s390: defective uses of va_arg in __debug_sprintf_event

s390: defective uses of va_arg in __debug_sprintf_event

[Joe Perches]

debug_sprintf_event calls __debug_sprintf_event
with a format and arguments.

There various types of arguments used in these
call, but __debug_sprintf_event uses va_arg
with only long as the type argument so random
errors could occur because the type and argument
are supposed to match.

    debug_entry_t *__debug_sprintf_event(debug_info_t *id, int level, char *string, ...)
    {
    	    [...]
    	    va_start(ap, string);
    	    curr_event->string = string;
    	    for (idx = 0; idx < min(numargs, (int)(id->buf_size / sizeof(long)) - 1); idx++)
    	    	    curr_event->args[idx] = va_arg(ap, long);
    	    va_end(ap);
    	    [...]
    }

from man va_arg

    va_arg()

    if type is not compatible with the type of the actual next argument
    (as promoted according to the default argument promotions),
    random errors will occur.

For instance, uses like:

arch/s390/kernel/perf_cpum_sf.c:919:    debug_sprintf_event(sfdbg, 6, "pmu_enable: es=%i cs=%i ed=%i cd=%i "
arch/s390/kernel/perf_cpum_sf.c-920-                        "tear=%p dear=%p\n", cpuhw->lsctl.es, cpuhw->lsctl.cs,
arch/s390/kernel/perf_cpum_sf.c-921-                        cpuhw->lsctl.ed, cpuhw->lsctl.cd,
arch/s390/kernel/perf_cpum_sf.c-922-                        (void *) cpuhw->lsctl.tear, (void *) cpuhw->lsctl.dear);

where the first 3 arguments are int but their type
as used by va_arg in __debug_sprintf_event is long
which could produce random errors.

Instead of adding complete format % decoding,
perhaps the easiest solution is to change all
the formats to use %lu or %ld and cast each
argument as appropriate.

And I found this when looking at another defect
in debug_sprintf_event where a %p<foo> extension
is unintentionally used via a string concatenation.
as pointed out by Rasmus Villemoes
---
 arch/s390/kernel/perf_cpum_sf.c | 4 +---
 1 file changed, 1 insertion(+), 3 deletions(-)

diff --git a/arch/s390/kernel/perf_cpum_sf.c b/arch/s390/kernel/perf_cpum_sf.c
index 1c9ddd7aa5ec..1c449a6f841a 100644
--- a/arch/s390/kernel/perf_cpum_sf.c
+++ b/arch/s390/kernel/perf_cpum_sf.c
@@ -212,9 +212,7 @@ static int realloc_sampling_buffer(struct sf_buffer *sfb,
 	 * the sampling buffer origin.
 	 */
 	if (sfb->sdbt != get_next_sdbt(tail)) {
-		debug_sprintf_event(sfdbg, 3, "realloc_sampling_buffer: "
-				    "sampling buffer is not linked: origin=%p"
-				    "tail=%p\n",
+		debug_sprintf_event(sfdbg, 3, "realloc_sampling_buffer: sampling buffer is not linked: origin=%p tail=%p\n",
 				    (void *) sfb->sdbt, (void *) tail);
 		return -EINVAL;
 	}

Re: s390: defective uses of va_arg in __debug_sprintf_event

[Martin Schwidefsky]

On Fri, 06 Apr 2018 12:08:43 -0700
Joe Perches <joe@perches.com> wrote:

> debug_sprintf_event calls __debug_sprintf_event
> with a format and arguments.
> 
> There various types of arguments used in these
> call, but __debug_sprintf_event uses va_arg
> with only long as the type argument so random
> errors could occur because the type and argument
> are supposed to match.
> 
>     debug_entry_t *__debug_sprintf_event(debug_info_t *id, int level, char *string, ...)
>     {
>     	    [...]
>     	    va_start(ap, string);
>     	    curr_event->string = string;
>     	    for (idx = 0; idx < min(numargs, (int)(id->buf_size / sizeof(long)) - 1); idx++)
>     	    	    curr_event->args[idx] = va_arg(ap, long);
>     	    va_end(ap);
>     	    [...]
>     }
> 
> from man va_arg
> 
>     va_arg()
> 
>     if type is not compatible with the type of the actual next argument
>     (as promoted according to the default argument promotions),
>     random errors will occur.
> 
> For instance, uses like:
> 
> arch/s390/kernel/perf_cpum_sf.c:919:    debug_sprintf_event(sfdbg, 6, "pmu_enable: es=%i cs=%i ed=%i cd=%i "
> arch/s390/kernel/perf_cpum_sf.c-920-                        "tear=%p dear=%p\n", cpuhw->lsctl.es, cpuhw->lsctl.cs,
> arch/s390/kernel/perf_cpum_sf.c-921-                        cpuhw->lsctl.ed, cpuhw->lsctl.cd,
> arch/s390/kernel/perf_cpum_sf.c-922-                        (void *) cpuhw->lsctl.tear, (void *) cpuhw->lsctl.dear);
> 
> where the first 3 arguments are int but their type
> as used by va_arg in __debug_sprintf_event is long
> which could produce random errors.

