[PATCH] Finish the task 'Fix corefilename race'

[Konstantin Belousov]

On Thu, Mar 19, 2015 at 03:13:43PM +0800, Tiwei Bie wrote:
> Hi, Mateusz!
> 
> I have finished the task: Fix corefilename race [1].
> 
> Following is my patch:
> 
> ---
>  sys/kern/kern_sig.c | 22 ++++++++++++++++++++--
>  1 file changed, 20 insertions(+), 2 deletions(-)
> 
> diff --git a/sys/kern/kern_sig.c b/sys/kern/kern_sig.c
> index 58d9707..a1421cb 100644
> --- a/sys/kern/kern_sig.c
> +++ b/sys/kern/kern_sig.c
> @@ -3090,8 +3090,24 @@ static int compress_user_cores = 0;
>  #endif
>  
>  static char corefilename[MAXPATHLEN] = {"%N.core"};
> -SYSCTL_STRING(_kern, OID_AUTO, corefile, CTLFLAG_RWTUN, corefilename,
> -    sizeof(corefilename), "Process corefile name format string");
> +
> +static struct sx corefilename_lock;
> +SX_SYSINIT(corefilename_init, &corefilename_lock, "corefilename lock");
> +
> +static int
> +sysctl_kern_corefile(SYSCTL_HANDLER_ARGS)
> +{
> +	int error;
> +
> +	sx_xlock(&corefilename_lock);
> +	error = sysctl_handle_string(oidp, corefilename, MAXPATHLEN, req);
> +	sx_xunlock(&corefilename_lock);
> +
> +	return (error);
> +}
> +SYSCTL_PROC(_kern, OID_AUTO, corefile, CTLTYPE_STRING | CTLFLAG_RWTUN |
> +    CTLFLAG_MPSAFE, 0, 0, sysctl_kern_corefile, "A",
> +    "Process corefile name format string");
>  
>  /*
>   * corefile_open(comm, uid, pid, td, compress, vpp, namep)
> @@ -3120,6 +3136,7 @@ corefile_open(const char *comm, uid_t uid, pid_t pid, struct thread *td,
>  	name = malloc(MAXPATHLEN, M_TEMP, M_WAITOK | M_ZERO);
>  	indexpos = -1;
>  	(void)sbuf_new(&sb, name, MAXPATHLEN, SBUF_FIXEDLEN);
> +	sx_slock(&corefilename_lock);
>  	for (i = 0; format[i] != '\0'; i++) {
>  		switch (format[i]) {
>  		case '%':	/* Format character */
> @@ -3162,6 +3179,7 @@ corefile_open(const char *comm, uid_t uid, pid_t pid, struct thread *td,
>  			break;
>  		}
>  	}
> +	sx_sunlock(&corefilename_lock);
>  	free(hostname, M_TEMP);
>  	if (compress)
>  		sbuf_printf(&sb, GZ_SUFFIX);

So the race is between somebody setting the core path string and another
process coredumping, am I right ?

If you, could you try to reuse some existing lock for the task ? It is a
waste to have sx dedicated to the task, which is probably never used by
99% of the machines in the world.