xref: /openbmc/linux/arch/x86/kernel/fpu/xstate.c (revision f8e17c17)
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  * Mask of xstate features supported by the CPU and the kernel:
58  */
59 u64 xfeatures_mask __read_mostly;
60 
61 static unsigned int xstate_offsets[XFEATURE_MAX] = { [ 0 ... XFEATURE_MAX - 1] = -1};
62 static unsigned int xstate_sizes[XFEATURE_MAX]   = { [ 0 ... XFEATURE_MAX - 1] = -1};
63 static unsigned int xstate_comp_offsets[XFEATURE_MAX] = { [ 0 ... XFEATURE_MAX - 1] = -1};
64 
65 /*
66  * The XSAVE area of kernel can be in standard or compacted format;
67  * it is always in standard format for user mode. This is the user
68  * mode standard format size used for signal and ptrace frames.
69  */
70 unsigned int fpu_user_xstate_size;
71 
72 /*
73  * Return whether the system supports a given xfeature.
74  *
75  * Also return the name of the (most advanced) feature that the caller requested:
76  */
77 int cpu_has_xfeatures(u64 xfeatures_needed, const char **feature_name)
78 {
79 	u64 xfeatures_missing = xfeatures_needed & ~xfeatures_mask;
80 
81 	if (unlikely(feature_name)) {
82 		long xfeature_idx, max_idx;
83 		u64 xfeatures_print;
84 		/*
85 		 * So we use FLS here to be able to print the most advanced
86 		 * feature that was requested but is missing. So if a driver
87 		 * asks about "XFEATURE_MASK_SSE | XFEATURE_MASK_YMM" we'll print the
88 		 * missing AVX feature - this is the most informative message
89 		 * to users:
90 		 */
91 		if (xfeatures_missing)
92 			xfeatures_print = xfeatures_missing;
93 		else
94 			xfeatures_print = xfeatures_needed;
95 
96 		xfeature_idx = fls64(xfeatures_print)-1;
97 		max_idx = ARRAY_SIZE(xfeature_names)-1;
98 		xfeature_idx = min(xfeature_idx, max_idx);
99 
100 		*feature_name = xfeature_names[xfeature_idx];
101 	}
102 
103 	if (xfeatures_missing)
104 		return 0;
105 
106 	return 1;
107 }
108 EXPORT_SYMBOL_GPL(cpu_has_xfeatures);
109 
110 static bool xfeature_is_supervisor(int xfeature_nr)
111 {
112 	/*
113 	 * Extended State Enumeration Sub-leaves (EAX = 0DH, ECX = n, n > 1)
114 	 * returns ECX[0] set to (1) for a supervisor state, and cleared (0)
115 	 * for a user state.
116 	 */
117 	u32 eax, ebx, ecx, edx;
118 
119 	cpuid_count(XSTATE_CPUID, xfeature_nr, &eax, &ebx, &ecx, &edx);
120 	return ecx & 1;
121 }
122 
123 static bool xfeature_is_user(int xfeature_nr)
124 {
125 	return !xfeature_is_supervisor(xfeature_nr);
126 }
127 
128 /*
129  * When executing XSAVEOPT (or other optimized XSAVE instructions), if
130  * a processor implementation detects that an FPU state component is still
131  * (or is again) in its initialized state, it may clear the corresponding
132  * bit in the header.xfeatures field, and can skip the writeout of registers
133  * to the corresponding memory layout.
134  *
135  * This means that when the bit is zero, the state component might still contain
136  * some previous - non-initialized register state.
137  *
138  * Before writing xstate information to user-space we sanitize those components,
139  * to always ensure that the memory layout of a feature will be in the init state
140  * if the corresponding header bit is zero. This is to ensure that user-space doesn't
141  * see some stale state in the memory layout during signal handling, debugging etc.
142  */
143 void fpstate_sanitize_xstate(struct fpu *fpu)
144 {
145 	struct fxregs_state *fx = &fpu->state.fxsave;
146 	int feature_bit;
147 	u64 xfeatures;
148 
149 	if (!use_xsaveopt())
150 		return;
151 
152 	xfeatures = fpu->state.xsave.header.xfeatures;
153 
154 	/*
155 	 * None of the feature bits are in init state. So nothing else
156 	 * to do for us, as the memory layout is up to date.
