[XFS] XFS propagates MS_NOATIME through two levels internally but doesn't
[linux-2.6] / fs / xfs / linux-2.6 / xfs_buf.c
1 /*
2  * Copyright (c) 2000-2005 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 <linux/stddef.h>
19 #include <linux/errno.h>
20 #include <linux/slab.h>
21 #include <linux/pagemap.h>
22 #include <linux/init.h>
23 #include <linux/vmalloc.h>
24 #include <linux/bio.h>
25 #include <linux/sysctl.h>
26 #include <linux/proc_fs.h>
27 #include <linux/workqueue.h>
28 #include <linux/percpu.h>
29 #include <linux/blkdev.h>
30 #include <linux/hash.h>
31 #include <linux/kthread.h>
32 #include "xfs_linux.h"
33
34 STATIC kmem_zone_t *xfs_buf_zone;
35 STATIC kmem_shaker_t xfs_buf_shake;
36 STATIC int xfsbufd(void *);
37 STATIC int xfsbufd_wakeup(int, gfp_t);
38 STATIC void xfs_buf_delwri_queue(xfs_buf_t *, int);
39
40 STATIC struct workqueue_struct *xfslogd_workqueue;
41 struct workqueue_struct *xfsdatad_workqueue;
42
43 #ifdef XFS_BUF_TRACE
44 void
45 xfs_buf_trace(
46         xfs_buf_t       *bp,
47         char            *id,
48         void            *data,
49         void            *ra)
50 {
51         ktrace_enter(xfs_buf_trace_buf,
52                 bp, id,
53                 (void *)(unsigned long)bp->b_flags,
54                 (void *)(unsigned long)bp->b_hold.counter,
55                 (void *)(unsigned long)bp->b_sema.count.counter,
56                 (void *)current,
57                 data, ra,
58                 (void *)(unsigned long)((bp->b_file_offset>>32) & 0xffffffff),
59                 (void *)(unsigned long)(bp->b_file_offset & 0xffffffff),
60                 (void *)(unsigned long)bp->b_buffer_length,
61                 NULL, NULL, NULL, NULL, NULL);
62 }
63 ktrace_t *xfs_buf_trace_buf;
64 #define XFS_BUF_TRACE_SIZE      4096
65 #define XB_TRACE(bp, id, data)  \
66         xfs_buf_trace(bp, id, (void *)data, (void *)__builtin_return_address(0))
67 #else
68 #define XB_TRACE(bp, id, data)  do { } while (0)
69 #endif
70
71 #ifdef XFS_BUF_LOCK_TRACKING
72 # define XB_SET_OWNER(bp)       ((bp)->b_last_holder = current->pid)
73 # define XB_CLEAR_OWNER(bp)     ((bp)->b_last_holder = -1)
74 # define XB_GET_OWNER(bp)       ((bp)->b_last_holder)
75 #else
76 # define XB_SET_OWNER(bp)       do { } while (0)
77 # define XB_CLEAR_OWNER(bp)     do { } while (0)
78 # define XB_GET_OWNER(bp)       do { } while (0)
79 #endif
80
81 #define xb_to_gfp(flags) \
82         ((((flags) & XBF_READ_AHEAD) ? __GFP_NORETRY : \
83           ((flags) & XBF_DONT_BLOCK) ? GFP_NOFS : GFP_KERNEL) | __GFP_NOWARN)
84
85 #define xb_to_km(flags) \
86          (((flags) & XBF_DONT_BLOCK) ? KM_NOFS : KM_SLEEP)
87
88 #define xfs_buf_allocate(flags) \
89         kmem_zone_alloc(xfs_buf_zone, xb_to_km(flags))
90 #define xfs_buf_deallocate(bp) \
91         kmem_zone_free(xfs_buf_zone, (bp));
92
93 /*
94  *      Page Region interfaces.
95  *
96  *      For pages in filesystems where the blocksize is smaller than the
97  *      pagesize, we use the page->private field (long) to hold a bitmap
98  *      of uptodate regions within the page.
99  *
100  *      Each such region is "bytes per page / bits per long" bytes long.
101  *
102  *      NBPPR == number-of-bytes-per-page-region
103  *      BTOPR == bytes-to-page-region (rounded up)
104  *      BTOPRT == bytes-to-page-region-truncated (rounded down)
105  */
106 #if (BITS_PER_LONG == 32)
107 #define PRSHIFT         (PAGE_CACHE_SHIFT - 5)  /* (32 == 1<<5) */
108 #elif (BITS_PER_LONG == 64)
109 #define PRSHIFT         (PAGE_CACHE_SHIFT - 6)  /* (64 == 1<<6) */
110 #else
111 #error BITS_PER_LONG must be 32 or 64
112 #endif
113 #define NBPPR           (PAGE_CACHE_SIZE/BITS_PER_LONG)
114 #define BTOPR(b)        (((unsigned int)(b) + (NBPPR - 1)) >> PRSHIFT)
115 #define BTOPRT(b)       (((unsigned int)(b) >> PRSHIFT))
116
117 STATIC unsigned long
118 page_region_mask(
119         size_t          offset,
120         size_t          length)
121 {
122         unsigned long   mask;
123         int             first, final;
124
125         first = BTOPR(offset);
126         final = BTOPRT(offset + length - 1);
127         first = min(first, final);
128
129         mask = ~0UL;
130         mask <<= BITS_PER_LONG - (final - first);
131         mask >>= BITS_PER_LONG - (final);
132
133         ASSERT(offset + length <= PAGE_CACHE_SIZE);
134         ASSERT((final - first) < BITS_PER_LONG && (final - first) >= 0);
135
136         return mask;
137 }
138
139 STATIC inline void
140 set_page_region(
141         struct page     *page,
142         size_t          offset,
143         size_t          length)
144 {
145         set_page_private(page,
146                 page_private(page) | page_region_mask(offset, length));
147         if (page_private(page) == ~0UL)
148                 SetPageUptodate(page);
149 }
150
151 STATIC inline int
152 test_page_region(
153         struct page     *page,
154         size_t          offset,
155         size_t          length)
156 {
157         unsigned long   mask = page_region_mask(offset, length);
158
159         return (mask && (page_private(page) & mask) == mask);
160 }
161
162 /*
163  *      Mapping of multi-page buffers into contiguous virtual space
164  */
165
166 typedef struct a_list {
167         void            *vm_addr;
168         struct a_list   *next;
169 } a_list_t;
170
171 STATIC a_list_t         *as_free_head;
172 STATIC int              as_list_len;
173 STATIC DEFINE_SPINLOCK(as_lock);
174
175 /*
176  *      Try to batch vunmaps because they are costly.
177  */
178 STATIC void
179 free_address(
180         void            *addr)
181 {
182         a_list_t        *aentry;
183
184         aentry = kmalloc(sizeof(a_list_t), GFP_ATOMIC & ~__GFP_HIGH);
185         if (likely(aentry)) {
186                 spin_lock(&as_lock);
187                 aentry->next = as_free_head;
188                 aentry->vm_addr = addr;
189                 as_free_head = aentry;
190                 as_list_len++;
191                 spin_unlock(&as_lock);
192         } else {
193                 vunmap(addr);
194         }
195 }
196
197 STATIC void
198 purge_addresses(void)
199 {
200         a_list_t        *aentry, *old;
201
202         if (as_free_head == NULL)
203                 return;
204
205         spin_lock(&as_lock);
206         aentry = as_free_head;
207         as_free_head = NULL;
208         as_list_len = 0;
209         spin_unlock(&as_lock);
210
211         while ((old = aentry) != NULL) {
212                 vunmap(aentry->vm_addr);
213                 aentry = aentry->next;
214                 kfree(old);
215         }
216 }
217
218 /*
219  *      Internal xfs_buf_t object manipulation
220  */
221
222 STATIC void
223 _xfs_buf_initialize(
224         xfs_buf_t               *bp,
225         xfs_buftarg_t           *target,
226         xfs_off_t               range_base,
227         size_t                  range_length,
228         xfs_buf_flags_t         flags)
229 {
230         /*
231          * We don't want certain flags to appear in b_flags.
232          */
233         flags &= ~(XBF_LOCK|XBF_MAPPED|XBF_DONT_BLOCK|XBF_READ_AHEAD);
234
235         memset(bp, 0, sizeof(xfs_buf_t));
236         atomic_set(&bp->b_hold, 1);
237         init_MUTEX_LOCKED(&bp->b_iodonesema);
238         INIT_LIST_HEAD(&bp->b_list);
239         INIT_LIST_HEAD(&bp->b_hash_list);
240         init_MUTEX_LOCKED(&bp->b_sema); /* held, no waiters */
241         XB_SET_OWNER(bp);
242         bp->b_target = target;
243         bp->b_file_offset = range_base;
244         /*
245          * Set buffer_length and count_desired to the same value initially.
