Re: [patch] cache flush bug in mm/filemap.c (all kernels >= 2.5.30(at least))

[Andrew Morton <akpm@digeo.com>]

"David S. Miller" <davem@redhat.com> wrote:
>
> On Thu, 2003-05-22 at 07:11, Russell King wrote:
> > We seem to have flush_icache_page() in install_page() - I wonder whether
> > we should also have flush_dcache_page() in there as well.
> ...
> > Maybe someone more knowledgeable of the VM layer can comment.
> 
> I am not sure of the exact environment install_page() is meant
> to run in, does it always know that no mapping exists at that
> address?

No, it does not - there could have been a different page mapped at that
virtual address.

> If not, something (either there or higher up) needs to be doing
> a flush_cache_page(...) at a minimum.

install_page() did a flush_cache_page() in zap_pte(), if it found there was
already a page mapped by the pte.

That flush_cache_page() was added 28 March 2003.  2.5.30 does not have it.

What's that flush_icache_page() doing in there?  The fine document makes
one suspect it is wrong.


install_page() is prefaulting pages into pagetables, so perhaps it should
have an update_mmu_cache()?

diff -puN mm/fremap.c~install_page-flushing mm/fremap.c
--- 25/mm/fremap.c~install_page-flushing	2003-05-23 02:01:52.000000000 -0700
+++ 25-akpm/mm/fremap.c	2003-05-23 02:09:03.000000000 -0700
@@ -84,7 +84,7 @@ int install_page(struct mm_struct *mm, s
 	pte_unmap(pte);
 	if (flush)
 		flush_tlb_page(vma, addr);
-
+	update_mmu_cache(vma, addr, *pte);
 	spin_unlock(&mm->page_table_lock);
 	pte_chain_free(pte_chain);
 	return 0;

_