xref: /openbmc/linux/arch/x86/kernel/fpu/xstate.c (revision 6dfcd296)
1 /*
2  * xsave/xrstor support.
3  *
4  * Author: Suresh Siddha <suresh.b.siddha@intel.com>
5  */
6 #include <linux/compat.h>
7 #include <linux/cpu.h>
8 #include <linux/mman.h>
9 #include <linux/pkeys.h>
10 
11 #include <asm/fpu/api.h>
12 #include <asm/fpu/internal.h>
13 #include <asm/fpu/signal.h>
14 #include <asm/fpu/regset.h>
15 #include <asm/fpu/xstate.h>
16 
17 #include <asm/tlbflush.h>
18 
19 /*
20  * Although we spell it out in here, the Processor Trace
21  * xfeature is completely unused.  We use other mechanisms
22  * to save/restore PT state in Linux.
23  */
24 static const char *xfeature_names[] =
25 {
26 	"x87 floating point registers"	,
27 	"SSE registers"			,
28 	"AVX registers"			,
29 	"MPX bounds registers"		,
30 	"MPX CSR"			,
31 	"AVX-512 opmask"		,
32 	"AVX-512 Hi256"			,
33 	"AVX-512 ZMM_Hi256"		,
34 	"Processor Trace (unused)"	,
35 	"Protection Keys User registers",
36 	"unknown xstate feature"	,
37 };
38 
39 /*
40  * Mask of xstate features supported by the CPU and the kernel:
41  */
42 u64 xfeatures_mask __read_mostly;
43 
44 static unsigned int xstate_offsets[XFEATURE_MAX] = { [ 0 ... XFEATURE_MAX - 1] = -1};
45 static unsigned int xstate_sizes[XFEATURE_MAX]   = { [ 0 ... XFEATURE_MAX - 1] = -1};
46 static unsigned int xstate_comp_offsets[sizeof(xfeatures_mask)*8];
47 
48 /*
49  * The XSAVE area of kernel can be in standard or compacted format;
50  * it is always in standard format for user mode. This is the user
51  * mode standard format size used for signal and ptrace frames.
52  */
53 unsigned int fpu_user_xstate_size;
54 
55 /*
56  * Clear all of the X86_FEATURE_* bits that are unavailable
57  * when the CPU has no XSAVE support.
58  */
59 void fpu__xstate_clear_all_cpu_caps(void)
60 {
61 	setup_clear_cpu_cap(X86_FEATURE_XSAVE);
62 	setup_clear_cpu_cap(X86_FEATURE_XSAVEOPT);
63 	setup_clear_cpu_cap(X86_FEATURE_XSAVEC);
64 	setup_clear_cpu_cap(X86_FEATURE_XSAVES);
65 	setup_clear_cpu_cap(X86_FEATURE_AVX);
66 	setup_clear_cpu_cap(X86_FEATURE_AVX2);
67 	setup_clear_cpu_cap(X86_FEATURE_AVX512F);
68 	setup_clear_cpu_cap(X86_FEATURE_AVX512PF);
69 	setup_clear_cpu_cap(X86_FEATURE_AVX512ER);
70 	setup_clear_cpu_cap(X86_FEATURE_AVX512CD);
71 	setup_clear_cpu_cap(X86_FEATURE_AVX512DQ);
72 	setup_clear_cpu_cap(X86_FEATURE_AVX512BW);
73 	setup_clear_cpu_cap(X86_FEATURE_AVX512VL);
74 	setup_clear_cpu_cap(X86_FEATURE_MPX);
75 	setup_clear_cpu_cap(X86_FEATURE_XGETBV1);
76 	setup_clear_cpu_cap(X86_FEATURE_PKU);
77 }
78 
79 /*
80  * Return whether the system supports a given xfeature.
81  *
82  * Also return the name of the (most advanced) feature that the caller requested:
83  */
84 int cpu_has_xfeatures(u64 xfeatures_needed, const char **feature_name)
85 {
86 	u64 xfeatures_missing = xfeatures_needed & ~xfeatures_mask;
87 
88 	if (unlikely(feature_name)) {
89 		long xfeature_idx, max_idx;
90 		u64 xfeatures_print;
91 		/*
92 		 * So we use FLS here to be able to print the most advanced
93 		 * feature that was requested but is missing. So if a driver
94 		 * asks about "XFEATURE_MASK_SSE | XFEATURE_MASK_YMM" we'll print the
95 		 * missing AVX feature - this is the most informative message
96 		 * to users:
97 		 */
98 		if (xfeatures_missing)
99 			xfeatures_print = xfeatures_missing;
100 		else
101 			xfeatures_print = xfeatures_needed;
102 
103 		xfeature_idx = fls64(xfeatures_print)-1;
104 		max_idx = ARRAY_SIZE(xfeature_names)-1;
105 		xfeature_idx = min(xfeature_idx, max_idx);
106 
107 		*feature_name = xfeature_names[xfeature_idx];
108 	}
109 
110 	if (xfeatures_missing)
111 		return 0;
112 
113 	return 1;
114 }
115 EXPORT_SYMBOL_GPL(cpu_has_xfeatures);
116 
117 static int xfeature_is_supervisor(int xfeature_nr)
118 {
119 	/*
120 	 * We currently do not support supervisor states, but if
121 	 * we did, we could find out like this.
