[PATCH 6/9] uprobes: flush cache after xol write
Ananth N Mavinakayanahalli
ananth at in.ibm.com
Fri Oct 26 01:52:39 EDT 2012
On Thu, Oct 25, 2012 at 04:58:39PM +0200, Oleg Nesterov wrote:
> On 10/16, Rabin Vincent wrote:
> > 2012/10/15 Oleg Nesterov <oleg at redhat.com>:
> > > On 10/14, Rabin Vincent wrote:
> > >> Flush the cache so that the instructions written to the XOL area are
> > >> visible.
> > >>
> > >> Signed-off-by: Rabin Vincent <rabin at rab.in>
> > >> ---
> > >> kernel/events/uprobes.c | 1 +
> > >> 1 file changed, 1 insertion(+)
> > >>
> > >> diff --git a/kernel/events/uprobes.c b/kernel/events/uprobes.c
> > >> index ca000a9..8c52f93 100644
> > >> --- a/kernel/events/uprobes.c
> > >> +++ b/kernel/events/uprobes.c
> > >> @@ -1246,6 +1246,7 @@ static unsigned long xol_get_insn_slot(struct uprobe *uprobe, unsigned long slot
> > >> offset = current->utask->xol_vaddr & ~PAGE_MASK;
> > >> vaddr = kmap_atomic(area->page);
> > >> arch_uprobe_xol_copy(&uprobe->arch, vaddr + offset);
> > >> + flush_dcache_page(area->page);
> > >> kunmap_atomic(vaddr);
> > >
> > > I agree... but why under kmap_atomic?
> > No real reason; I'll move it to after the unmap.
> OK. I assume you will send v2.
> But this patch looks like a bugfix, flush_dcache_page() is not a nop
> on powerpc. So perhaps we should apply this fix right now?
Starting Power5, all Power processers have coherent caches.
> OTOH, I do not understand this stuff, everything is nop on x86. And
> when I look into Documentation/cachetlb.txt I am starting to think
> that may be this needs flush_icache_user_range instead?
> Rabin, Ananth could you clarify this?
Yes. We need flush_icache_user_range(). Though for x86 its always been a
nop, one never knows if there is some Power4 or older machine out there
that is still being used. We are fine for Power5 and later.
More information about the linux-arm-kernel