[XFS] cleanup root inode handling in xfs_fs_fill_super
[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.counter,
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_MUTEX_LOCKED(&bp->b_iodonesema);
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                           bp->b_page_count * sizeof(struct page *));
315         }
316 }
317
318 /*
319  *      Releases the specified buffer.
320  *
321  *      The modification state of any associated pages is left unchanged.
322  *      The buffer most not be on any hash - use xfs_buf_rele instead for
323  *      hashed and refcounted buffers
324  */
325 void
326 xfs_buf_free(
327         xfs_buf_t               *bp)
328 {
329         XB_TRACE(bp, "free", 0);
330
331         ASSERT(list_empty(&bp->b_hash_list));
332
333         if (bp->b_flags & (_XBF_PAGE_CACHE|_XBF_PAGES)) {
334                 uint            i;
335
336                 if ((bp->b_flags & XBF_MAPPED) && (bp->b_page_count > 1))
337                         free_address(bp->b_addr - bp->b_offset);
338
339                 for (i = 0; i < bp->b_page_count; i++) {
340                         struct page     *page = bp->b_pages[i];
341
342                         if (bp->b_flags & _XBF_PAGE_CACHE)
343                                 ASSERT(!PagePrivate(page));
344                         page_cache_release(page);
345                 }
346                 _xfs_buf_free_pages(bp);
347         }
348
349         xfs_buf_deallocate(bp);
350 }
351
352 /*
353  *      Finds all pages for buffer in question and builds it's page list.
354  */
355 STATIC int
356 _xfs_buf_lookup_pages(
357         xfs_buf_t               *bp,
358         uint                    flags)
359 {
360         struct address_space    *mapping = bp->b_target->bt_mapping;
361         size_t                  blocksize = bp->b_target->bt_bsize;
362         size_t                  size = bp->b_count_desired;
363         size_t                  nbytes, offset;
364         gfp_t                   gfp_mask = xb_to_gfp(flags);
365         unsigned short          page_count, i;
366         pgoff_t                 first;
367         xfs_off_t               end;
368         int                     error;
369
370         end = bp->b_file_offset + bp->b_buffer_length;
371         page_count = xfs_buf_btoc(end) - xfs_buf_btoct(bp->b_file_offset);
372
373         error = _xfs_buf_get_pages(bp, page_count, flags);
374         if (unlikely(error))
375                 return error;
376         bp->b_flags |= _XBF_PAGE_CACHE;
377
378         offset = bp->b_offset;
379         first = bp->b_file_offset >> PAGE_CACHE_SHIFT;
380
381         for (i = 0; i < bp->b_page_count; i++) {
382                 struct page     *page;
383                 uint            retries = 0;
384
385               retry:
386                 page = find_or_create_page(mapping, first + i, gfp_mask);
387                 if (unlikely(page == NULL)) {
388                         if (flags & XBF_READ_AHEAD) {
389                                 bp->b_page_count = i;
390                                 return -ENOMEM;
391                         }
392
393                         /*
394                          * This could deadlock.
395                          *
396                          * But until all the XFS lowlevel code is revamped to
397                          * handle buffer allocation failures we can't do much.
398                          */
399                         if (!(++retries % 100))
400                                 printk(KERN_ERR
401                                         "XFS: possible memory allocation "
402                                         "deadlock in %s (mode:0x%x)\n",
403                                         __FUNCTION__, gfp_mask);
404
405                         XFS_STATS_INC(xb_page_retries);
406                         xfsbufd_wakeup(0, gfp_mask);
407                         congestion_wait(WRITE, HZ/50);
408                         goto retry;
409                 }
410
411                 XFS_STATS_INC(xb_page_found);
412
413                 nbytes = min_t(size_t, size, PAGE_CACHE_SIZE - offset);
414                 size -= nbytes;
415
416                 ASSERT(!PagePrivate(page));
417                 if (!PageUptodate(page)) {
418                         page_count--;
419                         if (blocksize < PAGE_CACHE_SIZE && !PagePrivate(page)) {
420                                 if (test_page_region(page, offset, nbytes))
421                                         page_count++;
422                         }
423                 }
424
425                 unlock_page(page);
426                 bp->b_pages[i] = page;
427                 offset = 0;
428         }
429
430         if (page_count == bp->b_page_count)
431                 bp->b_flags |= XBF_DONE;
432
433         XB_TRACE(bp, "lookup_pages", (long)page_count);
434         return error;
435 }
436
437 /*
438  *      Map buffer into kernel address-space if nessecary.
439  */
440 STATIC int
441 _xfs_buf_map_pages(
442         xfs_buf_t               *bp,
443         uint                    flags)
444 {
445         /* A single page buffer is always mappable */
446         if (bp->b_page_count == 1) {
447                 bp->b_addr = page_address(bp->b_pages[0]) + bp->b_offset;
448                 bp->b_flags |= XBF_MAPPED;
449         } else if (flags & XBF_MAPPED) {
450                 if (as_list_len > 64)
451                         purge_addresses();
452                 bp->b_addr = vmap(bp->b_pages, bp->b_page_count,
453                                         VM_MAP, PAGE_KERNEL);
454                 if (unlikely(bp->b_addr == NULL))
455                         return -ENOMEM;
456                 bp->b_addr += bp->b_offset;
457                 bp->b_flags |= XBF_MAPPED;
458         }
459
460         return 0;
461 }
462
463 /*
464  *      Finding and Reading Buffers
465  */
466
467 /*
468  *      Look up, and creates if absent, a lockable buffer for
469  *      a given range of an inode.  The buffer is returned
470  *      locked.  If other overlapping buffers exist, they are
471  *      released before the new buffer is created and locked,
472  *      which may imply that this call will block until those buffers
473  *      are unlocked.  No I/O is implied by this call.
