Merge pull request #15125 from ziglang/hcs-win-poc

coff: add hot-code swapping PoC
This commit is contained in:
Jakub Konka 2023-03-31 00:38:30 +02:00 committed by GitHub
commit 5b82b40043
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
6 changed files with 383 additions and 149 deletions

View File

@ -1514,6 +1514,24 @@ pub fn VirtualProtect(lpAddress: ?LPVOID, dwSize: SIZE_T, flNewProtect: DWORD, l
}
}
pub fn VirtualProtectEx(handle: HANDLE, addr: ?LPVOID, size: SIZE_T, new_prot: DWORD) VirtualProtectError!DWORD {
var old_prot: DWORD = undefined;
var out_addr = addr;
var out_size = size;
switch (ntdll.NtProtectVirtualMemory(
handle,
&out_addr,
&out_size,
new_prot,
&old_prot,
)) {
.SUCCESS => return old_prot,
.INVALID_ADDRESS => return error.InvalidAddress,
// TODO: map errors
else => |rc| return std.os.windows.unexpectedStatus(rc),
}
}
pub const VirtualQueryError = error{Unexpected};
pub fn VirtualQuery(lpAddress: ?LPVOID, lpBuffer: PMEMORY_BASIC_INFORMATION, dwLength: SIZE_T) VirtualQueryError!SIZE_T {
@ -4457,3 +4475,184 @@ pub const MODULEENTRY32 = extern struct {
szModule: [MAX_MODULE_NAME32 + 1]CHAR,
szExePath: [MAX_PATH]CHAR,
};
pub const THREADINFOCLASS = enum(c_int) {
ThreadBasicInformation,
ThreadTimes,
ThreadPriority,
ThreadBasePriority,
ThreadAffinityMask,
ThreadImpersonationToken,
ThreadDescriptorTableEntry,
ThreadEnableAlignmentFaultFixup,
ThreadEventPair_Reusable,
ThreadQuerySetWin32StartAddress,
ThreadZeroTlsCell,
ThreadPerformanceCount,
ThreadAmILastThread,
ThreadIdealProcessor,
ThreadPriorityBoost,
ThreadSetTlsArrayAddress,
ThreadIsIoPending,
// Windows 2000+ from here
ThreadHideFromDebugger,
// Windows XP+ from here
ThreadBreakOnTermination,
ThreadSwitchLegacyState,
ThreadIsTerminated,
// Windows Vista+ from here
ThreadLastSystemCall,
ThreadIoPriority,
ThreadCycleTime,
ThreadPagePriority,
ThreadActualBasePriority,
ThreadTebInformation,
ThreadCSwitchMon,
// Windows 7+ from here
ThreadCSwitchPmu,
ThreadWow64Context,
ThreadGroupInformation,
ThreadUmsInformation,
ThreadCounterProfiling,
ThreadIdealProcessorEx,
// Windows 8+ from here
ThreadCpuAccountingInformation,
// Windows 8.1+ from here
ThreadSuspendCount,
// Windows 10+ from here
ThreadHeterogeneousCpuPolicy,
ThreadContainerId,
ThreadNameInformation,
ThreadSelectedCpuSets,
ThreadSystemThreadInformation,
ThreadActualGroupAffinity,
};
pub const PROCESSINFOCLASS = enum(c_int) {
ProcessBasicInformation,
ProcessQuotaLimits,
ProcessIoCounters,
ProcessVmCounters,
ProcessTimes,
ProcessBasePriority,
ProcessRaisePriority,
ProcessDebugPort,
ProcessExceptionPort,
ProcessAccessToken,
ProcessLdtInformation,
ProcessLdtSize,
ProcessDefaultHardErrorMode,
ProcessIoPortHandlers,
ProcessPooledUsageAndLimits,
ProcessWorkingSetWatch,
