Re: [PATCH] MM: discard __GFP_ATOMIC

[Mel Gorman]

On Tue, Sep 06, 2022 at 09:35:41AM +0200, Michal Hocko wrote:
> > From: "NeilBrown" <neilb@xxxxxxx>
> > Subject: mm: discard __GFP_ATOMIC
> > 
> > __GFP_ATOMIC serves little purpose.  Its main effect is to set
> > ALLOC_HARDER which adds a few little boosts to increase the chance of an
> > allocation succeeding, one of which is to lower the water-mark at which it
> > will succeed.
> > 
> > It is *always* paired with __GFP_HIGH which sets ALLOC_HIGH which also
> > adjusts this watermark.  It is probable that other users of __GFP_HIGH
> > should benefit from the other little bonuses that __GFP_ATOMIC gets.
> > 
> > __GFP_ATOMIC also gives a warning if used with __GFP_DIRECT_RECLAIM. 
> > There is little point to this.  We already get a might_sleep() warning if
> > __GFP_DIRECT_RECLAIM is set.
> > 
> > __GFP_ATOMIC allows the "watermark_boost" to be side-stepped.  It is
> > probable that testing ALLOC_HARDER is a better fit here.
> > 
> > __GFP_ATOMIC is used by tegra-smmu.c to check if the allocation might
> > sleep.  This should test __GFP_DIRECT_RECLAIM instead.
> > 
> > This patch:
> >  - removes __GFP_ATOMIC
> >  - causes __GFP_HIGH to set ALLOC_HARDER unless __GFP_NOMEMALLOC is set
> >    (as well as ALLOC_HIGH).
> >  - makes other adjustments as suggested by the above.
> > 
> > The net result is not change to GFP_ATOMIC allocations.  Other
> > allocations that use __GFP_HIGH will benefit from a few different extra
> > privileges.  This affects:
> >   xen, dm, md, ntfs3
> >   the vermillion frame buffer
> >   hibernation
> >   ksm
> >   swap
> > all of which likely produce more benefit than cost if these selected
> > allocation are more likely to succeed quickly.
> 
> This is a good summary of the current usage and existing issues. It also
> shows that the naming is tricky and allows people to make wrong calls
> (tegra-smmu.c). I also thing that it is wrong to couple memory reserves
> access to the reclaim constrains/expectations of the caller.
> 

I think it's worth trying to get rid of __GFP_ATOMIC although this patch
needs to be rebased. Without rebasing it, I suspect there is a corner case
for reserving high order atomic blocks. A high-order atomic allocation
might get confused with a __GFP_HIGH high-order allocation that can sleep.
It would not be completely irrational to have such a caller if it was in a
path that can tolerate a stall but stalling might have visible consequences.
I'm also worried that the patch might allow __GFP_HIGH to ignore cpusets
which is probably not intended by direct users like ksm.

> > Link: https://lkml.kernel.org/r/163712397076.13692.4727608274002939094@xxxxxxxxxxxxxxxxxxxxx
> > Signed-off-by: NeilBrown <neilb@xxxxxxx>
> > Reviewed-by: Matthew Wilcox (Oracle) <willy@xxxxxxxxxxxxx>
> > Cc: Michal Hocko <mhocko@xxxxxxxx>
> > Cc: Thierry Reding <thierry.reding@xxxxxxxxx>
> > Cc: Mel Gorman <mgorman@xxxxxxxxxxxxxxxxxxx>
> > Signed-off-by: Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx>
> 
> Yes, I am all for dropping the gfp flag. One thing that is not really
> entirely clear to me, though, is whether we still need 3 levels of
> memory reserves access. Can we just drop ALLOC_HARDER? With this patch
> applied it serves RT tasks and conflates it with __GFP_HIGH users
> essentially. So why do we need that additional level of reserves?

I think this would fall under the "naming is hard". If __GFP_ATOMIC was
removed, the ALLOC_ flags might need renaming to detect differences in
high priority allocations (RT + GFP_ATOMIC), critical allocations (OOM)
and ones that can access special reserves (GFP_ATOMIC high-order).

-- 
Mel Gorman
SUSE Labs