sched: Fix unreleased llc_shared_mask bit during CPU hotplug
authorWanpeng Li <wanpeng.li@linux.intel.com>
Wed, 24 Sep 2014 08:38:05 +0000 (16:38 +0800)
committerIngo Molnar <mingo@kernel.org>
Wed, 24 Sep 2014 13:13:20 +0000 (15:13 +0200)
The following bug can be triggered by hot adding and removing a large number of
xen domain0's vcpus repeatedly:

BUG: unable to handle kernel NULL pointer dereference at 0000000000000004 IP: [..] find_busiest_group
PGD 5a9d5067 PUD 13067 PMD 0
Oops: 0000 [#3] SMP
[...]
Call Trace:
load_balance
? _raw_spin_unlock_irqrestore
idle_balance
__schedule
schedule
schedule_timeout
? lock_timer_base
schedule_timeout_uninterruptible
msleep
lock_device_hotplug_sysfs
online_store
dev_attr_store
sysfs_write_file
vfs_write
SyS_write
system_call_fastpath

Last level cache shared mask is built during CPU up and the
build_sched_domain() routine takes advantage of it to setup
the sched domain CPU topology.

However, llc_shared_mask is not released during CPU disable,
which leads to an invalid sched domainCPU topology.

This patch fix it by releasing the llc_shared_mask correctly
during CPU disable.

Yasuaki also reported that this can happen on real hardware:

  https://lkml.org/lkml/2014/7/22/1018

His case is here:

==
Here is an example on my system.
My system has 4 sockets and each socket has 15 cores and HT is
enabled. In this case, each core of sockes is numbered as
follows:

 | CPU#
Socket#0 | 0-14 , 60-74
Socket#1 | 15-29, 75-89
Socket#2 | 30-44, 90-104
Socket#3 | 45-59, 105-119

Then llc_shared_mask of CPU#30 has 0x3fff80000001fffc0000000.

It means that last level cache of Socket#2 is shared with
CPU#30-44 and 90-104.

When hot-removing socket#2 and #3, each core of sockets is
numbered as follows:

 | CPU#
Socket#0 | 0-14 , 60-74
Socket#1 | 15-29, 75-89

But llc_shared_mask is not cleared. So llc_shared_mask of CPU#30
remains having 0x3fff80000001fffc0000000.

After that, when hot-adding socket#2 and #3, each core of
sockets is numbered as follows:

 | CPU#
Socket#0 | 0-14 , 60-74
Socket#1 | 15-29, 75-89
Socket#2 | 30-59
Socket#3 | 90-119

Then llc_shared_mask of CPU#30 becomes
0x3fff8000fffffffc0000000. It means that last level cache of
Socket#2 is shared with CPU#30-59 and 90-104. So the mask has
the wrong value.

Signed-off-by: Wanpeng Li <wanpeng.li@linux.intel.com>
Tested-by: Linn Crosetto <linn@hp.com>
Reviewed-by: Borislav Petkov <bp@suse.de>
Reviewed-by: Toshi Kani <toshi.kani@hp.com>
Reviewed-by: Yasuaki Ishimatsu <isimatu.yasuaki@jp.fujitsu.com>
Cc: <stable@vger.kernel.org>
Cc: David Rientjes <rientjes@google.com>
Cc: Prarit Bhargava <prarit@redhat.com>
Cc: Steven Rostedt <srostedt@redhat.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: http://lkml.kernel.org/r/1411547885-48165-1-git-send-email-wanpeng.li@linux.intel.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
arch/x86/kernel/smpboot.c

index 2d872e08fab95f10e8b97c23b54ba1e886680369..42a2dca984b3d1816522684d97c8ac50d86193af 100644 (file)
@@ -1284,6 +1284,9 @@ static void remove_siblinginfo(int cpu)
 
        for_each_cpu(sibling, cpu_sibling_mask(cpu))
                cpumask_clear_cpu(cpu, cpu_sibling_mask(sibling));
+       for_each_cpu(sibling, cpu_llc_shared_mask(cpu))
+               cpumask_clear_cpu(cpu, cpu_llc_shared_mask(sibling));
+       cpumask_clear(cpu_llc_shared_mask(cpu));
        cpumask_clear(cpu_sibling_mask(cpu));
        cpumask_clear(cpu_core_mask(cpu));
        c->phys_proc_id = 0;