np.linalg.pinv() hangs for certain np.inf arrangements in the inversed matrix

If I try to do np.linalg.pinv() of a matrix that contains some np.infs, sometimes the function just hangs with 100% CPU usage, other times it returns something:

In [1]: import numpy as np
In [2]: x = np.array([[550.0, 1], [1, np.inf]])
In [3]: np.linalg.pinv(x)
Out[3]: 
array([[ 0.,  0.],
       [ 0.,  0.]])

In [4]: x = np.array([[np.inf, np.inf], [np.inf, np.inf]])
In [5]: np.linalg.pinv(x)
Out[5]: 
array([[ nan,  nan],
       [ nan,  nan]])

In [6]: x = np.array([[np.inf, np.inf], [550.0, 1]])
In [7]: np.linalg.pinv(x)
Out[7]: 
array([[ nan,  nan],
       [ nan,  nan]])

In [8]: x = np.array([[550.0, 1], [np.inf, np.inf]])
In [9]: np.linalg.pinv(x) # here it just hangs and I must kill it from outside
Killed

I think the correct behaviour should be to raise an exception or return nans or whatever, but definitely not hang.

My configuration:
Ubuntu 15.10

$ uname -srvmpio
Linux 4.2.0-34-generic #39-Ubuntu SMP Thu Mar 10 22:13:01 UTC 2016 x86_64 x86_64 x86_64 GNU/Linux

python version: 3.5.0+
numpy version: 1.10.4
installed via pip in a virtual environment, log of installation: pip-install-numpy.txt
gcc version: gcc (Ubuntu 5.2.1-22ubuntu2) 5.2.1 20151010

Start with the np.linalg.pinv() entry point and reproduce the reported 2x2 matrices, especially the case that reaches 100% CPU usage. Check the existing linear-algebra tests for pseudoinverse behavior with non-finite values. Done means the reproducer returns or raises instead of hanging, with regression coverage for the problematic arrangement.