Re: Kernel broken on processors without performance counters

[Borislav Petkov <bp@alien8.de>]

On Tue, Jul 21, 2015 at 06:12:15PM +0200, Peter Zijlstra wrote:
> Yes, if you start out false, you must be unlikely. If you start out
> true, you must be likely.
> 
> We could maybe try and untangle that if there really is a good use case,
> but this is the current state.
> 
> The whole reason this happened is because 'false' is like:
> 
> 
> 	...
> 	<nop>
> 1:
> 	...
> 
> 
> 
> label:
> 	<unlikely code>
> 	jmp 1b
> 
> 
> Where the code if out-of-line by default. The enable will rewrite the
> <nop> with a jmp label.

Btw, native_sched_clock() is kinda botched because of that, see below.

I'd want that RDTSC to come first with a NOP preceding it which can
become a JMP in case some idiotic CPU can't do RDTSC and needs to use
jiffies. Instead, we *unconditionally* jump to RDTSC which is WTF?! We
can just as well do a normal unlikely() without the static_key:

	.globl	native_sched_clock
	.type	native_sched_clock, @function
native_sched_clock:
	pushq	%rbp	#
	movq	%rsp, %rbp	#,
#APP
# 21 "./arch/x86/include/asm/jump_label.h" 1
	1:.byte 0x0f,0x1f,0x44,0x00,0
	.pushsection __jump_table,  "aw" 
	 .balign 8 
	 .quad 1b, .L122, __use_tsc 	#,
	.popsection 
	
# 0 "" 2
#NO_APP
	movabsq	$-4294667296000000, %rax	#, tmp118
	popq	%rbp	#
	imulq	$1000000, jiffies_64(%rip), %rdx	#, jiffies_64, D.28443
	addq	%rdx, %rax	# D.28443, D.28443
	ret
.L122:
#APP
# 118 "./arch/x86/include/asm/msr.h" 1
	rdtsc
# 0 "" 2

-- 
Regards/Gruss,
    Boris.

ECO tip #101: Trim your mails when you reply.
--