ProcessUserModeIOPL,
ProcessEnableAlignmentFaultFixup,
ProcessPriorityClass,
ProcessWx86Information,
ProcessHandleCount,
ProcessAffinityMask,
ProcessPriorityBoost,
ProcessDeviceMap,
ProcessSessionInformation,
ProcessForegroundInformation,
ProcessWow64Information,
ProcessImageFileName,
ProcessLUIDDeviceMapsEnabled,
ProcessBreakOnTermination,
ProcessDebugObjectHandle,
ProcessDebugFlags,
ProcessHandleTracing,
ProcessIoPriority,
ProcessExecuteFlags,
ProcessTlsInformation,
ProcessCookie,
ProcessImageInformation,
ProcessCycleTime,
ProcessPagePriority,
ProcessInstrumentationCallback,
ProcessThreadStackAllocation,
ProcessWorkingSetWatchEx,
ProcessImageFileNameWin32,
ProcessImageFileMapping,
ProcessAffinityUpdateMode,
ProcessMemoryAllocationMode,
ProcessGroupInformation,
ProcessTokenVirtualizationEnabled,
ProcessConsoleHostProcess,
ProcessWindowInformation,
MaxProcessInfoClass,
};
pub const PROCESS_BASIC_INFORMATION = extern struct {
ExitStatus: NTSTATUS,
PebBaseAddress: *PEB,
AffinityMask: ULONG_PTR,
BasePriority: KPRIORITY,
UniqueProcessId: ULONG_PTR,
InheritedFromUniqueProcessId: ULONG_PTR,
};
pub const ReadMemoryError = error{
Unexpected,
};
pub fn ReadProcessMemory(handle: HANDLE, addr: ?LPVOID, buffer: []u8) ReadMemoryError![]u8 {
var nread: usize = 0;
switch (ntdll.NtReadVirtualMemory(
handle,
addr,
buffer.ptr,
buffer.len,
&nread,
)) {
.SUCCESS => return buffer[0..nread],
// TODO: map errors
else => |rc| return unexpectedStatus(rc),
}
}
pub const WriteMemoryError = error{
Unexpected,
};
pub fn WriteProcessMemory(handle: HANDLE, addr: ?LPVOID, buffer: []const u8) WriteMemoryError!usize {
var nwritten: usize = 0;
switch (ntdll.NtWriteVirtualMemory(
handle,
addr,
@ptrCast(*const anyopaque, buffer.ptr),
buffer.len,
&nwritten,
)) {
.SUCCESS => return nwritten,
// TODO: map errors
else => |rc| return unexpectedStatus(rc),
}
}
pub const ProcessBaseAddressError = GetProcessMemoryInfoError || ReadMemoryError;
/// Returns the base address of the process loaded into memory.
pub fn ProcessBaseAddress(handle: HANDLE) ProcessBaseAddressError!HMODULE {
var info: PROCESS_BASIC_INFORMATION = undefined;
var nread: DWORD = 0;
const rc = ntdll.NtQueryInformationProcess(
handle,
.ProcessBasicInformation,
&info,
@sizeOf(PROCESS_BASIC_INFORMATION),
&nread,
);
switch (rc) {
.SUCCESS => {},
.ACCESS_DENIED => return error.AccessDenied,
.INVALID_HANDLE => return error.InvalidHandle,
.INVALID_PARAMETER => unreachable,
else => return unexpectedStatus(rc),
}
var peb_buf: [@sizeOf(PEB)]u8 align(@alignOf(PEB)) = undefined;
const peb_out = try ReadProcessMemory(handle, info.PebBaseAddress, &peb_buf);
const ppeb = @ptrCast(*const PEB, @alignCast(@alignOf(PEB), peb_out.ptr));
return ppeb.ImageBaseAddress;
}

