Message-ID: <92af6290-39a9-e623-9c8d-7c87bd92a821@ti.com>
Date:   Fri, 20 Dec 2019 13:24:25 +0200
From:   Tero Kristo <t-kristo@...com>
To:     Suman Anna <s-anna@...com>,
        Mathieu Poirier <mathieu.poirier@...aro.org>
CC:     <bjorn.andersson@...aro.org>, <ohad@...ery.com>,
        <linux-remoteproc@...r.kernel.org>, <linux-kernel@...r.kernel.org>,
        <linux-omap@...r.kernel.org>
Subject: Re: [PATCHv3 13/15] remoteproc/omap: add support for runtime
 auto-suspend/resume

On 20/12/2019 05:24, Suman Anna wrote:
> Hi Tero,
> 
> On 12/19/19 5:43 PM, Mathieu Poirier wrote:
>> On Fri, Dec 13, 2019 at 02:55:35PM +0200, Tero Kristo wrote:
>>> From: Suman Anna <s-anna@...com>
>>>
>>> This patch enhances the PM support in the OMAP remoteproc driver to
>>> support the runtime auto-suspend. A remoteproc may not be required to
>>> be running all the time, and typically will need to be active only
>>> during certain usecases. As such, to save power, it should be turned
>>> off during potential long periods of inactivity between usecases.
>>> This suspend and resume of the device is a relatively heavy process
>>> in terms of latencies, so a remoteproc should be suspended only after
>>> a certain period of prolonged inactivity. The OMAP remoteproc driver
>>> leverages the runtime pm framework's auto_suspend feature to accomplish
>>> this functionality. This feature is automatically enabled when a remote
>>> processor has successfully booted. The 'autosuspend_delay_ms' for each
>>> device dictates the inactivity period/time to wait for before
>>> suspending the device.
>>>
>>> The runtime auto-suspend design relies on marking the last busy time
>>> on every communication (virtqueue kick) to and from the remote processor.
>>> When there has been no activity for 'autosuspend_delay_ms' time, the
>>> runtime PM framework invokes the driver's runtime pm suspend callback
>>> to suspend the device. The remote processor will be woken up on the
>>> initiation of the next communication message through the runtime pm
>>> resume callback. The current auto-suspend design also allows a remote
>>> processor to deny a auto-suspend attempt, if it wishes to, by sending a
>>> NACK response to the initial suspend request message sent to the remote
>>> processor as part of the suspend process. The auto-suspend request is
>>> also only attempted if the remote processor is idled and in standby at
>>> the time of inactivity timer expiry. This choice is made to avoid
>>> unnecessary messaging, and the auto-suspend is simply rescheduled to
>>> be attempted again after a further lapse of autosuspend_delay_ms.
>>>
>>> The runtime pm callbacks functionality in this patch reuses most of the
>>> core logic from the suspend/resume support code, and make use of an
>>> additional auto_suspend flag to differentiate the logic in common code
>>> from system suspend. The system suspend/resume sequences are also updated
>>> to reflect the proper pm_runtime statuses, and also to really perform a
>>> suspend/resume only if the remoteproc has not been auto-suspended at the
>>> time of request. The remote processor is left in suspended state on a
>>> system resume if it has been auto-suspended before, and will be woken up
>>> only when a usecase needs to run. The other significant change in this
>>> patch is to reset the remoteproc device's pm_domain so as to avoid
>>> conflicts with the ordering sequences in the device pm_domain's runtime
>>> callbacks and the reset management and clock management implemented
>>> within the runtime callbacks in the driver.
>>>
>>> The OMAP remoteproc driver currently uses a default value of 10 seconds
>>> for all OMAP remoteprocs, and a different value can be chosen either by
>>> choosing a positive value for the 'autosuspend_delay' in the device's
>>> omap_rproc_fw_data in the driver match data or by updating the
>>> 'autosuspend_delay_ms' field at runtime through the sysfs interface.
>>>      Eg: To use 25 seconds for IPU2 on DRA7xx,
>>>        echo 25000 > /sys/bus/platform/devices/55020000.ipu/power/autosuspend_delay_ms
>>>
>>> The runtime suspend feature can also be similarly enabled or disabled by
>>> writing 'auto' or 'on' to the device's 'control' power field. The default
>>> is enabled.
>>>      Eg: To disable auto-suspend for IPU2 on DRA7xx SoC,
