block: Do away with the notion of hardsect_size
[linux-2.6.git] / 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 #include "xfs_sb.h"
38 #include "xfs_inum.h"
39 #include "xfs_ag.h"
40 #include "xfs_dmapi.h"
41 #include "xfs_mount.h"
42
43 static kmem_zone_t *xfs_buf_zone;
44 STATIC int xfsbufd(void *);
45 STATIC int xfsbufd_wakeup(int, gfp_t);
46 STATIC void xfs_buf_delwri_queue(xfs_buf_t *, int);
47 static struct shrinker xfs_buf_shake = {
48         .shrink = xfsbufd_wakeup,
49         .seeks = DEFAULT_SEEKS,
50 };
51
52 static struct workqueue_struct *xfslogd_workqueue;
53 struct workqueue_struct *xfsdatad_workqueue;
54 struct workqueue_struct *xfsconvertd_workqueue;
55
56 #ifdef XFS_BUF_TRACE
57 void
58 xfs_buf_trace(
59         xfs_buf_t       *bp,
60         char            *id,
61         void            *data,
62         void            *ra)
63 {
64         ktrace_enter(xfs_buf_trace_buf,
65                 bp, id,
66                 (void *)(unsigned long)bp->b_flags,
67                 (void *)(unsigned long)bp->b_hold.counter,
68                 (void *)(unsigned long)bp->b_sema.count,
69                 (void *)current,
70                 data, ra,
71                 (void *)(unsigned long)((bp->b_file_offset>>32) & 0xffffffff),
72                 (void *)(unsigned long)(bp->b_file_offset & 0xffffffff),
73                 (void *)(unsigned long)bp->b_buffer_length,
74                 NULL, NULL, NULL, NULL, NULL);
75 }
76 ktrace_t *xfs_buf_trace_buf;
77 #define XFS_BUF_TRACE_SIZE      4096
78 #define XB_TRACE(bp, id, data)  \
79         xfs_buf_trace(bp, id, (void *)data, (void *)__builtin_return_address(0))
80 #else
81 #define XB_TRACE(bp, id, data)  do { } while (0)
82 #endif
83
84 #ifdef XFS_BUF_LOCK_TRACKING
85 # define XB_SET_OWNER(bp)       ((bp)->b_last_holder = current->pid)
86 # define XB_CLEAR_OWNER(bp)     ((bp)->b_last_holder = -1)
87 # define XB_GET_OWNER(bp)       ((bp)->b_last_holder)
88 #else
89 # define XB_SET_OWNER(bp)       do { } while (0)
90 # define XB_CLEAR_OWNER(bp)     do { } while (0)
91 # define XB_GET_OWNER(bp)       do { } while (0)
92 #endif
93
94 #define xb_to_gfp(flags) \
95         ((((flags) & XBF_READ_AHEAD) ? __GFP_NORETRY : \
96           ((flags) & XBF_DONT_BLOCK) ? GFP_NOFS : GFP_KERNEL) | __GFP_NOWARN)
97
98 #define xb_to_km(flags) \
99          (((flags) & XBF_DONT_BLOCK) ? KM_NOFS : KM_SLEEP)
100
101 #define xfs_buf_allocate(flags) \
102         kmem_zone_alloc(xfs_buf_zone, xb_to_km(flags))
103 #define xfs_buf_deallocate(bp) \
104         kmem_zone_free(xfs_buf_zone, (bp));
105
106 /*
107  *      Page Region interfaces.
108  *
109  *      For pages in filesystems where the blocksize is smaller than the
110  *      pagesize, we use the page->private field (long) to hold a bitmap
111  *      of uptodate regions within the page.
112  *
113  *      Each such region is "bytes per page / bits per long" bytes long.
114  *
115  *      NBPPR == number-of-bytes-per-page-region
116  *      BTOPR == bytes-to-page-region (rounded up)
117  *      BTOPRT == bytes-to-page-region-truncated (rounded down)
118  */
119 #if (BITS_PER_LONG == 32)
120 #define PRSHIFT         (PAGE_CACHE_SHIFT - 5)  /* (32 == 1<<5) */
121 #elif (BITS_PER_LONG == 64)
122 #define PRSHIFT         (PAGE_CACHE_SHIFT - 6)  /* (64 == 1<<6) */
123 #else
124 #error BITS_PER_LONG must be 32 or 64
125 #endif
126 #define NBPPR           (PAGE_CACHE_SIZE/BITS_PER_LONG)
127 #define BTOPR(b)        (((unsigned int)(b) + (NBPPR - 1)) >> PRSHIFT)
128 #define BTOPRT(b)       (((unsigned int)(b) >> PRSHIFT))
129
130 STATIC unsigned long
131 page_region_mask(
132         size_t          offset,
133         size_t          length)
134 {
135         unsigned long   mask;
136         int             first, final;
137
138         first = BTOPR(offset);
139         final = BTOPRT(offset + length - 1);
140         first = min(first, final);
141
142         mask = ~0UL;
143         mask <<= BITS_PER_LONG - (final - first);
144         mask >>= BITS_PER_LONG - (final);
145
146         ASSERT(offset + length <= PAGE_CACHE_SIZE);
147         ASSERT((final - first) < BITS_PER_LONG && (final - first) >= 0);
148
149         return mask;
150 }
151
152 STATIC_INLINE void
153 set_page_region(
154         struct page     *page,
155         size_t          offset,
156         size_t          length)
157 {
158         set_page_private(page,
159                 page_private(page) | page_region_mask(offset, length));
160         if (page_private(page) == ~0UL)
161                 SetPageUptodate(page);
162 }
163
164 STATIC_INLINE int
165 test_page_region(
166         struct page     *page,
167         size_t          offset,
168         size_t          length)
169 {
170         unsigned long   mask = page_region_mask(offset, length);
171
172         return (mask && (page_private(page) & mask) == mask);
173 }
174
175 /*
176  *      Mapping of multi-page buffers into contiguous virtual space
177  */
178
179 typedef struct a_list {
180         void            *vm_addr;
181         struct a_list   *next;
182 } a_list_t;
183
184 static a_list_t         *as_free_head;
185 static int              as_list_len;
186 static DEFINE_SPINLOCK(as_lock);
187
188 /*
189  *      Try to batch vunmaps because they are costly.
190  */
191 STATIC void
192 free_address(
193         void            *addr)
194 {
195         a_list_t        *aentry;
196
197 #ifdef CONFIG_XEN
198         /*
199          * Xen needs to be able to make sure it can get an exclusive
200          * RO mapping of pages it wants to turn into a pagetable.  If
201          * a newly allocated page is also still being vmap()ed by xfs,
202          * it will cause pagetable construction to fail.  This is a
203          * quick workaround to always eagerly unmap pages so that Xen
204          * is happy.
205          */
206         vunmap(addr);
207         return;
208 #endif
209
210         aentry = kmalloc(sizeof(a_list_t), GFP_NOWAIT);
211         if (likely(aentry)) {
212                 spin_lock(&as_lock);
213                 aentry->next = as_free_head;
214                 aentry->vm_addr = addr;
215                 as_free_head = aentry;
216                 as_list_len++;
217                 spin_unlock(&as_lock);
218         } else {
219                 vunmap(addr);
220         }
221 }
222
223 STATIC void
224 purge_addresses(void)
225 {
226         a_list_t        *aentry, *old;
227
228         if (as_free_head == NULL)
229                 return;
230
231         spin_lock(&as_lock);
232         aentry = as_free_head;
233         as_free_head = NULL;
234         as_list_len = 0;
235         spin_unlock(&as_lock);
236
237         while ((old = aentry) != NULL) {
238                 vunmap(aentry->vm_addr);
239                 aentry = aentry->next;
240                 kfree(old);
241         }
242 }
243
244 /*
245  *      Internal xfs_buf_t object manipulation
246  */
247
248 STATIC void
249 _xfs_buf_initialize(
250         xfs_buf_t               *bp,
251         xfs_buftarg_t           *target,
252         xfs_off_t               range_base,
253         size_t                  range_length,
254         xfs_buf_flags_t         flags)
255 {
256         /*
257          * We don't want certain flags to appear in b_flags.
