[issue25003] os.urandom() should call getrandom(2) not getentropy(2) on Solaris 11.3 and newer

[Roundup Robot]

Roundup Robot added the comment:

New changeset 835085cc28cd by Victor Stinner in branch '3.5':
Issue #25003: On Solaris 11.3 or newer, os.urandom() now uses the getrandom()
https://hg.python.org/cpython/rev/835085cc28cd

--

___
Python tracker 

___
___
Python-bugs-list mailing list
Unsubscribe: 
https://mail.python.org/mailman/options/python-bugs-list/archive%40mail-archive.com

[issue25003] os.urandom() should call getrandom(2) not getentropy(2) on Solaris 11.3 and newer

[Roundup Robot]

Roundup Robot added the comment:

New changeset 202c827f86df by Victor Stinner in branch '2.7':
Issue #25003: os.urandom() doesn't use getentropy() on Solaris because
https://hg.python.org/cpython/rev/202c827f86df

New changeset 83dc79eeaf7f by Victor Stinner in branch '3.4':
Issue #25003: os.urandom() doesn't use getentropy() on Solaris because
https://hg.python.org/cpython/rev/83dc79eeaf7f

--

___
Python tracker 

___
___
Python-bugs-list mailing list
Unsubscribe: 
https://mail.python.org/mailman/options/python-bugs-list/archive%40mail-archive.com

[issue25003] os.urandom() should call getrandom(2) not getentropy(2) on Solaris 11.3 and newer

[STINNER Victor]

STINNER Victor added the comment:

Ok, I pushed fixes for Python 2.7, 3.4, 3.5 and 3.6.

Summary for Solaris:

- Python 2.7, 3.4: read from /dev/urandom (use a private file descriptor)
- Python 3.5, 3.6: use the getrandom() function on Solaris 11.3 and newer (no 
file descriptor), or fallback on reading from /dev/urandom (use a private file 
descriptor)

getentropy() is no more used on Solaris.

@John Beck: It would be great if you could run test_os on the development 
branches 2.7, 3.4, 3.5 and default (3.6).

--

___
Python tracker 

___
___
Python-bugs-list mailing list
Unsubscribe: 
https://mail.python.org/mailman/options/python-bugs-list/archive%40mail-archive.com

[issue25003] os.urandom() should call getrandom(2) not getentropy(2) on Solaris 11.3 and newer

[John Beck]

John Beck added the comment:

Confirmed that test_os runs cleanly on Solaris 12, for each of:
* 2.7.10 (plus your patch from 98454:202c827f86df)
* 3.4.3 (plus your patch from 98455:83dc79eeaf7f)
* 3.5.0 (plus your patch from 98452:835085cc28cd)
* 3.6 (tip)

--

___
Python tracker 

___
___
Python-bugs-list mailing list
Unsubscribe: 
https://mail.python.org/mailman/options/python-bugs-list/archive%40mail-archive.com

[issue25003] os.urandom() should call getrandom(2) not getentropy(2) on Solaris 11.3 and newer

[STINNER Victor]

STINNER Victor added the comment:

Thanks! I close the issue, it's now fixed.

--
resolution:  -> fixed
status: open -> closed

___
Python tracker 

___
___
Python-bugs-list mailing list
Unsubscribe: 
https://mail.python.org/mailman/options/python-bugs-list/archive%40mail-archive.com

[issue25003] os.urandom() should call getrandom(2) not getentropy(2) on Solaris 11.3 and newer

[STINNER Victor]

STINNER Victor added the comment:

"I have tested your patch with 3.5, and it works fine, although I did have to 
make a minor change to get test_os to pass.  In test_os.py you had: (...)"

Oh right, I fixed test_os too.

