Commit | Line | Data |
---|---|---|
dc1e35c6 SS |
1 | /* |
2 | * xsave/xrstor support. | |
3 | * | |
4 | * Author: Suresh Siddha <suresh.b.siddha@intel.com> | |
5 | */ | |
6 | #include <linux/bootmem.h> | |
7 | #include <linux/compat.h> | |
8 | #include <asm/i387.h> | |
1361b83a | 9 | #include <asm/fpu-internal.h> |
c37b5efe SS |
10 | #ifdef CONFIG_IA32_EMULATION |
11 | #include <asm/sigcontext32.h> | |
12 | #endif | |
6152e4b1 | 13 | #include <asm/xcr.h> |
dc1e35c6 SS |
14 | |
15 | /* | |
16 | * Supported feature mask by the CPU and the kernel. | |
17 | */ | |
6152e4b1 | 18 | u64 pcntxt_mask; |
dc1e35c6 | 19 | |
45c2d7f4 RR |
20 | /* |
21 | * Represents init state for the supported extended state. | |
22 | */ | |
23 | static struct xsave_struct *init_xstate_buf; | |
24 | ||
c37b5efe SS |
25 | struct _fpx_sw_bytes fx_sw_reserved; |
26 | #ifdef CONFIG_IA32_EMULATION | |
27 | struct _fpx_sw_bytes fx_sw_reserved_ia32; | |
28 | #endif | |
29 | ||
a1488f8b SS |
30 | static unsigned int *xstate_offsets, *xstate_sizes, xstate_features; |
31 | ||
29104e10 SS |
32 | /* |
33 | * If a processor implementation discern that a processor state component is | |
34 | * in its initialized state it may modify the corresponding bit in the | |
35 | * xsave_hdr.xstate_bv as '0', with out modifying the corresponding memory | |
36 | * layout in the case of xsaveopt. While presenting the xstate information to | |
37 | * the user, we always ensure that the memory layout of a feature will be in | |
38 | * the init state if the corresponding header bit is zero. This is to ensure | |
39 | * that the user doesn't see some stale state in the memory layout during | |
40 | * signal handling, debugging etc. | |
41 | */ | |
42 | void __sanitize_i387_state(struct task_struct *tsk) | |
43 | { | |
44 | u64 xstate_bv; | |
45 | int feature_bit = 0x2; | |
46 | struct i387_fxsave_struct *fx = &tsk->thread.fpu.state->fxsave; | |
47 | ||
48 | if (!fx) | |
49 | return; | |
50 | ||
29104e10 SS |
51 | xstate_bv = tsk->thread.fpu.state->xsave.xsave_hdr.xstate_bv; |
52 | ||
53 | /* | |
54 | * None of the feature bits are in init state. So nothing else | |
0d2eb44f | 55 | * to do for us, as the memory layout is up to date. |
29104e10 SS |
56 | */ |
57 | if ((xstate_bv & pcntxt_mask) == pcntxt_mask) | |
58 | return; | |
59 | ||
60 | /* | |
61 | * FP is in init state | |
62 | */ | |
63 | if (!(xstate_bv & XSTATE_FP)) { | |
64 | fx->cwd = 0x37f; | |
65 | fx->swd = 0; | |
66 | fx->twd = 0; | |
67 | fx->fop = 0; | |
68 | fx->rip = 0; | |
69 | fx->rdp = 0; | |
70 | memset(&fx->st_space[0], 0, 128); | |
71 | } | |
72 | ||
73 | /* | |
74 | * SSE is in init state | |
75 | */ | |
76 | if (!(xstate_bv & XSTATE_SSE)) | |
77 | memset(&fx->xmm_space[0], 0, 256); | |
78 | ||
79 | xstate_bv = (pcntxt_mask & ~xstate_bv) >> 2; | |
80 | ||
81 | /* | |
82 | * Update all the other memory layouts for which the corresponding | |
83 | * header bit is in the init state. | |
84 | */ | |
85 | while (xstate_bv) { | |
86 | if (xstate_bv & 0x1) { | |
87 | int offset = xstate_offsets[feature_bit]; | |
88 | int size = xstate_sizes[feature_bit]; | |
89 | ||
90 | memcpy(((void *) fx) + offset, | |
91 | ((void *) init_xstate_buf) + offset, | |
92 | size); | |
93 | } | |
94 | ||
95 | xstate_bv >>= 1; | |