246          * I/O routines should use count_desired, which will be the same in
247          * most cases but may be reset (e.g. XFS recovery).
248          */
249         bp->b_buffer_length = bp->b_count_desired = range_length;
250         bp->b_flags = flags;
251         bp->b_bn = XFS_BUF_DADDR_NULL;
252         atomic_set(&bp->b_pin_count, 0);
253         init_waitqueue_head(&bp->b_waiters);
254
255         XFS_STATS_INC(xb_create);
256         XB_TRACE(bp, "initialize", target);
257 }
258
259 /*
260  *      Allocate a page array capable of holding a specified number
261  *      of pages, and point the page buf at it.
262  */
263 STATIC int
264 _xfs_buf_get_pages(
265         xfs_buf_t               *bp,
266         int                     page_count,
267         xfs_buf_flags_t         flags)
268 {
269         /* Make sure that we have a page list */
270         if (bp->b_pages == NULL) {
271                 bp->b_offset = xfs_buf_poff(bp->b_file_offset);
272                 bp->b_page_count = page_count;
273                 if (page_count <= XB_PAGES) {
274                         bp->b_pages = bp->b_page_array;
275                 } else {
276                         bp->b_pages = kmem_alloc(sizeof(struct page *) *
277                                         page_count, xb_to_km(flags));
278                         if (bp->b_pages == NULL)
279                                 return -ENOMEM;
280                 }
281                 memset(bp->b_pages, 0, sizeof(struct page *) * page_count);
282         }
283         return 0;
284 }
285
286 /*
287  *      Frees b_pages if it was allocated.
288  */
289 STATIC void
290 _xfs_buf_free_pages(
291         xfs_buf_t       *bp)
292 {
293         if (bp->b_pages != bp->b_page_array) {
294                 kmem_free(bp->b_pages,
295                           bp->b_page_count * sizeof(struct page *));
296         }
297 }
298
299 /*
300  *      Releases the specified buffer.
301  *
302  *      The modification state of any associated pages is left unchanged.
303  *      The buffer most not be on any hash - use xfs_buf_rele instead for
304  *      hashed and refcounted buffers
305  */
306 void
307 xfs_buf_free(
308         xfs_buf_t               *bp)
309 {
310         XB_TRACE(bp, "free", 0);
311
312         ASSERT(list_empty(&bp->b_hash_list));
313
314         if (bp->b_flags & _XBF_PAGE_CACHE) {
315                 uint            i;
316
317                 if ((bp->b_flags & XBF_MAPPED) && (bp->b_page_count > 1))
318                         free_address(bp->b_addr - bp->b_offset);
319
320                 for (i = 0; i < bp->b_page_count; i++)
321                         page_cache_release(bp->b_pages[i]);
322                 _xfs_buf_free_pages(bp);
323         } else if (bp->b_flags & _XBF_KMEM_ALLOC) {
324                  /*
325                   * XXX(hch): bp->b_count_desired might be incorrect (see
326                   * xfs_buf_associate_memory for details), but fortunately
327                   * the Linux version of kmem_free ignores the len argument..
328                   */
329                 kmem_free(bp->b_addr, bp->b_count_desired);
330                 _xfs_buf_free_pages(bp);
331         }
332
333         xfs_buf_deallocate(bp);
334 }
335
336 /*
337  *      Finds all pages for buffer in question and builds it's page list.
338  */
339 STATIC int
340 _xfs_buf_lookup_pages(
341         xfs_buf_t               *bp,
342         uint                    flags)
343 {
344         struct address_space    *mapping = bp->b_target->bt_mapping;
345         size_t                  blocksize = bp->b_target->bt_bsize;
346         size_t                  size = bp->b_count_desired;
347         size_t                  nbytes, offset;
348         gfp_t                   gfp_mask = xb_to_gfp(flags);
349         unsigned short          page_count, i;
350         pgoff_t                 first;
351         xfs_off_t               end;
352         int                     error;
353
354         end = bp->b_file_offset + bp->b_buffer_length;
355         page_count = xfs_buf_btoc(end) - xfs_buf_btoct(bp->b_file_offset);
356
357         error = _xfs_buf_get_pages(bp, page_count, flags);
358         if (unlikely(error))
359                 return error;
360         bp->b_flags |= _XBF_PAGE_CACHE;
361
362         offset = bp->b_offset;
363         first = bp->b_file_offset >> PAGE_CACHE_SHIFT;
364
365         for (i = 0; i < bp->b_page_count; i++) {
366                 struct page     *page;
367                 uint            retries = 0;
368
369               retry:
370                 page = find_or_create_page(mapping, first + i, gfp_mask);
371                 if (unlikely(page == NULL)) {
372                         if (flags & XBF_READ_AHEAD) {
373                                 bp->b_page_count = i;
374                                 for (i = 0; i < bp->b_page_count; i++)
375                                         unlock_page(bp->b_pages[i]);
376                                 return -ENOMEM;
377                         }
378
379                         /*
380                          * This could deadlock.
381                          *
382                          * But until all the XFS lowlevel code is revamped to
383                          * handle buffer allocation failures we can't do much.
384                          */
385                         if (!(++retries % 100))
386                                 printk(KERN_ERR
387                                         "XFS: possible memory allocation "
388                                         "deadlock in %s (mode:0x%x)\n",
389                                         __FUNCTION__, gfp_mask);
390
391                         XFS_STATS_INC(xb_page_retries);
392                         xfsbufd_wakeup(0, gfp_mask);
393                         blk_congestion_wait(WRITE, HZ/50);
394                         goto retry;
395                 }
396
397                 XFS_STATS_INC(xb_page_found);
398
399                 nbytes = min_t(size_t, size, PAGE_CACHE_SIZE - offset);
400                 size -= nbytes;
401
402                 if (!PageUptodate(page)) {
403                         page_count--;
404                         if (blocksize >= PAGE_CACHE_SIZE) {
405                                 if (flags & XBF_READ)
406                                         bp->b_locked = 1;
407                         } else if (!PagePrivate(page)) {
408                                 if (test_page_region(page, offset, nbytes))
409                                         page_count++;
410                         }
411                 }
412
413                 bp->b_pages[i] = page;
414                 offset = 0;
415         }
416
417         if (!bp->b_locked) {
418                 for (i = 0; i < bp->b_page_count; i++)
419                         unlock_page(bp->b_pages[i]);
420         }
421
422         if (page_count == bp->b_page_count)
423                 bp->b_flags |= XBF_DONE;
424
425         XB_TRACE(bp, "lookup_pages", (long)page_count);
426         return error;
427 }
428
429 /*
430  *      Map buffer into kernel address-space if nessecary.
431  */
432 STATIC int
433 _xfs_buf_map_pages(
434         xfs_buf_t               *bp,
435         uint                    flags)
436 {
437         /* A single page buffer is always mappable */
438         if (bp->b_page_count == 1) {
439                 bp->b_addr = page_address(bp->b_pages[0]) + bp->b_offset;
440                 bp->b_flags |= XBF_MAPPED;
441         } else if (flags & XBF_MAPPED) {
442                 if (as_list_len > 64)
443                         purge_addresses();
444                 bp->b_addr = vmap(bp->b_pages, bp->b_page_count,
445                                         VM_MAP, PAGE_KERNEL);
446                 if (unlikely(bp->b_addr == NULL))
447                         return -ENOMEM;
448                 bp->b_addr += bp->b_offset;
449                 bp->b_flags |= XBF_MAPPED;
450         }
451
452         return 0;
453 }
454
455 /*
456  *      Finding and Reading Buffers
457  */
458
459 /*
460  *      Look up, and creates if absent, a lockable buffer for
461  *      a given range of an inode.  The buffer is returned
462  *      locked.  If other overlapping buffers exist, they are
463  *      released before the new buffer is created and locked,
464  *      which may imply that this call will block until those buffers
465  *      are unlocked.  No I/O is implied by this call.
