11da177e4SLinus Torvalds /* 21da177e4SLinus Torvalds * linux/fs/buffer.c 31da177e4SLinus Torvalds * 41da177e4SLinus Torvalds * Copyright (C) 1991, 1992, 2002 Linus Torvalds 51da177e4SLinus Torvalds */ 61da177e4SLinus Torvalds 71da177e4SLinus Torvalds /* 81da177e4SLinus Torvalds * Start bdflush() with kernel_thread not syscall - Paul Gortmaker, 12/95 91da177e4SLinus Torvalds * 101da177e4SLinus Torvalds * Removed a lot of unnecessary code and simplified things now that 111da177e4SLinus Torvalds * the buffer cache isn't our primary cache - Andrew Tridgell 12/96 121da177e4SLinus Torvalds * 131da177e4SLinus Torvalds * Speed up hash, lru, and free list operations. Use gfp() for allocating 141da177e4SLinus Torvalds * hash table, use SLAB cache for buffer heads. SMP threading. -DaveM 151da177e4SLinus Torvalds * 161da177e4SLinus Torvalds * Added 32k buffer block sizes - these are required older ARM systems. - RMK 171da177e4SLinus Torvalds * 181da177e4SLinus Torvalds * async buffer flushing, 1999 Andrea Arcangeli <andrea@suse.de> 191da177e4SLinus Torvalds */ 201da177e4SLinus Torvalds 211da177e4SLinus Torvalds #include <linux/kernel.h> 221da177e4SLinus Torvalds #include <linux/syscalls.h> 231da177e4SLinus Torvalds #include <linux/fs.h> 241da177e4SLinus Torvalds #include <linux/mm.h> 251da177e4SLinus Torvalds #include <linux/percpu.h> 261da177e4SLinus Torvalds #include <linux/slab.h> 2716f7e0feSRandy Dunlap #include <linux/capability.h> 281da177e4SLinus Torvalds #include <linux/blkdev.h> 291da177e4SLinus Torvalds #include <linux/file.h> 301da177e4SLinus Torvalds #include <linux/quotaops.h> 311da177e4SLinus Torvalds #include <linux/highmem.h> 321da177e4SLinus Torvalds #include <linux/module.h> 331da177e4SLinus Torvalds #include <linux/writeback.h> 341da177e4SLinus Torvalds #include <linux/hash.h> 351da177e4SLinus Torvalds #include <linux/suspend.h> 361da177e4SLinus Torvalds #include <linux/buffer_head.h> 3755e829afSAndrew Morton #include <linux/task_io_accounting_ops.h> 381da177e4SLinus Torvalds #include <linux/bio.h> 391da177e4SLinus Torvalds #include <linux/notifier.h> 401da177e4SLinus Torvalds #include <linux/cpu.h> 411da177e4SLinus Torvalds #include <linux/bitops.h> 421da177e4SLinus Torvalds #include <linux/mpage.h> 43fb1c8f93SIngo Molnar #include <linux/bit_spinlock.h> 44c515e1fdSDan Magenheimer #include <linux/cleancache.h> 451da177e4SLinus Torvalds 461da177e4SLinus Torvalds static int fsync_buffers_list(spinlock_t *lock, struct list_head *list); 471da177e4SLinus Torvalds 481da177e4SLinus Torvalds #define BH_ENTRY(list) list_entry((list), struct buffer_head, b_assoc_buffers) 491da177e4SLinus Torvalds 501da177e4SLinus Torvalds inline void 511da177e4SLinus Torvalds init_buffer(struct buffer_head *bh, bh_end_io_t *handler, void *private) 521da177e4SLinus Torvalds { 531da177e4SLinus Torvalds bh->b_end_io = handler; 541da177e4SLinus Torvalds bh->b_private = private; 551da177e4SLinus Torvalds } 561fe72eaaSH Hartley Sweeten EXPORT_SYMBOL(init_buffer); 571da177e4SLinus Torvalds 587eaceaccSJens Axboe static int sleep_on_buffer(void *word) 591da177e4SLinus Torvalds { 601da177e4SLinus Torvalds io_schedule(); 611da177e4SLinus Torvalds return 0; 621da177e4SLinus Torvalds } 631da177e4SLinus Torvalds 64fc9b52cdSHarvey Harrison void __lock_buffer(struct buffer_head *bh) 651da177e4SLinus Torvalds { 667eaceaccSJens Axboe wait_on_bit_lock(&bh->b_state, BH_Lock, sleep_on_buffer, 671da177e4SLinus Torvalds TASK_UNINTERRUPTIBLE); 681da177e4SLinus Torvalds } 691da177e4SLinus Torvalds EXPORT_SYMBOL(__lock_buffer); 701da177e4SLinus Torvalds 71fc9b52cdSHarvey Harrison void unlock_buffer(struct buffer_head *bh) 721da177e4SLinus Torvalds { 7351b07fc3SNick Piggin clear_bit_unlock(BH_Lock, &bh->b_state); 741da177e4SLinus Torvalds smp_mb__after_clear_bit(); 751da177e4SLinus Torvalds wake_up_bit(&bh->b_state, BH_Lock); 761da177e4SLinus Torvalds } 771fe72eaaSH Hartley Sweeten EXPORT_SYMBOL(unlock_buffer); 781da177e4SLinus Torvalds 791da177e4SLinus Torvalds /* 801da177e4SLinus Torvalds * Block until a buffer comes unlocked. This doesn't stop it 811da177e4SLinus Torvalds * from becoming locked again - you have to lock it yourself 821da177e4SLinus Torvalds * if you want to preserve its state. 831da177e4SLinus Torvalds */ 841da177e4SLinus Torvalds void __wait_on_buffer(struct buffer_head * bh) 851da177e4SLinus Torvalds { 867eaceaccSJens Axboe wait_on_bit(&bh->b_state, BH_Lock, sleep_on_buffer, TASK_UNINTERRUPTIBLE); 871da177e4SLinus Torvalds } 881fe72eaaSH Hartley Sweeten EXPORT_SYMBOL(__wait_on_buffer); 891da177e4SLinus Torvalds 901da177e4SLinus Torvalds static void 911da177e4SLinus Torvalds __clear_page_buffers(struct page *page) 921da177e4SLinus Torvalds { 931da177e4SLinus Torvalds ClearPagePrivate(page); 944c21e2f2SHugh Dickins set_page_private(page, 0); 951da177e4SLinus Torvalds page_cache_release(page); 961da177e4SLinus Torvalds } 971da177e4SLinus Torvalds 9808bafc03SKeith Mannthey 9908bafc03SKeith Mannthey static int quiet_error(struct buffer_head *bh) 10008bafc03SKeith Mannthey { 10108bafc03SKeith Mannthey if (!test_bit(BH_Quiet, &bh->b_state) && printk_ratelimit()) 10208bafc03SKeith Mannthey return 0; 10308bafc03SKeith Mannthey return 1; 10408bafc03SKeith Mannthey } 10508bafc03SKeith Mannthey 10608bafc03SKeith Mannthey 1071da177e4SLinus Torvalds static void buffer_io_error(struct buffer_head *bh) 1081da177e4SLinus Torvalds { 1091da177e4SLinus Torvalds char b[BDEVNAME_SIZE]; 1101da177e4SLinus Torvalds printk(KERN_ERR "Buffer I/O error on device %s, logical block %Lu\n", 1111da177e4SLinus Torvalds bdevname(bh->b_bdev, b), 1121da177e4SLinus Torvalds (unsigned long long)bh->b_blocknr); 1131da177e4SLinus Torvalds } 1141da177e4SLinus Torvalds 1151da177e4SLinus Torvalds /* 11668671f35SDmitry Monakhov * End-of-IO handler helper function which does not touch the bh after 11768671f35SDmitry Monakhov * unlocking it. 11868671f35SDmitry Monakhov * Note: unlock_buffer() sort-of does touch the bh after unlocking it, but 11968671f35SDmitry Monakhov * a race there is benign: unlock_buffer() only use the bh's address for 12068671f35SDmitry Monakhov * hashing after unlocking the buffer, so it doesn't actually touch the bh 12168671f35SDmitry Monakhov * itself. 1221da177e4SLinus Torvalds */ 12368671f35SDmitry Monakhov static void __end_buffer_read_notouch(struct buffer_head *bh, int uptodate) 1241da177e4SLinus Torvalds { 1251da177e4SLinus Torvalds if (uptodate) { 1261da177e4SLinus Torvalds set_buffer_uptodate(bh); 1271da177e4SLinus Torvalds } else { 1281da177e4SLinus Torvalds /* This happens, due to failed READA attempts. */ 1291da177e4SLinus Torvalds clear_buffer_uptodate(bh); 1301da177e4SLinus Torvalds } 1311da177e4SLinus Torvalds unlock_buffer(bh); 13268671f35SDmitry Monakhov } 13368671f35SDmitry Monakhov 13468671f35SDmitry Monakhov /* 13568671f35SDmitry Monakhov * Default synchronous end-of-IO handler.. Just mark it up-to-date and 13668671f35SDmitry Monakhov * unlock the buffer. This is what ll_rw_block uses too. 13768671f35SDmitry Monakhov */ 13868671f35SDmitry Monakhov void end_buffer_read_sync(struct buffer_head *bh, int uptodate) 13968671f35SDmitry Monakhov { 14068671f35SDmitry Monakhov __end_buffer_read_notouch(bh, uptodate); 1411da177e4SLinus Torvalds put_bh(bh); 1421da177e4SLinus Torvalds } 1431fe72eaaSH Hartley Sweeten EXPORT_SYMBOL(end_buffer_read_sync); 1441da177e4SLinus Torvalds 1451da177e4SLinus Torvalds void end_buffer_write_sync(struct buffer_head *bh, int uptodate) 1461da177e4SLinus Torvalds { 1471da177e4SLinus Torvalds char b[BDEVNAME_SIZE]; 1481da177e4SLinus Torvalds 1491da177e4SLinus Torvalds if (uptodate) { 1501da177e4SLinus Torvalds set_buffer_uptodate(bh); 1511da177e4SLinus Torvalds } else { 1520edd55faSChristoph Hellwig if (!quiet_error(bh)) { 1531da177e4SLinus Torvalds buffer_io_error(bh); 1541da177e4SLinus Torvalds printk(KERN_WARNING "lost page write due to " 1551da177e4SLinus Torvalds "I/O error on %s\n", 1561da177e4SLinus Torvalds bdevname(bh->b_bdev, b)); 1571da177e4SLinus Torvalds } 1581da177e4SLinus Torvalds set_buffer_write_io_error(bh); 1591da177e4SLinus Torvalds clear_buffer_uptodate(bh); 1601da177e4SLinus Torvalds } 1611da177e4SLinus Torvalds unlock_buffer(bh); 1621da177e4SLinus Torvalds put_bh(bh); 1631da177e4SLinus Torvalds } 1641fe72eaaSH Hartley Sweeten EXPORT_SYMBOL(end_buffer_write_sync); 1651da177e4SLinus Torvalds 1661da177e4SLinus Torvalds /* 1671da177e4SLinus Torvalds * Various filesystems appear to want __find_get_block to be non-blocking. 1681da177e4SLinus Torvalds * But it's the page lock which protects the buffers. To get around this, 1691da177e4SLinus Torvalds * we get exclusion from try_to_free_buffers with the blockdev mapping's 1701da177e4SLinus Torvalds * private_lock. 1711da177e4SLinus Torvalds * 1721da177e4SLinus Torvalds * Hack idea: for the blockdev mapping, i_bufferlist_lock contention 1731da177e4SLinus Torvalds * may be quite high. This code could TryLock the page, and if that 1741da177e4SLinus Torvalds * succeeds, there is no need to take private_lock. (But if 1751da177e4SLinus Torvalds * private_lock is contended then so is mapping->tree_lock). 1761da177e4SLinus Torvalds */ 1771da177e4SLinus Torvalds static struct buffer_head * 178385fd4c5SCoywolf Qi Hunt __find_get_block_slow(struct block_device *bdev, sector_t block) 1791da177e4SLinus Torvalds { 1801da177e4SLinus Torvalds struct inode *bd_inode = bdev->bd_inode; 1811da177e4SLinus Torvalds struct address_space *bd_mapping = bd_inode->i_mapping; 1821da177e4SLinus Torvalds struct buffer_head *ret = NULL; 1831da177e4SLinus Torvalds pgoff_t index; 1841da177e4SLinus Torvalds struct buffer_head *bh; 1851da177e4SLinus Torvalds struct buffer_head *head; 1861da177e4SLinus Torvalds struct page *page; 1871da177e4SLinus Torvalds int all_mapped = 1; 1881da177e4SLinus Torvalds 1891da177e4SLinus Torvalds index = block >> (PAGE_CACHE_SHIFT - bd_inode->i_blkbits); 1901da177e4SLinus Torvalds page = find_get_page(bd_mapping, index); 1911da177e4SLinus Torvalds if (!page) 1921da177e4SLinus Torvalds goto out; 1931da177e4SLinus Torvalds 1941da177e4SLinus Torvalds spin_lock(&bd_mapping->private_lock); 1951da177e4SLinus Torvalds if (!page_has_buffers(page)) 1961da177e4SLinus Torvalds goto out_unlock; 1971da177e4SLinus Torvalds head = page_buffers(page); 1981da177e4SLinus Torvalds bh = head; 1991da177e4SLinus Torvalds do { 20097f76d3dSNikanth Karthikesan if (!buffer_mapped(bh)) 20197f76d3dSNikanth Karthikesan all_mapped = 0; 20297f76d3dSNikanth Karthikesan else if (bh->b_blocknr == block) { 2031da177e4SLinus Torvalds ret = bh; 2041da177e4SLinus Torvalds get_bh(bh); 2051da177e4SLinus Torvalds goto out_unlock; 2061da177e4SLinus Torvalds } 2071da177e4SLinus Torvalds bh = bh->b_this_page; 2081da177e4SLinus Torvalds } while (bh != head); 2091da177e4SLinus Torvalds 2101da177e4SLinus Torvalds /* we might be here because some of the buffers on this page are 2111da177e4SLinus Torvalds * not mapped. This is due to various races between 2121da177e4SLinus Torvalds * file io on the block device and getblk. It gets dealt with 2131da177e4SLinus Torvalds * elsewhere, don't buffer_error if we had some unmapped buffers 2141da177e4SLinus Torvalds */ 2151da177e4SLinus Torvalds if (all_mapped) { 2161da177e4SLinus Torvalds printk("__find_get_block_slow() failed. " 2171da177e4SLinus Torvalds "block=%llu, b_blocknr=%llu\n", 218205f87f6SBadari Pulavarty (unsigned long long)block, 219205f87f6SBadari Pulavarty (unsigned long long)bh->b_blocknr); 220205f87f6SBadari Pulavarty printk("b_state=0x%08lx, b_size=%zu\n", 221205f87f6SBadari Pulavarty bh->b_state, bh->b_size); 2221da177e4SLinus Torvalds printk("device blocksize: %d\n", 1 << bd_inode->i_blkbits); 2231da177e4SLinus Torvalds } 2241da177e4SLinus Torvalds out_unlock: 2251da177e4SLinus Torvalds spin_unlock(&bd_mapping->private_lock); 2261da177e4SLinus Torvalds page_cache_release(page); 2271da177e4SLinus Torvalds out: 2281da177e4SLinus Torvalds return ret; 2291da177e4SLinus Torvalds } 2301da177e4SLinus Torvalds 2311da177e4SLinus Torvalds /* If invalidate_buffers() will trash dirty buffers, it means some kind 2321da177e4SLinus Torvalds of fs corruption is going on. Trashing dirty data always imply losing 2331da177e4SLinus Torvalds information that was supposed to be just stored on the physical layer 2341da177e4SLinus Torvalds by the user. 2351da177e4SLinus Torvalds 2361da177e4SLinus Torvalds Thus invalidate_buffers in general usage is not allwowed to trash 2371da177e4SLinus Torvalds dirty buffers. For example ioctl(FLSBLKBUF) expects dirty data to 2381da177e4SLinus Torvalds be preserved. These buffers are simply skipped. 2391da177e4SLinus Torvalds 2401da177e4SLinus Torvalds We also skip buffers which are still in use. For example this can 2411da177e4SLinus Torvalds happen if a userspace program is reading the block device. 2421da177e4SLinus Torvalds 2431da177e4SLinus Torvalds NOTE: In the case where the user removed a removable-media-disk even if 2441da177e4SLinus Torvalds there's still dirty data not synced on disk (due a bug in the device driver 2451da177e4SLinus Torvalds or due an error of the user), by not destroying the dirty buffers we could 2461da177e4SLinus Torvalds generate corruption also on the next media inserted, thus a parameter is 2471da177e4SLinus Torvalds necessary to handle this case in the most safe way possible (trying 2481da177e4SLinus Torvalds to not corrupt also the new disk inserted with the data belonging to 2491da177e4SLinus Torvalds the old now corrupted disk). Also for the ramdisk the natural thing 2501da177e4SLinus Torvalds to do in order to release the ramdisk memory is to destroy dirty buffers. 2511da177e4SLinus Torvalds 2521da177e4SLinus Torvalds These are two special cases. Normal usage imply the device driver 2531da177e4SLinus Torvalds to issue a sync on the device (without waiting I/O completion) and 2541da177e4SLinus Torvalds then an invalidate_buffers call that doesn't trash dirty buffers. 2551da177e4SLinus Torvalds 2561da177e4SLinus Torvalds For handling cache coherency with the blkdev pagecache the 'update' case 2571da177e4SLinus Torvalds is been introduced. It is needed to re-read from disk any pinned 2581da177e4SLinus Torvalds buffer. NOTE: re-reading from disk is destructive so we can do it only 2591da177e4SLinus Torvalds when we assume nobody is changing the buffercache under our I/O and when 2601da177e4SLinus Torvalds we think the disk contains more recent information than the buffercache. 2611da177e4SLinus Torvalds The update == 1 pass marks the buffers we need to update, the update == 2 2621da177e4SLinus Torvalds pass does the actual I/O. */ 263f98393a6SPeter Zijlstra void invalidate_bdev(struct block_device *bdev) 2641da177e4SLinus Torvalds { 2650e1dfc66SAndrew Morton struct address_space *mapping = bdev->bd_inode->i_mapping; 2660e1dfc66SAndrew Morton 2670e1dfc66SAndrew Morton if (mapping->nrpages == 0) 2680e1dfc66SAndrew Morton return; 2690e1dfc66SAndrew Morton 2701da177e4SLinus Torvalds invalidate_bh_lrus(); 271fa4b9074STejun Heo lru_add_drain_all(); /* make sure all lru add caches are flushed */ 272fc0ecff6SAndrew Morton invalidate_mapping_pages(mapping, 0, -1); 273c515e1fdSDan Magenheimer /* 99% of the time, we don't need to flush the cleancache on the bdev. 274c515e1fdSDan Magenheimer * But, for the strange corners, lets be cautious 275c515e1fdSDan Magenheimer */ 276c515e1fdSDan Magenheimer cleancache_flush_inode(mapping); 2771da177e4SLinus Torvalds } 2781fe72eaaSH Hartley Sweeten EXPORT_SYMBOL(invalidate_bdev); 2791da177e4SLinus Torvalds 2801da177e4SLinus Torvalds /* 2815b0830cbSJens Axboe * Kick the writeback threads then try to free up some ZONE_NORMAL memory. 2821da177e4SLinus Torvalds */ 2831da177e4SLinus Torvalds static void free_more_memory(void) 2841da177e4SLinus Torvalds { 28519770b32SMel Gorman struct zone *zone; 2860e88460dSMel Gorman int nid; 2871da177e4SLinus Torvalds 28803ba3782SJens Axboe wakeup_flusher_threads(1024); 2891da177e4SLinus Torvalds yield(); 2901da177e4SLinus Torvalds 2910e88460dSMel Gorman for_each_online_node(nid) { 29219770b32SMel Gorman (void)first_zones_zonelist(node_zonelist(nid, GFP_NOFS), 29319770b32SMel Gorman gfp_zone(GFP_NOFS), NULL, 29419770b32SMel Gorman &zone); 29519770b32SMel Gorman if (zone) 29654a6eb5cSMel Gorman try_to_free_pages(node_zonelist(nid, GFP_NOFS), 0, 297327c0e96SKAMEZAWA Hiroyuki GFP_NOFS, NULL); 2981da177e4SLinus Torvalds } 2991da177e4SLinus Torvalds } 3001da177e4SLinus Torvalds 3011da177e4SLinus Torvalds /* 3021da177e4SLinus Torvalds * I/O completion handler for block_read_full_page() - pages 3031da177e4SLinus Torvalds * which come unlocked at the end of I/O. 3041da177e4SLinus Torvalds */ 3051da177e4SLinus Torvalds static void end_buffer_async_read(struct buffer_head *bh, int uptodate) 3061da177e4SLinus Torvalds { 3071da177e4SLinus Torvalds unsigned long flags; 308a3972203SNick Piggin struct buffer_head *first; 3091da177e4SLinus Torvalds struct buffer_head *tmp; 3101da177e4SLinus Torvalds struct page *page; 3111da177e4SLinus Torvalds int page_uptodate = 1; 3121da177e4SLinus Torvalds 3131da177e4SLinus Torvalds BUG_ON(!buffer_async_read(bh)); 3141da177e4SLinus Torvalds 3151da177e4SLinus Torvalds page = bh->b_page; 3161da177e4SLinus Torvalds if (uptodate) { 3171da177e4SLinus Torvalds set_buffer_uptodate(bh); 3181da177e4SLinus Torvalds } else { 3191da177e4SLinus Torvalds clear_buffer_uptodate(bh); 32008bafc03SKeith Mannthey if (!quiet_error(bh)) 3211da177e4SLinus Torvalds buffer_io_error(bh); 3221da177e4SLinus Torvalds SetPageError(page); 3231da177e4SLinus Torvalds } 3241da177e4SLinus Torvalds 3251da177e4SLinus Torvalds /* 3261da177e4SLinus Torvalds * Be _very_ careful from here on. Bad things can happen if 3271da177e4SLinus Torvalds * two buffer heads end IO at almost the same time and both 3281da177e4SLinus Torvalds * decide that the page is now completely done. 3291da177e4SLinus Torvalds */ 330a3972203SNick Piggin first = page_buffers(page); 331a3972203SNick Piggin local_irq_save(flags); 332a3972203SNick Piggin bit_spin_lock(BH_Uptodate_Lock, &first->b_state); 3331da177e4SLinus Torvalds clear_buffer_async_read(bh); 3341da177e4SLinus Torvalds unlock_buffer(bh); 3351da177e4SLinus Torvalds tmp = bh; 3361da177e4SLinus Torvalds do { 3371da177e4SLinus Torvalds if (!buffer_uptodate(tmp)) 3381da177e4SLinus Torvalds page_uptodate = 0; 3391da177e4SLinus Torvalds if (buffer_async_read(tmp)) { 3401da177e4SLinus Torvalds BUG_ON(!buffer_locked(tmp)); 3411da177e4SLinus Torvalds goto still_busy; 3421da177e4SLinus Torvalds } 3431da177e4SLinus Torvalds tmp = tmp->b_this_page; 3441da177e4SLinus Torvalds } while (tmp != bh); 345a3972203SNick Piggin bit_spin_unlock(BH_Uptodate_Lock, &first->b_state); 346a3972203SNick Piggin local_irq_restore(flags); 3471da177e4SLinus Torvalds 3481da177e4SLinus Torvalds /* 3491da177e4SLinus Torvalds * If none of the buffers had errors and they are all 3501da177e4SLinus Torvalds * uptodate then we can set the page uptodate. 3511da177e4SLinus Torvalds */ 3521da177e4SLinus Torvalds if (page_uptodate && !PageError(page)) 3531da177e4SLinus Torvalds SetPageUptodate(page); 3541da177e4SLinus Torvalds unlock_page(page); 3551da177e4SLinus Torvalds return; 3561da177e4SLinus Torvalds 3571da177e4SLinus Torvalds still_busy: 358a3972203SNick Piggin bit_spin_unlock(BH_Uptodate_Lock, &first->b_state); 359a3972203SNick Piggin local_irq_restore(flags); 3601da177e4SLinus Torvalds return; 3611da177e4SLinus Torvalds } 3621da177e4SLinus Torvalds 3631da177e4SLinus Torvalds /* 3641da177e4SLinus Torvalds * Completion handler for block_write_full_page() - pages which are unlocked 3651da177e4SLinus Torvalds * during I/O, and which have PageWriteback cleared upon I/O completion. 3661da177e4SLinus Torvalds */ 36735c80d5fSChris Mason void end_buffer_async_write(struct buffer_head *bh, int uptodate) 3681da177e4SLinus Torvalds { 3691da177e4SLinus Torvalds char b[BDEVNAME_SIZE]; 3701da177e4SLinus Torvalds unsigned long flags; 371a3972203SNick Piggin struct buffer_head *first; 3721da177e4SLinus Torvalds struct buffer_head *tmp; 3731da177e4SLinus Torvalds struct page *page; 3741da177e4SLinus Torvalds 3751da177e4SLinus Torvalds BUG_ON(!buffer_async_write(bh)); 3761da177e4SLinus Torvalds 3771da177e4SLinus Torvalds page = bh->b_page; 3781da177e4SLinus Torvalds if (uptodate) { 3791da177e4SLinus Torvalds set_buffer_uptodate(bh); 3801da177e4SLinus Torvalds } else { 38108bafc03SKeith Mannthey if (!quiet_error(bh)) { 3821da177e4SLinus Torvalds buffer_io_error(bh); 3831da177e4SLinus Torvalds printk(KERN_WARNING "lost page write due to " 3841da177e4SLinus Torvalds "I/O error on %s\n", 3851da177e4SLinus Torvalds bdevname(bh->b_bdev, b)); 3861da177e4SLinus Torvalds } 3871da177e4SLinus Torvalds set_bit(AS_EIO, &page->mapping->flags); 38858ff407bSJan Kara set_buffer_write_io_error(bh); 3891da177e4SLinus Torvalds clear_buffer_uptodate(bh); 3901da177e4SLinus Torvalds SetPageError(page); 3911da177e4SLinus Torvalds } 3921da177e4SLinus Torvalds 393a3972203SNick Piggin first = page_buffers(page); 394a3972203SNick Piggin local_irq_save(flags); 395a3972203SNick Piggin bit_spin_lock(BH_Uptodate_Lock, &first->b_state); 396a3972203SNick Piggin 3971da177e4SLinus Torvalds clear_buffer_async_write(bh); 3981da177e4SLinus Torvalds unlock_buffer(bh); 3991da177e4SLinus Torvalds tmp = bh->b_this_page; 4001da177e4SLinus Torvalds while (tmp != bh) { 4011da177e4SLinus Torvalds if (buffer_async_write(tmp)) { 4021da177e4SLinus Torvalds BUG_ON(!buffer_locked(tmp)); 4031da177e4SLinus Torvalds goto still_busy; 4041da177e4SLinus Torvalds } 4051da177e4SLinus Torvalds tmp = tmp->b_this_page; 4061da177e4SLinus Torvalds } 407a3972203SNick Piggin bit_spin_unlock(BH_Uptodate_Lock, &first->b_state); 408a3972203SNick Piggin local_irq_restore(flags); 4091da177e4SLinus Torvalds end_page_writeback(page); 4101da177e4SLinus Torvalds return; 4111da177e4SLinus Torvalds 4121da177e4SLinus Torvalds still_busy: 413a3972203SNick Piggin bit_spin_unlock(BH_Uptodate_Lock, &first->b_state); 414a3972203SNick Piggin local_irq_restore(flags); 4151da177e4SLinus Torvalds return; 4161da177e4SLinus Torvalds } 4171fe72eaaSH Hartley Sweeten EXPORT_SYMBOL(end_buffer_async_write); 4181da177e4SLinus Torvalds 4191da177e4SLinus Torvalds /* 4201da177e4SLinus Torvalds * If a page's buffers are under async readin (end_buffer_async_read 4211da177e4SLinus Torvalds * completion) then there is a possibility that another thread of 4221da177e4SLinus Torvalds * control could lock one of the buffers after it has completed 4231da177e4SLinus Torvalds * but while some of the other buffers have not completed. This 4241da177e4SLinus Torvalds * locked buffer would confuse end_buffer_async_read() into not unlocking 4251da177e4SLinus Torvalds * the page. So the absence of BH_Async_Read tells end_buffer_async_read() 4261da177e4SLinus Torvalds * that this buffer is not under async I/O. 4271da177e4SLinus Torvalds * 4281da177e4SLinus Torvalds * The page comes unlocked when it has no locked buffer_async buffers 4291da177e4SLinus Torvalds * left. 4301da177e4SLinus Torvalds * 4311da177e4SLinus Torvalds * PageLocked prevents anyone starting new async I/O reads any of 4321da177e4SLinus Torvalds * the buffers. 