GFS2: Remove dirent_first() function
[linux-2.6.git] / mm / truncate.c
index af3dcf0..450cebd 100644 (file)
@@ -3,11 +3,12 @@
  *
  * Copyright (C) 2002, Linus Torvalds
  *
- * 10Sep2002   akpm@zip.com.au
+ * 10Sep2002   Andrew Morton
  *             Initial version.
  */
 
 #include <linux/kernel.h>
+#include <linux/backing-dev.h>
 #include <linux/mm.h>
 #include <linux/swap.h>
 #include <linux/module.h>
 #include <linux/task_io_accounting_ops.h>
 #include <linux/buffer_head.h> /* grr. try_to_release_page,
                                   do_invalidatepage */
+#include "internal.h"
 
 
 /**
- * do_invalidatepage - invalidate part of all of a page
+ * do_invalidatepage - invalidate part or all of a page
  * @page: the page which is affected
  * @offset: the index of the truncation point
  *
@@ -47,8 +49,8 @@ void do_invalidatepage(struct page *page, unsigned long offset)
 
 static inline void truncate_partial_page(struct page *page, unsigned partial)
 {
-       zero_user_page(page, partial, PAGE_CACHE_SIZE - partial, KM_USER0);
-       if (PagePrivate(page))
+       zero_user_segment(page, partial, PAGE_CACHE_SIZE);
+       if (page_has_private(page))
                do_invalidatepage(page, partial);
 }
 
@@ -72,6 +74,8 @@ void cancel_dirty_page(struct page *page, unsigned int account_size)
                struct address_space *mapping = page->mapping;
                if (mapping && mapping_cap_account_dirty(mapping)) {
                        dec_zone_page_state(page, NR_FILE_DIRTY);
+                       dec_bdi_stat(mapping->backing_dev_info,
+                                       BDI_RECLAIMABLE);
                        if (account_size)
                                task_io_account_cancelled_write(account_size);
                }
@@ -81,29 +85,30 @@ EXPORT_SYMBOL(cancel_dirty_page);
 
 /*
  * If truncate cannot remove the fs-private metadata from the page, the page
- * becomes anonymous.  It will be left on the LRU and may even be mapped into
- * user pagetables if we're racing with filemap_nopage().
+ * becomes orphaned.  It will be left on the LRU and may even be mapped into
+ * user pagetables if we're racing with filemap_fault().
  *
  * We need to bale out if page->mapping is no longer equal to the original
  * mapping.  This happens a) when the VM reclaimed the page while we waited on
  * its lock, b) when a concurrent invalidate_mapping_pages got there first and
  * c) when tmpfs swizzles a page between a tmpfs inode and swapper_space.
  */
-static void
+static int
 truncate_complete_page(struct address_space *mapping, struct page *page)
 {
        if (page->mapping != mapping)
-               return;
-
-       cancel_dirty_page(page, PAGE_CACHE_SIZE);
+               return -EIO;
 
-       if (PagePrivate(page))
+       if (page_has_private(page))
                do_invalidatepage(page, 0);
 
-       ClearPageUptodate(page);
-       ClearPageMappedToDisk(page);
+       cancel_dirty_page(page, PAGE_CACHE_SIZE);
+
+       clear_page_mlock(page);
        remove_from_page_cache(page);
+       ClearPageMappedToDisk(page);
        page_cache_release(page);       /* pagecache ref */
+       return 0;
 }
 
 /*
@@ -122,17 +127,62 @@ invalidate_complete_page(struct address_space *mapping, struct page *page)
        if (page->mapping != mapping)
                return 0;
 
-       if (PagePrivate(page) && !try_to_release_page(page, 0))
+       if (page_has_private(page) && !try_to_release_page(page, 0))
                return 0;
 
+       clear_page_mlock(page);
        ret = remove_mapping(mapping, page);
 
