xfs: Improve scalability of busy extent tracking
[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/gfp.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 #include <linux/list_sort.h>
37
38 #include "xfs_sb.h"
39 #include "xfs_inum.h"
40 #include "xfs_log.h"
41 #include "xfs_ag.h"
42 #include "xfs_dmapi.h"
43 #include "xfs_mount.h"
44 #include "xfs_trace.h"
45
46 static kmem_zone_t *xfs_buf_zone;
47 STATIC int xfsbufd(void *);
48 STATIC int xfsbufd_wakeup(int, gfp_t);
49 STATIC void xfs_buf_delwri_queue(xfs_buf_t *, int);
50 static struct shrinker xfs_buf_shake = {
51         .shrink = xfsbufd_wakeup,
52         .seeks = DEFAULT_SEEKS,
53 };
54
55 static struct workqueue_struct *xfslogd_workqueue;
56 struct workqueue_struct *xfsdatad_workqueue;
57 struct workqueue_struct *xfsconvertd_workqueue;
58
59 #ifdef XFS_BUF_LOCK_TRACKING
60 # define XB_SET_OWNER(bp)       ((bp)->b_last_holder = current->pid)
61 # define XB_CLEAR_OWNER(bp)     ((bp)->b_last_holder = -1)
62 # define XB_GET_OWNER(bp)       ((bp)->b_last_holder)
63 #else
64 # define XB_SET_OWNER(bp)       do { } while (0)
65 # define XB_CLEAR_OWNER(bp)     do { } while (0)
66 # define XB_GET_OWNER(bp)       do { } while (0)
67 #endif
68
69 #define xb_to_gfp(flags) \
70         ((((flags) & XBF_READ_AHEAD) ? __GFP_NORETRY : \
71           ((flags) & XBF_DONT_BLOCK) ? GFP_NOFS : GFP_KERNEL) | __GFP_NOWARN)
72
73 #define xb_to_km(flags) \
74          (((flags) & XBF_DONT_BLOCK) ? KM_NOFS : KM_SLEEP)
75
76 #define xfs_buf_allocate(flags) \
77         kmem_zone_alloc(xfs_buf_zone, xb_to_km(flags))
78 #define xfs_buf_deallocate(bp) \
79         kmem_zone_free(xfs_buf_zone, (bp));
80
81 static inline int
82 xfs_buf_is_vmapped(
83         struct xfs_buf  *bp)
84 {
85         /*
86          * Return true if the buffer is vmapped.
87          *
88          * The XBF_MAPPED flag is set if the buffer should be mapped, but the
89          * code is clever enough to know it doesn't have to map a single page,
90          * so the check has to be both for XBF_MAPPED and bp->b_page_count > 1.
91          */
92         return (bp->b_flags & XBF_MAPPED) && bp->b_page_count > 1;
93 }
94
95 static inline int
96 xfs_buf_vmap_len(
97         struct xfs_buf  *bp)
98 {
99         return (bp->b_page_count * PAGE_SIZE) - bp->b_offset;
100 }
101
102 /*
103  *      Page Region interfaces.
104  *
105  *      For pages in filesystems where the blocksize is smaller than the
106  *      pagesize, we use the page->private field (long) to hold a bitmap
107  *      of uptodate regions within the page.
108  *
109  *      Each such region is "bytes per page / bits per long" bytes long.
110  *
111  *      NBPPR == number-of-bytes-per-page-region
112  *      BTOPR == bytes-to-page-region (rounded up)
113  *      BTOPRT == bytes-to-page-region-truncated (rounded down)
114  */
115 #if (BITS_PER_LONG == 32)
116 #define PRSHIFT         (PAGE_CACHE_SHIFT - 5)  /* (32 == 1<<5) */
117 #elif (BITS_PER_LONG == 64)
118 #define PRSHIFT         (PAGE_CACHE_SHIFT - 6)  /* (64 == 1<<6) */
119 #else
120 #error BITS_PER_LONG must be 32 or 64
121 #endif
122 #define NBPPR           (PAGE_CACHE_SIZE/BITS_PER_LONG)
123 #define BTOPR(b)        (((unsigned int)(b) + (NBPPR - 1)) >> PRSHIFT)
124 #define BTOPRT(b)       (((unsigned int)(b) >> PRSHIFT))
125
126 STATIC unsigned long
127 page_region_mask(
128         size_t          offset,
129         size_t          length)
130 {
131         unsigned long   mask;
132         int             first, final;
133
134         first = BTOPR(offset);
135         final = BTOPRT(offset + length - 1);
136         first = min(first, final);
137
138         mask = ~0UL;
139         mask <<= BITS_PER_LONG - (final - first);
140         mask >>= BITS_PER_LONG - (final);
141
142         ASSERT(offset + length <= PAGE_CACHE_SIZE);
143         ASSERT((final - first) < BITS_PER_LONG && (final - first) >= 0);
144
145         return mask;
146 }
147
148 STATIC void
149 set_page_region(
150         struct page     *page,
151         size_t          offset,
152         size_t          length)
153 {
154         set_page_private(page,
155                 page_private(page) | page_region_mask(offset, length));
156         if (page_private(page) == ~0UL)
157                 SetPageUptodate(page);
158 }
159
160 STATIC int
161 test_page_region(
162         struct page     *page,
163         size_t          offset,
164         size_t          length)
165 {
166         unsigned long   mask = page_region_mask(offset, length);
167
168         return (mask && (page_private(page) & mask) == mask);
169 }
170
171 /*
172  *      Internal xfs_buf_t object manipulation
173  */
174
175 STATIC void
176 _xfs_buf_initialize(
177         xfs_buf_t               *bp,
178         xfs_buftarg_t           *target,
179         xfs_off_t               range_base,
180         size_t                  range_length,
181         xfs_buf_flags_t         flags)
182 {
183         /*
184          * We don't want certain flags to appear in b_flags.
185          */
186         flags &= ~(XBF_LOCK|XBF_MAPPED|XBF_DONT_BLOCK|XBF_READ_AHEAD);
187
188         memset(bp, 0, sizeof(xfs_buf_t));
189         atomic_set(&bp->b_hold, 1);
190         init_completion(&bp->b_iowait);
191         INIT_LIST_HEAD(&bp->b_list);
192         INIT_LIST_HEAD(&bp->b_hash_list);
193         init_MUTEX_LOCKED(&bp->b_sema); /* held, no waiters */
194         XB_SET_OWNER(bp);
195         bp->b_target = target;
196         bp->b_file_offset = range_base;
197         /*
198          * Set buffer_length and count_desired to the same value initially.
199          * I/O routines should use count_desired, which will be the same in
200          * most cases but may be reset (e.g. XFS recovery).
201          */
202         bp->b_buffer_length = bp->b_count_desired = range_length;
203         bp->b_flags = flags;
204         bp->b_bn = XFS_BUF_DADDR_NULL;
205         atomic_set(&bp->b_pin_count, 0);
206         init_waitqueue_head(&bp->b_waiters);
207
208         XFS_STATS_INC(xb_create);
209
210         trace_xfs_buf_init(bp, _RET_IP_);
211 }
212
213 /*
214  *      Allocate a page array capable of holding a specified number
215  *      of pages, and point the page buf at it.
216  */
217 STATIC int
218 _xfs_buf_get_pages(
219         xfs_buf_t               *bp,
220         int                     page_count,
221         xfs_buf_flags_t         flags)
222 {
223         /* Make sure that we have a page list */
224         if (bp->b_pages == NULL) {
225                 bp->b_offset = xfs_buf_poff(bp->b_file_offset);
226                 bp->b_page_count = page_count;
227                 if (page_count <= XB_PAGES) {
228                         bp->b_pages = bp->b_page_array;
229                 } else {
230                         bp->b_pages = kmem_alloc(sizeof(struct page *) *
231                                         page_count, xb_to_km(flags));
232                         if (bp->b_pages == NULL)
233                                 return -ENOMEM;
234                 }
235                 memset(bp->b_pages, 0, sizeof(struct page *) * page_count);
236         }
237         return 0;
238 }
239
240 /*
241  *      Frees b_pages if it was allocated.
242  */
243 STATIC void
244 _xfs_buf_free_pages(
245         xfs_buf_t       *bp)
246 {
247         if (bp->b_pages != bp->b_page_array) {
248                 kmem_free(bp->b_pages);
249                 bp->b_pages = NULL;
250         }
251 }
252
253 /*
254  *      Releases the specified buffer.
255  *
256  *      The modification state of any associated pages is left unchanged.
257  *      The buffer most not be on any hash - use xfs_buf_rele instead for
258  *      hashed and refcounted buffers
259  */
260 void
261 xfs_buf_free(
262         xfs_buf_t               *bp)
263 {
264         trace_xfs_buf_free(bp, _RET_IP_);
265
266         ASSERT(list_empty(&bp->b_hash_list));
267
268         if (bp->b_flags & (_XBF_PAGE_CACHE|_XBF_PAGES)) {
269                 uint            i;
270
271                 if (xfs_buf_is_vmapped(bp))
272                         vm_unmap_ram(bp->b_addr - bp->b_offset,
273                                         bp->b_page_count);
274
275                 for (i = 0; i < bp->b_page_count; i++) {
276                         struct page     *page = bp->b_pages[i];
277
278                         if (bp->b_flags & _XBF_PAGE_CACHE)
279                                 ASSERT(!PagePrivate(page));
280                         page_cache_release(page);
281                 }
282         }
283         _xfs_buf_free_pages(bp);
284         xfs_buf_deallocate(bp);
285 }
286
287 /*
288  *      Finds all pages for buffer in question and builds it's page list.
