Whenever we are changing frequency of a cpu, we are calling PRECHANGE and
POSTCHANGE notifiers. They must be serialized. i.e. PRECHANGE or POSTCHANGE
shouldn't be called twice continuously. Following examples show why this is
important:
Scenario 1:
-----------
One thread reading value of cpuinfo_cur_freq, which will call
__cpufreq_cpu_get()->cpufreq_out_of_sync()->cpufreq_notify_transition()..
And ondemand governor trying to change freq of cpu at the same time and so
sending notification via ->target()..
Notifiers are not serialized and suppose this is what happened
- PRECHANGE Notification for freq A (from cpuinfo_cur_freq)
- PRECHANGE Notification for freq B (from target())
- Freq changed by target() to B
- POSTCHANGE Notification for freq B
- POSTCHANGE Notification for freq A
Now the last POSTCHANGE Notification happened for freq A and hardware is
currently running at freq B :)
Where would we break then?: adjust_jiffies() in cpufreq.c and cpufreq_callback()
in arch/arm/kernel/smp.c (which is also adjusting jiffies).. All loops_per_jiffy
based stuff is broken..
Scenario 2:
-----------
Governor is changing freq and has called __cpufreq_driver_target(). At the same
time we are changing scaling_{min|max}_freq from sysfs, which would eventually
end up calling governor's: CPUFREQ_GOV_LIMITS notification, that will also call:
__cpufreq_driver_target().. And hence concurrent calls to ->target()
And Platform have something like this in their ->target() (Like: cpufreq-cpu0,
omap, exynos, etc)
A. If new freq is more than old: Increase voltage
B. Change freq
C. If new freq is less than old: decrease voltage
Now, two concurrent calls to target are X and Y, where X is trying to increase
freq and Y is trying to decrease it..
And this is the sequence that followed due to races..
X.A: voltage increased for larger freq
Y.A: nothing happened here
Y.B: freq decreased
Y.C: voltage decreased
X.B: freq increased
X.C: nothing happened..
We ended up setting a freq which is not supported by the voltage we have set..
That will probably make clock to CPU unstable and system might not be workable
anymore...
This patch adds protection in cpufreq_notify_transition() to make transitions
serialized. It runs WARN() if POSTCHANGE notification is sent when we are not in
middle of a transition. And PRECHANGE notification is forced to wait while the
current transition is in progress.
Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org>
---
This was discussed earlier here:
https://lkml.org/lkml/2013/9/25/402
Where Rafael asked for better fix, as he called the V1 fixes: "quick and dirty".
This is another approach, much simpler than previous one. Please see if this
looks fine. There is a TODO note in there as I wanted some suggestions on how
exactly should we wait for a transition to get over.
drivers/cpufreq/cpufreq.c | 39 +++++++++++++++++++++++++++++++++++++--
include/linux/cpufreq.h | 2 ++
2 files changed, 39 insertions(+), 2 deletions(-)
@@ -104,6 +104,8 @@ struct cpufreq_policy {*__cpufreq_governor(data,CPUFREQ_GOV_POLICY_EXIT);*/structrw_semaphorerwsem;+booltransition_ongoing;/* Tracks transition status */+structmutextransition_lock;};/* Only for ACPI */
From: Srivatsa S. Bhat <hidden> Date: 2014-03-18 12:50:25
On 03/14/2014 01:13 PM, Viresh Kumar wrote:
Whenever we are changing frequency of a cpu, we are calling PRECHANGE and
POSTCHANGE notifiers. They must be serialized. i.e. PRECHANGE or POSTCHANGE
shouldn't be called twice continuously. Following examples show why this is
important:
[...]
quoted hunk
This was discussed earlier here:
https://lkml.org/lkml/2013/9/25/402
Where Rafael asked for better fix, as he called the V1 fixes: "quick and dirty".
This is another approach, much simpler than previous one. Please see if this
looks fine. There is a TODO note in there as I wanted some suggestions on how
exactly should we wait for a transition to get over.
drivers/cpufreq/cpufreq.c | 39 +++++++++++++++++++++++++++++++++++++--
include/linux/cpufreq.h | 2 ++
2 files changed, 39 insertions(+), 2 deletions(-)
Nope, this is definitely not the way to go, IMHO. We should enforce that
the *callers* serialize the transitions, something like this:
cpufreq_transition_lock();
cpufreq_notify_transition(CPUFREQ_PRECHANGE);
//Perform the frequency change
cpufreq_notify_transition(CPUFREQ_POSTCHANGE);
cpufreq_transition_unlock();
That's it!
