Re: [GIT PULL] Please pull NFS client fixes for 4.12

[Trond Myklebust]

On Thu, 2017-05-11 at 14:26 +0200, Michal Hocko wrote:
> On Thu 11-05-17 12:16:37, Trond Myklebust wrote:
> > On Thu, 2017-05-11 at 09:59 +0200, Michal Hocko wrote:
> > > On Thu 11-05-17 10:53:27, Nikolay Borisov wrote:
> > > > 
> > > > 
> > > > On 10.05.2017 19:47, Trond Myklebust wrote:
> > > 
> > > [...]
> > > > > - Cleanup and removal of some memory failure paths now that
> > > > > Â GFP_NOFS is guaranteed to never fail.
> > > > 
> > > > What guarantees that? Since if this is the case then this can
> > > > result in
> > > > a lot of opportunities for cleanup across the whole kernel
> > > > tree.
> > > > After
> > > > discussing with mhocko (cc'ed) it seems that in practice
> > > > everything
> > > > below COSTLY_ORDER which are not GFP_NORETRY will never fail.
> > > > But
> > > > this
> > > > semantic is not the same as GFP_NOFAIL. E.g. nothing guarantees
> > > > that
> > > > this will stay like that in the future?
> > > 
> > > In practice it is hard to change the semantic of small
> > > allocations
> > > never
> > > fail _practically_. But this is absolutely not guaranteed! They
> > > can
> > > fail
> > > e.g. when the allocation context is the oom victim. Removing
> > > error
> > > paths
> > > for allocation failures is just wrong.
> > 
> > OK, this makes no fucking sense at all.
> > 
> > Either allocations can fail or they can't.
> > 1) If they can't fail, then we don't need the checks.
> > 2) If they can fail, then we do need them, and this hand wringing
> > in
> > the MM community about GFP_* semantics and how we need to prevent
> > failure is fucking pointless.
> 
> everything which is not __GFP_NOFAIL might fail. We try hard not to
> fail
> small allocations requests as much as we can in general but you
> _have_ to
> check for failures. There is simply no way to guarantee "never fail"
> semantic for all allocation requests. This has been like that
> basically
> since years. And even this try-to-be-nofailing for small allocations
> has
> been PITA for some corner cases.

I'll take that as a vote for (2), then.

I know that failures could occur in the past. That's why those code
paths were there. The problem is that the MM community has been making
lots of noise on mailing lists, conferences and LWN articles about how
we must not fail small allocations because the MM community believes
that nobody expects it. This is confusing everyone... It confused Neil
Brown, who contributed these patches, and it confused me and all the
other reviewers of these patches on the linux-nfs mailing list.

So if indeed (2) is correct, then please can we have a clear statement
_when discussing improvements to memory allocation semantics_ that
GFP_* still can fail, still will fail, and that callers should assume
it will fail and should test their code paths assuming the failure
case.

-- 
Trond Myklebust
Linux NFS client maintainer, PrimaryData
trond.myklebust@xxxxxxxxxxxxxxx