1 | // SPDX-License-Identifier: GPL-2.0 |
2 | /* |
3 | * FPU signal frame handling routines. |
4 | */ |
5 | |
6 | #include <linux/compat.h> |
7 | #include <linux/cpu.h> |
8 | #include <linux/pagemap.h> |
9 | |
10 | #include <asm/fpu/signal.h> |
11 | #include <asm/fpu/regset.h> |
12 | #include <asm/fpu/xstate.h> |
13 | |
14 | #include <asm/sigframe.h> |
15 | #include <asm/trapnr.h> |
16 | #include <asm/trace/fpu.h> |
17 | |
18 | #include "context.h" |
19 | #include "internal.h" |
20 | #include "legacy.h" |
21 | #include "xstate.h" |
22 | |
23 | /* |
24 | * Check for the presence of extended state information in the |
25 | * user fpstate pointer in the sigcontext. |
26 | */ |
27 | static inline bool check_xstate_in_sigframe(struct fxregs_state __user *fxbuf, |
28 | struct _fpx_sw_bytes *fx_sw) |
29 | { |
30 | int min_xstate_size = sizeof(struct fxregs_state) + |
31 | sizeof(struct xstate_header); |
32 | void __user *fpstate = fxbuf; |
33 | unsigned int magic2; |
34 | |
35 | if (__copy_from_user(to: fx_sw, from: &fxbuf->sw_reserved[0], n: sizeof(*fx_sw))) |
36 | return false; |
37 | |
38 | /* Check for the first magic field and other error scenarios. */ |
39 | if (fx_sw->magic1 != FP_XSTATE_MAGIC1 || |
40 | fx_sw->xstate_size < min_xstate_size || |
41 | fx_sw->xstate_size > current->thread.fpu.fpstate->user_size || |
42 | fx_sw->xstate_size > fx_sw->extended_size) |
43 | goto setfx; |
44 | |
45 | /* |
46 | * Check for the presence of second magic word at the end of memory |
47 | * layout. This detects the case where the user just copied the legacy |
48 | * fpstate layout with out copying the extended state information |
49 | * in the memory layout. |
50 | */ |
51 | if (__get_user(magic2, (__u32 __user *)(fpstate + fx_sw->xstate_size))) |
52 | return false; |
53 | |
54 | if (likely(magic2 == FP_XSTATE_MAGIC2)) |
55 | return true; |
56 | setfx: |
57 | trace_x86_fpu_xstate_check_failed(fpu: ¤t->thread.fpu); |
58 | |
59 | /* Set the parameters for fx only state */ |
60 | fx_sw->magic1 = 0; |
61 | fx_sw->xstate_size = sizeof(struct fxregs_state); |
62 | fx_sw->xfeatures = XFEATURE_MASK_FPSSE; |
63 | return true; |
64 | } |
65 | |
66 | /* |
67 | * Signal frame handlers. |
68 | */ |
69 | static inline bool (struct task_struct *tsk, void __user *buf) |
70 | { |
71 | if (use_fxsr()) { |
72 | struct xregs_state *xsave = &tsk->thread.fpu.fpstate->regs.xsave; |
73 | struct user_i387_ia32_struct env; |
74 | struct _fpstate_32 __user *fp = buf; |
75 | |
76 | fpregs_lock(); |
77 | if (!test_thread_flag(TIF_NEED_FPU_LOAD)) |
78 | fxsave(fx: &tsk->thread.fpu.fpstate->regs.fxsave); |
79 | fpregs_unlock(); |
80 | |
81 | convert_from_fxsr(env: &env, tsk); |
82 | |
83 | if (__copy_to_user(to: buf, from: &env, n: sizeof(env)) || |
84 | __put_user(xsave->i387.swd, &fp->status) || |
85 | __put_user(X86_FXSR_MAGIC, &fp->magic)) |
86 | return false; |
87 | } else { |
88 | struct fregs_state __user *fp = buf; |
89 | u32 swd; |
90 | |
91 | if (__get_user(swd, &fp->swd) || __put_user(swd, &fp->status)) |
92 | return false; |
93 | } |
94 | |
95 | return true; |
96 | } |
97 | |
98 | /* |