289  */
290 STATIC int
291 _xfs_buf_lookup_pages(
292         xfs_buf_t               *bp,
293         uint                    flags)
294 {
295         struct address_space    *mapping = bp->b_target->bt_mapping;
296         size_t                  blocksize = bp->b_target->bt_bsize;
297         size_t                  size = bp->b_count_desired;
298         size_t                  nbytes, offset;
299         gfp_t                   gfp_mask = xb_to_gfp(flags);
300         unsigned short          page_count, i;
301         pgoff_t                 first;
302         xfs_off_t               end;
303         int                     error;
304
305         end = bp->b_file_offset + bp->b_buffer_length;
306         page_count = xfs_buf_btoc(end) - xfs_buf_btoct(bp->b_file_offset);
307
308         error = _xfs_buf_get_pages(bp, page_count, flags);
309         if (unlikely(error))
310                 return error;
311         bp->b_flags |= _XBF_PAGE_CACHE;
312
313         offset = bp->b_offset;
314         first = bp->b_file_offset >> PAGE_CACHE_SHIFT;
315
316         for (i = 0; i < bp->b_page_count; i++) {
317                 struct page     *page;
318                 uint            retries = 0;
319
320               retry:
321                 page = find_or_create_page(mapping, first + i, gfp_mask);
322                 if (unlikely(page == NULL)) {
323                         if (flags & XBF_READ_AHEAD) {
324                                 bp->b_page_count = i;
325                                 for (i = 0; i < bp->b_page_count; i++)
326                                         unlock_page(bp->b_pages[i]);
327                                 return -ENOMEM;
328                         }
329
330                         /*
331                          * This could deadlock.
332                          *
333                          * But until all the XFS lowlevel code is revamped to
334                          * handle buffer allocation failures we can't do much.
335                          */
336                         if (!(++retries % 100))
337                                 printk(KERN_ERR
338                                         "XFS: possible memory allocation "
339                                         "deadlock in %s (mode:0x%x)\n",
340                                         __func__, gfp_mask);
341
342                         XFS_STATS_INC(xb_page_retries);
343                         xfsbufd_wakeup(0, gfp_mask);
344                         congestion_wait(BLK_RW_ASYNC, HZ/50);
345                         goto retry;
346                 }
347
348                 XFS_STATS_INC(xb_page_found);
349
350                 nbytes = min_t(size_t, size, PAGE_CACHE_SIZE - offset);
351                 size -= nbytes;
352
353                 ASSERT(!PagePrivate(page));
354                 if (!PageUptodate(page)) {
355                         page_count--;
356                         if (blocksize >= PAGE_CACHE_SIZE) {
357                                 if (flags & XBF_READ)
358                                         bp->b_flags |= _XBF_PAGE_LOCKED;
359                         } else if (!PagePrivate(page)) {
360                                 if (test_page_region(page, offset, nbytes))
361                                         page_count++;
362                         }
363                 }
364
365                 bp->b_pages[i] = page;
366                 offset = 0;
367         }
368
369         if (!(bp->b_flags & _XBF_PAGE_LOCKED)) {
370                 for (i = 0; i < bp->b_page_count; i++)
371                         unlock_page(bp->b_pages[i]);
372         }
373
374         if (page_count == bp->b_page_count)
375                 bp->b_flags |= XBF_DONE;
376
377         return error;
378 }
379
380 /*
381  *      Map buffer into kernel address-space if nessecary.
382  */
383 STATIC int
384 _xfs_buf_map_pages(
385         xfs_buf_t               *bp,
386         uint                    flags)
387 {
388         /* A single page buffer is always mappable */
389         if (bp->b_page_count == 1) {
390                 bp->b_addr = page_address(bp->b_pages[0]) + bp->b_offset;
391                 bp->b_flags |= XBF_MAPPED;
392         } else if (flags & XBF_MAPPED) {
393                 bp->b_addr = vm_map_ram(bp->b_pages, bp->b_page_count,
394                                         -1, PAGE_KERNEL);
395                 if (unlikely(bp->b_addr == NULL))
396                         return -ENOMEM;
397                 bp->b_addr += bp->b_offset;
398                 bp->b_flags |= XBF_MAPPED;
399         }
400
401         return 0;
402 }
403
404 /*
405  *      Finding and Reading Buffers
406  */
407
408 /*
409  *      Look up, and creates if absent, a lockable buffer for
410  *      a given range of an inode.  The buffer is returned
411  *      locked.  If other overlapping buffers exist, they are
412  *      released before the new buffer is created and locked,
413  *      which may imply that this call will block until those buffers
414  *      are unlocked.  No I/O is implied by this call.
415  */
416 xfs_buf_t *
417 _xfs_buf_find(
418         xfs_buftarg_t           *btp,   /* block device target          */
419         xfs_off_t               ioff,   /* starting offset of range     */
420         size_t                  isize,  /* length of range              */
421         xfs_buf_flags_t         flags,
422         xfs_buf_t               *new_bp)
423 {
424         xfs_off_t               range_base;
425         size_t                  range_length;
426         xfs_bufhash_t           *hash;
427         xfs_buf_t               *bp, *n;
428
429         range_base = (ioff << BBSHIFT);
430         range_length = (isize << BBSHIFT);
431
432         /* Check for IOs smaller than the sector size / not sector aligned */
433         ASSERT(!(range_length < (1 << btp->bt_sshift)));
434         ASSERT(!(range_base & (xfs_off_t)btp->bt_smask));
435
436         hash = &btp->bt_hash[hash_long((unsigned long)ioff, btp->bt_hashshift)];
437
438         spin_lock(&hash->bh_lock);
439
440         list_for_each_entry_safe(bp, n, &hash->bh_list, b_hash_list) {
441                 ASSERT(btp == bp->b_target);
442                 if (bp->b_file_offset == range_base &&
443                     bp->b_buffer_length == range_length) {
444                         /*
445                          * If we look at something, bring it to the
446                          * front of the list for next time.
447                          */
448                         atomic_inc(&bp->b_hold);
449                         list_move(&bp->b_hash_list, &hash->bh_list);
450                         goto found;
451                 }
452         }
453
454         /* No match found */
455         if (new_bp) {
456                 _xfs_buf_initialize(new_bp, btp, range_base,
457                                 range_length, flags);
458                 new_bp->b_hash = hash;
459                 list_add(&new_bp->b_hash_list, &hash->bh_list);
460         } else {
461                 XFS_STATS_INC(xb_miss_locked);
462         }
463
464         spin_unlock(&hash->bh_lock);
465         return new_bp;
466
467 found:
468         spin_unlock(&hash->bh_lock);
469
470         /* Attempt to get the semaphore without sleeping,
471          * if this does not work then we need to drop the
472          * spinlock and do a hard attempt on the semaphore.
473          */
474         if (down_trylock(&bp->b_sema)) {
475                 if (!(flags & XBF_TRYLOCK)) {
476                         /* wait for buffer ownership */
477                         xfs_buf_lock(bp);
478                         XFS_STATS_INC(xb_get_locked_waited);
479                 } else {
480                         /* We asked for a trylock and failed, no need
481                          * to look at file offset and length here, we
482                          * know that this buffer at least overlaps our
483                          * buffer and is locked, therefore our buffer
484                          * either does not exist, or is this buffer.
485                          */
486                         xfs_buf_rele(bp);
487                         XFS_STATS_INC(xb_busy_locked);
488                         return NULL;
489                 }
490         } else {
491                 /* trylock worked */
492                 XB_SET_OWNER(bp);
493         }
494
495         if (bp->b_flags & XBF_STALE) {
496                 ASSERT((bp->b_flags & _XBF_DELWRI_Q) == 0);
497                 bp->b_flags &= XBF_MAPPED;
498         }
499
500         trace_xfs_buf_find(bp, flags, _RET_IP_);
501         XFS_STATS_INC(xb_get_locked);
502         return bp;
503 }
504
505 /*
506  *      Assembles a buffer covering the specified range.
507  *      Storage in memory for all portions of the buffer will be allocated,
508  *      although backing storage may not be.
