2 * Copyright (c) 2000-2005 Silicon Graphics, Inc.
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.
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.
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
18 #include <linux/stddef.h>
19 #include <linux/errno.h>
20 #include <linux/slab.h>
21 #include <linux/pagemap.h>
22 #include <linux/init.h>
23 #include <linux/vmalloc.h>
24 #include <linux/bio.h>
25 #include <linux/sysctl.h>
26 #include <linux/proc_fs.h>
27 #include <linux/workqueue.h>
28 #include <linux/percpu.h>
29 #include <linux/blkdev.h>
30 #include <linux/hash.h>
31 #include <linux/kthread.h>
32 #include "xfs_linux.h"
34 STATIC kmem_zone_t *xfs_buf_zone;
35 STATIC kmem_shaker_t xfs_buf_shake;
36 STATIC int xfsbufd(void *);
37 STATIC int xfsbufd_wakeup(int, gfp_t);
38 STATIC void xfs_buf_delwri_queue(xfs_buf_t *, int);
40 STATIC struct workqueue_struct *xfslogd_workqueue;
41 struct workqueue_struct *xfsdatad_workqueue;
51 ktrace_enter(xfs_buf_trace_buf,
53 (void *)(unsigned long)bp->b_flags,
54 (void *)(unsigned long)bp->b_hold.counter,
55 (void *)(unsigned long)bp->b_sema.count.counter,
58 (void *)(unsigned long)((bp->b_file_offset>>32) & 0xffffffff),
59 (void *)(unsigned long)(bp->b_file_offset & 0xffffffff),
60 (void *)(unsigned long)bp->b_buffer_length,
61 NULL, NULL, NULL, NULL, NULL);
63 ktrace_t *xfs_buf_trace_buf;
64 #define XFS_BUF_TRACE_SIZE 4096
65 #define XB_TRACE(bp, id, data) \
66 xfs_buf_trace(bp, id, (void *)data, (void *)__builtin_return_address(0))
68 #define XB_TRACE(bp, id, data) do { } while (0)
71 #ifdef XFS_BUF_LOCK_TRACKING
72 # define XB_SET_OWNER(bp) ((bp)->b_last_holder = current->pid)
73 # define XB_CLEAR_OWNER(bp) ((bp)->b_last_holder = -1)
74 # define XB_GET_OWNER(bp) ((bp)->b_last_holder)
76 # define XB_SET_OWNER(bp) do { } while (0)
77 # define XB_CLEAR_OWNER(bp) do { } while (0)
78 # define XB_GET_OWNER(bp) do { } while (0)
81 #define xb_to_gfp(flags) \
82 ((((flags) & XBF_READ_AHEAD) ? __GFP_NORETRY : \
83 ((flags) & XBF_DONT_BLOCK) ? GFP_NOFS : GFP_KERNEL) | __GFP_NOWARN)
85 #define xb_to_km(flags) \
86 (((flags) & XBF_DONT_BLOCK) ? KM_NOFS : KM_SLEEP)
88 #define xfs_buf_allocate(flags) \
89 kmem_zone_alloc(xfs_buf_zone, xb_to_km(flags))
90 #define xfs_buf_deallocate(bp) \
91 kmem_zone_free(xfs_buf_zone, (bp));
94 * Page Region interfaces.
96 * For pages in filesystems where the blocksize is smaller than the
97 * pagesize, we use the page->private field (long) to hold a bitmap
98 * of uptodate regions within the page.
100 * Each such region is "bytes per page / bits per long" bytes long.
102 * NBPPR == number-of-bytes-per-page-region
103 * BTOPR == bytes-to-page-region (rounded up)
104 * BTOPRT == bytes-to-page-region-truncated (rounded down)
106 #if (BITS_PER_LONG == 32)
107 #define PRSHIFT (PAGE_CACHE_SHIFT - 5) /* (32 == 1<<5) */
108 #elif (BITS_PER_LONG == 64)
109 #define PRSHIFT (PAGE_CACHE_SHIFT - 6) /* (64 == 1<<6) */
111 #error BITS_PER_LONG must be 32 or 64
113 #define NBPPR (PAGE_CACHE_SIZE/BITS_PER_LONG)
114 #define BTOPR(b) (((unsigned int)(b) + (NBPPR - 1)) >> PRSHIFT)
115 #define BTOPRT(b) (((unsigned int)(b) >> PRSHIFT))
125 first = BTOPR(offset);
126 final = BTOPRT(offset + length - 1);
127 first = min(first, final);
130 mask <<= BITS_PER_LONG - (final - first);
131 mask >>= BITS_PER_LONG - (final);
133 ASSERT(offset + length <= PAGE_CACHE_SIZE);
134 ASSERT((final - first) < BITS_PER_LONG && (final - first) >= 0);
145 set_page_private(page,
146 page_private(page) | page_region_mask(offset, length));
147 if (page_private(page) == ~0UL)
148 SetPageUptodate(page);
157 unsigned long mask = page_region_mask(offset, length);
159 return (mask && (page_private(page) & mask) == mask);
163 * Mapping of multi-page buffers into contiguous virtual space
166 typedef struct a_list {
171 STATIC a_list_t *as_free_head;
172 STATIC int as_list_len;
173 STATIC DEFINE_SPINLOCK(as_lock);
176 * Try to batch vunmaps because they are costly.
184 aentry = kmalloc(sizeof(a_list_t), GFP_ATOMIC & ~__GFP_HIGH);
185 if (likely(aentry)) {
187 aentry->next = as_free_head;
188 aentry->vm_addr = addr;
189 as_free_head = aentry;
191 spin_unlock(&as_lock);
198 purge_addresses(void)
200 a_list_t *aentry, *old;
202 if (as_free_head == NULL)
206 aentry = as_free_head;
209 spin_unlock(&as_lock);
211 while ((old = aentry) != NULL) {
212 vunmap(aentry->vm_addr);
213 aentry = aentry->next;
219 * Internal xfs_buf_t object manipulation
225 xfs_buftarg_t *target,
228 xfs_buf_flags_t flags)
231 * We don't want certain flags to appear in b_flags.