466  */
467 xfs_buf_t *
468 _xfs_buf_find(
469         xfs_buftarg_t           *btp,   /* block device target          */
470         xfs_off_t               ioff,   /* starting offset of range     */
471         size_t                  isize,  /* length of range              */
472         xfs_buf_flags_t         flags,
473         xfs_buf_t               *new_bp)
474 {
475         xfs_off_t               range_base;
476         size_t                  range_length;
477         xfs_bufhash_t           *hash;
478         xfs_buf_t               *bp, *n;
479
480         range_base = (ioff << BBSHIFT);
481         range_length = (isize << BBSHIFT);
482
483         /* Check for IOs smaller than the sector size / not sector aligned */
484         ASSERT(!(range_length < (1 << btp->bt_sshift)));
485         ASSERT(!(range_base & (xfs_off_t)btp->bt_smask));
486
487         hash = &btp->bt_hash[hash_long((unsigned long)ioff, btp->bt_hashshift)];
488
489         spin_lock(&hash->bh_lock);
490
491         list_for_each_entry_safe(bp, n, &hash->bh_list, b_hash_list) {
492                 ASSERT(btp == bp->b_target);
493                 if (bp->b_file_offset == range_base &&
494                     bp->b_buffer_length == range_length) {
495                         /*
496                          * If we look at something, bring it to the
497                          * front of the list for next time.
498                          */
499                         atomic_inc(&bp->b_hold);
500                         list_move(&bp->b_hash_list, &hash->bh_list);
501                         goto found;
502                 }
503         }
504
505         /* No match found */
506         if (new_bp) {
507                 _xfs_buf_initialize(new_bp, btp, range_base,
508                                 range_length, flags);
509                 new_bp->b_hash = hash;
510                 list_add(&new_bp->b_hash_list, &hash->bh_list);
511         } else {
512                 XFS_STATS_INC(xb_miss_locked);
513         }
514
515         spin_unlock(&hash->bh_lock);
516         return new_bp;
517
518 found:
519         spin_unlock(&hash->bh_lock);
520
521         /* Attempt to get the semaphore without sleeping,
522          * if this does not work then we need to drop the
523          * spinlock and do a hard attempt on the semaphore.
524          */
525         if (down_trylock(&bp->b_sema)) {
526                 if (!(flags & XBF_TRYLOCK)) {
527                         /* wait for buffer ownership */
528                         XB_TRACE(bp, "get_lock", 0);
529                         xfs_buf_lock(bp);
530                         XFS_STATS_INC(xb_get_locked_waited);
531                 } else {
532                         /* We asked for a trylock and failed, no need
533                          * to look at file offset and length here, we
534                          * know that this buffer at least overlaps our
535                          * buffer and is locked, therefore our buffer
536                          * either does not exist, or is this buffer.
537                          */
538                         xfs_buf_rele(bp);
539                         XFS_STATS_INC(xb_busy_locked);
540                         return NULL;
541                 }
542         } else {
543                 /* trylock worked */
544                 XB_SET_OWNER(bp);
545         }
546
547         if (bp->b_flags & XBF_STALE) {
548                 ASSERT((bp->b_flags & _XBF_DELWRI_Q) == 0);
549                 bp->b_flags &= XBF_MAPPED;
550         }
551         XB_TRACE(bp, "got_lock", 0);
552         XFS_STATS_INC(xb_get_locked);
553         return bp;
554 }
555
556 /*
557  *      Assembles a buffer covering the specified range.
558  *      Storage in memory for all portions of the buffer will be allocated,
559  *      although backing storage may not be.
560  */
561 xfs_buf_t *
562 xfs_buf_get_flags(
563         xfs_buftarg_t           *target,/* target for buffer            */
564         xfs_off_t               ioff,   /* starting offset of range     */
565         size_t                  isize,  /* length of range              */
566         xfs_buf_flags_t         flags)
567 {
568         xfs_buf_t               *bp, *new_bp;
569         int                     error = 0, i;
570
571         new_bp = xfs_buf_allocate(flags);
572         if (unlikely(!new_bp))
573                 return NULL;
574
575         bp = _xfs_buf_find(target, ioff, isize, flags, new_bp);
576         if (bp == new_bp) {
577                 error = _xfs_buf_lookup_pages(bp, flags);
578                 if (error)
579                         goto no_buffer;
580         } else {
581                 xfs_buf_deallocate(new_bp);
582                 if (unlikely(bp == NULL))
583                         return NULL;
584         }
585
586         for (i = 0; i < bp->b_page_count; i++)
587                 mark_page_accessed(bp->b_pages[i]);
588
589         if (!(bp->b_flags & XBF_MAPPED)) {
590                 error = _xfs_buf_map_pages(bp, flags);
591                 if (unlikely(error)) {
592                         printk(KERN_WARNING "%s: failed to map pages\n",
593                                         __FUNCTION__);
594                         goto no_buffer;
595                 }
596         }
597
598         XFS_STATS_INC(xb_get);
599
600         /*
601          * Always fill in the block number now, the mapped cases can do
602          * their own overlay of this later.
603          */
604         bp->b_bn = ioff;
605         bp->b_count_desired = bp->b_buffer_length;
606
607         XB_TRACE(bp, "get", (unsigned long)flags);
608         return bp;
609
610  no_buffer:
611         if (flags & (XBF_LOCK | XBF_TRYLOCK))
612                 xfs_buf_unlock(bp);
613         xfs_buf_rele(bp);
614         return NULL;
615 }
616
617 xfs_buf_t *
618 xfs_buf_read_flags(
619         xfs_buftarg_t           *target,
620         xfs_off_t               ioff,
621         size_t                  isize,
622         xfs_buf_flags_t         flags)
623 {
624         xfs_buf_t               *bp;
625
626         flags |= XBF_READ;
627
628         bp = xfs_buf_get_flags(target, ioff, isize, flags);
629         if (bp) {
630                 if (!XFS_BUF_ISDONE(bp)) {
631                         XB_TRACE(bp, "read", (unsigned long)flags);
632                         XFS_STATS_INC(xb_get_read);
633                         xfs_buf_iostart(bp, flags);
634                 } else if (flags & XBF_ASYNC) {
635                         XB_TRACE(bp, "read_async", (unsigned long)flags);
636                         /*
637                          * Read ahead call which is already satisfied,
638                          * drop the buffer
639                          */
640                         goto no_buffer;
641                 } else {
642                         XB_TRACE(bp, "read_done", (unsigned long)flags);
643                         /* We do not want read in the flags */
644                         bp->b_flags &= ~XBF_READ;
645                 }
646         }
647
648         return bp;
649
650  no_buffer:
651         if (flags & (XBF_LOCK | XBF_TRYLOCK))
652                 xfs_buf_unlock(bp);
653         xfs_buf_rele(bp);
654         return NULL;
655 }
656
657 /*
658  *      If we are not low on memory then do the readahead in a deadlock
659  *      safe manner.