4331da177e4SLinus Torvalds * 4341da177e4SLinus Torvalds * PageWriteback is used to prevent simultaneous writeout of the same 4351da177e4SLinus Torvalds * page. 4361da177e4SLinus Torvalds * 4371da177e4SLinus Torvalds * PageLocked prevents anyone from starting writeback of a page which is 4381da177e4SLinus Torvalds * under read I/O (PageWriteback is only ever set against a locked page). 4391da177e4SLinus Torvalds */ 4401da177e4SLinus Torvalds static void mark_buffer_async_read(struct buffer_head *bh) 4411da177e4SLinus Torvalds { 4421da177e4SLinus Torvalds bh->b_end_io = end_buffer_async_read; 4431da177e4SLinus Torvalds set_buffer_async_read(bh); 4441da177e4SLinus Torvalds } 4451da177e4SLinus Torvalds 4461fe72eaaSH Hartley Sweeten static void mark_buffer_async_write_endio(struct buffer_head *bh, 44735c80d5fSChris Mason bh_end_io_t *handler) 44835c80d5fSChris Mason { 44935c80d5fSChris Mason bh->b_end_io = handler; 45035c80d5fSChris Mason set_buffer_async_write(bh); 45135c80d5fSChris Mason } 45235c80d5fSChris Mason 4531da177e4SLinus Torvalds void mark_buffer_async_write(struct buffer_head *bh) 4541da177e4SLinus Torvalds { 45535c80d5fSChris Mason mark_buffer_async_write_endio(bh, end_buffer_async_write); 4561da177e4SLinus Torvalds } 4571da177e4SLinus Torvalds EXPORT_SYMBOL(mark_buffer_async_write); 4581da177e4SLinus Torvalds 4591da177e4SLinus Torvalds 4601da177e4SLinus Torvalds /* 4611da177e4SLinus Torvalds * fs/buffer.c contains helper functions for buffer-backed address space's 4621da177e4SLinus Torvalds * fsync functions. A common requirement for buffer-based filesystems is 4631da177e4SLinus Torvalds * that certain data from the backing blockdev needs to be written out for 4641da177e4SLinus Torvalds * a successful fsync(). For example, ext2 indirect blocks need to be 4651da177e4SLinus Torvalds * written back and waited upon before fsync() returns. 4661da177e4SLinus Torvalds * 4671da177e4SLinus Torvalds * The functions mark_buffer_inode_dirty(), fsync_inode_buffers(), 4681da177e4SLinus Torvalds * inode_has_buffers() and invalidate_inode_buffers() are provided for the 4691da177e4SLinus Torvalds * management of a list of dependent buffers at ->i_mapping->private_list. 4701da177e4SLinus Torvalds * 4711da177e4SLinus Torvalds * Locking is a little subtle: try_to_free_buffers() will remove buffers 4721da177e4SLinus Torvalds * from their controlling inode's queue when they are being freed. But 4731da177e4SLinus Torvalds * try_to_free_buffers() will be operating against the *blockdev* mapping 4741da177e4SLinus Torvalds * at the time, not against the S_ISREG file which depends on those buffers. 4751da177e4SLinus Torvalds * So the locking for private_list is via the private_lock in the address_space 4761da177e4SLinus Torvalds * which backs the buffers. Which is different from the address_space 4771da177e4SLinus Torvalds * against which the buffers are listed. So for a particular address_space, 4781da177e4SLinus Torvalds * mapping->private_lock does *not* protect mapping->private_list! In fact, 4791da177e4SLinus Torvalds * mapping->private_list will always be protected by the backing blockdev's 4801da177e4SLinus Torvalds * ->private_lock. 4811da177e4SLinus Torvalds * 4821da177e4SLinus Torvalds * Which introduces a requirement: all buffers on an address_space's 4831da177e4SLinus Torvalds * ->private_list must be from the same address_space: the blockdev's. 4841da177e4SLinus Torvalds * 4851da177e4SLinus Torvalds * address_spaces which do not place buffers at ->private_list via these 4861da177e4SLinus Torvalds * utility functions are free to use private_lock and private_list for 4871da177e4SLinus Torvalds * whatever they want. The only requirement is that list_empty(private_list) 4881da177e4SLinus Torvalds * be true at clear_inode() time. 4891da177e4SLinus Torvalds * 4901da177e4SLinus Torvalds * FIXME: clear_inode should not call invalidate_inode_buffers(). The 4911da177e4SLinus Torvalds * filesystems should do that. invalidate_inode_buffers() should just go 4921da177e4SLinus Torvalds * BUG_ON(!list_empty). 4931da177e4SLinus Torvalds * 4941da177e4SLinus Torvalds * FIXME: mark_buffer_dirty_inode() is a data-plane operation. It should 4951da177e4SLinus Torvalds * take an address_space, not an inode. And it should be called 4961da177e4SLinus Torvalds * mark_buffer_dirty_fsync() to clearly define why those buffers are being 4971da177e4SLinus Torvalds * queued up. 4981da177e4SLinus Torvalds * 4991da177e4SLinus Torvalds * FIXME: mark_buffer_dirty_inode() doesn't need to add the buffer to the 5001da177e4SLinus Torvalds * list if it is already on a list. Because if the buffer is on a list, 5011da177e4SLinus Torvalds * it *must* already be on the right one. If not, the filesystem is being 5021da177e4SLinus Torvalds * silly. This will save a ton of locking. But first we have to ensure 5031da177e4SLinus Torvalds * that buffers are taken *off* the old inode's list when they are freed 5041da177e4SLinus Torvalds * (presumably in truncate). That requires careful auditing of all 5051da177e4SLinus Torvalds * filesystems (do it inside bforget()). It could also be done by bringing 5061da177e4SLinus Torvalds * b_inode back. 5071da177e4SLinus Torvalds */ 5081da177e4SLinus Torvalds 5091da177e4SLinus Torvalds /* 5101da177e4SLinus Torvalds * The buffer's backing address_space's private_lock must be held 5111da177e4SLinus Torvalds */ 512dbacefc9SThomas Petazzoni static void __remove_assoc_queue(struct buffer_head *bh) 5131da177e4SLinus Torvalds { 5141da177e4SLinus Torvalds list_del_init(&bh->b_assoc_buffers); 51558ff407bSJan Kara WARN_ON(!bh->b_assoc_map); 51658ff407bSJan Kara if (buffer_write_io_error(bh)) 51758ff407bSJan Kara set_bit(AS_EIO, &bh->b_assoc_map->flags); 51858ff407bSJan Kara bh->b_assoc_map = NULL; 5191da177e4SLinus Torvalds } 5201da177e4SLinus Torvalds 5211da177e4SLinus Torvalds int inode_has_buffers(struct inode *inode) 5221da177e4SLinus Torvalds { 5231da177e4SLinus Torvalds return !list_empty(&inode->i_data.private_list); 5241da177e4SLinus Torvalds } 5251da177e4SLinus Torvalds 5261da177e4SLinus Torvalds /* 5271da177e4SLinus Torvalds * osync is designed to support O_SYNC io. It waits synchronously for 5281da177e4SLinus Torvalds * all already-submitted IO to complete, but does not queue any new 5291da177e4SLinus Torvalds * writes to the disk. 5301da177e4SLinus Torvalds * 5311da177e4SLinus Torvalds * To do O_SYNC writes, just queue the buffer writes with ll_rw_block as 5321da177e4SLinus Torvalds * you dirty the buffers, and then use osync_inode_buffers to wait for 5331da177e4SLinus Torvalds * completion. Any other dirty buffers which are not yet queued for 5341da177e4SLinus Torvalds * write will not be flushed to disk by the osync. 5351da177e4SLinus Torvalds */ 5361da177e4SLinus Torvalds static int osync_buffers_list(spinlock_t *lock, struct list_head *list) 5371da177e4SLinus Torvalds { 5381da177e4SLinus Torvalds struct buffer_head *bh; 5391da177e4SLinus Torvalds struct list_head *p; 5401da177e4SLinus Torvalds int err = 0; 5411da177e4SLinus Torvalds 5421da177e4SLinus Torvalds spin_lock(lock); 5431da177e4SLinus Torvalds repeat: 5441da177e4SLinus Torvalds list_for_each_prev(p, list) { 5451da177e4SLinus Torvalds bh = BH_ENTRY(p); 5461da177e4SLinus Torvalds if (buffer_locked(bh)) { 5471da177e4SLinus Torvalds get_bh(bh); 5481da177e4SLinus Torvalds spin_unlock(lock); 5491da177e4SLinus Torvalds wait_on_buffer(bh); 5501da177e4SLinus Torvalds if (!buffer_uptodate(bh)) 5511da177e4SLinus Torvalds err = -EIO; 5521da177e4SLinus Torvalds brelse(bh); 5531da177e4SLinus Torvalds spin_lock(lock); 5541da177e4SLinus Torvalds goto repeat; 5551da177e4SLinus Torvalds } 5561da177e4SLinus Torvalds } 5571da177e4SLinus Torvalds spin_unlock(lock); 5581da177e4SLinus Torvalds return err; 5591da177e4SLinus Torvalds } 5601da177e4SLinus Torvalds 56101a05b33SAl Viro static void do_thaw_one(struct super_block *sb, void *unused) 562c2d75438SEric Sandeen { 563c2d75438SEric Sandeen char b[BDEVNAME_SIZE]; 564c2d75438SEric Sandeen while (sb->s_bdev && !thaw_bdev(sb->s_bdev, sb)) 565c2d75438SEric Sandeen printk(KERN_WARNING "Emergency Thaw on %s\n", 566c2d75438SEric Sandeen bdevname(sb->s_bdev, b)); 567c2d75438SEric Sandeen } 56801a05b33SAl Viro 56901a05b33SAl Viro static void do_thaw_all(struct work_struct *work) 57001a05b33SAl Viro { 57101a05b33SAl Viro iterate_supers(do_thaw_one, NULL); 572053c525fSJens Axboe kfree(work); 573c2d75438SEric Sandeen printk(KERN_WARNING "Emergency Thaw complete\n"); 574c2d75438SEric Sandeen } 575c2d75438SEric Sandeen 576c2d75438SEric Sandeen /** 577c2d75438SEric Sandeen * emergency_thaw_all -- forcibly thaw every frozen filesystem 578c2d75438SEric Sandeen * 579c2d75438SEric Sandeen * Used for emergency unfreeze of all filesystems via SysRq 580c2d75438SEric Sandeen */ 581c2d75438SEric Sandeen void emergency_thaw_all(void) 582c2d75438SEric Sandeen { 583053c525fSJens Axboe struct work_struct *work; 584053c525fSJens Axboe 585053c525fSJens Axboe work = kmalloc(sizeof(*work), GFP_ATOMIC); 586053c525fSJens Axboe if (work) { 587053c525fSJens Axboe INIT_WORK(work, do_thaw_all); 588053c525fSJens Axboe schedule_work(work); 589053c525fSJens Axboe } 590c2d75438SEric Sandeen } 591c2d75438SEric Sandeen 5921da177e4SLinus Torvalds /** 59378a4a50aSRandy Dunlap * sync_mapping_buffers - write out & wait upon a mapping's "associated" buffers 59467be2dd1SMartin Waitz * @mapping: the mapping which wants those buffers written 5951da177e4SLinus Torvalds * 5961da177e4SLinus Torvalds * Starts I/O against the buffers at mapping->private_list, and waits upon 5971da177e4SLinus Torvalds * that I/O. 5981da177e4SLinus Torvalds * 59967be2dd1SMartin Waitz * Basically, this is a convenience function for fsync(). 60067be2dd1SMartin Waitz * @mapping is a file or directory which needs those buffers to be written for 60167be2dd1SMartin Waitz * a successful fsync(). 6021da177e4SLinus Torvalds */ 6031da177e4SLinus Torvalds int sync_mapping_buffers(struct address_space *mapping) 6041da177e4SLinus Torvalds { 6051da177e4SLinus Torvalds struct address_space *buffer_mapping = mapping->assoc_mapping; 6061da177e4SLinus Torvalds 6071da177e4SLinus Torvalds if (buffer_mapping == NULL || list_empty(&mapping->private_list)) 6081da177e4SLinus Torvalds return 0; 6091da177e4SLinus Torvalds 6101da177e4SLinus Torvalds return fsync_buffers_list(&buffer_mapping->private_lock, 6111da177e4SLinus Torvalds &mapping->private_list); 6121da177e4SLinus Torvalds } 6131da177e4SLinus Torvalds EXPORT_SYMBOL(sync_mapping_buffers); 6141da177e4SLinus Torvalds 6151da177e4SLinus Torvalds /* 6161da177e4SLinus Torvalds * Called when we've recently written block `bblock', and it is known that 6171da177e4SLinus Torvalds * `bblock' was for a buffer_boundary() buffer. This means that the block at 6181da177e4SLinus Torvalds * `bblock + 1' is probably a dirty indirect block. Hunt it down and, if it's 6191da177e4SLinus Torvalds * dirty, schedule it for IO. So that indirects merge nicely with their data. 6201da177e4SLinus Torvalds */ 6211da177e4SLinus Torvalds void write_boundary_block(struct block_device *bdev, 6221da177e4SLinus Torvalds sector_t bblock, unsigned blocksize) 6231da177e4SLinus Torvalds { 6241da177e4SLinus Torvalds struct buffer_head *bh = __find_get_block(bdev, bblock + 1, blocksize); 6251da177e4SLinus Torvalds if (bh) { 6261da177e4SLinus Torvalds if (buffer_dirty(bh)) 6271da177e4SLinus Torvalds ll_rw_block(WRITE, 1, &bh); 6281da177e4SLinus Torvalds put_bh(bh); 6291da177e4SLinus Torvalds } 6301da177e4SLinus Torvalds } 6311da177e4SLinus Torvalds 6321da177e4SLinus Torvalds void mark_buffer_dirty_inode(struct buffer_head *bh, struct inode *inode) 6331da177e4SLinus Torvalds { 6341da177e4SLinus Torvalds struct address_space *mapping = inode->i_mapping; 6351da177e4SLinus Torvalds struct address_space *buffer_mapping = bh->b_page->mapping; 6361da177e4SLinus Torvalds 6371da177e4SLinus Torvalds mark_buffer_dirty(bh); 6381da177e4SLinus Torvalds if (!mapping->assoc_mapping) { 6391da177e4SLinus Torvalds mapping->assoc_mapping = buffer_mapping; 6401da177e4SLinus Torvalds } else { 641e827f923SEric Sesterhenn BUG_ON(mapping->assoc_mapping != buffer_mapping); 6421da177e4SLinus Torvalds } 643535ee2fbSJan Kara if (!bh->b_assoc_map) { 6441da177e4SLinus Torvalds spin_lock(&buffer_mapping->private_lock); 6451da177e4SLinus Torvalds list_move_tail(&bh->b_assoc_buffers, 6461da177e4SLinus Torvalds &mapping->private_list); 64758ff407bSJan Kara bh->b_assoc_map = mapping; 6481da177e4SLinus Torvalds spin_unlock(&buffer_mapping->private_lock); 6491da177e4SLinus Torvalds } 6501da177e4SLinus Torvalds } 6511da177e4SLinus Torvalds EXPORT_SYMBOL(mark_buffer_dirty_inode); 6521da177e4SLinus Torvalds 6531da177e4SLinus Torvalds /* 654787d2214SNick Piggin * Mark the page dirty, and set it dirty in the radix tree, and mark the inode 655787d2214SNick Piggin * dirty. 656787d2214SNick Piggin * 657787d2214SNick Piggin * If warn is true, then emit a warning if the page is not uptodate and has 658787d2214SNick Piggin * not been truncated. 659787d2214SNick Piggin */ 660a8e7d49aSLinus Torvalds static void __set_page_dirty(struct page *page, 661787d2214SNick Piggin struct address_space *mapping, int warn) 662787d2214SNick Piggin { 66319fd6231SNick Piggin spin_lock_irq(&mapping->tree_lock); 664787d2214SNick Piggin if (page->mapping) { /* Race with truncate? */ 665787d2214SNick Piggin WARN_ON_ONCE(warn && !PageUptodate(page)); 666e3a7cca1SEdward Shishkin account_page_dirtied(page, mapping); 667787d2214SNick Piggin radix_tree_tag_set(&mapping->page_tree, 668787d2214SNick Piggin page_index(page), PAGECACHE_TAG_DIRTY); 669787d2214SNick Piggin } 67019fd6231SNick Piggin spin_unlock_irq(&mapping->tree_lock); 671787d2214SNick Piggin __mark_inode_dirty(mapping->host, I_DIRTY_PAGES); 672787d2214SNick Piggin } 673787d2214SNick Piggin 674787d2214SNick Piggin /* 6751da177e4SLinus Torvalds * Add a page to the dirty page list. 6761da177e4SLinus Torvalds * 6771da177e4SLinus Torvalds * It is a sad fact of life that this function is called from several places 6781da177e4SLinus Torvalds * deeply under spinlocking. It may not sleep. 6791da177e4SLinus Torvalds * 6801da177e4SLinus Torvalds * If the page has buffers, the uptodate buffers are set dirty, to preserve 6811da177e4SLinus Torvalds * dirty-state coherency between the page and the buffers. It the page does 6821da177e4SLinus Torvalds * not have buffers then when they are later attached they will all be set 6831da177e4SLinus Torvalds * dirty. 6841da177e4SLinus Torvalds * 6851da177e4SLinus Torvalds * The buffers are dirtied before the page is dirtied. There's a small race 6861da177e4SLinus Torvalds * window in which a writepage caller may see the page cleanness but not the 6871da177e4SLinus Torvalds * buffer dirtiness. That's fine. If this code were to set the page dirty 6881da177e4SLinus Torvalds * before the buffers, a concurrent writepage caller could clear the page dirty 6891da177e4SLinus Torvalds * bit, see a bunch of clean buffers and we'd end up with dirty buffers/clean 6901da177e4SLinus Torvalds * page on the dirty page list. 6911da177e4SLinus Torvalds * 6921da177e4SLinus Torvalds * We use private_lock to lock against try_to_free_buffers while using the 6931da177e4SLinus Torvalds * page's buffer list. Also use this to protect against clean buffers being 6941da177e4SLinus Torvalds * added to the page after it was set dirty. 6951da177e4SLinus Torvalds * 6961da177e4SLinus Torvalds * FIXME: may need to call ->reservepage here as well. That's rather up to the 6971da177e4SLinus Torvalds * address_space though. 6981da177e4SLinus Torvalds */ 6991da177e4SLinus Torvalds int __set_page_dirty_buffers(struct page *page) 7001da177e4SLinus Torvalds { 701a8e7d49aSLinus Torvalds int newly_dirty; 702787d2214SNick Piggin struct address_space *mapping = page_mapping(page); 703ebf7a227SNick Piggin 704ebf7a227SNick Piggin if (unlikely(!mapping)) 705ebf7a227SNick Piggin return !TestSetPageDirty(page); 7061da177e4SLinus Torvalds 7071da177e4SLinus Torvalds spin_lock(&mapping->private_lock); 7081da177e4SLinus Torvalds if (page_has_buffers(page)) { 7091da177e4SLinus Torvalds struct buffer_head *head = page_buffers(page); 7101da177e4SLinus Torvalds struct buffer_head *bh = head; 7111da177e4SLinus Torvalds 7121da177e4SLinus Torvalds do { 7131da177e4SLinus Torvalds set_buffer_dirty(bh); 7141da177e4SLinus Torvalds bh = bh->b_this_page; 7151da177e4SLinus Torvalds } while (bh != head); 7161da177e4SLinus Torvalds } 717a8e7d49aSLinus Torvalds newly_dirty = !TestSetPageDirty(page); 7181da177e4SLinus Torvalds spin_unlock(&mapping->private_lock); 7191da177e4SLinus Torvalds 720a8e7d49aSLinus Torvalds if (newly_dirty) 721a8e7d49aSLinus Torvalds __set_page_dirty(page, mapping, 1); 722a8e7d49aSLinus Torvalds return newly_dirty; 7231da177e4SLinus Torvalds } 7241da177e4SLinus Torvalds EXPORT_SYMBOL(__set_page_dirty_buffers); 7251da177e4SLinus Torvalds 7261da177e4SLinus Torvalds /* 7271da177e4SLinus Torvalds * Write out and wait upon a list of buffers. 7281da177e4SLinus Torvalds * 7291da177e4SLinus Torvalds * We have conflicting pressures: we want to make sure that all 7301da177e4SLinus Torvalds * initially dirty buffers get waited on, but that any subsequently 7311da177e4SLinus Torvalds * dirtied buffers don't. After all, we don't want fsync to last 7321da177e4SLinus Torvalds * forever if somebody is actively writing to the file. 7331da177e4SLinus Torvalds * 7341da177e4SLinus Torvalds * Do this in two main stages: first we copy dirty buffers to a 7351da177e4SLinus Torvalds * temporary inode list, queueing the writes as we go. Then we clean 7361da177e4SLinus Torvalds * up, waiting for those writes to complete. 7371da177e4SLinus Torvalds * 7381da177e4SLinus Torvalds * During this second stage, any subsequent updates to the file may end 7391da177e4SLinus Torvalds * up refiling the buffer on the original inode's dirty list again, so 7401da177e4SLinus Torvalds * there is a chance we will end up with a buffer queued for write but 7411da177e4SLinus Torvalds * not yet completed on that list. So, as a final cleanup we go through 7421da177e4SLinus Torvalds * the osync code to catch these locked, dirty buffers without requeuing 7431da177e4SLinus Torvalds * any newly dirty buffers for write. 7441da177e4SLinus Torvalds */ 7451da177e4SLinus Torvalds static int fsync_buffers_list(spinlock_t *lock, struct list_head *list) 7461da177e4SLinus Torvalds { 7471da177e4SLinus Torvalds struct buffer_head *bh; 7481da177e4SLinus Torvalds struct list_head tmp; 7497eaceaccSJens Axboe struct address_space *mapping; 7501da177e4SLinus Torvalds int err = 0, err2; 7514ee2491eSJens Axboe struct blk_plug plug; 7521da177e4SLinus Torvalds 7531da177e4SLinus Torvalds INIT_LIST_HEAD(&tmp); 7544ee2491eSJens Axboe blk_start_plug(&plug); 7551da177e4SLinus Torvalds 7561da177e4SLinus Torvalds spin_lock(lock); 7571da177e4SLinus Torvalds while (!list_empty(list)) { 7581da177e4SLinus Torvalds bh = BH_ENTRY(list->next); 759535ee2fbSJan Kara mapping = bh->b_assoc_map; 76058ff407bSJan Kara __remove_assoc_queue(bh); 761535ee2fbSJan Kara /* Avoid race with mark_buffer_dirty_inode() which does 762535ee2fbSJan Kara * a lockless check and we rely on seeing the dirty bit */ 763535ee2fbSJan Kara smp_mb(); 7641da177e4SLinus Torvalds if (buffer_dirty(bh) || buffer_locked(bh)) { 7651da177e4SLinus Torvalds list_add(&bh->b_assoc_buffers, &tmp); 766535ee2fbSJan Kara bh->b_assoc_map = mapping; 7671da177e4SLinus Torvalds if (buffer_dirty(bh)) { 7681da177e4SLinus Torvalds get_bh(bh); 7691da177e4SLinus Torvalds spin_unlock(lock); 7701da177e4SLinus Torvalds /* 7711da177e4SLinus Torvalds * Ensure any pending I/O completes so that 7729cb569d6SChristoph Hellwig * write_dirty_buffer() actually writes the 7739cb569d6SChristoph Hellwig * current contents - it is a noop if I/O is 7749cb569d6SChristoph Hellwig * still in flight on potentially older 7759cb569d6SChristoph Hellwig * contents. 7761da177e4SLinus Torvalds */ 777721a9602SJens Axboe write_dirty_buffer(bh, WRITE_SYNC); 7789cf6b720SJens Axboe 7799cf6b720SJens Axboe /* 7809cf6b720SJens Axboe * Kick off IO for the previous mapping. Note 7819cf6b720SJens Axboe * that we will not run the very last mapping, 7829cf6b720SJens Axboe * wait_on_buffer() will do that for us 7839cf6b720SJens Axboe * through sync_buffer(). 