-
Notifications
You must be signed in to change notification settings - Fork 149
Segfault when testing on red hat 5, no useful output #120
Comments
Thanks very much testing this! I have a couple of thoughts:
Using
And then get the backtrace with:
That's all for now. I have been seeing segfaults in bloscpack recently also, so I am very much interested in drilling down into this, since I think these things might be related. Also I have experienced segfaults in bcolz when using a python build with I have some more ideas, but let's check the above first, that should give us a few more clues to work with. |
Regarding 2), I can't reproduce the segfault if I use nosetests in Regarding 1) I see a segfault for 0.7.3 as well, but again only if I run the test suite with the script:
and even then, it does not always segfaults. When it does not, it passes 3): does not look very useful I am afraid:
|
I actually also see segfaults on windows 64 bits (bcolz 0.8.0, numpy 1.8.1, MKL, built with vs 2008 on python 2.7.3). It does not always segfaults, but when it does, it does at |
meh :/ |
This is a long shot, but could you try running it from a git clone? Since I am not seeing the git info in the versions, I presume you are using a pip install or tarball? |
Can you check #121 and #122 and maybe use dynamic linking to test bcolz v0.8.0 with blosc v1.4.1? Instructions for linking dynamically to blosc can be found here: http://bcolz.blosc.org/install.html#installing-from-tarball-sources |
@esc sorry for the late reply. I just went back to the issue, and can confirm that using an external libblosc (1.4.1 in canopy's case) fix the issue on every platform we support. That may be irrelevant, but running the crashing bcolz on my mac VM worked flawlessly (the environment is exactly the same as we use automatic provisioning for our build environments at Enthought). This may suggest a CPU-specific issue, for example SSE/AVX support (my mac is the newest model, whereas my machine having the crash are older). |
@cournape thanks for the feedback, your findings are in-line with my gut feeling. |
I have put some time on this, and I came up with a small code that can trigger a problem in valgrind:
Part of the output in valgrind:
I have no completely nailed down the root of the double free, but I think it has to do with how views are released in carray objects, probably causing a double release of the referenced carray. |
Hmm, I think that my previous note was a red herring. I am thinking now that the problem might be an optimization that I introduced in blosclz for blosc 1.5. @cournape could you please apply the patch below and tell me how it goes? Unfortunately I cannot reproduce the crash in my machine :-/
|
I am going to close this due to inactivity, feel free to re-open if the issue persists. |
Hi,
While looking at bcolz 0.8.0, I got some segfaults when running the test suite:
I would have expected more output for verbose, if only to track down which test is failing. The test suite sometimes succeed, though.
The text was updated successfully, but these errors were encountered: