I have been using folder redirection (all folders that are relocatable) in win2k8r2 onto my samba 3.4.8. The underlying stack (of the working fs, not the samba binaries) is jfs/xfs-->dmcrypt-->mdraid-->blockdev.
Both with jfs and xfs, after only a few days of operation, the client would start complaining (sync issues, "disk full", etc), and closer examination would show that several smbd's were stuck (impervious to kill -9), apparently unable to sync. This would sync lock the entire fs, preventing unmounting and safe rebooting, often leading to file(-system) corruption.
After one incident with each fs and never having had similar issues before (though not using folder redirection before, thus only having light io loads on samba) I decided to ditch 3.4.8, but ran into a blocker on 3.5.4.
Since then I have upgraded the kernel from 2.6.32-hardened-r10 to 2.6.34-hardened-r1, but due to the risk of major data loss, I am quite reluctant to give 3.4.8 another try on the new kernel. It's also of note that -only- smbd's would lock up, no other processes would show similar behaviour.
Unkillable smbd processes (state D, survive kill -9) is a kernel bug. This is not a Samba problem.
Closing as invalid, please report to the kernel people.