96 | feature_bit++; | |
97 | } | |
98 | } | |
99 | ||
c37b5efe SS |
100 | /* |
101 | * Check for the presence of extended state information in the | |
102 | * user fpstate pointer in the sigcontext. | |
103 | */ | |
104 | int check_for_xstate(struct i387_fxsave_struct __user *buf, | |
105 | void __user *fpstate, | |
106 | struct _fpx_sw_bytes *fx_sw_user) | |
107 | { | |
108 | int min_xstate_size = sizeof(struct i387_fxsave_struct) + | |
109 | sizeof(struct xsave_hdr_struct); | |
110 | unsigned int magic2; | |
111 | int err; | |
112 | ||
113 | err = __copy_from_user(fx_sw_user, &buf->sw_reserved[0], | |
114 | sizeof(struct _fpx_sw_bytes)); | |
c37b5efe | 115 | if (err) |
d6d4d420 | 116 | return -EFAULT; |
c37b5efe SS |
117 | |
118 | /* | |
119 | * First Magic check failed. | |
120 | */ | |
121 | if (fx_sw_user->magic1 != FP_XSTATE_MAGIC1) | |
d6d4d420 | 122 | return -EINVAL; |
c37b5efe SS |
123 | |
124 | /* | |
125 | * Check for error scenarios. | |
126 | */ | |
127 | if (fx_sw_user->xstate_size < min_xstate_size || | |
128 | fx_sw_user->xstate_size > xstate_size || | |
129 | fx_sw_user->xstate_size > fx_sw_user->extended_size) | |
d6d4d420 | 130 | return -EINVAL; |
c37b5efe SS |
131 | |
132 | err = __get_user(magic2, (__u32 *) (((void *)fpstate) + | |
133 | fx_sw_user->extended_size - | |
134 | FP_XSTATE_MAGIC2_SIZE)); | |
d6d4d420 DC |
135 | if (err) |
136 | return err; | |
c37b5efe SS |
137 | /* |
138 | * Check for the presence of second magic word at the end of memory | |
139 | * layout. This detects the case where the user just copied the legacy | |
140 | * fpstate layout with out copying the extended state information | |
141 | * in the memory layout. | |
142 | */ | |
d6d4d420 DC |
143 | if (magic2 != FP_XSTATE_MAGIC2) |
144 | return -EFAULT; | |
c37b5efe SS |
145 | |
146 | return 0; | |
147 | } | |
148 | ||
ab513701 SS |
149 | #ifdef CONFIG_X86_64 |
150 | /* | |
151 | * Signal frame handlers. | |
152 | */ | |
153 | ||
154 | int save_i387_xstate(void __user *buf) | |
155 | { | |
156 | struct task_struct *tsk = current; | |
157 | int err = 0; | |
158 | ||
159 | if (!access_ok(VERIFY_WRITE, buf, sig_xstate_size)) | |
160 | return -EACCES; | |
161 | ||
f65bc214 | 162 | BUG_ON(sig_xstate_size < xstate_size); |
ab513701 | 163 | |
c37b5efe | 164 | if ((unsigned long)buf % 64) |
ab513701 SS |
165 | printk("save_i387_xstate: bad fpstate %p\n", buf); |
166 | ||
167 | if (!used_math()) | |
168 | return 0; | |
06c38d5e | 169 | |
15d8791c | 170 | if (user_has_fpu()) { |
c9ad4882 | 171 | if (use_xsave()) |
c37b5efe SS |
172 | err = xsave_user(buf); |
173 | else | |
174 | err = fxsave_user(buf); | |
175 | ||
ab513701 SS |
176 | if (err) |
177 | return err; | |
15d8791c | 178 | user_fpu_end(); |
ab513701 | 179 | } else { |
29104e10 | 180 | sanitize_i387_state(tsk); |
86603283 | 181 | if (__copy_to_user(buf, &tsk->thread.fpu.state->fxsave, |
ab513701 SS |
182 | xstate_size)) |
183 | return -1; | |
184 | } | |
c37b5efe | 185 | |
06c38d5e SS |
186 | clear_used_math(); /* trigger finit */ |
187 | ||
c9ad4882 | 188 | if (use_xsave()) { |
c37b5efe | 189 | struct _fpstate __user *fx = buf; |
04944b79 SS |
190 | struct _xstate __user *x = buf; |
191 | u64 xstate_bv; | |
c37b5efe SS |
192 | |
193 | err = __copy_to_user(&fx->sw_reserved, &fx_sw_reserved, | |
