Going low power in freertos app

[eyalasko]

Hi, 

I have an app based on freertos_hrs example (nRF52832, SDK14.2)

I'm trying to get it into low power(sleep) mode, however the idle task keep on running and the system consumes ~4.5mA constantly (ble is off most of the time)

I tried the following:

• Setting #define configUSE_TICKLESS_IDLE   1 in freeRTOSConfig.h but is does not seem to help.
• Disabled NRF_Log options in sdk_config.h as several posts suggested
• Verified that none of my app's tasks is running (All are blocked), neither is the timer task. The only task that seem to run is the idle task.
• Verified that the idle task handler portTASK_FUNCTION() (part of freertos port) do get to the line portSUPPRESS_TICKS_AND_SLEEP() that eventually calls sd_app_evt_wait()

• I tried to forcefully call sd_app_evt_wait() from within  the applications idle hook (vApplicationIdleHook(), #define configUSE_IDLE_HOOK  1)
  It does lower the consumed power to few uA, but I keep getting hard faults when ble activity is encountered (not sure exactly what is the cause) - e.g. advertising goes well but when connection is established ble_evt_handler() crashes in case BLE_GAP_EVT_CONNECTED block

Reading the doc it seems that implementing the 1st bullet above should take care of lowering the power but apparently something else is not set correctly.

Any help will be highly appreciated

Thanks in advance

[Susheel Nuguru]

which device are you running this on? 

instead of calling sd_app_evt_wait in vApplicationIdleHook, can you try to call below

                do{
                    __WFE();
                } while (0 == (NVIC->ISPR[0] | NVIC->ISPR[1]));
                

instead of 

            if (nrf_sdh_is_enabled())
            {
                uint32_t err_code = sd_app_evt_wait();
                APP_ERROR_CHECK(err_code);
            }

[eyalasko]

Hi Susheel, 

Thanks for your response. 

your 2nd code snippet is in the 'heart' of vPortSuppressTicksAndSleep() in port_cmsys_systick.c which is part of the freertos package (sort of 'system file' in higher OSes)

Do you mean to replace this in place (vPortSuppressTicksAndSleep()) or add the 1st snippet to vApplicationIdleHook() which is in the application file (main.c)

[eyalasko]

and its getting weirder …

On my post yesterday I've mentioned a 'binary search'. What I did was to run my app with the ble stuff disabled - idle task hook was not triggered. Than I've re-enabled the ble stuff and the comb in the photo above appeared.

Today I did the opposite - started with the ble stuff enabled and disabled all the application code. Everything looks fine.

Than I've started to add my code - first all initialization code (still not creating tasks). Everything still looks OK.

Now I'm adding the first createTask() statement and the pulse train comes out again.
I carefully comment out larger and larger sections of the 'offending task' code till I get to a single line in a function that is called as part of the task's initialization (before the while(1) loop)
This is the "suspicious line":

thisObj->_.numSamplesInBlock = (uint32_t)(TSAMPLE2SEC(thisObj->cfg->tInterval) * ACC_SAMPLE_RATE_Hz);

this long expression is merely a floating point calculation casted to uint32.

I've replaced this line with

thisObj->_.numSamplesInBlock = 7800;

and surprisingly (for me :) the system behave nice. No pulse train from the idle task hook !

Now I start to play around and 'strip down' the complex term above. nothing seems to help.
Just out of curiosity I added a 'dummy' float calculation -

 float val = (float)300000; 
 val *= (float)1e-3; 
 thisObj->_.numSamplesInBlock = 7800;

and the 'beast' rises again

So here I'm standing - a naïve floating point calculation in the initialization code in one of my tasks (I have ton of places with floating calculation) seem to drive the system crazy.

Furthermore, I've deleted the 2 offending lines (float val ..& val *= ...) and copied them to main.c , just before nrf_sdh_freertos_init() and the system behaves nicely (no pulses)

Now I'm not sure if this plethora of information is helpful or not, but I'm scratching my head (or better said banging).

It has been a long post. phew...

btw, same behavior when pulses are driven from vPortSuppressTicksAndSleep()

[Susheel Nuguru]

This seems to be due to the fact that we did not handle any floating point exceptions.

We can do one experiment

Can you implement the FPU_IRQHandler and see if the power consumption goes low. Add below to your code and make sure that your vector table has FPU_IRQHandler in its FPU IRQ address

void FPU_IRQHandler(void)
{
    uint32_t *fpscr = (uint32_t *)(FPU->FPCAR+0x40);
    (void)__get_FPSCR();

    *fpscr = *fpscr & ~(FPU_EXCEPTION_MASK);
}

[eyalasko]

Hi Susheel, 

Do you feel that this may be the problem ?

I can see that PWR_MGMT_FPU_SLEEP_PREPARE macro (that seems to workaround PAN87) is called in nrf_pwr_mgmt_run() just before sd_app_evt_wait() in 'standard' app but NOT in freertos based app....

Will update as soon as I get back to the office...

Thanks

btw, 

I feel silly to ask, but how do I register an ISR? (so it appears in the vector table)

I only used event handlers in Nordic SDK so far. 

[Susheel Nuguru]

eyalasko said:

how do I register an ISR? (so it appears in the vector table)

 make sure that that there is an entry in startup.s files in your project where we populate the vector table and it DOES NOT have default do nothiing logic with it. Since those entry in startup files are weak declaration any new declaration of the same function will override those implemenation.

[eyalasko]

Hi Susheel, 

For the sake of other fellow developers here is a summary of the issue and the solution that worked for me.

The problem stems from the FPU generating an exception when doing floating point math. This interrupt can not be masked out as described in Errata 87. 

As the FPU interrupt is pending, the 'go to sleep' call sd_app_evt_wait() which DO get called when freertos goes idle (in tickles mode), does practically nothing as it wakes up immediately due to the pending FPU interrupt.

To tackle this you should clear pending interrupt before calling sd_app_evt_wait()

There are 3 ways  to do so:

1. Clear the FPU interrupt immediately by hooking on FPU_IRQHandler() as described in your post above; OR
2. Clear it before idle task calls sd_app_evt_wait() the way described in Errata 87 in port_cmsis_systick.c in vPortSuppressTicksAndSleep() just before calling sd_app_evt_wait(); OR
3. Mimic the powering down sequence of non-freertos applications. This is done in nrf_power_mgmt_run() in nrf_pwr_mgmt.c

I choose to go with option 3 and add the PWR_MGMT_FPU_SLEEP_PREPARE() macro (with its 'implementation') to vPortSuppressTicksAndSleep() just before calling sd_app_evt_wait().

It worked extremely well for me, hopefully for other fellow developers.

I would highly recommend Nordic to implement this/similar fix on future version of the SDK\freertos examples

THANK YOU for your ongoing support.

[Susheel Nuguru]

Thanks for coming back with good suggestions. We really value users like you who do care for others and help them to avoid similar problems. One upvote to you from my side :)

[Jens Hauke]

This code fragment uses approach 3 of the verified answer, but without the need of editing SDK source code. It injects the workaround with the help of a "configPRE_SLEEP_PROCESSING" macro in config/FreeRTOSConfig.h:

/* Fix for Errata #87 nrf52832 */                                       \
#define configPRE_SLEEP_PROCESSING(xMIdleTime) if ( xMIdleTime > 0 ) {  \
        __set_FPSCR(__get_FPSCR() & ~(0x0000009F));                     \
        (void) __get_FPSCR();                                           \
         NVIC_ClearPendingIRQ(FPU_IRQn);                                \
}

Thanks for the summarizing answer!