xref: /linux/arch/x86/kernel/fpu/xstate.c (revision 7a309195d11cde854eb75559fbd6b48f9e518f25)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * xsave/xrstor support.
4  *
5  * Author: Suresh Siddha <suresh.b.siddha@intel.com>
6  */
7 #include <linux/compat.h>
8 #include <linux/cpu.h>
9 #include <linux/mman.h>
10 #include <linux/pkeys.h>
11 #include <linux/seq_file.h>
12 #include <linux/proc_fs.h>
13 
14 #include <asm/fpu/api.h>
15 #include <asm/fpu/internal.h>
16 #include <asm/fpu/signal.h>
17 #include <asm/fpu/regset.h>
18 #include <asm/fpu/xstate.h>
19 
20 #include <asm/tlbflush.h>
21 #include <asm/cpufeature.h>
22 
23 /*
24  * Although we spell it out in here, the Processor Trace
25  * xfeature is completely unused.  We use other mechanisms
26  * to save/restore PT state in Linux.
27  */
28 static const char *xfeature_names[] =
29 {
30 	"x87 floating point registers"	,
31 	"SSE registers"			,
32 	"AVX registers"			,
33 	"MPX bounds registers"		,
34 	"MPX CSR"			,
35 	"AVX-512 opmask"		,
36 	"AVX-512 Hi256"			,
37 	"AVX-512 ZMM_Hi256"		,
38 	"Processor Trace (unused)"	,
39 	"Protection Keys User registers",
40 	"unknown xstate feature"	,
41 };
42 
43 static short xsave_cpuid_features[] __initdata = {
44 	X86_FEATURE_FPU,
45 	X86_FEATURE_XMM,
46 	X86_FEATURE_AVX,
47 	X86_FEATURE_MPX,
48 	X86_FEATURE_MPX,
49 	X86_FEATURE_AVX512F,
50 	X86_FEATURE_AVX512F,
51 	X86_FEATURE_AVX512F,
52 	X86_FEATURE_INTEL_PT,
53 	X86_FEATURE_PKU,
54 };
55 
56 /*
57  * This represents the full set of bits that should ever be set in a kernel
58  * XSAVE buffer, both supervisor and user xstates.
59  */
60 u64 xfeatures_mask_all __read_mostly;
61 
62 static unsigned int xstate_offsets[XFEATURE_MAX] = { [ 0 ... XFEATURE_MAX - 1] = -1};
63 static unsigned int xstate_sizes[XFEATURE_MAX]   = { [ 0 ... XFEATURE_MAX - 1] = -1};
64 static unsigned int xstate_comp_offsets[XFEATURE_MAX] = { [ 0 ... XFEATURE_MAX - 1] = -1};
65 static unsigned int xstate_supervisor_only_offsets[XFEATURE_MAX] = { [ 0 ... XFEATURE_MAX - 1] = -1};
66 
67 /*
68  * The XSAVE area of kernel can be in standard or compacted format;
69  * it is always in standard format for user mode. This is the user
70  * mode standard format size used for signal and ptrace frames.
71  */
72 unsigned int fpu_user_xstate_size;
73 
74 /*
75  * Return whether the system supports a given xfeature.
76  *
77  * Also return the name of the (most advanced) feature that the caller requested:
78  */
79 int cpu_has_xfeatures(u64 xfeatures_needed, const char **feature_name)
80 {
81 	u64 xfeatures_missing = xfeatures_needed & ~xfeatures_mask_all;
82 
83 	if (unlikely(feature_name)) {
84 		long xfeature_idx, max_idx;
85 		u64 xfeatures_print;
86 		/*
87 		 * So we use FLS here to be able to print the most advanced
88 		 * feature that was requested but is missing. So if a driver
89 		 * asks about "XFEATURE_MASK_SSE | XFEATURE_MASK_YMM" we'll print the
90 		 * missing AVX feature - this is the most informative message
91 		 * to users:
92 		 */
93 		if (xfeatures_missing)
94 			xfeatures_print = xfeatures_missing;
95 		else
96 			xfeatures_print = xfeatures_needed;
97 
98 		xfeature_idx = fls64(xfeatures_print)-1;
99 		max_idx = ARRAY_SIZE(xfeature_names)-1;
100 		xfeature_idx = min(xfeature_idx, max_idx);
101 
102 		*feature_name = xfeature_names[xfeature_idx];
103 	}
104 
105 	if (xfeatures_missing)
106 		return 0;
107 
108 	return 1;
109 }
110 EXPORT_SYMBOL_GPL(cpu_has_xfeatures);
111 
112 static bool xfeature_is_supervisor(int xfeature_nr)
113 {
114 	/*
115 	 * Extended State Enumeration Sub-leaves (EAX = 0DH, ECX = n, n > 1)
116 	 * returns ECX[0] set to (1) for a supervisor state, and cleared (0)
117 	 * for a user state.
118 	 */
119 	u32 eax, ebx, ecx, edx;
120 
121 	cpuid_count(XSTATE_CPUID, xfeature_nr, &eax, &ebx, &ecx, &edx);
122 	return ecx & 1;
123 }
124 
125 /*
126  * When executing XSAVEOPT (or other optimized XSAVE instructions), if
127  * a processor implementation detects that an FPU state component is still
128  * (or is again) in its initialized state, it may clear the corresponding
129  * bit in the header.xfeatures field, and can skip the writeout of registers
130  * to the corresponding memory layout.
131  *
132  * This means that when the bit is zero, the state component might still contain
133  * some previous - non-initialized register state.
134  *
135  * Before writing xstate information to user-space we sanitize those components,
136  * to always ensure that the memory layout of a feature will be in the init state
137  * if the corresponding header bit is zero. This is to ensure that user-space doesn't
138  * see some stale state in the memory layout during signal handling, debugging etc.
139  */
140 void fpstate_sanitize_xstate(struct fpu *fpu)
141 {
142 	struct fxregs_state *fx = &fpu->state.fxsave;
143 	int feature_bit;
144 	u64 xfeatures;
145 
146 	if (!use_xsaveopt())
147 		return;
148 
149 	xfeatures = fpu->state.xsave.header.xfeatures;
150 
151 	/*
152 	 * None of the feature bits are in init state. So nothing else
153 	 * to do for us, as the memory layout is up to date.
154 	 */
155 	if ((xfeatures & xfeatures_mask_all) == xfeatures_mask_all)
156 		return;
157 
158 	/*
159 	 * FP is in init state
160 	 */
161 	if (!(xfeatures & XFEATURE_MASK_FP)) {
162 		fx->cwd = 0x37f;
163 		fx->swd = 0;
164 		fx->twd = 0;
165 		fx->fop = 0;
166 		fx->rip = 0;
167 		fx->rdp = 0;
168 		memset(&fx->st_space[0], 0, 128);
169 	}
170 
171 	/*
172 	 * SSE is in init state
173 	 */
174 	if (!(xfeatures & XFEATURE_MASK_SSE))
175 		memset(&fx->xmm_space[0], 0, 256);
176 
177 	/*
178 	 * First two features are FPU and SSE, which above we handled
179 	 * in a special way already:
180 	 */
181 	feature_bit = 0x2;
182 	xfeatures = (xfeatures_mask_user() & ~xfeatures) >> 2;
183 
184 	/*
185 	 * Update all the remaining memory layouts according to their
186 	 * standard xstate layout, if their header bit is in the init
187 	 * state:
188 	 */
189 	while (xfeatures) {
190 		if (xfeatures & 0x1) {
191 			int offset = xstate_comp_offsets[feature_bit];
192 			int size = xstate_sizes[feature_bit];
193 
194 			memcpy((void *)fx + offset,
195 			       (void *)&init_fpstate.xsave + offset,
196 			       size);
197 		}
198 
199 		xfeatures >>= 1;
200 		feature_bit++;
201 	}
202 }
203 
204 /*
205  * Enable the extended processor state save/restore feature.
