]>
Commit | Line | Data |
---|---|---|
c767a54b JP |
1 | #define pr_fmt(fmt) "SMP alternatives: " fmt |
2 | ||
9a0b5817 | 3 | #include <linux/module.h> |
f6a57033 | 4 | #include <linux/sched.h> |
2f1dafe5 | 5 | #include <linux/mutex.h> |
9a0b5817 | 6 | #include <linux/list.h> |
8b5a10fc | 7 | #include <linux/stringify.h> |
19d36ccd AK |
8 | #include <linux/mm.h> |
9 | #include <linux/vmalloc.h> | |
3945dab4 | 10 | #include <linux/memory.h> |
3d55cc8a | 11 | #include <linux/stop_machine.h> |
5a0e3ad6 | 12 | #include <linux/slab.h> |
fd4363ff | 13 | #include <linux/kdebug.h> |
35de5b06 | 14 | #include <asm/text-patching.h> |
9a0b5817 GH |
15 | #include <asm/alternative.h> |
16 | #include <asm/sections.h> | |
19d36ccd | 17 | #include <asm/pgtable.h> |
8f4e956b AK |
18 | #include <asm/mce.h> |
19 | #include <asm/nmi.h> | |
e587cadd | 20 | #include <asm/cacheflush.h> |
78ff7fae | 21 | #include <asm/tlbflush.h> |
e587cadd | 22 | #include <asm/io.h> |
78ff7fae | 23 | #include <asm/fixmap.h> |
9a0b5817 | 24 | |
5e907bb0 IM |
25 | int __read_mostly alternatives_patched; |
26 | ||
27 | EXPORT_SYMBOL_GPL(alternatives_patched); | |
28 | ||
ab144f5e AK |
29 | #define MAX_PATCH_LEN (255-1) |
30 | ||
8b5a10fc | 31 | static int __initdata_or_module debug_alternative; |
b7fb4af0 | 32 | |
d167a518 GH |
33 | static int __init debug_alt(char *str) |
34 | { | |
35 | debug_alternative = 1; | |
36 | return 1; | |
37 | } | |
d167a518 GH |
38 | __setup("debug-alternative", debug_alt); |
39 | ||
09488165 JB |
40 | static int noreplace_smp; |
41 | ||
b7fb4af0 JF |
42 | static int __init setup_noreplace_smp(char *str) |
43 | { | |
44 | noreplace_smp = 1; | |
45 | return 1; | |
46 | } | |
47 | __setup("noreplace-smp", setup_noreplace_smp); | |
48 | ||
959b4fdf | 49 | #ifdef CONFIG_PARAVIRT |
8b5a10fc | 50 | static int __initdata_or_module noreplace_paravirt = 0; |
959b4fdf JF |
51 | |
52 | static int __init setup_noreplace_paravirt(char *str) | |
53 | { | |
54 | noreplace_paravirt = 1; | |
55 | return 1; | |
56 | } | |
57 | __setup("noreplace-paravirt", setup_noreplace_paravirt); | |
58 | #endif | |
b7fb4af0 | 59 | |
db477a33 BP |
60 | #define DPRINTK(fmt, args...) \ |
61 | do { \ | |
62 | if (debug_alternative) \ | |
63 | printk(KERN_DEBUG "%s: " fmt "\n", __func__, ##args); \ | |
c767a54b | 64 | } while (0) |
d167a518 | 65 | |
48c7a250 BP |
66 | #define DUMP_BYTES(buf, len, fmt, args...) \ |
67 | do { \ | |
68 | if (unlikely(debug_alternative)) { \ | |
69 | int j; \ | |
70 | \ | |
71 | if (!(len)) \ | |
72 | break; \ | |
73 | \ | |
74 | printk(KERN_DEBUG fmt, ##args); \ | |
75 | for (j = 0; j < (len) - 1; j++) \ | |
76 | printk(KERN_CONT "%02hhx ", buf[j]); \ | |
77 | printk(KERN_CONT "%02hhx\n", buf[j]); \ | |
78 | } \ | |
79 | } while (0) | |
80 | ||
dc326fca PA |
81 | /* |
82 | * Each GENERIC_NOPX is of X bytes, and defined as an array of bytes | |
83 | * that correspond to that nop. Getting from one nop to the next, we | |
84 | * add to the array the offset that is equal to the sum of all sizes of | |
85 | * nops preceding the one we are after. | |
86 | * | |
87 | * Note: The GENERIC_NOP5_ATOMIC is at the end, as it breaks the | |
88 | * nice symmetry of sizes of the previous nops. | |
89 | */ | |
8b5a10fc | 90 | #if defined(GENERIC_NOP1) && !defined(CONFIG_X86_64) |
dc326fca PA |
91 | static const unsigned char intelnops[] = |
92 | { | |
93 | GENERIC_NOP1, | |
94 | GENERIC_NOP2, | |
95 | GENERIC_NOP3, | |
96 | GENERIC_NOP4, | |
97 | GENERIC_NOP5, | |
98 | GENERIC_NOP6, | |
99 | GENERIC_NOP7, | |
100 | GENERIC_NOP8, | |
101 | GENERIC_NOP5_ATOMIC | |
102 | }; | |
103 | static const unsigned char * const intel_nops[ASM_NOP_MAX+2] = | |
104 | { | |
9a0b5817 GH |
105 | NULL, |
106 | intelnops, | |
107 | intelnops + 1, | |
108 | intelnops + 1 + 2, | |
109 | intelnops + 1 + 2 + 3, | |
110 | intelnops + 1 + 2 + 3 + 4, | |
