[PATCH] arm64: net: bpf: don't BUG() on large shifts

[Will Deacon]

On Tue, Jan 12, 2016 at 08:45:43PM -0800, Z Lim wrote:
> On Tue, Jan 12, 2016 at 11:23 AM, Alexei Starovoitov
> <alexei.starovoitov at gmail.com> wrote:
> > On Tue, Jan 12, 2016 at 05:17:10PM +0000, Will Deacon wrote:
> >> On Fri, Jan 08, 2016 at 11:09:44AM -0800, Alexei Starovoitov wrote:
> >> > On Fri, Jan 08, 2016 at 03:44:23PM +0000, Will Deacon wrote:
> >> > > On Tue, Jan 05, 2016 at 06:39:03PM +0100, Rabin Vincent wrote:
> >> > > > Attempting to generate UBFM/SBFM instructions with shifts that can't be
> >> > > > encoded in the immediate fields of the opcodes leads to a trigger of a
> >> > > > BUG() in the instruction generation code.  As the ARMv8 ARM says: "The
> >> > > > shift amounts must be in the range 0 to one less than the register width
> >> > > > of the instruction, inclusive."  Make the JIT reject unencodable shifts
> >> > > > instead of crashing.
> >> > >
> >> > > I moaned about those BUG_ONs when they were introduced:
> >> > >
> >> > >   https://lkml.org/lkml/2014/7/17/438
> >> > >
> >> > > The response then was that the verifier would catch these issues so
> >> > > there was nothing to worry about. Has something changed so that is no
> >> > > longer the case? Do we need to consider a different way of rejecting
> >> > > invalid instructions at the encoding stage rather than bringing down the
> >> > > kernel?
> >> >
> >> > that discussion lead to replacement of all BUG_ONs in
> >> > arch/arm64/net/bpf_jit_comp.c with pr_err_once(), but looks like
> >> > arch/arm64/kernel/insn.c wasn't addressed.
> >> > The amount of BUG_ONs there is indeed overkill regardless of what
> >> > verifier and other JITs do. btw, x64 JIT doesn't have runtime BUG_ONs.
> >>
> >> Maybe, but insn.c is also used by the alternatives patching code, so we
> >> really need a way to communicate failure back to the BPF JIT when passed
> >> an invalid instruction description.
> >
> > agree. I think there are several options to achieve that after
> > all BUG_ONs are removed:
> > - change interface for all insn generating macros to check for
> >   AARCH64_BREAK_FAULT opcode as error.
> >   That will require all of emit*() functions in bpf_jit_comp.c to
> >   be changed to accept/return error.
> >   Overall that looks like massive change.
> > - ignore AARCH64_BREAK_FAULT during emit and add another pass after
> >   all code is generated. If such insn is found in a jited code,
> >   discard the jit.
> >   I think that's better option.
> >
> > Zi, any comments?
> >
> 
> Alexei, agreed. Second approach is cleaner. Full disclosure: I did not
> look at other callers beyond JIT.
> 
> Separately, sounds like there's now preference and consensus to
> removing all BUGs and BUG_ONs in insn.c. Did a quick grep of insn.c
> and noticed a legacy instance, followed by many introduced around the
> same time as JIT, and new additions since.
> 
> Will, any thoughts on the following replacement scheme?
> 
> BUG_ON() for codegen ==> pr_err(); return AARCH64_BREAK_FAULT;
> BUG() for decoding ==> leave as is.
> remaining BUG_ON() ==> leave as is.

That sounds good to me, thanks.

Will