[v10,3/3] soc: qcom: rpmh: Invoke rpmh_flush() for dirty caches

Commit Message

Maulik Shah March 3, 2020, 12:26 p.m. UTC

Add changes to invoke rpmh flush() from within cache_lock when the data
in cache is dirty.

This is done only if OSI is not supported in PSCI. If OSI is supported
rpmh_flush can get invoked when the last cpu going to power collapse
deepest low power mode.

Also remove "depends on COMPILE_TEST" for Kconfig option QCOM_RPMH so the
driver is only compiled for arm64 which supports psci_has_osi_support()
API.

Signed-off-by: Maulik Shah <[e-mail]>
Reviewed-by: Srinivas Rao L <[e-mail]>
---
 drivers/soc/qcom/Kconfig |  2 +-
 drivers/soc/qcom/rpmh.c  | 37 ++++++++++++++++++++++---------------
 2 files changed, 23 insertions(+), 16 deletions(-)

Comments

Doug Anderson March 4, 2020, 11:22 p.m. UTC | #1

Hi,

On Tue, Mar 3, 2020 at 4:27 AM Maulik Shah <[e-mail]> wrote:
>
> Add changes to invoke rpmh flush() from within cache_lock when the data
> in cache is dirty.
>
> This is done only if OSI is not supported in PSCI. If OSI is supported
> rpmh_flush can get invoked when the last cpu going to power collapse
> deepest low power mode.
>
> Also remove "depends on COMPILE_TEST" for Kconfig option QCOM_RPMH so the
> driver is only compiled for arm64 which supports psci_has_osi_support()
> API.
>
> Signed-off-by: Maulik Shah <[e-mail]>
> Reviewed-by: Srinivas Rao L <[e-mail]>
> ---
>  drivers/soc/qcom/Kconfig |  2 +-
>  drivers/soc/qcom/rpmh.c  | 37 ++++++++++++++++++++++---------------
>  2 files changed, 23 insertions(+), 16 deletions(-)
>
> diff --git a/drivers/soc/qcom/Kconfig b/drivers/soc/qcom/Kconfig
> index d0a73e7..2e581bc 100644
> --- a/drivers/soc/qcom/Kconfig
> +++ b/drivers/soc/qcom/Kconfig
> @@ -105,7 +105,7 @@ config QCOM_RMTFS_MEM
>
>  config QCOM_RPMH
>         bool "Qualcomm RPM-Hardened (RPMH) Communication"
> -       depends on ARCH_QCOM && ARM64 || COMPILE_TEST
> +       depends on ARCH_QCOM && ARM64
>         help
>           Support for communication with the hardened-RPM blocks in
>           Qualcomm Technologies Inc (QTI) SoCs. RPMH communication uses an
> diff --git a/drivers/soc/qcom/rpmh.c b/drivers/soc/qcom/rpmh.c
> index f28afe4..dafb0da 100644
> --- a/drivers/soc/qcom/rpmh.c
> +++ b/drivers/soc/qcom/rpmh.c
> @@ -12,6 +12,7 @@
>  #include <linux/module.h>
>  #include <linux/of.h>
>  #include <linux/platform_device.h>
> +#include <linux/psci.h>
>  #include <linux/slab.h>
>  #include <linux/spinlock.h>
>  #include <linux/types.h>
> @@ -158,6 +159,13 @@ static struct cache_req *cache_rpm_request(struct rpmh_ctrlr *ctrlr,
>         }
>
>  unlock:
> +       if (ctrlr->dirty && !psci_has_osi_support()) {
> +               if (rpmh_flush(ctrlr)) {
> +                       spin_unlock_irqrestore(&ctrlr->cache_lock, flags);
> +                       return ERR_PTR(-EINVAL);
> +               }
> +       }
> +
>         spin_unlock_irqrestore(&ctrlr->cache_lock, flags);
>
>         return req;
> @@ -285,26 +293,35 @@ int rpmh_write(const struct device *dev, enum rpmh_state state,
>  }
>  EXPORT_SYMBOL(rpmh_write);
>
> -static void cache_batch(struct rpmh_ctrlr *ctrlr, struct batch_cache_req *req)
> +static int cache_batch(struct rpmh_ctrlr *ctrlr, struct batch_cache_req *req)
>  {
>         unsigned long flags;
>
>         spin_lock_irqsave(&ctrlr->cache_lock, flags);
> +
>         list_add_tail(&req->list, &ctrlr->batch_cache);
>         ctrlr->dirty = true;
> +
> +       if (!psci_has_osi_support()) {
> +               if (rpmh_flush(ctrlr)) {

The whole API here is a bit unfortunate.  From what I can tell,
callers of this code almost always call rpmh_write_batch() in
triplicate, AKA:

rpmh_write_batch(active, ...)
rpmh_write_batch(wake, ...)
rpmh_write_batch(sleep, ...)