206  * Called once per CPU onlining.
207  */
208 void fpu__init_cpu_xstate(void)
209 {
210 	u64 unsup_bits;
211 
212 	if (!boot_cpu_has(X86_FEATURE_XSAVE) || !xfeatures_mask_all)
213 		return;
214 	/*
215 	 * Unsupported supervisor xstates should not be found in
216 	 * the xfeatures mask.
217 	 */
218 	unsup_bits = xfeatures_mask_all & XFEATURE_MASK_SUPERVISOR_UNSUPPORTED;
219 	WARN_ONCE(unsup_bits, "x86/fpu: Found unsupported supervisor xstates: 0x%llx\n",
220 		  unsup_bits);
221 
222 	xfeatures_mask_all &= ~XFEATURE_MASK_SUPERVISOR_UNSUPPORTED;
223 
224 	cr4_set_bits(X86_CR4_OSXSAVE);
225 
226 	/*
227 	 * XCR_XFEATURE_ENABLED_MASK (aka. XCR0) sets user features
228 	 * managed by XSAVE{C, OPT, S} and XRSTOR{S}.  Only XSAVE user
229 	 * states can be set here.
230 	 */
231 	xsetbv(XCR_XFEATURE_ENABLED_MASK, xfeatures_mask_user());
232 
233 	/*
234 	 * MSR_IA32_XSS sets supervisor states managed by XSAVES.
235 	 */
236 	if (boot_cpu_has(X86_FEATURE_XSAVES)) {
237 		wrmsrl(MSR_IA32_XSS, xfeatures_mask_supervisor() |
238 				     xfeatures_mask_dynamic());
239 	}
240 }
241 
242 static bool xfeature_enabled(enum xfeature xfeature)
243 {
244 	return xfeatures_mask_all & BIT_ULL(xfeature);
245 }
246 
247 /*
248  * Record the offsets and sizes of various xstates contained
249  * in the XSAVE state memory layout.
250  */
251 static void __init setup_xstate_features(void)
252 {
253 	u32 eax, ebx, ecx, edx, i;
254 	/* start at the beginnning of the "extended state" */
255 	unsigned int last_good_offset = offsetof(struct xregs_state,
256 						 extended_state_area);
257 	/*
258 	 * The FP xstates and SSE xstates are legacy states. They are always
259 	 * in the fixed offsets in the xsave area in either compacted form
260 	 * or standard form.
261 	 */
262 	xstate_offsets[XFEATURE_FP]	= 0;
263 	xstate_sizes[XFEATURE_FP]	= offsetof(struct fxregs_state,
264 						   xmm_space);
265 
266 	xstate_offsets[XFEATURE_SSE]	= xstate_sizes[XFEATURE_FP];
267 	xstate_sizes[XFEATURE_SSE]	= sizeof_field(struct fxregs_state,
268 						       xmm_space);
269 
270 	for (i = FIRST_EXTENDED_XFEATURE; i < XFEATURE_MAX; i++) {
271 		if (!xfeature_enabled(i))
272 			continue;
273 
274 		cpuid_count(XSTATE_CPUID, i, &eax, &ebx, &ecx, &edx);
275 
276 		xstate_sizes[i] = eax;
277 
278 		/*
279 		 * If an xfeature is supervisor state, the offset in EBX is
280 		 * invalid, leave it to -1.
281 		 */
282 		if (xfeature_is_supervisor(i))
283 			continue;
284 
285 		xstate_offsets[i] = ebx;
286 
287 		/*
288 		 * In our xstate size checks, we assume that the highest-numbered
289 		 * xstate feature has the highest offset in the buffer.  Ensure
290 		 * it does.
291 		 */
292 		WARN_ONCE(last_good_offset > xstate_offsets[i],
293 			  "x86/fpu: misordered xstate at %d\n", last_good_offset);
294 
295 		last_good_offset = xstate_offsets[i];
296 	}
297 }
298 
299 static void __init print_xstate_feature(u64 xstate_mask)
300 {
301 	const char *feature_name;
302 
303 	if (cpu_has_xfeatures(xstate_mask, &feature_name))
304 		pr_info("x86/fpu: Supporting XSAVE feature 0x%03Lx: '%s'\n", xstate_mask, feature_name);
305 }
306 
307 /*
308  * Print out all the supported xstate features:
309  */
310 static void __init print_xstate_features(void)
311 {
312 	print_xstate_feature(XFEATURE_MASK_FP);
313 	print_xstate_feature(XFEATURE_MASK_SSE);
314 	print_xstate_feature(XFEATURE_MASK_YMM);
315 	print_xstate_feature(XFEATURE_MASK_BNDREGS);
316 	print_xstate_feature(XFEATURE_MASK_BNDCSR);
317 	print_xstate_feature(XFEATURE_MASK_OPMASK);
318 	print_xstate_feature(XFEATURE_MASK_ZMM_Hi256);
319 	print_xstate_feature(XFEATURE_MASK_Hi16_ZMM);
320 	print_xstate_feature(XFEATURE_MASK_PKRU);
321 }
322 
323 /*
324  * This check is important because it is easy to get XSTATE_*
325  * confused with XSTATE_BIT_*.
326  */
327 #define CHECK_XFEATURE(nr) do {		\
328 	WARN_ON(nr < FIRST_EXTENDED_XFEATURE);	\
329 	WARN_ON(nr >= XFEATURE_MAX);	\
330 } while (0)
331 
332 /*
333  * We could cache this like xstate_size[], but we only use
334  * it here, so it would be a waste of space.
335  */
336 static int xfeature_is_aligned(int xfeature_nr)
337 {
338 	u32 eax, ebx, ecx, edx;
339 
340 	CHECK_XFEATURE(xfeature_nr);
341 
342 	if (!xfeature_enabled(xfeature_nr)) {
343 		WARN_ONCE(1, "Checking alignment of disabled xfeature %d\n",
344 			  xfeature_nr);
345 		return 0;
346 	}
347 
348 	cpuid_count(XSTATE_CPUID, xfeature_nr, &eax, &ebx, &ecx, &edx);
349 	/*
350 	 * The value returned by ECX[1] indicates the alignment
351 	 * of state component 'i' when the compacted format
352 	 * of the extended region of an XSAVE area is used:
353 	 */
354 	return !!(ecx & 2);
355 }
356 
357 /*
358  * This function sets up offsets and sizes of all extended states in
359  * xsave area. This supports both standard format and compacted format
360  * of the xsave area.
361  */
362 static void __init setup_xstate_comp_offsets(void)
363 {
364 	unsigned int next_offset;
365 	int i;
366 
367 	/*
368 	 * The FP xstates and SSE xstates are legacy states. They are always
369 	 * in the fixed offsets in the xsave area in either compacted form
370 	 * or standard form.