258          */
259         flags &= ~(XBF_LOCK|XBF_MAPPED|XBF_DONT_BLOCK|XBF_READ_AHEAD);
260
261         memset(bp, 0, sizeof(xfs_buf_t));
262         atomic_set(&bp->b_hold, 1);
263         init_completion(&bp->b_iowait);
264         INIT_LIST_HEAD(&bp->b_list);
265         INIT_LIST_HEAD(&bp->b_hash_list);
266         init_MUTEX_LOCKED(&bp->b_sema); /* held, no waiters */
267         XB_SET_OWNER(bp);
268         bp->b_target = target;
269         bp->b_file_offset = range_base;
270         /*
271          * Set buffer_length and count_desired to the same value initially.
272          * I/O routines should use count_desired, which will be the same in
273          * most cases but may be reset (e.g. XFS recovery).
274          */
275         bp->b_buffer_length = bp->b_count_desired = range_length;
276         bp->b_flags = flags;
277         bp->b_bn = XFS_BUF_DADDR_NULL;
278         atomic_set(&bp->b_pin_count, 0);
279         init_waitqueue_head(&bp->b_waiters);
280
281         XFS_STATS_INC(xb_create);
282         XB_TRACE(bp, "initialize", target);
283 }
284
285 /*
286  *      Allocate a page array capable of holding a specified number
287  *      of pages, and point the page buf at it.
288  */
289 STATIC int
290 _xfs_buf_get_pages(
291         xfs_buf_t               *bp,
292         int                     page_count,
293         xfs_buf_flags_t         flags)
294 {
295         /* Make sure that we have a page list */
296         if (bp->b_pages == NULL) {
297                 bp->b_offset = xfs_buf_poff(bp->b_file_offset);
298                 bp->b_page_count = page_count;
299                 if (page_count <= XB_PAGES) {
300                         bp->b_pages = bp->b_page_array;
301                 } else {
302                         bp->b_pages = kmem_alloc(sizeof(struct page *) *
303                                         page_count, xb_to_km(flags));
304                         if (bp->b_pages == NULL)
305                                 return -ENOMEM;
306                 }
307                 memset(bp->b_pages, 0, sizeof(struct page *) * page_count);
308         }
309         return 0;
310 }
311
312 /*
313  *      Frees b_pages if it was allocated.
314  */
315 STATIC void
316 _xfs_buf_free_pages(
317         xfs_buf_t       *bp)
318 {
319         if (bp->b_pages != bp->b_page_array) {
320                 kmem_free(bp->b_pages);
321         }
322 }
323
324 /*
325  *      Releases the specified buffer.
326  *
327  *      The modification state of any associated pages is left unchanged.
328  *      The buffer most not be on any hash - use xfs_buf_rele instead for
329  *      hashed and refcounted buffers
330  */
331 void
332 xfs_buf_free(
333         xfs_buf_t               *bp)
334 {
335         XB_TRACE(bp, "free", 0);
336
337         ASSERT(list_empty(&bp->b_hash_list));
338
339         if (bp->b_flags & (_XBF_PAGE_CACHE|_XBF_PAGES)) {
340                 uint            i;
341
342                 if ((bp->b_flags & XBF_MAPPED) && (bp->b_page_count > 1))
343                         free_address(bp->b_addr - bp->b_offset);
344
345                 for (i = 0; i < bp->b_page_count; i++) {
346                         struct page     *page = bp->b_pages[i];
347
348                         if (bp->b_flags & _XBF_PAGE_CACHE)
349                                 ASSERT(!PagePrivate(page));
350                         page_cache_release(page);
351                 }
352                 _xfs_buf_free_pages(bp);
353         }
354
355         xfs_buf_deallocate(bp);
356 }
357
358 /*
359  *      Finds all pages for buffer in question and builds it's page list.
360  */
361 STATIC int
362 _xfs_buf_lookup_pages(
363         xfs_buf_t               *bp,
364         uint                    flags)
365 {
366         struct address_space    *mapping = bp->b_target->bt_mapping;
367         size_t                  blocksize = bp->b_target->bt_bsize;
368         size_t                  size = bp->b_count_desired;
369         size_t                  nbytes, offset;
370         gfp_t                   gfp_mask = xb_to_gfp(flags);
371         unsigned short          page_count, i;
372         pgoff_t                 first;
373         xfs_off_t               end;
374         int                     error;
375
376         end = bp->b_file_offset + bp->b_buffer_length;
377         page_count = xfs_buf_btoc(end) - xfs_buf_btoct(bp->b_file_offset);
378
379         error = _xfs_buf_get_pages(bp, page_count, flags);
380         if (unlikely(error))
381                 return error;
382         bp->b_flags |= _XBF_PAGE_CACHE;
383
384         offset = bp->b_offset;
385         first = bp->b_file_offset >> PAGE_CACHE_SHIFT;
386
387         for (i = 0; i < bp->b_page_count; i++) {
388                 struct page     *page;
389                 uint            retries = 0;
390
391               retry:
392                 page = find_or_create_page(mapping, first + i, gfp_mask);
393                 if (unlikely(page == NULL)) {
394                         if (flags & XBF_READ_AHEAD) {
395                                 bp->b_page_count = i;
396                                 for (i = 0; i < bp->b_page_count; i++)
397                                         unlock_page(bp->b_pages[i]);
398                                 return -ENOMEM;
399                         }
400
401                         /*
402                          * This could deadlock.
403                          *
404                          * But until all the XFS lowlevel code is revamped to
405                          * handle buffer allocation failures we can't do much.
406                          */
407                         if (!(++retries % 100))
408                                 printk(KERN_ERR
409                                         "XFS: possible memory allocation "
410                                         "deadlock in %s (mode:0x%x)\n",
411                                         __func__, gfp_mask);
412
413                         XFS_STATS_INC(xb_page_retries);
414                         xfsbufd_wakeup(0, gfp_mask);
415                         congestion_wait(WRITE, HZ/50);
416                         goto retry;
417                 }
418
419                 XFS_STATS_INC(xb_page_found);
420
421                 nbytes = min_t(size_t, size, PAGE_CACHE_SIZE - offset);
422                 size -= nbytes;
423
424                 ASSERT(!PagePrivate(page));
425                 if (!PageUptodate(page)) {
426                         page_count--;
427                         if (blocksize >= PAGE_CACHE_SIZE) {
428                                 if (flags & XBF_READ)
429                                         bp->b_flags |= _XBF_PAGE_LOCKED;
430                         } else if (!PagePrivate(page)) {
431                                 if (test_page_region(page, offset, nbytes))
432                                         page_count++;
433                         }
434                 }
435
436                 bp->b_pages[i] = page;
437                 offset = 0;
438         }
439
440         if (!(bp->b_flags & _XBF_PAGE_LOCKED)) {
441                 for (i = 0; i < bp->b_page_count; i++)
442                         unlock_page(bp->b_pages[i]);
443         }
444
445         if (page_count == bp->b_page_count)
446                 bp->b_flags |= XBF_DONE;
447
448         XB_TRACE(bp, "lookup_pages", (long)page_count);
449         return error;
450 }
451
452 /*
453  *      Map buffer into kernel address-space if nessecary.
454  */
455 STATIC int
456 _xfs_buf_map_pages(
457         xfs_buf_t               *bp,
458         uint                    flags)
459 {
460         /* A single page buffer is always mappable */
461         if (bp->b_page_count == 1) {
462                 bp->b_addr = page_address(bp->b_pages[0]) + bp->b_offset;
463                 bp->b_flags |= XBF_MAPPED;
464         } else if (flags & XBF_MAPPED) {
465                 if (as_list_len > 64)
466                         purge_addresses();
467                 bp->b_addr = vmap(bp->b_pages, bp->b_page_count,
468                                         VM_MAP, PAGE_KERNEL);
469                 if (unlikely(bp->b_addr == NULL))
470                         return -ENOMEM;
471                 bp->b_addr += bp->b_offset;
472                 bp->b_flags |= XBF_MAPPED;
473         }
474
475         return 0;
476 }
477
478 /*
479  *      Finding and Reading Buffers
480  */
481
482 /*
483  *      Look up, and creates if absent, a lockable buffer for
484  *      a given range of an inode.  The buffer is returned
485  *      locked.  If other overlapping buffers exist, they are
486  *      released before the new buffer is created and locked,
487  *      which may imply that this call will block until those buffers
488  *      are unlocked.  No I/O is implied by this call.