122 	 *
123 	 * SDM says: If state component 'i' is a user state component,
124 	 * ECX[0] return 0; if state component i is a supervisor
125 	 * state component, ECX[0] returns 1.
126 	 */
127 	u32 eax, ebx, ecx, edx;
128 
129 	cpuid_count(XSTATE_CPUID, xfeature_nr, &eax, &ebx, &ecx, &edx);
130 	return !!(ecx & 1);
131 }
132 
133 static int xfeature_is_user(int xfeature_nr)
134 {
135 	return !xfeature_is_supervisor(xfeature_nr);
136 }
137 
138 /*
139  * When executing XSAVEOPT (or other optimized XSAVE instructions), if
140  * a processor implementation detects that an FPU state component is still
141  * (or is again) in its initialized state, it may clear the corresponding
142  * bit in the header.xfeatures field, and can skip the writeout of registers
143  * to the corresponding memory layout.
144  *
145  * This means that when the bit is zero, the state component might still contain
146  * some previous - non-initialized register state.
147  *
148  * Before writing xstate information to user-space we sanitize those components,
149  * to always ensure that the memory layout of a feature will be in the init state
150  * if the corresponding header bit is zero. This is to ensure that user-space doesn't
151  * see some stale state in the memory layout during signal handling, debugging etc.
152  */
153 void fpstate_sanitize_xstate(struct fpu *fpu)
154 {
155 	struct fxregs_state *fx = &fpu->state.fxsave;
156 	int feature_bit;
157 	u64 xfeatures;
158 
159 	if (!use_xsaveopt())
160 		return;
161 
162 	xfeatures = fpu->state.xsave.header.xfeatures;
163 
164 	/*
165 	 * None of the feature bits are in init state. So nothing else
166 	 * to do for us, as the memory layout is up to date.
167 	 */
168 	if ((xfeatures & xfeatures_mask) == xfeatures_mask)
169 		return;
170 
171 	/*
172 	 * FP is in init state
173 	 */
174 	if (!(xfeatures & XFEATURE_MASK_FP)) {
175 		fx->cwd = 0x37f;
176 		fx->swd = 0;
177 		fx->twd = 0;
178 		fx->fop = 0;
179 		fx->rip = 0;
180 		fx->rdp = 0;
181 		memset(&fx->st_space[0], 0, 128);
182 	}
183 
184 	/*
185 	 * SSE is in init state
186 	 */
187 	if (!(xfeatures & XFEATURE_MASK_SSE))
188 		memset(&fx->xmm_space[0], 0, 256);
189 
190 	/*
191 	 * First two features are FPU and SSE, which above we handled
192 	 * in a special way already:
193 	 */
194 	feature_bit = 0x2;
195 	xfeatures = (xfeatures_mask & ~xfeatures) >> 2;
196 
197 	/*
198 	 * Update all the remaining memory layouts according to their
199 	 * standard xstate layout, if their header bit is in the init
200 	 * state:
201 	 */
202 	while (xfeatures) {
203 		if (xfeatures & 0x1) {
204 			int offset = xstate_comp_offsets[feature_bit];
205 			int size = xstate_sizes[feature_bit];
206 
207 			memcpy((void *)fx + offset,
208 			       (void *)&init_fpstate.xsave + offset,
209 			       size);
210 		}
211 
212 		xfeatures >>= 1;
213 		feature_bit++;
214 	}
215 }
216 
217 /*
218  * Enable the extended processor state save/restore feature.
219  * Called once per CPU onlining.
220  */
221 void fpu__init_cpu_xstate(void)
222 {
223 	if (!boot_cpu_has(X86_FEATURE_XSAVE) || !xfeatures_mask)
224 		return;
225 	/*
226 	 * Make it clear that XSAVES supervisor states are not yet
227 	 * implemented should anyone expect it to work by changing
228 	 * bits in XFEATURE_MASK_* macros and XCR0.
229 	 */
230 	WARN_ONCE((xfeatures_mask & XFEATURE_MASK_SUPERVISOR),
231 		"x86/fpu: XSAVES supervisor states are not yet implemented.\n");
232 
233 	xfeatures_mask &= ~XFEATURE_MASK_SUPERVISOR;
234 
235 	cr4_set_bits(X86_CR4_OSXSAVE);
236 	xsetbv(XCR_XFEATURE_ENABLED_MASK, xfeatures_mask);
237 }
238 
239 /*
240  * Note that in the future we will likely need a pair of
241  * functions here: one for user xstates and the other for
242  * system xstates.  For now, they are the same.
243  */
244 static int xfeature_enabled(enum xfeature xfeature)
245 {
246 	return !!(xfeatures_mask & (1UL << xfeature));
247 }
248 
249 /*
250  * Record the offsets and sizes of various xstates contained
251  * in the XSAVE state memory layout.
252  */
253 static void __init setup_xstate_features(void)
254 {
255 	u32 eax, ebx, ecx, edx, i;
256 	/* start at the beginnning of the "extended state" */
257 	unsigned int last_good_offset = offsetof(struct xregs_state,
258 						 extended_state_area);
259 	/*
260 	 * The FP xstates and SSE xstates are legacy states. They are always
261 	 * in the fixed offsets in the xsave area in either compacted form
262 	 * or standard form.
