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> 441da177e4SLinus Torvalds 451da177e4SLinus Torvalds static int fsync_buffers_list(spinlock_t *lock, struct list_head *list); 461da177e4SLinus Torvalds 471da177e4SLinus Torvalds #define BH_ENTRY(list) list_entry((list), struct buffer_head, b_assoc_buffers) 481da177e4SLinus Torvalds 491da177e4SLinus Torvalds inline void 501da177e4SLinus Torvalds init_buffer(struct buffer_head *bh, bh_end_io_t *handler, void *private) 511da177e4SLinus Torvalds { 521da177e4SLinus Torvalds bh->b_end_io = handler; 531da177e4SLinus Torvalds bh->b_private = private; 541da177e4SLinus Torvalds } 551da177e4SLinus Torvalds 561da177e4SLinus Torvalds static int sync_buffer(void *word) 571da177e4SLinus Torvalds { 581da177e4SLinus Torvalds struct block_device *bd; 591da177e4SLinus Torvalds struct buffer_head *bh 601da177e4SLinus Torvalds = container_of(word, struct buffer_head, b_state); 611da177e4SLinus Torvalds 621da177e4SLinus Torvalds smp_mb(); 631da177e4SLinus Torvalds bd = bh->b_bdev; 641da177e4SLinus Torvalds if (bd) 651da177e4SLinus Torvalds blk_run_address_space(bd->bd_inode->i_mapping); 661da177e4SLinus Torvalds io_schedule(); 671da177e4SLinus Torvalds return 0; 681da177e4SLinus Torvalds } 691da177e4SLinus Torvalds 70fc9b52cdSHarvey Harrison void __lock_buffer(struct buffer_head *bh) 711da177e4SLinus Torvalds { 721da177e4SLinus Torvalds wait_on_bit_lock(&bh->b_state, BH_Lock, sync_buffer, 731da177e4SLinus Torvalds TASK_UNINTERRUPTIBLE); 741da177e4SLinus Torvalds } 751da177e4SLinus Torvalds EXPORT_SYMBOL(__lock_buffer); 761da177e4SLinus Torvalds 77fc9b52cdSHarvey Harrison void unlock_buffer(struct buffer_head *bh) 781da177e4SLinus Torvalds { 7951b07fc3SNick Piggin clear_bit_unlock(BH_Lock, &bh->b_state); 801da177e4SLinus Torvalds smp_mb__after_clear_bit(); 811da177e4SLinus Torvalds wake_up_bit(&bh->b_state, BH_Lock); 821da177e4SLinus Torvalds } 831da177e4SLinus Torvalds 841da177e4SLinus Torvalds /* 851da177e4SLinus Torvalds * Block until a buffer comes unlocked. This doesn't stop it 861da177e4SLinus Torvalds * from becoming locked again - you have to lock it yourself 871da177e4SLinus Torvalds * if you want to preserve its state. 881da177e4SLinus Torvalds */ 891da177e4SLinus Torvalds void __wait_on_buffer(struct buffer_head * bh) 901da177e4SLinus Torvalds { 911da177e4SLinus Torvalds wait_on_bit(&bh->b_state, BH_Lock, sync_buffer, TASK_UNINTERRUPTIBLE); 921da177e4SLinus Torvalds } 931da177e4SLinus Torvalds 941da177e4SLinus Torvalds static void 951da177e4SLinus Torvalds __clear_page_buffers(struct page *page) 961da177e4SLinus Torvalds { 971da177e4SLinus Torvalds ClearPagePrivate(page); 984c21e2f2SHugh Dickins set_page_private(page, 0); 991da177e4SLinus Torvalds page_cache_release(page); 1001da177e4SLinus Torvalds } 1011da177e4SLinus Torvalds 10208bafc03SKeith Mannthey 10308bafc03SKeith Mannthey static int quiet_error(struct buffer_head *bh) 10408bafc03SKeith Mannthey { 10508bafc03SKeith Mannthey if (!test_bit(BH_Quiet, &bh->b_state) && printk_ratelimit()) 10608bafc03SKeith Mannthey return 0; 10708bafc03SKeith Mannthey return 1; 10808bafc03SKeith Mannthey } 10908bafc03SKeith Mannthey 11008bafc03SKeith Mannthey 1111da177e4SLinus Torvalds static void buffer_io_error(struct buffer_head *bh) 1121da177e4SLinus Torvalds { 1131da177e4SLinus Torvalds char b[BDEVNAME_SIZE]; 1141da177e4SLinus Torvalds printk(KERN_ERR "Buffer I/O error on device %s, logical block %Lu\n", 1151da177e4SLinus Torvalds bdevname(bh->b_bdev, b), 1161da177e4SLinus Torvalds (unsigned long long)bh->b_blocknr); 1171da177e4SLinus Torvalds } 1181da177e4SLinus Torvalds 1191da177e4SLinus Torvalds /* 12068671f35SDmitry Monakhov * End-of-IO handler helper function which does not touch the bh after 12168671f35SDmitry Monakhov * unlocking it. 12268671f35SDmitry Monakhov * Note: unlock_buffer() sort-of does touch the bh after unlocking it, but 12368671f35SDmitry Monakhov * a race there is benign: unlock_buffer() only use the bh's address for 12468671f35SDmitry Monakhov * hashing after unlocking the buffer, so it doesn't actually touch the bh 12568671f35SDmitry Monakhov * itself. 1261da177e4SLinus Torvalds */ 12768671f35SDmitry Monakhov static void __end_buffer_read_notouch(struct buffer_head *bh, int uptodate) 1281da177e4SLinus Torvalds { 1291da177e4SLinus Torvalds if (uptodate) { 1301da177e4SLinus Torvalds set_buffer_uptodate(bh); 1311da177e4SLinus Torvalds } else { 1321da177e4SLinus Torvalds /* This happens, due to failed READA attempts. */ 1331da177e4SLinus Torvalds clear_buffer_uptodate(bh); 1341da177e4SLinus Torvalds } 1351da177e4SLinus Torvalds unlock_buffer(bh); 13668671f35SDmitry Monakhov } 13768671f35SDmitry Monakhov 13868671f35SDmitry Monakhov /* 13968671f35SDmitry Monakhov * Default synchronous end-of-IO handler.. Just mark it up-to-date and 14068671f35SDmitry Monakhov * unlock the buffer. This is what ll_rw_block uses too. 14168671f35SDmitry Monakhov */ 14268671f35SDmitry Monakhov void end_buffer_read_sync(struct buffer_head *bh, int uptodate) 14368671f35SDmitry Monakhov { 14468671f35SDmitry Monakhov __end_buffer_read_notouch(bh, uptodate); 1451da177e4SLinus Torvalds put_bh(bh); 1461da177e4SLinus Torvalds } 1471da177e4SLinus Torvalds 1481da177e4SLinus Torvalds void end_buffer_write_sync(struct buffer_head *bh, int uptodate) 1491da177e4SLinus Torvalds { 1501da177e4SLinus Torvalds char b[BDEVNAME_SIZE]; 1511da177e4SLinus Torvalds 1521da177e4SLinus Torvalds if (uptodate) { 1531da177e4SLinus Torvalds set_buffer_uptodate(bh); 1541da177e4SLinus Torvalds } else { 15508bafc03SKeith Mannthey if (!buffer_eopnotsupp(bh) && !quiet_error(bh)) { 1561da177e4SLinus Torvalds buffer_io_error(bh); 1571da177e4SLinus Torvalds printk(KERN_WARNING "lost page write due to " 1581da177e4SLinus Torvalds "I/O error on %s\n", 1591da177e4SLinus Torvalds bdevname(bh->b_bdev, b)); 1601da177e4SLinus Torvalds } 1611da177e4SLinus Torvalds set_buffer_write_io_error(bh); 1621da177e4SLinus Torvalds clear_buffer_uptodate(bh); 1631da177e4SLinus Torvalds } 1641da177e4SLinus Torvalds unlock_buffer(bh); 1651da177e4SLinus Torvalds put_bh(bh); 1661da177e4SLinus Torvalds } 1671da177e4SLinus Torvalds 1681da177e4SLinus Torvalds /* 1691da177e4SLinus Torvalds * Various filesystems appear to want __find_get_block to be non-blocking. 1701da177e4SLinus Torvalds * But it's the page lock which protects the buffers. To get around this, 1711da177e4SLinus Torvalds * we get exclusion from try_to_free_buffers with the blockdev mapping's 1721da177e4SLinus Torvalds * private_lock. 1731da177e4SLinus Torvalds * 1741da177e4SLinus Torvalds * Hack idea: for the blockdev mapping, i_bufferlist_lock contention 1751da177e4SLinus Torvalds * may be quite high. This code could TryLock the page, and if that 1761da177e4SLinus Torvalds * succeeds, there is no need to take private_lock. (But if 1771da177e4SLinus Torvalds * private_lock is contended then so is mapping->tree_lock). 1781da177e4SLinus Torvalds */ 1791da177e4SLinus Torvalds static struct buffer_head * 180385fd4c5SCoywolf Qi Hunt __find_get_block_slow(struct block_device *bdev, sector_t block) 1811da177e4SLinus Torvalds { 1821da177e4SLinus Torvalds struct inode *bd_inode = bdev->bd_inode; 1831da177e4SLinus Torvalds struct address_space *bd_mapping = bd_inode->i_mapping; 1841da177e4SLinus Torvalds struct buffer_head *ret = NULL; 1851da177e4SLinus Torvalds pgoff_t index; 1861da177e4SLinus Torvalds struct buffer_head *bh; 1871da177e4SLinus Torvalds struct buffer_head *head; 1881da177e4SLinus Torvalds struct page *page; 1891da177e4SLinus Torvalds int all_mapped = 1; 1901da177e4SLinus Torvalds 1911da177e4SLinus Torvalds index = block >> (PAGE_CACHE_SHIFT - bd_inode->i_blkbits); 1921da177e4SLinus Torvalds page = find_get_page(bd_mapping, index); 1931da177e4SLinus Torvalds if (!page) 1941da177e4SLinus Torvalds goto out; 1951da177e4SLinus Torvalds 1961da177e4SLinus Torvalds spin_lock(&bd_mapping->private_lock); 1971da177e4SLinus Torvalds if (!page_has_buffers(page)) 1981da177e4SLinus Torvalds goto out_unlock; 1991da177e4SLinus Torvalds head = page_buffers(page); 2001da177e4SLinus Torvalds bh = head; 2011da177e4SLinus Torvalds do { 20297f76d3dSNikanth Karthikesan if (!buffer_mapped(bh)) 20397f76d3dSNikanth Karthikesan all_mapped = 0; 20497f76d3dSNikanth Karthikesan else if (bh->b_blocknr == block) { 2051da177e4SLinus Torvalds ret = bh; 2061da177e4SLinus Torvalds get_bh(bh); 2071da177e4SLinus Torvalds goto out_unlock; 2081da177e4SLinus Torvalds } 2091da177e4SLinus Torvalds bh = bh->b_this_page; 2101da177e4SLinus Torvalds } while (bh != head); 2111da177e4SLinus Torvalds 2121da177e4SLinus Torvalds /* we might be here because some of the buffers on this page are 2131da177e4SLinus Torvalds * not mapped. This is due to various races between 2141da177e4SLinus Torvalds * file io on the block device and getblk. It gets dealt with 2151da177e4SLinus Torvalds * elsewhere, don't buffer_error if we had some unmapped buffers 2161da177e4SLinus Torvalds */ 2171da177e4SLinus Torvalds if (all_mapped) { 2181da177e4SLinus Torvalds printk("__find_get_block_slow() failed. " 2191da177e4SLinus Torvalds "block=%llu, b_blocknr=%llu\n", 220205f87f6SBadari Pulavarty (unsigned long long)block, 221205f87f6SBadari Pulavarty (unsigned long long)bh->b_blocknr); 222205f87f6SBadari Pulavarty printk("b_state=0x%08lx, b_size=%zu\n", 223205f87f6SBadari Pulavarty bh->b_state, bh->b_size); 2241da177e4SLinus Torvalds printk("device blocksize: %d\n", 1 << bd_inode->i_blkbits); 2251da177e4SLinus Torvalds } 2261da177e4SLinus Torvalds out_unlock: 2271da177e4SLinus Torvalds spin_unlock(&bd_mapping->private_lock); 2281da177e4SLinus Torvalds page_cache_release(page); 2291da177e4SLinus Torvalds out: 2301da177e4SLinus Torvalds return ret; 2311da177e4SLinus Torvalds } 2321da177e4SLinus Torvalds 2331da177e4SLinus Torvalds /* If invalidate_buffers() will trash dirty buffers, it means some kind 2341da177e4SLinus Torvalds of fs corruption is going on. Trashing dirty data always imply losing 2351da177e4SLinus Torvalds information that was supposed to be just stored on the physical layer 2361da177e4SLinus Torvalds by the user. 2371da177e4SLinus Torvalds 2381da177e4SLinus Torvalds Thus invalidate_buffers in general usage is not allwowed to trash 2391da177e4SLinus Torvalds dirty buffers. For example ioctl(FLSBLKBUF) expects dirty data to 2401da177e4SLinus Torvalds be preserved. These buffers are simply skipped. 2411da177e4SLinus Torvalds 2421da177e4SLinus Torvalds We also skip buffers which are still in use. For example this can 2431da177e4SLinus Torvalds happen if a userspace program is reading the block device. 2441da177e4SLinus Torvalds 2451da177e4SLinus Torvalds NOTE: In the case where the user removed a removable-media-disk even if 2461da177e4SLinus Torvalds there's still dirty data not synced on disk (due a bug in the device driver 2471da177e4SLinus Torvalds or due an error of the user), by not destroying the dirty buffers we could 2481da177e4SLinus Torvalds generate corruption also on the next media inserted, thus a parameter is 2491da177e4SLinus Torvalds necessary to handle this case in the most safe way possible (trying 2501da177e4SLinus Torvalds to not corrupt also the new disk inserted with the data belonging to 2511da177e4SLinus Torvalds the old now corrupted disk). Also for the ramdisk the natural thing 2521da177e4SLinus Torvalds to do in order to release the ramdisk memory is to destroy dirty buffers. 2531da177e4SLinus Torvalds 2541da177e4SLinus Torvalds These are two special cases. Normal usage imply the device driver 2551da177e4SLinus Torvalds to issue a sync on the device (without waiting I/O completion) and 2561da177e4SLinus Torvalds then an invalidate_buffers call that doesn't trash dirty buffers. 2571da177e4SLinus Torvalds 2581da177e4SLinus Torvalds For handling cache coherency with the blkdev pagecache the 'update' case 2591da177e4SLinus Torvalds is been introduced. It is needed to re-read from disk any pinned 2601da177e4SLinus Torvalds buffer. NOTE: re-reading from disk is destructive so we can do it only 2611da177e4SLinus Torvalds when we assume nobody is changing the buffercache under our I/O and when 2621da177e4SLinus Torvalds we think the disk contains more recent information than the buffercache. 2631da177e4SLinus Torvalds The update == 1 pass marks the buffers we need to update, the update == 2 2641da177e4SLinus Torvalds pass does the actual I/O. */ 265f98393a6SPeter Zijlstra void invalidate_bdev(struct block_device *bdev) 2661da177e4SLinus Torvalds { 2670e1dfc66SAndrew Morton struct address_space *mapping = bdev->bd_inode->i_mapping; 2680e1dfc66SAndrew Morton 2690e1dfc66SAndrew Morton if (mapping->nrpages == 0) 2700e1dfc66SAndrew Morton return; 2710e1dfc66SAndrew Morton 2721da177e4SLinus Torvalds invalidate_bh_lrus(); 273fc0ecff6SAndrew Morton invalidate_mapping_pages(mapping, 0, -1); 2741da177e4SLinus Torvalds } 2751da177e4SLinus Torvalds 2761da177e4SLinus Torvalds /* 2771da177e4SLinus Torvalds * Kick pdflush then try to free up some ZONE_NORMAL memory. 2781da177e4SLinus Torvalds */ 2791da177e4SLinus Torvalds static void free_more_memory(void) 2801da177e4SLinus Torvalds { 28119770b32SMel Gorman struct zone *zone; 2820e88460dSMel Gorman int nid; 2831da177e4SLinus Torvalds 284*03ba3782SJens Axboe wakeup_flusher_threads(1024); 2851da177e4SLinus Torvalds yield(); 2861da177e4SLinus Torvalds 2870e88460dSMel Gorman for_each_online_node(nid) { 28819770b32SMel Gorman (void)first_zones_zonelist(node_zonelist(nid, GFP_NOFS), 28919770b32SMel Gorman gfp_zone(GFP_NOFS), NULL, 29019770b32SMel Gorman &zone); 29119770b32SMel Gorman if (zone) 29254a6eb5cSMel Gorman try_to_free_pages(node_zonelist(nid, GFP_NOFS), 0, 293327c0e96SKAMEZAWA Hiroyuki GFP_NOFS, NULL); 2941da177e4SLinus Torvalds } 2951da177e4SLinus Torvalds } 2961da177e4SLinus Torvalds 2971da177e4SLinus Torvalds /* 2981da177e4SLinus Torvalds * I/O completion handler for block_read_full_page() - pages 2991da177e4SLinus Torvalds * which come unlocked at the end of I/O. 3001da177e4SLinus Torvalds */ 3011da177e4SLinus Torvalds static void end_buffer_async_read(struct buffer_head *bh, int uptodate) 3021da177e4SLinus Torvalds { 3031da177e4SLinus Torvalds unsigned long flags; 304a3972203SNick Piggin struct buffer_head *first; 3051da177e4SLinus Torvalds struct buffer_head *tmp; 3061da177e4SLinus Torvalds struct page *page; 3071da177e4SLinus Torvalds int page_uptodate = 1; 3081da177e4SLinus Torvalds 3091da177e4SLinus Torvalds BUG_ON(!buffer_async_read(bh)); 3101da177e4SLinus Torvalds 3111da177e4SLinus Torvalds page = bh->b_page; 3121da177e4SLinus Torvalds if (uptodate) { 3131da177e4SLinus Torvalds set_buffer_uptodate(bh); 3141da177e4SLinus Torvalds } else { 3151da177e4SLinus Torvalds clear_buffer_uptodate(bh); 31608bafc03SKeith Mannthey if (!quiet_error(bh)) 3171da177e4SLinus Torvalds buffer_io_error(bh); 3181da177e4SLinus Torvalds SetPageError(page); 3191da177e4SLinus Torvalds } 3201da177e4SLinus Torvalds 3211da177e4SLinus Torvalds /* 3221da177e4SLinus Torvalds * Be _very_ careful from here on. Bad things can happen if 3231da177e4SLinus Torvalds * two buffer heads end IO at almost the same time and both 3241da177e4SLinus Torvalds * decide that the page is now completely done. 3251da177e4SLinus Torvalds */ 326a3972203SNick Piggin first = page_buffers(page); 327a3972203SNick Piggin local_irq_save(flags); 328a3972203SNick Piggin bit_spin_lock(BH_Uptodate_Lock, &first->b_state); 3291da177e4SLinus Torvalds clear_buffer_async_read(bh); 3301da177e4SLinus Torvalds unlock_buffer(bh); 3311da177e4SLinus Torvalds tmp = bh; 3321da177e4SLinus Torvalds do { 3331da177e4SLinus Torvalds if (!buffer_uptodate(tmp)) 3341da177e4SLinus Torvalds page_uptodate = 0; 3351da177e4SLinus Torvalds if (buffer_async_read(tmp)) { 3361da177e4SLinus Torvalds BUG_ON(!buffer_locked(tmp)); 3371da177e4SLinus Torvalds goto still_busy; 3381da177e4SLinus Torvalds } 3391da177e4SLinus Torvalds tmp = tmp->b_this_page; 3401da177e4SLinus Torvalds } while (tmp != bh); 341a3972203SNick Piggin bit_spin_unlock(BH_Uptodate_Lock, &first->b_state); 342a3972203SNick Piggin local_irq_restore(flags); 3431da177e4SLinus Torvalds 3441da177e4SLinus Torvalds /* 3451da177e4SLinus Torvalds * If none of the buffers had errors and they are all 3461da177e4SLinus Torvalds * uptodate then we can set the page uptodate. 3471da177e4SLinus Torvalds */ 3481da177e4SLinus Torvalds if (page_uptodate && !PageError(page)) 3491da177e4SLinus Torvalds SetPageUptodate(page); 3501da177e4SLinus Torvalds unlock_page(page); 3511da177e4SLinus Torvalds return; 3521da177e4SLinus Torvalds 3531da177e4SLinus Torvalds still_busy: 354a3972203SNick Piggin bit_spin_unlock(BH_Uptodate_Lock, &first->b_state); 355a3972203SNick Piggin local_irq_restore(flags); 3561da177e4SLinus Torvalds return; 3571da177e4SLinus Torvalds } 3581da177e4SLinus Torvalds 3591da177e4SLinus Torvalds /* 3601da177e4SLinus Torvalds * Completion handler for block_write_full_page() - pages which are unlocked 3611da177e4SLinus Torvalds * during I/O, and which have PageWriteback cleared upon I/O completion. 3621da177e4SLinus Torvalds */ 36335c80d5fSChris Mason void end_buffer_async_write(struct buffer_head *bh, int uptodate) 3641da177e4SLinus Torvalds { 3651da177e4SLinus Torvalds char b[BDEVNAME_SIZE]; 3661da177e4SLinus Torvalds unsigned long flags; 367a3972203SNick Piggin struct buffer_head *first; 3681da177e4SLinus Torvalds struct buffer_head *tmp; 3691da177e4SLinus Torvalds struct page *page; 3701da177e4SLinus Torvalds 3711da177e4SLinus Torvalds BUG_ON(!buffer_async_write(bh)); 3721da177e4SLinus Torvalds 3731da177e4SLinus Torvalds page = bh->b_page; 3741da177e4SLinus Torvalds if (uptodate) { 3751da177e4SLinus Torvalds set_buffer_uptodate(bh); 3761da177e4SLinus Torvalds } else { 37708bafc03SKeith Mannthey if (!quiet_error(bh)) { 3781da177e4SLinus Torvalds buffer_io_error(bh); 3791da177e4SLinus Torvalds printk(KERN_WARNING "lost page write due to " 3801da177e4SLinus Torvalds "I/O error on %s\n", 3811da177e4SLinus Torvalds bdevname(bh->b_bdev, b)); 3821da177e4SLinus Torvalds } 3831da177e4SLinus Torvalds set_bit(AS_EIO, &page->mapping->flags); 38458ff407bSJan Kara set_buffer_write_io_error(bh); 3851da177e4SLinus Torvalds clear_buffer_uptodate(bh); 3861da177e4SLinus Torvalds SetPageError(page); 3871da177e4SLinus Torvalds } 3881da177e4SLinus Torvalds 389a3972203SNick Piggin first = page_buffers(page); 390a3972203SNick Piggin local_irq_save(flags); 391a3972203SNick Piggin bit_spin_lock(BH_Uptodate_Lock, &first->b_state); 392a3972203SNick Piggin 3931da177e4SLinus Torvalds clear_buffer_async_write(bh); 3941da177e4SLinus Torvalds unlock_buffer(bh); 3951da177e4SLinus Torvalds tmp = bh->b_this_page; 3961da177e4SLinus Torvalds while (tmp != bh) { 3971da177e4SLinus Torvalds if (buffer_async_write(tmp)) { 3981da177e4SLinus Torvalds BUG_ON(!buffer_locked(tmp)); 3991da177e4SLinus Torvalds goto still_busy; 4001da177e4SLinus Torvalds } 4011da177e4SLinus Torvalds tmp = tmp->b_this_page; 4021da177e4SLinus Torvalds } 403a3972203SNick Piggin bit_spin_unlock(BH_Uptodate_Lock, &first->b_state); 404a3972203SNick Piggin local_irq_restore(flags); 4051da177e4SLinus Torvalds end_page_writeback(page); 4061da177e4SLinus Torvalds return; 4071da177e4SLinus Torvalds 4081da177e4SLinus Torvalds still_busy: 409a3972203SNick Piggin bit_spin_unlock(BH_Uptodate_Lock, &first->b_state); 410a3972203SNick Piggin local_irq_restore(flags); 4111da177e4SLinus Torvalds return; 4121da177e4SLinus Torvalds } 4131da177e4SLinus Torvalds 4141da177e4SLinus Torvalds /* 4151da177e4SLinus Torvalds * If a page's buffers are under async readin (end_buffer_async_read 4161da177e4SLinus Torvalds * completion) then there is a possibility that another thread of 4171da177e4SLinus Torvalds * control could lock one of the buffers after it has completed 4181da177e4SLinus Torvalds * but while some of the other buffers have not completed. This 4191da177e4SLinus Torvalds * locked buffer would confuse end_buffer_async_read() into not unlocking 4201da177e4SLinus Torvalds * the page. So the absence of BH_Async_Read tells end_buffer_async_read() 4211da177e4SLinus Torvalds * that this buffer is not under async I/O. 4221da177e4SLinus Torvalds * 4231da177e4SLinus Torvalds * The page comes unlocked when it has no locked buffer_async buffers 4241da177e4SLinus Torvalds * left. 4251da177e4SLinus Torvalds * 4261da177e4SLinus Torvalds * PageLocked prevents anyone starting new async I/O reads any of 4271da177e4SLinus Torvalds * the buffers. 4281da177e4SLinus Torvalds * 4291da177e4SLinus Torvalds * PageWriteback is used to prevent simultaneous writeout of the same 4301da177e4SLinus Torvalds * page. 4311da177e4SLinus Torvalds * 4321da177e4SLinus Torvalds * PageLocked prevents anyone from starting writeback of a page which is 4331da177e4SLinus Torvalds * under read I/O (PageWriteback is only ever set against a locked page). 4341da177e4SLinus Torvalds */ 4351da177e4SLinus Torvalds static void mark_buffer_async_read(struct buffer_head *bh) 4361da177e4SLinus Torvalds { 4371da177e4SLinus Torvalds bh->b_end_io = end_buffer_async_read; 4381da177e4SLinus Torvalds set_buffer_async_read(bh); 4391da177e4SLinus Torvalds } 4401da177e4SLinus Torvalds 44135c80d5fSChris Mason void mark_buffer_async_write_endio(struct buffer_head *bh, 44235c80d5fSChris Mason bh_end_io_t *handler) 44335c80d5fSChris Mason { 44435c80d5fSChris Mason bh->b_end_io = handler; 44535c80d5fSChris Mason set_buffer_async_write(bh); 44635c80d5fSChris Mason } 44735c80d5fSChris Mason 4481da177e4SLinus Torvalds void mark_buffer_async_write(struct buffer_head *bh) 4491da177e4SLinus Torvalds { 45035c80d5fSChris Mason mark_buffer_async_write_endio(bh, end_buffer_async_write); 4511da177e4SLinus Torvalds } 4521da177e4SLinus Torvalds EXPORT_SYMBOL(mark_buffer_async_write); 4531da177e4SLinus Torvalds 4541da177e4SLinus Torvalds 4551da177e4SLinus Torvalds /* 4561da177e4SLinus Torvalds * fs/buffer.c contains helper functions for buffer-backed address space's 4571da177e4SLinus Torvalds * fsync functions. A common requirement for buffer-based filesystems is 4581da177e4SLinus Torvalds * that certain data from the backing blockdev needs to be written out for 4591da177e4SLinus Torvalds * a successful fsync(). For example, ext2 indirect blocks need to be 4601da177e4SLinus Torvalds * written back and waited upon before fsync() returns. 