Diagnostic verification of vnode VOCALLs

[Martin Husemann <martin%duskware.de@localhost>]

Hey folks,

I sometimes (like once every other month) see a kernel crash on sparc64
that looks like the kernel jumped through a bogus function pointer.
At the time of the crash, the stack seems to be corrupted, at least
ddb backtrace does not work. Register window information seems to point
at a call stack from the vnode cleaner thread via VCALL() -> VOCALL and
the issue might be anything from a compiler bug to random memory corruption
or just some use-after-free vnode/filesystem bug.

To catch it earlier (and have better ddb data available) I would like
to avoid the deadly jump and have the VOCALL macro verify integrity of
the ops vector.

I thought about a MD macro to KASSERT general function pointers (like on
sparc* check for % 4 == 0, test for >= kernel_start and maybe <= the
end of the module map region), and for DEBUG kernels that makro could
even do a pmap_extract().

Additionally (and maybe the easiest start) we could add some magic pointer
value to the start (and end?) of the ops vector and have VOCALL verify that
magic value.

Is that a sane idea?
Could we easily identify all ops vectors?


Martin
P.S.: in sys/sys/vnode.h, the comments about VOCALL not being used seem
to be exagerated, or I misread them.