CHROMIUM: watchdog: don't free perf_event on disable

When offlining a CPU, we currently free the perf_event used for
the hardlockup detector. This sometimes results in an allocation
failure when bringing the CPU back online. The hotplug code to
bring the CPU back online is run while gfp flags are masked by
pm_restrict_gfp_mask.

[10086.691854]  [<ffffffff810a4e80>] warn_alloc_failed+0x10f/0x123
[10086.691864]  [<ffffffff810a6560>] ? drain_local_pages+0x16/0x18
[10086.691872]  [<ffffffff810a75f8>] __alloc_pages_nodemask+0x697/0x752
[10086.691881]  [<ffffffff810a7738>] __get_free_pages+0x17/0x44
[10086.691891]  [<ffffffff810ccca0>] kmalloc_order_trace+0x2b/0x5b
[10086.691898]  [<ffffffff810cf095>] __kmalloc+0x37/0x151
[10086.691908]  [<ffffffff81011a0d>] ? kmalloc_node.isra.0.constprop.3+0xe/0x10
[10086.691917]  [<ffffffff81011a0d>] kmalloc_node.isra.0.constprop.3+0xe/0x10
[10086.691925]  [<ffffffff810120dc>] reserve_ds_buffers+0xaf/0x30e
[10086.691934]  [<ffffffff8100ec2c>] x86_pmu_event_init+0x2d2/0x31c
[10086.691941]  [<ffffffff8109e456>] perf_init_event+0x66/0xac
[10086.691948]  [<ffffffff8109e6fb>] perf_event_alloc+0x25f/0x38f
[10086.691957]  [<ffffffff81076488>] ? touch_nmi_watchdog+0x67/0x67
[10086.691965]  [<ffffffff8109eacd>] perf_event_create_kernel_counter+0x26/0xd9
[10086.691973]  [<ffffffff810762b0>] watchdog_enable+0x7e/0x1ef
[10086.691981]  [<ffffffff81076848>] cpu_callback+0x31/0x3f
[10086.691999]  [<ffffffff810769d4>] lockup_detector_bootcpu_resume+0x2e/0x32
[10086.692009]  [<ffffffff8105beab>] suspend_devices_and_enter+0x1a3/0x26e
[10086.692018]  [<ffffffff8105c075>] pm_suspend+0xff/0x1c2
[10086.692025]  [<ffffffff8105b3a1>] state_store+0x99/0xca
[10086.692034]  [<ffffffff811d6562>] kobj_attr_store+0xf/0x1b
[10086.692042]  [<ffffffff81123280>] sysfs_write_file+0xe9/0x121
[10086.692050]  [<ffffffff810d2fca>] vfs_write+0x98/0xda
[10086.692057]  [<ffffffff810d3199>] sys_write+0x43/0x73
[10086.692066]  [<ffffffff8146ac52>] system_call_fastpath+0x16/0x1b

The end result of the allocation failure is that the watchdog no
longer works sometimes after suspend/resume. The fix is to only
disable the event and not free it. This avoids the failed allocation.

BUG=chrome-os-partner:17522
TEST=Verify that detector still works after suspend/resume.

Change-Id: I78c90b13f718d660bd23884968be2f14c7c61860
Signed-off-by: Mandeep Singh Baines <msb@chromium.org>
Reviewed-on: https://gerrit.chromium.org/gerrit/42657
Reviewed-by: Sameer Nanda <snanda@chromium.org>

diff --git a/kernel/watchdog.c b/kernel/watchdog.c
index 6db3cb6ba071eaf7cdb9f7a23f392fed2cced8a7..f7a1568fbd44157c889567424da96d72a40cf814 100644 (file)
--- a/kernel/watchdog.c
+++ b/kernel/watchdog.c
@@ -408,13 +408,9 @@ static void watchdog_nmi_disable(int cpu)
 {
        struct perf_event *event = per_cpu(watchdog_ev, cpu);
 
-       if (event) {
+       if (event)
                perf_event_disable(event);
-               per_cpu(watchdog_ev, cpu) = NULL;
 
-               /* should be in cleanup, but blocks oprofile */
-               perf_event_release_kernel(event);
-       }
        return;
 }
 #else