4611da177e4SLinus Torvalds * 4621da177e4SLinus Torvalds * The functions mark_buffer_inode_dirty(), fsync_inode_buffers(), 4631da177e4SLinus Torvalds * inode_has_buffers() and invalidate_inode_buffers() are provided for the 4641da177e4SLinus Torvalds * management of a list of dependent buffers at ->i_mapping->private_list. 4651da177e4SLinus Torvalds * 4661da177e4SLinus Torvalds * Locking is a little subtle: try_to_free_buffers() will remove buffers 4671da177e4SLinus Torvalds * from their controlling inode's queue when they are being freed. But 4681da177e4SLinus Torvalds * try_to_free_buffers() will be operating against the *blockdev* mapping 4691da177e4SLinus Torvalds * at the time, not against the S_ISREG file which depends on those buffers. 4701da177e4SLinus Torvalds * So the locking for private_list is via the private_lock in the address_space 4711da177e4SLinus Torvalds * which backs the buffers. Which is different from the address_space 4721da177e4SLinus Torvalds * against which the buffers are listed. So for a particular address_space, 4731da177e4SLinus Torvalds * mapping->private_lock does *not* protect mapping->private_list! In fact, 4741da177e4SLinus Torvalds * mapping->private_list will always be protected by the backing blockdev's 4751da177e4SLinus Torvalds * ->private_lock. 4761da177e4SLinus Torvalds * 4771da177e4SLinus Torvalds * Which introduces a requirement: all buffers on an address_space's 4781da177e4SLinus Torvalds * ->private_list must be from the same address_space: the blockdev's. 4791da177e4SLinus Torvalds * 4801da177e4SLinus Torvalds * address_spaces which do not place buffers at ->private_list via these 4811da177e4SLinus Torvalds * utility functions are free to use private_lock and private_list for 4821da177e4SLinus Torvalds * whatever they want. The only requirement is that list_empty(private_list) 4831da177e4SLinus Torvalds * be true at clear_inode() time. 4841da177e4SLinus Torvalds * 4851da177e4SLinus Torvalds * FIXME: clear_inode should not call invalidate_inode_buffers(). The 4861da177e4SLinus Torvalds * filesystems should do that. invalidate_inode_buffers() should just go 4871da177e4SLinus Torvalds * BUG_ON(!list_empty). 4881da177e4SLinus Torvalds * 4891da177e4SLinus Torvalds * FIXME: mark_buffer_dirty_inode() is a data-plane operation. It should 4901da177e4SLinus Torvalds * take an address_space, not an inode. And it should be called 4911da177e4SLinus Torvalds * mark_buffer_dirty_fsync() to clearly define why those buffers are being 4921da177e4SLinus Torvalds * queued up. 4931da177e4SLinus Torvalds * 4941da177e4SLinus Torvalds * FIXME: mark_buffer_dirty_inode() doesn't need to add the buffer to the 4951da177e4SLinus Torvalds * list if it is already on a list. Because if the buffer is on a list, 4961da177e4SLinus Torvalds * it *must* already be on the right one. If not, the filesystem is being 4971da177e4SLinus Torvalds * silly. This will save a ton of locking. But first we have to ensure 4981da177e4SLinus Torvalds * that buffers are taken *off* the old inode's list when they are freed 4991da177e4SLinus Torvalds * (presumably in truncate). That requires careful auditing of all 5001da177e4SLinus Torvalds * filesystems (do it inside bforget()). It could also be done by bringing 5011da177e4SLinus Torvalds * b_inode back. 5021da177e4SLinus Torvalds */ 5031da177e4SLinus Torvalds 5041da177e4SLinus Torvalds /* 5051da177e4SLinus Torvalds * The buffer's backing address_space's private_lock must be held 5061da177e4SLinus Torvalds */ 507dbacefc9SThomas Petazzoni static void __remove_assoc_queue(struct buffer_head *bh) 5081da177e4SLinus Torvalds { 5091da177e4SLinus Torvalds list_del_init(&bh->b_assoc_buffers); 51058ff407bSJan Kara WARN_ON(!bh->b_assoc_map); 51158ff407bSJan Kara if (buffer_write_io_error(bh)) 51258ff407bSJan Kara set_bit(AS_EIO, &bh->b_assoc_map->flags); 51358ff407bSJan Kara bh->b_assoc_map = NULL; 5141da177e4SLinus Torvalds } 5151da177e4SLinus Torvalds 5161da177e4SLinus Torvalds int inode_has_buffers(struct inode *inode) 5171da177e4SLinus Torvalds { 5181da177e4SLinus Torvalds return !list_empty(&inode->i_data.private_list); 5191da177e4SLinus Torvalds } 5201da177e4SLinus Torvalds 5211da177e4SLinus Torvalds /* 5221da177e4SLinus Torvalds * osync is designed to support O_SYNC io. It waits synchronously for 5231da177e4SLinus Torvalds * all already-submitted IO to complete, but does not queue any new 5241da177e4SLinus Torvalds * writes to the disk. 5251da177e4SLinus Torvalds * 5261da177e4SLinus Torvalds * To do O_SYNC writes, just queue the buffer writes with ll_rw_block as 5271da177e4SLinus Torvalds * you dirty the buffers, and then use osync_inode_buffers to wait for 5281da177e4SLinus Torvalds * completion. Any other dirty buffers which are not yet queued for 5291da177e4SLinus Torvalds * write will not be flushed to disk by the osync. 5301da177e4SLinus Torvalds */ 5311da177e4SLinus Torvalds static int osync_buffers_list(spinlock_t *lock, struct list_head *list) 5321da177e4SLinus Torvalds { 5331da177e4SLinus Torvalds struct buffer_head *bh; 5341da177e4SLinus Torvalds struct list_head *p; 5351da177e4SLinus Torvalds int err = 0; 5361da177e4SLinus Torvalds 5371da177e4SLinus Torvalds spin_lock(lock); 5381da177e4SLinus Torvalds repeat: 5391da177e4SLinus Torvalds list_for_each_prev(p, list) { 5401da177e4SLinus Torvalds bh = BH_ENTRY(p); 5411da177e4SLinus Torvalds if (buffer_locked(bh)) { 5421da177e4SLinus Torvalds get_bh(bh); 5431da177e4SLinus Torvalds spin_unlock(lock); 5441da177e4SLinus Torvalds wait_on_buffer(bh); 5451da177e4SLinus Torvalds if (!buffer_uptodate(bh)) 5461da177e4SLinus Torvalds err = -EIO; 5471da177e4SLinus Torvalds brelse(bh); 5481da177e4SLinus Torvalds spin_lock(lock); 5491da177e4SLinus Torvalds goto repeat; 5501da177e4SLinus Torvalds } 5511da177e4SLinus Torvalds } 5521da177e4SLinus Torvalds spin_unlock(lock); 5531da177e4SLinus Torvalds return err; 5541da177e4SLinus Torvalds } 5551da177e4SLinus Torvalds 556053c525fSJens Axboe void do_thaw_all(struct work_struct *work) 557c2d75438SEric Sandeen { 558c2d75438SEric Sandeen struct super_block *sb; 559c2d75438SEric Sandeen char b[BDEVNAME_SIZE]; 560c2d75438SEric Sandeen 561c2d75438SEric Sandeen spin_lock(&sb_lock); 562c2d75438SEric Sandeen restart: 563c2d75438SEric Sandeen list_for_each_entry(sb, &super_blocks, s_list) { 564c2d75438SEric Sandeen sb->s_count++; 565c2d75438SEric Sandeen spin_unlock(&sb_lock); 566c2d75438SEric Sandeen down_read(&sb->s_umount); 567c2d75438SEric Sandeen while (sb->s_bdev && !thaw_bdev(sb->s_bdev, sb)) 568c2d75438SEric Sandeen printk(KERN_WARNING "Emergency Thaw on %s\n", 569c2d75438SEric Sandeen bdevname(sb->s_bdev, b)); 570c2d75438SEric Sandeen up_read(&sb->s_umount); 571c2d75438SEric Sandeen spin_lock(&sb_lock); 572c2d75438SEric Sandeen if (__put_super_and_need_restart(sb)) 573c2d75438SEric Sandeen goto restart; 574c2d75438SEric Sandeen } 575c2d75438SEric Sandeen spin_unlock(&sb_lock); 576053c525fSJens Axboe kfree(work); 577c2d75438SEric Sandeen printk(KERN_WARNING "Emergency Thaw complete\n"); 578c2d75438SEric Sandeen } 579c2d75438SEric Sandeen 580c2d75438SEric Sandeen /** 581c2d75438SEric Sandeen * emergency_thaw_all -- forcibly thaw every frozen filesystem 582c2d75438SEric Sandeen * 583c2d75438SEric Sandeen * Used for emergency unfreeze of all filesystems via SysRq 584c2d75438SEric Sandeen */ 585c2d75438SEric Sandeen void emergency_thaw_all(void) 586c2d75438SEric Sandeen { 587053c525fSJens Axboe struct work_struct *work; 588053c525fSJens Axboe 589053c525fSJens Axboe work = kmalloc(sizeof(*work), GFP_ATOMIC); 590053c525fSJens Axboe if (work) { 591053c525fSJens Axboe INIT_WORK(work, do_thaw_all); 592053c525fSJens Axboe schedule_work(work); 593053c525fSJens Axboe } 594c2d75438SEric Sandeen } 595c2d75438SEric Sandeen 5961da177e4SLinus Torvalds /** 59778a4a50aSRandy Dunlap * sync_mapping_buffers - write out & wait upon a mapping's "associated" buffers 59867be2dd1SMartin Waitz * @mapping: the mapping which wants those buffers written 5991da177e4SLinus Torvalds * 6001da177e4SLinus Torvalds * Starts I/O against the buffers at mapping->private_list, and waits upon 6011da177e4SLinus Torvalds * that I/O. 6021da177e4SLinus Torvalds * 60367be2dd1SMartin Waitz * Basically, this is a convenience function for fsync(). 60467be2dd1SMartin Waitz * @mapping is a file or directory which needs those buffers to be written for 60567be2dd1SMartin Waitz * a successful fsync(). 6061da177e4SLinus Torvalds */ 6071da177e4SLinus Torvalds int sync_mapping_buffers(struct address_space *mapping) 6081da177e4SLinus Torvalds { 6091da177e4SLinus Torvalds struct address_space *buffer_mapping = mapping->assoc_mapping; 6101da177e4SLinus Torvalds 6111da177e4SLinus Torvalds if (buffer_mapping == NULL || list_empty(&mapping->private_list)) 6121da177e4SLinus Torvalds return 0; 6131da177e4SLinus Torvalds 6141da177e4SLinus Torvalds return fsync_buffers_list(&buffer_mapping->private_lock, 6151da177e4SLinus Torvalds &mapping->private_list); 6161da177e4SLinus Torvalds } 6171da177e4SLinus Torvalds EXPORT_SYMBOL(sync_mapping_buffers); 6181da177e4SLinus Torvalds 6191da177e4SLinus Torvalds /* 6201da177e4SLinus Torvalds * Called when we've recently written block `bblock', and it is known that 6211da177e4SLinus Torvalds * `bblock' was for a buffer_boundary() buffer. This means that the block at 6221da177e4SLinus Torvalds * `bblock + 1' is probably a dirty indirect block. Hunt it down and, if it's 6231da177e4SLinus Torvalds * dirty, schedule it for IO. So that indirects merge nicely with their data. 6241da177e4SLinus Torvalds */ 6251da177e4SLinus Torvalds void write_boundary_block(struct block_device *bdev, 6261da177e4SLinus Torvalds sector_t bblock, unsigned blocksize) 6271da177e4SLinus Torvalds { 6281da177e4SLinus Torvalds struct buffer_head *bh = __find_get_block(bdev, bblock + 1, blocksize); 6291da177e4SLinus Torvalds if (bh) { 6301da177e4SLinus Torvalds if (buffer_dirty(bh)) 6311da177e4SLinus Torvalds ll_rw_block(WRITE, 1, &bh); 6321da177e4SLinus Torvalds put_bh(bh); 6331da177e4SLinus Torvalds } 6341da177e4SLinus Torvalds } 6351da177e4SLinus Torvalds 6361da177e4SLinus Torvalds void mark_buffer_dirty_inode(struct buffer_head *bh, struct inode *inode) 6371da177e4SLinus Torvalds { 6381da177e4SLinus Torvalds struct address_space *mapping = inode->i_mapping; 6391da177e4SLinus Torvalds struct address_space *buffer_mapping = bh->b_page->mapping; 6401da177e4SLinus Torvalds 6411da177e4SLinus Torvalds mark_buffer_dirty(bh); 6421da177e4SLinus Torvalds if (!mapping->assoc_mapping) { 6431da177e4SLinus Torvalds mapping->assoc_mapping = buffer_mapping; 6441da177e4SLinus Torvalds } else { 645e827f923SEric Sesterhenn BUG_ON(mapping->assoc_mapping != buffer_mapping); 6461da177e4SLinus Torvalds } 647535ee2fbSJan Kara if (!bh->b_assoc_map) { 6481da177e4SLinus Torvalds spin_lock(&buffer_mapping->private_lock); 6491da177e4SLinus Torvalds list_move_tail(&bh->b_assoc_buffers, 6501da177e4SLinus Torvalds &mapping->private_list); 65158ff407bSJan Kara bh->b_assoc_map = mapping; 6521da177e4SLinus Torvalds spin_unlock(&buffer_mapping->private_lock); 6531da177e4SLinus Torvalds } 6541da177e4SLinus Torvalds } 6551da177e4SLinus Torvalds EXPORT_SYMBOL(mark_buffer_dirty_inode); 6561da177e4SLinus Torvalds 6571da177e4SLinus Torvalds /* 658787d2214SNick Piggin * Mark the page dirty, and set it dirty in the radix tree, and mark the inode 659787d2214SNick Piggin * dirty. 660787d2214SNick Piggin * 661787d2214SNick Piggin * If warn is true, then emit a warning if the page is not uptodate and has 662787d2214SNick Piggin * not been truncated. 663787d2214SNick Piggin */ 664a8e7d49aSLinus Torvalds static void __set_page_dirty(struct page *page, 665787d2214SNick Piggin struct address_space *mapping, int warn) 666787d2214SNick Piggin { 66719fd6231SNick Piggin spin_lock_irq(&mapping->tree_lock); 668787d2214SNick Piggin if (page->mapping) { /* Race with truncate? */ 669787d2214SNick Piggin WARN_ON_ONCE(warn && !PageUptodate(page)); 670e3a7cca1SEdward Shishkin account_page_dirtied(page, mapping); 671787d2214SNick Piggin radix_tree_tag_set(&mapping->page_tree, 672787d2214SNick Piggin page_index(page), PAGECACHE_TAG_DIRTY); 673787d2214SNick Piggin } 67419fd6231SNick Piggin spin_unlock_irq(&mapping->tree_lock); 675787d2214SNick Piggin __mark_inode_dirty(mapping->host, I_DIRTY_PAGES); 676787d2214SNick Piggin } 677787d2214SNick Piggin 678787d2214SNick Piggin /* 6791da177e4SLinus Torvalds * Add a page to the dirty page list. 6801da177e4SLinus Torvalds * 6811da177e4SLinus Torvalds * It is a sad fact of life that this function is called from several places 6821da177e4SLinus Torvalds * deeply under spinlocking. It may not sleep. 6831da177e4SLinus Torvalds * 6841da177e4SLinus Torvalds * If the page has buffers, the uptodate buffers are set dirty, to preserve 6851da177e4SLinus Torvalds * dirty-state coherency between the page and the buffers. It the page does 6861da177e4SLinus Torvalds * not have buffers then when they are later attached they will all be set 6871da177e4SLinus Torvalds * dirty. 6881da177e4SLinus Torvalds * 6891da177e4SLinus Torvalds * The buffers are dirtied before the page is dirtied. There's a small race 6901da177e4SLinus Torvalds * window in which a writepage caller may see the page cleanness but not the 6911da177e4SLinus Torvalds * buffer dirtiness. That's fine. If this code were to set the page dirty 6921da177e4SLinus Torvalds * before the buffers, a concurrent writepage caller could clear the page dirty 6931da177e4SLinus Torvalds * bit, see a bunch of clean buffers and we'd end up with dirty buffers/clean 6941da177e4SLinus Torvalds * page on the dirty page list. 6951da177e4SLinus Torvalds * 6961da177e4SLinus Torvalds * We use private_lock to lock against try_to_free_buffers while using the 6971da177e4SLinus Torvalds * page's buffer list. Also use this to protect against clean buffers being 6981da177e4SLinus Torvalds * added to the page after it was set dirty. 6991da177e4SLinus Torvalds * 7001da177e4SLinus Torvalds * FIXME: may need to call ->reservepage here as well. That's rather up to the 7011da177e4SLinus Torvalds * address_space though. 7021da177e4SLinus Torvalds */ 7031da177e4SLinus Torvalds int __set_page_dirty_buffers(struct page *page) 7041da177e4SLinus Torvalds { 705a8e7d49aSLinus Torvalds int newly_dirty; 706787d2214SNick Piggin struct address_space *mapping = page_mapping(page); 707ebf7a227SNick Piggin 708ebf7a227SNick Piggin if (unlikely(!mapping)) 709ebf7a227SNick Piggin return !TestSetPageDirty(page); 7101da177e4SLinus Torvalds 7111da177e4SLinus Torvalds spin_lock(&mapping->private_lock); 7121da177e4SLinus Torvalds if (page_has_buffers(page)) { 7131da177e4SLinus Torvalds struct buffer_head *head = page_buffers(page); 7141da177e4SLinus Torvalds struct buffer_head *bh = head; 7151da177e4SLinus Torvalds 7161da177e4SLinus Torvalds do { 7171da177e4SLinus Torvalds set_buffer_dirty(bh); 7181da177e4SLinus Torvalds bh = bh->b_this_page; 7191da177e4SLinus Torvalds } while (bh != head); 7201da177e4SLinus Torvalds } 721a8e7d49aSLinus Torvalds newly_dirty = !TestSetPageDirty(page); 7221da177e4SLinus Torvalds spin_unlock(&mapping->private_lock); 7231da177e4SLinus Torvalds 724a8e7d49aSLinus Torvalds if (newly_dirty) 725a8e7d49aSLinus Torvalds __set_page_dirty(page, mapping, 1); 726a8e7d49aSLinus Torvalds return newly_dirty; 7271da177e4SLinus Torvalds } 7281da177e4SLinus Torvalds EXPORT_SYMBOL(__set_page_dirty_buffers); 7291da177e4SLinus Torvalds 7301da177e4SLinus Torvalds /* 7311da177e4SLinus Torvalds * Write out and wait upon a list of buffers. 7321da177e4SLinus Torvalds * 7331da177e4SLinus Torvalds * We have conflicting pressures: we want to make sure that all 7341da177e4SLinus Torvalds * initially dirty buffers get waited on, but that any subsequently 7351da177e4SLinus Torvalds * dirtied buffers don't. After all, we don't want fsync to last 7361da177e4SLinus Torvalds * forever if somebody is actively writing to the file. 7371da177e4SLinus Torvalds * 7381da177e4SLinus Torvalds * Do this in two main stages: first we copy dirty buffers to a 7391da177e4SLinus Torvalds * temporary inode list, queueing the writes as we go. Then we clean 7401da177e4SLinus Torvalds * up, waiting for those writes to complete. 7411da177e4SLinus Torvalds * 7421da177e4SLinus Torvalds * During this second stage, any subsequent updates to the file may end 7431da177e4SLinus Torvalds * up refiling the buffer on the original inode's dirty list again, so 7441da177e4SLinus Torvalds * there is a chance we will end up with a buffer queued for write but 7451da177e4SLinus Torvalds * not yet completed on that list. So, as a final cleanup we go through 7461da177e4SLinus Torvalds * the osync code to catch these locked, dirty buffers without requeuing 7471da177e4SLinus Torvalds * any newly dirty buffers for write. 7481da177e4SLinus Torvalds */ 7491da177e4SLinus Torvalds static int fsync_buffers_list(spinlock_t *lock, struct list_head *list) 7501da177e4SLinus Torvalds { 7511da177e4SLinus Torvalds struct buffer_head *bh; 7521da177e4SLinus Torvalds struct list_head tmp; 7539cf6b720SJens Axboe struct address_space *mapping, *prev_mapping = NULL; 7541da177e4SLinus Torvalds int err = 0, err2; 7551da177e4SLinus Torvalds 7561da177e4SLinus Torvalds INIT_LIST_HEAD(&tmp); 7571da177e4SLinus Torvalds 7581da177e4SLinus Torvalds spin_lock(lock); 7591da177e4SLinus Torvalds while (!list_empty(list)) { 7601da177e4SLinus Torvalds bh = BH_ENTRY(list->next); 761535ee2fbSJan Kara mapping = bh->b_assoc_map; 76258ff407bSJan Kara __remove_assoc_queue(bh); 763535ee2fbSJan Kara /* Avoid race with mark_buffer_dirty_inode() which does 764535ee2fbSJan Kara * a lockless check and we rely on seeing the dirty bit */ 765535ee2fbSJan Kara smp_mb(); 7661da177e4SLinus Torvalds if (buffer_dirty(bh) || buffer_locked(bh)) { 7671da177e4SLinus Torvalds list_add(&bh->b_assoc_buffers, &tmp); 768535ee2fbSJan Kara bh->b_assoc_map = mapping; 7691da177e4SLinus Torvalds if (buffer_dirty(bh)) { 7701da177e4SLinus Torvalds get_bh(bh); 7711da177e4SLinus Torvalds spin_unlock(lock); 7721da177e4SLinus Torvalds /* 7731da177e4SLinus Torvalds * Ensure any pending I/O completes so that 7741da177e4SLinus Torvalds * ll_rw_block() actually writes the current 7751da177e4SLinus Torvalds * contents - it is a noop if I/O is still in 7761da177e4SLinus Torvalds * flight on potentially older contents. 7771da177e4SLinus Torvalds */ 7789cf6b720SJens Axboe ll_rw_block(SWRITE_SYNC_PLUG, 1, &bh); 7799cf6b720SJens Axboe 7809cf6b720SJens Axboe /* 7819cf6b720SJens Axboe * Kick off IO for the previous mapping. Note 7829cf6b720SJens Axboe * that we will not run the very last mapping, 7839cf6b720SJens Axboe * wait_on_buffer() will do that for us 7849cf6b720SJens Axboe * through sync_buffer(). 7859cf6b720SJens Axboe */ 7869cf6b720SJens Axboe if (prev_mapping && prev_mapping != mapping) 7879cf6b720SJens Axboe blk_run_address_space(prev_mapping); 7889cf6b720SJens Axboe prev_mapping = mapping; 7899cf6b720SJens Axboe 7901da177e4SLinus Torvalds brelse(bh); 7911da177e4SLinus Torvalds spin_lock(lock); 7921da177e4SLinus Torvalds } 7931da177e4SLinus Torvalds } 7941da177e4SLinus Torvalds } 7951da177e4SLinus Torvalds 7961da177e4SLinus Torvalds while (!list_empty(&tmp)) { 7971da177e4SLinus Torvalds bh = BH_ENTRY(tmp.prev); 7981da177e4SLinus Torvalds get_bh(bh); 799535ee2fbSJan Kara mapping = bh->b_assoc_map; 800535ee2fbSJan Kara __remove_assoc_queue(bh); 801535ee2fbSJan Kara /* Avoid race with mark_buffer_dirty_inode() which does 802535ee2fbSJan Kara * a lockless check and we rely on seeing the dirty bit */ 803535ee2fbSJan Kara smp_mb(); 804535ee2fbSJan Kara if (buffer_dirty(bh)) { 805535ee2fbSJan Kara list_add(&bh->b_assoc_buffers, 806e3892296SJan Kara &mapping->private_list); 807535ee2fbSJan Kara bh->b_assoc_map = mapping; 808535ee2fbSJan Kara } 8091da177e4SLinus Torvalds spin_unlock(lock); 8101da177e4SLinus Torvalds wait_on_buffer(bh); 8111da177e4SLinus Torvalds if (!buffer_uptodate(bh)) 8121da177e4SLinus Torvalds err = -EIO; 8131da177e4SLinus Torvalds brelse(bh); 8141da177e4SLinus Torvalds spin_lock(lock); 8151da177e4SLinus Torvalds } 8161da177e4SLinus Torvalds 8171da177e4SLinus Torvalds spin_unlock(lock); 8181da177e4SLinus Torvalds err2 = osync_buffers_list(lock, list); 8191da177e4SLinus Torvalds if (err) 8201da177e4SLinus Torvalds return err; 8211da177e4SLinus Torvalds else 8221da177e4SLinus Torvalds return err2; 8231da177e4SLinus Torvalds } 8241da177e4SLinus Torvalds 8251da177e4SLinus Torvalds /* 8261da177e4SLinus Torvalds * Invalidate any and all dirty buffers on a given inode. We are 8271da177e4SLinus Torvalds * probably unmounting the fs, but that doesn't mean we have already 8281da177e4SLinus Torvalds * done a sync(). Just drop the buffers from the inode list. 8291da177e4SLinus Torvalds * 8301da177e4SLinus Torvalds * NOTE: we take the inode's blockdev's mapping's private_lock. Which 8311da177e4SLinus Torvalds * assumes that all the buffers are against the blockdev. Not true 8321da177e4SLinus Torvalds * for reiserfs. 