157 	 */
158 	if ((xfeatures & xfeatures_mask) == xfeatures_mask)
159 		return;
160 
161 	/*
162 	 * FP is in init state
163 	 */
164 	if (!(xfeatures & XFEATURE_MASK_FP)) {
165 		fx->cwd = 0x37f;
166 		fx->swd = 0;
167 		fx->twd = 0;
168 		fx->fop = 0;
169 		fx->rip = 0;
170 		fx->rdp = 0;
171 		memset(&fx->st_space[0], 0, 128);
172 	}
173 
174 	/*
175 	 * SSE is in init state
176 	 */
177 	if (!(xfeatures & XFEATURE_MASK_SSE))
178 		memset(&fx->xmm_space[0], 0, 256);
179 
180 	/*
181 	 * First two features are FPU and SSE, which above we handled
182 	 * in a special way already:
183 	 */
184 	feature_bit = 0x2;
185 	xfeatures = (xfeatures_mask & ~xfeatures) >> 2;
186 
187 	/*
188 	 * Update all the remaining memory layouts according to their
189 	 * standard xstate layout, if their header bit is in the init
190 	 * state:
191 	 */
192 	while (xfeatures) {
193 		if (xfeatures & 0x1) {
194 			int offset = xstate_comp_offsets[feature_bit];
195 			int size = xstate_sizes[feature_bit];
196 
197 			memcpy((void *)fx + offset,
198 			       (void *)&init_fpstate.xsave + offset,
199 			       size);
200 		}
201 
202 		xfeatures >>= 1;
203 		feature_bit++;
204 	}
205 }
206 
207 /*
208  * Enable the extended processor state save/restore feature.
209  * Called once per CPU onlining.
210  */
211 void fpu__init_cpu_xstate(void)
212 {
213 	if (!boot_cpu_has(X86_FEATURE_XSAVE) || !xfeatures_mask)
214 		return;
215 	/*
216 	 * Make it clear that XSAVES supervisor states are not yet
217 	 * implemented should anyone expect it to work by changing
218 	 * bits in XFEATURE_MASK_* macros and XCR0.
219 	 */
220 	WARN_ONCE((xfeatures_mask & XFEATURE_MASK_SUPERVISOR),
221 		"x86/fpu: XSAVES supervisor states are not yet implemented.\n");
222 
223 	xfeatures_mask &= ~XFEATURE_MASK_SUPERVISOR;
224 
225 	cr4_set_bits(X86_CR4_OSXSAVE);
226 	xsetbv(XCR_XFEATURE_ENABLED_MASK, xfeatures_mask);
227 }
228 
229 /*
230  * Note that in the future we will likely need a pair of
231  * functions here: one for user xstates and the other for
232  * system xstates.  For now, they are the same.
233  */
234 static int xfeature_enabled(enum xfeature xfeature)
235 {
236 	return !!(xfeatures_mask & (1UL << xfeature));
237 }
238 
239 /*
240  * Record the offsets and sizes of various xstates contained
241  * in the XSAVE state memory layout.
242  */
243 static void __init setup_xstate_features(void)
244 {
245 	u32 eax, ebx, ecx, edx, i;
246 	/* start at the beginnning of the "extended state" */
247 	unsigned int last_good_offset = offsetof(struct xregs_state,
248 						 extended_state_area);
249 	/*
250 	 * The FP xstates and SSE xstates are legacy states. They are always
251 	 * in the fixed offsets in the xsave area in either compacted form
252 	 * or standard form.
253 	 */
254 	xstate_offsets[XFEATURE_FP]	= 0;
255 	xstate_sizes[XFEATURE_FP]	= offsetof(struct fxregs_state,
256 						   xmm_space);
257 
258 	xstate_offsets[XFEATURE_SSE]	= xstate_sizes[XFEATURE_FP];
259 	xstate_sizes[XFEATURE_SSE]	= sizeof_field(struct fxregs_state,
260 						       xmm_space);
261 
262 	for (i = FIRST_EXTENDED_XFEATURE; i < XFEATURE_MAX; i++) {
263 		if (!xfeature_enabled(i))
264 			continue;
265 
266 		cpuid_count(XSTATE_CPUID, i, &eax, &ebx, &ecx, &edx);
267 
268 		/*
269 		 * If an xfeature is supervisor state, the offset
270 		 * in EBX is invalid. We leave it to -1.
271 		 */
272 		if (xfeature_is_user(i))
273 			xstate_offsets[i] = ebx;
274 
275 		xstate_sizes[i] = eax;
276 		/*
277 		 * In our xstate size checks, we assume that the
278 		 * highest-numbered xstate feature has the
279 		 * highest offset in the buffer.  Ensure it does.