View File

@ -31,61 +31,10 @@ const UNWIND_HISTORY_TABLE = windows.UNWIND_HISTORY_TABLE;
const RUNTIME_FUNCTION = windows.RUNTIME_FUNCTION;
const KNONVOLATILE_CONTEXT_POINTERS = windows.KNONVOLATILE_CONTEXT_POINTERS;
const EXCEPTION_ROUTINE = windows.EXCEPTION_ROUTINE;
pub const PROCESSINFOCLASS = enum(c_int) {
ProcessBasicInformation,
ProcessQuotaLimits,
ProcessIoCounters,
ProcessVmCounters,
ProcessTimes,
ProcessBasePriority,
ProcessRaisePriority,
ProcessDebugPort,
ProcessExceptionPort,
ProcessAccessToken,
ProcessLdtInformation,
ProcessLdtSize,
ProcessDefaultHardErrorMode,
ProcessIoPortHandlers,
ProcessPooledUsageAndLimits,
ProcessWorkingSetWatch,
ProcessUserModeIOPL,
ProcessEnableAlignmentFaultFixup,
ProcessPriorityClass,
ProcessWx86Information,
ProcessHandleCount,
ProcessAffinityMask,
ProcessPriorityBoost,
ProcessDeviceMap,
ProcessSessionInformation,
ProcessForegroundInformation,
ProcessWow64Information,
ProcessImageFileName,
ProcessLUIDDeviceMapsEnabled,
ProcessBreakOnTermination,
ProcessDebugObjectHandle,
ProcessDebugFlags,
ProcessHandleTracing,
ProcessIoPriority,
ProcessExecuteFlags,
ProcessTlsInformation,
ProcessCookie,
ProcessImageInformation,
ProcessCycleTime,
ProcessPagePriority,
ProcessInstrumentationCallback,
ProcessThreadStackAllocation,
ProcessWorkingSetWatchEx,
ProcessImageFileNameWin32,
ProcessImageFileMapping,
ProcessAffinityUpdateMode,
ProcessMemoryAllocationMode,
ProcessGroupInformation,
ProcessTokenVirtualizationEnabled,
ProcessConsoleHostProcess,
ProcessWindowInformation,
MaxProcessInfoClass,
};
const THREADINFOCLASS = windows.THREADINFOCLASS;
const PROCESSINFOCLASS = windows.PROCESSINFOCLASS;
const LPVOID = windows.LPVOID;
const LPCVOID = windows.LPCVOID;
pub extern "ntdll" fn NtQueryInformationProcess(
ProcessHandle: HANDLE,
@ -95,57 +44,6 @@ pub extern "ntdll" fn NtQueryInformationProcess(
ReturnLength: ?*ULONG,
) callconv(WINAPI) NTSTATUS;
pub const THREADINFOCLASS = enum(c_int) {
ThreadBasicInformation,
ThreadTimes,
ThreadPriority,
ThreadBasePriority,
ThreadAffinityMask,
ThreadImpersonationToken,
ThreadDescriptorTableEntry,
ThreadEnableAlignmentFaultFixup,
ThreadEventPair_Reusable,
ThreadQuerySetWin32StartAddress,
ThreadZeroTlsCell,
ThreadPerformanceCount,
ThreadAmILastThread,
ThreadIdealProcessor,
ThreadPriorityBoost,
ThreadSetTlsArrayAddress,
ThreadIsIoPending,
// Windows 2000+ from here
ThreadHideFromDebugger,
// Windows XP+ from here
ThreadBreakOnTermination,
ThreadSwitchLegacyState,
ThreadIsTerminated,
// Windows Vista+ from here
ThreadLastSystemCall,
ThreadIoPriority,
ThreadCycleTime,
ThreadPagePriority,
ThreadActualBasePriority,
ThreadTebInformation,
ThreadCSwitchMon,
// Windows 7+ from here
ThreadCSwitchPmu,
ThreadWow64Context,
ThreadGroupInformation,
ThreadUmsInformation,
ThreadCounterProfiling,
ThreadIdealProcessorEx,
// Windows 8+ from here
ThreadCpuAccountingInformation,
// Windows 8.1+ from here
ThreadSuspendCount,
// Windows 10+ from here
ThreadHeterogeneousCpuPolicy,
ThreadContainerId,
ThreadNameInformation,
ThreadSelectedCpuSets,
ThreadSystemThreadInformation,
ThreadActualGroupAffinity,
};
pub extern "ntdll" fn NtQueryInformationThread(
ThreadHandle: HANDLE,
ThreadInformationClass: THREADINFOCLASS,
@ -364,10 +262,26 @@ pub extern "ntdll" fn RtlQueryRegistryValues(
Environment: ?*anyopaque,
) callconv(WINAPI) NTSTATUS;
pub extern "ntdll" fn NtReadVirtualMemory(
ProcessHandle: HANDLE,
BaseAddress: ?PVOID,
Buffer: LPVOID,
NumberOfBytesToRead: SIZE_T,
NumberOfBytesRead: ?*SIZE_T,
) callconv(WINAPI) NTSTATUS;
pub extern "ntdll" fn NtWriteVirtualMemory(
ProcessHandle: HANDLE,
BaseAddress: ?PVOID,
Buffer: LPCVOID,
NumberOfBytesToWrite: SIZE_T,
NumberOfBytesWritten: ?*SIZE_T,
) callconv(WINAPI) NTSTATUS;
pub extern "ntdll" fn NtProtectVirtualMemory(
ProcessHandle: HANDLE,
BaseAddress: *PVOID,
NumberOfBytesToProtect: *ULONG,
BaseAddress: *?PVOID,
NumberOfBytesToProtect: *SIZE_T,
NewAccessProtection: ULONG,
OldAccessProtection: *ULONG,
) callconv(WINAPI) NTSTATUS;

