On Thu, Feb 27, 2014 at 01:28:22PM -0800, Linus Torvalds wrote:

On Thu, Feb 27, 2014 at 11:53 AM, Kirill A. Shutemov
[off-list ref] wrote:

quoted

Here's new version of faultaround patchset. It took a while to tune it and
collect performance data.

Andrew, mind taking this into -mm with my acks? It's based on top of
Kirill's cleanup patches that I think are also in your tree.

Kirill - no complaints from me. I do have two minor issues that you
might satisfy, but I think the patch is fine as-is.

The issues/questions are:

 (a) could you test this on a couple of different architectures? Even
if you just have access to intel machines, testing it across a couple
of generations of microarchitectures would be good. The reason I say
that is that from my profiles, it *looks* like the page fault costs
are relatively higher on Ivybridge/Haswell than on some earlier
uarchs.

These numbers were from Ivy Bridge.
I'll bring some numbers for Westmere and Haswell.

 (b) I suspect we should try to strongly discourage filesystems from
actually using map_pages unless they use the standard
filemap_map_pages function as-is. Even with the fairly clean
interface, and forcing people to use "do_set_pte()", I think the docs
might want to try to more explicitly discourage people from using this
to do their own hacks..

We would need ->map_pages() at least for shmem/tmpfs. It should be
benefitial there.

Also Matthew noticed that some drivers do ugly hacks like fault in whole
VMA on first page fault. IIUC, it's for performance reasons. See
psbfb_vm_fault() or ttm_bo_vm_fault().

I thought it could be reasonable to have ->map_pages() there and do VMA
population get_user_pages() on mmap() instead.

What do you think?

-- 
 Kirill A. Shutemov

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>