Merge commit '900cfa46191a7d87cf1891924cb90499287fd235'; branches 'timers/nohz',...
[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                                 for (i = 0; i < bp->b_page_count; i++)
391                                         unlock_page(bp->b_pages[i]);
392                                 return -ENOMEM;
393                         }
394
395                         /*
396                          * This could deadlock.
397                          *
398                          * But until all the XFS lowlevel code is revamped to
399                          * handle buffer allocation failures we can't do much.
400                          */
401                         if (!(++retries % 100))
402                                 printk(KERN_ERR
403                                         "XFS: possible memory allocation "
404                                         "deadlock in %s (mode:0x%x)\n",
405                                         __func__, gfp_mask);
406
407                         XFS_STATS_INC(xb_page_retries);
408                         xfsbufd_wakeup(0, gfp_mask);
409                         congestion_wait(WRITE, HZ/50);
410                         goto retry;
411                 }
412
413                 XFS_STATS_INC(xb_page_found);
414
415                 nbytes = min_t(size_t, size, PAGE_CACHE_SIZE - offset);
416                 size -= nbytes;
417
418                 ASSERT(!PagePrivate(page));
419                 if (!PageUptodate(page)) {
420                         page_count--;
421                         if (blocksize >= PAGE_CACHE_SIZE) {
422                                 if (flags & XBF_READ)
423                                         bp->b_flags |= _XBF_PAGE_LOCKED;
424                         } else if (!PagePrivate(page)) {
425                                 if (test_page_region(page, offset, nbytes))
426                                         page_count++;
427                         }
428                 }
429
430                 bp->b_pages[i] = page;
431                 offset = 0;
432         }
433
434         if (!(bp->b_flags & _XBF_PAGE_LOCKED)) {
435                 for (i = 0; i < bp->b_page_count; i++)
436                         unlock_page(bp->b_pages[i]);
437         }
438
439         if (page_count == bp->b_page_count)
440                 bp->b_flags |= XBF_DONE;
441
442         XB_TRACE(bp, "lookup_pages", (long)page_count);
443         return error;
444 }
445
446 /*
447  *      Map buffer into kernel address-space if nessecary.
448  */
449 STATIC int
450 _xfs_buf_map_pages(
451         xfs_buf_t               *bp,
452         uint                    flags)
453 {
454         /* A single page buffer is always mappable */
455         if (bp->b_page_count == 1) {
456                 bp->b_addr = page_address(bp->b_pages[0]) + bp->b_offset;
457                 bp->b_flags |= XBF_MAPPED;
458         } else if (flags & XBF_MAPPED) {
459                 if (as_list_len > 64)
460                         purge_addresses();
461                 bp->b_addr = vmap(bp->b_pages, bp->b_page_count,
462                                         VM_MAP, PAGE_KERNEL);
463                 if (unlikely(bp->b_addr == NULL))
464                         return -ENOMEM;
465                 bp->b_addr += bp->b_offset;
466                 bp->b_flags |= XBF_MAPPED;
467         }
468
469         return 0;
470 }
471
472 /*
473  *      Finding and Reading Buffers
474  */
475
476 /*
477  *      Look up, and creates if absent, a lockable buffer for
478  *      a given range of an inode.  The buffer is returned
479  *      locked.  If other overlapping buffers exist, they are
480  *      released before the new buffer is created and locked,
481  *      which may imply that this call will block until those buffers
482  *      are unlocked.  No I/O is implied by this call.
483  */
484 xfs_buf_t *
485 _xfs_buf_find(
486         xfs_buftarg_t           *btp,   /* block device target          */
487         xfs_off_t               ioff,   /* starting offset of range     */
488         size_t                  isize,  /* length of range              */
489         xfs_buf_flags_t         flags,
490         xfs_buf_t               *new_bp)
491 {
492         xfs_off_t               range_base;
493         size_t                  range_length;
494         xfs_bufhash_t           *hash;
495         xfs_buf_t               *bp, *n;
496
497         range_base = (ioff << BBSHIFT);
498         range_length = (isize << BBSHIFT);
499
500         /* Check for IOs smaller than the sector size / not sector aligned */
501         ASSERT(!(range_length < (1 << btp->bt_sshift)));
502         ASSERT(!(range_base & (xfs_off_t)btp->bt_smask));
503
504         hash = &btp->bt_hash[hash_long((unsigned long)ioff, btp->bt_hashshift)];
505
506         spin_lock(&hash->bh_lock);
507
508         list_for_each_entry_safe(bp, n, &hash->bh_list, b_hash_list) {
509                 ASSERT(btp == bp->b_target);
510                 if (bp->b_file_offset == range_base &&
511                     bp->b_buffer_length == range_length) {
512                         /*
513                          * If we look at something, bring it to the
514                          * front of the list for next time.
515                          */
516                         atomic_inc(&bp->b_hold);
517                         list_move(&bp->b_hash_list, &hash->bh_list);
518                         goto found;
519                 }
520         }
521
522         /* No match found */
523         if (new_bp) {
524                 _xfs_buf_initialize(new_bp, btp, range_base,
525                                 range_length, flags);
526                 new_bp->b_hash = hash;
527                 list_add(&new_bp->b_hash_list, &hash->bh_list);
528         } else {
529                 XFS_STATS_INC(xb_miss_locked);
530         }
531
532         spin_unlock(&hash->bh_lock);
533         return new_bp;
534
535 found:
536         spin_unlock(&hash->bh_lock);
537
538         /* Attempt to get the semaphore without sleeping,
539          * if this does not work then we need to drop the
540          * spinlock and do a hard attempt on the semaphore.
541          */
542         if (down_trylock(&bp->b_sema)) {
543                 if (!(flags & XBF_TRYLOCK)) {
544                         /* wait for buffer ownership */
545                         XB_TRACE(bp, "get_lock", 0);
546                         xfs_buf_lock(bp);
547                         XFS_STATS_INC(xb_get_locked_waited);
548                 } else {
549                         /* We asked for a trylock and failed, no need
550                          * to look at file offset and length here, we
551                          * know that this buffer at least overlaps our
552                          * buffer and is locked, therefore our buffer
553                          * either does not exist, or is this buffer.
554                          */
555                         xfs_buf_rele(bp);
556                         XFS_STATS_INC(xb_busy_locked);
557                         return NULL;
558                 }
559         } else {
560                 /* trylock worked */
561                 XB_SET_OWNER(bp);
562         }
563
564         if (bp->b_flags & XBF_STALE) {
565                 ASSERT((bp->b_flags & _XBF_DELWRI_Q) == 0);
566                 bp->b_flags &= XBF_MAPPED;
567         }
568         XB_TRACE(bp, "got_lock", 0);
569         XFS_STATS_INC(xb_get_locked);
570         return bp;
571 }
572
573 /*
574  *      Assembles a buffer covering the specified range.
