[PATCH RFC] add pthread_setname_np

[PATCH RFC] add pthread_setname_np

[Felix Janda]

gdb's "info threads" displays the thread name.
---
Should this function care about interrupted syscalls?
---
 include/pthread.h               |  1 +
 src/thread/pthread_setname_np.c | 21 +++++++++++++++++++++
 2 files changed, 22 insertions(+)
 create mode 100644 src/thread/pthread_setname_np.c

diff --git a/include/pthread.h b/include/pthread.h
index 3d2e0c4..94ef919 100644
--- a/include/pthread.h
+++ b/include/pthread.h
@@ -214,6 +214,7 @@ struct cpu_set_t;
 int pthread_getaffinity_np(pthread_t, size_t, struct cpu_set_t *);
 int pthread_setaffinity_np(pthread_t, size_t, const struct cpu_set_t *);
 int pthread_getattr_np(pthread_t, pthread_attr_t *);
+int pthread_setname_np(pthread_t, const char *);
 int pthread_tryjoin_np(pthread_t, void **);
 int pthread_timedjoin_np(pthread_t, void **, const struct timespec *);
 #endif
diff --git a/src/thread/pthread_setname_np.c b/src/thread/pthread_setname_np.c
new file mode 100644
index 0000000..b7f7a4b
--- /dev/null
+++ b/src/thread/pthread_setname_np.c
@@ -0,0 +1,21 @@
+#include <fcntl.h>
+#include <string.h>
+#include <unistd.h>
+
+#include "pthread_impl.h"
+
+int pthread_setname_np(pthread_t thread, const char *name)
+{
+	int fd, status = 0;
+	char f[sizeof "/proc/self/task//comm" + 7];
+	ssize_t len;
+
+	if ((len = strlen(name)) > 15) return ERANGE;
+
+	snprintf(f, sizeof f, "/proc/self/task/%d/comm", thread->tid);
+	if ((fd = open(f, O_WRONLY)) < 0) return errno;
+	if (write(fd, name, len) < 0) status = errno;
+
+	close(fd);
+	return status;
+}
-- 
2.7.3

Re: [PATCH RFC] add pthread_setname_np

[Rich Felker]

On Wed, Sep 14, 2016 at 11:02:16PM -0400, Felix Janda wrote:
> gdb's "info threads" displays the thread name.
> ---
> Should this function care about interrupted syscalls?

Unless it's specified not to return EINTR I'm fine with it returning
EINTR.

> ---
>  include/pthread.h               |  1 +
>  src/thread/pthread_setname_np.c | 21 +++++++++++++++++++++
>  2 files changed, 22 insertions(+)
>  create mode 100644 src/thread/pthread_setname_np.c
> 
> diff --git a/include/pthread.h b/include/pthread.h
> index 3d2e0c4..94ef919 100644
> --- a/include/pthread.h
> +++ b/include/pthread.h
> @@ -214,6 +214,7 @@ struct cpu_set_t;
>  int pthread_getaffinity_np(pthread_t, size_t, struct cpu_set_t *);
>  int pthread_setaffinity_np(pthread_t, size_t, const struct cpu_set_t *);
>  int pthread_getattr_np(pthread_t, pthread_attr_t *);
> +int pthread_setname_np(pthread_t, const char *);
>  int pthread_tryjoin_np(pthread_t, void **);
>  int pthread_timedjoin_np(pthread_t, void **, const struct timespec *);
>  #endif
> diff --git a/src/thread/pthread_setname_np.c b/src/thread/pthread_setname_np.c
> new file mode 100644
> index 0000000..b7f7a4b
> --- /dev/null
> +++ b/src/thread/pthread_setname_np.c
> @@ -0,0 +1,21 @@
> +#include <fcntl.h>
> +#include <string.h>
> +#include <unistd.h>
> +
> +#include "pthread_impl.h"
> +
> +int pthread_setname_np(pthread_t thread, const char *name)
> +{
> +	int fd, status = 0;
> +	char f[sizeof "/proc/self/task//comm" + 7];

Where does 7 come from? I don't think it's correct. 3*sizeof(int) is
the normal lazy bound we use for %d's.

> +	ssize_t len;
> +
> +	if ((len = strlen(name)) > 15) return ERANGE;

Storing the result of strlen in a ssize_t is odd. It works because no
object can be larger than SIZE_MAX/2, but it's ugly. Also using
strnlen would be preferable if there's a limit (so as not to read
unboundedly long strings only to reject them).