194 | sizeof(struct _fpx_sw_bytes)); | |
195 | ||
196 | err |= __put_user(FP_XSTATE_MAGIC2, | |
197 | (__u32 __user *) (buf + sig_xstate_size | |
198 | - FP_XSTATE_MAGIC2_SIZE)); | |
04944b79 SS |
199 | |
200 | /* | |
201 | * Read the xstate_bv which we copied (directly from the cpu or | |
202 | * from the state in task struct) to the user buffers and | |
203 | * set the FP/SSE bits. | |
204 | */ | |
205 | err |= __get_user(xstate_bv, &x->xstate_hdr.xstate_bv); | |
206 | ||
207 | /* | |
208 | * For legacy compatible, we always set FP/SSE bits in the bit | |
209 | * vector while saving the state to the user context. This will | |
210 | * enable us capturing any changes(during sigreturn) to | |
211 | * the FP/SSE bits by the legacy applications which don't touch | |
212 | * xstate_bv in the xsave header. | |
213 | * | |
214 | * xsave aware apps can change the xstate_bv in the xsave | |
215 | * header as well as change any contents in the memory layout. | |
216 | * xrestore as part of sigreturn will capture all the changes. | |
217 | */ | |
218 | xstate_bv |= XSTATE_FPSSE; | |
219 | ||
220 | err |= __put_user(xstate_bv, &x->xstate_hdr.xstate_bv); | |
221 | ||
f364eada SS |
222 | if (err) |
223 | return err; | |
c37b5efe SS |
224 | } |
225 | ||
ab513701 SS |
226 | return 1; |
227 | } | |
228 | ||
c37b5efe SS |
229 | /* |
230 | * Restore the extended state if present. Otherwise, restore the FP/SSE | |
231 | * state. | |
232 | */ | |
7820b756 | 233 | static int restore_user_xstate(void __user *buf) |
c37b5efe SS |
234 | { |
235 | struct _fpx_sw_bytes fx_sw_user; | |
6152e4b1 | 236 | u64 mask; |
c37b5efe SS |
237 | int err; |
238 | ||
239 | if (((unsigned long)buf % 64) || | |
240 | check_for_xstate(buf, buf, &fx_sw_user)) | |
241 | goto fx_only; | |
242 | ||
6152e4b1 | 243 | mask = fx_sw_user.xstate_bv; |
c37b5efe SS |
244 | |
245 | /* | |
246 | * restore the state passed by the user. | |
247 | */ | |
6152e4b1 | 248 | err = xrestore_user(buf, mask); |
c37b5efe SS |
249 | if (err) |
250 | return err; | |
251 | ||
252 | /* | |
253 | * init the state skipped by the user. | |
254 | */ | |
6152e4b1 | 255 | mask = pcntxt_mask & ~mask; |
8e221b6d SS |
256 | if (unlikely(mask)) |
257 | xrstor_state(init_xstate_buf, mask); | |
c37b5efe SS |
258 | |
259 | return 0; | |
260 | ||
261 | fx_only: | |
262 | /* | |
263 | * couldn't find the extended state information in the | |
264 | * memory layout. Restore just the FP/SSE and init all | |
265 | * the other extended state. | |
266 | */ | |
6152e4b1 | 267 | xrstor_state(init_xstate_buf, pcntxt_mask & ~XSTATE_FPSSE); |
c37b5efe SS |
268 | return fxrstor_checking((__force struct i387_fxsave_struct *)buf); |
269 | } | |
270 | ||
ab513701 SS |
271 | /* |
272 | * This restores directly out of user space. Exceptions are handled. | |
273 | */ | |
274 | int restore_i387_xstate(void __user *buf) | |
275 | { | |
276 | struct task_struct *tsk = current; | |
c37b5efe | 277 | int err = 0; |
ab513701 SS |
278 | |
279 | if (!buf) { | |
c37b5efe SS |
280 | if (used_math()) |
281 | goto clear; | |
ab513701 SS |
282 | return 0; |
283 | } else | |
284 | if (!access_ok(VERIFY_READ, buf, sig_xstate_size)) | |
285 | return -EACCES; | |
286 | ||
287 | if (!used_math()) { | |
288 | err = init_fpu(tsk); | |
289 | if (err) | |