7849cf6b720SJens Axboe */ 7851da177e4SLinus Torvalds brelse(bh); 7861da177e4SLinus Torvalds spin_lock(lock); 7871da177e4SLinus Torvalds } 7881da177e4SLinus Torvalds } 7891da177e4SLinus Torvalds } 7901da177e4SLinus Torvalds 7914ee2491eSJens Axboe spin_unlock(lock); 7924ee2491eSJens Axboe blk_finish_plug(&plug); 7934ee2491eSJens Axboe spin_lock(lock); 7944ee2491eSJens Axboe 7951da177e4SLinus Torvalds while (!list_empty(&tmp)) { 7961da177e4SLinus Torvalds bh = BH_ENTRY(tmp.prev); 7971da177e4SLinus Torvalds get_bh(bh); 798535ee2fbSJan Kara mapping = bh->b_assoc_map; 799535ee2fbSJan Kara __remove_assoc_queue(bh); 800535ee2fbSJan Kara /* Avoid race with mark_buffer_dirty_inode() which does 801535ee2fbSJan Kara * a lockless check and we rely on seeing the dirty bit */ 802535ee2fbSJan Kara smp_mb(); 803535ee2fbSJan Kara if (buffer_dirty(bh)) { 804535ee2fbSJan Kara list_add(&bh->b_assoc_buffers, 805e3892296SJan Kara &mapping->private_list); 806535ee2fbSJan Kara bh->b_assoc_map = mapping; 807535ee2fbSJan Kara } 8081da177e4SLinus Torvalds spin_unlock(lock); 8091da177e4SLinus Torvalds wait_on_buffer(bh); 8101da177e4SLinus Torvalds if (!buffer_uptodate(bh)) 8111da177e4SLinus Torvalds err = -EIO; 8121da177e4SLinus Torvalds brelse(bh); 8131da177e4SLinus Torvalds spin_lock(lock); 8141da177e4SLinus Torvalds } 8151da177e4SLinus Torvalds 8161da177e4SLinus Torvalds spin_unlock(lock); 8171da177e4SLinus Torvalds err2 = osync_buffers_list(lock, list); 8181da177e4SLinus Torvalds if (err) 8191da177e4SLinus Torvalds return err; 8201da177e4SLinus Torvalds else 8211da177e4SLinus Torvalds return err2; 8221da177e4SLinus Torvalds } 8231da177e4SLinus Torvalds 8241da177e4SLinus Torvalds /* 8251da177e4SLinus Torvalds * Invalidate any and all dirty buffers on a given inode. We are 8261da177e4SLinus Torvalds * probably unmounting the fs, but that doesn't mean we have already 8271da177e4SLinus Torvalds * done a sync(). Just drop the buffers from the inode list. 8281da177e4SLinus Torvalds * 8291da177e4SLinus Torvalds * NOTE: we take the inode's blockdev's mapping's private_lock. Which 8301da177e4SLinus Torvalds * assumes that all the buffers are against the blockdev. Not true 8311da177e4SLinus Torvalds * for reiserfs. 8321da177e4SLinus Torvalds */ 8331da177e4SLinus Torvalds void invalidate_inode_buffers(struct inode *inode) 8341da177e4SLinus Torvalds { 8351da177e4SLinus Torvalds if (inode_has_buffers(inode)) { 8361da177e4SLinus Torvalds struct address_space *mapping = &inode->i_data; 8371da177e4SLinus Torvalds struct list_head *list = &mapping->private_list; 8381da177e4SLinus Torvalds struct address_space *buffer_mapping = mapping->assoc_mapping; 8391da177e4SLinus Torvalds 8401da177e4SLinus Torvalds spin_lock(&buffer_mapping->private_lock); 8411da177e4SLinus Torvalds while (!list_empty(list)) 8421da177e4SLinus Torvalds __remove_assoc_queue(BH_ENTRY(list->next)); 8431da177e4SLinus Torvalds spin_unlock(&buffer_mapping->private_lock); 8441da177e4SLinus Torvalds } 8451da177e4SLinus Torvalds } 84652b19ac9SJan Kara EXPORT_SYMBOL(invalidate_inode_buffers); 8471da177e4SLinus Torvalds 8481da177e4SLinus Torvalds /* 8491da177e4SLinus Torvalds * Remove any clean buffers from the inode's buffer list. This is called 8501da177e4SLinus Torvalds * when we're trying to free the inode itself. Those buffers can pin it. 8511da177e4SLinus Torvalds * 8521da177e4SLinus Torvalds * Returns true if all buffers were removed. 8531da177e4SLinus Torvalds */ 8541da177e4SLinus Torvalds int remove_inode_buffers(struct inode *inode) 8551da177e4SLinus Torvalds { 8561da177e4SLinus Torvalds int ret = 1; 8571da177e4SLinus Torvalds 8581da177e4SLinus Torvalds if (inode_has_buffers(inode)) { 8591da177e4SLinus Torvalds struct address_space *mapping = &inode->i_data; 8601da177e4SLinus Torvalds struct list_head *list = &mapping->private_list; 8611da177e4SLinus Torvalds struct address_space *buffer_mapping = mapping->assoc_mapping; 8621da177e4SLinus Torvalds 8631da177e4SLinus Torvalds spin_lock(&buffer_mapping->private_lock); 8641da177e4SLinus Torvalds while (!list_empty(list)) { 8651da177e4SLinus Torvalds struct buffer_head *bh = BH_ENTRY(list->next); 8661da177e4SLinus Torvalds if (buffer_dirty(bh)) { 8671da177e4SLinus Torvalds ret = 0; 8681da177e4SLinus Torvalds break; 8691da177e4SLinus Torvalds } 8701da177e4SLinus Torvalds __remove_assoc_queue(bh); 8711da177e4SLinus Torvalds } 8721da177e4SLinus Torvalds spin_unlock(&buffer_mapping->private_lock); 8731da177e4SLinus Torvalds } 8741da177e4SLinus Torvalds return ret; 8751da177e4SLinus Torvalds } 8761da177e4SLinus Torvalds 8771da177e4SLinus Torvalds /* 8781da177e4SLinus Torvalds * Create the appropriate buffers when given a page for data area and 8791da177e4SLinus Torvalds * the size of each buffer.. Use the bh->b_this_page linked list to 8801da177e4SLinus Torvalds * follow the buffers created. Return NULL if unable to create more 8811da177e4SLinus Torvalds * buffers. 8821da177e4SLinus Torvalds * 8831da177e4SLinus Torvalds * The retry flag is used to differentiate async IO (paging, swapping) 8841da177e4SLinus Torvalds * which may not fail from ordinary buffer allocations. 8851da177e4SLinus Torvalds */ 8861da177e4SLinus Torvalds struct buffer_head *alloc_page_buffers(struct page *page, unsigned long size, 8871da177e4SLinus Torvalds int retry) 8881da177e4SLinus Torvalds { 8891da177e4SLinus Torvalds struct buffer_head *bh, *head; 8901da177e4SLinus Torvalds long offset; 8911da177e4SLinus Torvalds 8921da177e4SLinus Torvalds try_again: 8931da177e4SLinus Torvalds head = NULL; 8941da177e4SLinus Torvalds offset = PAGE_SIZE; 8951da177e4SLinus Torvalds while ((offset -= size) >= 0) { 8961da177e4SLinus Torvalds bh = alloc_buffer_head(GFP_NOFS); 8971da177e4SLinus Torvalds if (!bh) 8981da177e4SLinus Torvalds goto no_grow; 8991da177e4SLinus Torvalds 9001da177e4SLinus Torvalds bh->b_bdev = NULL; 9011da177e4SLinus Torvalds bh->b_this_page = head; 9021da177e4SLinus Torvalds bh->b_blocknr = -1; 9031da177e4SLinus Torvalds head = bh; 9041da177e4SLinus Torvalds 9051da177e4SLinus Torvalds bh->b_state = 0; 9061da177e4SLinus Torvalds atomic_set(&bh->b_count, 0); 9071da177e4SLinus Torvalds bh->b_size = size; 9081da177e4SLinus Torvalds 9091da177e4SLinus Torvalds /* Link the buffer to its page */ 9101da177e4SLinus Torvalds set_bh_page(bh, page, offset); 9111da177e4SLinus Torvalds 91201ffe339SNathan Scott init_buffer(bh, NULL, NULL); 9131da177e4SLinus Torvalds } 9141da177e4SLinus Torvalds return head; 9151da177e4SLinus Torvalds /* 9161da177e4SLinus Torvalds * In case anything failed, we just free everything we got. 9171da177e4SLinus Torvalds */ 9181da177e4SLinus Torvalds no_grow: 9191da177e4SLinus Torvalds if (head) { 9201da177e4SLinus Torvalds do { 9211da177e4SLinus Torvalds bh = head; 9221da177e4SLinus Torvalds head = head->b_this_page; 9231da177e4SLinus Torvalds free_buffer_head(bh); 9241da177e4SLinus Torvalds } while (head); 9251da177e4SLinus Torvalds } 9261da177e4SLinus Torvalds 9271da177e4SLinus Torvalds /* 9281da177e4SLinus Torvalds * Return failure for non-async IO requests. Async IO requests 9291da177e4SLinus Torvalds * are not allowed to fail, so we have to wait until buffer heads 9301da177e4SLinus Torvalds * become available. But we don't want tasks sleeping with 9311da177e4SLinus Torvalds * partially complete buffers, so all were released above. 9321da177e4SLinus Torvalds */ 9331da177e4SLinus Torvalds if (!retry) 9341da177e4SLinus Torvalds return NULL; 9351da177e4SLinus Torvalds 9361da177e4SLinus Torvalds /* We're _really_ low on memory. Now we just 9371da177e4SLinus Torvalds * wait for old buffer heads to become free due to 9381da177e4SLinus Torvalds * finishing IO. Since this is an async request and 9391da177e4SLinus Torvalds * the reserve list is empty, we're sure there are 9401da177e4SLinus Torvalds * async buffer heads in use. 9411da177e4SLinus Torvalds */ 9421da177e4SLinus Torvalds free_more_memory(); 9431da177e4SLinus Torvalds goto try_again; 9441da177e4SLinus Torvalds } 9451da177e4SLinus Torvalds EXPORT_SYMBOL_GPL(alloc_page_buffers); 9461da177e4SLinus Torvalds 9471da177e4SLinus Torvalds static inline void 9481da177e4SLinus Torvalds link_dev_buffers(struct page *page, struct buffer_head *head) 9491da177e4SLinus Torvalds { 9501da177e4SLinus Torvalds struct buffer_head *bh, *tail; 9511da177e4SLinus Torvalds 9521da177e4SLinus Torvalds bh = head; 9531da177e4SLinus Torvalds do { 9541da177e4SLinus Torvalds tail = bh; 9551da177e4SLinus Torvalds bh = bh->b_this_page; 9561da177e4SLinus Torvalds } while (bh); 9571da177e4SLinus Torvalds tail->b_this_page = head; 9581da177e4SLinus Torvalds attach_page_buffers(page, head); 9591da177e4SLinus Torvalds } 9601da177e4SLinus Torvalds 9611da177e4SLinus Torvalds /* 9621da177e4SLinus Torvalds * Initialise the state of a blockdev page's buffers. 9631da177e4SLinus Torvalds */ 9641da177e4SLinus Torvalds static void 9651da177e4SLinus Torvalds init_page_buffers(struct page *page, struct block_device *bdev, 9661da177e4SLinus Torvalds sector_t block, int size) 9671da177e4SLinus Torvalds { 9681da177e4SLinus Torvalds struct buffer_head *head = page_buffers(page); 9691da177e4SLinus Torvalds struct buffer_head *bh = head; 9701da177e4SLinus Torvalds int uptodate = PageUptodate(page); 9711da177e4SLinus Torvalds 9721da177e4SLinus Torvalds do { 9731da177e4SLinus Torvalds if (!buffer_mapped(bh)) { 9741da177e4SLinus Torvalds init_buffer(bh, NULL, NULL); 9751da177e4SLinus Torvalds bh->b_bdev = bdev; 9761da177e4SLinus Torvalds bh->b_blocknr = block; 9771da177e4SLinus Torvalds if (uptodate) 9781da177e4SLinus Torvalds set_buffer_uptodate(bh); 9791da177e4SLinus Torvalds set_buffer_mapped(bh); 9801da177e4SLinus Torvalds } 9811da177e4SLinus Torvalds block++; 9821da177e4SLinus Torvalds bh = bh->b_this_page; 9831da177e4SLinus Torvalds } while (bh != head); 9841da177e4SLinus Torvalds } 9851da177e4SLinus Torvalds 9861da177e4SLinus Torvalds /* 9871da177e4SLinus Torvalds * Create the page-cache page that contains the requested block. 9881da177e4SLinus Torvalds * 9891da177e4SLinus Torvalds * This is user purely for blockdev mappings. 9901da177e4SLinus Torvalds */ 9911da177e4SLinus Torvalds static struct page * 9921da177e4SLinus Torvalds grow_dev_page(struct block_device *bdev, sector_t block, 9931da177e4SLinus Torvalds pgoff_t index, int size) 9941da177e4SLinus Torvalds { 9951da177e4SLinus Torvalds struct inode *inode = bdev->bd_inode; 9961da177e4SLinus Torvalds struct page *page; 9971da177e4SLinus Torvalds struct buffer_head *bh; 9981da177e4SLinus Torvalds 999ea125892SChristoph Lameter page = find_or_create_page(inode->i_mapping, index, 1000769848c0SMel Gorman (mapping_gfp_mask(inode->i_mapping) & ~__GFP_FS)|__GFP_MOVABLE); 10011da177e4SLinus Torvalds if (!page) 10021da177e4SLinus Torvalds return NULL; 10031da177e4SLinus Torvalds 1004e827f923SEric Sesterhenn BUG_ON(!PageLocked(page)); 10051da177e4SLinus Torvalds 10061da177e4SLinus Torvalds if (page_has_buffers(page)) { 10071da177e4SLinus Torvalds bh = page_buffers(page); 10081da177e4SLinus Torvalds if (bh->b_size == size) { 10091da177e4SLinus Torvalds init_page_buffers(page, bdev, block, size); 10101da177e4SLinus Torvalds return page; 10111da177e4SLinus Torvalds } 10121da177e4SLinus Torvalds if (!try_to_free_buffers(page)) 10131da177e4SLinus Torvalds goto failed; 10141da177e4SLinus Torvalds } 10151da177e4SLinus Torvalds 10161da177e4SLinus Torvalds /* 10171da177e4SLinus Torvalds * Allocate some buffers for this page 10181da177e4SLinus Torvalds */ 10191da177e4SLinus Torvalds bh = alloc_page_buffers(page, size, 0); 10201da177e4SLinus Torvalds if (!bh) 10211da177e4SLinus Torvalds goto failed; 10221da177e4SLinus Torvalds 10231da177e4SLinus Torvalds /* 10241da177e4SLinus Torvalds * Link the page to the buffers and initialise them. Take the 10251da177e4SLinus Torvalds * lock to be atomic wrt __find_get_block(), which does not 10261da177e4SLinus Torvalds * run under the page lock. 10271da177e4SLinus Torvalds */ 10281da177e4SLinus Torvalds spin_lock(&inode->i_mapping->private_lock); 10291da177e4SLinus Torvalds link_dev_buffers(page, bh); 10301da177e4SLinus Torvalds init_page_buffers(page, bdev, block, size); 10311da177e4SLinus Torvalds spin_unlock(&inode->i_mapping->private_lock); 10321da177e4SLinus Torvalds return page; 10331da177e4SLinus Torvalds 10341da177e4SLinus Torvalds failed: 10351da177e4SLinus Torvalds BUG(); 10361da177e4SLinus Torvalds unlock_page(page); 10371da177e4SLinus Torvalds page_cache_release(page); 10381da177e4SLinus Torvalds return NULL; 10391da177e4SLinus Torvalds } 10401da177e4SLinus Torvalds 10411da177e4SLinus Torvalds /* 10421da177e4SLinus Torvalds * Create buffers for the specified block device block's page. If 10431da177e4SLinus Torvalds * that page was dirty, the buffers are set dirty also. 10441da177e4SLinus Torvalds */ 1045858119e1SArjan van de Ven static int 10461da177e4SLinus Torvalds grow_buffers(struct block_device *bdev, sector_t block, int size) 10471da177e4SLinus Torvalds { 10481da177e4SLinus Torvalds struct page *page; 10491da177e4SLinus Torvalds pgoff_t index; 10501da177e4SLinus Torvalds int sizebits; 10511da177e4SLinus Torvalds 10521da177e4SLinus Torvalds sizebits = -1; 10531da177e4SLinus Torvalds do { 10541da177e4SLinus Torvalds sizebits++; 10551da177e4SLinus Torvalds } while ((size << sizebits) < PAGE_SIZE); 10561da177e4SLinus Torvalds 10571da177e4SLinus Torvalds index = block >> sizebits; 10581da177e4SLinus Torvalds 1059e5657933SAndrew Morton /* 1060e5657933SAndrew Morton * Check for a block which wants to lie outside our maximum possible 1061e5657933SAndrew Morton * pagecache index. (this comparison is done using sector_t types). 1062e5657933SAndrew Morton */ 1063e5657933SAndrew Morton if (unlikely(index != block >> sizebits)) { 1064e5657933SAndrew Morton char b[BDEVNAME_SIZE]; 1065e5657933SAndrew Morton 1066e5657933SAndrew Morton printk(KERN_ERR "%s: requested out-of-range block %llu for " 1067e5657933SAndrew Morton "device %s\n", 10688e24eea7SHarvey Harrison __func__, (unsigned long long)block, 1069e5657933SAndrew Morton bdevname(bdev, b)); 1070e5657933SAndrew Morton return -EIO; 1071e5657933SAndrew Morton } 1072e5657933SAndrew Morton block = index << sizebits; 10731da177e4SLinus Torvalds /* Create a page with the proper size buffers.. */ 10741da177e4SLinus Torvalds page = grow_dev_page(bdev, block, index, size); 10751da177e4SLinus Torvalds if (!page) 10761da177e4SLinus Torvalds return 0; 10771da177e4SLinus Torvalds unlock_page(page); 10781da177e4SLinus Torvalds page_cache_release(page); 10791da177e4SLinus Torvalds return 1; 10801da177e4SLinus Torvalds } 10811da177e4SLinus Torvalds 108275c96f85SAdrian Bunk static struct buffer_head * 10831da177e4SLinus Torvalds __getblk_slow(struct block_device *bdev, sector_t block, int size) 10841da177e4SLinus Torvalds { 10851da177e4SLinus Torvalds /* Size must be multiple of hard sectorsize */ 1086e1defc4fSMartin K. Petersen if (unlikely(size & (bdev_logical_block_size(bdev)-1) || 10871da177e4SLinus Torvalds (size < 512 || size > PAGE_SIZE))) { 10881da177e4SLinus Torvalds printk(KERN_ERR "getblk(): invalid block size %d requested\n", 10891da177e4SLinus Torvalds size); 1090e1defc4fSMartin K. Petersen printk(KERN_ERR "logical block size: %d\n", 1091e1defc4fSMartin K. Petersen bdev_logical_block_size(bdev)); 10921da177e4SLinus Torvalds 10931da177e4SLinus Torvalds dump_stack(); 10941da177e4SLinus Torvalds return NULL; 10951da177e4SLinus Torvalds } 10961da177e4SLinus Torvalds 10971da177e4SLinus Torvalds for (;;) { 10981da177e4SLinus Torvalds struct buffer_head * bh; 1099e5657933SAndrew Morton int ret; 11001da177e4SLinus Torvalds 11011da177e4SLinus Torvalds bh = __find_get_block(bdev, block, size); 11021da177e4SLinus Torvalds if (bh) 11031da177e4SLinus Torvalds return bh; 11041da177e4SLinus Torvalds 1105e5657933SAndrew Morton ret = grow_buffers(bdev, block, size); 1106e5657933SAndrew Morton if (ret < 0) 1107e5657933SAndrew Morton return NULL; 1108e5657933SAndrew Morton if (ret == 0) 11091da177e4SLinus Torvalds free_more_memory(); 11101da177e4SLinus Torvalds } 11111da177e4SLinus Torvalds } 11121da177e4SLinus Torvalds 11131da177e4SLinus Torvalds /* 11141da177e4SLinus Torvalds * The relationship between dirty buffers and dirty pages: 11151da177e4SLinus Torvalds * 11161da177e4SLinus Torvalds * Whenever a page has any dirty buffers, the page's dirty bit is set, and 11171da177e4SLinus Torvalds * the page is tagged dirty in its radix tree. 11181da177e4SLinus Torvalds * 11191da177e4SLinus Torvalds * At all times, the dirtiness of the buffers represents the dirtiness of 11201da177e4SLinus Torvalds * subsections of the page. If the page has buffers, the page dirty bit is 11211da177e4SLinus Torvalds * merely a hint about the true dirty state. 11221da177e4SLinus Torvalds * 11231da177e4SLinus Torvalds * When a page is set dirty in its entirety, all its buffers are marked dirty 11241da177e4SLinus Torvalds * (if the page has buffers). 11251da177e4SLinus Torvalds * 11261da177e4SLinus Torvalds * When a buffer is marked dirty, its page is dirtied, but the page's other 11271da177e4SLinus Torvalds * buffers are not. 11281da177e4SLinus Torvalds * 11291da177e4SLinus Torvalds * Also. When blockdev buffers are explicitly read with bread(), they 11301da177e4SLinus Torvalds * individually become uptodate. But their backing page remains not 11311da177e4SLinus Torvalds * uptodate - even if all of its buffers are uptodate. A subsequent 11321da177e4SLinus Torvalds * block_read_full_page() against that page will discover all the uptodate 11331da177e4SLinus Torvalds * buffers, will set the page uptodate and will perform no I/O. 11341da177e4SLinus Torvalds */ 11351da177e4SLinus Torvalds 11361da177e4SLinus Torvalds /** 11371da177e4SLinus Torvalds * mark_buffer_dirty - mark a buffer_head as needing writeout 113867be2dd1SMartin Waitz * @bh: the buffer_head to mark dirty 11391da177e4SLinus Torvalds * 11401da177e4SLinus Torvalds * mark_buffer_dirty() will set the dirty bit against the buffer, then set its 11411da177e4SLinus Torvalds * backing page dirty, then tag the page as dirty in its address_space's radix 11421da177e4SLinus Torvalds * tree and then attach the address_space's inode to its superblock's dirty 11431da177e4SLinus Torvalds * inode list. 11441da177e4SLinus Torvalds * 11451da177e4SLinus Torvalds * mark_buffer_dirty() is atomic. It takes bh->b_page->mapping->private_lock, 1146250df6edSDave Chinner * mapping->tree_lock and mapping->host->i_lock. 11471da177e4SLinus Torvalds */ 1148fc9b52cdSHarvey Harrison void mark_buffer_dirty(struct buffer_head *bh) 11491da177e4SLinus Torvalds { 1150787d2214SNick Piggin WARN_ON_ONCE(!buffer_uptodate(bh)); 11511be62dc1SLinus Torvalds 11521be62dc1SLinus Torvalds /* 11531be62dc1SLinus Torvalds * Very *carefully* optimize the it-is-already-dirty case. 11541be62dc1SLinus Torvalds * 11551be62dc1SLinus Torvalds * Don't let the final "is it dirty" escape to before we 11561be62dc1SLinus Torvalds * perhaps modified the buffer. 11571be62dc1SLinus Torvalds */ 11581be62dc1SLinus Torvalds if (buffer_dirty(bh)) { 11591be62dc1SLinus Torvalds smp_mb(); 11601be62dc1SLinus Torvalds if (buffer_dirty(bh)) 11611be62dc1SLinus Torvalds return; 11621be62dc1SLinus Torvalds } 11631be62dc1SLinus Torvalds 1164a8e7d49aSLinus Torvalds if (!test_set_buffer_dirty(bh)) { 1165a8e7d49aSLinus Torvalds struct page *page = bh->b_page; 11668e9d78edSLinus Torvalds if (!TestSetPageDirty(page)) { 11678e9d78edSLinus Torvalds struct address_space *mapping = page_mapping(page); 11688e9d78edSLinus Torvalds if (mapping) 11698e9d78edSLinus Torvalds __set_page_dirty(page, mapping, 0); 11708e9d78edSLinus Torvalds } 1171a8e7d49aSLinus Torvalds } 11721da177e4SLinus Torvalds } 11731fe72eaaSH Hartley Sweeten EXPORT_SYMBOL(mark_buffer_dirty); 11741da177e4SLinus Torvalds 11751da177e4SLinus Torvalds /* 11761da177e4SLinus Torvalds * Decrement a buffer_head's reference count. If all buffers against a page 11771da177e4SLinus Torvalds * have zero reference count, are clean and unlocked, and if the page is clean 11781da177e4SLinus Torvalds * and unlocked then try_to_free_buffers() may strip the buffers from the page 11791da177e4SLinus Torvalds * in preparation for freeing it (sometimes, rarely, buffers are removed from 11801da177e4SLinus Torvalds * a page but it ends up not being freed, and buffers may later be reattached). 11811da177e4SLinus Torvalds */ 11821da177e4SLinus Torvalds void __brelse(struct buffer_head * buf) 11831da177e4SLinus Torvalds { 11841da177e4SLinus Torvalds if (atomic_read(&buf->b_count)) { 11851da177e4SLinus Torvalds put_bh(buf); 11861da177e4SLinus Torvalds return; 11871da177e4SLinus Torvalds } 11885c752ad9SArjan van de Ven WARN(1, KERN_ERR "VFS: brelse: Trying to free free buffer\n"); 11891da177e4SLinus Torvalds } 11901fe72eaaSH Hartley Sweeten EXPORT_SYMBOL(__brelse); 11911da177e4SLinus Torvalds 11921da177e4SLinus Torvalds /* 11931da177e4SLinus Torvalds * bforget() is like brelse(), except it discards any 11941da177e4SLinus Torvalds * potentially dirty data. 11951da177e4SLinus Torvalds */ 11961da177e4SLinus Torvalds void __bforget(struct buffer_head *bh) 11971da177e4SLinus Torvalds { 11981da177e4SLinus Torvalds clear_buffer_dirty(bh); 1199535ee2fbSJan Kara if (bh->b_assoc_map) { 12001da177e4SLinus Torvalds struct address_space *buffer_mapping = bh->b_page->mapping; 12011da177e4SLinus Torvalds 12021da177e4SLinus Torvalds spin_lock(&buffer_mapping->private_lock); 12031da177e4SLinus Torvalds list_del_init(&bh->b_assoc_buffers); 120458ff407bSJan Kara bh->b_assoc_map = NULL; 12051da177e4SLinus Torvalds spin_unlock(&buffer_mapping->private_lock); 12061da177e4SLinus Torvalds } 12071da177e4SLinus Torvalds __brelse(bh); 12081da177e4SLinus Torvalds } 12091fe72eaaSH Hartley Sweeten EXPORT_SYMBOL(__bforget); 12101da177e4SLinus Torvalds 12111da177e4SLinus Torvalds static struct buffer_head *__bread_slow(struct buffer_head *bh) 12121da177e4SLinus Torvalds { 12131da177e4SLinus Torvalds lock_buffer(bh); 12141da177e4SLinus Torvalds if (buffer_uptodate(bh)) { 12151da177e4SLinus Torvalds unlock_buffer(bh); 12161da177e4SLinus Torvalds return bh; 12171da177e4SLinus Torvalds } else { 12181da177e4SLinus Torvalds get_bh(bh); 12191da177e4SLinus Torvalds bh->b_end_io = end_buffer_read_sync; 12201da177e4SLinus Torvalds submit_bh(READ, bh); 12211da177e4SLinus Torvalds wait_on_buffer(bh); 12221da177e4SLinus Torvalds if (buffer_uptodate(bh)) 12231da177e4SLinus Torvalds return bh; 12241da177e4SLinus Torvalds } 12251da177e4SLinus Torvalds brelse(bh); 12261da177e4SLinus Torvalds return NULL; 12271da177e4SLinus Torvalds } 12281da177e4SLinus Torvalds 12291da177e4SLinus Torvalds /* 12301da177e4SLinus Torvalds * Per-cpu buffer LRU implementation. To reduce the cost of __find_get_block(). 