Merge branches 'topic/fix/hda' and 'topic/fix/misc' into for-linus
[linux-2.6] / fs / xfs / linux-2.6 / xfs_buf.c
1 /*
2  * Copyright (c) 2000-2006 Silicon Graphics, Inc.
3  * All Rights Reserved.
4  *
5  * This program is free software; you can redistribute it and/or
6  * modify it under the terms of the GNU General Public License as
7  * published by the Free Software Foundation.
8  *
9  * This program is distributed in the hope that it would be useful,
10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
12  * GNU General Public License for more details.
13  *
14  * You should have received a copy of the GNU General Public License
15  * along with this program; if not, write the Free Software Foundation,
16  * Inc.,  51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
17  */
18 #include "xfs.h"
19 #include <linux/stddef.h>
20 #include <linux/errno.h>
21 #include <linux/slab.h>
22 #include <linux/pagemap.h>
23 #include <linux/init.h>
24 #include <linux/vmalloc.h>
25 #include <linux/bio.h>
26 #include <linux/sysctl.h>
27 #include <linux/proc_fs.h>
28 #include <linux/workqueue.h>
29 #include <linux/percpu.h>
30 #include <linux/blkdev.h>
31 #include <linux/hash.h>
32 #include <linux/kthread.h>
33 #include <linux/migrate.h>
34 #include <linux/backing-dev.h>
35 #include <linux/freezer.h>
36
37 static kmem_zone_t *xfs_buf_zone;
38 STATIC int xfsbufd(void *);
39 STATIC int xfsbufd_wakeup(int, gfp_t);
40 STATIC void xfs_buf_delwri_queue(xfs_buf_t *, int);
41 static struct shrinker xfs_buf_shake = {
42         .shrink = xfsbufd_wakeup,
43         .seeks = DEFAULT_SEEKS,
44 };
45
46 static struct workqueue_struct *xfslogd_workqueue;
47 struct workqueue_struct *xfsdatad_workqueue;
48
49 #ifdef XFS_BUF_TRACE
50 void
51 xfs_buf_trace(
52         xfs_buf_t       *bp,
53         char            *id,
54         void            *data,
55         void            *ra)
56 {
57         ktrace_enter(xfs_buf_trace_buf,
58                 bp, id,
59                 (void *)(unsigned long)bp->b_flags,
60                 (void *)(unsigned long)bp->b_hold.counter,
61                 (void *)(unsigned long)bp->b_sema.count,
62                 (void *)current,
63                 data, ra,
64                 (void *)(unsigned long)((bp->b_file_offset>>32) & 0xffffffff),
65                 (void *)(unsigned long)(bp->b_file_offset & 0xffffffff),
66                 (void *)(unsigned long)bp->b_buffer_length,
67                 NULL, NULL, NULL, NULL, NULL);
68 }
69 ktrace_t *xfs_buf_trace_buf;
70 #define XFS_BUF_TRACE_SIZE      4096
71 #define XB_TRACE(bp, id, data)  \
72         xfs_buf_trace(bp, id, (void *)data, (void *)__builtin_return_address(0))
73 #else
74 #define XB_TRACE(bp, id, data)  do { } while (0)
75 #endif
76
77 #ifdef XFS_BUF_LOCK_TRACKING
78 # define XB_SET_OWNER(bp)       ((bp)->b_last_holder = current->pid)
79 # define XB_CLEAR_OWNER(bp)     ((bp)->b_last_holder = -1)
80 # define XB_GET_OWNER(bp)       ((bp)->b_last_holder)
81 #else
82 # define XB_SET_OWNER(bp)       do { } while (0)
83 # define XB_CLEAR_OWNER(bp)     do { } while (0)
84 # define XB_GET_OWNER(bp)       do { } while (0)
85 #endif
86
87 #define xb_to_gfp(flags) \
88         ((((flags) & XBF_READ_AHEAD) ? __GFP_NORETRY : \
89           ((flags) & XBF_DONT_BLOCK) ? GFP_NOFS : GFP_KERNEL) | __GFP_NOWARN)
90
91 #define xb_to_km(flags) \
92          (((flags) & XBF_DONT_BLOCK) ? KM_NOFS : KM_SLEEP)
93
94 #define xfs_buf_allocate(flags) \
95         kmem_zone_alloc(xfs_buf_zone, xb_to_km(flags))
96 #define xfs_buf_deallocate(bp) \
97         kmem_zone_free(xfs_buf_zone, (bp));
98
99 /*
100  *      Page Region interfaces.
101  *
102  *      For pages in filesystems where the blocksize is smaller than the
103  *      pagesize, we use the page->private field (long) to hold a bitmap
104  *      of uptodate regions within the page.
105  *
106  *      Each such region is "bytes per page / bits per long" bytes long.
107  *
108  *      NBPPR == number-of-bytes-per-page-region
109  *      BTOPR == bytes-to-page-region (rounded up)
110  *      BTOPRT == bytes-to-page-region-truncated (rounded down)
111  */
112 #if (BITS_PER_LONG == 32)
113 #define PRSHIFT         (PAGE_CACHE_SHIFT - 5)  /* (32 == 1<<5) */
114 #elif (BITS_PER_LONG == 64)
115 #define PRSHIFT         (PAGE_CACHE_SHIFT - 6)  /* (64 == 1<<6) */
116 #else
117 #error BITS_PER_LONG must be 32 or 64
118 #endif
119 #define NBPPR           (PAGE_CACHE_SIZE/BITS_PER_LONG)
120 #define BTOPR(b)        (((unsigned int)(b) + (NBPPR - 1)) >> PRSHIFT)
121 #define BTOPRT(b)       (((unsigned int)(b) >> PRSHIFT))
122
123 STATIC unsigned long
124 page_region_mask(
125         size_t          offset,
126         size_t          length)
127 {
128         unsigned long   mask;
129         int             first, final;
130
131         first = BTOPR(offset);
132         final = BTOPRT(offset + length - 1);
133         first = min(first, final);
134
135         mask = ~0UL;
136         mask <<= BITS_PER_LONG - (final - first);
137         mask >>= BITS_PER_LONG - (final);
138
139         ASSERT(offset + length <= PAGE_CACHE_SIZE);
140         ASSERT((final - first) < BITS_PER_LONG && (final - first) >= 0);
141
142         return mask;
143 }
144
145 STATIC_INLINE void
146 set_page_region(
147         struct page     *page,
148         size_t          offset,
149         size_t          length)
150 {
151         set_page_private(page,
152                 page_private(page) | page_region_mask(offset, length));
153         if (page_private(page) == ~0UL)
154                 SetPageUptodate(page);
155 }
156
157 STATIC_INLINE int
158 test_page_region(
159         struct page     *page,
160         size_t          offset,
161         size_t          length)
162 {
163         unsigned long   mask = page_region_mask(offset, length);
164
165         return (mask && (page_private(page) & mask) == mask);
166 }
167
168 /*
169  *      Mapping of multi-page buffers into contiguous virtual space
170  */
171
172 typedef struct a_list {
173         void            *vm_addr;
174         struct a_list   *next;
175 } a_list_t;
176
177 static a_list_t         *as_free_head;
178 static int              as_list_len;
179 static DEFINE_SPINLOCK(as_lock);
180
181 /*
182  *      Try to batch vunmaps because they are costly.
183  */
184 STATIC void
185 free_address(
186         void            *addr)
187 {
188         a_list_t        *aentry;
189
190 #ifdef CONFIG_XEN
191         /*
192          * Xen needs to be able to make sure it can get an exclusive
193          * RO mapping of pages it wants to turn into a pagetable.  If
194          * a newly allocated page is also still being vmap()ed by xfs,
195          * it will cause pagetable construction to fail.  This is a
196          * quick workaround to always eagerly unmap pages so that Xen
197          * is happy.
198          */
199         vunmap(addr);
200         return;
201 #endif
202
203         aentry = kmalloc(sizeof(a_list_t), GFP_NOWAIT);
204         if (likely(aentry)) {
205                 spin_lock(&as_lock);
206                 aentry->next = as_free_head;
207                 aentry->vm_addr = addr;
208                 as_free_head = aentry;
209                 as_list_len++;
210                 spin_unlock(&as_lock);
211         } else {
212                 vunmap(addr);
213         }
214 }
215
216 STATIC void
217 purge_addresses(void)
218 {
219         a_list_t        *aentry, *old;
220
221         if (as_free_head == NULL)
222                 return;
223
224         spin_lock(&as_lock);
225         aentry = as_free_head;
226         as_free_head = NULL;
227         as_list_len = 0;
228         spin_unlock(&as_lock);
229
230         while ((old = aentry) != NULL) {
231                 vunmap(aentry->vm_addr);
232                 aentry = aentry->next;
233                 kfree(old);
234         }
235 }
236
237 /*
238  *      Internal xfs_buf_t object manipulation
239  */
240
241 STATIC void
242 _xfs_buf_initialize(
243         xfs_buf_t               *bp,
244         xfs_buftarg_t           *target,
245         xfs_off_t               range_base,
246         size_t                  range_length,
247         xfs_buf_flags_t         flags)
248 {
249         /*
250          * We don't want certain flags to appear in b_flags.
