarg( | |
1619 | xfs_buftarg_t *btp, | |
1620 | unsigned int blocksize, | |
1621 | unsigned int sectorsize) | |
1622 | { | |
1623 | return xfs_setsize_buftarg_flags(btp, blocksize, sectorsize, 1); | |
1624 | } | |
1625 | ||
1626 | STATIC int | |
1627 | xfs_mapping_buftarg( | |
1628 | xfs_buftarg_t *btp, | |
1629 | struct block_device *bdev) | |
1630 | { | |
1631 | struct backing_dev_info *bdi; | |
1632 | struct inode *inode; | |
1633 | struct address_space *mapping; | |
1634 | static struct address_space_operations mapping_aops = { | |
1635 | .sync_page = block_sync_page, | |
1636 | }; | |
1637 | ||
1638 | inode = new_inode(bdev->bd_inode->i_sb); | |
1639 | if (!inode) { | |
1640 | printk(KERN_WARNING | |
1641 | "XFS: Cannot allocate mapping inode for device %s\n", | |
1642 | XFS_BUFTARG_NAME(btp)); | |
1643 | return ENOMEM; | |
1644 | } | |
1645 | inode->i_mode = S_IFBLK; | |
1646 | inode->i_bdev = bdev; | |
1647 | inode->i_rdev = bdev->bd_dev; | |
1648 | bdi = blk_get_backing_dev_info(bdev); | |
1649 | if (!bdi) | |
1650 | bdi = &default_backing_dev_info; | |
1651 | mapping = &inode->i_data; | |
1652 | mapping->a_ops = &mapping_aops; | |
1653 | mapping->backing_dev_info = bdi; | |
1654 | mapping_set_gfp_mask(mapping, GFP_NOFS); | |
1655 | btp->pbr_mapping = mapping; | |
1656 | return 0; | |
1657 | } | |
1658 | ||
1659 | xfs_buftarg_t * | |
1660 | xfs_alloc_buftarg( | |
1661 | struct block_device *bdev, | |
1662 | int external) | |
1663 | { | |
1664 | xfs_buftarg_t *btp; | |
1665 | ||
1666 | btp = kmem_zalloc(sizeof(*btp), KM_SLEEP); | |
1667 | ||
1668 | btp->pbr_dev = bdev->bd_dev; | |
1669 | btp->pbr_bdev = bdev; | |
1670 | if (xfs_setsize_buftarg_early(btp, bdev)) | |
1671 | goto error; | |
1672 | if (xfs_mapping_buftarg(btp, bdev)) | |
1673 | goto error; | |
1674 | xfs_alloc_bufhash(btp, external); | |
1675 | return btp; | |
1676 | ||
1677 | error: | |
1678 | kmem_free(btp, sizeof(*btp)); | |
1679 | return NULL; | |
1680 | } | |
1681 | ||
1682 | ||
1683 | /* | |
1684 | * Pagebuf delayed write buffer handling | |
1685 | */ | |
1686 | ||
1687 | STATIC LIST_HEAD(pbd_delwrite_queue); | |
1688 | STATIC DEFINE_SPINLOCK(pbd_delwrite_lock); | |
1689 | ||
1690 | STATIC void | |
1691 | pagebuf_delwri_queue( | |
1692 | xfs_buf_t *pb, | |
1693 | int unlock) | |
1694 | { | |
1695 | PB_TRACE(pb, "delwri_q", (long)unlock); | |
2f926587 DC |
1696 | ASSERT((pb->pb_flags & (PBF_DELWRI|PBF_ASYNC)) == |
1697 | (PBF_DELWRI|PBF_ASYNC)); | |
1da177e4 LT |
1698 | |
1699 | spin_lock(&pbd_delwrite_lock); | |
1700 | /* If already in the queue, dequeue and place at tail */ | |
1701 | if (!list_empty(&pb->pb_list)) { | |
2f926587 | 1702 | ASSERT(pb->pb_flags & _PBF_DELWRI_Q); |
1da177e4 LT |
1703 | if (unlock) { |
1704 | atomic_dec(&pb->pb_hold); | |
1705 | } | |
