Re: [RFC/PATCH, 2/4] readX_check() performance evaluation

[Linus Torvalds]

On Wed, 28 Jan 2004, Andi Kleen wrote:
>
> On Wed, 28 Jan 2004 12:28:56 -0800 (PST)
> Linus Torvalds <torvalds@xxxxxxxx> wrote:
> 
> > 
> > Alternatively, if you get a lot of information at MCE time (CPU that did
> > the access + some device data), just queue up the information in a per-CPU
> > queue. You don't have to worry about overflow - you can just drop if if 
> 
> That assumes that the access happened with preempt off ?

Yes. I assume you want _some_ locking anyway, at least within that 
particular device driver (you don't want to have an irq handler touch the 
device at the same time you are doing this thing), so any such spinlock 
would have disabled preemption anyway.

> That's fine if it's guaranteed that the MCE still happened inside 
> readl/writel. But if it's delayed longer for some reason there is no guarantee 
> that you can find back to the CPU that caused the fault.

Again, this is all depending on hardware implementation issues. It is 
entirely possible that "read_pcix_error()" has to do some kind of 
synchronization to make sure that any async operations have finished and 
all errors have been accounted for.

Again, this is the whole reason for doing the separate "clear/read"  
phases in error handling - exactly because hardware implementation may
make error handling fairly heavy-weight, so you don't want to do it on a
per-IO basis, but rather on a "per transaction" basis.

		Linus
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/