251          */
252         flags &= ~(XBF_LOCK|XBF_MAPPED|XBF_DONT_BLOCK|XBF_READ_AHEAD);
253
254         memset(bp, 0, sizeof(xfs_buf_t));
255         atomic_set(&bp->b_hold, 1);
256         init_completion(&bp->b_iowait);
257         INIT_LIST_HEAD(&bp->b_list);
258         INIT_LIST_HEAD(&bp->b_hash_list);
259         init_MUTEX_LOCKED(&bp->b_sema); /* held, no waiters */
260         XB_SET_OWNER(bp);
261         bp->b_target = target;
262         bp->b_file_offset = range_base;
263         /*
264          * Set buffer_length and count_desired to the same value initially.
265          * I/O routines should use count_desired, which will be the same in
266          * most cases but may be reset (e.g. XFS recovery).
267          */
268         bp->b_buffer_length = bp->b_count_desired = range_length;
269         bp->b_flags = flags;
270         bp->b_bn = XFS_BUF_DADDR_NULL;
271         atomic_set(&bp->b_pin_count, 0);
272         init_waitqueue_head(&bp->b_waiters);
273
274         XFS_STATS_INC(xb_create);
275         XB_TRACE(bp, "initialize", target);
276 }
277
278 /*
279  *      Allocate a page array capable of holding a specified number
280  *      of pages, and point the page buf at it.
281  */
282 STATIC int
283 _xfs_buf_get_pages(
284         xfs_buf_t               *bp,
285         int                     page_count,
286         xfs_buf_flags_t         flags)
287 {
288         /* Make sure that we have a page list */
289         if (bp->b_pages == NULL) {
290                 bp->b_offset = xfs_buf_poff(bp->b_file_offset);
291                 bp->b_page_count = page_count;
292                 if (page_count <= XB_PAGES) {
293                         bp->b_pages = bp->b_page_array;
294                 } else {
295                         bp->b_pages = kmem_alloc(sizeof(struct page *) *
296                                         page_count, xb_to_km(flags));
297                         if (bp->b_pages == NULL)
298                                 return -ENOMEM;
299                 }
300                 memset(bp->b_pages, 0, sizeof(struct page *) * page_count);
301         }
302         return 0;
303 }
304
305 /*
306  *      Frees b_pages if it was allocated.
307  */
308 STATIC void
309 _xfs_buf_free_pages(
310         xfs_buf_t       *bp)
311 {
312         if (bp->b_pages != bp->b_page_array) {
313                 kmem_free(bp->b_pages);
314         }
315 }
316
317 /*
318  *      Releases the specified buffer.
319  *
320  *      The modification state of any associated pages is left unchanged.
321  *      The buffer most not be on any hash - use xfs_buf_rele instead for
322  *      hashed and refcounted buffers
323  */
324 void
325 xfs_buf_free(
326         xfs_buf_t               *bp)
327 {
328         XB_TRACE(bp, "free", 0);
329
330         ASSERT(list_empty(&bp->b_hash_list));
331
332         if (bp->b_flags & (_XBF_PAGE_CACHE|_XBF_PAGES)) {
333                 uint            i;
334
335                 if ((bp->b_flags & XBF_MAPPED) && (bp->b_page_count > 1))
336                         free_address(bp->b_addr - bp->b_offset);
337
338                 for (i = 0; i < bp->b_page_count; i++) {
339                         struct page     *page = bp->b_pages[i];
340
341                         if (bp->b_flags & _XBF_PAGE_CACHE)
342                                 ASSERT(!PagePrivate(page));
343                         page_cache_release(page);
344                 }
345                 _xfs_buf_free_pages(bp);
346         }
347
348         xfs_buf_deallocate(bp);
349 }
350
351 /*
352  *      Finds all pages for buffer in question and builds it's page list.
353  */
354 STATIC int
355 _xfs_buf_lookup_pages(
356         xfs_buf_t               *bp,
357         uint                    flags)
358 {
359         struct address_space    *mapping = bp->b_target->bt_mapping;
360         size_t                  blocksize = bp->b_target->bt_bsize;
361         size_t                  size = bp->b_count_desired;
362         size_t                  nbytes, offset;
363         gfp_t                   gfp_mask = xb_to_gfp(flags);
364         unsigned short          page_count, i;
365         pgoff_t                 first;
366         xfs_off_t               end;
367         int                     error;
368
369         end = bp->b_file_offset + bp->b_buffer_length;
370         page_count = xfs_buf_btoc(end) - xfs_buf_btoct(bp->b_file_offset);
371
372         error = _xfs_buf_get_pages(bp, page_count, flags);
373         if (unlikely(error))
374                 return error;
375         bp->b_flags |= _XBF_PAGE_CACHE;
376
377         offset = bp->b_offset;
378         first = bp->b_file_offset >> PAGE_CACHE_SHIFT;
379
380         for (i = 0; i < bp->b_page_count; i++) {
381                 struct page     *page;
382                 uint            retries = 0;
383
384               retry:
385                 page = find_or_create_page(mapping, first + i, gfp_mask);
386                 if (unlikely(page == NULL)) {
387                         if (flags & XBF_READ_AHEAD) {
388                                 bp->b_page_count = i;
389                                 for (i = 0; i < bp->b_page_count; i++)
390                                         unlock_page(bp->b_pages[i]);
391                                 return -ENOMEM;
392                         }
393
394                         /*
395                          * This could deadlock.
396                          *
397                          * But until all the XFS lowlevel code is revamped to
398                          * handle buffer allocation failures we can't do much.
399                          */
400                         if (!(++retries % 100))
401                                 printk(KERN_ERR
402                                         "XFS: possible memory allocation "
403                                         "deadlock in %s (mode:0x%x)\n",
404                                         __func__, gfp_mask);
405
406                         XFS_STATS_INC(xb_page_retries);
407                         xfsbufd_wakeup(0, gfp_mask);
408                         congestion_wait(WRITE, HZ/50);
409                         goto retry;
410                 }
411
412                 XFS_STATS_INC(xb_page_found);
413
414                 nbytes = min_t(size_t, size, PAGE_CACHE_SIZE - offset);
415                 size -= nbytes;
416
417                 ASSERT(!PagePrivate(page));
418                 if (!PageUptodate(page)) {
419                         page_count--;
420                         if (blocksize >= PAGE_CACHE_SIZE) {
421                                 if (flags & XBF_READ)
422                                         bp->b_flags |= _XBF_PAGE_LOCKED;
423                         } else if (!PagePrivate(page)) {
424                                 if (test_page_region(page, offset, nbytes))
425                                         page_count++;
426                         }
427                 }
428
429                 bp->b_pages[i] = page;
430                 offset = 0;
431         }
432
433         if (!(bp->b_flags & _XBF_PAGE_LOCKED)) {
434                 for (i = 0; i < bp->b_page_count; i++)
435                         unlock_page(bp->b_pages[i]);
436         }
437
438         if (page_count == bp->b_page_count)
439                 bp->b_flags |= XBF_DONE;
440
441         XB_TRACE(bp, "lookup_pages", (long)page_count);
442         return error;
443 }
444
445 /*
446  *      Map buffer into kernel address-space if nessecary.
447  */
448 STATIC int
449 _xfs_buf_map_pages(
450         xfs_buf_t               *bp,
451         uint                    flags)
452 {
453         /* A single page buffer is always mappable */
454         if (bp->b_page_count == 1) {
455                 bp->b_addr = page_address(bp->b_pages[0]) + bp->b_offset;
456                 bp->b_flags |= XBF_MAPPED;
457         } else if (flags & XBF_MAPPED) {
458                 if (as_list_len > 64)
459                         purge_addresses();
460                 bp->b_addr = vmap(bp->b_pages, bp->b_page_count,
461                                         VM_MAP, PAGE_KERNEL);
462                 if (unlikely(bp->b_addr == NULL))
463                         return -ENOMEM;
464                 bp->b_addr += bp->b_offset;
465                 bp->b_flags |= XBF_MAPPED;
466         }
467
468         return 0;
469 }
470
471 /*
472  *      Finding and Reading Buffers
473  */
474
475 /*
476  *      Look up, and creates if absent, a lockable buffer for
477  *      a given range of an inode.  The buffer is returned
478  *      locked.  If other overlapping buffers exist, they are
479  *      released before the new buffer is created and locked,
480  *      which may imply that this call will block until those buffers
481  *      are unlocked.  No I/O is implied by this call.
