#132 PWM becomes inaccurate after stop() and start()

[Andrew Scheller]

While investigating problems with the GPIOZero unit-tests, I eventually managed to slim the problem down to this simple single test-script:

import RPi.GPIO as GPIO
import time

# GPIOs 22 and 27 need to be connected with a jumper
input_gpio = 27
output_gpio = 22

GPIO.setmode(GPIO.BCM)
GPIO.setup(input_gpio, GPIO.IN, pull_up_down=GPIO.PUD_DOWN)
GPIO.setup(output_gpio, GPIO.OUT)

p = GPIO.PWM(output_gpio, 100)
p.start(0)

do_start_stop = False

def test_duty_cycles(i):
    print(" test_duty_cycles iteration %d" % i)
    if do_start_stop:
        p.stop()
        p.start(0)
    for dc in (40, 60):
        p.ChangeDutyCycle(dc)
        total = sum(GPIO.input(input_gpio) for i in range(20000))
        avg = (total / 20000.0) * 100
        print("  testing %.1f Hz, measured %.2f" % (dc, avg))
        if abs(avg - dc) > 10:
            print("*  over 10 Hz difference!")

try:
    for i in range(5):
        test_duty_cycles(i+1)
finally:
    GPIO.cleanup()

With do_start_stop set to False, then the script always runs fine (i.e. the measured frequency is always within 10Hz of the expected frequency. (Tested on up-to-date Raspbian Jessie on a RPi3 using both python and python3)
However when do_start_stop is set to True, then it never runs completely successfully e.g.

 test_duty_cycles iteration 1
  testing 40.0 Hz, measured 44.27
  testing 60.0 Hz, measured 59.58
 test_duty_cycles iteration 2
  testing 40.0 Hz, measured 40.81
  testing 60.0 Hz, measured 20.71
*  over 10 Hz difference!
 test_duty_cycles iteration 3
  testing 40.0 Hz, measured 60.60
*  over 10 Hz difference!
  testing 60.0 Hz, measured 64.00
 test_duty_cycles iteration 4
  testing 40.0 Hz, measured 42.34
  testing 60.0 Hz, measured 40.79
*  over 10 Hz difference!
 test_duty_cycles iteration 5
  testing 40.0 Hz, measured 24.90
*  over 10 Hz difference!
  testing 60.0 Hz, measured 27.54
*  over 10 Hz difference!

and it frequently errors with "Segmentation fault" at the end of the script (which seems to be more common with python3 than python2).

The GPIOZero unit-tests involve lots of starting and stopping of PWM, and these segmentation faults mean the PWM unit-tests all have to be disabled :-(

[Paul S Kilroy]

I can confirm this behavoir as well, the work around was to use ChangeDutyCycle(0),ChangeDutyCycle(X) instead of stop()/start(X). From looking in gdb, start/stop create and destroy threads and should only be called once on a PWM instance. Calling multiple triggers some thread-unsafe logic that in my case lead to segmentation faults as well.