[ We can either introduce a new "transition" lock or perhaps even reuse
the cpufreq_driver_lock if it fits... but the point is, the _caller_ has
to perform the locking; trying to be smart inside cpufreq_notify_transition()
is a recipe for headache :-( ]
Is there any problem with this approach due to which you didn't take
this route?
+ if (unlikely(WARN_ON(!policy->transition_ongoing &&
+ (state == CPUFREQ_POSTCHANGE)))) {
+ mutex_unlock(&policy->transition_lock);
+ return;
+ }
+
+ if (state == CPUFREQ_PRECHANGE) {
+ while (policy->transition_ongoing) {
+ mutex_unlock(&policy->transition_lock);
+ /* TODO: Can we do something better here? */
+ cpu_relax();
+ mutex_lock(&policy->transition_lock);
If the caller takes care of the synchronization, we can avoid
these sorts of acrobatics ;-)
Regards,
Srivatsa S. Bhat
@@ -104,6 +104,8 @@ struct cpufreq_policy {*__cpufreq_governor(data,CPUFREQ_GOV_POLICY_EXIT);*/structrw_semaphorerwsem;+booltransition_ongoing;/* Tracks transition status */+structmutextransition_lock;};/* Only for ACPI */
On 18 March 2014 18:20, Srivatsa S. Bhat
[off-list ref] wrote:
On 03/14/2014 01:13 PM, Viresh Kumar wrote:
quoted
+ if ((state != CPUFREQ_PRECHANGE) && (state != CPUFREQ_POSTCHANGE))
Wait a min, when is this condition ever true? I mean, what else can
'state' ever be, apart from CPUFREQ_PRECHANGE and POSTCHANGE?
There were two more 'unused' states available:
CPUFREQ_RESUMECHANGE and CPUFREQ_SUSPENDCHANGE
I have sent a patch to remove them now and this code would go away..
Nope, this is definitely not the way to go, IMHO. We should enforce that
the *callers* serialize the transitions, something like this:
cpufreq_transition_lock();
cpufreq_notify_transition(CPUFREQ_PRECHANGE);
//Perform the frequency change
cpufreq_notify_transition(CPUFREQ_POSTCHANGE);
cpufreq_transition_unlock();
That's it!
[ We can either introduce a new "transition" lock or perhaps even reuse
the cpufreq_driver_lock if it fits... but the point is, the _caller_ has
to perform the locking; trying to be smart inside cpufreq_notify_transition()
is a recipe for headache :-( ]
Is there any problem with this approach due to which you didn't take
this route?
I didn't wanted drivers to handle this as core must make sure things are in
order. Over that it would have helped by not pasting redundant code
everywhere..
Drivers are anyway going to call cpufreq_notify_transition(), why increase
burden on them?
quoted
+ if (unlikely(WARN_ON(!policy->transition_ongoing &&
+ (state == CPUFREQ_POSTCHANGE)))) {
+ mutex_unlock(&policy->transition_lock);
+ return;
+ }
+
+ if (state == CPUFREQ_PRECHANGE) {
+ while (policy->transition_ongoing) {
+ mutex_unlock(&policy->transition_lock);
+ /* TODO: Can we do something better here? */
+ cpu_relax();
+ mutex_lock(&policy->transition_lock);
If the caller takes care of the synchronization, we can avoid
these sorts of acrobatics ;-)
If we are fine with taking a mutex for the entire transition, then
we can avoid above kind of acrobatics by just taking the mutex
from PRECHANGE and leaving it at POSTCHANGE..
It will look like this then, hope this looks fine :)
From: Srivatsa S. Bhat <hidden> Date: 2014-03-19 09:17:39
On 03/19/2014 11:38 AM, Viresh Kumar wrote:
On 18 March 2014 18:20, Srivatsa S. Bhat
[off-list ref] wrote:
quoted
On 03/14/2014 01:13 PM, Viresh Kumar wrote:
quoted
+ if ((state != CPUFREQ_PRECHANGE) && (state != CPUFREQ_POSTCHANGE))
Wait a min, when is this condition ever true? I mean, what else can
'state' ever be, apart from CPUFREQ_PRECHANGE and POSTCHANGE?
