On Tue, Jun 27, 2017 at 09:09:48AM -0600, Jens Axboe wrote:

On 06/27/2017 08:42 AM, Christoph Hellwig wrote:

quoted

The API looks ok, but the code could use some cleanups.  What do
you think about the incremental patch below:

It refactors various manipulations, and stores the write hint right
in the iocb as there is a 4 byte hole (this will need some minor
adjustments in the next patches):

How's this? Fixes for compile, and also squeeze an enum rw_hint into
a hole in the inode structure.

I'll refactor around this and squeeze into bio/rq holes as well, then
re-test it.

Looks good, minor nitpick below:

I didn't seem to need the move.  But if you want to move it can
we keep all the includes together at the very top?

+static inline enum rw_hint __inode_write_hint(struct inode *inode)
+{
+	return inode->i_write_hint;
+}
+
+static inline enum rw_hint inode_write_hint(struct inode *inode)
+{
+	enum rw_hint ret = __inode_write_hint(inode);
+	if (ret != WRITE_LIFE_NOT_SET)
+		return ret;
+	return WRITE_LIFE_NONE;
+}
+
+static inline enum rw_hint __file_write_hint(struct file *file)
+{
+	if (file->f_write_hint != WRITE_LIFE_NOT_SET)
+		return file->f_write_hint;
+
+	return __inode_write_hint(file_inode(file));
+}
+
+static inline enum rw_hint file_write_hint(struct file *file)
+{
+	enum rw_hint ret = __file_write_hint(file);
+	if (ret != WRITE_LIFE_NOT_SET)
+		return ret;
+	return WRITE_LIFE_NONE;
+}

I'd say kill all these helpers and just treat both WRITE_LIFE_NONE
and WRITE_LIFE_NOT_SET special all the way down in NVMe.