Re: [PATCH v4 05/16] riscv: Extend cpufeature.c to detect vendor extensions

[Conor Dooley]

On Wed, May 01, 2024 at 10:51:38AM -0700, Charlie Jenkins wrote:
> On Wed, May 01, 2024 at 09:44:15AM -0700, Evan Green wrote:
> > On Fri, Apr 26, 2024 at 2:29 PM Charlie Jenkins <charlie@xxxxxxxxxxxx> wrote:
> > > +       for (int i = 0; i < riscv_isa_vendor_ext_list_size; i++) {
> > > +               const struct riscv_isa_vendor_ext_data_list *ext_list = riscv_isa_vendor_ext_list[i];
> > > +
> > > +               if (bitmap_empty(ext_list->vendor_bitmap, ext_list->bitmap_size))
> > > +                       bitmap_copy(ext_list->vendor_bitmap,
> > > +                                   ext_list->per_hart_vendor_bitmap[cpu].isa,
> > > +                                   ext_list->bitmap_size);
> > 
> > Could you get into trouble here if the set of vendor extensions
> > reduces to zero, and then becomes non-zero? To illustrate, consider
> > these masks:
> > cpu 0: 0x0000C000
> > cpu 1: 0x00000003 <<< vendor_bitmap ANDs out to 0
> > cpu 2: 0x00000010 <<< oops, we end up copying this into vendor_bitmap
> > 
> 
> Huh that's a good point. The standard extensions have that same bug too?
> 
> 	if (bitmap_empty(riscv_isa, RISCV_ISA_EXT_MAX))
> 		bitmap_copy(riscv_isa, isainfo->isa, RISCV_ISA_EXT_MAX);
> 	else
> 		bitmap_and(riscv_isa, riscv_isa, isainfo->isa, RISCV_ISA_EXT_MAX);

I suppose it could in theory, but the boot hart needs ima to even get
this far. I think you'd only end up with this happening if there were
enabled harts that supported rvXXe, but I don't think we even add those
to the possible set of CPUs. I'll have to check.

Attachment: signature.asc
Description: PGP signature