1 | /* |
2 | ** Machine code management. |
3 | ** Copyright (C) 2005-2021 Mike Pall. See Copyright Notice in luajit.h |
4 | */ |
5 | |
6 | #define lj_mcode_c |
7 | #define LUA_CORE |
8 | |
9 | #include "lj_obj.h" |
10 | #if LJ_HASJIT |
11 | #include "lj_gc.h" |
12 | #include "lj_err.h" |
13 | #include "lj_jit.h" |
14 | #include "lj_mcode.h" |
15 | #include "lj_trace.h" |
16 | #include "lj_dispatch.h" |
17 | #include "lj_prng.h" |
18 | #endif |
19 | #if LJ_HASJIT || LJ_HASFFI |
20 | #include "lj_vm.h" |
21 | #endif |
22 | |
23 | /* -- OS-specific functions ----------------------------------------------- */ |
24 | |
25 | #if LJ_HASJIT || LJ_HASFFI |
26 | |
27 | /* Define this if you want to run LuaJIT with Valgrind. */ |
28 | #ifdef LUAJIT_USE_VALGRIND |
29 | #include <valgrind/valgrind.h> |
30 | #endif |
31 | |
32 | #if LJ_TARGET_IOS |
33 | void sys_icache_invalidate(void *start, size_t len); |
34 | #endif |
35 | |
36 | /* Synchronize data/instruction cache. */ |
37 | void lj_mcode_sync(void *start, void *end) |
38 | { |
39 | #ifdef LUAJIT_USE_VALGRIND |
40 | VALGRIND_DISCARD_TRANSLATIONS(start, (char *)end-(char *)start); |
41 | #endif |
42 | #if LJ_TARGET_X86ORX64 |
43 | UNUSED(start); UNUSED(end); |
44 | #elif LJ_TARGET_IOS |
45 | sys_icache_invalidate(start, (char *)end-(char *)start); |
46 | #elif LJ_TARGET_PPC |
47 | lj_vm_cachesync(start, end); |
48 | #elif defined(__GNUC__) || defined(__clang__) |
49 | __clear_cache(start, end); |
50 | #else |
51 | #error "Missing builtin to flush instruction cache" |
52 | #endif |
53 | } |
54 | |
55 | #endif |
56 | |
57 | #if LJ_HASJIT |
58 | |
59 | #if LJ_TARGET_WINDOWS |
60 | |
61 | #define WIN32_LEAN_AND_MEAN |
62 | #include <windows.h> |
63 | |
64 | #define MCPROT_RW PAGE_READWRITE |
65 | #define MCPROT_RX PAGE_EXECUTE_READ |
66 | #define MCPROT_RWX PAGE_EXECUTE_READWRITE |
67 | |
68 | static void *mcode_alloc_at(jit_State *J, uintptr_t hint, size_t sz, DWORD prot) |
69 | { |
70 | void *p = LJ_WIN_VALLOC((void *)hint, sz, |
71 | MEM_RESERVE|MEM_COMMIT|MEM_TOP_DOWN, prot); |
72 | if (!p && !hint) |
73 | lj_trace_err(J, LJ_TRERR_MCODEAL); |
74 | return p; |
75 | } |
76 | |
77 | static void mcode_free(jit_State *J, void *p, size_t sz) |
78 | { |
79 | UNUSED(J); UNUSED(sz); |
80 | VirtualFree(p, 0, MEM_RELEASE); |
81 | } |
82 | |
83 | static int mcode_setprot(void *p, size_t sz, DWORD prot) |
84 | { |
85 | DWORD oprot; |
86 | return !LJ_WIN_VPROTECT(p, sz, prot, &oprot); |
87 | } |
88 | |
89 | #elif LJ_TARGET_POSIX |
90 | |
91 | #include <sys/mman.h> |
92 | |
93 | #ifndef MAP_ANONYMOUS |
94 | #define MAP_ANONYMOUS MAP_ANON |
95 | #endif |
96 | |
97 | #define MCPROT_RW (PROT_READ|PROT_WRITE) |
98 | #define MCPROT_RX (PROT_READ|PROT_EXEC) |
99 | #define MCPROT_RWX (PROT_READ|PROT_WRITE|PROT_EXEC) |
100 | |
101 | static void *mcode_alloc_at(jit_State *J, uintptr_t hint, size_t sz, int prot) |
102 | { |
103 | void *p = mmap((void *)hint, sz, prot, MAP_PRIVATE|MAP_ANONYMOUS, -1, 0); |
104 | if (p == MAP_FAILED) { |
105 | if (!hint) lj_trace_err(J, LJ_TRERR_MCODEAL); |