280 		 */
281 		WARN_ONCE(last_good_offset > xstate_offsets[i],
282 			"x86/fpu: misordered xstate at %d\n", last_good_offset);
283 		last_good_offset = xstate_offsets[i];
284 	}
285 }
286 
287 static void __init print_xstate_feature(u64 xstate_mask)
288 {
289 	const char *feature_name;
290 
291 	if (cpu_has_xfeatures(xstate_mask, &feature_name))
292 		pr_info("x86/fpu: Supporting XSAVE feature 0x%03Lx: '%s'\n", xstate_mask, feature_name);
293 }
294 
295 /*
296  * Print out all the supported xstate features:
297  */
298 static void __init print_xstate_features(void)
299 {
300 	print_xstate_feature(XFEATURE_MASK_FP);
301 	print_xstate_feature(XFEATURE_MASK_SSE);
302 	print_xstate_feature(XFEATURE_MASK_YMM);
303 	print_xstate_feature(XFEATURE_MASK_BNDREGS);
304 	print_xstate_feature(XFEATURE_MASK_BNDCSR);
305 	print_xstate_feature(XFEATURE_MASK_OPMASK);
306 	print_xstate_feature(XFEATURE_MASK_ZMM_Hi256);
307 	print_xstate_feature(XFEATURE_MASK_Hi16_ZMM);
308 	print_xstate_feature(XFEATURE_MASK_PKRU);
309 }
310 
311 /*
312  * This check is important because it is easy to get XSTATE_*
313  * confused with XSTATE_BIT_*.
314  */
315 #define CHECK_XFEATURE(nr) do {		\
316 	WARN_ON(nr < FIRST_EXTENDED_XFEATURE);	\
317 	WARN_ON(nr >= XFEATURE_MAX);	\
318 } while (0)
319 
320 /*
321  * We could cache this like xstate_size[], but we only use
322  * it here, so it would be a waste of space.
323  */
324 static int xfeature_is_aligned(int xfeature_nr)
325 {
326 	u32 eax, ebx, ecx, edx;
327 
328 	CHECK_XFEATURE(xfeature_nr);
329 	cpuid_count(XSTATE_CPUID, xfeature_nr, &eax, &ebx, &ecx, &edx);
330 	/*
331 	 * The value returned by ECX[1] indicates the alignment
332 	 * of state component 'i' when the compacted format
333 	 * of the extended region of an XSAVE area is used:
334 	 */
335 	return !!(ecx & 2);
336 }
337 
338 /*
339  * This function sets up offsets and sizes of all extended states in
340  * xsave area. This supports both standard format and compacted format
341  * of the xsave aread.
342  */
343 static void __init setup_xstate_comp(void)
344 {
345 	unsigned int xstate_comp_sizes[XFEATURE_MAX];
346 	int i;
347 
348 	/*
349 	 * The FP xstates and SSE xstates are legacy states. They are always
350 	 * in the fixed offsets in the xsave area in either compacted form
351 	 * or standard form.
352 	 */
353 	xstate_comp_offsets[XFEATURE_FP] = 0;
354 	xstate_comp_offsets[XFEATURE_SSE] = offsetof(struct fxregs_state,
355 						     xmm_space);
356 
357 	if (!boot_cpu_has(X86_FEATURE_XSAVES)) {
358 		for (i = FIRST_EXTENDED_XFEATURE; i < XFEATURE_MAX; i++) {
359 			if (xfeature_enabled(i)) {
360 				xstate_comp_offsets[i] = xstate_offsets[i];
361 				xstate_comp_sizes[i] = xstate_sizes[i];
362 			}
363 		}
364 		return;
365 	}
366 
367 	xstate_comp_offsets[FIRST_EXTENDED_XFEATURE] =
368 		FXSAVE_SIZE + XSAVE_HDR_SIZE;
369 
370 	for (i = FIRST_EXTENDED_XFEATURE; i < XFEATURE_MAX; i++) {
371 		if (xfeature_enabled(i))
372 			xstate_comp_sizes[i] = xstate_sizes[i];
373 		else
374 			xstate_comp_sizes[i] = 0;
375 
376 		if (i > FIRST_EXTENDED_XFEATURE) {
377 			xstate_comp_offsets[i] = xstate_comp_offsets[i-1]