...that's going to end up writing the whole sleep/wake sets twice
every single time, right?  I know you talked about trying to keep
separate dirty bits for sleep/wake and maybe that would help, but it
might not be so easy due to the comparison of "sleep_val" and
"wake_val" in is_req_valid().

I guess we can keep the inefficiency for now and see how much it hits
us, but it feels ugly.


> +                       spin_unlock_irqrestore(&ctrlr->cache_lock, flags);
> +                       return -EINVAL;

nit: why not add "int ret = 0" to the top of the function, then here:

if (rpmh_flush(ctrl))
  ret = -EINVAL;

...then at the end "return ret".  It avoids the 2nd copy of the unlock?

Also: Why throw away the return value of rpmh_flush and replace it
with -EINVAL?  Trying to avoid -EBUSY?  ...oh, should you handle
-EBUSY?  AKA:

if (!psci_has_osi_support()) {
  do {
    ret = rpmh_flush(ctrl);
  } while (ret == -EBUSY);
}


> +               }
> +       }
> +
>         spin_unlock_irqrestore(&ctrlr->cache_lock, flags);
> +
> +       return 0;
>  }
>
>  static int flush_batch(struct rpmh_ctrlr *ctrlr)
>  {
>         struct batch_cache_req *req;
>         const struct rpmh_request *rpm_msg;
> -       unsigned long flags;
>         int ret = 0;
>         int i;
>
>         /* Send Sleep/Wake requests to the controller, expect no response */
> -       spin_lock_irqsave(&ctrlr->cache_lock, flags);
>         list_for_each_entry(req, &ctrlr->batch_cache, list) {
>                 for (i = 0; i < req->count; i++) {
>                         rpm_msg = req->rpm_msgs + i;
> @@ -314,7 +331,6 @@ static int flush_batch(struct rpmh_ctrlr *ctrlr)
>                                 break;
>                 }
>         }
> -       spin_unlock_irqrestore(&ctrlr->cache_lock, flags);
>
>         return ret;
>  }
> @@ -386,10 +402,8 @@ int rpmh_write_batch(const struct device *dev, enum rpmh_state state,
>                 cmd += n[i];
>         }
>
> -       if (state != RPMH_ACTIVE_ONLY_STATE) {
> -               cache_batch(ctrlr, req);
> -               return 0;
> -       }
> +       if (state != RPMH_ACTIVE_ONLY_STATE)
> +               return cache_batch(ctrlr, req);
>
>         for (i = 0; i < count; i++) {
>                 struct completion *compl = &compls[i];
> @@ -455,9 +469,6 @@ static int send_single(struct rpmh_ctrlr *ctrlr, enum rpmh_state state,
>   * Return: -EBUSY if the controller is busy, probably waiting on a response
>   * to a RPMH request sent earlier.
>   *
> - * This function is always called from the sleep code from the last CPU
> - * that is powering down the entire system. Since no other RPMH API would be
> - * executing at this time, it is safe to run lockless.

nit: you've now got an extra "blank" (just has a "*" on it) line at
the end of your comment block.

nit: in v9, Evan suggested "We should probably replace that with a
comment indicating that we assume ctrlr->cache_lock is already held".
Maybe you could do that?

Also: presumably you _will_ still be called by the sleep code from the
last CPU on systems with OSI.  Is that true?  If that's not true then
you should change your function to static.  If that is true, then your
comment should be something like "this function will either be called
from sleep code on the last CPU (thus no spinlock needed) or with the
spinlock already held".


