Re: Regression in 2.6.23-rc2-mm2, mounting cpusets causes a hang

[Serge E. Hallyn]

Quoting Serge E. Hallyn (serge@xxxxxxxxxx):
> Quoting Lee Schermerhorn (Lee.Schermerhorn@xxxxxx):
> > On Tue, 2007-08-14 at 13:03 -0500, Serge E. Hallyn wrote:
> > > Quoting Lee Schermerhorn (Lee.Schermerhorn@xxxxxx):
> > > > On Mon, 2007-08-13 at 15:12 -0500, Serge E. Hallyn wrote:
> > > > > Quoting Dhaval Giani (dhaval@xxxxxxxxxxxxxxxxxx):
> > > > > > Hi,
> > > > > > 
> > > > > > On mounting cpusets using containers, I have been hitting the following
> > > > > > bug.
> > > > > > 
> > > > > > 
> > > > > > -----------[ cut here ]------------
> > > > > > kernel BUG at kernel/cpuset.c:331!
> > > > <snip>
> > > > > > CONFIG_HIGHMEM=y
> > > > > > CONFIG_X86_PAE=y
> > > > > > # CONFIG_NUMA is not set
> > > > > > CONFIG_ARCH_POPULATES_NODE_MAP=y
> > > > > > CONFIG_SELECT_MEMORY_MODEL=y
> > > > > > CONFIG_FLATMEM_MANUAL=y
> > > > > > # CONFIG_DISCONTIGMEM_MANUAL is not set
> > > > > > # CONFIG_SPARSEMEM_MANUAL is not set
> > > > > > CONFIG_FLATMEM=y
> > > > > > CONFIG_FLAT_NODE_MEM_MAP=y
> > > > > > # CONFIG_SPARSEMEM_STATIC is not set
> > > > > > CONFIG_SPLIT_PTLOCK_CPUS=4
> > > > > > CONFIG_RESOURCES_64BIT=y
> > > > > > CONFIG_ZONE_DMA_FLAG=1
> > > > <snip>
> > > > > 
> > > > > Yeah, I'm seeing the same thing.  Oddly, my node_states[N_NORMAL_MEMORY]
> > > > > and node_states[N_HIGH_MEMORY] are empty, while node_states[N_ONLINE]
> > > > > contains my single cpu (on i386 kvm image).
> > > > > 
> > > > > -serge
> > > > 
> > > > Yes, you'll definitely hit that BUG if the N_HIGH_MEMORY mask is empty.
> > > > So far, I can't see how this could be, tho'.  __build_all_zonelists()
> > > > should be called for non-NUMA as well as NUMA.  It iterates over "all
> > > 
> > > Yup, and it is, and some debug statements insist that
> > > node_set_state(nid, N_HIGH_MEMORY) is being called for cpu 0,
> > > and the state is in fact being correctly set.
> > > 
> > > So it must get cleared later...
> > 
> > OK.  That helps.  I'll see what I can find...
> 
> Well apparently I lied?  I don't think it's getting cleared in node_states.
> I'm doing:
> 
> static void guarantee_online_mems(const struct cpuset *cs, nodemask_t *pmask)
> {
>         int nid=0;
>         printk(KERN_NOTICE "%s: at top, node %d is %d\n",
>                 __FUNCTION__, nid, node_state(nid, N_HIGH_MEMORY));
>         while (cs && !nodes_intersects(cs->mems_allowed,
>                                         node_states[N_HIGH_MEMORY])) {
>                 printk(KERN_NOTICE "%s: in loop\n", __FUNCTION__);
>                 cs = cs->parent;
>         }
>         if (cs) {
>                 printk(KERN_NOTICE "%s: cs was true\n", __FUNCTION__);
>                 nodes_and(*pmask, cs->mems_allowed,
>                                         node_states[N_HIGH_MEMORY]);
>         }
>         else
>         {
>                 printk(KERN_NOTICE "%s: cs was false\n", __FUNCTION__);
>                 *pmask = node_states[N_HIGH_MEMORY];
>         }
>         printk(KERN_NOTICE "%s: at bottom, node %d is %d\n",
>                 __FUNCTION__, nid, node_state(nid, N_HIGH_MEMORY));
>         printk(KERN_NOTICE "%s: at bottom, in pmask node %d is %d\n",
>                 __FUNCTION__, nid, node_isset(nid, *pmask));
>         BUG_ON(!nodes_intersects(*pmask, node_states[N_HIGH_MEMORY]));
> }
> 
> and getting:
> 
> guarantee_online_mems: in loop
> guarantee_online_mems: cs was false
> guarantee_online_mems: at bottom, node 0 is 1
> guarantee_online_mems: at bottom, in pmask node 0 is 0
> 
> 
> So I'm guess that
> 	*pmask = node_states[N_HIGH_MEMORY];
> isn't doing what it was intended to do?  That or I've got node_state()
> wrong...

Argh.

CONFIG_NODES_SHIFT was unset, so MAX_NUMNODES=1, so
node_state() and node_set_state() are dummies.

So __build_all_zonelists is using the dummy node_state() and
not actually setting node_states[N_HIGH_MEMORY], and my debug
statements were using the dumy node_state() which returns 1
if n==0  :)

Here is a patch which fixes the bug on my system.  As noted in
the patch description, there are other calls to node_set_state() in
mm/page_alloc.c which might also need to be switched to node_set(),
if this is deemed the right solution.

thanks,
-serge