>>>        echo on > /sys/bus/platform/devices/55020000.ipu/power/control
>>>
>>> Signed-off-by: Suman Anna <s-anna@...com>
>>> [t-kristo@...com: converted to use ti-sysc instead of hwmod]
>>> Signed-off-by: Tero Kristo <t-kristo@...com>
>>> ---
>>>   drivers/remoteproc/omap_remoteproc.c | 220 ++++++++++++++++++++++++++-
>>>   1 file changed, 214 insertions(+), 6 deletions(-)
>>>
>>> diff --git a/drivers/remoteproc/omap_remoteproc.c b/drivers/remoteproc/omap_remoteproc.c
>>> index 0a9b9f7d20da..463d6f60947a 100644
>>> --- a/drivers/remoteproc/omap_remoteproc.c
>>> +++ b/drivers/remoteproc/omap_remoteproc.c
>>> @@ -20,6 +20,7 @@
>>>   #include <linux/of_device.h>
>>>   #include <linux/of_reserved_mem.h>
>>>   #include <linux/platform_device.h>
>>> +#include <linux/pm_runtime.h>
>>>   #include <linux/dma-mapping.h>
>>>   #include <linux/remoteproc.h>
>>>   #include <linux/mailbox_client.h>
>>> @@ -37,6 +38,9 @@
>>>   #include "omap_remoteproc.h"
>>>   #include "remoteproc_internal.h"
>>>   
>>> +/* default auto-suspend delay (ms) */
>>> +#define DEFAULT_AUTOSUSPEND_DELAY		10000
>>> +
>>>   /**
>>>    * struct omap_rproc_boot_data - boot data structure for the DSP omap rprocs
>>>    * @syscon: regmap handle for the system control configuration module
>>> @@ -83,6 +87,8 @@ struct omap_rproc_timer {
>>>    * @num_mems: number of internal memory regions
>>>    * @num_timers: number of rproc timer(s)
>>>    * @timers: timer(s) info used by rproc
>>> + * @autosuspend_delay: auto-suspend delay value to be used for runtime pm
>>> + * @need_resume: if true a resume is needed in the system resume callback
>>>    * @rproc: rproc handle
>>>    * @reset: reset handle
>>>    * @pm_comp: completion primitive to sync for suspend response
>>> @@ -97,6 +103,8 @@ struct omap_rproc {
>>>   	int num_mems;
>>>   	int num_timers;
>>>   	struct omap_rproc_timer *timers;
>>> +	int autosuspend_delay;
>>> +	bool need_resume;
>>>   	struct rproc *rproc;
>>>   	struct reset_control *reset;
>>>   	struct completion pm_comp;
>>> @@ -111,6 +119,7 @@ struct omap_rproc {
>>>    * @boot_reg_shift: bit shift for the boot register mask
>>>    * @mem_names: memory names for this remote processor
>>>    * @dev_addrs: device addresses corresponding to the memory names
>>> + * @autosuspend_delay: custom auto-suspend delay value in milliseconds
>>>    */
>>>   struct omap_rproc_dev_data {
>>>   	const char *device_name;
>>> @@ -118,6 +127,7 @@ struct omap_rproc_dev_data {
>>>   	int boot_reg_shift;
>>>   	const char * const *mem_names;
>>>   	const u32 *dev_addrs;
>>> +	int autosuspend_delay;
>>>   };
>>>   
>>>   /**
>>> @@ -384,11 +394,23 @@ static void omap_rproc_kick(struct rproc *rproc, int vqid)
>>>   	struct device *dev = rproc->dev.parent;
>>>   	int ret;
>>>   
>>> +	/* wake up the rproc before kicking it */
>>> +	ret = pm_runtime_get_sync(dev);
>>> +	if (WARN_ON(ret < 0)) {
>>> +		dev_err(dev, "pm_runtime_get_sync() failed during kick, ret = %d\n",
>>> +			ret);
>>> +		pm_runtime_put_noidle(dev);
>>> +		return;
>>> +	}
>>> +
>>>   	/* send the index of the triggered virtqueue in the mailbox payload */
>>>   	ret = mbox_send_message(oproc->mbox, (void *)vqid);
>>>   	if (ret < 0)
>>>   		dev_err(dev, "failed to send mailbox message, status = %d\n",
>>>   			ret);
>>> +
>>> +	pm_runtime_mark_last_busy(dev);
>>> +	pm_runtime_put_autosuspend(dev);
>>>   }
>>>   
>>>   /**
>>> @@ -473,6 +495,19 @@ static int omap_rproc_start(struct rproc *rproc)
>>>   		goto put_mbox;
>>>   	}
>>>   
>>> +	/*
>>> +	 * remote processor is up, so update the runtime pm status and
>>> +	 * enable the auto-suspend. The device usage count is incremented
>>> +	 * manually for balancing it for auto-suspend
>>> +	 */
>>> +	pm_runtime_set_active(dev);
>>> +	pm_runtime_set_autosuspend_delay(dev, oproc->autosuspend_delay);
>>> +	pm_runtime_use_autosuspend(dev);
>>> +	pm_runtime_get(dev);
>>> +	pm_runtime_enable(dev);
>>> +	pm_runtime_mark_last_busy(dev);
>>> +	pm_runtime_put_autosuspend(dev);
>>> +
>>>   	reset_control_deassert(oproc->reset);
> 
> I see that you have flipped the reset call and all the pm_runtime calls
> (w.r.t my original code sequence) and pm_runtime_get instead of
> pm_runtime_get_noresume(). Is there a reason for it? What is the backend
> that gets exercised with pm_runtime?