111 | intelnops + 1 + 2 + 3 + 4 + 5, | |
112 | intelnops + 1 + 2 + 3 + 4 + 5 + 6, | |
113 | intelnops + 1 + 2 + 3 + 4 + 5 + 6 + 7, | |
dc326fca | 114 | intelnops + 1 + 2 + 3 + 4 + 5 + 6 + 7 + 8, |
9a0b5817 | 115 | }; |
d167a518 GH |
116 | #endif |
117 | ||
118 | #ifdef K8_NOP1 | |
dc326fca PA |
119 | static const unsigned char k8nops[] = |
120 | { | |
121 | K8_NOP1, | |
122 | K8_NOP2, | |
123 | K8_NOP3, | |
124 | K8_NOP4, | |
125 | K8_NOP5, | |
126 | K8_NOP6, | |
127 | K8_NOP7, | |
128 | K8_NOP8, | |
129 | K8_NOP5_ATOMIC | |
130 | }; | |
131 | static const unsigned char * const k8_nops[ASM_NOP_MAX+2] = | |
132 | { | |
9a0b5817 GH |
133 | NULL, |
134 | k8nops, | |
135 | k8nops + 1, | |
136 | k8nops + 1 + 2, | |
137 | k8nops + 1 + 2 + 3, | |
138 | k8nops + 1 + 2 + 3 + 4, | |
139 | k8nops + 1 + 2 + 3 + 4 + 5, | |
140 | k8nops + 1 + 2 + 3 + 4 + 5 + 6, | |
141 | k8nops + 1 + 2 + 3 + 4 + 5 + 6 + 7, | |
dc326fca | 142 | k8nops + 1 + 2 + 3 + 4 + 5 + 6 + 7 + 8, |
9a0b5817 | 143 | }; |
d167a518 GH |
144 | #endif |
145 | ||
8b5a10fc | 146 | #if defined(K7_NOP1) && !defined(CONFIG_X86_64) |
dc326fca PA |
147 | static const unsigned char k7nops[] = |
148 | { | |
149 | K7_NOP1, | |
150 | K7_NOP2, | |
151 | K7_NOP3, | |
152 | K7_NOP4, | |
153 | K7_NOP5, | |
154 | K7_NOP6, | |
155 | K7_NOP7, | |
156 | K7_NOP8, | |
157 | K7_NOP5_ATOMIC | |
158 | }; | |
159 | static const unsigned char * const k7_nops[ASM_NOP_MAX+2] = | |
160 | { | |
9a0b5817 GH |
161 | NULL, |
162 | k7nops, | |
163 | k7nops + 1, | |
164 | k7nops + 1 + 2, | |
165 | k7nops + 1 + 2 + 3, | |
166 | k7nops + 1 + 2 + 3 + 4, | |
167 | k7nops + 1 + 2 + 3 + 4 + 5, | |
168 | k7nops + 1 + 2 + 3 + 4 + 5 + 6, | |
169 | k7nops + 1 + 2 + 3 + 4 + 5 + 6 + 7, | |
dc326fca | 170 | k7nops + 1 + 2 + 3 + 4 + 5 + 6 + 7 + 8, |
9a0b5817 | 171 | }; |
d167a518 GH |
172 | #endif |
173 | ||
32c464f5 | 174 | #ifdef P6_NOP1 |
cb09cad4 | 175 | static const unsigned char p6nops[] = |
dc326fca PA |
176 | { |
177 | P6_NOP1, | |
178 | P6_NOP2, | |
179 | P6_NOP3, | |
180 | P6_NOP4, | |
181 | P6_NOP5, | |
182 | P6_NOP6, | |
183 | P6_NOP7, | |
184 | P6_NOP8, | |
185 | P6_NOP5_ATOMIC | |
186 | }; | |
187 | static const unsigned char * const p6_nops[ASM_NOP_MAX+2] = | |
188 | { | |
32c464f5 JB |
189 | NULL, |
190 | p6nops, | |
191 | p6nops + 1, | |
192 | p6nops + 1 + 2, | |
193 | p6nops + 1 + 2 + 3, | |
194 | p6nops + 1 + 2 + 3 + 4, | |
195 | p6nops + 1 + 2 + 3 + 4 + 5, | |
196 | p6nops + 1 + 2 + 3 + 4 + 5 + 6, | |
197 | p6nops + 1 + 2 + 3 + 4 + 5 + 6 + 7, | |
dc326fca | 198 | p6nops + 1 + 2 + 3 + 4 + 5 + 6 + 7 + 8, |
32c464f5 JB |
199 | }; |
200 | #endif | |
201 | ||
dc326fca | 202 | /* Initialize these to a safe default */ |
d167a518 | 203 | #ifdef CONFIG_X86_64 |
dc326fca PA |
204 | const unsigned char * const *ideal_nops = p6_nops; |
205 | #else | |
206 | const unsigned char * const *ideal_nops = intel_nops; | |
207 | #endif | |
d167a518 | 208 | |
dc326fca | 209 | void __init arch_init_ideal_nops(void) |
d167a518 | 210 | { |
dc326fca PA |
211 | switch (boot_cpu_data.x86_vendor) { |
212 | case X86_VENDOR_INTEL: | |
d8d9766c PA |
213 | /* |
214 | * Due to a decoder implementation quirk, some | |
215 | * specific Intel CPUs actually perform better with | |
216 | * the "k8_nops" than with the SDM-recommended NOPs. | |
217 | */ | |
218 | if (boot_cpu_data.x86 == 6 && | |
219 | boot_cpu_data.x86_model >= 0x0f && | |
220 | boot_cpu_data.x86_model != 0x1c && | |
221 | boot_cpu_data.x86_model != 0x26 && | |
222 | boot_cpu_data.x86_model != 0x27 && | |
223 | boot_cpu_data.x86_model < 0x30) { | |
224 | ideal_nops = k8_nops; | |
225 | } else if (boot_cpu_has(X86_FEATURE_NOPL)) { | |
dc326fca PA |
226 | ideal_nops = p6_nops; |
227 | } else { | |
228 | #ifdef CONFIG_X86_64 | |
229 | ideal_nops = k8_nops; | |
230 | #else | |
231 | ideal_nops = intel_nops; | |
232 | #endif | |
233 | } | |
