It seems the dust settles a bit: Found the relevant difference between my various filesystems, and how to reproduce the failure: Susceptible filesystems don't have the extent feature of ext4 enabled.
You can create such a filesystem using mke2fs -t ext4 -O ^extent /dev/... mount /mnt /dev/... Adapting the command line example provided above you can see rm -f /mnt/tmp.qcow2 cat $SRC_PATH > /mnt/tmp.qcow2 && qemu-img convert -O raw /mnt/tmp.qcow /mnt/tmp.qcow cksum /mnt/tmp.qcow creating corrupt (usually nullified) result images. By inserting a sleep of at least 33 seconds between the cat command and the qemu-img invocation I'm getting proper output. To me it's unclear now, where the actual defect is located. Creating ext4 filesystems with certain features disabled (such as the exetent tree) is apparently supported and ok. Is the fiemap ioctl supposed to handle this gracefully, for example by assuming FIEMAP_FLAG_SYNC in absence of an extent tree? Or are clients such as qemu-img supposed to always FIEMAP_FLAG_SYNC to be safe? -- You received this bug notification because you are a member of qemu- devel-ml, which is subscribed to QEMU. https://bugs.launchpad.net/bugs/1368815 Title: qemu-img convert intermittently corrupts output images Status in OpenStack Compute (Nova): Triaged Status in QEMU: New Bug description: -- Found in releases qemu-2.0.0, qemu-2.0.2, qemu-2.1.0. Tested on Ubuntu 14.04 using Ext4 filesystems. The command qemu-img convert -O raw inputimage.qcow2 outputimage.raw intermittently creates corrupted output images, when the input image is not yet fully synchronized to disk. While the issue has actually been discovered in operation of of OpenStack nova, it can be reproduced "easily" on command line using cat $SRC_PATH > $TMP_PATH && $QEMU_IMG_PATH convert -O raw $TMP_PATH $DST_PATH && cksum $DST_PATH on filesystems exposing this behavior. (The difficult part of this exercise is to prepare a filesystem to reliably trigger this race. On my test machine some filesystems are affected while other aren't, and unfortunately I haven't found the relevant difference between them, yet. Possible it's timing issues completely out of userspace control ...) The root cause, however, is the same as in http://lists.gnu.org/archive/html/coreutils/2011-04/msg00069.html and it can be solved the same way as suggested in http://lists.gnu.org/archive/html/coreutils/2011-04/msg00102.html In qemu, file block/raw-posix.c use the FIEMAP_FLAG_SYNC, i.e change f.fm.fm_flags = 0; to f.fm.fm_flags = FIEMAP_FLAG_SYNC; As discussed in the thread mentioned above, retrieving a page cache coherent map of file extents is possible only after fsync on that file. See also https://bugs.launchpad.net/nova/+bug/1350766 In that bug report filed against nova, fsync had been suggested to be performed by the framework invoking qemu-img. However, as the choice of fiemap -- implying this otherwise unneeded fsync of a temporary file -- is not made by the caller but by qemu-img, I agree with the nova bug reviewer's objection to put it into nova. The fsync should instead be triggered by qemu-img utilizing the FIEMAP_FLAG_SYNC, specifically intended for that purpose. To manage notifications about this bug go to: https://bugs.launchpad.net/nova/+bug/1368815/+subscriptions