power, sched: stop updating inside arch_update_cpu_topology() when nothing to be...
authorMichael Wang <wangyun@linux.vnet.ibm.com>
Tue, 8 Apr 2014 03:19:36 +0000 (11:19 +0800)
committerBenjamin Herrenschmidt <benh@kernel.crashing.org>
Wed, 9 Apr 2014 02:54:17 +0000 (12:54 +1000)
commit9a0133613e4412b2caaaf0d9dd81f213bcededf1
tree8298db7c2bfa2a4d22d6d9f0a8eee024d53265ac
parentd3d35d957a9d0733dc51f14b5abc0bff5d3c5f3a
power, sched: stop updating inside arch_update_cpu_topology() when nothing to be update

Since v1:
Edited the comment according to Srivatsa's suggestion.

During the testing, we encounter below WARN followed by Oops:

WARNING: at kernel/sched/core.c:6218
...
NIP [c000000000101660] .build_sched_domains+0x11d0/0x1200
LR [c000000000101358] .build_sched_domains+0xec8/0x1200
PACATMSCRATCH [800000000000f032]
Call Trace:
[c00000001b103850] [c000000000101358] .build_sched_domains+0xec8/0x1200
[c00000001b1039a0] [c00000000010aad4] .partition_sched_domains+0x484/0x510
[c00000001b103aa0] [c00000000016d0a8] .rebuild_sched_domains+0x68/0xa0
[c00000001b103b30] [c00000000005cbf0] .topology_work_fn+0x10/0x30
...
Oops: Kernel access of bad area, sig: 11 [#1]
...
NIP [c00000000045c000] .__bitmap_weight+0x60/0xf0
LR [c00000000010132c] .build_sched_domains+0xe9c/0x1200
PACATMSCRATCH [8000000000029032]
Call Trace:
[c00000001b1037a0] [c000000000288ff4] .kmem_cache_alloc_node_trace+0x184/0x3a0
[c00000001b103850] [c00000000010132c] .build_sched_domains+0xe9c/0x1200
[c00000001b1039a0] [c00000000010aad4] .partition_sched_domains+0x484/0x510
[c00000001b103aa0] [c00000000016d0a8] .rebuild_sched_domains+0x68/0xa0
[c00000001b103b30] [c00000000005cbf0] .topology_work_fn+0x10/0x30
...

This was caused by that 'sd->groups == NULL' after building groups, which
was caused by the empty 'sd->span'.

The cpu's domain contained nothing because the cpu was assigned to a wrong
node, due to the following unfortunate sequence of events:

1. The hypervisor sent a topology update to the guest OS, to notify changes
   to the cpu-node mapping. However, the update was actually redundant - i.e.,
   the "new" mapping was exactly the same as the old one.

2. Due to this, the 'updated_cpus' mask turned out to be empty after exiting
   the 'for-loop' in arch_update_cpu_topology().

3. So we ended up calling stop-machine() with an empty cpumask list, which made
   stop-machine internally elect cpumask_first(cpu_online_mask), i.e., CPU0 as
   the cpu to run the payload (the update_cpu_topology() function).

4. This causes update_cpu_topology() to be run by CPU0. And since 'updates'
   is kzalloc()'ed inside arch_update_cpu_topology(), update_cpu_topology()
   finds update->cpu as well as update->new_nid to be 0. In other words, we
   end up assigning CPU0 (and eventually its siblings) to node 0, incorrectly.

Along with the following wrong updating, it causes the sched-domain rebuild
code to break and crash the system.

Fix this by skipping the topology update in cases where we find that
the topology has not actually changed in reality (ie., spurious updates).

CC: Benjamin Herrenschmidt <benh@kernel.crashing.org>
CC: Paul Mackerras <paulus@samba.org>
CC: Nathan Fontenot <nfont@linux.vnet.ibm.com>
CC: Stephen Rothwell <sfr@canb.auug.org.au>
CC: Andrew Morton <akpm@linux-foundation.org>
CC: Robert Jennings <rcj@linux.vnet.ibm.com>
CC: Jesse Larrew <jlarrew@linux.vnet.ibm.com>
CC: "Srivatsa S. Bhat" <srivatsa.bhat@linux.vnet.ibm.com>
CC: Alistair Popple <alistair@popple.id.au>
Suggested-by: "Srivatsa S. Bhat" <srivatsa.bhat@linux.vnet.ibm.com>
Signed-off-by: Michael Wang <wangyun@linux.vnet.ibm.com>
Reviewed-by: Srivatsa S. Bhat <srivatsa.bhat@linux.vnet.ibm.com>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
arch/powerpc/mm/numa.c