d6250a3f | 234 | break; |
f21262b8 BP |
235 | |
236 | case X86_VENDOR_AMD: | |
237 | if (boot_cpu_data.x86 > 0xf) { | |
238 | ideal_nops = p6_nops; | |
239 | return; | |
240 | } | |
241 | ||
242 | /* fall through */ | |
243 | ||
dc326fca PA |
244 | default: |
245 | #ifdef CONFIG_X86_64 | |
246 | ideal_nops = k8_nops; | |
247 | #else | |
248 | if (boot_cpu_has(X86_FEATURE_K8)) | |
249 | ideal_nops = k8_nops; | |
250 | else if (boot_cpu_has(X86_FEATURE_K7)) | |
251 | ideal_nops = k7_nops; | |
252 | else | |
253 | ideal_nops = intel_nops; | |
254 | #endif | |
255 | } | |
9a0b5817 GH |
256 | } |
257 | ||
ab144f5e | 258 | /* Use this to add nops to a buffer, then text_poke the whole buffer. */ |
8b5a10fc | 259 | static void __init_or_module add_nops(void *insns, unsigned int len) |
139ec7c4 | 260 | { |
139ec7c4 RR |
261 | while (len > 0) { |
262 | unsigned int noplen = len; | |
263 | if (noplen > ASM_NOP_MAX) | |
264 | noplen = ASM_NOP_MAX; | |
dc326fca | 265 | memcpy(insns, ideal_nops[noplen], noplen); |
139ec7c4 RR |
266 | insns += noplen; |
267 | len -= noplen; | |
268 | } | |
269 | } | |
270 | ||
d167a518 | 271 | extern struct alt_instr __alt_instructions[], __alt_instructions_end[]; |
5967ed87 | 272 | extern s32 __smp_locks[], __smp_locks_end[]; |
fa6f2cc7 | 273 | void *text_poke_early(void *addr, const void *opcode, size_t len); |
d167a518 | 274 | |
48c7a250 BP |
275 | /* |
276 | * Are we looking at a near JMP with a 1 or 4-byte displacement. | |
277 | */ | |
278 | static inline bool is_jmp(const u8 opcode) | |
279 | { | |
280 | return opcode == 0xeb || opcode == 0xe9; | |
281 | } | |
282 | ||
283 | static void __init_or_module | |
284 | recompute_jump(struct alt_instr *a, u8 *orig_insn, u8 *repl_insn, u8 *insnbuf) | |
285 | { | |
286 | u8 *next_rip, *tgt_rip; | |
287 | s32 n_dspl, o_dspl; | |
288 | int repl_len; | |
289 | ||
290 | if (a->replacementlen != 5) | |
291 | return; | |
292 | ||
293 | o_dspl = *(s32 *)(insnbuf + 1); | |
294 | ||
295 | /* next_rip of the replacement JMP */ | |
296 | next_rip = repl_insn + a->replacementlen; | |
297 | /* target rip of the replacement JMP */ | |
298 | tgt_rip = next_rip + o_dspl; | |
299 | n_dspl = tgt_rip - orig_insn; | |
300 | ||
301 | DPRINTK("target RIP: %p, new_displ: 0x%x", tgt_rip, n_dspl); | |
302 | ||
303 | if (tgt_rip - orig_insn >= 0) { | |
304 | if (n_dspl - 2 <= 127) | |
305 | goto two_byte_jmp; | |
306 | else | |
307 | goto five_byte_jmp; | |
308 | /* negative offset */ | |
309 | } else { | |
310 | if (((n_dspl - 2) & 0xff) == (n_dspl - 2)) | |
311 | goto two_byte_jmp; | |
312 | else | |
313 | goto five_byte_jmp; | |
314 | } | |
315 | ||
316 | two_byte_jmp: | |
317 | n_dspl -= 2; | |
318 | ||
319 | insnbuf[0] = 0xeb; | |
320 | insnbuf[1] = (s8)n_dspl; | |
321 | add_nops(insnbuf + 2, 3); | |
322 | ||
323 | repl_len = 2; | |
324 | goto done; | |
325 | ||
326 | five_byte_jmp: | |
327 | n_dspl -= 5; | |
328 | ||
329 | insnbuf[0] = 0xe9; | |
330 | *(s32 *)&insnbuf[1] = n_dspl; | |
331 | ||
332 | repl_len = 5; | |
333 | ||
334 | done: | |
335 | ||
336 | DPRINTK("final displ: 0x%08x, JMP 0x%lx", | |
337 | n_dspl, (unsigned long)orig_insn + n_dspl + repl_len); | |
338 | } | |
339 | ||
34bfab0e BP |
340 | /* |
341 | * "noinline" to cause control flow change and thus invalidate I$ and | |
342 | * cause refetch after modification. | |
343 | */ | |
344 | static void __init_or_module noinline optimize_nops(struct alt_instr *a, u8 *instr) | |
4fd4b6e5 | 345 | { |
66c117d7 TG |
346 | unsigned long flags; |
347 | ||
69df353f BP |
348 | if (instr[0] != 0x90) |
349 | return; | |
350 | ||
66c117d7 | 351 | local_irq_save(flags); |
4fd4b6e5 | 352 | add_nops(instr + (a->instrlen - a->padlen), a->padlen); |
66c117d7 | 353 | local_irq_restore(flags); |
4fd4b6e5 BP |
354 | |
355 | DUMP_BYTES(instr, a->instrlen, "%p: [%d:%d) optimized NOPs: ", | |
