Error: symbol `__pastwait' is already defined

Error: symbol `__pastwait' is already defined

[Wu Zhangjin]

Hi, 

there is a guy reported a compiling problem in linux-2.6.29:

[...]
  CC      arch/mips/kernel/cpu-probe.o
{standard input}: Assembler messages:
{standard input}:3939: Error: symbol `__pastwait' is already defined
make[1]: *** [arch/mips/kernel/cpu-probe.o] Error 1
make: *** [arch/mips/kernel] Error 2

Seems I met this problem before, perhaps here is the reason:

arch/mips/kernel/cpu-probe.c:

void r4k_wait_irqoff(void)
{
    local_irq_disable();
    if (!need_resched())
        __asm__("   .set    push        \n"
            "   .set    mips3       \n"
            "   wait            \n"
            "   .set    pop     \n");
    local_irq_enable();
    __asm__("   .globl __pastwait   \n"
        "__pastwait:            \n");
    return;
}

there is a global symbol __pastwait defined at the end of
r4k_wait_irqoff, if r4k_wait_irqoff is called more than one time, the
__pastwait will be multi-defined. so, need to be fixed. does this fix
it?

arch/mips/kernel/cpu-probe.c:

void r4k_wait_irqoff(void)
{
    local_irq_disable();
    if (!need_resched())
        __asm__("   .set    push        \n"
            "   .set    mips3       \n"
            "   wait            \n"
            "   .set    pop     \n");
    local_irq_enable();
    return;
}
/* a dumy funciton for marking the end of r4k_wait_irqoff */
void __pastwait(void)
{
	;
}

but I am not sure the gcc compiler will tune the position of the
r4k_wait_irqoff and __pastwait or not, so seems not safe. perhaps we
should change something else instead.

perhaps we should tune the __pastwait solution directly, just spark it,
not look into it yet, seems __pastwait is only used here:

arch/mips/kernel/smtc.c:
smtc_send_ipi:

            if (cpu_wait == r4k_wait_irqoff) {
                tcrestart = read_tc_c0_tcrestart();
                if (tcrestart >= (unsigned long)r4k_wait_irqoff
                    && tcrestart < (unsigned long)__pastwait) {
                    write_tc_c0_tcrestart(__pastwait);
                    tcstatus &= ~TCSTATUS_IXMT;
                    write_tc_c0_tcstatus(tcstatus);
                    goto postdirect;
                }    
            } 

best wishes,
Wu Zhangjin 

Re: Error: symbol `__pastwait' is already defined

[Kevin D. Kissell]

Calling a function does not cause replication of its symbols.  That 
would happen if it were a macro, or an inline function, but not a simple 
global function, which r4k_wait_irqoff is supposed to be, since (at 
least the last time I worked with it), it is only called indirectly by 
having its address stored in the cpu_wait function pointer.  Either your 
compiler is doing something insane and replicating the function each 
time its address is taken (!), or someone has added another __pastwait 
symbol somewhere.

And you are correct that moving the symbol to another function risks 
breaking the functionality. Even if the compiler didn't reorder things - 
which you are correct to note that it might do - you would create a 
window during which the kernel would mistakenly believe that the CPU was 
in the interrupt-disabled wait state when in fact it had just fallen out 
of the loop and serviced an interrupt.  I don't think that would 
necessarily be fatal, but it would at least be inefficient.

          Regards,

          Kevin K.

