Lukewarm IQ detected in hotplug locking
BUG: warning at kernel/cpu.c:38/lock_cpu_hotplug()
[<b0134a42>] lock_cpu_hotplug+0x42/0x65
[<b02f8af1>] cpufreq_update_policy+0x25/0xad
[<b0358756>] kprobe_flush_task+0x18/0x40
[<b0355aab>] schedule+0x63f/0x68b
[<b01377c2>] __link_module+0x0/0x1f
[<b0119e7d>] __cond_resched+0x16/0x34
[<b03560bf>] cond_resched+0x26/0x31
[<b0355b0e>] wait_for_completion+0x17/0xb1
[<f965c547>] cpufreq_stat_cpu_callback+0x13/0x20 [cpufreq_stats]
[<f9670074>] cpufreq_stats_init+0x74/0x8b [cpufreq_stats]
[<b0137872>] sys_init_module+0x91/0x174
[<b0102c81>] sysenter_past_esp+0x56/0x79
As there are other places that call cpufreq_update_policy without
the hotplug lock, it seems better to keep the hotplug locking
at the lower level for the time being until this is revamped.
Signed-off-by: Dave Jones <davej@redhat.com>
This patch makes the needlessly global powersave_bias_target() static.
Signed-off-by: Adrian Bunk <bunk@stusta.de>
Signed-off-by: Dave Jones <davej@redhat.com>
ondemand selects the minimum frequency that can retire
a workload with negligible idle time -- ideally resulting in the highest
performance/power efficiency with negligible performance impact.
But on some systems and some workloads, this algorithm
is more performance biased than necessary, and
de-tuning it a bit to allow some performance impact
can save measurable power.
This patch adds a "powersave_bias" tunable to ondemand
to allow it to reduce its target frequency by a specified percent.
By default, the powersave_bias is 0 and has no effect.
powersave_bias is in units of 0.1%, so it has an effective range
of 1 through 1000, resulting in 0.1% to 100% impact.
In practice, users will not be able to detect a difference between
0.1% increments, but 1.0% increments turned out to be too large.
Also, the max value of 1000 (100%) would simply peg the system
in its deepest power saving P-state, unless the processor really has
a hardware P-state at 0Hz:-)
For example, If ondemand requests 2.0GHz based on utilization,
and powersave_bias=100, this code will knock 10% off the target
and seek a target of 1.8GHz instead of 2.0GHz until the
next sampling. If 1.8 is an exact match with an hardware frequency
we use it, otherwise we average our time between the frequency
next higher than 1.8 and next lower than 1.8.
Note that a user or administrative program can change powersave_bias
at run-time depending on how they expect the system to be used.
Signed-off-by: Venkatesh Pallipadi <venkatesh.pallipadi at intel.com>
Signed-off-by: Alexey Starikovskiy <alexey.y.starikovskiy at intel.com>
Signed-off-by: Dave Jones <davej@redhat.com>
Try to make dbs_check_cpu() call on all CPUs at the same jiffy.
This will help when multiple cores share P-states via Hardware Coordination.
Signed-off-by: Venkatesh Pallipadi <venkatesh.pallipadi at intel.com>
Signed-off-by: Alexey Starikovskiy <alexey.y.starikovskiy at intel.com>
Signed-off-by: Dave Jones <davej@redhat.com>
Adds a __find_governor() helper function to look up a governor by
name. Also restructures some error handling to conform to the
"single-exit" model which is generally preferred for kernel code.
Signed-off-by: Jeremy Fitzhardinge <jeremy@goop.org>
Signed-off-by: Dave Jones <davej@redhat.com>
I just stumbled on this bug/feature, this is how to reproduce it:
# echo 450000 > /sys/devices/system/cpu/cpu0/cpufreq/scaling_min_freq
# echo 450000 > /sys/devices/system/cpu/cpu0/cpufreq/scaling_max_freq
# echo powersave > /sys/devices/system/cpu/cpu0/cpufreq/scaling_governor
# cpufreq-info -p
450000 450000 powersave
# echo 1800000 > /sys/devices/system/cpu/cpu0/cpufreq/scaling_min_freq ; echo $?
0
# cpufreq-info -p
450000 450000 powersave
Here it is. The kernel refuses to set a min_freq higher than the
max_freq but it allows a max_freq lower than min_freq (lowering min_freq
also).