263 	 */
264 	xstate_offsets[0] = 0;
265 	xstate_sizes[0] = offsetof(struct fxregs_state, xmm_space);
266 	xstate_offsets[1] = xstate_sizes[0];
267 	xstate_sizes[1] = FIELD_SIZEOF(struct fxregs_state, xmm_space);
268 
269 	for (i = FIRST_EXTENDED_XFEATURE; i < XFEATURE_MAX; i++) {
270 		if (!xfeature_enabled(i))
271 			continue;
272 
273 		cpuid_count(XSTATE_CPUID, i, &eax, &ebx, &ecx, &edx);
274 
275 		/*
276 		 * If an xfeature is supervisor state, the offset
277 		 * in EBX is invalid. We leave it to -1.
278 		 */
279 		if (xfeature_is_user(i))
280 			xstate_offsets[i] = ebx;
281 
282 		xstate_sizes[i] = eax;
283 		/*
284 		 * In our xstate size checks, we assume that the
285 		 * highest-numbered xstate feature has the
286 		 * highest offset in the buffer.  Ensure it does.
287 		 */
288 		WARN_ONCE(last_good_offset > xstate_offsets[i],
289 			"x86/fpu: misordered xstate at %d\n", last_good_offset);
290 		last_good_offset = xstate_offsets[i];
291 	}
292 }
293 
294 static void __init print_xstate_feature(u64 xstate_mask)
295 {
296 	const char *feature_name;
297 
298 	if (cpu_has_xfeatures(xstate_mask, &feature_name))
299 		pr_info("x86/fpu: Supporting XSAVE feature 0x%03Lx: '%s'\n", xstate_mask, feature_name);
300 }
301 
302 /*
303  * Print out all the supported xstate features:
304  */
305 static void __init print_xstate_features(void)
306 {
307 	print_xstate_feature(XFEATURE_MASK_FP);
308 	print_xstate_feature(XFEATURE_MASK_SSE);
309 	print_xstate_feature(XFEATURE_MASK_YMM);
310 	print_xstate_feature(XFEATURE_MASK_BNDREGS);
311 	print_xstate_feature(XFEATURE_MASK_BNDCSR);
312 	print_xstate_feature(XFEATURE_MASK_OPMASK);
313 	print_xstate_feature(XFEATURE_MASK_ZMM_Hi256);
314 	print_xstate_feature(XFEATURE_MASK_Hi16_ZMM);
315 	print_xstate_feature(XFEATURE_MASK_PKRU);
316 }
317 
318 /*
319  * This check is important because it is easy to get XSTATE_*
320  * confused with XSTATE_BIT_*.
321  */
322 #define CHECK_XFEATURE(nr) do {		\
323 	WARN_ON(nr < FIRST_EXTENDED_XFEATURE);	\
324 	WARN_ON(nr >= XFEATURE_MAX);	\
325 } while (0)
326 
327 /*
328  * We could cache this like xstate_size[], but we only use
329  * it here, so it would be a waste of space.
330  */
331 static int xfeature_is_aligned(int xfeature_nr)
332 {
333 	u32 eax, ebx, ecx, edx;
334 
335 	CHECK_XFEATURE(xfeature_nr);
336 	cpuid_count(XSTATE_CPUID, xfeature_nr, &eax, &ebx, &ecx, &edx);
337 	/*
338 	 * The value returned by ECX[1] indicates the alignment
339 	 * of state component 'i' when the compacted format
340 	 * of the extended region of an XSAVE area is used:
341 	 */
342 	return !!(ecx & 2);
343 }
344 
345 /*
346  * This function sets up offsets and sizes of all extended states in
347  * xsave area. This supports both standard format and compacted format
348  * of the xsave aread.
349  */
350 static void __init setup_xstate_comp(void)
351 {
352 	unsigned int xstate_comp_sizes[sizeof(xfeatures_mask)*8];
353 	int i;
354 
355 	/*
356 	 * The FP xstates and SSE xstates are legacy states. They are always
357 	 * in the fixed offsets in the xsave area in either compacted form
358 	 * or standard form.
