Comment 3 for bug 1422307

For whatever reason, using an empty image now works for me, too:

$ ./qemu-img create -f vdi test.vdi 64M; ./qemu-nbd -c /dev/nbd0 test.vdi; dd if=/dev/urandom of=/dev/nbd0 bs=1K count=16384; md5sum /dev/nbd0; sync; echo 1 > /proc/sys/vm/drop_caches; md5sum /dev/nbd0; ./qemu-nbd -d /dev/nbd0
Formatting 'test.vdi', fmt=vdi size=67108864 static=off
16384+0 records in
16384+0 records out
16777216 bytes (17 MB) copied, 0.982225 s, 17.1 MB/s
216f7abbf90bf2539163396bdb7fd7b9 /dev/nbd0
a42faf71124c1f6102fa39cea82a1c86 /dev/nbd0
/dev/nbd0 disconnected

Writing less than 16384 kB, the issue is not always reproducible; for me, it disappears around 16160 kB (it's fuzzy, sometimes it appears, sometimes it doesn't).

So far I was only able to reproduce the issue by connecting qemu-nbd to the the Linux NBD interface; connecting to qemu-nbd via TCP worked fine.

So, a couple of test cases:

VDI and NBD over /dev/nbd0:
# for i in $(seq 0 9); do ./qemu-img create -f vdi test.vdi 64M > /dev/null; ./qemu-nbd -c /dev/nbd0 test.vdi; sleep 1; ./qemu-img convert -n blob.raw /dev/nbd0; ./qemu-img convert /dev/nbd0 test1.raw; sync; echo 1 > /proc/sys/vm/drop_caches; ./qemu-img convert /dev/nbd0 test2.raw; ./qemu-nbd -d /dev/nbd0 > /dev/null; if ! ./qemu-img compare -q test1.raw test2.raw; then md5sum test1.raw test2.raw; echo "$i failed"; break; fi; done; echo 'done'
e5185b807948d65bb4e837d992cea429 test1.raw
9907ca700f6ee4d4cdb136bb90fd8df1 test2.raw
6 failed
done

VDI and NBD over TCP:
# for i in $(seq 0 9); do ./qemu-img create -f vdi test.vdi 64M > /dev/null; (./qemu-nbd -t test.vdi &); sleep 1; ./qemu-img convert -n blob.raw nbd://localhost; ./qemu-img convert nbd://localhost test1.raw; sync; echo 1 > /proc/sys/vm/drop_caches; ./qemu-img convert nbd://localhost test2.raw; killall qemu-nbd; if ! ./qemu-img compare -q test1.raw test2.raw; then md5sum test1.raw test2.raw; echo "$i failed"; break; fi; done; echo 'done'
done

VDI and NBD over a Unix socket:
# for i in $(seq 0 9); do ./qemu-img create -f vdi test.vdi 64M > /dev/null; (./qemu-nbd -k /tmp/nbd -t test.vdi &); sleep 1; ./qemu-img convert -n blob.raw nbd+unix:///\?socket=/tmp/nbd; ./qemu-img convert nbd+unix:///\?socket=/tmp/nbd test1.raw; sync; echo 1 > /proc/sys/vm/drop_caches; ./qemu-img convert nbd+unix:///\?socket=/tmp/nbd test2.raw; killall qemu-nbd; if ! ./qemu-img compare -q test1.raw test2.raw; then md5sum test1.raw test2.raw; echo "$i failed"; break; fi; done; echo 'done'
done

VDI without NBD:
# for i in $(seq 0 9); do ./qemu-img create -f vdi test.vdi 64M > /dev/null; ./qemu-img convert -n -O vdi blob.raw test.vdi; ./qemu-img convert test.vdi test1.raw; sync; echo 1 > /proc/sys/vm/drop_caches; ./qemu-img convert test.vdi test2.raw; if ! ./qemu-img compare -q test1.raw test2.raw; then md5sum test1.raw test2.raw; echo "$i failed"; break; fi; done; echo 'done'
done

qcow2 and NBD over /dev/nbd0:
# for i in $(seq 0 9); do ./qemu-img create -f qcow2 test.qcow2 64M > /dev/null; ./qemu-nbd -c /dev/nbd0 test.qcow2; sleep 1; ./qemu-img convert -n blob.raw /dev/nbd0; ./qemu-img convert /dev/nbd0 test1.raw; sync; echo 1 > /proc/sys/vm/drop_caches; ./qemu-img convert /dev/nbd0 test2.raw; ./qemu-nbd -d /dev/nbd0 > /dev/null; if ! ./qemu-img compare -q test1.raw test2.raw; then md5sum test1.raw test2.raw; echo "$i failed"; break; fi; done; echo 'done'
done

raw and NBD over /dev/nbd0:
# for i in $(seq 0 9); do ./qemu-img create -f raw test.raw 64M > /dev/null; ./qemu-nbd -f raw -c /dev/nbd0 test.raw; sleep 1; ./qemu-img convert -n blob.raw /dev/nbd0; ./qemu-img convert /dev/nbd0 test1.raw; sync; echo 1 > /proc/sys/vm/drop_caches; ./qemu-img convert /dev/nbd0 test2.raw; ./qemu-nbd -d /dev/nbd0 > /dev/null; if ! ./qemu-img compare -q test1.raw test2.raw; then md5sum test1.raw test2.raw; echo "$i failed"; break; fi; done; echo 'done'
done

In conclusion, the only combination I can reproduce the issue with is VDI with NBD over the Linux NBD interface. It doesn't seem to be the kernel's fault because other file formats work fine; it doesn't seem to be qemu-nbd's fault because not using the kernel interface works fine; and it doesn't seem to be VDI's fault because not using NBD or at least using NBD over TCP or Unix sockets works fine, too.

I'll keep looking into it.

Max