233 flags &= ~(XBF_LOCK|XBF_MAPPED|XBF_DONT_BLOCK|XBF_READ_AHEAD);
235 memset(bp, 0, sizeof(xfs_buf_t));
236 atomic_set(&bp->b_hold, 1);
237 init_MUTEX_LOCKED(&bp->b_iodonesema);
238 INIT_LIST_HEAD(&bp->b_list);
239 INIT_LIST_HEAD(&bp->b_hash_list);
240 init_MUTEX_LOCKED(&bp->b_sema); /* held, no waiters */
242 bp->b_target = target;
243 bp->b_file_offset = range_base;
245 * Set buffer_length and count_desired to the same value initially.
246 * I/O routines should use count_desired, which will be the same in
247 * most cases but may be reset (e.g. XFS recovery).
249 bp->b_buffer_length = bp->b_count_desired = range_length;
251 bp->b_bn = XFS_BUF_DADDR_NULL;
252 atomic_set(&bp->b_pin_count, 0);
253 init_waitqueue_head(&bp->b_waiters);
255 XFS_STATS_INC(xb_create);
256 XB_TRACE(bp, "initialize", target);
260 * Allocate a page array capable of holding a specified number
261 * of pages, and point the page buf at it.
267 xfs_buf_flags_t flags)
269 /* Make sure that we have a page list */
270 if (bp->b_pages == NULL) {
271 bp->b_offset = xfs_buf_poff(bp->b_file_offset);
272 bp->b_page_count = page_count;
273 if (page_count <= XB_PAGES) {
274 bp->b_pages = bp->b_page_array;
276 bp->b_pages = kmem_alloc(sizeof(struct page *) *
277 page_count, xb_to_km(flags));
278 if (bp->b_pages == NULL)
281 memset(bp->b_pages, 0, sizeof(struct page *) * page_count);
287 * Frees b_pages if it was allocated.
293 if (bp->b_pages != bp->b_page_array) {
294 kmem_free(bp->b_pages,
295 bp->b_page_count * sizeof(struct page *));
300 * Releases the specified buffer.
302 * The modification state of any associated pages is left unchanged.
303 * The buffer most not be on any hash - use xfs_buf_rele instead for
304 * hashed and refcounted buffers
310 XB_TRACE(bp, "free", 0);
312 ASSERT(list_empty(&bp->b_hash_list));
314 if (bp->b_flags & _XBF_PAGE_CACHE) {
317 if ((bp->b_flags & XBF_MAPPED) && (bp->b_page_count > 1))
318 free_address(bp->b_addr - bp->b_offset);
320 for (i = 0; i < bp->b_page_count; i++)
321 page_cache_release(bp->b_pages[i]);
322 _xfs_buf_free_pages(bp);
323 } else if (bp->b_flags & _XBF_KMEM_ALLOC) {
325 * XXX(hch): bp->b_count_desired might be incorrect (see
326 * xfs_buf_associate_memory for details), but fortunately
327 * the Linux version of kmem_free ignores the len argument..
329 kmem_free(bp->b_addr, bp->b_count_desired);
330 _xfs_buf_free_pages(bp);
333 xfs_buf_deallocate(bp);
337 * Finds all pages for buffer in question and builds it's page list.
340 _xfs_buf_lookup_pages(
344 struct address_space *mapping = bp->b_target->bt_mapping;
345 size_t blocksize = bp->b_target->bt_bsize;
346 size_t size = bp->b_count_desired;
347 size_t nbytes, offset;
348 gfp_t gfp_mask = xb_to_gfp(flags);
349 unsigned short page_count, i;
354 end = bp->b_file_offset + bp->b_buffer_length;
355 page_count = xfs_buf_btoc(end) - xfs_buf_btoct(bp->b_file_offset);
357 error = _xfs_buf_get_pages(bp, page_count, flags);
360 bp->b_flags |= _XBF_PAGE_CACHE;
362 offset = bp->b_offset;
363 first = bp->b_file_offset >> PAGE_CACHE_SHIFT;
365 for (i = 0; i < bp->b_page_count; i++) {
370 page = find_or_create_page(mapping, first + i, gfp_mask);
371 if (unlikely(page == NULL)) {
372 if (flags & XBF_READ_AHEAD) {
373 bp->b_page_count = i;
374 for (i = 0; i < bp->b_page_count; i++)
375 unlock_page(bp->b_pages[i]);
380 * This could deadlock.
382 * But until all the XFS lowlevel code is revamped to
383 * handle buffer allocation failures we can't do much.
385 if (!(++retries % 100))
387 "XFS: possible memory allocation "
388 "deadlock in %s (mode:0x%x)\n",
389 __FUNCTION__, gfp_mask);
391 XFS_STATS_INC(xb_page_retries);
392 xfsbufd_wakeup(0, gfp_mask);
393 blk_congestion_wait(WRITE, HZ/50);
397 XFS_STATS_INC(xb_page_found);
399 nbytes = min_t(size_t, size, PAGE_CACHE_SIZE - offset);
402 if (!PageUptodate(page)) {
404 if (blocksize >= PAGE_CACHE_SIZE) {
405 if (flags & XBF_READ)
407 } else if (!PagePrivate(page)) {
408 if (test_page_region(page, offset, nbytes))
413 bp->b_pages[i] = page;
418 for (i = 0; i < bp->b_page_count; i++)
419 unlock_page(bp->b_pages[i]);
422 if (page_count == bp->b_page_count)
423 bp->b_flags |= XBF_DONE;
425 XB_TRACE(bp, "lookup_pages", (long)page_count);
430 * Map buffer into kernel address-space if nessecary.