There were two more 'unused' states available:
CPUFREQ_RESUMECHANGE and CPUFREQ_SUSPENDCHANGE
I have sent a patch to remove them now and this code would go away..
Nope, this is definitely not the way to go, IMHO. We should enforce that
the *callers* serialize the transitions, something like this:
cpufreq_transition_lock();
cpufreq_notify_transition(CPUFREQ_PRECHANGE);
//Perform the frequency change
cpufreq_notify_transition(CPUFREQ_POSTCHANGE);
cpufreq_transition_unlock();
That's it!
[ We can either introduce a new "transition" lock or perhaps even reuse
the cpufreq_driver_lock if it fits... but the point is, the _caller_ has
to perform the locking; trying to be smart inside cpufreq_notify_transition()
is a recipe for headache :-( ]
Is there any problem with this approach due to which you didn't take
this route?
Wait, I think I remember. The problem was about dealing with drivers that
do asynchronous notification (those that have the ASYNC_NOTIFICATION flag
set). In particular, exynos-5440 driver sends out the POSTCHANGE notification
from a workqueue worker, much later than sending the PRECHANGE notification.
From what I saw, this is how the exynos-5440 driver works:
1. ->target() is invoked, and the driver writes to a register and returns
to its caller.
2. An interrupt occurs that indicates that the frequency was changed.
3. The interrupt handler kicks off a worker thread which then sends out
the POSTCHANGE notification.
So the important question here is, how does the exynos-5440 driver
protect itself from say 2 ->target() calls which occur in close sequence
(before allowing the entire chain for the first call to complete)?
As far as I can see there is no such synchronization in the driver at
the moment. Adding Amit to CC for his comments.
Regards,
Srivatsa S. Bhat
quoted hunk
I didn't wanted drivers to handle this as core must make sure things are in
order. Over that it would have helped by not pasting redundant code
everywhere..
Drivers are anyway going to call cpufreq_notify_transition(), why increase
burden on them?
quoted
quoted
+ if (unlikely(WARN_ON(!policy->transition_ongoing &&
+ (state == CPUFREQ_POSTCHANGE)))) {
+ mutex_unlock(&policy->transition_lock);
+ return;
+ }
+
+ if (state == CPUFREQ_PRECHANGE) {
+ while (policy->transition_ongoing) {
+ mutex_unlock(&policy->transition_lock);
+ /* TODO: Can we do something better here? */
+ cpu_relax();
+ mutex_lock(&policy->transition_lock);
If the caller takes care of the synchronization, we can avoid
these sorts of acrobatics ;-)
If we are fine with taking a mutex for the entire transition, then
we can avoid above kind of acrobatics by just taking the mutex
from PRECHANGE and leaving it at POSTCHANGE..
It will look like this then, hope this looks fine :)
On 19 March 2014 14:47, Srivatsa S. Bhat
[off-list ref] wrote:
Wait, I think I remember. The problem was about dealing with drivers that
do asynchronous notification (those that have the ASYNC_NOTIFICATION flag
set). In particular, exynos-5440 driver sends out the POSTCHANGE notification
from a workqueue worker, much later than sending the PRECHANGE notification.
From what I saw, this is how the exynos-5440 driver works:
1. ->target() is invoked, and the driver writes to a register and returns
to its caller.
2. An interrupt occurs that indicates that the frequency was changed.
3. The interrupt handler kicks off a worker thread which then sends out
the POSTCHANGE notification.
Correct!!
So the important question here is, how does the exynos-5440 driver
protect itself from say 2 ->target() calls which occur in close sequence
(before allowing the entire chain for the first call to complete)?
As far as I can see there is no such synchronization in the driver at
the moment. Adding Amit to CC for his comments.
Yes, and that's what my patch is trying to fix. Where is the confusion?
