Re: [PATCH 09/11] include/exec: added functions to the stubs in exec-all.h

[Richard Henderson]

On 5/13/21 9:03 AM, Lucas Mateus Martins Araujo e Castro wrote:
> tlb_set_page is called by many ppc_hash64_handle_mmu_fault, 
> ppc_radix64_handle_mmu_fault and ppc_hash32_handle_mmu_fault, all of which from 
> what I've seen are only used inside #if defined(CONFIG_SOFTMMU).

tlb_set_page should only be called from one place: ppc_cpu_tlb_fill.  The other 
functions should fill in data, much like get_physical_address.


> So what is the 
> best way to deal with these tlb_set_page calls? Should these part of the 
> _handle_mmu_fault functions never be reached or should these functions never be 
> called?

There is some duplication between get_physical_address* and *handle_mmu_fault 
that should be fixed.

What should be happening is that you have one function (per mmu type) that 
takes a virtual address and resolves a physical address. This bit of code 
should be written so that it is usable by both 
CPUClass.get_phys_page_attrs_debug and TCGCPUOps.tlb_fill.  It appears as if 
ppc_radix64_xlate is the right interface for this.

It appears that real mode handling is duplicated between hash64 and radix64, 
which could be unified.

You should only call tlb_set_page from TCGCPUOps.tlb_fill, aka 
ppc_cpu_tlb_fill.  TCGCPUOps.tlb_fill is obviously TCG only.

The version you are looking at here is system emulation specific (sysemu, 
!defined(CONFIG_USER_ONLY)).  There is a second version of this function, with 
the same signature, that is used for user emulation in the helpfully named 
user_only_helper.c.


r~