437 /* A single page buffer is always mappable */
438 if (bp->b_page_count == 1) {
439 bp->b_addr = page_address(bp->b_pages[0]) + bp->b_offset;
440 bp->b_flags |= XBF_MAPPED;
441 } else if (flags & XBF_MAPPED) {
442 if (as_list_len > 64)
444 bp->b_addr = vmap(bp->b_pages, bp->b_page_count,
445 VM_MAP, PAGE_KERNEL);
446 if (unlikely(bp->b_addr == NULL))
448 bp->b_addr += bp->b_offset;
449 bp->b_flags |= XBF_MAPPED;
456 * Finding and Reading Buffers
460 * Look up, and creates if absent, a lockable buffer for
461 * a given range of an inode. The buffer is returned
462 * locked. If other overlapping buffers exist, they are
463 * released before the new buffer is created and locked,
464 * which may imply that this call will block until those buffers
465 * are unlocked. No I/O is implied by this call.
469 xfs_buftarg_t *btp, /* block device target */
470 loff_t ioff, /* starting offset of range */
471 size_t isize, /* length of range */
472 xfs_buf_flags_t flags,
480 range_base = (ioff << BBSHIFT);
481 range_length = (isize << BBSHIFT);
483 /* Check for IOs smaller than the sector size / not sector aligned */
484 ASSERT(!(range_length < (1 << btp->bt_sshift)));
485 ASSERT(!(range_base & (loff_t)btp->bt_smask));
487 hash = &btp->bt_hash[hash_long((unsigned long)ioff, btp->bt_hashshift)];
489 spin_lock(&hash->bh_lock);
491 list_for_each_entry_safe(bp, n, &hash->bh_list, b_hash_list) {
492 ASSERT(btp == bp->b_target);
493 if (bp->b_file_offset == range_base &&
494 bp->b_buffer_length == range_length) {
496 * If we look at something, bring it to the
497 * front of the list for next time.
499 atomic_inc(&bp->b_hold);
500 list_move(&bp->b_hash_list, &hash->bh_list);
507 _xfs_buf_initialize(new_bp, btp, range_base,
508 range_length, flags);
509 new_bp->b_hash = hash;
510 list_add(&new_bp->b_hash_list, &hash->bh_list);
512 XFS_STATS_INC(xb_miss_locked);
515 spin_unlock(&hash->bh_lock);
519 spin_unlock(&hash->bh_lock);
521 /* Attempt to get the semaphore without sleeping,
522 * if this does not work then we need to drop the
523 * spinlock and do a hard attempt on the semaphore.
525 if (down_trylock(&bp->b_sema)) {
526 if (!(flags & XBF_TRYLOCK)) {
527 /* wait for buffer ownership */
528 XB_TRACE(bp, "get_lock", 0);
530 XFS_STATS_INC(xb_get_locked_waited);
532 /* We asked for a trylock and failed, no need
533 * to look at file offset and length here, we
534 * know that this buffer at least overlaps our
535 * buffer and is locked, therefore our buffer
536 * either does not exist, or is this buffer.
539 XFS_STATS_INC(xb_busy_locked);
547 if (bp->b_flags & XBF_STALE) {
548 ASSERT((bp->b_flags & _XBF_DELWRI_Q) == 0);
549 bp->b_flags &= XBF_MAPPED;
551 XB_TRACE(bp, "got_lock", 0);
552 XFS_STATS_INC(xb_get_locked);
557 * Assembles a buffer covering the specified range.
558 * Storage in memory for all portions of the buffer will be allocated,
559 * although backing storage may not be.
563 xfs_buftarg_t *target,/* target for buffer */
564 loff_t ioff, /* starting offset of range */
565 size_t isize, /* length of range */
566 xfs_buf_flags_t flags)
568 xfs_buf_t *bp, *new_bp;
571 new_bp = xfs_buf_allocate(flags);
572 if (unlikely(!new_bp))
575 bp = _xfs_buf_find(target, ioff, isize, flags, new_bp);
577 error = _xfs_buf_lookup_pages(bp, flags);
581 xfs_buf_deallocate(new_bp);
582 if (unlikely(bp == NULL))
586 for (i = 0; i < bp->b_page_count; i++)
587 mark_page_accessed(bp->b_pages[i]);
589 if (!(bp->b_flags & XBF_MAPPED)) {
590 error = _xfs_buf_map_pages(bp, flags);
591 if (unlikely(error)) {
592 printk(KERN_WARNING "%s: failed to map pages\n",
598 XFS_STATS_INC(xb_get);
601 * Always fill in the block number now, the mapped cases can do
602 * their own overlay of this later.
