lkml.org 
[lkml]   [2008]   [Jun]   [25]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
Patch in this message
/
From
SubjectRe: [BUG] While changing the cpufreq governor, kernel hits a bug in workqueue.c
Date
Hi,

Nageswara R Sastry <rnsastry@linux.vnet.ibm.com> writes:

> Johannes Weiner wrote:
>>
>> From: Johannes Weiner <hannes@saeurebad.de>
>> Subject: cpufreq: Fix race in enabling ondemand/conservative governors
>>
>> Prevent double activation of the governor if two processes race on the
>> check for whether the governor is already active.
>>
>> Signed-off-by: Johannes Weiner <hannes@saeurebad.de>
>> ---
>>
>> diff --git a/drivers/cpufreq/cpufreq_conservative.c b/drivers/cpufreq/cpufreq_conservative.c
>> index 5d3a04b..a4902e4 100644
>> --- a/drivers/cpufreq/cpufreq_conservative.c
>> +++ b/drivers/cpufreq/cpufreq_conservative.c
>> @@ -486,10 +486,11 @@ static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
>> if ((!cpu_online(cpu)) || (!policy->cur))
>> return -EINVAL;
>>
>> - if (this_dbs_info->enable) /* Already enabled */
>> - break;
>> -
>> mutex_lock(&dbs_mutex);
>> + if (this_dbs_info->enable) {
>> + mutex_unlock(&dbs_mutex);
>> + break;
>> + }
>>
>> rc = sysfs_create_group(&policy->kobj, &dbs_attr_group);
>> if (rc) {
>> diff --git a/drivers/cpufreq/cpufreq_ondemand.c b/drivers/cpufreq/cpufreq_ondemand.c
>> index d2af20d..61705e1 100644
>> --- a/drivers/cpufreq/cpufreq_ondemand.c
>> +++ b/drivers/cpufreq/cpufreq_ondemand.c
>> @@ -508,10 +508,12 @@ static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
>> if ((!cpu_online(cpu)) || (!policy->cur))
>> return -EINVAL;
>>
>> - if (this_dbs_info->enable) /* Already enabled */
>> + mutex_lock(&dbs_mutex);
>> + if (this_dbs_info->enable) {
>> + mutex_unlock(&dbs_mutex);
>> break;
>> + }
>>
>> - mutex_lock(&dbs_mutex);
>> dbs_enable++;
>>
>> rc = sysfs_create_group(&policy->kobj, &dbs_attr_group);
>
> Applied the above patch against 2.6.25.7 kernel and tested the same,
> But unluckily the above patch didn't stop hitting kernel bug at
> kernel/workqueue.c:223
> Hitting bug time got delayed near to 2hrs with the above patch.
>
>>> Kernel stack trace:
> ------------[ cut here ]------------
> kernel BUG at kernel/workqueue.c:223!
> invalid opcode: 0000 [#2] SMP
> Modules linked in: cpufreq_powersave cpufreq_conservative
> cpufreq_userspace usbhid usb_storage ehci_hcd ohci_hcd uhci_hcd
> usbcore
>
> Pid: 232, comm: kondemand/1 Tainted: G D (2.6.25.7.cpufreqpatch #2)
> EIP: 0060:[<c012f61a>] EFLAGS: 00010286 CPU: 1
> EIP is at queue_delayed_work_on+0x20/0x97
> EAX: 00000000 EBX: c483ba94 ECX: c483ba94 EDX: 00000000
> ESI: c483bab0 EDI: f7a39708 EBP: 00000001 ESP: f7a69f40
> DS: 007b ES: 007b FS: 00d8 GS: 0000 SS: 0068
> Process kondemand/1 (pid: 232, ti=f7a68000 task=f79b4120 task.ti=f7a68000)
> Stack: 00000000 f7a377b0 c483ba80 f78bed80 c041e54d 00000040 00000000
> 00000001
> 00000040 00000246 00000000 00000002 00000000 c012ee7f c483ba98
> f7a377b0
> c483ba94 f7a69f9c c012eeba 00000000 00000002 c012ee7f c041e31e
> c099e2a8
> Call Trace:
> [<c041e54d>] do_dbs_timer+0x22f/0x24f
> [<c012ee7f>] run_workqueue+0x81/0x187
> [<c012eeba>] run_workqueue+0xbc/0x187
> [<c012ee7f>] run_workqueue+0x81/0x187
> [<c041e31e>] do_dbs_timer+0x0/0x24f
> [<c012f6fa>] worker_thread+0x0/0xbd
> [<c012f7ad>] worker_thread+0xb3/0xbd
> [<c0131acc>] autoremove_wake_function+0x0/0x2d
> [<c0131a1b>] kthread+0x38/0x5d
> [<c01319e3>] kthread+0x0/0x5d
> [<c0105527>] kernel_thread_helper+0x7/0x10
> =======================
> Code: c3 a1 dc da 6a c0 e9 78 ff ff ff 55 89 c5 57 89 d7 56 53 89 cb
> 8d 71 1c f0 0f ba 29 00 19 c0 31 d2 85 c0 75 76 83 79 1c 00 74 04 <0f>
> 0b eb fe 8d 41 04 39 41 04 74 04 0f 0b eb fe 89 f8 64 8b 15
> EIP: [<c012f61a>] queue_delayed_work_on+0x20/0x97 SS:ESP 0068:f7a69f40
> ---[ end trace 1f82bd8b1ee896cc ]---
>
> And when the script is running opened one more terminal and issued the
> following command to observe the state of governors.
>
> # watch -n 1 "cat /proc/cpuinfo | grep 'cpu MHz' ; cat
> /proc/acpi/processor/CPU*/performance; cat
> /sys/devices/system/cpu/*/cpufreq/scaling_governor; cat
> /sys/devices/system/cpu/*/cpufreq/scaling_cur_freq"
>
> The above command hit with an another kernel BUG at
> kernel/workqueue.c:272 the kernel stack trace is,
>
> ------------[ cut here ]------------
> kernel BUG at kernel/workqueue.c:272!
> invalid opcode: 0000 [#1] SMP
> Modules linked in: cpufreq_powersave cpufreq_conservative
> cpufreq_userspace usbhid usb_storage ehci_hcd ohci_hcd uhci_hcd
> usbcore
>
> Pid: 233, comm: kondemand/2 Not tainted (2.6.25.7.cpufreqpatch #2)
> EIP: 0060:[<c012ee7b>] EFLAGS: 00010212 CPU: 2
> EIP is at run_workqueue+0x7d/0x187
> EAX: f7a377b0 EBX: c4844a98 ECX: 00000000 EDX: 00000000
> ESI: f7a37718 EDI: c4844a94 EBP: f79f9f9c ESP: f79f9f98
> DS: 007b ES: 007b FS: 00d8 GS: 0000 SS: 0068
> Process kondemand/2 (pid: 233, ti=f79f8000 task=f79b5020 task.ti=f79f8000)
> Stack: c041e31e c099e2a8 00000000 c061ec35 f7a37718 c012f6fa f7a3773c
> 00000000
> c012f7ad 00000000 f79b5020 c0131acc f79f9fc8 f79f9fc8 f7a37718
> 00000000
> c0131a1b c01319e3 00000000 c0105527 f784bef0 00000000 00000000
> 00000000
> Call Trace:
> [<c041e31e>] do_dbs_timer+0x0/0x24f
> [<c012f6fa>] worker_thread+0x0/0xbd
> [<c012f7ad>] worker_thread+0xb3/0xbd
> [<c0131acc>] autoremove_wake_function+0x0/0x2d
> [<c0131a1b>] kthread+0x38/0x5d
> [<c01319e3>] kthread+0x0/0x5d
> [<c0105527>] kernel_thread_helper+0x7/0x10
> =======================
> Code: e8 d3 2f 16 00 8b 46 1c 89 7e 48 8b 08 8b 50 04 89 51 04 89 0a
> 89 40 04 89 00 89 f0 e8 59 ff 39 00 8b 43 fc 83 e0 fc 39 f0 74 04 <0f>
> 0b eb fe f0 0f ba 73 fc 00 8b 46 4c 31 c9 31 d2 68 7f ee 12
> EIP: [<c012ee7b>] run_workqueue+0x7d/0x187 SS:ESP 0068:f79f9f98
> ---[ end trace 1f82bd8b1ee896cc ]---