99 | * Prepare the SW reserved portion of the fxsave memory layout, indicating |
100 | * the presence of the extended state information in the memory layout |
101 | * pointed to by the fpstate pointer in the sigcontext. |
102 | * This is saved when ever the FP and extended state context is |
103 | * saved on the user stack during the signal handler delivery to the user. |
104 | */ |
105 | static inline void save_sw_bytes(struct _fpx_sw_bytes *sw_bytes, bool ia32_frame, |
106 | struct fpstate *fpstate) |
107 | { |
108 | sw_bytes->magic1 = FP_XSTATE_MAGIC1; |
109 | sw_bytes->extended_size = fpstate->user_size + FP_XSTATE_MAGIC2_SIZE; |
110 | sw_bytes->xfeatures = fpstate->user_xfeatures; |
111 | sw_bytes->xstate_size = fpstate->user_size; |
112 | |
113 | if (ia32_frame) |
114 | sw_bytes->extended_size += sizeof(struct fregs_state); |
115 | } |
116 | |
117 | static inline bool save_xstate_epilog(void __user *buf, int ia32_frame, |
118 | struct fpstate *fpstate) |
119 | { |
120 | struct xregs_state __user *x = buf; |
121 | struct _fpx_sw_bytes sw_bytes = {}; |
122 | u32 xfeatures; |
123 | int err; |
124 | |
125 | /* Setup the bytes not touched by the [f]xsave and reserved for SW. */ |
126 | save_sw_bytes(sw_bytes: &sw_bytes, ia32_frame, fpstate); |
127 | err = __copy_to_user(to: &x->i387.sw_reserved, from: &sw_bytes, n: sizeof(sw_bytes)); |
128 | |
129 | if (!use_xsave()) |
130 | return !err; |
131 | |
132 | err |= __put_user(FP_XSTATE_MAGIC2, |
133 | (__u32 __user *)(buf + fpstate->user_size)); |
134 | |
135 | /* |
136 | * Read the xfeatures which we copied (directly from the cpu or |
137 | * from the state in task struct) to the user buffers. |
138 | */ |
139 | err |= __get_user(xfeatures, (__u32 __user *)&x->header.xfeatures); |
140 | |
141 | /* |
142 | * For legacy compatible, we always set FP/SSE bits in the bit |
143 | * vector while saving the state to the user context. This will |
144 | * enable us capturing any changes(during sigreturn) to |
145 | * the FP/SSE bits by the legacy applications which don't touch |
146 | * xfeatures in the xsave header. |
147 | * |
148 | * xsave aware apps can change the xfeatures in the xsave |
149 | * header as well as change any contents in the memory layout. |
150 | * xrestore as part of sigreturn will capture all the changes. |
151 | */ |
152 | xfeatures |= XFEATURE_MASK_FPSSE; |
153 | |
154 | err |= __put_user(xfeatures, (__u32 __user *)&x->header.xfeatures); |
155 | |
156 | return !err; |
157 | } |
158 | |
159 | static inline int copy_fpregs_to_sigframe(struct xregs_state __user *buf) |
160 | { |
161 | if (use_xsave()) |
162 | return xsave_to_user_sigframe(buf); |
163 | if (use_fxsr()) |
164 | return fxsave_to_user_sigframe(fx: (struct fxregs_state __user *) buf); |
165 | else |
166 | return fnsave_to_user_sigframe(fx: (struct fregs_state __user *) buf); |
167 | } |
168 | |
169 | /* |
170 | * Save the fpu, extended register state to the user signal frame. |
171 | * |
172 | * 'buf_fx' is the 64-byte aligned pointer at which the [f|fx|x]save |
173 | * state is copied. |
174 | * 'buf' points to the 'buf_fx' or to the fsave header followed by 'buf_fx'. |
175 | * |