605 bp->b_count_desired = bp->b_buffer_length;
607 XB_TRACE(bp, "get", (unsigned long)flags);
611 if (flags & (XBF_LOCK | XBF_TRYLOCK))
619 xfs_buftarg_t *target,
622 xfs_buf_flags_t flags)
628 bp = xfs_buf_get_flags(target, ioff, isize, flags);
630 if (!XFS_BUF_ISDONE(bp)) {
631 XB_TRACE(bp, "read", (unsigned long)flags);
632 XFS_STATS_INC(xb_get_read);
633 xfs_buf_iostart(bp, flags);
634 } else if (flags & XBF_ASYNC) {
635 XB_TRACE(bp, "read_async", (unsigned long)flags);
637 * Read ahead call which is already satisfied,
642 XB_TRACE(bp, "read_done", (unsigned long)flags);
643 /* We do not want read in the flags */
644 bp->b_flags &= ~XBF_READ;
651 if (flags & (XBF_LOCK | XBF_TRYLOCK))
658 * If we are not low on memory then do the readahead in a deadlock
663 xfs_buftarg_t *target,
666 xfs_buf_flags_t flags)
668 struct backing_dev_info *bdi;
670 bdi = target->bt_mapping->backing_dev_info;
671 if (bdi_read_congested(bdi))
674 flags |= (XBF_TRYLOCK|XBF_ASYNC|XBF_READ_AHEAD);
675 xfs_buf_read_flags(target, ioff, isize, flags);
681 xfs_buftarg_t *target)
685 bp = xfs_buf_allocate(0);
687 _xfs_buf_initialize(bp, target, 0, len, 0);
691 static inline struct page *
695 if (((unsigned long)addr < VMALLOC_START) ||
696 ((unsigned long)addr >= VMALLOC_END)) {
697 return virt_to_page(addr);
699 return vmalloc_to_page(addr);
704 xfs_buf_associate_memory(
716 page_count = PAGE_CACHE_ALIGN(len) >> PAGE_CACHE_SHIFT;
717 offset = (off_t) mem - ((off_t)mem & PAGE_CACHE_MASK);
718 if (offset && (len > PAGE_CACHE_SIZE))
721 /* Free any previous set of page pointers */
723 _xfs_buf_free_pages(bp);
728 rval = _xfs_buf_get_pages(bp, page_count, 0);
732 bp->b_offset = offset;
733 ptr = (size_t) mem & PAGE_CACHE_MASK;
734 end = PAGE_CACHE_ALIGN((size_t) mem + len);
736 /* set up first page */
737 bp->b_pages[0] = mem_to_page(mem);
739 ptr += PAGE_CACHE_SIZE;
740 bp->b_page_count = ++i;
742 bp->b_pages[i] = mem_to_page((void *)ptr);
743 bp->b_page_count = ++i;
744 ptr += PAGE_CACHE_SIZE;
748 bp->b_count_desired = bp->b_buffer_length = len;
749 bp->b_flags |= XBF_MAPPED;
757 xfs_buftarg_t *target)
759 size_t malloc_len = len;
764 bp = xfs_buf_allocate(0);
765 if (unlikely(bp == NULL))
767 _xfs_buf_initialize(bp, target, 0, len, 0);
770 data = kmem_alloc(malloc_len, KM_SLEEP | KM_MAYFAIL);
771 if (unlikely(data == NULL))
774 /* check whether alignment matches.. */
775 if ((__psunsigned_t)data !=
776 ((__psunsigned_t)data & ~target->bt_smask)) {
777 /* .. else double the size and try again */
778 kmem_free(data, malloc_len);
783 error = xfs_buf_associate_memory(bp, data, len);
786 bp->b_flags |= _XBF_KMEM_ALLOC;
790 XB_TRACE(bp, "no_daddr", data);
793 kmem_free(data, malloc_len);
801 * Increment reference count on buffer, to hold the buffer concurrently
802 * with another thread which may release (free) the buffer asynchronously.
803 * Must hold the buffer already to call this function.
809 atomic_inc(&bp->b_hold);
810 XB_TRACE(bp, "hold", 0);
814 * Releases a hold on the specified buffer. If the
815 * the hold count is 1, calls xfs_buf_free.
821 xfs_bufhash_t *hash = bp->b_hash;
823 XB_TRACE(bp, "rele", bp->b_relse);
825 if (atomic_dec_and_lock(&bp->b_hold, &hash->bh_lock)) {
827 atomic_inc(&bp->b_hold);
828 spin_unlock(&hash->bh_lock);
829 (*(bp->b_relse)) (bp);
830 } else if (bp->b_flags & XBF_FS_MANAGED) {
831 spin_unlock(&hash->bh_lock);
833 ASSERT(!(bp->b_flags & (XBF_DELWRI|_XBF_DELWRI_Q)));
834 list_del_init(&bp->b_hash_list);
835 spin_unlock(&hash->bh_lock);
840 * Catch reference count leaks
842 ASSERT(atomic_read(&bp->b_hold) >= 0);
848 * Mutual exclusion on buffers. Locking model:
850 * Buffers associated with inodes for which buffer locking
851 * is not enabled are not protected by semaphores, and are
852 * assumed to be exclusively owned by the caller. There is a
853 * spinlock in the buffer, used by the caller when concurrent
854 * access is possible.
858 * Locks a buffer object, if it is not already locked.
859 * Note that this in no way locks the underlying pages, so it is only
860 * useful for synchronizing concurrent use of buffer objects, not for
861 * synchronizing independent access to the underlying pages.
869 locked = down_trylock(&bp->b_sema) == 0;
873 XB_TRACE(bp, "cond_lock", (long)locked);
874 return locked ? 0 : -EBUSY;
877 #if defined(DEBUG) || defined(XFS_BLI_TRACE)
882 return atomic_read(&bp->b_sema.count);
887 * Locks a buffer object.
888 * Note that this in no way locks the underlying pages, so it is only
889 * useful for synchronizing concurrent use of buffer objects, not for
890 * synchronizing independent access to the underlying pages.
896 XB_TRACE(bp, "lock", 0);
897 if (atomic_read(&bp->b_io_remaining))
898 blk_run_address_space(bp->b_target->bt_mapping);
901 XB_TRACE(bp, "locked", 0);
905 * Releases the lock on the buffer object.
906 * If the buffer is marked delwri but is not queued, do so before we
907 * unlock the buffer as we need to set flags correctly. We also need to
908 * take a reference for the delwri queue because the unlocker is going to
909 * drop their's and they don't know we just queued it.
915 if ((bp->b_flags & (XBF_DELWRI|_XBF_DELWRI_Q)) == XBF_DELWRI) {
916 atomic_inc(&bp->b_hold);
917 bp->b_flags |= XBF_ASYNC;
918 xfs_buf_delwri_queue(bp, 0);
923 XB_TRACE(bp, "unlock", 0);
928 * Pinning Buffer Storage in Memory
929 * Ensure that no attempt to force a buffer to disk will succeed.