-Doug

Maulik Shah March 5, 2020, 11:30 a.m. UTC | #2

On 3/5/2020 4:52 AM, Doug Anderson wrote:
> Hi,
>
> On Tue, Mar 3, 2020 at 4:27 AM Maulik Shah <[e-mail]> wrote:
>> Add changes to invoke rpmh flush() from within cache_lock when the data
>> in cache is dirty.
>>
>> This is done only if OSI is not supported in PSCI. If OSI is supported
>> rpmh_flush can get invoked when the last cpu going to power collapse
>> deepest low power mode.
>>
>> Also remove "depends on COMPILE_TEST" for Kconfig option QCOM_RPMH so the
>> driver is only compiled for arm64 which supports psci_has_osi_support()
>> API.
>>
>> Signed-off-by: Maulik Shah <[e-mail]>
>> Reviewed-by: Srinivas Rao L <[e-mail]>
>> ---
>>  drivers/soc/qcom/Kconfig |  2 +-
>>  drivers/soc/qcom/rpmh.c  | 37 ++++++++++++++++++++++---------------
>>  2 files changed, 23 insertions(+), 16 deletions(-)
>>
>> diff --git a/drivers/soc/qcom/Kconfig b/drivers/soc/qcom/Kconfig
>> index d0a73e7..2e581bc 100644
>> --- a/drivers/soc/qcom/Kconfig
>> +++ b/drivers/soc/qcom/Kconfig
>> @@ -105,7 +105,7 @@ config QCOM_RMTFS_MEM
>>
>>  config QCOM_RPMH
>>         bool "Qualcomm RPM-Hardened (RPMH) Communication"
>> -       depends on ARCH_QCOM && ARM64 || COMPILE_TEST
>> +       depends on ARCH_QCOM && ARM64
>>         help
>>           Support for communication with the hardened-RPM blocks in
>>           Qualcomm Technologies Inc (QTI) SoCs. RPMH communication uses an
>> diff --git a/drivers/soc/qcom/rpmh.c b/drivers/soc/qcom/rpmh.c
>> index f28afe4..dafb0da 100644
>> --- a/drivers/soc/qcom/rpmh.c
>> +++ b/drivers/soc/qcom/rpmh.c
>> @@ -12,6 +12,7 @@
>>  #include <linux/module.h>
>>  #include <linux/of.h>
>>  #include <linux/platform_device.h>
>> +#include <linux/psci.h>
>>  #include <linux/slab.h>
>>  #include <linux/spinlock.h>
>>  #include <linux/types.h>
>> @@ -158,6 +159,13 @@ static struct cache_req *cache_rpm_request(struct rpmh_ctrlr *ctrlr,
>>         }
>>
>>  unlock:
>> +       if (ctrlr->dirty && !psci_has_osi_support()) {
>> +               if (rpmh_flush(ctrlr)) {
>> +                       spin_unlock_irqrestore(&ctrlr->cache_lock, flags);
>> +                       return ERR_PTR(-EINVAL);
>> +               }
>> +       }
>> +
>>         spin_unlock_irqrestore(&ctrlr->cache_lock, flags);
>>
>>         return req;
>> @@ -285,26 +293,35 @@ int rpmh_write(const struct device *dev, enum rpmh_state state,
>>  }
>>  EXPORT_SYMBOL(rpmh_write);
>>
>> -static void cache_batch(struct rpmh_ctrlr *ctrlr, struct batch_cache_req *req)
>> +static int cache_batch(struct rpmh_ctrlr *ctrlr, struct batch_cache_req *req)
>>  {
>>         unsigned long flags;
>>
>>         spin_lock_irqsave(&ctrlr->cache_lock, flags);
>> +
>>         list_add_tail(&req->list, &ctrlr->batch_cache);
>>         ctrlr->dirty = true;
>> +
>> +       if (!psci_has_osi_support()) {
>> +               if (rpmh_flush(ctrlr)) {
> The whole API here is a bit unfortunate.  From what I can tell,
> callers of this code almost always call rpmh_write_batch() in
> triplicate, AKA:
>
> rpmh_write_batch(active, ...)
> rpmh_write_batch(wake, ...)
> rpmh_write_batch(sleep, ...)
>
> ...that's going to end up writing the whole sleep/wake sets twice
> every single time, right?  I know you talked about trying to keep
> separate dirty bits for sleep/wake and maybe that would help, but it
> might not be so easy due to the comparison of "sleep_val" and
> "wake_val" in is_req_valid().
>
> I guess we can keep the inefficiency for now and see how much it hits
> us, but it feels ugly.
>
>
>> +                       spin_unlock_irqrestore(&ctrlr->cache_lock, flags);
>> +                       return -EINVAL;
> nit: why not add "int ret = 0" to the top of the function, then here:
>
> if (rpmh_flush(ctrl))
>   ret = -EINVAL;
>
> ...then at the end "return ret".  It avoids the 2nd copy of the unlock?
Done.
>
> Also: Why throw away the return value of rpmh_flush and replace it
> with -EINVAL?  Trying to avoid -EBUSY?  ...oh, should you handle
> -EBUSY?  AKA:
>
> if (!psci_has_osi_support()) {
>   do {
>     ret = rpmh_flush(ctrl);
>   } while (ret == -EBUSY);
> }

Done, the return value from rpmh_flush() can be -EAGAIN, not -EBUSY.

i will update the comment accordingly and will include below change as well in next series.

https://patchwork.kernel.org/patch/11364067/

this should address for caller to not handle -EAGAIN.