106 | p = NULL; |
107 | } |
108 | return p; |
109 | } |
110 | |
111 | static void mcode_free(jit_State *J, void *p, size_t sz) |
112 | { |
113 | UNUSED(J); |
114 | munmap(p, sz); |
115 | } |
116 | |
117 | static int mcode_setprot(void *p, size_t sz, int prot) |
118 | { |
119 | return mprotect(p, sz, prot); |
120 | } |
121 | |
122 | #else |
123 | |
124 | #error "Missing OS support for explicit placement of executable memory" |
125 | |
126 | #endif |
127 | |
128 | /* -- MCode area protection ----------------------------------------------- */ |
129 | |
130 | #if LUAJIT_SECURITY_MCODE == 0 |
131 | |
132 | /* Define this ONLY if page protection twiddling becomes a bottleneck. |
133 | ** |
134 | ** It's generally considered to be a potential security risk to have |
135 | ** pages with simultaneous write *and* execute access in a process. |
136 | ** |
137 | ** Do not even think about using this mode for server processes or |
138 | ** apps handling untrusted external data. |
139 | ** |
140 | ** The security risk is not in LuaJIT itself -- but if an adversary finds |
141 | ** any *other* flaw in your C application logic, then any RWX memory pages |
142 | ** simplify writing an exploit considerably. |
143 | */ |
144 | #define MCPROT_GEN MCPROT_RWX |
145 | #define MCPROT_RUN MCPROT_RWX |
146 | |
147 | static void mcode_protect(jit_State *J, int prot) |
148 | { |
149 | UNUSED(J); UNUSED(prot); UNUSED(mcode_setprot); |
150 | } |
151 | |
152 | #else |
153 | |
154 | /* This is the default behaviour and much safer: |
155 | ** |
156 | ** Most of the time the memory pages holding machine code are executable, |
157 | ** but NONE of them is writable. |
158 | ** |
159 | ** The current memory area is marked read-write (but NOT executable) only |
160 | ** during the short time window while the assembler generates machine code. |
161 | */ |
162 | #define MCPROT_GEN MCPROT_RW |
163 | #define MCPROT_RUN MCPROT_RX |
164 | |
165 | /* Protection twiddling failed. Probably due to kernel security. */ |
166 | static LJ_NOINLINE void mcode_protfail(jit_State *J) |
167 | { |
168 | lua_CFunction panic = J2G(J)->panic; |
169 | if (panic) { |
170 | lua_State *L = J->L; |
171 | setstrV(L, L->top++, lj_err_str(L, LJ_ERR_JITPROT)); |
172 | panic(L); |
173 | } |
174 | } |
175 | |
176 | /* Change protection of MCode area. */ |
177 | static void mcode_protect(jit_State *J, int prot) |
178 | { |
179 | if (J->mcprot != prot) { |
180 | if (LJ_UNLIKELY(mcode_setprot(J->mcarea, J->szmcarea, prot))) |
181 | mcode_protfail(J); |
182 | J->mcprot = prot; |
183 | } |
184 | } |
185 | |
186 | #endif |
187 | |
188 | /* -- MCode area allocation ----------------------------------------------- */ |
189 | |
190 | #if LJ_64 |
191 | #define mcode_validptr(p) (p) |
192 | #else |
193 | #define mcode_validptr(p) ((p) && (uintptr_t)(p) < 0xffff0000) |
194 | #endif |
195 | |
196 | #ifdef LJ_TARGET_JUMPRANGE |
197 | |
198 | /* Get memory within relative jump distance of our code in 64 bit mode. */ |
199 | static void *mcode_alloc(jit_State *J, size_t sz) |
200 | { |