482  */
483 xfs_buf_t *
484 _xfs_buf_find(
485         xfs_buftarg_t           *btp,   /* block device target          */
486         xfs_off_t               ioff,   /* starting offset of range     */
487         size_t                  isize,  /* length of range              */
488         xfs_buf_flags_t         flags,
489         xfs_buf_t               *new_bp)
490 {
491         xfs_off_t               range_base;
492         size_t                  range_length;
493         xfs_bufhash_t           *hash;
494         xfs_buf_t               *bp, *n;
495
496         range_base = (ioff << BBSHIFT);
497         range_length = (isize << BBSHIFT);
498
499         /* Check for IOs smaller than the sector size / not sector aligned */
500         ASSERT(!(range_length < (1 << btp->bt_sshift)));
501         ASSERT(!(range_base & (xfs_off_t)btp->bt_smask));
502
503         hash = &btp->bt_hash[hash_long((unsigned long)ioff, btp->bt_hashshift)];
504
505         spin_lock(&hash->bh_lock);
506
507         list_for_each_entry_safe(bp, n, &hash->bh_list, b_hash_list) {
508                 ASSERT(btp == bp->b_target);
509                 if (bp->b_file_offset == range_base &&
510                     bp->b_buffer_length == range_length) {
511                         /*
512                          * If we look at something, bring it to the
513                          * front of the list for next time.
514                          */
515                         atomic_inc(&bp->b_hold);
516                         list_move(&bp->b_hash_list, &hash->bh_list);
517                         goto found;
518                 }
519         }
520
521         /* No match found */
522         if (new_bp) {
523                 _xfs_buf_initialize(new_bp, btp, range_base,
524                                 range_length, flags);
525                 new_bp->b_hash = hash;
526                 list_add(&new_bp->b_hash_list, &hash->bh_list);
527         } else {
528                 XFS_STATS_INC(xb_miss_locked);
529         }
530
531         spin_unlock(&hash->bh_lock);
532         return new_bp;
533
534 found:
535         spin_unlock(&hash->bh_lock);
536
537         /* Attempt to get the semaphore without sleeping,
538          * if this does not work then we need to drop the
539          * spinlock and do a hard attempt on the semaphore.
540          */
541         if (down_trylock(&bp->b_sema)) {
542                 if (!(flags & XBF_TRYLOCK)) {
543                         /* wait for buffer ownership */
544                         XB_TRACE(bp, "get_lock", 0);
545                         xfs_buf_lock(bp);
546                         XFS_STATS_INC(xb_get_locked_waited);
547                 } else {
548                         /* We asked for a trylock and failed, no need
549                          * to look at file offset and length here, we
550                          * know that this buffer at least overlaps our
551                          * buffer and is locked, therefore our buffer
552                          * either does not exist, or is this buffer.
553                          */
554                         xfs_buf_rele(bp);
555                         XFS_STATS_INC(xb_busy_locked);
556                         return NULL;
557                 }
558         } else {
559                 /* trylock worked */
560                 XB_SET_OWNER(bp);
561         }
562
563         if (bp->b_flags & XBF_STALE) {
564                 ASSERT((bp->b_flags & _XBF_DELWRI_Q) == 0);
565                 bp->b_flags &= XBF_MAPPED;
566         }
567         XB_TRACE(bp, "got_lock", 0);
568         XFS_STATS_INC(xb_get_locked);
569         return bp;
570 }
571
572 /*
573  *      Assembles a buffer covering the specified range.
574  *      Storage in memory for all portions of the buffer will be allocated,
575  *      although backing storage may not be.
576  */
577 xfs_buf_t *
578 xfs_buf_get_flags(
579         xfs_buftarg_t           *target,/* target for buffer            */
580         xfs_off_t               ioff,   /* starting offset of range     */
581         size_t                  isize,  /* length of range              */
582         xfs_buf_flags_t         flags)
583 {
584         xfs_buf_t               *bp, *new_bp;
585         int                     error = 0, i;
586
587         new_bp = xfs_buf_allocate(flags);
588         if (unlikely(!new_bp))
589                 return NULL;
590
591         bp = _xfs_buf_find(target, ioff, isize, flags, new_bp);
592         if (bp == new_bp) {
593                 error = _xfs_buf_lookup_pages(bp, flags);
594                 if (error)
595                         goto no_buffer;
596         } else {
597                 xfs_buf_deallocate(new_bp);
598                 if (unlikely(bp == NULL))
599                         return NULL;
600         }
601
602         for (i = 0; i < bp->b_page_count; i++)
603                 mark_page_accessed(bp->b_pages[i]);
604
605         if (!(bp->b_flags & XBF_MAPPED)) {
606                 error = _xfs_buf_map_pages(bp, flags);
607                 if (unlikely(error)) {
608                         printk(KERN_WARNING "%s: failed to map pages\n",
609                                         __func__);
610                         goto no_buffer;
611                 }
612         }
613
614         XFS_STATS_INC(xb_get);
615
616         /*
617          * Always fill in the block number now, the mapped cases can do
618          * their own overlay of this later.
619          */
620         bp->b_bn = ioff;
621         bp->b_count_desired = bp->b_buffer_length;
622
623         XB_TRACE(bp, "get", (unsigned long)flags);
624         return bp;
625
626  no_buffer:
627         if (flags & (XBF_LOCK | XBF_TRYLOCK))
628                 xfs_buf_unlock(bp);
629         xfs_buf_rele(bp);
630         return NULL;
631 }
632
633 xfs_buf_t *
634 xfs_buf_read_flags(
635         xfs_buftarg_t           *target,
636         xfs_off_t               ioff,
637         size_t                  isize,
638         xfs_buf_flags_t         flags)
639 {
640         xfs_buf_t               *bp;
641
642         flags |= XBF_READ;
643
644         bp = xfs_buf_get_flags(target, ioff, isize, flags);
645         if (bp) {
646                 if (!XFS_BUF_ISDONE(bp)) {
647                         XB_TRACE(bp, "read", (unsigned long)flags);
648                         XFS_STATS_INC(xb_get_read);
649                         xfs_buf_iostart(bp, flags);
650                 } else if (flags & XBF_ASYNC) {
651                         XB_TRACE(bp, "read_async", (unsigned long)flags);
652                         /*
653                          * Read ahead call which is already satisfied,
654                          * drop the buffer
655                          */
656                         goto no_buffer;
657                 } else {
658                         XB_TRACE(bp, "read_done", (unsigned long)flags);
659                         /* We do not want read in the flags */
660                         bp->b_flags &= ~XBF_READ;
661                 }
662         }
663
664         return bp;
665
666  no_buffer:
667         if (flags & (XBF_LOCK | XBF_TRYLOCK))
668                 xfs_buf_unlock(bp);
669         xfs_buf_rele(bp);
670         return NULL;
671 }
672
673 /*
674  *      If we are not low on memory then do the readahead in a deadlock
675  *      safe manner.