1706 | list_del(&pb->pb_list); | |
1707 | } | |
1708 | ||
2f926587 | 1709 | pb->pb_flags |= _PBF_DELWRI_Q; |
1da177e4 LT |
1710 | list_add_tail(&pb->pb_list, &pbd_delwrite_queue); |
1711 | pb->pb_queuetime = jiffies; | |
1712 | spin_unlock(&pbd_delwrite_lock); | |
1713 | ||
1714 | if (unlock) | |
1715 | pagebuf_unlock(pb); | |
1716 | } | |
1717 | ||
1718 | void | |
1719 | pagebuf_delwri_dequeue( | |
1720 | xfs_buf_t *pb) | |
1721 | { | |
1722 | int dequeued = 0; | |
1723 | ||
1724 | spin_lock(&pbd_delwrite_lock); | |
1725 | if ((pb->pb_flags & PBF_DELWRI) && !list_empty(&pb->pb_list)) { | |
2f926587 | 1726 | ASSERT(pb->pb_flags & _PBF_DELWRI_Q); |
1da177e4 LT |
1727 | list_del_init(&pb->pb_list); |
1728 | dequeued = 1; | |
1729 | } | |
2f926587 | 1730 | pb->pb_flags &= ~(PBF_DELWRI|_PBF_DELWRI_Q); |
1da177e4 LT |
1731 | spin_unlock(&pbd_delwrite_lock); |
1732 | ||
1733 | if (dequeued) | |
1734 | pagebuf_rele(pb); | |
1735 | ||
1736 | PB_TRACE(pb, "delwri_dq", (long)dequeued); | |
1737 | } | |
1738 | ||
1739 | STATIC void | |
1740 | pagebuf_runall_queues( | |
1741 | struct workqueue_struct *queue) | |
1742 | { | |
1743 | flush_workqueue(queue); | |
1744 | } | |
1745 | ||
1746 | /* Defines for pagebuf daemon */ | |
23ea4032 | 1747 | STATIC struct task_struct *xfsbufd_task; |
23ea4032 CH |
1748 | STATIC int xfsbufd_force_flush; |
1749 | STATIC int xfsbufd_force_sleep; | |
1da177e4 LT |
1750 | |
1751 | STATIC int | |
23ea4032 | 1752 | xfsbufd_wakeup( |
27496a8c AV |
1753 | int priority, |
1754 | gfp_t mask) | |
1da177e4 | 1755 | { |
23ea4032 | 1756 | if (xfsbufd_force_sleep) |
abd0cf7a | 1757 | return 0; |
23ea4032 | 1758 | xfsbufd_force_flush = 1; |
1da177e4 | 1759 | barrier(); |
23ea4032 | 1760 | wake_up_process(xfsbufd_task); |
1da177e4 LT |
1761 | return 0; |
1762 | } | |
1763 | ||
1764 | STATIC int | |
23ea4032 | 1765 | xfsbufd( |
1da177e4 LT |
1766 | void *data) |
1767 | { | |
1768 | struct list_head tmp; | |
1769 | unsigned long age; | |
1770 | xfs_buftarg_t *target; | |
1771 | xfs_buf_t *pb, *n; | |
1772 | ||
1da177e4 LT |
1773 | current->flags |= PF_MEMALLOC; |
1774 | ||
1da177e4 LT |
1775 | INIT_LIST_HEAD(&tmp); |
1776 | do { | |
3e1d1d28 | 1777 | if (unlikely(freezing(current))) { |
23ea4032 | 1778 | xfsbufd_force_sleep = 1; |
3e1d1d28 | 1779 | refrigerator(); |
abd0cf7a | 1780 | } else { |
23ea4032 | 1781 | xfsbufd_force_sleep = 0; |
abd0cf7a | 1782 | } |
1da177e4 | 1783 | |
041e0e3b NA |
1784 | schedule_timeout_interruptible |
1785 | (xfs_buf_timer_centisecs * msecs_to_jiffies(10)); | |
1da177e4 | 1786 | |
041e0e3b | 1787 | age = xfs_buf_age_centisecs * msecs_to_jiffies(10); |
1da177e4 LT |
1788 | spin_lock(&pbd_delwrite_lock); |
1789 | list_for_each_entry_safe(pb, n, &pbd_delwrite_queue, pb_list) { | |