>
>
>> +               }
>> +       }
>> +
>>         spin_unlock_irqrestore(&ctrlr->cache_lock, flags);
>> +
>> +       return 0;
>>  }
>>
>>  static int flush_batch(struct rpmh_ctrlr *ctrlr)
>>  {
>>         struct batch_cache_req *req;
>>         const struct rpmh_request *rpm_msg;
>> -       unsigned long flags;
>>         int ret = 0;
>>         int i;
>>
>>         /* Send Sleep/Wake requests to the controller, expect no response */
>> -       spin_lock_irqsave(&ctrlr->cache_lock, flags);
>>         list_for_each_entry(req, &ctrlr->batch_cache, list) {
>>                 for (i = 0; i < req->count; i++) {
>>                         rpm_msg = req->rpm_msgs + i;
>> @@ -314,7 +331,6 @@ static int flush_batch(struct rpmh_ctrlr *ctrlr)
>>                                 break;
>>                 }
>>         }
>> -       spin_unlock_irqrestore(&ctrlr->cache_lock, flags);
>>
>>         return ret;
>>  }
>> @@ -386,10 +402,8 @@ int rpmh_write_batch(const struct device *dev, enum rpmh_state state,
>>                 cmd += n[i];
>>         }
>>
>> -       if (state != RPMH_ACTIVE_ONLY_STATE) {
>> -               cache_batch(ctrlr, req);
>> -               return 0;
>> -       }
>> +       if (state != RPMH_ACTIVE_ONLY_STATE)
>> +               return cache_batch(ctrlr, req);
>>
>>         for (i = 0; i < count; i++) {
>>                 struct completion *compl = &compls[i];
>> @@ -455,9 +469,6 @@ static int send_single(struct rpmh_ctrlr *ctrlr, enum rpmh_state state,
>>   * Return: -EBUSY if the controller is busy, probably waiting on a response
>>   * to a RPMH request sent earlier.
>>   *
>> - * This function is always called from the sleep code from the last CPU
>> - * that is powering down the entire system. Since no other RPMH API would be
>> - * executing at this time, it is safe to run lockless.
> nit: you've now got an extra "blank" (just has a "*" on it) line at
> the end of your comment block.
Done.
> nit: in v9, Evan suggested "We should probably replace that with a
> comment indicating that we assume ctrlr->cache_lock is already held".
> Maybe you could do that?
yes i left it for below reason since we still can call it from sleep code.
i will mention same in v11.

Thanks,
Maulik
>
> Also: presumably you _will_ still be called by the sleep code from the
> last CPU on systems with OSI.  Is that true?  If that's not true then
> you should change your function to static.  If that is true, then your
> comment should be something like "this function will either be called
> from sleep code on the last CPU (thus no spinlock needed) or with the
> spinlock already held".
>
>
> -Doug

Doug Anderson March 5, 2020, 10:20 p.m. UTC | #3

Hi,

On Thu, Mar 5, 2020 at 3:30 AM Maulik Shah <[e-mail]> wrote:
>
> >> +                       spin_unlock_irqrestore(&ctrlr->cache_lock, flags);
> >> +                       return -EINVAL;
> > nit: why not add "int ret = 0" to the top of the function, then here:
> >
> > if (rpmh_flush(ctrl))
> >   ret = -EINVAL;
> >
> > ...then at the end "return ret".  It avoids the 2nd copy of the unlock?
> Done.
> >
> > Also: Why throw away the return value of rpmh_flush and replace it
> > with -EINVAL?  Trying to avoid -EBUSY?  ...oh, should you handle
> > -EBUSY?  AKA:
> >
> > if (!psci_has_osi_support()) {
> >   do {
> >     ret = rpmh_flush(ctrl);
> >   } while (ret == -EBUSY);
> > }
>
> Done, the return value from rpmh_flush() can be -EAGAIN, not -EBUSY.
>
> i will update the comment accordingly and will include below change as well in next series.
>
> https://patchwork.kernel.org/patch/11364067/
>
> this should address for caller to not handle -EAGAIN.

A few issues, I guess.

