On Solaris 11.3 (intel tested, but I assume issue is on SPARC as well),
I found the following fails:
  import os
  os.urandom(2500)

The above throws OSError: [Errno 22] Invalid argument.

It turns out that the Solaris version of getrandom() is limited to returning
no more than 1024 bytes, per the manpage:
  The getrandom() and getentropy() functions fail if:

  EINVAL    The flags are not set to GRND_RANDOM, GRND_NONBLOCK or  both,
            or bufsz is <= 0 or > 1024.

I've attached a possible patch for this issue, against the 3.5.1 source
tree.

See also the issue #25003 and the changeset 835085cc28cd:

Issue #25003: On Solaris 11.3 or newer, os.urandom() now uses the getrandom() function instead of the getentropy() function. The getentropy() function is blocking to generate very good quality entropy, os.urandom() doesn't need such high-quality entropy.

> I've attached a possible patch for this issue, against the 3.5.1 source
tree.

I guess that you are already using Python 3.5.1 which uses getrandom(). You should try to confirm using strace.

I updated your patch. I replaced "#if defined(__sun__)" with "#ifdef sun", since "#ifdef sun" looks more common in the Python code base, and I never saw "#if defined(__sun__)" in the Python code base.

I also avoided the new len variable, I reused the n variable.

I don't have Solaris, so I cannot test. I didn't find getrandom() manual page neither, I only found this blog post which doesn't mention the 1024 bytes limitation on Solaris:
https://blogs.oracle.com/darren/entry/solaris_new_system_calls_getentropy

The new patch looks fine; I used __sun__ rather than sun out of habit
(C standard requires
system specific macros be in the reserved namespace), but either will work.

I found the original problem through debugging with GDB, so I know
getrandom() was being
called, and the test case I provided (taken from Lib/Random.py) fails
without the patch,
and succeeds with it.

Oddly, the blog post you linked to describes getrandom as:
"Recent Linux kernels have a getrandom(2) system call that reads
between 1 and 1024 bytes of randomness"

But no such limit current exists in the Linux version that I can see;
however, the Solaris
version definitely does have that limit:
https://docs.oracle.com/cd/E53394_01/html/E54765/getrandom-2.html

On Tue, Apr 12, 2016 at 1:15 AM, STINNER Victor <report@bugs.python.org> wrote:
>
> STINNER Victor added the comment:
>
>> I've attached a possible patch for this issue, against the 3.5.1 source
> tree.
>
> I guess that you are already using Python 3.5.1 which uses getrandom(). You should try to confirm using strace.
>
> I updated your patch. I replaced "#if defined(__sun__)" with "#ifdef sun", since "#ifdef sun" looks more common in the Python code base, and I never saw "#if defined(__sun__)" in the Python code base.
>
> I also avoided the new len variable, I reused the n variable.
>
> I don't have Solaris, so I cannot test. I didn't find getrandom() manual page neither, I only found this blog post which doesn't mention the 1024 bytes limitation on Solaris:
> https://blogs.oracle.com/darren/entry/solaris_new_system_calls_getentropy
>
> ----------
> Added file: http://bugs.python.org/file42443/urandom_solaris.patch
>
> _______________________________________
> Python tracker <report@bugs.python.org>
> <http://bugs.python.org/issue26735>
> _______________________________________

> The new patch looks fine

Do you mean that it fixes your issue? Can it be applied to Python 3.5 & 3.6?

Yes, I've verified that:
* the issue existed in the default branch as of this morning.
* the patch applies cleanly against both 3.5 and default, and
addresses the issue in both branches.

New changeset fb7628e8dfef by Victor Stinner in branch '3.5':
Fix os.urandom() on Solaris 11.3
https://hg.python.org/cpython/rev/fb7628e8dfef

> Yes, I've verified that: (...)

Cool, thanks for the bug report and for the check.

It's now fixed. In the meanwhile, you can workaround the issue by limiting yourself calls to os.urandom() to 1024 bytes (and then concatenate the result).