From: Srivatsa S. Bhat <hidden> Date: 2014-03-19 09:51:48
On 03/19/2014 02:50 PM, Viresh Kumar wrote:
On 19 March 2014 14:47, Srivatsa S. Bhat
[off-list ref] wrote:
quoted
Wait, I think I remember. The problem was about dealing with drivers that
do asynchronous notification (those that have the ASYNC_NOTIFICATION flag
set). In particular, exynos-5440 driver sends out the POSTCHANGE notification
from a workqueue worker, much later than sending the PRECHANGE notification.
From what I saw, this is how the exynos-5440 driver works:
1. ->target() is invoked, and the driver writes to a register and returns
to its caller.
2. An interrupt occurs that indicates that the frequency was changed.
3. The interrupt handler kicks off a worker thread which then sends out
the POSTCHANGE notification.
Correct!!
quoted
So the important question here is, how does the exynos-5440 driver
protect itself from say 2 ->target() calls which occur in close sequence
(before allowing the entire chain for the first call to complete)?
As far as I can see there is no such synchronization in the driver at
the moment. Adding Amit to CC for his comments.
Yes, and that's what my patch is trying to fix. Where is the confusion?
Sorry, for a moment I got confused and thought that your patch addresses
the race conditions present in normal drivers alone, and not ASYNC_NOTIFICATION
drivers. But now I understand that your patch intends to fix both the
problems at once. I'll share my thoughts about the design in a separate
reply.
Regards,
Srivatsa S. Bhat
From: Srivatsa S. Bhat <hidden> Date: 2014-03-19 09:51:57
On 03/19/2014 11:38 AM, Viresh Kumar wrote:
On 18 March 2014 18:20, Srivatsa S. Bhat
[off-list ref] wrote:
quoted
On 03/14/2014 01:13 PM, Viresh Kumar wrote:
quoted
+ if ((state != CPUFREQ_PRECHANGE) && (state != CPUFREQ_POSTCHANGE))
Wait a min, when is this condition ever true? I mean, what else can
'state' ever be, apart from CPUFREQ_PRECHANGE and POSTCHANGE?
There were two more 'unused' states available:
CPUFREQ_RESUMECHANGE and CPUFREQ_SUSPENDCHANGE
I have sent a patch to remove them now and this code would go away..
Nope, this is definitely not the way to go, IMHO. We should enforce that
the *callers* serialize the transitions, something like this:
cpufreq_transition_lock();
cpufreq_notify_transition(CPUFREQ_PRECHANGE);
//Perform the frequency change
cpufreq_notify_transition(CPUFREQ_POSTCHANGE);
cpufreq_transition_unlock();
That's it!
[ We can either introduce a new "transition" lock or perhaps even reuse
the cpufreq_driver_lock if it fits... but the point is, the _caller_ has
to perform the locking; trying to be smart inside cpufreq_notify_transition()
is a recipe for headache :-( ]
Is there any problem with this approach due to which you didn't take
this route?
I didn't wanted drivers to handle this as core must make sure things are in
order. Over that it would have helped by not pasting redundant code
everywhere..
Drivers are anyway going to call cpufreq_notify_transition(), why increase
burden on them?
No, its not about burden. Its about the elegance of the design. We should
not be overly "smart" in the cpufreq core. Hiding the synchronization inside
the cpufreq core only encourages people to write buggy code in their drivers.
Why don't we go with what Rafael suggested? We can have dedicated
begin_transition() and end_transition() calls to demarcate the frequency
transitions. That way, it makes it very clear how the synchronization is
done. Of course, these functions would be provided (exported) by the cpufreq
core, by implementing them using locks/counters/whatever.
Basically what I'm arguing against, is the idea of having the cpufreq
core figure out what the driver _intended_ to do, from inside the
cpufreq_notify_transition() call.
What I would prefer instead is to have the cpufreq driver do something
like this:
cpufreq_freq_transition_begin();
cpufreq_notify_transition(CPUFREQ_PRECHANGE);
//perform the frequency change
cpufreq_notify_transition(CPUFREQ_POSTCHANGE);
cpufreq_freq_transition_end();
[ASYNC_NOTIFICATION drivers will invoke the last two functions in a
separate context/thread.]