> +
> +	snprintf(f, sizeof f, "/proc/self/task/%d/comm", thread->tid);
> +	if ((fd = open(f, O_WRONLY)) < 0) return errno;
> +	if (write(fd, name, len) < 0) status = errno;
> +
> +	close(fd);

This needs blocking of the cancellation points, I think.

Rich

Re: [PATCH RFC] add pthread_setname_np

[Felix Janda]

Rich Felker wrote:

Thanks for the review!

> On Wed, Sep 14, 2016 at 11:02:16PM -0400, Felix Janda wrote:
> > gdb's "info threads" displays the thread name.
> > ---
> > Should this function care about interrupted syscalls?
> 
> Unless it's specified not to return EINTR I'm fine with it returning
> EINTR.

ok. Is it possible that write() is interrupted with partial data
written (so that it won't return -1)?

> > ---
> >  include/pthread.h               |  1 +
> >  src/thread/pthread_setname_np.c | 21 +++++++++++++++++++++
> >  2 files changed, 22 insertions(+)
> >  create mode 100644 src/thread/pthread_setname_np.c
> > 
> > diff --git a/include/pthread.h b/include/pthread.h
> > index 3d2e0c4..94ef919 100644
> > --- a/include/pthread.h
> > +++ b/include/pthread.h
> > @@ -214,6 +214,7 @@ struct cpu_set_t;
> >  int pthread_getaffinity_np(pthread_t, size_t, struct cpu_set_t *);
> >  int pthread_setaffinity_np(pthread_t, size_t, const struct cpu_set_t *);
> >  int pthread_getattr_np(pthread_t, pthread_attr_t *);
> > +int pthread_setname_np(pthread_t, const char *);
> >  int pthread_tryjoin_np(pthread_t, void **);
> >  int pthread_timedjoin_np(pthread_t, void **, const struct timespec *);
> >  #endif
> > diff --git a/src/thread/pthread_setname_np.c b/src/thread/pthread_setname_np.c
> > new file mode 100644
> > index 0000000..b7f7a4b
> > --- /dev/null
> > +++ b/src/thread/pthread_setname_np.c
> > @@ -0,0 +1,21 @@
> > +#include <fcntl.h>
> > +#include <string.h>
> > +#include <unistd.h>
> > +
> > +#include "pthread_impl.h"
> > +
> > +int pthread_setname_np(pthread_t thread, const char *name)
> > +{
> > +	int fd, status = 0;
> > +	char f[sizeof "/proc/self/task//comm" + 7];
> 
> Where does 7 come from? I don't think it's correct. 3*sizeof(int) is
> the normal lazy bound we use for %d's.

man 5 proc

says that the maximum value for a pid is about 4 million. I'm fine with
to changing it to the usual lazy bound, though.

> > +	ssize_t len;
> > +
> > +	if ((len = strlen(name)) > 15) return ERANGE;
> 
> Storing the result of strlen in a ssize_t is odd. It works because no
> object can be larger than SIZE_MAX/2, but it's ugly. Also using
> strnlen would be preferable if there's a limit (so as not to read
> unboundedly long strings only to reject them).

ssize_t is a remnant from reusing len for the return value of write().
Will fix both issues.

> > +
> > +	snprintf(f, sizeof f, "/proc/self/task/%d/comm", thread->tid);
> > +	if ((fd = open(f, O_WRONLY)) < 0) return errno;
> > +	if (write(fd, name, len) < 0) status = errno;
> > +
> > +	close(fd);
> 
> This needs blocking of the cancellation points, I think.

Will look into it.

Thanks,
Felix

Re: [PATCH RFC] add pthread_setname_np

[Rich Felker]

On Wed, Sep 14, 2016 at 11:43:51PM -0400, Felix Janda wrote:
> Rich Felker wrote:
> 
> Thanks for the review!
> 
> > On Wed, Sep 14, 2016 at 11:02:16PM -0400, Felix Janda wrote:
> > > gdb's "info threads" displays the thread name.
> > > ---
> > > Should this function care about interrupted syscalls?
> > 
> > Unless it's specified not to return EINTR I'm fine with it returning
> > EINTR.
> 
> ok. Is it possible that write() is interrupted with partial data
> written (so that it won't return -1)?

