On 10/22/2010 04:38 PM, Richard Purdie wrote:

On Fri, 2010-10-22 at 11:22 -0700, Darren Hart wrote:

Here is an example of the do_fetch/do_unpack failure that I believe is
representative of what Dirk saw and what I have been seeing periodically.

I saw the following while building libproxy in poky-image-sdk-live on a
Fedora 13 box behind the firewall. I saw identical failures with opkg on
a different machine - also behind the firewall. Both SRC_URIs point to
googlecode.

log.do_fetch:
-------------
NOTE: fetch http://libproxy.googlecode.com/files/libproxy-0.4.3.tar.gz
NOTE: libproxy-0.4.3:
http://libproxy.googlecode.com/files/libproxy-0.4.3.tar.gz has no entry
in conf/checksums.ini, not checking URI

log.do_unpack:
--------------
gzip: stdin: unexpected end of file
tar: Child returned status 1
tar: Exiting with failure status due to previous errors
NOTE: Unpacking
/vol/1/dvhart/poky.git/build/downloads/libproxy-0.4.3.tar.gz to
/vol/1/dvhart/poky.git/build/tmp/work/i586-poky-linux/libproxy-0.4.3-r2/
ERROR: Task failed:

With opkg, I tried the following which resulted in the same error.

$ bitbake -c clean;
$ rm downloads/*opkg*
$ bitbake opkg

In both cases, using wget to manually download the file and place it in
the downloads directory allowed the build to continue.

We've talked about this a lot over jabber. We don't have a reproducer.
What we do see is its always the wget fetcher that hits this (for mirror
urls in the case of git repos) and its failing creating a zero length
file.

If we make the assumption that wget is creating the empty file and
exiting with an error code and look at what bitbake would do, it would
give the behaviour described in these bugs.

Why? If the original fetch fails, it falls back to the mirror code. That
looks, sees a file on the disk and says "nothing to do".

Solution is therefore to make sure if the wget fetcher fails we wipe any
file that may be present.

Is is possible the wget fetcher is creating a zero length file and no
setting an error code but I really hope its not that broken.

So I've pushed this patch as a bandaid for the problem:

http://git.pokylinux.org/cgit/cgit.cgi/poky/commit/?id=c5fab99a6f979a4a0ce246c6395b35a3082aec0d

I've applied and am running another build. This looks sane to me. If I still run into this, I think the next step is to do a brute force zero-length test, and throw the FetchError and take this modified path from this patch in that case too.

wget will generated a 0 length file when using -O (which we confirmed we do not) and there are old bugs against it for creating 0 length files when a valid host fails to connect (without the -O option) but I wasn't able to reproduce that with a current wget.

Here's to hoping this fixes it.

--
Darren Hart
Embedded Linux Kernel