8331da177e4SLinus Torvalds */ 8341da177e4SLinus Torvalds void invalidate_inode_buffers(struct inode *inode) 8351da177e4SLinus Torvalds { 8361da177e4SLinus Torvalds if (inode_has_buffers(inode)) { 8371da177e4SLinus Torvalds struct address_space *mapping = &inode->i_data; 8381da177e4SLinus Torvalds struct list_head *list = &mapping->private_list; 8391da177e4SLinus Torvalds struct address_space *buffer_mapping = mapping->assoc_mapping; 8401da177e4SLinus Torvalds 8411da177e4SLinus Torvalds spin_lock(&buffer_mapping->private_lock); 8421da177e4SLinus Torvalds while (!list_empty(list)) 8431da177e4SLinus Torvalds __remove_assoc_queue(BH_ENTRY(list->next)); 8441da177e4SLinus Torvalds spin_unlock(&buffer_mapping->private_lock); 8451da177e4SLinus Torvalds } 8461da177e4SLinus Torvalds } 84752b19ac9SJan Kara EXPORT_SYMBOL(invalidate_inode_buffers); 8481da177e4SLinus Torvalds 8491da177e4SLinus Torvalds /* 8501da177e4SLinus Torvalds * Remove any clean buffers from the inode's buffer list. This is called 8511da177e4SLinus Torvalds * when we're trying to free the inode itself. Those buffers can pin it. 8521da177e4SLinus Torvalds * 8531da177e4SLinus Torvalds * Returns true if all buffers were removed. 8541da177e4SLinus Torvalds */ 8551da177e4SLinus Torvalds int remove_inode_buffers(struct inode *inode) 8561da177e4SLinus Torvalds { 8571da177e4SLinus Torvalds int ret = 1; 8581da177e4SLinus Torvalds 8591da177e4SLinus Torvalds if (inode_has_buffers(inode)) { 8601da177e4SLinus Torvalds struct address_space *mapping = &inode->i_data; 8611da177e4SLinus Torvalds struct list_head *list = &mapping->private_list; 8621da177e4SLinus Torvalds struct address_space *buffer_mapping = mapping->assoc_mapping; 8631da177e4SLinus Torvalds 8641da177e4SLinus Torvalds spin_lock(&buffer_mapping->private_lock); 8651da177e4SLinus Torvalds while (!list_empty(list)) { 8661da177e4SLinus Torvalds struct buffer_head *bh = BH_ENTRY(list->next); 8671da177e4SLinus Torvalds if (buffer_dirty(bh)) { 8681da177e4SLinus Torvalds ret = 0; 8691da177e4SLinus Torvalds break; 8701da177e4SLinus Torvalds } 8711da177e4SLinus Torvalds __remove_assoc_queue(bh); 8721da177e4SLinus Torvalds } 8731da177e4SLinus Torvalds spin_unlock(&buffer_mapping->private_lock); 8741da177e4SLinus Torvalds } 8751da177e4SLinus Torvalds return ret; 8761da177e4SLinus Torvalds } 8771da177e4SLinus Torvalds 8781da177e4SLinus Torvalds /* 8791da177e4SLinus Torvalds * Create the appropriate buffers when given a page for data area and 8801da177e4SLinus Torvalds * the size of each buffer.. Use the bh->b_this_page linked list to 8811da177e4SLinus Torvalds * follow the buffers created. Return NULL if unable to create more 8821da177e4SLinus Torvalds * buffers. 8831da177e4SLinus Torvalds * 8841da177e4SLinus Torvalds * The retry flag is used to differentiate async IO (paging, swapping) 8851da177e4SLinus Torvalds * which may not fail from ordinary buffer allocations. 8861da177e4SLinus Torvalds */ 8871da177e4SLinus Torvalds struct buffer_head *alloc_page_buffers(struct page *page, unsigned long size, 8881da177e4SLinus Torvalds int retry) 8891da177e4SLinus Torvalds { 8901da177e4SLinus Torvalds struct buffer_head *bh, *head; 8911da177e4SLinus Torvalds long offset; 8921da177e4SLinus Torvalds 8931da177e4SLinus Torvalds try_again: 8941da177e4SLinus Torvalds head = NULL; 8951da177e4SLinus Torvalds offset = PAGE_SIZE; 8961da177e4SLinus Torvalds while ((offset -= size) >= 0) { 8971da177e4SLinus Torvalds bh = alloc_buffer_head(GFP_NOFS); 8981da177e4SLinus Torvalds if (!bh) 8991da177e4SLinus Torvalds goto no_grow; 9001da177e4SLinus Torvalds 9011da177e4SLinus Torvalds bh->b_bdev = NULL; 9021da177e4SLinus Torvalds bh->b_this_page = head; 9031da177e4SLinus Torvalds bh->b_blocknr = -1; 9041da177e4SLinus Torvalds head = bh; 9051da177e4SLinus Torvalds 9061da177e4SLinus Torvalds bh->b_state = 0; 9071da177e4SLinus Torvalds atomic_set(&bh->b_count, 0); 908fc5cd582SChris Mason bh->b_private = NULL; 9091da177e4SLinus Torvalds bh->b_size = size; 9101da177e4SLinus Torvalds 9111da177e4SLinus Torvalds /* Link the buffer to its page */ 9121da177e4SLinus Torvalds set_bh_page(bh, page, offset); 9131da177e4SLinus Torvalds 91401ffe339SNathan Scott init_buffer(bh, NULL, NULL); 9151da177e4SLinus Torvalds } 9161da177e4SLinus Torvalds return head; 9171da177e4SLinus Torvalds /* 9181da177e4SLinus Torvalds * In case anything failed, we just free everything we got. 9191da177e4SLinus Torvalds */ 9201da177e4SLinus Torvalds no_grow: 9211da177e4SLinus Torvalds if (head) { 9221da177e4SLinus Torvalds do { 9231da177e4SLinus Torvalds bh = head; 9241da177e4SLinus Torvalds head = head->b_this_page; 9251da177e4SLinus Torvalds free_buffer_head(bh); 9261da177e4SLinus Torvalds } while (head); 9271da177e4SLinus Torvalds } 9281da177e4SLinus Torvalds 9291da177e4SLinus Torvalds /* 9301da177e4SLinus Torvalds * Return failure for non-async IO requests. Async IO requests 9311da177e4SLinus Torvalds * are not allowed to fail, so we have to wait until buffer heads 9321da177e4SLinus Torvalds * become available. But we don't want tasks sleeping with 9331da177e4SLinus Torvalds * partially complete buffers, so all were released above. 9341da177e4SLinus Torvalds */ 9351da177e4SLinus Torvalds if (!retry) 9361da177e4SLinus Torvalds return NULL; 9371da177e4SLinus Torvalds 9381da177e4SLinus Torvalds /* We're _really_ low on memory. Now we just 9391da177e4SLinus Torvalds * wait for old buffer heads to become free due to 9401da177e4SLinus Torvalds * finishing IO. Since this is an async request and 9411da177e4SLinus Torvalds * the reserve list is empty, we're sure there are 9421da177e4SLinus Torvalds * async buffer heads in use. 9431da177e4SLinus Torvalds */ 9441da177e4SLinus Torvalds free_more_memory(); 9451da177e4SLinus Torvalds goto try_again; 9461da177e4SLinus Torvalds } 9471da177e4SLinus Torvalds EXPORT_SYMBOL_GPL(alloc_page_buffers); 9481da177e4SLinus Torvalds 9491da177e4SLinus Torvalds static inline void 9501da177e4SLinus Torvalds link_dev_buffers(struct page *page, struct buffer_head *head) 9511da177e4SLinus Torvalds { 9521da177e4SLinus Torvalds struct buffer_head *bh, *tail; 9531da177e4SLinus Torvalds 9541da177e4SLinus Torvalds bh = head; 9551da177e4SLinus Torvalds do { 9561da177e4SLinus Torvalds tail = bh; 9571da177e4SLinus Torvalds bh = bh->b_this_page; 9581da177e4SLinus Torvalds } while (bh); 9591da177e4SLinus Torvalds tail->b_this_page = head; 9601da177e4SLinus Torvalds attach_page_buffers(page, head); 9611da177e4SLinus Torvalds } 9621da177e4SLinus Torvalds 9631da177e4SLinus Torvalds /* 9641da177e4SLinus Torvalds * Initialise the state of a blockdev page's buffers. 9651da177e4SLinus Torvalds */ 9661da177e4SLinus Torvalds static void 9671da177e4SLinus Torvalds init_page_buffers(struct page *page, struct block_device *bdev, 9681da177e4SLinus Torvalds sector_t block, int size) 9691da177e4SLinus Torvalds { 9701da177e4SLinus Torvalds struct buffer_head *head = page_buffers(page); 9711da177e4SLinus Torvalds struct buffer_head *bh = head; 9721da177e4SLinus Torvalds int uptodate = PageUptodate(page); 9731da177e4SLinus Torvalds 9741da177e4SLinus Torvalds do { 9751da177e4SLinus Torvalds if (!buffer_mapped(bh)) { 9761da177e4SLinus Torvalds init_buffer(bh, NULL, NULL); 9771da177e4SLinus Torvalds bh->b_bdev = bdev; 9781da177e4SLinus Torvalds bh->b_blocknr = block; 9791da177e4SLinus Torvalds if (uptodate) 9801da177e4SLinus Torvalds set_buffer_uptodate(bh); 9811da177e4SLinus Torvalds set_buffer_mapped(bh); 9821da177e4SLinus Torvalds } 9831da177e4SLinus Torvalds block++; 9841da177e4SLinus Torvalds bh = bh->b_this_page; 9851da177e4SLinus Torvalds } while (bh != head); 9861da177e4SLinus Torvalds } 9871da177e4SLinus Torvalds 9881da177e4SLinus Torvalds /* 9891da177e4SLinus Torvalds * Create the page-cache page that contains the requested block. 9901da177e4SLinus Torvalds * 9911da177e4SLinus Torvalds * This is user purely for blockdev mappings. 9921da177e4SLinus Torvalds */ 9931da177e4SLinus Torvalds static struct page * 9941da177e4SLinus Torvalds grow_dev_page(struct block_device *bdev, sector_t block, 9951da177e4SLinus Torvalds pgoff_t index, int size) 9961da177e4SLinus Torvalds { 9971da177e4SLinus Torvalds struct inode *inode = bdev->bd_inode; 9981da177e4SLinus Torvalds struct page *page; 9991da177e4SLinus Torvalds struct buffer_head *bh; 10001da177e4SLinus Torvalds 1001ea125892SChristoph Lameter page = find_or_create_page(inode->i_mapping, index, 1002769848c0SMel Gorman (mapping_gfp_mask(inode->i_mapping) & ~__GFP_FS)|__GFP_MOVABLE); 10031da177e4SLinus Torvalds if (!page) 10041da177e4SLinus Torvalds return NULL; 10051da177e4SLinus Torvalds 1006e827f923SEric Sesterhenn BUG_ON(!PageLocked(page)); 10071da177e4SLinus Torvalds 10081da177e4SLinus Torvalds if (page_has_buffers(page)) { 10091da177e4SLinus Torvalds bh = page_buffers(page); 10101da177e4SLinus Torvalds if (bh->b_size == size) { 10111da177e4SLinus Torvalds init_page_buffers(page, bdev, block, size); 10121da177e4SLinus Torvalds return page; 10131da177e4SLinus Torvalds } 10141da177e4SLinus Torvalds if (!try_to_free_buffers(page)) 10151da177e4SLinus Torvalds goto failed; 10161da177e4SLinus Torvalds } 10171da177e4SLinus Torvalds 10181da177e4SLinus Torvalds /* 10191da177e4SLinus Torvalds * Allocate some buffers for this page 10201da177e4SLinus Torvalds */ 10211da177e4SLinus Torvalds bh = alloc_page_buffers(page, size, 0); 10221da177e4SLinus Torvalds if (!bh) 10231da177e4SLinus Torvalds goto failed; 10241da177e4SLinus Torvalds 10251da177e4SLinus Torvalds /* 10261da177e4SLinus Torvalds * Link the page to the buffers and initialise them. Take the 10271da177e4SLinus Torvalds * lock to be atomic wrt __find_get_block(), which does not 10281da177e4SLinus Torvalds * run under the page lock. 10291da177e4SLinus Torvalds */ 10301da177e4SLinus Torvalds spin_lock(&inode->i_mapping->private_lock); 10311da177e4SLinus Torvalds link_dev_buffers(page, bh); 10321da177e4SLinus Torvalds init_page_buffers(page, bdev, block, size); 10331da177e4SLinus Torvalds spin_unlock(&inode->i_mapping->private_lock); 10341da177e4SLinus Torvalds return page; 10351da177e4SLinus Torvalds 10361da177e4SLinus Torvalds failed: 10371da177e4SLinus Torvalds BUG(); 10381da177e4SLinus Torvalds unlock_page(page); 10391da177e4SLinus Torvalds page_cache_release(page); 10401da177e4SLinus Torvalds return NULL; 10411da177e4SLinus Torvalds } 10421da177e4SLinus Torvalds 10431da177e4SLinus Torvalds /* 10441da177e4SLinus Torvalds * Create buffers for the specified block device block's page. If 10451da177e4SLinus Torvalds * that page was dirty, the buffers are set dirty also. 10461da177e4SLinus Torvalds */ 1047858119e1SArjan van de Ven static int 10481da177e4SLinus Torvalds grow_buffers(struct block_device *bdev, sector_t block, int size) 10491da177e4SLinus Torvalds { 10501da177e4SLinus Torvalds struct page *page; 10511da177e4SLinus Torvalds pgoff_t index; 10521da177e4SLinus Torvalds int sizebits; 10531da177e4SLinus Torvalds 10541da177e4SLinus Torvalds sizebits = -1; 10551da177e4SLinus Torvalds do { 10561da177e4SLinus Torvalds sizebits++; 10571da177e4SLinus Torvalds } while ((size << sizebits) < PAGE_SIZE); 10581da177e4SLinus Torvalds 10591da177e4SLinus Torvalds index = block >> sizebits; 10601da177e4SLinus Torvalds 1061e5657933SAndrew Morton /* 1062e5657933SAndrew Morton * Check for a block which wants to lie outside our maximum possible 1063e5657933SAndrew Morton * pagecache index. (this comparison is done using sector_t types). 1064e5657933SAndrew Morton */ 1065e5657933SAndrew Morton if (unlikely(index != block >> sizebits)) { 1066e5657933SAndrew Morton char b[BDEVNAME_SIZE]; 1067e5657933SAndrew Morton 1068e5657933SAndrew Morton printk(KERN_ERR "%s: requested out-of-range block %llu for " 1069e5657933SAndrew Morton "device %s\n", 10708e24eea7SHarvey Harrison __func__, (unsigned long long)block, 1071e5657933SAndrew Morton bdevname(bdev, b)); 1072e5657933SAndrew Morton return -EIO; 1073e5657933SAndrew Morton } 1074e5657933SAndrew Morton block = index << sizebits; 10751da177e4SLinus Torvalds /* Create a page with the proper size buffers.. */ 10761da177e4SLinus Torvalds page = grow_dev_page(bdev, block, index, size); 10771da177e4SLinus Torvalds if (!page) 10781da177e4SLinus Torvalds return 0; 10791da177e4SLinus Torvalds unlock_page(page); 10801da177e4SLinus Torvalds page_cache_release(page); 10811da177e4SLinus Torvalds return 1; 10821da177e4SLinus Torvalds } 10831da177e4SLinus Torvalds 108475c96f85SAdrian Bunk static struct buffer_head * 10851da177e4SLinus Torvalds __getblk_slow(struct block_device *bdev, sector_t block, int size) 10861da177e4SLinus Torvalds { 10871da177e4SLinus Torvalds /* Size must be multiple of hard sectorsize */ 1088e1defc4fSMartin K. Petersen if (unlikely(size & (bdev_logical_block_size(bdev)-1) || 10891da177e4SLinus Torvalds (size < 512 || size > PAGE_SIZE))) { 10901da177e4SLinus Torvalds printk(KERN_ERR "getblk(): invalid block size %d requested\n", 10911da177e4SLinus Torvalds size); 1092e1defc4fSMartin K. Petersen printk(KERN_ERR "logical block size: %d\n", 1093e1defc4fSMartin K. Petersen bdev_logical_block_size(bdev)); 10941da177e4SLinus Torvalds 10951da177e4SLinus Torvalds dump_stack(); 10961da177e4SLinus Torvalds return NULL; 10971da177e4SLinus Torvalds } 10981da177e4SLinus Torvalds 10991da177e4SLinus Torvalds for (;;) { 11001da177e4SLinus Torvalds struct buffer_head * bh; 1101e5657933SAndrew Morton int ret; 11021da177e4SLinus Torvalds 11031da177e4SLinus Torvalds bh = __find_get_block(bdev, block, size); 11041da177e4SLinus Torvalds if (bh) 11051da177e4SLinus Torvalds return bh; 11061da177e4SLinus Torvalds 1107e5657933SAndrew Morton ret = grow_buffers(bdev, block, size); 1108e5657933SAndrew Morton if (ret < 0) 1109e5657933SAndrew Morton return NULL; 1110e5657933SAndrew Morton if (ret == 0) 11111da177e4SLinus Torvalds free_more_memory(); 11121da177e4SLinus Torvalds } 11131da177e4SLinus Torvalds } 11141da177e4SLinus Torvalds 11151da177e4SLinus Torvalds /* 11161da177e4SLinus Torvalds * The relationship between dirty buffers and dirty pages: 11171da177e4SLinus Torvalds * 11181da177e4SLinus Torvalds * Whenever a page has any dirty buffers, the page's dirty bit is set, and 11191da177e4SLinus Torvalds * the page is tagged dirty in its radix tree. 11201da177e4SLinus Torvalds * 11211da177e4SLinus Torvalds * At all times, the dirtiness of the buffers represents the dirtiness of 11221da177e4SLinus Torvalds * subsections of the page. If the page has buffers, the page dirty bit is 11231da177e4SLinus Torvalds * merely a hint about the true dirty state. 11241da177e4SLinus Torvalds * 11251da177e4SLinus Torvalds * When a page is set dirty in its entirety, all its buffers are marked dirty 11261da177e4SLinus Torvalds * (if the page has buffers). 11271da177e4SLinus Torvalds * 11281da177e4SLinus Torvalds * When a buffer is marked dirty, its page is dirtied, but the page's other 11291da177e4SLinus Torvalds * buffers are not. 11301da177e4SLinus Torvalds * 11311da177e4SLinus Torvalds * Also. When blockdev buffers are explicitly read with bread(), they 11321da177e4SLinus Torvalds * individually become uptodate. But their backing page remains not 11331da177e4SLinus Torvalds * uptodate - even if all of its buffers are uptodate. A subsequent 11341da177e4SLinus Torvalds * block_read_full_page() against that page will discover all the uptodate 11351da177e4SLinus Torvalds * buffers, will set the page uptodate and will perform no I/O. 11361da177e4SLinus Torvalds */ 11371da177e4SLinus Torvalds 11381da177e4SLinus Torvalds /** 11391da177e4SLinus Torvalds * mark_buffer_dirty - mark a buffer_head as needing writeout 114067be2dd1SMartin Waitz * @bh: the buffer_head to mark dirty 11411da177e4SLinus Torvalds * 11421da177e4SLinus Torvalds * mark_buffer_dirty() will set the dirty bit against the buffer, then set its 11431da177e4SLinus Torvalds * backing page dirty, then tag the page as dirty in its address_space's radix 11441da177e4SLinus Torvalds * tree and then attach the address_space's inode to its superblock's dirty 11451da177e4SLinus Torvalds * inode list. 11461da177e4SLinus Torvalds * 11471da177e4SLinus Torvalds * mark_buffer_dirty() is atomic. It takes bh->b_page->mapping->private_lock, 11481da177e4SLinus Torvalds * mapping->tree_lock and the global inode_lock. 11491da177e4SLinus Torvalds */ 1150fc9b52cdSHarvey Harrison void mark_buffer_dirty(struct buffer_head *bh) 11511da177e4SLinus Torvalds { 1152787d2214SNick Piggin WARN_ON_ONCE(!buffer_uptodate(bh)); 11531be62dc1SLinus Torvalds 11541be62dc1SLinus Torvalds /* 11551be62dc1SLinus Torvalds * Very *carefully* optimize the it-is-already-dirty case. 11561be62dc1SLinus Torvalds * 11571be62dc1SLinus Torvalds * Don't let the final "is it dirty" escape to before we 11581be62dc1SLinus Torvalds * perhaps modified the buffer. 11591be62dc1SLinus Torvalds */ 11601be62dc1SLinus Torvalds if (buffer_dirty(bh)) { 11611be62dc1SLinus Torvalds smp_mb(); 11621be62dc1SLinus Torvalds if (buffer_dirty(bh)) 11631be62dc1SLinus Torvalds return; 11641be62dc1SLinus Torvalds } 11651be62dc1SLinus Torvalds 1166a8e7d49aSLinus Torvalds if (!test_set_buffer_dirty(bh)) { 1167a8e7d49aSLinus Torvalds struct page *page = bh->b_page; 11688e9d78edSLinus Torvalds if (!TestSetPageDirty(page)) { 11698e9d78edSLinus Torvalds struct address_space *mapping = page_mapping(page); 11708e9d78edSLinus Torvalds if (mapping) 11718e9d78edSLinus Torvalds __set_page_dirty(page, mapping, 0); 11728e9d78edSLinus Torvalds } 1173a8e7d49aSLinus Torvalds } 11741da177e4SLinus Torvalds } 11751da177e4SLinus Torvalds 11761da177e4SLinus Torvalds /* 11771da177e4SLinus Torvalds * Decrement a buffer_head's reference count. If all buffers against a page 11781da177e4SLinus Torvalds * have zero reference count, are clean and unlocked, and if the page is clean 11791da177e4SLinus Torvalds * and unlocked then try_to_free_buffers() may strip the buffers from the page 11801da177e4SLinus Torvalds * in preparation for freeing it (sometimes, rarely, buffers are removed from 11811da177e4SLinus Torvalds * a page but it ends up not being freed, and buffers may later be reattached). 11821da177e4SLinus Torvalds */ 11831da177e4SLinus Torvalds void __brelse(struct buffer_head * buf) 11841da177e4SLinus Torvalds { 11851da177e4SLinus Torvalds if (atomic_read(&buf->b_count)) { 11861da177e4SLinus Torvalds put_bh(buf); 11871da177e4SLinus Torvalds return; 11881da177e4SLinus Torvalds } 11895c752ad9SArjan van de Ven WARN(1, KERN_ERR "VFS: brelse: Trying to free free buffer\n"); 11901da177e4SLinus Torvalds } 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 } 12091da177e4SLinus Torvalds 12101da177e4SLinus Torvalds static struct buffer_head *__bread_slow(struct buffer_head *bh) 12111da177e4SLinus Torvalds { 12121da177e4SLinus Torvalds lock_buffer(bh); 12131da177e4SLinus Torvalds if (buffer_uptodate(bh)) { 12141da177e4SLinus Torvalds unlock_buffer(bh); 12151da177e4SLinus Torvalds return bh; 12161da177e4SLinus Torvalds } else { 12171da177e4SLinus Torvalds get_bh(bh); 12181da177e4SLinus Torvalds bh->b_end_io = end_buffer_read_sync; 12191da177e4SLinus Torvalds submit_bh(READ, bh); 12201da177e4SLinus Torvalds wait_on_buffer(bh); 12211da177e4SLinus Torvalds if (buffer_uptodate(bh)) 12221da177e4SLinus Torvalds return bh; 12231da177e4SLinus Torvalds } 12241da177e4SLinus Torvalds brelse(bh); 12251da177e4SLinus Torvalds return NULL; 12261da177e4SLinus Torvalds } 12271da177e4SLinus Torvalds 12281da177e4SLinus Torvalds /* 12291da177e4SLinus Torvalds * Per-cpu buffer LRU implementation. To reduce the cost of __find_get_block(). 12301da177e4SLinus Torvalds * The bhs[] array is sorted - newest buffer is at bhs[0]. Buffers have their 12311da177e4SLinus Torvalds * refcount elevated by one when they're in an LRU. A buffer can only appear 12321da177e4SLinus Torvalds * once in a particular CPU's LRU. A single buffer can be present in multiple 12331da177e4SLinus Torvalds * CPU's LRUs at the same time. 12341da177e4SLinus Torvalds * 12351da177e4SLinus Torvalds * This is a transparent caching front-end to sb_bread(), sb_getblk() and 12361da177e4SLinus Torvalds * sb_find_get_block(). 12371da177e4SLinus Torvalds * 12381da177e4SLinus Torvalds * The LRUs themselves only need locking against invalidate_bh_lrus. We use 12391da177e4SLinus Torvalds * a local interrupt disable for that. 12401da177e4SLinus Torvalds */ 12411da177e4SLinus Torvalds 12421da177e4SLinus Torvalds #define BH_LRU_SIZE 8 12431da177e4SLinus Torvalds 12441da177e4SLinus Torvalds struct bh_lru { 12451da177e4SLinus Torvalds struct buffer_head *bhs[BH_LRU_SIZE]; 12461da177e4SLinus Torvalds }; 12471da177e4SLinus Torvalds 12481da177e4SLinus Torvalds static DEFINE_PER_CPU(struct bh_lru, bh_lrus) = {{ NULL }}; 12491da177e4SLinus Torvalds 12501da177e4SLinus Torvalds #ifdef CONFIG_SMP 12511da177e4SLinus Torvalds #define bh_lru_lock() local_irq_disable() 12521da177e4SLinus Torvalds #define bh_lru_unlock() local_irq_enable() 12531da177e4SLinus Torvalds #else 12541da177e4SLinus Torvalds #define bh_lru_lock() preempt_disable() 12551da177e4SLinus Torvalds #define bh_lru_unlock() preempt_enable() 12561da177e4SLinus Torvalds #endif 12571da177e4SLinus Torvalds 12581da177e4SLinus Torvalds static inline void check_irqs_on(void) 12591da177e4SLinus Torvalds { 12601da177e4SLinus Torvalds #ifdef irqs_disabled 12611da177e4SLinus Torvalds BUG_ON(irqs_disabled()); 12621da177e4SLinus Torvalds #endif 12631da177e4SLinus Torvalds } 12641da177e4SLinus Torvalds 12651da177e4SLinus Torvalds /* 12661da177e4SLinus Torvalds * The LRU management algorithm is dopey-but-simple. Sorry. 