// SPDX-License-Identifier: ISC // SPDX-FileCopyrightText: 2024 Michael Smith // SPDX-FileCopyrightText: 2022 Willian Henrique // SPDX-FileCopyrightText: 2024 Matthew Wozniak #include #include "3p/sst/x86.h" #include "intdef.h" #include "log.h" #include "os.h" #ifdef _WIN32 // try to avoid pulling in all of Windows.h for this... (redundant dllimport // avoids warnings in hook.test.c where Windows.h is included via test.h) __declspec(dllimport) int __stdcall FlushInstructionCache( void *, const void *, usize); #endif // Warning: half-arsed hacky implementation (because that's all we really need) // Almost certainly breaks in some weird cases. Oh well! Most of the time, // vtable hooking is more reliable, this is only for, uh, emergencies. __attribute__((aligned(4096))) static uchar trampolines[4096]; static uchar *nexttrampoline = trampolines; bool hook_init(void) { return os_mprot(trampolines, sizeof(trampolines), PAGE_EXECUTE_READWRITE); } static inline void iflush(void *p, int len) { #if defined(_WIN32) // -1 is the current process, and it's a constant in the WDK, so it's // assumed we can safely avoid the useless GetCurrentProcess call FlushInstructionCache((void *)-1, p, len); #elif defined(__GNUC__) __builtin___clear_cache((char *)p, (char *)p + len); #else warn("no way to clear instruction cache!"); #endif } void *hook_inline(void *func_, void *target) { uchar *func = func_; // dumb hack: if we hit some thunk that immediately jumps elsewhere (which // seems common for win32 API functions), hook the underlying thing instead. while (*func == X86_JMPIW) func += *(i32 *)(func + 1) + 5; if (!os_mprot(func, 5, PAGE_EXECUTE_READWRITE)) return 0; int len = 0; for (;;) { // FIXME: these cases may result in somewhat dodgy error messaging. They // shouldn't happen anyway though. Maybe if we're confident we just // compile 'em out of release builds some day, but that sounds a little // scary. For now preferring confusing messages over crashes, I guess. if (func[len] == X86_CALL) { warn("can't trampoline call instructions\n"); return 0; } int ilen = x86_len(func + len); if (ilen == -1) { warn("unknown or invalid instruction\n"); return 0; } len += ilen; if (len >= 5) break; if (func[len] == X86_JMPIW) { warn("can't trampoline jmp instructions\n"); return 0; } } // for simplicity, just bump alloc the trampoline. no need to free anyway if (nexttrampoline - trampolines > (int)sizeof(trampolines) - len - 6) { warn("out of trampoline space\n"); return 0; } uchar *trampoline = nexttrampoline; nexttrampoline += len + 6; // NOT thread-safe. we don't need that anyway! *trampoline++ = len; // stick length in front for quicker unhooking memcpy(trampoline, func, len); trampoline[len] = X86_JMPIW; u32 diff = func - (trampoline + 5); // goto the continuation memcpy(trampoline + len + 1, &diff, 4); diff = (uchar *)target - (func + 5); // goto the hook target func[0] = X86_JMPIW; memcpy(func + 1, &diff, 4); iflush(func, 5); return trampoline; } void unhook_inline(void *orig) { uchar *p = orig; int len = p[-1]; int off = *(i32 *)(p + len + 1); uchar *q = p + off + 5; memcpy(q, p, 5); // XXX: not atomic atm! (does any of it even need to be?) iflush(q, 5); } void *hook_dllapi(const char *module, const char *name, void *target) { void *func = os_dlsym(os_dlopen(module), name); debug("%s = %p", name, func); if (!func) warn("couldn't find function %s in %s", name, module); else return hook_inline(func, target); return NULL; } // vi: sw=4 ts=4 noet tw=80 cc=80