575  *      Storage in memory for all portions of the buffer will be allocated,
576  *      although backing storage may not be.
577  */
578 xfs_buf_t *
579 xfs_buf_get_flags(
580         xfs_buftarg_t           *target,/* target for buffer            */
581         xfs_off_t               ioff,   /* starting offset of range     */
582         size_t                  isize,  /* length of range              */
583         xfs_buf_flags_t         flags)
584 {
585         xfs_buf_t               *bp, *new_bp;
586         int                     error = 0, i;
587
588         new_bp = xfs_buf_allocate(flags);
589         if (unlikely(!new_bp))
590                 return NULL;
591
592         bp = _xfs_buf_find(target, ioff, isize, flags, new_bp);
593         if (bp == new_bp) {
594                 error = _xfs_buf_lookup_pages(bp, flags);
595                 if (error)
596                         goto no_buffer;
597         } else {
598                 xfs_buf_deallocate(new_bp);
599                 if (unlikely(bp == NULL))
600                         return NULL;
601         }
602
603         for (i = 0; i < bp->b_page_count; i++)
604                 mark_page_accessed(bp->b_pages[i]);
605
606         if (!(bp->b_flags & XBF_MAPPED)) {
607                 error = _xfs_buf_map_pages(bp, flags);
608                 if (unlikely(error)) {
609                         printk(KERN_WARNING "%s: failed to map pages\n",
610                                         __func__);
611                         goto no_buffer;
612                 }
613         }
614
615         XFS_STATS_INC(xb_get);
616
617         /*
618          * Always fill in the block number now, the mapped cases can do
619          * their own overlay of this later.
620          */
621         bp->b_bn = ioff;
622         bp->b_count_desired = bp->b_buffer_length;
623
624         XB_TRACE(bp, "get", (unsigned long)flags);
625         return bp;
626
627  no_buffer:
628         if (flags & (XBF_LOCK | XBF_TRYLOCK))
629                 xfs_buf_unlock(bp);
630         xfs_buf_rele(bp);
631         return NULL;
632 }
633
634 xfs_buf_t *
635 xfs_buf_read_flags(
636         xfs_buftarg_t           *target,
637         xfs_off_t               ioff,
638         size_t                  isize,
639         xfs_buf_flags_t         flags)
640 {
641         xfs_buf_t               *bp;
642
643         flags |= XBF_READ;
644
645         bp = xfs_buf_get_flags(target, ioff, isize, flags);
646         if (bp) {
647                 if (!XFS_BUF_ISDONE(bp)) {
648                         XB_TRACE(bp, "read", (unsigned long)flags);
649                         XFS_STATS_INC(xb_get_read);
650                         xfs_buf_iostart(bp, flags);
651                 } else if (flags & XBF_ASYNC) {
652                         XB_TRACE(bp, "read_async", (unsigned long)flags);
653                         /*
654                          * Read ahead call which is already satisfied,
655                          * drop the buffer
656                          */
657                         goto no_buffer;
658                 } else {
659                         XB_TRACE(bp, "read_done", (unsigned long)flags);
660                         /* We do not want read in the flags */
661                         bp->b_flags &= ~XBF_READ;
662                 }
663         }
664
665         return bp;
666
667  no_buffer:
668         if (flags & (XBF_LOCK | XBF_TRYLOCK))
669                 xfs_buf_unlock(bp);
670         xfs_buf_rele(bp);
671         return NULL;
672 }
673
674 /*
675  *      If we are not low on memory then do the readahead in a deadlock
676  *      safe manner.
677  */
678 void
679 xfs_buf_readahead(
680         xfs_buftarg_t           *target,
681         xfs_off_t               ioff,
682         size_t                  isize,
683         xfs_buf_flags_t         flags)
684 {
685         struct backing_dev_info *bdi;
686
687         bdi = target->bt_mapping->backing_dev_info;
688         if (bdi_read_congested(bdi))
689                 return;
690
691         flags |= (XBF_TRYLOCK|XBF_ASYNC|XBF_READ_AHEAD);
692         xfs_buf_read_flags(target, ioff, isize, flags);
693 }
694
695 xfs_buf_t *
696 xfs_buf_get_empty(
697         size_t                  len,
698         xfs_buftarg_t           *target)
699 {
700         xfs_buf_t               *bp;
701
702         bp = xfs_buf_allocate(0);
703         if (bp)
704                 _xfs_buf_initialize(bp, target, 0, len, 0);
705         return bp;
706 }
707
708 static inline struct page *
709 mem_to_page(
710         void                    *addr)
711 {
712         if ((!is_vmalloc_addr(addr))) {
713                 return virt_to_page(addr);
714         } else {
715                 return vmalloc_to_page(addr);
716         }
717 }
718
719 int
720 xfs_buf_associate_memory(
721         xfs_buf_t               *bp,
722         void                    *mem,
723         size_t                  len)
724 {
725         int                     rval;
726         int                     i = 0;
727         unsigned long           pageaddr;
728         unsigned long           offset;
729         size_t                  buflen;
730         int                     page_count;
731
732         pageaddr = (unsigned long)mem & PAGE_CACHE_MASK;
733         offset = (unsigned long)mem - pageaddr;
734         buflen = PAGE_CACHE_ALIGN(len + offset);
735         page_count = buflen >> PAGE_CACHE_SHIFT;
736
737         /* Free any previous set of page pointers */
738         if (bp->b_pages)
739                 _xfs_buf_free_pages(bp);
740
741         bp->b_pages = NULL;
742         bp->b_addr = mem;
743
744         rval = _xfs_buf_get_pages(bp, page_count, 0);
745         if (rval)
746                 return rval;
747
748         bp->b_offset = offset;
749
750         for (i = 0; i < bp->b_page_count; i++) {
751                 bp->b_pages[i] = mem_to_page((void *)pageaddr);
752                 pageaddr += PAGE_CACHE_SIZE;
753         }
754
755         bp->b_count_desired = len;
756         bp->b_buffer_length = buflen;
757         bp->b_flags |= XBF_MAPPED;
758         bp->b_flags &= ~_XBF_PAGE_LOCKED;
759
760         return 0;
761 }
762
763 xfs_buf_t *
764 xfs_buf_get_noaddr(
765         size_t                  len,
766         xfs_buftarg_t           *target)
767 {
768         unsigned long           page_count = PAGE_ALIGN(len) >> PAGE_SHIFT;
769         int                     error, i;
770         xfs_buf_t               *bp;
771
772         bp = xfs_buf_allocate(0);
773         if (unlikely(bp == NULL))
774                 goto fail;
775         _xfs_buf_initialize(bp, target, 0, len, 0);
776
777         error = _xfs_buf_get_pages(bp, page_count, 0);
778         if (error)
779                 goto fail_free_buf;
780
781         for (i = 0; i < page_count; i++) {
782                 bp->b_pages[i] = alloc_page(GFP_KERNEL);
783                 if (!bp->b_pages[i])
784                         goto fail_free_mem;
785         }
786         bp->b_flags |= _XBF_PAGES;
787
788         error = _xfs_buf_map_pages(bp, XBF_MAPPED);
789         if (unlikely(error)) {
790                 printk(KERN_WARNING "%s: failed to map pages\n",
791                                 __func__);
792                 goto fail_free_mem;
793         }
794
795         xfs_buf_unlock(bp);
796
797         XB_TRACE(bp, "no_daddr", len);
798         return bp;
799
800  fail_free_mem:
801         while (--i >= 0)
802                 __free_page(bp->b_pages[i]);
803         _xfs_buf_free_pages(bp);
804  fail_free_buf:
805         xfs_buf_deallocate(bp);
806  fail:
807         return NULL;
808 }
809
810 /*
811  *      Increment reference count on buffer, to hold the buffer concurrently
812  *      with another thread which may release (free) the buffer asynchronously.