12671da177e4SLinus Torvalds */ 12681da177e4SLinus Torvalds static void bh_lru_install(struct buffer_head *bh) 12691da177e4SLinus Torvalds { 12701da177e4SLinus Torvalds struct buffer_head *evictee = NULL; 12711da177e4SLinus Torvalds struct bh_lru *lru; 12721da177e4SLinus Torvalds 12731da177e4SLinus Torvalds check_irqs_on(); 12741da177e4SLinus Torvalds bh_lru_lock(); 12751da177e4SLinus Torvalds lru = &__get_cpu_var(bh_lrus); 12761da177e4SLinus Torvalds if (lru->bhs[0] != bh) { 12771da177e4SLinus Torvalds struct buffer_head *bhs[BH_LRU_SIZE]; 12781da177e4SLinus Torvalds int in; 12791da177e4SLinus Torvalds int out = 0; 12801da177e4SLinus Torvalds 12811da177e4SLinus Torvalds get_bh(bh); 12821da177e4SLinus Torvalds bhs[out++] = bh; 12831da177e4SLinus Torvalds for (in = 0; in < BH_LRU_SIZE; in++) { 12841da177e4SLinus Torvalds struct buffer_head *bh2 = lru->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; 12991da177e4SLinus Torvalds memcpy(lru->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; 13141da177e4SLinus Torvalds struct bh_lru *lru; 13153991d3bdSTomasz Kvarsin unsigned int i; 13161da177e4SLinus Torvalds 13171da177e4SLinus Torvalds check_irqs_on(); 13181da177e4SLinus Torvalds bh_lru_lock(); 13191da177e4SLinus Torvalds lru = &__get_cpu_var(bh_lrus); 13201da177e4SLinus Torvalds for (i = 0; i < BH_LRU_SIZE; i++) { 13211da177e4SLinus Torvalds struct buffer_head *bh = lru->bhs[i]; 13221da177e4SLinus Torvalds 13231da177e4SLinus Torvalds if (bh && bh->b_bdev == bdev && 13241da177e4SLinus Torvalds bh->b_blocknr == block && bh->b_size == size) { 13251da177e4SLinus Torvalds if (i) { 13261da177e4SLinus Torvalds while (i) { 13271da177e4SLinus Torvalds lru->bhs[i] = lru->bhs[i - 1]; 13281da177e4SLinus Torvalds i--; 13291da177e4SLinus Torvalds } 13301da177e4SLinus Torvalds lru->bhs[0] = bh; 13311da177e4SLinus Torvalds } 13321da177e4SLinus Torvalds get_bh(bh); 13331da177e4SLinus Torvalds ret = bh; 13341da177e4SLinus Torvalds break; 13351da177e4SLinus Torvalds } 13361da177e4SLinus Torvalds } 13371da177e4SLinus Torvalds bh_lru_unlock(); 13381da177e4SLinus Torvalds return ret; 13391da177e4SLinus Torvalds } 13401da177e4SLinus Torvalds 13411da177e4SLinus Torvalds /* 13421da177e4SLinus Torvalds * Perform a pagecache lookup for the matching buffer. If it's there, refresh 13431da177e4SLinus Torvalds * it in the LRU and mark it as accessed. If it is not present then return 13441da177e4SLinus Torvalds * NULL 13451da177e4SLinus Torvalds */ 13461da177e4SLinus Torvalds struct buffer_head * 13473991d3bdSTomasz Kvarsin __find_get_block(struct block_device *bdev, sector_t block, unsigned size) 13481da177e4SLinus Torvalds { 13491da177e4SLinus Torvalds struct buffer_head *bh = lookup_bh_lru(bdev, block, size); 13501da177e4SLinus Torvalds 13511da177e4SLinus Torvalds if (bh == NULL) { 1352385fd4c5SCoywolf Qi Hunt bh = __find_get_block_slow(bdev, block); 13531da177e4SLinus Torvalds if (bh) 13541da177e4SLinus Torvalds bh_lru_install(bh); 13551da177e4SLinus Torvalds } 13561da177e4SLinus Torvalds if (bh) 13571da177e4SLinus Torvalds touch_buffer(bh); 13581da177e4SLinus Torvalds return bh; 13591da177e4SLinus Torvalds } 13601da177e4SLinus Torvalds EXPORT_SYMBOL(__find_get_block); 13611da177e4SLinus Torvalds 13621da177e4SLinus Torvalds /* 13631da177e4SLinus Torvalds * __getblk will locate (and, if necessary, create) the buffer_head 13641da177e4SLinus Torvalds * which corresponds to the passed block_device, block and size. The 13651da177e4SLinus Torvalds * returned buffer has its reference count incremented. 13661da177e4SLinus Torvalds * 13671da177e4SLinus Torvalds * __getblk() cannot fail - it just keeps trying. If you pass it an 13681da177e4SLinus Torvalds * illegal block number, __getblk() will happily return a buffer_head 13691da177e4SLinus Torvalds * which represents the non-existent block. Very weird. 13701da177e4SLinus Torvalds * 13711da177e4SLinus Torvalds * __getblk() will lock up the machine if grow_dev_page's try_to_free_buffers() 13721da177e4SLinus Torvalds * attempt is failing. FIXME, perhaps? 13731da177e4SLinus Torvalds */ 13741da177e4SLinus Torvalds struct buffer_head * 13753991d3bdSTomasz Kvarsin __getblk(struct block_device *bdev, sector_t block, unsigned size) 13761da177e4SLinus Torvalds { 13771da177e4SLinus Torvalds struct buffer_head *bh = __find_get_block(bdev, block, size); 13781da177e4SLinus Torvalds 13791da177e4SLinus Torvalds might_sleep(); 13801da177e4SLinus Torvalds if (bh == NULL) 13811da177e4SLinus Torvalds bh = __getblk_slow(bdev, block, size); 13821da177e4SLinus Torvalds return bh; 13831da177e4SLinus Torvalds } 13841da177e4SLinus Torvalds EXPORT_SYMBOL(__getblk); 13851da177e4SLinus Torvalds 13861da177e4SLinus Torvalds /* 13871da177e4SLinus Torvalds * Do async read-ahead on a buffer.. 13881da177e4SLinus Torvalds */ 13893991d3bdSTomasz Kvarsin void __breadahead(struct block_device *bdev, sector_t block, unsigned size) 13901da177e4SLinus Torvalds { 13911da177e4SLinus Torvalds struct buffer_head *bh = __getblk(bdev, block, size); 1392a3e713b5SAndrew Morton if (likely(bh)) { 13931da177e4SLinus Torvalds ll_rw_block(READA, 1, &bh); 13941da177e4SLinus Torvalds brelse(bh); 13951da177e4SLinus Torvalds } 1396a3e713b5SAndrew Morton } 13971da177e4SLinus Torvalds EXPORT_SYMBOL(__breadahead); 13981da177e4SLinus Torvalds 13991da177e4SLinus Torvalds /** 14001da177e4SLinus Torvalds * __bread() - reads a specified block and returns the bh 140167be2dd1SMartin Waitz * @bdev: the block_device to read from 14021da177e4SLinus Torvalds * @block: number of block 14031da177e4SLinus Torvalds * @size: size (in bytes) to read 14041da177e4SLinus Torvalds * 14051da177e4SLinus Torvalds * Reads a specified block, and returns buffer head that contains it. 14061da177e4SLinus Torvalds * It returns NULL if the block was unreadable. 14071da177e4SLinus Torvalds */ 14081da177e4SLinus Torvalds struct buffer_head * 14093991d3bdSTomasz Kvarsin __bread(struct block_device *bdev, sector_t block, unsigned size) 14101da177e4SLinus Torvalds { 14111da177e4SLinus Torvalds struct buffer_head *bh = __getblk(bdev, block, size); 14121da177e4SLinus Torvalds 1413a3e713b5SAndrew Morton if (likely(bh) && !buffer_uptodate(bh)) 14141da177e4SLinus Torvalds bh = __bread_slow(bh); 14151da177e4SLinus Torvalds return bh; 14161da177e4SLinus Torvalds } 14171da177e4SLinus Torvalds EXPORT_SYMBOL(__bread); 14181da177e4SLinus Torvalds 14191da177e4SLinus Torvalds /* 14201da177e4SLinus Torvalds * invalidate_bh_lrus() is called rarely - but not only at unmount. 14211da177e4SLinus Torvalds * This doesn't race because it runs in each cpu either in irq 14221da177e4SLinus Torvalds * or with preempt disabled. 14231da177e4SLinus Torvalds */ 14241da177e4SLinus Torvalds static void invalidate_bh_lru(void *arg) 14251da177e4SLinus Torvalds { 14261da177e4SLinus Torvalds struct bh_lru *b = &get_cpu_var(bh_lrus); 14271da177e4SLinus Torvalds int i; 14281da177e4SLinus Torvalds 14291da177e4SLinus Torvalds for (i = 0; i < BH_LRU_SIZE; i++) { 14301da177e4SLinus Torvalds brelse(b->bhs[i]); 14311da177e4SLinus Torvalds b->bhs[i] = NULL; 14321da177e4SLinus Torvalds } 14331da177e4SLinus Torvalds put_cpu_var(bh_lrus); 14341da177e4SLinus Torvalds } 14351da177e4SLinus Torvalds 1436f9a14399SPeter Zijlstra void invalidate_bh_lrus(void) 14371da177e4SLinus Torvalds { 143815c8b6c1SJens Axboe on_each_cpu(invalidate_bh_lru, NULL, 1); 14391da177e4SLinus Torvalds } 14409db5579bSNick Piggin EXPORT_SYMBOL_GPL(invalidate_bh_lrus); 14411da177e4SLinus Torvalds 14421da177e4SLinus Torvalds void set_bh_page(struct buffer_head *bh, 14431da177e4SLinus Torvalds struct page *page, unsigned long offset) 14441da177e4SLinus Torvalds { 14451da177e4SLinus Torvalds bh->b_page = page; 1446e827f923SEric Sesterhenn BUG_ON(offset >= PAGE_SIZE); 14471da177e4SLinus Torvalds if (PageHighMem(page)) 14481da177e4SLinus Torvalds /* 14491da177e4SLinus Torvalds * This catches illegal uses and preserves the offset: 14501da177e4SLinus Torvalds */ 14511da177e4SLinus Torvalds bh->b_data = (char *)(0 + offset); 14521da177e4SLinus Torvalds else 14531da177e4SLinus Torvalds bh->b_data = page_address(page) + offset; 14541da177e4SLinus Torvalds } 14551da177e4SLinus Torvalds EXPORT_SYMBOL(set_bh_page); 14561da177e4SLinus Torvalds 14571da177e4SLinus Torvalds /* 14581da177e4SLinus Torvalds * Called when truncating a buffer on a page completely. 14591da177e4SLinus Torvalds */ 1460858119e1SArjan van de Ven static void discard_buffer(struct buffer_head * bh) 14611da177e4SLinus Torvalds { 14621da177e4SLinus Torvalds lock_buffer(bh); 14631da177e4SLinus Torvalds clear_buffer_dirty(bh); 14641da177e4SLinus Torvalds bh->b_bdev = NULL; 14651da177e4SLinus Torvalds clear_buffer_mapped(bh); 14661da177e4SLinus Torvalds clear_buffer_req(bh); 14671da177e4SLinus Torvalds clear_buffer_new(bh); 14681da177e4SLinus Torvalds clear_buffer_delay(bh); 146933a266ddSDavid Chinner clear_buffer_unwritten(bh); 14701da177e4SLinus Torvalds unlock_buffer(bh); 14711da177e4SLinus Torvalds } 14721da177e4SLinus Torvalds 14731da177e4SLinus Torvalds /** 14741da177e4SLinus Torvalds * block_invalidatepage - invalidate part of all of a buffer-backed page 14751da177e4SLinus Torvalds * 14761da177e4SLinus Torvalds * @page: the page which is affected 14771da177e4SLinus Torvalds * @offset: the index of the truncation point 14781da177e4SLinus Torvalds * 14791da177e4SLinus Torvalds * block_invalidatepage() is called when all or part of the page has become 14801da177e4SLinus Torvalds * invalidatedby a truncate operation. 14811da177e4SLinus Torvalds * 14821da177e4SLinus Torvalds * block_invalidatepage() does not have to release all buffers, but it must 14831da177e4SLinus Torvalds * ensure that no dirty buffer is left outside @offset and that no I/O 14841da177e4SLinus Torvalds * is underway against any of the blocks which are outside the truncation 14851da177e4SLinus Torvalds * point. Because the caller is about to free (and possibly reuse) those 14861da177e4SLinus Torvalds * blocks on-disk. 14871da177e4SLinus Torvalds */ 14882ff28e22SNeilBrown void block_invalidatepage(struct page *page, unsigned long offset) 14891da177e4SLinus Torvalds { 14901da177e4SLinus Torvalds struct buffer_head *head, *bh, *next; 14911da177e4SLinus Torvalds unsigned int curr_off = 0; 14921da177e4SLinus Torvalds 14931da177e4SLinus Torvalds BUG_ON(!PageLocked(page)); 14941da177e4SLinus Torvalds if (!page_has_buffers(page)) 14951da177e4SLinus Torvalds goto out; 14961da177e4SLinus Torvalds 14971da177e4SLinus Torvalds head = page_buffers(page); 14981da177e4SLinus Torvalds bh = head; 14991da177e4SLinus Torvalds do { 15001da177e4SLinus Torvalds unsigned int next_off = curr_off + bh->b_size; 15011da177e4SLinus Torvalds next = bh->b_this_page; 15021da177e4SLinus Torvalds 15031da177e4SLinus Torvalds /* 15041da177e4SLinus Torvalds * is this block fully invalidated? 15051da177e4SLinus Torvalds */ 15061da177e4SLinus Torvalds if (offset <= curr_off) 15071da177e4SLinus Torvalds discard_buffer(bh); 15081da177e4SLinus Torvalds curr_off = next_off; 15091da177e4SLinus Torvalds bh = next; 15101da177e4SLinus Torvalds } while (bh != head); 15111da177e4SLinus Torvalds 15121da177e4SLinus Torvalds /* 15131da177e4SLinus Torvalds * We release buffers only if the entire page is being invalidated. 15141da177e4SLinus Torvalds * The get_block cached value has been unconditionally invalidated, 15151da177e4SLinus Torvalds * so real IO is not possible anymore. 15161da177e4SLinus Torvalds */ 15171da177e4SLinus Torvalds if (offset == 0) 15182ff28e22SNeilBrown try_to_release_page(page, 0); 15191da177e4SLinus Torvalds out: 15202ff28e22SNeilBrown return; 15211da177e4SLinus Torvalds } 15221da177e4SLinus Torvalds EXPORT_SYMBOL(block_invalidatepage); 15231da177e4SLinus Torvalds 15241da177e4SLinus Torvalds /* 15251da177e4SLinus Torvalds * We attach and possibly dirty the buffers atomically wrt 15261da177e4SLinus Torvalds * __set_page_dirty_buffers() via private_lock. try_to_free_buffers 15271da177e4SLinus Torvalds * is already excluded via the page lock. 15281da177e4SLinus Torvalds */ 15291da177e4SLinus Torvalds void create_empty_buffers(struct page *page, 15301da177e4SLinus Torvalds unsigned long blocksize, unsigned long b_state) 15311da177e4SLinus Torvalds { 15321da177e4SLinus Torvalds struct buffer_head *bh, *head, *tail; 15331da177e4SLinus Torvalds 15341da177e4SLinus Torvalds head = alloc_page_buffers(page, blocksize, 1); 15351da177e4SLinus Torvalds bh = head; 15361da177e4SLinus Torvalds do { 15371da177e4SLinus Torvalds bh->b_state |= b_state; 15381da177e4SLinus Torvalds tail = bh; 15391da177e4SLinus Torvalds bh = bh->b_this_page; 15401da177e4SLinus Torvalds } while (bh); 15411da177e4SLinus Torvalds tail->b_this_page = head; 15421da177e4SLinus Torvalds 15431da177e4SLinus Torvalds spin_lock(&page->mapping->private_lock); 15441da177e4SLinus Torvalds if (PageUptodate(page) || PageDirty(page)) { 15451da177e4SLinus Torvalds bh = head; 15461da177e4SLinus Torvalds do { 15471da177e4SLinus Torvalds if (PageDirty(page)) 15481da177e4SLinus Torvalds set_buffer_dirty(bh); 15491da177e4SLinus Torvalds if (PageUptodate(page)) 15501da177e4SLinus Torvalds set_buffer_uptodate(bh); 15511da177e4SLinus Torvalds bh = bh->b_this_page; 15521da177e4SLinus Torvalds } while (bh != head); 15531da177e4SLinus Torvalds } 15541da177e4SLinus Torvalds attach_page_buffers(page, head); 15551da177e4SLinus Torvalds spin_unlock(&page->mapping->private_lock); 15561da177e4SLinus Torvalds } 15571da177e4SLinus Torvalds EXPORT_SYMBOL(create_empty_buffers); 15581da177e4SLinus Torvalds 15591da177e4SLinus Torvalds /* 15601da177e4SLinus Torvalds * We are taking a block for data and we don't want any output from any 15611da177e4SLinus Torvalds * buffer-cache aliases starting from return from that function and 15621da177e4SLinus Torvalds * until the moment when something will explicitly mark the buffer 15631da177e4SLinus Torvalds * dirty (hopefully that will not happen until we will free that block ;-) 15641da177e4SLinus Torvalds * We don't even need to mark it not-uptodate - nobody can expect 15651da177e4SLinus Torvalds * anything from a newly allocated buffer anyway. We used to used 15661da177e4SLinus Torvalds * unmap_buffer() for such invalidation, but that was wrong. We definitely 15671da177e4SLinus Torvalds * don't want to mark the alias unmapped, for example - it would confuse 15681da177e4SLinus Torvalds * anyone who might pick it with bread() afterwards... 15691da177e4SLinus Torvalds * 15701da177e4SLinus Torvalds * Also.. Note that bforget() doesn't lock the buffer. So there can 15711da177e4SLinus Torvalds * be writeout I/O going on against recently-freed buffers. We don't 15721da177e4SLinus Torvalds * wait on that I/O in bforget() - it's more efficient to wait on the I/O 15731da177e4SLinus Torvalds * only if we really need to. That happens here. 15741da177e4SLinus Torvalds */ 15751da177e4SLinus Torvalds void unmap_underlying_metadata(struct block_device *bdev, sector_t block) 15761da177e4SLinus Torvalds { 15771da177e4SLinus Torvalds struct buffer_head *old_bh; 15781da177e4SLinus Torvalds 15791da177e4SLinus Torvalds might_sleep(); 15801da177e4SLinus Torvalds 1581385fd4c5SCoywolf Qi Hunt old_bh = __find_get_block_slow(bdev, block); 15821da177e4SLinus Torvalds if (old_bh) { 15831da177e4SLinus Torvalds clear_buffer_dirty(old_bh); 15841da177e4SLinus Torvalds wait_on_buffer(old_bh); 15851da177e4SLinus Torvalds clear_buffer_req(old_bh); 15861da177e4SLinus Torvalds __brelse(old_bh); 15871da177e4SLinus Torvalds } 15881da177e4SLinus Torvalds } 15891da177e4SLinus Torvalds EXPORT_SYMBOL(unmap_underlying_metadata); 15901da177e4SLinus Torvalds 15911da177e4SLinus Torvalds /* 15921da177e4SLinus Torvalds * NOTE! All mapped/uptodate combinations are valid: 15931da177e4SLinus Torvalds * 15941da177e4SLinus Torvalds * Mapped Uptodate Meaning 15951da177e4SLinus Torvalds * 15961da177e4SLinus Torvalds * No No "unknown" - must do get_block() 15971da177e4SLinus Torvalds * No Yes "hole" - zero-filled 15981da177e4SLinus Torvalds * Yes No "allocated" - allocated on disk, not read in 15991da177e4SLinus Torvalds * Yes Yes "valid" - allocated and up-to-date in memory. 16001da177e4SLinus Torvalds * 16011da177e4SLinus Torvalds * "Dirty" is valid only with the last case (mapped+uptodate). 16021da177e4SLinus Torvalds */ 16031da177e4SLinus Torvalds 16041da177e4SLinus Torvalds /* 16051da177e4SLinus Torvalds * While block_write_full_page is writing back the dirty buffers under 16061da177e4SLinus Torvalds * the page lock, whoever dirtied the buffers may decide to clean them 16071da177e4SLinus Torvalds * again at any time. We handle that by only looking at the buffer 16081da177e4SLinus Torvalds * state inside lock_buffer(). 16091da177e4SLinus Torvalds * 16101da177e4SLinus Torvalds * If block_write_full_page() is called for regular writeback 16111da177e4SLinus Torvalds * (wbc->sync_mode == WB_SYNC_NONE) then it will redirty a page which has a 16121da177e4SLinus Torvalds * locked buffer. This only can happen if someone has written the buffer 16131da177e4SLinus Torvalds * directly, with submit_bh(). At the address_space level PageWriteback 16141da177e4SLinus Torvalds * prevents this contention from occurring. 16156e34eeddSTheodore Ts'o * 16166e34eeddSTheodore Ts'o * If block_write_full_page() is called with wbc->sync_mode == 16176e34eeddSTheodore Ts'o * WB_SYNC_ALL, the writes are posted using WRITE_SYNC_PLUG; this 16186e34eeddSTheodore Ts'o * causes the writes to be flagged as synchronous writes, but the 16196e34eeddSTheodore Ts'o * block device queue will NOT be unplugged, since usually many pages 16206e34eeddSTheodore Ts'o * will be pushed to the out before the higher-level caller actually 16216e34eeddSTheodore Ts'o * waits for the writes to be completed. The various wait functions, 16226e34eeddSTheodore Ts'o * such as wait_on_writeback_range() will ultimately call sync_page() 16236e34eeddSTheodore Ts'o * which will ultimately call blk_run_backing_dev(), which will end up 16246e34eeddSTheodore Ts'o * unplugging the device queue. 16251da177e4SLinus Torvalds */ 16261da177e4SLinus Torvalds static int __block_write_full_page(struct inode *inode, struct page *page, 162735c80d5fSChris Mason get_block_t *get_block, struct writeback_control *wbc, 162835c80d5fSChris Mason bh_end_io_t *handler) 16291da177e4SLinus Torvalds { 16301da177e4SLinus Torvalds int err; 16311da177e4SLinus Torvalds sector_t block; 16321da177e4SLinus Torvalds sector_t last_block; 1633f0fbd5fcSAndrew Morton struct buffer_head *bh, *head; 1634b0cf2321SBadari Pulavarty const unsigned blocksize = 1 << inode->i_blkbits; 16351da177e4SLinus Torvalds int nr_underway = 0; 16366e34eeddSTheodore Ts'o int write_op = (wbc->sync_mode == WB_SYNC_ALL ? 16376e34eeddSTheodore Ts'o WRITE_SYNC_PLUG : WRITE); 16381da177e4SLinus Torvalds 16391da177e4SLinus Torvalds BUG_ON(!PageLocked(page)); 16401da177e4SLinus Torvalds 16411da177e4SLinus Torvalds last_block = (i_size_read(inode) - 1) >> inode->i_blkbits; 16421da177e4SLinus Torvalds 16431da177e4SLinus Torvalds if (!page_has_buffers(page)) { 1644b0cf2321SBadari Pulavarty create_empty_buffers(page, blocksize, 16451da177e4SLinus Torvalds (1 << BH_Dirty)|(1 << BH_Uptodate)); 16461da177e4SLinus Torvalds } 16471da177e4SLinus Torvalds 16481da177e4SLinus Torvalds /* 16491da177e4SLinus Torvalds * Be very careful. We have no exclusion from __set_page_dirty_buffers 16501da177e4SLinus Torvalds * here, and the (potentially unmapped) buffers may become dirty at 16511da177e4SLinus Torvalds * any time. If a buffer becomes dirty here after we've inspected it 16521da177e4SLinus Torvalds * then we just miss that fact, and the page stays dirty. 16531da177e4SLinus Torvalds * 16541da177e4SLinus Torvalds * Buffers outside i_size may be dirtied by __set_page_dirty_buffers; 16551da177e4SLinus Torvalds * handle that here by just cleaning them. 16561da177e4SLinus Torvalds */ 16571da177e4SLinus Torvalds 165854b21a79SAndrew Morton block = (sector_t)page->index << (PAGE_CACHE_SHIFT - inode->i_blkbits); 16591da177e4SLinus Torvalds head = page_buffers(page); 16601da177e4SLinus Torvalds bh = head; 16611da177e4SLinus Torvalds 16621da177e4SLinus Torvalds /* 16631da177e4SLinus Torvalds * Get all the dirty buffers mapped to disk addresses and 16641da177e4SLinus Torvalds * handle any aliases from the underlying blockdev's mapping. 16651da177e4SLinus Torvalds */ 16661da177e4SLinus Torvalds do { 16671da177e4SLinus Torvalds if (block > last_block) { 16681da177e4SLinus Torvalds /* 16691da177e4SLinus Torvalds * mapped buffers outside i_size will occur, because 16701da177e4SLinus Torvalds * this page can be outside i_size when there is a 16711da177e4SLinus Torvalds * truncate in progress. 16721da177e4SLinus Torvalds */ 16731da177e4SLinus Torvalds /* 16741da177e4SLinus Torvalds * The buffer was zeroed by block_write_full_page() 16751da177e4SLinus Torvalds */ 16761da177e4SLinus Torvalds clear_buffer_dirty(bh); 16771da177e4SLinus Torvalds set_buffer_uptodate(bh); 167829a814d2SAlex Tomas } else if ((!buffer_mapped(bh) || buffer_delay(bh)) && 167929a814d2SAlex Tomas buffer_dirty(bh)) { 1680b0cf2321SBadari Pulavarty WARN_ON(bh->b_size != blocksize); 16811da177e4SLinus Torvalds err = get_block(inode, block, bh, 1); 16821da177e4SLinus Torvalds if (err) 16831da177e4SLinus Torvalds goto recover; 168429a814d2SAlex Tomas clear_buffer_delay(bh); 16851da177e4SLinus Torvalds if (buffer_new(bh)) { 16861da177e4SLinus Torvalds /* blockdev mappings never come here */ 16871da177e4SLinus Torvalds clear_buffer_new(bh); 16881da177e4SLinus Torvalds unmap_underlying_metadata(bh->b_bdev, 16891da177e4SLinus Torvalds bh->b_blocknr); 16901da177e4SLinus Torvalds } 16911da177e4SLinus Torvalds } 16921da177e4SLinus Torvalds bh = bh->b_this_page; 16931da177e4SLinus Torvalds block++; 16941da177e4SLinus Torvalds } while (bh != head); 16951da177e4SLinus Torvalds 16961da177e4SLinus Torvalds do { 16971da177e4SLinus Torvalds if (!buffer_mapped(bh)) 16981da177e4SLinus Torvalds continue; 16991da177e4SLinus Torvalds /* 17001da177e4SLinus Torvalds * If it's a fully non-blocking write attempt and we cannot 17011da177e4SLinus Torvalds * lock the buffer then redirty the page. Note that this can 17021da177e4SLinus Torvalds * potentially cause a busy-wait loop from pdflush and kswapd 17031da177e4SLinus Torvalds * activity, but those code paths have their own higher-level 17041da177e4SLinus Torvalds * throttling. 