Ralf Baechle wrote:
On Wed, Oct 22, 2008 at 02:35:21AM +0100, Maciej W. Rozycki wrote:
This probe is necessary, because for a VIVT I-cache, code from there may
be executed even if there is no mapping stored for the virtual address of
the instruction in the TLB anymore. However this trap handler wants to
read the instruction word from the memory and obviously this goes through
the D-cache which is not virtually tagged. As such a TLB refill exception
would happen if the mapping was indeed absent.
However, please note that this piece of code runs at the exception level
and therefore such a scenario would qualify as a nested exception. Which
means the general exception vector would be used and the TLBL or TLBS
handler invoked as appropriate. Neither of which are currently prepared
to do a refill. Changing that would be rather trivial as it boils down to
checking the value of cp0.index.p and executiong TLBWR rather than TLBWI
as usual, but that is in the fast path, so we do not want to waste cycles
for such a corner case as RDHWR emulation.
To clarify, this behaviour of hitting in an VIVT I-cache even though there
is no address translation in the TLB is allowed but not required by the
the MIPS architecture spec. From a software perspective it's a bit
quirky but it allows faster pipeline implementations. The currently
supported VIVT I-cache processors are the SB1, 20K and 25K. The SB1
has this behaviour; of the 20K and 25K I don't know.
And to perhaps clarify even more, the octeon too has this property, so
the probe of the TLB is needed there.