Regards,
Srivatsa S. Bhat
quoted hunk
quoted
quoted
+ if (unlikely(WARN_ON(!policy->transition_ongoing &&
+ (state == CPUFREQ_POSTCHANGE)))) {
+ mutex_unlock(&policy->transition_lock);
+ return;
+ }
+
+ if (state == CPUFREQ_PRECHANGE) {
+ while (policy->transition_ongoing) {
+ mutex_unlock(&policy->transition_lock);
+ /* TODO: Can we do something better here? */
+ cpu_relax();
+ mutex_lock(&policy->transition_lock);
If the caller takes care of the synchronization, we can avoid
these sorts of acrobatics ;-)
If we are fine with taking a mutex for the entire transition, then
we can avoid above kind of acrobatics by just taking the mutex
from PRECHANGE and leaving it at POSTCHANGE..
It will look like this then, hope this looks fine :)
On 19 March 2014 15:20, Srivatsa S. Bhat
[off-list ref] wrote:
No, its not about burden. Its about the elegance of the design. We should
not be overly "smart" in the cpufreq core. Hiding the synchronization inside
the cpufreq core only encourages people to write buggy code in their drivers.
What kind of buggy code can be there? They are supposed to call notifiers
in the order mentioned and so it shouldn't be a problem at all.. Don't know..
Why don't we go with what Rafael suggested? We can have dedicated
begin_transition() and end_transition() calls to demarcate the frequency
transitions. That way, it makes it very clear how the synchronization is
done. Of course, these functions would be provided (exported) by the cpufreq
core, by implementing them using locks/counters/whatever.
Basically what I'm arguing against, is the idea of having the cpufreq
core figure out what the driver _intended_ to do, from inside the
cpufreq_notify_transition() call.
What I would prefer instead is to have the cpufreq driver do something
like this:
cpufreq_freq_transition_begin();
cpufreq_notify_transition(CPUFREQ_PRECHANGE);
Why do we need two routines then? What about doing notification from
inside cpufreq_freq_transition_begin()?
This is a burden for driver writers, who don't normally understand the
relevance of these calls in detail and may think, only the first one is
enough or the second one is..
Its better if they simply let the core that they are starting to do transitions,
i.e. cpufreq_freq_transition_begin() and then the core should send
notifications.
//perform the frequency change
cpufreq_notify_transition(CPUFREQ_POSTCHANGE);
cpufreq_freq_transition_end();
[ASYNC_NOTIFICATION drivers will invoke the last two functions in a
separate context/thread.]
Same for the last two routines and yes they would be called from
separate thread for ASYNC_NOTIFICATION drivers..
From: Srivatsa S. Bhat <hidden> Date: 2014-03-19 12:16:19
On 03/19/2014 03:39 PM, Viresh Kumar wrote:
On 19 March 2014 15:20, Srivatsa S. Bhat
[off-list ref] wrote:
quoted
No, its not about burden. Its about the elegance of the design. We should
not be overly "smart" in the cpufreq core. Hiding the synchronization inside
the cpufreq core only encourages people to write buggy code in their drivers.
What kind of buggy code can be there? They are supposed to call notifiers
in the order mentioned and so it shouldn't be a problem at all.. Don't know..
quoted
Why don't we go with what Rafael suggested? We can have dedicated
begin_transition() and end_transition() calls to demarcate the frequency
transitions. That way, it makes it very clear how the synchronization is
done. Of course, these functions would be provided (exported) by the cpufreq
core, by implementing them using locks/counters/whatever.
Basically what I'm arguing against, is the idea of having the cpufreq
core figure out what the driver _intended_ to do, from inside the
cpufreq_notify_transition() call.
What I would prefer instead is to have the cpufreq driver do something
like this:
cpufreq_freq_transition_begin();
cpufreq_notify_transition(CPUFREQ_PRECHANGE);
Why do we need two routines then? What about doing notification from
inside cpufreq_freq_transition_begin()?
Hmm, that's a good idea. I thought about ways to simplify the synchronization
and this is what I came up with. Its completely untested though. Let me know
what you think!
---------------------------------------------------------------------------
From: Srivatsa S. Bhat <redacted>
Subject: [PATCH v3] cpufreq: Make sure frequency transitions are serialized
Whenever we change the frequency of a CPU, we call the PRECHANGE and
POSTCHANGE notifiers. They must be serialized, i.e. PRECHANGE and
POSTCHANGE notifiers should strictly alternate, thereby preventing
two different sets of PRECHANGE or POSTCHANGE notifiers from
interleaving arbitrarily.