474  */
475 xfs_buf_t *
476 _xfs_buf_find(
477         xfs_buftarg_t           *btp,   /* block device target          */
478         xfs_off_t               ioff,   /* starting offset of range     */
479         size_t                  isize,  /* length of range              */
480         xfs_buf_flags_t         flags,
481         xfs_buf_t               *new_bp)
482 {
483         xfs_off_t               range_base;
484         size_t                  range_length;
485         xfs_bufhash_t           *hash;
486         xfs_buf_t               *bp, *n;
487
488         range_base = (ioff << BBSHIFT);
489         range_length = (isize << BBSHIFT);
490
491         /* Check for IOs smaller than the sector size / not sector aligned */
492         ASSERT(!(range_length < (1 << btp->bt_sshift)));
493         ASSERT(!(range_base & (xfs_off_t)btp->bt_smask));
494
495         hash = &btp->bt_hash[hash_long((unsigned long)ioff, btp->bt_hashshift)];
496
497         spin_lock(&hash->bh_lock);
498
499         list_for_each_entry_safe(bp, n, &hash->bh_list, b_hash_list) {
500                 ASSERT(btp == bp->b_target);
501                 if (bp->b_file_offset == range_base &&
502                     bp->b_buffer_length == range_length) {
503                         /*
504                          * If we look at something, bring it to the
505                          * front of the list for next time.
506                          */
507                         atomic_inc(&bp->b_hold);
508                         list_move(&bp->b_hash_list, &hash->bh_list);
509                         goto found;
510                 }
511         }
512
513         /* No match found */
514         if (new_bp) {
515                 _xfs_buf_initialize(new_bp, btp, range_base,
516                                 range_length, flags);
517                 new_bp->b_hash = hash;
518                 list_add(&new_bp->b_hash_list, &hash->bh_list);
519         } else {
520                 XFS_STATS_INC(xb_miss_locked);
521         }
522
523         spin_unlock(&hash->bh_lock);
524         return new_bp;
525
526 found:
527         spin_unlock(&hash->bh_lock);
528
529         /* Attempt to get the semaphore without sleeping,
530          * if this does not work then we need to drop the
531          * spinlock and do a hard attempt on the semaphore.
532          */
533         if (down_trylock(&bp->b_sema)) {
534                 if (!(flags & XBF_TRYLOCK)) {
535                         /* wait for buffer ownership */
536                         XB_TRACE(bp, "get_lock", 0);
537                         xfs_buf_lock(bp);
538                         XFS_STATS_INC(xb_get_locked_waited);
539                 } else {
540                         /* We asked for a trylock and failed, no need
541                          * to look at file offset and length here, we
542                          * know that this buffer at least overlaps our
543                          * buffer and is locked, therefore our buffer
544                          * either does not exist, or is this buffer.
545                          */
546                         xfs_buf_rele(bp);
547                         XFS_STATS_INC(xb_busy_locked);
548                         return NULL;
549                 }
550         } else {
551                 /* trylock worked */
552                 XB_SET_OWNER(bp);
553         }
554
555         if (bp->b_flags & XBF_STALE) {
556                 ASSERT((bp->b_flags & _XBF_DELWRI_Q) == 0);
557                 bp->b_flags &= XBF_MAPPED;
558         }
559         XB_TRACE(bp, "got_lock", 0);
560         XFS_STATS_INC(xb_get_locked);
561         return bp;
562 }
563
564 /*
565  *      Assembles a buffer covering the specified range.
566  *      Storage in memory for all portions of the buffer will be allocated,
567  *      although backing storage may not be.
568  */
569 xfs_buf_t *
570 xfs_buf_get_flags(
571         xfs_buftarg_t           *target,/* target for buffer            */
572         xfs_off_t               ioff,   /* starting offset of range     */
573         size_t                  isize,  /* length of range              */
574         xfs_buf_flags_t         flags)
575 {
576         xfs_buf_t               *bp, *new_bp;
577         int                     error = 0, i;
578
579         new_bp = xfs_buf_allocate(flags);
580         if (unlikely(!new_bp))
581                 return NULL;
582
583         bp = _xfs_buf_find(target, ioff, isize, flags, new_bp);
584         if (bp == new_bp) {
585                 error = _xfs_buf_lookup_pages(bp, flags);
586                 if (error)
587                         goto no_buffer;
588         } else {
589                 xfs_buf_deallocate(new_bp);
590                 if (unlikely(bp == NULL))
591                         return NULL;
592         }
593
594         for (i = 0; i < bp->b_page_count; i++)
595                 mark_page_accessed(bp->b_pages[i]);
596
597         if (!(bp->b_flags & XBF_MAPPED)) {
598                 error = _xfs_buf_map_pages(bp, flags);
599                 if (unlikely(error)) {
600                         printk(KERN_WARNING "%s: failed to map pages\n",
601                                         __FUNCTION__);
602                         goto no_buffer;
603                 }
604         }
605
606         XFS_STATS_INC(xb_get);
607
608         /*
609          * Always fill in the block number now, the mapped cases can do
610          * their own overlay of this later.
611          */
612         bp->b_bn = ioff;
613         bp->b_count_desired = bp->b_buffer_length;
614
615         XB_TRACE(bp, "get", (unsigned long)flags);
616         return bp;
617
618  no_buffer:
619         if (flags & (XBF_LOCK | XBF_TRYLOCK))
620                 xfs_buf_unlock(bp);
621         xfs_buf_rele(bp);
622         return NULL;
623 }
624
625 xfs_buf_t *
626 xfs_buf_read_flags(
627         xfs_buftarg_t           *target,
628         xfs_off_t               ioff,
629         size_t                  isize,
630         xfs_buf_flags_t         flags)
631 {
632         xfs_buf_t               *bp;
633
634         flags |= XBF_READ;
635
636         bp = xfs_buf_get_flags(target, ioff, isize, flags);
637         if (bp) {
638                 if (!XFS_BUF_ISDONE(bp)) {
639                         XB_TRACE(bp, "read", (unsigned long)flags);
640                         XFS_STATS_INC(xb_get_read);
641                         xfs_buf_iostart(bp, flags);
642                 } else if (flags & XBF_ASYNC) {
643                         XB_TRACE(bp, "read_async", (unsigned long)flags);
644                         /*
645                          * Read ahead call which is already satisfied,
646                          * drop the buffer
647                          */
648                         goto no_buffer;
649                 } else {
650                         XB_TRACE(bp, "read_done", (unsigned long)flags);
651                         /* We do not want read in the flags */
652                         bp->b_flags &= ~XBF_READ;
653                 }
654         }
655
656         return bp;
657
658  no_buffer:
659         if (flags & (XBF_LOCK | XBF_TRYLOCK))
660                 xfs_buf_unlock(bp);
661         xfs_buf_rele(bp);
662         return NULL;
663 }
664
665 /*
666  *      If we are not low on memory then do the readahead in a deadlock
667  *      safe manner.