813  *      Must hold the buffer already to call this function.
814  */
815 void
816 xfs_buf_hold(
817         xfs_buf_t               *bp)
818 {
819         atomic_inc(&bp->b_hold);
820         XB_TRACE(bp, "hold", 0);
821 }
822
823 /*
824  *      Releases a hold on the specified buffer.  If the
825  *      the hold count is 1, calls xfs_buf_free.
826  */
827 void
828 xfs_buf_rele(
829         xfs_buf_t               *bp)
830 {
831         xfs_bufhash_t           *hash = bp->b_hash;
832
833         XB_TRACE(bp, "rele", bp->b_relse);
834
835         if (unlikely(!hash)) {
836                 ASSERT(!bp->b_relse);
837                 if (atomic_dec_and_test(&bp->b_hold))
838                         xfs_buf_free(bp);
839                 return;
840         }
841
842         if (atomic_dec_and_lock(&bp->b_hold, &hash->bh_lock)) {
843                 if (bp->b_relse) {
844                         atomic_inc(&bp->b_hold);
845                         spin_unlock(&hash->bh_lock);
846                         (*(bp->b_relse)) (bp);
847                 } else if (bp->b_flags & XBF_FS_MANAGED) {
848                         spin_unlock(&hash->bh_lock);
849                 } else {
850                         ASSERT(!(bp->b_flags & (XBF_DELWRI|_XBF_DELWRI_Q)));
851                         list_del_init(&bp->b_hash_list);
852                         spin_unlock(&hash->bh_lock);
853                         xfs_buf_free(bp);
854                 }
855         } else {
856                 /*
857                  * Catch reference count leaks
858                  */
859                 ASSERT(atomic_read(&bp->b_hold) >= 0);
860         }
861 }
862
863
864 /*
865  *      Mutual exclusion on buffers.  Locking model:
866  *
867  *      Buffers associated with inodes for which buffer locking
868  *      is not enabled are not protected by semaphores, and are
869  *      assumed to be exclusively owned by the caller.  There is a
870  *      spinlock in the buffer, used by the caller when concurrent
871  *      access is possible.
872  */
873
874 /*
875  *      Locks a buffer object, if it is not already locked.
876  *      Note that this in no way locks the underlying pages, so it is only
877  *      useful for synchronizing concurrent use of buffer objects, not for
878  *      synchronizing independent access to the underlying pages.
879  */
880 int
881 xfs_buf_cond_lock(
882         xfs_buf_t               *bp)
883 {
884         int                     locked;
885
886         locked = down_trylock(&bp->b_sema) == 0;
887         if (locked) {
888                 XB_SET_OWNER(bp);
889         }
890         XB_TRACE(bp, "cond_lock", (long)locked);
891         return locked ? 0 : -EBUSY;
892 }
893
894 #if defined(DEBUG) || defined(XFS_BLI_TRACE)
895 int
896 xfs_buf_lock_value(
897         xfs_buf_t               *bp)
898 {
899         return bp->b_sema.count;
900 }
901 #endif
902
903 /*
904  *      Locks a buffer object.
905  *      Note that this in no way locks the underlying pages, so it is only
906  *      useful for synchronizing concurrent use of buffer objects, not for
907  *      synchronizing independent access to the underlying pages.
908  */
909 void
910 xfs_buf_lock(
911         xfs_buf_t               *bp)
912 {
913         XB_TRACE(bp, "lock", 0);
914         if (atomic_read(&bp->b_io_remaining))
915                 blk_run_address_space(bp->b_target->bt_mapping);
916         down(&bp->b_sema);
917         XB_SET_OWNER(bp);
918         XB_TRACE(bp, "locked", 0);
919 }
920
921 /*
922  *      Releases the lock on the buffer object.
923  *      If the buffer is marked delwri but is not queued, do so before we
924  *      unlock the buffer as we need to set flags correctly.  We also need to
925  *      take a reference for the delwri queue because the unlocker is going to
926  *      drop their's and they don't know we just queued it.
927  */
928 void
929 xfs_buf_unlock(
930         xfs_buf_t               *bp)
931 {
932         if ((bp->b_flags & (XBF_DELWRI|_XBF_DELWRI_Q)) == XBF_DELWRI) {
933                 atomic_inc(&bp->b_hold);
934                 bp->b_flags |= XBF_ASYNC;
935                 xfs_buf_delwri_queue(bp, 0);
936         }
937
938         XB_CLEAR_OWNER(bp);
939         up(&bp->b_sema);
940         XB_TRACE(bp, "unlock", 0);
941 }
942
943
944 /*
945  *      Pinning Buffer Storage in Memory
946  *      Ensure that no attempt to force a buffer to disk will succeed.
947  */
948 void
949 xfs_buf_pin(
950         xfs_buf_t               *bp)
951 {
952         atomic_inc(&bp->b_pin_count);
953         XB_TRACE(bp, "pin", (long)bp->b_pin_count.counter);
954 }
955
956 void
957 xfs_buf_unpin(
958         xfs_buf_t               *bp)
959 {
960         if (atomic_dec_and_test(&bp->b_pin_count))
961                 wake_up_all(&bp->b_waiters);
962         XB_TRACE(bp, "unpin", (long)bp->b_pin_count.counter);
963 }
964
965 int
966 xfs_buf_ispin(
967         xfs_buf_t               *bp)
968 {
969         return atomic_read(&bp->b_pin_count);
970 }
971
972 STATIC void
973 xfs_buf_wait_unpin(
974         xfs_buf_t               *bp)
975 {
976         DECLARE_WAITQUEUE       (wait, current);
977
978         if (atomic_read(&bp->b_pin_count) == 0)
979                 return;
980
981         add_wait_queue(&bp->b_waiters, &wait);
982         for (;;) {
983                 set_current_state(TASK_UNINTERRUPTIBLE);
984                 if (atomic_read(&bp->b_pin_count) == 0)
985                         break;
986                 if (atomic_read(&bp->b_io_remaining))
987                         blk_run_address_space(bp->b_target->bt_mapping);
988                 schedule();
989         }
990         remove_wait_queue(&bp->b_waiters, &wait);
991         set_current_state(TASK_RUNNING);
992 }
993
994 /*
995  *      Buffer Utility Routines
996  */
997
998 STATIC void
999 xfs_buf_iodone_work(
1000         struct work_struct      *work)
1001 {
1002         xfs_buf_t               *bp =
1003                 container_of(work, xfs_buf_t, b_iodone_work);
1004
1005         /*
1006          * We can get an EOPNOTSUPP to ordered writes.  Here we clear the
1007          * ordered flag and reissue them.  Because we can't tell the higher
1008          * layers directly that they should not issue ordered I/O anymore, they
1009          * need to check if the ordered flag was cleared during I/O completion.
