Re: [PATCH 6/6] Use one zonelist that is filtered by nodemask

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

On 09.11.2007 [12:18:52 -0500], Lee Schermerhorn wrote:
> On Fri, 2007-11-09 at 08:45 -0800, Nishanth Aravamudan wrote:
> > On 09.11.2007 [16:14:55 +0000], Mel Gorman wrote:
> > > On (09/11/07 07:45), Christoph Lameter didst pronounce:
> > > > On Fri, 9 Nov 2007, Mel Gorman wrote:
> > > > 
> > > > >  struct page * fastcall
> > > > >  __alloc_pages(gfp_t gfp_mask, unsigned int order,
> > > > >  		struct zonelist *zonelist)
> > > > >  {
> > > > > +	/*
> > > > > +	 * Use a temporary nodemask for __GFP_THISNODE allocations. If the
> > > > > +	 * cost of allocating on the stack or the stack usage becomes
> > > > > +	 * noticable, allocate the nodemasks per node at boot or compile time
> > > > > +	 */
> > > > > +	if (unlikely(gfp_mask & __GFP_THISNODE)) {
> > > > > +		nodemask_t nodemask;
> > > > 
> > > > Hmmm.. This places a potentially big structure on the stack. nodemask can 
> > > > contain up to 1024 bits which means 128 bytes. Maybe keep an array of 
> > > > gfp_thisnode nodemasks (node_nodemask?) and use node_nodemask[nid]?
> > > > 
> > > 
> > > That is what I was hinting at in the comment as a possible solution.
> > > 
> > > > > +
> > > > > +		return __alloc_pages_internal(gfp_mask, order,
> > > > > +			zonelist, nodemask_thisnode(numa_node_id(), &nodemask));
> > > > 
> > > > Argh.... GFP_THISNODE must use the nid passed to alloc_pages_node
> > > > and *not* the local numa node id. Only if the node specified to
> > > > alloc_pages nodes is -1 will this work.
> > > > 
> > > 
> > > alloc_pages_node() calls __alloc_pages_nodemask() though where in this
> > > function if I'm reading it right is called without a node id. Given no
> > > other details on the nid, the current one seemed a logical choice.
> > 
> > Yeah, I guess the context here matters (and is a little hard to follow
> > because thare are a few places that change in different ways here):
> > 
> > For allocating pages from a particular node (GFP_THISNODE with nid),
> > the nid clearly must be specified. This only happens with
> > alloc_pages_node(), AFAICT. So, in that interface, the right thing is
> > done and the appropriate nodemask will be built.
> 
> I agree.  In an earlier patch, Mel was ignoring nid and using
> numa_node_id() here.  This was causing your [Nish's] hugetlb pool
> allocation patches to fail.  Mel fixed that ~9oct07.  

Yep, and that's why I'm on the Cc, I think :)

> > On the other hand, if we call alloc_pages() with GFP_THISNODE set, there
> > is no nid to base the allocation on, so we "fallback" to numa_node_id()
> > [ almost like the nid had been specified as -1 ].
> > 
> > So I guess this is logical -- but I wonder, do we have any callers of
> > alloc_pages(GFP_THISNODE) ? It seems like an odd thing to do, when
> > alloc_pages_node() exists?
> 
> I don't know if we have any current callers that do this, but absent any
> documentation specifying otherwise, Mel's implementation matches what
> I'd expect the behavior to be if I DID call alloc_pages with 'THISNODE.
> However, we could specify that THISNODE is ignored in __alloc_pages()
> and recommend the use of alloc_pages_node() passing numa_node_id() as
> the nid parameter to achieve the behavior.  This would eliminate the
> check for 'THISNODE in __alloc_pages().  Just mask it off before calling
> down to __alloc_pages_internal().
> 
> Does this make sense?

The caller could also just use -1 as the nid, since then
alloc_pages_node() should do the numa_node_id() for the caller... But I
agree, there is no documentation saying GFP_THISNODE is *not* allowed
for alloc_pages(), so we should probably handle it

-Nish

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