668  */
669 void
670 xfs_buf_readahead(
671         xfs_buftarg_t           *target,
672         xfs_off_t               ioff,
673         size_t                  isize,
674         xfs_buf_flags_t         flags)
675 {
676         struct backing_dev_info *bdi;
677
678         bdi = target->bt_mapping->backing_dev_info;
679         if (bdi_read_congested(bdi))
680                 return;
681
682         flags |= (XBF_TRYLOCK|XBF_ASYNC|XBF_READ_AHEAD);
683         xfs_buf_read_flags(target, ioff, isize, flags);
684 }
685
686 xfs_buf_t *
687 xfs_buf_get_empty(
688         size_t                  len,
689         xfs_buftarg_t           *target)
690 {
691         xfs_buf_t               *bp;
692
693         bp = xfs_buf_allocate(0);
694         if (bp)
695                 _xfs_buf_initialize(bp, target, 0, len, 0);
696         return bp;
697 }
698
699 static inline struct page *
700 mem_to_page(
701         void                    *addr)
702 {
703         if ((!is_vmalloc_addr(addr))) {
704                 return virt_to_page(addr);
705         } else {
706                 return vmalloc_to_page(addr);
707         }
708 }
709
710 int
711 xfs_buf_associate_memory(
712         xfs_buf_t               *bp,
713         void                    *mem,
714         size_t                  len)
715 {
716         int                     rval;
717         int                     i = 0;
718         unsigned long           pageaddr;
719         unsigned long           offset;
720         size_t                  buflen;
721         int                     page_count;
722
723         pageaddr = (unsigned long)mem & PAGE_CACHE_MASK;
724         offset = (unsigned long)mem - pageaddr;
725         buflen = PAGE_CACHE_ALIGN(len + offset);
726         page_count = buflen >> PAGE_CACHE_SHIFT;
727
728         /* Free any previous set of page pointers */
729         if (bp->b_pages)
730                 _xfs_buf_free_pages(bp);
731
732         bp->b_pages = NULL;
733         bp->b_addr = mem;
734
735         rval = _xfs_buf_get_pages(bp, page_count, 0);
736         if (rval)
737                 return rval;
738
739         bp->b_offset = offset;
740
741         for (i = 0; i < bp->b_page_count; i++) {
742                 bp->b_pages[i] = mem_to_page((void *)pageaddr);
743                 pageaddr += PAGE_CACHE_SIZE;
744         }
745
746         bp->b_count_desired = len;
747         bp->b_buffer_length = buflen;
748         bp->b_flags |= XBF_MAPPED;
749
750         return 0;
751 }
752
753 xfs_buf_t *
754 xfs_buf_get_noaddr(
755         size_t                  len,
756         xfs_buftarg_t           *target)
757 {
758         unsigned long           page_count = PAGE_ALIGN(len) >> PAGE_SHIFT;
759         int                     error, i;
760         xfs_buf_t               *bp;
761
762         bp = xfs_buf_allocate(0);
763         if (unlikely(bp == NULL))
764                 goto fail;
765         _xfs_buf_initialize(bp, target, 0, len, 0);
766
767         error = _xfs_buf_get_pages(bp, page_count, 0);
768         if (error)
769                 goto fail_free_buf;
770
771         for (i = 0; i < page_count; i++) {
772                 bp->b_pages[i] = alloc_page(GFP_KERNEL);
773                 if (!bp->b_pages[i])
774                         goto fail_free_mem;
775         }
776         bp->b_flags |= _XBF_PAGES;
777
778         error = _xfs_buf_map_pages(bp, XBF_MAPPED);
779         if (unlikely(error)) {
780                 printk(KERN_WARNING "%s: failed to map pages\n",
781                                 __FUNCTION__);
782                 goto fail_free_mem;
783         }
784
785         xfs_buf_unlock(bp);
786
787         XB_TRACE(bp, "no_daddr", len);
788         return bp;
789
790  fail_free_mem:
791         while (--i >= 0)
792                 __free_page(bp->b_pages[i]);
793         _xfs_buf_free_pages(bp);
794  fail_free_buf:
795         xfs_buf_deallocate(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         struct work_struct      *work)
991 {
992         xfs_buf_t               *bp =
993                 container_of(work, xfs_buf_t, b_iodone_work);
994
995         /*
996          * We can get an EOPNOTSUPP to ordered writes.  Here we clear the
997          * ordered flag and reissue them.  Because we can't tell the higher
998          * layers directly that they should not issue ordered I/O anymore, they
999          * need to check if the ordered flag was cleared during I/O completion.
1000          */
1001         if ((bp->b_error == EOPNOTSUPP) &&
1002             (bp->b_flags & (XBF_ORDERED|XBF_ASYNC)) == (XBF_ORDERED|XBF_ASYNC)) {
1003                 XB_TRACE(bp, "ordered_retry", bp->b_iodone);
1004                 bp->b_flags &= ~XBF_ORDERED;
1005                 xfs_buf_iorequest(bp);
1006         } else if (bp->b_iodone)
1007                 (*(bp->b_iodone))(bp);
1008         else if (bp->b_flags & XBF_ASYNC)
1009                 xfs_buf_relse(bp);
1010 }
1011
1012 void
1013 xfs_buf_ioend(
1014         xfs_buf_t               *bp,
1015         int                     schedule)
1016 {
1017         bp->b_flags &= ~(XBF_READ | XBF_WRITE | XBF_READ_AHEAD);
1018         if (bp->b_error == 0)
1019                 bp->b_flags |= XBF_DONE;
1020
1021         XB_TRACE(bp, "iodone", bp->b_iodone);
1022
1023         if ((bp->b_iodone) || (bp->b_flags & XBF_ASYNC)) {
1024                 if (schedule) {
1025                         INIT_WORK(&bp->b_iodone_work, xfs_buf_iodone_work);
1026                         queue_work(xfslogd_workqueue, &bp->b_iodone_work);
1027                 } else {
1028                         xfs_buf_iodone_work(&bp->b_iodone_work);
1029                 }
1030         } else {
1031                 up(&bp->b_iodonesema);
1032         }
1033 }
1034
1035 void
1036 xfs_buf_ioerror(
1037         xfs_buf_t               *bp,
1038         int                     error)
1039 {
1040         ASSERT(error >= 0 && error <= 0xffff);
1041         bp->b_error = (unsigned short)error;
1042         XB_TRACE(bp, "ioerror", (unsigned long)error);
1043 }
1044
1045 /*
1046  *      Initiate I/O on a buffer, based on the flags supplied.
