@@ -6008,9 +6008,10 @@ vm_fault_t ext4_page_mkwrite(struct vm_fault *vmf)
len = PAGE_SIZE;
/*
* Return if we have all the buffers mapped. This avoids the need to do
- * journal_start/journal_stop which can block and take a long time
+ * journal_start/journal_stop which can block and take a long time. But
+ * not on data journalling, as we have to add the inode to the txn list.
*/
- if (page_has_buffers(page)) {
+ if (page_has_buffers(page) && !ext4_should_journal_data(inode)) {
if (!ext4_walk_page_buffers(NULL, page_buffers(page),
0, len, NULL,
ext4_bh_unmapped)) {
@@ -6043,6 +6044,12 @@ vm_fault_t ext4_page_mkwrite(struct vm_fault *vmf)
goto out;
}
ext4_set_inode_state(inode, EXT4_STATE_JDATA);
+ if (ext4_jbd2_inode_add_write(handle, inode, 0, PAGE_SIZE)) {
+ unlock_page(page);
+ ret = VM_FAULT_SIGBUS;
+ ext4_journal_stop(handle);
+ goto out;
+ }
}
ext4_journal_stop(handle);
if (err == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
Since we only add the inode to the transaction's inode list in __ext4_journalled_writepage(), we depend on msync() or writeback work (which call it) for the write-protect mechanism to work. This test snippet shows that, as pwrite() gets the inode into a transaction (!= than into transaction's inode list), and addr[] write access gets the page writeably mapped. fd = open("file"); addr = mmap(fd); pwrite(fd, "a", 1, 0); // journals inode via ext4_write_begin() addr[0] = 'a'; // page is writeably mapped to user space. // periodic journal commit / jbd2 thread runs now. // __ext4_journalled_writepage() was not called yet. Now it's possible for a subsequent addr[] write access to race with the commit function, and possibly hit the window to cause invalid checksums. --- fs/ext4/inode.c | 11 +++++++++-- 1 file changed, 9 insertions(+), 2 deletions(-)