489  */
490 xfs_buf_t *
491 _xfs_buf_find(
492         xfs_buftarg_t           *btp,   /* block device target          */
493         xfs_off_t               ioff,   /* starting offset of range     */
494         size_t                  isize,  /* length of range              */
495         xfs_buf_flags_t         flags,
496         xfs_buf_t               *new_bp)
497 {
498         xfs_off_t               range_base;
499         size_t                  range_length;
500         xfs_bufhash_t           *hash;
501         xfs_buf_t               *bp, *n;
502
503         range_base = (ioff << BBSHIFT);
504         range_length = (isize << BBSHIFT);
505
506         /* Check for IOs smaller than the sector size / not sector aligned */
507         ASSERT(!(range_length < (1 << btp->bt_sshift)));
508         ASSERT(!(range_base & (xfs_off_t)btp->bt_smask));
509
510         hash = &btp->bt_hash[hash_long((unsigned long)ioff, btp->bt_hashshift)];
511
512         spin_lock(&hash->bh_lock);
513
514         list_for_each_entry_safe(bp, n, &hash->bh_list, b_hash_list) {
515                 ASSERT(btp == bp->b_target);
516                 if (bp->b_file_offset == range_base &&
517                     bp->b_buffer_length == range_length) {
518                         /*
519                          * If we look at something, bring it to the
520                          * front of the list for next time.
521                          */
522                         atomic_inc(&bp->b_hold);
523                         list_move(&bp->b_hash_list, &hash->bh_list);
524                         goto found;
525                 }
526         }
527
528         /* No match found */
529         if (new_bp) {
530                 _xfs_buf_initialize(new_bp, btp, range_base,
531                                 range_length, flags);
532                 new_bp->b_hash = hash;
533                 list_add(&new_bp->b_hash_list, &hash->bh_list);
534         } else {
535                 XFS_STATS_INC(xb_miss_locked);
536         }
537
538         spin_unlock(&hash->bh_lock);
539         return new_bp;
540
541 found:
542         spin_unlock(&hash->bh_lock);
543
544         /* Attempt to get the semaphore without sleeping,
545          * if this does not work then we need to drop the
546          * spinlock and do a hard attempt on the semaphore.
547          */
548         if (down_trylock(&bp->b_sema)) {
549                 if (!(flags & XBF_TRYLOCK)) {
550                         /* wait for buffer ownership */
551                         XB_TRACE(bp, "get_lock", 0);
552                         xfs_buf_lock(bp);
553                         XFS_STATS_INC(xb_get_locked_waited);
554                 } else {
555                         /* We asked for a trylock and failed, no need
556                          * to look at file offset and length here, we
557                          * know that this buffer at least overlaps our
558                          * buffer and is locked, therefore our buffer
559                          * either does not exist, or is this buffer.
560                          */
561                         xfs_buf_rele(bp);
562                         XFS_STATS_INC(xb_busy_locked);
563                         return NULL;
564                 }
565         } else {
566                 /* trylock worked */
567                 XB_SET_OWNER(bp);
568         }
569
570         if (bp->b_flags & XBF_STALE) {
571                 ASSERT((bp->b_flags & _XBF_DELWRI_Q) == 0);
572                 bp->b_flags &= XBF_MAPPED;
573         }
574         XB_TRACE(bp, "got_lock", 0);
575         XFS_STATS_INC(xb_get_locked);
576         return bp;
577 }
578
579 /*
580  *      Assembles a buffer covering the specified range.
581  *      Storage in memory for all portions of the buffer will be allocated,
582  *      although backing storage may not be.
583  */
584 xfs_buf_t *
585 xfs_buf_get_flags(
586         xfs_buftarg_t           *target,/* target for buffer            */
587         xfs_off_t               ioff,   /* starting offset of range     */
588         size_t                  isize,  /* length of range              */
589         xfs_buf_flags_t         flags)
590 {
591         xfs_buf_t               *bp, *new_bp;
592         int                     error = 0, i;
593
594         new_bp = xfs_buf_allocate(flags);
595         if (unlikely(!new_bp))
596                 return NULL;
597
598         bp = _xfs_buf_find(target, ioff, isize, flags, new_bp);
599         if (bp == new_bp) {
600                 error = _xfs_buf_lookup_pages(bp, flags);
601                 if (error)
602                         goto no_buffer;
603         } else {
604                 xfs_buf_deallocate(new_bp);
605                 if (unlikely(bp == NULL))
606                         return NULL;
607         }
608
609         for (i = 0; i < bp->b_page_count; i++)
610                 mark_page_accessed(bp->b_pages[i]);
611
612         if (!(bp->b_flags & XBF_MAPPED)) {
613                 error = _xfs_buf_map_pages(bp, flags);
614                 if (unlikely(error)) {
615                         printk(KERN_WARNING "%s: failed to map pages\n",
616                                         __func__);
617                         goto no_buffer;
618                 }
619         }
620
621         XFS_STATS_INC(xb_get);
622
623         /*
624          * Always fill in the block number now, the mapped cases can do
625          * their own overlay of this later.
626          */
627         bp->b_bn = ioff;
628         bp->b_count_desired = bp->b_buffer_length;
629
630         XB_TRACE(bp, "get", (unsigned long)flags);
631         return bp;
632
633  no_buffer:
634         if (flags & (XBF_LOCK | XBF_TRYLOCK))
635                 xfs_buf_unlock(bp);
636         xfs_buf_rele(bp);
637         return NULL;
638 }
639
640 STATIC int
641 _xfs_buf_read(
642         xfs_buf_t               *bp,
643         xfs_buf_flags_t         flags)
644 {
645         int                     status;
646
647         XB_TRACE(bp, "_xfs_buf_read", (unsigned long)flags);
648
649         ASSERT(!(flags & (XBF_DELWRI|XBF_WRITE)));
650         ASSERT(bp->b_bn != XFS_BUF_DADDR_NULL);
651
652         bp->b_flags &= ~(XBF_WRITE | XBF_ASYNC | XBF_DELWRI | \
653                         XBF_READ_AHEAD | _XBF_RUN_QUEUES);
654         bp->b_flags |= flags & (XBF_READ | XBF_ASYNC | \
655                         XBF_READ_AHEAD | _XBF_RUN_QUEUES);
656
657         status = xfs_buf_iorequest(bp);
658         if (!status && !(flags & XBF_ASYNC))
659                 status = xfs_buf_iowait(bp);
660         return status;
661 }
662
663 xfs_buf_t *
664 xfs_buf_read_flags(
665         xfs_buftarg_t           *target,
666         xfs_off_t               ioff,
667         size_t                  isize,
668         xfs_buf_flags_t         flags)
669 {
670         xfs_buf_t               *bp;
671
672         flags |= XBF_READ;
673
674         bp = xfs_buf_get_flags(target, ioff, isize, flags);
675         if (bp) {
676                 if (!XFS_BUF_ISDONE(bp)) {
677                         XB_TRACE(bp, "read", (unsigned long)flags);
678                         XFS_STATS_INC(xb_get_read);
679                         _xfs_buf_read(bp, flags);
680                 } else if (flags & XBF_ASYNC) {
681                         XB_TRACE(bp, "read_async", (unsigned long)flags);
682                         /*
683                          * Read ahead call which is already satisfied,
684                          * drop the buffer
685                          */
686                         goto no_buffer;
687                 } else {
688                         XB_TRACE(bp, "read_done", (unsigned long)flags);
689                         /* We do not want read in the flags */
690                         bp->b_flags &= ~XBF_READ;
691                 }
692         }
693
694         return bp;
695
696  no_buffer:
697         if (flags & (XBF_LOCK | XBF_TRYLOCK))
698                 xfs_buf_unlock(bp);
699         xfs_buf_rele(bp);
700         return NULL;
701 }
702
703 /*
704  *      If we are not low on memory then do the readahead in a deadlock
705  *      safe manner.