356 | instr, a->instrlen - a->padlen, a->padlen); | |
357 | } | |
358 | ||
db477a33 BP |
359 | /* |
360 | * Replace instructions with better alternatives for this CPU type. This runs | |
361 | * before SMP is initialized to avoid SMP problems with self modifying code. | |
362 | * This implies that asymmetric systems where APs have less capabilities than | |
363 | * the boot processor are not handled. Tough. Make sure you disable such | |
364 | * features by hand. | |
34bfab0e BP |
365 | * |
366 | * Marked "noinline" to cause control flow change and thus insn cache | |
367 | * to refetch changed I$ lines. | |
db477a33 | 368 | */ |
34bfab0e BP |
369 | void __init_or_module noinline apply_alternatives(struct alt_instr *start, |
370 | struct alt_instr *end) | |
9a0b5817 | 371 | { |
9a0b5817 | 372 | struct alt_instr *a; |
59e97e4d | 373 | u8 *instr, *replacement; |
1b1d9258 | 374 | u8 insnbuf[MAX_PATCH_LEN]; |
9a0b5817 | 375 | |
db477a33 | 376 | DPRINTK("alt table %p -> %p", start, end); |
50973133 FY |
377 | /* |
378 | * The scan order should be from start to end. A later scanned | |
db477a33 | 379 | * alternative code can overwrite previously scanned alternative code. |
50973133 FY |
380 | * Some kernel functions (e.g. memcpy, memset, etc) use this order to |
381 | * patch code. | |
382 | * | |
383 | * So be careful if you want to change the scan order to any other | |
384 | * order. | |
385 | */ | |
9a0b5817 | 386 | for (a = start; a < end; a++) { |
48c7a250 BP |
387 | int insnbuf_sz = 0; |
388 | ||
59e97e4d AL |
389 | instr = (u8 *)&a->instr_offset + a->instr_offset; |
390 | replacement = (u8 *)&a->repl_offset + a->repl_offset; | |
ab144f5e | 391 | BUG_ON(a->instrlen > sizeof(insnbuf)); |
65fc985b | 392 | BUG_ON(a->cpuid >= (NCAPINTS + NBUGINTS) * 32); |
4fd4b6e5 BP |
393 | if (!boot_cpu_has(a->cpuid)) { |
394 | if (a->padlen > 1) | |
395 | optimize_nops(a, instr); | |
396 | ||
9a0b5817 | 397 | continue; |
4fd4b6e5 | 398 | } |
59e97e4d | 399 | |
dbe4058a | 400 | DPRINTK("feat: %d*32+%d, old: (%p, len: %d), repl: (%p, len: %d), pad: %d", |
db477a33 BP |
401 | a->cpuid >> 5, |
402 | a->cpuid & 0x1f, | |
403 | instr, a->instrlen, | |
dbe4058a | 404 | replacement, a->replacementlen, a->padlen); |
db477a33 | 405 | |
48c7a250 BP |
406 | DUMP_BYTES(instr, a->instrlen, "%p: old_insn: ", instr); |
407 | DUMP_BYTES(replacement, a->replacementlen, "%p: rpl_insn: ", replacement); | |
408 | ||
59e97e4d | 409 | memcpy(insnbuf, replacement, a->replacementlen); |
48c7a250 | 410 | insnbuf_sz = a->replacementlen; |
59e97e4d | 411 | |
fc152d22 MJ |
412 | /* |
413 | * 0xe8 is a relative jump; fix the offset. | |
414 | * | |
415 | * Instruction length is checked before the opcode to avoid | |
416 | * accessing uninitialized bytes for zero-length replacements. | |
417 | */ | |
418 | if (a->replacementlen == 5 && *insnbuf == 0xe8) { | |
db477a33 | 419 | *(s32 *)(insnbuf + 1) += replacement - instr; |
48c7a250 BP |
420 | DPRINTK("Fix CALL offset: 0x%x, CALL 0x%lx", |
421 | *(s32 *)(insnbuf + 1), | |
422 | (unsigned long)instr + *(s32 *)(insnbuf + 1) + 5); | |
db477a33 | 423 | } |
59e97e4d | 424 | |
48c7a250 BP |
425 | if (a->replacementlen && is_jmp(replacement[0])) |
426 | recompute_jump(a, instr, replacement, insnbuf); | |
427 | ||
428 | if (a->instrlen > a->replacementlen) { | |
4332195c BP |
429 | add_nops(insnbuf + a->replacementlen, |
430 | a->instrlen - a->replacementlen); | |
48c7a250 BP |
431 | insnbuf_sz += a->instrlen - a->replacementlen; |
432 | } | |
433 | DUMP_BYTES(insnbuf, insnbuf_sz, "%p: final_insn: ", instr); | |
59e97e4d | 434 | |
48c7a250 | 435 | text_poke_early(instr, insnbuf, insnbuf_sz); |
9a0b5817 GH |
436 | } |
437 | } | |
438 | ||
8ec4d41f | 439 | #ifdef CONFIG_SMP |
5967ed87 JB |
440 | static void alternatives_smp_lock(const s32 *start, const s32 *end, |
441 | u8 *text, u8 *text_end) | |
9a0b5817 | 442 | { |