359 	 */
360 	xstate_comp_offsets[0] = 0;
361 	xstate_comp_offsets[1] = offsetof(struct fxregs_state, xmm_space);
362 
363 	if (!boot_cpu_has(X86_FEATURE_XSAVES)) {
364 		for (i = FIRST_EXTENDED_XFEATURE; i < XFEATURE_MAX; i++) {
365 			if (xfeature_enabled(i)) {
366 				xstate_comp_offsets[i] = xstate_offsets[i];
367 				xstate_comp_sizes[i] = xstate_sizes[i];
368 			}
369 		}
370 		return;
371 	}
372 
373 	xstate_comp_offsets[FIRST_EXTENDED_XFEATURE] =
374 		FXSAVE_SIZE + XSAVE_HDR_SIZE;
375 
376 	for (i = FIRST_EXTENDED_XFEATURE; i < XFEATURE_MAX; i++) {
377 		if (xfeature_enabled(i))
378 			xstate_comp_sizes[i] = xstate_sizes[i];
379 		else
380 			xstate_comp_sizes[i] = 0;
381 
382 		if (i > FIRST_EXTENDED_XFEATURE) {
383 			xstate_comp_offsets[i] = xstate_comp_offsets[i-1]
384 					+ xstate_comp_sizes[i-1];
385 
386 			if (xfeature_is_aligned(i))
387 				xstate_comp_offsets[i] =
388 					ALIGN(xstate_comp_offsets[i], 64);
389 		}
390 	}
391 }
392 
393 /*
394  * Print out xstate component offsets and sizes
395  */
396 static void __init print_xstate_offset_size(void)
397 {
398 	int i;
399 
400 	for (i = FIRST_EXTENDED_XFEATURE; i < XFEATURE_MAX; i++) {
401 		if (!xfeature_enabled(i))
402 			continue;
403 		pr_info("x86/fpu: xstate_offset[%d]: %4d, xstate_sizes[%d]: %4d\n",
404 			 i, xstate_comp_offsets[i], i, xstate_sizes[i]);
405 	}
406 }
407 
408 /*
409  * setup the xstate image representing the init state
410  */
411 static void __init setup_init_fpu_buf(void)
412 {
413 	static int on_boot_cpu __initdata = 1;
414 
415 	WARN_ON_FPU(!on_boot_cpu);
416 	on_boot_cpu = 0;
417 
418 	if (!boot_cpu_has(X86_FEATURE_XSAVE))
419 		return;
420 
421 	setup_xstate_features();
422 	print_xstate_features();
423 
424 	if (boot_cpu_has(X86_FEATURE_XSAVES))
425 		init_fpstate.xsave.header.xcomp_bv = (u64)1 << 63 | xfeatures_mask;
426 
427 	/*
428 	 * Init all the features state with header.xfeatures being 0x0
429 	 */
430 	copy_kernel_to_xregs_booting(&init_fpstate.xsave);
431 
432 	/*
433 	 * Dump the init state again. This is to identify the init state
434 	 * of any feature which is not represented by all zero's.
435 	 */
436 	copy_xregs_to_kernel_booting(&init_fpstate.xsave);
437 }
438 
439 static int xfeature_uncompacted_offset(int xfeature_nr)
440 {
441 	u32 eax, ebx, ecx, edx;
442 
443 	/*
444 	 * Only XSAVES supports supervisor states and it uses compacted
445 	 * format. Checking a supervisor state's uncompacted offset is
446 	 * an error.
447 	 */
448 	if (XFEATURE_MASK_SUPERVISOR & (1 << xfeature_nr)) {
449 		WARN_ONCE(1, "No fixed offset for xstate %d\n", xfeature_nr);
450 		return -1;
451 	}
452 
453 	CHECK_XFEATURE(xfeature_nr);
454 	cpuid_count(XSTATE_CPUID, xfeature_nr, &eax, &ebx, &ecx, &edx);
455 	return ebx;
456 }
457 
458 static int xfeature_size(int xfeature_nr)
459 {
460 	u32 eax, ebx, ecx, edx;
461 
462 	CHECK_XFEATURE(xfeature_nr);
463 	cpuid_count(XSTATE_CPUID, xfeature_nr, &eax, &ebx, &ecx, &edx);
464 	return eax;
465 }
466 
467 /*
468  * 'XSAVES' implies two different things:
469  * 1. saving of supervisor/system state
470  * 2. using the compacted format
471  *
472  * Use this function when dealing with the compacted format so
473  * that it is obvious which aspect of 'XSAVES' is being handled
474  * by the calling code.
475  */
476 int using_compacted_format(void)
477 {
478 	return boot_cpu_has(X86_FEATURE_XSAVES);
479 }
480 
481 static void __xstate_dump_leaves(void)
482 {
483 	int i;
484 	u32 eax, ebx, ecx, edx;
485 	static int should_dump = 1;
486 
487 	if (!should_dump)
488 		return;
489 	should_dump = 0;
490 	/*
491 	 * Dump out a few leaves past the ones that we support
492 	 * just in case there are some goodies up there
493 	 */
494 	for (i = 0; i < XFEATURE_MAX + 10; i++) {
495 		cpuid_count(XSTATE_CPUID, i, &eax, &ebx, &ecx, &edx);
496 		pr_warn("CPUID[%02x, %02x]: eax=%08x ebx=%08x ecx=%08x edx=%08x\n",
497 			XSTATE_CPUID, i, eax, ebx, ecx, edx);
498 	}
499 }
500 
501 #define XSTATE_WARN_ON(x) do {							\
502 	if (WARN_ONCE(x, "XSAVE consistency problem, dumping leaves")) {	\
503 		__xstate_dump_leaves();						\
504 	}									\
505 } while (0)
506 
507 #define XCHECK_SZ(sz, nr, nr_macro, __struct) do {			\
508 	if ((nr == nr_macro) &&						\
509 	    WARN_ONCE(sz != sizeof(__struct),				\
510 		"%s: struct is %zu bytes, cpu state %d bytes\n",	\
511 		__stringify(nr_macro), sizeof(__struct), sz)) {		\
512 		__xstate_dump_leaves();					\
513 	}								\
514 } while (0)
515 
516 /*
517  * We have a C struct for each 'xstate'.  We need to ensure
518  * that our software representation matches what the CPU
519  * tells us about the state's size.
520  */
521 static void check_xstate_against_struct(int nr)
522 {
523 	/*
524 	 * Ask the CPU for the size of the state.
525 	 */
526 	int sz = xfeature_size(nr);
527 	/*
528 	 * Match each CPU state with the corresponding software
529 	 * structure.
