On Wednesday 14 November 2007, Thomas Gleixner wrote:
> On Mon, 12 Nov 2007, David Brownell wrote:
> > > I'm still trying to understand what you've observed here. Is it the case
> > > that a single gpio operation went from 6.4 up to 11.2 usecs?
> >
> > That was a single bitbanged I2C bit transfer, with embedded udelay()s.
> > I believe that was four gpio operations, as summarized at the end of
> > that email above. Enabling preempt + debug increased the cost of
> > each GPIO call from whatever it was (reasonable) by 1.2 usecs.
>
> This raw lock change is just pampering over the design problem of the
> gpio lib:
>
> There is no need to check for every single access to a GPIO pin,
> whether the pin has a valid number and the chip, which provides access
> to the pin, is still registered.
As Haavard had noted. The "requested" flag is actually serving
as a longterm bit-level lock, which -- assuming well-behaved
callers, and no debug instrumentation -- obviates any need to
grab a spinlock in hot paths.
> Each driver, which wants to access a pin, needs to make sure that
>
> - the pin is available
> - the pin is associated to this driver
> - the chip reference count is incremented
>
> _before_ it starts to do anything with the pin. Once this is done the
> access to the pin is completely lock free except for the protection of
> the chip hardware itself.
That's what the gpio_request() call does, although it's using
something isomorphic to a refcount, not an actual refcount.
The key observation here is that we already *have* a bit which
is serving as a per-gpio lock. It's just never been viewed as
a lock before. :)
> The protection of the chip list can be converted to a mutex and
> does not need to be a spinlock at all.
No, we still need to use a spinlock to protect table changes.
The reason for that is briefly:
- gpio_request()/gpio_free() have so far been optional. Most
platforms implement them as NOPs, not all drivers use them.
(Having gpiolib in place should help change that ...)
- gpio_direction_input()/gpio_direction_output() implicitly
request the pins, if they weren't already requested.
- Those input/output direction-setting calls may be called
in IRQ contexts, which means (on non-RT kernels) no mutex.
So we're actually in good shape; just take out a bit of code
(or turn it into debugging instrumentation) and I don't think
anyone will complain about the locking any more.
- Dave
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to [EMAIL PROTECTED]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
