Re: [PATCH 2.6.21] cramfs: add cramfs Linear XIP

[Carsten Otte <cotte@de.ibm.com>]

Jörn Engel wrote:
> Nitbit: Sooner or later the point/unpoint should get replaced by
> something with page granularity.  Something also needs to keep lists of
> mapped pages and invalidate them whenever the device is written to.
> That could be done in the filesystem or device driver.  I believe the
> device driver would be a better solution.
I think it needs to work like this:
- temporary references (for read/write syscalls and friends) get 
retrieved via get_xip_page and returned again via to-be-invented 
put_page/page_cache_release
- permanent references (for mapping to userland) get retrieved via 
get_xip_page and don't get returned until unmap
- the device driver can access page->count via a helper function 
provided by mm. This way, it can identify which pages are in use.
- In order to get references back, the device driver can call a 
callback provided by the file system. A default implementation will go 
to filemap_xip.c. This callback would use rmap to find all mappings, 
and unmap the page via xip_file_unmap()[mm/filemap_xip.c].

The nice thing about this approach is: we use page->count and rmap, 
both already exist and are perfectly suited for our purpose.
The downside: We need mem_map[] struct page entries behind all memory 
segments. Nowerdays we can easily create those via vmem_map/sparsemem.

Opinions?