Cool, thanks for testing!

I think the above patch is right but not enough.

Dave, what do you think about the following?

---
From: Johannes Weiner <hannes@saeurebad.de>
Subject: cpufreq: cancel self-rearming work synchroneuously

The ondemand and conservative governor workers are self-rearming.
Cancel them synchroneously to avoid nasty races.

Reported-by: Nageswara R Sastry <rnsastry@linux.vnet.ibm.com>
Signed-off-by: Johannes Weiner <hannes@saeurebad.de>
---

diff --git a/drivers/cpufreq/cpufreq_conservative.c b/drivers/cpufreq/cpufreq_conservative.c
index 5d3a04b..78bac06 100644
--- a/drivers/cpufreq/cpufreq_conservative.c
+++ b/drivers/cpufreq/cpufreq_conservative.c
@@ -467,7 +467,7 @@ static inline void dbs_timer_init(void)

static inline void dbs_timer_exit(void)
{
- cancel_delayed_work(&dbs_work);
+ cancel_delayed_work_sync(&dbs_work);
return;
}

diff --git a/drivers/cpufreq/cpufreq_ondemand.c b/drivers/cpufreq/cpufreq_ondemand.c
index d2af20d..1eb8c58 100644
--- a/drivers/cpufreq/cpufreq_ondemand.c
+++ b/drivers/cpufreq/cpufreq_ondemand.c
@@ -490,7 +490,7 @@ static inline void dbs_timer_init(struct cpu_dbs_info_s *dbs_info)
static inline void dbs_timer_exit(struct cpu_dbs_info_s *dbs_info)
{
dbs_info->enable = 0;
- cancel_delayed_work(&dbs_info->work);
+ cancel_delayed_work_sync(&dbs_info->work);
}

static int cpufreq_governor_dbs(struct cpufreq_policy *policy,

\
 
 \ /
  Last update: 2008-06-25 21:51    [W:0.173 / U:0.788 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site