Re: [PATCH 3/7] PCI PM: Fix saving of device state inpci_legacy_suspend

[Michael Ellerman]

On Thu, 2009-02-05 at 12:29 +1100, Benjamin Herrenschmidt wrote:
> > > Yup, missed that change in the patch. In fact, I even missed the
> > > existence of a WARN_ONCE that takes those text arguments... baaaah. I
> > > suppose I should try to read more of lkml :-)
> > 
> > But don't expect to see them on powerpc, they never make it to the
> > console.
> 
> Allright, that's the problem with our implementation of WARN using a
> conditional trap, we don't carry all the variable arguments for the
> printf over.
> 
> I'm tempted to turn our implementation back to normal C code with an if
> and a branch out of line, but Michael seems to say that even with
> appropriate use of unlikely() etc... gcc decides to generate the worst
> possible code every single time ...
> 
> <rant>
> such as always inlining the whole printf mumbo jumbo and having the
> not-warning case branch over it, not -all- CPUs have good branch
> prediction dude, so the icache is going to scream but gcc folks know
> better of course....
> </rant>
> 
> Anyway, I'll see Michael what exactly the situation is here and if we
> can fix it a way or another.

Well I'm glad you listened to me when I told you not to fire off a rant
email before I had another look at it ;)

Because Arjan has patched the generic bug.h so that we at least get the
printk on powerpc, and then use our version of __WARN().

So you should see the message at least.


As far as the code gen, with gcc 4.3.1 (debian).

WARN looks like:

#define WARN(condition, format...) ({                                           \
        int __ret_warn_on = !!(condition);                              \
        if (unlikely(__ret_warn_on))                                    \
                __WARN_printf(format);                                  \
        unlikely(__ret_warn_on);                                        \
})


Called from:

int smp_request_message_ipi(int virq, int msg)
{
        int err;
...
        err = request_irq(virq, smp_ipi_action[msg], IRQF_DISABLED|IRQF_PERCPU,
                          smp_ipi_name[msg], 0);
        WARN(err < 0, "unable to request_irq %d for %s (rc %d)\n",
                virq, smp_ipi_name[msg], err);

        return err;
}


Comes out as:

c00000000002c9e8:       48 0a 88 c9     bl      c0000000000d52b0 <.request_irq>
c00000000002c9ec:       60 00 00 00     nop
c00000000002c9f0:       2f 83 00 00     cmpwi   cr7,r3,0
c00000000002c9f4:       7c 7f 1b 78     mr      r31,r3
c00000000002c9f8:       40 bc 00 20     bge+    cr7,c00000000002ca18 <.smp_request_message_ipi+0x84>
c00000000002c9fc:       7c bc e8 2a     ldx     r5,r28,r29
c00000000002ca00:       e8 7e 80 d0     ld      r3,-32560(r30)
c00000000002ca04:       7f 64 db 78     mr      r4,r27
c00000000002ca08:       7f e6 fb 78     mr      r6,r31
c00000000002ca0c:       48 5f 53 7d     bl      c000000000621d88 <.printk>
c00000000002ca10:       60 00 00 00     nop     
c00000000002ca14:       0f e0 00 00     twi     31,r0,0
c00000000002ca18:       38 21 00 a0     addi    r1,r1,160
c00000000002ca1c:       7f e3 fb 78     mr      r3,r31
c00000000002ca20:       e8 01 00 10     ld      r0,16(r1)
c00000000002ca24:       eb 61 ff d8     ld      r27,-40(r1)
c00000000002ca28:       eb 81 ff e0     ld      r28,-32(r1)
c00000000002ca2c:       eb a1 ff e8     ld      r29,-24(r1)
c00000000002ca30:       eb c1 ff f0     ld      r30,-16(r1)
c00000000002ca34:       eb e1 ff f8     ld      r31,-8(r1)
c00000000002ca38:       7c 08 03 a6     mtlr    r0
c00000000002ca3c:       4e 80 00 20     blr    


So it's jumping over the if body, rather than moving it out of the
straight line. It is hinting it right, so we should check if that helps
us at all.

cheers

-- 
Michael Ellerman
OzLabs, IBM Australia Development Lab

wwweb: http://michael.ellerman.id.au
phone: +61 2 6212 1183 (tie line 70 21183)

We do not inherit the earth from our ancestors,
we borrow it from our children. - S.M.A.R.T Person

Attachment: signature.asc
Description: This is a digitally signed message part