Re: [PATCH 2/2] riscv: mm: Update mmap_rnd_bits_max

[Sami Tolvanen]

On Mon, Oct 2, 2023 at 12:02 AM Alexandre Ghiti <alex@xxxxxxxx> wrote:
>
> On 01/10/2023 17:19, Pedro Falcato wrote:
> > On Sun, Oct 1, 2023 at 2:51 AM Kees Cook <keescook@xxxxxxxxxxxx> wrote:
> >> On Sat, Sep 30, 2023 at 10:02:35AM +0100, Conor Dooley wrote:
> >>> On Fri, Sep 29, 2023 at 03:52:22PM -0700, Sami Tolvanen wrote:
> >>>> On Fri, Sep 29, 2023 at 2:54 PM Kees Cook <keescook@xxxxxxxxxxxx> wrote:
> >>>>> On Fri, Sep 29, 2023 at 09:11:58PM +0000, Sami Tolvanen wrote:
> >>>>>> ARCH_MMAP_RND_BITS_MAX is based on Sv39, which leaves a few
> >>>>>> potential bits of mmap randomness on the table if we end up enabling
> >>>>>> 4/5-level paging. Update mmap_rnd_bits_max to take the final address
> >>>>>> space size into account. This increases mmap_rnd_bits_max from 24 to
> >>>>>> 33 with Sv48/57.
> >>>>>>
> >>>>>> Signed-off-by: Sami Tolvanen <samitolvanen@xxxxxxxxxx>
> >>>>> I like this. Is RISCV the only arch where the paging level can be chosen
> >>>>> at boot time?
> >>>> I haven't seen this elsewhere, but I also haven't looked at all the
> >>>> other architectures that closely. arm64 does something interesting
> >>>> with ARM64_VA_BITS_52, but I think we can still handle that in
> >>>> Kconfig.
> >>> AFAIU, x86-64 can do this also:
> >>>
> >>>        no4lvl          [RISCV] Disable 4-level and 5-level paging modes. Forces
> >>>                        kernel to use 3-level paging instead.
> >>>
> >>>        no5lvl          [X86-64,RISCV] Disable 5-level paging mode. Forces
> >>>                        kernel to use 4-level paging instead.
> >> Ah-ha! Okay, well, then let's track this idea:
> >> https://github.com/KSPP/linux/issues/346
> > (Replying here for visibility, tell me if you want to move this
> > discussion to github)
> >
> > AIUI, x86 cannot do this for compat reasons. Even if you enable LA57,
> > mmap only gives you < 48-bit addresses, for compatibility with things
> > like JITs, etc that stash information in the upper 16 bits. You need
> > to pass a > 48-bit mmap hint to get 57-bit addresses.
> >
> > I imagine riscv does not have this issue yet, due to little
> > accumulated cruft, but it may be wise to check against popular JITters
> > for these problems on riscv code.
> >
>
> We already encountered those issues and the same solution was recently
> merged (restrict to sv48 unless otherwise specified):
> https://lore.kernel.org/all/20230809232218.849726-1-charlie@xxxxxxxxxxxx/

We recently ran into this issue when bringing up Android as well
because qemu defaults to Sv57 and some userspace bits weren't happy
with >48-bit mmap addresses.

Note that this patch uses MMAP_VA_BITS, which is 48 for both Sv48 and
Sv57, which is why mmap_rnd_bits_max will be 33 even with Sv57.

Sami