Comment 4 for bug 1422307

First insight: Having the VDI image in tmpfs or using --cache=none for qemu-nbd changes nothing, therefore the reason why dropping the caches affects the result is probably Linux's cache for the NBD block device.

Second insight: Using blkverify makes it look very much like qemu's VDI implementation is the culprit:

# for i in $(seq 0 9); do ./qemu-img create -f vdi /tmp/test.vdi 64M > /dev/null; ./qemu-img create -f raw /tmp/test.raw 64M > /dev/null; (./qemu-nbd -v -f raw -c /dev/nbd0 blkverify:/tmp/test.raw:/tmp/test.vdi &); sleep 1; ./qemu-img convert -n blob.raw /dev/nbd0; ./qemu-img convert /dev/nbd0 /tmp/test1.raw; sync; echo 1 > /proc/sys/vm/drop_caches; ./qemu-img convert /dev/nbd0 /tmp/test2.raw; ./qemu-nbd -d /dev/nbd0 > /dev/null; if ! ./qemu-img compare -q /tmp/test1.raw /tmp/test2.raw; then md5sum /tmp/test1.raw /tmp/test2.raw; echo "$i failed"; break; fi; done; echo 'done'
NBD device /dev/nbd0 is now connected to blkverify:/tmp/test.raw:/tmp/test.vdi
NBD device /dev/nbd0 is now connected to blkverify:/tmp/test.raw:/tmp/test.vdi
NBD device /dev/nbd0 is now connected to blkverify:/tmp/test.raw:/tmp/test.vdi
blkverify: read sector_num=27872 nb_sectors=256 contents mismatch in sector 27904
blkverify: read sector_num=28128 nb_sectors=256 contents mismatch in sector 28128
qemu-img: error while reading sector 24576: Input/output error
e5185b807948d65bb4e837d992cea429 /tmp/test1.raw
b8a195424a240ed77c849d89002fa23b /tmp/test2.raw
2 failed
done

Max