Am 12.12.2012 15:09, schrieb Philipp Hahn: > Hello Kevin, > > Am Mittwoch 12 Dezember 2012 14:41:49 schrieb Kevin Wolf: >> As you can see in the commit message of that patch I was convinced that >> no bug did exist in practice and this was only dangerous with respect to >> future changes. Therefore my first question is if you're using an >> unmodified upstream qemu or if some backported patches are applied to >> it? If it's indeed unmodified, we should probably review the code once >> again to understand why it makes a difference. > > This were all unmodified versions directly from git between "qemu-kvm-1.1.0" > and "qemu-kvm-1.2.0" > > "git checkout b7ab0fea37c15ca9e249c42c46f5c48fd1a0943c" works, > "git checkout b7ab0fea37c15ca9e249c42c46f5c48fd1a0943c~1" is broken. > "git checkout qemu-kvm-1.1.2" is broken, > "git checkout qemu-kvm-1.1.2 ; git cherry-pick > b7ab0fea37c15ca9e249c42c46f5c48fd1a0943c" works
Ok, thanks for clarifying. Then I must have missed some interesting case while doing the patch. Ideally we would find a sequence of qemu-io commands to reliably reproduce this. First thing worth trying would be running the current qemu-iotests suite on the old versions. If we don't find it this way, I guess we need to catch it with code review. I'm not sure if I can get to it this week, and starting next week I'll be on vacation, so any help with finding a reproducer would be appreciated. >> In any case, this is the cluster allocation code. It's probably not >> related to rereading things from disk, but rather to the writeout of the >> page cache. > > Yes, the problem is probably write related. But as the write "doens't explode > with some spectacular error", I only notice the error on the following read > by comparing md5 sums. > I just re-checked it: After a reboot the md5sums are still invalid, so I > guess > the data is corrupted on writeout. Yes, it's really the only thing that makes sense in connection with this patch. Kevin