That won't happen with these proc pseudo-files because the write is
required by the kernel to be atomic. If it's broken up into multiple
parts, the kernel only uses the last one.

> > >  include/pthread.h               |  1 +
> > >  src/thread/pthread_setname_np.c | 21 +++++++++++++++++++++
> > >  2 files changed, 22 insertions(+)
> > >  create mode 100644 src/thread/pthread_setname_np.c
> > > 
> > > diff --git a/include/pthread.h b/include/pthread.h
> > > index 3d2e0c4..94ef919 100644
> > > --- a/include/pthread.h
> > > +++ b/include/pthread.h
> > > @@ -214,6 +214,7 @@ struct cpu_set_t;
> > >  int pthread_getaffinity_np(pthread_t, size_t, struct cpu_set_t *);
> > >  int pthread_setaffinity_np(pthread_t, size_t, const struct cpu_set_t *);
> > >  int pthread_getattr_np(pthread_t, pthread_attr_t *);
> > > +int pthread_setname_np(pthread_t, const char *);
> > >  int pthread_tryjoin_np(pthread_t, void **);
> > >  int pthread_timedjoin_np(pthread_t, void **, const struct timespec *);
> > >  #endif
> > > diff --git a/src/thread/pthread_setname_np.c b/src/thread/pthread_setname_np.c
> > > new file mode 100644
> > > index 0000000..b7f7a4b
> > > --- /dev/null
> > > +++ b/src/thread/pthread_setname_np.c
> > > @@ -0,0 +1,21 @@
> > > +#include <fcntl.h>
> > > +#include <string.h>
> > > +#include <unistd.h>
> > > +
> > > +#include "pthread_impl.h"
> > > +
> > > +int pthread_setname_np(pthread_t thread, const char *name)
> > > +{
> > > +	int fd, status = 0;
> > > +	char f[sizeof "/proc/self/task//comm" + 7];
> > 
> > Where does 7 come from? I don't think it's correct. 3*sizeof(int) is
> > the normal lazy bound we use for %d's.
> 
> man 5 proc
> 
> says that the maximum value for a pid is about 4 million. I'm fine with
> to changing it to the usual lazy bound, though.

Ah. That may be true for all current versions of Linux, but I don't
think it counts as part of the API/ABI. The ABI limit on pids is 512M
IIRC (because high bits have special meaning in robust/PI futex ABI).

Rich

Re: [PATCH RFC] add pthread_setname_np

[Alexander Monakov]

Hi,

On Wed, 14 Sep 2016, Felix Janda wrote:
> --- /dev/null
> +++ b/src/thread/pthread_setname_np.c
> @@ -0,0 +1,21 @@
[...]
> +	snprintf(f, sizeof f, "/proc/self/task/%d/comm", thread->tid);
> +	if ((fd = open(f, O_WRONLY)) < 0) return errno;
> +	if (write(fd, name, len) < 0) status = errno;


Is there a reason to do it this way rather than via prctl(PR_SET_NAME, ...)?
If so, it should be mentioned in the commit message. (I think using prctl
would be more appropriate)

Thanks.
Alexander

Re: [PATCH RFC] add pthread_setname_np

[Felix Janda]

Alexander Monakov wrote:
> Hi,
> 
> On Wed, 14 Sep 2016, Felix Janda wrote:
> > --- /dev/null
> > +++ b/src/thread/pthread_setname_np.c
> > @@ -0,0 +1,21 @@
> [...]
> > +	snprintf(f, sizeof f, "/proc/self/task/%d/comm", thread->tid);
> > +	if ((fd = open(f, O_WRONLY)) < 0) return errno;
> > +	if (write(fd, name, len) < 0) status = errno;
> 
> 
> Is there a reason to do it this way rather than via prctl(PR_SET_NAME, ...)?
> If so, it should be mentioned in the commit message. (I think using prctl
> would be more appropriate)

pthread_setname_np() is more general than prctl(PR_SET_NAME, ...): It can
change the name of any thread, not just the calling one.