Wu Zhangjin wrote:
> Hi, 
>
> there is a guy reported a compiling problem in linux-2.6.29:
>
> [...]
>   CC      arch/mips/kernel/cpu-probe.o
> {standard input}: Assembler messages:
> {standard input}:3939: Error: symbol `__pastwait' is already defined
> make[1]: *** [arch/mips/kernel/cpu-probe.o] Error 1
> make: *** [arch/mips/kernel] Error 2
>
> Seems I met this problem before, perhaps here is the reason:
>
> arch/mips/kernel/cpu-probe.c:
>
> void r4k_wait_irqoff(void)
> {
>     local_irq_disable();
>     if (!need_resched())
>         __asm__("   .set    push        \n"
>             "   .set    mips3       \n"
>             "   wait            \n"
>             "   .set    pop     \n");
>     local_irq_enable();
>     __asm__("   .globl __pastwait   \n"
>         "__pastwait:            \n");
>     return;
> }
>
> there is a global symbol __pastwait defined at the end of
> r4k_wait_irqoff, if r4k_wait_irqoff is called more than one time, the
> __pastwait will be multi-defined. so, need to be fixed. does this fix
> it?
>
> arch/mips/kernel/cpu-probe.c:
>
> void r4k_wait_irqoff(void)
> {
>     local_irq_disable();
>     if (!need_resched())
>         __asm__("   .set    push        \n"
>             "   .set    mips3       \n"
>             "   wait            \n"
>             "   .set    pop     \n");
>     local_irq_enable();
>     return;
> }
> /* a dumy funciton for marking the end of r4k_wait_irqoff */
> void __pastwait(void)
> {
> 	;
> }
>
> but I am not sure the gcc compiler will tune the position of the
> r4k_wait_irqoff and __pastwait or not, so seems not safe. perhaps we
> should change something else instead.
>
> perhaps we should tune the __pastwait solution directly, just spark it,
> not look into it yet, seems __pastwait is only used here:
>
> arch/mips/kernel/smtc.c:
> smtc_send_ipi:
>
>             if (cpu_wait == r4k_wait_irqoff) {
>                 tcrestart = read_tc_c0_tcrestart();
>                 if (tcrestart >= (unsigned long)r4k_wait_irqoff
>                     && tcrestart < (unsigned long)__pastwait) {
>                     write_tc_c0_tcrestart(__pastwait);
>                     tcstatus &= ~TCSTATUS_IXMT;
>                     write_tc_c0_tcstatus(tcstatus);
>                     goto postdirect;
>                 }    
>             } 
>
> best wishes,
> Wu Zhangjin 
>
>
>   

Re: Error: symbol `__pastwait' is already defined

[Ralf Baechle]

On Sat, Jun 13, 2009 at 06:25:14PM +0200, Kevin D. Kissell wrote:

> Calling a function does not cause replication of its symbols.  That  
> would happen if it were a macro, or an inline function, but not a simple  
> global function, which r4k_wait_irqoff is supposed to be, since (at  
> least the last time I worked with it), it is only called indirectly by  
> having its address stored in the cpu_wait function pointer.  Either your  
> compiler is doing something insane and replicating the function each  
> time its address is taken (!), or someone has added another __pastwait  
> symbol somewhere.
>
> And you are correct that moving the symbol to another function risks  
> breaking the functionality. Even if the compiler didn't reorder things -  
> which you are correct to note that it might do - you would create a  
> window during which the kernel would mistakenly believe that the CPU was  
> in the interrupt-disabled wait state when in fact it had just fallen out  
> of the loop and serviced an interrupt.  I don't think that would  
> necessarily be fatal, but it would at least be inefficient.

It depends on how gcc optimized the if statement.  Gcc might compile the
function as if it was written like this:

void r4k_wait_irqoff(void)
{
	local_irq_disable();
	if (need_resched())
		goto nowait;

	__asm__(
	"	.set	push		\n"
	"	.set	mips3		\n"
	"	wait			\n"
	"	.set	pop		\n");
	local_irq_enable();
	__asm__(
	"	.globl	__pastwait	\n"
	"__pastwait:			\n");
	return;

nowait
	local_irq_enable();
	__asm__(
	"	.globl	__pastwait	\n"
	"__pastwait:			\n");
}

Which isn't quite the brightest thing to do but perfectly legal.  As for
gcc follow the old motto trust is futile, suspicion breeds confidence.

  Ralf