PM runtime nowadays exercises ti-sysc, and we can't deassert reset 
before we enable the clocks for the device. Thus, these are flipped.

In the old incarnation of this code, PM runtime just exercised the 
hacked up functionality implemented for omap rproc alone.

> 
>>>   
>>>   	return 0;
>>> @@ -485,9 +520,26 @@ static int omap_rproc_start(struct rproc *rproc)
>>>   /* power off the remote processor */
>>>   static int omap_rproc_stop(struct rproc *rproc)
>>>   {
>>> +	struct device *dev = rproc->dev.parent;
>>>   	struct omap_rproc *oproc = rproc->priv;
>>>   	int ret;
>>>   
>>> +	/*
>>> +	 * cancel any possible scheduled runtime suspend by incrementing
>>> +	 * the device usage count, and resuming the device. The remoteproc
>>> +	 * also needs to be woken up if suspended, to avoid the remoteproc
>>> +	 * OS to continue to remember any context that it has saved, and
>>> +	 * avoid potential issues in misindentifying a subsequent device
>>> +	 * reboot as a power restore boot
>>> +	 */
>>> +	ret = pm_runtime_get_sync(dev);
>>> +	if (ret < 0) {
>>> +		pm_runtime_put_noidle(dev);
>>> +		return ret;
>>> +	}
>>> +
>>> +	pm_runtime_put_sync(dev);
>>> +
> 
> I didn't have this call either. And get_sync() followed by put_sync() is
> essentially a no-op. Am I missing something here?

Hmm right, but you did have the device_shutdown() here, somehow this 
appears to simulate that, but you are right, it is just a no-op. I don't 
know if this code has actually ever done anything useful. Let me 
experiment with this one a bit and see what happens. It looks like it 
should be removed completely.

