Developer blog: marco

After a few days of having my house packed with my wife's family I am hoping to get back to hacking. During my absence Miod could no longer help himself and he bailed me out of the hppa rfork agony. I had it mostly working but it was not good enough for commit. The reason was pretty dang simple, I was doing the stack setup wrong. No idea what possessed me to setup the stack first and then call rfork; let me make some random excuse, I'm an idiot! The positive is that I now have a basic understanding of hppa assembly and ABI.

Here is what miod came up with:

ENTRY(rfork_thread, 0) 
        SYSCALL(rfork) 
        comb,<> r0, ret0, 1f 
        nop
 
        /* 
         * In child process: switch stack, invoke function, then exit.
         */ 
 
        /* 
         * PIC code expects 32 bytes of room available below sp.
         * Then the regular procedure invocation requires us to allocate
         * 64 bytes as well. 
         */ 
        copy    arg1, sp 
        ldo     0(sp), r3 
        stw,ma  r0, HPPA_FRAME_SIZE(sp)
        stw     r0, HPPA_FRAME_CRP(sp)
        stw     r0, HPPA_FRAME_PSP(sp)
 
        copy    arg3, arg0              /* arg */
        copy    arg2, t1 
        bl      $$dyncall, r31 
        copy    r31, rp 
 
        copy    r0, arg0 
        SYSCALL(threxit) 
 
1: 
        bv      r0(rp) 
        nop
 
EXIT(rfork_thread) 

A couple of comments here. General purpose register 0 (r0) is hardwired to always return 0x00. The hard part of this code was to call the thread. We had to emulate the $$dyncall (dynamic call) mechanism that is defined in the ABI and used by gcc. There is a downside to this code and that is the offset can not be more than a 256k range. Consensus is that this should be sufficient.

What I was doing wrong was that I was setting up the stackframe *before* rfork and fixing it back in the parent before returning to librthread. Another thing I was doing that was unnecessary was to create an additional stack frame; it didn't matter just an extra step.

A couple of people were involved in making this tiny piece of code work. Credit goes to Miod, Kettenis & Mickey. Amazing how simple things can take so long to get right.