1. I _think_ it's important that you enable interrupts between
retries.  If you're on the same CPU that the interrupt is routed to
and you were waiting for 'tcs_in_use' to be cleared you'll be in
trouble otherwise.  ...I think we need to audit all of the places that
are looping based on -EAGAIN and confirm that interrupts are enabled
between retries.  Before your patch series the only looping I see was
in rpmh_invalidate() and the lock wasn't held.  After your series it's
also in rpmh_flush() which is called under spin_lock_irqsave() which
will be a problem.

2. The RPMH code uses both -EBUSY and -EAGAIN so I looked carefully at
this again.  You're right that -EBUSY seems to be exclusively returned
by things only called by rpmh_rsc_send_data() and that function
handles the retries.  ...but looking at this made me find a broken
corner case with the "zero active tcs" case (assuming you care about
this case as per your other thread).  Specifically if you have "zero
active tcs" then get_tcs_for_msg() can call rpmh_rsc_invalidate()
which can return -EAGAIN.  That will return the -EAGAIN out of
tcs_write() into rpmh_rsc_send_data().  rpmh_rsc_send_data() only
handles -EBUSY, not -EAGAIN.

-Doug

Maulik Shah March 10, 2020, 11 a.m. UTC | #4

On 3/6/2020 3:50 AM, Doug Anderson wrote:
> Hi,
>
> On Thu, Mar 5, 2020 at 3:30 AM Maulik Shah <[e-mail]> wrote:
>>>> +                       spin_unlock_irqrestore(&ctrlr->cache_lock, flags);
>>>> +                       return -EINVAL;
>>> nit: why not add "int ret = 0" to the top of the function, then here:
>>>
>>> if (rpmh_flush(ctrl))
>>>   ret = -EINVAL;
>>>
>>> ...then at the end "return ret".  It avoids the 2nd copy of the unlock?
>> Done.
>>> Also: Why throw away the return value of rpmh_flush and replace it
>>> with -EINVAL?  Trying to avoid -EBUSY?  ...oh, should you handle
>>> -EBUSY?  AKA:
>>>
>>> if (!psci_has_osi_support()) {
>>>   do {
>>>     ret = rpmh_flush(ctrl);
>>>   } while (ret == -EBUSY);
>>> }
>> Done, the return value from rpmh_flush() can be -EAGAIN, not -EBUSY.
>>
>> i will update the comment accordingly and will include below change as well in next series.
>>
>> https://patchwork.kernel.org/patch/11364067/
>>
>> this should address for caller to not handle -EAGAIN.
> A few issues, I guess.
>
> 1. I _think_ it's important that you enable interrupts between
> retries.  If you're on the same CPU that the interrupt is routed to
> and you were waiting for 'tcs_in_use' to be cleared you'll be in
> trouble otherwise.  ...I think we need to audit all of the places that
> are looping based on -EAGAIN and confirm that interrupts are enabled
> between retries.  Before your patch series the only looping I see was
> in rpmh_invalidate() and the lock wasn't held.  After your series it's
> also in rpmh_flush() which is called under spin_lock_irqsave() which
> will be a problem.
I will take a look at interrupts part.
>
> 2. The RPMH code uses both -EBUSY and -EAGAIN so I looked carefully at
> this again.  You're right that -EBUSY seems to be exclusively returned
> by things only called by rpmh_rsc_send_data() and that function
> handles the retries.  ...but looking at this made me find a broken
> corner case with the "zero active tcs" case (assuming you care about
> this case as per your other thread).  Specifically if you have "zero
> active tcs" then get_tcs_for_msg() can call rpmh_rsc_invalidate()
> which can return -EAGAIN.  That will return the -EAGAIN out of
> tcs_write() into rpmh_rsc_send_data().  rpmh_rsc_send_data() only
> handles -EBUSY, not -EAGAIN.
>
> -Doug

Thanks Doug. I will have a patch to fix this.

