Subject: Re: ksh lossage on sparc64
To: Todd Vierling <>
From: Richard Earnshaw <>
List: tech-toolchain
Date: 11/20/2000 18:09:23
> On Mon, 20 Nov 2000, Martin Husemann wrote:
> : What now happens on sparc64 is: the sigchld function modifies the struct
> : pointed to by "j" in this loop, returns, thus waking up the sigsuspend a
> : few lines below, continues to the "while" check and still has the old value
> : of j->state in a register, checks against that and continues the loop one
> : more time - thus blocking in "sigsuspend" again.
> Which means that this is a bug in gcc, not a bug in the ksh code.
> I'll see what I can find out about the codegen, and may submit a gcc PR.

It doesn't look particularly like a gcc bug to me.  If j->state is being 
updated from within a signal handler, then it must be declared volatile in 
the original declaration *and in all references to it*.  Adding casts is 
not good enough.

But in jobs.c we see:

struct job {
        Job     *next;          /* next job in list */
        int     job;            /* job number: %n */
        int     flags;          /* see JF_* */
        int     state;          /* job state */
        int     status;         /* exit status of last process */

So the declaration here, for starters, is invalid.

The same also applies to any other object that is being touched from 
within the signal handler.