Re: Why is text_mutex used in jump_label_transform for x86_64

[Will Deacon]

On Mon, Apr 06, 2020 at 11:15:51AM +0200, Peter Zijlstra wrote:
> On Mon, Apr 06, 2020 at 04:39:11PM +0800, chengjian (D) wrote:
> > 
> > On 2020/3/20 18:27, Peter Zijlstra wrote:
> > > It depends on the architecture details of how self-modifying code works.
> > > In particular, x86 is a variable instruction length architecture and
> > > needs extreme care -- it's implementation requires there only be a
> > > single text modifier at any one time, hence the use of text_mutex.
> > > 
> > > ARM64 OTOH is, like most RISC based architectures, a fixed width
> > > instruction architecture. And in particular it can re-write certain
> > > (branch) instructions with impunity (see their
> > > aarch64_insn_patch_text_nosync()). Which is why they don't need
> > > additional serialization.
> > 
> > Hi, Peter
> > 
> > Thank you very much for your reply.
> > 
> > X86 is a variable-length instruction, only one byte modification of the
> > instruction
> > can be regarded as atomic. so we must be very careful when modifying
> > instructions
> > concurrently.
> 
> Close enough.
> 
> > For other architectures such as ARM64, the modification of some instructions
> > can be
> > considered atomic, (Eg. nop -> jmp/b). The set of instructions that can be
> > executed
> > by one thread of execution as they are being modified by another thread of
> > execution
> > without requiring explicit synchronization.
> > 
> > In ARM64 Architecture Reference Manual, I find that:
> > ÂÂÂ Concurrent modification and execution of instructions can lead to the
> > resulting instruction performing any behavior
> > ÂÂÂ that can be achieved by executing any sequence of instructions that can
> > be executed from the same Exception level,
> > ÂÂÂ except where each of the instruction before modification and the
> > instruction after modification is one of a B, BL, BRK,
> > ÂÂÂ HVC, ISB, NOP, SMC, or SVC instruction.
> > ÂÂÂ For the B, BL, BRK, HVC, ISB, NOP, SMC, and SVC instructions the
> > architecture guarantees that, after modification of the
> > ÂÂÂ instruction, behavior is consistent with execution of either:
> > ÂÂÂ â The instruction originally fetched.
> > ÂÂÂ â A fetch of the modified instruction
> > 
> > So we can safely modify jump_label for ARM64(from NOP to b or form b to
> > NOP).
> > 
> > Is my understanding correct?
> 
> I think so; but I'm really not much of an ARM64 person. FWIW I think I
> remember Will saying the same is true of ARM (32bit) and they could
> implement the same optimization, but so far nobody has bothered doing
> so. But please, ask an ARM64 maintainer and don't take my word for this.

On 32-bit there are complications with Thumb-2 instructions where you can
have a mixture of 16-bit and 32-bit encodings, so you have to be pretty
careful there.

For arm64, we have aarch64_insn_patch_text_nosync() which we use to toggle
jump labels.

Will