378 					+ xstate_comp_sizes[i-1];
379 
380 			if (xfeature_is_aligned(i))
381 				xstate_comp_offsets[i] =
382 					ALIGN(xstate_comp_offsets[i], 64);
383 		}
384 	}
385 }
386 
387 /*
388  * Print out xstate component offsets and sizes
389  */
390 static void __init print_xstate_offset_size(void)
391 {
392 	int i;
393 
394 	for (i = FIRST_EXTENDED_XFEATURE; i < XFEATURE_MAX; i++) {
395 		if (!xfeature_enabled(i))
396 			continue;
397 		pr_info("x86/fpu: xstate_offset[%d]: %4d, xstate_sizes[%d]: %4d\n",
398 			 i, xstate_comp_offsets[i], i, xstate_sizes[i]);
399 	}
400 }
401 
402 /*
403  * setup the xstate image representing the init state
404  */
405 static void __init setup_init_fpu_buf(void)
406 {
407 	static int on_boot_cpu __initdata = 1;
408 
409 	WARN_ON_FPU(!on_boot_cpu);
410 	on_boot_cpu = 0;
411 
412 	if (!boot_cpu_has(X86_FEATURE_XSAVE))
413 		return;
414 
415 	setup_xstate_features();
416 	print_xstate_features();
417 
418 	if (boot_cpu_has(X86_FEATURE_XSAVES))
419 		init_fpstate.xsave.header.xcomp_bv = XCOMP_BV_COMPACTED_FORMAT |
420 						     xfeatures_mask;
421 
422 	/*
423 	 * Init all the features state with header.xfeatures being 0x0
424 	 */
425 	copy_kernel_to_xregs_booting(&init_fpstate.xsave);
426 
427 	/*
428 	 * Dump the init state again. This is to identify the init state
429 	 * of any feature which is not represented by all zero's.
430 	 */
431 	copy_xregs_to_kernel_booting(&init_fpstate.xsave);
432 }
433 
434 static int xfeature_uncompacted_offset(int xfeature_nr)
435 {
436 	u32 eax, ebx, ecx, edx;
437 
438 	/*
439 	 * Only XSAVES supports supervisor states and it uses compacted
440 	 * format. Checking a supervisor state's uncompacted offset is
441 	 * an error.
442 	 */
443 	if (XFEATURE_MASK_SUPERVISOR & BIT_ULL(xfeature_nr)) {
444 		WARN_ONCE(1, "No fixed offset for xstate %d\n", xfeature_nr);
445 		return -1;
446 	}
447 
448 	CHECK_XFEATURE(xfeature_nr);
449 	cpuid_count(XSTATE_CPUID, xfeature_nr, &eax, &ebx, &ecx, &edx);
450 	return ebx;
451 }
452 
453 static int xfeature_size(int xfeature_nr)
454 {
455 	u32 eax, ebx, ecx, edx;
456 
457 	CHECK_XFEATURE(xfeature_nr);
458 	cpuid_count(XSTATE_CPUID, xfeature_nr, &eax, &ebx, &ecx, &edx);
459 	return eax;
460 }
461 
462 /*
463  * 'XSAVES' implies two different things:
464  * 1. saving of supervisor/system state
465  * 2. using the compacted format
466  *
467  * Use this function when dealing with the compacted format so
468  * that it is obvious which aspect of 'XSAVES' is being handled
469  * by the calling code.
470  */
471 int using_compacted_format(void)
472 {
473 	return boot_cpu_has(X86_FEATURE_XSAVES);
474 }
475 
476 /* Validate an xstate header supplied by userspace (ptrace or sigreturn) */
477 int validate_xstate_header(const struct xstate_header *hdr)
478 {
479 	/* No unknown or supervisor features may be set */
480 	if (hdr->xfeatures & (~xfeatures_mask | XFEATURE_MASK_SUPERVISOR))
481 		return -EINVAL;
482 
483 	/* Userspace must use the uncompacted format */
484 	if (hdr->xcomp_bv)
485 		return -EINVAL;
486 
487 	/*
488 	 * If 'reserved' is shrunken to add a new field, make sure to validate
489 	 * that new field here!
