On Mon, Jan 04, 2021 at 09:44:15AM +0100, Christoph Hellwig wrote:

On Wed, Dec 30, 2020 at 08:08:15AM +0800, Ming Lei wrote:

quoted

It is observed that __block_write_full_page() always submit bio with size of block size,
which is often 512 bytes.

In case of sequential IO, or >=4k BS random/seq writeback IO, most of times IO
represented by all buffer_head in each page can be done in single bio. It is actually
done in single request IO by block layer's plug merge too.

So check if IO represented by buffer_head can be merged to single page
IO, if yes, just submit single bio instead of submitting one bio for each buffer_head.

There is some very weird formatting in here.  From a very quick look
the changes look sensible, but I wonder if we should spend so much
time optimizing the legacy buffer_head I/O path, rather than switching
callers to saner helpers.

It may take long to convert fs code into iomap, and I understand fs/block_dev.c
can't be converted to iomap until all FS removes buffer_head, correct me
if it is wrong.


Thanks,
Ming