509  */
510 xfs_buf_t *
511 xfs_buf_get(
512         xfs_buftarg_t           *target,/* target for buffer            */
513         xfs_off_t               ioff,   /* starting offset of range     */
514         size_t                  isize,  /* length of range              */
515         xfs_buf_flags_t         flags)
516 {
517         xfs_buf_t               *bp, *new_bp;
518         int                     error = 0, i;
519
520         new_bp = xfs_buf_allocate(flags);
521         if (unlikely(!new_bp))
522                 return NULL;
523
524         bp = _xfs_buf_find(target, ioff, isize, flags, new_bp);
525         if (bp == new_bp) {
526                 error = _xfs_buf_lookup_pages(bp, flags);
527                 if (error)
528                         goto no_buffer;
529         } else {
530                 xfs_buf_deallocate(new_bp);
531                 if (unlikely(bp == NULL))
532                         return NULL;
533         }
534
535         for (i = 0; i < bp->b_page_count; i++)
536                 mark_page_accessed(bp->b_pages[i]);
537
538         if (!(bp->b_flags & XBF_MAPPED)) {
539                 error = _xfs_buf_map_pages(bp, flags);
540                 if (unlikely(error)) {
541                         printk(KERN_WARNING "%s: failed to map pages\n",
542                                         __func__);
543                         goto no_buffer;
544                 }
545         }
546
547         XFS_STATS_INC(xb_get);
548
549         /*
550          * Always fill in the block number now, the mapped cases can do
551          * their own overlay of this later.
552          */
553         bp->b_bn = ioff;
554         bp->b_count_desired = bp->b_buffer_length;
555
556         trace_xfs_buf_get(bp, flags, _RET_IP_);
557         return bp;
558
559  no_buffer:
560         if (flags & (XBF_LOCK | XBF_TRYLOCK))
561                 xfs_buf_unlock(bp);
562         xfs_buf_rele(bp);
563         return NULL;
564 }
565
566 STATIC int
567 _xfs_buf_read(
568         xfs_buf_t               *bp,
569         xfs_buf_flags_t         flags)
570 {
571         int                     status;
572
573         ASSERT(!(flags & (XBF_DELWRI|XBF_WRITE)));
574         ASSERT(bp->b_bn != XFS_BUF_DADDR_NULL);
575
576         bp->b_flags &= ~(XBF_WRITE | XBF_ASYNC | XBF_DELWRI | \
577                         XBF_READ_AHEAD | _XBF_RUN_QUEUES);
578         bp->b_flags |= flags & (XBF_READ | XBF_ASYNC | \
579                         XBF_READ_AHEAD | _XBF_RUN_QUEUES);
580
581         status = xfs_buf_iorequest(bp);
582         if (!status && !(flags & XBF_ASYNC))
583                 status = xfs_buf_iowait(bp);
584         return status;
585 }
586
587 xfs_buf_t *
588 xfs_buf_read(
589         xfs_buftarg_t           *target,
590         xfs_off_t               ioff,
591         size_t                  isize,
592         xfs_buf_flags_t         flags)
593 {
594         xfs_buf_t               *bp;
595
596         flags |= XBF_READ;
597
598         bp = xfs_buf_get(target, ioff, isize, flags);
599         if (bp) {
600                 trace_xfs_buf_read(bp, flags, _RET_IP_);
601
602                 if (!XFS_BUF_ISDONE(bp)) {
603                         XFS_STATS_INC(xb_get_read);
604                         _xfs_buf_read(bp, flags);
605                 } else if (flags & XBF_ASYNC) {
606                         /*
607                          * Read ahead call which is already satisfied,
608                          * drop the buffer
609                          */
610                         goto no_buffer;
611                 } else {
612                         /* We do not want read in the flags */
613                         bp->b_flags &= ~XBF_READ;
614                 }
615         }
616
617         return bp;
618
619  no_buffer:
620         if (flags & (XBF_LOCK | XBF_TRYLOCK))
621                 xfs_buf_unlock(bp);
622         xfs_buf_rele(bp);
623         return NULL;
624 }
625
626 /*
627  *      If we are not low on memory then do the readahead in a deadlock
628  *      safe manner.
629  */
630 void
631 xfs_buf_readahead(
632         xfs_buftarg_t           *target,
633         xfs_off_t               ioff,
634         size_t                  isize,
635         xfs_buf_flags_t         flags)
636 {
637         struct backing_dev_info *bdi;
638
639         bdi = target->bt_mapping->backing_dev_info;
640         if (bdi_read_congested(bdi))
641                 return;
642
643         flags |= (XBF_TRYLOCK|XBF_ASYNC|XBF_READ_AHEAD);
644         xfs_buf_read(target, ioff, isize, flags);
645 }
646
647 xfs_buf_t *
648 xfs_buf_get_empty(
649         size_t                  len,
650         xfs_buftarg_t           *target)
651 {
652         xfs_buf_t               *bp;
653
654         bp = xfs_buf_allocate(0);
655         if (bp)
656                 _xfs_buf_initialize(bp, target, 0, len, 0);
657         return bp;
658 }
659
660 static inline struct page *
661 mem_to_page(
662         void                    *addr)
663 {
664         if ((!is_vmalloc_addr(addr))) {
665                 return virt_to_page(addr);
666         } else {
667                 return vmalloc_to_page(addr);
668         }
669 }
670
671 int
672 xfs_buf_associate_memory(
673         xfs_buf_t               *bp,
674         void                    *mem,
675         size_t                  len)
676 {
677         int                     rval;
678         int                     i = 0;
679         unsigned long           pageaddr;
680         unsigned long           offset;
681         size_t                  buflen;
682         int                     page_count;
683
684         pageaddr = (unsigned long)mem & PAGE_CACHE_MASK;
685         offset = (unsigned long)mem - pageaddr;
686         buflen = PAGE_CACHE_ALIGN(len + offset);
687         page_count = buflen >> PAGE_CACHE_SHIFT;
688
689         /* Free any previous set of page pointers */
690         if (bp->b_pages)
691                 _xfs_buf_free_pages(bp);
692
693         bp->b_pages = NULL;
694         bp->b_addr = mem;
695
696         rval = _xfs_buf_get_pages(bp, page_count, XBF_DONT_BLOCK);
697         if (rval)
698                 return rval;
699
700         bp->b_offset = offset;
701
702         for (i = 0; i < bp->b_page_count; i++) {
703                 bp->b_pages[i] = mem_to_page((void *)pageaddr);
704                 pageaddr += PAGE_CACHE_SIZE;
705         }
706
707         bp->b_count_desired = len;
708         bp->b_buffer_length = buflen;
709         bp->b_flags |= XBF_MAPPED;
710         bp->b_flags &= ~_XBF_PAGE_LOCKED;
711
712         return 0;
713 }
714
715 xfs_buf_t *
716 xfs_buf_get_noaddr(
717         size_t                  len,
718         xfs_buftarg_t           *target)
719 {
720         unsigned long           page_count = PAGE_ALIGN(len) >> PAGE_SHIFT;
721         int                     error, i;
722         xfs_buf_t               *bp;
723
724         bp = xfs_buf_allocate(0);
725         if (unlikely(bp == NULL))
726                 goto fail;
727         _xfs_buf_initialize(bp, target, 0, len, 0);
728
729         error = _xfs_buf_get_pages(bp, page_count, 0);
730         if (error)
731                 goto fail_free_buf;
732
733         for (i = 0; i < page_count; i++) {
734                 bp->b_pages[i] = alloc_page(GFP_KERNEL);
735                 if (!bp->b_pages[i])
736                         goto fail_free_mem;
737         }
738         bp->b_flags |= _XBF_PAGES;
739
740         error = _xfs_buf_map_pages(bp, XBF_MAPPED);
741         if (unlikely(error)) {
742                 printk(KERN_WARNING "%s: failed to map pages\n",
743                                 __func__);
744                 goto fail_free_mem;
745         }
746
747         xfs_buf_unlock(bp);
748
749         trace_xfs_buf_get_noaddr(bp, _RET_IP_);
750         return bp;
751
752  fail_free_mem:
753         while (--i >= 0)
754                 __free_page(bp->b_pages[i]);
755         _xfs_buf_free_pages(bp);
756  fail_free_buf:
757         xfs_buf_deallocate(bp);
758  fail:
759         return NULL;
760 }
761
762 /*
763  *      Increment reference count on buffer, to hold the buffer concurrently
764  *      with another thread which may release (free) the buffer asynchronously.
765  *      Must hold the buffer already to call this function.
766  */
767 void
768 xfs_buf_hold(
769         xfs_buf_t               *bp)
770 {
771         trace_xfs_buf_hold(bp, _RET_IP_);
772         atomic_inc(&bp->b_hold);
773 }
774
775 /*
776  *      Releases a hold on the specified buffer.  If the
777  *      the hold count is 1, calls xfs_buf_free.
778  */
779 void
780 xfs_buf_rele(
781         xfs_buf_t               *bp)
782 {
783         xfs_bufhash_t           *hash = bp->b_hash;
784
785         trace_xfs_buf_rele(bp, _RET_IP_);
786
787         if (unlikely(!hash)) {
788                 ASSERT(!bp->b_relse);
789                 if (atomic_dec_and_test(&bp->b_hold))
790                         xfs_buf_free(bp);
791                 return;
792         }
793
794         ASSERT(atomic_read(&bp->b_hold) > 0);
795         if (atomic_dec_and_lock(&bp->b_hold, &hash->bh_lock)) {
796                 if (bp->b_relse) {
797                         atomic_inc(&bp->b_hold);
798                         spin_unlock(&hash->bh_lock);
799                         (*(bp->b_relse)) (bp);
800                 } else if (bp->b_flags & XBF_FS_MANAGED) {
801                         spin_unlock(&hash->bh_lock);
802                 } else {
803                         ASSERT(!(bp->b_flags & (XBF_DELWRI|_XBF_DELWRI_Q)));
804                         list_del_init(&bp->b_hash_list);
805                         spin_unlock(&hash->bh_lock);
806                         xfs_buf_free(bp);
807                 }
808         }
809 }
810
811
812 /*
813  *      Mutual exclusion on buffers.  Locking model:
814  *
815  *      Buffers associated with inodes for which buffer locking
816  *      is not enabled are not protected by semaphores, and are
817  *      assumed to be exclusively owned by the caller.  There is a
818  *      spinlock in the buffer, used by the caller when concurrent
819  *      access is possible.