The following examples illustrate why this is important:
Scenario 1:
-----------
A thread reading the value of cpuinfo_cur_freq, will call
__cpufreq_cpu_get()->cpufreq_out_of_sync()->cpufreq_notify_transition()
The ondemand governor can decide to change the frequency of the CPU at the
same time and hence it can end up sending the notifications via ->target().
If the notifiers are not serialized, the following sequence can occur:
- PRECHANGE Notification for freq A (from cpuinfo_cur_freq)
- PRECHANGE Notification for freq B (from target())
- Freq changed by target() to B
- POSTCHANGE Notification for freq B
- POSTCHANGE Notification for freq A
We can see from the above that the last POSTCHANGE Notification happens for
freq A but the hardware is set to run at freq B.
Where would we break then?: adjust_jiffies() in cpufreq.c & cpufreq_callback()
in arch/arm/kernel/smp.c (which also adjusts the jiffies). All the
loops_per_jiffy calculations will get messed up.
Scenario 2:
-----------
The governor calls __cpufreq_driver_target() to change the frequency. At the
same time, if we change scaling_{min|max}_freq from sysfs, it will end up
calling the governor's CPUFREQ_GOV_LIMITS notification, which will also call
__cpufreq_driver_target(). And hence we end up issuing concurrent calls to
->target().
Typically, platforms have the following logic in their ->target() routines:
(Eg: cpufreq-cpu0, omap, exynos, etc)
A. If new freq is more than old: Increase voltage
B. Change freq
C. If new freq is less than old: decrease voltage
Now, if the two concurrent calls to ->target() are X and Y, where X is trying
to increase the freq and Y is trying to decrease it, we get the following
race condition:
X.A: voltage gets increased for larger freq
Y.A: nothing happens
Y.B: freq gets decreased
Y.C: voltage gets decreased
X.B: freq gets increased
X.C: nothing happens
Thus we can end up setting a freq which is not supported by the voltage we
have set. That will probably make the clock to the CPU unstable and the system
might not work properly anymore.
This patch introduces a set of synchronization primitives to serialize
frequency transitions, which are to be used as shown below:
cpufreq_freq_transition_begin();
//Perform the frequency change
cpufreq_freq_transition_end();
The _begin() call sends the PRECHANGE notification whereas the _end() call
sends the POSTCHANGE notification. Also, all the necessary synchronization
is handled within these calls. In particular, even drivers which set the
ASYNC_NOTIFICATION flag can also use these APIs for performing frequency
transitions (ie., you can call _begin() from one task, and call the
corresponding _end() from a different task).
The actual synchronization underneath is not that complicated:
The key challenge is to allow drivers to begin the transition from one thread
and end it in a completely different thread (this is to enable drivers that do
asynchronous POSTCHANGE notification from bottom-halves, to also use the same
interface).
To achieve this, a 'transition_ongoing' flag, a 'transition_lock' mutex and a
wait-queue are added per-policy. The flag and the wait-queue are used in
conjunction to create an "uninterrupted flow" from _begin() to _end(). The
mutex-lock is used to ensure that only one such "flow" is in flight at any
given time. Put together, this provides us all the necessary synchronization.
Based-on-patch-by: Viresh Kumar [off-list ref]
Signed-off-by: Srivatsa S. Bhat <redacted>
---
drivers/cpufreq/cpufreq.c | 34 ++++++++++++++++++++++++++++++++++
include/linux/cpufreq.h | 5 +++++
2 files changed, 39 insertions(+)
@@ -101,6 +101,11 @@ struct cpufreq_policy {*__cpufreq_governor(data,CPUFREQ_GOV_POLICY_EXIT);*/structrw_semaphorerwsem;++/* Synchronization for frequency transitions */+booltransition_ongoing;/* Tracks transition status */+structmutextransition_lock;+wait_queue_head_ttransition_wait;};/* Only for ACPI */
Similar to what I have done in my last version, why do you need
transition_ongoing and transition_wait? Simply work with
transition_lock? i.e. Acquire it for the complete transition sequence.