371 	 */
372 	xstate_comp_offsets[XFEATURE_FP] = 0;
373 	xstate_comp_offsets[XFEATURE_SSE] = offsetof(struct fxregs_state,
374 						     xmm_space);
375 
376 	if (!boot_cpu_has(X86_FEATURE_XSAVES)) {
377 		for (i = FIRST_EXTENDED_XFEATURE; i < XFEATURE_MAX; i++) {
378 			if (xfeature_enabled(i))
379 				xstate_comp_offsets[i] = xstate_offsets[i];
380 		}
381 		return;
382 	}
383 
384 	next_offset = FXSAVE_SIZE + XSAVE_HDR_SIZE;
385 
386 	for (i = FIRST_EXTENDED_XFEATURE; i < XFEATURE_MAX; i++) {
387 		if (!xfeature_enabled(i))
388 			continue;
389 
390 		if (xfeature_is_aligned(i))
391 			next_offset = ALIGN(next_offset, 64);
392 
393 		xstate_comp_offsets[i] = next_offset;
394 		next_offset += xstate_sizes[i];
395 	}
396 }
397 
398 /*
399  * Setup offsets of a supervisor-state-only XSAVES buffer:
400  *
401  * The offsets stored in xstate_comp_offsets[] only work for one specific
402  * value of the Requested Feature BitMap (RFBM).  In cases where a different
403  * RFBM value is used, a different set of offsets is required.  This set of
404  * offsets is for when RFBM=xfeatures_mask_supervisor().
405  */
406 static void __init setup_supervisor_only_offsets(void)
407 {
408 	unsigned int next_offset;
409 	int i;
410 
411 	next_offset = FXSAVE_SIZE + XSAVE_HDR_SIZE;
412 
413 	for (i = FIRST_EXTENDED_XFEATURE; i < XFEATURE_MAX; i++) {
414 		if (!xfeature_enabled(i) || !xfeature_is_supervisor(i))
415 			continue;
416 
417 		if (xfeature_is_aligned(i))
418 			next_offset = ALIGN(next_offset, 64);
419 
420 		xstate_supervisor_only_offsets[i] = next_offset;
421 		next_offset += xstate_sizes[i];
422 	}
423 }
424 
425 /*
426  * Print out xstate component offsets and sizes
427  */
428 static void __init print_xstate_offset_size(void)
429 {
430 	int i;
431 
432 	for (i = FIRST_EXTENDED_XFEATURE; i < XFEATURE_MAX; i++) {
433 		if (!xfeature_enabled(i))
434 			continue;
435 		pr_info("x86/fpu: xstate_offset[%d]: %4d, xstate_sizes[%d]: %4d\n",
436 			 i, xstate_comp_offsets[i], i, xstate_sizes[i]);
437 	}
438 }
439 
440 /*
441  * setup the xstate image representing the init state
442  */
443 static void __init setup_init_fpu_buf(void)
444 {
445 	static int on_boot_cpu __initdata = 1;
446 
447 	WARN_ON_FPU(!on_boot_cpu);
448 	on_boot_cpu = 0;
449 
450 	if (!boot_cpu_has(X86_FEATURE_XSAVE))
451 		return;
452 
453 	setup_xstate_features();
454 	print_xstate_features();
455 
456 	if (boot_cpu_has(X86_FEATURE_XSAVES))
457 		init_fpstate.xsave.header.xcomp_bv = XCOMP_BV_COMPACTED_FORMAT |
458 						     xfeatures_mask_all;
459 
460 	/*
461 	 * Init all the features state with header.xfeatures being 0x0
462 	 */
463 	copy_kernel_to_xregs_booting(&init_fpstate.xsave);
464 
465 	/*
466 	 * Dump the init state again. This is to identify the init state
467 	 * of any feature which is not represented by all zero's.
468 	 */
469 	copy_xregs_to_kernel_booting(&init_fpstate.xsave);
470 }
471 
472 static int xfeature_uncompacted_offset(int xfeature_nr)
473 {
474 	u32 eax, ebx, ecx, edx;
475 
476 	/*
477 	 * Only XSAVES supports supervisor states and it uses compacted
478 	 * format. Checking a supervisor state's uncompacted offset is
479 	 * an error.
480 	 */
481 	if (XFEATURE_MASK_SUPERVISOR_ALL & BIT_ULL(xfeature_nr)) {
482 		WARN_ONCE(1, "No fixed offset for xstate %d\n", xfeature_nr);
483 		return -1;
484 	}
485 
486 	CHECK_XFEATURE(xfeature_nr);
487 	cpuid_count(XSTATE_CPUID, xfeature_nr, &eax, &ebx, &ecx, &edx);
488 	return ebx;
489 }
490 
491 int xfeature_size(int xfeature_nr)
492 {
493 	u32 eax, ebx, ecx, edx;
494 
495 	CHECK_XFEATURE(xfeature_nr);
496 	cpuid_count(XSTATE_CPUID, xfeature_nr, &eax, &ebx, &ecx, &edx);
497 	return eax;
498 }
499 
500 /*
501  * 'XSAVES' implies two different things:
502  * 1. saving of supervisor/system state
503  * 2. using the compacted format
504  *
505  * Use this function when dealing with the compacted format so
506  * that it is obvious which aspect of 'XSAVES' is being handled
507  * by the calling code.
508  */
509 int using_compacted_format(void)
510 {
511 	return boot_cpu_has(X86_FEATURE_XSAVES);
512 }
513 
514 /* Validate an xstate header supplied by userspace (ptrace or sigreturn) */
515 int validate_user_xstate_header(const struct xstate_header *hdr)
516 {
517 	/* No unknown or supervisor features may be set */
518 	if (hdr->xfeatures & ~xfeatures_mask_user())
519 		return -EINVAL;
520 
521 	/* Userspace must use the uncompacted format */
522 	if (hdr->xcomp_bv)
523 		return -EINVAL;
524 
525 	/*
526 	 * If 'reserved' is shrunken to add a new field, make sure to validate
527 	 * that new field here!
528 	 */
529 	BUILD_BUG_ON(sizeof(hdr->reserved) != 48);
530 
531 	/* No reserved bits may be set */
532 	if (memchr_inv(hdr->reserved, 0, sizeof(hdr->reserved)))
533 		return -EINVAL;
534 
535 	return 0;
536 }
537 
538 static void __xstate_dump_leaves(void)
539 {
540 	int i;
541 	u32 eax, ebx, ecx, edx;
542 	static int should_dump = 1;
543 
544 	if (!should_dump)
545 		return;
546 	should_dump = 0;
547 	/*
548 	 * Dump out a few leaves past the ones that we support
549 	 * just in case there are some goodies up there
550 	 */
551 	for (i = 0; i < XFEATURE_MAX + 10; i++) {
552 		cpuid_count(XSTATE_CPUID, i, &eax, &ebx, &ecx, &edx);
553 		pr_warn("CPUID[%02x, %02x]: eax=%08x ebx=%08x ecx=%08x edx=%08x\n",
554 			XSTATE_CPUID, i, eax, ebx, ecx, edx);
555 	}
556 }
557 
558 #define XSTATE_WARN_ON(x) do {							\
559 	if (WARN_ONCE(x, "XSAVE consistency problem, dumping leaves")) {	\
560 		__xstate_dump_leaves();						\
561 	}									\
562 } while (0)
563 
564 #define XCHECK_SZ(sz, nr, nr_macro, __struct) do {			\
565 	if ((nr == nr_macro) &&						\
566 	    WARN_ONCE(sz != sizeof(__struct),				\
567 		"%s: struct is %zu bytes, cpu state %d bytes\n",	\
568 		__stringify(nr_macro), sizeof(__struct), sz)) {		\
569 		__xstate_dump_leaves();					\
570 	}								\
571 } while (0)
572 
573 /*
574  * We have a C struct for each 'xstate'.  We need to ensure
575  * that our software representation matches what the CPU
576  * tells us about the state's size.