1790 | PB_TRACE(pb, "walkq1", (long)pagebuf_ispin(pb)); | |
1791 | ASSERT(pb->pb_flags & PBF_DELWRI); | |
1792 | ||
1793 | if (!pagebuf_ispin(pb) && !pagebuf_cond_lock(pb)) { | |
23ea4032 | 1794 | if (!xfsbufd_force_flush && |
1da177e4 LT |
1795 | time_before(jiffies, |
1796 | pb->pb_queuetime + age)) { | |
1797 | pagebuf_unlock(pb); | |
1798 | break; | |
1799 | } | |
1800 | ||
2f926587 | 1801 | pb->pb_flags &= ~(PBF_DELWRI|_PBF_DELWRI_Q); |
1da177e4 LT |
1802 | pb->pb_flags |= PBF_WRITE; |
1803 | list_move(&pb->pb_list, &tmp); | |
1804 | } | |
1805 | } | |
1806 | spin_unlock(&pbd_delwrite_lock); | |
1807 | ||
1808 | while (!list_empty(&tmp)) { | |
1809 | pb = list_entry(tmp.next, xfs_buf_t, pb_list); | |
1810 | target = pb->pb_target; | |
1811 | ||
1812 | list_del_init(&pb->pb_list); | |
1813 | pagebuf_iostrategy(pb); | |
1814 | ||
1815 | blk_run_address_space(target->pbr_mapping); | |
1816 | } | |
1817 | ||
1818 | if (as_list_len > 0) | |
1819 | purge_addresses(); | |
1820 | ||
23ea4032 | 1821 | xfsbufd_force_flush = 0; |
4df08c52 | 1822 | } while (!kthread_should_stop()); |
1da177e4 | 1823 | |
4df08c52 | 1824 | return 0; |
1da177e4 LT |
1825 | } |
1826 | ||
1827 | /* | |
1828 | * Go through all incore buffers, and release buffers if they belong to | |
1829 | * the given device. This is used in filesystem error handling to | |
1830 | * preserve the consistency of its metadata. | |
1831 | */ | |
1832 | int | |
1833 | xfs_flush_buftarg( | |
1834 | xfs_buftarg_t *target, | |
1835 | int wait) | |
1836 | { | |
1837 | struct list_head tmp; | |
1838 | xfs_buf_t *pb, *n; | |
1839 | int pincount = 0; | |
1840 | ||
23ea4032 CH |
1841 | pagebuf_runall_queues(xfsdatad_workqueue); |
1842 | pagebuf_runall_queues(xfslogd_workqueue); | |
1da177e4 LT |
1843 | |
1844 | INIT_LIST_HEAD(&tmp); | |
1845 | spin_lock(&pbd_delwrite_lock); | |
1846 | list_for_each_entry_safe(pb, n, &pbd_delwrite_queue, pb_list) { | |
1847 | ||
1848 | if (pb->pb_target != target) | |
1849 | continue; | |
1850 | ||
2f926587 | 1851 | ASSERT(pb->pb_flags & (PBF_DELWRI|_PBF_DELWRI_Q)); |
1da177e4 LT |
1852 | PB_TRACE(pb, "walkq2", (long)pagebuf_ispin(pb)); |
1853 | if (pagebuf_ispin(pb)) { | |
1854 | pincount++; | |
1855 | continue; | |
1856 | } | |
1857 | ||
1da177e4 LT |
1858 | list_move(&pb->pb_list, &tmp); |
1859 | } | |
1860 | spin_unlock(&pbd_delwrite_lock); | |
1861 | ||
1862 | /* | |
1863 | * Dropped the delayed write list lock, now walk the temporary list | |
1864 | */ | |
1865 | list_for_each_entry_safe(pb, n, &tmp, pb_list) { | |
2f926587 DC |
1866 | pagebuf_lock(pb); |
1867 | pb->pb_flags &= ~(PBF_DELWRI|_PBF_DELWRI_Q); | |
1868 | pb->pb_flags |= PBF_WRITE; | |
1da177e4 LT |
1869 | if (wait) |
1870 | pb->pb_flags &= ~PBF_ASYNC; | |
1871 | else | |
1872 | list_del_init(&pb->pb_list); | |
1873 | ||