17051da177e4SLinus Torvalds */ 17061da177e4SLinus Torvalds if (wbc->sync_mode != WB_SYNC_NONE || !wbc->nonblocking) { 17071da177e4SLinus Torvalds lock_buffer(bh); 1708ca5de404SNick Piggin } else if (!trylock_buffer(bh)) { 17091da177e4SLinus Torvalds redirty_page_for_writepage(wbc, page); 17101da177e4SLinus Torvalds continue; 17111da177e4SLinus Torvalds } 17121da177e4SLinus Torvalds if (test_clear_buffer_dirty(bh)) { 171335c80d5fSChris Mason mark_buffer_async_write_endio(bh, handler); 17141da177e4SLinus Torvalds } else { 17151da177e4SLinus Torvalds unlock_buffer(bh); 17161da177e4SLinus Torvalds } 17171da177e4SLinus Torvalds } while ((bh = bh->b_this_page) != head); 17181da177e4SLinus Torvalds 17191da177e4SLinus Torvalds /* 17201da177e4SLinus Torvalds * The page and its buffers are protected by PageWriteback(), so we can 17211da177e4SLinus Torvalds * drop the bh refcounts early. 17221da177e4SLinus Torvalds */ 17231da177e4SLinus Torvalds BUG_ON(PageWriteback(page)); 17241da177e4SLinus Torvalds set_page_writeback(page); 17251da177e4SLinus Torvalds 17261da177e4SLinus Torvalds do { 17271da177e4SLinus Torvalds struct buffer_head *next = bh->b_this_page; 17281da177e4SLinus Torvalds if (buffer_async_write(bh)) { 1729a64c8610STheodore Ts'o submit_bh(write_op, bh); 17301da177e4SLinus Torvalds nr_underway++; 1731ad576e63SNick Piggin } 17321da177e4SLinus Torvalds bh = next; 17331da177e4SLinus Torvalds } while (bh != head); 173405937baaSAndrew Morton unlock_page(page); 17351da177e4SLinus Torvalds 17361da177e4SLinus Torvalds err = 0; 17371da177e4SLinus Torvalds done: 17381da177e4SLinus Torvalds if (nr_underway == 0) { 17391da177e4SLinus Torvalds /* 17401da177e4SLinus Torvalds * The page was marked dirty, but the buffers were 17411da177e4SLinus Torvalds * clean. Someone wrote them back by hand with 17421da177e4SLinus Torvalds * ll_rw_block/submit_bh. A rare case. 17431da177e4SLinus Torvalds */ 17441da177e4SLinus Torvalds end_page_writeback(page); 17453d67f2d7SNick Piggin 17461da177e4SLinus Torvalds /* 17471da177e4SLinus Torvalds * The page and buffer_heads can be released at any time from 17481da177e4SLinus Torvalds * here on. 17491da177e4SLinus Torvalds */ 17501da177e4SLinus Torvalds } 17511da177e4SLinus Torvalds return err; 17521da177e4SLinus Torvalds 17531da177e4SLinus Torvalds recover: 17541da177e4SLinus Torvalds /* 17551da177e4SLinus Torvalds * ENOSPC, or some other error. We may already have added some 17561da177e4SLinus Torvalds * blocks to the file, so we need to write these out to avoid 17571da177e4SLinus Torvalds * exposing stale data. 17581da177e4SLinus Torvalds * The page is currently locked and not marked for writeback 17591da177e4SLinus Torvalds */ 17601da177e4SLinus Torvalds bh = head; 17611da177e4SLinus Torvalds /* Recovery: lock and submit the mapped buffers */ 17621da177e4SLinus Torvalds do { 176329a814d2SAlex Tomas if (buffer_mapped(bh) && buffer_dirty(bh) && 176429a814d2SAlex Tomas !buffer_delay(bh)) { 17651da177e4SLinus Torvalds lock_buffer(bh); 176635c80d5fSChris Mason mark_buffer_async_write_endio(bh, handler); 17671da177e4SLinus Torvalds } else { 17681da177e4SLinus Torvalds /* 17691da177e4SLinus Torvalds * The buffer may have been set dirty during 17701da177e4SLinus Torvalds * attachment to a dirty page. 17711da177e4SLinus Torvalds */ 17721da177e4SLinus Torvalds clear_buffer_dirty(bh); 17731da177e4SLinus Torvalds } 17741da177e4SLinus Torvalds } while ((bh = bh->b_this_page) != head); 17751da177e4SLinus Torvalds SetPageError(page); 17761da177e4SLinus Torvalds BUG_ON(PageWriteback(page)); 17777e4c3690SAndrew Morton mapping_set_error(page->mapping, err); 17781da177e4SLinus Torvalds set_page_writeback(page); 17791da177e4SLinus Torvalds do { 17801da177e4SLinus Torvalds struct buffer_head *next = bh->b_this_page; 17811da177e4SLinus Torvalds if (buffer_async_write(bh)) { 17821da177e4SLinus Torvalds clear_buffer_dirty(bh); 1783a64c8610STheodore Ts'o submit_bh(write_op, bh); 17841da177e4SLinus Torvalds nr_underway++; 1785ad576e63SNick Piggin } 17861da177e4SLinus Torvalds bh = next; 17871da177e4SLinus Torvalds } while (bh != head); 1788ffda9d30SNick Piggin unlock_page(page); 17891da177e4SLinus Torvalds goto done; 17901da177e4SLinus Torvalds } 17911da177e4SLinus Torvalds 1792afddba49SNick Piggin /* 1793afddba49SNick Piggin * If a page has any new buffers, zero them out here, and mark them uptodate 1794afddba49SNick Piggin * and dirty so they'll be written out (in order to prevent uninitialised 1795afddba49SNick Piggin * block data from leaking). And clear the new bit. 1796afddba49SNick Piggin */ 1797afddba49SNick Piggin void page_zero_new_buffers(struct page *page, unsigned from, unsigned to) 1798afddba49SNick Piggin { 1799afddba49SNick Piggin unsigned int block_start, block_end; 1800afddba49SNick Piggin struct buffer_head *head, *bh; 1801afddba49SNick Piggin 1802afddba49SNick Piggin BUG_ON(!PageLocked(page)); 1803afddba49SNick Piggin if (!page_has_buffers(page)) 1804afddba49SNick Piggin return; 1805afddba49SNick Piggin 1806afddba49SNick Piggin bh = head = page_buffers(page); 1807afddba49SNick Piggin block_start = 0; 1808afddba49SNick Piggin do { 1809afddba49SNick Piggin block_end = block_start + bh->b_size; 1810afddba49SNick Piggin 1811afddba49SNick Piggin if (buffer_new(bh)) { 1812afddba49SNick Piggin if (block_end > from && block_start < to) { 1813afddba49SNick Piggin if (!PageUptodate(page)) { 1814afddba49SNick Piggin unsigned start, size; 1815afddba49SNick Piggin 1816afddba49SNick Piggin start = max(from, block_start); 1817afddba49SNick Piggin size = min(to, block_end) - start; 1818afddba49SNick Piggin 1819eebd2aa3SChristoph Lameter zero_user(page, start, size); 1820afddba49SNick Piggin set_buffer_uptodate(bh); 1821afddba49SNick Piggin } 1822afddba49SNick Piggin 1823afddba49SNick Piggin clear_buffer_new(bh); 1824afddba49SNick Piggin mark_buffer_dirty(bh); 1825afddba49SNick Piggin } 1826afddba49SNick Piggin } 1827afddba49SNick Piggin 1828afddba49SNick Piggin block_start = block_end; 1829afddba49SNick Piggin bh = bh->b_this_page; 1830afddba49SNick Piggin } while (bh != head); 1831afddba49SNick Piggin } 1832afddba49SNick Piggin EXPORT_SYMBOL(page_zero_new_buffers); 1833afddba49SNick Piggin 18341da177e4SLinus Torvalds static int __block_prepare_write(struct inode *inode, struct page *page, 18351da177e4SLinus Torvalds unsigned from, unsigned to, get_block_t *get_block) 18361da177e4SLinus Torvalds { 18371da177e4SLinus Torvalds unsigned block_start, block_end; 18381da177e4SLinus Torvalds sector_t block; 18391da177e4SLinus Torvalds int err = 0; 18401da177e4SLinus Torvalds unsigned blocksize, bbits; 18411da177e4SLinus Torvalds struct buffer_head *bh, *head, *wait[2], **wait_bh=wait; 18421da177e4SLinus Torvalds 18431da177e4SLinus Torvalds BUG_ON(!PageLocked(page)); 18441da177e4SLinus Torvalds BUG_ON(from > PAGE_CACHE_SIZE); 18451da177e4SLinus Torvalds BUG_ON(to > PAGE_CACHE_SIZE); 18461da177e4SLinus Torvalds BUG_ON(from > to); 18471da177e4SLinus Torvalds 18481da177e4SLinus Torvalds blocksize = 1 << inode->i_blkbits; 18491da177e4SLinus Torvalds if (!page_has_buffers(page)) 18501da177e4SLinus Torvalds create_empty_buffers(page, blocksize, 0); 18511da177e4SLinus Torvalds head = page_buffers(page); 18521da177e4SLinus Torvalds 18531da177e4SLinus Torvalds bbits = inode->i_blkbits; 18541da177e4SLinus Torvalds block = (sector_t)page->index << (PAGE_CACHE_SHIFT - bbits); 18551da177e4SLinus Torvalds 18561da177e4SLinus Torvalds for(bh = head, block_start = 0; bh != head || !block_start; 18571da177e4SLinus Torvalds block++, block_start=block_end, bh = bh->b_this_page) { 18581da177e4SLinus Torvalds block_end = block_start + blocksize; 18591da177e4SLinus Torvalds if (block_end <= from || block_start >= to) { 18601da177e4SLinus Torvalds if (PageUptodate(page)) { 18611da177e4SLinus Torvalds if (!buffer_uptodate(bh)) 18621da177e4SLinus Torvalds set_buffer_uptodate(bh); 18631da177e4SLinus Torvalds } 18641da177e4SLinus Torvalds continue; 18651da177e4SLinus Torvalds } 18661da177e4SLinus Torvalds if (buffer_new(bh)) 18671da177e4SLinus Torvalds clear_buffer_new(bh); 18681da177e4SLinus Torvalds if (!buffer_mapped(bh)) { 1869b0cf2321SBadari Pulavarty WARN_ON(bh->b_size != blocksize); 18701da177e4SLinus Torvalds err = get_block(inode, block, bh, 1); 18711da177e4SLinus Torvalds if (err) 1872f3ddbdc6SNick Piggin break; 18731da177e4SLinus Torvalds if (buffer_new(bh)) { 18741da177e4SLinus Torvalds unmap_underlying_metadata(bh->b_bdev, 18751da177e4SLinus Torvalds bh->b_blocknr); 18761da177e4SLinus Torvalds if (PageUptodate(page)) { 1877637aff46SNick Piggin clear_buffer_new(bh); 18781da177e4SLinus Torvalds set_buffer_uptodate(bh); 1879637aff46SNick Piggin mark_buffer_dirty(bh); 18801da177e4SLinus Torvalds continue; 18811da177e4SLinus Torvalds } 1882eebd2aa3SChristoph Lameter if (block_end > to || block_start < from) 1883eebd2aa3SChristoph Lameter zero_user_segments(page, 1884eebd2aa3SChristoph Lameter to, block_end, 1885eebd2aa3SChristoph Lameter block_start, from); 18861da177e4SLinus Torvalds continue; 18871da177e4SLinus Torvalds } 18881da177e4SLinus Torvalds } 18891da177e4SLinus Torvalds if (PageUptodate(page)) { 18901da177e4SLinus Torvalds if (!buffer_uptodate(bh)) 18911da177e4SLinus Torvalds set_buffer_uptodate(bh); 18921da177e4SLinus Torvalds continue; 18931da177e4SLinus Torvalds } 18941da177e4SLinus Torvalds if (!buffer_uptodate(bh) && !buffer_delay(bh) && 189533a266ddSDavid Chinner !buffer_unwritten(bh) && 18961da177e4SLinus Torvalds (block_start < from || block_end > to)) { 18971da177e4SLinus Torvalds ll_rw_block(READ, 1, &bh); 18981da177e4SLinus Torvalds *wait_bh++=bh; 18991da177e4SLinus Torvalds } 19001da177e4SLinus Torvalds } 19011da177e4SLinus Torvalds /* 19021da177e4SLinus Torvalds * If we issued read requests - let them complete. 19031da177e4SLinus Torvalds */ 19041da177e4SLinus Torvalds while(wait_bh > wait) { 19051da177e4SLinus Torvalds wait_on_buffer(*--wait_bh); 19061da177e4SLinus Torvalds if (!buffer_uptodate(*wait_bh)) 1907f3ddbdc6SNick Piggin err = -EIO; 19081da177e4SLinus Torvalds } 1909afddba49SNick Piggin if (unlikely(err)) 1910afddba49SNick Piggin page_zero_new_buffers(page, from, to); 19111da177e4SLinus Torvalds return err; 19121da177e4SLinus Torvalds } 19131da177e4SLinus Torvalds 19141da177e4SLinus Torvalds static int __block_commit_write(struct inode *inode, struct page *page, 19151da177e4SLinus Torvalds unsigned from, unsigned to) 19161da177e4SLinus Torvalds { 19171da177e4SLinus Torvalds unsigned block_start, block_end; 19181da177e4SLinus Torvalds int partial = 0; 19191da177e4SLinus Torvalds unsigned blocksize; 19201da177e4SLinus Torvalds struct buffer_head *bh, *head; 19211da177e4SLinus Torvalds 19221da177e4SLinus Torvalds blocksize = 1 << inode->i_blkbits; 19231da177e4SLinus Torvalds 19241da177e4SLinus Torvalds for(bh = head = page_buffers(page), block_start = 0; 19251da177e4SLinus Torvalds bh != head || !block_start; 19261da177e4SLinus Torvalds block_start=block_end, bh = bh->b_this_page) { 19271da177e4SLinus Torvalds block_end = block_start + blocksize; 19281da177e4SLinus Torvalds if (block_end <= from || block_start >= to) { 19291da177e4SLinus Torvalds if (!buffer_uptodate(bh)) 19301da177e4SLinus Torvalds partial = 1; 19311da177e4SLinus Torvalds } else { 19321da177e4SLinus Torvalds set_buffer_uptodate(bh); 19331da177e4SLinus Torvalds mark_buffer_dirty(bh); 19341da177e4SLinus Torvalds } 1935afddba49SNick Piggin clear_buffer_new(bh); 19361da177e4SLinus Torvalds } 19371da177e4SLinus Torvalds 19381da177e4SLinus Torvalds /* 19391da177e4SLinus Torvalds * If this is a partial write which happened to make all buffers 19401da177e4SLinus Torvalds * uptodate then we can optimize away a bogus readpage() for 19411da177e4SLinus Torvalds * the next read(). Here we 'discover' whether the page went 19421da177e4SLinus Torvalds * uptodate as a result of this (potentially partial) write. 19431da177e4SLinus Torvalds */ 19441da177e4SLinus Torvalds if (!partial) 19451da177e4SLinus Torvalds SetPageUptodate(page); 19461da177e4SLinus Torvalds return 0; 19471da177e4SLinus Torvalds } 19481da177e4SLinus Torvalds 19491da177e4SLinus Torvalds /* 1950afddba49SNick Piggin * block_write_begin takes care of the basic task of block allocation and 1951afddba49SNick Piggin * bringing partial write blocks uptodate first. 1952afddba49SNick Piggin * 1953afddba49SNick Piggin * If *pagep is not NULL, then block_write_begin uses the locked page 1954afddba49SNick Piggin * at *pagep rather than allocating its own. In this case, the page will 1955afddba49SNick Piggin * not be unlocked or deallocated on failure. 1956afddba49SNick Piggin */ 1957afddba49SNick Piggin int block_write_begin(struct file *file, struct address_space *mapping, 1958afddba49SNick Piggin loff_t pos, unsigned len, unsigned flags, 1959afddba49SNick Piggin struct page **pagep, void **fsdata, 1960afddba49SNick Piggin get_block_t *get_block) 1961afddba49SNick Piggin { 1962afddba49SNick Piggin struct inode *inode = mapping->host; 1963afddba49SNick Piggin int status = 0; 1964afddba49SNick Piggin struct page *page; 1965afddba49SNick Piggin pgoff_t index; 1966afddba49SNick Piggin unsigned start, end; 1967afddba49SNick Piggin int ownpage = 0; 1968afddba49SNick Piggin 1969afddba49SNick Piggin index = pos >> PAGE_CACHE_SHIFT; 1970afddba49SNick Piggin start = pos & (PAGE_CACHE_SIZE - 1); 1971afddba49SNick Piggin end = start + len; 1972afddba49SNick Piggin 1973afddba49SNick Piggin page = *pagep; 1974afddba49SNick Piggin if (page == NULL) { 1975afddba49SNick Piggin ownpage = 1; 197654566b2cSNick Piggin page = grab_cache_page_write_begin(mapping, index, flags); 1977afddba49SNick Piggin if (!page) { 1978afddba49SNick Piggin status = -ENOMEM; 1979afddba49SNick Piggin goto out; 1980afddba49SNick Piggin } 1981afddba49SNick Piggin *pagep = page; 1982afddba49SNick Piggin } else 1983afddba49SNick Piggin BUG_ON(!PageLocked(page)); 1984afddba49SNick Piggin 1985afddba49SNick Piggin status = __block_prepare_write(inode, page, start, end, get_block); 1986afddba49SNick Piggin if (unlikely(status)) { 1987afddba49SNick Piggin ClearPageUptodate(page); 1988afddba49SNick Piggin 1989afddba49SNick Piggin if (ownpage) { 1990afddba49SNick Piggin unlock_page(page); 1991afddba49SNick Piggin page_cache_release(page); 1992afddba49SNick Piggin *pagep = NULL; 1993afddba49SNick Piggin 1994afddba49SNick Piggin /* 1995afddba49SNick Piggin * prepare_write() may have instantiated a few blocks 1996afddba49SNick Piggin * outside i_size. Trim these off again. Don't need 1997afddba49SNick Piggin * i_size_read because we hold i_mutex. 1998afddba49SNick Piggin */ 1999afddba49SNick Piggin if (pos + len > inode->i_size) 2000afddba49SNick Piggin vmtruncate(inode, inode->i_size); 2001afddba49SNick Piggin } 2002afddba49SNick Piggin } 2003afddba49SNick Piggin 2004afddba49SNick Piggin out: 2005afddba49SNick Piggin return status; 2006afddba49SNick Piggin } 2007afddba49SNick Piggin EXPORT_SYMBOL(block_write_begin); 2008afddba49SNick Piggin 2009afddba49SNick Piggin int block_write_end(struct file *file, struct address_space *mapping, 2010afddba49SNick Piggin loff_t pos, unsigned len, unsigned copied, 2011afddba49SNick Piggin struct page *page, void *fsdata) 2012afddba49SNick Piggin { 2013afddba49SNick Piggin struct inode *inode = mapping->host; 2014afddba49SNick Piggin unsigned start; 2015afddba49SNick Piggin 2016afddba49SNick Piggin start = pos & (PAGE_CACHE_SIZE - 1); 2017afddba49SNick Piggin 2018afddba49SNick Piggin if (unlikely(copied < len)) { 2019afddba49SNick Piggin /* 2020afddba49SNick Piggin * The buffers that were written will now be uptodate, so we 2021afddba49SNick Piggin * don't have to worry about a readpage reading them and 2022afddba49SNick Piggin * overwriting a partial write. However if we have encountered 2023afddba49SNick Piggin * a short write and only partially written into a buffer, it 2024afddba49SNick Piggin * will not be marked uptodate, so a readpage might come in and 2025afddba49SNick Piggin * destroy our partial write. 2026afddba49SNick Piggin * 2027afddba49SNick Piggin * Do the simplest thing, and just treat any short write to a 2028afddba49SNick Piggin * non uptodate page as a zero-length write, and force the 2029afddba49SNick Piggin * caller to redo the whole thing. 2030afddba49SNick Piggin */ 2031afddba49SNick Piggin if (!PageUptodate(page)) 2032afddba49SNick Piggin copied = 0; 2033afddba49SNick Piggin 2034afddba49SNick Piggin page_zero_new_buffers(page, start+copied, start+len); 2035afddba49SNick Piggin } 2036afddba49SNick Piggin flush_dcache_page(page); 2037afddba49SNick Piggin 2038afddba49SNick Piggin /* This could be a short (even 0-length) commit */ 2039afddba49SNick Piggin __block_commit_write(inode, page, start, start+copied); 2040afddba49SNick Piggin 2041afddba49SNick Piggin return copied; 2042afddba49SNick Piggin } 2043afddba49SNick Piggin EXPORT_SYMBOL(block_write_end); 2044afddba49SNick Piggin 2045afddba49SNick Piggin int generic_write_end(struct file *file, struct address_space *mapping, 2046afddba49SNick Piggin loff_t pos, unsigned len, unsigned copied, 2047afddba49SNick Piggin struct page *page, void *fsdata) 2048afddba49SNick Piggin { 2049afddba49SNick Piggin struct inode *inode = mapping->host; 2050c7d206b3SJan Kara int i_size_changed = 0; 2051afddba49SNick Piggin 2052afddba49SNick Piggin copied = block_write_end(file, mapping, pos, len, copied, page, fsdata); 2053afddba49SNick Piggin 2054afddba49SNick Piggin /* 2055afddba49SNick Piggin * No need to use i_size_read() here, the i_size 2056afddba49SNick Piggin * cannot change under us because we hold i_mutex. 2057afddba49SNick Piggin * 2058afddba49SNick Piggin * But it's important to update i_size while still holding page lock: 2059afddba49SNick Piggin * page writeout could otherwise come in and zero beyond i_size. 2060afddba49SNick Piggin */ 2061afddba49SNick Piggin if (pos+copied > inode->i_size) { 2062afddba49SNick Piggin i_size_write(inode, pos+copied); 2063c7d206b3SJan Kara i_size_changed = 1; 2064afddba49SNick Piggin } 2065afddba49SNick Piggin 2066afddba49SNick Piggin unlock_page(page); 2067afddba49SNick Piggin page_cache_release(page); 2068afddba49SNick Piggin 2069c7d206b3SJan Kara /* 2070c7d206b3SJan Kara * Don't mark the inode dirty under page lock. First, it unnecessarily 2071c7d206b3SJan Kara * makes the holding time of page lock longer. Second, it forces lock 2072c7d206b3SJan Kara * ordering of page lock and transaction start for journaling 2073c7d206b3SJan Kara * filesystems. 2074c7d206b3SJan Kara */ 2075c7d206b3SJan Kara if (i_size_changed) 2076c7d206b3SJan Kara mark_inode_dirty(inode); 2077c7d206b3SJan Kara 2078afddba49SNick Piggin return copied; 2079afddba49SNick Piggin } 2080afddba49SNick Piggin EXPORT_SYMBOL(generic_write_end); 2081afddba49SNick Piggin 2082afddba49SNick Piggin /* 20838ab22b9aSHisashi Hifumi * block_is_partially_uptodate checks whether buffers within a page are 20848ab22b9aSHisashi Hifumi * uptodate or not. 20858ab22b9aSHisashi Hifumi * 20868ab22b9aSHisashi Hifumi * Returns true if all buffers which correspond to a file portion 20878ab22b9aSHisashi Hifumi * we want to read are uptodate. 20888ab22b9aSHisashi Hifumi */ 20898ab22b9aSHisashi Hifumi int block_is_partially_uptodate(struct page *page, read_descriptor_t *desc, 20908ab22b9aSHisashi Hifumi unsigned long from) 20918ab22b9aSHisashi Hifumi { 20928ab22b9aSHisashi Hifumi struct inode *inode = page->mapping->host; 20938ab22b9aSHisashi Hifumi unsigned block_start, block_end, blocksize; 20948ab22b9aSHisashi Hifumi unsigned to; 20958ab22b9aSHisashi Hifumi struct buffer_head *bh, *head; 20968ab22b9aSHisashi Hifumi int ret = 1; 20978ab22b9aSHisashi Hifumi 20988ab22b9aSHisashi Hifumi if (!page_has_buffers(page)) 20998ab22b9aSHisashi Hifumi return 0; 21008ab22b9aSHisashi Hifumi 21018ab22b9aSHisashi Hifumi blocksize = 1 << inode->i_blkbits; 21028ab22b9aSHisashi Hifumi to = min_t(unsigned, PAGE_CACHE_SIZE - from, desc->count); 21038ab22b9aSHisashi Hifumi to = from + to; 21048ab22b9aSHisashi Hifumi if (from < blocksize && to > PAGE_CACHE_SIZE - blocksize) 21058ab22b9aSHisashi Hifumi return 0; 21068ab22b9aSHisashi Hifumi 21078ab22b9aSHisashi Hifumi head = page_buffers(page); 21088ab22b9aSHisashi Hifumi bh = head; 21098ab22b9aSHisashi Hifumi block_start = 0; 21108ab22b9aSHisashi Hifumi do { 21118ab22b9aSHisashi Hifumi block_end = block_start + blocksize; 21128ab22b9aSHisashi Hifumi if (block_end > from && block_start < to) { 21138ab22b9aSHisashi Hifumi if (!buffer_uptodate(bh)) { 21148ab22b9aSHisashi Hifumi ret = 0; 21158ab22b9aSHisashi Hifumi break; 21168ab22b9aSHisashi Hifumi } 21178ab22b9aSHisashi Hifumi if (block_end >= to) 21188ab22b9aSHisashi Hifumi break; 21198ab22b9aSHisashi Hifumi } 21208ab22b9aSHisashi Hifumi block_start = block_end; 21218ab22b9aSHisashi Hifumi bh = bh->b_this_page; 21228ab22b9aSHisashi Hifumi } while (bh != head); 21238ab22b9aSHisashi Hifumi 21248ab22b9aSHisashi Hifumi return ret; 21258ab22b9aSHisashi Hifumi } 21268ab22b9aSHisashi Hifumi EXPORT_SYMBOL(block_is_partially_uptodate); 21278ab22b9aSHisashi Hifumi 21288ab22b9aSHisashi Hifumi /* 21291da177e4SLinus Torvalds * Generic "read page" function for block devices that have the normal 21301da177e4SLinus Torvalds * get_block functionality. This is most of the block device filesystems. 