watchdog warnings on 3.4 triggered by perf

watchdog warnings on 3.4 triggered by perf

[Vinson Lee]

Hi.

We are seeing kernel watchdog warnings on Linux 3.4.78 on several
different machines. The warning occurs occasionally when there are
multiple instances of perf stat with an event list each tracing a
different process. This isn't a recent regression as we saw this
happen with earlier versions of 3.4.y too.

Here is the typical stack trace.

------------[ cut here ]------------
WARNING: at kernel/watchdog.c:241 watchdog_overflow_callback+0x9b/0xa6()
Watchdog detected hard LOCKUP on cpu 4
Modules linked in: netconsole configfs ipv6 dm_multipath video sbs
sbshc acpi_pad acpi_memhotplug acpi_ipmi parport_pc lp parport
tcp_diag inet_diag ipmi_si ipmi_d
andler igb dcdbas hed sb_edac i2c_i801 ioatdma i2c_core iTCO_wdt
shpchp iTCO_vendor_support edac_core wmi dca microcode
Pid: 20128, comm: java Not tainted 3.4.78 #1
Call Trace:
 <NMI>  [<ffffffff8103c20c>] warn_slowpath_common+0x85/0x9d
 [<ffffffff8103c2c7>] warn_slowpath_fmt+0x46/0x48
 [<ffffffff810689db>] ? sched_clock_cpu+0x44/0xce
 [<ffffffff810ad7c4>] watchdog_overflow_callback+0x9b/0xa6
 [<ffffffff810dbae7>] __perf_event_overflow+0x137/0x1c1
 [<ffffffff8100fa36>] ? x86_perf_event_set_period+0x107/0x113
 [<ffffffff810dbeeb>] perf_event_overflow+0x14/0x16
 [<ffffffff81014406>] intel_pmu_handle_irq+0x263/0x2e0
 [<ffffffff814bcc8f>] perf_event_nmi_handler+0x19/0x1b
 [<ffffffff814bc663>] default_do_nmi+0x66/0x1d4
 [<ffffffff814bc841>] do_nmi+0x70/0xbb
 [<ffffffff814bbd0c>] end_repeat_nmi+0x1a/0x1e
 <<EOE>>
---[ end trace 6a76b814d8600490 ]---

Cheers,
Vinson