706  */
707 void
708 xfs_buf_readahead(
709         xfs_buftarg_t           *target,
710         xfs_off_t               ioff,
711         size_t                  isize,
712         xfs_buf_flags_t         flags)
713 {
714         struct backing_dev_info *bdi;
715
716         bdi = target->bt_mapping->backing_dev_info;
717         if (bdi_read_congested(bdi))
718                 return;
719
720         flags |= (XBF_TRYLOCK|XBF_ASYNC|XBF_READ_AHEAD);
721         xfs_buf_read_flags(target, ioff, isize, flags);
722 }
723
724 xfs_buf_t *
725 xfs_buf_get_empty(
726         size_t                  len,
727         xfs_buftarg_t           *target)
728 {
729         xfs_buf_t               *bp;
730
731         bp = xfs_buf_allocate(0);
732         if (bp)
733                 _xfs_buf_initialize(bp, target, 0, len, 0);
734         return bp;
735 }
736
737 static inline struct page *
738 mem_to_page(
739         void                    *addr)
740 {
741         if ((!is_vmalloc_addr(addr))) {
742                 return virt_to_page(addr);
743         } else {
744                 return vmalloc_to_page(addr);
745         }
746 }
747
748 int
749 xfs_buf_associate_memory(
750         xfs_buf_t               *bp,
751         void                    *mem,
752         size_t                  len)
753 {
754         int                     rval;
755         int                     i = 0;
756         unsigned long           pageaddr;
757         unsigned long           offset;
758         size_t                  buflen;
759         int                     page_count;
760
761         pageaddr = (unsigned long)mem & PAGE_CACHE_MASK;
762         offset = (unsigned long)mem - pageaddr;
763         buflen = PAGE_CACHE_ALIGN(len + offset);
764         page_count = buflen >> PAGE_CACHE_SHIFT;
765
766         /* Free any previous set of page pointers */
767         if (bp->b_pages)
768                 _xfs_buf_free_pages(bp);
769
770         bp->b_pages = NULL;
771         bp->b_addr = mem;
772
773         rval = _xfs_buf_get_pages(bp, page_count, 0);
774         if (rval)
775                 return rval;
776
777         bp->b_offset = offset;
778
779         for (i = 0; i < bp->b_page_count; i++) {
780                 bp->b_pages[i] = mem_to_page((void *)pageaddr);
781                 pageaddr += PAGE_CACHE_SIZE;
782         }
783
784         bp->b_count_desired = len;
785         bp->b_buffer_length = buflen;
786         bp->b_flags |= XBF_MAPPED;
787         bp->b_flags &= ~_XBF_PAGE_LOCKED;
788
789         return 0;
790 }
791
792 xfs_buf_t *
793 xfs_buf_get_noaddr(
794         size_t                  len,
795         xfs_buftarg_t           *target)
796 {
797         unsigned long           page_count = PAGE_ALIGN(len) >> PAGE_SHIFT;
798         int                     error, i;
799         xfs_buf_t               *bp;
800
801         bp = xfs_buf_allocate(0);
802         if (unlikely(bp == NULL))
803                 goto fail;
804         _xfs_buf_initialize(bp, target, 0, len, 0);
805
806         error = _xfs_buf_get_pages(bp, page_count, 0);
807         if (error)
808                 goto fail_free_buf;
809
810         for (i = 0; i < page_count; i++) {
811                 bp->b_pages[i] = alloc_page(GFP_KERNEL);
812                 if (!bp->b_pages[i])
813                         goto fail_free_mem;
814         }
815         bp->b_flags |= _XBF_PAGES;
816
817         error = _xfs_buf_map_pages(bp, XBF_MAPPED);
818         if (unlikely(error)) {
819                 printk(KERN_WARNING "%s: failed to map pages\n",
820                                 __func__);
821                 goto fail_free_mem;
822         }
823
824         xfs_buf_unlock(bp);
825
826         XB_TRACE(bp, "no_daddr", len);
827         return bp;
828
829  fail_free_mem:
830         while (--i >= 0)
831                 __free_page(bp->b_pages[i]);
832         _xfs_buf_free_pages(bp);
833  fail_free_buf:
834         xfs_buf_deallocate(bp);
835  fail:
836         return NULL;
837 }
838
839 /*
840  *      Increment reference count on buffer, to hold the buffer concurrently
841  *      with another thread which may release (free) the buffer asynchronously.
842  *      Must hold the buffer already to call this function.
843  */
844 void
845 xfs_buf_hold(
846         xfs_buf_t               *bp)
847 {
848         atomic_inc(&bp->b_hold);
849         XB_TRACE(bp, "hold", 0);
850 }
851
852 /*
853  *      Releases a hold on the specified buffer.  If the
854  *      the hold count is 1, calls xfs_buf_free.
855  */
856 void
857 xfs_buf_rele(
858         xfs_buf_t               *bp)
859 {
860         xfs_bufhash_t           *hash = bp->b_hash;
861
862         XB_TRACE(bp, "rele", bp->b_relse);
863
864         if (unlikely(!hash)) {
865                 ASSERT(!bp->b_relse);
866                 if (atomic_dec_and_test(&bp->b_hold))
867                         xfs_buf_free(bp);
868                 return;
869         }
870
871         ASSERT(atomic_read(&bp->b_hold) > 0);
872         if (atomic_dec_and_lock(&bp->b_hold, &hash->bh_lock)) {
873                 if (bp->b_relse) {
874                         atomic_inc(&bp->b_hold);
875                         spin_unlock(&hash->bh_lock);
876                         (*(bp->b_relse)) (bp);
877                 } else if (bp->b_flags & XBF_FS_MANAGED) {
878                         spin_unlock(&hash->bh_lock);
879                 } else {
880                         ASSERT(!(bp->b_flags & (XBF_DELWRI|_XBF_DELWRI_Q)));
881                         list_del_init(&bp->b_hash_list);
882                         spin_unlock(&hash->bh_lock);
883                         xfs_buf_free(bp);
884                 }
885         }
886 }
887
888
889 /*
890  *      Mutual exclusion on buffers.  Locking model:
891  *
892  *      Buffers associated with inodes for which buffer locking
893  *      is not enabled are not protected by semaphores, and are
894  *      assumed to be exclusively owned by the caller.  There is a
895  *      spinlock in the buffer, used by the caller when concurrent
896  *      access is possible.
897  */
898
899 /*
900  *      Locks a buffer object, if it is not already locked.
901  *      Note that this in no way locks the underlying pages, so it is only
902  *      useful for synchronizing concurrent use of buffer objects, not for
903  *      synchronizing independent access to the underlying pages.
904  */
905 int
906 xfs_buf_cond_lock(
907         xfs_buf_t               *bp)
908 {
909         int                     locked;
910
911         locked = down_trylock(&bp->b_sema) == 0;
912         if (locked) {
913                 XB_SET_OWNER(bp);
914         }
915         XB_TRACE(bp, "cond_lock", (long)locked);
916         return locked ? 0 : -EBUSY;
917 }
918
919 #if defined(DEBUG) || defined(XFS_BLI_TRACE)
920 int
921 xfs_buf_lock_value(
922         xfs_buf_t               *bp)
923 {
924         return bp->b_sema.count;
925 }
926 #endif
927
928 /*
929  *      Locks a buffer object.
930  *      Note that this in no way locks the underlying pages, so it is only
931  *      useful for synchronizing concurrent use of buffer objects, not for
932  *      synchronizing independent access to the underlying pages.
933  */
934 void
935 xfs_buf_lock(
936         xfs_buf_t               *bp)
937 {
938         XB_TRACE(bp, "lock", 0);
939         if (atomic_read(&bp->b_io_remaining))
940                 blk_run_address_space(bp->b_target->bt_mapping);
941         down(&bp->b_sema);
942         XB_SET_OWNER(bp);
943         XB_TRACE(bp, "locked", 0);
944 }
945
946 /*
947  *      Releases the lock on the buffer object.
948  *      If the buffer is marked delwri but is not queued, do so before we
949  *      unlock the buffer as we need to set flags correctly.  We also need to
950  *      take a reference for the delwri queue because the unlocker is going to
951  *      drop their's and they don't know we just queued it.
952  */
953 void
954 xfs_buf_unlock(
955         xfs_buf_t               *bp)
956 {
957         if ((bp->b_flags & (XBF_DELWRI|_XBF_DELWRI_Q)) == XBF_DELWRI) {
958                 atomic_inc(&bp->b_hold);
959                 bp->b_flags |= XBF_ASYNC;
960                 xfs_buf_delwri_queue(bp, 0);
961         }
962
963         XB_CLEAR_OWNER(bp);
964         up(&bp->b_sema);
965         XB_TRACE(bp, "unlock", 0);
966 }
967
968
969 /*
970  *      Pinning Buffer Storage in Memory
971  *      Ensure that no attempt to force a buffer to disk will succeed.
