search

riscv / riscv-fast-interrupt

Proposal for a RISC-V Core-Local Interrupt Controller (CLIC)
https://jira.riscv.org/browse/RVG-63
Creative Commons Attribution 4.0 International
245 stars 49 forks source link

xcause.inhv conflicts with enabling CLIC per mode #348

Open JamesKenneyImperas opened 1 year ago

JamesKenneyImperas commented 1 year ago

A recent change to the specification redefines the inhv field so that it is held with the modeX exception mode instead of the modeY interrupt mode. This makes sense because this information is required by the exception handler, not the interrupt handler.

Currently the CLIC can only be globally enabled, but the specification states that in future this might change to allow CLIC to be enabled per privilege level. The new definition will therefore impose an extra constraint that if CLIC is to be enabled at any privilege level, it must also be enabled at all higher privilege levels (because the inhv bit only exists for a privilege level if the CLIC is enabled for that privilege level).

It feels to me that with this change the inhv bit is in the wrong place: it should really be in a fixed position per mode in mstatus/mstatush instead (like uie/mie/sie etc)

kasanovic commented 1 year ago

inhv is written in the higher mode if the lower mode has CLIC enabled and a table trap occurs. The higher mode does not need to have CLIC enabled itself to use inhv when servicing traps for a lower mode. If we later (post-1.0) proceed with optional CLIC per mode, we will define extensions specifying the control/visibility of lower-mode CLIC activity in the higher mode when CLIC is not available to the higher mode.

JamesKenneyImperas commented 1 year ago

Hi Krste,

Okay, but some sections are going to need some careful rework later. For example, section 7.2.4 says this:

For backwards compatibility in systems supporting both CLINT and CLIC modes, when switching to CLINT mode the new CLIC xcause state fields (xinhv and xpil) are zeroed.

This doesn't work when CLIC is enabled per privilege mode, because the xinhv field is potentially owned by a different privilege mode.

In CLIC mode, when a trap is taken, xcause has the CLIC format and the xcause fields are updated. On the other hand, when not in CLIC mode, xcause has the CLINT mode format.

This clearly won't be true when the CLIC is enabled per privilege mode if the higher privilege level that "owns" xinhv is in CLINT mode.

Thanks.