Re: [RFC] ARC: mm: Restrict definition of pfn_valid() macro for CONFIG_FLATMEM

From: Vineet Gupta
Date: Wed Nov 30 2016 - 11:55:22 EST

On 11/30/2016 06:21 AM, Yuriy Kolerov wrote:
>> On Tue 29-11-16 18:29:06, Yuriy Kolerov wrote:
>>> > > Despite the fact that subtraction of unsigned integers is a defined
>>> > > behaviour however such operations can lead to unexpected results. Thus
>>> > > it is better to check both left and right boundaries to avoid
>>> > > potential bugs as it done in the generic page.h.
>> >
>> > Why and which code would use an out of range pfn? Why other arches do
>> > not need to care?
> Actually some arches do care about checking of both left and right boundaries (e.g. avr32, sparc, etc). The problem is that a value of pfn may be calculated incorrectly in some places of the kernel. E.g. not long ago I sent a patch which fixes truncation of the most significant byte in pfn/pte in some cases (in the kernel with PAE40, however it is not a FLATMEM case). So such situations can happens in the most unexpected places.

So the point is - is this a preventive fix (desired thing) or it being there would
have helped find the PAE40 bug earlier / easier. Woudl it have prevented the
kernel crash. If so then this is a nobrainer fix.

BTW did you try to gauge the code gen impact - this function gets pulled all over
the place in mm code. So build kernel with and w/o change and do a