676  */
677 void
678 xfs_buf_readahead(
679         xfs_buftarg_t           *target,
680         xfs_off_t               ioff,
681         size_t                  isize,
682         xfs_buf_flags_t         flags)
683 {
684         struct backing_dev_info *bdi;
685
686         bdi = target->bt_mapping->backing_dev_info;
687         if (bdi_read_congested(bdi))
688                 return;
689
690         flags |= (XBF_TRYLOCK|XBF_ASYNC|XBF_READ_AHEAD);
691         xfs_buf_read_flags(target, ioff, isize, flags);
692 }
693
694 xfs_buf_t *
695 xfs_buf_get_empty(
696         size_t                  len,
697         xfs_buftarg_t           *target)
698 {
699         xfs_buf_t               *bp;
700
701         bp = xfs_buf_allocate(0);
702         if (bp)
703                 _xfs_buf_initialize(bp, target, 0, len, 0);
704         return bp;
705 }
706
707 static inline struct page *
708 mem_to_page(
709         void                    *addr)
710 {
711         if ((!is_vmalloc_addr(addr))) {
712                 return virt_to_page(addr);
713         } else {
714                 return vmalloc_to_page(addr);
715         }
716 }
717
718 int
719 xfs_buf_associate_memory(
720         xfs_buf_t               *bp,
721         void                    *mem,
722         size_t                  len)
723 {
724         int                     rval;
725         int                     i = 0;
726         unsigned long           pageaddr;
727         unsigned long           offset;
728         size_t                  buflen;
729         int                     page_count;
730
731         pageaddr = (unsigned long)mem & PAGE_CACHE_MASK;
732         offset = (unsigned long)mem - pageaddr;
733         buflen = PAGE_CACHE_ALIGN(len + offset);
734         page_count = buflen >> PAGE_CACHE_SHIFT;
735
736         /* Free any previous set of page pointers */
737         if (bp->b_pages)
738                 _xfs_buf_free_pages(bp);
739
740         bp->b_pages = NULL;
741         bp->b_addr = mem;
742
743         rval = _xfs_buf_get_pages(bp, page_count, 0);
744         if (rval)
745                 return rval;
746
747         bp->b_offset = offset;
748
749         for (i = 0; i < bp->b_page_count; i++) {
750                 bp->b_pages[i] = mem_to_page((void *)pageaddr);
751                 pageaddr += PAGE_CACHE_SIZE;
752         }
753
754         bp->b_count_desired = len;
755         bp->b_buffer_length = buflen;
756         bp->b_flags |= XBF_MAPPED;
757         bp->b_flags &= ~_XBF_PAGE_LOCKED;
758
759         return 0;
760 }
761
762 xfs_buf_t *
763 xfs_buf_get_noaddr(
764         size_t                  len,
765         xfs_buftarg_t           *target)
766 {
767         unsigned long           page_count = PAGE_ALIGN(len) >> PAGE_SHIFT;
768         int                     error, i;
769         xfs_buf_t               *bp;
770
771         bp = xfs_buf_allocate(0);
772         if (unlikely(bp == NULL))
773                 goto fail;
774         _xfs_buf_initialize(bp, target, 0, len, 0);
775
776         error = _xfs_buf_get_pages(bp, page_count, 0);
777         if (error)
778                 goto fail_free_buf;
779
780         for (i = 0; i < page_count; i++) {
781                 bp->b_pages[i] = alloc_page(GFP_KERNEL);
782                 if (!bp->b_pages[i])
783                         goto fail_free_mem;
784         }
785         bp->b_flags |= _XBF_PAGES;
786
787         error = _xfs_buf_map_pages(bp, XBF_MAPPED);
788         if (unlikely(error)) {
789                 printk(KERN_WARNING "%s: failed to map pages\n",
790                                 __func__);
791                 goto fail_free_mem;
792         }
793
794         xfs_buf_unlock(bp);
795
796         XB_TRACE(bp, "no_daddr", len);
797         return bp;
798
799  fail_free_mem:
800         while (--i >= 0)
801                 __free_page(bp->b_pages[i]);
802         _xfs_buf_free_pages(bp);
803  fail_free_buf:
804         xfs_buf_deallocate(bp);
805  fail:
806         return NULL;
807 }
808
809 /*
810  *      Increment reference count on buffer, to hold the buffer concurrently
811  *      with another thread which may release (free) the buffer asynchronously.
812  *      Must hold the buffer already to call this function.
813  */
814 void
815 xfs_buf_hold(
816         xfs_buf_t               *bp)
817 {
818         atomic_inc(&bp->b_hold);
819         XB_TRACE(bp, "hold", 0);
820 }
821
822 /*
823  *      Releases a hold on the specified buffer.  If the
824  *      the hold count is 1, calls xfs_buf_free.
825  */
826 void
827 xfs_buf_rele(
828         xfs_buf_t               *bp)
829 {
830         xfs_bufhash_t           *hash = bp->b_hash;
831
832         XB_TRACE(bp, "rele", bp->b_relse);
833
834         if (unlikely(!hash)) {
835                 ASSERT(!bp->b_relse);
836                 if (atomic_dec_and_test(&bp->b_hold))
837                         xfs_buf_free(bp);
838                 return;
839         }
840
841         ASSERT(atomic_read(&bp->b_hold) > 0);
842         if (atomic_dec_and_lock(&bp->b_hold, &hash->bh_lock)) {
843                 if (bp->b_relse) {
844                         atomic_inc(&bp->b_hold);
845                         spin_unlock(&hash->bh_lock);
846                         (*(bp->b_relse)) (bp);
847                 } else if (bp->b_flags & XBF_FS_MANAGED) {
848                         spin_unlock(&hash->bh_lock);
849                 } else {
850                         ASSERT(!(bp->b_flags & (XBF_DELWRI|_XBF_DELWRI_Q)));
851                         list_del_init(&bp->b_hash_list);
852                         spin_unlock(&hash->bh_lock);
853                         xfs_buf_free(bp);
854                 }
855         }
856 }
857
858
859 /*
860  *      Mutual exclusion on buffers.  Locking model:
861  *
862  *      Buffers associated with inodes for which buffer locking
863  *      is not enabled are not protected by semaphores, and are
864  *      assumed to be exclusively owned by the caller.  There is a
865  *      spinlock in the buffer, used by the caller when concurrent
866  *      access is possible.
867  */
868
869 /*
870  *      Locks a buffer object, if it is not already locked.
871  *      Note that this in no way locks the underlying pages, so it is only
872  *      useful for synchronizing concurrent use of buffer objects, not for
873  *      synchronizing independent access to the underlying pages.
874  */
875 int
876 xfs_buf_cond_lock(
877         xfs_buf_t               *bp)
878 {
879         int                     locked;
880
881         locked = down_trylock(&bp->b_sema) == 0;
882         if (locked) {
883                 XB_SET_OWNER(bp);
884         }
885         XB_TRACE(bp, "cond_lock", (long)locked);
886         return locked ? 0 : -EBUSY;
887 }
888
889 #if defined(DEBUG) || defined(XFS_BLI_TRACE)
890 int
891 xfs_buf_lock_value(
892         xfs_buf_t               *bp)
893 {
894         return bp->b_sema.count;
895 }
896 #endif
897
898 /*
899  *      Locks a buffer object.
900  *      Note that this in no way locks the underlying pages, so it is only
901  *      useful for synchronizing concurrent use of buffer objects, not for
902  *      synchronizing independent access to the underlying pages.
903  */
904 void
905 xfs_buf_lock(
906         xfs_buf_t               *bp)
907 {
908         XB_TRACE(bp, "lock", 0);
909         if (atomic_read(&bp->b_io_remaining))
910                 blk_run_address_space(bp->b_target->bt_mapping);
911         down(&bp->b_sema);
912         XB_SET_OWNER(bp);
913         XB_TRACE(bp, "locked", 0);
914 }
915
916 /*
917  *      Releases the lock on the buffer object.
918  *      If the buffer is marked delwri but is not queued, do so before we
919  *      unlock the buffer as we need to set flags correctly.  We also need to
920  *      take a reference for the delwri queue because the unlocker is going to
921  *      drop their's and they don't know we just queued it.
922  */
923 void
924 xfs_buf_unlock(
925         xfs_buf_t               *bp)
926 {
927         if ((bp->b_flags & (XBF_DELWRI|_XBF_DELWRI_Q)) == XBF_DELWRI) {
928                 atomic_inc(&bp->b_hold);
929                 bp->b_flags |= XBF_ASYNC;
930                 xfs_buf_delwri_queue(bp, 0);
931         }
932
933         XB_CLEAR_OWNER(bp);
934         up(&bp->b_sema);
935         XB_TRACE(bp, "unlock", 0);
936 }
937
938
939 /*
940  *      Pinning Buffer Storage in Memory
941  *      Ensure that no attempt to force a buffer to disk will succeed.