1010          */
1011         if ((bp->b_error == EOPNOTSUPP) &&
1012             (bp->b_flags & (XBF_ORDERED|XBF_ASYNC)) == (XBF_ORDERED|XBF_ASYNC)) {
1013                 XB_TRACE(bp, "ordered_retry", bp->b_iodone);
1014                 bp->b_flags &= ~XBF_ORDERED;
1015                 xfs_buf_iorequest(bp);
1016         } else if (bp->b_iodone)
1017                 (*(bp->b_iodone))(bp);
1018         else if (bp->b_flags & XBF_ASYNC)
1019                 xfs_buf_relse(bp);
1020 }
1021
1022 void
1023 xfs_buf_ioend(
1024         xfs_buf_t               *bp,
1025         int                     schedule)
1026 {
1027         bp->b_flags &= ~(XBF_READ | XBF_WRITE | XBF_READ_AHEAD);
1028         if (bp->b_error == 0)
1029                 bp->b_flags |= XBF_DONE;
1030
1031         XB_TRACE(bp, "iodone", bp->b_iodone);
1032
1033         if ((bp->b_iodone) || (bp->b_flags & XBF_ASYNC)) {
1034                 if (schedule) {
1035                         INIT_WORK(&bp->b_iodone_work, xfs_buf_iodone_work);
1036                         queue_work(xfslogd_workqueue, &bp->b_iodone_work);
1037                 } else {
1038                         xfs_buf_iodone_work(&bp->b_iodone_work);
1039                 }
1040         } else {
1041                 up(&bp->b_iodonesema);
1042         }
1043 }
1044
1045 void
1046 xfs_buf_ioerror(
1047         xfs_buf_t               *bp,
1048         int                     error)
1049 {
1050         ASSERT(error >= 0 && error <= 0xffff);
1051         bp->b_error = (unsigned short)error;
1052         XB_TRACE(bp, "ioerror", (unsigned long)error);
1053 }
1054
1055 /*
1056  *      Initiate I/O on a buffer, based on the flags supplied.
1057  *      The b_iodone routine in the buffer supplied will only be called
1058  *      when all of the subsidiary I/O requests, if any, have been completed.
1059  */
1060 int
1061 xfs_buf_iostart(
1062         xfs_buf_t               *bp,
1063         xfs_buf_flags_t         flags)
1064 {
1065         int                     status = 0;
1066
1067         XB_TRACE(bp, "iostart", (unsigned long)flags);
1068
1069         if (flags & XBF_DELWRI) {
1070                 bp->b_flags &= ~(XBF_READ | XBF_WRITE | XBF_ASYNC);
1071                 bp->b_flags |= flags & (XBF_DELWRI | XBF_ASYNC);
1072                 xfs_buf_delwri_queue(bp, 1);
1073                 return 0;
1074         }
1075
1076         bp->b_flags &= ~(XBF_READ | XBF_WRITE | XBF_ASYNC | XBF_DELWRI | \
1077                         XBF_READ_AHEAD | _XBF_RUN_QUEUES);
1078         bp->b_flags |= flags & (XBF_READ | XBF_WRITE | XBF_ASYNC | \
1079                         XBF_READ_AHEAD | _XBF_RUN_QUEUES);
1080
1081         BUG_ON(bp->b_bn == XFS_BUF_DADDR_NULL);
1082
1083         /* For writes allow an alternate strategy routine to precede
1084          * the actual I/O request (which may not be issued at all in
1085          * a shutdown situation, for example).
1086          */
1087         status = (flags & XBF_WRITE) ?
1088                 xfs_buf_iostrategy(bp) : xfs_buf_iorequest(bp);
1089
1090         /* Wait for I/O if we are not an async request.
1091          * Note: async I/O request completion will release the buffer,
1092          * and that can already be done by this point.  So using the
1093          * buffer pointer from here on, after async I/O, is invalid.
1094          */
1095         if (!status && !(flags & XBF_ASYNC))
1096                 status = xfs_buf_iowait(bp);
1097
1098         return status;
1099 }
1100
1101 STATIC_INLINE void
1102 _xfs_buf_ioend(
1103         xfs_buf_t               *bp,
1104         int                     schedule)
1105 {
1106         if (atomic_dec_and_test(&bp->b_io_remaining) == 1) {
1107                 bp->b_flags &= ~_XBF_PAGE_LOCKED;
1108                 xfs_buf_ioend(bp, schedule);
1109         }
1110 }
1111
1112 STATIC void
1113 xfs_buf_bio_end_io(
1114         struct bio              *bio,
1115         int                     error)
1116 {
1117         xfs_buf_t               *bp = (xfs_buf_t *)bio->bi_private;
1118         unsigned int            blocksize = bp->b_target->bt_bsize;
1119         struct bio_vec          *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
1120
1121         if (!test_bit(BIO_UPTODATE, &bio->bi_flags))
1122                 bp->b_error = EIO;
1123
1124         do {
1125                 struct page     *page = bvec->bv_page;
1126
1127                 ASSERT(!PagePrivate(page));
1128                 if (unlikely(bp->b_error)) {
1129                         if (bp->b_flags & XBF_READ)
1130                                 ClearPageUptodate(page);
1131                 } else if (blocksize >= PAGE_CACHE_SIZE) {
1132                         SetPageUptodate(page);
1133                 } else if (!PagePrivate(page) &&
1134                                 (bp->b_flags & _XBF_PAGE_CACHE)) {
1135                         set_page_region(page, bvec->bv_offset, bvec->bv_len);
1136                 }
1137
1138                 if (--bvec >= bio->bi_io_vec)
1139                         prefetchw(&bvec->bv_page->flags);
1140
1141                 if (bp->b_flags & _XBF_PAGE_LOCKED)
1142                         unlock_page(page);
1143         } while (bvec >= bio->bi_io_vec);
1144
1145         _xfs_buf_ioend(bp, 1);
1146         bio_put(bio);
1147 }
1148
1149 STATIC void
1150 _xfs_buf_ioapply(
1151         xfs_buf_t               *bp)
1152 {
1153         int                     rw, map_i, total_nr_pages, nr_pages;
1154         struct bio              *bio;
1155         int                     offset = bp->b_offset;
1156         int                     size = bp->b_count_desired;
1157         sector_t                sector = bp->b_bn;
1158         unsigned int            blocksize = bp->b_target->bt_bsize;
1159
1160         total_nr_pages = bp->b_page_count;
1161         map_i = 0;
1162
1163         if (bp->b_flags & XBF_ORDERED) {
1164                 ASSERT(!(bp->b_flags & XBF_READ));
1165                 rw = WRITE_BARRIER;
1166         } else if (bp->b_flags & _XBF_RUN_QUEUES) {
1167                 ASSERT(!(bp->b_flags & XBF_READ_AHEAD));
1168                 bp->b_flags &= ~_XBF_RUN_QUEUES;
1169                 rw = (bp->b_flags & XBF_WRITE) ? WRITE_SYNC : READ_SYNC;
1170         } else {
1171                 rw = (bp->b_flags & XBF_WRITE) ? WRITE :
1172                      (bp->b_flags & XBF_READ_AHEAD) ? READA : READ;
1173         }
1174
1175         /* Special code path for reading a sub page size buffer in --
1176          * we populate up the whole page, and hence the other metadata
1177          * in the same page.  This optimization is only valid when the
1178          * filesystem block size is not smaller than the page size.