820  */
821
822 /*
823  *      Locks a buffer object, if it is not already locked.
824  *      Note that this in no way locks the underlying pages, so it is only
825  *      useful for synchronizing concurrent use of buffer objects, not for
826  *      synchronizing independent access to the underlying pages.
827  */
828 int
829 xfs_buf_cond_lock(
830         xfs_buf_t               *bp)
831 {
832         int                     locked;
833
834         locked = down_trylock(&bp->b_sema) == 0;
835         if (locked)
836                 XB_SET_OWNER(bp);
837
838         trace_xfs_buf_cond_lock(bp, _RET_IP_);
839         return locked ? 0 : -EBUSY;
840 }
841
842 int
843 xfs_buf_lock_value(
844         xfs_buf_t               *bp)
845 {
846         return bp->b_sema.count;
847 }
848
849 /*
850  *      Locks a buffer object.
851  *      Note that this in no way locks the underlying pages, so it is only
852  *      useful for synchronizing concurrent use of buffer objects, not for
853  *      synchronizing independent access to the underlying pages.
854  *
855  *      If we come across a stale, pinned, locked buffer, we know that we
856  *      are being asked to lock a buffer that has been reallocated. Because
857  *      it is pinned, we know that the log has not been pushed to disk and
858  *      hence it will still be locked. Rather than sleeping until someone
859  *      else pushes the log, push it ourselves before trying to get the lock.
860  */
861 void
862 xfs_buf_lock(
863         xfs_buf_t               *bp)
864 {
865         trace_xfs_buf_lock(bp, _RET_IP_);
866
867         if (atomic_read(&bp->b_pin_count) && (bp->b_flags & XBF_STALE))
868                 xfs_log_force(bp->b_mount, 0);
869         if (atomic_read(&bp->b_io_remaining))
870                 blk_run_address_space(bp->b_target->bt_mapping);
871         down(&bp->b_sema);
872         XB_SET_OWNER(bp);
873
874         trace_xfs_buf_lock_done(bp, _RET_IP_);
875 }
876
877 /*
878  *      Releases the lock on the buffer object.
879  *      If the buffer is marked delwri but is not queued, do so before we
880  *      unlock the buffer as we need to set flags correctly.  We also need to
881  *      take a reference for the delwri queue because the unlocker is going to
882  *      drop their's and they don't know we just queued it.
883  */
884 void
885 xfs_buf_unlock(
886         xfs_buf_t               *bp)
887 {
888         if ((bp->b_flags & (XBF_DELWRI|_XBF_DELWRI_Q)) == XBF_DELWRI) {
889                 atomic_inc(&bp->b_hold);
890                 bp->b_flags |= XBF_ASYNC;
891                 xfs_buf_delwri_queue(bp, 0);
892         }
893
894         XB_CLEAR_OWNER(bp);
895         up(&bp->b_sema);
896
897         trace_xfs_buf_unlock(bp, _RET_IP_);
898 }
899
900
901 /*
902  *      Pinning Buffer Storage in Memory
903  *      Ensure that no attempt to force a buffer to disk will succeed.
904  */
905 void
906 xfs_buf_pin(
907         xfs_buf_t               *bp)
908 {
909         trace_xfs_buf_pin(bp, _RET_IP_);
910         atomic_inc(&bp->b_pin_count);
911 }
912
913 void
914 xfs_buf_unpin(
915         xfs_buf_t               *bp)
916 {
917         trace_xfs_buf_unpin(bp, _RET_IP_);
918
919         if (atomic_dec_and_test(&bp->b_pin_count))
920                 wake_up_all(&bp->b_waiters);
921 }
922
923 int
924 xfs_buf_ispin(
925         xfs_buf_t               *bp)
926 {
927         return atomic_read(&bp->b_pin_count);
928 }
929
930 STATIC void
931 xfs_buf_wait_unpin(
932         xfs_buf_t               *bp)
933 {
934         DECLARE_WAITQUEUE       (wait, current);
935
936         if (atomic_read(&bp->b_pin_count) == 0)
937                 return;
938
939         add_wait_queue(&bp->b_waiters, &wait);
940         for (;;) {
941                 set_current_state(TASK_UNINTERRUPTIBLE);
942                 if (atomic_read(&bp->b_pin_count) == 0)
943                         break;
944                 if (atomic_read(&bp->b_io_remaining))
945                         blk_run_address_space(bp->b_target->bt_mapping);
946                 schedule();
947         }
948         remove_wait_queue(&bp->b_waiters, &wait);
949         set_current_state(TASK_RUNNING);
950 }
951
952 /*
953  *      Buffer Utility Routines
954  */
955
956 STATIC void
957 xfs_buf_iodone_work(
958         struct work_struct      *work)
959 {
960         xfs_buf_t               *bp =
961                 container_of(work, xfs_buf_t, b_iodone_work);
962
963         /*
964          * We can get an EOPNOTSUPP to ordered writes.  Here we clear the
965          * ordered flag and reissue them.  Because we can't tell the higher
966          * layers directly that they should not issue ordered I/O anymore, they
967          * need to check if the _XFS_BARRIER_FAILED flag was set during I/O completion.
968          */
969         if ((bp->b_error == EOPNOTSUPP) &&
970             (bp->b_flags & (XBF_ORDERED|XBF_ASYNC)) == (XBF_ORDERED|XBF_ASYNC)) {
971                 trace_xfs_buf_ordered_retry(bp, _RET_IP_);
972                 bp->b_flags &= ~XBF_ORDERED;
973                 bp->b_flags |= _XFS_BARRIER_FAILED;
974                 xfs_buf_iorequest(bp);
975         } else if (bp->b_iodone)
976                 (*(bp->b_iodone))(bp);
977         else if (bp->b_flags & XBF_ASYNC)
978                 xfs_buf_relse(bp);
979 }
980
981 void
982 xfs_buf_ioend(
983         xfs_buf_t               *bp,
984         int                     schedule)
985 {
986         trace_xfs_buf_iodone(bp, _RET_IP_);
987
988         bp->b_flags &= ~(XBF_READ | XBF_WRITE | XBF_READ_AHEAD);
989         if (bp->b_error == 0)
990                 bp->b_flags |= XBF_DONE;
991
992         if ((bp->b_iodone) || (bp->b_flags & XBF_ASYNC)) {
993                 if (schedule) {
994                         INIT_WORK(&bp->b_iodone_work, xfs_buf_iodone_work);
995                         queue_work(xfslogd_workqueue, &bp->b_iodone_work);
996                 } else {
997                         xfs_buf_iodone_work(&bp->b_iodone_work);
998                 }
999         } else {
1000                 complete(&bp->b_iowait);
1001         }
1002 }
1003
1004 void
1005 xfs_buf_ioerror(
1006         xfs_buf_t               *bp,
1007         int                     error)
1008 {
1009         ASSERT(error >= 0 && error <= 0xffff);
1010         bp->b_error = (unsigned short)error;
1011         trace_xfs_buf_ioerror(bp, error, _RET_IP_);
1012 }
1013
1014 int
1015 xfs_bwrite(
1016         struct xfs_mount        *mp,
1017         struct xfs_buf          *bp)
1018 {
1019         int                     error;
1020
1021         bp->b_strat = xfs_bdstrat_cb;
1022         bp->b_mount = mp;
1023         bp->b_flags |= XBF_WRITE;
1024         bp->b_flags &= ~(XBF_ASYNC | XBF_READ);
1025
1026         xfs_buf_delwri_dequeue(bp);
1027         xfs_buf_iostrategy(bp);
1028
1029         error = xfs_buf_iowait(bp);
1030         if (error)
1031                 xfs_force_shutdown(mp, SHUTDOWN_META_IO_ERROR);
1032         xfs_buf_relse(bp);
1033         return error;
1034 }
1035
1036 void
1037 xfs_bdwrite(
1038         void                    *mp,
1039         struct xfs_buf          *bp)
1040 {
1041         trace_xfs_buf_bdwrite(bp, _RET_IP_);
1042
1043         bp->b_strat = xfs_bdstrat_cb;
1044         bp->b_mount = mp;
1045
1046         bp->b_flags &= ~XBF_READ;
1047         bp->b_flags |= (XBF_DELWRI | XBF_ASYNC);
1048
1049         xfs_buf_delwri_queue(bp, 1);
1050 }
1051
1052 /*
1053  * Called when we want to stop a buffer from getting written or read.
1054  * We attach the EIO error, muck with its flags, and call biodone
1055  * so that the proper iodone callbacks get called.
