Re: [PATCH] [ARM] force dcache flush if dcache_dirty bit set

[Catalin Marinas <catalin.marinas@arm.com>]

On Wed, 2009-10-14 at 18:10 +0100, Catalin Marinas wrote:
> On Tue, 2009-10-13 at 11:07 +0900, Paul Mundt wrote:
> > On Mon, Oct 12, 2009 at 06:03:12PM +0100, Russell King wrote:
> > > I don't see that there should have been any bearing on whether a page
> > > has a mapping or not when we get to update_mmu_cache.  The issue here
> > > is that > if PG_arch_1 is set on a page, then we didn't flush it at
> > > the time when we believed it was appropriate to do so. <
> > > 
> > > Tell me I'm wrong (having only just sent it to Linus...)
> > 
> > Having looked at the ARM fix, in the !mapping case do you not need the
> > I-cache flush on vma->vm_flags & VM_EXEC? Or is the presumption that
> > flush_icache_page()-type action doesn't need to be undertaken by
> > flush_dcache_page()/update_mmu_cache() when there is no page_mapping()?
> 
> If I understand Nitin's scenario correctly, I think it should also
> invalidate the I-cache.
> 
> For executable anonymous pages containing, it's the user app writing the
> code (JIT etc.) and it calls an ARM-specific syscall for I and D cache
> maintenance. If such page is read back from swap, following Nitin's
> scenario, the I-cache would need to be invalidated as well otherwise it
> can have stale entries.

We can have a scenario where I-cache invalidation would not help on ARM.
Some apps may temporarily change protection from RX to RW to write some
data (not instructions) to a page containing code. If Nitin's scenario
happens when the page is RW, the VM_EXEC wouldn't be set, hence no
I-cache invalidation.

The app would later do mprotect(RX) but on some ARM processors
flush_cache_range() is a no-op and therefore no cache flushing. What
other architectures with Harvard caches do for mprotect(RX)? Is this
assumed to invalidate the I-cache?

(we have another I-D cache coherency problem on ARM with COW text pages
following a RX -> RW -> write data -> RX scenario and there are a few
solutions for this, though none of them optimal with the current cache
flushing API, especially with the read-implies-exec ELF personality)

-- 
Catalin