FW: cmpxchg hard lockup on AMD64 - ASUS(A8V-MX)

FW: cmpxchg hard lockup on AMD64 - ASUS(A8V-MX)

[chaitanya Huilgol]

I am seeing a hard lockup when the lock-free lifo 
implementation below is run on a AMD Athlon 64
and ASUS A8V-MX motherboard. GCC version is
gcc version 3.2.3 20030502 (Red Hat Linux 3.2.3-47.fc4)

I have run this code on Pentiums without any issue
till now. Also the same code works fine in userland. 
I fail to understand as to why the problem occurs
only in kernel mode. 

Thanks,
Chaitanya 



http://lalists.stanford.edu/lad/2000/Jul/0319.html

#ifdef __SMP__
#define LOCK "lock ; "
#else
#define LOCK ""
#endif

typedef struct cell {
       struct cell* link; /* next cell in the lifo */
       /*...*/ /* any data here */
} cell;

typedef struct lifo {
        volatile cell* top; /* top of the stack */
        volatile unsigned long cnt; /* used to avoid ABA problem */
} lifo;
void init(lifo* lf)
{
        lf->top = 0;
        lf->cnt = 0;
}

void push (lifo * lf, cell * cl)
{
        __asm__ __volatile__ (
                "# LFPUSH \n\t"
                "1:\t"
                "movl %2, (%1) \n"
                LOCK "cmpxchg %1, %0 \n\t"
                "jnz 1b \n\t"
                :
                :"m" (*lf), "r" (cl), "a" (lf->top)
                );
}

cell* pop (lifo * lf)
{
        cell* v=0;
        __asm__ __volatile__ (
                "# LFPOP \n\t"
                "testl %%eax, %%eax \n\t"
                "jz 20f \n"
                "10:\t"
                "movl (%%eax), %%ebx \n\t"
                "movl %%edx, %%ecx \n\t"
                "incl %%ecx \n\t"
                LOCK "cmpxchg8b %1 \n\t"
                "jz 20f \n\t"
                "testl %%eax, %%eax \n\t"
                "jnz 10b \n"
                "20:\t"
                :"=a" (v)
                :"m" (*lf), "a" (lf->top), "d" (lf->cnt)
                :"ecx", "ebx" );
        return v;
}

Re: FW: cmpxchg hard lockup on AMD64 - ASUS(A8V-MX)

[Andi Kleen]

"chaitanya Huilgol" <chaitanya@tidaldata.com> writes:


> {
>         __asm__ __volatile__ (
>                 "# LFPUSH \n\t"
>                 "1:\t"
>                 "movl %2, (%1) \n"
>                 LOCK "cmpxchg %1, %0 \n\t"
>                 "jnz 1b \n\t"
>                 :
>                 :"m" (*lf), "r" (cl), "a" (lf->top)

You don't tell gcc *lf is modified.

>                 );
> }
> 
> cell* pop (lifo * lf)
> {
>         cell* v=0;
>         __asm__ __volatile__ (
>                 "# LFPOP \n\t"
>                 "testl %%eax, %%eax \n\t"
>                 "jz 20f \n"
>                 "10:\t"
>                 "movl (%%eax), %%ebx \n\t"
>                 "movl %%edx, %%ecx \n\t"
>                 "incl %%ecx \n\t"

Nothing uses the incremented %ecx

>                 LOCK "cmpxchg8b %1 \n\t"
>                 "jz 20f \n\t"
>                 "testl %%eax, %%eax \n\t"
>                 "jnz 10b \n"
>                 "20:\t"
>                 :"=a" (v)
>                 :"m" (*lf),


And you don't tell *lf is modified again.

It could be just a miscompilation caused by your wrong asm constraints.

-Andi