Fix #104

1. Refactor shebang parsing logic: Extract the duplicate binfmt_script parsing logic from main.c and exec.c into a shared helper function elf_parse_shebang defined in elf.c and declared in elf.h.
2. Provide SEQPACKET test: Add test cases to test-socket.c to cover UNIX domain SOCK_SEQPACKET sockets and socketpairs, verifying that getsockopt SO_TYPE correctly returns SOCK_SEQPACKET.

Changes

[MODIFY] elf.h

• Declare the shared helper function:

  int elf_parse_shebang(const char *host_path, char *interp_out, size_t interp_sz, char *arg_out, size_t arg_sz);

[MODIFY] elf.c

• Implement elf_parse_shebang to:
  • Open the shebang file on the host.
  • Read the script header (up to 512 bytes).
  • Verify it starts with #!.
  • Parse the interpreter path and optional single argument, trimming leading/trailing whitespaces and carriage returns correctly.
  • Copy results to output buffers.
  • Return 1 on success, 0 if not a shebang script, or negative error code (like -ENOENT, -ENOEXEC) on failure.

[MODIFY] main.c

• Refactor shebang resolution loop to call elf_parse_shebang and reuse its outputs.

[MODIFY] exec.c

• Refactor fallback logic to call elf_parse_shebang and reuse its outputs.

[MODIFY] test-socket.c

• Add Test 12: socketpair(AF_UNIX, SOCK_SEQPACKET, 0, seq_sv) to verify creation, data exchange, and getsockopt SO_TYPE returning SOCK_SEQPACKET.
• Add Test 13: socket(AF_UNIX, SOCK_SEQPACKET, 0) to verify creation and getsockopt SO_TYPE returning SOCK_SEQPACKET.

Verification Plan

Automated Tests

• Run the compiled test binary under elfuse to verify all tests (including the new SEQPACKET test cases) pass:

  build/elfuse build/test-socket

sys_socket (src/syscall/net.c:97) and sys_socketpair (:174) now downgrade
every AF_UNIX SOCK_SEQPACKET request to SOCK_STREAM, while the option
cache keeps the original type (:127, :156, :210-211) so
getsockopt(SO_TYPE) still reports SOCK_SEQPACKET.

On AF_UNIX the only difference between the two types is that SEQPACKET
preserves message boundaries and STREAM does not (both are connection-oriented,
reliable, ordered). So a guest that assumes "one recv = one whole message" —
the reason to pick SEQPACKET in the first place — breaks via coalescing/short
reads, a recvmsg-until-MSG_EOR loop that never ends, or flipped small-buffer
truncation. The breakage is timing-dependent (intermittent) and, because
SO_TYPE lies, undetectable from inside the guest. Affected: systemd
ListenSequentialPacket= services, socketpair(SEQPACKET) privsep/worker
channels, any daemon using SEQPACKET for framing. Strict ping-pong users are
fine (why STREAM works for the common case). The old behavior failed cleanly at
socket() (EPROTOTYPE); this trades that for silently corrupting the
framing-dependent minority — defensible for a compat layer, but it should be a
conscious choice, and boundaries can be kept.

Two possible way to implement SEQPACKET:

• socketpair(SEQPACKET) → SOCK_DGRAM, not SOCK_STREAM (~1 line at
  net.c:174). macOS supports AF_UNIX DGRAM socketpairs; local datagrams are
  reliable, ordered, and keep message boundaries, and the pairing already
  supplies the connection semantics SEQPACKET adds over DGRAM. Fixes the most
  common pattern at zero framing cost.
• listen/accept servers need connection semantics DGRAM lacks → SOCK_STREAM
  • userspace length-prefix framing. Larger effort; only if a real server-pattern
    guest appears.

Changes Made

shebang / binfmt_script Refactoring

• elf.h: Declared the helper function elf_parse_shebang to parse script files starting with a shebang (#!).
• elf.c: Implemented elf_parse_shebang using safe buffer/string utilities (str_copy_trunc) to avoid security violations (such as strcpy).
• main.c: Replaced duplicate shebang parsing logic with a call to elf_parse_shebang.
• exec.c: Replaced duplicate shebang parsing logic with a call to elf_parse_shebang.

SOCK_SEQPACKET Emulation & Boundary Preservation Tests

• net.c: Updated sys_socketpair to downgrade UNIX domain SOCK_SEQPACKET socketpairs to SOCK_DGRAM instead of SOCK_STREAM. Since macOS fully supports AF_UNIX datagram socketpairs (which are reliable, ordered, and preserve message boundaries), this preserves sequential packet connection semantics at zero framing cost.
• test-socket.c: Added/enhanced the test cases:
  • Test 12: Verifies creation of UNIX domain SOCK_SEQPACKET socketpair using consecutive writes, checking that reading from the socketpair preserves message boundaries (returns distinct messages and sizes instead of coalescing/short reads) and that getsockopt SO_TYPE correctly returns SOCK_SEQPACKET.
  • Test 13: Verifies creation of UNIX domain SOCK_SEQPACKET socket, and ensures getsockopt SO_TYPE correctly returns SOCK_SEQPACKET.

Verification & Testing

Security Check

All security checks passed:

bash .ci/check-security.sh
# Output: Security checks passed.

Compilation

1. Compiled the refactored elfuse binary natively on macOS:

   make elfuse

Execution Results

Executed test-socket under elfuse showing that both the standard tests and the updated SOCK_SEQPACKET boundary tests pass cleanly:

test-socket: 1. socketpair(AF_UNIX)... PASS (fds=3,4)
test-socket: 2. write/read through socketpair... PASS
test-socket: 3. getsockopt(SO_TYPE)... PASS
test-socket: 4. getsockname/getpeername... PASS
test-socket: 5. sendmsg/recvmsg single iovec... PASS
test-socket: 6. sendmsg/recvmsg two iovecs... PASS
test-socket: 7. sendmmsg/recvmmsg single message... PASS
test-socket: 8. sendto/recvfrom AF_UNIX dgram... PASS
test-socket: 9. recvmsg single iovec preserves MSG_TRUNC... PASS
test-socket: 10. getsockopt(SO_ERROR) read-and-clear... PASS
test-socket: 11. shutdown(SHUT_WR) -> EOF... PASS
test-socket: 12. socketpair(AF_UNIX, SOCK_SEQPACKET)... PASS
test-socket: 13. socket(AF_UNIX, SOCK_SEQPACKET)... PASS
test-socket: all tests passed -- PASS