1047  *      The b_iodone routine in the buffer supplied will only be called
1048  *      when all of the subsidiary I/O requests, if any, have been completed.
1049  */
1050 int
1051 xfs_buf_iostart(
1052         xfs_buf_t               *bp,
1053         xfs_buf_flags_t         flags)
1054 {
1055         int                     status = 0;
1056
1057         XB_TRACE(bp, "iostart", (unsigned long)flags);
1058
1059         if (flags & XBF_DELWRI) {
1060                 bp->b_flags &= ~(XBF_READ | XBF_WRITE | XBF_ASYNC);
1061                 bp->b_flags |= flags & (XBF_DELWRI | XBF_ASYNC);
1062                 xfs_buf_delwri_queue(bp, 1);
1063                 return status;
1064         }
1065
1066         bp->b_flags &= ~(XBF_READ | XBF_WRITE | XBF_ASYNC | XBF_DELWRI | \
1067                         XBF_READ_AHEAD | _XBF_RUN_QUEUES);
1068         bp->b_flags |= flags & (XBF_READ | XBF_WRITE | XBF_ASYNC | \
1069                         XBF_READ_AHEAD | _XBF_RUN_QUEUES);
1070
1071         BUG_ON(bp->b_bn == XFS_BUF_DADDR_NULL);
1072
1073         /* For writes allow an alternate strategy routine to precede
1074          * the actual I/O request (which may not be issued at all in
1075          * a shutdown situation, for example).
1076          */
1077         status = (flags & XBF_WRITE) ?
1078                 xfs_buf_iostrategy(bp) : xfs_buf_iorequest(bp);
1079
1080         /* Wait for I/O if we are not an async request.
1081          * Note: async I/O request completion will release the buffer,
1082          * and that can already be done by this point.  So using the
1083          * buffer pointer from here on, after async I/O, is invalid.
1084          */
1085         if (!status && !(flags & XBF_ASYNC))
1086                 status = xfs_buf_iowait(bp);
1087
1088         return status;
1089 }
1090
1091 STATIC_INLINE void
1092 _xfs_buf_ioend(
1093         xfs_buf_t               *bp,
1094         int                     schedule)
1095 {
1096         if (atomic_dec_and_test(&bp->b_io_remaining) == 1)
1097                 xfs_buf_ioend(bp, schedule);
1098 }
1099
1100 STATIC void
1101 xfs_buf_bio_end_io(
1102         struct bio              *bio,
1103         int                     error)
1104 {
1105         xfs_buf_t               *bp = (xfs_buf_t *)bio->bi_private;
1106         unsigned int            blocksize = bp->b_target->bt_bsize;
1107         struct bio_vec          *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
1108
1109         if (!test_bit(BIO_UPTODATE, &bio->bi_flags))
1110                 bp->b_error = EIO;
1111
1112         do {
1113                 struct page     *page = bvec->bv_page;
1114
1115                 ASSERT(!PagePrivate(page));
1116                 if (unlikely(bp->b_error)) {
1117                         if (bp->b_flags & XBF_READ)
1118                                 ClearPageUptodate(page);
1119                 } else if (blocksize >= PAGE_CACHE_SIZE) {
1120                         SetPageUptodate(page);
1121                 } else if (!PagePrivate(page) &&
1122                                 (bp->b_flags & _XBF_PAGE_CACHE)) {
1123                         set_page_region(page, bvec->bv_offset, bvec->bv_len);
1124                 }
1125
1126                 if (--bvec >= bio->bi_io_vec)
1127                         prefetchw(&bvec->bv_page->flags);
1128         } while (bvec >= bio->bi_io_vec);
1129
1130         _xfs_buf_ioend(bp, 1);
1131         bio_put(bio);
1132 }
1133
1134 STATIC void
1135 _xfs_buf_ioapply(
1136         xfs_buf_t               *bp)
1137 {
1138         int                     rw, map_i, total_nr_pages, nr_pages;
1139         struct bio              *bio;
1140         int                     offset = bp->b_offset;
1141         int                     size = bp->b_count_desired;
1142         sector_t                sector = bp->b_bn;
1143         unsigned int            blocksize = bp->b_target->bt_bsize;
1144
1145         total_nr_pages = bp->b_page_count;
1146         map_i = 0;
1147
1148         if (bp->b_flags & XBF_ORDERED) {
1149                 ASSERT(!(bp->b_flags & XBF_READ));
1150                 rw = WRITE_BARRIER;
1151         } else if (bp->b_flags & _XBF_RUN_QUEUES) {
1152                 ASSERT(!(bp->b_flags & XBF_READ_AHEAD));
1153                 bp->b_flags &= ~_XBF_RUN_QUEUES;
1154                 rw = (bp->b_flags & XBF_WRITE) ? WRITE_SYNC : READ_SYNC;
1155         } else {
1156                 rw = (bp->b_flags & XBF_WRITE) ? WRITE :
1157                      (bp->b_flags & XBF_READ_AHEAD) ? READA : READ;
1158         }
1159
1160         /* Special code path for reading a sub page size buffer in --
1161          * we populate up the whole page, and hence the other metadata
1162          * in the same page.  This optimization is only valid when the
1163          * filesystem block size is not smaller than the page size.