660  */
661 void
662 xfs_buf_readahead(
663         xfs_buftarg_t           *target,
664         xfs_off_t               ioff,
665         size_t                  isize,
666         xfs_buf_flags_t         flags)
667 {
668         struct backing_dev_info *bdi;
669
670         bdi = target->bt_mapping->backing_dev_info;
671         if (bdi_read_congested(bdi))
672                 return;
673
674         flags |= (XBF_TRYLOCK|XBF_ASYNC|XBF_READ_AHEAD);
675         xfs_buf_read_flags(target, ioff, isize, flags);
676 }
677
678 xfs_buf_t *
679 xfs_buf_get_empty(
680         size_t                  len,
681         xfs_buftarg_t           *target)
682 {
683         xfs_buf_t               *bp;
684
685         bp = xfs_buf_allocate(0);
686         if (bp)
687                 _xfs_buf_initialize(bp, target, 0, len, 0);
688         return bp;
689 }
690
691 static inline struct page *
692 mem_to_page(
693         void                    *addr)
694 {
695         if (((unsigned long)addr < VMALLOC_START) ||
696             ((unsigned long)addr >= VMALLOC_END)) {
697                 return virt_to_page(addr);
698         } else {
699                 return vmalloc_to_page(addr);
700         }
701 }
702
703 int
704 xfs_buf_associate_memory(
705         xfs_buf_t               *bp,
706         void                    *mem,
707         size_t                  len)
708 {
709         int                     rval;
710         int                     i = 0;
711         size_t                  ptr;
712         size_t                  end, end_cur;
713         off_t                   offset;
714         int                     page_count;
715
716         page_count = PAGE_CACHE_ALIGN(len) >> PAGE_CACHE_SHIFT;
717         offset = (off_t) mem - ((off_t)mem & PAGE_CACHE_MASK);
718         if (offset && (len > PAGE_CACHE_SIZE))
719                 page_count++;
720
721         /* Free any previous set of page pointers */
722         if (bp->b_pages)
723                 _xfs_buf_free_pages(bp);
724
725         bp->b_pages = NULL;
726         bp->b_addr = mem;
727
728         rval = _xfs_buf_get_pages(bp, page_count, 0);
729         if (rval)
730                 return rval;
731
732         bp->b_offset = offset;
733         ptr = (size_t) mem & PAGE_CACHE_MASK;
734         end = PAGE_CACHE_ALIGN((size_t) mem + len);
735         end_cur = end;
736         /* set up first page */
737         bp->b_pages[0] = mem_to_page(mem);
738
739         ptr += PAGE_CACHE_SIZE;
740         bp->b_page_count = ++i;
741         while (ptr < end) {
742                 bp->b_pages[i] = mem_to_page((void *)ptr);
743                 bp->b_page_count = ++i;
744                 ptr += PAGE_CACHE_SIZE;
745         }
746         bp->b_locked = 0;
747
748         bp->b_count_desired = bp->b_buffer_length = len;
749         bp->b_flags |= XBF_MAPPED;
750
751         return 0;
752 }
753
754 xfs_buf_t *
755 xfs_buf_get_noaddr(
756         size_t                  len,
757         xfs_buftarg_t           *target)
758 {
759         size_t                  malloc_len = len;
760         xfs_buf_t               *bp;
761         void                    *data;
762         int                     error;
763
764         bp = xfs_buf_allocate(0);
765         if (unlikely(bp == NULL))
766                 goto fail;
767         _xfs_buf_initialize(bp, target, 0, len, 0);
768
769  try_again:
770         data = kmem_alloc(malloc_len, KM_SLEEP | KM_MAYFAIL);
771         if (unlikely(data == NULL))
772                 goto fail_free_buf;
773
774         /* check whether alignment matches.. */
775         if ((__psunsigned_t)data !=
776             ((__psunsigned_t)data & ~target->bt_smask)) {
777                 /* .. else double the size and try again */
778                 kmem_free(data, malloc_len);
779                 malloc_len <<= 1;
780                 goto try_again;
781         }
782
783         error = xfs_buf_associate_memory(bp, data, len);
784         if (error)
785                 goto fail_free_mem;
786         bp->b_flags |= _XBF_KMEM_ALLOC;
787
788         xfs_buf_unlock(bp);
789
790         XB_TRACE(bp, "no_daddr", data);
791         return bp;
792  fail_free_mem:
793         kmem_free(data, malloc_len);
794  fail_free_buf:
795         xfs_buf_free(bp);
796  fail:
797         return NULL;
798 }
799
800 /*
801  *      Increment reference count on buffer, to hold the buffer concurrently
802  *      with another thread which may release (free) the buffer asynchronously.
803  *      Must hold the buffer already to call this function.
804  */
805 void
806 xfs_buf_hold(
807         xfs_buf_t               *bp)
808 {
809         atomic_inc(&bp->b_hold);
810         XB_TRACE(bp, "hold", 0);
811 }
812
813 /*
814  *      Releases a hold on the specified buffer.  If the
815  *      the hold count is 1, calls xfs_buf_free.
816  */
817 void
818 xfs_buf_rele(
819         xfs_buf_t               *bp)
820 {
821         xfs_bufhash_t           *hash = bp->b_hash;
822
823         XB_TRACE(bp, "rele", bp->b_relse);
824
825         if (unlikely(!hash)) {
826                 ASSERT(!bp->b_relse);
827                 if (atomic_dec_and_test(&bp->b_hold))
828                         xfs_buf_free(bp);
829                 return;
830         }
831
832         if (atomic_dec_and_lock(&bp->b_hold, &hash->bh_lock)) {
833                 if (bp->b_relse) {
834                         atomic_inc(&bp->b_hold);
835                         spin_unlock(&hash->bh_lock);
836                         (*(bp->b_relse)) (bp);
837                 } else if (bp->b_flags & XBF_FS_MANAGED) {
838                         spin_unlock(&hash->bh_lock);
839                 } else {
840                         ASSERT(!(bp->b_flags & (XBF_DELWRI|_XBF_DELWRI_Q)));
841                         list_del_init(&bp->b_hash_list);
842                         spin_unlock(&hash->bh_lock);
843                         xfs_buf_free(bp);
844                 }
845         } else {
846                 /*
847                  * Catch reference count leaks
848                  */
849                 ASSERT(atomic_read(&bp->b_hold) >= 0);
850         }
851 }
852
853
854 /*
855  *      Mutual exclusion on buffers.  Locking model:
856  *
857  *      Buffers associated with inodes for which buffer locking
858  *      is not enabled are not protected by semaphores, and are
859  *      assumed to be exclusively owned by the caller.  There is a
860  *      spinlock in the buffer, used by the caller when concurrent
861  *      access is possible.
862  */
863
864 /*
865  *      Locks a buffer object, if it is not already locked.
866  *      Note that this in no way locks the underlying pages, so it is only
867  *      useful for synchronizing concurrent use of buffer objects, not for
868  *      synchronizing independent access to the underlying pages.
869  */
870 int
871 xfs_buf_cond_lock(
872         xfs_buf_t               *bp)
873 {
874         int                     locked;
875
876         locked = down_trylock(&bp->b_sema) == 0;
877         if (locked) {
878                 XB_SET_OWNER(bp);
879         }
880         XB_TRACE(bp, "cond_lock", (long)locked);
881         return locked ? 0 : -EBUSY;
882 }
883
884 #if defined(DEBUG) || defined(XFS_BLI_TRACE)
885 int
886 xfs_buf_lock_value(
887         xfs_buf_t               *bp)
888 {
889         return atomic_read(&bp->b_sema.count);
890 }
891 #endif
892
893 /*
894  *      Locks a buffer object.
895  *      Note that this in no way locks the underlying pages, so it is only
896  *      useful for synchronizing concurrent use of buffer objects, not for
897  *      synchronizing independent access to the underlying pages.
898  */
899 void
900 xfs_buf_lock(
901         xfs_buf_t               *bp)
902 {
903         XB_TRACE(bp, "lock", 0);
904         if (atomic_read(&bp->b_io_remaining))
905                 blk_run_address_space(bp->b_target->bt_mapping);
906         down(&bp->b_sema);
907         XB_SET_OWNER(bp);
908         XB_TRACE(bp, "locked", 0);
909 }
910
911 /*
912  *      Releases the lock on the buffer object.
913  *      If the buffer is marked delwri but is not queued, do so before we
914  *      unlock the buffer as we need to set flags correctly.  We also need to
915  *      take a reference for the delwri queue because the unlocker is going to
916  *      drop their's and they don't know we just queued it.
917  */
918 void
919 xfs_buf_unlock(
920         xfs_buf_t               *bp)
921 {
922         if ((bp->b_flags & (XBF_DELWRI|_XBF_DELWRI_Q)) == XBF_DELWRI) {
923                 atomic_inc(&bp->b_hold);
924                 bp->b_flags |= XBF_ASYNC;
925                 xfs_buf_delwri_queue(bp, 0);
926         }
927
928         XB_CLEAR_OWNER(bp);
929         up(&bp->b_sema);
930         XB_TRACE(bp, "unlock", 0);
931 }
932
933
934 /*
935  *      Pinning Buffer Storage in Memory
936  *      Ensure that no attempt to force a buffer to disk will succeed.