12311da177e4SLinus Torvalds * The bhs[] array is sorted - newest buffer is at bhs[0]. Buffers have their 12321da177e4SLinus Torvalds * refcount elevated by one when they're in an LRU. A buffer can only appear 12331da177e4SLinus Torvalds * once in a particular CPU's LRU. A single buffer can be present in multiple 12341da177e4SLinus Torvalds * CPU's LRUs at the same time. 12351da177e4SLinus Torvalds * 12361da177e4SLinus Torvalds * This is a transparent caching front-end to sb_bread(), sb_getblk() and 12371da177e4SLinus Torvalds * sb_find_get_block(). 12381da177e4SLinus Torvalds * 12391da177e4SLinus Torvalds * The LRUs themselves only need locking against invalidate_bh_lrus. We use 12401da177e4SLinus Torvalds * a local interrupt disable for that. 12411da177e4SLinus Torvalds */ 12421da177e4SLinus Torvalds 12431da177e4SLinus Torvalds #define BH_LRU_SIZE 8 12441da177e4SLinus Torvalds 12451da177e4SLinus Torvalds struct bh_lru { 12461da177e4SLinus Torvalds struct buffer_head *bhs[BH_LRU_SIZE]; 12471da177e4SLinus Torvalds }; 12481da177e4SLinus Torvalds 12491da177e4SLinus Torvalds static DEFINE_PER_CPU(struct bh_lru, bh_lrus) = {{ NULL }}; 12501da177e4SLinus Torvalds 12511da177e4SLinus Torvalds #ifdef CONFIG_SMP 12521da177e4SLinus Torvalds #define bh_lru_lock() local_irq_disable() 12531da177e4SLinus Torvalds #define bh_lru_unlock() local_irq_enable() 12541da177e4SLinus Torvalds #else 12551da177e4SLinus Torvalds #define bh_lru_lock() preempt_disable() 12561da177e4SLinus Torvalds #define bh_lru_unlock() preempt_enable() 12571da177e4SLinus Torvalds #endif 12581da177e4SLinus Torvalds 12591da177e4SLinus Torvalds static inline void check_irqs_on(void) 12601da177e4SLinus Torvalds { 12611da177e4SLinus Torvalds #ifdef irqs_disabled 12621da177e4SLinus Torvalds BUG_ON(irqs_disabled()); 12631da177e4SLinus Torvalds #endif 12641da177e4SLinus Torvalds } 12651da177e4SLinus Torvalds 12661da177e4SLinus Torvalds /* 12671da177e4SLinus Torvalds * The LRU management algorithm is dopey-but-simple. Sorry. 12681da177e4SLinus Torvalds */ 12691da177e4SLinus Torvalds static void bh_lru_install(struct buffer_head *bh) 12701da177e4SLinus Torvalds { 12711da177e4SLinus Torvalds struct buffer_head *evictee = NULL; 12721da177e4SLinus Torvalds 12731da177e4SLinus Torvalds check_irqs_on(); 12741da177e4SLinus Torvalds bh_lru_lock(); 1275c7b92516SChristoph Lameter if (__this_cpu_read(bh_lrus.bhs[0]) != bh) { 12761da177e4SLinus Torvalds struct buffer_head *bhs[BH_LRU_SIZE]; 12771da177e4SLinus Torvalds int in; 12781da177e4SLinus Torvalds int out = 0; 12791da177e4SLinus Torvalds 12801da177e4SLinus Torvalds get_bh(bh); 12811da177e4SLinus Torvalds bhs[out++] = bh; 12821da177e4SLinus Torvalds for (in = 0; in < BH_LRU_SIZE; in++) { 1283c7b92516SChristoph Lameter struct buffer_head *bh2 = 1284c7b92516SChristoph Lameter __this_cpu_read(bh_lrus.bhs[in]); 12851da177e4SLinus Torvalds 12861da177e4SLinus Torvalds if (bh2 == bh) { 12871da177e4SLinus Torvalds __brelse(bh2); 12881da177e4SLinus Torvalds } else { 12891da177e4SLinus Torvalds if (out >= BH_LRU_SIZE) { 12901da177e4SLinus Torvalds BUG_ON(evictee != NULL); 12911da177e4SLinus Torvalds evictee = bh2; 12921da177e4SLinus Torvalds } else { 12931da177e4SLinus Torvalds bhs[out++] = bh2; 12941da177e4SLinus Torvalds } 12951da177e4SLinus Torvalds } 12961da177e4SLinus Torvalds } 12971da177e4SLinus Torvalds while (out < BH_LRU_SIZE) 12981da177e4SLinus Torvalds bhs[out++] = NULL; 1299c7b92516SChristoph Lameter memcpy(__this_cpu_ptr(&bh_lrus.bhs), bhs, sizeof(bhs)); 13001da177e4SLinus Torvalds } 13011da177e4SLinus Torvalds bh_lru_unlock(); 13021da177e4SLinus Torvalds 13031da177e4SLinus Torvalds if (evictee) 13041da177e4SLinus Torvalds __brelse(evictee); 13051da177e4SLinus Torvalds } 13061da177e4SLinus Torvalds 13071da177e4SLinus Torvalds /* 13081da177e4SLinus Torvalds * Look up the bh in this cpu's LRU. If it's there, move it to the head. 13091da177e4SLinus Torvalds */ 1310858119e1SArjan van de Ven static struct buffer_head * 13113991d3bdSTomasz Kvarsin lookup_bh_lru(struct block_device *bdev, sector_t block, unsigned size) 13121da177e4SLinus Torvalds { 13131da177e4SLinus Torvalds struct buffer_head *ret = NULL; 13143991d3bdSTomasz Kvarsin unsigned int i; 13151da177e4SLinus Torvalds 13161da177e4SLinus Torvalds check_irqs_on(); 13171da177e4SLinus Torvalds bh_lru_lock(); 13181da177e4SLinus Torvalds for (i = 0; i < BH_LRU_SIZE; i++) { 1319c7b92516SChristoph Lameter struct buffer_head *bh = __this_cpu_read(bh_lrus.bhs[i]); 13201da177e4SLinus Torvalds 13211da177e4SLinus Torvalds if (bh && bh->b_bdev == bdev && 13221da177e4SLinus Torvalds bh->b_blocknr == block && bh->b_size == size) { 13231da177e4SLinus Torvalds if (i) { 13241da177e4SLinus Torvalds while (i) { 1325c7b92516SChristoph Lameter __this_cpu_write(bh_lrus.bhs[i], 1326c7b92516SChristoph Lameter __this_cpu_read(bh_lrus.bhs[i - 1])); 13271da177e4SLinus Torvalds i--; 13281da177e4SLinus Torvalds } 1329c7b92516SChristoph Lameter __this_cpu_write(bh_lrus.bhs[0], bh); 13301da177e4SLinus Torvalds } 13311da177e4SLinus Torvalds get_bh(bh); 13321da177e4SLinus Torvalds ret = bh; 13331da177e4SLinus Torvalds break; 13341da177e4SLinus Torvalds } 13351da177e4SLinus Torvalds } 13361da177e4SLinus Torvalds bh_lru_unlock(); 13371da177e4SLinus Torvalds return ret; 13381da177e4SLinus Torvalds } 13391da177e4SLinus Torvalds 13401da177e4SLinus Torvalds /* 13411da177e4SLinus Torvalds * Perform a pagecache lookup for the matching buffer. If it's there, refresh 13421da177e4SLinus Torvalds * it in the LRU and mark it as accessed. If it is not present then return 13431da177e4SLinus Torvalds * NULL 13441da177e4SLinus Torvalds */ 13451da177e4SLinus Torvalds struct buffer_head * 13463991d3bdSTomasz Kvarsin __find_get_block(struct block_device *bdev, sector_t block, unsigned size) 13471da177e4SLinus Torvalds { 13481da177e4SLinus Torvalds struct buffer_head *bh = lookup_bh_lru(bdev, block, size); 13491da177e4SLinus Torvalds 13501da177e4SLinus Torvalds if (bh == NULL) { 1351385fd4c5SCoywolf Qi Hunt bh = __find_get_block_slow(bdev, block); 13521da177e4SLinus Torvalds if (bh) 13531da177e4SLinus Torvalds bh_lru_install(bh); 13541da177e4SLinus Torvalds } 13551da177e4SLinus Torvalds if (bh) 13561da177e4SLinus Torvalds touch_buffer(bh); 13571da177e4SLinus Torvalds return bh; 13581da177e4SLinus Torvalds } 13591da177e4SLinus Torvalds EXPORT_SYMBOL(__find_get_block); 13601da177e4SLinus Torvalds 13611da177e4SLinus Torvalds /* 13621da177e4SLinus Torvalds * __getblk will locate (and, if necessary, create) the buffer_head 13631da177e4SLinus Torvalds * which corresponds to the passed block_device, block and size. The 13641da177e4SLinus Torvalds * returned buffer has its reference count incremented. 13651da177e4SLinus Torvalds * 13661da177e4SLinus Torvalds * __getblk() cannot fail - it just keeps trying. If you pass it an 13671da177e4SLinus Torvalds * illegal block number, __getblk() will happily return a buffer_head 13681da177e4SLinus Torvalds * which represents the non-existent block. Very weird. 13691da177e4SLinus Torvalds * 13701da177e4SLinus Torvalds * __getblk() will lock up the machine if grow_dev_page's try_to_free_buffers() 13711da177e4SLinus Torvalds * attempt is failing. FIXME, perhaps? 13721da177e4SLinus Torvalds */ 13731da177e4SLinus Torvalds struct buffer_head * 13743991d3bdSTomasz Kvarsin __getblk(struct block_device *bdev, sector_t block, unsigned size) 13751da177e4SLinus Torvalds { 13761da177e4SLinus Torvalds struct buffer_head *bh = __find_get_block(bdev, block, size); 13771da177e4SLinus Torvalds 13781da177e4SLinus Torvalds might_sleep(); 13791da177e4SLinus Torvalds if (bh == NULL) 13801da177e4SLinus Torvalds bh = __getblk_slow(bdev, block, size); 13811da177e4SLinus Torvalds return bh; 13821da177e4SLinus Torvalds } 13831da177e4SLinus Torvalds EXPORT_SYMBOL(__getblk); 13841da177e4SLinus Torvalds 13851da177e4SLinus Torvalds /* 13861da177e4SLinus Torvalds * Do async read-ahead on a buffer.. 13871da177e4SLinus Torvalds */ 13883991d3bdSTomasz Kvarsin void __breadahead(struct block_device *bdev, sector_t block, unsigned size) 13891da177e4SLinus Torvalds { 13901da177e4SLinus Torvalds struct buffer_head *bh = __getblk(bdev, block, size); 1391a3e713b5SAndrew Morton if (likely(bh)) { 13921da177e4SLinus Torvalds ll_rw_block(READA, 1, &bh); 13931da177e4SLinus Torvalds brelse(bh); 13941da177e4SLinus Torvalds } 1395a3e713b5SAndrew Morton } 13961da177e4SLinus Torvalds EXPORT_SYMBOL(__breadahead); 13971da177e4SLinus Torvalds 13981da177e4SLinus Torvalds /** 13991da177e4SLinus Torvalds * __bread() - reads a specified block and returns the bh 140067be2dd1SMartin Waitz * @bdev: the block_device to read from 14011da177e4SLinus Torvalds * @block: number of block 14021da177e4SLinus Torvalds * @size: size (in bytes) to read 14031da177e4SLinus Torvalds * 14041da177e4SLinus Torvalds * Reads a specified block, and returns buffer head that contains it. 14051da177e4SLinus Torvalds * It returns NULL if the block was unreadable. 14061da177e4SLinus Torvalds */ 14071da177e4SLinus Torvalds struct buffer_head * 14083991d3bdSTomasz Kvarsin __bread(struct block_device *bdev, sector_t block, unsigned size) 14091da177e4SLinus Torvalds { 14101da177e4SLinus Torvalds struct buffer_head *bh = __getblk(bdev, block, size); 14111da177e4SLinus Torvalds 1412a3e713b5SAndrew Morton if (likely(bh) && !buffer_uptodate(bh)) 14131da177e4SLinus Torvalds bh = __bread_slow(bh); 14141da177e4SLinus Torvalds return bh; 14151da177e4SLinus Torvalds } 14161da177e4SLinus Torvalds EXPORT_SYMBOL(__bread); 14171da177e4SLinus Torvalds 14181da177e4SLinus Torvalds /* 14191da177e4SLinus Torvalds * invalidate_bh_lrus() is called rarely - but not only at unmount. 14201da177e4SLinus Torvalds * This doesn't race because it runs in each cpu either in irq 14211da177e4SLinus Torvalds * or with preempt disabled. 14221da177e4SLinus Torvalds */ 14231da177e4SLinus Torvalds static void invalidate_bh_lru(void *arg) 14241da177e4SLinus Torvalds { 14251da177e4SLinus Torvalds struct bh_lru *b = &get_cpu_var(bh_lrus); 14261da177e4SLinus Torvalds int i; 14271da177e4SLinus Torvalds 14281da177e4SLinus Torvalds for (i = 0; i < BH_LRU_SIZE; i++) { 14291da177e4SLinus Torvalds brelse(b->bhs[i]); 14301da177e4SLinus Torvalds b->bhs[i] = NULL; 14311da177e4SLinus Torvalds } 14321da177e4SLinus Torvalds put_cpu_var(bh_lrus); 14331da177e4SLinus Torvalds } 14341da177e4SLinus Torvalds 1435f9a14399SPeter Zijlstra void invalidate_bh_lrus(void) 14361da177e4SLinus Torvalds { 143715c8b6c1SJens Axboe on_each_cpu(invalidate_bh_lru, NULL, 1); 14381da177e4SLinus Torvalds } 14399db5579bSNick Piggin EXPORT_SYMBOL_GPL(invalidate_bh_lrus); 14401da177e4SLinus Torvalds 14411da177e4SLinus Torvalds void set_bh_page(struct buffer_head *bh, 14421da177e4SLinus Torvalds struct page *page, unsigned long offset) 14431da177e4SLinus Torvalds { 14441da177e4SLinus Torvalds bh->b_page = page; 1445e827f923SEric Sesterhenn BUG_ON(offset >= PAGE_SIZE); 14461da177e4SLinus Torvalds if (PageHighMem(page)) 14471da177e4SLinus Torvalds /* 14481da177e4SLinus Torvalds * This catches illegal uses and preserves the offset: 14491da177e4SLinus Torvalds */ 14501da177e4SLinus Torvalds bh->b_data = (char *)(0 + offset); 14511da177e4SLinus Torvalds else 14521da177e4SLinus Torvalds bh->b_data = page_address(page) + offset; 14531da177e4SLinus Torvalds } 14541da177e4SLinus Torvalds EXPORT_SYMBOL(set_bh_page); 14551da177e4SLinus Torvalds 14561da177e4SLinus Torvalds /* 14571da177e4SLinus Torvalds * Called when truncating a buffer on a page completely. 14581da177e4SLinus Torvalds */ 1459858119e1SArjan van de Ven static void discard_buffer(struct buffer_head * bh) 14601da177e4SLinus Torvalds { 14611da177e4SLinus Torvalds lock_buffer(bh); 14621da177e4SLinus Torvalds clear_buffer_dirty(bh); 14631da177e4SLinus Torvalds bh->b_bdev = NULL; 14641da177e4SLinus Torvalds clear_buffer_mapped(bh); 14651da177e4SLinus Torvalds clear_buffer_req(bh); 14661da177e4SLinus Torvalds clear_buffer_new(bh); 14671da177e4SLinus Torvalds clear_buffer_delay(bh); 146833a266ddSDavid Chinner clear_buffer_unwritten(bh); 14691da177e4SLinus Torvalds unlock_buffer(bh); 14701da177e4SLinus Torvalds } 14711da177e4SLinus Torvalds 14721da177e4SLinus Torvalds /** 14731da177e4SLinus Torvalds * block_invalidatepage - invalidate part of all of a buffer-backed page 14741da177e4SLinus Torvalds * 14751da177e4SLinus Torvalds * @page: the page which is affected 14761da177e4SLinus Torvalds * @offset: the index of the truncation point 14771da177e4SLinus Torvalds * 14781da177e4SLinus Torvalds * block_invalidatepage() is called when all or part of the page has become 14791da177e4SLinus Torvalds * invalidatedby a truncate operation. 14801da177e4SLinus Torvalds * 14811da177e4SLinus Torvalds * block_invalidatepage() does not have to release all buffers, but it must 14821da177e4SLinus Torvalds * ensure that no dirty buffer is left outside @offset and that no I/O 14831da177e4SLinus Torvalds * is underway against any of the blocks which are outside the truncation 14841da177e4SLinus Torvalds * point. Because the caller is about to free (and possibly reuse) those 14851da177e4SLinus Torvalds * blocks on-disk. 14861da177e4SLinus Torvalds */ 14872ff28e22SNeilBrown void block_invalidatepage(struct page *page, unsigned long offset) 14881da177e4SLinus Torvalds { 14891da177e4SLinus Torvalds struct buffer_head *head, *bh, *next; 14901da177e4SLinus Torvalds unsigned int curr_off = 0; 14911da177e4SLinus Torvalds 14921da177e4SLinus Torvalds BUG_ON(!PageLocked(page)); 14931da177e4SLinus Torvalds if (!page_has_buffers(page)) 14941da177e4SLinus Torvalds goto out; 14951da177e4SLinus Torvalds 14961da177e4SLinus Torvalds head = page_buffers(page); 14971da177e4SLinus Torvalds bh = head; 14981da177e4SLinus Torvalds do { 14991da177e4SLinus Torvalds unsigned int next_off = curr_off + bh->b_size; 15001da177e4SLinus Torvalds next = bh->b_this_page; 15011da177e4SLinus Torvalds 15021da177e4SLinus Torvalds /* 15031da177e4SLinus Torvalds * is this block fully invalidated? 15041da177e4SLinus Torvalds */ 15051da177e4SLinus Torvalds if (offset <= curr_off) 15061da177e4SLinus Torvalds discard_buffer(bh); 15071da177e4SLinus Torvalds curr_off = next_off; 15081da177e4SLinus Torvalds bh = next; 15091da177e4SLinus Torvalds } while (bh != head); 15101da177e4SLinus Torvalds 15111da177e4SLinus Torvalds /* 15121da177e4SLinus Torvalds * We release buffers only if the entire page is being invalidated. 15131da177e4SLinus Torvalds * The get_block cached value has been unconditionally invalidated, 15141da177e4SLinus Torvalds * so real IO is not possible anymore. 15151da177e4SLinus Torvalds */ 15161da177e4SLinus Torvalds if (offset == 0) 15172ff28e22SNeilBrown try_to_release_page(page, 0); 15181da177e4SLinus Torvalds out: 15192ff28e22SNeilBrown return; 15201da177e4SLinus Torvalds } 15211da177e4SLinus Torvalds EXPORT_SYMBOL(block_invalidatepage); 15221da177e4SLinus Torvalds 15231da177e4SLinus Torvalds /* 15241da177e4SLinus Torvalds * We attach and possibly dirty the buffers atomically wrt 15251da177e4SLinus Torvalds * __set_page_dirty_buffers() via private_lock. try_to_free_buffers 15261da177e4SLinus Torvalds * is already excluded via the page lock. 15271da177e4SLinus Torvalds */ 15281da177e4SLinus Torvalds void create_empty_buffers(struct page *page, 15291da177e4SLinus Torvalds unsigned long blocksize, unsigned long b_state) 15301da177e4SLinus Torvalds { 15311da177e4SLinus Torvalds struct buffer_head *bh, *head, *tail; 15321da177e4SLinus Torvalds 15331da177e4SLinus Torvalds head = alloc_page_buffers(page, blocksize, 1); 15341da177e4SLinus Torvalds bh = head; 15351da177e4SLinus Torvalds do { 15361da177e4SLinus Torvalds bh->b_state |= b_state; 15371da177e4SLinus Torvalds tail = bh; 15381da177e4SLinus Torvalds bh = bh->b_this_page; 15391da177e4SLinus Torvalds } while (bh); 15401da177e4SLinus Torvalds tail->b_this_page = head; 15411da177e4SLinus Torvalds 15421da177e4SLinus Torvalds spin_lock(&page->mapping->private_lock); 15431da177e4SLinus Torvalds if (PageUptodate(page) || PageDirty(page)) { 15441da177e4SLinus Torvalds bh = head; 15451da177e4SLinus Torvalds do { 15461da177e4SLinus Torvalds if (PageDirty(page)) 15471da177e4SLinus Torvalds set_buffer_dirty(bh); 15481da177e4SLinus Torvalds if (PageUptodate(page)) 15491da177e4SLinus Torvalds set_buffer_uptodate(bh); 15501da177e4SLinus Torvalds bh = bh->b_this_page; 15511da177e4SLinus Torvalds } while (bh != head); 15521da177e4SLinus Torvalds } 15531da177e4SLinus Torvalds attach_page_buffers(page, head); 15541da177e4SLinus Torvalds spin_unlock(&page->mapping->private_lock); 15551da177e4SLinus Torvalds } 15561da177e4SLinus Torvalds EXPORT_SYMBOL(create_empty_buffers); 15571da177e4SLinus Torvalds 15581da177e4SLinus Torvalds /* 15591da177e4SLinus Torvalds * We are taking a block for data and we don't want any output from any 15601da177e4SLinus Torvalds * buffer-cache aliases starting from return from that function and 15611da177e4SLinus Torvalds * until the moment when something will explicitly mark the buffer 15621da177e4SLinus Torvalds * dirty (hopefully that will not happen until we will free that block ;-) 15631da177e4SLinus Torvalds * We don't even need to mark it not-uptodate - nobody can expect 15641da177e4SLinus Torvalds * anything from a newly allocated buffer anyway. We used to used 15651da177e4SLinus Torvalds * unmap_buffer() for such invalidation, but that was wrong. We definitely 15661da177e4SLinus Torvalds * don't want to mark the alias unmapped, for example - it would confuse 15671da177e4SLinus Torvalds * anyone who might pick it with bread() afterwards... 15681da177e4SLinus Torvalds * 15691da177e4SLinus Torvalds * Also.. Note that bforget() doesn't lock the buffer. So there can 15701da177e4SLinus Torvalds * be writeout I/O going on against recently-freed buffers. We don't 15711da177e4SLinus Torvalds * wait on that I/O in bforget() - it's more efficient to wait on the I/O 15721da177e4SLinus Torvalds * only if we really need to. That happens here. 15731da177e4SLinus Torvalds */ 15741da177e4SLinus Torvalds void unmap_underlying_metadata(struct block_device *bdev, sector_t block) 15751da177e4SLinus Torvalds { 15761da177e4SLinus Torvalds struct buffer_head *old_bh; 15771da177e4SLinus Torvalds 15781da177e4SLinus Torvalds might_sleep(); 15791da177e4SLinus Torvalds 1580385fd4c5SCoywolf Qi Hunt old_bh = __find_get_block_slow(bdev, block); 15811da177e4SLinus Torvalds if (old_bh) { 15821da177e4SLinus Torvalds clear_buffer_dirty(old_bh); 15831da177e4SLinus Torvalds wait_on_buffer(old_bh); 15841da177e4SLinus Torvalds clear_buffer_req(old_bh); 15851da177e4SLinus Torvalds __brelse(old_bh); 15861da177e4SLinus Torvalds } 15871da177e4SLinus Torvalds } 15881da177e4SLinus Torvalds EXPORT_SYMBOL(unmap_underlying_metadata); 15891da177e4SLinus Torvalds 15901da177e4SLinus Torvalds /* 15911da177e4SLinus Torvalds * NOTE! All mapped/uptodate combinations are valid: 15921da177e4SLinus Torvalds * 15931da177e4SLinus Torvalds * Mapped Uptodate Meaning 15941da177e4SLinus Torvalds * 15951da177e4SLinus Torvalds * No No "unknown" - must do get_block() 15961da177e4SLinus Torvalds * No Yes "hole" - zero-filled 15971da177e4SLinus Torvalds * Yes No "allocated" - allocated on disk, not read in 15981da177e4SLinus Torvalds * Yes Yes "valid" - allocated and up-to-date in memory. 15991da177e4SLinus Torvalds * 16001da177e4SLinus Torvalds * "Dirty" is valid only with the last case (mapped+uptodate). 16011da177e4SLinus Torvalds */ 16021da177e4SLinus Torvalds 16031da177e4SLinus Torvalds /* 16041da177e4SLinus Torvalds * While block_write_full_page is writing back the dirty buffers under 16051da177e4SLinus Torvalds * the page lock, whoever dirtied the buffers may decide to clean them 16061da177e4SLinus Torvalds * again at any time. We handle that by only looking at the buffer 16071da177e4SLinus Torvalds * state inside lock_buffer(). 16081da177e4SLinus Torvalds * 16091da177e4SLinus Torvalds * If block_write_full_page() is called for regular writeback 16101da177e4SLinus Torvalds * (wbc->sync_mode == WB_SYNC_NONE) then it will redirty a page which has a 16111da177e4SLinus Torvalds * locked buffer. This only can happen if someone has written the buffer 16121da177e4SLinus Torvalds * directly, with submit_bh(). At the address_space level PageWriteback 16131da177e4SLinus Torvalds * prevents this contention from occurring. 16146e34eeddSTheodore Ts'o * 16156e34eeddSTheodore Ts'o * If block_write_full_page() is called with wbc->sync_mode == 1616721a9602SJens Axboe * WB_SYNC_ALL, the writes are posted using WRITE_SYNC; this 1617721a9602SJens Axboe * causes the writes to be flagged as synchronous writes. 16181da177e4SLinus Torvalds */ 16191da177e4SLinus Torvalds static int __block_write_full_page(struct inode *inode, struct page *page, 162035c80d5fSChris Mason get_block_t *get_block, struct writeback_control *wbc, 162135c80d5fSChris Mason bh_end_io_t *handler) 16221da177e4SLinus Torvalds { 16231da177e4SLinus Torvalds int err; 16241da177e4SLinus Torvalds sector_t block; 16251da177e4SLinus Torvalds sector_t last_block; 1626f0fbd5fcSAndrew Morton struct buffer_head *bh, *head; 1627b0cf2321SBadari Pulavarty const unsigned blocksize = 1 << inode->i_blkbits; 16281da177e4SLinus Torvalds int nr_underway = 0; 16296e34eeddSTheodore Ts'o int write_op = (wbc->sync_mode == WB_SYNC_ALL ? 