935 atomic_inc(&bp->b_pin_count);
936 XB_TRACE(bp, "pin", (long)bp->b_pin_count.counter);
943 if (atomic_dec_and_test(&bp->b_pin_count))
944 wake_up_all(&bp->b_waiters);
945 XB_TRACE(bp, "unpin", (long)bp->b_pin_count.counter);
952 return atomic_read(&bp->b_pin_count);
959 DECLARE_WAITQUEUE (wait, current);
961 if (atomic_read(&bp->b_pin_count) == 0)
964 add_wait_queue(&bp->b_waiters, &wait);
966 set_current_state(TASK_UNINTERRUPTIBLE);
967 if (atomic_read(&bp->b_pin_count) == 0)
969 if (atomic_read(&bp->b_io_remaining))
970 blk_run_address_space(bp->b_target->bt_mapping);
973 remove_wait_queue(&bp->b_waiters, &wait);
974 set_current_state(TASK_RUNNING);
978 * Buffer Utility Routines
985 xfs_buf_t *bp = (xfs_buf_t *)v;
988 (*(bp->b_iodone))(bp);
989 else if (bp->b_flags & XBF_ASYNC)
998 bp->b_flags &= ~(XBF_READ | XBF_WRITE);
999 if (bp->b_error == 0)
1000 bp->b_flags |= XBF_DONE;
1002 XB_TRACE(bp, "iodone", bp->b_iodone);
1004 if ((bp->b_iodone) || (bp->b_flags & XBF_ASYNC)) {
1006 INIT_WORK(&bp->b_iodone_work, xfs_buf_iodone_work, bp);
1007 queue_work(xfslogd_workqueue, &bp->b_iodone_work);
1009 xfs_buf_iodone_work(bp);
1012 up(&bp->b_iodonesema);
1021 ASSERT(error >= 0 && error <= 0xffff);
1022 bp->b_error = (unsigned short)error;
1023 XB_TRACE(bp, "ioerror", (unsigned long)error);
1027 * Initiate I/O on a buffer, based on the flags supplied.
1028 * The b_iodone routine in the buffer supplied will only be called
1029 * when all of the subsidiary I/O requests, if any, have been completed.
1034 xfs_buf_flags_t flags)
1038 XB_TRACE(bp, "iostart", (unsigned long)flags);
1040 if (flags & XBF_DELWRI) {
1041 bp->b_flags &= ~(XBF_READ | XBF_WRITE | XBF_ASYNC);
1042 bp->b_flags |= flags & (XBF_DELWRI | XBF_ASYNC);
1043 xfs_buf_delwri_queue(bp, 1);
1047 bp->b_flags &= ~(XBF_READ | XBF_WRITE | XBF_ASYNC | XBF_DELWRI | \
1048 XBF_READ_AHEAD | _XBF_RUN_QUEUES);
1049 bp->b_flags |= flags & (XBF_READ | XBF_WRITE | XBF_ASYNC | \
1050 XBF_READ_AHEAD | _XBF_RUN_QUEUES);
1052 BUG_ON(bp->b_bn == XFS_BUF_DADDR_NULL);
1054 /* For writes allow an alternate strategy routine to precede
1055 * the actual I/O request (which may not be issued at all in
1056 * a shutdown situation, for example).
1058 status = (flags & XBF_WRITE) ?
1059 xfs_buf_iostrategy(bp) : xfs_buf_iorequest(bp);
1061 /* Wait for I/O if we are not an async request.
1062 * Note: async I/O request completion will release the buffer,
1063 * and that can already be done by this point. So using the
1064 * buffer pointer from here on, after async I/O, is invalid.
1066 if (!status && !(flags & XBF_ASYNC))
1067 status = xfs_buf_iowait(bp);
1072 STATIC __inline__ int
1076 ASSERT(bp->b_flags & (XBF_READ | XBF_WRITE));
1077 if (bp->b_flags & XBF_READ)
1078 return bp->b_locked;
1082 STATIC __inline__ void
1087 if (atomic_dec_and_test(&bp->b_io_remaining) == 1) {
1089 xfs_buf_ioend(bp, schedule);
1096 unsigned int bytes_done,
1099 xfs_buf_t *bp = (xfs_buf_t *)bio->bi_private;
1100 unsigned int blocksize = bp->b_target->bt_bsize;
1101 struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
1106 if (!test_bit(BIO_UPTODATE, &bio->bi_flags))
1110 struct page *page = bvec->bv_page;
1112 if (unlikely(bp->b_error)) {
1113 if (bp->b_flags & XBF_READ)
1114 ClearPageUptodate(page);
1116 } else if (blocksize >= PAGE_CACHE_SIZE) {
1117 SetPageUptodate(page);
1118 } else if (!PagePrivate(page) &&
1119 (bp->b_flags & _XBF_PAGE_CACHE)) {
1120 set_page_region(page, bvec->bv_offset, bvec->bv_len);
1123 if (--bvec >= bio->bi_io_vec)
1124 prefetchw(&bvec->bv_page->flags);
1126 if (_xfs_buf_iolocked(bp)) {
1129 } while (bvec >= bio->bi_io_vec);
1131 _xfs_buf_ioend(bp, 1);
1140 int i, rw, map_i, total_nr_pages, nr_pages;
1142 int offset = bp->b_offset;
1143 int size = bp->b_count_desired;
1144 sector_t sector = bp->b_bn;
1145 unsigned int blocksize = bp->b_target->bt_bsize;
1146 int locking = _xfs_buf_iolocked(bp);
1148 total_nr_pages = bp->b_page_count;
1151 if (bp->b_flags & _XBF_RUN_QUEUES) {
1152 bp->b_flags &= ~_XBF_RUN_QUEUES;
1153 rw = (bp->b_flags & XBF_READ) ? READ_SYNC : WRITE_SYNC;
1155 rw = (bp->b_flags & XBF_READ) ? READ : WRITE;
1158 if (bp->b_flags & XBF_ORDERED) {
1159 ASSERT(!(bp->b_flags & XBF_READ));
1163 /* Special code path for reading a sub page size buffer in --
1164 * we populate up the whole page, and hence the other metadata
1165 * in the same page. This optimization is only valid when the
1166 * filesystem block size is not smaller than the page size.
