Btrfs: Use async helpers to deal with pages that have been improperly dirtied

Higher layers sometimes call set_page_dirty without asking the filesystem
to help.  This causes many problems for the data=ordered and cow code.
This commit detects pages that haven't been properly setup for IO and
kicks off an async helper to deal with them.

Signed-off-by: Chris Mason <chris.mason@oracle.com>

fs/btrfs/ctree.h

@@ -546,6 +546,12 @@ struct btrfs_fs_info {
 	struct btrfs_workers endio_workers;
 	struct btrfs_workers endio_write_workers;
 	struct btrfs_workers submit_workers;
+	/*
+	 * fixup workers take dirty pages that didn't properly go through
+	 * the cow mechanism and make them safe to write.  It happens
+	 * for the sys_munmap function call path
+	 */
+	struct btrfs_workers fixup_workers;
 	struct task_struct *transaction_kthread;
 	struct task_struct *cleaner_kthread;
 	int thread_pool_size;

fs/btrfs/disk-io.c

@@ -1329,11 +1329,13 @@ struct btrfs_root *open_ctree(struct super_block *sb,
 	 */
 	btrfs_init_workers(&fs_info->workers, fs_info->thread_pool_size);
 	btrfs_init_workers(&fs_info->submit_workers, fs_info->thread_pool_size);
+	btrfs_init_workers(&fs_info->fixup_workers, 1);
 	btrfs_init_workers(&fs_info->endio_workers, fs_info->thread_pool_size);
 	btrfs_init_workers(&fs_info->endio_write_workers,
 			   fs_info->thread_pool_size);
 	btrfs_start_workers(&fs_info->workers, 1);
 	btrfs_start_workers(&fs_info->submit_workers, 1);
+	btrfs_start_workers(&fs_info->fixup_workers, 1);
 	btrfs_start_workers(&fs_info->endio_workers, fs_info->thread_pool_size);
 	btrfs_start_workers(&fs_info->endio_write_workers,
 			    fs_info->thread_pool_size);
@@ -1454,6 +1456,7 @@ struct btrfs_root *open_ctree(struct super_block *sb,
 fail_sys_array:
 fail_sb_buffer:
 	extent_io_tree_empty_lru(&BTRFS_I(fs_info->btree_inode)->io_tree);
+	btrfs_stop_workers(&fs_info->fixup_workers);
 	btrfs_stop_workers(&fs_info->workers);
 	btrfs_stop_workers(&fs_info->endio_workers);
 	btrfs_stop_workers(&fs_info->endio_write_workers);
@@ -1710,6 +1713,7 @@ int close_ctree(struct btrfs_root *root)
 
 	truncate_inode_pages(fs_info->btree_inode->i_mapping, 0);
 
+	btrfs_stop_workers(&fs_info->fixup_workers);
 	btrfs_stop_workers(&fs_info->workers);
 	btrfs_stop_workers(&fs_info->endio_workers);
 	btrfs_stop_workers(&fs_info->endio_write_workers);

fs/btrfs/extent_io.c

@@ -2050,6 +2050,16 @@ static int __extent_writepage(struct page *page, struct writeback_control *wbc,
 	lock_extent(tree, start, page_end, GFP_NOFS);
 	unlock_start = start;
 
+	if (tree->ops && tree->ops->writepage_start_hook) {
+		ret = tree->ops->writepage_start_hook(page, start, page_end);
+		if (ret == -EAGAIN) {
+			unlock_extent(tree, start, page_end, GFP_NOFS);
+			redirty_page_for_writepage(wbc, page);
+			unlock_page(page);
+			return 0;
+		}
+	}
+
 	end = page_end;
 	if (test_range_bit(tree, start, page_end, EXTENT_DELALLOC, 0)) {
 		printk("found delalloc bits after lock_extent\n");

fs/btrfs/extent_io.h

@@ -30,6 +30,7 @@ typedef	int (extent_submit_bio_hook_t)(struct inode *inode, int rw,
 				       struct bio *bio, int mirror_num);
 struct extent_io_ops {
 	int (*fill_delalloc)(struct inode *inode, u64 start, u64 end);
+	int (*writepage_start_hook)(struct page *page, u64 start, u64 end);
 	int (*writepage_io_hook)(struct page *page, u64 start, u64 end);
 	extent_submit_bio_hook_t *submit_bio_hook;
 	int (*merge_bio_hook)(struct page *page, unsigned long offset,

fs/btrfs/file.c

@@ -313,6 +313,7 @@ static int noinline dirty_and_release_pages(struct btrfs_trans_handle *trans,
 		for (i = 0; i < num_pages; i++) {
 			struct page *p = pages[i];
 			SetPageUptodate(p);
+			ClearPageChecked(p);
 			set_page_dirty(p);
 		}
 	} else {

fs/btrfs/inode.c

@@ -403,6 +403,87 @@ static int add_pending_csums(struct btrfs_trans_handle *trans,
 	return 0;
 }
 