1da177e4 LT |
1874 | pagebuf_iostrategy(pb); |
1875 | } | |
1876 | ||
1877 | /* | |
1878 | * Remaining list items must be flushed before returning | |
1879 | */ | |
1880 | while (!list_empty(&tmp)) { | |
1881 | pb = list_entry(tmp.next, xfs_buf_t, pb_list); | |
1882 | ||
1883 | list_del_init(&pb->pb_list); | |
1884 | xfs_iowait(pb); | |
1885 | xfs_buf_relse(pb); | |
1886 | } | |
1887 | ||
1888 | if (wait) | |
1889 | blk_run_address_space(target->pbr_mapping); | |
1890 | ||
1891 | return pincount; | |
1892 | } | |
1893 | ||
04d8b284 CH |
1894 | int __init |
1895 | pagebuf_init(void) | |
1da177e4 | 1896 | { |
23ea4032 | 1897 | int error = -ENOMEM; |
1da177e4 | 1898 | |
04d8b284 CH |
1899 | #ifdef PAGEBUF_TRACE |
1900 | pagebuf_trace_buf = ktrace_alloc(PAGEBUF_TRACE_SIZE, KM_SLEEP); | |
1901 | #endif | |
1902 | ||
1903 | pagebuf_zone = kmem_zone_init(sizeof(xfs_buf_t), "xfs_buf"); | |
1904 | if (!pagebuf_zone) | |
1905 | goto out_free_trace_buf; | |
1906 | ||
23ea4032 CH |
1907 | xfslogd_workqueue = create_workqueue("xfslogd"); |
1908 | if (!xfslogd_workqueue) | |
04d8b284 | 1909 | goto out_free_buf_zone; |
1da177e4 | 1910 | |
23ea4032 CH |
1911 | xfsdatad_workqueue = create_workqueue("xfsdatad"); |
1912 | if (!xfsdatad_workqueue) | |
1913 | goto out_destroy_xfslogd_workqueue; | |
1da177e4 | 1914 | |
4df08c52 CH |
1915 | xfsbufd_task = kthread_run(xfsbufd, NULL, "xfsbufd"); |
1916 | if (IS_ERR(xfsbufd_task)) { | |
1917 | error = PTR_ERR(xfsbufd_task); | |
23ea4032 | 1918 | goto out_destroy_xfsdatad_workqueue; |
4df08c52 | 1919 | } |
04d8b284 CH |
1920 | |
1921 | pagebuf_shake = kmem_shake_register(xfsbufd_wakeup); | |
1922 | if (!pagebuf_shake) | |
1923 | goto out_stop_xfsbufd; | |
1924 | ||
23ea4032 | 1925 | return 0; |
1da177e4 | 1926 | |
04d8b284 CH |
1927 | out_stop_xfsbufd: |
1928 | kthread_stop(xfsbufd_task); | |
23ea4032 CH |
1929 | out_destroy_xfsdatad_workqueue: |
1930 | destroy_workqueue(xfsdatad_workqueue); | |
1931 | out_destroy_xfslogd_workqueue: | |
1932 | destroy_workqueue(xfslogd_workqueue); | |
23ea4032 | 1933 | out_free_buf_zone: |
04d8b284 CH |
1934 | kmem_zone_destroy(pagebuf_zone); |
1935 | out_free_trace_buf: | |
23ea4032 CH |
1936 | #ifdef PAGEBUF_TRACE |
1937 | ktrace_free(pagebuf_trace_buf); | |
1938 | #endif | |
23ea4032 | 1939 | return error; |
1da177e4 LT |
1940 | } |
1941 | ||
1da177e4 LT |
1942 | void |
1943 | pagebuf_terminate(void) | |
1944 | { | |
04d8b284 CH |
1945 | kmem_shake_deregister(pagebuf_shake); |
1946 | kthread_stop(xfsbufd_task); | |
1947 | destroy_workqueue(xfsdatad_workqueue); | |
1948 | destroy_workqueue(xfslogd_workqueue); | |
1949 | kmem_zone_destroy(pagebuf_zone); | |
1da177e4 LT |
1950 | #ifdef PAGEBUF_TRACE |
1951 | ktrace_free(pagebuf_trace_buf); | |
1952 | #endif | |
1da177e4 | 1953 | } |