tmpfs: copy truncate_inode_pages_range
[linux-2.6.git] / mm / truncate.c
index f47e46d..232eb27 100644 (file)
@@ -3,11 +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>
 #include <linux/task_io_accounting_ops.h>
 #include <linux/buffer_head.h> /* grr. try_to_release_page,
                                   do_invalidatepage */
+#include <linux/cleancache.h>
+#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 +51,9 @@ 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);
+       cleancache_flush_page(page->mapping, page);
+       if (page_has_private(page))
                do_invalidatepage(page, partial);
 }
 
@@ -72,6 +77,8 @@ 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 +88,29 @@ 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);
 
-       remove_from_page_cache(page);
-       ClearPageUptodate(page);
+       cancel_dirty_page(page, PAGE_CACHE_SIZE);
+
+       clear_page_mlock(page);
        ClearPageMappedToDisk(page);
-       page_cache_release(page);       /* pagecache ref */
+       delete_from_page_cache(page);
+       return 0;
 }
 
 /*
@@ -122,17 +129,62 @@ 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
@@ -147,9 +199,6 @@ invalidate_complete_page(struct address_space *mapping, struct page *page)
  * The first pass will remove most pages, so the search cost of the second pass
  * is low.
  *
- * When looking at page->index outside the page lock we need to be careful to
- * copy it into a local to avoid races (it could change at any time).
- *
  * We pass down the cache-hot hint to the page freeing code.  Even if the
  * mapping is large, it is probably the case that the final pages are the most
  * recently touched, and freeing happens in ascending file offset order.
@@ -158,12 +207,13 @@ void truncate_inode_pages_range(struct address_space *mapping,
                                loff_t lstart, loff_t lend)
 {
        const pgoff_t start = (lstart + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT;
-       pgoff_t end;
        const unsigned partial = lstart & (PAGE_CACHE_SIZE - 1);
        struct pagevec pvec;
-       pgoff_t next;
+       pgoff_t index;
+       pgoff_t end;
        int i;
 
+       cleancache_flush_inode(mapping);
        if (mapping->nrpages == 0)
                return;
 
@@ -171,32 +221,32 @@ void truncate_inode_pages_range(struct address_space *mapping,
        end = (lend >> PAGE_CACHE_SHIFT);
 
        pagevec_init(&pvec, 0);
-       next = start;
-       while (next <= end &&
-              pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
+       index = start;
+       while (index <= end && pagevec_lookup(&pvec, mapping, index,
+                       min(end - index, (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 = page->index;
 
-                       if (page_index > end) {
-                               next = page_index;
+                       /* We rely upon deletion not changing page->index */
+                       index = page->index;
+                       if (index > end)
                                break;
-                       }
 
-                       if (page_index > next)
-                               next = page_index;
-                       next++;
-                       if (TestSetPageLocked(page))
+                       if (!trylock_page(page))
                                continue;
+                       WARN_ON(page->index != index);
                        if (PageWriteback(page)) {
                                unlock_page(page);
                                continue;
                        }
-                       truncate_complete_page(mapping, page);
+                       truncate_inode_page(mapping, page);
                        unlock_page(page);
                }
                pagevec_release(&pvec);
+               mem_cgroup_uncharge_end();
                cond_resched();
+               index++;
        }
 
        if (partial) {
@@ -209,34 +259,40 @@ void truncate_inode_pages_range(struct address_space *mapping,
                }
        }
 
-       next = start;
+       index = start;
        for ( ; ; ) {
                cond_resched();
-               if (!pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
-                       if (next == start)
+               if (!pagevec_lookup(&pvec, mapping, index,
+                       min(end - index, (pgoff_t)PAGEVEC_SIZE - 1) + 1)) {
+                       if (index == start)
                                break;
-                       next = start;
+                       index = start;
                        continue;
                }
-               if (pvec.pages[0]->index > end) {
+               if (index == start && pvec.pages[0]->index > end) {
                        pagevec_release(&pvec);
                        break;
                }
+               mem_cgroup_uncharge_start();
                for (i = 0; i < pagevec_count(&pvec); i++) {
                        struct page *page = pvec.pages[i];
 
-                       if (page->index > end)
+                       /* We rely upon deletion not changing page->index */
+                       index = page->index;
+                       if (index > end)
                                break;
+
                        lock_page(page);
+                       WARN_ON(page->index != index);
                        wait_on_page_writeback(page);
-                       if (page->index > next)
-                               next = page->index;
-                       next++;
-                       truncate_complete_page(mapping, page);
+                       truncate_inode_page(mapping, page);
                        unlock_page(page);
                }
                pagevec_release(&pvec);
+               mem_cgroup_uncharge_end();
+               index++;
        }
+       cleancache_flush_inode(mapping);
 }
 EXPORT_SYMBOL(truncate_inode_pages_range);
 