942  */
943 void
944 xfs_buf_pin(
945         xfs_buf_t               *bp)
946 {
947         atomic_inc(&bp->b_pin_count);
948         XB_TRACE(bp, "pin", (long)bp->b_pin_count.counter);
949 }
950
951 void
952 xfs_buf_unpin(
953         xfs_buf_t               *bp)
954 {
955         if (atomic_dec_and_test(&bp->b_pin_count))
956                 wake_up_all(&bp->b_waiters);
957         XB_TRACE(bp, "unpin", (long)bp->b_pin_count.counter);
958 }
959
960 int
961 xfs_buf_ispin(
962         xfs_buf_t               *bp)
963 {
964         return atomic_read(&bp->b_pin_count);
965 }
966
967 STATIC void
968 xfs_buf_wait_unpin(
969         xfs_buf_t               *bp)
970 {
971         DECLARE_WAITQUEUE       (wait, current);
972
973         if (atomic_read(&bp->b_pin_count) == 0)
974                 return;
975
976         add_wait_queue(&bp->b_waiters, &wait);
977         for (;;) {
978                 set_current_state(TASK_UNINTERRUPTIBLE);
979                 if (atomic_read(&bp->b_pin_count) == 0)
980                         break;
981                 if (atomic_read(&bp->b_io_remaining))
982                         blk_run_address_space(bp->b_target->bt_mapping);
983                 schedule();
984         }
985         remove_wait_queue(&bp->b_waiters, &wait);
986         set_current_state(TASK_RUNNING);
987 }
988
989 /*
990  *      Buffer Utility Routines
991  */
992
993 STATIC void
994 xfs_buf_iodone_work(
995         struct work_struct      *work)
996 {
997         xfs_buf_t               *bp =
998                 container_of(work, xfs_buf_t, b_iodone_work);
999
1000         /*
1001          * We can get an EOPNOTSUPP to ordered writes.  Here we clear the
1002          * ordered flag and reissue them.  Because we can't tell the higher
1003          * layers directly that they should not issue ordered I/O anymore, they
1004          * need to check if the _XFS_BARRIER_FAILED flag was set during I/O completion.
1005          */
1006         if ((bp->b_error == EOPNOTSUPP) &&
1007             (bp->b_flags & (XBF_ORDERED|XBF_ASYNC)) == (XBF_ORDERED|XBF_ASYNC)) {
1008                 XB_TRACE(bp, "ordered_retry", bp->b_iodone);
1009                 bp->b_flags &= ~XBF_ORDERED;
1010                 bp->b_flags |= _XFS_BARRIER_FAILED;
1011                 xfs_buf_iorequest(bp);
1012         } else if (bp->b_iodone)
1013                 (*(bp->b_iodone))(bp);
1014         else if (bp->b_flags & XBF_ASYNC)
1015                 xfs_buf_relse(bp);
1016 }
1017
1018 void
1019 xfs_buf_ioend(
1020         xfs_buf_t               *bp,
1021         int                     schedule)
1022 {
1023         bp->b_flags &= ~(XBF_READ | XBF_WRITE | XBF_READ_AHEAD);
1024         if (bp->b_error == 0)
1025                 bp->b_flags |= XBF_DONE;
1026
1027         XB_TRACE(bp, "iodone", bp->b_iodone);
1028
1029         if ((bp->b_iodone) || (bp->b_flags & XBF_ASYNC)) {
1030                 if (schedule) {
1031                         INIT_WORK(&bp->b_iodone_work, xfs_buf_iodone_work);
1032                         queue_work(xfslogd_workqueue, &bp->b_iodone_work);
1033                 } else {
1034                         xfs_buf_iodone_work(&bp->b_iodone_work);
1035                 }
1036         } else {
1037                 complete(&bp->b_iowait);
1038         }
1039 }
1040
1041 void
1042 xfs_buf_ioerror(
1043         xfs_buf_t               *bp,
1044         int                     error)
1045 {
1046         ASSERT(error >= 0 && error <= 0xffff);
1047         bp->b_error = (unsigned short)error;
1048         XB_TRACE(bp, "ioerror", (unsigned long)error);
1049 }
1050
1051 /*
1052  *      Initiate I/O on a buffer, based on the flags supplied.
1053  *      The b_iodone routine in the buffer supplied will only be called
1054  *      when all of the subsidiary I/O requests, if any, have been completed.
1055  */
1056 int
1057 xfs_buf_iostart(
1058         xfs_buf_t               *bp,
1059         xfs_buf_flags_t         flags)
1060 {
1061         int                     status = 0;
1062
1063         XB_TRACE(bp, "iostart", (unsigned long)flags);
1064
1065         if (flags & XBF_DELWRI) {
1066                 bp->b_flags &= ~(XBF_READ | XBF_WRITE | XBF_ASYNC);
1067                 bp->b_flags |= flags & (XBF_DELWRI | XBF_ASYNC);
1068                 xfs_buf_delwri_queue(bp, 1);
1069                 return 0;
1070         }
1071
1072         bp->b_flags &= ~(XBF_READ | XBF_WRITE | XBF_ASYNC | XBF_DELWRI | \
1073                         XBF_READ_AHEAD | _XBF_RUN_QUEUES);
1074         bp->b_flags |= flags & (XBF_READ | XBF_WRITE | XBF_ASYNC | \
1075                         XBF_READ_AHEAD | _XBF_RUN_QUEUES);
1076
1077         BUG_ON(bp->b_bn == XFS_BUF_DADDR_NULL);
1078
1079         /* For writes allow an alternate strategy routine to precede
1080          * the actual I/O request (which may not be issued at all in
1081          * a shutdown situation, for example).
1082          */
1083         status = (flags & XBF_WRITE) ?
1084                 xfs_buf_iostrategy(bp) : xfs_buf_iorequest(bp);
1085
1086         /* Wait for I/O if we are not an async request.
1087          * Note: async I/O request completion will release the buffer,
1088          * and that can already be done by this point.  So using the
1089          * buffer pointer from here on, after async I/O, is invalid.
1090          */
1091         if (!status && !(flags & XBF_ASYNC))
1092                 status = xfs_buf_iowait(bp);
1093
1094         return status;
1095 }
1096
1097 STATIC_INLINE void
1098 _xfs_buf_ioend(
1099         xfs_buf_t               *bp,
1100         int                     schedule)
1101 {
1102         if (atomic_dec_and_test(&bp->b_io_remaining) == 1) {
1103                 bp->b_flags &= ~_XBF_PAGE_LOCKED;
1104                 xfs_buf_ioend(bp, schedule);
1105         }
1106 }
1107
1108 STATIC void
1109 xfs_buf_bio_end_io(
1110         struct bio              *bio,
1111         int                     error)
1112 {
1113         xfs_buf_t               *bp = (xfs_buf_t *)bio->bi_private;
1114         unsigned int            blocksize = bp->b_target->bt_bsize;
1115         struct bio_vec          *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
1116
1117         if (!test_bit(BIO_UPTODATE, &bio->bi_flags))
1118                 bp->b_error = EIO;
1119
1120         do {
1121                 struct page     *page = bvec->bv_page;
1122
1123                 ASSERT(!PagePrivate(page));
1124                 if (unlikely(bp->b_error)) {
1125                         if (bp->b_flags & XBF_READ)
1126                                 ClearPageUptodate(page);
1127                 } else if (blocksize >= PAGE_CACHE_SIZE) {
1128                         SetPageUptodate(page);
1129                 } else if (!PagePrivate(page) &&
1130                                 (bp->b_flags & _XBF_PAGE_CACHE)) {
1131                         set_page_region(page, bvec->bv_offset, bvec->bv_len);
1132                 }
1133
1134                 if (--bvec >= bio->bi_io_vec)
1135                         prefetchw(&bvec->bv_page->flags);
1136
1137                 if (bp->b_flags & _XBF_PAGE_LOCKED)
1138                         unlock_page(page);
1139         } while (bvec >= bio->bi_io_vec);
1140
1141         _xfs_buf_ioend(bp, 1);
1142         bio_put(bio);
1143 }
1144
1145 STATIC void
1146 _xfs_buf_ioapply(
1147         xfs_buf_t               *bp)
1148 {
1149         int                     rw, map_i, total_nr_pages, nr_pages;
1150         struct bio              *bio;
1151         int                     offset = bp->b_offset;
1152         int                     size = bp->b_count_desired;
1153         sector_t                sector = bp->b_bn;
1154         unsigned int            blocksize = bp->b_target->bt_bsize;
1155
1156         total_nr_pages = bp->b_page_count;
1157         map_i = 0;
1158
1159         if (bp->b_flags & XBF_ORDERED) {
1160                 ASSERT(!(bp->b_flags & XBF_READ));
1161                 rw = WRITE_BARRIER;
1162         } else if (bp->b_flags & _XBF_RUN_QUEUES) {
1163                 ASSERT(!(bp->b_flags & XBF_READ_AHEAD));
1164                 bp->b_flags &= ~_XBF_RUN_QUEUES;
1165                 rw = (bp->b_flags & XBF_WRITE) ? WRITE_SYNC : READ_SYNC;
1166         } else {
1167                 rw = (bp->b_flags & XBF_WRITE) ? WRITE :
1168                      (bp->b_flags & XBF_READ_AHEAD) ? READA : READ;
1169         }
1170
1171         /* Special code path for reading a sub page size buffer in --
1172          * we populate up the whole page, and hence the other metadata
1173          * in the same page.  This optimization is only valid when the
1174          * filesystem block size is not smaller than the page size.
