Re: [RFC PATCH 2/2] erofs: directly traverse pages in z_erofs_readahead()

[Gao Xiang <xiang@kernel.org>]

Hi Matthew,

On Mon, Jul 05, 2021 at 09:15:26PM +0100, Matthew Wilcox wrote:
> On Tue, Jul 06, 2021 at 02:32:53AM +0800, Gao Xiang wrote:
> > In that way, pages can be accessed directly with xarray.
> 
> I didn't mean "open code readahead_page()".  I meant "Wouldn't it be
> great if z_erofs_do_read_page() used readahead_expand() in order to
> allocate the extra pages in the extents that cover the start and end of
> the requested chunk".  That way the pages would already be in the page
> cache for subsequent reads.

Yes, I understand that idea. readahead_expand() can be used to
cache the whole requested decompressed chunk. Currently, EROFS tends
to cache compressed data for incomplete requested chunks instead in the
managed inode if cache mode is enabled since LZ4 compressed data is
more effective for caching to save I/O than uncompressed data
(considering LZ4 decompression speed, which is somewhat like zcache.)

For now, EROFS LZMA support is also on the way (but recently I
have to work on the DAX support in advance.) LZMA is somewhat slow
algorithm but with much better compression ratio. I think it should
decompress all requested chunks to save the overall decompression
time. So readahead_expand() is quite useful with EROFS LZMA cases,
I will investigate it for LZMA-specific cases...

BTW, from the patch [2/2] itself, it there some chance to extend
readahead_page() or some new API like this? So the original dance
can be avoided.

Some comments for the iteration behavior itself is always useful,
anyway...

Thanks,
Gao Xiang

>