Re: Properly implement flush_dcache_page in 2.4? (Or is it possible?)

[Jun Sun]

On Sat, May 31, 2003 at 12:14:58AM +0100, Russell King wrote:
> > In addition, I am not sure if the vma struct will show up in the
> > "shared" list _if_ the page is only mapped in one user process and
> > in kernel (for example, those pages you obtain through get_user_pages()
> > call).
> 
> If a mapping is using MAP_SHARED, my understanding is that the pages should
> appear on the i_mmap_shared list.
>

That is my understanding too.
 
> I don't see a reason to worry about privately mapped pages on the i_mmap
> list since they are private, and therefore shouldn't be updated with
> modifications to other mappings, 

Actually there is, at least in 2.4.  Whenever kernel calls get_user_pages()
it maps a user page into kernel virtual address space.  If kernel modifies
that page, flush_dcache_page() needs to make sure any stale cache data
at user virtual address is flush in order user to see kernel changes.

I have a test case at 

	http://linux.junsun.net/test-programs

(Note, sometimes even if you pass the test, you _may_ still have a wrong
flush_dcache_page() implementation, because stale cache could be flushed 
due to other execution sequences)

I took a brief look of 2.5 code.  It seems this problem should still
exist (of course, assuming the CPU has cache aliasing problem and the
flush_dcache_page() is not properly implemented).  

Actually in 2.5 you may fail the test even if you have a properly implemented
flush_dcache_page().  It appears it lacks another flush_dcache_page()
after the direct_IO is done.  I don't have a working 2.5 on MIPS.  Can't 
verify that. 

Jun