Thanks,
Maulik

Patch

diff --git a/drivers/soc/qcom/Kconfig b/drivers/soc/qcom/Kconfig
index d0a73e7..2e581bc 100644
--- a/drivers/soc/qcom/Kconfig
+++ b/drivers/soc/qcom/Kconfig
@@ -105,7 +105,7 @@  config QCOM_RMTFS_MEM
 
 config QCOM_RPMH
 	bool "Qualcomm RPM-Hardened (RPMH) Communication"
-	depends on ARCH_QCOM && ARM64 || COMPILE_TEST
+	depends on ARCH_QCOM && ARM64
 	help
 	  Support for communication with the hardened-RPM blocks in
 	  Qualcomm Technologies Inc (QTI) SoCs. RPMH communication uses an
diff --git a/drivers/soc/qcom/rpmh.c b/drivers/soc/qcom/rpmh.c
index f28afe4..dafb0da 100644
--- a/drivers/soc/qcom/rpmh.c
+++ b/drivers/soc/qcom/rpmh.c
@@ -12,6 +12,7 @@ 
 #include <linux/module.h>
 #include <linux/of.h>
 #include <linux/platform_device.h>
+#include <linux/psci.h>
 #include <linux/slab.h>
 #include <linux/spinlock.h>
 #include <linux/types.h>
@@ -158,6 +159,13 @@  static struct cache_req *cache_rpm_request(struct rpmh_ctrlr *ctrlr,
 	}
 
 unlock:
+	if (ctrlr->dirty && !psci_has_osi_support()) {
+		if (rpmh_flush(ctrlr)) {
+			spin_unlock_irqrestore(&ctrlr->cache_lock, flags);
+			return ERR_PTR(-EINVAL);
+		}
+	}
+
 	spin_unlock_irqrestore(&ctrlr->cache_lock, flags);
 
 	return req;
@@ -285,26 +293,35 @@  int rpmh_write(const struct device *dev, enum rpmh_state state,
 }
 EXPORT_SYMBOL(rpmh_write);
 
-static void cache_batch(struct rpmh_ctrlr *ctrlr, struct batch_cache_req *req)
+static int cache_batch(struct rpmh_ctrlr *ctrlr, struct batch_cache_req *req)
 {
 	unsigned long flags;
 
 	spin_lock_irqsave(&ctrlr->cache_lock, flags);
+
 	list_add_tail(&req->list, &ctrlr->batch_cache);
 	ctrlr->dirty = true;
+
+	if (!psci_has_osi_support()) {
+		if (rpmh_flush(ctrlr)) {
+			spin_unlock_irqrestore(&ctrlr->cache_lock, flags);
+			return -EINVAL;
+		}
+	}
+
 	spin_unlock_irqrestore(&ctrlr->cache_lock, flags);
+
+	return 0;
 }
 
 static int flush_batch(struct rpmh_ctrlr *ctrlr)
 {
 	struct batch_cache_req *req;
 	const struct rpmh_request *rpm_msg;
-	unsigned long flags;
 	int ret = 0;
 	int i;
 
 	/* Send Sleep/Wake requests to the controller, expect no response */
-	spin_lock_irqsave(&ctrlr->cache_lock, flags);
 	list_for_each_entry(req, &ctrlr->batch_cache, list) {
 		for (i = 0; i < req->count; i++) {
 			rpm_msg = req->rpm_msgs + i;
@@ -314,7 +331,6 @@  static int flush_batch(struct rpmh_ctrlr *ctrlr)
 				break;
 		}
 	}
-	spin_unlock_irqrestore(&ctrlr->cache_lock, flags);
 
 	return ret;
 }
@@ -386,10 +402,8 @@  int rpmh_write_batch(const struct device *dev, enum rpmh_state state,
 		cmd += n[i];
 	}
 
-	if (state != RPMH_ACTIVE_ONLY_STATE) {
-		cache_batch(ctrlr, req);
-		return 0;
-	}
+	if (state != RPMH_ACTIVE_ONLY_STATE)
+		return cache_batch(ctrlr, req);
 
 	for (i = 0; i < count; i++) {
 		struct completion *compl = &compls[i];
@@ -455,9 +469,6 @@  static int send_single(struct rpmh_ctrlr *ctrlr, enum rpmh_state state,
  * Return: -EBUSY if the controller is busy, probably waiting on a response
  * to a RPMH request sent earlier.
  *
- * This function is always called from the sleep code from the last CPU
- * that is powering down the entire system. Since no other RPMH API would be
- * executing at this time, it is safe to run lockless.
  */
 int rpmh_flush(struct rpmh_ctrlr *ctrlr)
 {
@@ -474,10 +485,6 @@  int rpmh_flush(struct rpmh_ctrlr *ctrlr)
 	if (ret)
 		return ret;
 
-	/*
-	 * Nobody else should be calling this function other than system PM,
-	 * hence we can run without locks.
-	 */
 	list_for_each_entry(p, &ctrlr->cache, list) {
 		if (!is_req_valid(p)) {
 			pr_debug("%s: skipping RPMH req: a:%#x s:%#x w:%#x",