View File

@ -379,24 +379,30 @@ pub const File = struct {
if (base.file != null) return;
const emit = base.options.emit orelse return;
if (base.child_pid) |pid| {
// If we try to open the output file in write mode while it is running,
// it will return ETXTBSY. So instead, we copy the file, atomically rename it
// over top of the exe path, and then proceed normally. This changes the inode,
// avoiding the error.
const tmp_sub_path = try std.fmt.allocPrint(base.allocator, "{s}-{x}", .{
emit.sub_path, std.crypto.random.int(u32),
});
try emit.directory.handle.copyFile(emit.sub_path, emit.directory.handle, tmp_sub_path, .{});
try emit.directory.handle.rename(tmp_sub_path, emit.sub_path);
switch (builtin.os.tag) {
.linux => std.os.ptrace(std.os.linux.PTRACE.ATTACH, pid, 0, 0) catch |err| {
log.warn("ptrace failure: {s}", .{@errorName(err)});
},
.macos => base.cast(MachO).?.ptraceAttach(pid) catch |err| {
if (builtin.os.tag == .windows) {
base.cast(Coff).?.ptraceAttach(pid) catch |err| {
log.warn("attaching failed with error: {s}", .{@errorName(err)});
},
.windows => {},
else => return error.HotSwapUnavailableOnHostOperatingSystem,
};
} else {
// If we try to open the output file in write mode while it is running,
// it will return ETXTBSY. So instead, we copy the file, atomically rename it
// over top of the exe path, and then proceed normally. This changes the inode,
// avoiding the error.
const tmp_sub_path = try std.fmt.allocPrint(base.allocator, "{s}-{x}", .{
emit.sub_path, std.crypto.random.int(u32),
});
try emit.directory.handle.copyFile(emit.sub_path, emit.directory.handle, tmp_sub_path, .{});
try emit.directory.handle.rename(tmp_sub_path, emit.sub_path);
switch (builtin.os.tag) {
.linux => std.os.ptrace(std.os.linux.PTRACE.ATTACH, pid, 0, 0) catch |err| {
log.warn("ptrace failure: {s}", .{@errorName(err)});
},
.macos => base.cast(MachO).?.ptraceAttach(pid) catch |err| {
log.warn("attaching failed with error: {s}", .{@errorName(err)});
},
.windows => unreachable,
else => return error.HotSwapUnavailableOnHostOperatingSystem,
}
}
}
base.file = try emit.directory.handle.createFile(emit.sub_path, .{
@ -437,7 +443,7 @@ pub const File = struct {
.macos => base.cast(MachO).?.ptraceDetach(pid) catch |err| {
log.warn("detaching failed with error: {s}", .{@errorName(err)});
},
.windows => {},
.windows => base.cast(Coff).?.ptraceDetach(pid),
else => return error.HotSwapUnavailableOnHostOperatingSystem,
}
}