1056  */
1057 STATIC int
1058 xfs_bioerror(
1059         xfs_buf_t *bp)
1060 {
1061 #ifdef XFSERRORDEBUG
1062         ASSERT(XFS_BUF_ISREAD(bp) || bp->b_iodone);
1063 #endif
1064
1065         /*
1066          * No need to wait until the buffer is unpinned, we aren't flushing it.
1067          */
1068         XFS_BUF_ERROR(bp, EIO);
1069
1070         /*
1071          * We're calling biodone, so delete XBF_DONE flag.
1072          */
1073         XFS_BUF_UNREAD(bp);
1074         XFS_BUF_UNDELAYWRITE(bp);
1075         XFS_BUF_UNDONE(bp);
1076         XFS_BUF_STALE(bp);
1077
1078         XFS_BUF_CLR_BDSTRAT_FUNC(bp);
1079         xfs_biodone(bp);
1080
1081         return EIO;
1082 }
1083
1084 /*
1085  * Same as xfs_bioerror, except that we are releasing the buffer
1086  * here ourselves, and avoiding the biodone call.
1087  * This is meant for userdata errors; metadata bufs come with
1088  * iodone functions attached, so that we can track down errors.
1089  */
1090 STATIC int
1091 xfs_bioerror_relse(
1092         struct xfs_buf  *bp)
1093 {
1094         int64_t         fl = XFS_BUF_BFLAGS(bp);
1095         /*
1096          * No need to wait until the buffer is unpinned.
1097          * We aren't flushing it.
1098          *
1099          * chunkhold expects B_DONE to be set, whether
1100          * we actually finish the I/O or not. We don't want to
1101          * change that interface.
1102          */
1103         XFS_BUF_UNREAD(bp);
1104         XFS_BUF_UNDELAYWRITE(bp);
1105         XFS_BUF_DONE(bp);
1106         XFS_BUF_STALE(bp);
1107         XFS_BUF_CLR_IODONE_FUNC(bp);
1108         XFS_BUF_CLR_BDSTRAT_FUNC(bp);
1109         if (!(fl & XBF_ASYNC)) {
1110                 /*
1111                  * Mark b_error and B_ERROR _both_.
1112                  * Lot's of chunkcache code assumes that.
1113                  * There's no reason to mark error for
1114                  * ASYNC buffers.
1115                  */
1116                 XFS_BUF_ERROR(bp, EIO);
1117                 XFS_BUF_FINISH_IOWAIT(bp);
1118         } else {
1119                 xfs_buf_relse(bp);
1120         }
1121
1122         return EIO;
1123 }
1124
1125
1126 /*
1127  * All xfs metadata buffers except log state machine buffers
1128  * get this attached as their b_bdstrat callback function.
1129  * This is so that we can catch a buffer
1130  * after prematurely unpinning it to forcibly shutdown the filesystem.
1131  */
1132 int
1133 xfs_bdstrat_cb(
1134         struct xfs_buf  *bp)
1135 {
1136         if (XFS_FORCED_SHUTDOWN(bp->b_mount)) {
1137                 trace_xfs_bdstrat_shut(bp, _RET_IP_);
1138                 /*
1139                  * Metadata write that didn't get logged but
1140                  * written delayed anyway. These aren't associated
1141                  * with a transaction, and can be ignored.
1142                  */
1143                 if (!bp->b_iodone && !XFS_BUF_ISREAD(bp))
1144                         return xfs_bioerror_relse(bp);
1145                 else
1146                         return xfs_bioerror(bp);
1147         }
1148
1149         xfs_buf_iorequest(bp);
1150         return 0;
1151 }
1152
1153 /*
1154  * Wrapper around bdstrat so that we can stop data from going to disk in case
1155  * we are shutting down the filesystem.  Typically user data goes thru this
1156  * path; one of the exceptions is the superblock.
1157  */
1158 void
1159 xfsbdstrat(
1160         struct xfs_mount        *mp,
1161         struct xfs_buf          *bp)
1162 {
1163         if (XFS_FORCED_SHUTDOWN(mp)) {
1164                 trace_xfs_bdstrat_shut(bp, _RET_IP_);
1165                 xfs_bioerror_relse(bp);
1166                 return;
1167         }
1168
1169         xfs_buf_iorequest(bp);
1170 }
1171
1172 STATIC void
1173 _xfs_buf_ioend(
1174         xfs_buf_t               *bp,
1175         int                     schedule)
1176 {
1177         if (atomic_dec_and_test(&bp->b_io_remaining) == 1) {
1178                 bp->b_flags &= ~_XBF_PAGE_LOCKED;
1179                 xfs_buf_ioend(bp, schedule);
1180         }
1181 }
1182
1183 STATIC void
1184 xfs_buf_bio_end_io(
1185         struct bio              *bio,
1186         int                     error)
1187 {
1188         xfs_buf_t               *bp = (xfs_buf_t *)bio->bi_private;
1189         unsigned int            blocksize = bp->b_target->bt_bsize;
1190         struct bio_vec          *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
1191
1192         xfs_buf_ioerror(bp, -error);
1193
1194         if (!error && xfs_buf_is_vmapped(bp) && (bp->b_flags & XBF_READ))
1195                 invalidate_kernel_vmap_range(bp->b_addr, xfs_buf_vmap_len(bp));
1196
1197         do {
1198                 struct page     *page = bvec->bv_page;
1199
1200                 ASSERT(!PagePrivate(page));
1201                 if (unlikely(bp->b_error)) {
1202                         if (bp->b_flags & XBF_READ)
1203                                 ClearPageUptodate(page);
1204                 } else if (blocksize >= PAGE_CACHE_SIZE) {
1205                         SetPageUptodate(page);
1206                 } else if (!PagePrivate(page) &&
1207                                 (bp->b_flags & _XBF_PAGE_CACHE)) {
1208                         set_page_region(page, bvec->bv_offset, bvec->bv_len);
1209                 }
1210
1211                 if (--bvec >= bio->bi_io_vec)
1212                         prefetchw(&bvec->bv_page->flags);
1213
1214                 if (bp->b_flags & _XBF_PAGE_LOCKED)
1215                         unlock_page(page);
1216         } while (bvec >= bio->bi_io_vec);
1217
1218         _xfs_buf_ioend(bp, 1);
1219         bio_put(bio);
1220 }
1221
1222 STATIC void
1223 _xfs_buf_ioapply(
1224         xfs_buf_t               *bp)
1225 {
1226         int                     rw, map_i, total_nr_pages, nr_pages;
1227         struct bio              *bio;
1228         int                     offset = bp->b_offset;
1229         int                     size = bp->b_count_desired;
1230         sector_t                sector = bp->b_bn;
1231         unsigned int            blocksize = bp->b_target->bt_bsize;
1232
1233         total_nr_pages = bp->b_page_count;
1234         map_i = 0;
1235
1236         if (bp->b_flags & XBF_ORDERED) {
1237                 ASSERT(!(bp->b_flags & XBF_READ));
1238                 rw = WRITE_BARRIER;
1239         } else if (bp->b_flags & XBF_LOG_BUFFER) {
1240                 ASSERT(!(bp->b_flags & XBF_READ_AHEAD));
1241                 bp->b_flags &= ~_XBF_RUN_QUEUES;
1242                 rw = (bp->b_flags & XBF_WRITE) ? WRITE_SYNC : READ_SYNC;
1243         } else if (bp->b_flags & _XBF_RUN_QUEUES) {
1244                 ASSERT(!(bp->b_flags & XBF_READ_AHEAD));
1245                 bp->b_flags &= ~_XBF_RUN_QUEUES;
1246                 rw = (bp->b_flags & XBF_WRITE) ? WRITE_META : READ_META;
1247         } else {
1248                 rw = (bp->b_flags & XBF_WRITE) ? WRITE :
1249                      (bp->b_flags & XBF_READ_AHEAD) ? READA : READ;
1250         }
1251
1252         /* Special code path for reading a sub page size buffer in --
1253          * we populate up the whole page, and hence the other metadata
1254          * in the same page.  This optimization is only valid when the
1255          * filesystem block size is not smaller than the page size.
