Re: [PATCH 3/3] mm: Enable Buddy allocation isolation for CDM nodes

[Vlastimil Babka]

On 02/09/2017 06:05 AM, Anshuman Khandual wrote:
> On 02/08/2017 10:48 PM, Vlastimil Babka wrote:
> > On 02/08/2017 03:01 PM, Anshuman Khandual wrote:
> > > This implements allocation isolation for CDM nodes in buddy allocator by
> > > discarding CDM memory zones all the time except in the cases where the
> > > gfp
> > > flag has got __GFP_THISNODE or the nodemask contains CDM nodes in cases
> > > where it is non NULL (explicit allocation request in the kernel or user
> > > process MPOL_BIND policy based requests).
> > >
> > > Signed-off-by: Anshuman Khandual <khandual@xxxxxxxxxxxxxxxxxx>
> > > ---
> > >  mm/page_alloc.c | 19 +++++++++++++++++++
> > >  1 file changed, 19 insertions(+)
> > >
> > > diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> > > index 40908de..7d8c82a 100644
> > > --- a/mm/page_alloc.c
> > > +++ b/mm/page_alloc.c
> > > @@ -64,6 +64,7 @@
> > >  #include <linux/page_owner.h>
> > >  #include <linux/kthread.h>
> > >  #include <linux/memcontrol.h>
> > > +#include <linux/node.h>
> > >
> > >  #include <asm/sections.h>
> > >  #include <asm/tlbflush.h>
> > > @@ -2908,6 +2909,24 @@ get_page_from_freelist(gfp_t gfp_mask, unsigned
> > > int order, int alloc_flags,
> > >          struct page *page;
> > >          unsigned long mark;
> > >
> > > +        /*
> > > +         * CDM nodes get skipped if the requested gfp flag
> > > +         * does not have __GFP_THISNODE set or the nodemask
> > > +         * does not have any CDM nodes in case the nodemask
> > > +         * is non NULL (explicit allocation requests from
> > > +         * kernel or user process MPOL_BIND policy which has
> > > +         * CDM nodes).
> > > +         */
> > > +        if (is_cdm_node(zone->zone_pgdat->node_id)) {
> > > +            if (!(gfp_mask & __GFP_THISNODE)) {
> > > +                if (!ac->nodemask)
> > > +                    continue;
> > > +
> > > +                if (!nodemask_has_cdm(*ac->nodemask))
> > > +                    continue;
> >
> > nodemask_has_cdm() looks quite expensive, combined with the loop here
> > that's O(n^2). But I don't understand why you need it. If there is no
> > cdm node in the nodemask, then we never reach this code with a cdm node,
> > because the zonelist iterator already checks the nodemask? Am I missing
> > something?
>
> A CDM zone can be selected during zonelist iteration if
>
> 	(1) If nodemask is NULL (where all zones are eligible)
>
> 		(1) Skip it if __GFP_THISNODE is not mentioned
> 		(2) Pick it if __GFP_THISNODE is mentioned
>
> 	(2) If nodemask has CDM (where CDM zones are eligible)
>
> 		(1) Pick it if nodemask has CDM
> 		(2) Pick it if __GFP_THISNODE is mentioned
>
> (1) (1) Enforces the primary isolation
> (2) (1) Is the only option which could be O(n^2) as the worst case
>
> Checking for both the zone being a CDM zone and the nodemask containing
> CDM node has to happen together for (2) (1). But we dont run into this
> option unless we have first checked if request contains __GFP_THISNODE
> and that nodemask is really a non NULL value. Hence the number cases
> getting into (2) (1) should be less. IIUC only the user space MPOL_BIND
> ones will come here.

Maybe I'm still missing something, but when you do nodemask_has_cdm() above then 
we already passed "if (!ac->nodemask) continue" which means ac->nodemask is not 
null, which means the zonelist iterator already did the filtering on 
ac->nodemask, and if this zone passed the filter and it's a cdm zone, then it 
has to be set in the nodemask?