Question

I’m writing a linux kernel module that provides a (virtual) block device (so no actual hardware IO is performed).

Currently I’m using spin_lock_irqsave / spin_unlock_irqrestore to handle locks.

There is only one function running in non-process context, and this is the make_request function of the block device.

Would it be safe to use spin_lock_bh / spin_unlock_bh to handle locks? I guess that simple spin_lock is insufficient, since make_request is not run by a process (is this correct?).

Thanks in advance.

Answer 1

Your make_request function actually does run in process context. The only caller of q->make_request_fn is the block layer’s generic_make_request(), which assumes process context (eg look at its use of current). And as another example, drivers/md/md.c has md_make_request() which explicitly sleeps on a waitqueue.

So you are completely safe using plain spin_lock()/spin_unlock() as long as all the rest of your code is process context too.