937  */
938 void
939 xfs_buf_pin(
940         xfs_buf_t               *bp)
941 {
942         atomic_inc(&bp->b_pin_count);
943         XB_TRACE(bp, "pin", (long)bp->b_pin_count.counter);
944 }
945
946 void
947 xfs_buf_unpin(
948         xfs_buf_t               *bp)
949 {
950         if (atomic_dec_and_test(&bp->b_pin_count))
951                 wake_up_all(&bp->b_waiters);
952         XB_TRACE(bp, "unpin", (long)bp->b_pin_count.counter);
953 }
954
955 int
956 xfs_buf_ispin(
957         xfs_buf_t               *bp)
958 {
959         return atomic_read(&bp->b_pin_count);
960 }
961
962 STATIC void
963 xfs_buf_wait_unpin(
964         xfs_buf_t               *bp)
965 {
966         DECLARE_WAITQUEUE       (wait, current);
967
968         if (atomic_read(&bp->b_pin_count) == 0)
969                 return;
970
971         add_wait_queue(&bp->b_waiters, &wait);
972         for (;;) {
973                 set_current_state(TASK_UNINTERRUPTIBLE);
974                 if (atomic_read(&bp->b_pin_count) == 0)
975                         break;
976                 if (atomic_read(&bp->b_io_remaining))
977                         blk_run_address_space(bp->b_target->bt_mapping);
978                 schedule();
979         }
980         remove_wait_queue(&bp->b_waiters, &wait);
981         set_current_state(TASK_RUNNING);
982 }
983
984 /*
985  *      Buffer Utility Routines
986  */
987
988 STATIC void
989 xfs_buf_iodone_work(
990         void                    *v)
991 {
992         xfs_buf_t               *bp = (xfs_buf_t *)v;
993
994         if (bp->b_iodone)
995                 (*(bp->b_iodone))(bp);
996         else if (bp->b_flags & XBF_ASYNC)
997                 xfs_buf_relse(bp);
998 }
999
1000 void
1001 xfs_buf_ioend(
1002         xfs_buf_t               *bp,
1003         int                     schedule)
1004 {
1005         bp->b_flags &= ~(XBF_READ | XBF_WRITE);
1006         if (bp->b_error == 0)
1007                 bp->b_flags |= XBF_DONE;
1008
1009         XB_TRACE(bp, "iodone", bp->b_iodone);
1010
1011         if ((bp->b_iodone) || (bp->b_flags & XBF_ASYNC)) {
1012                 if (schedule) {
1013                         INIT_WORK(&bp->b_iodone_work, xfs_buf_iodone_work, bp);
1014                         queue_work(xfslogd_workqueue, &bp->b_iodone_work);
1015                 } else {
1016                         xfs_buf_iodone_work(bp);
1017                 }
1018         } else {
1019                 up(&bp->b_iodonesema);
1020         }
1021 }
1022
1023 void
1024 xfs_buf_ioerror(
1025         xfs_buf_t               *bp,
1026         int                     error)
1027 {
1028         ASSERT(error >= 0 && error <= 0xffff);
1029         bp->b_error = (unsigned short)error;
1030         XB_TRACE(bp, "ioerror", (unsigned long)error);
1031 }
1032
1033 /*
1034  *      Initiate I/O on a buffer, based on the flags supplied.
1035  *      The b_iodone routine in the buffer supplied will only be called
1036  *      when all of the subsidiary I/O requests, if any, have been completed.
1037  */
1038 int
1039 xfs_buf_iostart(
1040         xfs_buf_t               *bp,
1041         xfs_buf_flags_t         flags)
1042 {
1043         int                     status = 0;
1044
1045         XB_TRACE(bp, "iostart", (unsigned long)flags);
1046
1047         if (flags & XBF_DELWRI) {
1048                 bp->b_flags &= ~(XBF_READ | XBF_WRITE | XBF_ASYNC);
1049                 bp->b_flags |= flags & (XBF_DELWRI | XBF_ASYNC);
1050                 xfs_buf_delwri_queue(bp, 1);
1051                 return status;
1052         }
1053
1054         bp->b_flags &= ~(XBF_READ | XBF_WRITE | XBF_ASYNC | XBF_DELWRI | \
1055                         XBF_READ_AHEAD | _XBF_RUN_QUEUES);
1056         bp->b_flags |= flags & (XBF_READ | XBF_WRITE | XBF_ASYNC | \
1057                         XBF_READ_AHEAD | _XBF_RUN_QUEUES);
1058
1059         BUG_ON(bp->b_bn == XFS_BUF_DADDR_NULL);
1060
1061         /* For writes allow an alternate strategy routine to precede
1062          * the actual I/O request (which may not be issued at all in
1063          * a shutdown situation, for example).
1064          */
1065         status = (flags & XBF_WRITE) ?
1066                 xfs_buf_iostrategy(bp) : xfs_buf_iorequest(bp);
1067
1068         /* Wait for I/O if we are not an async request.
1069          * Note: async I/O request completion will release the buffer,
1070          * and that can already be done by this point.  So using the
1071          * buffer pointer from here on, after async I/O, is invalid.
1072          */
1073         if (!status && !(flags & XBF_ASYNC))
1074                 status = xfs_buf_iowait(bp);
1075
1076         return status;
1077 }
1078
1079 STATIC __inline__ int
1080 _xfs_buf_iolocked(
1081         xfs_buf_t               *bp)
1082 {
1083         ASSERT(bp->b_flags & (XBF_READ | XBF_WRITE));
1084         if (bp->b_flags & XBF_READ)
1085                 return bp->b_locked;
1086         return 0;
1087 }
1088
1089 STATIC __inline__ void
1090 _xfs_buf_ioend(
1091         xfs_buf_t               *bp,
1092         int                     schedule)
1093 {
1094         if (atomic_dec_and_test(&bp->b_io_remaining) == 1) {
1095                 bp->b_locked = 0;
1096                 xfs_buf_ioend(bp, schedule);
1097         }
1098 }
1099
1100 STATIC int
1101 xfs_buf_bio_end_io(
1102         struct bio              *bio,
1103         unsigned int            bytes_done,
1104         int                     error)
1105 {
1106         xfs_buf_t               *bp = (xfs_buf_t *)bio->bi_private;
1107         unsigned int            blocksize = bp->b_target->bt_bsize;
1108         struct bio_vec          *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
1109
1110         if (bio->bi_size)
1111                 return 1;
1112
1113         if (!test_bit(BIO_UPTODATE, &bio->bi_flags))
1114                 bp->b_error = EIO;
1115
1116         do {
1117                 struct page     *page = bvec->bv_page;
1118
1119                 if (unlikely(bp->b_error)) {
1120                         if (bp->b_flags & XBF_READ)
1121                                 ClearPageUptodate(page);
1122                         SetPageError(page);
1123                 } else if (blocksize >= PAGE_CACHE_SIZE) {
1124                         SetPageUptodate(page);
1125                 } else if (!PagePrivate(page) &&
1126                                 (bp->b_flags & _XBF_PAGE_CACHE)) {
1127                         set_page_region(page, bvec->bv_offset, bvec->bv_len);
1128                 }
1129
1130                 if (--bvec >= bio->bi_io_vec)
1131                         prefetchw(&bvec->bv_page->flags);
1132
1133                 if (_xfs_buf_iolocked(bp)) {
1134                         unlock_page(page);
1135                 }
1136         } while (bvec >= bio->bi_io_vec);
1137
1138         _xfs_buf_ioend(bp, 1);
1139         bio_put(bio);
1140         return 0;
1141 }
1142
1143 STATIC void
1144 _xfs_buf_ioapply(
1145         xfs_buf_t               *bp)
1146 {
1147         int                     i, rw, map_i, total_nr_pages, nr_pages;
1148         struct bio              *bio;
1149         int                     offset = bp->b_offset;
1150         int                     size = bp->b_count_desired;
1151         sector_t                sector = bp->b_bn;
1152         unsigned int            blocksize = bp->b_target->bt_bsize;
1153         int                     locking = _xfs_buf_iolocked(bp);
1154
1155         total_nr_pages = bp->b_page_count;
1156         map_i = 0;
1157
1158         if (bp->b_flags & _XBF_RUN_QUEUES) {
1159                 bp->b_flags &= ~_XBF_RUN_QUEUES;
1160                 rw = (bp->b_flags & XBF_READ) ? READ_SYNC : WRITE_SYNC;
1161         } else {
1162                 rw = (bp->b_flags & XBF_READ) ? READ : WRITE;
1163         }
1164
1165         if (bp->b_flags & XBF_ORDERED) {
1166                 ASSERT(!(bp->b_flags & XBF_READ));
1167                 rw = WRITE_BARRIER;
1168         }
1169
1170         /* Special code path for reading a sub page size buffer in --
1171          * we populate up the whole page, and hence the other metadata
1172          * in the same page.  This optimization is only valid when the
1173          * filesystem block size is not smaller than the page size.