1168 if ((bp->b_buffer_length < PAGE_CACHE_SIZE) &&
1169 (bp->b_flags & XBF_READ) && locking &&
1170 (blocksize >= PAGE_CACHE_SIZE)) {
1171 bio = bio_alloc(GFP_NOIO, 1);
1173 bio->bi_bdev = bp->b_target->bt_bdev;
1174 bio->bi_sector = sector - (offset >> BBSHIFT);
1175 bio->bi_end_io = xfs_buf_bio_end_io;
1176 bio->bi_private = bp;
1178 bio_add_page(bio, bp->b_pages[0], PAGE_CACHE_SIZE, 0);
1181 atomic_inc(&bp->b_io_remaining);
1186 /* Lock down the pages which we need to for the request */
1187 if (locking && (bp->b_flags & XBF_WRITE) && (bp->b_locked == 0)) {
1188 for (i = 0; size; i++) {
1189 int nbytes = PAGE_CACHE_SIZE - offset;
1190 struct page *page = bp->b_pages[i];
1200 offset = bp->b_offset;
1201 size = bp->b_count_desired;
1205 atomic_inc(&bp->b_io_remaining);
1206 nr_pages = BIO_MAX_SECTORS >> (PAGE_SHIFT - BBSHIFT);
1207 if (nr_pages > total_nr_pages)
1208 nr_pages = total_nr_pages;
1210 bio = bio_alloc(GFP_NOIO, nr_pages);
1211 bio->bi_bdev = bp->b_target->bt_bdev;
1212 bio->bi_sector = sector;
1213 bio->bi_end_io = xfs_buf_bio_end_io;
1214 bio->bi_private = bp;
1216 for (; size && nr_pages; nr_pages--, map_i++) {
1217 int rbytes, nbytes = PAGE_CACHE_SIZE - offset;
1222 rbytes = bio_add_page(bio, bp->b_pages[map_i], nbytes, offset);
1223 if (rbytes < nbytes)
1227 sector += nbytes >> BBSHIFT;
1233 if (likely(bio->bi_size)) {
1234 submit_bio(rw, bio);
1239 xfs_buf_ioerror(bp, EIO);
1247 XB_TRACE(bp, "iorequest", 0);
1249 if (bp->b_flags & XBF_DELWRI) {
1250 xfs_buf_delwri_queue(bp, 1);
1254 if (bp->b_flags & XBF_WRITE) {
1255 xfs_buf_wait_unpin(bp);
1260 /* Set the count to 1 initially, this will stop an I/O
1261 * completion callout which happens before we have started
1262 * all the I/O from calling xfs_buf_ioend too early.
1264 atomic_set(&bp->b_io_remaining, 1);
1265 _xfs_buf_ioapply(bp);
1266 _xfs_buf_ioend(bp, 0);
1273 * Waits for I/O to complete on the buffer supplied.
1274 * It returns immediately if no I/O is pending.
1275 * It returns the I/O error code, if any, or 0 if there was no error.
1281 XB_TRACE(bp, "iowait", 0);
1282 if (atomic_read(&bp->b_io_remaining))
1283 blk_run_address_space(bp->b_target->bt_mapping);
1284 down(&bp->b_iodonesema);
1285 XB_TRACE(bp, "iowaited", (long)bp->b_error);
1296 if (bp->b_flags & XBF_MAPPED)
1297 return XFS_BUF_PTR(bp) + offset;
1299 offset += bp->b_offset;
1300 page = bp->b_pages[offset >> PAGE_CACHE_SHIFT];
1301 return (xfs_caddr_t)page_address(page) + (offset & (PAGE_CACHE_SIZE-1));
1305 * Move data into or out of a buffer.
1309 xfs_buf_t *bp, /* buffer to process */
1310 size_t boff, /* starting buffer offset */
1311 size_t bsize, /* length to copy */
1312 caddr_t data, /* data address */
1313 xfs_buf_rw_t mode) /* read/write/zero flag */
1315 size_t bend, cpoff, csize;
1318 bend = boff + bsize;
1319 while (boff < bend) {
1320 page = bp->b_pages[xfs_buf_btoct(boff + bp->b_offset)];
1321 cpoff = xfs_buf_poff(boff + bp->b_offset);
1322 csize = min_t(size_t,
1323 PAGE_CACHE_SIZE-cpoff, bp->b_count_desired-boff);
1325 ASSERT(((csize + cpoff) <= PAGE_CACHE_SIZE));
1329 memset(page_address(page) + cpoff, 0, csize);
1332 memcpy(data, page_address(page) + cpoff, csize);
1335 memcpy(page_address(page) + cpoff, data, csize);
1344 * Handling of buffer targets (buftargs).
1348 * Wait for any bufs with callbacks that have been submitted but
1349 * have not yet returned... walk the hash list for the target.
1356 xfs_bufhash_t *hash;
1359 for (i = 0; i < (1 << btp->bt_hashshift); i++) {
1360 hash = &btp->bt_hash[i];
1362 spin_lock(&hash->bh_lock);
1363 list_for_each_entry_safe(bp, n, &hash->bh_list, b_hash_list) {
1364 ASSERT(btp == bp->b_target);
1365 if (!(bp->b_flags & XBF_FS_MANAGED)) {
1366 spin_unlock(&hash->bh_lock);
1368 * Catch superblock reference count leaks
1371 BUG_ON(bp->b_bn == 0);
1376 spin_unlock(&hash->bh_lock);
1381 * Allocate buffer hash table for a given target.
1382 * For devices containing metadata (i.e. not the log/realtime devices)
1383 * we need to allocate a much larger hash table.
