Re: [PATCH] add pthread_setname_np

[Rich Felker <dalias@libc.org>]

On Fri, Sep 16, 2016 at 12:03:16PM +0300, Alexander Monakov wrote:
> Hi,
> 
> On Thu, 15 Sep 2016, Felix Janda wrote:
> > --- /dev/null
> > +++ b/src/thread/pthread_setname_np.c
> > @@ -0,0 +1,25 @@
> > +#include <fcntl.h>
> > +#include <string.h>
> > +#include <unistd.h>
> > +#include <sys/prctl.h>
> > +
> > +#include "pthread_impl.h"
> 
> This should bring in the declaration of pthread_setname_np from pthread.h by
> #defin'ing _GNU_SOURCE prior to inclusion.
> (today, this rule is not enforced with warnings and thus not always followed in
> musl, for example the recent pthread_tryjoin_np patch missed that as well)

Yep.

> > +int pthread_setname_np(pthread_t thread, const char *name)
> > +{
> > +	int fd, cs, status = 0;
> > +	char f[sizeof "/proc/self/task//comm" + 3*sizeof(int)];
> > +	size_t len;
> > +
> > +	if ((len = strnlen(name, 16)) > 15) return ERANGE;
> > +
> > +	if (thread == pthread_self())
> 
> This likely should use the static inline function __pthread_self()?

My preference when implementing nonstandard extension/junk functions
that are not performance-critical is to write them using nothing but
public APIs so that they're musl-agnostic and thereby
maintenance-free. Certainly saving a few bytes/cycles for the function
call here is not worthwhile so I'd stick with pthread_self().

> > +		return prctl(PR_SET_NAME, (unsigned long)name, 0, 0, 0) ? errno : 0;
> 
> First, prctl is declared as a variadic function, so zeros should be passed with
> the right type (0ul); passing fewer than 5 arguments will cause musl's
> implementation of prctl to invoke UB, since it always retrieves 4 variadic arguments.
> 
> Second, while I don't have a strong opinion on whether musl wants to use prctl
> here (my previous comment in this thread was based on the wrong idea that prctl
> would be sufficient in all cases - sorry about that), I think some tuning would
> be nice in case the decision is to use prctl. So, either
> 
>   return -__syscall(SYS_prctl, PR_SET_NAME, name);
> 
> (not sure if Rich would like it), or

See above. This would be smaller but if we (or someone else using the
code) wanted to change the conventions of how __syscall is used, this
would require them to dig info nonstandard functions (which don't
_need_ to be using __syscall) in addition to all the standard ones
(which might need to do this for namespace reasons, etc.). So I'd
prefer just using the public interface.

>   if (thread == pthread_self()) {
>     status = prctl(...);
>   } else {
>     ...
>     fd = status = open(f, O_WRONLY);
>     if (fd >= 0) {
>       status = write(fd, name, len);
>       close(fd);
>     }
>     pthread_setcancelstate(cs, 0);
>   }
>   return status ? errno : 0;
> 
> > +	snprintf(f, sizeof f, "/proc/self/task/%d/comm", thread->tid);
> > +	pthread_setcancelstate(PTHREAD_CANCEL_DISABLE, &cs);
> > +	if ((fd = open(f, O_WRONLY)) < 0 || write(fd, name, len) < 0) status = errno;
> > +	if (fd >= 0) close(fd);
> > +	pthread_setcancelstate(cs, 0);
> > +	return status;
> > +}

I don't have a strong opinion on how this part is structured.

Rich