5967ed87 | 443 | const s32 *poff; |
9a0b5817 | 444 | |
5967ed87 JB |
445 | for (poff = start; poff < end; poff++) { |
446 | u8 *ptr = (u8 *)poff + *poff; | |
447 | ||
448 | if (!*poff || ptr < text || ptr >= text_end) | |
9a0b5817 | 449 | continue; |
f88f07e0 | 450 | /* turn DS segment override prefix into lock prefix */ |
d9c5841e PA |
451 | if (*ptr == 0x3e) |
452 | text_poke(ptr, ((unsigned char []){0xf0}), 1); | |
4b8073e4 | 453 | } |
9a0b5817 GH |
454 | } |
455 | ||
5967ed87 JB |
456 | static void alternatives_smp_unlock(const s32 *start, const s32 *end, |
457 | u8 *text, u8 *text_end) | |
9a0b5817 | 458 | { |
5967ed87 | 459 | const s32 *poff; |
9a0b5817 | 460 | |
5967ed87 JB |
461 | for (poff = start; poff < end; poff++) { |
462 | u8 *ptr = (u8 *)poff + *poff; | |
463 | ||
464 | if (!*poff || ptr < text || ptr >= text_end) | |
9a0b5817 | 465 | continue; |
f88f07e0 | 466 | /* turn lock prefix into DS segment override prefix */ |
d9c5841e PA |
467 | if (*ptr == 0xf0) |
468 | text_poke(ptr, ((unsigned char []){0x3E}), 1); | |
4b8073e4 | 469 | } |
9a0b5817 GH |
470 | } |
471 | ||
472 | struct smp_alt_module { | |
473 | /* what is this ??? */ | |
474 | struct module *mod; | |
475 | char *name; | |
476 | ||
477 | /* ptrs to lock prefixes */ | |
5967ed87 JB |
478 | const s32 *locks; |
479 | const s32 *locks_end; | |
9a0b5817 GH |
480 | |
481 | /* .text segment, needed to avoid patching init code ;) */ | |
482 | u8 *text; | |
483 | u8 *text_end; | |
484 | ||
485 | struct list_head next; | |
486 | }; | |
487 | static LIST_HEAD(smp_alt_modules); | |
e846d139 | 488 | static bool uniproc_patched = false; /* protected by text_mutex */ |
9a0b5817 | 489 | |
8b5a10fc JB |
490 | void __init_or_module alternatives_smp_module_add(struct module *mod, |
491 | char *name, | |
492 | void *locks, void *locks_end, | |
493 | void *text, void *text_end) | |
9a0b5817 GH |
494 | { |
495 | struct smp_alt_module *smp; | |
9a0b5817 | 496 | |
e846d139 | 497 | mutex_lock(&text_mutex); |
816afe4f RR |
498 | if (!uniproc_patched) |
499 | goto unlock; | |
b7fb4af0 | 500 | |
816afe4f RR |
501 | if (num_possible_cpus() == 1) |
502 | /* Don't bother remembering, we'll never have to undo it. */ | |
503 | goto smp_unlock; | |
9a0b5817 GH |
504 | |
505 | smp = kzalloc(sizeof(*smp), GFP_KERNEL); | |
506 | if (NULL == smp) | |
816afe4f RR |
507 | /* we'll run the (safe but slow) SMP code then ... */ |
508 | goto unlock; | |
9a0b5817 GH |
509 | |
510 | smp->mod = mod; | |
511 | smp->name = name; | |
512 | smp->locks = locks; | |
513 | smp->locks_end = locks_end; | |
514 | smp->text = text; | |
515 | smp->text_end = text_end; | |
db477a33 BP |
516 | DPRINTK("locks %p -> %p, text %p -> %p, name %s\n", |
517 | smp->locks, smp->locks_end, | |
9a0b5817 GH |
518 | smp->text, smp->text_end, smp->name); |
519 | ||
9a0b5817 | 520 | list_add_tail(&smp->next, &smp_alt_modules); |
816afe4f RR |
521 | smp_unlock: |
522 | alternatives_smp_unlock(locks, locks_end, text, text_end); | |
523 | unlock: | |
e846d139 | 524 | mutex_unlock(&text_mutex); |
9a0b5817 GH |
525 | } |
526 | ||
8b5a10fc | 527 | void __init_or_module alternatives_smp_module_del(struct module *mod) |
9a0b5817 GH |
528 | { |
529 | struct smp_alt_module *item; | |
9a0b5817 | 530 | |
e846d139 | 531 | mutex_lock(&text_mutex); |
9a0b5817 GH |
532 | list_for_each_entry(item, &smp_alt_modules, next) { |
533 | if (mod != item->mod) | |
534 | continue; | |
535 | list_del(&item->next); | |
9a0b5817 | 536 | kfree(item); |
816afe4f | 537 | break; |
9a0b5817 | 538 | } |
e846d139 | 539 | mutex_unlock(&text_mutex); |
9a0b5817 GH |
540 | } |
541 | ||
816afe4f | 542 | void alternatives_enable_smp(void) |
9a0b5817 GH |
543 | { |
544 | struct smp_alt_module *mod; | |
9a0b5817 | 545 | |
816afe4f RR |
546 | /* Why bother if there are no other CPUs? */ |
547 | BUG_ON(num_possible_cpus() == 1); | |
