[RFC PATCH 2/2] powerpc: nop trap instruction after WARN_ONCE fires

[David Laight]

From: Nicholas Piggin
> Sent: 23 September 2022 16:42
>
> WARN_ONCE and similar are often used in frequently executed code, and
> should not crash the system. The program check interrupt caused by
> WARN_ON_ONCE can be a significant overhead even when nothing is being
> printed. This can cause performance to become unacceptable, having the
> same effective impact to the user as a BUG_ON().
> 
> Avoid this overhead by patching the trap with a nop instruction after a
> "once" trap fires. Conditional warnings that return a result must have
> equivalent compare and branch instructions after the trap, so when it is
> nopped the statement will behave the same way. It's possible the asm
> goto should be removed entirely and this comparison just done in C now.
> 
> XXX: possibly this should schedule the patching to run in a different
> context than the program check.

I'm pretty sure WARN_ON_ONCE() is valid everywhere printk() is allowed.
In many cases this means you can't call mutex_enter().

So you need a different scheme.

	David

-
Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK
Registration No: 1397386 (Wales)