[lkml]   [2008]   [Mar]   [10]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
SubjectRe: Regression: Re: [patch -mm 2/4] mempolicy: create mempolicy_operations structure
On Sat, 2008-03-08 at 14:09 -0800, David Rientjes wrote: 
> On Sat, 8 Mar 2008, Lee Schermerhorn wrote:
> > > Excuse me, but there was significant discussion about this on LKML and I
> > > eventually did force MPOL_DEFAULT to require a non-empty nodemask
> Correction: s/non-empty/empty

That makes more sense. I agree. more below...
> > > specifically because of your demand that it should. It didn't originally
> > > require this in my patchset, and now you're removing the exact same
> > > requirement that you demanded.
> > >
> > > You said on February 13:
> > >
> > > 1) we've discussed the issue of returning EINVAL for non-empty
> > > nodemasks with MPOL_DEFAULT. By removing this restriction, we run
> > > the risk of breaking applications if we should ever want to define
> > > a semantic to non-empty node mask for MPOL_DEFAULT.
> > >
> > > If you want to remove this requirement now (please get agreement from
> > > Paul) and are sure of your position, you'll at least need an update to
> > > Documentation/vm/numa-memory-policy.txt.
> >
> > Excuse me. I thought that the discussion--my position, anyway--was
> > about preserving existing behavior for MPOL_DEFAULT which is to require
> > an EMPTY [or NULL--same effect] nodemask. Not a NON-EMPTY one. See:
> >
> > It does appear that your patches now require a non-empty nodemask. This
> > was intentional?
> >
> The first and second set did not have this requirement, but the third set
> does (not currently in -mm), so I've changed it back. Hopefully there's
> no confusion and we can settle on a solution without continuously
> revisiting the topic.
> My position was originally to allow any type of nodemask to be passed with
> MPOL_DEFAULT since its not used. You asked for strict argument checking
> and so after some debate I changed it to require an empty nodemask mainly
> because I didn't want the patchset to stall on such a minor point. But in
> your regression fix, you expressed the desire once again to allow it to
> accept any nodemask because the testsuite does not check for it.

Not a desire. Just that when I fixed the MPOL_PREFERRED with empty node
mask regression, I also fixed mpol_new() not to require a non-empty
nodemask with MPOL_DEFAULT. I didn't go the extra step to require an
empty one. I'm tiring of the subject, as I think you are, and didn't
want to argue it anymore. So, I was willing to "cave" on that point.

> So if you'd like to do that, I'd encourage you to submit it as a separate
> patch and open it up for review.

No, I'm quite happy if, after your patches, the APIs retain the previous
behavior w/rt nodemask error checking.

> What is currently in -mm and what I will be posting shortly is the updated
> regression fix. All of these patches require that MPOL_DEFAULT include a
> NULL pointer or empty nodemask passed via the two syscalls.
> > Note: in the subject patch, I didn't enforce this behavior because your
> > patch didn't [it enforced just the opposite], and I've pretty much given
> > up. Although I prefer current behavior [before your series], if we
> > change it, we will need to change the man pages to remove the error
> > condition for non-empty nodemasks with MPOL_DEFAULT.
> >
> With my patches it still requires a NULL pointer or empty nodemask and
> I've updated Documentation/vm/numa_memory_policy.txt to explicitly say its
> an error if a non-empty nodemask is passed.


Do you intend for your patch entitled "[patch -mm v2] mempolicy:
disallow static or relative flags for local preferred mode" to replace
the patch that I sent in to repair the regression? Looks that way.
I'll replace it in my tree and retest.


 \ /
  Last update: 2008-03-10 16:01    [W:0.066 / U:8.184 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site