Re: [RFC/PATCH 2/4] fs/mm: execute in place (3rd version)

[Christoph Hellwig <hch@infradead.org>]

On Tue, May 24, 2005 at 07:01:21PM +0200, Carsten Otte wrote:
> When looking at patch v2, the read split is done in do_generic_mapping_read
> vs do_xip_mapping read. In the write path, the split is at
> generic_file_xip_write,
> generic_file_buffered_write and generic_file_direct_write.
> How about abstracting on that interface? Like make those become address
> space operations.

No, these aren't related to the address_space at all.  Please keep the
abstractions clean or we'll end up with a total mess.

> This way, the filesystems could select the corresponding
> function. No need to distinguish between xip, direct_IO, and classic
> readpage/writepage in the generic code anymore.
> Would this go in the direction you're thinking Suparna? Is it worth a
> try to see
> how it comes out? Opinions anyone?

I don't think this split makes a whole lot of sense either.  The normal
codepath needs to go through a lot of hops to make sure the odd behaviour
of falling back from direct to buffered I/O for some types of requests
works.  That needs intimate knowledge of how the direct and buffered I/O
path works.  Your XIP codepath has the luxury of not having to care for
all this, so you can keep it really simple.