972  */
973 void
974 xfs_buf_pin(
975         xfs_buf_t               *bp)
976 {
977         atomic_inc(&bp->b_pin_count);
978         XB_TRACE(bp, "pin", (long)bp->b_pin_count.counter);
979 }
980
981 void
982 xfs_buf_unpin(
983         xfs_buf_t               *bp)
984 {
985         if (atomic_dec_and_test(&bp->b_pin_count))
986                 wake_up_all(&bp->b_waiters);
987         XB_TRACE(bp, "unpin", (long)bp->b_pin_count.counter);
988 }
989
990 int
991 xfs_buf_ispin(
992         xfs_buf_t               *bp)
993 {
994         return atomic_read(&bp->b_pin_count);
995 }
996
997 STATIC void
998 xfs_buf_wait_unpin(
999         xfs_buf_t               *bp)
1000 {
1001         DECLARE_WAITQUEUE       (wait, current);
1002
1003         if (atomic_read(&bp->b_pin_count) == 0)
1004                 return;
1005
1006         add_wait_queue(&bp->b_waiters, &wait);
1007         for (;;) {
1008                 set_current_state(TASK_UNINTERRUPTIBLE);
1009                 if (atomic_read(&bp->b_pin_count) == 0)
1010                         break;
1011                 if (atomic_read(&bp->b_io_remaining))
1012                         blk_run_address_space(bp->b_target->bt_mapping);
1013                 schedule();
1014         }
1015         remove_wait_queue(&bp->b_waiters, &wait);
1016         set_current_state(TASK_RUNNING);
1017 }
1018
1019 /*
1020  *      Buffer Utility Routines
1021  */
1022
1023 STATIC void
1024 xfs_buf_iodone_work(
1025         struct work_struct      *work)
1026 {
1027         xfs_buf_t               *bp =
1028                 container_of(work, xfs_buf_t, b_iodone_work);
1029
1030         /*
1031          * We can get an EOPNOTSUPP to ordered writes.  Here we clear the
1032          * ordered flag and reissue them.  Because we can't tell the higher
1033          * layers directly that they should not issue ordered I/O anymore, they
1034          * need to check if the _XFS_BARRIER_FAILED flag was set during I/O completion.
1035          */
1036         if ((bp->b_error == EOPNOTSUPP) &&
1037             (bp->b_flags & (XBF_ORDERED|XBF_ASYNC)) == (XBF_ORDERED|XBF_ASYNC)) {
1038                 XB_TRACE(bp, "ordered_retry", bp->b_iodone);
1039                 bp->b_flags &= ~XBF_ORDERED;
1040                 bp->b_flags |= _XFS_BARRIER_FAILED;
1041                 xfs_buf_iorequest(bp);
1042         } else if (bp->b_iodone)
1043                 (*(bp->b_iodone))(bp);
1044         else if (bp->b_flags & XBF_ASYNC)
1045                 xfs_buf_relse(bp);
1046 }
1047
1048 void
1049 xfs_buf_ioend(
1050         xfs_buf_t               *bp,
1051         int                     schedule)
1052 {
1053         bp->b_flags &= ~(XBF_READ | XBF_WRITE | XBF_READ_AHEAD);
1054         if (bp->b_error == 0)
1055                 bp->b_flags |= XBF_DONE;
1056
1057         XB_TRACE(bp, "iodone", bp->b_iodone);
1058
1059         if ((bp->b_iodone) || (bp->b_flags & XBF_ASYNC)) {
1060                 if (schedule) {
1061                         INIT_WORK(&bp->b_iodone_work, xfs_buf_iodone_work);
1062                         queue_work(xfslogd_workqueue, &bp->b_iodone_work);
1063                 } else {
1064                         xfs_buf_iodone_work(&bp->b_iodone_work);
1065                 }
1066         } else {
1067                 complete(&bp->b_iowait);
1068         }
1069 }
1070
1071 void
1072 xfs_buf_ioerror(
1073         xfs_buf_t               *bp,
1074         int                     error)
1075 {
1076         ASSERT(error >= 0 && error <= 0xffff);
1077         bp->b_error = (unsigned short)error;
1078         XB_TRACE(bp, "ioerror", (unsigned long)error);
1079 }
1080
1081 int
1082 xfs_bawrite(
1083         void                    *mp,
1084         struct xfs_buf          *bp)
1085 {
1086         XB_TRACE(bp, "bawrite", 0);
1087
1088         ASSERT(bp->b_bn != XFS_BUF_DADDR_NULL);
1089
1090         xfs_buf_delwri_dequeue(bp);
1091
1092         bp->b_flags &= ~(XBF_READ | XBF_DELWRI | XBF_READ_AHEAD);
1093         bp->b_flags |= (XBF_WRITE | XBF_ASYNC | _XBF_RUN_QUEUES);
1094
1095         bp->b_mount = mp;
1096         bp->b_strat = xfs_bdstrat_cb;
1097         return xfs_bdstrat_cb(bp);
1098 }
1099
1100 void
1101 xfs_bdwrite(
1102         void                    *mp,
1103         struct xfs_buf          *bp)
1104 {
1105         XB_TRACE(bp, "bdwrite", 0);
1106
1107         bp->b_strat = xfs_bdstrat_cb;
1108         bp->b_mount = mp;
1109
1110         bp->b_flags &= ~XBF_READ;
1111         bp->b_flags |= (XBF_DELWRI | XBF_ASYNC);
1112
1113         xfs_buf_delwri_queue(bp, 1);
1114 }
1115
1116 STATIC_INLINE void
1117 _xfs_buf_ioend(
1118         xfs_buf_t               *bp,
1119         int                     schedule)
1120 {
1121         if (atomic_dec_and_test(&bp->b_io_remaining) == 1) {
1122                 bp->b_flags &= ~_XBF_PAGE_LOCKED;
1123                 xfs_buf_ioend(bp, schedule);
1124         }
1125 }
1126
1127 STATIC void
1128 xfs_buf_bio_end_io(
1129         struct bio              *bio,
1130         int                     error)
1131 {
1132         xfs_buf_t               *bp = (xfs_buf_t *)bio->bi_private;
1133         unsigned int            blocksize = bp->b_target->bt_bsize;
1134         struct bio_vec          *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
1135
1136         xfs_buf_ioerror(bp, -error);
1137
1138         do {
1139                 struct page     *page = bvec->bv_page;
1140
1141                 ASSERT(!PagePrivate(page));
1142                 if (unlikely(bp->b_error)) {
1143                         if (bp->b_flags & XBF_READ)
1144                                 ClearPageUptodate(page);
1145                 } else if (blocksize >= PAGE_CACHE_SIZE) {
1146                         SetPageUptodate(page);
1147                 } else if (!PagePrivate(page) &&
1148                                 (bp->b_flags & _XBF_PAGE_CACHE)) {
1149                         set_page_region(page, bvec->bv_offset, bvec->bv_len);
1150                 }
1151
1152                 if (--bvec >= bio->bi_io_vec)
1153                         prefetchw(&bvec->bv_page->flags);
1154
1155                 if (bp->b_flags & _XBF_PAGE_LOCKED)
1156                         unlock_page(page);
1157         } while (bvec >= bio->bi_io_vec);
1158
1159         _xfs_buf_ioend(bp, 1);
1160         bio_put(bio);
1161 }
1162
1163 STATIC void
1164 _xfs_buf_ioapply(
1165         xfs_buf_t               *bp)
1166 {
1167         int                     rw, map_i, total_nr_pages, nr_pages;
1168         struct bio              *bio;
1169         int                     offset = bp->b_offset;
1170         int                     size = bp->b_count_desired;
1171         sector_t                sector = bp->b_bn;
1172         unsigned int            blocksize = bp->b_target->bt_bsize;
1173
1174         total_nr_pages = bp->b_page_count;
1175         map_i = 0;
1176
1177         if (bp->b_flags & XBF_ORDERED) {
1178                 ASSERT(!(bp->b_flags & XBF_READ));
1179                 rw = WRITE_BARRIER;
1180         } else if (bp->b_flags & _XBF_RUN_QUEUES) {
1181                 ASSERT(!(bp->b_flags & XBF_READ_AHEAD));
1182                 bp->b_flags &= ~_XBF_RUN_QUEUES;
1183                 rw = (bp->b_flags & XBF_WRITE) ? WRITE_SYNC : READ_SYNC;
1184         } else {
1185                 rw = (bp->b_flags & XBF_WRITE) ? WRITE :
1186                      (bp->b_flags & XBF_READ_AHEAD) ? READA : READ;
1187         }
1188
1189         /* Special code path for reading a sub page size buffer in --
1190          * we populate up the whole page, and hence the other metadata
1191          * in the same page.  This optimization is only valid when the
1192          * filesystem block size is not smaller than the page size.