577  */
578 static void check_xstate_against_struct(int nr)
579 {
580 	/*
581 	 * Ask the CPU for the size of the state.
582 	 */
583 	int sz = xfeature_size(nr);
584 	/*
585 	 * Match each CPU state with the corresponding software
586 	 * structure.
587 	 */
588 	XCHECK_SZ(sz, nr, XFEATURE_YMM,       struct ymmh_struct);
589 	XCHECK_SZ(sz, nr, XFEATURE_BNDREGS,   struct mpx_bndreg_state);
590 	XCHECK_SZ(sz, nr, XFEATURE_BNDCSR,    struct mpx_bndcsr_state);
591 	XCHECK_SZ(sz, nr, XFEATURE_OPMASK,    struct avx_512_opmask_state);
592 	XCHECK_SZ(sz, nr, XFEATURE_ZMM_Hi256, struct avx_512_zmm_uppers_state);
593 	XCHECK_SZ(sz, nr, XFEATURE_Hi16_ZMM,  struct avx_512_hi16_state);
594 	XCHECK_SZ(sz, nr, XFEATURE_PKRU,      struct pkru_state);
595 
596 	/*
597 	 * Make *SURE* to add any feature numbers in below if
598 	 * there are "holes" in the xsave state component
599 	 * numbers.
600 	 */
601 	if ((nr < XFEATURE_YMM) ||
602 	    (nr >= XFEATURE_MAX) ||
603 	    (nr == XFEATURE_PT_UNIMPLEMENTED_SO_FAR) ||
604 	    ((nr >= XFEATURE_RSRVD_COMP_10) && (nr <= XFEATURE_LBR))) {
605 		WARN_ONCE(1, "no structure for xstate: %d\n", nr);
606 		XSTATE_WARN_ON(1);
607 	}
608 }
609 
610 /*
611  * This essentially double-checks what the cpu told us about
612  * how large the XSAVE buffer needs to be.  We are recalculating
613  * it to be safe.
614  */
615 static void do_extra_xstate_size_checks(void)
616 {
617 	int paranoid_xstate_size = FXSAVE_SIZE + XSAVE_HDR_SIZE;
618 	int i;
619 
620 	for (i = FIRST_EXTENDED_XFEATURE; i < XFEATURE_MAX; i++) {
621 		if (!xfeature_enabled(i))
622 			continue;
623 
624 		check_xstate_against_struct(i);
625 		/*
626 		 * Supervisor state components can be managed only by
627 		 * XSAVES, which is compacted-format only.
628 		 */
629 		if (!using_compacted_format())
630 			XSTATE_WARN_ON(xfeature_is_supervisor(i));
631 
632 		/* Align from the end of the previous feature */
633 		if (xfeature_is_aligned(i))
634 			paranoid_xstate_size = ALIGN(paranoid_xstate_size, 64);
635 		/*
636 		 * The offset of a given state in the non-compacted
637 		 * format is given to us in a CPUID leaf.  We check
638 		 * them for being ordered (increasing offsets) in
639 		 * setup_xstate_features().
640 		 */
641 		if (!using_compacted_format())
642 			paranoid_xstate_size = xfeature_uncompacted_offset(i);
643 		/*
644 		 * The compacted-format offset always depends on where
645 		 * the previous state ended.
646 		 */
647 		paranoid_xstate_size += xfeature_size(i);
648 	}
649 	XSTATE_WARN_ON(paranoid_xstate_size != fpu_kernel_xstate_size);
650 }
651 
652 
653 /*
654  * Get total size of enabled xstates in XCR0 | IA32_XSS.
655  *
656  * Note the SDM's wording here.  "sub-function 0" only enumerates
657  * the size of the *user* states.  If we use it to size a buffer
658  * that we use 'XSAVES' on, we could potentially overflow the
659  * buffer because 'XSAVES' saves system states too.
660  */
661 static unsigned int __init get_xsaves_size(void)
662 {
663 	unsigned int eax, ebx, ecx, edx;
664 	/*
665 	 * - CPUID function 0DH, sub-function 1:
666 	 *    EBX enumerates the size (in bytes) required by
667 	 *    the XSAVES instruction for an XSAVE area
668 	 *    containing all the state components
669 	 *    corresponding to bits currently set in
670 	 *    XCR0 | IA32_XSS.
671 	 */
672 	cpuid_count(XSTATE_CPUID, 1, &eax, &ebx, &ecx, &edx);
673 	return ebx;
674 }
675 
676 static unsigned int __init get_xsave_size(void)
677 {
678 	unsigned int eax, ebx, ecx, edx;
679 	/*
680 	 * - CPUID function 0DH, sub-function 0:
681 	 *    EBX enumerates the size (in bytes) required by
682 	 *    the XSAVE instruction for an XSAVE area
683 	 *    containing all the *user* state components
684 	 *    corresponding to bits currently set in XCR0.
685 	 */
686 	cpuid_count(XSTATE_CPUID, 0, &eax, &ebx, &ecx, &edx);
687 	return ebx;
688 }
689 
690 /*
691  * Will the runtime-enumerated 'xstate_size' fit in the init
692  * task's statically-allocated buffer?
693  */
694 static bool is_supported_xstate_size(unsigned int test_xstate_size)
695 {
696 	if (test_xstate_size <= sizeof(union fpregs_state))
697 		return true;
698 
699 	pr_warn("x86/fpu: xstate buffer too small (%zu < %d), disabling xsave\n",
700 			sizeof(union fpregs_state), test_xstate_size);
701 	return false;
702 }
703 
704 static int __init init_xstate_size(void)
705 {
706 	/* Recompute the context size for enabled features: */
707 	unsigned int possible_xstate_size;
708 	unsigned int xsave_size;
709 
710 	xsave_size = get_xsave_size();
711 
712 	if (boot_cpu_has(X86_FEATURE_XSAVES))
713 		possible_xstate_size = get_xsaves_size();
714 	else
715 		possible_xstate_size = xsave_size;
716 
717 	/* Ensure we have the space to store all enabled: */
718 	if (!is_supported_xstate_size(possible_xstate_size))
719 		return -EINVAL;
720 
721 	/*
722 	 * The size is OK, we are definitely going to use xsave,
723 	 * make it known to the world that we need more space.
724 	 */
725 	fpu_kernel_xstate_size = possible_xstate_size;
726 	do_extra_xstate_size_checks();
727 
728 	/*
729 	 * User space is always in standard format.
730 	 */
731 	fpu_user_xstate_size = xsave_size;
732 	return 0;
733 }
734 
735 /*
736  * We enabled the XSAVE hardware, but something went wrong and
737  * we can not use it.  Disable it.
738  */
739 static void fpu__init_disable_system_xstate(void)
740 {
741 	xfeatures_mask_all = 0;
742 	cr4_clear_bits(X86_CR4_OSXSAVE);
743 	setup_clear_cpu_cap(X86_FEATURE_XSAVE);
744 }
745 
746 /*
747  * Enable and initialize the xsave feature.
