Re: lockdep and threaded IRQs (was: ...)

On Friday 27 February 2009, Andrew Morton wrote:

On Fri, 27 Feb 2009 15:32:04 -0800
Andrew Morton [off-list ref] wrote:

quoted

Why does this function:

static irqreturn_t powerbutton_irq(int irq, void *dev_id)
{
	... threaded irq handler body elided ...
}

Which is connected up via this statement:

	err = request_irq(irq, powerbutton_irq,
			IRQF_TRIGGER_FALLING | IRQF_TRIGGER_RISING,
			"twl4030-pwrbutton", NULL);

reenable local interrupts?

Because threaded IRQ handlers are, well, threaded.  And all
the twl4030 IRQ handlers are threaded -- must be.  But when
CONFIG_LOCKDEP is enabled, it goofs such handlers ... as well
as a bunch of other perfectly functional driver code.

In the absense of the lockdep IRQF_DISABLED goofage, the IRQs
are properly dispatched -- IRQs stay enabled while these
handlers run, all the relevant locking invariants are obeyed.

ah, OK, twl4030_i2c_read_u8() does i2c I/O.

Can't do that. 

Threaded IRQ handlers *can* do that.  That's the point.

Now, if you were to say "keep waiting a few more years
until some threaded IRQ framework finally merges" ...
the question comes up, "What to do in the meanwhile".
(Ditto, "well, we've been waiting a long time now to
see those threaded IRQs, what's up with them?")

"Nothing" is not an option.  The "something" being done
here is a reasonably clean approach, and doesn't call
for any surgery to kernel/irq/* ... the *only* problem
is the lockdep bug, which causes trouble for a variety
of other drivers too.

If some random process currently holds 
mutex_lock(&twl->xfer_lock) and an interrupt occurs then this interrupt
handler will try to acquire mutex_lock(&twl->xfer_lock).  Deadlock.

No, no, no.  *THREADED IRQ HANDLER* at work here.  Bzzt.

Threaded IRQ handler.

The relevant mutexes are *never* accessed outside of a
thread context.  Not by this code.  Not by any other code.