--

___
Python tracker 

___
___
Python-bugs-list mailing list
Unsubscribe: 
https://mail.python.org/mailman/options/python-bugs-list/archive%40mail-archive.com

[issue25003] os.urandom() should call getrandom(2) not getentropy(2) on Solaris 11.3 and newer

[Roundup Robot]

Roundup Robot added the comment:

New changeset 221e09b89186 by Victor Stinner in branch 'default':
Issue #25003: Skip test_os.URandomFDTests on Solaris 11.3 and newer
https://hg.python.org/cpython/rev/221e09b89186

--

___
Python tracker 

___
___
Python-bugs-list mailing list
Unsubscribe: 
https://mail.python.org/mailman/options/python-bugs-list/archive%40mail-archive.com

[issue25003] os.urandom() should call getrandom(2) not getentropy(2) on Solaris 11.3 and newer

[John Beck]

John Beck added the comment:

I have tested your patch with 3.5, and it works fine, although I did have to 
make a minor change to get test_os to pass.  In test_os.py you had:

...
USE_GETENTROPY = ((sysconfig.get_config_var('HAVE_GETENTROPY') == 1)
  and not sys.platform.startswith("sunos"))
HAVE_GETRANDOM = (sysconfig.get_config_var('HAVE_GETRANDOM_SYSCALL') == 1)
 
@unittest.skipIf(USE_GETENTROPY,
 "getentropy() does not use a file descriptor")
@unittest.skipIf(HAVE_GETRANDOM,
 "getrandom() does not use a file descriptor")
...

whereas I came up with this:

...
USE_GETENTROPY = ((sysconfig.get_config_var('HAVE_GETENTROPY') == 1)
  and not sys.platform.startswith("sunos"))
HAVE_GETRANDOM = (sysconfig.get_config_var('HAVE_GETRANDOM') == 1)
HAVE_GETRANDOM_SYSCALL = (sysconfig.get_config_var('HAVE_GETRANDOM_SYSCALL') == 
1)

@unittest.skipIf(USE_GETENTROPY,
 "getentropy() does not use a file descriptor")
@unittest.skipIf(HAVE_GETRANDOM,
 "getrandom() does not use a file descriptor")
@unittest.skipIf(HAVE_GETRANDOM_SYSCALL,
 "getrandom() does not use a file descriptor")
...

--

___
Python tracker 

___
___
Python-bugs-list mailing list
Unsubscribe: 
https://mail.python.org/mailman/options/python-bugs-list/archive%40mail-archive.com

[issue25003] os.urandom() should call getrandom(2) not getentropy(2) on Solaris 11.3 and newer

[STINNER Victor]

STINNER Victor added the comment:

>  While testing this, I found out why I had needed EINVAL earlier (and still 
> do, for now): there is a bug in the Solaris implementation of getrandom(2).  
> If flags are 0 and the buffer size > 1024, then it fails with EINVAL.  That 
> is only supposed to happen for a buffer that big if GNRD_RANDOM is set in 
> flags but GNRD_NONBLOCK is not set.  So until that bug is fixed, ...

Ah! So it was useful to dig the EINVAL issue, it's a bug in the kernel :-)

You wrote that the final version of Solaris 11.3 and 12 was not released yet. 
Can we expect a fix in the kernel before the final version?

--

___
Python tracker 

___
___
Python-bugs-list mailing list
Unsubscribe: 
https://mail.python.org/mailman/options/python-bugs-list/archive%40mail-archive.com

[issue25003] os.urandom() should call getrandom(2) not getentropy(2) on Solaris 11.3 and newer

[John Beck]

John Beck added the comment:

I have queried the engineer who owns the kernel bug and will post an update 
once I hear back from him.  But as it is already almost midnight on Friday in 
his geo, it may well be Monday before I hear back.

--

___
Python tracker 

___
___
Python-bugs-list mailing list
Unsubscribe: 
https://mail.python.org/mailman/options/python-bugs-list/archive%40mail-archive.com

[issue25003] os.urandom() should call getrandom(2) not getentropy(2) on Solaris 11.3 and newer

[STINNER Victor]

STINNER Victor added the comment:

> But EINVAL should also be checked for to make sure the system call was 
> invoked with proper parameters.

py_getrandom() calls Py_FatalError() or raises an OSError on EINVAL. The error 
is not silently ignored.