9a0b5817 | 548 | |
e846d139 | 549 | mutex_lock(&text_mutex); |
ca74a6f8 | 550 | |
816afe4f | 551 | if (uniproc_patched) { |
c767a54b | 552 | pr_info("switching to SMP code\n"); |
816afe4f | 553 | BUG_ON(num_online_cpus() != 1); |
53756d37 JF |
554 | clear_cpu_cap(&boot_cpu_data, X86_FEATURE_UP); |
555 | clear_cpu_cap(&cpu_data(0), X86_FEATURE_UP); | |
9a0b5817 GH |
556 | list_for_each_entry(mod, &smp_alt_modules, next) |
557 | alternatives_smp_lock(mod->locks, mod->locks_end, | |
558 | mod->text, mod->text_end); | |
816afe4f | 559 | uniproc_patched = false; |
9a0b5817 | 560 | } |
e846d139 | 561 | mutex_unlock(&text_mutex); |
9a0b5817 GH |
562 | } |
563 | ||
e846d139 ZC |
564 | /* |
565 | * Return 1 if the address range is reserved for SMP-alternatives. | |
566 | * Must hold text_mutex. | |
567 | */ | |
2cfa1978 MH |
568 | int alternatives_text_reserved(void *start, void *end) |
569 | { | |
570 | struct smp_alt_module *mod; | |
5967ed87 | 571 | const s32 *poff; |
076dc4a6 MH |
572 | u8 *text_start = start; |
573 | u8 *text_end = end; | |
2cfa1978 | 574 | |
e846d139 ZC |
575 | lockdep_assert_held(&text_mutex); |
576 | ||
2cfa1978 | 577 | list_for_each_entry(mod, &smp_alt_modules, next) { |
076dc4a6 | 578 | if (mod->text > text_end || mod->text_end < text_start) |
2cfa1978 | 579 | continue; |
5967ed87 JB |
580 | for (poff = mod->locks; poff < mod->locks_end; poff++) { |
581 | const u8 *ptr = (const u8 *)poff + *poff; | |
582 | ||
583 | if (text_start <= ptr && text_end > ptr) | |
2cfa1978 | 584 | return 1; |
5967ed87 | 585 | } |
2cfa1978 MH |
586 | } |
587 | ||
588 | return 0; | |
589 | } | |
48c7a250 | 590 | #endif /* CONFIG_SMP */ |
8ec4d41f | 591 | |
139ec7c4 | 592 | #ifdef CONFIG_PARAVIRT |
8b5a10fc JB |
593 | void __init_or_module apply_paravirt(struct paravirt_patch_site *start, |
594 | struct paravirt_patch_site *end) | |
139ec7c4 | 595 | { |
98de032b | 596 | struct paravirt_patch_site *p; |
ab144f5e | 597 | char insnbuf[MAX_PATCH_LEN]; |
139ec7c4 | 598 | |
959b4fdf JF |
599 | if (noreplace_paravirt) |
600 | return; | |
601 | ||
139ec7c4 RR |
602 | for (p = start; p < end; p++) { |
603 | unsigned int used; | |
604 | ||
ab144f5e | 605 | BUG_ON(p->len > MAX_PATCH_LEN); |
d34fda4a CW |
606 | /* prep the buffer with the original instructions */ |
607 | memcpy(insnbuf, p->instr, p->len); | |
93b1eab3 JF |
608 | used = pv_init_ops.patch(p->instrtype, p->clobbers, insnbuf, |
609 | (unsigned long)p->instr, p->len); | |
7f63c41c | 610 | |
63f70270 JF |
611 | BUG_ON(used > p->len); |
612 | ||
139ec7c4 | 613 | /* Pad the rest with nops */ |
ab144f5e | 614 | add_nops(insnbuf + used, p->len - used); |
e587cadd | 615 | text_poke_early(p->instr, insnbuf, p->len); |
139ec7c4 | 616 | } |
139ec7c4 | 617 | } |
98de032b | 618 | extern struct paravirt_patch_site __start_parainstructions[], |
139ec7c4 RR |
619 | __stop_parainstructions[]; |
620 | #endif /* CONFIG_PARAVIRT */ | |
621 | ||
9a0b5817 GH |
622 | void __init alternative_instructions(void) |
623 | { | |
8f4e956b AK |
624 | /* The patching is not fully atomic, so try to avoid local interruptions |
625 | that might execute the to be patched code. | |
626 | Other CPUs are not running. */ | |
627 | stop_nmi(); | |
123aa76e AK |
628 | |
629 | /* | |
630 | * Don't stop machine check exceptions while patching. | |
631 | * MCEs only happen when something got corrupted and in this | |
632 | * case we must do something about the corruption. | |
633 | * Ignoring it is worse than a unlikely patching race. | |
634 | * Also machine checks tend to be broadcast and if one CPU | |
635 | * goes into machine check the others follow quickly, so we don't | |
636 | * expect a machine check to cause undue problems during to code | |
637 | * patching. | |
638 | */ | |
8f4e956b | 639 | |
9a0b5817 GH |
640 | apply_alternatives(__alt_instructions, __alt_instructions_end); |