21311da177e4SLinus Torvalds * Reads the page asynchronously --- the unlock_buffer() and 21321da177e4SLinus Torvalds * set/clear_buffer_uptodate() functions propagate buffer state into the 21331da177e4SLinus Torvalds * page struct once IO has completed. 21341da177e4SLinus Torvalds */ 21351da177e4SLinus Torvalds int block_read_full_page(struct page *page, get_block_t *get_block) 21361da177e4SLinus Torvalds { 21371da177e4SLinus Torvalds struct inode *inode = page->mapping->host; 21381da177e4SLinus Torvalds sector_t iblock, lblock; 21391da177e4SLinus Torvalds struct buffer_head *bh, *head, *arr[MAX_BUF_PER_PAGE]; 21401da177e4SLinus Torvalds unsigned int blocksize; 21411da177e4SLinus Torvalds int nr, i; 21421da177e4SLinus Torvalds int fully_mapped = 1; 21431da177e4SLinus Torvalds 2144cd7619d6SMatt Mackall BUG_ON(!PageLocked(page)); 21451da177e4SLinus Torvalds blocksize = 1 << inode->i_blkbits; 21461da177e4SLinus Torvalds if (!page_has_buffers(page)) 21471da177e4SLinus Torvalds create_empty_buffers(page, blocksize, 0); 21481da177e4SLinus Torvalds head = page_buffers(page); 21491da177e4SLinus Torvalds 21501da177e4SLinus Torvalds iblock = (sector_t)page->index << (PAGE_CACHE_SHIFT - inode->i_blkbits); 21511da177e4SLinus Torvalds lblock = (i_size_read(inode)+blocksize-1) >> inode->i_blkbits; 21521da177e4SLinus Torvalds bh = head; 21531da177e4SLinus Torvalds nr = 0; 21541da177e4SLinus Torvalds i = 0; 21551da177e4SLinus Torvalds 21561da177e4SLinus Torvalds do { 21571da177e4SLinus Torvalds if (buffer_uptodate(bh)) 21581da177e4SLinus Torvalds continue; 21591da177e4SLinus Torvalds 21601da177e4SLinus Torvalds if (!buffer_mapped(bh)) { 2161c64610baSAndrew Morton int err = 0; 2162c64610baSAndrew Morton 21631da177e4SLinus Torvalds fully_mapped = 0; 21641da177e4SLinus Torvalds if (iblock < lblock) { 2165b0cf2321SBadari Pulavarty WARN_ON(bh->b_size != blocksize); 2166c64610baSAndrew Morton err = get_block(inode, iblock, bh, 0); 2167c64610baSAndrew Morton if (err) 21681da177e4SLinus Torvalds SetPageError(page); 21691da177e4SLinus Torvalds } 21701da177e4SLinus Torvalds if (!buffer_mapped(bh)) { 2171eebd2aa3SChristoph Lameter zero_user(page, i * blocksize, blocksize); 2172c64610baSAndrew Morton if (!err) 21731da177e4SLinus Torvalds set_buffer_uptodate(bh); 21741da177e4SLinus Torvalds continue; 21751da177e4SLinus Torvalds } 21761da177e4SLinus Torvalds /* 21771da177e4SLinus Torvalds * get_block() might have updated the buffer 21781da177e4SLinus Torvalds * synchronously 21791da177e4SLinus Torvalds */ 21801da177e4SLinus Torvalds if (buffer_uptodate(bh)) 21811da177e4SLinus Torvalds continue; 21821da177e4SLinus Torvalds } 21831da177e4SLinus Torvalds arr[nr++] = bh; 21841da177e4SLinus Torvalds } while (i++, iblock++, (bh = bh->b_this_page) != head); 21851da177e4SLinus Torvalds 21861da177e4SLinus Torvalds if (fully_mapped) 21871da177e4SLinus Torvalds SetPageMappedToDisk(page); 21881da177e4SLinus Torvalds 21891da177e4SLinus Torvalds if (!nr) { 21901da177e4SLinus Torvalds /* 21911da177e4SLinus Torvalds * All buffers are uptodate - we can set the page uptodate 21921da177e4SLinus Torvalds * as well. But not if get_block() returned an error. 21931da177e4SLinus Torvalds */ 21941da177e4SLinus Torvalds if (!PageError(page)) 21951da177e4SLinus Torvalds SetPageUptodate(page); 21961da177e4SLinus Torvalds unlock_page(page); 21971da177e4SLinus Torvalds return 0; 21981da177e4SLinus Torvalds } 21991da177e4SLinus Torvalds 22001da177e4SLinus Torvalds /* Stage two: lock the buffers */ 22011da177e4SLinus Torvalds for (i = 0; i < nr; i++) { 22021da177e4SLinus Torvalds bh = arr[i]; 22031da177e4SLinus Torvalds lock_buffer(bh); 22041da177e4SLinus Torvalds mark_buffer_async_read(bh); 22051da177e4SLinus Torvalds } 22061da177e4SLinus Torvalds 22071da177e4SLinus Torvalds /* 22081da177e4SLinus Torvalds * Stage 3: start the IO. Check for uptodateness 22091da177e4SLinus Torvalds * inside the buffer lock in case another process reading 22101da177e4SLinus Torvalds * the underlying blockdev brought it uptodate (the sct fix). 22111da177e4SLinus Torvalds */ 22121da177e4SLinus Torvalds for (i = 0; i < nr; i++) { 22131da177e4SLinus Torvalds bh = arr[i]; 22141da177e4SLinus Torvalds if (buffer_uptodate(bh)) 22151da177e4SLinus Torvalds end_buffer_async_read(bh, 1); 22161da177e4SLinus Torvalds else 22171da177e4SLinus Torvalds submit_bh(READ, bh); 22181da177e4SLinus Torvalds } 22191da177e4SLinus Torvalds return 0; 22201da177e4SLinus Torvalds } 22211da177e4SLinus Torvalds 22221da177e4SLinus Torvalds /* utility function for filesystems that need to do work on expanding 222389e10787SNick Piggin * truncates. Uses filesystem pagecache writes to allow the filesystem to 22241da177e4SLinus Torvalds * deal with the hole. 22251da177e4SLinus Torvalds */ 222689e10787SNick Piggin int generic_cont_expand_simple(struct inode *inode, loff_t size) 22271da177e4SLinus Torvalds { 22281da177e4SLinus Torvalds struct address_space *mapping = inode->i_mapping; 22291da177e4SLinus Torvalds struct page *page; 223089e10787SNick Piggin void *fsdata; 223105eb0b51SOGAWA Hirofumi unsigned long limit; 22321da177e4SLinus Torvalds int err; 22331da177e4SLinus Torvalds 22341da177e4SLinus Torvalds err = -EFBIG; 22351da177e4SLinus Torvalds limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur; 22361da177e4SLinus Torvalds if (limit != RLIM_INFINITY && size > (loff_t)limit) { 22371da177e4SLinus Torvalds send_sig(SIGXFSZ, current, 0); 22381da177e4SLinus Torvalds goto out; 22391da177e4SLinus Torvalds } 22401da177e4SLinus Torvalds if (size > inode->i_sb->s_maxbytes) 22411da177e4SLinus Torvalds goto out; 22421da177e4SLinus Torvalds 224389e10787SNick Piggin err = pagecache_write_begin(NULL, mapping, size, 0, 224489e10787SNick Piggin AOP_FLAG_UNINTERRUPTIBLE|AOP_FLAG_CONT_EXPAND, 224589e10787SNick Piggin &page, &fsdata); 224689e10787SNick Piggin if (err) 224705eb0b51SOGAWA Hirofumi goto out; 224805eb0b51SOGAWA Hirofumi 224989e10787SNick Piggin err = pagecache_write_end(NULL, mapping, size, 0, 0, page, fsdata); 225089e10787SNick Piggin BUG_ON(err > 0); 225105eb0b51SOGAWA Hirofumi 225205eb0b51SOGAWA Hirofumi out: 225305eb0b51SOGAWA Hirofumi return err; 225405eb0b51SOGAWA Hirofumi } 225505eb0b51SOGAWA Hirofumi 2256f1e3af72SAdrian Bunk static int cont_expand_zero(struct file *file, struct address_space *mapping, 225789e10787SNick Piggin loff_t pos, loff_t *bytes) 225805eb0b51SOGAWA Hirofumi { 225989e10787SNick Piggin struct inode *inode = mapping->host; 226089e10787SNick Piggin unsigned blocksize = 1 << inode->i_blkbits; 226189e10787SNick Piggin struct page *page; 226289e10787SNick Piggin void *fsdata; 226389e10787SNick Piggin pgoff_t index, curidx; 226489e10787SNick Piggin loff_t curpos; 226589e10787SNick Piggin unsigned zerofrom, offset, len; 226689e10787SNick Piggin int err = 0; 226705eb0b51SOGAWA Hirofumi 226889e10787SNick Piggin index = pos >> PAGE_CACHE_SHIFT; 226989e10787SNick Piggin offset = pos & ~PAGE_CACHE_MASK; 227089e10787SNick Piggin 227189e10787SNick Piggin while (index > (curidx = (curpos = *bytes)>>PAGE_CACHE_SHIFT)) { 227289e10787SNick Piggin zerofrom = curpos & ~PAGE_CACHE_MASK; 227389e10787SNick Piggin if (zerofrom & (blocksize-1)) { 227489e10787SNick Piggin *bytes |= (blocksize-1); 227589e10787SNick Piggin (*bytes)++; 227689e10787SNick Piggin } 227789e10787SNick Piggin len = PAGE_CACHE_SIZE - zerofrom; 227889e10787SNick Piggin 227989e10787SNick Piggin err = pagecache_write_begin(file, mapping, curpos, len, 228089e10787SNick Piggin AOP_FLAG_UNINTERRUPTIBLE, 228189e10787SNick Piggin &page, &fsdata); 228289e10787SNick Piggin if (err) 228389e10787SNick Piggin goto out; 2284eebd2aa3SChristoph Lameter zero_user(page, zerofrom, len); 228589e10787SNick Piggin err = pagecache_write_end(file, mapping, curpos, len, len, 228689e10787SNick Piggin page, fsdata); 228789e10787SNick Piggin if (err < 0) 228889e10787SNick Piggin goto out; 228989e10787SNick Piggin BUG_ON(err != len); 229089e10787SNick Piggin err = 0; 2291061e9746SOGAWA Hirofumi 2292061e9746SOGAWA Hirofumi balance_dirty_pages_ratelimited(mapping); 229389e10787SNick Piggin } 229489e10787SNick Piggin 229589e10787SNick Piggin /* page covers the boundary, find the boundary offset */ 229689e10787SNick Piggin if (index == curidx) { 229789e10787SNick Piggin zerofrom = curpos & ~PAGE_CACHE_MASK; 229889e10787SNick Piggin /* if we will expand the thing last block will be filled */ 229989e10787SNick Piggin if (offset <= zerofrom) { 230089e10787SNick Piggin goto out; 230189e10787SNick Piggin } 230289e10787SNick Piggin if (zerofrom & (blocksize-1)) { 230389e10787SNick Piggin *bytes |= (blocksize-1); 230489e10787SNick Piggin (*bytes)++; 230589e10787SNick Piggin } 230689e10787SNick Piggin len = offset - zerofrom; 230789e10787SNick Piggin 230889e10787SNick Piggin err = pagecache_write_begin(file, mapping, curpos, len, 230989e10787SNick Piggin AOP_FLAG_UNINTERRUPTIBLE, 231089e10787SNick Piggin &page, &fsdata); 231189e10787SNick Piggin if (err) 231289e10787SNick Piggin goto out; 2313eebd2aa3SChristoph Lameter zero_user(page, zerofrom, len); 231489e10787SNick Piggin err = pagecache_write_end(file, mapping, curpos, len, len, 231589e10787SNick Piggin page, fsdata); 231689e10787SNick Piggin if (err < 0) 231789e10787SNick Piggin goto out; 231889e10787SNick Piggin BUG_ON(err != len); 231989e10787SNick Piggin err = 0; 232089e10787SNick Piggin } 232189e10787SNick Piggin out: 232289e10787SNick Piggin return err; 23231da177e4SLinus Torvalds } 23241da177e4SLinus Torvalds 23251da177e4SLinus Torvalds /* 23261da177e4SLinus Torvalds * For moronic filesystems that do not allow holes in file. 23271da177e4SLinus Torvalds * We may have to extend the file. 23281da177e4SLinus Torvalds */ 232989e10787SNick Piggin int cont_write_begin(struct file *file, struct address_space *mapping, 233089e10787SNick Piggin loff_t pos, unsigned len, unsigned flags, 233189e10787SNick Piggin struct page **pagep, void **fsdata, 233289e10787SNick Piggin get_block_t *get_block, loff_t *bytes) 23331da177e4SLinus Torvalds { 23341da177e4SLinus Torvalds struct inode *inode = mapping->host; 23351da177e4SLinus Torvalds unsigned blocksize = 1 << inode->i_blkbits; 233689e10787SNick Piggin unsigned zerofrom; 233789e10787SNick Piggin int err; 23381da177e4SLinus Torvalds 233989e10787SNick Piggin err = cont_expand_zero(file, mapping, pos, bytes); 234089e10787SNick Piggin if (err) 23411da177e4SLinus Torvalds goto out; 23421da177e4SLinus Torvalds 23431da177e4SLinus Torvalds zerofrom = *bytes & ~PAGE_CACHE_MASK; 234489e10787SNick Piggin if (pos+len > *bytes && zerofrom & (blocksize-1)) { 23451da177e4SLinus Torvalds *bytes |= (blocksize-1); 23461da177e4SLinus Torvalds (*bytes)++; 23471da177e4SLinus Torvalds } 23481da177e4SLinus Torvalds 234989e10787SNick Piggin *pagep = NULL; 235089e10787SNick Piggin err = block_write_begin(file, mapping, pos, len, 235189e10787SNick Piggin flags, pagep, fsdata, get_block); 23521da177e4SLinus Torvalds out: 235389e10787SNick Piggin return err; 23541da177e4SLinus Torvalds } 23551da177e4SLinus Torvalds 23561da177e4SLinus Torvalds int block_prepare_write(struct page *page, unsigned from, unsigned to, 23571da177e4SLinus Torvalds get_block_t *get_block) 23581da177e4SLinus Torvalds { 23591da177e4SLinus Torvalds struct inode *inode = page->mapping->host; 23601da177e4SLinus Torvalds int err = __block_prepare_write(inode, page, from, to, get_block); 23611da177e4SLinus Torvalds if (err) 23621da177e4SLinus Torvalds ClearPageUptodate(page); 23631da177e4SLinus Torvalds return err; 23641da177e4SLinus Torvalds } 23651da177e4SLinus Torvalds 23661da177e4SLinus Torvalds int block_commit_write(struct page *page, unsigned from, unsigned to) 23671da177e4SLinus Torvalds { 23681da177e4SLinus Torvalds struct inode *inode = page->mapping->host; 23691da177e4SLinus Torvalds __block_commit_write(inode,page,from,to); 23701da177e4SLinus Torvalds return 0; 23711da177e4SLinus Torvalds } 23721da177e4SLinus Torvalds 237354171690SDavid Chinner /* 237454171690SDavid Chinner * block_page_mkwrite() is not allowed to change the file size as it gets 237554171690SDavid Chinner * called from a page fault handler when a page is first dirtied. Hence we must 237654171690SDavid Chinner * be careful to check for EOF conditions here. We set the page up correctly 237754171690SDavid Chinner * for a written page which means we get ENOSPC checking when writing into 237854171690SDavid Chinner * holes and correct delalloc and unwritten extent mapping on filesystems that 237954171690SDavid Chinner * support these features. 238054171690SDavid Chinner * 238154171690SDavid Chinner * We are not allowed to take the i_mutex here so we have to play games to 238254171690SDavid Chinner * protect against truncate races as the page could now be beyond EOF. Because 238354171690SDavid Chinner * vmtruncate() writes the inode size before removing pages, once we have the 238454171690SDavid Chinner * page lock we can determine safely if the page is beyond EOF. If it is not 238554171690SDavid Chinner * beyond EOF, then the page is guaranteed safe against truncation until we 238654171690SDavid Chinner * unlock the page. 238754171690SDavid Chinner */ 238854171690SDavid Chinner int 2389c2ec175cSNick Piggin block_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf, 239054171690SDavid Chinner get_block_t get_block) 239154171690SDavid Chinner { 2392c2ec175cSNick Piggin struct page *page = vmf->page; 239354171690SDavid Chinner struct inode *inode = vma->vm_file->f_path.dentry->d_inode; 239454171690SDavid Chinner unsigned long end; 239554171690SDavid Chinner loff_t size; 239656a76f82SNick Piggin int ret = VM_FAULT_NOPAGE; /* make the VM retry the fault */ 239754171690SDavid Chinner 239854171690SDavid Chinner lock_page(page); 239954171690SDavid Chinner size = i_size_read(inode); 240054171690SDavid Chinner if ((page->mapping != inode->i_mapping) || 240118336338SNick Piggin (page_offset(page) > size)) { 240254171690SDavid Chinner /* page got truncated out from underneath us */ 2403b827e496SNick Piggin unlock_page(page); 2404b827e496SNick Piggin goto out; 240554171690SDavid Chinner } 240654171690SDavid Chinner 240754171690SDavid Chinner /* page is wholly or partially inside EOF */ 240854171690SDavid Chinner if (((page->index + 1) << PAGE_CACHE_SHIFT) > size) 240954171690SDavid Chinner end = size & ~PAGE_CACHE_MASK; 241054171690SDavid Chinner else 241154171690SDavid Chinner end = PAGE_CACHE_SIZE; 241254171690SDavid Chinner 241354171690SDavid Chinner ret = block_prepare_write(page, 0, end, get_block); 241454171690SDavid Chinner if (!ret) 241554171690SDavid Chinner ret = block_commit_write(page, 0, end); 241654171690SDavid Chinner 241756a76f82SNick Piggin if (unlikely(ret)) { 2418b827e496SNick Piggin unlock_page(page); 241956a76f82SNick Piggin if (ret == -ENOMEM) 242056a76f82SNick Piggin ret = VM_FAULT_OOM; 242156a76f82SNick Piggin else /* -ENOSPC, -EIO, etc */ 2422c2ec175cSNick Piggin ret = VM_FAULT_SIGBUS; 2423b827e496SNick Piggin } else 2424b827e496SNick Piggin ret = VM_FAULT_LOCKED; 2425c2ec175cSNick Piggin 2426b827e496SNick Piggin out: 242754171690SDavid Chinner return ret; 242854171690SDavid Chinner } 24291da177e4SLinus Torvalds 24301da177e4SLinus Torvalds /* 243103158cd7SNick Piggin * nobh_write_begin()'s prereads are special: the buffer_heads are freed 24321da177e4SLinus Torvalds * immediately, while under the page lock. So it needs a special end_io 24331da177e4SLinus Torvalds * handler which does not touch the bh after unlocking it. 24341da177e4SLinus Torvalds */ 24351da177e4SLinus Torvalds static void end_buffer_read_nobh(struct buffer_head *bh, int uptodate) 24361da177e4SLinus Torvalds { 243768671f35SDmitry Monakhov __end_buffer_read_notouch(bh, uptodate); 24381da177e4SLinus Torvalds } 24391da177e4SLinus Torvalds 24401da177e4SLinus Torvalds /* 244103158cd7SNick Piggin * Attach the singly-linked list of buffers created by nobh_write_begin, to 244203158cd7SNick Piggin * the page (converting it to circular linked list and taking care of page 244303158cd7SNick Piggin * dirty races). 244403158cd7SNick Piggin */ 244503158cd7SNick Piggin static void attach_nobh_buffers(struct page *page, struct buffer_head *head) 244603158cd7SNick Piggin { 244703158cd7SNick Piggin struct buffer_head *bh; 244803158cd7SNick Piggin 244903158cd7SNick Piggin BUG_ON(!PageLocked(page)); 245003158cd7SNick Piggin 245103158cd7SNick Piggin spin_lock(&page->mapping->private_lock); 245203158cd7SNick Piggin bh = head; 245303158cd7SNick Piggin do { 245403158cd7SNick Piggin if (PageDirty(page)) 245503158cd7SNick Piggin set_buffer_dirty(bh); 245603158cd7SNick Piggin if (!bh->b_this_page) 245703158cd7SNick Piggin bh->b_this_page = head; 245803158cd7SNick Piggin bh = bh->b_this_page; 245903158cd7SNick Piggin } while (bh != head); 246003158cd7SNick Piggin attach_page_buffers(page, head); 246103158cd7SNick Piggin spin_unlock(&page->mapping->private_lock); 246203158cd7SNick Piggin } 246303158cd7SNick Piggin 246403158cd7SNick Piggin /* 24651da177e4SLinus Torvalds * On entry, the page is fully not uptodate. 24661da177e4SLinus Torvalds * On exit the page is fully uptodate in the areas outside (from,to) 24671da177e4SLinus Torvalds */ 246803158cd7SNick Piggin int nobh_write_begin(struct file *file, struct address_space *mapping, 246903158cd7SNick Piggin loff_t pos, unsigned len, unsigned flags, 247003158cd7SNick Piggin struct page **pagep, void **fsdata, 24711da177e4SLinus Torvalds get_block_t *get_block) 24721da177e4SLinus Torvalds { 247303158cd7SNick Piggin struct inode *inode = mapping->host; 24741da177e4SLinus Torvalds const unsigned blkbits = inode->i_blkbits; 24751da177e4SLinus Torvalds const unsigned blocksize = 1 << blkbits; 2476a4b0672dSNick Piggin struct buffer_head *head, *bh; 247703158cd7SNick Piggin struct page *page; 247803158cd7SNick Piggin pgoff_t index; 247903158cd7SNick Piggin unsigned from, to; 24801da177e4SLinus Torvalds unsigned block_in_page; 2481a4b0672dSNick Piggin unsigned block_start, block_end; 24821da177e4SLinus Torvalds sector_t block_in_file; 24831da177e4SLinus Torvalds int nr_reads = 0; 24841da177e4SLinus Torvalds int ret = 0; 24851da177e4SLinus Torvalds int is_mapped_to_disk = 1; 24861da177e4SLinus Torvalds 248703158cd7SNick Piggin index = pos >> PAGE_CACHE_SHIFT; 248803158cd7SNick Piggin from = pos & (PAGE_CACHE_SIZE - 1); 248903158cd7SNick Piggin to = from + len; 249003158cd7SNick Piggin 249154566b2cSNick Piggin page = grab_cache_page_write_begin(mapping, index, flags); 249203158cd7SNick Piggin if (!page) 249303158cd7SNick Piggin return -ENOMEM; 249403158cd7SNick Piggin *pagep = page; 249503158cd7SNick Piggin *fsdata = NULL; 249603158cd7SNick Piggin 249703158cd7SNick Piggin if (page_has_buffers(page)) { 249803158cd7SNick Piggin unlock_page(page); 249903158cd7SNick Piggin page_cache_release(page); 250003158cd7SNick Piggin *pagep = NULL; 250103158cd7SNick Piggin return block_write_begin(file, mapping, pos, len, flags, pagep, 250203158cd7SNick Piggin fsdata, get_block); 250303158cd7SNick Piggin } 2504a4b0672dSNick Piggin 25051da177e4SLinus Torvalds if (PageMappedToDisk(page)) 25061da177e4SLinus Torvalds return 0; 25071da177e4SLinus Torvalds 2508a4b0672dSNick Piggin /* 2509a4b0672dSNick Piggin * Allocate buffers so that we can keep track of state, and potentially 2510a4b0672dSNick Piggin * attach them to the page if an error occurs. In the common case of 2511a4b0672dSNick Piggin * no error, they will just be freed again without ever being attached 2512a4b0672dSNick Piggin * to the page (which is all OK, because we're under the page lock). 2513a4b0672dSNick Piggin * 2514a4b0672dSNick Piggin * Be careful: the buffer linked list is a NULL terminated one, rather 2515a4b0672dSNick Piggin * than the circular one we're used to. 2516a4b0672dSNick Piggin */ 2517a4b0672dSNick Piggin head = alloc_page_buffers(page, blocksize, 0); 251803158cd7SNick Piggin if (!head) { 251903158cd7SNick Piggin ret = -ENOMEM; 252003158cd7SNick Piggin goto out_release; 252103158cd7SNick Piggin } 2522a4b0672dSNick Piggin 25231da177e4SLinus Torvalds block_in_file = (sector_t)page->index << (PAGE_CACHE_SHIFT - blkbits); 25241da177e4SLinus Torvalds 25251da177e4SLinus Torvalds /* 25261da177e4SLinus Torvalds * We loop across all blocks in the page, whether or not they are 25271da177e4SLinus Torvalds * part of the affected region. This is so we can discover if the 25281da177e4SLinus Torvalds * page is fully mapped-to-disk. 25291da177e4SLinus Torvalds */ 2530a4b0672dSNick Piggin for (block_start = 0, block_in_page = 0, bh = head; 25311da177e4SLinus Torvalds block_start < PAGE_CACHE_SIZE; 2532a4b0672dSNick Piggin block_in_page++, block_start += blocksize, bh = bh->b_this_page) { 25331da177e4SLinus Torvalds int create; 25341da177e4SLinus Torvalds 2535a4b0672dSNick Piggin block_end = block_start + blocksize; 2536a4b0672dSNick Piggin bh->b_state = 0; 25371da177e4SLinus Torvalds create = 1; 25381da177e4SLinus Torvalds if (block_start >= to) 25391da177e4SLinus Torvalds create = 0; 25401da177e4SLinus Torvalds ret = get_block(inode, block_in_file + block_in_page, 2541a4b0672dSNick Piggin bh, create); 25421da177e4SLinus Torvalds if (ret) 25431da177e4SLinus Torvalds goto failed; 2544a4b0672dSNick Piggin if (!buffer_mapped(bh)) 25451da177e4SLinus Torvalds is_mapped_to_disk = 0; 2546a4b0672dSNick Piggin if (buffer_new(bh)) 2547a4b0672dSNick Piggin unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr); 2548a4b0672dSNick Piggin if (PageUptodate(page)) { 2549a4b0672dSNick Piggin set_buffer_uptodate(bh); 25501da177e4SLinus Torvalds continue; 2551a4b0672dSNick Piggin } 2552a4b0672dSNick Piggin if (buffer_new(bh) || !buffer_mapped(bh)) { 2553eebd2aa3SChristoph Lameter zero_user_segments(page, block_start, from, 2554eebd2aa3SChristoph Lameter to, block_end); 25551da177e4SLinus Torvalds continue; 25561da177e4SLinus Torvalds } 2557a4b0672dSNick Piggin if (buffer_uptodate(bh)) 25581da177e4SLinus Torvalds continue; /* reiserfs does this */ 25591da177e4SLinus Torvalds if (block_start < from || block_end > to) { 2560a4b0672dSNick Piggin lock_buffer(bh); 2561a4b0672dSNick Piggin bh->b_end_io = end_buffer_read_nobh; 2562a4b0672dSNick Piggin submit_bh(READ, bh); 2563a4b0672dSNick Piggin nr_reads++; 25641da177e4SLinus Torvalds } 25651da177e4SLinus Torvalds } 25661da177e4SLinus Torvalds 25671da177e4SLinus Torvalds if (nr_reads) { 25681da177e4SLinus Torvalds /* 25691da177e4SLinus Torvalds * The page is locked, so these buffers are protected from 25701da177e4SLinus Torvalds * any VM or truncate activity. Hence we don't need to care 25711da177e4SLinus Torvalds * for the buffer_head refcounts. 