mm: revert "vmscan: get_scan_ratio() cleanup"
[linux-2.6.git] / mm / truncate.c
index 6650c1d..f42675a 100644 (file)
@@ -3,12 +3,13 @@
  *
  * 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/gfp.h>
 #include <linux/mm.h>
 #include <linux/swap.h>
 #include <linux/module.h>
@@ -18,6 +19,7 @@
 #include <linux/task_io_accounting_ops.h>
 #include <linux/buffer_head.h> /* grr. try_to_release_page,
                                   do_invalidatepage */
+#include "internal.h"
 
 
 /**
@@ -49,7 +51,7 @@ void do_invalidatepage(struct page *page, unsigned long offset)
 static inline void truncate_partial_page(struct page *page, unsigned partial)
 {
        zero_user_segment(page, partial, PAGE_CACHE_SIZE);
-       if (PagePrivate(page))
+       if (page_has_private(page))
                do_invalidatepage(page, partial);
 }
 
@@ -92,20 +94,22 @@ EXPORT_SYMBOL(cancel_dirty_page);
  * 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;
+               return -EIO;
 
-       if (PagePrivate(page))
+       if (page_has_private(page))
                do_invalidatepage(page, 0);
 
        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;
 }
 
 /*
@@ -124,14 +128,60 @@ 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 & end byte offsets
  * @mapping: mapping to truncate
@@ -193,12 +243,7 @@ void truncate_inode_pages_range(struct address_space *mapping,
                                unlock_page(page);
                                continue;
                        }
-                       if (page_mapped(page)) {
-                               unmap_mapping_range(mapping,
-                                 (loff_t)page_index<<PAGE_CACHE_SHIFT,
-                                 PAGE_CACHE_SIZE, 0);
-                       }
-                       truncate_complete_page(mapping, page);
+                       truncate_inode_page(mapping, page);
                        unlock_page(page);
                }
                pagevec_release(&pvec);
@@ -228,6 +273,7 @@ void truncate_inode_pages_range(struct address_space *mapping,
                        pagevec_release(&pvec);
                        break;
                }
+               mem_cgroup_uncharge_start();
                for (i = 0; i < pagevec_count(&pvec); i++) {
                        struct page *page = pvec.pages[i];
 
@@ -235,18 +281,14 @@ void truncate_inode_pages_range(struct address_space *mapping,
                                break;
                        lock_page(page);
                        wait_on_page_writeback(page);
-                       if (page_mapped(page)) {
-                               unmap_mapping_range(mapping,
-                                 (loff_t)page->index<<PAGE_CACHE_SHIFT,
-                                 PAGE_CACHE_SIZE, 0);
-                       }
+                       truncate_inode_page(mapping, page);
                        if (page->index > next)
                                next = page->index;
                        next++;
-                       truncate_complete_page(mapping, page);
                        unlock_page(page);
                }
                pagevec_release(&pvec);
+               mem_cgroup_uncharge_end();
        }
 }
 EXPORT_SYMBOL(truncate_inode_pages_range);
@@ -264,8 +306,21 @@ void truncate_inode_pages(struct address_space *mapping, loff_t lstart)
 }
 EXPORT_SYMBOL(truncate_inode_pages);
 
-unsigned long __invalidate_mapping_pages(struct address_space *mapping,
-                               pgoff_t start, pgoff_t end, bool be_atomic)
+/**
+ * invalidate_mapping_pages - Invalidate all the unlocked pages of one inode
+ * @mapping: the address_space which holds the pages to invalidate
+ * @start: the offset 'from' which to invalidate
+ * @end: the offset 'to' which to invalidate (inclusive)
+ *
+ * This function only removes the unlocked pages, if you want to
+ * remove all the pages of one inode, you must call truncate_inode_pages.
+ *
+ * invalidate_mapping_pages() will not block on IO activity. It will not
+ * invalidate pages which are dirty, locked, under writeback or mapped into
+ * pagetables.
+ */
+unsigned long invalidate_mapping_pages(struct address_space *mapping,
+                                      pgoff_t start, pgoff_t end)
 {
        struct pagevec pvec;
        pgoff_t next = start;
@@ -275,6 +330,7 @@ unsigned long __invalidate_mapping_pages(struct address_space *mapping,
        pagevec_init(&pvec, 0);
        while (next <= end &&
                        pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
+               mem_cgroup_uncharge_start();
                for (i = 0; i < pagevec_count(&pvec); i++) {
                        struct page *page = pvec.pages[i];
                        pgoff_t index;
@@ -295,41 +351,18 @@ 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);
-               if (likely(!be_atomic))
-                       cond_resched();
+               mem_cgroup_uncharge_end();
+               cond_resched();
        }
        return ret;
 }
-
-/**
- * invalidate_mapping_pages - Invalidate all the unlocked pages of one inode
- * @mapping: the address_space which holds the pages to invalidate
- * @start: the offset 'from' which to invalidate
- * @end: the offset 'to' which to invalidate (inclusive)
- *
- * This function only removes the unlocked pages, if you want to
- * remove all the pages of one inode, you must call truncate_inode_pages.
- *
- * invalidate_mapping_pages() will not block on IO activity. It will not
- * invalidate pages which are dirty, locked, under writeback or mapped into
- * pagetables.
- */
-unsigned long invalidate_mapping_pages(struct address_space *mapping,
-                               pgoff_t start, pgoff_t end)
-{
-       return __invalidate_mapping_pages(mapping, start, end, false);
-}
 EXPORT_SYMBOL(invalidate_mapping_pages);
 
 /*
@@ -345,16 +378,18 @@ 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;
 
        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);
        spin_unlock_irq(&mapping->tree_lock);
+       mem_cgroup_uncharge_cache_page(page);
        page_cache_release(page);       /* pagecache ref */
        return 1;
 failed:
@@ -398,6 +433,7 @@ int invalidate_inode_pages2_range(struct address_space *mapping,
        while (next <= end && !wrapped &&
                pagevec_lookup(&pvec, mapping, next,
                        min(end - next, (pgoff_t)PAGEVEC_SIZE - 1) + 1)) {
+               mem_cgroup_uncharge_start();
                for (i = 0; i < pagevec_count(&pvec); i++) {
                        struct page *page = pvec.pages[i];
                        pgoff_t page_index;
@@ -447,6 +483,7 @@ int invalidate_inode_pages2_range(struct address_space *mapping,
                        unlock_page(page);
                }
                pagevec_release(&pvec);
+               mem_cgroup_uncharge_end();
                cond_resched();
        }
        return ret;
@@ -460,10 +497,72 @@ EXPORT_SYMBOL_GPL(invalidate_inode_pages2_range);
  * 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(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)
+{
+       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);