748  * Called once per system bootup.
749  */
750 void __init fpu__init_system_xstate(void)
751 {
752 	unsigned int eax, ebx, ecx, edx;
753 	static int on_boot_cpu __initdata = 1;
754 	int err;
755 	int i;
756 
757 	WARN_ON_FPU(!on_boot_cpu);
758 	on_boot_cpu = 0;
759 
760 	if (!boot_cpu_has(X86_FEATURE_FPU)) {
761 		pr_info("x86/fpu: No FPU detected\n");
762 		return;
763 	}
764 
765 	if (!boot_cpu_has(X86_FEATURE_XSAVE)) {
766 		pr_info("x86/fpu: x87 FPU will use %s\n",
767 			boot_cpu_has(X86_FEATURE_FXSR) ? "FXSAVE" : "FSAVE");
768 		return;
769 	}
770 
771 	if (boot_cpu_data.cpuid_level < XSTATE_CPUID) {
772 		WARN_ON_FPU(1);
773 		return;
774 	}
775 
776 	/*
777 	 * Find user xstates supported by the processor.
778 	 */
779 	cpuid_count(XSTATE_CPUID, 0, &eax, &ebx, &ecx, &edx);
780 	xfeatures_mask_all = eax + ((u64)edx << 32);
781 
782 	/*
783 	 * Find supervisor xstates supported by the processor.
784 	 */
785 	cpuid_count(XSTATE_CPUID, 1, &eax, &ebx, &ecx, &edx);
786 	xfeatures_mask_all |= ecx + ((u64)edx << 32);
787 
788 	if ((xfeatures_mask_user() & XFEATURE_MASK_FPSSE) != XFEATURE_MASK_FPSSE) {
789 		/*
790 		 * This indicates that something really unexpected happened
791 		 * with the enumeration.  Disable XSAVE and try to continue
792 		 * booting without it.  This is too early to BUG().
793 		 */
794 		pr_err("x86/fpu: FP/SSE not present amongst the CPU's xstate features: 0x%llx.\n",
795 		       xfeatures_mask_all);
796 		goto out_disable;
797 	}
798 
799 	/*
800 	 * Clear XSAVE features that are disabled in the normal CPUID.
801 	 */
802 	for (i = 0; i < ARRAY_SIZE(xsave_cpuid_features); i++) {
803 		if (!boot_cpu_has(xsave_cpuid_features[i]))
804 			xfeatures_mask_all &= ~BIT_ULL(i);
805 	}
806 
807 	xfeatures_mask_all &= fpu__get_supported_xfeatures_mask();
808 
809 	/* Enable xstate instructions to be able to continue with initialization: */
810 	fpu__init_cpu_xstate();
811 	err = init_xstate_size();
812 	if (err)
813 		goto out_disable;
814 
815 	/*
816 	 * Update info used for ptrace frames; use standard-format size and no
817 	 * supervisor xstates:
818 	 */
819 	update_regset_xstate_info(fpu_user_xstate_size, xfeatures_mask_user());
820 
821 	fpu__init_prepare_fx_sw_frame();
822 	setup_init_fpu_buf();
823 	setup_xstate_comp_offsets();
824 	setup_supervisor_only_offsets();
825 	print_xstate_offset_size();
826 
827 	pr_info("x86/fpu: Enabled xstate features 0x%llx, context size is %d bytes, using '%s' format.\n",
828 		xfeatures_mask_all,
829 		fpu_kernel_xstate_size,
830 		boot_cpu_has(X86_FEATURE_XSAVES) ? "compacted" : "standard");
831 	return;
832 
833 out_disable:
834 	/* something went wrong, try to boot without any XSAVE support */
835 	fpu__init_disable_system_xstate();
836 }
837 
838 /*
839  * Restore minimal FPU state after suspend:
840  */
841 void fpu__resume_cpu(void)
842 {
843 	/*
844 	 * Restore XCR0 on xsave capable CPUs:
845 	 */
846 	if (boot_cpu_has(X86_FEATURE_XSAVE))
847 		xsetbv(XCR_XFEATURE_ENABLED_MASK, xfeatures_mask_user());
848 
849 	/*
850 	 * Restore IA32_XSS. The same CPUID bit enumerates support
851 	 * of XSAVES and MSR_IA32_XSS.
852 	 */
853 	if (boot_cpu_has(X86_FEATURE_XSAVES)) {
854 		wrmsrl(MSR_IA32_XSS, xfeatures_mask_supervisor()  |
855 				     xfeatures_mask_dynamic());
856 	}
857 }
858 
859 /*
860  * Given an xstate feature nr, calculate where in the xsave
861  * buffer the state is.  Callers should ensure that the buffer
862  * is valid.
863  */
864 static void *__raw_xsave_addr(struct xregs_state *xsave, int xfeature_nr)
865 {
866 	if (!xfeature_enabled(xfeature_nr)) {
867 		WARN_ON_FPU(1);
868 		return NULL;
869 	}
870 
871 	return (void *)xsave + xstate_comp_offsets[xfeature_nr];
872 }
873 /*
874  * Given the xsave area and a state inside, this function returns the
875  * address of the state.
876  *
877  * This is the API that is called to get xstate address in either
878  * standard format or compacted format of xsave area.
879  *
880  * Note that if there is no data for the field in the xsave buffer
881  * this will return NULL.
882  *
883  * Inputs:
884  *	xstate: the thread's storage area for all FPU data
885  *	xfeature_nr: state which is defined in xsave.h (e.g. XFEATURE_FP,
886  *	XFEATURE_SSE, etc...)
887  * Output:
888  *	address of the state in the xsave area, or NULL if the
889  *	field is not present in the xsave buffer.
890  */
891 void *get_xsave_addr(struct xregs_state *xsave, int xfeature_nr)
892 {
893 	/*
894 	 * Do we even *have* xsave state?
895 	 */
896 	if (!boot_cpu_has(X86_FEATURE_XSAVE))
897 		return NULL;
898 
899 	/*
900 	 * We should not ever be requesting features that we
901 	 * have not enabled.
902 	 */
903 	WARN_ONCE(!(xfeatures_mask_all & BIT_ULL(xfeature_nr)),
904 		  "get of unsupported state");
905 	/*
906 	 * This assumes the last 'xsave*' instruction to
907 	 * have requested that 'xfeature_nr' be saved.
908 	 * If it did not, we might be seeing and old value
909 	 * of the field in the buffer.
910 	 *
911 	 * This can happen because the last 'xsave' did not
912 	 * request that this feature be saved (unlikely)
913 	 * or because the "init optimization" caused it
914 	 * to not be saved.
915 	 */
916 	if (!(xsave->header.xfeatures & BIT_ULL(xfeature_nr)))
917 		return NULL;
918 
919 	return __raw_xsave_addr(xsave, xfeature_nr);
920 }
921 EXPORT_SYMBOL_GPL(get_xsave_addr);
922 
923 /*
924  * This wraps up the common operations that need to occur when retrieving
925  * data from xsave state.  It first ensures that the current task was
926  * using the FPU and retrieves the data in to a buffer.  It then calculates
927  * the offset of the requested field in the buffer.
928  *
929  * This function is safe to call whether the FPU is in use or not.
930  *
931  * Note that this only works on the current task.
932  *
933  * Inputs:
934  *	@xfeature_nr: state which is defined in xsave.h (e.g. XFEATURE_FP,
935  *	XFEATURE_SSE, etc...)