25721da177e4SLinus Torvalds */ 2573a4b0672dSNick Piggin for (bh = head; bh; bh = bh->b_this_page) { 25741da177e4SLinus Torvalds wait_on_buffer(bh); 25751da177e4SLinus Torvalds if (!buffer_uptodate(bh)) 25761da177e4SLinus Torvalds ret = -EIO; 25771da177e4SLinus Torvalds } 25781da177e4SLinus Torvalds if (ret) 25791da177e4SLinus Torvalds goto failed; 25801da177e4SLinus Torvalds } 25811da177e4SLinus Torvalds 25821da177e4SLinus Torvalds if (is_mapped_to_disk) 25831da177e4SLinus Torvalds SetPageMappedToDisk(page); 25841da177e4SLinus Torvalds 258503158cd7SNick Piggin *fsdata = head; /* to be released by nobh_write_end */ 2586a4b0672dSNick Piggin 25871da177e4SLinus Torvalds return 0; 25881da177e4SLinus Torvalds 25891da177e4SLinus Torvalds failed: 259003158cd7SNick Piggin BUG_ON(!ret); 25911da177e4SLinus Torvalds /* 2592a4b0672dSNick Piggin * Error recovery is a bit difficult. We need to zero out blocks that 2593a4b0672dSNick Piggin * were newly allocated, and dirty them to ensure they get written out. 2594a4b0672dSNick Piggin * Buffers need to be attached to the page at this point, otherwise 2595a4b0672dSNick Piggin * the handling of potential IO errors during writeout would be hard 2596a4b0672dSNick Piggin * (could try doing synchronous writeout, but what if that fails too?) 25971da177e4SLinus Torvalds */ 259803158cd7SNick Piggin attach_nobh_buffers(page, head); 259903158cd7SNick Piggin page_zero_new_buffers(page, from, to); 2600a4b0672dSNick Piggin 260103158cd7SNick Piggin out_release: 260203158cd7SNick Piggin unlock_page(page); 260303158cd7SNick Piggin page_cache_release(page); 260403158cd7SNick Piggin *pagep = NULL; 2605a4b0672dSNick Piggin 260603158cd7SNick Piggin if (pos + len > inode->i_size) 260703158cd7SNick Piggin vmtruncate(inode, inode->i_size); 2608a4b0672dSNick Piggin 26091da177e4SLinus Torvalds return ret; 26101da177e4SLinus Torvalds } 261103158cd7SNick Piggin EXPORT_SYMBOL(nobh_write_begin); 26121da177e4SLinus Torvalds 261303158cd7SNick Piggin int nobh_write_end(struct file *file, struct address_space *mapping, 261403158cd7SNick Piggin loff_t pos, unsigned len, unsigned copied, 261503158cd7SNick Piggin struct page *page, void *fsdata) 26161da177e4SLinus Torvalds { 26171da177e4SLinus Torvalds struct inode *inode = page->mapping->host; 2618efdc3131SNick Piggin struct buffer_head *head = fsdata; 261903158cd7SNick Piggin struct buffer_head *bh; 26205b41e74aSDmitri Monakhov BUG_ON(fsdata != NULL && page_has_buffers(page)); 26211da177e4SLinus Torvalds 2622d4cf109fSDave Kleikamp if (unlikely(copied < len) && head) 262303158cd7SNick Piggin attach_nobh_buffers(page, head); 2624a4b0672dSNick Piggin if (page_has_buffers(page)) 262503158cd7SNick Piggin return generic_write_end(file, mapping, pos, len, 262603158cd7SNick Piggin copied, page, fsdata); 2627a4b0672dSNick Piggin 262822c8ca78SNick Piggin SetPageUptodate(page); 26291da177e4SLinus Torvalds set_page_dirty(page); 263003158cd7SNick Piggin if (pos+copied > inode->i_size) { 263103158cd7SNick Piggin i_size_write(inode, pos+copied); 26321da177e4SLinus Torvalds mark_inode_dirty(inode); 26331da177e4SLinus Torvalds } 263403158cd7SNick Piggin 263503158cd7SNick Piggin unlock_page(page); 263603158cd7SNick Piggin page_cache_release(page); 263703158cd7SNick Piggin 263803158cd7SNick Piggin while (head) { 263903158cd7SNick Piggin bh = head; 264003158cd7SNick Piggin head = head->b_this_page; 264103158cd7SNick Piggin free_buffer_head(bh); 26421da177e4SLinus Torvalds } 264303158cd7SNick Piggin 264403158cd7SNick Piggin return copied; 264503158cd7SNick Piggin } 264603158cd7SNick Piggin EXPORT_SYMBOL(nobh_write_end); 26471da177e4SLinus Torvalds 26481da177e4SLinus Torvalds /* 26491da177e4SLinus Torvalds * nobh_writepage() - based on block_full_write_page() except 26501da177e4SLinus Torvalds * that it tries to operate without attaching bufferheads to 26511da177e4SLinus Torvalds * the page. 26521da177e4SLinus Torvalds */ 26531da177e4SLinus Torvalds int nobh_writepage(struct page *page, get_block_t *get_block, 26541da177e4SLinus Torvalds struct writeback_control *wbc) 26551da177e4SLinus Torvalds { 26561da177e4SLinus Torvalds struct inode * const inode = page->mapping->host; 26571da177e4SLinus Torvalds loff_t i_size = i_size_read(inode); 26581da177e4SLinus Torvalds const pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT; 26591da177e4SLinus Torvalds unsigned offset; 26601da177e4SLinus Torvalds int ret; 26611da177e4SLinus Torvalds 26621da177e4SLinus Torvalds /* Is the page fully inside i_size? */ 26631da177e4SLinus Torvalds if (page->index < end_index) 26641da177e4SLinus Torvalds goto out; 26651da177e4SLinus Torvalds 26661da177e4SLinus Torvalds /* Is the page fully outside i_size? (truncate in progress) */ 26671da177e4SLinus Torvalds offset = i_size & (PAGE_CACHE_SIZE-1); 26681da177e4SLinus Torvalds if (page->index >= end_index+1 || !offset) { 26691da177e4SLinus Torvalds /* 26701da177e4SLinus Torvalds * The page may have dirty, unmapped buffers. For example, 26711da177e4SLinus Torvalds * they may have been added in ext3_writepage(). Make them 26721da177e4SLinus Torvalds * freeable here, so the page does not leak. 26731da177e4SLinus Torvalds */ 26741da177e4SLinus Torvalds #if 0 26751da177e4SLinus Torvalds /* Not really sure about this - do we need this ? */ 26761da177e4SLinus Torvalds if (page->mapping->a_ops->invalidatepage) 26771da177e4SLinus Torvalds page->mapping->a_ops->invalidatepage(page, offset); 26781da177e4SLinus Torvalds #endif 26791da177e4SLinus Torvalds unlock_page(page); 26801da177e4SLinus Torvalds return 0; /* don't care */ 26811da177e4SLinus Torvalds } 26821da177e4SLinus Torvalds 26831da177e4SLinus Torvalds /* 26841da177e4SLinus Torvalds * The page straddles i_size. It must be zeroed out on each and every 26851da177e4SLinus Torvalds * writepage invocation because it may be mmapped. "A file is mapped 26861da177e4SLinus Torvalds * in multiples of the page size. For a file that is not a multiple of 26871da177e4SLinus Torvalds * the page size, the remaining memory is zeroed when mapped, and 26881da177e4SLinus Torvalds * writes to that region are not written out to the file." 26891da177e4SLinus Torvalds */ 2690eebd2aa3SChristoph Lameter zero_user_segment(page, offset, PAGE_CACHE_SIZE); 26911da177e4SLinus Torvalds out: 26921da177e4SLinus Torvalds ret = mpage_writepage(page, get_block, wbc); 26931da177e4SLinus Torvalds if (ret == -EAGAIN) 269435c80d5fSChris Mason ret = __block_write_full_page(inode, page, get_block, wbc, 269535c80d5fSChris Mason end_buffer_async_write); 26961da177e4SLinus Torvalds return ret; 26971da177e4SLinus Torvalds } 26981da177e4SLinus Torvalds EXPORT_SYMBOL(nobh_writepage); 26991da177e4SLinus Torvalds 270003158cd7SNick Piggin int nobh_truncate_page(struct address_space *mapping, 270103158cd7SNick Piggin loff_t from, get_block_t *get_block) 27021da177e4SLinus Torvalds { 27031da177e4SLinus Torvalds pgoff_t index = from >> PAGE_CACHE_SHIFT; 27041da177e4SLinus Torvalds unsigned offset = from & (PAGE_CACHE_SIZE-1); 270503158cd7SNick Piggin unsigned blocksize; 270603158cd7SNick Piggin sector_t iblock; 270703158cd7SNick Piggin unsigned length, pos; 270803158cd7SNick Piggin struct inode *inode = mapping->host; 27091da177e4SLinus Torvalds struct page *page; 271003158cd7SNick Piggin struct buffer_head map_bh; 271103158cd7SNick Piggin int err; 27121da177e4SLinus Torvalds 271303158cd7SNick Piggin blocksize = 1 << inode->i_blkbits; 271403158cd7SNick Piggin length = offset & (blocksize - 1); 27151da177e4SLinus Torvalds 271603158cd7SNick Piggin /* Block boundary? Nothing to do */ 271703158cd7SNick Piggin if (!length) 271803158cd7SNick Piggin return 0; 271903158cd7SNick Piggin 272003158cd7SNick Piggin length = blocksize - length; 272103158cd7SNick Piggin iblock = (sector_t)index << (PAGE_CACHE_SHIFT - inode->i_blkbits); 272203158cd7SNick Piggin 27231da177e4SLinus Torvalds page = grab_cache_page(mapping, index); 272403158cd7SNick Piggin err = -ENOMEM; 27251da177e4SLinus Torvalds if (!page) 27261da177e4SLinus Torvalds goto out; 27271da177e4SLinus Torvalds 272803158cd7SNick Piggin if (page_has_buffers(page)) { 272903158cd7SNick Piggin has_buffers: 273003158cd7SNick Piggin unlock_page(page); 273103158cd7SNick Piggin page_cache_release(page); 273203158cd7SNick Piggin return block_truncate_page(mapping, from, get_block); 27331da177e4SLinus Torvalds } 273403158cd7SNick Piggin 273503158cd7SNick Piggin /* Find the buffer that contains "offset" */ 273603158cd7SNick Piggin pos = blocksize; 273703158cd7SNick Piggin while (offset >= pos) { 273803158cd7SNick Piggin iblock++; 273903158cd7SNick Piggin pos += blocksize; 274003158cd7SNick Piggin } 274103158cd7SNick Piggin 2742460bcf57STheodore Ts'o map_bh.b_size = blocksize; 2743460bcf57STheodore Ts'o map_bh.b_state = 0; 274403158cd7SNick Piggin err = get_block(inode, iblock, &map_bh, 0); 274503158cd7SNick Piggin if (err) 274603158cd7SNick Piggin goto unlock; 274703158cd7SNick Piggin /* unmapped? It's a hole - nothing to do */ 274803158cd7SNick Piggin if (!buffer_mapped(&map_bh)) 274903158cd7SNick Piggin goto unlock; 275003158cd7SNick Piggin 275103158cd7SNick Piggin /* Ok, it's mapped. Make sure it's up-to-date */ 275203158cd7SNick Piggin if (!PageUptodate(page)) { 275303158cd7SNick Piggin err = mapping->a_ops->readpage(NULL, page); 275403158cd7SNick Piggin if (err) { 275503158cd7SNick Piggin page_cache_release(page); 275603158cd7SNick Piggin goto out; 275703158cd7SNick Piggin } 275803158cd7SNick Piggin lock_page(page); 275903158cd7SNick Piggin if (!PageUptodate(page)) { 276003158cd7SNick Piggin err = -EIO; 276103158cd7SNick Piggin goto unlock; 276203158cd7SNick Piggin } 276303158cd7SNick Piggin if (page_has_buffers(page)) 276403158cd7SNick Piggin goto has_buffers; 276503158cd7SNick Piggin } 2766eebd2aa3SChristoph Lameter zero_user(page, offset, length); 276703158cd7SNick Piggin set_page_dirty(page); 276803158cd7SNick Piggin err = 0; 276903158cd7SNick Piggin 277003158cd7SNick Piggin unlock: 27711da177e4SLinus Torvalds unlock_page(page); 27721da177e4SLinus Torvalds page_cache_release(page); 27731da177e4SLinus Torvalds out: 277403158cd7SNick Piggin return err; 27751da177e4SLinus Torvalds } 27761da177e4SLinus Torvalds EXPORT_SYMBOL(nobh_truncate_page); 27771da177e4SLinus Torvalds 27781da177e4SLinus Torvalds int block_truncate_page(struct address_space *mapping, 27791da177e4SLinus Torvalds loff_t from, get_block_t *get_block) 27801da177e4SLinus Torvalds { 27811da177e4SLinus Torvalds pgoff_t index = from >> PAGE_CACHE_SHIFT; 27821da177e4SLinus Torvalds unsigned offset = from & (PAGE_CACHE_SIZE-1); 27831da177e4SLinus Torvalds unsigned blocksize; 278454b21a79SAndrew Morton sector_t iblock; 27851da177e4SLinus Torvalds unsigned length, pos; 27861da177e4SLinus Torvalds struct inode *inode = mapping->host; 27871da177e4SLinus Torvalds struct page *page; 27881da177e4SLinus Torvalds struct buffer_head *bh; 27891da177e4SLinus Torvalds int err; 27901da177e4SLinus Torvalds 27911da177e4SLinus Torvalds blocksize = 1 << inode->i_blkbits; 27921da177e4SLinus Torvalds length = offset & (blocksize - 1); 27931da177e4SLinus Torvalds 27941da177e4SLinus Torvalds /* Block boundary? Nothing to do */ 27951da177e4SLinus Torvalds if (!length) 27961da177e4SLinus Torvalds return 0; 27971da177e4SLinus Torvalds 27981da177e4SLinus Torvalds length = blocksize - length; 279954b21a79SAndrew Morton iblock = (sector_t)index << (PAGE_CACHE_SHIFT - inode->i_blkbits); 28001da177e4SLinus Torvalds 28011da177e4SLinus Torvalds page = grab_cache_page(mapping, index); 28021da177e4SLinus Torvalds err = -ENOMEM; 28031da177e4SLinus Torvalds if (!page) 28041da177e4SLinus Torvalds goto out; 28051da177e4SLinus Torvalds 28061da177e4SLinus Torvalds if (!page_has_buffers(page)) 28071da177e4SLinus Torvalds create_empty_buffers(page, blocksize, 0); 28081da177e4SLinus Torvalds 28091da177e4SLinus Torvalds /* Find the buffer that contains "offset" */ 28101da177e4SLinus Torvalds bh = page_buffers(page); 28111da177e4SLinus Torvalds pos = blocksize; 28121da177e4SLinus Torvalds while (offset >= pos) { 28131da177e4SLinus Torvalds bh = bh->b_this_page; 28141da177e4SLinus Torvalds iblock++; 28151da177e4SLinus Torvalds pos += blocksize; 28161da177e4SLinus Torvalds } 28171da177e4SLinus Torvalds 28181da177e4SLinus Torvalds err = 0; 28191da177e4SLinus Torvalds if (!buffer_mapped(bh)) { 2820b0cf2321SBadari Pulavarty WARN_ON(bh->b_size != blocksize); 28211da177e4SLinus Torvalds err = get_block(inode, iblock, bh, 0); 28221da177e4SLinus Torvalds if (err) 28231da177e4SLinus Torvalds goto unlock; 28241da177e4SLinus Torvalds /* unmapped? It's a hole - nothing to do */ 28251da177e4SLinus Torvalds if (!buffer_mapped(bh)) 28261da177e4SLinus Torvalds goto unlock; 28271da177e4SLinus Torvalds } 28281da177e4SLinus Torvalds 28291da177e4SLinus Torvalds /* Ok, it's mapped. Make sure it's up-to-date */ 28301da177e4SLinus Torvalds if (PageUptodate(page)) 28311da177e4SLinus Torvalds set_buffer_uptodate(bh); 28321da177e4SLinus Torvalds 283333a266ddSDavid Chinner if (!buffer_uptodate(bh) && !buffer_delay(bh) && !buffer_unwritten(bh)) { 28341da177e4SLinus Torvalds err = -EIO; 28351da177e4SLinus Torvalds ll_rw_block(READ, 1, &bh); 28361da177e4SLinus Torvalds wait_on_buffer(bh); 28371da177e4SLinus Torvalds /* Uhhuh. Read error. Complain and punt. */ 28381da177e4SLinus Torvalds if (!buffer_uptodate(bh)) 28391da177e4SLinus Torvalds goto unlock; 28401da177e4SLinus Torvalds } 28411da177e4SLinus Torvalds 2842eebd2aa3SChristoph Lameter zero_user(page, offset, length); 28431da177e4SLinus Torvalds mark_buffer_dirty(bh); 28441da177e4SLinus Torvalds err = 0; 28451da177e4SLinus Torvalds 28461da177e4SLinus Torvalds unlock: 28471da177e4SLinus Torvalds unlock_page(page); 28481da177e4SLinus Torvalds page_cache_release(page); 28491da177e4SLinus Torvalds out: 28501da177e4SLinus Torvalds return err; 28511da177e4SLinus Torvalds } 28521da177e4SLinus Torvalds 28531da177e4SLinus Torvalds /* 28541da177e4SLinus Torvalds * The generic ->writepage function for buffer-backed address_spaces 285535c80d5fSChris Mason * this form passes in the end_io handler used to finish the IO. 28561da177e4SLinus Torvalds */ 285735c80d5fSChris Mason int block_write_full_page_endio(struct page *page, get_block_t *get_block, 285835c80d5fSChris Mason struct writeback_control *wbc, bh_end_io_t *handler) 28591da177e4SLinus Torvalds { 28601da177e4SLinus Torvalds struct inode * const inode = page->mapping->host; 28611da177e4SLinus Torvalds loff_t i_size = i_size_read(inode); 28621da177e4SLinus Torvalds const pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT; 28631da177e4SLinus Torvalds unsigned offset; 28641da177e4SLinus Torvalds 28651da177e4SLinus Torvalds /* Is the page fully inside i_size? */ 28661da177e4SLinus Torvalds if (page->index < end_index) 286735c80d5fSChris Mason return __block_write_full_page(inode, page, get_block, wbc, 286835c80d5fSChris Mason handler); 28691da177e4SLinus Torvalds 28701da177e4SLinus Torvalds /* Is the page fully outside i_size? (truncate in progress) */ 28711da177e4SLinus Torvalds offset = i_size & (PAGE_CACHE_SIZE-1); 28721da177e4SLinus Torvalds if (page->index >= end_index+1 || !offset) { 28731da177e4SLinus Torvalds /* 28741da177e4SLinus Torvalds * The page may have dirty, unmapped buffers. For example, 28751da177e4SLinus Torvalds * they may have been added in ext3_writepage(). Make them 28761da177e4SLinus Torvalds * freeable here, so the page does not leak. 28771da177e4SLinus Torvalds */ 2878aaa4059bSJan Kara do_invalidatepage(page, 0); 28791da177e4SLinus Torvalds unlock_page(page); 28801da177e4SLinus Torvalds return 0; /* don't care */ 28811da177e4SLinus Torvalds } 28821da177e4SLinus Torvalds 28831da177e4SLinus Torvalds /* 28841da177e4SLinus Torvalds * The page straddles i_size. It must be zeroed out on each and every 28851da177e4SLinus Torvalds * writepage invokation because it may be mmapped. "A file is mapped 28861da177e4SLinus Torvalds * in multiples of the page size. For a file that is not a multiple of 28871da177e4SLinus Torvalds * the page size, the remaining memory is zeroed when mapped, and 28881da177e4SLinus Torvalds * writes to that region are not written out to the file." 28891da177e4SLinus Torvalds */ 2890eebd2aa3SChristoph Lameter zero_user_segment(page, offset, PAGE_CACHE_SIZE); 289135c80d5fSChris Mason return __block_write_full_page(inode, page, get_block, wbc, handler); 28921da177e4SLinus Torvalds } 28931da177e4SLinus Torvalds 289435c80d5fSChris Mason /* 289535c80d5fSChris Mason * The generic ->writepage function for buffer-backed address_spaces 289635c80d5fSChris Mason */ 289735c80d5fSChris Mason int block_write_full_page(struct page *page, get_block_t *get_block, 289835c80d5fSChris Mason struct writeback_control *wbc) 289935c80d5fSChris Mason { 290035c80d5fSChris Mason return block_write_full_page_endio(page, get_block, wbc, 290135c80d5fSChris Mason end_buffer_async_write); 290235c80d5fSChris Mason } 290335c80d5fSChris Mason 290435c80d5fSChris Mason 29051da177e4SLinus Torvalds sector_t generic_block_bmap(struct address_space *mapping, sector_t block, 29061da177e4SLinus Torvalds get_block_t *get_block) 29071da177e4SLinus Torvalds { 29081da177e4SLinus Torvalds struct buffer_head tmp; 29091da177e4SLinus Torvalds struct inode *inode = mapping->host; 29101da177e4SLinus Torvalds tmp.b_state = 0; 29111da177e4SLinus Torvalds tmp.b_blocknr = 0; 2912b0cf2321SBadari Pulavarty tmp.b_size = 1 << inode->i_blkbits; 29131da177e4SLinus Torvalds get_block(inode, block, &tmp, 0); 29141da177e4SLinus Torvalds return tmp.b_blocknr; 29151da177e4SLinus Torvalds } 29161da177e4SLinus Torvalds 29176712ecf8SNeilBrown static void end_bio_bh_io_sync(struct bio *bio, int err) 29181da177e4SLinus Torvalds { 29191da177e4SLinus Torvalds struct buffer_head *bh = bio->bi_private; 29201da177e4SLinus Torvalds 29211da177e4SLinus Torvalds if (err == -EOPNOTSUPP) { 29221da177e4SLinus Torvalds set_bit(BIO_EOPNOTSUPP, &bio->bi_flags); 29231da177e4SLinus Torvalds set_bit(BH_Eopnotsupp, &bh->b_state); 29241da177e4SLinus Torvalds } 29251da177e4SLinus Torvalds 292608bafc03SKeith Mannthey if (unlikely (test_bit(BIO_QUIET,&bio->bi_flags))) 292708bafc03SKeith Mannthey set_bit(BH_Quiet, &bh->b_state); 292808bafc03SKeith Mannthey 29291da177e4SLinus Torvalds bh->b_end_io(bh, test_bit(BIO_UPTODATE, &bio->bi_flags)); 29301da177e4SLinus Torvalds bio_put(bio); 29311da177e4SLinus Torvalds } 29321da177e4SLinus Torvalds 29331da177e4SLinus Torvalds int submit_bh(int rw, struct buffer_head * bh) 29341da177e4SLinus Torvalds { 29351da177e4SLinus Torvalds struct bio *bio; 29361da177e4SLinus Torvalds int ret = 0; 29371da177e4SLinus Torvalds 29381da177e4SLinus Torvalds BUG_ON(!buffer_locked(bh)); 29391da177e4SLinus Torvalds BUG_ON(!buffer_mapped(bh)); 29401da177e4SLinus Torvalds BUG_ON(!bh->b_end_io); 29418fb0e342SAneesh Kumar K.V BUG_ON(buffer_delay(bh)); 29428fb0e342SAneesh Kumar K.V BUG_ON(buffer_unwritten(bh)); 29431da177e4SLinus Torvalds 294448fd4f93SJens Axboe /* 294548fd4f93SJens Axboe * Mask in barrier bit for a write (could be either a WRITE or a 294648fd4f93SJens Axboe * WRITE_SYNC 294748fd4f93SJens Axboe */ 294848fd4f93SJens Axboe if (buffer_ordered(bh) && (rw & WRITE)) 294948fd4f93SJens Axboe rw |= WRITE_BARRIER; 29501da177e4SLinus Torvalds 29511da177e4SLinus Torvalds /* 295248fd4f93SJens Axboe * Only clear out a write error when rewriting 29531da177e4SLinus Torvalds */ 295448fd4f93SJens Axboe if (test_set_buffer_req(bh) && (rw & WRITE)) 29551da177e4SLinus Torvalds clear_buffer_write_io_error(bh); 29561da177e4SLinus Torvalds 29571da177e4SLinus Torvalds /* 29581da177e4SLinus Torvalds * from here on down, it's all bio -- do the initial mapping, 29591da177e4SLinus Torvalds * submit_bio -> generic_make_request may further map this bio around 29601da177e4SLinus Torvalds */ 29611da177e4SLinus Torvalds bio = bio_alloc(GFP_NOIO, 1); 29621da177e4SLinus Torvalds 29631da177e4SLinus Torvalds bio->bi_sector = bh->b_blocknr * (bh->b_size >> 9); 29641da177e4SLinus Torvalds bio->bi_bdev = bh->b_bdev; 29651da177e4SLinus Torvalds bio->bi_io_vec[0].bv_page = bh->b_page; 29661da177e4SLinus Torvalds bio->bi_io_vec[0].bv_len = bh->b_size; 29671da177e4SLinus Torvalds bio->bi_io_vec[0].bv_offset = bh_offset(bh); 29681da177e4SLinus Torvalds 29691da177e4SLinus Torvalds bio->bi_vcnt = 1; 29701da177e4SLinus Torvalds bio->bi_idx = 0; 29711da177e4SLinus Torvalds bio->bi_size = bh->b_size; 29721da177e4SLinus Torvalds 29731da177e4SLinus Torvalds bio->bi_end_io = end_bio_bh_io_sync; 29741da177e4SLinus Torvalds bio->bi_private = bh; 29751da177e4SLinus Torvalds 29761da177e4SLinus Torvalds bio_get(bio); 29771da177e4SLinus Torvalds submit_bio(rw, bio); 29781da177e4SLinus Torvalds 29791da177e4SLinus Torvalds if (bio_flagged(bio, BIO_EOPNOTSUPP)) 29801da177e4SLinus Torvalds ret = -EOPNOTSUPP; 29811da177e4SLinus Torvalds 29821da177e4SLinus Torvalds bio_put(bio); 29831da177e4SLinus Torvalds return ret; 29841da177e4SLinus Torvalds } 29851da177e4SLinus Torvalds 29861da177e4SLinus Torvalds /** 29871da177e4SLinus Torvalds * ll_rw_block: low-level access to block devices (DEPRECATED) 2988a7662236SJan Kara * @rw: whether to %READ or %WRITE or %SWRITE or maybe %READA (readahead) 29891da177e4SLinus Torvalds * @nr: number of &struct buffer_heads in the array 29901da177e4SLinus Torvalds * @bhs: array of pointers to &struct buffer_head 29911da177e4SLinus Torvalds * 2992a7662236SJan Kara * ll_rw_block() takes an array of pointers to &struct buffer_heads, and 2993a7662236SJan Kara * requests an I/O operation on them, either a %READ or a %WRITE. The third 2994a7662236SJan Kara * %SWRITE is like %WRITE only we make sure that the *current* data in buffers 2995a7662236SJan Kara * are sent to disk. The fourth %READA option is described in the documentation 2996a7662236SJan Kara * for generic_make_request() which ll_rw_block() calls. 