CPU hotplug: fix cpu_is_offline() on !CONFIG_HOTPLUG_CPU
make randconfig bootup testing found that the cpufreq code
crashes on bootup, if the powernow-k8 driver is enabled and
if maxcpus=1 passed on the boot line to a !CONFIG_HOTPLUG_CPU
kernel.
First lockdep found out that there's an inconsistent unlock
sequence:
=====================================
[ BUG: bad unlock balance detected! ]
-------------------------------------
swapper/1 is trying to release lock (&per_cpu(cpu_policy_rwsem, cpu)) at:
[<ffffffff806ffd8e>] unlock_policy_rwsem_write+0x3c/0x42
but there are no more locks to release!
Call Trace:
[<ffffffff806ffd8e>] unlock_policy_rwsem_write+0x3c/0x42
[<ffffffff80251c29>] print_unlock_inbalance_bug+0x104/0x12c
[<ffffffff80252f3a>] mark_held_locks+0x56/0x94
[<ffffffff806ffd8e>] unlock_policy_rwsem_write+0x3c/0x42
[<ffffffff807008b6>] cpufreq_add_dev+0x2a8/0x5c4
...
then shortly afterwards the cpufreq code crashed on an assert:
------------[ cut here ]------------
kernel BUG at drivers/cpufreq/cpufreq.c:1068!
invalid opcode: 0000 [1] SMP
[...]
Call Trace:
[<ffffffff805145d6>] sysdev_driver_unregister+0x5b/0x91
[<ffffffff806ff520>] cpufreq_register_driver+0x15d/0x1a2
[<ffffffff80cc0596>] powernowk8_init+0x86/0x94
[...]
---[ end trace 1e9219be2b4431de ]---
the bug was caused by maxcpus=1 bootup, which brought up the
secondary core as !cpu_online() but !cpu_is_offline() either,
which on on !CONFIG_HOTPLUG_CPU is always 0 (include/linux/cpu.h):
/* CPUs don't go offline once they're online w/o CONFIG_HOTPLUG_CPU */
static inline int cpu_is_offline(int cpu) { return 0; }
but the cpufreq code uses cpu_online() and cpu_is_offline() in
a mixed way - the low-level drivers use cpu_online(), while
the cpufreq core uses cpu_is_offline(). This opened up the
possibility to add the non-initialized sysdev device of the
secondary core:
cpufreq-core: trying to register driver powernow-k8
cpufreq-core: adding CPU 0
powernow-k8: BIOS error - no PSB or ACPI _PSS objects
cpufreq-core: initialization failed
cpufreq-core: adding CPU 1
cpufreq-core: initialization failed
which then blew up. The fix is to make cpu_is_offline() always
the negation of cpu_online(). With that fix applied the kernel
boots up fine without crashing:
Calling initcall 0xffffffff80cc0510: powernowk8_init+0x0/0x94()
powernow-k8: Found 1 AMD Athlon(tm) 64 X2 Dual Core Processor 3800+ processors (1 cpu cores) (version 2.20.00)
powernow-k8: BIOS error - no PSB or ACPI _PSS objects
initcall 0xffffffff80cc0510: powernowk8_init+0x0/0x94() returned -19.
initcall 0xffffffff80cc0510 ran for 19 msecs: powernowk8_init+0x0/0x94()
Calling initcall 0xffffffff80cc328f: init_lapic_nmi_sysfs+0x0/0x39()
We could fix this by making CPU enumeration aware of max_cpus, but that
would be more fragile IMO, and the cpu_online(cpu) != cpu_is_offline(cpu)
possibility was quite confusing and a continuous source of bugs too.
Most distributions have kernels with CPU hotplug enabled, so this bug
remained hidden for a long time.
Bug forensics:
The broken cpu_is_offline() API variant was introduced via:
commit a59d2e4e6977e7b94e003c96a41f07e96cddc340
Author: Rusty Russell <rusty@rustcorp.com.au>
Date: Mon Mar 8 06:06:03 2004 -0800
[PATCH] minor cleanups for hotplug CPUs
( this predates linux-2.6.git, this commit is available from Thomas's
historic git tree. )
Then 1.5 years later the cpufreq code made use of it:
commit c32b6b8e524d2c337767d312814484d9289550cf
Author: Ashok Raj <ashok.raj@intel.com>
Date: Sun Oct 30 14:59:54 2005 -0800
[PATCH] create and destroy cpufreq sysfs entries based on cpu notifiers
+ if (cpu_is_offline(cpu))
+ return 0;
which is a correct use of the subtly broken new API. v2.6.15 then
shipped with this bug included.
then it took two more years for random-kernel qa to hit it.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2 files changed