> My builds of Python 3.5.0.X (don't recall whether X was a late Beta or 
> release candidate) were core dumping because Python was making that syscall 
> but not checking for EINVAL,

Py_FatalError() calls abort(). Usually, operating systems dump a core dump in 
this case. But it is the expected behaviour. Python now refuses to start if the 
OS random number generator doesn't work. There are similar checks on the system 
and monotonic clocks for example.

> ... and thus assuming the call was valid, when it was not.

Ah! Finally you explain the problem :-) I wrote py_getrandom() for Linux. I 
didn't expect other operating systems to implement the getrandom() syscall. I 
hardcoded the flags (0) for example.

py_getrandom() calls directly the syscall, because I like the new cool 
getrandom() syscall of Linux: it avoids the need of a private file descriptor. 
It would be much better to call a function of the C library, but the GNU C 
library didn't expose the function yet...

On Solaris, the function is available in C, no need to call directly the 
syscall. It would be better to call the C function, and check if it's available 
in configure.

Can you please try remove_get_entropy.patch + urandom_solaris.patch?

--
Added file: http://bugs.python.org/file40437/getrandom_solaris.patch

___
Python tracker 

___
___
Python-bugs-list mailing list
Unsubscribe: 
https://mail.python.org/mailman/options/python-bugs-list/archive%40mail-archive.com

[issue25003] os.urandom() should call getrandom(2) not getentropy(2) on Solaris 11.3 and newer

[John Beck]

John Beck added the comment:

Yes, those patches work, with a caveat.  While testing this, I found out why I 
had needed EINVAL earlier (and still do, for now): there is a bug in the 
Solaris implementation of getrandom(2).  If flags are 0 and the buffer size > 
1024, then it fails with EINVAL.  That is only supposed to happen for a buffer 
that big if GNRD_RANDOM is set in flags but GNRD_NONBLOCK is not set.  So until 
that bug is fixed, I have to patch py_getrandom() to treat EINVAL like ENOSYS 
and fall back to using /dev/urandom as if getrandom(2) were not supported.

--

___
Python tracker 

___
___
Python-bugs-list mailing list
Unsubscribe: 
https://mail.python.org/mailman/options/python-bugs-list/archive%40mail-archive.com

[issue25003] os.urandom() should call getrandom(2) not getentropy(2) on Solaris 11.3 and newer

[STINNER Victor]

Changes by STINNER Victor :


--
title: os.urandom() should call getrandom(2) not getentropy(2) on Solaris -> 
os.urandom() should call getrandom(2) not getentropy(2) on Solaris 11.3 and 
newer

___
Python tracker 

___
___
Python-bugs-list mailing list
Unsubscribe: 
https://mail.python.org/mailman/options/python-bugs-list/archive%40mail-archive.com

[issue25003] os.urandom() should call getrandom(2) not getentropy(2) on Solaris 11.3 and newer

[John Beck]

John Beck added the comment:

Sorry, let me try to clarify.  ENOSYS is the correct errno value to check for, 
for the situation (e.g., Solaris < 11.3) of a non-existent system call. But 
EINVAL should also be checked for to make sure the system call was invoked with 
proper parameters.  My builds of Python 3.5.0.X (don't recall whether X was a 
late Beta or release candidate) were core dumping because Python was making 
that syscall but not checking for EINVAL, and thus assuming the call was valid, 
when it was not.  Sorry, I did not try to debug why it was invalid.  But once I 
added EINVAL, alongside ENOSYS, as a reason to set getrandom_works to 0, then 
python started behaving properly.

--

___
Python tracker 

___
___
Python-bugs-list mailing list
Unsubscribe: 
https://mail.python.org/mailman/options/python-bugs-list/archive%40mail-archive.com