feat(libkernel): implement getcontext, setcontext and makecontext - #1841

What

getcontext (dar03AOn+nM) and setcontext (2VRU7xiqLO8) in libkernel and makecontext (GipcbdDM5cI) in libSceLibcInternal are missing, so the 15 titles that #1578 lists for them stop at start-up. Part of #1578; no open PR implements them.

In PPSA99039 (EVO Player 0.10.0) every call site is in the statically linked OpenSSL's ASYNC code (crypto/async/arch/async_posix.c): ASYNC_is_capable, async_fibre_makecontext, and async_fibre_swapcontext, which enters a new fibre with setcontext and switches afterwards with _setjmp/_longjmp.

getcontext is written in assembly, so it saves every register as it was at the call. It stores the context as FreeBSD's sys_getcontext does: rax and rdx as 0, the carry flag clear, rip and rsp as after the return, the FXSAVE image, mc_len, and the signal mask (through sigprocmask). Like UC_COPY_SIZE, it writes only the 0x4c0 bytes before uc_link, so uc_link, uc_stack and uc_flags keep what the caller stored.

setcontext fails with EINVAL on a null context, a wrong mc_len or unknown mc_flags bits, as set_mcontext does. It installs the context's signal mask, masks MXCSR with the host's MXCSR_MASK as fpusetregs does (FXRSTOR faults on reserved bits), and returns through iretq: every general register, rip, rsp and the PSL_USERCHANGE flags come from the context, and nothing is written below the target stack pointer, which matters for a context taken in a signal handler.

makecontext follows FreeBSD 11's lib/libc/amd64/gen/makecontext.c: at most six arguments and a stack of at least MINSIGSTKSZ (2048) bytes, otherwise mc_len becomes 0. Its wrapper calls the function, then setcontext(uc_link), or exit(0) when there is no link. It lives in the libSceLibcInternal module, which links libkernel.

The three use the layout of the exception handler's context (mcontext at 0x40, 0x500 bytes), which System/include/GuestContext.hpp now declares for them. shadPS4 passes the same layout to sigaction handlers, and Unity's PS5Util reads mc_rsp at 0xf8. Homebrew built against FreeBSD's headers expects FreeBSD's 0x370-byte layout instead: PPSA99039's ASYNC_is_capable passes a 0x370-byte stack buffer. This is recorded in TechnicalDebt.

Exception.cpp keeps its own copy of the structures for now: #1431, #1562 and #1802 all change them, and moving them here would conflict with all three. Once those land, Exception.cpp can include the header in a follow-up.

Windows stack bounds. Windows checks guard-page growth against StackBase and DeallocationStack in the TEB, and exception dispatch and unwinding against StackLimit and StackBase, so the TEB must describe the stack the thread runs on, as libSceFiber keeps it for fiber stacks. A new libc helper, GuestStacks, does this for setcontext and longjmp. makecontext registers its stack and pins it writable, as fiber stacks are pinned. A switch to a registered stack takes its bounds; any other switch takes the committed region around the target rsp from VirtualQuery, which for a thread stack is its current StackLimit to StackBase. The bounds change after the last write to the old stack. longjmp needs this because OpenSSL leaves a fibre with _longjmp: without it the thread would run on its own stack with the fibre's bounds, and its next guard-page hit would raise STATUS_GUARD_PAGE_VIOLATION. Linux gets the three functions and no other change.

TechnicalDebt records the unverified layout and the FreeBSD-header mismatch, the FreeBSD source of makecontext, what the FXSAVE-only state and setcontext leave out, the Windows pin, and the bounds fallback.

A test merge of this branch with every other open pull request on 364a56b6 conflicts only with #1128, which reformats the whole tree, including the setjmp/longjmp assembly. The overlap report's __error_nid_postfix entries are declarations of the existing libc export, as in nine files on main; this PR does not export it.

Tested

Windows 11 (26200), MinGW-w64 GCC 15.2.0 (WinLibs r7), Release, full build with BUILD_TESTING=ON on 364a56b6: ctest 483/483 passed. guest_user_context and guest_user_context_exit passed 200 runs in a row (--repeat until-fail:200).

guest_user_context covers:

• errors: null contexts, a wrong mc_len, unknown mc_flags; makecontext with seven arguments, a 2047-byte stack, no stack, and an invalid context;
• capture: an assembly caller loads known values into every register and sets the carry flag; the test checks each saved register, rax/rdx 0, carry clear, rip, rsp, MXCSR, the x87 control word, the signal mask, the zeroed fields, and that nothing from uc_link on changes;
• restore: a context with patterns in every general register, XMM0-15, MXCSR (with a reserved bit set), the x87 control word and the flags lands in an assembly stub that records them;
• makecontext with six 64-bit arguments, an aligned entry stack and a return through uc_link;
• OpenSSL's pattern: setcontext into a job, then four rounds of setjmp/longjmp both ways;
• the signal mask restored by setcontext;
• on Windows: the TEB bounds after every switch, a C++ exception thrown and caught on the context stacks, and about 512 KiB of new host stack after returning.

guest_user_context_exit checks that a context without uc_link exits with 0.

Mutation check, all caught: setcontext or longjmp keeping the old bounds, makecontext not registering its stack, getcontext keeping the carry flag, writing past uc_link or saving rcx as rbx, setcontext keeping reserved MXCSR bits, ignoring the context's flags, not loading r11, keeping the current signal mask or accepting a wrong mc_len, makecontext shifting its arguments or misaligning the entry stack, and a context without uc_link aborting instead of exiting.

tools/import_audit.py on PPSA99039 (EVO Player 0.10.0): 24 absent imports on main, 21 with this branch; the three are resolved, and the built prx export the NIDs above. No title run: PPSA99039 still has other absent imports.

Not run on Linux (no Linux host here). The Linux path is the same assembly without the TEB call; CI runs it.

Checklist

• Based on current main; no other open PR implements the same functions
• One topic per PR; follow-ups go in a new PR
• No comments in code except technical debt
• Unimplemented paths throw (NotImplemented_nid_no_patch); silent stubs are listed in TechnicalDebt
• No notes, investigation .md files or images added to the repository (attach them to this PR)
• New third-party code is a submodule built from source
• Generic behaviour, not specific to one title
• Depends on: none
• AI-assisted: yes