1174          */
1175         if ((bp->b_buffer_length < PAGE_CACHE_SIZE) &&
1176             (bp->b_flags & XBF_READ) && locking &&
1177             (blocksize >= PAGE_CACHE_SIZE)) {
1178                 bio = bio_alloc(GFP_NOIO, 1);
1179
1180                 bio->bi_bdev = bp->b_target->bt_bdev;
1181                 bio->bi_sector = sector - (offset >> BBSHIFT);
1182                 bio->bi_end_io = xfs_buf_bio_end_io;
1183                 bio->bi_private = bp;
1184
1185                 bio_add_page(bio, bp->b_pages[0], PAGE_CACHE_SIZE, 0);
1186                 size = 0;
1187
1188                 atomic_inc(&bp->b_io_remaining);
1189
1190                 goto submit_io;
1191         }
1192
1193         /* Lock down the pages which we need to for the request */
1194         if (locking && (bp->b_flags & XBF_WRITE) && (bp->b_locked == 0)) {
1195                 for (i = 0; size; i++) {
1196                         int             nbytes = PAGE_CACHE_SIZE - offset;
1197                         struct page     *page = bp->b_pages[i];
1198
1199                         if (nbytes > size)
1200                                 nbytes = size;
1201
1202                         lock_page(page);
1203
1204                         size -= nbytes;
1205                         offset = 0;
1206                 }
1207                 offset = bp->b_offset;
1208                 size = bp->b_count_desired;
1209         }
1210
1211 next_chunk:
1212         atomic_inc(&bp->b_io_remaining);
1213         nr_pages = BIO_MAX_SECTORS >> (PAGE_SHIFT - BBSHIFT);
1214         if (nr_pages > total_nr_pages)
1215                 nr_pages = total_nr_pages;
1216
1217         bio = bio_alloc(GFP_NOIO, nr_pages);
1218         bio->bi_bdev = bp->b_target->bt_bdev;
1219         bio->bi_sector = sector;
1220         bio->bi_end_io = xfs_buf_bio_end_io;
1221         bio->bi_private = bp;
1222
1223         for (; size && nr_pages; nr_pages--, map_i++) {
1224                 int     rbytes, nbytes = PAGE_CACHE_SIZE - offset;
1225
1226                 if (nbytes > size)
1227                         nbytes = size;
1228
1229                 rbytes = bio_add_page(bio, bp->b_pages[map_i], nbytes, offset);
1230                 if (rbytes < nbytes)
1231                         break;
1232
1233                 offset = 0;
1234                 sector += nbytes >> BBSHIFT;
1235                 size -= nbytes;
1236                 total_nr_pages--;
1237         }
1238
1239 submit_io:
1240         if (likely(bio->bi_size)) {
1241                 submit_bio(rw, bio);
1242                 if (size)
1243                         goto next_chunk;
1244         } else {
1245                 bio_put(bio);
1246                 xfs_buf_ioerror(bp, EIO);
1247         }
1248 }
1249
1250 int
1251 xfs_buf_iorequest(
1252         xfs_buf_t               *bp)
1253 {
1254         XB_TRACE(bp, "iorequest", 0);
1255
1256         if (bp->b_flags & XBF_DELWRI) {
1257                 xfs_buf_delwri_queue(bp, 1);
1258                 return 0;
1259         }
1260
1261         if (bp->b_flags & XBF_WRITE) {
1262                 xfs_buf_wait_unpin(bp);
1263         }
1264
1265         xfs_buf_hold(bp);
1266
1267         /* Set the count to 1 initially, this will stop an I/O
1268          * completion callout which happens before we have started
1269          * all the I/O from calling xfs_buf_ioend too early.
1270          */
1271         atomic_set(&bp->b_io_remaining, 1);
1272         _xfs_buf_ioapply(bp);
1273         _xfs_buf_ioend(bp, 0);
1274
1275         xfs_buf_rele(bp);
1276         return 0;
1277 }
1278
1279 /*
1280  *      Waits for I/O to complete on the buffer supplied.
1281  *      It returns immediately if no I/O is pending.
1282  *      It returns the I/O error code, if any, or 0 if there was no error.
1283  */
1284 int
1285 xfs_buf_iowait(
1286         xfs_buf_t               *bp)
1287 {
1288         XB_TRACE(bp, "iowait", 0);
1289         if (atomic_read(&bp->b_io_remaining))
1290                 blk_run_address_space(bp->b_target->bt_mapping);
1291         down(&bp->b_iodonesema);
1292         XB_TRACE(bp, "iowaited", (long)bp->b_error);
1293         return bp->b_error;
1294 }
1295
1296 xfs_caddr_t
1297 xfs_buf_offset(
1298         xfs_buf_t               *bp,
1299         size_t                  offset)
1300 {
1301         struct page             *page;
1302
1303         if (bp->b_flags & XBF_MAPPED)
1304                 return XFS_BUF_PTR(bp) + offset;
1305
1306         offset += bp->b_offset;
1307         page = bp->b_pages[offset >> PAGE_CACHE_SHIFT];
1308         return (xfs_caddr_t)page_address(page) + (offset & (PAGE_CACHE_SIZE-1));
1309 }
1310
1311 /*
1312  *      Move data into or out of a buffer.
1313  */
1314 void
1315 xfs_buf_iomove(
1316         xfs_buf_t               *bp,    /* buffer to process            */
1317         size_t                  boff,   /* starting buffer offset       */
1318         size_t                  bsize,  /* length to copy               */
1319         caddr_t                 data,   /* data address                 */
1320         xfs_buf_rw_t            mode)   /* read/write/zero flag         */
1321 {
1322         size_t                  bend, cpoff, csize;
1323         struct page             *page;
1324
1325         bend = boff + bsize;
1326         while (boff < bend) {
1327                 page = bp->b_pages[xfs_buf_btoct(boff + bp->b_offset)];
1328                 cpoff = xfs_buf_poff(boff + bp->b_offset);
1329                 csize = min_t(size_t,
1330                               PAGE_CACHE_SIZE-cpoff, bp->b_count_desired-boff);
1331
1332                 ASSERT(((csize + cpoff) <= PAGE_CACHE_SIZE));
1333
1334                 switch (mode) {
1335                 case XBRW_ZERO:
1336                         memset(page_address(page) + cpoff, 0, csize);
1337                         break;
1338                 case XBRW_READ:
1339                         memcpy(data, page_address(page) + cpoff, csize);
1340                         break;
1341                 case XBRW_WRITE:
1342                         memcpy(page_address(page) + cpoff, data, csize);
1343                 }
1344
1345                 boff += csize;
1346                 data += csize;
1347         }
1348 }
1349
1350 /*
1351  *      Handling of buffer targets (buftargs).
1352  */
1353
1354 /*
1355  *      Wait for any bufs with callbacks that have been submitted but
1356  *      have not yet returned... walk the hash list for the target.
1357  */
1358 void
1359 xfs_wait_buftarg(
1360         xfs_buftarg_t   *btp)
1361 {
1362         xfs_buf_t       *bp, *n;
1363         xfs_bufhash_t   *hash;
1364         uint            i;
1365
1366         for (i = 0; i < (1 << btp->bt_hashshift); i++) {
1367                 hash = &btp->bt_hash[i];
1368 again:
1369                 spin_lock(&hash->bh_lock);
1370                 list_for_each_entry_safe(bp, n, &hash->bh_list, b_hash_list) {
1371                         ASSERT(btp == bp->b_target);
1372                         if (!(bp->b_flags & XBF_FS_MANAGED)) {
1373                                 spin_unlock(&hash->bh_lock);
1374                                 /*
1375                                  * Catch superblock reference count leaks
1376                                  * immediately
1377                                  */
1378                                 BUG_ON(bp->b_bn == 0);
1379                                 delay(100);
1380                                 goto again;
1381                         }
1382                 }
1383                 spin_unlock(&hash->bh_lock);
1384         }
1385 }
1386
1387 /*
1388  *      Allocate buffer hash table for a given target.
1389  *      For devices containing metadata (i.e. not the log/realtime devices)
1390  *      we need to allocate a much larger hash table.
