Non-linear Remap Vs. VM Cleanup: Performance Hit

Non-linear Remap Vs. VM Cleanup: Performance Hit

[D. Cooper Stevenson]

Hello All,

I read the Linus's release for RC1. One of his favorite features in the 
release are, "actually some vm cleanups, where this release is getting 
rid of the largely unused non-linear remapping code (replaced with just 
emulating it with lots of smaller mappings) and unifies the NUMA and 
PROTNONE handling for page tables."

I understand that non-linear page remapping is unused (and, presumably, 
reduces the code base size) but according to the remap_file_pages man 
page (http://man7.org/linux/man-pages/man2/remap_file_pages.2.html) 
non-linear system will be, "eventually be replaced by a slower in-kernel 
emulation."

Was this a change of necessity? What gives?

Also, would you please Cc: me on your replies as I am not a member of 
the mailing list?


Best Regards,


-Cooper

Re: Non-linear Remap Vs. VM Cleanup: Performance Hit

[Kirill A. Shutemov]

On Tue, Feb 24, 2015 at 02:36:25PM -0800, D. Cooper Stevenson wrote:
> Hello All,
> 
> I read the Linus's release for RC1. One of his favorite features in the
> release are, "actually some vm cleanups, where this release is getting rid
> of the largely unused non-linear remapping code (replaced with just
> emulating it with lots of smaller mappings) and unifies the NUMA and
> PROTNONE handling for page tables."
> 
> I understand that non-linear page remapping is unused (and, presumably,
> reduces the code base size) but according to the remap_file_pages man page
> (http://man7.org/linux/man-pages/man2/remap_file_pages.2.html) non-linear
> system will be, "eventually be replaced by a slower in-kernel emulation."
> 
> Was this a change of necessity? What gives?

As with all cleanups it's not necessary, but it makes life easier.

It removed >1.5k lines of non-trivial code from critical path of the
kernel. It's rarely used and tend to be error-prone. And developers need
to be aware about exists of non-linear mapping while implementing new
features.

For instance, faultaround feature I've implemented was broken for
non-linear mapping for some time and it took time to track it down.

-- 
 Kirill A. Shutemov