530 	 */
531 	XCHECK_SZ(sz, nr, XFEATURE_YMM,       struct ymmh_struct);
532 	XCHECK_SZ(sz, nr, XFEATURE_BNDREGS,   struct mpx_bndreg_state);
533 	XCHECK_SZ(sz, nr, XFEATURE_BNDCSR,    struct mpx_bndcsr_state);
534 	XCHECK_SZ(sz, nr, XFEATURE_OPMASK,    struct avx_512_opmask_state);
535 	XCHECK_SZ(sz, nr, XFEATURE_ZMM_Hi256, struct avx_512_zmm_uppers_state);
536 	XCHECK_SZ(sz, nr, XFEATURE_Hi16_ZMM,  struct avx_512_hi16_state);
537 	XCHECK_SZ(sz, nr, XFEATURE_PKRU,      struct pkru_state);
538 
539 	/*
540 	 * Make *SURE* to add any feature numbers in below if
541 	 * there are "holes" in the xsave state component
542 	 * numbers.
543 	 */
544 	if ((nr < XFEATURE_YMM) ||
545 	    (nr >= XFEATURE_MAX) ||
546 	    (nr == XFEATURE_PT_UNIMPLEMENTED_SO_FAR)) {
547 		WARN_ONCE(1, "no structure for xstate: %d\n", nr);
548 		XSTATE_WARN_ON(1);
549 	}
550 }
551 
552 /*
553  * This essentially double-checks what the cpu told us about
554  * how large the XSAVE buffer needs to be.  We are recalculating
555  * it to be safe.
556  */
557 static void do_extra_xstate_size_checks(void)
558 {
559 	int paranoid_xstate_size = FXSAVE_SIZE + XSAVE_HDR_SIZE;
560 	int i;
561 
562 	for (i = FIRST_EXTENDED_XFEATURE; i < XFEATURE_MAX; i++) {
563 		if (!xfeature_enabled(i))
564 			continue;
565 
566 		check_xstate_against_struct(i);
567 		/*
568 		 * Supervisor state components can be managed only by
569 		 * XSAVES, which is compacted-format only.
570 		 */
571 		if (!using_compacted_format())
572 			XSTATE_WARN_ON(xfeature_is_supervisor(i));
573 
574 		/* Align from the end of the previous feature */
575 		if (xfeature_is_aligned(i))
576 			paranoid_xstate_size = ALIGN(paranoid_xstate_size, 64);
577 		/*
578 		 * The offset of a given state in the non-compacted
579 		 * format is given to us in a CPUID leaf.  We check
580 		 * them for being ordered (increasing offsets) in
581 		 * setup_xstate_features().
582 		 */
583 		if (!using_compacted_format())
584 			paranoid_xstate_size = xfeature_uncompacted_offset(i);
585 		/*
586 		 * The compacted-format offset always depends on where
587 		 * the previous state ended.
588 		 */
589 		paranoid_xstate_size += xfeature_size(i);
590 	}
591 	XSTATE_WARN_ON(paranoid_xstate_size != fpu_kernel_xstate_size);
592 }
593 
594 
595 /*
596  * Get total size of enabled xstates in XCR0/xfeatures_mask.
597  *
598  * Note the SDM's wording here.  "sub-function 0" only enumerates
599  * the size of the *user* states.  If we use it to size a buffer
600  * that we use 'XSAVES' on, we could potentially overflow the
601  * buffer because 'XSAVES' saves system states too.
602  *
603  * Note that we do not currently set any bits on IA32_XSS so
604  * 'XCR0 | IA32_XSS == XCR0' for now.
605  */
606 static unsigned int __init get_xsaves_size(void)
607 {
608 	unsigned int eax, ebx, ecx, edx;
609 	/*
610 	 * - CPUID function 0DH, sub-function 1:
611 	 *    EBX enumerates the size (in bytes) required by
612 	 *    the XSAVES instruction for an XSAVE area
613 	 *    containing all the state components
614 	 *    corresponding to bits currently set in
615 	 *    XCR0 | IA32_XSS.
616 	 */
617 	cpuid_count(XSTATE_CPUID, 1, &eax, &ebx, &ecx, &edx);
618 	return ebx;
619 }
620 
621 static unsigned int __init get_xsave_size(void)
622 {
623 	unsigned int eax, ebx, ecx, edx;
624 	/*
625 	 * - CPUID function 0DH, sub-function 0:
626 	 *    EBX enumerates the size (in bytes) required by
627 	 *    the XSAVE instruction for an XSAVE area
628 	 *    containing all the *user* state components
629 	 *    corresponding to bits currently set in XCR0.
630 	 */
631 	cpuid_count(XSTATE_CPUID, 0, &eax, &ebx, &ecx, &edx);
632 	return ebx;
633 }
634 
635 /*
636  * Will the runtime-enumerated 'xstate_size' fit in the init
637  * task's statically-allocated buffer?
638  */
639 static bool is_supported_xstate_size(unsigned int test_xstate_size)
640 {
641 	if (test_xstate_size <= sizeof(union fpregs_state))
642 		return true;
643 
644 	pr_warn("x86/fpu: xstate buffer too small (%zu < %d), disabling xsave\n",
645 			sizeof(union fpregs_state), test_xstate_size);
646 	return false;
647 }
648 
649 static int init_xstate_size(void)
650 {
651 	/* Recompute the context size for enabled features: */
652 	unsigned int possible_xstate_size;
653 	unsigned int xsave_size;
654 
655 	xsave_size = get_xsave_size();
656 
657 	if (boot_cpu_has(X86_FEATURE_XSAVES))
658 		possible_xstate_size = get_xsaves_size();
659 	else
660 		possible_xstate_size = xsave_size;
661 
662 	/* Ensure we have the space to store all enabled: */
663 	if (!is_supported_xstate_size(possible_xstate_size))
664 		return -EINVAL;
665 
666 	/*
667 	 * The size is OK, we are definitely going to use xsave,
668 	 * make it known to the world that we need more space.
