On Thu, Jun 20, 2013 at 05:50:16AM -0400, Martin Sivak wrote:
Hi,
recently I discovered an issue with our Jenkins test environment. It was failing in
testHooks.py because my Gerrit name contains diacritics and our code tried to decode it as
ascii.
Traceback (most recent call last):
File "/usr/lib64/python2.6/unittest.py", line 278, in run
testMethod()
File "/ephemeral0/vdsm_unit_tests_gerrit_el/tests/hooksTests.py", line 125,
in test_deviceCustomProperties
params={'customProperty': ' rocks!'})
File "/ephemeral0/vdsm_unit_tests_gerrit_el/vdsm/hooks.py", line 70, in
_runHooksDir
scriptenv[k] = unicode(v).encode('utf-8')
UnicodeDecodeError: 'ascii' codec can't decode byte 0xc3 in position 12:
ordinal not in range(128)
The relevant code is here:
hooks.py:
60 scriptenv = os.environ.copy()
...
69 for k, v in scriptenv.iteritems():
70 scriptenv[k] = unicode(v).encode('utf-8')
My first instinct was to decode it using the proper encoding:
source_encoding = sys.stdin.encoding or locale.getpreferredencoding()
for k, v in scriptenv.iteritems():
scriptenv[k] = v.decode(source_encoding).encode('utf-8')
But it still did not work. So I tried to print out the environment and encodings that are
used when make check is being run and got this:
sys.stdin.encoding == None
locale.getpreferredencoding() -> ANSI_X3.4-1968
os.environ['LC_ALL'] == 'C'
os.environ['LANG'] == 'en_US.UTF-8'
Please notice the encoding part, my system and terminal are using utf-8, but vdsm reads
the environment values using ANSI. That is obviously wrong and can't work.
So i tried to investigate it further and found out we force LC_ALL to C in vdsmd.init,
run_tests.sh.in and run_tests_local.sh.in.
I also found the commit that introduced this - 107644dbad9af250c00e7f25fc51a92c6250d442 -
and finally understood where the issue was.
Although I understand the reasons for the patch, I do not agree with
it. If we are executing other tools and parse their output, we should
be preparing and passing the updated locale _only_ to those tools. We
should not be setting the locale we need for parsing stuff to the
whole vdsm daemon.
Since vdsm is not intended for direct human control, I actually like the
idea of turning off all locale noise by a global LC_ALL=C. The
alternative, of setting it to C before each application with parsed
output seems tedious and easily forgotten.
Our current practice of setting LC_ALL to C no matter on what terminal
or system we are starting vdsmd is causing us the above mentioned
issue, because the environment can (and does) contain data in the
system encoding. This essentially prevents anybody with utf-8 chars in
their names to submit anything to vdsm.
No doubt that we have to fix it. The easiest hack is to ask our Jenkins
job to clear the Jenkins env vars before calling `make check`. I'm sure
David (CCed) can do it quite easily.
So I would like to start a discussion about this that will lead to the
necessary fixes and change in our current practice :)
Unfortunately, I have no idea beyond exterminating non-7-bit chars from
the environment and setting LC_ALL=C in n+1 places.
The first approach may not be so horrible as it seems: I'm not sure we
should pass every vdsm env variable to the hook scripts. Passing only
ascii ones may be good enough.
Obviously, unicode custom properties should continue to be explicitly
added, with utf-8 encoding, to the script environment, as this is a
documented vdsm API.
Dan.