1630721a9602SJens Axboe WRITE_SYNC : WRITE); 16311da177e4SLinus Torvalds 16321da177e4SLinus Torvalds BUG_ON(!PageLocked(page)); 16331da177e4SLinus Torvalds 16341da177e4SLinus Torvalds last_block = (i_size_read(inode) - 1) >> inode->i_blkbits; 16351da177e4SLinus Torvalds 16361da177e4SLinus Torvalds if (!page_has_buffers(page)) { 1637b0cf2321SBadari Pulavarty create_empty_buffers(page, blocksize, 16381da177e4SLinus Torvalds (1 << BH_Dirty)|(1 << BH_Uptodate)); 16391da177e4SLinus Torvalds } 16401da177e4SLinus Torvalds 16411da177e4SLinus Torvalds /* 16421da177e4SLinus Torvalds * Be very careful. We have no exclusion from __set_page_dirty_buffers 16431da177e4SLinus Torvalds * here, and the (potentially unmapped) buffers may become dirty at 16441da177e4SLinus Torvalds * any time. If a buffer becomes dirty here after we've inspected it 16451da177e4SLinus Torvalds * then we just miss that fact, and the page stays dirty. 16461da177e4SLinus Torvalds * 16471da177e4SLinus Torvalds * Buffers outside i_size may be dirtied by __set_page_dirty_buffers; 16481da177e4SLinus Torvalds * handle that here by just cleaning them. 16491da177e4SLinus Torvalds */ 16501da177e4SLinus Torvalds 165154b21a79SAndrew Morton block = (sector_t)page->index << (PAGE_CACHE_SHIFT - inode->i_blkbits); 16521da177e4SLinus Torvalds head = page_buffers(page); 16531da177e4SLinus Torvalds bh = head; 16541da177e4SLinus Torvalds 16551da177e4SLinus Torvalds /* 16561da177e4SLinus Torvalds * Get all the dirty buffers mapped to disk addresses and 16571da177e4SLinus Torvalds * handle any aliases from the underlying blockdev's mapping. 16581da177e4SLinus Torvalds */ 16591da177e4SLinus Torvalds do { 16601da177e4SLinus Torvalds if (block > last_block) { 16611da177e4SLinus Torvalds /* 16621da177e4SLinus Torvalds * mapped buffers outside i_size will occur, because 16631da177e4SLinus Torvalds * this page can be outside i_size when there is a 16641da177e4SLinus Torvalds * truncate in progress. 16651da177e4SLinus Torvalds */ 16661da177e4SLinus Torvalds /* 16671da177e4SLinus Torvalds * The buffer was zeroed by block_write_full_page() 16681da177e4SLinus Torvalds */ 16691da177e4SLinus Torvalds clear_buffer_dirty(bh); 16701da177e4SLinus Torvalds set_buffer_uptodate(bh); 167129a814d2SAlex Tomas } else if ((!buffer_mapped(bh) || buffer_delay(bh)) && 167229a814d2SAlex Tomas buffer_dirty(bh)) { 1673b0cf2321SBadari Pulavarty WARN_ON(bh->b_size != blocksize); 16741da177e4SLinus Torvalds err = get_block(inode, block, bh, 1); 16751da177e4SLinus Torvalds if (err) 16761da177e4SLinus Torvalds goto recover; 167729a814d2SAlex Tomas clear_buffer_delay(bh); 16781da177e4SLinus Torvalds if (buffer_new(bh)) { 16791da177e4SLinus Torvalds /* blockdev mappings never come here */ 16801da177e4SLinus Torvalds clear_buffer_new(bh); 16811da177e4SLinus Torvalds unmap_underlying_metadata(bh->b_bdev, 16821da177e4SLinus Torvalds bh->b_blocknr); 16831da177e4SLinus Torvalds } 16841da177e4SLinus Torvalds } 16851da177e4SLinus Torvalds bh = bh->b_this_page; 16861da177e4SLinus Torvalds block++; 16871da177e4SLinus Torvalds } while (bh != head); 16881da177e4SLinus Torvalds 16891da177e4SLinus Torvalds do { 16901da177e4SLinus Torvalds if (!buffer_mapped(bh)) 16911da177e4SLinus Torvalds continue; 16921da177e4SLinus Torvalds /* 16931da177e4SLinus Torvalds * If it's a fully non-blocking write attempt and we cannot 16941da177e4SLinus Torvalds * lock the buffer then redirty the page. Note that this can 16955b0830cbSJens Axboe * potentially cause a busy-wait loop from writeback threads 16965b0830cbSJens Axboe * and kswapd activity, but those code paths have their own 16975b0830cbSJens Axboe * higher-level throttling. 16981da177e4SLinus Torvalds */ 16991b430beeSWu Fengguang if (wbc->sync_mode != WB_SYNC_NONE) { 17001da177e4SLinus Torvalds lock_buffer(bh); 1701ca5de404SNick Piggin } else if (!trylock_buffer(bh)) { 17021da177e4SLinus Torvalds redirty_page_for_writepage(wbc, page); 17031da177e4SLinus Torvalds continue; 17041da177e4SLinus Torvalds } 17051da177e4SLinus Torvalds if (test_clear_buffer_dirty(bh)) { 170635c80d5fSChris Mason mark_buffer_async_write_endio(bh, handler); 17071da177e4SLinus Torvalds } else { 17081da177e4SLinus Torvalds unlock_buffer(bh); 17091da177e4SLinus Torvalds } 17101da177e4SLinus Torvalds } while ((bh = bh->b_this_page) != head); 17111da177e4SLinus Torvalds 17121da177e4SLinus Torvalds /* 17131da177e4SLinus Torvalds * The page and its buffers are protected by PageWriteback(), so we can 17141da177e4SLinus Torvalds * drop the bh refcounts early. 17151da177e4SLinus Torvalds */ 17161da177e4SLinus Torvalds BUG_ON(PageWriteback(page)); 17171da177e4SLinus Torvalds set_page_writeback(page); 17181da177e4SLinus Torvalds 17191da177e4SLinus Torvalds do { 17201da177e4SLinus Torvalds struct buffer_head *next = bh->b_this_page; 17211da177e4SLinus Torvalds if (buffer_async_write(bh)) { 1722a64c8610STheodore Ts'o submit_bh(write_op, bh); 17231da177e4SLinus Torvalds nr_underway++; 1724ad576e63SNick Piggin } 17251da177e4SLinus Torvalds bh = next; 17261da177e4SLinus Torvalds } while (bh != head); 172705937baaSAndrew Morton unlock_page(page); 17281da177e4SLinus Torvalds 17291da177e4SLinus Torvalds err = 0; 17301da177e4SLinus Torvalds done: 17311da177e4SLinus Torvalds if (nr_underway == 0) { 17321da177e4SLinus Torvalds /* 17331da177e4SLinus Torvalds * The page was marked dirty, but the buffers were 17341da177e4SLinus Torvalds * clean. Someone wrote them back by hand with 17351da177e4SLinus Torvalds * ll_rw_block/submit_bh. A rare case. 17361da177e4SLinus Torvalds */ 17371da177e4SLinus Torvalds end_page_writeback(page); 17383d67f2d7SNick Piggin 17391da177e4SLinus Torvalds /* 17401da177e4SLinus Torvalds * The page and buffer_heads can be released at any time from 17411da177e4SLinus Torvalds * here on. 17421da177e4SLinus Torvalds */ 17431da177e4SLinus Torvalds } 17441da177e4SLinus Torvalds return err; 17451da177e4SLinus Torvalds 17461da177e4SLinus Torvalds recover: 17471da177e4SLinus Torvalds /* 17481da177e4SLinus Torvalds * ENOSPC, or some other error. We may already have added some 17491da177e4SLinus Torvalds * blocks to the file, so we need to write these out to avoid 17501da177e4SLinus Torvalds * exposing stale data. 17511da177e4SLinus Torvalds * The page is currently locked and not marked for writeback 17521da177e4SLinus Torvalds */ 17531da177e4SLinus Torvalds bh = head; 17541da177e4SLinus Torvalds /* Recovery: lock and submit the mapped buffers */ 17551da177e4SLinus Torvalds do { 175629a814d2SAlex Tomas if (buffer_mapped(bh) && buffer_dirty(bh) && 175729a814d2SAlex Tomas !buffer_delay(bh)) { 17581da177e4SLinus Torvalds lock_buffer(bh); 175935c80d5fSChris Mason mark_buffer_async_write_endio(bh, handler); 17601da177e4SLinus Torvalds } else { 17611da177e4SLinus Torvalds /* 17621da177e4SLinus Torvalds * The buffer may have been set dirty during 17631da177e4SLinus Torvalds * attachment to a dirty page. 17641da177e4SLinus Torvalds */ 17651da177e4SLinus Torvalds clear_buffer_dirty(bh); 17661da177e4SLinus Torvalds } 17671da177e4SLinus Torvalds } while ((bh = bh->b_this_page) != head); 17681da177e4SLinus Torvalds SetPageError(page); 17691da177e4SLinus Torvalds BUG_ON(PageWriteback(page)); 17707e4c3690SAndrew Morton mapping_set_error(page->mapping, err); 17711da177e4SLinus Torvalds set_page_writeback(page); 17721da177e4SLinus Torvalds do { 17731da177e4SLinus Torvalds struct buffer_head *next = bh->b_this_page; 17741da177e4SLinus Torvalds if (buffer_async_write(bh)) { 17751da177e4SLinus Torvalds clear_buffer_dirty(bh); 1776a64c8610STheodore Ts'o submit_bh(write_op, bh); 17771da177e4SLinus Torvalds nr_underway++; 1778ad576e63SNick Piggin } 17791da177e4SLinus Torvalds bh = next; 17801da177e4SLinus Torvalds } while (bh != head); 1781ffda9d30SNick Piggin unlock_page(page); 17821da177e4SLinus Torvalds goto done; 17831da177e4SLinus Torvalds } 17841da177e4SLinus Torvalds 1785afddba49SNick Piggin /* 1786afddba49SNick Piggin * If a page has any new buffers, zero them out here, and mark them uptodate 1787afddba49SNick Piggin * and dirty so they'll be written out (in order to prevent uninitialised 1788afddba49SNick Piggin * block data from leaking). And clear the new bit. 1789afddba49SNick Piggin */ 1790afddba49SNick Piggin void page_zero_new_buffers(struct page *page, unsigned from, unsigned to) 1791afddba49SNick Piggin { 1792afddba49SNick Piggin unsigned int block_start, block_end; 1793afddba49SNick Piggin struct buffer_head *head, *bh; 1794afddba49SNick Piggin 1795afddba49SNick Piggin BUG_ON(!PageLocked(page)); 1796afddba49SNick Piggin if (!page_has_buffers(page)) 1797afddba49SNick Piggin return; 1798afddba49SNick Piggin 1799afddba49SNick Piggin bh = head = page_buffers(page); 1800afddba49SNick Piggin block_start = 0; 1801afddba49SNick Piggin do { 1802afddba49SNick Piggin block_end = block_start + bh->b_size; 1803afddba49SNick Piggin 1804afddba49SNick Piggin if (buffer_new(bh)) { 1805afddba49SNick Piggin if (block_end > from && block_start < to) { 1806afddba49SNick Piggin if (!PageUptodate(page)) { 1807afddba49SNick Piggin unsigned start, size; 1808afddba49SNick Piggin 1809afddba49SNick Piggin start = max(from, block_start); 1810afddba49SNick Piggin size = min(to, block_end) - start; 1811afddba49SNick Piggin 1812eebd2aa3SChristoph Lameter zero_user(page, start, size); 1813afddba49SNick Piggin set_buffer_uptodate(bh); 1814afddba49SNick Piggin } 1815afddba49SNick Piggin 1816afddba49SNick Piggin clear_buffer_new(bh); 1817afddba49SNick Piggin mark_buffer_dirty(bh); 1818afddba49SNick Piggin } 1819afddba49SNick Piggin } 1820afddba49SNick Piggin 1821afddba49SNick Piggin block_start = block_end; 1822afddba49SNick Piggin bh = bh->b_this_page; 1823afddba49SNick Piggin } while (bh != head); 1824afddba49SNick Piggin } 1825afddba49SNick Piggin EXPORT_SYMBOL(page_zero_new_buffers); 1826afddba49SNick Piggin 1827ebdec241SChristoph Hellwig int __block_write_begin(struct page *page, loff_t pos, unsigned len, 18286e1db88dSChristoph Hellwig get_block_t *get_block) 18291da177e4SLinus Torvalds { 1830ebdec241SChristoph Hellwig unsigned from = pos & (PAGE_CACHE_SIZE - 1); 1831ebdec241SChristoph Hellwig unsigned to = from + len; 18326e1db88dSChristoph Hellwig struct inode *inode = page->mapping->host; 18331da177e4SLinus Torvalds unsigned block_start, block_end; 18341da177e4SLinus Torvalds sector_t block; 18351da177e4SLinus Torvalds int err = 0; 18361da177e4SLinus Torvalds unsigned blocksize, bbits; 18371da177e4SLinus Torvalds struct buffer_head *bh, *head, *wait[2], **wait_bh=wait; 18381da177e4SLinus Torvalds 18391da177e4SLinus Torvalds BUG_ON(!PageLocked(page)); 18401da177e4SLinus Torvalds BUG_ON(from > PAGE_CACHE_SIZE); 18411da177e4SLinus Torvalds BUG_ON(to > PAGE_CACHE_SIZE); 18421da177e4SLinus Torvalds BUG_ON(from > to); 18431da177e4SLinus Torvalds 18441da177e4SLinus Torvalds blocksize = 1 << inode->i_blkbits; 18451da177e4SLinus Torvalds if (!page_has_buffers(page)) 18461da177e4SLinus Torvalds create_empty_buffers(page, blocksize, 0); 18471da177e4SLinus Torvalds head = page_buffers(page); 18481da177e4SLinus Torvalds 18491da177e4SLinus Torvalds bbits = inode->i_blkbits; 18501da177e4SLinus Torvalds block = (sector_t)page->index << (PAGE_CACHE_SHIFT - bbits); 18511da177e4SLinus Torvalds 18521da177e4SLinus Torvalds for(bh = head, block_start = 0; bh != head || !block_start; 18531da177e4SLinus Torvalds block++, block_start=block_end, bh = bh->b_this_page) { 18541da177e4SLinus Torvalds block_end = block_start + blocksize; 18551da177e4SLinus Torvalds if (block_end <= from || block_start >= to) { 18561da177e4SLinus Torvalds if (PageUptodate(page)) { 18571da177e4SLinus Torvalds if (!buffer_uptodate(bh)) 18581da177e4SLinus Torvalds set_buffer_uptodate(bh); 18591da177e4SLinus Torvalds } 18601da177e4SLinus Torvalds continue; 18611da177e4SLinus Torvalds } 18621da177e4SLinus Torvalds if (buffer_new(bh)) 18631da177e4SLinus Torvalds clear_buffer_new(bh); 18641da177e4SLinus Torvalds if (!buffer_mapped(bh)) { 1865b0cf2321SBadari Pulavarty WARN_ON(bh->b_size != blocksize); 18661da177e4SLinus Torvalds err = get_block(inode, block, bh, 1); 18671da177e4SLinus Torvalds if (err) 1868f3ddbdc6SNick Piggin break; 18691da177e4SLinus Torvalds if (buffer_new(bh)) { 18701da177e4SLinus Torvalds unmap_underlying_metadata(bh->b_bdev, 18711da177e4SLinus Torvalds bh->b_blocknr); 18721da177e4SLinus Torvalds if (PageUptodate(page)) { 1873637aff46SNick Piggin clear_buffer_new(bh); 18741da177e4SLinus Torvalds set_buffer_uptodate(bh); 1875637aff46SNick Piggin mark_buffer_dirty(bh); 18761da177e4SLinus Torvalds continue; 18771da177e4SLinus Torvalds } 1878eebd2aa3SChristoph Lameter if (block_end > to || block_start < from) 1879eebd2aa3SChristoph Lameter zero_user_segments(page, 1880eebd2aa3SChristoph Lameter to, block_end, 1881eebd2aa3SChristoph Lameter block_start, from); 18821da177e4SLinus Torvalds continue; 18831da177e4SLinus Torvalds } 18841da177e4SLinus Torvalds } 18851da177e4SLinus Torvalds if (PageUptodate(page)) { 18861da177e4SLinus Torvalds if (!buffer_uptodate(bh)) 18871da177e4SLinus Torvalds set_buffer_uptodate(bh); 18881da177e4SLinus Torvalds continue; 18891da177e4SLinus Torvalds } 18901da177e4SLinus Torvalds if (!buffer_uptodate(bh) && !buffer_delay(bh) && 189133a266ddSDavid Chinner !buffer_unwritten(bh) && 18921da177e4SLinus Torvalds (block_start < from || block_end > to)) { 18931da177e4SLinus Torvalds ll_rw_block(READ, 1, &bh); 18941da177e4SLinus Torvalds *wait_bh++=bh; 18951da177e4SLinus Torvalds } 18961da177e4SLinus Torvalds } 18971da177e4SLinus Torvalds /* 18981da177e4SLinus Torvalds * If we issued read requests - let them complete. 18991da177e4SLinus Torvalds */ 19001da177e4SLinus Torvalds while(wait_bh > wait) { 19011da177e4SLinus Torvalds wait_on_buffer(*--wait_bh); 19021da177e4SLinus Torvalds if (!buffer_uptodate(*wait_bh)) 1903f3ddbdc6SNick Piggin err = -EIO; 19041da177e4SLinus Torvalds } 19056e1db88dSChristoph Hellwig if (unlikely(err)) { 1906afddba49SNick Piggin page_zero_new_buffers(page, from, to); 19076e1db88dSChristoph Hellwig ClearPageUptodate(page); 19086e1db88dSChristoph Hellwig } 19091da177e4SLinus Torvalds return err; 19101da177e4SLinus Torvalds } 1911ebdec241SChristoph Hellwig EXPORT_SYMBOL(__block_write_begin); 19121da177e4SLinus Torvalds 19131da177e4SLinus Torvalds static int __block_commit_write(struct inode *inode, struct page *page, 19141da177e4SLinus Torvalds unsigned from, unsigned to) 19151da177e4SLinus Torvalds { 19161da177e4SLinus Torvalds unsigned block_start, block_end; 19171da177e4SLinus Torvalds int partial = 0; 19181da177e4SLinus Torvalds unsigned blocksize; 19191da177e4SLinus Torvalds struct buffer_head *bh, *head; 19201da177e4SLinus Torvalds 19211da177e4SLinus Torvalds blocksize = 1 << inode->i_blkbits; 19221da177e4SLinus Torvalds 19231da177e4SLinus Torvalds for(bh = head = page_buffers(page), block_start = 0; 19241da177e4SLinus Torvalds bh != head || !block_start; 19251da177e4SLinus Torvalds block_start=block_end, bh = bh->b_this_page) { 19261da177e4SLinus Torvalds block_end = block_start + blocksize; 19271da177e4SLinus Torvalds if (block_end <= from || block_start >= to) { 19281da177e4SLinus Torvalds if (!buffer_uptodate(bh)) 19291da177e4SLinus Torvalds partial = 1; 19301da177e4SLinus Torvalds } else { 19311da177e4SLinus Torvalds set_buffer_uptodate(bh); 19321da177e4SLinus Torvalds mark_buffer_dirty(bh); 19331da177e4SLinus Torvalds } 1934afddba49SNick Piggin clear_buffer_new(bh); 19351da177e4SLinus Torvalds } 19361da177e4SLinus Torvalds 19371da177e4SLinus Torvalds /* 19381da177e4SLinus Torvalds * If this is a partial write which happened to make all buffers 19391da177e4SLinus Torvalds * uptodate then we can optimize away a bogus readpage() for 19401da177e4SLinus Torvalds * the next read(). Here we 'discover' whether the page went 19411da177e4SLinus Torvalds * uptodate as a result of this (potentially partial) write. 19421da177e4SLinus Torvalds */ 19431da177e4SLinus Torvalds if (!partial) 19441da177e4SLinus Torvalds SetPageUptodate(page); 19451da177e4SLinus Torvalds return 0; 19461da177e4SLinus Torvalds } 19471da177e4SLinus Torvalds 19481da177e4SLinus Torvalds /* 1949155130a4SChristoph Hellwig * block_write_begin takes care of the basic task of block allocation and 1950155130a4SChristoph Hellwig * bringing partial write blocks uptodate first. 1951155130a4SChristoph Hellwig * 19527bb46a67Snpiggin@suse.de * The filesystem needs to handle block truncation upon failure. 1953afddba49SNick Piggin */ 1954155130a4SChristoph Hellwig int block_write_begin(struct address_space *mapping, loff_t pos, unsigned len, 1955155130a4SChristoph Hellwig unsigned flags, struct page **pagep, get_block_t *get_block) 1956afddba49SNick Piggin { 19576e1db88dSChristoph Hellwig pgoff_t index = pos >> PAGE_CACHE_SHIFT; 1958afddba49SNick Piggin struct page *page; 19596e1db88dSChristoph Hellwig int status; 1960afddba49SNick Piggin 196154566b2cSNick Piggin page = grab_cache_page_write_begin(mapping, index, flags); 19626e1db88dSChristoph Hellwig if (!page) 19636e1db88dSChristoph Hellwig return -ENOMEM; 1964afddba49SNick Piggin 19656e1db88dSChristoph Hellwig status = __block_write_begin(page, pos, len, get_block); 1966afddba49SNick Piggin if (unlikely(status)) { 1967afddba49SNick Piggin unlock_page(page); 1968afddba49SNick Piggin page_cache_release(page); 19696e1db88dSChristoph Hellwig page = NULL; 1970afddba49SNick Piggin } 1971afddba49SNick Piggin 19726e1db88dSChristoph Hellwig *pagep = page; 1973afddba49SNick Piggin return status; 1974afddba49SNick Piggin } 1975afddba49SNick Piggin EXPORT_SYMBOL(block_write_begin); 1976afddba49SNick Piggin 1977afddba49SNick Piggin int block_write_end(struct file *file, struct address_space *mapping, 1978afddba49SNick Piggin loff_t pos, unsigned len, unsigned copied, 1979afddba49SNick Piggin struct page *page, void *fsdata) 1980afddba49SNick Piggin { 1981afddba49SNick Piggin struct inode *inode = mapping->host; 1982afddba49SNick Piggin unsigned start; 1983afddba49SNick Piggin 1984afddba49SNick Piggin start = pos & (PAGE_CACHE_SIZE - 1); 1985afddba49SNick Piggin 1986afddba49SNick Piggin if (unlikely(copied < len)) { 1987afddba49SNick Piggin /* 1988afddba49SNick Piggin * The buffers that were written will now be uptodate, so we 1989afddba49SNick Piggin * don't have to worry about a readpage reading them and 1990afddba49SNick Piggin * overwriting a partial write. However if we have encountered 1991afddba49SNick Piggin * a short write and only partially written into a buffer, it 1992afddba49SNick Piggin * will not be marked uptodate, so a readpage might come in and 1993afddba49SNick Piggin * destroy our partial write. 1994afddba49SNick Piggin * 1995afddba49SNick Piggin * Do the simplest thing, and just treat any short write to a 1996afddba49SNick Piggin * non uptodate page as a zero-length write, and force the 1997afddba49SNick Piggin * caller to redo the whole thing. 1998afddba49SNick Piggin */ 1999afddba49SNick Piggin if (!PageUptodate(page)) 2000afddba49SNick Piggin copied = 0; 2001afddba49SNick Piggin 2002afddba49SNick Piggin page_zero_new_buffers(page, start+copied, start+len); 2003afddba49SNick Piggin } 2004afddba49SNick Piggin flush_dcache_page(page); 2005afddba49SNick Piggin 2006afddba49SNick Piggin /* This could be a short (even 0-length) commit */ 2007afddba49SNick Piggin __block_commit_write(inode, page, start, start+copied); 2008afddba49SNick Piggin 2009afddba49SNick Piggin return copied; 2010afddba49SNick Piggin } 2011afddba49SNick Piggin EXPORT_SYMBOL(block_write_end); 2012afddba49SNick Piggin 2013afddba49SNick Piggin int generic_write_end(struct file *file, struct address_space *mapping, 2014afddba49SNick Piggin loff_t pos, unsigned len, unsigned copied, 2015afddba49SNick Piggin struct page *page, void *fsdata) 2016afddba49SNick Piggin { 2017afddba49SNick Piggin struct inode *inode = mapping->host; 2018c7d206b3SJan Kara int i_size_changed = 0; 2019afddba49SNick Piggin 2020afddba49SNick Piggin copied = block_write_end(file, mapping, pos, len, copied, page, fsdata); 2021afddba49SNick Piggin 2022afddba49SNick Piggin /* 2023afddba49SNick Piggin * No need to use i_size_read() here, the i_size 2024afddba49SNick Piggin * cannot change under us because we hold i_mutex. 2025afddba49SNick Piggin * 2026afddba49SNick Piggin * But it's important to update i_size while still holding page lock: 2027afddba49SNick Piggin * page writeout could otherwise come in and zero beyond i_size. 2028afddba49SNick Piggin */ 2029afddba49SNick Piggin if (pos+copied > inode->i_size) { 2030afddba49SNick Piggin i_size_write(inode, pos+copied); 2031c7d206b3SJan Kara i_size_changed = 1; 2032afddba49SNick Piggin } 2033afddba49SNick Piggin 2034afddba49SNick Piggin unlock_page(page); 2035afddba49SNick Piggin page_cache_release(page); 2036afddba49SNick Piggin 2037c7d206b3SJan Kara /* 2038c7d206b3SJan Kara * Don't mark the inode dirty under page lock. First, it unnecessarily 2039c7d206b3SJan Kara * makes the holding time of page lock longer. Second, it forces lock 2040c7d206b3SJan Kara * ordering of page lock and transaction start for journaling 2041c7d206b3SJan Kara * filesystems. 2042c7d206b3SJan Kara */ 2043c7d206b3SJan Kara if (i_size_changed) 2044c7d206b3SJan Kara mark_inode_dirty(inode); 2045c7d206b3SJan Kara 2046afddba49SNick Piggin return copied; 2047afddba49SNick Piggin } 2048afddba49SNick Piggin EXPORT_SYMBOL(generic_write_end); 2049afddba49SNick Piggin 2050afddba49SNick Piggin /* 20518ab22b9aSHisashi Hifumi * block_is_partially_uptodate checks whether buffers within a page are 20528ab22b9aSHisashi Hifumi * uptodate or not. 20538ab22b9aSHisashi Hifumi * 20548ab22b9aSHisashi Hifumi * Returns true if all buffers which correspond to a file portion 20558ab22b9aSHisashi Hifumi * we want to read are uptodate. 