        return ret;
 }
 
+int truncate_inode_page(struct address_space *mapping, struct page *page)
+{
+       if (page_mapped(page)) {
+               unmap_mapping_range(mapping,
+                                  (loff_t)page->index << PAGE_CACHE_SHIFT,
+                                  PAGE_CACHE_SIZE, 0);
+       }
+       return truncate_complete_page(mapping, page);
+}
+
+/*
+ * Used to get rid of pages on hardware memory corruption.
+ */
+int generic_error_remove_page(struct address_space *mapping, struct page *page)
+{
+       if (!mapping)
+               return -EINVAL;
+       /*
+        * Only punch for normal data pages for now.
+        * Handling other types like directories would need more auditing.
+        */
+       if (!S_ISREG(mapping->host->i_mode))
+               return -EIO;
+       return truncate_inode_page(mapping, page);
+}
+EXPORT_SYMBOL(generic_error_remove_page);
+
+/*
+ * Safely invalidate one page from its pagecache mapping.
+ * It only drops clean, unused pages. The page must be locked.
+ *
+ * Returns 1 if the page is successfully invalidated, otherwise 0.
+ */
+int invalidate_inode_page(struct page *page)
+{
+       struct address_space *mapping = page_mapping(page);
+       if (!mapping)
+               return 0;
+       if (PageDirty(page) || PageWriteback(page))
+               return 0;
+       if (page_mapped(page))
+               return 0;
+       return invalidate_complete_page(mapping, page);
+}
+
 /**
- * truncate_inode_pages - truncate range of pages specified by start and
- * end byte offsets
+ * truncate_inode_pages - truncate range of pages specified by start & end byte offsets
  * @mapping: mapping to truncate
  * @lstart: offset from which to truncate
  * @lend: offset to which to truncate
@@ -186,13 +236,13 @@ void truncate_inode_pages_range(struct address_space *mapping,
                        if (page_index > next)
                                next = page_index;
                        next++;
-                       if (TestSetPageLocked(page))
+                       if (!trylock_page(page))
                                continue;
                        if (PageWriteback(page)) {
                                unlock_page(page);
                                continue;
                        }
-                       truncate_complete_page(mapping, page);
+                       truncate_inode_page(mapping, page);
                        unlock_page(page);
                }
                pagevec_release(&pvec);
@@ -229,10 +279,10 @@ void truncate_inode_pages_range(struct address_space *mapping,
                                break;
                        lock_page(page);
                        wait_on_page_writeback(page);
+                       truncate_inode_page(mapping, page);
                        if (page->index > next)
                                next = page->index;
                        next++;
-                       truncate_complete_page(mapping, page);
                        unlock_page(page);
                }
                pagevec_release(&pvec);
@@ -267,7 +317,7 @@ EXPORT_SYMBOL(truncate_inode_pages);
  * pagetables.
  */
 unsigned long invalidate_mapping_pages(struct address_space *mapping,
-                               pgoff_t start, pgoff_t end)
+                                      pgoff_t start, pgoff_t end)
 {
        struct pagevec pvec;
        pgoff_t next = start;
@@ -282,7 +332,7 @@ unsigned long invalidate_mapping_pages(struct address_space *mapping,
                        pgoff_t index;
                        int lock_failed;
 
-                       lock_failed = TestSetPageLocked(page);
+                       lock_failed = !trylock_page(page);
 
                        /*
                         * We really shouldn't be looking at the ->index of an
@@ -297,17 +347,14 @@ unsigned long invalidate_mapping_pages(struct address_space *mapping,
                        if (lock_failed)
                                continue;
 
-                       if (PageDirty(page) || PageWriteback(page))
-                               goto unlock;
-                       if (page_mapped(page))
-                               goto unlock;
-                       ret += invalidate_complete_page(mapping, page);
-unlock:
+                       ret += invalidate_inode_page(page);
+
                        unlock_page(page);
                        if (next > end)
                                break;
                }
                pagevec_release(&pvec);
+               cond_resched();
        }
        return ret;
 }
@@ -326,21 +373,22 @@ invalidate_complete_page2(struct address_space *mapping, struct page *page)
        if (page->mapping != mapping)
                return 0;
 
-       if (PagePrivate(page) && !try_to_release_page(page, GFP_KERNEL))
+       if (page_has_private(page) && !try_to_release_page(page, GFP_KERNEL))
                return 0;
 
-       write_lock_irq(&mapping->tree_lock);
+       spin_lock_irq(&mapping->tree_lock);
        if (PageDirty(page))
                goto failed;
 
-       BUG_ON(PagePrivate(page));
+       clear_page_mlock(page);
+       BUG_ON(page_has_private(page));
        __remove_from_page_cache(page);
-       write_unlock_irq(&mapping->tree_lock);
-       ClearPageUptodate(page);
+       spin_unlock_irq(&mapping->tree_lock);
+       mem_cgroup_uncharge_cache_page(page);
        page_cache_release(page);       /* pagecache ref */
        return 1;
 failed:
-       write_unlock_irq(&mapping->tree_lock);
+       spin_unlock_irq(&mapping->tree_lock);
        return 0;
 }
 