1392 btp->bt_hashshift = external ? 3 : 8; /* 8 or 256 buckets */
1393 btp->bt_hashmask = (1 << btp->bt_hashshift) - 1;
1394 btp->bt_hash = kmem_zalloc((1 << btp->bt_hashshift) *
1395 sizeof(xfs_bufhash_t), KM_SLEEP);
1396 for (i = 0; i < (1 << btp->bt_hashshift); i++) {
1397 spin_lock_init(&btp->bt_hash[i].bh_lock);
1398 INIT_LIST_HEAD(&btp->bt_hash[i].bh_list);
1406 kmem_free(btp->bt_hash, (1<<btp->bt_hashshift) * sizeof(xfs_bufhash_t));
1407 btp->bt_hash = NULL;
1411 * buftarg list for delwrite queue processing
1413 STATIC LIST_HEAD(xfs_buftarg_list);
1414 STATIC DEFINE_SPINLOCK(xfs_buftarg_lock);
1417 xfs_register_buftarg(
1420 spin_lock(&xfs_buftarg_lock);
1421 list_add(&btp->bt_list, &xfs_buftarg_list);
1422 spin_unlock(&xfs_buftarg_lock);
1426 xfs_unregister_buftarg(
1429 spin_lock(&xfs_buftarg_lock);
1430 list_del(&btp->bt_list);
1431 spin_unlock(&xfs_buftarg_lock);
1439 xfs_flush_buftarg(btp, 1);
1441 xfs_blkdev_put(btp->bt_bdev);
1442 xfs_free_bufhash(btp);
1443 iput(btp->bt_mapping->host);
1445 /* Unregister the buftarg first so that we don't get a
1446 * wakeup finding a non-existent task
1448 xfs_unregister_buftarg(btp);
1449 kthread_stop(btp->bt_task);
1451 kmem_free(btp, sizeof(*btp));
1455 xfs_setsize_buftarg_flags(
1457 unsigned int blocksize,
1458 unsigned int sectorsize,
1461 btp->bt_bsize = blocksize;
1462 btp->bt_sshift = ffs(sectorsize) - 1;
1463 btp->bt_smask = sectorsize - 1;
1465 if (set_blocksize(btp->bt_bdev, sectorsize)) {
1467 "XFS: Cannot set_blocksize to %u on device %s\n",
1468 sectorsize, XFS_BUFTARG_NAME(btp));
1473 (PAGE_CACHE_SIZE / BITS_PER_LONG) > sectorsize) {
1475 "XFS: %u byte sectors in use on device %s. "
1476 "This is suboptimal; %u or greater is ideal.\n",
1477 sectorsize, XFS_BUFTARG_NAME(btp),
1478 (unsigned int)PAGE_CACHE_SIZE / BITS_PER_LONG);
1485 * When allocating the initial buffer target we have not yet
1486 * read in the superblock, so don't know what sized sectors
1487 * are being used is at this early stage. Play safe.
1490 xfs_setsize_buftarg_early(
1492 struct block_device *bdev)
1494 return xfs_setsize_buftarg_flags(btp,
1495 PAGE_CACHE_SIZE, bdev_hardsect_size(bdev), 0);
1499 xfs_setsize_buftarg(
1501 unsigned int blocksize,
1502 unsigned int sectorsize)
1504 return xfs_setsize_buftarg_flags(btp, blocksize, sectorsize, 1);
1508 xfs_mapping_buftarg(
1510 struct block_device *bdev)
1512 struct backing_dev_info *bdi;
1513 struct inode *inode;
1514 struct address_space *mapping;
1515 static struct address_space_operations mapping_aops = {
1516 .sync_page = block_sync_page,
1519 inode = new_inode(bdev->bd_inode->i_sb);
1522 "XFS: Cannot allocate mapping inode for device %s\n",
1523 XFS_BUFTARG_NAME(btp));
1526 inode->i_mode = S_IFBLK;
1527 inode->i_bdev = bdev;
1528 inode->i_rdev = bdev->bd_dev;
1529 bdi = blk_get_backing_dev_info(bdev);
1531 bdi = &default_backing_dev_info;
1532 mapping = &inode->i_data;
1533 mapping->a_ops = &mapping_aops;
1534 mapping->backing_dev_info = bdi;
1535 mapping_set_gfp_mask(mapping, GFP_NOFS);
1536 btp->bt_mapping = mapping;
1541 xfs_alloc_delwrite_queue(
1546 INIT_LIST_HEAD(&btp->bt_list);
1547 INIT_LIST_HEAD(&btp->bt_delwrite_queue);
1548 spinlock_init(&btp->bt_delwrite_lock, "delwri_lock");
1550 btp->bt_task = kthread_run(xfsbufd, btp, "xfsbufd");
1551 if (IS_ERR(btp->bt_task)) {
1552 error = PTR_ERR(btp->bt_task);
1555 xfs_register_buftarg(btp);
1562 struct block_device *bdev,
1567 btp = kmem_zalloc(sizeof(*btp), KM_SLEEP);
1569 btp->bt_dev = bdev->bd_dev;
1570 btp->bt_bdev = bdev;
1571 if (xfs_setsize_buftarg_early(btp, bdev))
1573 if (xfs_mapping_buftarg(btp, bdev))
1575 if (xfs_alloc_delwrite_queue(btp))
1577 xfs_alloc_bufhash(btp, external);
1581 kmem_free(btp, sizeof(*btp));
1587 * Delayed write buffer handling
1590 xfs_buf_delwri_queue(
1594 struct list_head *dwq = &bp->b_target->bt_delwrite_queue;
1595 spinlock_t *dwlk = &bp->b_target->bt_delwrite_lock;
1597 XB_TRACE(bp, "delwri_q", (long)unlock);
1598 ASSERT((bp->b_flags&(XBF_DELWRI|XBF_ASYNC)) == (XBF_DELWRI|XBF_ASYNC));
1601 /* If already in the queue, dequeue and place at tail */
1602 if (!list_empty(&bp->b_list)) {
1603 ASSERT(bp->b_flags & _XBF_DELWRI_Q);
1605 atomic_dec(&bp->b_hold);
1606 list_del(&bp->b_list);
1609 bp->b_flags |= _XBF_DELWRI_Q;
1610 list_add_tail(&bp->b_list, dwq);
1611 bp->b_queuetime = jiffies;
1619 xfs_buf_delwri_dequeue(