641 | ||
8ec4d41f | 642 | #ifdef CONFIG_SMP |
816afe4f RR |
643 | /* Patch to UP if other cpus not imminent. */ |
644 | if (!noreplace_smp && (num_present_cpus() == 1 || setup_max_cpus <= 1)) { | |
645 | uniproc_patched = true; | |
9a0b5817 GH |
646 | alternatives_smp_module_add(NULL, "core kernel", |
647 | __smp_locks, __smp_locks_end, | |
648 | _text, _etext); | |
9a0b5817 | 649 | } |
8f4e956b | 650 | |
816afe4f | 651 | if (!uniproc_patched || num_possible_cpus() == 1) |
f68fd5f4 FW |
652 | free_init_pages("SMP alternatives", |
653 | (unsigned long)__smp_locks, | |
654 | (unsigned long)__smp_locks_end); | |
816afe4f RR |
655 | #endif |
656 | ||
657 | apply_paravirt(__parainstructions, __parainstructions_end); | |
f68fd5f4 | 658 | |
8f4e956b | 659 | restart_nmi(); |
5e907bb0 | 660 | alternatives_patched = 1; |
9a0b5817 | 661 | } |
19d36ccd | 662 | |
e587cadd MD |
663 | /** |
664 | * text_poke_early - Update instructions on a live kernel at boot time | |
665 | * @addr: address to modify | |
666 | * @opcode: source of the copy | |
667 | * @len: length to copy | |
668 | * | |
19d36ccd AK |
669 | * When you use this code to patch more than one byte of an instruction |
670 | * you need to make sure that other CPUs cannot execute this code in parallel. | |
e587cadd MD |
671 | * Also no thread must be currently preempted in the middle of these |
672 | * instructions. And on the local CPU you need to be protected again NMI or MCE | |
673 | * handlers seeing an inconsistent instruction while you patch. | |
19d36ccd | 674 | */ |
fa6f2cc7 | 675 | void *__init_or_module text_poke_early(void *addr, const void *opcode, |
8b5a10fc | 676 | size_t len) |
19d36ccd | 677 | { |
e587cadd MD |
678 | unsigned long flags; |
679 | local_irq_save(flags); | |
19d36ccd | 680 | memcpy(addr, opcode, len); |
5367b688 | 681 | local_irq_restore(flags); |
e587cadd MD |
682 | /* Could also do a CLFLUSH here to speed up CPU recovery; but |
683 | that causes hangs on some VIA CPUs. */ | |
684 | return addr; | |
685 | } | |
686 | ||
687 | /** | |
688 | * text_poke - Update instructions on a live kernel | |
689 | * @addr: address to modify | |
690 | * @opcode: source of the copy | |
691 | * @len: length to copy | |
692 | * | |
693 | * Only atomic text poke/set should be allowed when not doing early patching. | |
694 | * It means the size must be writable atomically and the address must be aligned | |
695 | * in a way that permits an atomic write. It also makes sure we fit on a single | |
696 | * page. | |
78ff7fae MH |
697 | * |
698 | * Note: Must be called under text_mutex. | |
e587cadd | 699 | */ |
9c54b616 | 700 | void *text_poke(void *addr, const void *opcode, size_t len) |
e587cadd | 701 | { |
78ff7fae | 702 | unsigned long flags; |
e587cadd | 703 | char *vaddr; |
b7b66baa MD |
704 | struct page *pages[2]; |
705 | int i; | |
e587cadd | 706 | |
b7b66baa MD |
707 | if (!core_kernel_text((unsigned long)addr)) { |
708 | pages[0] = vmalloc_to_page(addr); | |
709 | pages[1] = vmalloc_to_page(addr + PAGE_SIZE); | |
15a601eb | 710 | } else { |
b7b66baa | 711 | pages[0] = virt_to_page(addr); |
00c6b2d5 | 712 | WARN_ON(!PageReserved(pages[0])); |
b7b66baa | 713 | pages[1] = virt_to_page(addr + PAGE_SIZE); |
e587cadd | 714 | } |
b7b66baa | 715 | BUG_ON(!pages[0]); |
7cf49427 | 716 | local_irq_save(flags); |
78ff7fae MH |
717 | set_fixmap(FIX_TEXT_POKE0, page_to_phys(pages[0])); |
718 | if (pages[1]) | |
719 | set_fixmap(FIX_TEXT_POKE1, page_to_phys(pages[1])); | |
720 | vaddr = (char *)fix_to_virt(FIX_TEXT_POKE0); | |
b7b66baa | 721 | memcpy(&vaddr[(unsigned long)addr & ~PAGE_MASK], opcode, len); |
78ff7fae MH |
722 | clear_fixmap(FIX_TEXT_POKE0); |
723 | if (pages[1]) | |
724 | clear_fixmap(FIX_TEXT_POKE1); | |
725 | local_flush_tlb(); | |
19d36ccd | 726 | sync_core(); |
a534b679 AK |
727 | /* Could also do a CLFLUSH here to speed up CPU recovery; but |
