Re: [PATCH] populated_map: fix !NUMA case, remove comment

[Nishanth Aravamudan <nacc@us.ibm.com>]

On 13.06.2007 [16:12:49 -0700], Christoph Lameter wrote:
> On Wed, 13 Jun 2007, Nishanth Aravamudan wrote:
> 
> > > > That way, if the node has no memory whatsoever, we don't bother checking
> > > > the pgdat of the relevant zone?
> > > 
> > > Checking the pgdat is already done in __alloc_pages. No need to repeat
> > > it here.
> > 
> > Except that check is broken in the same way it is for memoryless nodes,
> > right?
> > 
> > from get_page_from_freelist():
> > 
> >                 if (unlikely(NUMA_BUILD && (gfp_mask & __GFP_THISNODE) &&
> >                         zone->zone_pgdat != zonelist->zones[0]->zone_pgdat))
> > 
> > Which asks if for this zone, is the first node the same as each node we look at
> > for THISNODE requests. But if the first node for the zone is a
> > *different* node, we still satisfy the request, but go off-node?
> > 
> > Just trying to see if that maybe is the problem here?
> 
> Right. But we do not have the pgdat pointer available in alloc_pages.
> Thus Lee's check works in alloc_pages_node().

Yep, exactly.

> Hmmm... This gets pretty difficult to comprehend. Maybe there is
> another easier way to implement GFP_THISNODE?

Well...maybe we can do better by just adding another GFP flag?

GFP_ONLYTHISNODE?

THISNODE has the current semantics, that the "closest" node is
preferred, which may be local, and it will succeed if memory exists
somewhere for the allocation you want (I think).

ONLYTHISNODE will return NULL if it has to go off-node for any reason.

> The breakage of SLUB makes it pretty evident that if GFP_THISNODE
> returns NULL for a memoryless node then lots of
> 
> for_each_online_node()
> 
> loops in the VM that assume that an online node contain memory are no
> longer working properly. We need to review the VM and convert those
> loops to use the node_memory_map.

That would avoid having to make these changes too.

Maybe with time, we can audit the users of THISNODE and move them over
to ONLYTHISNODE, as appropriate?

Thanks,
Nish

-- 
Nishanth Aravamudan <nacc@us.ibm.com>
IBM Linux Technology Center

--
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>