176 | * buf == buf_fx for 64-bit frames and 32-bit fsave frame. |
177 | * buf != buf_fx for 32-bit frames with fxstate. |
178 | * |
179 | * Save it directly to the user frame with disabled page fault handler. If |
180 | * that faults, try to clear the frame which handles the page fault. |
181 | * |
182 | * If this is a 32-bit frame with fxstate, put a fsave header before |
183 | * the aligned state at 'buf_fx'. |
184 | * |
185 | * For [f]xsave state, update the SW reserved fields in the [f]xsave frame |
186 | * indicating the absence/presence of the extended state to the user. |
187 | */ |
188 | bool copy_fpstate_to_sigframe(void __user *buf, void __user *buf_fx, int size) |
189 | { |
190 | struct task_struct *tsk = current; |
191 | struct fpstate *fpstate = tsk->thread.fpu.fpstate; |
192 | bool ia32_fxstate = (buf != buf_fx); |
193 | int ret; |
194 | |
195 | ia32_fxstate &= (IS_ENABLED(CONFIG_X86_32) || |
196 | IS_ENABLED(CONFIG_IA32_EMULATION)); |
197 | |
198 | if (!static_cpu_has(X86_FEATURE_FPU)) { |
199 | struct user_i387_ia32_struct fp; |
200 | |
201 | fpregs_soft_get(current, NULL, (struct membuf){.p = &fp, |
202 | .left = sizeof(fp)}); |
203 | return !copy_to_user(to: buf, from: &fp, n: sizeof(fp)); |
204 | } |
205 | |
206 | if (!access_ok(buf, size)) |
207 | return false; |
208 | |
209 | if (use_xsave()) { |
210 | struct xregs_state __user *xbuf = buf_fx; |
211 | |
212 | /* |
213 | * Clear the xsave header first, so that reserved fields are |
214 | * initialized to zero. |
215 | */ |
216 | if (__clear_user(addr: &xbuf->header, size: sizeof(xbuf->header))) |
217 | return false; |
218 | } |
219 | retry: |
220 | /* |
221 | * Load the FPU registers if they are not valid for the current task. |
222 | * With a valid FPU state we can attempt to save the state directly to |
223 | * userland's stack frame which will likely succeed. If it does not, |
224 | * resolve the fault in the user memory and try again. |
225 | */ |
226 | fpregs_lock(); |
227 | if (test_thread_flag(TIF_NEED_FPU_LOAD)) |
228 | fpregs_restore_userregs(); |
229 | |
230 | pagefault_disable(); |
231 | ret = copy_fpregs_to_sigframe(buf: buf_fx); |
232 | pagefault_enable(); |
233 | fpregs_unlock(); |
234 | |
235 | if (ret) { |
236 | if (!__clear_user(addr: buf_fx, size: fpstate->user_size)) |
237 | goto retry; |
238 | return false; |
239 | } |
240 | |
241 | /* Save the fsave header for the 32-bit frames. */ |
242 | if ((ia32_fxstate || !use_fxsr()) && !save_fsave_header(tsk, buf)) |
243 | return false; |
244 | |
245 | if (use_fxsr() && !save_xstate_epilog(buf: buf_fx, ia32_frame: ia32_fxstate, fpstate)) |
246 | return false; |
247 | |
248 | return true; |
249 | } |
250 | |
251 | static int __restore_fpregs_from_user(void __user *buf, u64 ufeatures, |
252 | u64 xrestore, bool fx_only) |
253 | { |
254 | if (use_xsave()) { |
255 | u64 init_bv = ufeatures & ~xrestore; |
256 | int ret; |
257 | |
258 | if (likely(!fx_only)) |
259 | ret = xrstor_from_user_sigframe(buf, mask: xrestore); |
260 | else |
261 | ret = fxrstor_from_user_sigframe(fx: buf); |
262 | |
263 | if (!ret && unlikely(init_bv)) |
264 | os_xrstor(fpstate: &init_fpstate, mask: init_bv); |