Felix

Re: [PATCH RFC] add pthread_setname_np

[Felix Janda]

Rich Felker wrote:
> On Wed, Sep 14, 2016 at 11:43:51PM -0400, Felix Janda wrote:
> > Rich Felker wrote:
> > 
> > Thanks for the review!
> > 
> > > On Wed, Sep 14, 2016 at 11:02:16PM -0400, Felix Janda wrote:
> > > > gdb's "info threads" displays the thread name.
> > > > ---
> > > > Should this function care about interrupted syscalls?
> > > 
> > > Unless it's specified not to return EINTR I'm fine with it returning
> > > EINTR.
> > 
> > ok. Is it possible that write() is interrupted with partial data
> > written (so that it won't return -1)?
> 
> That won't happen with these proc pseudo-files because the write is
> required by the kernel to be atomic. If it's broken up into multiple
> parts, the kernel only uses the last one.

Thanks for explaining this!

So the current function will notice that it is interrupted, and return
EINTR.

> > > >  include/pthread.h               |  1 +
> > > >  src/thread/pthread_setname_np.c | 21 +++++++++++++++++++++
> > > >  2 files changed, 22 insertions(+)
> > > >  create mode 100644 src/thread/pthread_setname_np.c
> > > > 
> > > > diff --git a/include/pthread.h b/include/pthread.h
> > > > index 3d2e0c4..94ef919 100644
> > > > --- a/include/pthread.h
> > > > +++ b/include/pthread.h
> > > > @@ -214,6 +214,7 @@ struct cpu_set_t;
> > > >  int pthread_getaffinity_np(pthread_t, size_t, struct cpu_set_t *);
> > > >  int pthread_setaffinity_np(pthread_t, size_t, const struct cpu_set_t *);
> > > >  int pthread_getattr_np(pthread_t, pthread_attr_t *);
> > > > +int pthread_setname_np(pthread_t, const char *);
> > > >  int pthread_tryjoin_np(pthread_t, void **);
> > > >  int pthread_timedjoin_np(pthread_t, void **, const struct timespec *);
> > > >  #endif
> > > > diff --git a/src/thread/pthread_setname_np.c b/src/thread/pthread_setname_np.c
> > > > new file mode 100644
> > > > index 0000000..b7f7a4b
> > > > --- /dev/null
> > > > +++ b/src/thread/pthread_setname_np.c
> > > > @@ -0,0 +1,21 @@
> > > > +#include <fcntl.h>
> > > > +#include <string.h>
> > > > +#include <unistd.h>
> > > > +
> > > > +#include "pthread_impl.h"
> > > > +
> > > > +int pthread_setname_np(pthread_t thread, const char *name)
> > > > +{
> > > > +	int fd, status = 0;
> > > > +	char f[sizeof "/proc/self/task//comm" + 7];
> > > 
> > > Where does 7 come from? I don't think it's correct. 3*sizeof(int) is
> > > the normal lazy bound we use for %d's.
> > 
> > man 5 proc
> > 
> > says that the maximum value for a pid is about 4 million. I'm fine with
> > to changing it to the usual lazy bound, though.
> 
> Ah. That may be true for all current versions of Linux, but I don't
> think it counts as part of the API/ABI. The ABI limit on pids is 512M
> IIRC (because high bits have special meaning in robust/PI futex ABI).

Ok, better play it safe.


Will post an updated patch later

Felix

Re: [PATCH RFC] add pthread_setname_np

[Rich Felker]

On Thu, Sep 15, 2016 at 07:57:03AM -0400, Felix Janda wrote:
> Alexander Monakov wrote:
> > Hi,
> > 
> > On Wed, 14 Sep 2016, Felix Janda wrote:
> > > --- /dev/null
> > > +++ b/src/thread/pthread_setname_np.c
> > > @@ -0,0 +1,21 @@
> > [...]
> > > +	snprintf(f, sizeof f, "/proc/self/task/%d/comm", thread->tid);
> > > +	if ((fd = open(f, O_WRONLY)) < 0) return errno;
> > > +	if (write(fd, name, len) < 0) status = errno;
> > 
> > Is there a reason to do it this way rather than via prctl(PR_SET_NAME, ...)?
> > If so, it should be mentioned in the commit message. (I think using prctl
> > would be more appropriate)
> 
> pthread_setname_np() is more general than prctl(PR_SET_NAME, ...): It can
> change the name of any thread, not just the calling one.

It might be nice to use prctl if thread==pthread_self(), just to avoid
dependency on /proc in this case. But I think that's a minor detail.
This is a non-critical feature anyway.

Rich