1256          */
1257         if ((bp->b_buffer_length < PAGE_CACHE_SIZE) &&
1258             ((bp->b_flags & (XBF_READ|_XBF_PAGE_LOCKED)) ==
1259               (XBF_READ|_XBF_PAGE_LOCKED)) &&
1260             (blocksize >= PAGE_CACHE_SIZE)) {
1261                 bio = bio_alloc(GFP_NOIO, 1);
1262
1263                 bio->bi_bdev = bp->b_target->bt_bdev;
1264                 bio->bi_sector = sector - (offset >> BBSHIFT);
1265                 bio->bi_end_io = xfs_buf_bio_end_io;
1266                 bio->bi_private = bp;
1267
1268                 bio_add_page(bio, bp->b_pages[0], PAGE_CACHE_SIZE, 0);
1269                 size = 0;
1270
1271                 atomic_inc(&bp->b_io_remaining);
1272
1273                 goto submit_io;
1274         }
1275
1276 next_chunk:
1277         atomic_inc(&bp->b_io_remaining);
1278         nr_pages = BIO_MAX_SECTORS >> (PAGE_SHIFT - BBSHIFT);
1279         if (nr_pages > total_nr_pages)
1280                 nr_pages = total_nr_pages;
1281
1282         bio = bio_alloc(GFP_NOIO, nr_pages);
1283         bio->bi_bdev = bp->b_target->bt_bdev;
1284         bio->bi_sector = sector;
1285         bio->bi_end_io = xfs_buf_bio_end_io;
1286         bio->bi_private = bp;
1287
1288         for (; size && nr_pages; nr_pages--, map_i++) {
1289                 int     rbytes, nbytes = PAGE_CACHE_SIZE - offset;
1290
1291                 if (nbytes > size)
1292                         nbytes = size;
1293
1294                 rbytes = bio_add_page(bio, bp->b_pages[map_i], nbytes, offset);
1295                 if (rbytes < nbytes)
1296                         break;
1297
1298                 offset = 0;
1299                 sector += nbytes >> BBSHIFT;
1300                 size -= nbytes;
1301                 total_nr_pages--;
1302         }
1303
1304 submit_io:
1305         if (likely(bio->bi_size)) {
1306                 if (xfs_buf_is_vmapped(bp)) {
1307                         flush_kernel_vmap_range(bp->b_addr,
1308                                                 xfs_buf_vmap_len(bp));
1309                 }
1310                 submit_bio(rw, bio);
1311                 if (size)
1312                         goto next_chunk;
1313         } else {
1314                 bio_put(bio);
1315                 xfs_buf_ioerror(bp, EIO);
1316         }
1317 }
1318
1319 int
1320 xfs_buf_iorequest(
1321         xfs_buf_t               *bp)
1322 {
1323         trace_xfs_buf_iorequest(bp, _RET_IP_);
1324
1325         if (bp->b_flags & XBF_DELWRI) {
1326                 xfs_buf_delwri_queue(bp, 1);
1327                 return 0;
1328         }
1329
1330         if (bp->b_flags & XBF_WRITE) {
1331                 xfs_buf_wait_unpin(bp);
1332         }
1333
1334         xfs_buf_hold(bp);
1335
1336         /* Set the count to 1 initially, this will stop an I/O
1337          * completion callout which happens before we have started
1338          * all the I/O from calling xfs_buf_ioend too early.
1339          */
1340         atomic_set(&bp->b_io_remaining, 1);
1341         _xfs_buf_ioapply(bp);
1342         _xfs_buf_ioend(bp, 0);
1343
1344         xfs_buf_rele(bp);
1345         return 0;
1346 }
1347
1348 /*
1349  *      Waits for I/O to complete on the buffer supplied.
1350  *      It returns immediately if no I/O is pending.
1351  *      It returns the I/O error code, if any, or 0 if there was no error.
1352  */
1353 int
1354 xfs_buf_iowait(
1355         xfs_buf_t               *bp)
1356 {
1357         trace_xfs_buf_iowait(bp, _RET_IP_);
1358
1359         if (atomic_read(&bp->b_io_remaining))
1360                 blk_run_address_space(bp->b_target->bt_mapping);
1361         wait_for_completion(&bp->b_iowait);
1362
1363         trace_xfs_buf_iowait_done(bp, _RET_IP_);
1364         return bp->b_error;
1365 }
1366
1367 xfs_caddr_t
1368 xfs_buf_offset(
1369         xfs_buf_t               *bp,
1370         size_t                  offset)
1371 {
1372         struct page             *page;
1373
1374         if (bp->b_flags & XBF_MAPPED)
1375                 return XFS_BUF_PTR(bp) + offset;
1376
1377         offset += bp->b_offset;
1378         page = bp->b_pages[offset >> PAGE_CACHE_SHIFT];
1379         return (xfs_caddr_t)page_address(page) + (offset & (PAGE_CACHE_SIZE-1));
1380 }
1381
1382 /*
1383  *      Move data into or out of a buffer.
1384  */
1385 void
1386 xfs_buf_iomove(
1387         xfs_buf_t               *bp,    /* buffer to process            */
1388         size_t                  boff,   /* starting buffer offset       */
1389         size_t                  bsize,  /* length to copy               */
1390         void                    *data,  /* data address                 */
1391         xfs_buf_rw_t            mode)   /* read/write/zero flag         */
1392 {
1393         size_t                  bend, cpoff, csize;
1394         struct page             *page;
1395
1396         bend = boff + bsize;
1397         while (boff < bend) {
1398                 page = bp->b_pages[xfs_buf_btoct(boff + bp->b_offset)];
1399                 cpoff = xfs_buf_poff(boff + bp->b_offset);
1400                 csize = min_t(size_t,
1401                               PAGE_CACHE_SIZE-cpoff, bp->b_count_desired-boff);
1402
1403                 ASSERT(((csize + cpoff) <= PAGE_CACHE_SIZE));
1404
1405                 switch (mode) {
1406                 case XBRW_ZERO:
1407                         memset(page_address(page) + cpoff, 0, csize);
1408                         break;
1409                 case XBRW_READ:
1410                         memcpy(data, page_address(page) + cpoff, csize);
1411                         break;
1412                 case XBRW_WRITE:
1413                         memcpy(page_address(page) + cpoff, data, csize);
1414                 }
1415
1416                 boff += csize;
1417                 data += csize;
1418         }
1419 }
1420
1421 /*
1422  *      Handling of buffer targets (buftargs).
1423  */
1424
1425 /*
1426  *      Wait for any bufs with callbacks that have been submitted but
1427  *      have not yet returned... walk the hash list for the target.
1428  */
1429 void
1430 xfs_wait_buftarg(
1431         xfs_buftarg_t   *btp)
1432 {
1433         xfs_buf_t       *bp, *n;
1434         xfs_bufhash_t   *hash;
1435         uint            i;
1436
1437         for (i = 0; i < (1 << btp->bt_hashshift); i++) {
1438                 hash = &btp->bt_hash[i];
1439 again:
1440                 spin_lock(&hash->bh_lock);
1441                 list_for_each_entry_safe(bp, n, &hash->bh_list, b_hash_list) {
1442                         ASSERT(btp == bp->b_target);
1443                         if (!(bp->b_flags & XBF_FS_MANAGED)) {
1444                                 spin_unlock(&hash->bh_lock);
1445                                 /*
1446                                  * Catch superblock reference count leaks
1447                                  * immediately
1448                                  */
1449                                 BUG_ON(bp->b_bn == 0);
1450                                 delay(100);
1451                                 goto again;
1452                         }
1453                 }
1454                 spin_unlock(&hash->bh_lock);
1455         }
1456 }
1457
1458 /*
1459  *      Allocate buffer hash table for a given target.
1460  *      For devices containing metadata (i.e. not the log/realtime devices)
1461  *      we need to allocate a much larger hash table.
1462  */
1463 STATIC void
1464 xfs_alloc_bufhash(
1465         xfs_buftarg_t           *btp,
1466         int                     external)
1467 {
1468         unsigned int            i;
1469
1470         btp->bt_hashshift = external ? 3 : 8;   /* 8 or 256 buckets */
1471         btp->bt_hashmask = (1 << btp->bt_hashshift) - 1;
1472         btp->bt_hash = kmem_zalloc_large((1 << btp->bt_hashshift) *
1473                                          sizeof(xfs_bufhash_t));
1474         for (i = 0; i < (1 << btp->bt_hashshift); i++) {
1475                 spin_lock_init(&btp->bt_hash[i].bh_lock);
1476                 INIT_LIST_HEAD(&btp->bt_hash[i].bh_list);
1477         }
1478 }
1479
1480 STATIC void
1481 xfs_free_bufhash(
1482         xfs_buftarg_t           *btp)
1483 {
1484         kmem_free_large(btp->bt_hash);
1485         btp->bt_hash = NULL;
1486 }
1487
1488 /*
1489  *      buftarg list for delwrite queue processing
1490  */
1491 static LIST_HEAD(xfs_buftarg_list);
1492 static DEFINE_SPINLOCK(xfs_buftarg_lock);
1493
1494 STATIC void
1495 xfs_register_buftarg(
1496         xfs_buftarg_t           *btp)
1497 {
1498         spin_lock(&xfs_buftarg_lock);
1499         list_add(&btp->bt_list, &xfs_buftarg_list);
1500         spin_unlock(&xfs_buftarg_lock);
1501 }
1502
1503 STATIC void
1504 xfs_unregister_buftarg(
1505         xfs_buftarg_t           *btp)
1506 {
1507         spin_lock(&xfs_buftarg_lock);
1508         list_del(&btp->bt_list);
1509         spin_unlock(&xfs_buftarg_lock);
1510 }
1511
1512 void
1513 xfs_free_buftarg(
1514         struct xfs_mount        *mp,
1515         struct xfs_buftarg      *btp)
1516 {
1517         xfs_flush_buftarg(btp, 1);
1518         if (mp->m_flags & XFS_MOUNT_BARRIER)
1519                 xfs_blkdev_issue_flush(btp);
1520         xfs_free_bufhash(btp);
1521         iput(btp->bt_mapping->host);
1522
1523         /* Unregister the buftarg first so that we don't get a
1524          * wakeup finding a non-existent task
1525          */
1526         xfs_unregister_buftarg(btp);
1527         kthread_stop(btp->bt_task);
1528
1529         kmem_free(btp);
1530 }
1531
1532 STATIC int
1533 xfs_setsize_buftarg_flags(
1534         xfs_buftarg_t           *btp,
1535         unsigned int            blocksize,
1536         unsigned int            sectorsize,
1537         int                     verbose)
1538 {
1539         btp->bt_bsize = blocksize;
1540         btp->bt_sshift = ffs(sectorsize) - 1;
1541         btp->bt_smask = sectorsize - 1;
1542
1543         if (set_blocksize(btp->bt_bdev, sectorsize)) {
1544                 printk(KERN_WARNING
1545                         "XFS: Cannot set_blocksize to %u on device %s\n",
1546                         sectorsize, XFS_BUFTARG_NAME(btp));
1547                 return EINVAL;
1548         }
1549
1550         if (verbose &&
1551             (PAGE_CACHE_SIZE / BITS_PER_LONG) > sectorsize) {
1552                 printk(KERN_WARNING
1553                         "XFS: %u byte sectors in use on device %s.  "
1554                         "This is suboptimal; %u or greater is ideal.\n",
1555                         sectorsize, XFS_BUFTARG_NAME(btp),
1556                         (unsigned int)PAGE_CACHE_SIZE / BITS_PER_LONG);
1557         }
1558
1559         return 0;
1560 }
1561
1562 /*
1563  *      When allocating the initial buffer target we have not yet
1564  *      read in the superblock, so don't know what sized sectors
1565  *      are being used is at this early stage.  Play safe.