29971da177e4SLinus Torvalds * 29981da177e4SLinus Torvalds * This function drops any buffer that it cannot get a lock on (with the 2999a7662236SJan Kara * BH_Lock state bit) unless SWRITE is required, any buffer that appears to be 3000a7662236SJan Kara * clean when doing a write request, and any buffer that appears to be 3001a7662236SJan Kara * up-to-date when doing read request. Further it marks as clean buffers that 3002a7662236SJan Kara * are processed for writing (the buffer cache won't assume that they are 3003a7662236SJan Kara * actually clean until the buffer gets unlocked). 30041da177e4SLinus Torvalds * 30051da177e4SLinus Torvalds * ll_rw_block sets b_end_io to simple completion handler that marks 30061da177e4SLinus Torvalds * the buffer up-to-date (if approriate), unlocks the buffer and wakes 30071da177e4SLinus Torvalds * any waiters. 30081da177e4SLinus Torvalds * 30091da177e4SLinus Torvalds * All of the buffers must be for the same device, and must also be a 30101da177e4SLinus Torvalds * multiple of the current approved size for the device. 30111da177e4SLinus Torvalds */ 30121da177e4SLinus Torvalds void ll_rw_block(int rw, int nr, struct buffer_head *bhs[]) 30131da177e4SLinus Torvalds { 30141da177e4SLinus Torvalds int i; 30151da177e4SLinus Torvalds 30161da177e4SLinus Torvalds for (i = 0; i < nr; i++) { 30171da177e4SLinus Torvalds struct buffer_head *bh = bhs[i]; 30181da177e4SLinus Torvalds 30199cf6b720SJens Axboe if (rw == SWRITE || rw == SWRITE_SYNC || rw == SWRITE_SYNC_PLUG) 3020a7662236SJan Kara lock_buffer(bh); 3021ca5de404SNick Piggin else if (!trylock_buffer(bh)) 30221da177e4SLinus Torvalds continue; 30231da177e4SLinus Torvalds 30249cf6b720SJens Axboe if (rw == WRITE || rw == SWRITE || rw == SWRITE_SYNC || 30259cf6b720SJens Axboe rw == SWRITE_SYNC_PLUG) { 30261da177e4SLinus Torvalds if (test_clear_buffer_dirty(bh)) { 302776c3073aSakpm@osdl.org bh->b_end_io = end_buffer_write_sync; 3028e60e5c50SOGAWA Hirofumi get_bh(bh); 302918ce3751SJens Axboe if (rw == SWRITE_SYNC) 303018ce3751SJens Axboe submit_bh(WRITE_SYNC, bh); 303118ce3751SJens Axboe else 30321da177e4SLinus Torvalds submit_bh(WRITE, bh); 30331da177e4SLinus Torvalds continue; 30341da177e4SLinus Torvalds } 30351da177e4SLinus Torvalds } else { 30361da177e4SLinus Torvalds if (!buffer_uptodate(bh)) { 303776c3073aSakpm@osdl.org bh->b_end_io = end_buffer_read_sync; 3038e60e5c50SOGAWA Hirofumi get_bh(bh); 30391da177e4SLinus Torvalds submit_bh(rw, bh); 30401da177e4SLinus Torvalds continue; 30411da177e4SLinus Torvalds } 30421da177e4SLinus Torvalds } 30431da177e4SLinus Torvalds unlock_buffer(bh); 30441da177e4SLinus Torvalds } 30451da177e4SLinus Torvalds } 30461da177e4SLinus Torvalds 30471da177e4SLinus Torvalds /* 30481da177e4SLinus Torvalds * For a data-integrity writeout, we need to wait upon any in-progress I/O 30491da177e4SLinus Torvalds * and then start new I/O and then wait upon it. The caller must have a ref on 30501da177e4SLinus Torvalds * the buffer_head. 30511da177e4SLinus Torvalds */ 30521da177e4SLinus Torvalds int sync_dirty_buffer(struct buffer_head *bh) 30531da177e4SLinus Torvalds { 30541da177e4SLinus Torvalds int ret = 0; 30551da177e4SLinus Torvalds 30561da177e4SLinus Torvalds WARN_ON(atomic_read(&bh->b_count) < 1); 30571da177e4SLinus Torvalds lock_buffer(bh); 30581da177e4SLinus Torvalds if (test_clear_buffer_dirty(bh)) { 30591da177e4SLinus Torvalds get_bh(bh); 30601da177e4SLinus Torvalds bh->b_end_io = end_buffer_write_sync; 30611aa2a7ccSJens Axboe ret = submit_bh(WRITE_SYNC, bh); 30621da177e4SLinus Torvalds wait_on_buffer(bh); 30631da177e4SLinus Torvalds if (buffer_eopnotsupp(bh)) { 30641da177e4SLinus Torvalds clear_buffer_eopnotsupp(bh); 30651da177e4SLinus Torvalds ret = -EOPNOTSUPP; 30661da177e4SLinus Torvalds } 30671da177e4SLinus Torvalds if (!ret && !buffer_uptodate(bh)) 30681da177e4SLinus Torvalds ret = -EIO; 30691da177e4SLinus Torvalds } else { 30701da177e4SLinus Torvalds unlock_buffer(bh); 30711da177e4SLinus Torvalds } 30721da177e4SLinus Torvalds return ret; 30731da177e4SLinus Torvalds } 30741da177e4SLinus Torvalds 30751da177e4SLinus Torvalds /* 30761da177e4SLinus Torvalds * try_to_free_buffers() checks if all the buffers on this particular page 30771da177e4SLinus Torvalds * are unused, and releases them if so. 30781da177e4SLinus Torvalds * 30791da177e4SLinus Torvalds * Exclusion against try_to_free_buffers may be obtained by either 30801da177e4SLinus Torvalds * locking the page or by holding its mapping's private_lock. 30811da177e4SLinus Torvalds * 30821da177e4SLinus Torvalds * If the page is dirty but all the buffers are clean then we need to 30831da177e4SLinus Torvalds * be sure to mark the page clean as well. This is because the page 30841da177e4SLinus Torvalds * may be against a block device, and a later reattachment of buffers 30851da177e4SLinus Torvalds * to a dirty page will set *all* buffers dirty. Which would corrupt 30861da177e4SLinus Torvalds * filesystem data on the same device. 30871da177e4SLinus Torvalds * 30881da177e4SLinus Torvalds * The same applies to regular filesystem pages: if all the buffers are 30891da177e4SLinus Torvalds * clean then we set the page clean and proceed. To do that, we require 30901da177e4SLinus Torvalds * total exclusion from __set_page_dirty_buffers(). That is obtained with 30911da177e4SLinus Torvalds * private_lock. 30921da177e4SLinus Torvalds * 30931da177e4SLinus Torvalds * try_to_free_buffers() is non-blocking. 30941da177e4SLinus Torvalds */ 30951da177e4SLinus Torvalds static inline int buffer_busy(struct buffer_head *bh) 30961da177e4SLinus Torvalds { 30971da177e4SLinus Torvalds return atomic_read(&bh->b_count) | 30981da177e4SLinus Torvalds (bh->b_state & ((1 << BH_Dirty) | (1 << BH_Lock))); 30991da177e4SLinus Torvalds } 31001da177e4SLinus Torvalds 31011da177e4SLinus Torvalds static int 31021da177e4SLinus Torvalds drop_buffers(struct page *page, struct buffer_head **buffers_to_free) 31031da177e4SLinus Torvalds { 31041da177e4SLinus Torvalds struct buffer_head *head = page_buffers(page); 31051da177e4SLinus Torvalds struct buffer_head *bh; 31061da177e4SLinus Torvalds 31071da177e4SLinus Torvalds bh = head; 31081da177e4SLinus Torvalds do { 3109de7d5a3bSakpm@osdl.org if (buffer_write_io_error(bh) && page->mapping) 31101da177e4SLinus Torvalds set_bit(AS_EIO, &page->mapping->flags); 31111da177e4SLinus Torvalds if (buffer_busy(bh)) 31121da177e4SLinus Torvalds goto failed; 31131da177e4SLinus Torvalds bh = bh->b_this_page; 31141da177e4SLinus Torvalds } while (bh != head); 31151da177e4SLinus Torvalds 31161da177e4SLinus Torvalds do { 31171da177e4SLinus Torvalds struct buffer_head *next = bh->b_this_page; 31181da177e4SLinus Torvalds 3119535ee2fbSJan Kara if (bh->b_assoc_map) 31201da177e4SLinus Torvalds __remove_assoc_queue(bh); 31211da177e4SLinus Torvalds bh = next; 31221da177e4SLinus Torvalds } while (bh != head); 31231da177e4SLinus Torvalds *buffers_to_free = head; 31241da177e4SLinus Torvalds __clear_page_buffers(page); 31251da177e4SLinus Torvalds return 1; 31261da177e4SLinus Torvalds failed: 31271da177e4SLinus Torvalds return 0; 31281da177e4SLinus Torvalds } 31291da177e4SLinus Torvalds 31301da177e4SLinus Torvalds int try_to_free_buffers(struct page *page) 31311da177e4SLinus Torvalds { 31321da177e4SLinus Torvalds struct address_space * const mapping = page->mapping; 31331da177e4SLinus Torvalds struct buffer_head *buffers_to_free = NULL; 31341da177e4SLinus Torvalds int ret = 0; 31351da177e4SLinus Torvalds 31361da177e4SLinus Torvalds BUG_ON(!PageLocked(page)); 3137ecdfc978SLinus Torvalds if (PageWriteback(page)) 31381da177e4SLinus Torvalds return 0; 31391da177e4SLinus Torvalds 31401da177e4SLinus Torvalds if (mapping == NULL) { /* can this still happen? */ 31411da177e4SLinus Torvalds ret = drop_buffers(page, &buffers_to_free); 31421da177e4SLinus Torvalds goto out; 31431da177e4SLinus Torvalds } 31441da177e4SLinus Torvalds 31451da177e4SLinus Torvalds spin_lock(&mapping->private_lock); 31461da177e4SLinus Torvalds ret = drop_buffers(page, &buffers_to_free); 3147ecdfc978SLinus Torvalds 3148ecdfc978SLinus Torvalds /* 3149ecdfc978SLinus Torvalds * If the filesystem writes its buffers by hand (eg ext3) 3150ecdfc978SLinus Torvalds * then we can have clean buffers against a dirty page. We 3151ecdfc978SLinus Torvalds * clean the page here; otherwise the VM will never notice 3152ecdfc978SLinus Torvalds * that the filesystem did any IO at all. 3153ecdfc978SLinus Torvalds * 3154ecdfc978SLinus Torvalds * Also, during truncate, discard_buffer will have marked all 3155ecdfc978SLinus Torvalds * the page's buffers clean. We discover that here and clean 3156ecdfc978SLinus Torvalds * the page also. 315787df7241SNick Piggin * 315887df7241SNick Piggin * private_lock must be held over this entire operation in order 315987df7241SNick Piggin * to synchronise against __set_page_dirty_buffers and prevent the 316087df7241SNick Piggin * dirty bit from being lost. 3161ecdfc978SLinus Torvalds */ 3162ecdfc978SLinus Torvalds if (ret) 3163ecdfc978SLinus Torvalds cancel_dirty_page(page, PAGE_CACHE_SIZE); 316487df7241SNick Piggin spin_unlock(&mapping->private_lock); 31651da177e4SLinus Torvalds out: 31661da177e4SLinus Torvalds if (buffers_to_free) { 31671da177e4SLinus Torvalds struct buffer_head *bh = buffers_to_free; 31681da177e4SLinus Torvalds 31691da177e4SLinus Torvalds do { 31701da177e4SLinus Torvalds struct buffer_head *next = bh->b_this_page; 31711da177e4SLinus Torvalds free_buffer_head(bh); 31721da177e4SLinus Torvalds bh = next; 31731da177e4SLinus Torvalds } while (bh != buffers_to_free); 31741da177e4SLinus Torvalds } 31751da177e4SLinus Torvalds return ret; 31761da177e4SLinus Torvalds } 31771da177e4SLinus Torvalds EXPORT_SYMBOL(try_to_free_buffers); 31781da177e4SLinus Torvalds 31793978d717SNeilBrown void block_sync_page(struct page *page) 31801da177e4SLinus Torvalds { 31811da177e4SLinus Torvalds struct address_space *mapping; 31821da177e4SLinus Torvalds 31831da177e4SLinus Torvalds smp_mb(); 31841da177e4SLinus Torvalds mapping = page_mapping(page); 31851da177e4SLinus Torvalds if (mapping) 31861da177e4SLinus Torvalds blk_run_backing_dev(mapping->backing_dev_info, page); 31871da177e4SLinus Torvalds } 31881da177e4SLinus Torvalds 31891da177e4SLinus Torvalds /* 31901da177e4SLinus Torvalds * There are no bdflush tunables left. But distributions are 31911da177e4SLinus Torvalds * still running obsolete flush daemons, so we terminate them here. 31921da177e4SLinus Torvalds * 31931da177e4SLinus Torvalds * Use of bdflush() is deprecated and will be removed in a future kernel. 31941da177e4SLinus Torvalds * The `pdflush' kernel threads fully replace bdflush daemons and this call. 31951da177e4SLinus Torvalds */ 3196bdc480e3SHeiko Carstens SYSCALL_DEFINE2(bdflush, int, func, long, data) 31971da177e4SLinus Torvalds { 31981da177e4SLinus Torvalds static int msg_count; 31991da177e4SLinus Torvalds 32001da177e4SLinus Torvalds if (!capable(CAP_SYS_ADMIN)) 32011da177e4SLinus Torvalds return -EPERM; 32021da177e4SLinus Torvalds 32031da177e4SLinus Torvalds if (msg_count < 5) { 32041da177e4SLinus Torvalds msg_count++; 32051da177e4SLinus Torvalds printk(KERN_INFO 32061da177e4SLinus Torvalds "warning: process `%s' used the obsolete bdflush" 32071da177e4SLinus Torvalds " system call\n", current->comm); 32081da177e4SLinus Torvalds printk(KERN_INFO "Fix your initscripts?\n"); 32091da177e4SLinus Torvalds } 32101da177e4SLinus Torvalds 32111da177e4SLinus Torvalds if (func == 1) 32121da177e4SLinus Torvalds do_exit(0); 32131da177e4SLinus Torvalds return 0; 32141da177e4SLinus Torvalds } 32151da177e4SLinus Torvalds 32161da177e4SLinus Torvalds /* 32171da177e4SLinus Torvalds * Buffer-head allocation 32181da177e4SLinus Torvalds */ 3219e18b890bSChristoph Lameter static struct kmem_cache *bh_cachep; 32201da177e4SLinus Torvalds 32211da177e4SLinus Torvalds /* 32221da177e4SLinus Torvalds * Once the number of bh's in the machine exceeds this level, we start 32231da177e4SLinus Torvalds * stripping them in writeback. 32241da177e4SLinus Torvalds */ 32251da177e4SLinus Torvalds static int max_buffer_heads; 32261da177e4SLinus Torvalds 32271da177e4SLinus Torvalds int buffer_heads_over_limit; 32281da177e4SLinus Torvalds 32291da177e4SLinus Torvalds struct bh_accounting { 32301da177e4SLinus Torvalds int nr; /* Number of live bh's */ 32311da177e4SLinus Torvalds int ratelimit; /* Limit cacheline bouncing */ 32321da177e4SLinus Torvalds }; 32331da177e4SLinus Torvalds 32341da177e4SLinus Torvalds static DEFINE_PER_CPU(struct bh_accounting, bh_accounting) = {0, 0}; 32351da177e4SLinus Torvalds 32361da177e4SLinus Torvalds static void recalc_bh_state(void) 32371da177e4SLinus Torvalds { 32381da177e4SLinus Torvalds int i; 32391da177e4SLinus Torvalds int tot = 0; 32401da177e4SLinus Torvalds 32411da177e4SLinus Torvalds if (__get_cpu_var(bh_accounting).ratelimit++ < 4096) 32421da177e4SLinus Torvalds return; 32431da177e4SLinus Torvalds __get_cpu_var(bh_accounting).ratelimit = 0; 32448a143426SEric Dumazet for_each_online_cpu(i) 32451da177e4SLinus Torvalds tot += per_cpu(bh_accounting, i).nr; 32461da177e4SLinus Torvalds buffer_heads_over_limit = (tot > max_buffer_heads); 32471da177e4SLinus Torvalds } 32481da177e4SLinus Torvalds 3249dd0fc66fSAl Viro struct buffer_head *alloc_buffer_head(gfp_t gfp_flags) 32501da177e4SLinus Torvalds { 3251488514d1SChristoph Lameter struct buffer_head *ret = kmem_cache_alloc(bh_cachep, gfp_flags); 32521da177e4SLinus Torvalds if (ret) { 3253a35afb83SChristoph Lameter INIT_LIST_HEAD(&ret->b_assoc_buffers); 3254736c7b80SCoywolf Qi Hunt get_cpu_var(bh_accounting).nr++; 32551da177e4SLinus Torvalds recalc_bh_state(); 3256736c7b80SCoywolf Qi Hunt put_cpu_var(bh_accounting); 32571da177e4SLinus Torvalds } 32581da177e4SLinus Torvalds return ret; 32591da177e4SLinus Torvalds } 32601da177e4SLinus Torvalds EXPORT_SYMBOL(alloc_buffer_head); 32611da177e4SLinus Torvalds 32621da177e4SLinus Torvalds void free_buffer_head(struct buffer_head *bh) 32631da177e4SLinus Torvalds { 32641da177e4SLinus Torvalds BUG_ON(!list_empty(&bh->b_assoc_buffers)); 32651da177e4SLinus Torvalds kmem_cache_free(bh_cachep, bh); 3266736c7b80SCoywolf Qi Hunt get_cpu_var(bh_accounting).nr--; 32671da177e4SLinus Torvalds recalc_bh_state(); 3268736c7b80SCoywolf Qi Hunt put_cpu_var(bh_accounting); 32691da177e4SLinus Torvalds } 32701da177e4SLinus Torvalds EXPORT_SYMBOL(free_buffer_head); 32711da177e4SLinus Torvalds 32721da177e4SLinus Torvalds static void buffer_exit_cpu(int cpu) 32731da177e4SLinus Torvalds { 32741da177e4SLinus Torvalds int i; 32751da177e4SLinus Torvalds struct bh_lru *b = &per_cpu(bh_lrus, cpu); 32761da177e4SLinus Torvalds 32771da177e4SLinus Torvalds for (i = 0; i < BH_LRU_SIZE; i++) { 32781da177e4SLinus Torvalds brelse(b->bhs[i]); 32791da177e4SLinus Torvalds b->bhs[i] = NULL; 32801da177e4SLinus Torvalds } 32818a143426SEric Dumazet get_cpu_var(bh_accounting).nr += per_cpu(bh_accounting, cpu).nr; 32828a143426SEric Dumazet per_cpu(bh_accounting, cpu).nr = 0; 32838a143426SEric Dumazet put_cpu_var(bh_accounting); 32841da177e4SLinus Torvalds } 32851da177e4SLinus Torvalds 32861da177e4SLinus Torvalds static int buffer_cpu_notify(struct notifier_block *self, 32871da177e4SLinus Torvalds unsigned long action, void *hcpu) 32881da177e4SLinus Torvalds { 32898bb78442SRafael J. Wysocki if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) 32901da177e4SLinus Torvalds buffer_exit_cpu((unsigned long)hcpu); 32911da177e4SLinus Torvalds return NOTIFY_OK; 32921da177e4SLinus Torvalds } 32931da177e4SLinus Torvalds 3294389d1b08SAneesh Kumar K.V /** 3295a6b91919SRandy Dunlap * bh_uptodate_or_lock - Test whether the buffer is uptodate 3296389d1b08SAneesh Kumar K.V * @bh: struct buffer_head 3297389d1b08SAneesh Kumar K.V * 3298389d1b08SAneesh Kumar K.V * Return true if the buffer is up-to-date and false, 3299389d1b08SAneesh Kumar K.V * with the buffer locked, if not. 3300389d1b08SAneesh Kumar K.V */ 3301389d1b08SAneesh Kumar K.V int bh_uptodate_or_lock(struct buffer_head *bh) 3302389d1b08SAneesh Kumar K.V { 3303389d1b08SAneesh Kumar K.V if (!buffer_uptodate(bh)) { 3304389d1b08SAneesh Kumar K.V lock_buffer(bh); 3305389d1b08SAneesh Kumar K.V if (!buffer_uptodate(bh)) 3306389d1b08SAneesh Kumar K.V return 0; 3307389d1b08SAneesh Kumar K.V unlock_buffer(bh); 3308389d1b08SAneesh Kumar K.V } 3309389d1b08SAneesh Kumar K.V return 1; 3310389d1b08SAneesh Kumar K.V } 3311389d1b08SAneesh Kumar K.V EXPORT_SYMBOL(bh_uptodate_or_lock); 3312389d1b08SAneesh Kumar K.V 3313389d1b08SAneesh Kumar K.V /** 3314a6b91919SRandy Dunlap * bh_submit_read - Submit a locked buffer for reading 3315389d1b08SAneesh Kumar K.V * @bh: struct buffer_head 3316389d1b08SAneesh Kumar K.V * 3317389d1b08SAneesh Kumar K.V * Returns zero on success and -EIO on error. 3318389d1b08SAneesh Kumar K.V */ 3319389d1b08SAneesh Kumar K.V int bh_submit_read(struct buffer_head *bh) 3320389d1b08SAneesh Kumar K.V { 3321389d1b08SAneesh Kumar K.V BUG_ON(!buffer_locked(bh)); 3322389d1b08SAneesh Kumar K.V 3323389d1b08SAneesh Kumar K.V if (buffer_uptodate(bh)) { 3324389d1b08SAneesh Kumar K.V unlock_buffer(bh); 3325389d1b08SAneesh Kumar K.V return 0; 3326389d1b08SAneesh Kumar K.V } 3327389d1b08SAneesh Kumar K.V 3328389d1b08SAneesh Kumar K.V get_bh(bh); 3329389d1b08SAneesh Kumar K.V bh->b_end_io = end_buffer_read_sync; 3330389d1b08SAneesh Kumar K.V submit_bh(READ, bh); 3331389d1b08SAneesh Kumar K.V wait_on_buffer(bh); 3332389d1b08SAneesh Kumar K.V if (buffer_uptodate(bh)) 3333389d1b08SAneesh Kumar K.V return 0; 3334389d1b08SAneesh Kumar K.V return -EIO; 3335389d1b08SAneesh Kumar K.V } 3336389d1b08SAneesh Kumar K.V EXPORT_SYMBOL(bh_submit_read); 3337389d1b08SAneesh Kumar K.V 3338b98938c3SChristoph Lameter static void 333951cc5068SAlexey Dobriyan init_buffer_head(void *data) 3340b98938c3SChristoph Lameter { 3341b98938c3SChristoph Lameter struct buffer_head *bh = data; 3342b98938c3SChristoph Lameter 3343b98938c3SChristoph Lameter memset(bh, 0, sizeof(*bh)); 3344b98938c3SChristoph Lameter INIT_LIST_HEAD(&bh->b_assoc_buffers); 3345b98938c3SChristoph Lameter } 3346b98938c3SChristoph Lameter 33471da177e4SLinus Torvalds void __init buffer_init(void) 33481da177e4SLinus Torvalds { 33491da177e4SLinus Torvalds int nrpages; 33501da177e4SLinus Torvalds 3351b98938c3SChristoph Lameter bh_cachep = kmem_cache_create("buffer_head", 3352b98938c3SChristoph Lameter sizeof(struct buffer_head), 0, 3353b98938c3SChristoph Lameter (SLAB_RECLAIM_ACCOUNT|SLAB_PANIC| 3354b98938c3SChristoph Lameter SLAB_MEM_SPREAD), 3355b98938c3SChristoph Lameter init_buffer_head); 33561da177e4SLinus Torvalds 33571da177e4SLinus Torvalds /* 33581da177e4SLinus Torvalds * Limit the bh occupancy to 10% of ZONE_NORMAL 33591da177e4SLinus Torvalds */ 33601da177e4SLinus Torvalds nrpages = (nr_free_buffer_pages() * 10) / 100; 33611da177e4SLinus Torvalds max_buffer_heads = nrpages * (PAGE_SIZE / sizeof(struct buffer_head)); 33621da177e4SLinus Torvalds hotcpu_notifier(buffer_cpu_notify, 0); 33631da177e4SLinus Torvalds } 33641da177e4SLinus Torvalds 33651da177e4SLinus Torvalds EXPORT_SYMBOL(__bforget); 33661da177e4SLinus Torvalds EXPORT_SYMBOL(__brelse); 33671da177e4SLinus Torvalds EXPORT_SYMBOL(__wait_on_buffer); 33681da177e4SLinus Torvalds EXPORT_SYMBOL(block_commit_write); 33691da177e4SLinus Torvalds EXPORT_SYMBOL(block_prepare_write); 337054171690SDavid Chinner EXPORT_SYMBOL(block_page_mkwrite); 33711da177e4SLinus Torvalds EXPORT_SYMBOL(block_read_full_page); 33721da177e4SLinus Torvalds EXPORT_SYMBOL(block_sync_page); 33731da177e4SLinus Torvalds EXPORT_SYMBOL(block_truncate_page); 33741da177e4SLinus Torvalds EXPORT_SYMBOL(block_write_full_page); 337535c80d5fSChris Mason EXPORT_SYMBOL(block_write_full_page_endio); 337689e10787SNick Piggin EXPORT_SYMBOL(cont_write_begin); 33771da177e4SLinus Torvalds EXPORT_SYMBOL(end_buffer_read_sync); 33781da177e4SLinus Torvalds EXPORT_SYMBOL(end_buffer_write_sync); 337935c80d5fSChris Mason EXPORT_SYMBOL(end_buffer_async_write); 33801da177e4SLinus Torvalds EXPORT_SYMBOL(file_fsync); 33811da177e4SLinus Torvalds EXPORT_SYMBOL(generic_block_bmap); 338205eb0b51SOGAWA Hirofumi EXPORT_SYMBOL(generic_cont_expand_simple); 33831da177e4SLinus Torvalds EXPORT_SYMBOL(init_buffer); 33841da177e4SLinus Torvalds EXPORT_SYMBOL(invalidate_bdev); 33851da177e4SLinus Torvalds EXPORT_SYMBOL(ll_rw_block); 33861da177e4SLinus Torvalds EXPORT_SYMBOL(mark_buffer_dirty); 33871da177e4SLinus Torvalds EXPORT_SYMBOL(submit_bh); 33881da177e4SLinus Torvalds EXPORT_SYMBOL(sync_dirty_buffer); 33891da177e4SLinus Torvalds EXPORT_SYMBOL(unlock_buffer); 3390