Re: [PATCH 07/10] IOCHK interface for I/O error handling/detecting

From: David Mosberger <>
Date: 2005-06-11 03:25:46
>>>>> On Fri, 10 Jun 2005 19:29:58 +0900, Hidetoshi Seto <> said:

  Hidetoshi> Hi David,
  Hidetoshi> David Mosberger wrote:
  >>>>>>> On Thu, 09 Jun 2005 21:58:26 +0900, Hidetoshi Seto <> said:
  Hidetoshi> +/*
  Hidetoshi> + * Some I/O bridges may poison the data read, instead of
  Hidetoshi> + * signaling a BERR.  The consummation of poisoned data
  Hidetoshi> + * triggers a local, imprecise MCA.
  Hidetoshi> + * Note that the read operation by itself does not consume
  Hidetoshi> + * the bad data, you have to do something with it, e.g.:
  Hidetoshi> + *
  Hidetoshi> + *	ld.8	r9=[r10];;	// r10 == I/O address
  Hidetoshi> + *	add.8	r8=r9,r9;;	// fake operation
  Hidetoshi> + */
  Hidetoshi> +#define ia64_poison_check(val)					\
  Hidetoshi> +{	register unsigned long gr8 asm("r8");			\
  Hidetoshi> +	asm volatile ("add %0=%1,r0" : "=r"(gr8) : "r"(val)); }
  Hidetoshi> +
  Hidetoshi> #endif /* CONFIG_IOMAP_CHECK  */

  >> I have only looked that this briefly and I didn't see off hand where you get
  >> the "r9=[r10]" sequence from --- I hope you're not relying on the compiler
  >> happening to generate this sequence!

  Hidetoshi> +static inline unsigned char
  Hidetoshi> +___ia64_readb (const volatile void __iomem *addr)
  Hidetoshi> +{
  Hidetoshi> +    unsigned char val;
  Hidetoshi> +
  Hidetoshi> +    val = *(volatile unsigned char __force *)addr;
  Hidetoshi> +    ia64_poison_check(val);
  Hidetoshi> +
  Hidetoshi> +    return val;
  Hidetoshi> +}

Ah, I see now what you're trying to do.  I think it's really a
machine-check barrier that you want there.

I'm doubtful whether this is the right approach, though: your
ia64_poison_check() will cause _every single_ readX() operation to
stall the CPU for 1,000+ cycles.  Why not define an explicit
iochk_barrier() instead?  Then you could do things like this:

	a = readb(X);
	b = readb(Y);
	c = readb(Z);
	iochk_barrier(a + b + c);

That is, if it's unimportant to know whether the read of X, Y, or Z
caused the MCA, you can amortize the cost of iochk_barrier() over 3

I'd probably make iochk_barrier() an out-of-line no-op assembly
routine.  The cost of two branches compared to stalling for hundreds
of cycles is rather trivial.

To unsubscribe from this list: send the line "unsubscribe linux-ia64" in
the body of a message to
More majordomo info at
Received on Fri Jun 10 13:26:25 2005

This archive was generated by hypermail 2.1.8 : 2005-08-02 09:20:39 EST