@@ -246,6 +302,11 @@ EXPORT_SYMBOL(truncate_inode_pages_range);
  * @lstart: offset from which to truncate
  *
  * Called under (and serialised by) inode->i_mutex.
+ *
+ * Note: When this function returns, there can be a page in the process of
+ * deletion (inside __delete_from_page_cache()) in the specified range.  Thus
+ * mapping->nrpages can be non-zero when this function returns even after
+ * truncation of the whole mapping.
  */
 void truncate_inode_pages(struct address_space *mapping, loff_t lstart)
 {
@@ -253,54 +314,6 @@ 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)
-{
-       struct pagevec pvec;
-       pgoff_t next = start;
-       unsigned long ret = 0;
-       int i;
-
-       pagevec_init(&pvec, 0);
-       while (next <= end &&
-                       pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
-               for (i = 0; i < pagevec_count(&pvec); i++) {
-                       struct page *page = pvec.pages[i];
-                       pgoff_t index;
-                       int lock_failed;
-
-                       lock_failed = TestSetPageLocked(page);
-
-                       /*
-                        * We really shouldn't be looking at the ->index of an
-                        * unlocked page.  But we're not allowed to lock these
-                        * pages.  So we rely upon nobody altering the ->index
-                        * of this (pinned-by-us) page.
-                        */
-                       index = page->index;
-                       if (index > next)
-                               next = index;
-                       next++;
-                       if (lock_failed)
-                               continue;
-
-                       if (PageDirty(page) || PageWriteback(page))
-                               goto unlock;
-                       if (page_mapped(page))
-                               goto unlock;
-                       ret += invalidate_complete_page(mapping, page);
-unlock:
-                       unlock_page(page);
-                       if (next > end)
-                               break;
-               }
-               pagevec_release(&pvec);
-               if (likely(!be_atomic))
-                       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
@@ -315,9 +328,45 @@ unlock:
  * pagetables.
  */
 unsigned long invalidate_mapping_pages(struct address_space *mapping,
-                               pgoff_t start, pgoff_t end)
+               pgoff_t start, pgoff_t end)
 {
-       return __invalidate_mapping_pages(mapping, start, end, false);
+       struct pagevec pvec;
+       pgoff_t index = start;
+       unsigned long ret;
+       unsigned long count = 0;
+       int i;
+
+       pagevec_init(&pvec, 0);
+       while (index <= end && pagevec_lookup(&pvec, mapping, index,
+                       min(end - index, (pgoff_t)PAGEVEC_SIZE - 1) + 1)) {
+               mem_cgroup_uncharge_start();
+               for (i = 0; i < pagevec_count(&pvec); i++) {
+                       struct page *page = pvec.pages[i];
+
+                       /* We rely upon deletion not changing page->index */
+                       index = page->index;
+                       if (index > end)
+                               break;
+
+                       if (!trylock_page(page))
+                               continue;
+                       WARN_ON(page->index != index);
+                       ret = invalidate_inode_page(page);
+                       unlock_page(page);
+                       /*
+                        * Invalidation is a hint that the page is no longer
+                        * of interest and try to speed up its reclaim.
+                        */
+                       if (!ret)
+                               deactivate_page(page);
+                       count += ret;
+               }
+               pagevec_release(&pvec);
+               mem_cgroup_uncharge_end();
+               cond_resched();
+               index++;
+       }
+       return count;
 }
 EXPORT_SYMBOL(invalidate_mapping_pages);
 
@@ -334,21 +383,26 @@ 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));
-       __remove_from_page_cache(page);
-       write_unlock_irq(&mapping->tree_lock);
-       ClearPageUptodate(page);
+       clear_page_mlock(page);
+       BUG_ON(page_has_private(page));
+       __delete_from_page_cache(page);
+       spin_unlock_irq(&mapping->tree_lock);
+       mem_cgroup_uncharge_cache_page(page);
+
+       if (mapping->a_ops->freepage)
+               mapping->a_ops->freepage(page);
+
        page_cache_release(page);       /* pagecache ref */
        return 1;
 failed:
-       write_unlock_irq(&mapping->tree_lock);
+       spin_unlock_irq(&mapping->tree_lock);
        return 0;
 }
 
@@ -370,50 +424,48 @@ 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)
 {
        struct pagevec pvec;
-       pgoff_t next;
+       pgoff_t index;
        int i;
        int ret = 0;
+       int ret2 = 0;
        int did_range_unmap = 0;
-       int wrapped = 0;
 
+       cleancache_flush_inode(mapping);
        pagevec_init(&pvec, 0);
-       next = start;
-       while (next <= end && !wrapped &&
-               pagevec_lookup(&pvec, mapping, next,
-                       min(end - next, (pgoff_t)PAGEVEC_SIZE - 1) + 1)) {
+       index = start;
+       while (index <= end && pagevec_lookup(&pvec, mapping, index,
+                       min(end - index, (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;
+
+                       /* We rely upon deletion not changing page->index */
+                       index = page->index;
+                       if (index > end)
+                               break;
 
