Re: Hang on large copy_from_user with PREEMPT_NONE

[Borislav Petkov <bp@alien8.de>]

On Mon, Apr 06, 2015 at 10:51:42AM -0400, Sasha Levin wrote:
> > Btw, what's happening to that user pointer in %rdi, is it kosher?
> 
> Probably not. Trinity passes random addresses into that syscall, so it's probably one of it's magic
> values.

So what is the expectation? If that pointer goes away from under us,
userspace is supposed to get some signal instead of hanging?

> I've attached dmesg with debug-alternative, and my .config. vmlinux is pretty
> large (1GB compressed) so hopefully the config would be enough - let me know
> if it isn't.

Thanks, I'll give it a try and do some staring tomorrow. Still a holiday
today here :-)

> > Btw, that zerorest is being tested on every loop iteration! AFAICT,
> > 
> > 	if (!zerorest) {
> > 		clac();
> > 		return len;
> > 	}
> > 
> > before the loop should be nicer. Or am I missing something?
> 
> Looks good to me. I suspect the compiler optimizes that anyway though.

Yeah, doesn't look like it. See the

	e:   85 c9                   test   %ecx,%ecx

?

That's:

---
	testl	%esi, %esi	# __pu_err
	je	.L34	#,
.L21:
	testl	%ecx, %ecx	# zerorest		<---
	je	.L33	#,
	movl	%edx, %eax	# len, len
	xorl	%r8d, %r8d	# tmp194
	jmp	.L27	#
	.p2align 4,,7
	.p2align 3
.L35:
	decl	%eax	# len
	movq	%rsi, %rdi	# to, to
	je	.L33	#,
.L27:
	leaq	1(%rdi), %rsi	#, to
---

and my asm is not so different from yours so that I can match your asm
with the .s output here.

Anyway, that's a different story, I'll do a fix for it.

-- 
Regards/Gruss,
    Boris.

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