Is I2C device interface thread-safe on Linux?

Question

I have a single board computer with a I2C GPIO expander device, listed as /dev/i2c-0. In my multithreaded program I open() this character device and read/write to it on separate threads using i2c_smbus_write_byte_data and friends from the libi2c library, but I am not sure if that is thread-safe.

I have seen kernel space GPIO driver implementations that bind over the I2C character device and allow to interface with the device as if it was a GPIO device (pca953x driver) and they have internal mutexes within every gpio_set/get function. In another question people say i2c_smbus_* functions have internal mutexes so I cannot be sure.

Is this kind of concurrent access thread-safe?

void *thread_fn(void *arg){
 int fd = *(int*)arg;
 uint16_t d;
 for(uint16_t i=0;i<1000;i++) {
 if (i%2==0){
 d = i2c_smbus_read_word_data(fd, REG);
 } else {
 d |= 1 << (i%16);
 i2c_smbus_write_word_data(fd, REG, d);
 }
 }
 return NULL;
}
int main(void){
 int fd = open("/dev/i2c-0");
 pthread_t t;
 pthread_create(&t, NULL, thread_fn, &fd);
 for(uint16_t i=0;i<1000;i++)
 i2c_smbus_write_word_data(fd, REG, i);
 pthread_join(t, NULL);
 return 0;
}

• 0andriy: Why don’t you use the kernel driver for that expander which is definitely thread-safe?
• j-hc: @0andriy On the IO, I need a precise square wave output at high frequencies up to 8kHz. That is not feasible with the gpio driver (I measured with an oscilloscope) and I assume the reason is the constant locking/unlocking. On the contrary, when I interface with it as a I2C device, I can go up to high frequencies even with a userspace mutex and careful scheduling of the IO access. That's what made me think that maybe the mutex in the driver is redundant.
• 0andriy: No, the problem is the choosen (SW) tool. Linux kernel is General Purpose multi-tasking OS, doing what you want requires quite a high priority task to guarantee the desired latency. You may consider PREEMPT_RT kernel, but still the application needs to be properly written for that. What you have working is just by luck (mostly idling system, with likely a single user which also does nothing). That said, using in-kernel driver is a good start (and user space driver for I²C component like that in GP OS is quite controversial move).