> 
>>>   	reset_control_assert(oproc->reset);
>>>   
>>>   	ret = omap_rproc_disable_timers(rproc, true);
>>> @@ -496,6 +548,15 @@ static int omap_rproc_stop(struct rproc *rproc)
>>>   
>>>   	mbox_free_channel(oproc->mbox);
>>>   
>>> +	/*
>>> +	 * update the runtime pm states and status now that the remoteproc
>>> +	 * has stopped
>>> +	 */
>>> +	pm_runtime_disable(dev);
>>> +	pm_runtime_dont_use_autosuspend(dev);
>>> +	pm_runtime_put_noidle(dev);
>>> +	pm_runtime_set_suspended(dev);
>>> +
>>>   	return 0;
>>>   }
>>>   
>>> @@ -552,17 +613,19 @@ static bool _is_rproc_in_standby(struct omap_rproc *oproc)
>>>   
>>>   /* 1 sec is long enough time to let the remoteproc side suspend the device */
>>>   #define DEF_SUSPEND_TIMEOUT 1000
>>> -static int _omap_rproc_suspend(struct rproc *rproc)
>>> +static int _omap_rproc_suspend(struct rproc *rproc, bool auto_suspend)
>>>   {
>>>   	struct device *dev = rproc->dev.parent;
>>>   	struct omap_rproc *oproc = rproc->priv;
>>>   	unsigned long to = msecs_to_jiffies(DEF_SUSPEND_TIMEOUT);
>>>   	unsigned long ta = jiffies + to;
>>> +	u32 suspend_msg = auto_suspend ?
>>> +				RP_MBOX_SUSPEND_AUTO : RP_MBOX_SUSPEND_SYSTEM;
>>>   	int ret;
>>>   
>>>   	reinit_completion(&oproc->pm_comp);
>>>   	oproc->suspend_acked = false;
>>> -	ret = mbox_send_message(oproc->mbox, (void *)RP_MBOX_SUSPEND_SYSTEM);
>>> +	ret = mbox_send_message(oproc->mbox, (void *)suspend_msg);
>>>   	if (ret < 0) {
>>>   		dev_err(dev, "PM mbox_send_message failed: %d\n", ret);
>>>   		return ret;
>>> @@ -602,25 +665,55 @@ static int _omap_rproc_suspend(struct rproc *rproc)
>>>   		goto enable_device;
>>>   	}
>>>   
>>> +	/*
>>> +	 * IOMMUs would have to be disabled specifically for runtime suspend.
>>> +	 * They are handled automatically through System PM callbacks for
>>> +	 * regular system suspend
>>> +	 */
>>> +	if (auto_suspend) {
>>> +		ret = omap_iommu_domain_deactivate(rproc->domain);
>>> +		if (ret) {
>>> +			dev_err(dev, "iommu domain deactivate failed %d\n",
>>> +				ret);
>>> +			goto enable_timers;
>>> +		}
>>> +	}
>>> +
>>>   	return 0;
> 
> blank line here, and remove the one before enable_device.

Hmm why? I don't think there is any coding guideline that would dictate 
this. However, I'll fix this as this is effectively your file anyways.

