ext4: pre-zero allocated blocks for DAX IO

Currently ext4 treats DAX IO the same way as direct IO. I.e., it
allocates unwritten extents before IO is done and converts unwritten
extents afterwards. However this way DAX IO can race with page fault to
the same area:

ext4_ext_direct_IO() dax_fault()
dax_io()
get_block() - allocates unwritten extent
copy_from_iter_pmem()
get_block() - converts
unwritten block to
written and zeroes it
out
ext4_convert_unwritten_extents()

So data written with DAX IO gets lost. Similarly dax_new_buf() called
from dax_io() can overwrite data that has been already written to the
block via mmap.

Fix the problem by using pre-zeroed blocks for DAX IO the same way as we
use them for DAX mmap. The downside of this solution is that every
allocating write writes each block twice (once zeros, once data). Fixing
the race with locking is possible as well however we would need to
lock-out faults for the whole range written to by DAX IO. And that is
not easy to do without locking-out faults for the whole file which seems
too aggressive.

Signed-off-by: Jan Kara <[e-mail]>
Signed-off-by: Theodore Ts'o <[e-mail]>

fs/ext4/ext4.h

···
 struct buffer_head *ext4_bread(handle_t *, struct inode *, ext4_lblk_t, int);
 int ext4_get_block_unwritten(struct inode *inode, sector_t iblock,
 			     struct buffer_head *bh_result, int create);
-int ext4_dax_mmap_get_block(struct inode *inode, sector_t iblock,
-			    struct buffer_head *bh_result, int create);
+int ext4_dax_get_block(struct inode *inode, sector_t iblock,
+		       struct buffer_head *bh_result, int create);
 int ext4_get_block(struct inode *inode, sector_t iblock,
 		   struct buffer_head *bh_result, int create);
 int ext4_dio_get_block(struct inode *inode, sector_t iblock,
···
 		if (atomic_dec_and_test(&EXT4_I(inode)->i_unwritten))
 			wake_up_all(ext4_ioend_wq(inode));
 	}
+}
+
+static inline bool ext4_aligned_io(struct inode *inode, loff_t off, loff_t len)
+{
+	int blksize = 1 << inode->i_blkbits;
+
+	return IS_ALIGNED(off, blksize) && IS_ALIGNED(len, blksize);
 }
 
 #endif	/* __KERNEL__ */

fs/ext4/file.c

···
 	if (IS_ERR(handle))
 		result = VM_FAULT_SIGBUS;
 	else
-		result = __dax_fault(vma, vmf, ext4_dax_mmap_get_block, NULL);
+		result = __dax_fault(vma, vmf, ext4_dax_get_block, NULL);
 
 	if (write) {
 		if (!IS_ERR(handle))
···
 		result = VM_FAULT_SIGBUS;
 	else
 		result = __dax_pmd_fault(vma, addr, pmd, flags,
-				ext4_dax_mmap_get_block, NULL);
+					 ext4_dax_get_block, NULL);
 
 	if (write) {
 		if (!IS_ERR(handle))

fs/ext4/inode.c

···
 }
 
 #ifdef CONFIG_FS_DAX
-int ext4_dax_mmap_get_block(struct inode *inode, sector_t iblock,
-			    struct buffer_head *bh_result, int create)
+/*
+ * Get block function for DAX IO and mmap faults. It takes care of converting
+ * unwritten extents to written ones and initializes new / converted blocks
+ * to zeros.
+ */
+int ext4_dax_get_block(struct inode *inode, sector_t iblock,
+		       struct buffer_head *bh_result, int create)
 {
 	int ret;
 
-	ext4_debug("ext4_dax_mmap_get_block: inode %lu, create flag %d\n",
-		   inode->i_ino, create);
+	ext4_debug("inode %lu, create flag %d\n", inode->i_ino, create);
 	if (!create)
 		return _ext4_get_block(inode, iblock, bh_result, 0);
 
···
 
 	if (buffer_unwritten(bh_result)) {
 		/*
-		 * We are protected by i_mmap_sem so we know block cannot go
-		 * away from under us even though we dropped i_data_sem.
-		 * Convert extent to written and write zeros there.
+		 * We are protected by i_mmap_sem or i_mutex so we know block
+		 * cannot go away from under us even though we dropped
+		 * i_data_sem. Convert extent to written and write zeros there.
 		 */
 		ret = ext4_get_block_trans(inode, iblock, bh_result,
 					   EXT4_GET_BLOCKS_CONVERT |
···
 	 * doesn't attempt to zero blocks again in a racy way.
 	 */
 	clear_buffer_new(bh_result);
+	return 0;
+}
+#else
+/* Just define empty function, it will never get called. */
+int ext4_dax_get_block(struct inode *inode, sector_t iblock,
+		       struct buffer_head *bh_result, int create)
+{
+	BUG();
 	return 0;
 }
 #endif
···
 	iocb->private = NULL;
 	if (overwrite)
 		get_block_func = ext4_dio_get_block_overwrite;
-	else if (!ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS) ||
-		 round_down(offset, 1 << inode->i_blkbits) >= inode->i_size) {
+	else if (IS_DAX(inode)) {
+		/*
+		 * We can avoid zeroing for aligned DAX writes beyond EOF. Other
+		 * writes need zeroing either because they can race with page
+		 * faults or because they use partial blocks.
+		 */
+		if (round_down(offset, 1<<inode->i_blkbits) >= inode->i_size &&
+		    ext4_aligned_io(inode, offset, count))
+			get_block_func = ext4_dio_get_block;
+		else
+			get_block_func = ext4_dax_get_block;
+		dio_flags = DIO_LOCKING;
+	} else if (!ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS) ||
+		   round_down(offset, 1 << inode->i_blkbits) >= inode->i_size) {
 		get_block_func = ext4_dio_get_block;
 		dio_flags = DIO_LOCKING | DIO_SKIP_HOLES;
 	} else if (is_sync_kiocb(iocb)) {
···
 	BUG_ON(ext4_encrypted_inode(inode) && S_ISREG(inode->i_mode));
 #endif
 	if (IS_DAX(inode)) {
-		dio_flags &= ~DIO_SKIP_HOLES;
 		ret = dax_do_io(iocb, inode, iter, offset, get_block_func,
 				ext4_end_io_dio, dio_flags);
 	} else