1391  */
1392 STATIC void
1393 xfs_alloc_bufhash(
1394         xfs_buftarg_t           *btp,
1395         int                     external)
1396 {
1397         unsigned int            i;
1398
1399         btp->bt_hashshift = external ? 3 : 8;   /* 8 or 256 buckets */
1400         btp->bt_hashmask = (1 << btp->bt_hashshift) - 1;
1401         btp->bt_hash = kmem_zalloc((1 << btp->bt_hashshift) *
1402                                         sizeof(xfs_bufhash_t), KM_SLEEP);
1403         for (i = 0; i < (1 << btp->bt_hashshift); i++) {
1404                 spin_lock_init(&btp->bt_hash[i].bh_lock);
1405                 INIT_LIST_HEAD(&btp->bt_hash[i].bh_list);
1406         }
1407 }
1408
1409 STATIC void
1410 xfs_free_bufhash(
1411         xfs_buftarg_t           *btp)
1412 {
1413         kmem_free(btp->bt_hash, (1<<btp->bt_hashshift) * sizeof(xfs_bufhash_t));
1414         btp->bt_hash = NULL;
1415 }
1416
1417 /*
1418  *      buftarg list for delwrite queue processing
1419  */
1420 STATIC LIST_HEAD(xfs_buftarg_list);
1421 STATIC DEFINE_SPINLOCK(xfs_buftarg_lock);
1422
1423 STATIC void
1424 xfs_register_buftarg(
1425         xfs_buftarg_t           *btp)
1426 {
1427         spin_lock(&xfs_buftarg_lock);
1428         list_add(&btp->bt_list, &xfs_buftarg_list);
1429         spin_unlock(&xfs_buftarg_lock);
1430 }
1431
1432 STATIC void
1433 xfs_unregister_buftarg(
1434         xfs_buftarg_t           *btp)
1435 {
1436         spin_lock(&xfs_buftarg_lock);
1437         list_del(&btp->bt_list);
1438         spin_unlock(&xfs_buftarg_lock);
1439 }
1440
1441 void
1442 xfs_free_buftarg(
1443         xfs_buftarg_t           *btp,
1444         int                     external)
1445 {
1446         xfs_flush_buftarg(btp, 1);
1447         if (external)
1448                 xfs_blkdev_put(btp->bt_bdev);
1449         xfs_free_bufhash(btp);
1450         iput(btp->bt_mapping->host);
1451
1452         /* Unregister the buftarg first so that we don't get a
1453          * wakeup finding a non-existent task
1454          */
1455         xfs_unregister_buftarg(btp);
1456         kthread_stop(btp->bt_task);
1457
1458         kmem_free(btp, sizeof(*btp));
1459 }
1460
1461 STATIC int
1462 xfs_setsize_buftarg_flags(
1463         xfs_buftarg_t           *btp,
1464         unsigned int            blocksize,
1465         unsigned int            sectorsize,
1466         int                     verbose)
1467 {
1468         btp->bt_bsize = blocksize;
1469         btp->bt_sshift = ffs(sectorsize) - 1;
1470         btp->bt_smask = sectorsize - 1;
1471
1472         if (set_blocksize(btp->bt_bdev, sectorsize)) {
1473                 printk(KERN_WARNING
1474                         "XFS: Cannot set_blocksize to %u on device %s\n",
1475                         sectorsize, XFS_BUFTARG_NAME(btp));
1476                 return EINVAL;
1477         }
1478
1479         if (verbose &&
1480             (PAGE_CACHE_SIZE / BITS_PER_LONG) > sectorsize) {
1481                 printk(KERN_WARNING
1482                         "XFS: %u byte sectors in use on device %s.  "
1483                         "This is suboptimal; %u or greater is ideal.\n",
1484                         sectorsize, XFS_BUFTARG_NAME(btp),
1485                         (unsigned int)PAGE_CACHE_SIZE / BITS_PER_LONG);
1486         }
1487
1488         return 0;
1489 }
1490
1491 /*
1492  *      When allocating the initial buffer target we have not yet
1493  *      read in the superblock, so don't know what sized sectors
1494  *      are being used is at this early stage.  Play safe.
1495  */
1496 STATIC int
1497 xfs_setsize_buftarg_early(
1498         xfs_buftarg_t           *btp,
1499         struct block_device     *bdev)
1500 {
1501         return xfs_setsize_buftarg_flags(btp,
1502                         PAGE_CACHE_SIZE, bdev_hardsect_size(bdev), 0);
1503 }
1504
1505 int
1506 xfs_setsize_buftarg(
1507         xfs_buftarg_t           *btp,
1508         unsigned int            blocksize,
1509         unsigned int            sectorsize)
1510 {
1511         return xfs_setsize_buftarg_flags(btp, blocksize, sectorsize, 1);
1512 }
1513
1514 STATIC int
1515 xfs_mapping_buftarg(
1516         xfs_buftarg_t           *btp,
1517         struct block_device     *bdev)
1518 {
1519         struct backing_dev_info *bdi;
1520         struct inode            *inode;
1521         struct address_space    *mapping;
1522         static struct address_space_operations mapping_aops = {
1523                 .sync_page = block_sync_page,
1524                 .migratepage = fail_migrate_page,
1525         };
1526
1527         inode = new_inode(bdev->bd_inode->i_sb);
1528         if (!inode) {
1529                 printk(KERN_WARNING
1530                         "XFS: Cannot allocate mapping inode for device %s\n",
1531                         XFS_BUFTARG_NAME(btp));
1532                 return ENOMEM;
1533         }
1534         inode->i_mode = S_IFBLK;
1535         inode->i_bdev = bdev;
1536         inode->i_rdev = bdev->bd_dev;
1537         bdi = blk_get_backing_dev_info(bdev);
1538         if (!bdi)
1539                 bdi = &default_backing_dev_info;
1540         mapping = &inode->i_data;
1541         mapping->a_ops = &mapping_aops;
1542         mapping->backing_dev_info = bdi;
1543         mapping_set_gfp_mask(mapping, GFP_NOFS);
1544         btp->bt_mapping = mapping;
1545         return 0;
1546 }
1547
1548 STATIC int
1549 xfs_alloc_delwrite_queue(
1550         xfs_buftarg_t           *btp)
1551 {
1552         int     error = 0;
1553
1554         INIT_LIST_HEAD(&btp->bt_list);
1555         INIT_LIST_HEAD(&btp->bt_delwrite_queue);
1556         spinlock_init(&btp->bt_delwrite_lock, "delwri_lock");
1557         btp->bt_flags = 0;
1558         btp->bt_task = kthread_run(xfsbufd, btp, "xfsbufd");
1559         if (IS_ERR(btp->bt_task)) {
1560                 error = PTR_ERR(btp->bt_task);
1561                 goto out_error;
1562         }
1563         xfs_register_buftarg(btp);
1564 out_error:
1565         return error;
1566 }
1567
1568 xfs_buftarg_t *
1569 xfs_alloc_buftarg(
1570         struct block_device     *bdev,
1571         int                     external)
1572 {
1573         xfs_buftarg_t           *btp;
1574
1575         btp = kmem_zalloc(sizeof(*btp), KM_SLEEP);
1576
1577         btp->bt_dev =  bdev->bd_dev;
1578         btp->bt_bdev = bdev;
1579         if (xfs_setsize_buftarg_early(btp, bdev))
1580                 goto error;
1581         if (xfs_mapping_buftarg(btp, bdev))
1582                 goto error;
1583         if (xfs_alloc_delwrite_queue(btp))
1584                 goto error;
1585         xfs_alloc_bufhash(btp, external);
1586         return btp;
1587
1588 error:
1589         kmem_free(btp, sizeof(*btp));
1590         return NULL;
1591 }
1592
1593
1594 /*
1595  *      Delayed write buffer handling
1596  */
1597 STATIC void
1598 xfs_buf_delwri_queue(
1599         xfs_buf_t               *bp,
1600         int                     unlock)
1601 {
1602         struct list_head        *dwq = &bp->b_target->bt_delwrite_queue;
1603         spinlock_t              *dwlk = &bp->b_target->bt_delwrite_lock;
1604
1605         XB_TRACE(bp, "delwri_q", (long)unlock);
1606         ASSERT((bp->b_flags&(XBF_DELWRI|XBF_ASYNC)) == (XBF_DELWRI|XBF_ASYNC));
1607
1608         spin_lock(dwlk);
1609         /* If already in the queue, dequeue and place at tail */
1610         if (!list_empty(&bp->b_list)) {
1611                 ASSERT(bp->b_flags & _XBF_DELWRI_Q);
1612                 if (unlock)
1613                         atomic_dec(&bp->b_hold);
1614                 list_del(&bp->b_list);
1615         }
1616
1617         bp->b_flags |= _XBF_DELWRI_Q;
1618         list_add_tail(&bp->b_list, dwq);
1619         bp->b_queuetime = jiffies;
1620         spin_unlock(dwlk);
1621
1622         if (unlock)
1623                 xfs_buf_unlock(bp);
1624 }
1625
1626 void
1627 xfs_buf_delwri_dequeue(
1628         xfs_buf_t               *bp)
1629 {
1630         spinlock_t              *dwlk = &bp->b_target->bt_delwrite_lock;
1631         int                     dequeued = 0;