1179          */
1180         if ((bp->b_buffer_length < PAGE_CACHE_SIZE) &&
1181             ((bp->b_flags & (XBF_READ|_XBF_PAGE_LOCKED)) ==
1182               (XBF_READ|_XBF_PAGE_LOCKED)) &&
1183             (blocksize >= PAGE_CACHE_SIZE)) {
1184                 bio = bio_alloc(GFP_NOIO, 1);
1185
1186                 bio->bi_bdev = bp->b_target->bt_bdev;
1187                 bio->bi_sector = sector - (offset >> BBSHIFT);
1188                 bio->bi_end_io = xfs_buf_bio_end_io;
1189                 bio->bi_private = bp;
1190
1191                 bio_add_page(bio, bp->b_pages[0], PAGE_CACHE_SIZE, 0);
1192                 size = 0;
1193
1194                 atomic_inc(&bp->b_io_remaining);
1195
1196                 goto submit_io;
1197         }
1198
1199 next_chunk:
1200         atomic_inc(&bp->b_io_remaining);
1201         nr_pages = BIO_MAX_SECTORS >> (PAGE_SHIFT - BBSHIFT);
1202         if (nr_pages > total_nr_pages)
1203                 nr_pages = total_nr_pages;
1204
1205         bio = bio_alloc(GFP_NOIO, nr_pages);
1206         bio->bi_bdev = bp->b_target->bt_bdev;
1207         bio->bi_sector = sector;
1208         bio->bi_end_io = xfs_buf_bio_end_io;
1209         bio->bi_private = bp;
1210
1211         for (; size && nr_pages; nr_pages--, map_i++) {
1212                 int     rbytes, nbytes = PAGE_CACHE_SIZE - offset;
1213
1214                 if (nbytes > size)
1215                         nbytes = size;
1216
1217                 rbytes = bio_add_page(bio, bp->b_pages[map_i], nbytes, offset);
1218                 if (rbytes < nbytes)
1219                         break;
1220
1221                 offset = 0;
1222                 sector += nbytes >> BBSHIFT;
1223                 size -= nbytes;
1224                 total_nr_pages--;
1225         }
1226
1227 submit_io:
1228         if (likely(bio->bi_size)) {
1229                 submit_bio(rw, bio);
1230                 if (size)
1231                         goto next_chunk;
1232         } else {
1233                 bio_put(bio);
1234                 xfs_buf_ioerror(bp, EIO);
1235         }
1236 }
1237
1238 int
1239 xfs_buf_iorequest(
1240         xfs_buf_t               *bp)
1241 {
1242         XB_TRACE(bp, "iorequest", 0);
1243
1244         if (bp->b_flags & XBF_DELWRI) {
1245                 xfs_buf_delwri_queue(bp, 1);
1246                 return 0;
1247         }
1248
1249         if (bp->b_flags & XBF_WRITE) {
1250                 xfs_buf_wait_unpin(bp);
1251         }
1252
1253         xfs_buf_hold(bp);
1254
1255         /* Set the count to 1 initially, this will stop an I/O
1256          * completion callout which happens before we have started
1257          * all the I/O from calling xfs_buf_ioend too early.
1258          */
1259         atomic_set(&bp->b_io_remaining, 1);
1260         _xfs_buf_ioapply(bp);
1261         _xfs_buf_ioend(bp, 0);
1262
1263         xfs_buf_rele(bp);
1264         return 0;
1265 }
1266
1267 /*
1268  *      Waits for I/O to complete on the buffer supplied.
1269  *      It returns immediately if no I/O is pending.
1270  *      It returns the I/O error code, if any, or 0 if there was no error.
1271  */
1272 int
1273 xfs_buf_iowait(
1274         xfs_buf_t               *bp)
1275 {
1276         XB_TRACE(bp, "iowait", 0);
1277         if (atomic_read(&bp->b_io_remaining))
1278                 blk_run_address_space(bp->b_target->bt_mapping);
1279         down(&bp->b_iodonesema);
1280         XB_TRACE(bp, "iowaited", (long)bp->b_error);
1281         return bp->b_error;
1282 }
1283
1284 xfs_caddr_t
1285 xfs_buf_offset(
1286         xfs_buf_t               *bp,
1287         size_t                  offset)
1288 {
1289         struct page             *page;
1290
1291         if (bp->b_flags & XBF_MAPPED)
1292                 return XFS_BUF_PTR(bp) + offset;
1293
1294         offset += bp->b_offset;
1295         page = bp->b_pages[offset >> PAGE_CACHE_SHIFT];
1296         return (xfs_caddr_t)page_address(page) + (offset & (PAGE_CACHE_SIZE-1));
1297 }
1298
1299 /*
1300  *      Move data into or out of a buffer.
