Add exec option to mmap

[fatteneder]

mmap(exec=true) should allow one to generate a buffer that is writeable and (with enough precautions taken) can be executed.

We only implement exec=true for non-file-backed mmap, in which we request a new file handle from the OS, as otherwise we would have to adjust permissions on whatever file handle we received, which might not be possible in all cases.

---

MWE. Don't try this at home

# mwe_exec.jl
using Mmap

buf = mmap(Vector{UInt8}, 6, shared=false, exec=true)
function myidentity(x::Int32)
    ux = unsigned(x)
    buf[1] = 0xb8
    buf[2:5] .= reinterpret(UInt8, [ux])
    buf[6] = 0xc3
    return ccall(pointer(buf), Cint, ())
end

for x in rand(Int32, 5)
    @show x == myidentity(x)
end

julia> versioninfo()
Julia Version 1.12.0-DEV.5
Commit 1c13b8b90f (2024-02-17 22:20 UTC)
Platform Info:
  OS: Linux (x86_64-unknown-linux-gnu)
  CPU: 12 × Intel(R) Core(TM) i7-5820K CPU @ 3.30GHz
  WORD_SIZE: 64
  LLVM: libLLVM-16.0.6 (ORCJIT, haswell)
Threads: 1 default, 0 interactive, 1 GC (on 12 virtual cores)

julia> include("mwe_exec.jl")
x == myidentity(x) = true
x == myidentity(x) = true
x == myidentity(x) = true
x == myidentity(x) = true
x == myidentity(x) = true

• Anyone knows why this segfaults when either buf is marked const or buf initialization is moved into myidentity?

[vtjnash]

Most CPUs define that it is UB to execute code immediately after writing to the memory. They usually require some sort of flush, fence, and icache invalidation calls before that is legal

[fatteneder]

TODO:

• windows failure: SystemError: mmap: Access is denied.
  My guess is that the file handle is missing GENERIC_EXECUTE permission. (https://learn.microsoft.com/en-us/windows/win32/api/winbase/nf-winbase-createfilemappinga)
  Am I correct that we need to modify ios_file to make that work?
• macos failure: SystemError: memory mapping failed: Permission denied Don't see why this succeeds only 3/4 times.
• Might want to implement isexecutable for IOs?

[fatteneder]

Regarding windows failures:
I did some more reading about windows' CreatFileMapping and friends.
Assuming GENERIC_EXECUTE is really the problem, I would have to mess with the ACLs of the file handle that underlies the IO that we pass to mmap.
Not sure if that's even feasible to implement, but I rather don't.

Instead, I was thinking about limiting the exec option to only work with the Anonymous() dispatch, e.g. none file backed mmap.
This way I could just request a file handle using VirtualAlloc and the right access rights and all should be good (trademark).
No need for VirtualAlloc, because the existing setup just works with non-file-backed mmap.

Need to dust off my windows VM to see if I can implement that locally.

[fatteneder]

Regarding macOS failures:
I think the issue is a missing MAP_JIT flag, for which I can't find its value online.
Refs:

• https://developer.apple.com/documentation/bundleresources/entitlements/com_apple_security_cs_allow-jit
• https://developer.apple.com/documentation/apple-silicon/porting-just-in-time-compilers-to-apple-silicon#Restore-Write-Protections-Before-You-Execute-Instructions
• Expose mprotect in mmap python/cpython#114233 (comment).

[semarie]

I would like to note that WX permissions could be restricted

[fatteneder]

I would like to note that WX permissions could be restricted

I updated the PR description and I think this is no longer an issue, because we only implement exec for non-file-backed mmap, e.g. we query the handle with the desired rights ourselves. (but atm there is still an edge case I need to patch)