1164          */
1165         if ((bp->b_buffer_length < PAGE_CACHE_SIZE) &&
1166             (bp->b_flags & XBF_READ) &&
1167             (blocksize >= PAGE_CACHE_SIZE)) {
1168                 bio = bio_alloc(GFP_NOIO, 1);
1169
1170                 bio->bi_bdev = bp->b_target->bt_bdev;
1171                 bio->bi_sector = sector - (offset >> BBSHIFT);
1172                 bio->bi_end_io = xfs_buf_bio_end_io;
1173                 bio->bi_private = bp;
1174
1175                 bio_add_page(bio, bp->b_pages[0], PAGE_CACHE_SIZE, 0);
1176                 size = 0;
1177
1178                 atomic_inc(&bp->b_io_remaining);
1179
1180                 goto submit_io;
1181         }
1182
1183 next_chunk:
1184         atomic_inc(&bp->b_io_remaining);
1185         nr_pages = BIO_MAX_SECTORS >> (PAGE_SHIFT - BBSHIFT);
1186         if (nr_pages > total_nr_pages)
1187                 nr_pages = total_nr_pages;
1188
1189         bio = bio_alloc(GFP_NOIO, nr_pages);
1190         bio->bi_bdev = bp->b_target->bt_bdev;
1191         bio->bi_sector = sector;
1192         bio->bi_end_io = xfs_buf_bio_end_io;
1193         bio->bi_private = bp;
1194
1195         for (; size && nr_pages; nr_pages--, map_i++) {
1196                 int     rbytes, nbytes = PAGE_CACHE_SIZE - offset;
1197
1198                 if (nbytes > size)
1199                         nbytes = size;
1200
1201                 rbytes = bio_add_page(bio, bp->b_pages[map_i], nbytes, offset);
1202                 if (rbytes < nbytes)
1203                         break;
1204
1205                 offset = 0;
1206                 sector += nbytes >> BBSHIFT;
1207                 size -= nbytes;
1208                 total_nr_pages--;
1209         }
1210
1211 submit_io:
1212         if (likely(bio->bi_size)) {
1213                 submit_bio(rw, bio);
1214                 if (size)
1215                         goto next_chunk;
1216         } else {
1217                 bio_put(bio);
1218                 xfs_buf_ioerror(bp, EIO);
1219         }
1220 }
1221
1222 int
1223 xfs_buf_iorequest(
1224         xfs_buf_t               *bp)
1225 {
1226         XB_TRACE(bp, "iorequest", 0);
1227
1228         if (bp->b_flags & XBF_DELWRI) {
1229                 xfs_buf_delwri_queue(bp, 1);
1230                 return 0;
1231         }
1232
1233         if (bp->b_flags & XBF_WRITE) {
1234                 xfs_buf_wait_unpin(bp);
1235         }
1236
1237         xfs_buf_hold(bp);
1238
1239         /* Set the count to 1 initially, this will stop an I/O
1240          * completion callout which happens before we have started
1241          * all the I/O from calling xfs_buf_ioend too early.
1242          */
1243         atomic_set(&bp->b_io_remaining, 1);
1244         _xfs_buf_ioapply(bp);
1245         _xfs_buf_ioend(bp, 0);
1246
1247         xfs_buf_rele(bp);
1248         return 0;
1249 }
1250
1251 /*
1252  *      Waits for I/O to complete on the buffer supplied.
1253  *      It returns immediately if no I/O is pending.
1254  *      It returns the I/O error code, if any, or 0 if there was no error.
1255  */
1256 int
1257 xfs_buf_iowait(
1258         xfs_buf_t               *bp)
1259 {
1260         XB_TRACE(bp, "iowait", 0);
1261         if (atomic_read(&bp->b_io_remaining))
1262                 blk_run_address_space(bp->b_target->bt_mapping);
1263         down(&bp->b_iodonesema);
1264         XB_TRACE(bp, "iowaited", (long)bp->b_error);
1265         return bp->b_error;
1266 }
1267
1268 xfs_caddr_t
1269 xfs_buf_offset(
1270         xfs_buf_t               *bp,
1271         size_t                  offset)
1272 {
1273         struct page             *page;
1274
1275         if (bp->b_flags & XBF_MAPPED)
1276                 return XFS_BUF_PTR(bp) + offset;
1277
1278         offset += bp->b_offset;
1279         page = bp->b_pages[offset >> PAGE_CACHE_SHIFT];
1280         return (xfs_caddr_t)page_address(page) + (offset & (PAGE_CACHE_SIZE-1));
1281 }
1282
1283 /*
1284  *      Move data into or out of a buffer.
1285  */
1286 void
1287 xfs_buf_iomove(
1288         xfs_buf_t               *bp,    /* buffer to process            */
1289         size_t                  boff,   /* starting buffer offset       */
1290         size_t                  bsize,  /* length to copy               */
1291         caddr_t                 data,   /* data address                 */
1292         xfs_buf_rw_t            mode)   /* read/write/zero flag         */
1293 {
1294         size_t                  bend, cpoff, csize;
1295         struct page             *page;
1296
1297         bend = boff + bsize;
1298         while (boff < bend) {
1299                 page = bp->b_pages[xfs_buf_btoct(boff + bp->b_offset)];
1300                 cpoff = xfs_buf_poff(boff + bp->b_offset);
1301                 csize = min_t(size_t,
1302                               PAGE_CACHE_SIZE-cpoff, bp->b_count_desired-boff);
1303
1304                 ASSERT(((csize + cpoff) <= PAGE_CACHE_SIZE));
1305
1306                 switch (mode) {
1307                 case XBRW_ZERO:
1308                         memset(page_address(page) + cpoff, 0, csize);
1309                         break;
1310                 case XBRW_READ:
1311                         memcpy(data, page_address(page) + cpoff, csize);
1312                         break;
1313                 case XBRW_WRITE:
1314                         memcpy(page_address(page) + cpoff, data, csize);
1315                 }
1316
1317                 boff += csize;
1318                 data += csize;
1319         }
1320 }
1321
1322 /*
1323  *      Handling of buffer targets (buftargs).
1324  */
1325
1326 /*
1327  *      Wait for any bufs with callbacks that have been submitted but
1328  *      have not yet returned... walk the hash list for the target.