201 | /* Target an address in the static assembler code (64K aligned). |
202 | ** Try addresses within a distance of target-range/2+1MB..target+range/2-1MB. |
203 | ** Use half the jump range so every address in the range can reach any other. |
204 | */ |
205 | #if LJ_TARGET_MIPS |
206 | /* Use the middle of the 256MB-aligned region. */ |
207 | uintptr_t target = ((uintptr_t)(void *)lj_vm_exit_handler & |
208 | ~(uintptr_t)0x0fffffffu) + 0x08000000u; |
209 | #else |
210 | uintptr_t target = (uintptr_t)(void *)lj_vm_exit_handler & ~(uintptr_t)0xffff; |
211 | #endif |
212 | const uintptr_t range = (1u << (LJ_TARGET_JUMPRANGE-1)) - (1u << 21); |
213 | /* First try a contiguous area below the last one. */ |
214 | uintptr_t hint = J->mcarea ? (uintptr_t)J->mcarea - sz : 0; |
215 | int i; |
216 | /* Limit probing iterations, depending on the available pool size. */ |
217 | for (i = 0; i < LJ_TARGET_JUMPRANGE; i++) { |
218 | if (mcode_validptr(hint)) { |
219 | void *p = mcode_alloc_at(J, hint, sz, MCPROT_GEN); |
220 | |
221 | if (mcode_validptr(p) && |
222 | ((uintptr_t)p + sz - target < range || target - (uintptr_t)p < range)) |
223 | return p; |
224 | if (p) mcode_free(J, p, sz); /* Free badly placed area. */ |
225 | } |
226 | /* Next try probing 64K-aligned pseudo-random addresses. */ |
227 | do { |
228 | hint = lj_prng_u64(&J2G(J)->prng) & ((1u<<LJ_TARGET_JUMPRANGE)-0x10000); |
229 | } while (!(hint + sz < range+range)); |
230 | hint = target + hint - range; |
231 | } |
232 | lj_trace_err(J, LJ_TRERR_MCODEAL); /* Give up. OS probably ignores hints? */ |
233 | return NULL; |
234 | } |
235 | |
236 | #else |
237 | |
238 | /* All memory addresses are reachable by relative jumps. */ |
239 | static void *mcode_alloc(jit_State *J, size_t sz) |
240 | { |
241 | #if defined(__OpenBSD__) || LJ_TARGET_UWP |
242 | /* Allow better executable memory allocation for OpenBSD W^X mode. */ |
243 | void *p = mcode_alloc_at(J, 0, sz, MCPROT_RUN); |
244 | if (p && mcode_setprot(p, sz, MCPROT_GEN)) { |
245 | mcode_free(J, p, sz); |
246 | return NULL; |
247 | } |
248 | return p; |
249 | #else |
250 | return mcode_alloc_at(J, 0, sz, MCPROT_GEN); |
251 | #endif |
252 | } |
253 | |
254 | #endif |
255 | |
256 | /* -- MCode area management ----------------------------------------------- */ |
257 | |
258 | /* Allocate a new MCode area. */ |
259 | static void mcode_allocarea(jit_State *J) |
260 | { |
261 | MCode *oldarea = J->mcarea; |
262 | size_t sz = (size_t)J->param[JIT_P_sizemcode] << 10; |
263 | sz = (sz + LJ_PAGESIZE-1) & ~(size_t)(LJ_PAGESIZE - 1); |
264 | J->mcarea = (MCode *)mcode_alloc(J, sz); |
265 | J->szmcarea = sz; |
266 | J->mcprot = MCPROT_GEN; |
267 | J->mctop = (MCode *)((char *)J->mcarea + J->szmcarea); |
268 | J->mcbot = (MCode *)((char *)J->mcarea + sizeof(MCLink)); |
269 | ((MCLink *)J->mcarea)->next = oldarea; |
270 | ((MCLink *)J->mcarea)->size = sz; |
271 | J->szallmcarea += sz; |
272 | J->mcbot = (MCode *)lj_err_register_mcode(J->mcarea, sz, (uint8_t *)J->mcbot); |
273 | } |
274 | |
275 | /* Free all MCode areas. */ |