-Tero

> 
> regards
> Suman
> 
>>> +enable_timers:
>>> +	/* ignore errors on re-enabling code */
>>> +	omap_rproc_enable_timers(rproc, false);
>>>   
>>>   enable_device:
>>>   	reset_control_deassert(oproc->reset);
>>>   	return ret;
>>>   }
>>>   
>>> -static int _omap_rproc_resume(struct rproc *rproc)
>>> +static int _omap_rproc_resume(struct rproc *rproc, bool auto_suspend)
>>>   {
>>>   	struct device *dev = rproc->dev.parent;
>>>   	struct omap_rproc *oproc = rproc->priv;
>>>   	int ret;
>>>   
>>> +	/*
>>> +	 * IOMMUs would have to be enabled specifically for runtime resume.
>>> +	 * They would have been already enabled automatically through System
>>> +	 * PM callbacks for regular system resume
>>> +	 */
>>> +	if (auto_suspend) {
>>> +		ret = omap_iommu_domain_activate(rproc->domain);
>>> +		if (ret) {
>>> +			dev_err(dev, "omap_iommu activate failed %d\n", ret);
>>> +			goto out;
>>> +		}
>>> +	}
>>> +
>>>   	/* boot address could be lost after suspend, so restore it */
>>>   	if (oproc->boot_data) {
>>>   		ret = omap_rproc_write_dsp_boot_addr(rproc);
>>>   		if (ret) {
>>>   			dev_err(dev, "boot address restore failed %d\n", ret);
>>> -			goto out;
>>> +			goto suspend_iommu;
>>
>> The same needs to be done if omap_rproc_enable_timers() fails.
>>
>>>   		}
>>>   	}
>>>   
>>> @@ -633,6 +726,12 @@ static int _omap_rproc_resume(struct rproc *rproc)
>>>   
>>>   	reset_control_deassert(oproc->reset);
>>>   
>>> +	return 0;
>>> +
>>> +suspend_iommu:
>>> +	if (auto_suspend)
>>> +		omap_iommu_domain_deactivate(rproc->domain);
>>> +
>>>   out:
>>>   	return ret;
>>>   }
>>> @@ -641,6 +740,7 @@ static int __maybe_unused omap_rproc_suspend(struct device *dev)
>>>   {
>>>   	struct platform_device *pdev = to_platform_device(dev);
>>>   	struct rproc *rproc = platform_get_drvdata(pdev);
>>> +	struct omap_rproc *oproc = rproc->priv;
>>>   	int ret = 0;
>>>   
>>>   	mutex_lock(&rproc->lock);
>>> @@ -655,13 +755,25 @@ static int __maybe_unused omap_rproc_suspend(struct device *dev)
>>>   		goto out;
>>>   	}
>>>   
>>> -	ret = _omap_rproc_suspend(rproc);
>>> +	ret = _omap_rproc_suspend(rproc, false);
>>>   	if (ret) {
>>>   		dev_err(dev, "suspend failed %d\n", ret);
>>>   		goto out;
>>>   	}
>>>   
>>> +	/*
>>> +	 * remoteproc is running at the time of system suspend, so remember
>>> +	 * it so as to wake it up during system resume
>>> +	 */
>>> +	oproc->need_resume = 1;
>>
>> Please use 'true' to be consistent with the type and omap_rproc_resume().
>>
>>>   	rproc->state = RPROC_SUSPENDED;
>>> +
>>> +	/*
>>> +	 * update the runtime pm status to be suspended, without decrementing
>>> +	 * the device usage count
>>> +	 */
>>> +	pm_runtime_disable(dev);
>>> +	pm_runtime_set_suspended(dev);
>>>   out:
>>>   	mutex_unlock(&rproc->lock);
>>>   	return ret;
>>> @@ -671,6 +783,7 @@ static int __maybe_unused omap_rproc_resume(struct device *dev)
>>>   {
>>>   	struct platform_device *pdev = to_platform_device(dev);
>>>   	struct rproc *rproc = platform_get_drvdata(pdev);
>>> +	struct omap_rproc *oproc = rproc->priv;
>>>   	int ret = 0;
>>>   
>>>   	mutex_lock(&rproc->lock);
>>> @@ -682,17 +795,91 @@ static int __maybe_unused omap_rproc_resume(struct device *dev)
>>>   		goto out;
>>>   	}
>>>   
>>> -	ret = _omap_rproc_resume(rproc);
>>> +	/*
>>> +	 * remoteproc was auto-suspended at the time of system suspend,
>>> +	 * so no need to wake-up the processor (leave it in suspended
>>> +	 * state, will be woken up during a subsequent runtime_resume)
>>> +	 */
>>> +	if (!oproc->need_resume)
>>> +		goto out;
>>> +
>>> +	ret = _omap_rproc_resume(rproc, false);
>>>   	if (ret) {
>>>   		dev_err(dev, "resume failed %d\n", ret);
>>>   		goto out;
>>>   	}
>>> +	oproc->need_resume = false;
>>>   
>>>   	rproc->state = RPROC_RUNNING;
>>> +
>>> +	/*
>>> +	 * update the runtime pm status to be active, without incrementing