1175          */
1176         if ((bp->b_buffer_length < PAGE_CACHE_SIZE) &&
1177             ((bp->b_flags & (XBF_READ|_XBF_PAGE_LOCKED)) ==
1178               (XBF_READ|_XBF_PAGE_LOCKED)) &&
1179             (blocksize >= PAGE_CACHE_SIZE)) {
1180                 bio = bio_alloc(GFP_NOIO, 1);
1181
1182                 bio->bi_bdev = bp->b_target->bt_bdev;
1183                 bio->bi_sector = sector - (offset >> BBSHIFT);
1184                 bio->bi_end_io = xfs_buf_bio_end_io;
1185                 bio->bi_private = bp;
1186
1187                 bio_add_page(bio, bp->b_pages[0], PAGE_CACHE_SIZE, 0);
1188                 size = 0;
1189
1190                 atomic_inc(&bp->b_io_remaining);
1191
1192                 goto submit_io;
1193         }
1194
1195 next_chunk:
1196         atomic_inc(&bp->b_io_remaining);
1197         nr_pages = BIO_MAX_SECTORS >> (PAGE_SHIFT - BBSHIFT);
1198         if (nr_pages > total_nr_pages)
1199                 nr_pages = total_nr_pages;
1200
1201         bio = bio_alloc(GFP_NOIO, nr_pages);
1202         bio->bi_bdev = bp->b_target->bt_bdev;
1203         bio->bi_sector = sector;
1204         bio->bi_end_io = xfs_buf_bio_end_io;
1205         bio->bi_private = bp;
1206
1207         for (; size && nr_pages; nr_pages--, map_i++) {
1208                 int     rbytes, nbytes = PAGE_CACHE_SIZE - offset;
1209
1210                 if (nbytes > size)
1211                         nbytes = size;
1212
1213                 rbytes = bio_add_page(bio, bp->b_pages[map_i], nbytes, offset);
1214                 if (rbytes < nbytes)
1215                         break;
1216
1217                 offset = 0;
1218                 sector += nbytes >> BBSHIFT;
1219                 size -= nbytes;
1220                 total_nr_pages--;
1221         }
1222
1223 submit_io:
1224         if (likely(bio->bi_size)) {
1225                 submit_bio(rw, bio);
1226                 if (size)
1227                         goto next_chunk;
1228         } else {
1229                 bio_put(bio);
1230                 xfs_buf_ioerror(bp, EIO);
1231         }
1232 }
1233
1234 int
1235 xfs_buf_iorequest(
1236         xfs_buf_t               *bp)
1237 {
1238         XB_TRACE(bp, "iorequest", 0);
1239
1240         if (bp->b_flags & XBF_DELWRI) {
1241                 xfs_buf_delwri_queue(bp, 1);
1242                 return 0;
1243         }
1244
1245         if (bp->b_flags & XBF_WRITE) {
1246                 xfs_buf_wait_unpin(bp);
1247         }
1248
1249         xfs_buf_hold(bp);
1250
1251         /* Set the count to 1 initially, this will stop an I/O
1252          * completion callout which happens before we have started
1253          * all the I/O from calling xfs_buf_ioend too early.
1254          */
1255         atomic_set(&bp->b_io_remaining, 1);
1256         _xfs_buf_ioapply(bp);
1257         _xfs_buf_ioend(bp, 0);
1258
1259         xfs_buf_rele(bp);
1260         return 0;
1261 }
1262
1263 /*
1264  *      Waits for I/O to complete on the buffer supplied.
1265  *      It returns immediately if no I/O is pending.
1266  *      It returns the I/O error code, if any, or 0 if there was no error.
1267  */
1268 int
1269 xfs_buf_iowait(
1270         xfs_buf_t               *bp)
1271 {
1272         XB_TRACE(bp, "iowait", 0);
1273         if (atomic_read(&bp->b_io_remaining))
1274                 blk_run_address_space(bp->b_target->bt_mapping);
1275         wait_for_completion(&bp->b_iowait);
1276         XB_TRACE(bp, "iowaited", (long)bp->b_error);
1277         return bp->b_error;
1278 }
1279
1280 xfs_caddr_t
1281 xfs_buf_offset(
1282         xfs_buf_t               *bp,
1283         size_t                  offset)
1284 {
1285         struct page             *page;
1286
1287         if (bp->b_flags & XBF_MAPPED)
1288                 return XFS_BUF_PTR(bp) + offset;
1289
1290         offset += bp->b_offset;
1291         page = bp->b_pages[offset >> PAGE_CACHE_SHIFT];
1292         return (xfs_caddr_t)page_address(page) + (offset & (PAGE_CACHE_SIZE-1));
1293 }
1294
1295 /*
1296  *      Move data into or out of a buffer.
1297  */
1298 void
1299 xfs_buf_iomove(
1300         xfs_buf_t               *bp,    /* buffer to process            */
1301         size_t                  boff,   /* starting buffer offset       */
1302         size_t                  bsize,  /* length to copy               */
1303         caddr_t                 data,   /* data address                 */
1304         xfs_buf_rw_t            mode)   /* read/write/zero flag         */
1305 {
1306         size_t                  bend, cpoff, csize;
1307         struct page             *page;
1308
1309         bend = boff + bsize;
1310         while (boff < bend) {
1311                 page = bp->b_pages[xfs_buf_btoct(boff + bp->b_offset)];
1312                 cpoff = xfs_buf_poff(boff + bp->b_offset);
1313                 csize = min_t(size_t,
1314                               PAGE_CACHE_SIZE-cpoff, bp->b_count_desired-boff);
1315
1316                 ASSERT(((csize + cpoff) <= PAGE_CACHE_SIZE));
1317
1318                 switch (mode) {
1319                 case XBRW_ZERO:
1320                         memset(page_address(page) + cpoff, 0, csize);
1321                         break;
1322                 case XBRW_READ:
1323                         memcpy(data, page_address(page) + cpoff, csize);
1324                         break;
1325                 case XBRW_WRITE:
1326                         memcpy(page_address(page) + cpoff, data, csize);
1327                 }
1328
1329                 boff += csize;
1330                 data += csize;
1331         }
1332 }
1333
1334 /*
1335  *      Handling of buffer targets (buftargs).
1336  */
1337
1338 /*
1339  *      Wait for any bufs with callbacks that have been submitted but
1340  *      have not yet returned... walk the hash list for the target.
1341  */
1342 void
1343 xfs_wait_buftarg(
1344         xfs_buftarg_t   *btp)
1345 {
1346         xfs_buf_t       *bp, *n;
1347         xfs_bufhash_t   *hash;
1348         uint            i;
1349
1350         for (i = 0; i < (1 << btp->bt_hashshift); i++) {
1351                 hash = &btp->bt_hash[i];
1352 again:
1353                 spin_lock(&hash->bh_lock);
1354                 list_for_each_entry_safe(bp, n, &hash->bh_list, b_hash_list) {
1355                         ASSERT(btp == bp->b_target);
1356                         if (!(bp->b_flags & XBF_FS_MANAGED)) {
1357                                 spin_unlock(&hash->bh_lock);
1358                                 /*
1359                                  * Catch superblock reference count leaks
1360                                  * immediately
1361                                  */
1362                                 BUG_ON(bp->b_bn == 0);
1363                                 delay(100);
1364                                 goto again;
1365                         }
1366                 }
1367                 spin_unlock(&hash->bh_lock);
1368         }
1369 }
1370
1371 /*
1372  *      Allocate buffer hash table for a given target.
1373  *      For devices containing metadata (i.e. not the log/realtime devices)
1374  *      we need to allocate a much larger hash table.