1329  */
1330 void
1331 xfs_wait_buftarg(
1332         xfs_buftarg_t   *btp)
1333 {
1334         xfs_buf_t       *bp, *n;
1335         xfs_bufhash_t   *hash;
1336         uint            i;
1337
1338         for (i = 0; i < (1 << btp->bt_hashshift); i++) {
1339                 hash = &btp->bt_hash[i];
1340 again:
1341                 spin_lock(&hash->bh_lock);
1342                 list_for_each_entry_safe(bp, n, &hash->bh_list, b_hash_list) {
1343                         ASSERT(btp == bp->b_target);
1344                         if (!(bp->b_flags & XBF_FS_MANAGED)) {
1345                                 spin_unlock(&hash->bh_lock);
1346                                 /*
1347                                  * Catch superblock reference count leaks
1348                                  * immediately
1349                                  */
1350                                 BUG_ON(bp->b_bn == 0);
1351                                 delay(100);
1352                                 goto again;
1353                         }
1354                 }
1355                 spin_unlock(&hash->bh_lock);
1356         }
1357 }
1358
1359 /*
1360  *      Allocate buffer hash table for a given target.
1361  *      For devices containing metadata (i.e. not the log/realtime devices)
1362  *      we need to allocate a much larger hash table.
1363  */
1364 STATIC void
1365 xfs_alloc_bufhash(
1366         xfs_buftarg_t           *btp,
1367         int                     external)
1368 {
1369         unsigned int            i;
1370
1371         btp->bt_hashshift = external ? 3 : 8;   /* 8 or 256 buckets */
1372         btp->bt_hashmask = (1 << btp->bt_hashshift) - 1;
1373         btp->bt_hash = kmem_zalloc((1 << btp->bt_hashshift) *
1374                                         sizeof(xfs_bufhash_t), KM_SLEEP | KM_LARGE);
1375         for (i = 0; i < (1 << btp->bt_hashshift); i++) {
1376                 spin_lock_init(&btp->bt_hash[i].bh_lock);
1377                 INIT_LIST_HEAD(&btp->bt_hash[i].bh_list);
1378         }
1379 }
1380
1381 STATIC void
1382 xfs_free_bufhash(
1383         xfs_buftarg_t           *btp)
1384 {
1385         kmem_free(btp->bt_hash, (1<<btp->bt_hashshift) * sizeof(xfs_bufhash_t));
1386         btp->bt_hash = NULL;
1387 }
1388
1389 /*
1390  *      buftarg list for delwrite queue processing
1391  */
1392 static LIST_HEAD(xfs_buftarg_list);
1393 static DEFINE_SPINLOCK(xfs_buftarg_lock);
1394
1395 STATIC void
1396 xfs_register_buftarg(
1397         xfs_buftarg_t           *btp)
1398 {
1399         spin_lock(&xfs_buftarg_lock);
1400         list_add(&btp->bt_list, &xfs_buftarg_list);
1401         spin_unlock(&xfs_buftarg_lock);
1402 }
1403
1404 STATIC void
1405 xfs_unregister_buftarg(
1406         xfs_buftarg_t           *btp)
1407 {
1408         spin_lock(&xfs_buftarg_lock);
1409         list_del(&btp->bt_list);
1410         spin_unlock(&xfs_buftarg_lock);
1411 }
1412
1413 void
1414 xfs_free_buftarg(
1415         xfs_buftarg_t           *btp,
1416         int                     external)
1417 {
1418         xfs_flush_buftarg(btp, 1);
1419         xfs_blkdev_issue_flush(btp);
1420         if (external)
1421                 xfs_blkdev_put(btp->bt_bdev);
1422         xfs_free_bufhash(btp);
1423         iput(btp->bt_mapping->host);
1424
1425         /* Unregister the buftarg first so that we don't get a
1426          * wakeup finding a non-existent task
1427          */
1428         xfs_unregister_buftarg(btp);
1429         kthread_stop(btp->bt_task);
1430
1431         kmem_free(btp, sizeof(*btp));
1432 }
1433
1434 STATIC int
1435 xfs_setsize_buftarg_flags(
1436         xfs_buftarg_t           *btp,
1437         unsigned int            blocksize,
1438         unsigned int            sectorsize,
1439         int                     verbose)
1440 {
1441         btp->bt_bsize = blocksize;
1442         btp->bt_sshift = ffs(sectorsize) - 1;
1443         btp->bt_smask = sectorsize - 1;
1444
1445         if (set_blocksize(btp->bt_bdev, sectorsize)) {
1446                 printk(KERN_WARNING
1447                         "XFS: Cannot set_blocksize to %u on device %s\n",
1448                         sectorsize, XFS_BUFTARG_NAME(btp));
1449                 return EINVAL;
1450         }
1451
1452         if (verbose &&
1453             (PAGE_CACHE_SIZE / BITS_PER_LONG) > sectorsize) {
1454                 printk(KERN_WARNING
1455                         "XFS: %u byte sectors in use on device %s.  "
1456                         "This is suboptimal; %u or greater is ideal.\n",
1457                         sectorsize, XFS_BUFTARG_NAME(btp),
1458                         (unsigned int)PAGE_CACHE_SIZE / BITS_PER_LONG);
1459         }
1460
1461         return 0;
1462 }
1463
1464 /*
1465  *      When allocating the initial buffer target we have not yet
1466  *      read in the superblock, so don't know what sized sectors
1467  *      are being used is at this early stage.  Play safe.