>>> +	 * the device usage count
>>> +	 */
>>> +	pm_runtime_set_active(dev);
>>> +	pm_runtime_enable(dev);
>>> +	pm_runtime_mark_last_busy(dev);
>>>   out:
>>>   	mutex_unlock(&rproc->lock);
>>>   	return ret;
>>>   }
>>> +
>>> +static int omap_rproc_runtime_suspend(struct device *dev)
>>> +{
>>> +	struct rproc *rproc = dev_get_drvdata(dev);
>>> +	struct omap_rproc *oproc = rproc->priv;
>>> +	int ret;
>>> +
>>> +	if (rproc->state == RPROC_CRASHED) {
>>> +		dev_dbg(dev, "rproc cannot be runtime suspended when crashed!\n");
>>> +		return -EBUSY;
>>> +	}
>>> +
>>> +	if (WARN_ON(rproc->state != RPROC_RUNNING)) {
>>> +		dev_err(dev, "rproc cannot be runtime suspended when not running!\n");
>>> +		return -EBUSY;
>>> +	}
>>> +
>>> +	/*
>>> +	 * do not even attempt suspend if the remote processor is not
>>> +	 * idled for runtime auto-suspend
>>> +	 */
>>> +	if (!_is_rproc_in_standby(oproc)) {
>>> +		ret = -EBUSY;
>>> +		goto abort;
>>> +	}
>>> +
>>> +	ret = _omap_rproc_suspend(rproc, true);
>>> +	if (ret)
>>> +		goto abort;
>>> +
>>> +	rproc->state = RPROC_SUSPENDED;
>>> +	return 0;
>>> +
>>> +abort:
>>> +	pm_runtime_mark_last_busy(dev);
>>> +	return ret;
>>> +}
>>> +
>>> +static int omap_rproc_runtime_resume(struct device *dev)
>>> +{
>>> +	struct rproc *rproc = dev_get_drvdata(dev);
>>> +	int ret;
>>> +
>>> +	if (WARN_ON(rproc->state != RPROC_SUSPENDED)) {
>>> +		dev_err(dev, "rproc cannot be runtime resumed if not suspended!\n");
>>> +		return -EBUSY;
>>> +	}
>>> +
>>> +	ret = _omap_rproc_resume(rproc, true);
>>> +	if (ret) {
>>> +		dev_err(dev, "runtime resume failed %d\n", ret);
>>> +		return ret;
>>> +	}
>>> +
>>> +	rproc->state = RPROC_RUNNING;
>>> +	return 0;
>>> +}
>>>   #endif /* CONFIG_PM */
>>>   
>>>   static const char * const ipu_mem_names[] = {
>>> @@ -778,6 +965,20 @@ static const struct of_device_id omap_rproc_of_match[] = {
>>>   };
>>>   MODULE_DEVICE_TABLE(of, omap_rproc_of_match);
>>>   
>>> +static int omap_rproc_get_autosuspend_delay(struct platform_device *pdev)
>>> +{
>>> +	const struct omap_rproc_dev_data *data;
>>> +	int delay;
>>> +
>>> +	data = of_device_get_match_data(&pdev->dev);
>>> +	if (!data)
>>> +		return -ENODEV;
>>
>> This check is done in omap_rproc_of_get_internal_memories() and
>> omap_rproc_get_boot_data().  I think it would be best to do it once at the top
>> of the probe() function and be done with it.
>>
>> That being said and as noted in a previous comment, I would push all tuneables
>> to the DT.  If the property is missing then things default to
>> DEFAULT_AUTOSUSPEND_DELAY.
>>
>>> +
>>> +	delay = data->autosuspend_delay;
>>> +
>>> +	return (delay > 0) ? delay : DEFAULT_AUTOSUSPEND_DELAY;
>>> +}
>>> +
>>>   static const char *omap_rproc_get_firmware(struct platform_device *pdev)
>>>   {
>>>   	const char *fw_name;
>>> @@ -953,6 +1154,11 @@ static int omap_rproc_probe(struct platform_device *pdev)
>>>   	}
>>>   
>>>   	init_completion(&oproc->pm_comp);
>>> +	oproc->autosuspend_delay = omap_rproc_get_autosuspend_delay(pdev);
>>> +	if (oproc->autosuspend_delay < 0) {
>>> +		ret = oproc->autosuspend_delay;
>>> +		goto free_rproc;
>>> +	}
>>>   
>>>   	oproc->fck = devm_clk_get(&pdev->dev, 0);
>>>   	if (IS_ERR(oproc->fck)) {
>>> @@ -994,6 +1200,8 @@ static int omap_rproc_remove(struct platform_device *pdev)
>>>   
>>>   static const struct dev_pm_ops omap_rproc_pm_ops = {
>>>   	SET_SYSTEM_SLEEP_PM_OPS(omap_rproc_suspend, omap_rproc_resume)
>>> +	SET_RUNTIME_PM_OPS(omap_rproc_runtime_suspend,
>>> +			   omap_rproc_runtime_resume, NULL)
>>>   };
>>>   
>>>   static struct platform_driver omap_rproc_driver = {
>>> -- 
>>> 2.17.1
>>>
>>> --
> 

--
Texas Instruments Finland Oy, Porkkalankatu 22, 00180 Helsinki. Y-tunnus/Business ID: 0615521-4. Kotipaikka/Domicile: Helsinki

Powered by blists - more mailing lists