290 | return err; | |
291 | } | |
292 | ||
15d8791c | 293 | user_fpu_begin(); |
c9ad4882 | 294 | if (use_xsave()) |
c37b5efe SS |
295 | err = restore_user_xstate(buf); |
296 | else | |
297 | err = fxrstor_checking((__force struct i387_fxsave_struct *) | |
298 | buf); | |
ab513701 SS |
299 | if (unlikely(err)) { |
300 | /* | |
301 | * Encountered an error while doing the restore from the | |
302 | * user buffer, clear the fpu state. | |
303 | */ | |
c37b5efe | 304 | clear: |
ab513701 SS |
305 | clear_fpu(tsk); |
306 | clear_used_math(); | |
307 | } | |
308 | return err; | |
309 | } | |
310 | #endif | |
311 | ||
c37b5efe SS |
312 | /* |
313 | * Prepare the SW reserved portion of the fxsave memory layout, indicating | |
314 | * the presence of the extended state information in the memory layout | |
315 | * pointed by the fpstate pointer in the sigcontext. | |
316 | * This will be saved when ever the FP and extended state context is | |
317 | * saved on the user stack during the signal handler delivery to the user. | |
318 | */ | |
8bcad30f | 319 | static void prepare_fx_sw_frame(void) |
c37b5efe SS |
320 | { |
321 | int size_extended = (xstate_size - sizeof(struct i387_fxsave_struct)) + | |
322 | FP_XSTATE_MAGIC2_SIZE; | |
323 | ||
324 | sig_xstate_size = sizeof(struct _fpstate) + size_extended; | |
325 | ||
326 | #ifdef CONFIG_IA32_EMULATION | |
327 | sig_xstate_ia32_size = sizeof(struct _fpstate_ia32) + size_extended; | |
328 | #endif | |
329 | ||
330 | memset(&fx_sw_reserved, 0, sizeof(fx_sw_reserved)); | |
331 | ||
332 | fx_sw_reserved.magic1 = FP_XSTATE_MAGIC1; | |
333 | fx_sw_reserved.extended_size = sig_xstate_size; | |
6152e4b1 | 334 | fx_sw_reserved.xstate_bv = pcntxt_mask; |
c37b5efe SS |
335 | fx_sw_reserved.xstate_size = xstate_size; |
336 | #ifdef CONFIG_IA32_EMULATION | |
337 | memcpy(&fx_sw_reserved_ia32, &fx_sw_reserved, | |
338 | sizeof(struct _fpx_sw_bytes)); | |
339 | fx_sw_reserved_ia32.extended_size = sig_xstate_ia32_size; | |
340 | #endif | |
341 | } | |
342 | ||
3c1c7f10 SS |
343 | #ifdef CONFIG_X86_64 |
344 | unsigned int sig_xstate_size = sizeof(struct _fpstate); | |
345 | #endif | |
346 | ||
dc1e35c6 SS |
347 | /* |
348 | * Enable the extended processor state save/restore feature | |
349 | */ | |
1cff92d8 | 350 | static inline void xstate_enable(void) |
dc1e35c6 | 351 | { |
dc1e35c6 | 352 | set_in_cr4(X86_CR4_OSXSAVE); |
6152e4b1 | 353 | xsetbv(XCR_XFEATURE_ENABLED_MASK, pcntxt_mask); |
dc1e35c6 SS |
354 | } |
355 | ||
a1488f8b SS |
356 | /* |
357 | * Record the offsets and sizes of different state managed by the xsave | |
358 | * memory layout. | |
359 | */ | |
4995b9db | 360 | static void __init setup_xstate_features(void) |
a1488f8b SS |
361 | { |
362 | int eax, ebx, ecx, edx, leaf = 0x2; | |
363 | ||
364 | xstate_features = fls64(pcntxt_mask); | |
365 | xstate_offsets = alloc_bootmem(xstate_features * sizeof(int)); | |
366 | xstate_sizes = alloc_bootmem(xstate_features * sizeof(int)); | |
367 | ||
368 | do { | |
ee813d53 | 369 | cpuid_count(XSTATE_CPUID, leaf, &eax, &ebx, &ecx, &edx); |
a1488f8b SS |
370 | |
371 | if (eax == 0) | |
372 | break; | |
373 | ||
374 | xstate_offsets[leaf] = ebx; | |
375 | xstate_sizes[leaf] = eax; | |
376 | ||
377 | leaf++; | |
378 | } while (1); | |
379 | } | |
380 | ||