+struct btrfs_writepage_fixup {
+	struct page *page;
+	struct btrfs_work work;
+};
+
+/* see btrfs_writepage_start_hook for details on why this is required */
+void btrfs_writepage_fixup_worker(struct btrfs_work *work)
+{
+	struct btrfs_writepage_fixup *fixup;
+	struct btrfs_ordered_extent *ordered;
+	struct page *page;
+	struct inode *inode;
+	u64 page_start;
+	u64 page_end;
+
+	fixup = container_of(work, struct btrfs_writepage_fixup, work);
+	page = fixup->page;
+
+	lock_page(page);
+	if (!page->mapping || !PageDirty(page) || !PageChecked(page)) {
+		ClearPageChecked(page);
+		goto out_page;
+	}
+
+	inode = page->mapping->host;
+	page_start = page_offset(page);
+	page_end = page_offset(page) + PAGE_CACHE_SIZE - 1;
+
+	lock_extent(&BTRFS_I(inode)->io_tree, page_start, page_end, GFP_NOFS);
+	ordered = btrfs_lookup_ordered_extent(inode, page_start);
+	if (ordered)
+		goto out;
+
+	set_extent_delalloc(&BTRFS_I(inode)->io_tree, page_start, page_end,
+			    GFP_NOFS);
+	ClearPageChecked(page);
+out:
+	unlock_extent(&BTRFS_I(inode)->io_tree, page_start, page_end, GFP_NOFS);
+out_page:
+	unlock_page(page);
+	page_cache_release(page);
+}
+
+/*
+ * There are a few paths in the higher layers of the kernel that directly
+ * set the page dirty bit without asking the filesystem if it is a
+ * good idea.  This causes problems because we want to make sure COW
+ * properly happens and the data=ordered rules are followed.
+ *
+ * In our case any range that doesn't have the EXTENT_ORDERED bit set
+ * hasn't been properly setup for IO.  We kick off an async process
+ * to fix it up.  The async helper will wait for ordered extents, set
+ * the delalloc bit and make it safe to write the page.
+ */
+int btrfs_writepage_start_hook(struct page *page, u64 start, u64 end)
+{
+	struct inode *inode = page->mapping->host;
+	struct btrfs_writepage_fixup *fixup;
+	struct btrfs_root *root = BTRFS_I(inode)->root;
+	int ret;
+
+	ret = test_range_bit(&BTRFS_I(inode)->io_tree, start, end,
+			     EXTENT_ORDERED, 0);
+	if (ret)
+		return 0;
+
+	if (PageChecked(page))
+		return -EAGAIN;
+
+	fixup = kzalloc(sizeof(*fixup), GFP_NOFS);
+	if (!fixup)
+		return -EAGAIN;
+printk("queueing worker to fixup page %lu %Lu\n", inode->i_ino, page_offset(page));
+	SetPageChecked(page);
+	page_cache_get(page);
+	fixup->work.func = btrfs_writepage_fixup_worker;
+	fixup->page = page;
+	btrfs_queue_worker(&root->fs_info->fixup_workers, &fixup->work);
+	return -EAGAIN;
+}
+
 int btrfs_writepage_end_io_hook(struct page *page, u64 start, u64 end,
 				struct extent_state *state, int uptodate)
 {
@@ -1263,6 +1344,7 @@ static int btrfs_truncate_page(struct address_space *mapping, loff_t from)
 		flush_dcache_page(page);
 		kunmap(page);
 	}
+	ClearPageChecked(page);
 	set_page_dirty(page);
 	unlock_extent(io_tree, page_start, page_end, GFP_NOFS);
 
@@ -2658,6 +2740,7 @@ int btrfs_page_mkwrite(struct vm_area_struct *vma, struct page *page)
 		flush_dcache_page(page);
 		kunmap(page);
 	}
+	ClearPageChecked(page);
 	set_page_dirty(page);
 	unlock_extent(io_tree, page_start, page_end, GFP_NOFS);
 
@@ -3039,15 +3122,6 @@ static int btrfs_symlink(struct inode *dir, struct dentry *dentry,
 
 static int btrfs_set_page_dirty(struct page *page)
 {
-	struct inode *inode = page->mapping->host;
-	u64 page_start = page_offset(page);
-	u64 page_end = page_start + PAGE_CACHE_SIZE - 1;
-
-	if (!test_range_bit(&BTRFS_I(inode)->io_tree, page_start, page_end,
-			    EXTENT_DELALLOC, 0)) {
-printk("inode %lu page %Lu not delalloc\n", inode->i_ino, page_offset(page));
-WARN_ON(1);
-	}
 	return __set_page_dirty_nobuffers(page);
 }
 
@@ -3098,6 +3172,7 @@ static struct extent_io_ops btrfs_extent_io_ops = {
 	.readpage_io_hook = btrfs_readpage_io_hook,
 	.readpage_end_io_hook = btrfs_readpage_end_io_hook,
 	.writepage_end_io_hook = btrfs_writepage_end_io_hook,
+	.writepage_start_hook = btrfs_writepage_start_hook,
 	.readpage_io_failed_hook = btrfs_io_failed_hook,
 	.set_bit_hook = btrfs_set_bit_hook,
 	.clear_bit_hook = btrfs_clear_bit_hook,