1375  */
1376 STATIC void
1377 xfs_alloc_bufhash(
1378         xfs_buftarg_t           *btp,
1379         int                     external)
1380 {
1381         unsigned int            i;
1382
1383         btp->bt_hashshift = external ? 3 : 8;   /* 8 or 256 buckets */
1384         btp->bt_hashmask = (1 << btp->bt_hashshift) - 1;
1385         btp->bt_hash = kmem_zalloc((1 << btp->bt_hashshift) *
1386                                         sizeof(xfs_bufhash_t), KM_SLEEP | KM_LARGE);
1387         for (i = 0; i < (1 << btp->bt_hashshift); i++) {
1388                 spin_lock_init(&btp->bt_hash[i].bh_lock);
1389                 INIT_LIST_HEAD(&btp->bt_hash[i].bh_list);
1390         }
1391 }
1392
1393 STATIC void
1394 xfs_free_bufhash(
1395         xfs_buftarg_t           *btp)
1396 {
1397         kmem_free(btp->bt_hash);
1398         btp->bt_hash = NULL;
1399 }
1400
1401 /*
1402  *      buftarg list for delwrite queue processing
1403  */
1404 static LIST_HEAD(xfs_buftarg_list);
1405 static DEFINE_SPINLOCK(xfs_buftarg_lock);
1406
1407 STATIC void
1408 xfs_register_buftarg(
1409         xfs_buftarg_t           *btp)
1410 {
1411         spin_lock(&xfs_buftarg_lock);
1412         list_add(&btp->bt_list, &xfs_buftarg_list);
1413         spin_unlock(&xfs_buftarg_lock);
1414 }
1415
1416 STATIC void
1417 xfs_unregister_buftarg(
1418         xfs_buftarg_t           *btp)
1419 {
1420         spin_lock(&xfs_buftarg_lock);
1421         list_del(&btp->bt_list);
1422         spin_unlock(&xfs_buftarg_lock);
1423 }
1424
1425 void
1426 xfs_free_buftarg(
1427         xfs_buftarg_t           *btp)
1428 {
1429         xfs_flush_buftarg(btp, 1);
1430         xfs_blkdev_issue_flush(btp);
1431         xfs_free_bufhash(btp);
1432         iput(btp->bt_mapping->host);
1433
1434         /* Unregister the buftarg first so that we don't get a
1435          * wakeup finding a non-existent task
1436          */
1437         xfs_unregister_buftarg(btp);
1438         kthread_stop(btp->bt_task);
1439
1440         kmem_free(btp);
1441 }
1442
1443 STATIC int
1444 xfs_setsize_buftarg_flags(
1445         xfs_buftarg_t           *btp,
1446         unsigned int            blocksize,
1447         unsigned int            sectorsize,
1448         int                     verbose)
1449 {
1450         btp->bt_bsize = blocksize;
1451         btp->bt_sshift = ffs(sectorsize) - 1;
1452         btp->bt_smask = sectorsize - 1;
1453
1454         if (set_blocksize(btp->bt_bdev, sectorsize)) {
1455                 printk(KERN_WARNING
1456                         "XFS: Cannot set_blocksize to %u on device %s\n",
1457                         sectorsize, XFS_BUFTARG_NAME(btp));
1458                 return EINVAL;
1459         }
1460
1461         if (verbose &&
1462             (PAGE_CACHE_SIZE / BITS_PER_LONG) > sectorsize) {
1463                 printk(KERN_WARNING
1464                         "XFS: %u byte sectors in use on device %s.  "
1465                         "This is suboptimal; %u or greater is ideal.\n",
1466                         sectorsize, XFS_BUFTARG_NAME(btp),
1467                         (unsigned int)PAGE_CACHE_SIZE / BITS_PER_LONG);
1468         }
1469
1470         return 0;
1471 }
1472
1473 /*
1474  *      When allocating the initial buffer target we have not yet
1475  *      read in the superblock, so don't know what sized sectors
1476  *      are being used is at this early stage.  Play safe.
1477  */
1478 STATIC int
1479 xfs_setsize_buftarg_early(
1480         xfs_buftarg_t           *btp,
1481         struct block_device     *bdev)
1482 {
1483         return xfs_setsize_buftarg_flags(btp,
1484                         PAGE_CACHE_SIZE, bdev_hardsect_size(bdev), 0);
1485 }
1486
1487 int
1488 xfs_setsize_buftarg(
1489         xfs_buftarg_t           *btp,
1490         unsigned int            blocksize,
1491         unsigned int            sectorsize)
1492 {
1493         return xfs_setsize_buftarg_flags(btp, blocksize, sectorsize, 1);
1494 }
1495
1496 STATIC int
1497 xfs_mapping_buftarg(
1498         xfs_buftarg_t           *btp,
1499         struct block_device     *bdev)
1500 {
1501         struct backing_dev_info *bdi;
1502         struct inode            *inode;
1503         struct address_space    *mapping;
1504         static const struct address_space_operations mapping_aops = {
1505                 .sync_page = block_sync_page,
1506                 .migratepage = fail_migrate_page,
1507         };
1508
1509         inode = new_inode(bdev->bd_inode->i_sb);
1510         if (!inode) {
1511                 printk(KERN_WARNING
1512                         "XFS: Cannot allocate mapping inode for device %s\n",
1513                         XFS_BUFTARG_NAME(btp));
1514                 return ENOMEM;
1515         }
1516         inode->i_mode = S_IFBLK;
1517         inode->i_bdev = bdev;
1518         inode->i_rdev = bdev->bd_dev;
1519         bdi = blk_get_backing_dev_info(bdev);
1520         if (!bdi)
1521                 bdi = &default_backing_dev_info;
1522         mapping = &inode->i_data;
1523         mapping->a_ops = &mapping_aops;
1524         mapping->backing_dev_info = bdi;
1525         mapping_set_gfp_mask(mapping, GFP_NOFS);
1526         btp->bt_mapping = mapping;
1527         return 0;
1528 }
1529
1530 STATIC int
1531 xfs_alloc_delwrite_queue(
1532         xfs_buftarg_t           *btp)
1533 {
1534         int     error = 0;
1535
1536         INIT_LIST_HEAD(&btp->bt_list);
1537         INIT_LIST_HEAD(&btp->bt_delwrite_queue);
1538         spin_lock_init(&btp->bt_delwrite_lock);
1539         btp->bt_flags = 0;
1540         btp->bt_task = kthread_run(xfsbufd, btp, "xfsbufd");
1541         if (IS_ERR(btp->bt_task)) {
1542                 error = PTR_ERR(btp->bt_task);
1543                 goto out_error;
1544         }
1545         xfs_register_buftarg(btp);
1546 out_error:
1547         return error;
1548 }
1549
1550 xfs_buftarg_t *
1551 xfs_alloc_buftarg(
1552         struct block_device     *bdev,
1553         int                     external)
1554 {
1555         xfs_buftarg_t           *btp;
1556
1557         btp = kmem_zalloc(sizeof(*btp), KM_SLEEP);
1558
1559         btp->bt_dev =  bdev->bd_dev;
1560         btp->bt_bdev = bdev;
1561         if (xfs_setsize_buftarg_early(btp, bdev))
1562                 goto error;
1563         if (xfs_mapping_buftarg(btp, bdev))
1564                 goto error;
1565         if (xfs_alloc_delwrite_queue(btp))
1566                 goto error;
1567         xfs_alloc_bufhash(btp, external);
1568         return btp;
1569
1570 error:
1571         kmem_free(btp);
1572         return NULL;
1573 }
1574
1575
1576 /*
1577  *      Delayed write buffer handling
1578  */
1579 STATIC void
1580 xfs_buf_delwri_queue(
1581         xfs_buf_t               *bp,
1582         int                     unlock)
1583 {
1584         struct list_head        *dwq = &bp->b_target->bt_delwrite_queue;
1585         spinlock_t              *dwlk = &bp->b_target->bt_delwrite_lock;
1586
1587         XB_TRACE(bp, "delwri_q", (long)unlock);
1588         ASSERT((bp->b_flags&(XBF_DELWRI|XBF_ASYNC)) == (XBF_DELWRI|XBF_ASYNC));
1589
1590         spin_lock(dwlk);
1591         /* If already in the queue, dequeue and place at tail */
1592         if (!list_empty(&bp->b_list)) {
1593                 ASSERT(bp->b_flags & _XBF_DELWRI_Q);
1594                 if (unlock)
1595                         atomic_dec(&bp->b_hold);
1596                 list_del(&bp->b_list);
1597         }
1598
1599         bp->b_flags |= _XBF_DELWRI_Q;
1600         list_add_tail(&bp->b_list, dwq);
1601         bp->b_queuetime = jiffies;
1602         spin_unlock(dwlk);
1603
1604         if (unlock)
1605                 xfs_buf_unlock(bp);
1606 }
1607
1608 void
1609 xfs_buf_delwri_dequeue(
1610         xfs_buf_t               *bp)
1611 {
1612         spinlock_t              *dwlk = &bp->b_target->bt_delwrite_lock;
1613         int                     dequeued = 0;
1614
1615         spin_lock(dwlk);
1616         if ((bp->b_flags & XBF_DELWRI) && !list_empty(&bp->b_list)) {
1617                 ASSERT(bp->b_flags & _XBF_DELWRI_Q);
1618                 list_del_init(&bp->b_list);
1619                 dequeued = 1;
1620         }
1621         bp->b_flags &= ~(XBF_DELWRI|_XBF_DELWRI_Q);
1622         spin_unlock(dwlk);
1623
1624         if (dequeued)
1625                 xfs_buf_rele(bp);
1626
1627         XB_TRACE(bp, "delwri_dq", (long)dequeued);
1628 }
1629
1630 STATIC void
1631 xfs_buf_runall_queues(
1632         struct workqueue_struct *queue)
1633 {
1634         flush_workqueue(queue);
1635 }
1636
1637 STATIC int
1638 xfsbufd_wakeup(
1639         int                     priority,
1640         gfp_t                   mask)
1641 {
1642         xfs_buftarg_t           *btp;
1643
1644         spin_lock(&xfs_buftarg_lock);
1645         list_for_each_entry(btp, &xfs_buftarg_list, bt_list) {
1646                 if (test_bit(XBT_FORCE_SLEEP, &btp->bt_flags))
1647                         continue;
1648                 set_bit(XBT_FORCE_FLUSH, &btp->bt_flags);
1649                 wake_up_process(btp->bt_task);
1650         }
1651         spin_unlock(&xfs_buftarg_lock);
1652         return 0;
1653 }
1654
1655 /*
1656  * Move as many buffers as specified to the supplied list
1657  * idicating if we skipped any buffers to prevent deadlocks.
