On Wed, Aug 01, 2018 at 01:56:19PM -0700, Linus Torvalds wrote:

On Wed, Aug 1, 2018 at 1:52 PM Kirill A. Shutemov [off-list ref] wrote:

quoted

Is there a reason why we pass vma to flush_tlb_range?

Yes. It's even in that patch.

The fact is, real MM users *have* a vma, and passing it in to the TLB
flushing is the right thing to do. That allows architectures that care
(mainly powerpc, I think) to notice that "hey, this range only had
execute permissions, so I only need to flush the ITLB".

The people who use tlb_flush_range() any other way are doing an
arch-specific hack.  It's not how tlb_flush_range() was defined, and
it's not how you can use it in general.

Okay, I see.

ARM, unicore32 and xtensa avoid iTLB flush for non-executable VMAs.

quoted

It's not obvious to me what information from VMA can be useful for an
implementation.

See the patch I sent, which had this as part of it:

-                * XXX fix me: flush_tlb_range() should take an mm
pointer instead of a
-                * vma pointer.
+                * flush_tlb_range() takes a vma instead of a mm pointer because
+                * some architectures want the vm_flags for ITLB/DTLB flush.

because I wanted to educate people about why the interface was what it
was, and the "fixme" was bogus shit.

I didn't noticied this. Sorry.

-- 
 Kirill A. Shutemov