Re: [PATCH] numa: fix slab_node(MPOL_BIND)

[Eric Dumazet <eric.dumazet@gmail.com>]

Le jeudi 28 octobre 2010 à 08:59 -0700, Linus Torvalds a écrit :
> Hmm. More people added to the discussion..
> 
> This code seems to go back all the way to commit 19770b32609b: "mm:
> filter based on a nodemask as well as a gfp_mask". Which was back in
> April 2008. and got merged into 2.6.26.
> 
> And I'd be happy to commit it (in fact, I was going to), but when
> looking for other uses of first_zones_zonelist(), I found
> local_memory_node() which does the exact same thing: ignore the return
> value, and unconditionally dereference the resulting 'zone' variable.
> 
> And so does - although less obviously - mm/vmscan.c for the
> wait_iff_confgested() thing.
> 
> So are those buggy too, since first_zones_zonelist() can apparently return NULL?
> 
> Please advise...
> 

local_memory_node() is for ia64 only, and it is a bit different :

(void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
	gfp_zone(GFP_KERNEL),
	NULL,
	&zone);

Apparently node_zonelist(node, GFP_KERNEL) returns a pointer to
something that must have a zone suitable for GFP_KERNEL

While the code I tried to fix uses :

zonelist = &NODE_DATA(numa_node_id())->node_zonelists[0];
without any guarantee it contains a LOWMEM zone


Also check commit 7eb54824b76793dd86afb54f182ef9aa64b3a45a
for a similar fix in the past.

About do_try_to_free_pages(), I can not comment on this one, sorry.