On Friday 16 November 2012 09:38 AM, Al Viro wrote:

On Wed, Nov 07, 2012 at 10:47:53AM +0100, Vineet Gupta wrote:

quoted

+; When we land here, pt_regs have already been updated in-place correctly
+; A pointer to them is also passed by kernel_execve, we just need to make sure
+; that SP is set to point to them.
+ARC_ENTRY ret_from_kernel_execve
+	; Force SP to "normal" pt_regs just populated.
+	b.d   ret_from_system_call
+	mov   sp, r0

won't that splatter crap into regs->r0?  IOW, why not branch to
ret_from_exception here?

Yes it does - although I didn't notice the ill-effect since execve as an
API doesn't expect r0 to have a certain value (like fork for child).
Fixed in next revision.

quoted

+ARC_EXIT ret_from_kernel_execve

Another thing: why not fold that branch to ret_from_exception into the end of
ret_from_kernel_thread() (instead of calling sys_exit()), select
GENERIC_KERNEL_EXECVE and lose __ARCH_WANT_KERNEL_EXECVE.

If you noticed I only had 3 patches and didn't have the equivalent of
"switch to saner kernel_execve() semantics" which would achieve above
because the bug you elucidate to below was giving me grief in bring up
(and I couldn't spend enough time debugging it).

Actually, now that I look at your ret_from_kernel_thread...  How the hell
will it cope with kernel_thread() payload trying to return?  AFAICS, this
j.d [r1] will lose the return address, won't it? 

Exactly so. I was missing the equivalent of following

ARC_ENTRY ret_from_kernel_thread
	bl  @schedule_tail
	ld  r1, [sp, PT_r1]
-	j.d [r1]
+	jl.d [r1]
        ld  r0, [sp, PT_r0]
	j   @sys_exit
ARC_EXIT ret_from_kernel_thread

And since none of the bootup kernel threads took the return path I had a
working system - the problem only showed up when kernel_execve( )
started returning to exercise the broken path above.

 And while we are at it,
I would suggest passing callback and its argument via callee-saved registers -
makes for simpler life in ret_from_kernel_thread(), since switch_to() itself
will take care of loading those...

Absolutely - done that too !

Would you prefer to take a look at the updated patches now or will you
rather wait for v2 series which might take a week or two.

Thanks for taking the time to review this stuff.

-Vineet