Re: [RFC PATCH 0/9] iio: Fix ABBA deadlock in inv-mpu6050

From: Daniel Baluta
Date: Fri Feb 26 2016 - 10:52:17 EST

On Mon, Feb 22, 2016 at 1:17 AM, Wolfram Sang <wsa@xxxxxxxxxxxxx> wrote:
> On Thu, Feb 18, 2016 at 05:53:05PM +0200, Daniel Baluta wrote:
>> Sending this as an RFC because I don't know if style fixes are appropriate
>> for this driver and also not sure if deadlock fix is the best solution.
>> I2C people should only look at patches 8/9 and 9/9.
>> The mpu6050 accel+gyro combo has an auxiliary I2C bus, allowing for
>> an auxiliary sensor to be connected (eg. a magnetometer).
>> The mpu can either act as an I2C master (functionality not currently
>> implemented in the driver) or it can use a bypass multiplexer which
>> directly connects the auxiliary I2C bus pins to the main I2C bus pins [1]
>> Currently the driver implements the bypass mode via an I2C mux.
>> This patchset fixes a deadlock in inv-mpu6050 IMU which happens when
>> magnetometer (on auxiliary I2C bus) runs in parallel with accel or gyro.
>> First 7 patches do some cleanup in order to make INV MPU6050 code easier
>> to read.
>> Patch number 8 allows passing NULL select callbacks to i2c_mux_master_xfer
>> and i2c_mux_smbus_xfer.
>> Patch number 9 actually fixes the deadlock.
> We recently had a bigger patch series fixing locking problems related to
> muxes. I sadly didn't have the time to review it. Can you have a look if
> it helps your case?

Hi Wolfram,

Tested this and the deadlock is still there :(.

It can be easily reproduced with following patch which enlarges
the race window:

iff --git a/drivers/iio/imu/inv_mpu6050/inv_mpu_core.c
index a3f5070..279d6e0 100644
--- a/drivers/iio/imu/inv_mpu6050/inv_mpu_core.c
+++ b/drivers/iio/imu/inv_mpu6050/inv_mpu_core.c
@@ -232,6 +232,7 @@ inv_mpu6050_read_raw(struct iio_dev *indio_dev,
ret = IIO_VAL_INT;
result = 0;
+ msleep(12000);
if (!st->chip_config.enable) {
result = inv_mpu6050_set_power_itg(st, true);
if (result)

using the following commands:

$ cat /sys/bus/iio/devices/iio:device0/in_accel_x_raw
$ cat /sys/bus/iio/devices/iio:device1/in_magn_x_raw

Off topic: I'm very curious why I didn't get a lockdep splat while running this,
might be because adapter lock is a rt_mutex.