                        lock_page(page);
+                       WARN_ON(page->index != index);
                        if (page->mapping != mapping) {
                                unlock_page(page);
                                continue;
                        }
-                       page_index = page->index;
-                       next = page_index + 1;
-                       if (next == 0)
-                               wrapped = 1;
-                       if (page_index > end) {
-                               unlock_page(page);
-                               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.
                                         */
                                        unmap_mapping_range(mapping,
-                                          (loff_t)page_index<<PAGE_CACHE_SHIFT,
-                                          (loff_t)(end - page_index + 1)
-                                                       << PAGE_CACHE_SHIFT,
+                                          (loff_t)index << PAGE_CACHE_SHIFT,
+                                          (loff_t)(1 + end - index)
+                                                        << PAGE_CACHE_SHIFT,
                                            0);
                                        did_range_unmap = 1;
                                } else {
@@ -421,18 +473,26 @@ int invalidate_inode_pages2_range(struct address_space *mapping,
                                         * Just zap this page
                                         */
                                        unmap_mapping_range(mapping,
-                                         (loff_t)page_index<<PAGE_CACHE_SHIFT,
-                                         PAGE_CACHE_SIZE, 0);
+                                          (loff_t)index << PAGE_CACHE_SHIFT,
+                                          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);
+               mem_cgroup_uncharge_end();
                cond_resched();
+               index++;
        }
+       cleancache_flush_inode(mapping);
        return ret;
 }
 EXPORT_SYMBOL_GPL(invalidate_inode_pages2_range);
@@ -444,10 +504,117 @@ 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
+ * @oldsize: old file size
+ * @newsize: new file size
+ *
+ * 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 oldsize, loff_t newsize)
+{
+       struct address_space *mapping = inode->i_mapping;
+       loff_t holebegin = round_up(newsize, PAGE_SIZE);
+
+       /*
+        * 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, holebegin, 0, 1);
+       truncate_inode_pages(mapping, newsize);
+       unmap_mapping_range(mapping, holebegin, 0, 1);
+}
+EXPORT_SYMBOL(truncate_pagecache);
+
+/**
+ * truncate_setsize - update inode and pagecache for a new file size
+ * @inode: inode
+ * @newsize: new file size
+ *
+ * truncate_setsize updates i_size and performs pagecache truncation (if
+ * necessary) to @newsize. It will be typically be called from the filesystem's
+ * setattr function when ATTR_SIZE is passed in.
+ *
+ * Must be called with inode_mutex held and before all filesystem specific
+ * block truncation has been performed.
+ */
+void truncate_setsize(struct inode *inode, loff_t newsize)
+{
+       loff_t oldsize;
+
+       oldsize = inode->i_size;
+       i_size_write(inode, newsize);
+
+       truncate_pagecache(inode, oldsize, newsize);
+}
+EXPORT_SYMBOL(truncate_setsize);
+
+/**
+ * vmtruncate - unmap mappings "freed" by truncate() syscall
+ * @inode: inode of the file used
+ * @newsize: file offset to start truncating
+ *
+ * This function is deprecated and truncate_setsize or truncate_pagecache
+ * should be used instead, together with filesystem specific block truncation.
+ */
+int vmtruncate(struct inode *inode, loff_t newsize)
+{
+       int error;
+
+       error = inode_newsize_ok(inode, newsize);
+       if (error)
+               return error;
+
+       truncate_setsize(inode, newsize);
+       if (inode->i_op->truncate)
+               inode->i_op->truncate(inode);
+       return 0;
+}
+EXPORT_SYMBOL(vmtruncate);
+
+int vmtruncate_range(struct inode *inode, loff_t lstart, loff_t lend)
+{
+       struct address_space *mapping = inode->i_mapping;
+       loff_t holebegin = round_up(lstart, PAGE_SIZE);
+       loff_t holelen = 1 + lend - holebegin;
+
+       /*
+        * If the underlying filesystem is not going to provide
+        * a way to truncate a range of blocks (punch a hole) -
+        * we should return failure right now.
+        */
+       if (!inode->i_op->truncate_range)
+               return -ENOSYS;
+
+       mutex_lock(&inode->i_mutex);
+       inode_dio_wait(inode);
+       unmap_mapping_range(mapping, holebegin, holelen, 1);
+       inode->i_op->truncate_range(inode, lstart, lend);
+       /* unmap again to remove racily COWed private pages */
+       unmap_mapping_range(mapping, holebegin, holelen, 1);
+       mutex_unlock(&inode->i_mutex);
+
+       return 0;
+}