Re: Random panic in load_balance() with 3.16-rc

[Linus Torvalds <torvalds@linux-foundation.org>]

On Thu, Jul 24, 2014 at 6:25 PM, Michel Dänzer <michel@daenzer.net> wrote:
>
> Attached is fair.s from Debian gcc 4.8.3-5. Does that look better? I'm
> going to try reproducing the problem with a kernel built by that now.

This looks better. For roughly that same code sequence it does
(ignoring the debug line and cfi information):

        subq    $184, %rsp      #,
        movq    (%r12), %rax    # sd_22(D)->parent, sd_parent
        movl    %edi, -156(%rbp)        # this_cpu, %sfp
        movl    %ecx, -160(%rbp)        # idle, %sfp
        movq    %r8, -184(%rbp) # continue_balancing, %sfp
        movq    %rax, -176(%rbp)        # sd_parent, %sfp
        movq    $load_balance_mask, %rax        #, tcp_ptr__
#APP
        add %gs:this_cpu_off, %rax      # this_cpu_off, tcp_ptr__
#NO_APP

so it updates the stack pointer before any spills, and it also doesn't
spill that constant value.

I still have no idea why it does the 4-byte rep stosl/movsl thing, but
that's a whole separate guessing game and might have something to do
with the fact that you do CONFIG_CC_OPTIMIZE_FOR_SIZE and the 4-byte
form is one byte smaller.

I'm a big believer in not blowing up the I$ footprint, and I have to
admit to pushing that myself a few years ago, but gcc does some rather
bad things with '-Os', so it's not actually suggested for the kernel
any more. I wish there was some middle ground model that cared about
size, but not to exclusion of everything else. The string instructions
are not good for performance when it's a compile-time known small
size.

                 Linus