20568ab22b9aSHisashi Hifumi */ 20578ab22b9aSHisashi Hifumi int block_is_partially_uptodate(struct page *page, read_descriptor_t *desc, 20588ab22b9aSHisashi Hifumi unsigned long from) 20598ab22b9aSHisashi Hifumi { 20608ab22b9aSHisashi Hifumi struct inode *inode = page->mapping->host; 20618ab22b9aSHisashi Hifumi unsigned block_start, block_end, blocksize; 20628ab22b9aSHisashi Hifumi unsigned to; 20638ab22b9aSHisashi Hifumi struct buffer_head *bh, *head; 20648ab22b9aSHisashi Hifumi int ret = 1; 20658ab22b9aSHisashi Hifumi 20668ab22b9aSHisashi Hifumi if (!page_has_buffers(page)) 20678ab22b9aSHisashi Hifumi return 0; 20688ab22b9aSHisashi Hifumi 20698ab22b9aSHisashi Hifumi blocksize = 1 << inode->i_blkbits; 20708ab22b9aSHisashi Hifumi to = min_t(unsigned, PAGE_CACHE_SIZE - from, desc->count); 20718ab22b9aSHisashi Hifumi to = from + to; 20728ab22b9aSHisashi Hifumi if (from < blocksize && to > PAGE_CACHE_SIZE - blocksize) 20738ab22b9aSHisashi Hifumi return 0; 20748ab22b9aSHisashi Hifumi 20758ab22b9aSHisashi Hifumi head = page_buffers(page); 20768ab22b9aSHisashi Hifumi bh = head; 20778ab22b9aSHisashi Hifumi block_start = 0; 20788ab22b9aSHisashi Hifumi do { 20798ab22b9aSHisashi Hifumi block_end = block_start + blocksize; 20808ab22b9aSHisashi Hifumi if (block_end > from && block_start < to) { 20818ab22b9aSHisashi Hifumi if (!buffer_uptodate(bh)) { 20828ab22b9aSHisashi Hifumi ret = 0; 20838ab22b9aSHisashi Hifumi break; 20848ab22b9aSHisashi Hifumi } 20858ab22b9aSHisashi Hifumi if (block_end >= to) 20868ab22b9aSHisashi Hifumi break; 20878ab22b9aSHisashi Hifumi } 20888ab22b9aSHisashi Hifumi block_start = block_end; 20898ab22b9aSHisashi Hifumi bh = bh->b_this_page; 20908ab22b9aSHisashi Hifumi } while (bh != head); 20918ab22b9aSHisashi Hifumi 20928ab22b9aSHisashi Hifumi return ret; 20938ab22b9aSHisashi Hifumi } 20948ab22b9aSHisashi Hifumi EXPORT_SYMBOL(block_is_partially_uptodate); 20958ab22b9aSHisashi Hifumi 20968ab22b9aSHisashi Hifumi /* 20971da177e4SLinus Torvalds * Generic "read page" function for block devices that have the normal 20981da177e4SLinus Torvalds * get_block functionality. This is most of the block device filesystems. 20991da177e4SLinus Torvalds * Reads the page asynchronously --- the unlock_buffer() and 21001da177e4SLinus Torvalds * set/clear_buffer_uptodate() functions propagate buffer state into the 21011da177e4SLinus Torvalds * page struct once IO has completed. 21021da177e4SLinus Torvalds */ 21031da177e4SLinus Torvalds int block_read_full_page(struct page *page, get_block_t *get_block) 21041da177e4SLinus Torvalds { 21051da177e4SLinus Torvalds struct inode *inode = page->mapping->host; 21061da177e4SLinus Torvalds sector_t iblock, lblock; 21071da177e4SLinus Torvalds struct buffer_head *bh, *head, *arr[MAX_BUF_PER_PAGE]; 21081da177e4SLinus Torvalds unsigned int blocksize; 21091da177e4SLinus Torvalds int nr, i; 21101da177e4SLinus Torvalds int fully_mapped = 1; 21111da177e4SLinus Torvalds 2112cd7619d6SMatt Mackall BUG_ON(!PageLocked(page)); 21131da177e4SLinus Torvalds blocksize = 1 << inode->i_blkbits; 21141da177e4SLinus Torvalds if (!page_has_buffers(page)) 21151da177e4SLinus Torvalds create_empty_buffers(page, blocksize, 0); 21161da177e4SLinus Torvalds head = page_buffers(page); 21171da177e4SLinus Torvalds 21181da177e4SLinus Torvalds iblock = (sector_t)page->index << (PAGE_CACHE_SHIFT - inode->i_blkbits); 21191da177e4SLinus Torvalds lblock = (i_size_read(inode)+blocksize-1) >> inode->i_blkbits; 21201da177e4SLinus Torvalds bh = head; 21211da177e4SLinus Torvalds nr = 0; 21221da177e4SLinus Torvalds i = 0; 21231da177e4SLinus Torvalds 21241da177e4SLinus Torvalds do { 21251da177e4SLinus Torvalds if (buffer_uptodate(bh)) 21261da177e4SLinus Torvalds continue; 21271da177e4SLinus Torvalds 21281da177e4SLinus Torvalds if (!buffer_mapped(bh)) { 2129c64610baSAndrew Morton int err = 0; 2130c64610baSAndrew Morton 21311da177e4SLinus Torvalds fully_mapped = 0; 21321da177e4SLinus Torvalds if (iblock < lblock) { 2133b0cf2321SBadari Pulavarty WARN_ON(bh->b_size != blocksize); 2134c64610baSAndrew Morton err = get_block(inode, iblock, bh, 0); 2135c64610baSAndrew Morton if (err) 21361da177e4SLinus Torvalds SetPageError(page); 21371da177e4SLinus Torvalds } 21381da177e4SLinus Torvalds if (!buffer_mapped(bh)) { 2139eebd2aa3SChristoph Lameter zero_user(page, i * blocksize, blocksize); 2140c64610baSAndrew Morton if (!err) 21411da177e4SLinus Torvalds set_buffer_uptodate(bh); 21421da177e4SLinus Torvalds continue; 21431da177e4SLinus Torvalds } 21441da177e4SLinus Torvalds /* 21451da177e4SLinus Torvalds * get_block() might have updated the buffer 21461da177e4SLinus Torvalds * synchronously 21471da177e4SLinus Torvalds */ 21481da177e4SLinus Torvalds if (buffer_uptodate(bh)) 21491da177e4SLinus Torvalds continue; 21501da177e4SLinus Torvalds } 21511da177e4SLinus Torvalds arr[nr++] = bh; 21521da177e4SLinus Torvalds } while (i++, iblock++, (bh = bh->b_this_page) != head); 21531da177e4SLinus Torvalds 21541da177e4SLinus Torvalds if (fully_mapped) 21551da177e4SLinus Torvalds SetPageMappedToDisk(page); 21561da177e4SLinus Torvalds 21571da177e4SLinus Torvalds if (!nr) { 21581da177e4SLinus Torvalds /* 21591da177e4SLinus Torvalds * All buffers are uptodate - we can set the page uptodate 21601da177e4SLinus Torvalds * as well. But not if get_block() returned an error. 21611da177e4SLinus Torvalds */ 21621da177e4SLinus Torvalds if (!PageError(page)) 21631da177e4SLinus Torvalds SetPageUptodate(page); 21641da177e4SLinus Torvalds unlock_page(page); 21651da177e4SLinus Torvalds return 0; 21661da177e4SLinus Torvalds } 21671da177e4SLinus Torvalds 21681da177e4SLinus Torvalds /* Stage two: lock the buffers */ 21691da177e4SLinus Torvalds for (i = 0; i < nr; i++) { 21701da177e4SLinus Torvalds bh = arr[i]; 21711da177e4SLinus Torvalds lock_buffer(bh); 21721da177e4SLinus Torvalds mark_buffer_async_read(bh); 21731da177e4SLinus Torvalds } 21741da177e4SLinus Torvalds 21751da177e4SLinus Torvalds /* 21761da177e4SLinus Torvalds * Stage 3: start the IO. Check for uptodateness 21771da177e4SLinus Torvalds * inside the buffer lock in case another process reading 21781da177e4SLinus Torvalds * the underlying blockdev brought it uptodate (the sct fix). 21791da177e4SLinus Torvalds */ 21801da177e4SLinus Torvalds for (i = 0; i < nr; i++) { 21811da177e4SLinus Torvalds bh = arr[i]; 21821da177e4SLinus Torvalds if (buffer_uptodate(bh)) 21831da177e4SLinus Torvalds end_buffer_async_read(bh, 1); 21841da177e4SLinus Torvalds else 21851da177e4SLinus Torvalds submit_bh(READ, bh); 21861da177e4SLinus Torvalds } 21871da177e4SLinus Torvalds return 0; 21881da177e4SLinus Torvalds } 21891fe72eaaSH Hartley Sweeten EXPORT_SYMBOL(block_read_full_page); 21901da177e4SLinus Torvalds 21911da177e4SLinus Torvalds /* utility function for filesystems that need to do work on expanding 219289e10787SNick Piggin * truncates. Uses filesystem pagecache writes to allow the filesystem to 21931da177e4SLinus Torvalds * deal with the hole. 21941da177e4SLinus Torvalds */ 219589e10787SNick Piggin int generic_cont_expand_simple(struct inode *inode, loff_t size) 21961da177e4SLinus Torvalds { 21971da177e4SLinus Torvalds struct address_space *mapping = inode->i_mapping; 21981da177e4SLinus Torvalds struct page *page; 219989e10787SNick Piggin void *fsdata; 22001da177e4SLinus Torvalds int err; 22011da177e4SLinus Torvalds 2202c08d3b0eSnpiggin@suse.de err = inode_newsize_ok(inode, size); 2203c08d3b0eSnpiggin@suse.de if (err) 22041da177e4SLinus Torvalds goto out; 22051da177e4SLinus Torvalds 220689e10787SNick Piggin err = pagecache_write_begin(NULL, mapping, size, 0, 220789e10787SNick Piggin AOP_FLAG_UNINTERRUPTIBLE|AOP_FLAG_CONT_EXPAND, 220889e10787SNick Piggin &page, &fsdata); 220989e10787SNick Piggin if (err) 221005eb0b51SOGAWA Hirofumi goto out; 221105eb0b51SOGAWA Hirofumi 221289e10787SNick Piggin err = pagecache_write_end(NULL, mapping, size, 0, 0, page, fsdata); 221389e10787SNick Piggin BUG_ON(err > 0); 221405eb0b51SOGAWA Hirofumi 221505eb0b51SOGAWA Hirofumi out: 221605eb0b51SOGAWA Hirofumi return err; 221705eb0b51SOGAWA Hirofumi } 22181fe72eaaSH Hartley Sweeten EXPORT_SYMBOL(generic_cont_expand_simple); 221905eb0b51SOGAWA Hirofumi 2220f1e3af72SAdrian Bunk static int cont_expand_zero(struct file *file, struct address_space *mapping, 222189e10787SNick Piggin loff_t pos, loff_t *bytes) 222205eb0b51SOGAWA Hirofumi { 222389e10787SNick Piggin struct inode *inode = mapping->host; 222489e10787SNick Piggin unsigned blocksize = 1 << inode->i_blkbits; 222589e10787SNick Piggin struct page *page; 222689e10787SNick Piggin void *fsdata; 222789e10787SNick Piggin pgoff_t index, curidx; 222889e10787SNick Piggin loff_t curpos; 222989e10787SNick Piggin unsigned zerofrom, offset, len; 223089e10787SNick Piggin int err = 0; 223105eb0b51SOGAWA Hirofumi 223289e10787SNick Piggin index = pos >> PAGE_CACHE_SHIFT; 223389e10787SNick Piggin offset = pos & ~PAGE_CACHE_MASK; 223489e10787SNick Piggin 223589e10787SNick Piggin while (index > (curidx = (curpos = *bytes)>>PAGE_CACHE_SHIFT)) { 223689e10787SNick Piggin zerofrom = curpos & ~PAGE_CACHE_MASK; 223789e10787SNick Piggin if (zerofrom & (blocksize-1)) { 223889e10787SNick Piggin *bytes |= (blocksize-1); 223989e10787SNick Piggin (*bytes)++; 224089e10787SNick Piggin } 224189e10787SNick Piggin len = PAGE_CACHE_SIZE - zerofrom; 224289e10787SNick Piggin 224389e10787SNick Piggin err = pagecache_write_begin(file, mapping, curpos, len, 224489e10787SNick Piggin AOP_FLAG_UNINTERRUPTIBLE, 224589e10787SNick Piggin &page, &fsdata); 224689e10787SNick Piggin if (err) 224789e10787SNick Piggin goto out; 2248eebd2aa3SChristoph Lameter zero_user(page, zerofrom, len); 224989e10787SNick Piggin err = pagecache_write_end(file, mapping, curpos, len, len, 225089e10787SNick Piggin page, fsdata); 225189e10787SNick Piggin if (err < 0) 225289e10787SNick Piggin goto out; 225389e10787SNick Piggin BUG_ON(err != len); 225489e10787SNick Piggin err = 0; 2255061e9746SOGAWA Hirofumi 2256061e9746SOGAWA Hirofumi balance_dirty_pages_ratelimited(mapping); 225789e10787SNick Piggin } 225889e10787SNick Piggin 225989e10787SNick Piggin /* page covers the boundary, find the boundary offset */ 226089e10787SNick Piggin if (index == curidx) { 226189e10787SNick Piggin zerofrom = curpos & ~PAGE_CACHE_MASK; 226289e10787SNick Piggin /* if we will expand the thing last block will be filled */ 226389e10787SNick Piggin if (offset <= zerofrom) { 226489e10787SNick Piggin goto out; 226589e10787SNick Piggin } 226689e10787SNick Piggin if (zerofrom & (blocksize-1)) { 226789e10787SNick Piggin *bytes |= (blocksize-1); 226889e10787SNick Piggin (*bytes)++; 226989e10787SNick Piggin } 227089e10787SNick Piggin len = offset - zerofrom; 227189e10787SNick Piggin 227289e10787SNick Piggin err = pagecache_write_begin(file, mapping, curpos, len, 227389e10787SNick Piggin AOP_FLAG_UNINTERRUPTIBLE, 227489e10787SNick Piggin &page, &fsdata); 227589e10787SNick Piggin if (err) 227689e10787SNick Piggin goto out; 2277eebd2aa3SChristoph Lameter zero_user(page, zerofrom, len); 227889e10787SNick Piggin err = pagecache_write_end(file, mapping, curpos, len, len, 227989e10787SNick Piggin page, fsdata); 228089e10787SNick Piggin if (err < 0) 228189e10787SNick Piggin goto out; 228289e10787SNick Piggin BUG_ON(err != len); 228389e10787SNick Piggin err = 0; 228489e10787SNick Piggin } 228589e10787SNick Piggin out: 228689e10787SNick Piggin return err; 22871da177e4SLinus Torvalds } 22881da177e4SLinus Torvalds 22891da177e4SLinus Torvalds /* 22901da177e4SLinus Torvalds * For moronic filesystems that do not allow holes in file. 22911da177e4SLinus Torvalds * We may have to extend the file. 22921da177e4SLinus Torvalds */ 2293282dc178SChristoph Hellwig int cont_write_begin(struct file *file, struct address_space *mapping, 229489e10787SNick Piggin loff_t pos, unsigned len, unsigned flags, 229589e10787SNick Piggin struct page **pagep, void **fsdata, 229689e10787SNick Piggin get_block_t *get_block, loff_t *bytes) 22971da177e4SLinus Torvalds { 22981da177e4SLinus Torvalds struct inode *inode = mapping->host; 22991da177e4SLinus Torvalds unsigned blocksize = 1 << inode->i_blkbits; 230089e10787SNick Piggin unsigned zerofrom; 230189e10787SNick Piggin int err; 23021da177e4SLinus Torvalds 230389e10787SNick Piggin err = cont_expand_zero(file, mapping, pos, bytes); 230489e10787SNick Piggin if (err) 2305155130a4SChristoph Hellwig return err; 23061da177e4SLinus Torvalds 23071da177e4SLinus Torvalds zerofrom = *bytes & ~PAGE_CACHE_MASK; 230889e10787SNick Piggin if (pos+len > *bytes && zerofrom & (blocksize-1)) { 23091da177e4SLinus Torvalds *bytes |= (blocksize-1); 23101da177e4SLinus Torvalds (*bytes)++; 23111da177e4SLinus Torvalds } 23121da177e4SLinus Torvalds 2313155130a4SChristoph Hellwig return block_write_begin(mapping, pos, len, flags, pagep, get_block); 23141da177e4SLinus Torvalds } 23151fe72eaaSH Hartley Sweeten EXPORT_SYMBOL(cont_write_begin); 23161da177e4SLinus Torvalds 23171da177e4SLinus Torvalds int block_commit_write(struct page *page, unsigned from, unsigned to) 23181da177e4SLinus Torvalds { 23191da177e4SLinus Torvalds struct inode *inode = page->mapping->host; 23201da177e4SLinus Torvalds __block_commit_write(inode,page,from,to); 23211da177e4SLinus Torvalds return 0; 23221da177e4SLinus Torvalds } 23231fe72eaaSH Hartley Sweeten EXPORT_SYMBOL(block_commit_write); 23241da177e4SLinus Torvalds 232554171690SDavid Chinner /* 232654171690SDavid Chinner * block_page_mkwrite() is not allowed to change the file size as it gets 232754171690SDavid Chinner * called from a page fault handler when a page is first dirtied. Hence we must 232854171690SDavid Chinner * be careful to check for EOF conditions here. We set the page up correctly 232954171690SDavid Chinner * for a written page which means we get ENOSPC checking when writing into 233054171690SDavid Chinner * holes and correct delalloc and unwritten extent mapping on filesystems that 233154171690SDavid Chinner * support these features. 233254171690SDavid Chinner * 233354171690SDavid Chinner * We are not allowed to take the i_mutex here so we have to play games to 233454171690SDavid Chinner * protect against truncate races as the page could now be beyond EOF. Because 23357bb46a67Snpiggin@suse.de * truncate writes the inode size before removing pages, once we have the 233654171690SDavid Chinner * page lock we can determine safely if the page is beyond EOF. If it is not 233754171690SDavid Chinner * beyond EOF, then the page is guaranteed safe against truncation until we 233854171690SDavid Chinner * unlock the page. 2339ea13a864SJan Kara * 2340ea13a864SJan Kara * Direct callers of this function should call vfs_check_frozen() so that page 2341ea13a864SJan Kara * fault does not busyloop until the fs is thawed. 234254171690SDavid Chinner */ 234324da4fabSJan Kara int __block_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf, 234454171690SDavid Chinner get_block_t get_block) 234554171690SDavid Chinner { 2346c2ec175cSNick Piggin struct page *page = vmf->page; 234754171690SDavid Chinner struct inode *inode = vma->vm_file->f_path.dentry->d_inode; 234854171690SDavid Chinner unsigned long end; 234954171690SDavid Chinner loff_t size; 235024da4fabSJan Kara int ret; 235154171690SDavid Chinner 235254171690SDavid Chinner lock_page(page); 235354171690SDavid Chinner size = i_size_read(inode); 235454171690SDavid Chinner if ((page->mapping != inode->i_mapping) || 235518336338SNick Piggin (page_offset(page) > size)) { 235624da4fabSJan Kara /* We overload EFAULT to mean page got truncated */ 235724da4fabSJan Kara ret = -EFAULT; 235824da4fabSJan Kara goto out_unlock; 235954171690SDavid Chinner } 236054171690SDavid Chinner 236154171690SDavid Chinner /* page is wholly or partially inside EOF */ 236254171690SDavid Chinner if (((page->index + 1) << PAGE_CACHE_SHIFT) > size) 236354171690SDavid Chinner end = size & ~PAGE_CACHE_MASK; 236454171690SDavid Chinner else 236554171690SDavid Chinner end = PAGE_CACHE_SIZE; 236654171690SDavid Chinner 2367ebdec241SChristoph Hellwig ret = __block_write_begin(page, 0, end, get_block); 236854171690SDavid Chinner if (!ret) 236954171690SDavid Chinner ret = block_commit_write(page, 0, end); 237054171690SDavid Chinner 237124da4fabSJan Kara if (unlikely(ret < 0)) 237224da4fabSJan Kara goto out_unlock; 2373ea13a864SJan Kara /* 2374ea13a864SJan Kara * Freezing in progress? We check after the page is marked dirty and 2375ea13a864SJan Kara * with page lock held so if the test here fails, we are sure freezing 2376ea13a864SJan Kara * code will wait during syncing until the page fault is done - at that 2377ea13a864SJan Kara * point page will be dirty and unlocked so freezing code will write it 2378ea13a864SJan Kara * and writeprotect it again. 2379ea13a864SJan Kara */ 2380ea13a864SJan Kara set_page_dirty(page); 2381ea13a864SJan Kara if (inode->i_sb->s_frozen != SB_UNFROZEN) { 2382ea13a864SJan Kara ret = -EAGAIN; 2383ea13a864SJan Kara goto out_unlock; 2384ea13a864SJan Kara } 2385*d76ee18aSDarrick J. Wong wait_on_page_writeback(page); 238624da4fabSJan Kara return 0; 238724da4fabSJan Kara out_unlock: 2388b827e496SNick Piggin unlock_page(page); 238954171690SDavid Chinner return ret; 239054171690SDavid Chinner } 239124da4fabSJan Kara EXPORT_SYMBOL(__block_page_mkwrite); 239224da4fabSJan Kara 239324da4fabSJan Kara int block_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf, 239424da4fabSJan Kara get_block_t get_block) 239524da4fabSJan Kara { 2396ea13a864SJan Kara int ret; 2397ea13a864SJan Kara struct super_block *sb = vma->vm_file->f_path.dentry->d_inode->i_sb; 239824da4fabSJan Kara 2399ea13a864SJan Kara /* 2400ea13a864SJan Kara * This check is racy but catches the common case. The check in 2401ea13a864SJan Kara * __block_page_mkwrite() is reliable. 2402ea13a864SJan Kara */ 2403ea13a864SJan Kara vfs_check_frozen(sb, SB_FREEZE_WRITE); 2404ea13a864SJan Kara ret = __block_page_mkwrite(vma, vmf, get_block); 240524da4fabSJan Kara return block_page_mkwrite_return(ret); 240624da4fabSJan Kara } 24071fe72eaaSH Hartley Sweeten EXPORT_SYMBOL(block_page_mkwrite); 24081da177e4SLinus Torvalds 24091da177e4SLinus Torvalds /* 241003158cd7SNick Piggin * nobh_write_begin()'s prereads are special: the buffer_heads are freed 24111da177e4SLinus Torvalds * immediately, while under the page lock. So it needs a special end_io 24121da177e4SLinus Torvalds * handler which does not touch the bh after unlocking it. 24131da177e4SLinus Torvalds */ 24141da177e4SLinus Torvalds static void end_buffer_read_nobh(struct buffer_head *bh, int uptodate) 24151da177e4SLinus Torvalds { 241668671f35SDmitry Monakhov __end_buffer_read_notouch(bh, uptodate); 24171da177e4SLinus Torvalds } 24181da177e4SLinus Torvalds 24191da177e4SLinus Torvalds /* 242003158cd7SNick Piggin * Attach the singly-linked list of buffers created by nobh_write_begin, to 242103158cd7SNick Piggin * the page (converting it to circular linked list and taking care of page 242203158cd7SNick Piggin * dirty races). 242303158cd7SNick Piggin */ 242403158cd7SNick Piggin static void attach_nobh_buffers(struct page *page, struct buffer_head *head) 242503158cd7SNick Piggin { 242603158cd7SNick Piggin struct buffer_head *bh; 242703158cd7SNick Piggin 242803158cd7SNick Piggin BUG_ON(!PageLocked(page)); 242903158cd7SNick Piggin 243003158cd7SNick Piggin spin_lock(&page->mapping->private_lock); 243103158cd7SNick Piggin bh = head; 243203158cd7SNick Piggin do { 243303158cd7SNick Piggin if (PageDirty(page)) 243403158cd7SNick Piggin set_buffer_dirty(bh); 243503158cd7SNick Piggin if (!bh->b_this_page) 243603158cd7SNick Piggin bh->b_this_page = head; 243703158cd7SNick Piggin bh = bh->b_this_page; 243803158cd7SNick Piggin } while (bh != head); 243903158cd7SNick Piggin attach_page_buffers(page, head); 244003158cd7SNick Piggin spin_unlock(&page->mapping->private_lock); 244103158cd7SNick Piggin } 244203158cd7SNick Piggin 244303158cd7SNick Piggin /* 2444ea0f04e5SChristoph Hellwig * On entry, the page is fully not uptodate. 2445ea0f04e5SChristoph Hellwig * On exit the page is fully uptodate in the areas outside (from,to) 24467bb46a67Snpiggin@suse.de * The filesystem needs to handle block truncation upon failure. 24471da177e4SLinus Torvalds */ 2448ea0f04e5SChristoph Hellwig int nobh_write_begin(struct address_space *mapping, 244903158cd7SNick Piggin loff_t pos, unsigned len, unsigned flags, 245003158cd7SNick Piggin struct page **pagep, void **fsdata, 24511da177e4SLinus Torvalds get_block_t *get_block) 24521da177e4SLinus Torvalds { 245303158cd7SNick Piggin struct inode *inode = mapping->host; 24541da177e4SLinus Torvalds const unsigned blkbits = inode->i_blkbits; 24551da177e4SLinus Torvalds const unsigned blocksize = 1 << blkbits; 2456a4b0672dSNick Piggin struct buffer_head *head, *bh; 245703158cd7SNick Piggin struct page *page; 245803158cd7SNick Piggin pgoff_t index; 245903158cd7SNick Piggin unsigned from, to; 24601da177e4SLinus Torvalds unsigned block_in_page; 2461a4b0672dSNick Piggin unsigned block_start, block_end; 24621da177e4SLinus Torvalds sector_t block_in_file; 24631da177e4SLinus Torvalds int nr_reads = 0; 24641da177e4SLinus Torvalds int ret = 0; 24651da177e4SLinus Torvalds int is_mapped_to_disk = 1; 24661da177e4SLinus Torvalds 246703158cd7SNick Piggin index = pos >> PAGE_CACHE_SHIFT; 246803158cd7SNick Piggin from = pos & (PAGE_CACHE_SIZE - 1); 246903158cd7SNick Piggin to = from + len; 247003158cd7SNick Piggin 247154566b2cSNick Piggin page = grab_cache_page_write_begin(mapping, index, flags); 247203158cd7SNick Piggin if (!page) 247303158cd7SNick Piggin return -ENOMEM; 247403158cd7SNick Piggin *pagep = page; 247503158cd7SNick Piggin *fsdata = NULL; 247603158cd7SNick Piggin 247703158cd7SNick Piggin if (page_has_buffers(page)) { 2478309f77adSNamhyung Kim ret = __block_write_begin(page, pos, len, get_block); 2479309f77adSNamhyung Kim if (unlikely(ret)) 2480309f77adSNamhyung Kim goto out_release; 2481309f77adSNamhyung Kim return ret; 248203158cd7SNick Piggin } 2483a4b0672dSNick Piggin 24841da177e4SLinus Torvalds if (PageMappedToDisk(page)) 24851da177e4SLinus Torvalds return 0; 24861da177e4SLinus Torvalds 2487a4b0672dSNick Piggin /* 2488a4b0672dSNick Piggin * Allocate buffers so that we can keep track of state, and potentially 2489a4b0672dSNick Piggin * attach them to the page if an error occurs. In the common case of 2490a4b0672dSNick Piggin * no error, they will just be freed again without ever being attached 2491a4b0672dSNick Piggin * to the page (which is all OK, because we're under the page lock). 