[PATCH 3/3] firmware: fw_base.S: Add common NMI trap handler

[Nylon Chen]

Radim Krcmar <rkrcmar at qti.qualcomm.com> 於 2026年2月2日週一 下午10:49寫道：
>
> 2026-01-29T00:13:43-08:00, Nylon Chen <nylon.chen at sifive.com>:
> > Add rnmi_handler assembly wrapper and sbi_rnmi_trap_handler default
> > C handler. The default handler prints diagnostics and hangs the hart.
> >
> > Co-developed-by: Zong Li <zong.li at sifive.com>
> > Signed-off-by: Zong Li <zong.li at sifive.com>
> > Suggested-by: Nick Hu <nick.hu at sifive.com>
> > Suggested-by: Samuel Holland <samuel.holland at sifive.com>
> > Signed-off-by: Nylon Chen <nylon.chen at sifive.com>
> > Signed-off-by: Yong-Xuan Wang <yongxuan.wang at sifive.com>
> > ---
> > diff --git a/firmware/fw_base.S b/firmware/fw_base.S
> > @@ -643,6 +643,70 @@ memcmp:
> > +     .section .entry, "ax", %progbits
> > +     .align 4
> > +     .globl sbi_rnmi_vector
> > +sbi_rnmi_vector:
> > +     /* Swap SP with MNSCRATCH */
> > +     csrrw   sp, CSR_MNSCRATCH, sp
> > +
> > +     /* Allocate space for full trap registers structure */
> > +     addi    sp, sp, -(SBI_TRAP_REGS_SIZE)
>
> [2/3] did:
>
>         rnmi_context_offset = sbi_scratch_alloc_offset(SBI_TRAP_REGS_SIZE);
>         rnmi_ctx = sbi_scratch_offset_ptr(scratch, rnmi_context_offset);
>         rnmi_sp  = (unsigned long)rnmi_ctx + SBI_TRAP_REGS_SIZE;
>         csr_write(CSR_MNSCRATCH, rnmi_sp);
>
> We can avoid the add/subtract dance, since it produces rnmi_ctx.
> A bigger issue is that it leaves 0 bytes for the stack of the callback.
>
> Is there a reason not to reuse the existing M-mode stack?
>
> Thanks.
Hi Radim

Thank you for the review.

Regarding your question "Is there a reason not to reuse the existing
M-mode stack?":

We cannot reuse the M-mode stack because RNMI can interrupt M-mode code itself
If RNMI reuses the M-mode stack, it would corrupt the interrupted
M-mode function's stack frame.

You're absolutely right about the stack space issue.

The current code allocates only SBI_TRAP_REGS_SIZE (280 bytes),
leaving 0 bytes for the C handler.

I'll fix this by allocating additional stack space:

#define RNMI_STACK_SIZE  2048
#define RNMI_CONTEXT_SIZE  (SBI_TRAP_REGS_SIZE + RNMI_STACK_SIZE)
rnmi_context_offset = sbi_scratch_alloc_offset(RNMI_CONTEXT_SIZE);
rnmi_sp = (unsigned long)rnmi_ctx + RNMI_CONTEXT_SIZE;

This gives the C handler 2048 bytes of stack space while keeping the
register save area separate.

Does this approach sound reasonable to you?