1468  */
1469 STATIC int
1470 xfs_setsize_buftarg_early(
1471         xfs_buftarg_t           *btp,
1472         struct block_device     *bdev)
1473 {
1474         return xfs_setsize_buftarg_flags(btp,
1475                         PAGE_CACHE_SIZE, bdev_hardsect_size(bdev), 0);
1476 }
1477
1478 int
1479 xfs_setsize_buftarg(
1480         xfs_buftarg_t           *btp,
1481         unsigned int            blocksize,
1482         unsigned int            sectorsize)
1483 {
1484         return xfs_setsize_buftarg_flags(btp, blocksize, sectorsize, 1);
1485 }
1486
1487 STATIC int
1488 xfs_mapping_buftarg(
1489         xfs_buftarg_t           *btp,
1490         struct block_device     *bdev)
1491 {
1492         struct backing_dev_info *bdi;
1493         struct inode            *inode;
1494         struct address_space    *mapping;
1495         static const struct address_space_operations mapping_aops = {
1496                 .sync_page = block_sync_page,
1497                 .migratepage = fail_migrate_page,
1498         };
1499
1500         inode = new_inode(bdev->bd_inode->i_sb);
1501         if (!inode) {
1502                 printk(KERN_WARNING
1503                         "XFS: Cannot allocate mapping inode for device %s\n",
1504                         XFS_BUFTARG_NAME(btp));
1505                 return ENOMEM;
1506         }
1507         inode->i_mode = S_IFBLK;
1508         inode->i_bdev = bdev;
1509         inode->i_rdev = bdev->bd_dev;
1510         bdi = blk_get_backing_dev_info(bdev);
1511         if (!bdi)
1512                 bdi = &default_backing_dev_info;
1513         mapping = &inode->i_data;
1514         mapping->a_ops = &mapping_aops;
1515         mapping->backing_dev_info = bdi;
1516         mapping_set_gfp_mask(mapping, GFP_NOFS);
1517         btp->bt_mapping = mapping;
1518         return 0;
1519 }
1520
1521 STATIC int
1522 xfs_alloc_delwrite_queue(
1523         xfs_buftarg_t           *btp)
1524 {
1525         int     error = 0;
1526
1527         INIT_LIST_HEAD(&btp->bt_list);
1528         INIT_LIST_HEAD(&btp->bt_delwrite_queue);
1529         spin_lock_init(&btp->bt_delwrite_lock);
1530         btp->bt_flags = 0;
1531         btp->bt_task = kthread_run(xfsbufd, btp, "xfsbufd");
1532         if (IS_ERR(btp->bt_task)) {
1533                 error = PTR_ERR(btp->bt_task);
1534                 goto out_error;
1535         }
1536         xfs_register_buftarg(btp);
1537 out_error:
1538         return error;
1539 }
1540
1541 xfs_buftarg_t *
1542 xfs_alloc_buftarg(
1543         struct block_device     *bdev,
1544         int                     external)
1545 {
1546         xfs_buftarg_t           *btp;
1547
1548         btp = kmem_zalloc(sizeof(*btp), KM_SLEEP);
1549
1550         btp->bt_dev =  bdev->bd_dev;
1551         btp->bt_bdev = bdev;
1552         if (xfs_setsize_buftarg_early(btp, bdev))
1553                 goto error;
1554         if (xfs_mapping_buftarg(btp, bdev))
1555                 goto error;
1556         if (xfs_alloc_delwrite_queue(btp))
1557                 goto error;
1558         xfs_alloc_bufhash(btp, external);
1559         return btp;
1560
1561 error:
1562         kmem_free(btp, sizeof(*btp));
1563         return NULL;
1564 }
1565
1566
1567 /*
1568  *      Delayed write buffer handling
1569  */
1570 STATIC void
1571 xfs_buf_delwri_queue(
1572         xfs_buf_t               *bp,
1573         int                     unlock)
1574 {
1575         struct list_head        *dwq = &bp->b_target->bt_delwrite_queue;
1576         spinlock_t              *dwlk = &bp->b_target->bt_delwrite_lock;
1577
1578         XB_TRACE(bp, "delwri_q", (long)unlock);
1579         ASSERT((bp->b_flags&(XBF_DELWRI|XBF_ASYNC)) == (XBF_DELWRI|XBF_ASYNC));
1580
1581         spin_lock(dwlk);
1582         /* If already in the queue, dequeue and place at tail */
1583         if (!list_empty(&bp->b_list)) {
1584                 ASSERT(bp->b_flags & _XBF_DELWRI_Q);
1585                 if (unlock)
1586                         atomic_dec(&bp->b_hold);
1587                 list_del(&bp->b_list);
1588         }
1589
1590         bp->b_flags |= _XBF_DELWRI_Q;
1591         list_add_tail(&bp->b_list, dwq);
1592         bp->b_queuetime = jiffies;
1593         spin_unlock(dwlk);
1594
1595         if (unlock)
1596                 xfs_buf_unlock(bp);
1597 }
1598
1599 void
1600 xfs_buf_delwri_dequeue(
1601         xfs_buf_t               *bp)
1602 {
1603         spinlock_t              *dwlk = &bp->b_target->bt_delwrite_lock;
1604         int                     dequeued = 0;
1605
1606         spin_lock(dwlk);
1607         if ((bp->b_flags & XBF_DELWRI) && !list_empty(&bp->b_list)) {
1608                 ASSERT(bp->b_flags & _XBF_DELWRI_Q);
1609                 list_del_init(&bp->b_list);
1610                 dequeued = 1;
1611         }
1612         bp->b_flags &= ~(XBF_DELWRI|_XBF_DELWRI_Q);
1613         spin_unlock(dwlk);
1614
1615         if (dequeued)
1616                 xfs_buf_rele(bp);
1617
1618         XB_TRACE(bp, "delwri_dq", (long)dequeued);
1619 }
1620
1621 STATIC void
1622 xfs_buf_runall_queues(
1623         struct workqueue_struct *queue)
1624 {
1625         flush_workqueue(queue);
1626 }
1627
1628 STATIC int
1629 xfsbufd_wakeup(
1630         int                     priority,
1631         gfp_t                   mask)
1632 {
1633         xfs_buftarg_t           *btp;
1634
1635         spin_lock(&xfs_buftarg_lock);
1636         list_for_each_entry(btp, &xfs_buftarg_list, bt_list) {
1637                 if (test_bit(XBT_FORCE_SLEEP, &btp->bt_flags))
1638                         continue;
1639                 set_bit(XBT_FORCE_FLUSH, &btp->bt_flags);
1640                 wake_up_process(btp->bt_task);
1641         }
1642         spin_unlock(&xfs_buftarg_lock);
1643         return 0;
1644 }
1645
1646 /*
1647  * Move as many buffers as specified to the supplied list
1648  * idicating if we skipped any buffers to prevent deadlocks.
