On Thu, May 17, 2018 at 12:56 PM, Matthew Wilcox <wi...@infradead.org> wrote:
> On Thu, May 17, 2018 at 12:32:07PM -0700, Dan Williams wrote:
>> On Thu, May 17, 2018 at 11:37 AM, Matthew Wilcox <wi...@infradead.org> wrote:
>> > I plucked this patch from my XArray work. It seems self-contained enough
>> > that it could go into the DAX tree for merging this cycle.
>> > From 8cb56f4ba36af38814ca7b8ba030a66384e59a21 Mon Sep 17 00:00:00 2001
>> > From: Matthew Wilcox <mawil...@microsoft.com>
>> > Date: Thu, 29 Mar 2018 22:41:18 -0400
>> > Subject: [PATCH] dax: Fix use of zero page
>> > Use my_zero_pfn instead of ZERO_PAGE, and pass the vaddr to it so it
>> > works on MIPS and s390.
>> > Signed-off-by: Matthew Wilcox <mawil...@microsoft.com>
>> I'm being thick and / or lazy, what's the user visible effect of this fix?
> For s390 it appears to be a performance issue:
> Author: Martin Schwidefsky <schwidef...@de.ibm.com>
> Date: Mon Oct 25 16:10:07 2010 +0200
> [S390] zero page cache synonyms
> If the zero page is mapped to virtual user space addresses that differ
> only in bit 2^12 or 2^13 we get L1 cache synonyms which can affect
> performance. Follow the mips model and use multiple zero pages to avoid
> the synonyms.
> MIPS' use of multiple ZERO_PAGEs predates git history. Given the
> history of MIPS' caches behaving in incredibly weird ways, I'd assume
> that getting this wrong results in miniature black holes forming and/or
> the CPU calculating the largest prime number.
Unless I am missing something I think this sounds like 4.18-rc1
material with a cc: stable. Last I heard no one is really using
dccssblk + dax, and MIPS has no way to describe pmem outside of
memmap= which is only a development tool.
Linux-nvdimm mailing list