Re: [GIT PATCH] more Driver core patches for 2.6.19

[Thomas Gleixner <tglx@linutronix.de>]

On Wed, 2006-12-13 at 12:58 -0800, Linus Torvalds wrote:
> In other words, I'd like to see code that uses this that is actually 
> _better_ than an in-kernel driver in some way.
> 
> For USB, the user-mode thing made sense. You have tons of random devices, 
> and the abstraction level is higher to begin with. Quite frankly, I simply 
> don't even see the same being true for something like this.

We started to work on this for industrial I/O devices. Many of them are
dual port memory based, others are dedicated chips for motion control or
field busses.

The design requires to have an in kernel stub driver with interrupt
handler which is capable to handle shared interrupts. User space
_cannot_ override an irq_disable(), it just has access to the chip
registers of the device, which is possible right now as well.

The risk, that such a driver stalls the kernel is exactly the same as
the risk you have with any other badly written driver.

This is a real world example of such a drivers interrupt handler:

/*
 * The chip specific portion of the interrupt handler. The framework code
 * takes care of userspace notification when we return IRQ_HANDLED
 */
static irqreturn_t sercos_handler(int irq, void *dev_id, struct pt_regs *reg)
{
        /* Check, if this interrupt is originated from the SERCOS chip */
        if (!(sercos_read(IRQ_STATUS) & SERCOS_INTERRUPT_MASK))
                return IRQ_NONE;

        /* Acknowledge the chip interrupts */
        sercos_write(IRQ_ACK1, SERCOS_INTERRUPT_ACK1);
        sercos_write(IRQ_ACK2, SERCOS_INTERRUPT_ACK2);

        return IRQ_HANDLED;
}

With a full kernel driver we need:

1. Interrupt handler
	check interrupt
	acknowledge interrupt
	copy data from/to chip into a kernel buffer
	wakeup user space task
2. read data from driver, which goes through copy to user
3. do calculations
4. write data to driver, which goes through copy from user

After changing the driver concept we have only:
1. Interrupt handler
	check interrupt
	acknowledge interrupt
	wakeup user space task
2. User space task handles the mmaped chip directly

The change gave a serious performance gain in the range of 20% after the
application was optimized for dealing with the chip directly.

There are tons of such exotic hardware devices out there, which now have
either a closed source driver or an out of tree patch with an horrible
amount of individual ioctl functions to get to the same point with less
performance.

> Btw: there's one driver we _know_ we want to support in user space, and 
> that's the X kind of direct-rendering thing. So if you can show that this 
> driver infrastructure actually makes sense as a replacement for the DRI 
> layer, then _that_ would be a hell of a convincing argument.

I did not look closely into that, but I think that it is a valid usage
candidate. The interface of graphic cards is user space mappable and it
probably needs some interrupt handling + notification mechanism as well
as the devices mentioned above.

	tglx