1566  */
1567 STATIC int
1568 xfs_setsize_buftarg_early(
1569         xfs_buftarg_t           *btp,
1570         struct block_device     *bdev)
1571 {
1572         return xfs_setsize_buftarg_flags(btp,
1573                         PAGE_CACHE_SIZE, bdev_logical_block_size(bdev), 0);
1574 }
1575
1576 int
1577 xfs_setsize_buftarg(
1578         xfs_buftarg_t           *btp,
1579         unsigned int            blocksize,
1580         unsigned int            sectorsize)
1581 {
1582         return xfs_setsize_buftarg_flags(btp, blocksize, sectorsize, 1);
1583 }
1584
1585 STATIC int
1586 xfs_mapping_buftarg(
1587         xfs_buftarg_t           *btp,
1588         struct block_device     *bdev)
1589 {
1590         struct backing_dev_info *bdi;
1591         struct inode            *inode;
1592         struct address_space    *mapping;
1593         static const struct address_space_operations mapping_aops = {
1594                 .sync_page = block_sync_page,
1595                 .migratepage = fail_migrate_page,
1596         };
1597
1598         inode = new_inode(bdev->bd_inode->i_sb);
1599         if (!inode) {
1600                 printk(KERN_WARNING
1601                         "XFS: Cannot allocate mapping inode for device %s\n",
1602                         XFS_BUFTARG_NAME(btp));
1603                 return ENOMEM;
1604         }
1605         inode->i_mode = S_IFBLK;
1606         inode->i_bdev = bdev;
1607         inode->i_rdev = bdev->bd_dev;
1608         bdi = blk_get_backing_dev_info(bdev);
1609         if (!bdi)
1610                 bdi = &default_backing_dev_info;
1611         mapping = &inode->i_data;
1612         mapping->a_ops = &mapping_aops;
1613         mapping->backing_dev_info = bdi;
1614         mapping_set_gfp_mask(mapping, GFP_NOFS);
1615         btp->bt_mapping = mapping;
1616         return 0;
1617 }
1618
1619 STATIC int
1620 xfs_alloc_delwrite_queue(
1621         xfs_buftarg_t           *btp,
1622         const char              *fsname)
1623 {
1624         int     error = 0;
1625
1626         INIT_LIST_HEAD(&btp->bt_list);
1627         INIT_LIST_HEAD(&btp->bt_delwrite_queue);
1628         spin_lock_init(&btp->bt_delwrite_lock);
1629         btp->bt_flags = 0;
1630         btp->bt_task = kthread_run(xfsbufd, btp, "xfsbufd/%s", fsname);
1631         if (IS_ERR(btp->bt_task)) {
1632                 error = PTR_ERR(btp->bt_task);
1633                 goto out_error;
1634         }
1635         xfs_register_buftarg(btp);
1636 out_error:
1637         return error;
1638 }
1639
1640 xfs_buftarg_t *
1641 xfs_alloc_buftarg(
1642         struct block_device     *bdev,
1643         int                     external,
1644         const char              *fsname)
1645 {
1646         xfs_buftarg_t           *btp;
1647
1648         btp = kmem_zalloc(sizeof(*btp), KM_SLEEP);
1649
1650         btp->bt_dev =  bdev->bd_dev;
1651         btp->bt_bdev = bdev;
1652         if (xfs_setsize_buftarg_early(btp, bdev))
1653                 goto error;
1654         if (xfs_mapping_buftarg(btp, bdev))
1655                 goto error;
1656         if (xfs_alloc_delwrite_queue(btp, fsname))
1657                 goto error;
1658         xfs_alloc_bufhash(btp, external);
1659         return btp;
1660
1661 error:
1662         kmem_free(btp);
1663         return NULL;
1664 }
1665
1666
1667 /*
1668  *      Delayed write buffer handling
1669  */
1670 STATIC void
1671 xfs_buf_delwri_queue(
1672         xfs_buf_t               *bp,
1673         int                     unlock)
1674 {
1675         struct list_head        *dwq = &bp->b_target->bt_delwrite_queue;
1676         spinlock_t              *dwlk = &bp->b_target->bt_delwrite_lock;
1677
1678         trace_xfs_buf_delwri_queue(bp, _RET_IP_);
1679
1680         ASSERT((bp->b_flags&(XBF_DELWRI|XBF_ASYNC)) == (XBF_DELWRI|XBF_ASYNC));
1681
1682         spin_lock(dwlk);
1683         /* If already in the queue, dequeue and place at tail */
1684         if (!list_empty(&bp->b_list)) {
1685                 ASSERT(bp->b_flags & _XBF_DELWRI_Q);
1686                 if (unlock)
1687                         atomic_dec(&bp->b_hold);
1688                 list_del(&bp->b_list);
1689         }
1690
1691         if (list_empty(dwq)) {
1692                 /* start xfsbufd as it is about to have something to do */
1693                 wake_up_process(bp->b_target->bt_task);
1694         }
1695
1696         bp->b_flags |= _XBF_DELWRI_Q;
1697         list_add_tail(&bp->b_list, dwq);
1698         bp->b_queuetime = jiffies;
1699         spin_unlock(dwlk);
1700
1701         if (unlock)
1702                 xfs_buf_unlock(bp);
1703 }
1704
1705 void
1706 xfs_buf_delwri_dequeue(
1707         xfs_buf_t               *bp)
1708 {
1709         spinlock_t              *dwlk = &bp->b_target->bt_delwrite_lock;
1710         int                     dequeued = 0;
1711
1712         spin_lock(dwlk);
1713         if ((bp->b_flags & XBF_DELWRI) && !list_empty(&bp->b_list)) {
1714                 ASSERT(bp->b_flags & _XBF_DELWRI_Q);
1715                 list_del_init(&bp->b_list);
1716                 dequeued = 1;
1717         }
1718         bp->b_flags &= ~(XBF_DELWRI|_XBF_DELWRI_Q);
1719         spin_unlock(dwlk);
1720
1721         if (dequeued)
1722                 xfs_buf_rele(bp);
1723
1724         trace_xfs_buf_delwri_dequeue(bp, _RET_IP_);
1725 }
1726
1727 /*
1728  * If a delwri buffer needs to be pushed before it has aged out, then promote
1729  * it to the head of the delwri queue so that it will be flushed on the next
1730  * xfsbufd run. We do this by resetting the queuetime of the buffer to be older
1731  * than the age currently needed to flush the buffer. Hence the next time the
1732  * xfsbufd sees it is guaranteed to be considered old enough to flush.
1733  */
1734 void
1735 xfs_buf_delwri_promote(
1736         struct xfs_buf  *bp)
1737 {
1738         struct xfs_buftarg *btp = bp->b_target;
1739         long            age = xfs_buf_age_centisecs * msecs_to_jiffies(10) + 1;
1740
1741         ASSERT(bp->b_flags & XBF_DELWRI);
1742         ASSERT(bp->b_flags & _XBF_DELWRI_Q);
1743
1744         /*
1745          * Check the buffer age before locking the delayed write queue as we
1746          * don't need to promote buffers that are already past the flush age.