265 | return ret; |
266 | } else if (use_fxsr()) { |
267 | return fxrstor_from_user_sigframe(fx: buf); |
268 | } else { |
269 | return frstor_from_user_sigframe(fx: buf); |
270 | } |
271 | } |
272 | |
273 | /* |
274 | * Attempt to restore the FPU registers directly from user memory. |
275 | * Pagefaults are handled and any errors returned are fatal. |
276 | */ |
277 | static bool restore_fpregs_from_user(void __user *buf, u64 xrestore, bool fx_only) |
278 | { |
279 | struct fpu *fpu = ¤t->thread.fpu; |
280 | int ret; |
281 | |
282 | /* Restore enabled features only. */ |
283 | xrestore &= fpu->fpstate->user_xfeatures; |
284 | retry: |
285 | fpregs_lock(); |
286 | /* Ensure that XFD is up to date */ |
287 | xfd_update_state(fpstate: fpu->fpstate); |
288 | pagefault_disable(); |
289 | ret = __restore_fpregs_from_user(buf, ufeatures: fpu->fpstate->user_xfeatures, |
290 | xrestore, fx_only); |
291 | pagefault_enable(); |
292 | |
293 | if (unlikely(ret)) { |
294 | /* |
295 | * The above did an FPU restore operation, restricted to |
296 | * the user portion of the registers, and failed, but the |
297 | * microcode might have modified the FPU registers |
298 | * nevertheless. |
299 | * |
300 | * If the FPU registers do not belong to current, then |
301 | * invalidate the FPU register state otherwise the task |
302 | * might preempt current and return to user space with |
303 | * corrupted FPU registers. |
304 | */ |
305 | if (test_thread_flag(TIF_NEED_FPU_LOAD)) |
306 | __cpu_invalidate_fpregs_state(); |
307 | fpregs_unlock(); |
308 | |
309 | /* Try to handle #PF, but anything else is fatal. */ |
310 | if (ret != X86_TRAP_PF) |
311 | return false; |
312 | |
313 | if (!fault_in_readable(uaddr: buf, size: fpu->fpstate->user_size)) |
314 | goto retry; |
315 | return false; |
316 | } |
317 | |
318 | /* |
319 | * Restore supervisor states: previous context switch etc has done |
320 | * XSAVES and saved the supervisor states in the kernel buffer from |
321 | * which they can be restored now. |
322 | * |
323 | * It would be optimal to handle this with a single XRSTORS, but |
324 | * this does not work because the rest of the FPU registers have |
325 | * been restored from a user buffer directly. |
326 | */ |
327 | if (test_thread_flag(TIF_NEED_FPU_LOAD) && xfeatures_mask_supervisor()) |
328 | os_xrstor_supervisor(fpstate: fpu->fpstate); |
329 | |
330 | fpregs_mark_activate(); |
331 | fpregs_unlock(); |
332 | return true; |
333 | } |
334 | |
335 | static bool __fpu_restore_sig(void __user *buf, void __user *buf_fx, |
336 | bool ia32_fxstate) |
337 | { |
338 | struct task_struct *tsk = current; |
339 | struct fpu *fpu = &tsk->thread.fpu; |
340 | struct user_i387_ia32_struct env; |
341 | bool success, fx_only = false; |
342 | union fpregs_state *fpregs; |
343 | u64 user_xfeatures = 0; |
344 | |
345 | if (use_xsave()) { |
346 | struct _fpx_sw_bytes fx_sw_user; |
347 | |
348 | if (!check_xstate_in_sigframe(fxbuf: buf_fx, fx_sw: &fx_sw_user)) |
349 | return false; |
350 | |
351 | fx_only = !fx_sw_user.magic1; |
352 | user_xfeatures = fx_sw_user.xfeatures; |
353 | } else { |