728 | that causes hangs on some VIA CPUs. */ | |
b7b66baa MD |
729 | for (i = 0; i < len; i++) |
730 | BUG_ON(((char *)addr)[i] != ((char *)opcode)[i]); | |
7cf49427 | 731 | local_irq_restore(flags); |
e587cadd | 732 | return addr; |
19d36ccd | 733 | } |
3d55cc8a | 734 | |
fd4363ff JK |
735 | static void do_sync_core(void *info) |
736 | { | |
737 | sync_core(); | |
738 | } | |
739 | ||
740 | static bool bp_patching_in_progress; | |
741 | static void *bp_int3_handler, *bp_int3_addr; | |
742 | ||
17f41571 | 743 | int poke_int3_handler(struct pt_regs *regs) |
fd4363ff | 744 | { |
01651324 PZ |
745 | /* |
746 | * Having observed our INT3 instruction, we now must observe | |
747 | * bp_patching_in_progress. | |
748 | * | |
749 | * in_progress = TRUE INT3 | |
750 | * WMB RMB | |
751 | * write INT3 if (in_progress) | |
752 | * | |
753 | * Idem for bp_int3_handler. | |
754 | */ | |
fd4363ff JK |
755 | smp_rmb(); |
756 | ||
757 | if (likely(!bp_patching_in_progress)) | |
17f41571 | 758 | return 0; |
fd4363ff | 759 | |
f39b6f0e | 760 | if (user_mode(regs) || regs->ip != (unsigned long)bp_int3_addr) |
17f41571 | 761 | return 0; |
fd4363ff JK |
762 | |
763 | /* set up the specified breakpoint handler */ | |
17f41571 JK |
764 | regs->ip = (unsigned long) bp_int3_handler; |
765 | ||
766 | return 1; | |
fd4363ff | 767 | |
fd4363ff | 768 | } |
17f41571 | 769 | |
fd4363ff JK |
770 | /** |
771 | * text_poke_bp() -- update instructions on live kernel on SMP | |
772 | * @addr: address to patch | |
773 | * @opcode: opcode of new instruction | |
774 | * @len: length to copy | |
775 | * @handler: address to jump to when the temporary breakpoint is hit | |
776 | * | |
777 | * Modify multi-byte instruction by using int3 breakpoint on SMP. | |
ea8596bb MH |
778 | * We completely avoid stop_machine() here, and achieve the |
779 | * synchronization using int3 breakpoint. | |
fd4363ff JK |
780 | * |
781 | * The way it is done: | |
782 | * - add a int3 trap to the address that will be patched | |
783 | * - sync cores | |
784 | * - update all but the first byte of the patched range | |
785 | * - sync cores | |
786 | * - replace the first byte (int3) by the first byte of | |
787 | * replacing opcode | |
788 | * - sync cores | |
789 | * | |
790 | * Note: must be called under text_mutex. | |
791 | */ | |
792 | void *text_poke_bp(void *addr, const void *opcode, size_t len, void *handler) | |
793 | { | |
794 | unsigned char int3 = 0xcc; | |
795 | ||
796 | bp_int3_handler = handler; | |
797 | bp_int3_addr = (u8 *)addr + sizeof(int3); | |
798 | bp_patching_in_progress = true; | |
799 | /* | |
01651324 PZ |
800 | * Corresponding read barrier in int3 notifier for making sure the |
801 | * in_progress and handler are correctly ordered wrt. patching. | |
fd4363ff JK |
802 | */ |
803 | smp_wmb(); | |
804 | ||
805 | text_poke(addr, &int3, sizeof(int3)); | |
806 | ||
807 | on_each_cpu(do_sync_core, NULL, 1); | |
808 | ||
809 | if (len - sizeof(int3) > 0) { | |
810 | /* patch all but the first byte */ | |
811 | text_poke((char *)addr + sizeof(int3), | |
812 | (const char *) opcode + sizeof(int3), | |
813 | len - sizeof(int3)); | |
814 | /* | |
815 | * According to Intel, this core syncing is very likely | |
816 | * not necessary and we'd be safe even without it. But | |
817 | * better safe than sorry (plus there's not only Intel). | |
818 | */ | |
819 | on_each_cpu(do_sync_core, NULL, 1); | |
820 | } | |
821 | ||
822 | /* patch the first byte */ | |
823 | text_poke(addr, opcode, sizeof(int3)); | |
824 | ||
825 | on_each_cpu(do_sync_core, NULL, 1); | |
01651324 PZ |
826 | /* |
827 | * sync_core() implies an smp_mb() and orders this store against | |
828 | * the writing of the new instruction. | |
829 | */ | |
fd4363ff | 830 | bp_patching_in_progress = false; |
fd4363ff JK |
831 | |
832 | return addr; | |
833 | } | |
834 |