In principle you are right that the va_arg handling is not 100%
correct. It works though because the C ABI for s390x requires
that arguments are sign- or zero-extended to 64 bits by the caller
of a function. This is true for values passed in registers and for
the variable argument list.

-- 
blue skies,
   Martin.

"Reality continues to ruin my life." - Calvin.

Re: s390: defective uses of va_arg in __debug_sprintf_event

[Joe Perches]

On Mon, 2018-04-09 at 14:39 +0200, Martin Schwidefsky wrote:
> On Fri, 06 Apr 2018 12:08:43 -0700
> Joe Perches <joe@perches.com> wrote:
> 
> > debug_sprintf_event calls __debug_sprintf_event
> > with a format and arguments.
> > 
> > There various types of arguments used in these
> > call, but __debug_sprintf_event uses va_arg
> > with only long as the type argument so random
> > errors could occur because the type and argument
> > are supposed to match.
> > 
> >     debug_entry_t *__debug_sprintf_event(debug_info_t *id, int level, char *string, ...)
> >     {
> >     	    [...]
> >     	    va_start(ap, string);
> >     	    curr_event->string = string;
> >     	    for (idx = 0; idx < min(numargs, (int)(id->buf_size / sizeof(long)) - 1); idx++)
> >     	    	    curr_event->args[idx] = va_arg(ap, long);
> >     	    va_end(ap);
> >     	    [...]
> >     }
> > 
> > from man va_arg
> > 
> >     va_arg()
> > 
> >     if type is not compatible with the type of the actual next argument
> >     (as promoted according to the default argument promotions),
> >     random errors will occur.
> > 
> > For instance, uses like:
> > 
> > arch/s390/kernel/perf_cpum_sf.c:919:    debug_sprintf_event(sfdbg, 6, "pmu_enable: es=%i cs=%i ed=%i cd=%i "
> > arch/s390/kernel/perf_cpum_sf.c-920-                        "tear=%p dear=%p\n", cpuhw->lsctl.es, cpuhw->lsctl.cs,
> > arch/s390/kernel/perf_cpum_sf.c-921-                        cpuhw->lsctl.ed, cpuhw->lsctl.cd,
> > arch/s390/kernel/perf_cpum_sf.c-922-                        (void *) cpuhw->lsctl.tear, (void *) cpuhw->lsctl.dear);
> > 
> > where the first 3 arguments are int but their type
> > as used by va_arg in __debug_sprintf_event is long
> > which could produce random errors.
> 
> In principle you are right that the va_arg handling is not 100%
> correct. It works though because the C ABI for s390x requires
> that arguments are sign- or zero-extended to 64 bits by the caller
> of a function. This is true for values passed in registers and for
> the variable argument list.

Thanks.

Reference:
http://legacy.redhat.com/pub/redhat/linux/7.1/es/os/s390x/doc/lzsabi0.pdf
(found via your explanation)

It might be nice to explain that somewhere in
Documentation/s390 if it's not there already.

Re: s390: defective uses of va_arg in __debug_sprintf_event

[Petr Mladek]

On Fri 2018-04-06 12:08:43, Joe Perches wrote:
> And I found this when looking at another defect
> in debug_sprintf_event where a %p<foo> extension
> is unintentionally used via a string concatenation.
> as pointed out by Rasmus Villemoes
> ---
>  arch/s390/kernel/perf_cpum_sf.c | 4 +---
>  1 file changed, 1 insertion(+), 3 deletions(-)
> 
> diff --git a/arch/s390/kernel/perf_cpum_sf.c b/arch/s390/kernel/perf_cpum_sf.c
> index 1c9ddd7aa5ec..1c449a6f841a 100644
> --- a/arch/s390/kernel/perf_cpum_sf.c
> +++ b/arch/s390/kernel/perf_cpum_sf.c
> @@ -212,9 +212,7 @@ static int realloc_sampling_buffer(struct sf_buffer *sfb,
>  	 * the sampling buffer origin.
>  	 */
>  	if (sfb->sdbt != get_next_sdbt(tail)) {
> -		debug_sprintf_event(sfdbg, 3, "realloc_sampling_buffer: "
> -				    "sampling buffer is not linked: origin=%p"
> -				    "tail=%p\n",
> +		debug_sprintf_event(sfdbg, 3, "realloc_sampling_buffer: sampling buffer is not linked: origin=%p tail=%p\n",
>  				    (void *) sfb->sdbt, (void *) tail);
>  		return -EINVAL;
>  	}

This concatenation fix is correct. %p has many possible specifiers.
The code takes a simple approach and substitute "%ptail" with the
pointer value. See the following code in vsnprintf():

		case FORMAT_TYPE_PTR:
			str = pointer(fmt, str, end, va_arg(args, void *),
				      spec);
			while (isalnum(*fmt))
				fmt++;
			break;

We might debate if vsprintf() should handle this better. Anyway, there
is a missing space in the above debug_sprintf_event() call.

Joe, are you going to send this as a separate patch, your SOB, ...?

Best Regards,
Petr