Re: [PATCH] x86 page fault handler not interrupt safe

[Linus Torvalds]

On Mon, 7 May 2001, Alan Cox wrote:
> 
> I dont see that we will get a valid value in both cases.
> 
> 	get_user
> 		fault - set %cr2
> 					IRQ
> 					vmalloc
> 					fault
> 						set %cr2
> 						fixup runs
> 					end IRQ
> 		cr2 is corrupt

Wrong. "%cr2" is _not_ "corrupt". It has a well-defined value.

So what happens is

	get_user (or user-mode access)
		fault - set %cr2 to fault1
			irq
			vmalloc fault - set %cr2 to fault2
				fixup runs, iret
			irq runs, iret
		%cr2 is still %fault2
		vmalloc fault - nothing to do
		"false fixup" runs, iret
	get_user (or user-mode access)
		fault - set %cr2 to fault1
		... get the right behaviour now ...
		
> There are a whole set of races with the vmalloc fixups.

As far as I can tell, there are no races anywhere. Just silly bugs that
are hard to see.

		Linus