1649  */
1650 STATIC int
1651 xfs_buf_delwri_split(
1652         xfs_buftarg_t   *target,
1653         struct list_head *list,
1654         unsigned long   age)
1655 {
1656         xfs_buf_t       *bp, *n;
1657         struct list_head *dwq = &target->bt_delwrite_queue;
1658         spinlock_t      *dwlk = &target->bt_delwrite_lock;
1659         int             skipped = 0;
1660         int             force;
1661
1662         force = test_and_clear_bit(XBT_FORCE_FLUSH, &target->bt_flags);
1663         INIT_LIST_HEAD(list);
1664         spin_lock(dwlk);
1665         list_for_each_entry_safe(bp, n, dwq, b_list) {
1666                 XB_TRACE(bp, "walkq1", (long)xfs_buf_ispin(bp));
1667                 ASSERT(bp->b_flags & XBF_DELWRI);
1668
1669                 if (!xfs_buf_ispin(bp) && !xfs_buf_cond_lock(bp)) {
1670                         if (!force &&
1671                             time_before(jiffies, bp->b_queuetime + age)) {
1672                                 xfs_buf_unlock(bp);
1673                                 break;
1674                         }
1675
1676                         bp->b_flags &= ~(XBF_DELWRI|_XBF_DELWRI_Q|
1677                                          _XBF_RUN_QUEUES);
1678                         bp->b_flags |= XBF_WRITE;
1679                         list_move_tail(&bp->b_list, list);
1680                 } else
1681                         skipped++;
1682         }
1683         spin_unlock(dwlk);
1684
1685         return skipped;
1686
1687 }
1688
1689 STATIC int
1690 xfsbufd(
1691         void            *data)
1692 {
1693         struct list_head tmp;
1694         xfs_buftarg_t   *target = (xfs_buftarg_t *)data;
1695         int             count;
1696         xfs_buf_t       *bp;
1697
1698         current->flags |= PF_MEMALLOC;
1699
1700         set_freezable();
1701
1702         do {
1703                 if (unlikely(freezing(current))) {
1704                         set_bit(XBT_FORCE_SLEEP, &target->bt_flags);
1705                         refrigerator();
1706                 } else {
1707                         clear_bit(XBT_FORCE_SLEEP, &target->bt_flags);
1708                 }
1709
1710                 schedule_timeout_interruptible(
1711                         xfs_buf_timer_centisecs * msecs_to_jiffies(10));
1712
1713                 xfs_buf_delwri_split(target, &tmp,
1714                                 xfs_buf_age_centisecs * msecs_to_jiffies(10));
1715
1716                 count = 0;
1717                 while (!list_empty(&tmp)) {
1718                         bp = list_entry(tmp.next, xfs_buf_t, b_list);
1719                         ASSERT(target == bp->b_target);
1720
1721                         list_del_init(&bp->b_list);
1722                         xfs_buf_iostrategy(bp);
1723                         count++;
1724                 }
1725
1726                 if (as_list_len > 0)
1727                         purge_addresses();
1728                 if (count)
1729                         blk_run_address_space(target->bt_mapping);
1730
1731         } while (!kthread_should_stop());
1732
1733         return 0;
1734 }
1735
1736 /*
1737  *      Go through all incore buffers, and release buffers if they belong to
1738  *      the given device. This is used in filesystem error handling to
1739  *      preserve the consistency of its metadata.
1740  */
1741 int
1742 xfs_flush_buftarg(
1743         xfs_buftarg_t   *target,
1744         int             wait)
1745 {
1746         struct list_head tmp;
1747         xfs_buf_t       *bp, *n;
1748         int             pincount = 0;
1749
1750         xfs_buf_runall_queues(xfsdatad_workqueue);
1751         xfs_buf_runall_queues(xfslogd_workqueue);
1752
1753         set_bit(XBT_FORCE_FLUSH, &target->bt_flags);
1754         pincount = xfs_buf_delwri_split(target, &tmp, 0);
1755
1756         /*
1757          * Dropped the delayed write list lock, now walk the temporary list
1758          */
1759         list_for_each_entry_safe(bp, n, &tmp, b_list) {
1760                 ASSERT(target == bp->b_target);
1761                 if (wait)
1762                         bp->b_flags &= ~XBF_ASYNC;
1763                 else
1764                         list_del_init(&bp->b_list);
1765
1766                 xfs_buf_iostrategy(bp);
1767         }
1768
1769         if (wait)
1770                 blk_run_address_space(target->bt_mapping);
1771
1772         /*
1773          * Remaining list items must be flushed before returning
1774          */
1775         while (!list_empty(&tmp)) {
1776                 bp = list_entry(tmp.next, xfs_buf_t, b_list);
1777
1778                 list_del_init(&bp->b_list);
1779                 xfs_iowait(bp);
1780                 xfs_buf_relse(bp);
1781         }
1782
1783         return pincount;
1784 }
1785
1786 int __init
1787 xfs_buf_init(void)
1788 {
1789 #ifdef XFS_BUF_TRACE
1790         xfs_buf_trace_buf = ktrace_alloc(XFS_BUF_TRACE_SIZE, KM_SLEEP);
1791 #endif
1792
1793         xfs_buf_zone = kmem_zone_init_flags(sizeof(xfs_buf_t), "xfs_buf",
1794                                                 KM_ZONE_HWALIGN, NULL);
1795         if (!xfs_buf_zone)
1796                 goto out_free_trace_buf;
1797
1798         xfslogd_workqueue = create_workqueue("xfslogd");
1799         if (!xfslogd_workqueue)
1800                 goto out_free_buf_zone;
1801
1802         xfsdatad_workqueue = create_workqueue("xfsdatad");
1803         if (!xfsdatad_workqueue)
1804                 goto out_destroy_xfslogd_workqueue;
1805
1806         register_shrinker(&xfs_buf_shake);
1807         return 0;
1808
1809  out_destroy_xfslogd_workqueue:
1810         destroy_workqueue(xfslogd_workqueue);
1811  out_free_buf_zone:
1812         kmem_zone_destroy(xfs_buf_zone);
1813  out_free_trace_buf:
1814 #ifdef XFS_BUF_TRACE
1815         ktrace_free(xfs_buf_trace_buf);
1816 #endif
1817         return -ENOMEM;
1818 }
1819
1820 void
1821 xfs_buf_terminate(void)
1822 {
1823         unregister_shrinker(&xfs_buf_shake);
1824         destroy_workqueue(xfsdatad_workqueue);
1825         destroy_workqueue(xfslogd_workqueue);
1826         kmem_zone_destroy(xfs_buf_zone);
1827 #ifdef XFS_BUF_TRACE
1828         ktrace_free(xfs_buf_trace_buf);
1829 #endif
1830 }
1831
1832 #ifdef CONFIG_KDB_MODULES
1833 struct list_head *
1834 xfs_get_buftarg_list(void)
1835 {
1836         return &xfs_buftarg_list;
1837 }
1838 #endif