1622 spinlock_t *dwlk = &bp->b_target->bt_delwrite_lock;
1626 if ((bp->b_flags & XBF_DELWRI) && !list_empty(&bp->b_list)) {
1627 ASSERT(bp->b_flags & _XBF_DELWRI_Q);
1628 list_del_init(&bp->b_list);
1631 bp->b_flags &= ~(XBF_DELWRI|_XBF_DELWRI_Q);
1637 XB_TRACE(bp, "delwri_dq", (long)dequeued);
1641 xfs_buf_runall_queues(
1642 struct workqueue_struct *queue)
1644 flush_workqueue(queue);
1652 xfs_buftarg_t *btp, *n;
1654 spin_lock(&xfs_buftarg_lock);
1655 list_for_each_entry_safe(btp, n, &xfs_buftarg_list, bt_list) {
1656 if (test_bit(XBT_FORCE_SLEEP, &btp->bt_flags))
1658 set_bit(XBT_FORCE_FLUSH, &btp->bt_flags);
1660 wake_up_process(btp->bt_task);
1662 spin_unlock(&xfs_buftarg_lock);
1670 struct list_head tmp;
1672 xfs_buftarg_t *target = (xfs_buftarg_t *)data;
1674 struct list_head *dwq = &target->bt_delwrite_queue;
1675 spinlock_t *dwlk = &target->bt_delwrite_lock;
1677 current->flags |= PF_MEMALLOC;
1679 INIT_LIST_HEAD(&tmp);
1681 if (unlikely(freezing(current))) {
1682 set_bit(XBT_FORCE_SLEEP, &target->bt_flags);
1685 clear_bit(XBT_FORCE_SLEEP, &target->bt_flags);
1688 schedule_timeout_interruptible(
1689 xfs_buf_timer_centisecs * msecs_to_jiffies(10));
1691 age = xfs_buf_age_centisecs * msecs_to_jiffies(10);
1693 list_for_each_entry_safe(bp, n, dwq, b_list) {
1694 XB_TRACE(bp, "walkq1", (long)xfs_buf_ispin(bp));
1695 ASSERT(bp->b_flags & XBF_DELWRI);
1697 if (!xfs_buf_ispin(bp) && !xfs_buf_cond_lock(bp)) {
1698 if (!test_bit(XBT_FORCE_FLUSH,
1699 &target->bt_flags) &&
1700 time_before(jiffies,
1701 bp->b_queuetime + age)) {
1706 bp->b_flags &= ~(XBF_DELWRI|_XBF_DELWRI_Q);
1707 bp->b_flags |= XBF_WRITE;
1708 list_move(&bp->b_list, &tmp);
1713 while (!list_empty(&tmp)) {
1714 bp = list_entry(tmp.next, xfs_buf_t, b_list);
1715 ASSERT(target == bp->b_target);
1717 list_del_init(&bp->b_list);
1718 xfs_buf_iostrategy(bp);
1720 blk_run_address_space(target->bt_mapping);
1723 if (as_list_len > 0)
1726 clear_bit(XBT_FORCE_FLUSH, &target->bt_flags);
1727 } while (!kthread_should_stop());
1733 * Go through all incore buffers, and release buffers if they belong to
1734 * the given device. This is used in filesystem error handling to
1735 * preserve the consistency of its metadata.
1739 xfs_buftarg_t *target,
1742 struct list_head tmp;
1745 struct list_head *dwq = &target->bt_delwrite_queue;
1746 spinlock_t *dwlk = &target->bt_delwrite_lock;
1748 xfs_buf_runall_queues(xfsdatad_workqueue);
1749 xfs_buf_runall_queues(xfslogd_workqueue);
1751 INIT_LIST_HEAD(&tmp);
1753 list_for_each_entry_safe(bp, n, dwq, b_list) {
1754 ASSERT(bp->b_target == target);
1755 ASSERT(bp->b_flags & (XBF_DELWRI | _XBF_DELWRI_Q));
1756 XB_TRACE(bp, "walkq2", (long)xfs_buf_ispin(bp));
1757 if (xfs_buf_ispin(bp)) {
1762 list_move(&bp->b_list, &tmp);
1767 * Dropped the delayed write list lock, now walk the temporary list
1769 list_for_each_entry_safe(bp, n, &tmp, b_list) {
1771 bp->b_flags &= ~(XBF_DELWRI|_XBF_DELWRI_Q);
1772 bp->b_flags |= XBF_WRITE;
1774 bp->b_flags &= ~XBF_ASYNC;
1776 list_del_init(&bp->b_list);
1778 xfs_buf_iostrategy(bp);
1782 * Remaining list items must be flushed before returning
1784 while (!list_empty(&tmp)) {
1785 bp = list_entry(tmp.next, xfs_buf_t, b_list);
1787 list_del_init(&bp->b_list);
1793 blk_run_address_space(target->bt_mapping);
1801 int error = -ENOMEM;
1803 #ifdef XFS_BUF_TRACE
1804 xfs_buf_trace_buf = ktrace_alloc(XFS_BUF_TRACE_SIZE, KM_SLEEP);
1807 xfs_buf_zone = kmem_zone_init(sizeof(xfs_buf_t), "xfs_buf");
1809 goto out_free_trace_buf;
1811 xfslogd_workqueue = create_workqueue("xfslogd");
1812 if (!xfslogd_workqueue)
1813 goto out_free_buf_zone;
1815 xfsdatad_workqueue = create_workqueue("xfsdatad");
1816 if (!xfsdatad_workqueue)
1817 goto out_destroy_xfslogd_workqueue;
1819 xfs_buf_shake = kmem_shake_register(xfsbufd_wakeup);
1821 goto out_destroy_xfsdatad_workqueue;
1825 out_destroy_xfsdatad_workqueue:
1826 destroy_workqueue(xfsdatad_workqueue);
1827 out_destroy_xfslogd_workqueue:
1828 destroy_workqueue(xfslogd_workqueue);
1830 kmem_zone_destroy(xfs_buf_zone);
1832 #ifdef XFS_BUF_TRACE
1833 ktrace_free(xfs_buf_trace_buf);
1839 xfs_buf_terminate(void)
1841 kmem_shake_deregister(xfs_buf_shake);
1842 destroy_workqueue(xfsdatad_workqueue);
1843 destroy_workqueue(xfslogd_workqueue);
1844 kmem_zone_destroy(xfs_buf_zone);
1845 #ifdef XFS_BUF_TRACE
1846 ktrace_free(xfs_buf_trace_buf);