Thank you very much for the clarification!
I found tlb_set_page with size != TARGET_PAGE_SIZE makes the translation way too slow; the Linux doesn't seem to boot.
If that's the only way to reduce PMP granularity to less than TARGET_PAGE_SIZE,
Can we just set the PMP default granularity to TARGET_PAGE_SIZE as it did before?
OR
Can we bypass the partial match violation when size is unknown? (check the starting address only)
I think both of the options does not exactly match with the ISA specification,
but given that size=0 always causes the problem, I want it to be fixed as soon as possible.
Any thoughts would be appreciated!
On Mon, Oct 7, 2019, 6:00 AM Richard Henderson <
address@hidden> wrote:
On 10/6/19 10:28 PM, Dayeol Lee wrote:
> riscv_cpu_tlb_fill() uses the `size` parameter to check PMP violation
> using pmp_hart_has_privs().
> However, the size passed from tlb_fill(), which is called by
> get_page_addr_code(), is always a hard-coded value 0.
> This causes a false PMP violation if the instruction presents on a
> PMP boundary.
>
> In order to fix, simply correct the size to 4 if the access_type is
> MMU_INST_FETCH.
That's not correct.
In general, size 0 means "unknown size". In this case, the one tlb lookup is
going to be used by lots of instructions -- everything that fits on the page.
If you want to support PMP on things that are not page boundaries, then you
will also have to call tlb_set_page with size != TARGET_PAGE_SIZE.
Fixing that will cause instructions within that page to be executed one at a
time, which also means they will be tlb_fill'd one at a time, which means that
you'll get the correct size value.
Which will be 2 or 4, depending on whether the configuration supports the
Compressed extension, and not just 4.
r~