Similar to what I have done in my last version, why do you need
transition_ongoing and transition_wait? Simply work with
transition_lock? i.e. Acquire it for the complete transition sequence.
We *can't* acquire it for the complete transition sequence
in case of drivers that do asynchronous notification, because
PRECHANGE is done in one thread and POSTCHANGE is done in a
totally different thread! You can't acquire a lock in one
task and release it in a different task. That would be a
fundamental violation of locking.
That's why I introduced the wait queue to help us create
a "flow" which encompasses 2 different, but co-ordinating
tasks. You simply can't do that elegantly by using plain
locks alone.
Regards,
Srivatsa S. Bhat
Similar to what I have done in my last version, why do you need
transition_ongoing and transition_wait? Simply work with
transition_lock? i.e. Acquire it for the complete transition sequence.
We *can't* acquire it for the complete transition sequence
in case of drivers that do asynchronous notification, because
PRECHANGE is done in one thread and POSTCHANGE is done in a
totally different thread! You can't acquire a lock in one
task and release it in a different task. That would be a
fundamental violation of locking.
That's why I introduced the wait queue to help us create
a "flow" which encompasses 2 different, but co-ordinating
tasks. You simply can't do that elegantly by using plain
locks alone.
By the way, note the updated changelog in my patch. It includes a brief
overview of the synchronization design, which is copy-pasted below for
reference. I forgot to mention this earlier!
-----
This patch introduces a set of synchronization primitives to serialize
frequency transitions, which are to be used as shown below:
cpufreq_freq_transition_begin();
//Perform the frequency change
cpufreq_freq_transition_end();
The _begin() call sends the PRECHANGE notification whereas the _end() call
sends the POSTCHANGE notification. Also, all the necessary synchronization
is handled within these calls. In particular, even drivers which set the
ASYNC_NOTIFICATION flag can also use these APIs for performing frequency
transitions (ie., you can call _begin() from one task, and call the
corresponding _end() from a different task).
The actual synchronization underneath is not that complicated:
The key challenge is to allow drivers to begin the transition from one thread
and end it in a completely different thread (this is to enable drivers that do
asynchronous POSTCHANGE notification from bottom-halves, to also use the same
interface).
To achieve this, a 'transition_ongoing' flag, a 'transition_lock' mutex and a
wait-queue are added per-policy. The flag and the wait-queue are used in
conjunction to create an "uninterrupted flow" from _begin() to _end(). The
mutex-lock is used to ensure that only one such "flow" is in flight at any
given time. Put together, this provides us all the necessary synchronization.
Regards,
Srivatsa S. Bhat
I think its broken here. At this point another thread can come take lock,
update transition_ongoing, send notification and finally unlock..
And after that we can take the lock and send another notification..
Correct?
+ if (!mutex_trylock(&policy->transition_lock))
+ goto wait;
+
+ policy->transition_ongoing++;
I think its broken here. At this point another thread can come take lock,
update transition_ongoing, send notification and finally unlock..
And after that we can take the lock and send another notification..
Correct?
Good catch! I missed that yesterday. Please find the updated patch below,
with all your suggestions incorporated. Does this version look any better?
------------------------------------------------------------------------
@@ -101,6 +101,11 @@ struct cpufreq_policy {*__cpufreq_governor(data,CPUFREQ_GOV_POLICY_EXIT);*/structrw_semaphorerwsem;++/* Synchronization for frequency transitions */+booltransition_ongoing;/* Tracks transition status */+structmutextransition_lock;+wait_queue_head_ttransition_wait;};/* Only for ACPI */
Why do we need locking here? You explained that earlier :)
Hmm.. I had thought of some complex race condition which would make
tasks miss the wake-up event and sleep forever, and hence added
the locking there to prevent that. But now that I think more closely,
I'm not able to recall that race... I will give some more thought to
it and if I can't find any loopholes in doing the second update to
the ongoing flag without locks, then I'll post the patchset with
that lockless version itself.