2492a4b0672dSNick Piggin * 2493a4b0672dSNick Piggin * Be careful: the buffer linked list is a NULL terminated one, rather 2494a4b0672dSNick Piggin * than the circular one we're used to. 2495a4b0672dSNick Piggin */ 2496a4b0672dSNick Piggin head = alloc_page_buffers(page, blocksize, 0); 249703158cd7SNick Piggin if (!head) { 249803158cd7SNick Piggin ret = -ENOMEM; 249903158cd7SNick Piggin goto out_release; 250003158cd7SNick Piggin } 2501a4b0672dSNick Piggin 25021da177e4SLinus Torvalds block_in_file = (sector_t)page->index << (PAGE_CACHE_SHIFT - blkbits); 25031da177e4SLinus Torvalds 25041da177e4SLinus Torvalds /* 25051da177e4SLinus Torvalds * We loop across all blocks in the page, whether or not they are 25061da177e4SLinus Torvalds * part of the affected region. This is so we can discover if the 25071da177e4SLinus Torvalds * page is fully mapped-to-disk. 25081da177e4SLinus Torvalds */ 2509a4b0672dSNick Piggin for (block_start = 0, block_in_page = 0, bh = head; 25101da177e4SLinus Torvalds block_start < PAGE_CACHE_SIZE; 2511a4b0672dSNick Piggin block_in_page++, block_start += blocksize, bh = bh->b_this_page) { 25121da177e4SLinus Torvalds int create; 25131da177e4SLinus Torvalds 2514a4b0672dSNick Piggin block_end = block_start + blocksize; 2515a4b0672dSNick Piggin bh->b_state = 0; 25161da177e4SLinus Torvalds create = 1; 25171da177e4SLinus Torvalds if (block_start >= to) 25181da177e4SLinus Torvalds create = 0; 25191da177e4SLinus Torvalds ret = get_block(inode, block_in_file + block_in_page, 2520a4b0672dSNick Piggin bh, create); 25211da177e4SLinus Torvalds if (ret) 25221da177e4SLinus Torvalds goto failed; 2523a4b0672dSNick Piggin if (!buffer_mapped(bh)) 25241da177e4SLinus Torvalds is_mapped_to_disk = 0; 2525a4b0672dSNick Piggin if (buffer_new(bh)) 2526a4b0672dSNick Piggin unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr); 2527a4b0672dSNick Piggin if (PageUptodate(page)) { 2528a4b0672dSNick Piggin set_buffer_uptodate(bh); 25291da177e4SLinus Torvalds continue; 2530a4b0672dSNick Piggin } 2531a4b0672dSNick Piggin if (buffer_new(bh) || !buffer_mapped(bh)) { 2532eebd2aa3SChristoph Lameter zero_user_segments(page, block_start, from, 2533eebd2aa3SChristoph Lameter to, block_end); 25341da177e4SLinus Torvalds continue; 25351da177e4SLinus Torvalds } 2536a4b0672dSNick Piggin if (buffer_uptodate(bh)) 25371da177e4SLinus Torvalds continue; /* reiserfs does this */ 25381da177e4SLinus Torvalds if (block_start < from || block_end > to) { 2539a4b0672dSNick Piggin lock_buffer(bh); 2540a4b0672dSNick Piggin bh->b_end_io = end_buffer_read_nobh; 2541a4b0672dSNick Piggin submit_bh(READ, bh); 2542a4b0672dSNick Piggin nr_reads++; 25431da177e4SLinus Torvalds } 25441da177e4SLinus Torvalds } 25451da177e4SLinus Torvalds 25461da177e4SLinus Torvalds if (nr_reads) { 25471da177e4SLinus Torvalds /* 25481da177e4SLinus Torvalds * The page is locked, so these buffers are protected from 25491da177e4SLinus Torvalds * any VM or truncate activity. Hence we don't need to care 25501da177e4SLinus Torvalds * for the buffer_head refcounts. 25511da177e4SLinus Torvalds */ 2552a4b0672dSNick Piggin for (bh = head; bh; bh = bh->b_this_page) { 25531da177e4SLinus Torvalds wait_on_buffer(bh); 25541da177e4SLinus Torvalds if (!buffer_uptodate(bh)) 25551da177e4SLinus Torvalds ret = -EIO; 25561da177e4SLinus Torvalds } 25571da177e4SLinus Torvalds if (ret) 25581da177e4SLinus Torvalds goto failed; 25591da177e4SLinus Torvalds } 25601da177e4SLinus Torvalds 25611da177e4SLinus Torvalds if (is_mapped_to_disk) 25621da177e4SLinus Torvalds SetPageMappedToDisk(page); 25631da177e4SLinus Torvalds 256403158cd7SNick Piggin *fsdata = head; /* to be released by nobh_write_end */ 2565a4b0672dSNick Piggin 25661da177e4SLinus Torvalds return 0; 25671da177e4SLinus Torvalds 25681da177e4SLinus Torvalds failed: 256903158cd7SNick Piggin BUG_ON(!ret); 25701da177e4SLinus Torvalds /* 2571a4b0672dSNick Piggin * Error recovery is a bit difficult. We need to zero out blocks that 2572a4b0672dSNick Piggin * were newly allocated, and dirty them to ensure they get written out. 2573a4b0672dSNick Piggin * Buffers need to be attached to the page at this point, otherwise 2574a4b0672dSNick Piggin * the handling of potential IO errors during writeout would be hard 2575a4b0672dSNick Piggin * (could try doing synchronous writeout, but what if that fails too?) 25761da177e4SLinus Torvalds */ 257703158cd7SNick Piggin attach_nobh_buffers(page, head); 257803158cd7SNick Piggin page_zero_new_buffers(page, from, to); 2579a4b0672dSNick Piggin 258003158cd7SNick Piggin out_release: 258103158cd7SNick Piggin unlock_page(page); 258203158cd7SNick Piggin page_cache_release(page); 258303158cd7SNick Piggin *pagep = NULL; 2584a4b0672dSNick Piggin 25857bb46a67Snpiggin@suse.de return ret; 25867bb46a67Snpiggin@suse.de } 258703158cd7SNick Piggin EXPORT_SYMBOL(nobh_write_begin); 25881da177e4SLinus Torvalds 258903158cd7SNick Piggin int nobh_write_end(struct file *file, struct address_space *mapping, 259003158cd7SNick Piggin loff_t pos, unsigned len, unsigned copied, 259103158cd7SNick Piggin struct page *page, void *fsdata) 25921da177e4SLinus Torvalds { 25931da177e4SLinus Torvalds struct inode *inode = page->mapping->host; 2594efdc3131SNick Piggin struct buffer_head *head = fsdata; 259503158cd7SNick Piggin struct buffer_head *bh; 25965b41e74aSDmitri Monakhov BUG_ON(fsdata != NULL && page_has_buffers(page)); 25971da177e4SLinus Torvalds 2598d4cf109fSDave Kleikamp if (unlikely(copied < len) && head) 259903158cd7SNick Piggin attach_nobh_buffers(page, head); 2600a4b0672dSNick Piggin if (page_has_buffers(page)) 260103158cd7SNick Piggin return generic_write_end(file, mapping, pos, len, 260203158cd7SNick Piggin copied, page, fsdata); 2603a4b0672dSNick Piggin 260422c8ca78SNick Piggin SetPageUptodate(page); 26051da177e4SLinus Torvalds set_page_dirty(page); 260603158cd7SNick Piggin if (pos+copied > inode->i_size) { 260703158cd7SNick Piggin i_size_write(inode, pos+copied); 26081da177e4SLinus Torvalds mark_inode_dirty(inode); 26091da177e4SLinus Torvalds } 261003158cd7SNick Piggin 261103158cd7SNick Piggin unlock_page(page); 261203158cd7SNick Piggin page_cache_release(page); 261303158cd7SNick Piggin 261403158cd7SNick Piggin while (head) { 261503158cd7SNick Piggin bh = head; 261603158cd7SNick Piggin head = head->b_this_page; 261703158cd7SNick Piggin free_buffer_head(bh); 26181da177e4SLinus Torvalds } 261903158cd7SNick Piggin 262003158cd7SNick Piggin return copied; 262103158cd7SNick Piggin } 262203158cd7SNick Piggin EXPORT_SYMBOL(nobh_write_end); 26231da177e4SLinus Torvalds 26241da177e4SLinus Torvalds /* 26251da177e4SLinus Torvalds * nobh_writepage() - based on block_full_write_page() except 26261da177e4SLinus Torvalds * that it tries to operate without attaching bufferheads to 26271da177e4SLinus Torvalds * the page. 26281da177e4SLinus Torvalds */ 26291da177e4SLinus Torvalds int nobh_writepage(struct page *page, get_block_t *get_block, 26301da177e4SLinus Torvalds struct writeback_control *wbc) 26311da177e4SLinus Torvalds { 26321da177e4SLinus Torvalds struct inode * const inode = page->mapping->host; 26331da177e4SLinus Torvalds loff_t i_size = i_size_read(inode); 26341da177e4SLinus Torvalds const pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT; 26351da177e4SLinus Torvalds unsigned offset; 26361da177e4SLinus Torvalds int ret; 26371da177e4SLinus Torvalds 26381da177e4SLinus Torvalds /* Is the page fully inside i_size? */ 26391da177e4SLinus Torvalds if (page->index < end_index) 26401da177e4SLinus Torvalds goto out; 26411da177e4SLinus Torvalds 26421da177e4SLinus Torvalds /* Is the page fully outside i_size? (truncate in progress) */ 26431da177e4SLinus Torvalds offset = i_size & (PAGE_CACHE_SIZE-1); 26441da177e4SLinus Torvalds if (page->index >= end_index+1 || !offset) { 26451da177e4SLinus Torvalds /* 26461da177e4SLinus Torvalds * The page may have dirty, unmapped buffers. For example, 26471da177e4SLinus Torvalds * they may have been added in ext3_writepage(). Make them 26481da177e4SLinus Torvalds * freeable here, so the page does not leak. 26491da177e4SLinus Torvalds */ 26501da177e4SLinus Torvalds #if 0 26511da177e4SLinus Torvalds /* Not really sure about this - do we need this ? */ 26521da177e4SLinus Torvalds if (page->mapping->a_ops->invalidatepage) 26531da177e4SLinus Torvalds page->mapping->a_ops->invalidatepage(page, offset); 26541da177e4SLinus Torvalds #endif 26551da177e4SLinus Torvalds unlock_page(page); 26561da177e4SLinus Torvalds return 0; /* don't care */ 26571da177e4SLinus Torvalds } 26581da177e4SLinus Torvalds 26591da177e4SLinus Torvalds /* 26601da177e4SLinus Torvalds * The page straddles i_size. It must be zeroed out on each and every 26611da177e4SLinus Torvalds * writepage invocation because it may be mmapped. "A file is mapped 26621da177e4SLinus Torvalds * in multiples of the page size. For a file that is not a multiple of 26631da177e4SLinus Torvalds * the page size, the remaining memory is zeroed when mapped, and 26641da177e4SLinus Torvalds * writes to that region are not written out to the file." 26651da177e4SLinus Torvalds */ 2666eebd2aa3SChristoph Lameter zero_user_segment(page, offset, PAGE_CACHE_SIZE); 26671da177e4SLinus Torvalds out: 26681da177e4SLinus Torvalds ret = mpage_writepage(page, get_block, wbc); 26691da177e4SLinus Torvalds if (ret == -EAGAIN) 267035c80d5fSChris Mason ret = __block_write_full_page(inode, page, get_block, wbc, 267135c80d5fSChris Mason end_buffer_async_write); 26721da177e4SLinus Torvalds return ret; 26731da177e4SLinus Torvalds } 26741da177e4SLinus Torvalds EXPORT_SYMBOL(nobh_writepage); 26751da177e4SLinus Torvalds 267603158cd7SNick Piggin int nobh_truncate_page(struct address_space *mapping, 267703158cd7SNick Piggin loff_t from, get_block_t *get_block) 26781da177e4SLinus Torvalds { 26791da177e4SLinus Torvalds pgoff_t index = from >> PAGE_CACHE_SHIFT; 26801da177e4SLinus Torvalds unsigned offset = from & (PAGE_CACHE_SIZE-1); 268103158cd7SNick Piggin unsigned blocksize; 268203158cd7SNick Piggin sector_t iblock; 268303158cd7SNick Piggin unsigned length, pos; 268403158cd7SNick Piggin struct inode *inode = mapping->host; 26851da177e4SLinus Torvalds struct page *page; 268603158cd7SNick Piggin struct buffer_head map_bh; 268703158cd7SNick Piggin int err; 26881da177e4SLinus Torvalds 268903158cd7SNick Piggin blocksize = 1 << inode->i_blkbits; 269003158cd7SNick Piggin length = offset & (blocksize - 1); 26911da177e4SLinus Torvalds 269203158cd7SNick Piggin /* Block boundary? Nothing to do */ 269303158cd7SNick Piggin if (!length) 269403158cd7SNick Piggin return 0; 269503158cd7SNick Piggin 269603158cd7SNick Piggin length = blocksize - length; 269703158cd7SNick Piggin iblock = (sector_t)index << (PAGE_CACHE_SHIFT - inode->i_blkbits); 269803158cd7SNick Piggin 26991da177e4SLinus Torvalds page = grab_cache_page(mapping, index); 270003158cd7SNick Piggin err = -ENOMEM; 27011da177e4SLinus Torvalds if (!page) 27021da177e4SLinus Torvalds goto out; 27031da177e4SLinus Torvalds 270403158cd7SNick Piggin if (page_has_buffers(page)) { 270503158cd7SNick Piggin has_buffers: 270603158cd7SNick Piggin unlock_page(page); 270703158cd7SNick Piggin page_cache_release(page); 270803158cd7SNick Piggin return block_truncate_page(mapping, from, get_block); 27091da177e4SLinus Torvalds } 271003158cd7SNick Piggin 271103158cd7SNick Piggin /* Find the buffer that contains "offset" */ 271203158cd7SNick Piggin pos = blocksize; 271303158cd7SNick Piggin while (offset >= pos) { 271403158cd7SNick Piggin iblock++; 271503158cd7SNick Piggin pos += blocksize; 271603158cd7SNick Piggin } 271703158cd7SNick Piggin 2718460bcf57STheodore Ts'o map_bh.b_size = blocksize; 2719460bcf57STheodore Ts'o map_bh.b_state = 0; 272003158cd7SNick Piggin err = get_block(inode, iblock, &map_bh, 0); 272103158cd7SNick Piggin if (err) 272203158cd7SNick Piggin goto unlock; 272303158cd7SNick Piggin /* unmapped? It's a hole - nothing to do */ 272403158cd7SNick Piggin if (!buffer_mapped(&map_bh)) 272503158cd7SNick Piggin goto unlock; 272603158cd7SNick Piggin 272703158cd7SNick Piggin /* Ok, it's mapped. Make sure it's up-to-date */ 272803158cd7SNick Piggin if (!PageUptodate(page)) { 272903158cd7SNick Piggin err = mapping->a_ops->readpage(NULL, page); 273003158cd7SNick Piggin if (err) { 273103158cd7SNick Piggin page_cache_release(page); 273203158cd7SNick Piggin goto out; 273303158cd7SNick Piggin } 273403158cd7SNick Piggin lock_page(page); 273503158cd7SNick Piggin if (!PageUptodate(page)) { 273603158cd7SNick Piggin err = -EIO; 273703158cd7SNick Piggin goto unlock; 273803158cd7SNick Piggin } 273903158cd7SNick Piggin if (page_has_buffers(page)) 274003158cd7SNick Piggin goto has_buffers; 274103158cd7SNick Piggin } 2742eebd2aa3SChristoph Lameter zero_user(page, offset, length); 274303158cd7SNick Piggin set_page_dirty(page); 274403158cd7SNick Piggin err = 0; 274503158cd7SNick Piggin 274603158cd7SNick Piggin unlock: 27471da177e4SLinus Torvalds unlock_page(page); 27481da177e4SLinus Torvalds page_cache_release(page); 27491da177e4SLinus Torvalds out: 275003158cd7SNick Piggin return err; 27511da177e4SLinus Torvalds } 27521da177e4SLinus Torvalds EXPORT_SYMBOL(nobh_truncate_page); 27531da177e4SLinus Torvalds 27541da177e4SLinus Torvalds int block_truncate_page(struct address_space *mapping, 27551da177e4SLinus Torvalds loff_t from, get_block_t *get_block) 27561da177e4SLinus Torvalds { 27571da177e4SLinus Torvalds pgoff_t index = from >> PAGE_CACHE_SHIFT; 27581da177e4SLinus Torvalds unsigned offset = from & (PAGE_CACHE_SIZE-1); 27591da177e4SLinus Torvalds unsigned blocksize; 276054b21a79SAndrew Morton sector_t iblock; 27611da177e4SLinus Torvalds unsigned length, pos; 27621da177e4SLinus Torvalds struct inode *inode = mapping->host; 27631da177e4SLinus Torvalds struct page *page; 27641da177e4SLinus Torvalds struct buffer_head *bh; 27651da177e4SLinus Torvalds int err; 27661da177e4SLinus Torvalds 27671da177e4SLinus Torvalds blocksize = 1 << inode->i_blkbits; 27681da177e4SLinus Torvalds length = offset & (blocksize - 1); 27691da177e4SLinus Torvalds 27701da177e4SLinus Torvalds /* Block boundary? Nothing to do */ 27711da177e4SLinus Torvalds if (!length) 27721da177e4SLinus Torvalds return 0; 27731da177e4SLinus Torvalds 27741da177e4SLinus Torvalds length = blocksize - length; 277554b21a79SAndrew Morton iblock = (sector_t)index << (PAGE_CACHE_SHIFT - inode->i_blkbits); 27761da177e4SLinus Torvalds 27771da177e4SLinus Torvalds page = grab_cache_page(mapping, index); 27781da177e4SLinus Torvalds err = -ENOMEM; 27791da177e4SLinus Torvalds if (!page) 27801da177e4SLinus Torvalds goto out; 27811da177e4SLinus Torvalds 27821da177e4SLinus Torvalds if (!page_has_buffers(page)) 27831da177e4SLinus Torvalds create_empty_buffers(page, blocksize, 0); 27841da177e4SLinus Torvalds 27851da177e4SLinus Torvalds /* Find the buffer that contains "offset" */ 27861da177e4SLinus Torvalds bh = page_buffers(page); 27871da177e4SLinus Torvalds pos = blocksize; 27881da177e4SLinus Torvalds while (offset >= pos) { 27891da177e4SLinus Torvalds bh = bh->b_this_page; 27901da177e4SLinus Torvalds iblock++; 27911da177e4SLinus Torvalds pos += blocksize; 27921da177e4SLinus Torvalds } 27931da177e4SLinus Torvalds 27941da177e4SLinus Torvalds err = 0; 27951da177e4SLinus Torvalds if (!buffer_mapped(bh)) { 2796b0cf2321SBadari Pulavarty WARN_ON(bh->b_size != blocksize); 27971da177e4SLinus Torvalds err = get_block(inode, iblock, bh, 0); 27981da177e4SLinus Torvalds if (err) 27991da177e4SLinus Torvalds goto unlock; 28001da177e4SLinus Torvalds /* unmapped? It's a hole - nothing to do */ 28011da177e4SLinus Torvalds if (!buffer_mapped(bh)) 28021da177e4SLinus Torvalds goto unlock; 28031da177e4SLinus Torvalds } 28041da177e4SLinus Torvalds 28051da177e4SLinus Torvalds /* Ok, it's mapped. Make sure it's up-to-date */ 28061da177e4SLinus Torvalds if (PageUptodate(page)) 28071da177e4SLinus Torvalds set_buffer_uptodate(bh); 28081da177e4SLinus Torvalds 280933a266ddSDavid Chinner if (!buffer_uptodate(bh) && !buffer_delay(bh) && !buffer_unwritten(bh)) { 28101da177e4SLinus Torvalds err = -EIO; 28111da177e4SLinus Torvalds ll_rw_block(READ, 1, &bh); 28121da177e4SLinus Torvalds wait_on_buffer(bh); 28131da177e4SLinus Torvalds /* Uhhuh. Read error. Complain and punt. */ 28141da177e4SLinus Torvalds if (!buffer_uptodate(bh)) 28151da177e4SLinus Torvalds goto unlock; 28161da177e4SLinus Torvalds } 28171da177e4SLinus Torvalds 2818eebd2aa3SChristoph Lameter zero_user(page, offset, length); 28191da177e4SLinus Torvalds mark_buffer_dirty(bh); 28201da177e4SLinus Torvalds err = 0; 28211da177e4SLinus Torvalds 28221da177e4SLinus Torvalds unlock: 28231da177e4SLinus Torvalds unlock_page(page); 28241da177e4SLinus Torvalds page_cache_release(page); 28251da177e4SLinus Torvalds out: 28261da177e4SLinus Torvalds return err; 28271da177e4SLinus Torvalds } 28281fe72eaaSH Hartley Sweeten EXPORT_SYMBOL(block_truncate_page); 28291da177e4SLinus Torvalds 28301da177e4SLinus Torvalds /* 28311da177e4SLinus Torvalds * The generic ->writepage function for buffer-backed address_spaces 283235c80d5fSChris Mason * this form passes in the end_io handler used to finish the IO. 28331da177e4SLinus Torvalds */ 283435c80d5fSChris Mason int block_write_full_page_endio(struct page *page, get_block_t *get_block, 283535c80d5fSChris Mason struct writeback_control *wbc, bh_end_io_t *handler) 28361da177e4SLinus Torvalds { 28371da177e4SLinus Torvalds struct inode * const inode = page->mapping->host; 28381da177e4SLinus Torvalds loff_t i_size = i_size_read(inode); 28391da177e4SLinus Torvalds const pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT; 28401da177e4SLinus Torvalds unsigned offset; 28411da177e4SLinus Torvalds 28421da177e4SLinus Torvalds /* Is the page fully inside i_size? */ 28431da177e4SLinus Torvalds if (page->index < end_index) 284435c80d5fSChris Mason return __block_write_full_page(inode, page, get_block, wbc, 284535c80d5fSChris Mason handler); 28461da177e4SLinus Torvalds 28471da177e4SLinus Torvalds /* Is the page fully outside i_size? (truncate in progress) */ 28481da177e4SLinus Torvalds offset = i_size & (PAGE_CACHE_SIZE-1); 28491da177e4SLinus Torvalds if (page->index >= end_index+1 || !offset) { 28501da177e4SLinus Torvalds /* 28511da177e4SLinus Torvalds * The page may have dirty, unmapped buffers. For example, 28521da177e4SLinus Torvalds * they may have been added in ext3_writepage(). Make them 28531da177e4SLinus Torvalds * freeable here, so the page does not leak. 28541da177e4SLinus Torvalds */ 2855aaa4059bSJan Kara do_invalidatepage(page, 0); 28561da177e4SLinus Torvalds unlock_page(page); 28571da177e4SLinus Torvalds return 0; /* don't care */ 28581da177e4SLinus Torvalds } 28591da177e4SLinus Torvalds 28601da177e4SLinus Torvalds /* 28611da177e4SLinus Torvalds * The page straddles i_size. It must be zeroed out on each and every 28622a61aa40SAdam Buchbinder * writepage invocation because it may be mmapped. "A file is mapped 28631da177e4SLinus Torvalds * in multiples of the page size. For a file that is not a multiple of 28641da177e4SLinus Torvalds * the page size, the remaining memory is zeroed when mapped, and 28651da177e4SLinus Torvalds * writes to that region are not written out to the file." 28661da177e4SLinus Torvalds */ 2867eebd2aa3SChristoph Lameter zero_user_segment(page, offset, PAGE_CACHE_SIZE); 286835c80d5fSChris Mason return __block_write_full_page(inode, page, get_block, wbc, handler); 28691da177e4SLinus Torvalds } 28701fe72eaaSH Hartley Sweeten EXPORT_SYMBOL(block_write_full_page_endio); 28711da177e4SLinus Torvalds 287235c80d5fSChris Mason /* 287335c80d5fSChris Mason * The generic ->writepage function for buffer-backed address_spaces 287435c80d5fSChris Mason */ 287535c80d5fSChris Mason int block_write_full_page(struct page *page, get_block_t *get_block, 287635c80d5fSChris Mason struct writeback_control *wbc) 287735c80d5fSChris Mason { 287835c80d5fSChris Mason return block_write_full_page_endio(page, get_block, wbc, 287935c80d5fSChris Mason end_buffer_async_write); 288035c80d5fSChris Mason } 28811fe72eaaSH Hartley Sweeten EXPORT_SYMBOL(block_write_full_page); 288235c80d5fSChris Mason 28831da177e4SLinus Torvalds sector_t generic_block_bmap(struct address_space *mapping, sector_t block, 28841da177e4SLinus Torvalds get_block_t *get_block) 28851da177e4SLinus Torvalds { 28861da177e4SLinus Torvalds struct buffer_head tmp; 28871da177e4SLinus Torvalds struct inode *inode = mapping->host; 28881da177e4SLinus Torvalds tmp.b_state = 0; 28891da177e4SLinus Torvalds tmp.b_blocknr = 0; 2890b0cf2321SBadari Pulavarty tmp.b_size = 1 << inode->i_blkbits; 28911da177e4SLinus Torvalds get_block(inode, block, &tmp, 0); 28921da177e4SLinus Torvalds return tmp.b_blocknr; 28931da177e4SLinus Torvalds } 28941fe72eaaSH Hartley Sweeten EXPORT_SYMBOL(generic_block_bmap); 28951da177e4SLinus Torvalds 28966712ecf8SNeilBrown static void end_bio_bh_io_sync(struct bio *bio, int err) 28971da177e4SLinus Torvalds { 28981da177e4SLinus Torvalds struct buffer_head *bh = bio->bi_private; 28991da177e4SLinus Torvalds 29001da177e4SLinus Torvalds if (err == -EOPNOTSUPP) { 29011da177e4SLinus Torvalds set_bit(BIO_EOPNOTSUPP, &bio->bi_flags); 29021da177e4SLinus Torvalds } 29031da177e4SLinus Torvalds 290408bafc03SKeith Mannthey if (unlikely (test_bit(BIO_QUIET,&bio->bi_flags))) 290508bafc03SKeith Mannthey set_bit(BH_Quiet, &bh->b_state); 290608bafc03SKeith Mannthey 29071da177e4SLinus Torvalds bh->b_end_io(bh, test_bit(BIO_UPTODATE, &bio->bi_flags)); 29081da177e4SLinus Torvalds bio_put(bio); 29091da177e4SLinus Torvalds } 29101da177e4SLinus Torvalds 29111da177e4SLinus Torvalds int submit_bh(int rw, struct buffer_head * bh) 29121da177e4SLinus Torvalds { 29131da177e4SLinus Torvalds struct bio *bio; 29141da177e4SLinus Torvalds int ret = 0; 29151da177e4SLinus Torvalds 29161da177e4SLinus Torvalds BUG_ON(!buffer_locked(bh)); 29171da177e4SLinus Torvalds BUG_ON(!buffer_mapped(bh)); 29181da177e4SLinus Torvalds BUG_ON(!bh->b_end_io); 29198fb0e342SAneesh Kumar K.V BUG_ON(buffer_delay(bh)); 29208fb0e342SAneesh Kumar K.V BUG_ON(buffer_unwritten(bh)); 29211da177e4SLinus Torvalds 292248fd4f93SJens Axboe /* 292348fd4f93SJens Axboe * Only clear out a write error when rewriting 29241da177e4SLinus Torvalds */ 292548fd4f93SJens Axboe if (test_set_buffer_req(bh) && (rw & WRITE)) 29261da177e4SLinus Torvalds clear_buffer_write_io_error(bh); 29271da177e4SLinus Torvalds 29281da177e4SLinus Torvalds /* 29291da177e4SLinus Torvalds * from here on down, it's all bio -- do the initial mapping, 29301da177e4SLinus Torvalds * submit_bio -> generic_make_request may further map this bio around 29311da177e4SLinus Torvalds */ 29321da177e4SLinus Torvalds bio = bio_alloc(GFP_NOIO, 1); 29331da177e4SLinus Torvalds 29341da177e4SLinus Torvalds bio->bi_sector = bh->b_blocknr * (bh->b_size >> 9); 29351da177e4SLinus Torvalds bio->bi_bdev = bh->b_bdev; 29361da177e4SLinus Torvalds bio->bi_io_vec[0].bv_page = bh->b_page; 29371da177e4SLinus Torvalds bio->bi_io_vec[0].bv_len = bh->b_size; 29381da177e4SLinus Torvalds bio->bi_io_vec[0].bv_offset = bh_offset(bh); 29391da177e4SLinus Torvalds 29401da177e4SLinus Torvalds bio->bi_vcnt = 1; 29411da177e4SLinus Torvalds bio->bi_idx = 0; 29421da177e4SLinus Torvalds bio->bi_size = bh->b_size; 29431da177e4SLinus Torvalds 29441da177e4SLinus Torvalds bio->bi_end_io = end_bio_bh_io_sync; 29451da177e4SLinus Torvalds bio->bi_private = bh; 29461da177e4SLinus Torvalds 29471da177e4SLinus Torvalds bio_get(bio); 29481da177e4SLinus Torvalds submit_bio(rw, bio); 29491da177e4SLinus Torvalds 29501da177e4SLinus Torvalds if (bio_flagged(bio, BIO_EOPNOTSUPP)) 29511da177e4SLinus Torvalds ret = -EOPNOTSUPP; 29521da177e4SLinus Torvalds 29531da177e4SLinus Torvalds bio_put(bio); 29541da177e4SLinus Torvalds return ret; 29551da177e4SLinus Torvalds } 29561fe72eaaSH Hartley Sweeten EXPORT_SYMBOL(submit_bh); 29571da177e4SLinus Torvalds 29581da177e4SLinus Torvalds /** 29591da177e4SLinus Torvalds * ll_rw_block: low-level access to block devices (DEPRECATED) 29609cb569d6SChristoph Hellwig * @rw: whether to %READ or %WRITE or maybe %READA (readahead) 29611da177e4SLinus Torvalds * @nr: number of &struct buffer_heads in the array 29621da177e4SLinus Torvalds * @bhs: array of pointers to &struct buffer_head 29631da177e4SLinus Torvalds * 2964a7662236SJan Kara * ll_rw_block() takes an array of pointers to &struct buffer_heads, and 2965a7662236SJan Kara * requests an I/O operation on them, either a %READ or a %WRITE. The third 29669cb569d6SChristoph Hellwig * %READA option is described in the documentation for generic_make_request() 29679cb569d6SChristoph Hellwig * which ll_rw_block() calls. 