dc1e35c6 SS |
381 | /* |
382 | * setup the xstate image representing the init state | |
383 | */ | |
a19aac85 | 384 | static void __init setup_xstate_init(void) |
dc1e35c6 | 385 | { |
29104e10 SS |
386 | setup_xstate_features(); |
387 | ||
388 | /* | |
389 | * Setup init_xstate_buf to represent the init state of | |
390 | * all the features managed by the xsave | |
391 | */ | |
10340ae1 SS |
392 | init_xstate_buf = alloc_bootmem_align(xstate_size, |
393 | __alignof__(struct xsave_struct)); | |
dc1e35c6 | 394 | init_xstate_buf->i387.mxcsr = MXCSR_DEFAULT; |
a1488f8b | 395 | |
29104e10 SS |
396 | clts(); |
397 | /* | |
398 | * Init all the features state with header_bv being 0x0 | |
399 | */ | |
400 | xrstor_state(init_xstate_buf, -1); | |
401 | /* | |
402 | * Dump the init state again. This is to identify the init state | |
403 | * of any feature which is not represented by all zero's. | |
404 | */ | |
405 | xsave_state(init_xstate_buf, -1); | |
406 | stts(); | |
dc1e35c6 SS |
407 | } |
408 | ||
409 | /* | |
410 | * Enable and initialize the xsave feature. | |
411 | */ | |
1cff92d8 | 412 | static void __init xstate_enable_boot_cpu(void) |
dc1e35c6 SS |
413 | { |
414 | unsigned int eax, ebx, ecx, edx; | |
415 | ||
ee813d53 RR |
416 | if (boot_cpu_data.cpuid_level < XSTATE_CPUID) { |
417 | WARN(1, KERN_ERR "XSTATE_CPUID missing\n"); | |
418 | return; | |
419 | } | |
420 | ||
421 | cpuid_count(XSTATE_CPUID, 0, &eax, &ebx, &ecx, &edx); | |
6152e4b1 | 422 | pcntxt_mask = eax + ((u64)edx << 32); |
dc1e35c6 | 423 | |
6152e4b1 PA |
424 | if ((pcntxt_mask & XSTATE_FPSSE) != XSTATE_FPSSE) { |
425 | printk(KERN_ERR "FP/SSE not shown under xsave features 0x%llx\n", | |
426 | pcntxt_mask); | |
dc1e35c6 SS |
427 | BUG(); |
428 | } | |
429 | ||
430 | /* | |
a30469e7 | 431 | * Support only the state known to OS. |
dc1e35c6 | 432 | */ |
6152e4b1 | 433 | pcntxt_mask = pcntxt_mask & XCNTXT_MASK; |
97e80a70 | 434 | |
1cff92d8 | 435 | xstate_enable(); |
dc1e35c6 SS |
436 | |
437 | /* | |
438 | * Recompute the context size for enabled features | |
439 | */ | |
ee813d53 | 440 | cpuid_count(XSTATE_CPUID, 0, &eax, &ebx, &ecx, &edx); |
dc1e35c6 SS |
441 | xstate_size = ebx; |
442 | ||
5b3efd50 | 443 | update_regset_xstate_info(xstate_size, pcntxt_mask); |
c37b5efe SS |
444 | prepare_fx_sw_frame(); |
445 | ||
dc1e35c6 SS |
446 | setup_xstate_init(); |
447 | ||
6152e4b1 | 448 | printk(KERN_INFO "xsave/xrstor: enabled xstate_bv 0x%llx, " |
dc1e35c6 | 449 | "cntxt size 0x%x\n", |
6152e4b1 | 450 | pcntxt_mask, xstate_size); |
dc1e35c6 | 451 | } |
82d4150c | 452 | |
1cff92d8 PA |
453 | /* |
454 | * For the very first instance, this calls xstate_enable_boot_cpu(); | |
455 | * for all subsequent instances, this calls xstate_enable(). | |
456 | * | |
457 | * This is somewhat obfuscated due to the lack of powerful enough | |
458 | * overrides for the section checks. | |
459 | */ | |
82d4150c RR |
460 | void __cpuinit xsave_init(void) |
461 | { | |
1cff92d8 PA |
462 | static __refdata void (*next_func)(void) = xstate_enable_boot_cpu; |
463 | void (*this_func)(void); | |
464 | ||
0e49bf66 RR |
465 | if (!cpu_has_xsave) |
466 | return; | |
467 | ||
1cff92d8 PA |
468 | this_func = next_func; |
469 | next_func = xstate_enable; | |
470 | this_func(); | |
82d4150c | 471 | } |