1658  */
1659 STATIC int
1660 xfs_buf_delwri_split(
1661         xfs_buftarg_t   *target,
1662         struct list_head *list,
1663         unsigned long   age)
1664 {
1665         xfs_buf_t       *bp, *n;
1666         struct list_head *dwq = &target->bt_delwrite_queue;
1667         spinlock_t      *dwlk = &target->bt_delwrite_lock;
1668         int             skipped = 0;
1669         int             force;
1670
1671         force = test_and_clear_bit(XBT_FORCE_FLUSH, &target->bt_flags);
1672         INIT_LIST_HEAD(list);
1673         spin_lock(dwlk);
1674         list_for_each_entry_safe(bp, n, dwq, b_list) {
1675                 XB_TRACE(bp, "walkq1", (long)xfs_buf_ispin(bp));
1676                 ASSERT(bp->b_flags & XBF_DELWRI);
1677
1678                 if (!xfs_buf_ispin(bp) && !xfs_buf_cond_lock(bp)) {
1679                         if (!force &&
1680                             time_before(jiffies, bp->b_queuetime + age)) {
1681                                 xfs_buf_unlock(bp);
1682                                 break;
1683                         }
1684
1685                         bp->b_flags &= ~(XBF_DELWRI|_XBF_DELWRI_Q|
1686                                          _XBF_RUN_QUEUES);
1687                         bp->b_flags |= XBF_WRITE;
1688                         list_move_tail(&bp->b_list, list);
1689                 } else
1690                         skipped++;
1691         }
1692         spin_unlock(dwlk);
1693
1694         return skipped;
1695
1696 }
1697
1698 STATIC int
1699 xfsbufd(
1700         void            *data)
1701 {
1702         struct list_head tmp;
1703         xfs_buftarg_t   *target = (xfs_buftarg_t *)data;
1704         int             count;
1705         xfs_buf_t       *bp;
1706
1707         current->flags |= PF_MEMALLOC;
1708
1709         set_freezable();
1710
1711         do {
1712                 if (unlikely(freezing(current))) {
1713                         set_bit(XBT_FORCE_SLEEP, &target->bt_flags);
1714                         refrigerator();
1715                 } else {
1716                         clear_bit(XBT_FORCE_SLEEP, &target->bt_flags);
1717                 }
1718
1719                 schedule_timeout_interruptible(
1720                         xfs_buf_timer_centisecs * msecs_to_jiffies(10));
1721
1722                 xfs_buf_delwri_split(target, &tmp,
1723                                 xfs_buf_age_centisecs * msecs_to_jiffies(10));
1724
1725                 count = 0;
1726                 while (!list_empty(&tmp)) {
1727                         bp = list_entry(tmp.next, xfs_buf_t, b_list);
1728                         ASSERT(target == bp->b_target);
1729
1730                         list_del_init(&bp->b_list);
1731                         xfs_buf_iostrategy(bp);
1732                         count++;
1733                 }
1734
1735                 if (as_list_len > 0)
1736                         purge_addresses();
1737                 if (count)
1738                         blk_run_address_space(target->bt_mapping);
1739
1740         } while (!kthread_should_stop());
1741
1742         return 0;
1743 }
1744
1745 /*
1746  *      Go through all incore buffers, and release buffers if they belong to
1747  *      the given device. This is used in filesystem error handling to
1748  *      preserve the consistency of its metadata.
1749  */
1750 int
1751 xfs_flush_buftarg(
1752         xfs_buftarg_t   *target,
1753         int             wait)
1754 {
1755         struct list_head tmp;
1756         xfs_buf_t       *bp, *n;
1757         int             pincount = 0;
1758
1759         xfs_buf_runall_queues(xfsdatad_workqueue);
1760         xfs_buf_runall_queues(xfslogd_workqueue);
1761
1762         set_bit(XBT_FORCE_FLUSH, &target->bt_flags);
1763         pincount = xfs_buf_delwri_split(target, &tmp, 0);
1764
1765         /*
1766          * Dropped the delayed write list lock, now walk the temporary list
1767          */
1768         list_for_each_entry_safe(bp, n, &tmp, b_list) {
1769                 ASSERT(target == bp->b_target);
1770                 if (wait)
1771                         bp->b_flags &= ~XBF_ASYNC;
1772                 else
1773                         list_del_init(&bp->b_list);
1774
1775                 xfs_buf_iostrategy(bp);
1776         }
1777
1778         if (wait)
1779                 blk_run_address_space(target->bt_mapping);
1780
1781         /*
1782          * Remaining list items must be flushed before returning
1783          */
1784         while (!list_empty(&tmp)) {
1785                 bp = list_entry(tmp.next, xfs_buf_t, b_list);
1786
1787                 list_del_init(&bp->b_list);
1788                 xfs_iowait(bp);
1789                 xfs_buf_relse(bp);
1790         }
1791
1792         return pincount;
1793 }
1794
1795 int __init
1796 xfs_buf_init(void)
1797 {
1798 #ifdef XFS_BUF_TRACE
1799         xfs_buf_trace_buf = ktrace_alloc(XFS_BUF_TRACE_SIZE, KM_NOFS);
1800 #endif
1801
1802         xfs_buf_zone = kmem_zone_init_flags(sizeof(xfs_buf_t), "xfs_buf",
1803                                                 KM_ZONE_HWALIGN, NULL);
1804         if (!xfs_buf_zone)
1805                 goto out_free_trace_buf;
1806
1807         xfslogd_workqueue = create_workqueue("xfslogd");
1808         if (!xfslogd_workqueue)
1809                 goto out_free_buf_zone;
1810
1811         xfsdatad_workqueue = create_workqueue("xfsdatad");
1812         if (!xfsdatad_workqueue)
1813                 goto out_destroy_xfslogd_workqueue;
1814
1815         register_shrinker(&xfs_buf_shake);
1816         return 0;
1817
1818  out_destroy_xfslogd_workqueue:
1819         destroy_workqueue(xfslogd_workqueue);
1820  out_free_buf_zone:
1821         kmem_zone_destroy(xfs_buf_zone);
1822  out_free_trace_buf:
1823 #ifdef XFS_BUF_TRACE
1824         ktrace_free(xfs_buf_trace_buf);
1825 #endif
1826         return -ENOMEM;
1827 }
1828
1829 void
1830 xfs_buf_terminate(void)
1831 {
1832         unregister_shrinker(&xfs_buf_shake);
1833         destroy_workqueue(xfsdatad_workqueue);
1834         destroy_workqueue(xfslogd_workqueue);
1835         kmem_zone_destroy(xfs_buf_zone);
1836 #ifdef XFS_BUF_TRACE
1837         ktrace_free(xfs_buf_trace_buf);
1838 #endif
1839 }
1840
1841 #ifdef CONFIG_KDB_MODULES
1842 struct list_head *
1843 xfs_get_buftarg_list(void)
1844 {
1845         return &xfs_buftarg_list;
1846 }
1847 #endif