1632
1633         spin_lock(dwlk);
1634         if ((bp->b_flags & XBF_DELWRI) && !list_empty(&bp->b_list)) {
1635                 ASSERT(bp->b_flags & _XBF_DELWRI_Q);
1636                 list_del_init(&bp->b_list);
1637                 dequeued = 1;
1638         }
1639         bp->b_flags &= ~(XBF_DELWRI|_XBF_DELWRI_Q);
1640         spin_unlock(dwlk);
1641
1642         if (dequeued)
1643                 xfs_buf_rele(bp);
1644
1645         XB_TRACE(bp, "delwri_dq", (long)dequeued);
1646 }
1647
1648 STATIC void
1649 xfs_buf_runall_queues(
1650         struct workqueue_struct *queue)
1651 {
1652         flush_workqueue(queue);
1653 }
1654
1655 STATIC int
1656 xfsbufd_wakeup(
1657         int                     priority,
1658         gfp_t                   mask)
1659 {
1660         xfs_buftarg_t           *btp;
1661
1662         spin_lock(&xfs_buftarg_lock);
1663         list_for_each_entry(btp, &xfs_buftarg_list, bt_list) {
1664                 if (test_bit(XBT_FORCE_SLEEP, &btp->bt_flags))
1665                         continue;
1666                 set_bit(XBT_FORCE_FLUSH, &btp->bt_flags);
1667                 wake_up_process(btp->bt_task);
1668         }
1669         spin_unlock(&xfs_buftarg_lock);
1670         return 0;
1671 }
1672
1673 STATIC int
1674 xfsbufd(
1675         void                    *data)
1676 {
1677         struct list_head        tmp;
1678         unsigned long           age;
1679         xfs_buftarg_t           *target = (xfs_buftarg_t *)data;
1680         xfs_buf_t               *bp, *n;
1681         struct list_head        *dwq = &target->bt_delwrite_queue;
1682         spinlock_t              *dwlk = &target->bt_delwrite_lock;
1683
1684         current->flags |= PF_MEMALLOC;
1685
1686         INIT_LIST_HEAD(&tmp);
1687         do {
1688                 if (unlikely(freezing(current))) {
1689                         set_bit(XBT_FORCE_SLEEP, &target->bt_flags);
1690                         refrigerator();
1691                 } else {
1692                         clear_bit(XBT_FORCE_SLEEP, &target->bt_flags);
1693                 }
1694
1695                 schedule_timeout_interruptible(
1696                         xfs_buf_timer_centisecs * msecs_to_jiffies(10));
1697
1698                 age = xfs_buf_age_centisecs * msecs_to_jiffies(10);
1699                 spin_lock(dwlk);
1700                 list_for_each_entry_safe(bp, n, dwq, b_list) {
1701                         XB_TRACE(bp, "walkq1", (long)xfs_buf_ispin(bp));
1702                         ASSERT(bp->b_flags & XBF_DELWRI);
1703
1704                         if (!xfs_buf_ispin(bp) && !xfs_buf_cond_lock(bp)) {
1705                                 if (!test_bit(XBT_FORCE_FLUSH,
1706                                                 &target->bt_flags) &&
1707                                     time_before(jiffies,
1708                                                 bp->b_queuetime + age)) {
1709                                         xfs_buf_unlock(bp);
1710                                         break;
1711                                 }
1712
1713                                 bp->b_flags &= ~(XBF_DELWRI|_XBF_DELWRI_Q);
1714                                 bp->b_flags |= XBF_WRITE;
1715                                 list_move(&bp->b_list, &tmp);
1716                         }
1717                 }
1718                 spin_unlock(dwlk);
1719
1720                 while (!list_empty(&tmp)) {
1721                         bp = list_entry(tmp.next, xfs_buf_t, b_list);
1722                         ASSERT(target == bp->b_target);
1723
1724                         list_del_init(&bp->b_list);
1725                         xfs_buf_iostrategy(bp);
1726
1727                         blk_run_address_space(target->bt_mapping);
1728                 }
1729
1730                 if (as_list_len > 0)
1731                         purge_addresses();
1732
1733                 clear_bit(XBT_FORCE_FLUSH, &target->bt_flags);
1734         } while (!kthread_should_stop());
1735
1736         return 0;
1737 }
1738
1739 /*
1740  *      Go through all incore buffers, and release buffers if they belong to
1741  *      the given device. This is used in filesystem error handling to
1742  *      preserve the consistency of its metadata.
1743  */
1744 int
1745 xfs_flush_buftarg(
1746         xfs_buftarg_t           *target,
1747         int                     wait)
1748 {
1749         struct list_head        tmp;
1750         xfs_buf_t               *bp, *n;
1751         int                     pincount = 0;
1752         struct list_head        *dwq = &target->bt_delwrite_queue;
1753         spinlock_t              *dwlk = &target->bt_delwrite_lock;
1754
1755         xfs_buf_runall_queues(xfsdatad_workqueue);
1756         xfs_buf_runall_queues(xfslogd_workqueue);
1757
1758         INIT_LIST_HEAD(&tmp);
1759         spin_lock(dwlk);
1760         list_for_each_entry_safe(bp, n, dwq, b_list) {
1761                 ASSERT(bp->b_target == target);
1762                 ASSERT(bp->b_flags & (XBF_DELWRI | _XBF_DELWRI_Q));
1763                 XB_TRACE(bp, "walkq2", (long)xfs_buf_ispin(bp));
1764                 if (xfs_buf_ispin(bp)) {
1765                         pincount++;
1766                         continue;
1767                 }
1768
1769                 list_move(&bp->b_list, &tmp);
1770         }
1771         spin_unlock(dwlk);
1772
1773         /*
1774          * Dropped the delayed write list lock, now walk the temporary list
1775          */
1776         list_for_each_entry_safe(bp, n, &tmp, b_list) {
1777                 xfs_buf_lock(bp);
1778                 bp->b_flags &= ~(XBF_DELWRI|_XBF_DELWRI_Q);
1779                 bp->b_flags |= XBF_WRITE;
1780                 if (wait)
1781                         bp->b_flags &= ~XBF_ASYNC;
1782                 else
1783                         list_del_init(&bp->b_list);
1784
1785                 xfs_buf_iostrategy(bp);
1786         }
1787
1788         /*
1789          * Remaining list items must be flushed before returning
1790          */
1791         while (!list_empty(&tmp)) {
1792                 bp = list_entry(tmp.next, xfs_buf_t, b_list);
1793
1794                 list_del_init(&bp->b_list);
1795                 xfs_iowait(bp);
1796                 xfs_buf_relse(bp);
1797         }
1798
1799         if (wait)
1800                 blk_run_address_space(target->bt_mapping);
1801
1802         return pincount;
1803 }
1804
1805 int __init
1806 xfs_buf_init(void)
1807 {
1808         int             error = -ENOMEM;
1809
1810 #ifdef XFS_BUF_TRACE
1811         xfs_buf_trace_buf = ktrace_alloc(XFS_BUF_TRACE_SIZE, KM_SLEEP);
1812 #endif
1813
1814         xfs_buf_zone = kmem_zone_init(sizeof(xfs_buf_t), "xfs_buf");
1815         if (!xfs_buf_zone)
1816                 goto out_free_trace_buf;
1817
1818         xfslogd_workqueue = create_workqueue("xfslogd");
1819         if (!xfslogd_workqueue)
1820                 goto out_free_buf_zone;
1821
1822         xfsdatad_workqueue = create_workqueue("xfsdatad");
1823         if (!xfsdatad_workqueue)
1824                 goto out_destroy_xfslogd_workqueue;
1825
1826         xfs_buf_shake = kmem_shake_register(xfsbufd_wakeup);
1827         if (!xfs_buf_shake)
1828                 goto out_destroy_xfsdatad_workqueue;
1829
1830         return 0;
1831
1832  out_destroy_xfsdatad_workqueue:
1833         destroy_workqueue(xfsdatad_workqueue);
1834  out_destroy_xfslogd_workqueue:
1835         destroy_workqueue(xfslogd_workqueue);
1836  out_free_buf_zone:
1837         kmem_zone_destroy(xfs_buf_zone);
1838  out_free_trace_buf:
1839 #ifdef XFS_BUF_TRACE
1840         ktrace_free(xfs_buf_trace_buf);
1841 #endif
1842         return error;
1843 }
1844
1845 void
1846 xfs_buf_terminate(void)
1847 {
1848         kmem_shake_deregister(xfs_buf_shake);
1849         destroy_workqueue(xfsdatad_workqueue);
1850         destroy_workqueue(xfslogd_workqueue);
1851         kmem_zone_destroy(xfs_buf_zone);
1852 #ifdef XFS_BUF_TRACE
1853         ktrace_free(xfs_buf_trace_buf);
1854 #endif
1855 }