This behaviour is pretty straightforward (but undocumented) and it
doesn't return an error altough failing to accomplish the requested
action (set min_freq).
The problem (IMO) is basically that userspace is not allowed to set a
full policy atomically while the kernel always does that thus it must
enforce an ordering on operations.
The attached patch returns -EINVAL if trying to increase frequencies
starting from scaling_min_freq and documents the correct ordering of writes.
Signed-off-by: Mattia Dongili <malattia@linux.it>
Signed-off-by: Dominik Brodowski <linux at dominikbrodowski.net>
Signed-off-by: Dave Jones <davej@redhat.com>
--
The patch below moves the cpu hotplugging higher up in the cpufreq
layering; this is needed to avoid recursive taking of the cpu hotplug
lock and to otherwise detangle the mess.
The new rules are:
1. you must do lock_cpu_hotplug() around the following functions:
__cpufreq_driver_target
__cpufreq_governor (for CPUFREQ_GOV_LIMITS operation only)
__cpufreq_set_policy
2. governer methods (.governer) must NOT take the lock_cpu_hotplug()
lock in any way; they are called with the lock taken already
3. if your governer spawns a thread that does things, like calling
__cpufreq_driver_target, your thread must honor rule #1.
4. the policy lock and other cpufreq internal locks nest within
the lock_cpu_hotplug() lock.
I'm not entirely happy about how the __cpufreq_governor rule ended up
(conditional locking rule depending on the argument) but basically all
callers pass this as a constant so it's not too horrible.
The patch also removes the cpufreq_governor() function since during the
locking audit it turned out to be entirely unused (so no need to fix it)
The patch works on my testbox, but it could use more testing
(otoh... it can't be much worse than the current code)
Signed-off-by: Arjan van de Ven <arjan@linux.intel.com>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Shutting down the ondemand policy was fraught with potential
problems, causing issues for SMP suspend (which wants to hot-
unplug) all but the last CPU.
This should fix at least the worst problems (divide-by-zero
and infinite wait for the workqueue to shut down).
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
[ There's some not quite baked bits in cpufreq-git right now
so sending this on as a patch instead ]
On Thu, 2006-07-06 at 07:58 -0700, Tom London wrote:
> After installing .2356 I get this each time I boot:
> =======================================================
> [ INFO: possible circular locking dependency detected ]
> -------------------------------------------------------
> S06cpuspeed/1620 is trying to acquire lock:
> (dbs_mutex){--..}, at: [<c060d6bb>] mutex_lock+0x21/0x24
>
> but task is already holding lock:
> (cpucontrol){--..}, at: [<c060d6bb>] mutex_lock+0x21/0x24
>
> which lock already depends on the new lock.
>
make sure the cpu hotplug recursive mutex (yuck) is taken early in the
cpufreq codepaths to avoid a AB-BA deadlock.
Signed-off-by: Arjan van de Ven <arjan@linux.intel.com>
Signed-off-by: Dave Jones <davej@redhat.com>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
* master.kernel.org:/pub/scm/linux/kernel/git/davej/cpufreq:
Move workqueue exports to where the functions are defined.
[CPUFREQ] Misc cleanups in ondemand.
[CPUFREQ] Make ondemand sampling per CPU and remove the mutex usage in sampling path.
[CPUFREQ] Add queue_delayed_work_on() interface for workqueues.
[CPUFREQ] Remove slowdown from ondemand sampling path.
Remove slowdown from ondemand sampling path. This reduces the code path length
in dbs_check_cpu() by half. slowdown was not used by ondemand by default.