29681da177e4SLinus Torvalds * 29691da177e4SLinus Torvalds * This function drops any buffer that it cannot get a lock on (with the 29709cb569d6SChristoph Hellwig * BH_Lock state bit), any buffer that appears to be clean when doing a write 29719cb569d6SChristoph Hellwig * request, and any buffer that appears to be up-to-date when doing read 29729cb569d6SChristoph Hellwig * request. Further it marks as clean buffers that are processed for 29739cb569d6SChristoph Hellwig * writing (the buffer cache won't assume that they are actually clean 29749cb569d6SChristoph Hellwig * until the buffer gets unlocked). 29751da177e4SLinus Torvalds * 29761da177e4SLinus Torvalds * ll_rw_block sets b_end_io to simple completion handler that marks 29771da177e4SLinus Torvalds * the buffer up-to-date (if approriate), unlocks the buffer and wakes 29781da177e4SLinus Torvalds * any waiters. 29791da177e4SLinus Torvalds * 29801da177e4SLinus Torvalds * All of the buffers must be for the same device, and must also be a 29811da177e4SLinus Torvalds * multiple of the current approved size for the device. 29821da177e4SLinus Torvalds */ 29831da177e4SLinus Torvalds void ll_rw_block(int rw, int nr, struct buffer_head *bhs[]) 29841da177e4SLinus Torvalds { 29851da177e4SLinus Torvalds int i; 29861da177e4SLinus Torvalds 29871da177e4SLinus Torvalds for (i = 0; i < nr; i++) { 29881da177e4SLinus Torvalds struct buffer_head *bh = bhs[i]; 29891da177e4SLinus Torvalds 29909cb569d6SChristoph Hellwig if (!trylock_buffer(bh)) 29911da177e4SLinus Torvalds continue; 29929cb569d6SChristoph Hellwig if (rw == WRITE) { 29931da177e4SLinus Torvalds if (test_clear_buffer_dirty(bh)) { 299476c3073aSakpm@osdl.org bh->b_end_io = end_buffer_write_sync; 2995e60e5c50SOGAWA Hirofumi get_bh(bh); 29961da177e4SLinus Torvalds submit_bh(WRITE, bh); 29971da177e4SLinus Torvalds continue; 29981da177e4SLinus Torvalds } 29991da177e4SLinus Torvalds } else { 30001da177e4SLinus Torvalds if (!buffer_uptodate(bh)) { 300176c3073aSakpm@osdl.org bh->b_end_io = end_buffer_read_sync; 3002e60e5c50SOGAWA Hirofumi get_bh(bh); 30031da177e4SLinus Torvalds submit_bh(rw, bh); 30041da177e4SLinus Torvalds continue; 30051da177e4SLinus Torvalds } 30061da177e4SLinus Torvalds } 30071da177e4SLinus Torvalds unlock_buffer(bh); 30081da177e4SLinus Torvalds } 30091da177e4SLinus Torvalds } 30101fe72eaaSH Hartley Sweeten EXPORT_SYMBOL(ll_rw_block); 30111da177e4SLinus Torvalds 30129cb569d6SChristoph Hellwig void write_dirty_buffer(struct buffer_head *bh, int rw) 30139cb569d6SChristoph Hellwig { 30149cb569d6SChristoph Hellwig lock_buffer(bh); 30159cb569d6SChristoph Hellwig if (!test_clear_buffer_dirty(bh)) { 30169cb569d6SChristoph Hellwig unlock_buffer(bh); 30179cb569d6SChristoph Hellwig return; 30189cb569d6SChristoph Hellwig } 30199cb569d6SChristoph Hellwig bh->b_end_io = end_buffer_write_sync; 30209cb569d6SChristoph Hellwig get_bh(bh); 30219cb569d6SChristoph Hellwig submit_bh(rw, bh); 30229cb569d6SChristoph Hellwig } 30239cb569d6SChristoph Hellwig EXPORT_SYMBOL(write_dirty_buffer); 30249cb569d6SChristoph Hellwig 30251da177e4SLinus Torvalds /* 30261da177e4SLinus Torvalds * For a data-integrity writeout, we need to wait upon any in-progress I/O 30271da177e4SLinus Torvalds * and then start new I/O and then wait upon it. The caller must have a ref on 30281da177e4SLinus Torvalds * the buffer_head. 30291da177e4SLinus Torvalds */ 303087e99511SChristoph Hellwig int __sync_dirty_buffer(struct buffer_head *bh, int rw) 30311da177e4SLinus Torvalds { 30321da177e4SLinus Torvalds int ret = 0; 30331da177e4SLinus Torvalds 30341da177e4SLinus Torvalds WARN_ON(atomic_read(&bh->b_count) < 1); 30351da177e4SLinus Torvalds lock_buffer(bh); 30361da177e4SLinus Torvalds if (test_clear_buffer_dirty(bh)) { 30371da177e4SLinus Torvalds get_bh(bh); 30381da177e4SLinus Torvalds bh->b_end_io = end_buffer_write_sync; 303987e99511SChristoph Hellwig ret = submit_bh(rw, bh); 30401da177e4SLinus Torvalds wait_on_buffer(bh); 30411da177e4SLinus Torvalds if (!ret && !buffer_uptodate(bh)) 30421da177e4SLinus Torvalds ret = -EIO; 30431da177e4SLinus Torvalds } else { 30441da177e4SLinus Torvalds unlock_buffer(bh); 30451da177e4SLinus Torvalds } 30461da177e4SLinus Torvalds return ret; 30471da177e4SLinus Torvalds } 304887e99511SChristoph Hellwig EXPORT_SYMBOL(__sync_dirty_buffer); 304987e99511SChristoph Hellwig 305087e99511SChristoph Hellwig int sync_dirty_buffer(struct buffer_head *bh) 305187e99511SChristoph Hellwig { 305287e99511SChristoph Hellwig return __sync_dirty_buffer(bh, WRITE_SYNC); 305387e99511SChristoph Hellwig } 30541fe72eaaSH Hartley Sweeten EXPORT_SYMBOL(sync_dirty_buffer); 30551da177e4SLinus Torvalds 30561da177e4SLinus Torvalds /* 30571da177e4SLinus Torvalds * try_to_free_buffers() checks if all the buffers on this particular page 30581da177e4SLinus Torvalds * are unused, and releases them if so. 30591da177e4SLinus Torvalds * 30601da177e4SLinus Torvalds * Exclusion against try_to_free_buffers may be obtained by either 30611da177e4SLinus Torvalds * locking the page or by holding its mapping's private_lock. 30621da177e4SLinus Torvalds * 30631da177e4SLinus Torvalds * If the page is dirty but all the buffers are clean then we need to 30641da177e4SLinus Torvalds * be sure to mark the page clean as well. This is because the page 30651da177e4SLinus Torvalds * may be against a block device, and a later reattachment of buffers 30661da177e4SLinus Torvalds * to a dirty page will set *all* buffers dirty. Which would corrupt 30671da177e4SLinus Torvalds * filesystem data on the same device. 30681da177e4SLinus Torvalds * 30691da177e4SLinus Torvalds * The same applies to regular filesystem pages: if all the buffers are 30701da177e4SLinus Torvalds * clean then we set the page clean and proceed. To do that, we require 30711da177e4SLinus Torvalds * total exclusion from __set_page_dirty_buffers(). That is obtained with 30721da177e4SLinus Torvalds * private_lock. 30731da177e4SLinus Torvalds * 30741da177e4SLinus Torvalds * try_to_free_buffers() is non-blocking. 30751da177e4SLinus Torvalds */ 30761da177e4SLinus Torvalds static inline int buffer_busy(struct buffer_head *bh) 30771da177e4SLinus Torvalds { 30781da177e4SLinus Torvalds return atomic_read(&bh->b_count) | 30791da177e4SLinus Torvalds (bh->b_state & ((1 << BH_Dirty) | (1 << BH_Lock))); 30801da177e4SLinus Torvalds } 30811da177e4SLinus Torvalds 30821da177e4SLinus Torvalds static int 30831da177e4SLinus Torvalds drop_buffers(struct page *page, struct buffer_head **buffers_to_free) 30841da177e4SLinus Torvalds { 30851da177e4SLinus Torvalds struct buffer_head *head = page_buffers(page); 30861da177e4SLinus Torvalds struct buffer_head *bh; 30871da177e4SLinus Torvalds 30881da177e4SLinus Torvalds bh = head; 30891da177e4SLinus Torvalds do { 3090de7d5a3bSakpm@osdl.org if (buffer_write_io_error(bh) && page->mapping) 30911da177e4SLinus Torvalds set_bit(AS_EIO, &page->mapping->flags); 30921da177e4SLinus Torvalds if (buffer_busy(bh)) 30931da177e4SLinus Torvalds goto failed; 30941da177e4SLinus Torvalds bh = bh->b_this_page; 30951da177e4SLinus Torvalds } while (bh != head); 30961da177e4SLinus Torvalds 30971da177e4SLinus Torvalds do { 30981da177e4SLinus Torvalds struct buffer_head *next = bh->b_this_page; 30991da177e4SLinus Torvalds 3100535ee2fbSJan Kara if (bh->b_assoc_map) 31011da177e4SLinus Torvalds __remove_assoc_queue(bh); 31021da177e4SLinus Torvalds bh = next; 31031da177e4SLinus Torvalds } while (bh != head); 31041da177e4SLinus Torvalds *buffers_to_free = head; 31051da177e4SLinus Torvalds __clear_page_buffers(page); 31061da177e4SLinus Torvalds return 1; 31071da177e4SLinus Torvalds failed: 31081da177e4SLinus Torvalds return 0; 31091da177e4SLinus Torvalds } 31101da177e4SLinus Torvalds 31111da177e4SLinus Torvalds int try_to_free_buffers(struct page *page) 31121da177e4SLinus Torvalds { 31131da177e4SLinus Torvalds struct address_space * const mapping = page->mapping; 31141da177e4SLinus Torvalds struct buffer_head *buffers_to_free = NULL; 31151da177e4SLinus Torvalds int ret = 0; 31161da177e4SLinus Torvalds 31171da177e4SLinus Torvalds BUG_ON(!PageLocked(page)); 3118ecdfc978SLinus Torvalds if (PageWriteback(page)) 31191da177e4SLinus Torvalds return 0; 31201da177e4SLinus Torvalds 31211da177e4SLinus Torvalds if (mapping == NULL) { /* can this still happen? */ 31221da177e4SLinus Torvalds ret = drop_buffers(page, &buffers_to_free); 31231da177e4SLinus Torvalds goto out; 31241da177e4SLinus Torvalds } 31251da177e4SLinus Torvalds 31261da177e4SLinus Torvalds spin_lock(&mapping->private_lock); 31271da177e4SLinus Torvalds ret = drop_buffers(page, &buffers_to_free); 3128ecdfc978SLinus Torvalds 3129ecdfc978SLinus Torvalds /* 3130ecdfc978SLinus Torvalds * If the filesystem writes its buffers by hand (eg ext3) 3131ecdfc978SLinus Torvalds * then we can have clean buffers against a dirty page. We 3132ecdfc978SLinus Torvalds * clean the page here; otherwise the VM will never notice 3133ecdfc978SLinus Torvalds * that the filesystem did any IO at all. 3134ecdfc978SLinus Torvalds * 3135ecdfc978SLinus Torvalds * Also, during truncate, discard_buffer will have marked all 3136ecdfc978SLinus Torvalds * the page's buffers clean. We discover that here and clean 3137ecdfc978SLinus Torvalds * the page also. 313887df7241SNick Piggin * 313987df7241SNick Piggin * private_lock must be held over this entire operation in order 314087df7241SNick Piggin * to synchronise against __set_page_dirty_buffers and prevent the 314187df7241SNick Piggin * dirty bit from being lost. 3142ecdfc978SLinus Torvalds */ 3143ecdfc978SLinus Torvalds if (ret) 3144ecdfc978SLinus Torvalds cancel_dirty_page(page, PAGE_CACHE_SIZE); 314587df7241SNick Piggin spin_unlock(&mapping->private_lock); 31461da177e4SLinus Torvalds out: 31471da177e4SLinus Torvalds if (buffers_to_free) { 31481da177e4SLinus Torvalds struct buffer_head *bh = buffers_to_free; 31491da177e4SLinus Torvalds 31501da177e4SLinus Torvalds do { 31511da177e4SLinus Torvalds struct buffer_head *next = bh->b_this_page; 31521da177e4SLinus Torvalds free_buffer_head(bh); 31531da177e4SLinus Torvalds bh = next; 31541da177e4SLinus Torvalds } while (bh != buffers_to_free); 31551da177e4SLinus Torvalds } 31561da177e4SLinus Torvalds return ret; 31571da177e4SLinus Torvalds } 31581da177e4SLinus Torvalds EXPORT_SYMBOL(try_to_free_buffers); 31591da177e4SLinus Torvalds 31601da177e4SLinus Torvalds /* 31611da177e4SLinus Torvalds * There are no bdflush tunables left. But distributions are 31621da177e4SLinus Torvalds * still running obsolete flush daemons, so we terminate them here. 31631da177e4SLinus Torvalds * 31641da177e4SLinus Torvalds * Use of bdflush() is deprecated and will be removed in a future kernel. 31655b0830cbSJens Axboe * The `flush-X' kernel threads fully replace bdflush daemons and this call. 31661da177e4SLinus Torvalds */ 3167bdc480e3SHeiko Carstens SYSCALL_DEFINE2(bdflush, int, func, long, data) 31681da177e4SLinus Torvalds { 31691da177e4SLinus Torvalds static int msg_count; 31701da177e4SLinus Torvalds 31711da177e4SLinus Torvalds if (!capable(CAP_SYS_ADMIN)) 31721da177e4SLinus Torvalds return -EPERM; 31731da177e4SLinus Torvalds 31741da177e4SLinus Torvalds if (msg_count < 5) { 31751da177e4SLinus Torvalds msg_count++; 31761da177e4SLinus Torvalds printk(KERN_INFO 31771da177e4SLinus Torvalds "warning: process `%s' used the obsolete bdflush" 31781da177e4SLinus Torvalds " system call\n", current->comm); 31791da177e4SLinus Torvalds printk(KERN_INFO "Fix your initscripts?\n"); 31801da177e4SLinus Torvalds } 31811da177e4SLinus Torvalds 31821da177e4SLinus Torvalds if (func == 1) 31831da177e4SLinus Torvalds do_exit(0); 31841da177e4SLinus Torvalds return 0; 31851da177e4SLinus Torvalds } 31861da177e4SLinus Torvalds 31871da177e4SLinus Torvalds /* 31881da177e4SLinus Torvalds * Buffer-head allocation 31891da177e4SLinus Torvalds */ 3190e18b890bSChristoph Lameter static struct kmem_cache *bh_cachep; 31911da177e4SLinus Torvalds 31921da177e4SLinus Torvalds /* 31931da177e4SLinus Torvalds * Once the number of bh's in the machine exceeds this level, we start 31941da177e4SLinus Torvalds * stripping them in writeback. 31951da177e4SLinus Torvalds */ 31961da177e4SLinus Torvalds static int max_buffer_heads; 31971da177e4SLinus Torvalds 31981da177e4SLinus Torvalds int buffer_heads_over_limit; 31991da177e4SLinus Torvalds 32001da177e4SLinus Torvalds struct bh_accounting { 32011da177e4SLinus Torvalds int nr; /* Number of live bh's */ 32021da177e4SLinus Torvalds int ratelimit; /* Limit cacheline bouncing */ 32031da177e4SLinus Torvalds }; 32041da177e4SLinus Torvalds 32051da177e4SLinus Torvalds static DEFINE_PER_CPU(struct bh_accounting, bh_accounting) = {0, 0}; 32061da177e4SLinus Torvalds 32071da177e4SLinus Torvalds static void recalc_bh_state(void) 32081da177e4SLinus Torvalds { 32091da177e4SLinus Torvalds int i; 32101da177e4SLinus Torvalds int tot = 0; 32111da177e4SLinus Torvalds 3212ee1be862SChristoph Lameter if (__this_cpu_inc_return(bh_accounting.ratelimit) - 1 < 4096) 32131da177e4SLinus Torvalds return; 3214c7b92516SChristoph Lameter __this_cpu_write(bh_accounting.ratelimit, 0); 32158a143426SEric Dumazet for_each_online_cpu(i) 32161da177e4SLinus Torvalds tot += per_cpu(bh_accounting, i).nr; 32171da177e4SLinus Torvalds buffer_heads_over_limit = (tot > max_buffer_heads); 32181da177e4SLinus Torvalds } 32191da177e4SLinus Torvalds 3220dd0fc66fSAl Viro struct buffer_head *alloc_buffer_head(gfp_t gfp_flags) 32211da177e4SLinus Torvalds { 3222019b4d12SRichard Kennedy struct buffer_head *ret = kmem_cache_zalloc(bh_cachep, gfp_flags); 32231da177e4SLinus Torvalds if (ret) { 3224a35afb83SChristoph Lameter INIT_LIST_HEAD(&ret->b_assoc_buffers); 3225c7b92516SChristoph Lameter preempt_disable(); 3226c7b92516SChristoph Lameter __this_cpu_inc(bh_accounting.nr); 32271da177e4SLinus Torvalds recalc_bh_state(); 3228c7b92516SChristoph Lameter preempt_enable(); 32291da177e4SLinus Torvalds } 32301da177e4SLinus Torvalds return ret; 32311da177e4SLinus Torvalds } 32321da177e4SLinus Torvalds EXPORT_SYMBOL(alloc_buffer_head); 32331da177e4SLinus Torvalds 32341da177e4SLinus Torvalds void free_buffer_head(struct buffer_head *bh) 32351da177e4SLinus Torvalds { 32361da177e4SLinus Torvalds BUG_ON(!list_empty(&bh->b_assoc_buffers)); 32371da177e4SLinus Torvalds kmem_cache_free(bh_cachep, bh); 3238c7b92516SChristoph Lameter preempt_disable(); 3239c7b92516SChristoph Lameter __this_cpu_dec(bh_accounting.nr); 32401da177e4SLinus Torvalds recalc_bh_state(); 3241c7b92516SChristoph Lameter preempt_enable(); 32421da177e4SLinus Torvalds } 32431da177e4SLinus Torvalds EXPORT_SYMBOL(free_buffer_head); 32441da177e4SLinus Torvalds 32451da177e4SLinus Torvalds static void buffer_exit_cpu(int cpu) 32461da177e4SLinus Torvalds { 32471da177e4SLinus Torvalds int i; 32481da177e4SLinus Torvalds struct bh_lru *b = &per_cpu(bh_lrus, cpu); 32491da177e4SLinus Torvalds 32501da177e4SLinus Torvalds for (i = 0; i < BH_LRU_SIZE; i++) { 32511da177e4SLinus Torvalds brelse(b->bhs[i]); 32521da177e4SLinus Torvalds b->bhs[i] = NULL; 32531da177e4SLinus Torvalds } 3254c7b92516SChristoph Lameter this_cpu_add(bh_accounting.nr, per_cpu(bh_accounting, cpu).nr); 32558a143426SEric Dumazet per_cpu(bh_accounting, cpu).nr = 0; 32561da177e4SLinus Torvalds } 32571da177e4SLinus Torvalds 32581da177e4SLinus Torvalds static int buffer_cpu_notify(struct notifier_block *self, 32591da177e4SLinus Torvalds unsigned long action, void *hcpu) 32601da177e4SLinus Torvalds { 32618bb78442SRafael J. Wysocki if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) 32621da177e4SLinus Torvalds buffer_exit_cpu((unsigned long)hcpu); 32631da177e4SLinus Torvalds return NOTIFY_OK; 32641da177e4SLinus Torvalds } 32651da177e4SLinus Torvalds 3266389d1b08SAneesh Kumar K.V /** 3267a6b91919SRandy Dunlap * bh_uptodate_or_lock - Test whether the buffer is uptodate 3268389d1b08SAneesh Kumar K.V * @bh: struct buffer_head 3269389d1b08SAneesh Kumar K.V * 3270389d1b08SAneesh Kumar K.V * Return true if the buffer is up-to-date and false, 3271389d1b08SAneesh Kumar K.V * with the buffer locked, if not. 3272389d1b08SAneesh Kumar K.V */ 3273389d1b08SAneesh Kumar K.V int bh_uptodate_or_lock(struct buffer_head *bh) 3274389d1b08SAneesh Kumar K.V { 3275389d1b08SAneesh Kumar K.V if (!buffer_uptodate(bh)) { 3276389d1b08SAneesh Kumar K.V lock_buffer(bh); 3277389d1b08SAneesh Kumar K.V if (!buffer_uptodate(bh)) 3278389d1b08SAneesh Kumar K.V return 0; 3279389d1b08SAneesh Kumar K.V unlock_buffer(bh); 3280389d1b08SAneesh Kumar K.V } 3281389d1b08SAneesh Kumar K.V return 1; 3282389d1b08SAneesh Kumar K.V } 3283389d1b08SAneesh Kumar K.V EXPORT_SYMBOL(bh_uptodate_or_lock); 3284389d1b08SAneesh Kumar K.V 3285389d1b08SAneesh Kumar K.V /** 3286a6b91919SRandy Dunlap * bh_submit_read - Submit a locked buffer for reading 3287389d1b08SAneesh Kumar K.V * @bh: struct buffer_head 3288389d1b08SAneesh Kumar K.V * 3289389d1b08SAneesh Kumar K.V * Returns zero on success and -EIO on error. 3290389d1b08SAneesh Kumar K.V */ 3291389d1b08SAneesh Kumar K.V int bh_submit_read(struct buffer_head *bh) 3292389d1b08SAneesh Kumar K.V { 3293389d1b08SAneesh Kumar K.V BUG_ON(!buffer_locked(bh)); 3294389d1b08SAneesh Kumar K.V 3295389d1b08SAneesh Kumar K.V if (buffer_uptodate(bh)) { 3296389d1b08SAneesh Kumar K.V unlock_buffer(bh); 3297389d1b08SAneesh Kumar K.V return 0; 3298389d1b08SAneesh Kumar K.V } 3299389d1b08SAneesh Kumar K.V 3300389d1b08SAneesh Kumar K.V get_bh(bh); 3301389d1b08SAneesh Kumar K.V bh->b_end_io = end_buffer_read_sync; 3302389d1b08SAneesh Kumar K.V submit_bh(READ, bh); 3303389d1b08SAneesh Kumar K.V wait_on_buffer(bh); 3304389d1b08SAneesh Kumar K.V if (buffer_uptodate(bh)) 3305389d1b08SAneesh Kumar K.V return 0; 3306389d1b08SAneesh Kumar K.V return -EIO; 3307389d1b08SAneesh Kumar K.V } 3308389d1b08SAneesh Kumar K.V EXPORT_SYMBOL(bh_submit_read); 3309389d1b08SAneesh Kumar K.V 33101da177e4SLinus Torvalds void __init buffer_init(void) 33111da177e4SLinus Torvalds { 33121da177e4SLinus Torvalds int nrpages; 33131da177e4SLinus Torvalds 3314b98938c3SChristoph Lameter bh_cachep = kmem_cache_create("buffer_head", 3315b98938c3SChristoph Lameter sizeof(struct buffer_head), 0, 3316b98938c3SChristoph Lameter (SLAB_RECLAIM_ACCOUNT|SLAB_PANIC| 3317b98938c3SChristoph Lameter SLAB_MEM_SPREAD), 3318019b4d12SRichard Kennedy NULL); 33191da177e4SLinus Torvalds 33201da177e4SLinus Torvalds /* 33211da177e4SLinus Torvalds * Limit the bh occupancy to 10% of ZONE_NORMAL 33221da177e4SLinus Torvalds */ 33231da177e4SLinus Torvalds nrpages = (nr_free_buffer_pages() * 10) / 100; 33241da177e4SLinus Torvalds max_buffer_heads = nrpages * (PAGE_SIZE / sizeof(struct buffer_head)); 33251da177e4SLinus Torvalds hotcpu_notifier(buffer_cpu_notify, 0); 33261da177e4SLinus Torvalds } 3327