1193          */
1194         if ((bp->b_buffer_length < PAGE_CACHE_SIZE) &&
1195             ((bp->b_flags & (XBF_READ|_XBF_PAGE_LOCKED)) ==
1196               (XBF_READ|_XBF_PAGE_LOCKED)) &&
1197             (blocksize >= PAGE_CACHE_SIZE)) {
1198                 bio = bio_alloc(GFP_NOIO, 1);
1199
1200                 bio->bi_bdev = bp->b_target->bt_bdev;
1201                 bio->bi_sector = sector - (offset >> BBSHIFT);
1202                 bio->bi_end_io = xfs_buf_bio_end_io;
1203                 bio->bi_private = bp;
1204
1205                 bio_add_page(bio, bp->b_pages[0], PAGE_CACHE_SIZE, 0);
1206                 size = 0;
1207
1208                 atomic_inc(&bp->b_io_remaining);
1209
1210                 goto submit_io;
1211         }
1212
1213 next_chunk:
1214         atomic_inc(&bp->b_io_remaining);
1215         nr_pages = BIO_MAX_SECTORS >> (PAGE_SHIFT - BBSHIFT);
1216         if (nr_pages > total_nr_pages)
1217                 nr_pages = total_nr_pages;
1218
1219         bio = bio_alloc(GFP_NOIO, nr_pages);
1220         bio->bi_bdev = bp->b_target->bt_bdev;
1221         bio->bi_sector = sector;
1222         bio->bi_end_io = xfs_buf_bio_end_io;
1223         bio->bi_private = bp;
1224
1225         for (; size && nr_pages; nr_pages--, map_i++) {
1226                 int     rbytes, nbytes = PAGE_CACHE_SIZE - offset;
1227
1228                 if (nbytes > size)
1229                         nbytes = size;
1230
1231                 rbytes = bio_add_page(bio, bp->b_pages[map_i], nbytes, offset);
1232                 if (rbytes < nbytes)
1233                         break;
1234
1235                 offset = 0;
1236                 sector += nbytes >> BBSHIFT;
1237                 size -= nbytes;
1238                 total_nr_pages--;
1239         }
1240
1241 submit_io:
1242         if (likely(bio->bi_size)) {
1243                 submit_bio(rw, bio);
1244                 if (size)
1245                         goto next_chunk;
1246         } else {
1247                 bio_put(bio);
1248                 xfs_buf_ioerror(bp, EIO);
1249         }
1250 }
1251
1252 int
1253 xfs_buf_iorequest(
1254         xfs_buf_t               *bp)
1255 {
1256         XB_TRACE(bp, "iorequest", 0);
1257
1258         if (bp->b_flags & XBF_DELWRI) {
1259                 xfs_buf_delwri_queue(bp, 1);
1260                 return 0;
1261         }
1262
1263         if (bp->b_flags & XBF_WRITE) {
1264                 xfs_buf_wait_unpin(bp);
1265         }
1266
1267         xfs_buf_hold(bp);
1268
1269         /* Set the count to 1 initially, this will stop an I/O
1270          * completion callout which happens before we have started
1271          * all the I/O from calling xfs_buf_ioend too early.
1272          */
1273         atomic_set(&bp->b_io_remaining, 1);
1274         _xfs_buf_ioapply(bp);
1275         _xfs_buf_ioend(bp, 0);
1276
1277         xfs_buf_rele(bp);
1278         return 0;
1279 }
1280
1281 /*
1282  *      Waits for I/O to complete on the buffer supplied.
1283  *      It returns immediately if no I/O is pending.
1284  *      It returns the I/O error code, if any, or 0 if there was no error.
1285  */
1286 int
1287 xfs_buf_iowait(
1288         xfs_buf_t               *bp)
1289 {
1290         XB_TRACE(bp, "iowait", 0);
1291         if (atomic_read(&bp->b_io_remaining))
1292                 blk_run_address_space(bp->b_target->bt_mapping);
1293         wait_for_completion(&bp->b_iowait);
1294         XB_TRACE(bp, "iowaited", (long)bp->b_error);
1295         return bp->b_error;
1296 }
1297
1298 xfs_caddr_t
1299 xfs_buf_offset(
1300         xfs_buf_t               *bp,
1301         size_t                  offset)
1302 {
1303         struct page             *page;
1304
1305         if (bp->b_flags & XBF_MAPPED)
1306                 return XFS_BUF_PTR(bp) + offset;
1307
1308         offset += bp->b_offset;
1309         page = bp->b_pages[offset >> PAGE_CACHE_SHIFT];
1310         return (xfs_caddr_t)page_address(page) + (offset & (PAGE_CACHE_SIZE-1));
1311 }
1312
1313 /*
1314  *      Move data into or out of a buffer.
1315  */
1316 void
1317 xfs_buf_iomove(
1318         xfs_buf_t               *bp,    /* buffer to process            */
1319         size_t                  boff,   /* starting buffer offset       */
1320         size_t                  bsize,  /* length to copy               */
1321         caddr_t                 data,   /* data address                 */
1322         xfs_buf_rw_t            mode)   /* read/write/zero flag         */
1323 {
1324         size_t                  bend, cpoff, csize;
1325         struct page             *page;
1326
1327         bend = boff + bsize;
1328         while (boff < bend) {
1329                 page = bp->b_pages[xfs_buf_btoct(boff + bp->b_offset)];
1330                 cpoff = xfs_buf_poff(boff + bp->b_offset);
1331                 csize = min_t(size_t,
1332                               PAGE_CACHE_SIZE-cpoff, bp->b_count_desired-boff);
1333
1334                 ASSERT(((csize + cpoff) <= PAGE_CACHE_SIZE));
1335
1336                 switch (mode) {
1337                 case XBRW_ZERO:
1338                         memset(page_address(page) + cpoff, 0, csize);
1339                         break;
1340                 case XBRW_READ:
1341                         memcpy(data, page_address(page) + cpoff, csize);
1342                         break;
1343                 case XBRW_WRITE:
1344                         memcpy(page_address(page) + cpoff, data, csize);
1345                 }
1346
1347                 boff += csize;
1348                 data += csize;
1349         }
1350 }
1351
1352 /*
1353  *      Handling of buffer targets (buftargs).
1354  */
1355
1356 /*
1357  *      Wait for any bufs with callbacks that have been submitted but
1358  *      have not yet returned... walk the hash list for the target.
1359  */
1360 void
1361 xfs_wait_buftarg(
1362         xfs_buftarg_t   *btp)
1363 {
1364         xfs_buf_t       *bp, *n;
1365         xfs_bufhash_t   *hash;
1366         uint            i;
1367
1368         for (i = 0; i < (1 << btp->bt_hashshift); i++) {
1369                 hash = &btp->bt_hash[i];
1370 again:
1371                 spin_lock(&hash->bh_lock);
1372                 list_for_each_entry_safe(bp, n, &hash->bh_list, b_hash_list) {
1373                         ASSERT(btp == bp->b_target);
1374                         if (!(bp->b_flags & XBF_FS_MANAGED)) {
1375                                 spin_unlock(&hash->bh_lock);
1376                                 /*
1377                                  * Catch superblock reference count leaks
1378                                  * immediately
1379                                  */
1380                                 BUG_ON(bp->b_bn == 0);
1381                                 delay(100);
1382                                 goto again;
1383                         }
1384                 }
1385                 spin_unlock(&hash->bh_lock);
1386         }
1387 }
1388
1389 /*
1390  *      Allocate buffer hash table for a given target.
1391  *      For devices containing metadata (i.e. not the log/realtime devices)
1392  *      we need to allocate a much larger hash table.
