Hi, On Mon, Jan 05, 2015 at 01:15:52AM +1100, Russell Sim wrote: > > I can reproduce this in my test environment on i386 and amd64. It only > > happens > > when the builddir and /tmp are on different filesystems. It seems the local > > clone tries to create a hard link, which fails across filesystems (the fact > > that this happens without fallback is an error in itself, so the test > > actually > > discovered a problem here). When setting the TMPDIR to a directory on the > > same > > filesystem, the test doesn't hit this issue, and the build works fine. > > I've had a go at reproducing this locally, but it seems to be a pretty > subtle bug. My TMP_DIR on one of my machines is on a different file > system to my builddir but it seems to switch to copying instead of > linking during the clone. I can see this with strace. They do stat > /tmp and . before beginning the copy/link. I realise it's probably > futile but could you please run stat and compare your device id's. That > is how they are checking that they are the same FS in the clone code.
Both the build dir and the tmp dir are bind mounts of the same tmpfs. So the device id is the same, but they are different mountpoints. You can probably reproduce it by just bind-mounting (mount --bind) the filesystem of your build-dir to some place, and use that as tmpdir. The device id's are the same, but you are not allowed to use hard-links (because they are not the same mountpoint). If you look in strace for XDEV, you'll find the error. Cheers, Ivo -- To UNSUBSCRIBE, email to debian-bugs-rc-requ...@lists.debian.org with a subject of "unsubscribe". Trouble? Contact listmas...@lists.debian.org