1301  */
1302 void
1303 xfs_buf_iomove(
1304         xfs_buf_t               *bp,    /* buffer to process            */
1305         size_t                  boff,   /* starting buffer offset       */
1306         size_t                  bsize,  /* length to copy               */
1307         caddr_t                 data,   /* data address                 */
1308         xfs_buf_rw_t            mode)   /* read/write/zero flag         */
1309 {
1310         size_t                  bend, cpoff, csize;
1311         struct page             *page;
1312
1313         bend = boff + bsize;
1314         while (boff < bend) {
1315                 page = bp->b_pages[xfs_buf_btoct(boff + bp->b_offset)];
1316                 cpoff = xfs_buf_poff(boff + bp->b_offset);
1317                 csize = min_t(size_t,
1318                               PAGE_CACHE_SIZE-cpoff, bp->b_count_desired-boff);
1319
1320                 ASSERT(((csize + cpoff) <= PAGE_CACHE_SIZE));
1321
1322                 switch (mode) {
1323                 case XBRW_ZERO:
1324                         memset(page_address(page) + cpoff, 0, csize);
1325                         break;
1326                 case XBRW_READ:
1327                         memcpy(data, page_address(page) + cpoff, csize);
1328                         break;
1329                 case XBRW_WRITE:
1330                         memcpy(page_address(page) + cpoff, data, csize);
1331                 }
1332
1333                 boff += csize;
1334                 data += csize;
1335         }
1336 }
1337
1338 /*
1339  *      Handling of buffer targets (buftargs).
1340  */
1341
1342 /*
1343  *      Wait for any bufs with callbacks that have been submitted but
1344  *      have not yet returned... walk the hash list for the target.
1345  */
1346 void
1347 xfs_wait_buftarg(
1348         xfs_buftarg_t   *btp)
1349 {
1350         xfs_buf_t       *bp, *n;
1351         xfs_bufhash_t   *hash;
1352         uint            i;
1353
1354         for (i = 0; i < (1 << btp->bt_hashshift); i++) {
1355                 hash = &btp->bt_hash[i];
1356 again:
1357                 spin_lock(&hash->bh_lock);
1358                 list_for_each_entry_safe(bp, n, &hash->bh_list, b_hash_list) {
1359                         ASSERT(btp == bp->b_target);
1360                         if (!(bp->b_flags & XBF_FS_MANAGED)) {
1361                                 spin_unlock(&hash->bh_lock);
1362                                 /*
1363                                  * Catch superblock reference count leaks
1364                                  * immediately
1365                                  */
1366                                 BUG_ON(bp->b_bn == 0);
1367                                 delay(100);
1368                                 goto again;
1369                         }
1370                 }
1371                 spin_unlock(&hash->bh_lock);
1372         }
1373 }
1374
1375 /*
1376  *      Allocate buffer hash table for a given target.
1377  *      For devices containing metadata (i.e. not the log/realtime devices)
1378  *      we need to allocate a much larger hash table.
1379  */
1380 STATIC void
1381 xfs_alloc_bufhash(
1382         xfs_buftarg_t           *btp,
1383         int                     external)
1384 {
1385         unsigned int            i;
1386
1387         btp->bt_hashshift = external ? 3 : 8;   /* 8 or 256 buckets */
1388         btp->bt_hashmask = (1 << btp->bt_hashshift) - 1;
1389         btp->bt_hash = kmem_zalloc((1 << btp->bt_hashshift) *
1390                                         sizeof(xfs_bufhash_t), KM_SLEEP | KM_LARGE);
1391         for (i = 0; i < (1 << btp->bt_hashshift); i++) {
1392                 spin_lock_init(&btp->bt_hash[i].bh_lock);
1393                 INIT_LIST_HEAD(&btp->bt_hash[i].bh_list);
1394         }
1395 }
1396
1397 STATIC void
1398 xfs_free_bufhash(
1399         xfs_buftarg_t           *btp)
1400 {
1401         kmem_free(btp->bt_hash, (1<<btp->bt_hashshift) * sizeof(xfs_bufhash_t));
1402         btp->bt_hash = NULL;
1403 }
1404
1405 /*
1406  *      buftarg list for delwrite queue processing
1407  */
1408 static LIST_HEAD(xfs_buftarg_list);
1409 static DEFINE_SPINLOCK(xfs_buftarg_lock);
1410
1411 STATIC void
1412 xfs_register_buftarg(
1413         xfs_buftarg_t           *btp)
1414 {
1415         spin_lock(&xfs_buftarg_lock);
1416         list_add(&btp->bt_list, &xfs_buftarg_list);
1417         spin_unlock(&xfs_buftarg_lock);
1418 }
1419
1420 STATIC void
1421 xfs_unregister_buftarg(
1422         xfs_buftarg_t           *btp)
1423 {
1424         spin_lock(&xfs_buftarg_lock);
1425         list_del(&btp->bt_list);
1426         spin_unlock(&xfs_buftarg_lock);
1427 }
1428
1429 void
1430 xfs_free_buftarg(
1431         xfs_buftarg_t           *btp,
1432         int                     external)
1433 {
1434         xfs_flush_buftarg(btp, 1);
1435         xfs_blkdev_issue_flush(btp);
1436         if (external)
1437                 xfs_blkdev_put(btp->bt_bdev);
1438         xfs_free_bufhash(btp);
1439         iput(btp->bt_mapping->host);
1440
1441         /* Unregister the buftarg first so that we don't get a
1442          * wakeup finding a non-existent task
1443          */
1444         xfs_unregister_buftarg(btp);
1445         kthread_stop(btp->bt_task);
1446
1447         kmem_free(btp, sizeof(*btp));
1448 }
1449
1450 STATIC int
1451 xfs_setsize_buftarg_flags(
1452         xfs_buftarg_t           *btp,
1453         unsigned int            blocksize,
1454         unsigned int            sectorsize,
1455         int                     verbose)
1456 {
1457         btp->bt_bsize = blocksize;
1458         btp->bt_sshift = ffs(sectorsize) - 1;
1459         btp->bt_smask = sectorsize - 1;
1460
1461         if (set_blocksize(btp->bt_bdev, sectorsize)) {
1462                 printk(KERN_WARNING
1463                         "XFS: Cannot set_blocksize to %u on device %s\n",
1464                         sectorsize, XFS_BUFTARG_NAME(btp));
1465                 return EINVAL;
1466         }
1467
1468         if (verbose &&
1469             (PAGE_CACHE_SIZE / BITS_PER_LONG) > sectorsize) {
1470                 printk(KERN_WARNING
1471                         "XFS: %u byte sectors in use on device %s.  "
1472                         "This is suboptimal; %u or greater is ideal.\n",
1473                         sectorsize, XFS_BUFTARG_NAME(btp),
1474                         (unsigned int)PAGE_CACHE_SIZE / BITS_PER_LONG);
1475         }
1476
1477         return 0;
1478 }
1479
1480 /*
1481  *      When allocating the initial buffer target we have not yet
1482  *      read in the superblock, so don't know what sized sectors
1483  *      are being used is at this early stage.  Play safe.