View File

@ -89,6 +89,20 @@ relocs: RelocTable = .{},
/// this will be a table indexed by index into the list of Atoms.
base_relocs: BaseRelocationTable = .{},
/// Hot-code swapping state.
hot_state: if (is_hot_update_compatible) HotUpdateState else struct {} = .{},
const is_hot_update_compatible = switch (builtin.target.os.tag) {
.windows => true,
else => false,
};
const HotUpdateState = struct {
/// Base address at which the process (image) got loaded.
/// We need this info to correctly slide pointers when relocating.
loaded_base_address: ?std.os.windows.HMODULE = null,
};
const Entry = struct {
target: SymbolWithLoc,
// Index into the synthetic symbol table (i.e., file == null).
@ -772,13 +786,87 @@ fn writeAtom(self: *Coff, atom_index: Atom.Index, code: []u8) !void {
const sym = atom.getSymbol(self);
const section = self.sections.get(@enumToInt(sym.section_number) - 1);
const file_offset = section.header.pointer_to_raw_data + sym.value - section.header.virtual_address;
log.debug("writing atom for symbol {s} at file offset 0x{x} to 0x{x}", .{
atom.getName(self),
file_offset,
file_offset + code.len,
});
self.resolveRelocs(atom_index, code);
const gpa = self.base.allocator;
// Gather relocs which can be resolved.
// We need to do this as we will be applying different slide values depending
// if we are running in hot-code swapping mode or not.
// TODO: how crazy would it be to try and apply the actual image base of the loaded
// process for the in-file values rather than the Windows defaults?
var relocs = std.ArrayList(*Relocation).init(gpa);
defer relocs.deinit();
if (self.relocs.getPtr(atom_index)) |rels| {
try relocs.ensureTotalCapacityPrecise(rels.items.len);
for (rels.items) |*reloc| {
if (reloc.isResolvable(self)) relocs.appendAssumeCapacity(reloc);
}
}
if (is_hot_update_compatible) {
if (self.base.child_pid) |handle| {
const slide = @ptrToInt(self.hot_state.loaded_base_address.?);
const mem_code = try gpa.dupe(u8, code);
defer gpa.free(mem_code);
self.resolveRelocs(atom_index, relocs.items, mem_code, slide);
const vaddr = sym.value + slide;
const pvaddr = @intToPtr(*anyopaque, vaddr);
log.debug("writing to memory at address {x}", .{vaddr});
if (build_options.enable_logging) {
try debugMem(gpa, handle, pvaddr, mem_code);
}
if (section.header.flags.MEM_WRITE == 0) {
writeMemProtected(handle, pvaddr, mem_code) catch |err| {
log.warn("writing to protected memory failed with error: {s}", .{@errorName(err)});
};
} else {
writeMem(handle, pvaddr, mem_code) catch |err| {
log.warn("writing to protected memory failed with error: {s}", .{@errorName(err)});
};
}
}
}
self.resolveRelocs(atom_index, relocs.items, code, self.getImageBase());
try self.base.file.?.pwriteAll(code, file_offset);
// Now we can mark the relocs as resolved.
while (relocs.popOrNull()) |reloc| {
reloc.dirty = false;
}
}
fn debugMem(allocator: Allocator, handle: std.ChildProcess.Id, pvaddr: std.os.windows.LPVOID, code: []const u8) !void {
var buffer = try allocator.alloc(u8, code.len);
defer allocator.free(buffer);
const memread = try std.os.windows.ReadProcessMemory(handle, pvaddr, buffer);
log.debug("to write: {x}", .{std.fmt.fmtSliceHexLower(code)});
log.debug("in memory: {x}", .{std.fmt.fmtSliceHexLower(memread)});
}
fn writeMemProtected(handle: std.ChildProcess.Id, pvaddr: std.os.windows.LPVOID, code: []const u8) !void {
const old_prot = try std.os.windows.VirtualProtectEx(handle, pvaddr, code.len, std.os.windows.PAGE_EXECUTE_WRITECOPY);
try writeMem(handle, pvaddr, code);
// TODO: We can probably just set the pages writeable and leave it at that without having to restore the attributes.
// For that though, we want to track which page has already been modified.
_ = try std.os.windows.VirtualProtectEx(handle, pvaddr, code.len, old_prot);
}
fn writeMem(handle: std.ChildProcess.Id, pvaddr: std.os.windows.LPVOID, code: []const u8) !void {
const amt = try std.os.windows.WriteProcessMemory(handle, pvaddr, code);
if (amt != code.len) return error.InputOutput;
}
fn writePtrWidthAtom(self: *Coff, atom_index: Atom.Index) !void {
@ -814,19 +902,30 @@ fn markRelocsDirtyByAddress(self: *Coff, addr: u32) void {
}
}
fn resolveRelocs(self: *Coff, atom_index: Atom.Index, code: []u8) void {
const relocs = self.relocs.getPtr(atom_index) orelse return;
fn resolveRelocs(self: *Coff, atom_index: Atom.Index, relocs: []*const Relocation, code: []u8, image_base: u64) void {
log.debug("relocating '{s}'", .{self.getAtom(atom_index).getName(self)});
for (relocs.items) |*reloc| {
if (!reloc.dirty) continue;
if (reloc.resolve(atom_index, code, self)) {
reloc.dirty = false;
}
for (relocs) |reloc| {
reloc.resolve(atom_index, code, image_base, self);
}
}
pub fn ptraceAttach(self: *Coff, handle: std.ChildProcess.Id) !void {
if (!is_hot_update_compatible) return;
log.debug("attaching to process with handle {*}", .{handle});
self.hot_state.loaded_base_address = std.os.windows.ProcessBaseAddress(handle) catch |err| {
log.warn("failed to get base address for the process with error: {s}", .{@errorName(err)});
return;
};
}
pub fn ptraceDetach(self: *Coff, handle: std.ChildProcess.Id) void {
if (!is_hot_update_compatible) return;
log.debug("detaching from process with handle {*}", .{handle});
self.hot_state.loaded_base_address = null;
}
fn freeAtom(self: *Coff, atom_index: Atom.Index) void {
log.debug("freeAtom {d}", .{atom_index});
@ -1421,7 +1520,7 @@ pub fn flushModule(self: *Coff, comp: *Compilation, prog_node: *std.Progress.Nod
for (self.relocs.keys(), self.relocs.values()) |atom_index, relocs| {
const needs_update = for (relocs.items) |reloc| {
if (reloc.dirty) break true;
if (reloc.isResolvable(self)) break true;
} else false;
if (!needs_update) continue;