936  * Output:
937  *	address of the state in the xsave area or NULL if the state
938  *	is not present or is in its 'init state'.
939  */
940 const void *get_xsave_field_ptr(int xfeature_nr)
941 {
942 	struct fpu *fpu = &current->thread.fpu;
943 
944 	/*
945 	 * fpu__save() takes the CPU's xstate registers
946 	 * and saves them off to the 'fpu memory buffer.
947 	 */
948 	fpu__save(fpu);
949 
950 	return get_xsave_addr(&fpu->state.xsave, xfeature_nr);
951 }
952 
953 #ifdef CONFIG_ARCH_HAS_PKEYS
954 
955 /*
956  * This will go out and modify PKRU register to set the access
957  * rights for @pkey to @init_val.
958  */
959 int arch_set_user_pkey_access(struct task_struct *tsk, int pkey,
960 		unsigned long init_val)
961 {
962 	u32 old_pkru;
963 	int pkey_shift = (pkey * PKRU_BITS_PER_PKEY);
964 	u32 new_pkru_bits = 0;
965 
966 	/*
967 	 * This check implies XSAVE support.  OSPKE only gets
968 	 * set if we enable XSAVE and we enable PKU in XCR0.
969 	 */
970 	if (!boot_cpu_has(X86_FEATURE_OSPKE))
971 		return -EINVAL;
972 
973 	/*
974 	 * This code should only be called with valid 'pkey'
975 	 * values originating from in-kernel users.  Complain
976 	 * if a bad value is observed.
977 	 */
978 	WARN_ON_ONCE(pkey >= arch_max_pkey());
979 
980 	/* Set the bits we need in PKRU:  */
981 	if (init_val & PKEY_DISABLE_ACCESS)
982 		new_pkru_bits |= PKRU_AD_BIT;
983 	if (init_val & PKEY_DISABLE_WRITE)
984 		new_pkru_bits |= PKRU_WD_BIT;
985 
986 	/* Shift the bits in to the correct place in PKRU for pkey: */
987 	new_pkru_bits <<= pkey_shift;
988 
989 	/* Get old PKRU and mask off any old bits in place: */
990 	old_pkru = read_pkru();
991 	old_pkru &= ~((PKRU_AD_BIT|PKRU_WD_BIT) << pkey_shift);
992 
993 	/* Write old part along with new part: */
994 	write_pkru(old_pkru | new_pkru_bits);
995 
996 	return 0;
997 }
998 #endif /* ! CONFIG_ARCH_HAS_PKEYS */
999 
1000 /*
1001  * Weird legacy quirk: SSE and YMM states store information in the
1002  * MXCSR and MXCSR_FLAGS fields of the FP area. That means if the FP
1003  * area is marked as unused in the xfeatures header, we need to copy
1004  * MXCSR and MXCSR_FLAGS if either SSE or YMM are in use.
1005  */
1006 static inline bool xfeatures_mxcsr_quirk(u64 xfeatures)
1007 {
1008 	if (!(xfeatures & (XFEATURE_MASK_SSE|XFEATURE_MASK_YMM)))
1009 		return false;
1010 
1011 	if (xfeatures & XFEATURE_MASK_FP)
1012 		return false;
1013 
1014 	return true;
1015 }
1016 
1017 static void fill_gap(unsigned to, void **kbuf, unsigned *pos, unsigned *count)
1018 {
1019 	if (*pos < to) {
1020 		unsigned size = to - *pos;
1021 
1022 		if (size > *count)
1023 			size = *count;
1024 		memcpy(*kbuf, (void *)&init_fpstate.xsave + *pos, size);
1025 		*kbuf += size;
1026 		*pos += size;
1027 		*count -= size;
1028 	}
1029 }
1030 
1031 static void copy_part(unsigned offset, unsigned size, void *from,
1032 			void **kbuf, unsigned *pos, unsigned *count)
1033 {
1034 	fill_gap(offset, kbuf, pos, count);
1035 	if (size > *count)
1036 		size = *count;
1037 	if (size) {
1038 		memcpy(*kbuf, from, size);
1039 		*kbuf += size;
1040 		*pos += size;
1041 		*count -= size;
1042 	}
1043 }
1044 
1045 /*
1046  * Convert from kernel XSAVES compacted format to standard format and copy
1047  * to a kernel-space ptrace buffer.
1048  *
1049  * It supports partial copy but pos always starts from zero. This is called
1050  * from xstateregs_get() and there we check the CPU has XSAVES.
1051  */
1052 int copy_xstate_to_kernel(void *kbuf, struct xregs_state *xsave, unsigned int offset_start, unsigned int size_total)
1053 {
1054 	struct xstate_header header;
1055 	const unsigned off_mxcsr = offsetof(struct fxregs_state, mxcsr);
1056 	unsigned count = size_total;
1057 	int i;
1058 
1059 	/*
1060 	 * Currently copy_regset_to_user() starts from pos 0:
1061 	 */
1062 	if (unlikely(offset_start != 0))
1063 		return -EFAULT;
1064 
1065 	/*
1066 	 * The destination is a ptrace buffer; we put in only user xstates:
1067 	 */
1068 	memset(&header, 0, sizeof(header));
1069 	header.xfeatures = xsave->header.xfeatures;
1070 	header.xfeatures &= xfeatures_mask_user();
1071 
1072 	if (header.xfeatures & XFEATURE_MASK_FP)
1073 		copy_part(0, off_mxcsr,
1074 			  &xsave->i387, &kbuf, &offset_start, &count);
1075 	if (header.xfeatures & (XFEATURE_MASK_SSE | XFEATURE_MASK_YMM))
1076 		copy_part(off_mxcsr, MXCSR_AND_FLAGS_SIZE,
1077 			  &xsave->i387.mxcsr, &kbuf, &offset_start, &count);
1078 	if (header.xfeatures & XFEATURE_MASK_FP)
1079 		copy_part(offsetof(struct fxregs_state, st_space), 128,
1080 			  &xsave->i387.st_space, &kbuf, &offset_start, &count);
1081 	if (header.xfeatures & XFEATURE_MASK_SSE)
1082 		copy_part(xstate_offsets[XFEATURE_SSE], 256,
1083 			  &xsave->i387.xmm_space, &kbuf, &offset_start, &count);
1084 	/*
1085 	 * Fill xsave->i387.sw_reserved value for ptrace frame:
1086 	 */
1087 	copy_part(offsetof(struct fxregs_state, sw_reserved), 48,
1088 		  xstate_fx_sw_bytes, &kbuf, &offset_start, &count);
1089 	/*
1090 	 * Copy xregs_state->header:
1091 	 */
1092 	copy_part(offsetof(struct xregs_state, header), sizeof(header),
1093 		  &header, &kbuf, &offset_start, &count);
1094 
1095 	for (i = FIRST_EXTENDED_XFEATURE; i < XFEATURE_MAX; i++) {
1096 		/*
1097 		 * Copy only in-use xstates:
1098 		 */
1099 		if ((header.xfeatures >> i) & 1) {
1100 			void *src = __raw_xsave_addr(xsave, i);
1101 
1102 			copy_part(xstate_offsets[i], xstate_sizes[i],
1103 				  src, &kbuf, &offset_start, &count);
1104 		}
1105 
1106 	}
1107 	fill_gap(size_total, &kbuf, &offset_start, &count);
1108 
1109 	return 0;
1110 }
1111 
1112 static inline int
1113 __copy_xstate_to_user(void __user *ubuf, const void *data, unsigned int offset, unsigned int size, unsigned int size_total)
1114 {
1115 	if (!size)
1116 		return 0;
1117 
1118 	if (offset < size_total) {
1119 		unsigned int copy = min(size, size_total - offset);
1120 
1121 		if (__copy_to_user(ubuf + offset, data, copy))
1122 			return -EFAULT;
1123 	}
1124 	return 0;
1125 }
1126 
1127 /*
1128  * Convert from kernel XSAVES compacted format to standard format and copy
1129  * to a user-space buffer. It supports partial copy but pos always starts from
1130  * zero. This is called from xstateregs_get() and there we check the CPU
1131  * has XSAVES.