354 | user_xfeatures = XFEATURE_MASK_FPSSE; |
355 | } |
356 | |
357 | if (likely(!ia32_fxstate)) { |
358 | /* Restore the FPU registers directly from user memory. */ |
359 | return restore_fpregs_from_user(buf: buf_fx, xrestore: user_xfeatures, fx_only); |
360 | } |
361 | |
362 | /* |
363 | * Copy the legacy state because the FP portion of the FX frame has |
364 | * to be ignored for histerical raisins. The legacy state is folded |
365 | * in once the larger state has been copied. |
366 | */ |
367 | if (__copy_from_user(to: &env, from: buf, n: sizeof(env))) |
368 | return false; |
369 | |
370 | /* |
371 | * By setting TIF_NEED_FPU_LOAD it is ensured that our xstate is |
372 | * not modified on context switch and that the xstate is considered |
373 | * to be loaded again on return to userland (overriding last_cpu avoids |
374 | * the optimisation). |
375 | */ |
376 | fpregs_lock(); |
377 | if (!test_thread_flag(TIF_NEED_FPU_LOAD)) { |
378 | /* |
379 | * If supervisor states are available then save the |
380 | * hardware state in current's fpstate so that the |
381 | * supervisor state is preserved. Save the full state for |
382 | * simplicity. There is no point in optimizing this by only |
383 | * saving the supervisor states and then shuffle them to |
384 | * the right place in memory. It's ia32 mode. Shrug. |
385 | */ |
386 | if (xfeatures_mask_supervisor()) |
387 | os_xsave(fpstate: fpu->fpstate); |
388 | set_thread_flag(TIF_NEED_FPU_LOAD); |
389 | } |
390 | __fpu_invalidate_fpregs_state(fpu); |
391 | __cpu_invalidate_fpregs_state(); |
392 | fpregs_unlock(); |
393 | |
394 | fpregs = &fpu->fpstate->regs; |
395 | if (use_xsave() && !fx_only) { |
396 | if (copy_sigframe_from_user_to_xstate(tsk, ubuf: buf_fx)) |
397 | return false; |
398 | } else { |
399 | if (__copy_from_user(to: &fpregs->fxsave, from: buf_fx, |
400 | n: sizeof(fpregs->fxsave))) |
401 | return false; |
402 | |
403 | if (IS_ENABLED(CONFIG_X86_64)) { |
404 | /* Reject invalid MXCSR values. */ |
405 | if (fpregs->fxsave.mxcsr & ~mxcsr_feature_mask) |
406 | return false; |
407 | } else { |
408 | /* Mask invalid bits out for historical reasons (broken hardware). */ |
409 | fpregs->fxsave.mxcsr &= mxcsr_feature_mask; |
410 | } |
411 | |
412 | /* Enforce XFEATURE_MASK_FPSSE when XSAVE is enabled */ |
413 | if (use_xsave()) |
414 | fpregs->xsave.header.xfeatures |= XFEATURE_MASK_FPSSE; |
415 | } |
416 | |
417 | /* Fold the legacy FP storage */ |
418 | convert_to_fxsr(fxsave: &fpregs->fxsave, env: &env); |
419 | |
420 | fpregs_lock(); |
421 | if (use_xsave()) { |
422 | /* |
423 | * Remove all UABI feature bits not set in user_xfeatures |
424 | * from the memory xstate header which makes the full |
425 | * restore below bring them into init state. This works for |
426 | * fx_only mode as well because that has only FP and SSE |
427 | * set in user_xfeatures. |
428 | * |
429 | * Preserve supervisor states! |
430 | */ |
431 | u64 mask = user_xfeatures | xfeatures_mask_supervisor(); |
432 | |
433 | fpregs->xsave.header.xfeatures &= mask; |
434 | success = !os_xrstor_safe(fpstate: fpu->fpstate, |
435 | mask: fpu_kernel_cfg.max_features); |