View File

@ -72,14 +72,18 @@ pub fn getTargetAddress(self: Relocation, coff_file: *const Coff) ?u32 {
}
}
/// Returns `false` if obtaining the target address has been deferred until `flushModule`.
/// This can happen when trying to resolve address of an import table entry ahead of time.
pub fn resolve(self: Relocation, atom_index: Atom.Index, code: []u8, coff_file: *Coff) bool {
/// Returns true if and only if the reloc is dirty AND the target address is available.
pub fn isResolvable(self: Relocation, coff_file: *Coff) bool {
_ = self.getTargetAddress(coff_file) orelse return false;
return self.dirty;
}
pub fn resolve(self: Relocation, atom_index: Atom.Index, code: []u8, image_base: u64, coff_file: *Coff) void {
const atom = coff_file.getAtom(atom_index);
const source_sym = atom.getSymbol(coff_file);
const source_vaddr = source_sym.value + self.offset;
const target_vaddr = self.getTargetAddress(coff_file) orelse return false;
const target_vaddr = self.getTargetAddress(coff_file).?; // Oops, you didn't check if the relocation can be resolved with isResolvable().
const target_vaddr_with_addend = target_vaddr + self.addend;
log.debug(" ({x}: [() => 0x{x} ({s})) ({s}) ", .{
@ -92,7 +96,7 @@ pub fn resolve(self: Relocation, atom_index: Atom.Index, code: []u8, coff_file:
const ctx: Context = .{
.source_vaddr = source_vaddr,
.target_vaddr = target_vaddr_with_addend,
.image_base = coff_file.getImageBase(),
.image_base = image_base,
.code = code,
.ptr_width = coff_file.ptr_width,
};
@ -102,8 +106,6 @@ pub fn resolve(self: Relocation, atom_index: Atom.Index, code: []u8, coff_file:
.x86, .x86_64 => self.resolveX86(ctx),
else => unreachable, // unhandled target architecture
}
return true;
}
const Context = struct {

View File

@ -3817,11 +3817,25 @@ fn runOrTestHotSwap(
runtime_args_start: ?usize,
) !std.ChildProcess.Id {
const exe_emit = comp.bin_file.options.emit.?;
// A naive `directory.join` here will indeed get the correct path to the binary,
// however, in the case of cwd, we actually want `./foo` so that the path can be executed.
const exe_path = try fs.path.join(gpa, &[_][]const u8{
exe_emit.directory.path orelse ".", exe_emit.sub_path,
});
const exe_path = switch (builtin.target.os.tag) {
// On Windows it seems impossible to perform an atomic rename of a file that is currently
// running in a process. Therefore, we do the opposite. We create a copy of the file in
// tmp zig-cache and use it to spawn the child process. This way we are free to update
// the binary with each requested hot update.
.windows => blk: {
try exe_emit.directory.handle.copyFile(exe_emit.sub_path, comp.local_cache_directory.handle, exe_emit.sub_path, .{});
break :blk try fs.path.join(gpa, &[_][]const u8{
comp.local_cache_directory.path orelse ".", exe_emit.sub_path,
});
},
// A naive `directory.join` here will indeed get the correct path to the binary,
// however, in the case of cwd, we actually want `./foo` so that the path can be executed.
else => try fs.path.join(gpa, &[_][]const u8{
exe_emit.directory.path orelse ".", exe_emit.sub_path,
}),
};
defer gpa.free(exe_path);
var argv = std.ArrayList([]const u8).init(gpa);