@@ -362,7 +410,7 @@ static int do_launder_page(struct address_space *mapping, struct page *page)
  * Any pages which are found to be mapped into pagetables are unmapped prior to
  * invalidation.
  *
- * Returns -EIO if any pages could not be invalidated.
+ * Returns -EBUSY if any pages could not be invalidated.
  */
 int invalidate_inode_pages2_range(struct address_space *mapping,
                                  pgoff_t start, pgoff_t end)
@@ -371,6 +419,7 @@ int invalidate_inode_pages2_range(struct address_space *mapping,
        pgoff_t next;
        int i;
        int ret = 0;
+       int ret2 = 0;
        int did_range_unmap = 0;
        int wrapped = 0;
 
@@ -397,7 +446,7 @@ int invalidate_inode_pages2_range(struct address_space *mapping,
                                break;
                        }
                        wait_on_page_writeback(page);
-                       while (page_mapped(page)) {
+                       if (page_mapped(page)) {
                                if (!did_range_unmap) {
                                        /*
                                         * Zap the rest of the file in one hit.
@@ -417,9 +466,14 @@ int invalidate_inode_pages2_range(struct address_space *mapping,
                                          PAGE_CACHE_SIZE, 0);
                                }
                        }
-                       ret = do_launder_page(mapping, page);
-                       if (ret == 0 && !invalidate_complete_page2(mapping, page))
-                               ret = -EIO;
+                       BUG_ON(page_mapped(page));
+                       ret2 = do_launder_page(mapping, page);
+                       if (ret2 == 0) {
+                               if (!invalidate_complete_page2(mapping, page))
+                                       ret2 = -EBUSY;
+                       }
+                       if (ret2 < 0)
+                               ret = ret2;
                        unlock_page(page);
                }
                pagevec_release(&pvec);
@@ -443,3 +497,67 @@ int invalidate_inode_pages2(struct address_space *mapping)
        return invalidate_inode_pages2_range(mapping, 0, -1);
 }
 EXPORT_SYMBOL_GPL(invalidate_inode_pages2);
+
+/**
+ * truncate_pagecache - unmap and remove pagecache that has been truncated
+ * @inode: inode
+ * @old: old file offset
+ * @new: new file offset
+ *
+ * inode's new i_size must already be written before truncate_pagecache
+ * is called.
+ *
+ * This function should typically be called before the filesystem
+ * releases resources associated with the freed range (eg. deallocates
+ * blocks). This way, pagecache will always stay logically coherent
+ * with on-disk format, and the filesystem would not have to deal with
+ * situations such as writepage being called for a page that has already
+ * had its underlying blocks deallocated.
+ */
+void truncate_pagecache(struct inode *inode, loff_t old, loff_t new)
+{
+       if (new < old) {
+               struct address_space *mapping = inode->i_mapping;
+
+               /*
+                * unmap_mapping_range is called twice, first simply for
+                * efficiency so that truncate_inode_pages does fewer
+                * single-page unmaps.  However after this first call, and
+                * before truncate_inode_pages finishes, it is possible for
+                * private pages to be COWed, which remain after
+                * truncate_inode_pages finishes, hence the second
+                * unmap_mapping_range call must be made for correctness.
+                */
+               unmap_mapping_range(mapping, new + PAGE_SIZE - 1, 0, 1);
+               truncate_inode_pages(mapping, new);
+               unmap_mapping_range(mapping, new + PAGE_SIZE - 1, 0, 1);
+       }
+}
+EXPORT_SYMBOL(truncate_pagecache);
+
+/**
+ * vmtruncate - unmap mappings "freed" by truncate() syscall
+ * @inode: inode of the file used
+ * @offset: file offset to start truncating
+ *
+ * NOTE! We have to be ready to update the memory sharing
+ * between the file and the memory map for a potential last
+ * incomplete page.  Ugly, but necessary.
+ */
+int vmtruncate(struct inode *inode, loff_t offset)
+{
+       loff_t oldsize;
+       int error;
+
+       error = inode_newsize_ok(inode, offset);
+       if (error)
+               return error;
+       oldsize = inode->i_size;
+       i_size_write(inode, offset);
+       truncate_pagecache(inode, oldsize, offset);
+       if (inode->i_op->truncate)
+               inode->i_op->truncate(inode);
+
+       return error;
+}
+EXPORT_SYMBOL(vmtruncate);