669 	 */
670 	fpu_kernel_xstate_size = possible_xstate_size;
671 	do_extra_xstate_size_checks();
672 
673 	/*
674 	 * User space is always in standard format.
675 	 */
676 	fpu_user_xstate_size = xsave_size;
677 	return 0;
678 }
679 
680 /*
681  * We enabled the XSAVE hardware, but something went wrong and
682  * we can not use it.  Disable it.
683  */
684 static void fpu__init_disable_system_xstate(void)
685 {
686 	xfeatures_mask = 0;
687 	cr4_clear_bits(X86_CR4_OSXSAVE);
688 	fpu__xstate_clear_all_cpu_caps();
689 }
690 
691 /*
692  * Enable and initialize the xsave feature.
693  * Called once per system bootup.
694  */
695 void __init fpu__init_system_xstate(void)
696 {
697 	unsigned int eax, ebx, ecx, edx;
698 	static int on_boot_cpu __initdata = 1;
699 	int err;
700 
701 	WARN_ON_FPU(!on_boot_cpu);
702 	on_boot_cpu = 0;
703 
704 	if (!boot_cpu_has(X86_FEATURE_XSAVE)) {
705 		pr_info("x86/fpu: Legacy x87 FPU detected.\n");
706 		return;
707 	}
708 
709 	if (boot_cpu_data.cpuid_level < XSTATE_CPUID) {
710 		WARN_ON_FPU(1);
711 		return;
712 	}
713 
714 	cpuid_count(XSTATE_CPUID, 0, &eax, &ebx, &ecx, &edx);
715 	xfeatures_mask = eax + ((u64)edx << 32);
716 
717 	if ((xfeatures_mask & XFEATURE_MASK_FPSSE) != XFEATURE_MASK_FPSSE) {
718 		/*
719 		 * This indicates that something really unexpected happened
720 		 * with the enumeration.  Disable XSAVE and try to continue
721 		 * booting without it.  This is too early to BUG().
722 		 */
723 		pr_err("x86/fpu: FP/SSE not present amongst the CPU's xstate features: 0x%llx.\n", xfeatures_mask);
724 		goto out_disable;
725 	}
726 
727 	xfeatures_mask &= fpu__get_supported_xfeatures_mask();
728 
729 	/* Enable xstate instructions to be able to continue with initialization: */
730 	fpu__init_cpu_xstate();
731 	err = init_xstate_size();
732 	if (err)
733 		goto out_disable;
734 
735 	/*
736 	 * Update info used for ptrace frames; use standard-format size and no
737 	 * supervisor xstates:
738 	 */
739 	update_regset_xstate_info(fpu_user_xstate_size,	xfeatures_mask & ~XFEATURE_MASK_SUPERVISOR);
740 
741 	fpu__init_prepare_fx_sw_frame();
742 	setup_init_fpu_buf();
743 	setup_xstate_comp();
744 	print_xstate_offset_size();
745 
746 	pr_info("x86/fpu: Enabled xstate features 0x%llx, context size is %d bytes, using '%s' format.\n",
747 		xfeatures_mask,
748 		fpu_kernel_xstate_size,
749 		boot_cpu_has(X86_FEATURE_XSAVES) ? "compacted" : "standard");
750 	return;
751 
752 out_disable:
753 	/* something went wrong, try to boot without any XSAVE support */
754 	fpu__init_disable_system_xstate();
755 }
756 
757 /*
758  * Restore minimal FPU state after suspend:
759  */
760 void fpu__resume_cpu(void)
761 {
762 	/*
763 	 * Restore XCR0 on xsave capable CPUs:
764 	 */
765 	if (boot_cpu_has(X86_FEATURE_XSAVE))
766 		xsetbv(XCR_XFEATURE_ENABLED_MASK, xfeatures_mask);
767 }
768 
769 /*
770  * Given an xstate feature mask, calculate where in the xsave
771  * buffer the state is.  Callers should ensure that the buffer
772  * is valid.
773  *
774  * Note: does not work for compacted buffers.
775  */
776 void *__raw_xsave_addr(struct xregs_state *xsave, int xstate_feature_mask)
777 {
778 	int feature_nr = fls64(xstate_feature_mask) - 1;
779 
780 	if (!xfeature_enabled(feature_nr)) {
781 		WARN_ON_FPU(1);
782 		return NULL;
783 	}
784 
785 	return (void *)xsave + xstate_comp_offsets[feature_nr];
786 }
787 /*
788  * Given the xsave area and a state inside, this function returns the
789  * address of the state.
790  *
791  * This is the API that is called to get xstate address in either
792  * standard format or compacted format of xsave area.
793  *
794  * Note that if there is no data for the field in the xsave buffer
795  * this will return NULL.
796  *
797  * Inputs:
798  *	xstate: the thread's storage area for all FPU data
799  *	xstate_feature: state which is defined in xsave.h (e.g.
