Re: [PATCH 5/6] driver core: Add __alloc_size hint to devm allocators

[Sami Tolvanen]

On Thu, Feb 2, 2023 at 11:31 AM Nick Desaulniers
<ndesaulniers@xxxxxxxxxx> wrote:
>
> On Thu, Feb 2, 2023 at 11:20 AM Ard Biesheuvel <ardb@xxxxxxxxxx> wrote:
> >
> > On Thu, 2 Feb 2023 at 20:10, Kees Cook <keescook@xxxxxxxxxxxx> wrote:
> > >
> > > On Thu, Feb 02, 2023 at 10:56:29AM -0800, John Stultz wrote:
> > > > That said, making sense of the error message isn't completely trivial
> > > > either. I've been seeing a few cases recently of some of the new
> > > > compiler tooling (I pinged you earlier on a CFI one) causing errors
> > > > that developers aren't really sure how to address.  I know sometimes
> > > > it's not easy to surface the errors with context to what was wrong,
> > > > but at the risk of intense bike shedding, is there some way to provide
> > > > something like "Likely array bounds error" instead of just "BRK
> > > > handler: Fatal exception"?
> > >
> > > Yeah, this is a result of the size trade-off that resulted in config
> > > CONFIG_UBSAN_TRAP -- there ends up being no message about what went
> > > wrong. I'd really like to have cleaner handling of this -- perhaps what
> > > was done for KCFI could be applied to UBSAN as well, though this is an
> > > area I don't know well myself. (i.e. encoding "this was a UBSAN trap"
> > > in the trap itself.)
> > >
> > > Sami or Ard, is this something that could be improved for arm64?
> > >
> >
> > -ENOCONTEXT, so I am going to assume this is about runtime
> > instrumentation that needs some kind of 'panic' function which it will
> > invoke if some condition is met that should never occur?
> >
> > We already use brk with different immediate values in the opcode, so
> > the arch layer already has what we need. Is this a limitation in the
> > compiler, perhaps, where it always emits the same brk opcode?
>
> Yeah, we'd need to update both the compiler to produce the encoding,
> and the kernel to recognize the encoding and do something special.

A quick look at Clang's source code suggests that Intrinsic::ubsantrap
already accepts the handler ID (from the SanitizerHandler enum) as an
argument and the arm64 LLVM back-end appears to encode the value as an
immediate for the brk instruction. I didn't confirm that this actually
works, but perhaps we just need to teach the kernel about the possible
values?

Sami