1484  */
1485 STATIC int
1486 xfs_setsize_buftarg_early(
1487         xfs_buftarg_t           *btp,
1488         struct block_device     *bdev)
1489 {
1490         return xfs_setsize_buftarg_flags(btp,
1491                         PAGE_CACHE_SIZE, bdev_hardsect_size(bdev), 0);
1492 }
1493
1494 int
1495 xfs_setsize_buftarg(
1496         xfs_buftarg_t           *btp,
1497         unsigned int            blocksize,
1498         unsigned int            sectorsize)
1499 {
1500         return xfs_setsize_buftarg_flags(btp, blocksize, sectorsize, 1);
1501 }
1502
1503 STATIC int
1504 xfs_mapping_buftarg(
1505         xfs_buftarg_t           *btp,
1506         struct block_device     *bdev)
1507 {
1508         struct backing_dev_info *bdi;
1509         struct inode            *inode;
1510         struct address_space    *mapping;
1511         static const struct address_space_operations mapping_aops = {
1512                 .sync_page = block_sync_page,
1513                 .migratepage = fail_migrate_page,
1514         };
1515
1516         inode = new_inode(bdev->bd_inode->i_sb);
1517         if (!inode) {
1518                 printk(KERN_WARNING
1519                         "XFS: Cannot allocate mapping inode for device %s\n",
1520                         XFS_BUFTARG_NAME(btp));
1521                 return ENOMEM;
1522         }
1523         inode->i_mode = S_IFBLK;
1524         inode->i_bdev = bdev;
1525         inode->i_rdev = bdev->bd_dev;
1526         bdi = blk_get_backing_dev_info(bdev);
1527         if (!bdi)
1528                 bdi = &default_backing_dev_info;
1529         mapping = &inode->i_data;
1530         mapping->a_ops = &mapping_aops;
1531         mapping->backing_dev_info = bdi;
1532         mapping_set_gfp_mask(mapping, GFP_NOFS);
1533         btp->bt_mapping = mapping;
1534         return 0;
1535 }
1536
1537 STATIC int
1538 xfs_alloc_delwrite_queue(
1539         xfs_buftarg_t           *btp)
1540 {
1541         int     error = 0;
1542
1543         INIT_LIST_HEAD(&btp->bt_list);
1544         INIT_LIST_HEAD(&btp->bt_delwrite_queue);
1545         spin_lock_init(&btp->bt_delwrite_lock);
1546         btp->bt_flags = 0;
1547         btp->bt_task = kthread_run(xfsbufd, btp, "xfsbufd");
1548         if (IS_ERR(btp->bt_task)) {
1549                 error = PTR_ERR(btp->bt_task);
1550                 goto out_error;
1551         }
1552         xfs_register_buftarg(btp);
1553 out_error:
1554         return error;
1555 }
1556
1557 xfs_buftarg_t *
1558 xfs_alloc_buftarg(
1559         struct block_device     *bdev,
1560         int                     external)
1561 {
1562         xfs_buftarg_t           *btp;
1563
1564         btp = kmem_zalloc(sizeof(*btp), KM_SLEEP);
1565
1566         btp->bt_dev =  bdev->bd_dev;
1567         btp->bt_bdev = bdev;
1568         if (xfs_setsize_buftarg_early(btp, bdev))
1569                 goto error;
1570         if (xfs_mapping_buftarg(btp, bdev))
1571                 goto error;
1572         if (xfs_alloc_delwrite_queue(btp))
1573                 goto error;
1574         xfs_alloc_bufhash(btp, external);
1575         return btp;
1576
1577 error:
1578         kmem_free(btp, sizeof(*btp));
1579         return NULL;
1580 }
1581
1582
1583 /*
1584  *      Delayed write buffer handling
1585  */
1586 STATIC void
1587 xfs_buf_delwri_queue(
1588         xfs_buf_t               *bp,
1589         int                     unlock)
1590 {
1591         struct list_head        *dwq = &bp->b_target->bt_delwrite_queue;
1592         spinlock_t              *dwlk = &bp->b_target->bt_delwrite_lock;
1593
1594         XB_TRACE(bp, "delwri_q", (long)unlock);
1595         ASSERT((bp->b_flags&(XBF_DELWRI|XBF_ASYNC)) == (XBF_DELWRI|XBF_ASYNC));
1596
1597         spin_lock(dwlk);
1598         /* If already in the queue, dequeue and place at tail */
1599         if (!list_empty(&bp->b_list)) {
1600                 ASSERT(bp->b_flags & _XBF_DELWRI_Q);
1601                 if (unlock)
1602                         atomic_dec(&bp->b_hold);
1603                 list_del(&bp->b_list);
1604         }
1605
1606         bp->b_flags |= _XBF_DELWRI_Q;
1607         list_add_tail(&bp->b_list, dwq);
1608         bp->b_queuetime = jiffies;
1609         spin_unlock(dwlk);
1610
1611         if (unlock)
1612                 xfs_buf_unlock(bp);
1613 }
1614
1615 void
1616 xfs_buf_delwri_dequeue(
1617         xfs_buf_t               *bp)
1618 {
1619         spinlock_t              *dwlk = &bp->b_target->bt_delwrite_lock;
1620         int                     dequeued = 0;
1621
1622         spin_lock(dwlk);
1623         if ((bp->b_flags & XBF_DELWRI) && !list_empty(&bp->b_list)) {
1624                 ASSERT(bp->b_flags & _XBF_DELWRI_Q);
1625                 list_del_init(&bp->b_list);
1626                 dequeued = 1;
1627         }
1628         bp->b_flags &= ~(XBF_DELWRI|_XBF_DELWRI_Q);
1629         spin_unlock(dwlk);
1630
1631         if (dequeued)
1632                 xfs_buf_rele(bp);
1633
1634         XB_TRACE(bp, "delwri_dq", (long)dequeued);
1635 }
1636
1637 STATIC void
1638 xfs_buf_runall_queues(
1639         struct workqueue_struct *queue)
1640 {
1641         flush_workqueue(queue);
1642 }
1643
1644 STATIC int
1645 xfsbufd_wakeup(
1646         int                     priority,
1647         gfp_t                   mask)
1648 {
1649         xfs_buftarg_t           *btp;
1650
1651         spin_lock(&xfs_buftarg_lock);
1652         list_for_each_entry(btp, &xfs_buftarg_list, bt_list) {
1653                 if (test_bit(XBT_FORCE_SLEEP, &btp->bt_flags))
1654                         continue;
1655                 set_bit(XBT_FORCE_FLUSH, &btp->bt_flags);
1656                 wake_up_process(btp->bt_task);
1657         }
1658         spin_unlock(&xfs_buftarg_lock);
1659         return 0;
1660 }
1661
1662 /*
1663  * Move as many buffers as specified to the supplied list
1664  * idicating if we skipped any buffers to prevent deadlocks.