436 | } else { |
437 | success = !fxrstor_safe(fx: &fpregs->fxsave); |
438 | } |
439 | |
440 | if (likely(success)) |
441 | fpregs_mark_activate(); |
442 | |
443 | fpregs_unlock(); |
444 | return success; |
445 | } |
446 | |
447 | static inline unsigned int xstate_sigframe_size(struct fpstate *fpstate) |
448 | { |
449 | unsigned int size = fpstate->user_size; |
450 | |
451 | return use_xsave() ? size + FP_XSTATE_MAGIC2_SIZE : size; |
452 | } |
453 | |
454 | /* |
455 | * Restore FPU state from a sigframe: |
456 | */ |
457 | bool fpu__restore_sig(void __user *buf, int ia32_frame) |
458 | { |
459 | struct fpu *fpu = ¤t->thread.fpu; |
460 | void __user *buf_fx = buf; |
461 | bool ia32_fxstate = false; |
462 | bool success = false; |
463 | unsigned int size; |
464 | |
465 | if (unlikely(!buf)) { |
466 | fpu__clear_user_states(fpu); |
467 | return true; |
468 | } |
469 | |
470 | size = xstate_sigframe_size(fpstate: fpu->fpstate); |
471 | |
472 | ia32_frame &= (IS_ENABLED(CONFIG_X86_32) || |
473 | IS_ENABLED(CONFIG_IA32_EMULATION)); |
474 | |
475 | /* |
476 | * Only FXSR enabled systems need the FX state quirk. |
477 | * FRSTOR does not need it and can use the fast path. |
478 | */ |
479 | if (ia32_frame && use_fxsr()) { |
480 | buf_fx = buf + sizeof(struct fregs_state); |
481 | size += sizeof(struct fregs_state); |
482 | ia32_fxstate = true; |
483 | } |
484 | |
485 | if (!access_ok(buf, size)) |
486 | goto out; |
487 | |
488 | if (!IS_ENABLED(CONFIG_X86_64) && !cpu_feature_enabled(X86_FEATURE_FPU)) { |
489 | success = !fpregs_soft_set(current, NULL, 0, |
490 | sizeof(struct user_i387_ia32_struct), |
491 | NULL, buf); |
492 | } else { |
493 | success = __fpu_restore_sig(buf, buf_fx, ia32_fxstate); |
494 | } |
495 | |
496 | out: |
497 | if (unlikely(!success)) |
498 | fpu__clear_user_states(fpu); |
499 | return success; |
500 | } |
501 | |
502 | unsigned long |
503 | fpu__alloc_mathframe(unsigned long sp, int ia32_frame, |
504 | unsigned long *buf_fx, unsigned long *size) |
505 | { |
506 | unsigned long frame_size = xstate_sigframe_size(current->thread.fpu.fpstate); |
507 | |
508 | *buf_fx = sp = round_down(sp - frame_size, 64); |
509 | if (ia32_frame && use_fxsr()) { |
510 | frame_size += sizeof(struct fregs_state); |
511 | sp -= sizeof(struct fregs_state); |
512 | } |
513 | |
514 | *size = frame_size; |
515 | |
516 | return sp; |
517 | } |
518 | |
519 | unsigned long __init fpu__get_fpstate_size(void) |
520 | { |
521 | unsigned long ret = fpu_user_cfg.max_size; |
522 | |
523 | if (use_xsave()) |
524 | ret += FP_XSTATE_MAGIC2_SIZE; |
525 | |
526 | /* |
527 | * This space is needed on (most) 32-bit kernels, or when a 32-bit |
528 | * app is running on a 64-bit kernel. To keep things simple, just |
529 | * assume the worst case and always include space for 'freg_state', |
530 | * even for 64-bit apps on 64-bit kernels. This wastes a bit of |
531 | * space, but keeps the code simple. |
532 | */ |
533 | if ((IS_ENABLED(CONFIG_IA32_EMULATION) || |
534 | IS_ENABLED(CONFIG_X86_32)) && use_fxsr()) |
535 | ret += sizeof(struct fregs_state); |
536 | |
537 | return ret; |
538 | } |
539 | |
540 | |