If there are any user level tools that were using this tunable, they
may report error now.
Signed-off-by: Alexey Starikovskiy <alexey.y.starikovskiy@intel.com>
Signed-off-by: Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>
Signed-off-by: Dave Jones <davej@redhat.com>
Make notifier_blocks associated with cpu_notifier as __cpuinitdata.
__cpuinitdata makes sure that the data is init time only unless
CONFIG_HOTPLUG_CPU is defined.
Signed-off-by: Chandra Seetharaman <sekharan@us.ibm.com>
Cc: Ashok Raj <ashok.raj@intel.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
CPUs come online only at init time (unless CONFIG_HOTPLUG_CPU is defined).
So, cpu_notifier functionality need to be available only at init time.
This patch makes register_cpu_notifier() available only at init time, unless
CONFIG_HOTPLUG_CPU is defined.
This patch exports register_cpu_notifier() and unregister_cpu_notifier() only
if CONFIG_HOTPLUG_CPU is defined.
Signed-off-by: Chandra Seetharaman <sekharan@us.ibm.com>
Cc: Ashok Raj <ashok.raj@intel.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
In 2.6.17, there was a problem with cpu_notifiers and XFS. I provided a
band-aid solution to solve that problem. In the process, i undid all the
changes you both were making to ensure that these notifiers were available
only at init time (unless CONFIG_HOTPLUG_CPU is defined).
We deferred the real fix to 2.6.18. Here is a set of patches that fixes the
XFS problem cleanly and makes the cpu notifiers available only at init time
(unless CONFIG_HOTPLUG_CPU is defined).
If CONFIG_HOTPLUG_CPU is defined then cpu notifiers are available at run
time.
This patch reverts the notifier_call changes made in 2.6.17
Signed-off-by: Chandra Seetharaman <sekharan@us.ibm.com>
Cc: Ashok Raj <ashok.raj@intel.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
drivers/cpufreq/cpufreq_ondemand.c: In function 'do_dbs_timer':
drivers/cpufreq/cpufreq_ondemand.c:374: warning: implicit declaration of function 'lock_cpu_hotplug'
drivers/cpufreq/cpufreq_ondemand.c:381: warning: implicit declaration of function 'unlock_cpu_hotplug'
drivers/cpufreq/cpufreq_conservative.c: In function 'do_dbs_timer':
drivers/cpufreq/cpufreq_conservative.c:425: warning: implicit declaration of function 'lock_cpu_hotplug'
drivers/cpufreq/cpufreq_conservative.c:432: warning: implicit declaration of function 'unlock_cpu_hotplug'
Cc: Dave Jones <davej@codemonkey.org.uk>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Rootcaused the bug to a deadlock in cpufreq and ondemand. Due to non-existent
ordering between cpu_hotplug lock and dbs_mutex. Basically a race condition
between cpu_down() and do_dbs_timer().
cpu_down() flow:
* cpu_down() call for CPU 1
* Takes hot plug lock
* Calls pre down notifier
* cpufreq notifier handler calls cpufreq_driver_target() which takes
cpu_hotplug lock again. OK as cpu_hotplug lock is recursive in same
process context
* CPU 1 goes down
* Calls post down notifier
* cpufreq notifier handler calls ondemand event stop which takes dbs_mutex
So, cpu_hotplug lock is taken before dbs_mutex in this flow.
do_dbs_timer is triggerred by a periodic timer event.
It first takes dbs_mutex and then takes cpu_hotplug lock in
cpufreq_driver_target().
Note the reverse order here compared to above. So, if this timer event happens
at right moment during cpu_down, system will deadlok.
Attached patch fixes the issue for both ondemand and conservative.
Signed-off-by: Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>
Signed-off-by: Dave Jones <davej@redhat.com>
Remove KERN_* suffixes from some cpufreq driver's dprintk-s.
Signed-off-by: Jan Beulich <jbeulich@novell.com>
Signed-off-by: Dave Jones <davej@redhat.com>