276 | void lj_mcode_free(jit_State *J) |
277 | { |
278 | MCode *mc = J->mcarea; |
279 | J->mcarea = NULL; |
280 | J->szallmcarea = 0; |
281 | while (mc) { |
282 | MCode *next = ((MCLink *)mc)->next; |
283 | size_t sz = ((MCLink *)mc)->size; |
284 | lj_err_deregister_mcode(mc, sz, (uint8_t *)mc + sizeof(MCLink)); |
285 | mcode_free(J, mc, sz); |
286 | mc = next; |
287 | } |
288 | } |
289 | |
290 | /* -- MCode transactions -------------------------------------------------- */ |
291 | |
292 | /* Reserve the remainder of the current MCode area. */ |
293 | MCode *lj_mcode_reserve(jit_State *J, MCode **lim) |
294 | { |
295 | if (!J->mcarea) |
296 | mcode_allocarea(J); |
297 | else |
298 | mcode_protect(J, MCPROT_GEN); |
299 | *lim = J->mcbot; |
300 | return J->mctop; |
301 | } |
302 | |
303 | /* Commit the top part of the current MCode area. */ |
304 | void lj_mcode_commit(jit_State *J, MCode *top) |
305 | { |
306 | J->mctop = top; |
307 | mcode_protect(J, MCPROT_RUN); |
308 | } |
309 | |
310 | /* Abort the reservation. */ |
311 | void lj_mcode_abort(jit_State *J) |
312 | { |
313 | if (J->mcarea) |
314 | mcode_protect(J, MCPROT_RUN); |
315 | } |
316 | |
317 | /* Set/reset protection to allow patching of MCode areas. */ |
318 | MCode *lj_mcode_patch(jit_State *J, MCode *ptr, int finish) |
319 | { |
320 | if (finish) { |
321 | #if LUAJIT_SECURITY_MCODE |
322 | if (J->mcarea == ptr) |
323 | mcode_protect(J, MCPROT_RUN); |
324 | else if (LJ_UNLIKELY(mcode_setprot(ptr, ((MCLink *)ptr)->size, MCPROT_RUN))) |
325 | mcode_protfail(J); |
326 | #endif |
327 | return NULL; |
328 | } else { |
329 | MCode *mc = J->mcarea; |
330 | /* Try current area first to use the protection cache. */ |
331 | if (ptr >= mc && ptr < (MCode *)((char *)mc + J->szmcarea)) { |
332 | #if LUAJIT_SECURITY_MCODE |
333 | mcode_protect(J, MCPROT_GEN); |
334 | #endif |
335 | return mc; |
336 | } |
337 | /* Otherwise search through the list of MCode areas. */ |
338 | for (;;) { |
339 | mc = ((MCLink *)mc)->next; |
340 | lj_assertJ(mc != NULL, "broken MCode area chain" ); |
341 | if (ptr >= mc && ptr < (MCode *)((char *)mc + ((MCLink *)mc)->size)) { |
342 | #if LUAJIT_SECURITY_MCODE |
343 | if (LJ_UNLIKELY(mcode_setprot(mc, ((MCLink *)mc)->size, MCPROT_GEN))) |
344 | mcode_protfail(J); |
345 | #endif |
346 | return mc; |
347 | } |
348 | } |
349 | } |
350 | } |
351 | |
352 | /* Limit of MCode reservation reached. */ |
353 | void lj_mcode_limiterr(jit_State *J, size_t need) |
354 | { |
355 | size_t sizemcode, maxmcode; |
356 | lj_mcode_abort(J); |
357 | sizemcode = (size_t)J->param[JIT_P_sizemcode] << 10; |
358 | sizemcode = (sizemcode + LJ_PAGESIZE-1) & ~(size_t)(LJ_PAGESIZE - 1); |
359 | maxmcode = (size_t)J->param[JIT_P_maxmcode] << 10; |
360 | if ((size_t)need > sizemcode) |
361 | lj_trace_err(J, LJ_TRERR_MCODEOV); /* Too long for any area. */ |
362 | if (J->szallmcarea + sizemcode > maxmcode) |
363 | lj_trace_err(J, LJ_TRERR_MCODEAL); |
364 | mcode_allocarea(J); |
365 | lj_trace_err(J, LJ_TRERR_MCODELM); /* Retry with new area. */ |
366 | } |
367 | |
368 | #endif |
369 | |