1132  */
1133 int copy_xstate_to_user(void __user *ubuf, struct xregs_state *xsave, unsigned int offset_start, unsigned int size_total)
1134 {
1135 	unsigned int offset, size;
1136 	int ret, i;
1137 	struct xstate_header header;
1138 
1139 	/*
1140 	 * Currently copy_regset_to_user() starts from pos 0:
1141 	 */
1142 	if (unlikely(offset_start != 0))
1143 		return -EFAULT;
1144 
1145 	/*
1146 	 * The destination is a ptrace buffer; we put in only user xstates:
1147 	 */
1148 	memset(&header, 0, sizeof(header));
1149 	header.xfeatures = xsave->header.xfeatures;
1150 	header.xfeatures &= xfeatures_mask_user();
1151 
1152 	/*
1153 	 * Copy xregs_state->header:
1154 	 */
1155 	offset = offsetof(struct xregs_state, header);
1156 	size = sizeof(header);
1157 
1158 	ret = __copy_xstate_to_user(ubuf, &header, offset, size, size_total);
1159 	if (ret)
1160 		return ret;
1161 
1162 	for (i = 0; i < XFEATURE_MAX; i++) {
1163 		/*
1164 		 * Copy only in-use xstates:
1165 		 */
1166 		if ((header.xfeatures >> i) & 1) {
1167 			void *src = __raw_xsave_addr(xsave, i);
1168 
1169 			offset = xstate_offsets[i];
1170 			size = xstate_sizes[i];
1171 
1172 			/* The next component has to fit fully into the output buffer: */
1173 			if (offset + size > size_total)
1174 				break;
1175 
1176 			ret = __copy_xstate_to_user(ubuf, src, offset, size, size_total);
1177 			if (ret)
1178 				return ret;
1179 		}
1180 
1181 	}
1182 
1183 	if (xfeatures_mxcsr_quirk(header.xfeatures)) {
1184 		offset = offsetof(struct fxregs_state, mxcsr);
1185 		size = MXCSR_AND_FLAGS_SIZE;
1186 		__copy_xstate_to_user(ubuf, &xsave->i387.mxcsr, offset, size, size_total);
1187 	}
1188 
1189 	/*
1190 	 * Fill xsave->i387.sw_reserved value for ptrace frame:
1191 	 */
1192 	offset = offsetof(struct fxregs_state, sw_reserved);
1193 	size = sizeof(xstate_fx_sw_bytes);
1194 
1195 	ret = __copy_xstate_to_user(ubuf, xstate_fx_sw_bytes, offset, size, size_total);
1196 	if (ret)
1197 		return ret;
1198 
1199 	return 0;
1200 }
1201 
1202 /*
1203  * Convert from a ptrace standard-format kernel buffer to kernel XSAVES format
1204  * and copy to the target thread. This is called from xstateregs_set().
1205  */
1206 int copy_kernel_to_xstate(struct xregs_state *xsave, const void *kbuf)
1207 {
1208 	unsigned int offset, size;
1209 	int i;
1210 	struct xstate_header hdr;
1211 
1212 	offset = offsetof(struct xregs_state, header);
1213 	size = sizeof(hdr);
1214 
1215 	memcpy(&hdr, kbuf + offset, size);
1216 
1217 	if (validate_user_xstate_header(&hdr))
1218 		return -EINVAL;
1219 
1220 	for (i = 0; i < XFEATURE_MAX; i++) {
1221 		u64 mask = ((u64)1 << i);
1222 
1223 		if (hdr.xfeatures & mask) {
1224 			void *dst = __raw_xsave_addr(xsave, i);
1225 
1226 			offset = xstate_offsets[i];
1227 			size = xstate_sizes[i];
1228 
1229 			memcpy(dst, kbuf + offset, size);
1230 		}
1231 	}
1232 
1233 	if (xfeatures_mxcsr_quirk(hdr.xfeatures)) {
1234 		offset = offsetof(struct fxregs_state, mxcsr);
1235 		size = MXCSR_AND_FLAGS_SIZE;
1236 		memcpy(&xsave->i387.mxcsr, kbuf + offset, size);
1237 	}
1238 
1239 	/*
1240 	 * The state that came in from userspace was user-state only.
1241 	 * Mask all the user states out of 'xfeatures':
1242 	 */
1243 	xsave->header.xfeatures &= XFEATURE_MASK_SUPERVISOR_ALL;
1244 
1245 	/*
1246 	 * Add back in the features that came in from userspace:
1247 	 */
1248 	xsave->header.xfeatures |= hdr.xfeatures;
1249 
1250 	return 0;
1251 }
1252 
1253 /*
1254  * Convert from a ptrace or sigreturn standard-format user-space buffer to
1255  * kernel XSAVES format and copy to the target thread. This is called from
1256  * xstateregs_set(), as well as potentially from the sigreturn() and
1257  * rt_sigreturn() system calls.
1258  */
1259 int copy_user_to_xstate(struct xregs_state *xsave, const void __user *ubuf)
1260 {
1261 	unsigned int offset, size;
1262 	int i;
1263 	struct xstate_header hdr;
1264 
1265 	offset = offsetof(struct xregs_state, header);
1266 	size = sizeof(hdr);
1267 
1268 	if (__copy_from_user(&hdr, ubuf + offset, size))
1269 		return -EFAULT;
1270 
1271 	if (validate_user_xstate_header(&hdr))
1272 		return -EINVAL;
1273 
1274 	for (i = 0; i < XFEATURE_MAX; i++) {
1275 		u64 mask = ((u64)1 << i);
1276 
1277 		if (hdr.xfeatures & mask) {
1278 			void *dst = __raw_xsave_addr(xsave, i);
1279 
1280 			offset = xstate_offsets[i];
1281 			size = xstate_sizes[i];
1282 
1283 			if (__copy_from_user(dst, ubuf + offset, size))
1284 				return -EFAULT;
1285 		}
1286 	}
1287 
1288 	if (xfeatures_mxcsr_quirk(hdr.xfeatures)) {
1289 		offset = offsetof(struct fxregs_state, mxcsr);
1290 		size = MXCSR_AND_FLAGS_SIZE;
1291 		if (__copy_from_user(&xsave->i387.mxcsr, ubuf + offset, size))
1292 			return -EFAULT;
1293 	}
1294 
1295 	/*
1296 	 * The state that came in from userspace was user-state only.
