The bug in CTR has been resolved! Herbert wrote a better patch than the one I did. After some minor fixes to the patch, CTR now does multi-page processing properly. The same idea should be applied to Salsa20, which also has the same bug.
A large test vectors (large in the sense that it forces multi-page access) of 4100 bytes was added to tcrypt.h to test the code.
The patches are available here. Unfortunately, the patches results in substantial bloat of tcrypt.ko.