1665  */
1666 STATIC int
1667 xfs_buf_delwri_split(
1668         xfs_buftarg_t   *target,
1669         struct list_head *list,
1670         unsigned long   age)
1671 {
1672         xfs_buf_t       *bp, *n;
1673         struct list_head *dwq = &target->bt_delwrite_queue;
1674         spinlock_t      *dwlk = &target->bt_delwrite_lock;
1675         int             skipped = 0;
1676         int             force;
1677
1678         force = test_and_clear_bit(XBT_FORCE_FLUSH, &target->bt_flags);
1679         INIT_LIST_HEAD(list);
1680         spin_lock(dwlk);
1681         list_for_each_entry_safe(bp, n, dwq, b_list) {
1682                 XB_TRACE(bp, "walkq1", (long)xfs_buf_ispin(bp));
1683                 ASSERT(bp->b_flags & XBF_DELWRI);
1684
1685                 if (!xfs_buf_ispin(bp) && !xfs_buf_cond_lock(bp)) {
1686                         if (!force &&
1687                             time_before(jiffies, bp->b_queuetime + age)) {
1688                                 xfs_buf_unlock(bp);
1689                                 break;
1690                         }
1691
1692                         bp->b_flags &= ~(XBF_DELWRI|_XBF_DELWRI_Q|
1693                                          _XBF_RUN_QUEUES);
1694                         bp->b_flags |= XBF_WRITE;
1695                         list_move_tail(&bp->b_list, list);
1696                 } else
1697                         skipped++;
1698         }
1699         spin_unlock(dwlk);
1700
1701         return skipped;
1702
1703 }
1704
1705 STATIC int
1706 xfsbufd(
1707         void            *data)
1708 {
1709         struct list_head tmp;
1710         xfs_buftarg_t   *target = (xfs_buftarg_t *)data;
1711         int             count;
1712         xfs_buf_t       *bp;
1713
1714         current->flags |= PF_MEMALLOC;
1715
1716         set_freezable();
1717
1718         do {
1719                 if (unlikely(freezing(current))) {
1720                         set_bit(XBT_FORCE_SLEEP, &target->bt_flags);
1721                         refrigerator();
1722                 } else {
1723                         clear_bit(XBT_FORCE_SLEEP, &target->bt_flags);
1724                 }
1725
1726                 schedule_timeout_interruptible(
1727                         xfs_buf_timer_centisecs * msecs_to_jiffies(10));
1728
1729                 xfs_buf_delwri_split(target, &tmp,
1730                                 xfs_buf_age_centisecs * msecs_to_jiffies(10));
1731
1732                 count = 0;
1733                 while (!list_empty(&tmp)) {
1734                         bp = list_entry(tmp.next, xfs_buf_t, b_list);
1735                         ASSERT(target == bp->b_target);
1736
1737                         list_del_init(&bp->b_list);
1738                         xfs_buf_iostrategy(bp);
1739                         count++;
1740                 }
1741
1742                 if (as_list_len > 0)
1743                         purge_addresses();
1744                 if (count)
1745                         blk_run_address_space(target->bt_mapping);
1746
1747         } while (!kthread_should_stop());
1748
1749         return 0;
1750 }
1751
1752 /*
1753  *      Go through all incore buffers, and release buffers if they belong to
1754  *      the given device. This is used in filesystem error handling to
1755  *      preserve the consistency of its metadata.
1756  */
1757 int
1758 xfs_flush_buftarg(
1759         xfs_buftarg_t   *target,
1760         int             wait)
1761 {
1762         struct list_head tmp;
1763         xfs_buf_t       *bp, *n;
1764         int             pincount = 0;
1765
1766         xfs_buf_runall_queues(xfsdatad_workqueue);
1767         xfs_buf_runall_queues(xfslogd_workqueue);
1768
1769         set_bit(XBT_FORCE_FLUSH, &target->bt_flags);
1770         pincount = xfs_buf_delwri_split(target, &tmp, 0);
1771
1772         /*
1773          * Dropped the delayed write list lock, now walk the temporary list
1774          */
1775         list_for_each_entry_safe(bp, n, &tmp, b_list) {
1776                 ASSERT(target == bp->b_target);
1777                 if (wait)
1778                         bp->b_flags &= ~XBF_ASYNC;
1779                 else
1780                         list_del_init(&bp->b_list);
1781
1782                 xfs_buf_iostrategy(bp);
1783         }
1784
1785         if (wait)
1786                 blk_run_address_space(target->bt_mapping);
1787
1788         /*
1789          * Remaining list items must be flushed before returning
1790          */
1791         while (!list_empty(&tmp)) {
1792                 bp = list_entry(tmp.next, xfs_buf_t, b_list);
1793
1794                 list_del_init(&bp->b_list);
1795                 xfs_iowait(bp);
1796                 xfs_buf_relse(bp);
1797         }
1798
1799         return pincount;
1800 }
1801
1802 int __init
1803 xfs_buf_init(void)
1804 {
1805 #ifdef XFS_BUF_TRACE
1806         xfs_buf_trace_buf = ktrace_alloc(XFS_BUF_TRACE_SIZE, KM_SLEEP);
1807 #endif
1808
1809         xfs_buf_zone = kmem_zone_init_flags(sizeof(xfs_buf_t), "xfs_buf",
1810                                                 KM_ZONE_HWALIGN, NULL);
1811         if (!xfs_buf_zone)
1812                 goto out_free_trace_buf;
1813
1814         xfslogd_workqueue = create_workqueue("xfslogd");
1815         if (!xfslogd_workqueue)
1816                 goto out_free_buf_zone;
1817
1818         xfsdatad_workqueue = create_workqueue("xfsdatad");
1819         if (!xfsdatad_workqueue)
1820                 goto out_destroy_xfslogd_workqueue;
1821
1822         register_shrinker(&xfs_buf_shake);
1823         return 0;
1824
1825  out_destroy_xfslogd_workqueue:
1826         destroy_workqueue(xfslogd_workqueue);
1827  out_free_buf_zone:
1828         kmem_zone_destroy(xfs_buf_zone);
1829  out_free_trace_buf:
1830 #ifdef XFS_BUF_TRACE
1831         ktrace_free(xfs_buf_trace_buf);
1832 #endif
1833         return -ENOMEM;
1834 }
1835
1836 void
1837 xfs_buf_terminate(void)
1838 {
1839         unregister_shrinker(&xfs_buf_shake);
1840         destroy_workqueue(xfsdatad_workqueue);
1841         destroy_workqueue(xfslogd_workqueue);
1842         kmem_zone_destroy(xfs_buf_zone);
1843 #ifdef XFS_BUF_TRACE
1844         ktrace_free(xfs_buf_trace_buf);
1845 #endif
1846 }
1847
1848 #ifdef CONFIG_KDB_MODULES
1849 struct list_head *
1850 xfs_get_buftarg_list(void)
1851 {
1852         return &xfs_buftarg_list;
1853 }
1854 #endif