800  *	XFEATURE_MASK_FP, XFEATURE_MASK_SSE, etc...)
801  * Output:
802  *	address of the state in the xsave area, or NULL if the
803  *	field is not present in the xsave buffer.
804  */
805 void *get_xsave_addr(struct xregs_state *xsave, int xstate_feature)
806 {
807 	/*
808 	 * Do we even *have* xsave state?
809 	 */
810 	if (!boot_cpu_has(X86_FEATURE_XSAVE))
811 		return NULL;
812 
813 	/*
814 	 * We should not ever be requesting features that we
815 	 * have not enabled.  Remember that pcntxt_mask is
816 	 * what we write to the XCR0 register.
817 	 */
818 	WARN_ONCE(!(xfeatures_mask & xstate_feature),
819 		  "get of unsupported state");
820 	/*
821 	 * This assumes the last 'xsave*' instruction to
822 	 * have requested that 'xstate_feature' be saved.
823 	 * If it did not, we might be seeing and old value
824 	 * of the field in the buffer.
825 	 *
826 	 * This can happen because the last 'xsave' did not
827 	 * request that this feature be saved (unlikely)
828 	 * or because the "init optimization" caused it
829 	 * to not be saved.
830 	 */
831 	if (!(xsave->header.xfeatures & xstate_feature))
832 		return NULL;
833 
834 	return __raw_xsave_addr(xsave, xstate_feature);
835 }
836 EXPORT_SYMBOL_GPL(get_xsave_addr);
837 
838 /*
839  * This wraps up the common operations that need to occur when retrieving
840  * data from xsave state.  It first ensures that the current task was
841  * using the FPU and retrieves the data in to a buffer.  It then calculates
842  * the offset of the requested field in the buffer.
843  *
844  * This function is safe to call whether the FPU is in use or not.
845  *
846  * Note that this only works on the current task.
847  *
848  * Inputs:
849  *	@xsave_state: state which is defined in xsave.h (e.g. XFEATURE_MASK_FP,
850  *	XFEATURE_MASK_SSE, etc...)
851  * Output:
852  *	address of the state in the xsave area or NULL if the state
853  *	is not present or is in its 'init state'.
854  */
855 const void *get_xsave_field_ptr(int xsave_state)
856 {
857 	struct fpu *fpu = &current->thread.fpu;
858 
859 	if (!fpu->fpstate_active)
860 		return NULL;
861 	/*
862 	 * fpu__save() takes the CPU's xstate registers
863 	 * and saves them off to the 'fpu memory buffer.
864 	 */
865 	fpu__save(fpu);
866 
867 	return get_xsave_addr(&fpu->state.xsave, xsave_state);
868 }
869 
870 #ifdef CONFIG_ARCH_HAS_PKEYS
871 
872 #define NR_VALID_PKRU_BITS (CONFIG_NR_PROTECTION_KEYS * 2)
873 #define PKRU_VALID_MASK (NR_VALID_PKRU_BITS - 1)
874 /*
875  * This will go out and modify PKRU register to set the access
876  * rights for @pkey to @init_val.
877  */
878 int arch_set_user_pkey_access(struct task_struct *tsk, int pkey,
879 		unsigned long init_val)
880 {
881 	u32 old_pkru;
882 	int pkey_shift = (pkey * PKRU_BITS_PER_PKEY);
883 	u32 new_pkru_bits = 0;
884 
885 	/*
886 	 * This check implies XSAVE support.  OSPKE only gets
887 	 * set if we enable XSAVE and we enable PKU in XCR0.
888 	 */
889 	if (!boot_cpu_has(X86_FEATURE_OSPKE))
890 		return -EINVAL;
891 	/*
892 	 * For most XSAVE components, this would be an arduous task:
893 	 * brining fpstate up to date with fpregs, updating fpstate,
894 	 * then re-populating fpregs.  But, for components that are
895 	 * never lazily managed, we can just access the fpregs
896 	 * directly.  PKRU is never managed lazily, so we can just
897 	 * manipulate it directly.  Make sure it stays that way.
898 	 */
899 	WARN_ON_ONCE(!use_eager_fpu());
900 
901 	/* Set the bits we need in PKRU:  */
902 	if (init_val & PKEY_DISABLE_ACCESS)
903 		new_pkru_bits |= PKRU_AD_BIT;
904 	if (init_val & PKEY_DISABLE_WRITE)
905 		new_pkru_bits |= PKRU_WD_BIT;
906 
907 	/* Shift the bits in to the correct place in PKRU for pkey: */
908 	new_pkru_bits <<= pkey_shift;
909 
910 	/* Get old PKRU and mask off any old bits in place: */
911 	old_pkru = read_pkru();
912 	old_pkru &= ~((PKRU_AD_BIT|PKRU_WD_BIT) << pkey_shift);
913 
914 	/* Write old part along with new part: */
915 	write_pkru(old_pkru | new_pkru_bits);
916 
917 	return 0;
918 }
919 #endif /* ! CONFIG_ARCH_HAS_PKEYS */
920 
921 /*
922  * This is similar to user_regset_copyout(), but will not add offset to
923  * the source data pointer or increment pos, count, kbuf, and ubuf.
