1&2) I removed the try/except around the import. I have no clue if os 
might be unavailable. Maybe leave out handling that until we see that 
breaking.

I added the try/except because I saw that in other tests in the same 
file when importing gc.

3) Done.
4) The EOFError exceptions are tested in test_tofromfile.

> I tried to apply both the patches on the trunk but the tests don't
> pass. With the latest patch I get an EOFError instead of IOError in
> the assertRaises.
That is the also behaviour before the patch, so the ferror(fp) does not 
fire. Either the test is inappropriate or your system doesn't regard 
reading a closed filehandle an error. I'm not able to investigate this 
as it works fine on my os x 10.6. What system did you test it on?

Reliable ways of producing IOErrors are harder to find than I thought. 
Deleting the file between to reads makes it just look truncated. Another 
method I tried was crashing a process which holds the other end of a 
pipe, but that's messy and complicated.