1297 	 * Mask all the user states out of 'xfeatures':
1298 	 */
1299 	xsave->header.xfeatures &= XFEATURE_MASK_SUPERVISOR_ALL;
1300 
1301 	/*
1302 	 * Add back in the features that came in from userspace:
1303 	 */
1304 	xsave->header.xfeatures |= hdr.xfeatures;
1305 
1306 	return 0;
1307 }
1308 
1309 /*
1310  * Save only supervisor states to the kernel buffer.  This blows away all
1311  * old states, and is intended to be used only in __fpu__restore_sig(), where
1312  * user states are restored from the user buffer.
1313  */
1314 void copy_supervisor_to_kernel(struct xregs_state *xstate)
1315 {
1316 	struct xstate_header *header;
1317 	u64 max_bit, min_bit;
1318 	u32 lmask, hmask;
1319 	int err, i;
1320 
1321 	if (WARN_ON(!boot_cpu_has(X86_FEATURE_XSAVES)))
1322 		return;
1323 
1324 	if (!xfeatures_mask_supervisor())
1325 		return;
1326 
1327 	max_bit = __fls(xfeatures_mask_supervisor());
1328 	min_bit = __ffs(xfeatures_mask_supervisor());
1329 
1330 	lmask = xfeatures_mask_supervisor();
1331 	hmask = xfeatures_mask_supervisor() >> 32;
1332 	XSTATE_OP(XSAVES, xstate, lmask, hmask, err);
1333 
1334 	/* We should never fault when copying to a kernel buffer: */
1335 	if (WARN_ON_FPU(err))
1336 		return;
1337 
1338 	/*
1339 	 * At this point, the buffer has only supervisor states and must be
1340 	 * converted back to normal kernel format.
1341 	 */
1342 	header = &xstate->header;
1343 	header->xcomp_bv |= xfeatures_mask_all;
1344 
1345 	/*
1346 	 * This only moves states up in the buffer.  Start with
1347 	 * the last state and move backwards so that states are
1348 	 * not overwritten until after they are moved.  Note:
1349 	 * memmove() allows overlapping src/dst buffers.
1350 	 */
1351 	for (i = max_bit; i >= min_bit; i--) {
1352 		u8 *xbuf = (u8 *)xstate;
1353 
1354 		if (!((header->xfeatures >> i) & 1))
1355 			continue;
1356 
1357 		/* Move xfeature 'i' into its normal location */
1358 		memmove(xbuf + xstate_comp_offsets[i],
1359 			xbuf + xstate_supervisor_only_offsets[i],
1360 			xstate_sizes[i]);
1361 	}
1362 }
1363 
1364 /**
1365  * copy_dynamic_supervisor_to_kernel() - Save dynamic supervisor states to
1366  *                                       an xsave area
1367  * @xstate: A pointer to an xsave area
1368  * @mask: Represent the dynamic supervisor features saved into the xsave area
1369  *
1370  * Only the dynamic supervisor states sets in the mask are saved into the xsave
1371  * area (See the comment in XFEATURE_MASK_DYNAMIC for the details of dynamic
1372  * supervisor feature). Besides the dynamic supervisor states, the legacy
1373  * region and XSAVE header are also saved into the xsave area. The supervisor
1374  * features in the XFEATURE_MASK_SUPERVISOR_SUPPORTED and
1375  * XFEATURE_MASK_SUPERVISOR_UNSUPPORTED are not saved.
1376  *
1377  * The xsave area must be 64-bytes aligned.
1378  */
1379 void copy_dynamic_supervisor_to_kernel(struct xregs_state *xstate, u64 mask)
1380 {
1381 	u64 dynamic_mask = xfeatures_mask_dynamic() & mask;
1382 	u32 lmask, hmask;
1383 	int err;
1384 
1385 	if (WARN_ON_FPU(!boot_cpu_has(X86_FEATURE_XSAVES)))
1386 		return;
1387 
1388 	if (WARN_ON_FPU(!dynamic_mask))
1389 		return;
1390 
1391 	lmask = dynamic_mask;
1392 	hmask = dynamic_mask >> 32;
1393 
1394 	XSTATE_OP(XSAVES, xstate, lmask, hmask, err);
1395 
1396 	/* Should never fault when copying to a kernel buffer */
1397 	WARN_ON_FPU(err);
1398 }
1399 
1400 /**
1401  * copy_kernel_to_dynamic_supervisor() - Restore dynamic supervisor states from
1402  *                                       an xsave area
1403  * @xstate: A pointer to an xsave area
1404  * @mask: Represent the dynamic supervisor features restored from the xsave area
1405  *
1406  * Only the dynamic supervisor states sets in the mask are restored from the
1407  * xsave area (See the comment in XFEATURE_MASK_DYNAMIC for the details of
1408  * dynamic supervisor feature). Besides the dynamic supervisor states, the
1409  * legacy region and XSAVE header are also restored from the xsave area. The
1410  * supervisor features in the XFEATURE_MASK_SUPERVISOR_SUPPORTED and
1411  * XFEATURE_MASK_SUPERVISOR_UNSUPPORTED are not restored.
1412  *
1413  * The xsave area must be 64-bytes aligned.
1414  */
1415 void copy_kernel_to_dynamic_supervisor(struct xregs_state *xstate, u64 mask)
1416 {
1417 	u64 dynamic_mask = xfeatures_mask_dynamic() & mask;
1418 	u32 lmask, hmask;
1419 	int err;
1420 
1421 	if (WARN_ON_FPU(!boot_cpu_has(X86_FEATURE_XSAVES)))
1422 		return;
1423 
1424 	if (WARN_ON_FPU(!dynamic_mask))
1425 		return;
1426 
1427 	lmask = dynamic_mask;
1428 	hmask = dynamic_mask >> 32;
1429 
1430 	XSTATE_OP(XRSTORS, xstate, lmask, hmask, err);
1431 
1432 	/* Should never fault when copying from a kernel buffer */
1433 	WARN_ON_FPU(err);
1434 }
1435 
1436 #ifdef CONFIG_PROC_PID_ARCH_STATUS
1437 /*
1438  * Report the amount of time elapsed in millisecond since last AVX512
1439  * use in the task.
1440  */
1441 static void avx512_status(struct seq_file *m, struct task_struct *task)
1442 {
1443 	unsigned long timestamp = READ_ONCE(task->thread.fpu.avx512_timestamp);
1444 	long delta;
1445 
1446 	if (!timestamp) {
1447 		/*
1448 		 * Report -1 if no AVX512 usage
1449 		 */
1450 		delta = -1;
1451 	} else {
1452 		delta = (long)(jiffies - timestamp);
1453 		/*
1454 		 * Cap to LONG_MAX if time difference > LONG_MAX
1455 		 */
1456 		if (delta < 0)
1457 			delta = LONG_MAX;
1458 		delta = jiffies_to_msecs(delta);
1459 	}
1460 
1461 	seq_put_decimal_ll(m, "AVX512_elapsed_ms:\t", delta);
1462 	seq_putc(m, '\n');
1463 }
1464 
1465 /*
1466  * Report architecture specific information
1467  */
1468 int proc_pid_arch_status(struct seq_file *m, struct pid_namespace *ns,
1469 			struct pid *pid, struct task_struct *task)
1470 {
1471 	/*
1472 	 * Report AVX512 state if the processor and build option supported.
1473 	 */
1474 	if (cpu_feature_enabled(X86_FEATURE_AVX512F))
1475 		avx512_status(m, task);
1476 
1477 	return 0;
1478 }
1479 #endif /* CONFIG_PROC_PID_ARCH_STATUS */
1480