Also, I would like to add this here:
WARN_ON(policy->transition_ongoing);
Hmm? Won't it always be true? We are the ones who set that flag to
true earlier, right? I guess you meant WARN_ON(!policy->transition_ongoing)
perhaps? I'm not sure whether its really worth it, because it kinda looks
obvious. Not sure what kind of bugs it would catch. I can't think of any
such scenario :-(
On 20 March 2014 14:54, Srivatsa S. Bhat
[off-list ref] wrote:
On 03/20/2014 02:07 PM, Viresh Kumar wrote:
quoted
WARN_ON(policy->transition_ongoing);
I guess you meant WARN_ON(!policy->transition_ongoing)
perhaps?
Ooops!!
I'm not sure whether its really worth it, because it kinda looks
obvious. Not sure what kind of bugs it would catch. I can't think of any
such scenario :-(
Just to catch if somebody is sending a POSTCHANGE one without first
sending a PRECHANGE one.. Just another check to make sure things are
in order.
From: Srivatsa S. Bhat <hidden> Date: 2014-03-20 09:46:21
On 03/20/2014 03:03 PM, Viresh Kumar wrote:
On 20 March 2014 14:54, Srivatsa S. Bhat
[off-list ref] wrote:
quoted
On 03/20/2014 02:07 PM, Viresh Kumar wrote:
quoted
WARN_ON(policy->transition_ongoing);
I guess you meant WARN_ON(!policy->transition_ongoing)
perhaps?
Ooops!!
quoted
I'm not sure whether its really worth it, because it kinda looks
obvious. Not sure what kind of bugs it would catch. I can't think of any
such scenario :-(
Just to catch if somebody is sending a POSTCHANGE one without first
sending a PRECHANGE one.. Just another check to make sure things are
in order.
Well, that's unlikely, since they will have to call _end() before
_begin() :-) That's the power of having great function names - they make
it impossible to use them incorrectly ;-) But anyway, I can add the check,
just in case somebody misses even such an obvious cue! :-)
By the way, I'm also thinking of using a spinlock instead of a mutex.
The critical section is tiny and we don't sleep inside the critical
section - sounds like the perfect case for a spinlock.
Regards,
Srivatsa S. Bhat
On 20 March 2014 15:15, Srivatsa S. Bhat
[off-list ref] wrote:
By the way, I'm also thinking of using a spinlock instead of a mutex.
The critical section is tiny and we don't sleep inside the critical
section - sounds like the perfect case for a spinlock.
From: Rafael J. Wysocki <hidden> Date: 2014-03-19 13:38:32
On Wednesday, March 19, 2014 03:39:16 PM Viresh Kumar wrote:
On 19 March 2014 15:20, Srivatsa S. Bhat
[off-list ref] wrote:
quoted
No, its not about burden. Its about the elegance of the design. We should
not be overly "smart" in the cpufreq core. Hiding the synchronization inside
the cpufreq core only encourages people to write buggy code in their drivers.
What kind of buggy code can be there? They are supposed to call notifiers
in the order mentioned and so it shouldn't be a problem at all.. Don't know..
quoted
Why don't we go with what Rafael suggested? We can have dedicated
begin_transition() and end_transition() calls to demarcate the frequency
transitions. That way, it makes it very clear how the synchronization is
done. Of course, these functions would be provided (exported) by the cpufreq
core, by implementing them using locks/counters/whatever.
Basically what I'm arguing against, is the idea of having the cpufreq
core figure out what the driver _intended_ to do, from inside the
cpufreq_notify_transition() call.
What I would prefer instead is to have the cpufreq driver do something
like this:
cpufreq_freq_transition_begin();
cpufreq_notify_transition(CPUFREQ_PRECHANGE);
Why do we need two routines then? What about doing notification from
inside cpufreq_freq_transition_begin()?
We can do that in my opinion.
This is a burden for driver writers, who don't normally understand the
relevance of these calls in detail and may think, only the first one is
enough or the second one is..
Its better if they simply let the core that they are starting to do transitions,
i.e. cpufreq_freq_transition_begin() and then the core should send
notifications.
quoted
//perform the frequency change
cpufreq_notify_transition(CPUFREQ_POSTCHANGE);
cpufreq_freq_transition_end();
[ASYNC_NOTIFICATION drivers will invoke the last two functions in a
separate context/thread.]
Same for the last two routines and yes they would be called from
separate thread for ASYNC_NOTIFICATION drivers..
That'd be fine by me in principle.
--
I speak only for myself.
Rafael J. Wysocki, Intel Open Source Technology Center.