On 2-Jun-13, at 2:02 PM, James Bottomley wrote:
On Sun, 2013-06-02 at 12:09 -0400, John David Anglin wrote:
As noted by James E.J. Bottomley, we need to purge the kernel TLB
entries used when copying/clearing
to user pages on PA 2.0 systems that require cache coherency.
Otherwise, we generate inequivalent
aliases and incorrect cache operation.
What I meant was we have to remove the tmpalias tlb entry at the end
of
the tmpalias operation. This means in the asm of pacache.S. Like the
patch below (untested because all my boxes are currently in a shipping
container on the docks at Seattle).
The entries are removed in flush_dcache_page_asm and
flush_icache_page_asm.
They are not currently removed in copy_user_page_asm and
clear_user_page_asm.
The latter two routines are not currently used.
CONFIG_PARISC_TMPALIAS needs
to be defined to switch to using the latter two routines.
copy_user_page_asm has
never been enabled in the linux source and I haven't tested in the
highpage code in
over a year.
diff --git a/arch/parisc/kernel/cache.c b/arch/parisc/kernel/cache.c
index 65fb4cb..a16ede352 100644
--- a/arch/parisc/kernel/cache.c
+++ b/arch/parisc/kernel/cache.c
@@ -383,7 +383,7 @@ void clear_user_page(void *vto, unsigned long
vaddr, struct page *page)
{
clear_page_asm(vto);
if (!parisc_requires_coherency())
- flush_kernel_dcache_page_asm(vto);
+ flush_kernel_dcache_page_addr(vto);
}
EXPORT_SYMBOL(clear_user_page);
@@ -397,10 +397,10 @@ void copy_user_page(void *vto, void *vfrom,
unsigned long vaddr,
the kernel mapping. */
preempt_disable();
flush_dcache_page_asm(__pa(vfrom), vaddr);
- preempt_enable();
copy_page_asm(vto, vfrom);
if (!parisc_requires_coherency())
- flush_kernel_dcache_page_asm(vto);
+ flush_kernel_dcache_page_addr(vto);
+ preempt_enable();
}
EXPORT_SYMBOL(copy_user_page);
These two changes should be ineffective. The only time
clear_user_page() and copy_user_page() are used is if the architecture
doesn't supply clear_user_highpage() and copy_user_highpage() along
with
the necessary defines for highmem.h, which we do.
Yes, I added them.
If we go back to the non highpage defines, we've effectively double
purged, because the kunmap_atomic will do the kernel tlb purge for
us as
well.
Actually, no. The checks in clear_user_page and copy_user_page are
!parisc_requires_coherency(). They can't affect rp3440. So, the
original
code was correct. So, the improved stability has to have come from
moving
the preempt lines.
Dave
--
John David Anglin dave.anglin@xxxxxxxx
--
To unsubscribe from this list: send the line "unsubscribe linux-parisc" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