1393  */
1394 STATIC void
1395 xfs_alloc_bufhash(
1396         xfs_buftarg_t           *btp,
1397         int                     external)
1398 {
1399         unsigned int            i;
1400
1401         btp->bt_hashshift = external ? 3 : 8;   /* 8 or 256 buckets */
1402         btp->bt_hashmask = (1 << btp->bt_hashshift) - 1;
1403         btp->bt_hash = kmem_zalloc((1 << btp->bt_hashshift) *
1404                                         sizeof(xfs_bufhash_t), KM_SLEEP | KM_LARGE);
1405         for (i = 0; i < (1 << btp->bt_hashshift); i++) {
1406                 spin_lock_init(&btp->bt_hash[i].bh_lock);
1407                 INIT_LIST_HEAD(&btp->bt_hash[i].bh_list);
1408         }
1409 }
1410
1411 STATIC void
1412 xfs_free_bufhash(
1413         xfs_buftarg_t           *btp)
1414 {
1415         kmem_free(btp->bt_hash);
1416         btp->bt_hash = NULL;
1417 }
1418
1419 /*
1420  *      buftarg list for delwrite queue processing
1421  */
1422 static LIST_HEAD(xfs_buftarg_list);
1423 static DEFINE_SPINLOCK(xfs_buftarg_lock);
1424
1425 STATIC void
1426 xfs_register_buftarg(
1427         xfs_buftarg_t           *btp)
1428 {
1429         spin_lock(&xfs_buftarg_lock);
1430         list_add(&btp->bt_list, &xfs_buftarg_list);
1431         spin_unlock(&xfs_buftarg_lock);
1432 }
1433
1434 STATIC void
1435 xfs_unregister_buftarg(
1436         xfs_buftarg_t           *btp)
1437 {
1438         spin_lock(&xfs_buftarg_lock);
1439         list_del(&btp->bt_list);
1440         spin_unlock(&xfs_buftarg_lock);
1441 }
1442
1443 void
1444 xfs_free_buftarg(
1445         struct xfs_mount        *mp,
1446         struct xfs_buftarg      *btp)
1447 {
1448         xfs_flush_buftarg(btp, 1);
1449         if (mp->m_flags & XFS_MOUNT_BARRIER)
1450                 xfs_blkdev_issue_flush(btp);
1451         xfs_free_bufhash(btp);
1452         iput(btp->bt_mapping->host);
1453
1454         /* Unregister the buftarg first so that we don't get a
1455          * wakeup finding a non-existent task
1456          */
1457         xfs_unregister_buftarg(btp);
1458         kthread_stop(btp->bt_task);
1459
1460         kmem_free(btp);
1461 }
1462
1463 STATIC int
1464 xfs_setsize_buftarg_flags(
1465         xfs_buftarg_t           *btp,
1466         unsigned int            blocksize,
1467         unsigned int            sectorsize,
1468         int                     verbose)
1469 {
1470         btp->bt_bsize = blocksize;
1471         btp->bt_sshift = ffs(sectorsize) - 1;
1472         btp->bt_smask = sectorsize - 1;
1473
1474         if (set_blocksize(btp->bt_bdev, sectorsize)) {
1475                 printk(KERN_WARNING
1476                         "XFS: Cannot set_blocksize to %u on device %s\n",
1477                         sectorsize, XFS_BUFTARG_NAME(btp));
1478                 return EINVAL;
1479         }
1480
1481         if (verbose &&
1482             (PAGE_CACHE_SIZE / BITS_PER_LONG) > sectorsize) {
1483                 printk(KERN_WARNING
1484                         "XFS: %u byte sectors in use on device %s.  "
1485                         "This is suboptimal; %u or greater is ideal.\n",
1486                         sectorsize, XFS_BUFTARG_NAME(btp),
1487                         (unsigned int)PAGE_CACHE_SIZE / BITS_PER_LONG);
1488         }
1489
1490         return 0;
1491 }
1492
1493 /*
1494  *      When allocating the initial buffer target we have not yet
1495  *      read in the superblock, so don't know what sized sectors
1496  *      are being used is at this early stage.  Play safe.
1497  */
1498 STATIC int
1499 xfs_setsize_buftarg_early(
1500         xfs_buftarg_t           *btp,
1501         struct block_device     *bdev)
1502 {
1503         return xfs_setsize_buftarg_flags(btp,
1504                         PAGE_CACHE_SIZE, bdev_logical_block_size(bdev), 0);
1505 }
1506
1507 int
1508 xfs_setsize_buftarg(
1509         xfs_buftarg_t           *btp,
1510         unsigned int            blocksize,
1511         unsigned int            sectorsize)
1512 {
1513         return xfs_setsize_buftarg_flags(btp, blocksize, sectorsize, 1);
1514 }
1515
1516 STATIC int
1517 xfs_mapping_buftarg(
1518         xfs_buftarg_t           *btp,
1519         struct block_device     *bdev)
1520 {
1521         struct backing_dev_info *bdi;
1522         struct inode            *inode;
1523         struct address_space    *mapping;
1524         static const struct address_space_operations mapping_aops = {
1525                 .sync_page = block_sync_page,
1526                 .migratepage = fail_migrate_page,
1527         };
1528
1529         inode = new_inode(bdev->bd_inode->i_sb);
1530         if (!inode) {
1531                 printk(KERN_WARNING
1532                         "XFS: Cannot allocate mapping inode for device %s\n",
1533                         XFS_BUFTARG_NAME(btp));
1534                 return ENOMEM;
1535         }
1536         inode->i_mode = S_IFBLK;
1537         inode->i_bdev = bdev;
1538         inode->i_rdev = bdev->bd_dev;
1539         bdi = blk_get_backing_dev_info(bdev);
1540         if (!bdi)
1541                 bdi = &default_backing_dev_info;
1542         mapping = &inode->i_data;
1543         mapping->a_ops = &mapping_aops;
1544         mapping->backing_dev_info = bdi;
1545         mapping_set_gfp_mask(mapping, GFP_NOFS);
1546         btp->bt_mapping = mapping;
1547         return 0;
1548 }
1549
1550 STATIC int
1551 xfs_alloc_delwrite_queue(
1552         xfs_buftarg_t           *btp)
1553 {
1554         int     error = 0;
1555
1556         INIT_LIST_HEAD(&btp->bt_list);
1557         INIT_LIST_HEAD(&btp->bt_delwrite_queue);
1558         spin_lock_init(&btp->bt_delwrite_lock);
1559         btp->bt_flags = 0;
1560         btp->bt_task = kthread_run(xfsbufd, btp, "xfsbufd");
1561         if (IS_ERR(btp->bt_task)) {
1562                 error = PTR_ERR(btp->bt_task);
1563                 goto out_error;
1564         }
1565         xfs_register_buftarg(btp);
1566 out_error:
1567         return error;
1568 }
1569
1570 xfs_buftarg_t *
1571 xfs_alloc_buftarg(
1572         struct block_device     *bdev,
1573         int                     external)
1574 {
1575         xfs_buftarg_t           *btp;
1576
1577         btp = kmem_zalloc(sizeof(*btp), KM_SLEEP);
1578
1579         btp->bt_dev =  bdev->bd_dev;
1580         btp->bt_bdev = bdev;
1581         if (xfs_setsize_buftarg_early(btp, bdev))
1582                 goto error;
1583         if (xfs_mapping_buftarg(btp, bdev))
1584                 goto error;
1585         if (xfs_alloc_delwrite_queue(btp))
1586                 goto error;
1587         xfs_alloc_bufhash(btp, external);
1588         return btp;
1589
1590 error:
1591         kmem_free(btp);
1592         return NULL;
1593 }
1594
1595
1596 /*
1597  *      Delayed write buffer handling
1598  */
1599 STATIC void
1600 xfs_buf_delwri_queue(
1601         xfs_buf_t               *bp,
1602         int                     unlock)
1603 {
1604         struct list_head        *dwq = &bp->b_target->bt_delwrite_queue;
1605         spinlock_t              *dwlk = &bp->b_target->bt_delwrite_lock;
1606
1607         XB_TRACE(bp, "delwri_q", (long)unlock);
1608         ASSERT((bp->b_flags&(XBF_DELWRI|XBF_ASYNC)) == (XBF_DELWRI|XBF_ASYNC));
1609
1610         spin_lock(dwlk);
1611         /* If already in the queue, dequeue and place at tail */
1612         if (!list_empty(&bp->b_list)) {
1613                 ASSERT(bp->b_flags & _XBF_DELWRI_Q);
1614                 if (unlock)
1615                         atomic_dec(&bp->b_hold);
1616                 list_del(&bp->b_list);
1617         }
1618
1619         bp->b_flags |= _XBF_DELWRI_Q;
1620         list_add_tail(&bp->b_list, dwq);
1621         bp->b_queuetime = jiffies;
1622         spin_unlock(dwlk);
1623
1624         if (unlock)
1625                 xfs_buf_unlock(bp);
1626 }
1627
1628 void
1629 xfs_buf_delwri_dequeue(
1630         xfs_buf_t               *bp)
1631 {
1632         spinlock_t              *dwlk = &bp->b_target->bt_delwrite_lock;
1633         int                     dequeued = 0;
1634
1635         spin_lock(dwlk);
1636         if ((bp->b_flags & XBF_DELWRI) && !list_empty(&bp->b_list)) {
1637                 ASSERT(bp->b_flags & _XBF_DELWRI_Q);
1638                 list_del_init(&bp->b_list);
1639                 dequeued = 1;
1640         }
1641         bp->b_flags &= ~(XBF_DELWRI|_XBF_DELWRI_Q);
1642         spin_unlock(dwlk);
1643
1644         if (dequeued)
1645                 xfs_buf_rele(bp);
1646
1647         XB_TRACE(bp, "delwri_dq", (long)dequeued);
1648 }
1649
1650 STATIC void
1651 xfs_buf_runall_queues(
1652         struct workqueue_struct *queue)
1653 {
1654         flush_workqueue(queue);
1655 }
1656
1657 STATIC int
1658 xfsbufd_wakeup(
1659         int                     priority,
1660         gfp_t                   mask)
1661 {
1662         xfs_buftarg_t           *btp;
1663
1664         spin_lock(&xfs_buftarg_lock);
1665         list_for_each_entry(btp, &xfs_buftarg_list, bt_list) {
1666                 if (test_bit(XBT_FORCE_SLEEP, &btp->bt_flags))
1667                         continue;
1668                 set_bit(XBT_FORCE_FLUSH, &btp->bt_flags);
1669                 wake_up_process(btp->bt_task);
1670         }
1671         spin_unlock(&xfs_buftarg_lock);
1672         return 0;
1673 }
1674
1675 /*
1676  * Move as many buffers as specified to the supplied list
1677  * idicating if we skipped any buffers to prevent deadlocks.
