RE: [PATCH 3/3] x86: add local_tlb_flush_kernel_range()

[Dan Magenheimer <dan.magenheimer@oracle.com>]

> From: Seth Jennings [mailto:sjenning@linux.vnet.ibm.com]
> Sent: Wednesday, June 27, 2012 12:34 PM
> To: Dan Magenheimer
> Cc: Minchan Kim; Alex Shi; Greg Kroah-Hartman; devel@driverdev.osuosl.org; Konrad Wilk; linux-
> kernel@vger.kernel.org; linux-mm@kvack.org; Andrew Morton; Robert Jennings; Nitin Gupta
> Subject: Re: [PATCH 3/3] x86: add local_tlb_flush_kernel_range()
> 
> On 06/27/2012 10:12 AM, Dan Magenheimer wrote:
> >> From: Minchan Kim [mailto:minchan@kernel.org]
> >> Subject: Re: [PATCH 3/3] x86: add local_tlb_flush_kernel_range()
> >>
> >> On 06/27/2012 03:14 PM, Alex Shi wrote:
> >>>
> >>> On 06/27/2012 01:53 PM, Minchan Kim wrote:
> >>> Different CPU type has different balance point on the invlpg replacing
> >>> flush all. and some CPU never get benefit from invlpg, So, it's better
> >>> to use different value for different CPU, not a fixed
> >>> INVLPG_BREAK_EVEN_PAGES.
> >>
> >> I think it could be another patch as further step and someone who are
> >> very familiar with architecture could do better than.
> >> So I hope it could be merged if it doesn't have real big problem.
> >>
> >> Thanks for the comment, Alex.
> >
> > Just my opinion, but I have to agree with Alex.  Hardcoding
> > behavior that is VERY processor-specific is a bad idea.  TLBs should
> > only be messed with when absolutely necessary, not for the
> > convenience of defending an abstraction that is nice-to-have
> > but, in current OS kernel code, unnecessary.
> 
> I agree that it's not optimal.  The selection based on CPUID
> is part of Alex's patchset, and I'll be glad to use that
> code when it gets integrated.
> 
> But the real discussion is are we going to:
> 1) wait until Alex's patches to be integrated, degrading
> zsmalloc in the meantime or
> 2) put in some simple temporary logic that works well (not
> best) for most cases
> 
> > IIUC, zsmalloc only cares that the breakeven point is greater
> > than two.  An arch-specific choice of (A) two page flushes
> > vs (B) one all-TLB flush should be all that is necessary right
> > now.  (And, per separate discussion, even this isn't really
> > necessary either.)
> >
> > If zsmalloc _ever_ gets extended to support items that might
> > span three or more pages, a more generic TLB flush-pages-vs-flush-all
> > approach may be warranted and, by then, may already exist in some
> > future kernel.  Until then, IMHO, keep it simple.
> 
> I guess I'm not following.  Are you supporting the removal
> of the "break even" logic?  I added that logic as a
> compromise for Peter's feedback:
> 
> http://lkml.org/lkml/2012/5/17/177

Yes, as long as I am correct that zsmalloc never has to map/flush
more than two pages at a time, I think dealing with the break-even
logic is overkill.  I see Peter isn't on this dist list... maybe
you should ask him if he agrees, as long as we are only always
talking about flush-two-TLB-pages vs flush-all.

(And, of course, per previous discussion, I think even mapping/flushing
two TLB pages is unnecessary and overkill required only for protecting an
abstraction, but will stop beating that dead horse. ;-)

Dan

--
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>