Dear list, we have this very old bug [1] that I just keep pushing in front of me. I made several attempts to fix it. However, none of them made into the tree. I guess it's time to have discussion what to do about it. IIRC, the algorithm that I implemented last was to keep original label in XATTRs (among with some ref counter) and the last one to restore the label will look there and find the original label. There was a problem with two libvirtds fighting over a file on shared FS.

So I guess my question is can we come up with a design that would work? Or at least work to the extent that we're satisfied with? Personally, I like the XATTRs approach. And to resolve the NFS race we can invent yet another lockspace to guard labelling - I also have a bug for that [2] (although, I'm not that familiar with lockspaces). I guess doing disk metadata locking is not going to be trivial, is it?

On 07/27/2018 11:01 AM, Daniel P. Berrangé wrote: > On Fri, Jul 27, 2018 at 09:56:23AM +0200, Michal Privoznik wrote: >> Dear list, >> >> we have this very old bug [1] that I just keep pushing in front of me. I >> made several attempts to fix it. However, none of them made into the >> tree. I guess it's time to have discussion what to do about it. IIRC, >> the algorithm that I implemented last was to keep original label in >> XATTRs (among with some ref counter) and the last one to restore the >> label will look there and find the original label. There was a problem >> with two libvirtds fighting over a file on shared FS. > > IIRC there was a second problem around security of XATTRs. eg, if we set > an XATTR it was possible for QEMU to change it once we given QEMU privs > on the image. Thus a malicious QEMU can cause libvirtd to change the > image to an arbitrary user on shutdown. > > I think it was possible to deal with that on Linux, because the kernel > hsa some xattr namespacing that is reserved for only root. Yes, there are basically 4 levels (defined by prefix of attribute name): security. system. trusted. user. For the first two there is no restriction on side of VFS (but there might be some coming from underlying FS and/or security module). The trusted. can be get/set only by CAP_SYS_ADMIN process. And finally user. can be set only one regular files (plus some other restrictions), but it's available for basically everybody. IIRC, I used trusted. namespace because of CAP_SYS_ADMIN. Nota bene, if a file is stored on NFS and only CAP_SYS_ADMIN can change the attribute's value they are CAP_SYS_ADMIN already - they might change seclabel too. Why bother mangling libvirt's records. IOW, if somebody runs qemu with CAP_SYS_ADMIN all bets are off anyway.

This would of course mean that there's no label restore for session daemon, but that can hardly ever work. Do we even initialize DAC/SELinux drivers for session daemon?

> This protection > is worthless though when using NFS images as you can't trust the remote NFS > client to honour the same restriction. > >> So I guess my question is can we come up with a design that would work? >> Or at least work to the extent that we're satisfied with? >> >> Personally, I like the XATTRs approach. And to resolve the NFS race we >> can invent yet another lockspace to guard labelling - I also have a bug >> for that [2] (although, I'm not that familiar with lockspaces). I guess >> doing disk metadata locking is not going to be trivial, is it? > > Yeah, for sure we need two distinct locking regimes. Our current locking > aims to protect disk content, and the lifetime of the lock is the entire > duration of the QEMU process. For XATTRs, we only need write protection > for the short time window in which the security drivers execute, which > is at start, during hotplug/unplug, during shutdown, and finally migration. I guess migration would be covered by start. Since these locks would be held only for very short period (basically they would be acquired just before we try to set seclabel and released after disk content lock is successfully acquired) they can be exclusive. However, since we want to protect more than just disk seclabels, we need to acquire this new lock from more places. > > I think we could achieve this with the virtlockd if we make it use the > same locking file, but a different byte offset within the file. Just > need to make sure it doesn't clash with the byte offset QEMU has chosen, > nor the current offset we use. Makes sense. So the first step is to introduce metadata locking. I'll look into that.

On 07/27/2018 11:01 AM, Daniel P. Berrangé wrote: > On Fri, Jul 27, 2018 at 09:56:23AM +0200, Michal Privoznik wrote: >> Dear list, >> >> we have this very old bug [1] that I just keep pushing in front of me. I >> made several attempts to fix it. However, none of them made into the >> tree. I guess it's time to have discussion what to do about it. IIRC, >> the algorithm that I implemented last was to keep original label in >> XATTRs (among with some ref counter) and the last one to restore the >> label will look there and find the original label. There was a problem >> with two libvirtds fighting over a file on shared FS. > > IIRC there was a second problem around security of XATTRs. eg, if we set > an XATTR it was possible for QEMU to change it once we given QEMU privs > on the image. Thus a malicious QEMU can cause libvirtd to change the > image to an arbitrary user on shutdown. > > I think it was possible to deal with that on Linux, because the kernel > hsa some xattr namespacing that is reserved for only root. This protection > is worthless though when using NFS images as you can't trust the remote NFS > client to honour the same restriction. > >> So I guess my question is can we come up with a design that would work? >> Or at least work to the extent that we're satisfied with? >> >> Personally, I like the XATTRs approach. And to resolve the NFS race we >> can invent yet another lockspace to guard labelling - I also have a bug >> for that [2] (although, I'm not that familiar with lockspaces). I guess >> doing disk metadata locking is not going to be trivial, is it? > > Yeah, for sure we need two distinct locking regimes. Our current locking > aims to protect disk content, and the lifetime of the lock is the entire > duration of the QEMU process. For XATTRs, we only need write protection > for the short time window in which the security drivers execute, which > is at start, during hotplug/unplug, during shutdown, and finally migration. > > I think we could achieve this with the virtlockd if we make it use the > same locking file, but a different byte offset within the file. Just > need to make sure it doesn't clash with the byte offset QEMU has chosen, > nor the current offset we use. So do we really need virtlockd for this? I mean, the whole purpose of it is to hold locks so that libvirtd can restart independently, without losing the lock attached. However, since the metadata lock will be held only for a fraction of a second and will be not held through more than a single API aren't we just fine with libvirtd holding the lock? The way I imagine this to work is the following:

Frankly, I've started implementing this with virtlockd already, but the changes I made so far simply do not feel right, e.g. I have to change lock_protocol.x so that virLockSpaceProtocolAcquireResourceArgs has this 'type' argument which tells virtlockd which byte we want to lock.

On 08/06/2018 10:30 AM, Daniel P. Berrangé wrote: >> >> So do we really need virtlockd for this? I mean, the whole purpose of it >> is to hold locks so that libvirtd can restart independently, without >> losing the lock attached. However, since the metadata lock will be held >> only for a fraction of a second and will be not held through more than a >> single API aren't we just fine with libvirtd holding the lock? The way I >> imagine this to work is the following: > > There were two reasons for virtlockd. The first was the persistent holding > of locks, but the other reasons is that POSIX fcntl() locks are horrific. > If one thread has an FD open with a lock held, if another thread opens and > closes the same underlying file, the first thread's lock will get dropped. > > We can't be confident that other threads won't open the file, so the only > way to be safe was to put locking in to a separate process where we know > exactly what all threads are doing. Ah. That's terrible. But IIUC avoidable by using OFDs instead (which are not available on BSD I presume).