1678  */
1679 STATIC int
1680 xfs_buf_delwri_split(
1681         xfs_buftarg_t   *target,
1682         struct list_head *list,
1683         unsigned long   age)
1684 {
1685         xfs_buf_t       *bp, *n;
1686         struct list_head *dwq = &target->bt_delwrite_queue;
1687         spinlock_t      *dwlk = &target->bt_delwrite_lock;
1688         int             skipped = 0;
1689         int             force;
1690
1691         force = test_and_clear_bit(XBT_FORCE_FLUSH, &target->bt_flags);
1692         INIT_LIST_HEAD(list);
1693         spin_lock(dwlk);
1694         list_for_each_entry_safe(bp, n, dwq, b_list) {
1695                 XB_TRACE(bp, "walkq1", (long)xfs_buf_ispin(bp));
1696                 ASSERT(bp->b_flags & XBF_DELWRI);
1697
1698                 if (!xfs_buf_ispin(bp) && !xfs_buf_cond_lock(bp)) {
1699                         if (!force &&
1700                             time_before(jiffies, bp->b_queuetime + age)) {
1701                                 xfs_buf_unlock(bp);
1702                                 break;
1703                         }
1704
1705                         bp->b_flags &= ~(XBF_DELWRI|_XBF_DELWRI_Q|
1706                                          _XBF_RUN_QUEUES);
1707                         bp->b_flags |= XBF_WRITE;
1708                         list_move_tail(&bp->b_list, list);
1709                 } else
1710                         skipped++;
1711         }
1712         spin_unlock(dwlk);
1713
1714         return skipped;
1715
1716 }
1717
1718 STATIC int
1719 xfsbufd(
1720         void            *data)
1721 {
1722         struct list_head tmp;
1723         xfs_buftarg_t   *target = (xfs_buftarg_t *)data;
1724         int             count;
1725         xfs_buf_t       *bp;
1726
1727         current->flags |= PF_MEMALLOC;
1728
1729         set_freezable();
1730
1731         do {
1732                 if (unlikely(freezing(current))) {
1733                         set_bit(XBT_FORCE_SLEEP, &target->bt_flags);
1734                         refrigerator();
1735                 } else {
1736                         clear_bit(XBT_FORCE_SLEEP, &target->bt_flags);
1737                 }
1738
1739                 schedule_timeout_interruptible(
1740                         xfs_buf_timer_centisecs * msecs_to_jiffies(10));
1741
1742                 xfs_buf_delwri_split(target, &tmp,
1743                                 xfs_buf_age_centisecs * msecs_to_jiffies(10));
1744
1745                 count = 0;
1746                 while (!list_empty(&tmp)) {
1747                         bp = list_entry(tmp.next, xfs_buf_t, b_list);
1748                         ASSERT(target == bp->b_target);
1749
1750                         list_del_init(&bp->b_list);
1751                         xfs_buf_iostrategy(bp);
1752                         count++;
1753                 }
1754
1755                 if (as_list_len > 0)
1756                         purge_addresses();
1757                 if (count)
1758                         blk_run_address_space(target->bt_mapping);
1759
1760         } while (!kthread_should_stop());
1761
1762         return 0;
1763 }
1764
1765 /*
1766  *      Go through all incore buffers, and release buffers if they belong to
1767  *      the given device. This is used in filesystem error handling to
1768  *      preserve the consistency of its metadata.
1769  */
1770 int
1771 xfs_flush_buftarg(
1772         xfs_buftarg_t   *target,
1773         int             wait)
1774 {
1775         struct list_head tmp;
1776         xfs_buf_t       *bp, *n;
1777         int             pincount = 0;
1778
1779         xfs_buf_runall_queues(xfsconvertd_workqueue);
1780         xfs_buf_runall_queues(xfsdatad_workqueue);
1781         xfs_buf_runall_queues(xfslogd_workqueue);
1782
1783         set_bit(XBT_FORCE_FLUSH, &target->bt_flags);
1784         pincount = xfs_buf_delwri_split(target, &tmp, 0);
1785
1786         /*
1787          * Dropped the delayed write list lock, now walk the temporary list
1788          */
1789         list_for_each_entry_safe(bp, n, &tmp, b_list) {
1790                 ASSERT(target == bp->b_target);
1791                 if (wait)
1792                         bp->b_flags &= ~XBF_ASYNC;
1793                 else
1794                         list_del_init(&bp->b_list);
1795
1796                 xfs_buf_iostrategy(bp);
1797         }
1798
1799         if (wait)
1800                 blk_run_address_space(target->bt_mapping);
1801
1802         /*
1803          * Remaining list items must be flushed before returning
1804          */
1805         while (!list_empty(&tmp)) {
1806                 bp = list_entry(tmp.next, xfs_buf_t, b_list);
1807
1808                 list_del_init(&bp->b_list);
1809                 xfs_iowait(bp);
1810                 xfs_buf_relse(bp);
1811         }
1812
1813         return pincount;
1814 }
1815
1816 int __init
1817 xfs_buf_init(void)
1818 {
1819 #ifdef XFS_BUF_TRACE
1820         xfs_buf_trace_buf = ktrace_alloc(XFS_BUF_TRACE_SIZE, KM_NOFS);
1821 #endif
1822
1823         xfs_buf_zone = kmem_zone_init_flags(sizeof(xfs_buf_t), "xfs_buf",
1824                                                 KM_ZONE_HWALIGN, NULL);
1825         if (!xfs_buf_zone)
1826                 goto out_free_trace_buf;
1827
1828         xfslogd_workqueue = create_workqueue("xfslogd");
1829         if (!xfslogd_workqueue)
1830                 goto out_free_buf_zone;
1831
1832         xfsdatad_workqueue = create_workqueue("xfsdatad");
1833         if (!xfsdatad_workqueue)
1834                 goto out_destroy_xfslogd_workqueue;
1835
1836         xfsconvertd_workqueue = create_workqueue("xfsconvertd");
1837         if (!xfsconvertd_workqueue)
1838                 goto out_destroy_xfsdatad_workqueue;
1839
1840         register_shrinker(&xfs_buf_shake);
1841         return 0;
1842
1843  out_destroy_xfsdatad_workqueue:
1844         destroy_workqueue(xfsdatad_workqueue);
1845  out_destroy_xfslogd_workqueue:
1846         destroy_workqueue(xfslogd_workqueue);
1847  out_free_buf_zone:
1848         kmem_zone_destroy(xfs_buf_zone);
1849  out_free_trace_buf:
1850 #ifdef XFS_BUF_TRACE
1851         ktrace_free(xfs_buf_trace_buf);
1852 #endif
1853         return -ENOMEM;
1854 }
1855
1856 void
1857 xfs_buf_terminate(void)
1858 {
1859         unregister_shrinker(&xfs_buf_shake);
1860         destroy_workqueue(xfsconvertd_workqueue);
1861         destroy_workqueue(xfsdatad_workqueue);
1862         destroy_workqueue(xfslogd_workqueue);
1863         kmem_zone_destroy(xfs_buf_zone);
1864 #ifdef XFS_BUF_TRACE
1865         ktrace_free(xfs_buf_trace_buf);
1866 #endif
1867 }
1868
1869 #ifdef CONFIG_KDB_MODULES
1870 struct list_head *
1871 xfs_get_buftarg_list(void)
1872 {
1873         return &xfs_buftarg_list;
1874 }
1875 #endif