1747          */
1748         if (bp->b_queuetime < jiffies - age)
1749                 return;
1750         bp->b_queuetime = jiffies - age;
1751         spin_lock(&btp->bt_delwrite_lock);
1752         list_move(&bp->b_list, &btp->bt_delwrite_queue);
1753         spin_unlock(&btp->bt_delwrite_lock);
1754 }
1755
1756 STATIC void
1757 xfs_buf_runall_queues(
1758         struct workqueue_struct *queue)
1759 {
1760         flush_workqueue(queue);
1761 }
1762
1763 STATIC int
1764 xfsbufd_wakeup(
1765         int                     priority,
1766         gfp_t                   mask)
1767 {
1768         xfs_buftarg_t           *btp;
1769
1770         spin_lock(&xfs_buftarg_lock);
1771         list_for_each_entry(btp, &xfs_buftarg_list, bt_list) {
1772                 if (test_bit(XBT_FORCE_SLEEP, &btp->bt_flags))
1773                         continue;
1774                 if (list_empty(&btp->bt_delwrite_queue))
1775                         continue;
1776                 set_bit(XBT_FORCE_FLUSH, &btp->bt_flags);
1777                 wake_up_process(btp->bt_task);
1778         }
1779         spin_unlock(&xfs_buftarg_lock);
1780         return 0;
1781 }
1782
1783 /*
1784  * Move as many buffers as specified to the supplied list
1785  * idicating if we skipped any buffers to prevent deadlocks.
1786  */
1787 STATIC int
1788 xfs_buf_delwri_split(
1789         xfs_buftarg_t   *target,
1790         struct list_head *list,
1791         unsigned long   age)
1792 {
1793         xfs_buf_t       *bp, *n;
1794         struct list_head *dwq = &target->bt_delwrite_queue;
1795         spinlock_t      *dwlk = &target->bt_delwrite_lock;
1796         int             skipped = 0;
1797         int             force;
1798
1799         force = test_and_clear_bit(XBT_FORCE_FLUSH, &target->bt_flags);
1800         INIT_LIST_HEAD(list);
1801         spin_lock(dwlk);
1802         list_for_each_entry_safe(bp, n, dwq, b_list) {
1803                 trace_xfs_buf_delwri_split(bp, _RET_IP_);
1804                 ASSERT(bp->b_flags & XBF_DELWRI);
1805
1806                 if (!xfs_buf_ispin(bp) && !xfs_buf_cond_lock(bp)) {
1807                         if (!force &&
1808                             time_before(jiffies, bp->b_queuetime + age)) {
1809                                 xfs_buf_unlock(bp);
1810                                 break;
1811                         }
1812
1813                         bp->b_flags &= ~(XBF_DELWRI|_XBF_DELWRI_Q|
1814                                          _XBF_RUN_QUEUES);
1815                         bp->b_flags |= XBF_WRITE;
1816                         list_move_tail(&bp->b_list, list);
1817                 } else
1818                         skipped++;
1819         }
1820         spin_unlock(dwlk);
1821
1822         return skipped;
1823
1824 }
1825
1826 /*
1827  * Compare function is more complex than it needs to be because
1828  * the return value is only 32 bits and we are doing comparisons
1829  * on 64 bit values
1830  */
1831 static int
1832 xfs_buf_cmp(
1833         void            *priv,
1834         struct list_head *a,
1835         struct list_head *b)
1836 {
1837         struct xfs_buf  *ap = container_of(a, struct xfs_buf, b_list);
1838         struct xfs_buf  *bp = container_of(b, struct xfs_buf, b_list);
1839         xfs_daddr_t             diff;
1840
1841         diff = ap->b_bn - bp->b_bn;
1842         if (diff < 0)
1843                 return -1;
1844         if (diff > 0)
1845                 return 1;
1846         return 0;
1847 }
1848
1849 void
1850 xfs_buf_delwri_sort(
1851         xfs_buftarg_t   *target,
1852         struct list_head *list)
1853 {
1854         list_sort(NULL, list, xfs_buf_cmp);
1855 }
1856
1857 STATIC int
1858 xfsbufd(
1859         void            *data)
1860 {
1861         xfs_buftarg_t   *target = (xfs_buftarg_t *)data;
1862
1863         current->flags |= PF_MEMALLOC;
1864
1865         set_freezable();
1866
1867         do {
1868                 long    age = xfs_buf_age_centisecs * msecs_to_jiffies(10);
1869                 long    tout = xfs_buf_timer_centisecs * msecs_to_jiffies(10);
1870                 int     count = 0;
1871                 struct list_head tmp;
1872
1873                 if (unlikely(freezing(current))) {
1874                         set_bit(XBT_FORCE_SLEEP, &target->bt_flags);
1875                         refrigerator();
1876                 } else {
1877                         clear_bit(XBT_FORCE_SLEEP, &target->bt_flags);
1878                 }
1879
1880                 /* sleep for a long time if there is nothing to do. */
1881                 if (list_empty(&target->bt_delwrite_queue))
1882                         tout = MAX_SCHEDULE_TIMEOUT;
1883                 schedule_timeout_interruptible(tout);
1884
1885                 xfs_buf_delwri_split(target, &tmp, age);
1886                 list_sort(NULL, &tmp, xfs_buf_cmp);
1887                 while (!list_empty(&tmp)) {
1888                         struct xfs_buf *bp;
1889                         bp = list_first_entry(&tmp, struct xfs_buf, b_list);
1890                         list_del_init(&bp->b_list);
1891                         xfs_buf_iostrategy(bp);
1892                         count++;
1893                 }
1894                 if (count)
1895                         blk_run_address_space(target->bt_mapping);
1896
1897         } while (!kthread_should_stop());
1898
1899         return 0;
1900 }
1901
1902 /*
1903  *      Go through all incore buffers, and release buffers if they belong to
1904  *      the given device. This is used in filesystem error handling to
1905  *      preserve the consistency of its metadata.
1906  */
1907 int
1908 xfs_flush_buftarg(
1909         xfs_buftarg_t   *target,
1910         int             wait)
1911 {
1912         xfs_buf_t       *bp;
1913         int             pincount = 0;
1914         LIST_HEAD(tmp_list);
1915         LIST_HEAD(wait_list);
1916
1917         xfs_buf_runall_queues(xfsconvertd_workqueue);
1918         xfs_buf_runall_queues(xfsdatad_workqueue);
1919         xfs_buf_runall_queues(xfslogd_workqueue);
1920
1921         set_bit(XBT_FORCE_FLUSH, &target->bt_flags);
1922         pincount = xfs_buf_delwri_split(target, &tmp_list, 0);
1923
1924         /*
1925          * Dropped the delayed write list lock, now walk the temporary list.
1926          * All I/O is issued async and then if we need to wait for completion
1927          * we do that after issuing all the IO.
1928          */
1929         list_sort(NULL, &tmp_list, xfs_buf_cmp);
1930         while (!list_empty(&tmp_list)) {
1931                 bp = list_first_entry(&tmp_list, struct xfs_buf, b_list);
1932                 ASSERT(target == bp->b_target);
1933                 list_del_init(&bp->b_list);
1934                 if (wait) {
1935                         bp->b_flags &= ~XBF_ASYNC;
1936                         list_add(&bp->b_list, &wait_list);
1937                 }
1938                 xfs_buf_iostrategy(bp);
1939         }
1940
1941         if (wait) {
1942                 /* Expedite and wait for IO to complete. */
1943                 blk_run_address_space(target->bt_mapping);
1944                 while (!list_empty(&wait_list)) {
1945                         bp = list_first_entry(&wait_list, struct xfs_buf, b_list);
1946
1947                         list_del_init(&bp->b_list);
1948                         xfs_iowait(bp);
1949                         xfs_buf_relse(bp);
1950                 }
1951         }
1952
1953         return pincount;
1954 }
1955
1956 int __init
1957 xfs_buf_init(void)
1958 {
1959         xfs_buf_zone = kmem_zone_init_flags(sizeof(xfs_buf_t), "xfs_buf",
1960                                                 KM_ZONE_HWALIGN, NULL);
1961         if (!xfs_buf_zone)
1962                 goto out;
1963
1964         xfslogd_workqueue = create_workqueue("xfslogd");
1965         if (!xfslogd_workqueue)
1966                 goto out_free_buf_zone;
1967
1968         xfsdatad_workqueue = create_workqueue("xfsdatad");
1969         if (!xfsdatad_workqueue)
1970                 goto out_destroy_xfslogd_workqueue;
1971
1972         xfsconvertd_workqueue = create_workqueue("xfsconvertd");
1973         if (!xfsconvertd_workqueue)
1974                 goto out_destroy_xfsdatad_workqueue;
1975
1976         register_shrinker(&xfs_buf_shake);
1977         return 0;
1978
1979  out_destroy_xfsdatad_workqueue:
1980         destroy_workqueue(xfsdatad_workqueue);
1981  out_destroy_xfslogd_workqueue:
1982         destroy_workqueue(xfslogd_workqueue);
1983  out_free_buf_zone:
1984         kmem_zone_destroy(xfs_buf_zone);
1985  out:
1986         return -ENOMEM;
1987 }
1988
1989 void
1990 xfs_buf_terminate(void)
1991 {
1992         unregister_shrinker(&xfs_buf_shake);
1993         destroy_workqueue(xfsconvertd_workqueue);
1994         destroy_workqueue(xfsdatad_workqueue);
1995         destroy_workqueue(xfslogd_workqueue);
1996         kmem_zone_destroy(xfs_buf_zone);
1997 }
1998
1999 #ifdef CONFIG_KDB_MODULES
2000 struct list_head *
2001 xfs_get_buftarg_list(void)
2002 {
2003         return &xfs_buftarg_list;
2004 }
2005 #endif