924  */
925 static inline int xstate_copyout(unsigned int pos, unsigned int count,
926 				 void *kbuf, void __user *ubuf,
927 				 const void *data, const int start_pos,
928 				 const int end_pos)
929 {
930 	if ((count == 0) || (pos < start_pos))
931 		return 0;
932 
933 	if (end_pos < 0 || pos < end_pos) {
934 		unsigned int copy = (end_pos < 0 ? count : min(count, end_pos - pos));
935 
936 		if (kbuf) {
937 			memcpy(kbuf + pos, data, copy);
938 		} else {
939 			if (__copy_to_user(ubuf + pos, data, copy))
940 				return -EFAULT;
941 		}
942 	}
943 	return 0;
944 }
945 
946 /*
947  * Convert from kernel XSAVES compacted format to standard format and copy
948  * to a ptrace buffer. It supports partial copy but pos always starts from
949  * zero. This is called from xstateregs_get() and there we check the CPU
950  * has XSAVES.
951  */
952 int copyout_from_xsaves(unsigned int pos, unsigned int count, void *kbuf,
953 			void __user *ubuf, struct xregs_state *xsave)
954 {
955 	unsigned int offset, size;
956 	int ret, i;
957 	struct xstate_header header;
958 
959 	/*
960 	 * Currently copy_regset_to_user() starts from pos 0:
961 	 */
962 	if (unlikely(pos != 0))
963 		return -EFAULT;
964 
965 	/*
966 	 * The destination is a ptrace buffer; we put in only user xstates:
967 	 */
968 	memset(&header, 0, sizeof(header));
969 	header.xfeatures = xsave->header.xfeatures;
970 	header.xfeatures &= ~XFEATURE_MASK_SUPERVISOR;
971 
972 	/*
973 	 * Copy xregs_state->header:
974 	 */
975 	offset = offsetof(struct xregs_state, header);
976 	size = sizeof(header);
977 
978 	ret = xstate_copyout(offset, size, kbuf, ubuf, &header, 0, count);
979 
980 	if (ret)
981 		return ret;
982 
983 	for (i = 0; i < XFEATURE_MAX; i++) {
984 		/*
985 		 * Copy only in-use xstates:
986 		 */
987 		if ((header.xfeatures >> i) & 1) {
988 			void *src = __raw_xsave_addr(xsave, 1 << i);
989 
990 			offset = xstate_offsets[i];
991 			size = xstate_sizes[i];
992 
993 			ret = xstate_copyout(offset, size, kbuf, ubuf, src, 0, count);
994 
995 			if (ret)
996 				return ret;
997 
998 			if (offset + size >= count)
999 				break;
1000 		}
1001 
1002 	}
1003 
1004 	/*
1005 	 * Fill xsave->i387.sw_reserved value for ptrace frame:
1006 	 */
1007 	offset = offsetof(struct fxregs_state, sw_reserved);
1008 	size = sizeof(xstate_fx_sw_bytes);
1009 
1010 	ret = xstate_copyout(offset, size, kbuf, ubuf, xstate_fx_sw_bytes, 0, count);
1011 
1012 	if (ret)
1013 		return ret;
1014 
1015 	return 0;
1016 }
1017 
1018 /*
1019  * Convert from a ptrace standard-format buffer to kernel XSAVES format
1020  * and copy to the target thread. This is called from xstateregs_set() and
1021  * there we check the CPU has XSAVES and a whole standard-sized buffer
1022  * exists.
1023  */
1024 int copyin_to_xsaves(const void *kbuf, const void __user *ubuf,
1025 		     struct xregs_state *xsave)
1026 {
1027 	unsigned int offset, size;
1028 	int i;
1029 	u64 xfeatures;
1030 	u64 allowed_features;
1031 
1032 	offset = offsetof(struct xregs_state, header);
1033 	size = sizeof(xfeatures);
1034 
1035 	if (kbuf) {
1036 		memcpy(&xfeatures, kbuf + offset, size);
1037 	} else {
1038 		if (__copy_from_user(&xfeatures, ubuf + offset, size))
1039 			return -EFAULT;
1040 	}
1041 
1042 	/*
1043 	 * Reject if the user sets any disabled or supervisor features:
1044 	 */
1045 	allowed_features = xfeatures_mask & ~XFEATURE_MASK_SUPERVISOR;
1046 
1047 	if (xfeatures & ~allowed_features)
1048 		return -EINVAL;
1049 
1050 	for (i = 0; i < XFEATURE_MAX; i++) {
1051 		u64 mask = ((u64)1 << i);
1052 
1053 		if (xfeatures & mask) {
1054 			void *dst = __raw_xsave_addr(xsave, 1 << i);
1055 
1056 			offset = xstate_offsets[i];
1057 			size = xstate_sizes[i];
1058 
1059 			if (kbuf) {
1060 				memcpy(dst, kbuf + offset, size);
1061 			} else {
1062 				if (__copy_from_user(dst, ubuf + offset, size))
1063 					return -EFAULT;
1064 			}
1065 		}
1066 	}
1067 
1068 	/*
1069 	 * The state that came in from userspace was user-state only.
1070 	 * Mask all the user states out of 'xfeatures':
1071 	 */
1072 	xsave->header.xfeatures &= XFEATURE_MASK_SUPERVISOR;
1073 
1074 	/*
1075 	 * Add back in the features that came in from userspace:
1076 	 */
1077 	xsave->header.xfeatures |= xfeatures;
1078 
1079 	return 0;
1080 }
1081