490 	 */
491 	BUILD_BUG_ON(sizeof(hdr->reserved) != 48);
492 
493 	/* No reserved bits may be set */
494 	if (memchr_inv(hdr->reserved, 0, sizeof(hdr->reserved)))
495 		return -EINVAL;
496 
497 	return 0;
498 }
499 
500 static void __xstate_dump_leaves(void)
501 {
502 	int i;
503 	u32 eax, ebx, ecx, edx;
504 	static int should_dump = 1;
505 
506 	if (!should_dump)
507 		return;
508 	should_dump = 0;
509 	/*
510 	 * Dump out a few leaves past the ones that we support
511 	 * just in case there are some goodies up there
512 	 */
513 	for (i = 0; i < XFEATURE_MAX + 10; i++) {
514 		cpuid_count(XSTATE_CPUID, i, &eax, &ebx, &ecx, &edx);
515 		pr_warn("CPUID[%02x, %02x]: eax=%08x ebx=%08x ecx=%08x edx=%08x\n",
516 			XSTATE_CPUID, i, eax, ebx, ecx, edx);
517 	}
518 }
519 
520 #define XSTATE_WARN_ON(x) do {							\
521 	if (WARN_ONCE(x, "XSAVE consistency problem, dumping leaves")) {	\
522 		__xstate_dump_leaves();						\
523 	}									\
524 } while (0)
525 
526 #define XCHECK_SZ(sz, nr, nr_macro, __struct) do {			\
527 	if ((nr == nr_macro) &&						\
528 	    WARN_ONCE(sz != sizeof(__struct),				\
529 		"%s: struct is %zu bytes, cpu state %d bytes\n",	\
530 		__stringify(nr_macro), sizeof(__struct), sz)) {		\
531 		__xstate_dump_leaves();					\
532 	}								\
533 } while (0)
534 
535 /*
536  * We have a C struct for each 'xstate'.  We need to ensure
537  * that our software representation matches what the CPU
538  * tells us about the state's size.
539  */
540 static void check_xstate_against_struct(int nr)
541 {
542 	/*
543 	 * Ask the CPU for the size of the state.
544 	 */
545 	int sz = xfeature_size(nr);
546 	/*
547 	 * Match each CPU state with the corresponding software
548 	 * structure.
549 	 */
550 	XCHECK_SZ(sz, nr, XFEATURE_YMM,       struct ymmh_struct);
551 	XCHECK_SZ(sz, nr, XFEATURE_BNDREGS,   struct mpx_bndreg_state);
552 	XCHECK_SZ(sz, nr, XFEATURE_BNDCSR,    struct mpx_bndcsr_state);
553 	XCHECK_SZ(sz, nr, XFEATURE_OPMASK,    struct avx_512_opmask_state);
554 	XCHECK_SZ(sz, nr, XFEATURE_ZMM_Hi256, struct avx_512_zmm_uppers_state);
555 	XCHECK_SZ(sz, nr, XFEATURE_Hi16_ZMM,  struct avx_512_hi16_state);
556 	XCHECK_SZ(sz, nr, XFEATURE_PKRU,      struct pkru_state);
557 
558 	/*
559 	 * Make *SURE* to add any feature numbers in below if
560 	 * there are "holes" in the xsave state component
561 	 * numbers.
562 	 */
563 	if ((nr < XFEATURE_YMM) ||
564 	    (nr >= XFEATURE_MAX) ||
565 	    (nr == XFEATURE_PT_UNIMPLEMENTED_SO_FAR)) {
566 		WARN_ONCE(1, "no structure for xstate: %d\n", nr);
567 		XSTATE_WARN_ON(1);
568 	}
569 }
570 
571 /*
572  * This essentially double-checks what the cpu told us about
573  * how large the XSAVE buffer needs to be.  We are recalculating
574  * it to be safe.
575  */
576 static void do_extra_xstate_size_checks(void)
577 {
578 	int paranoid_xstate_size = FXSAVE_SIZE + XSAVE_HDR_SIZE;
579 	int i;
580 
581 	for (i = FIRST_EXTENDED_XFEATURE; i < XFEATURE_MAX; i++) {
582 		if (!xfeature_enabled(i))
583 			continue;
584 
585 		check_xstate_against_struct(i);
586 		/*
587 		 * Supervisor state components can be managed only by
588 		 * XSAVES, which is compacted-format only.
589 		 */
590 		if (!using_compacted_format())
591 			XSTATE_WARN_ON(xfeature_is_supervisor(i));
592 
593 		/* Align from the end of the previous feature */
594 		if (xfeature_is_aligned(i))
595 			paranoid_xstate_size = ALIGN(paranoid_xstate_size, 64);
596 		/*
597 		 * The offset of a given state in the non-compacted
598 		 * format is given to us in a CPUID leaf.  We check
599 		 * them for being ordered (increasing offsets) in
600 		 * setup_xstate_features().
601 		 */
602 		if (!using_compacted_format())
603 			paranoid_xstate_size = xfeature_uncompacted_offset(i);
604 		/*
605 		 * The compacted-format offset always depends on where
606 		 * the previous state ended.
607 		 */
608 		paranoid_xstate_size += xfeature_size(i);
609 	}
610 	XSTATE_WARN_ON(paranoid_xstate_size != fpu_kernel_xstate_size);
611 }
612 
613 
614 /*
615  * Get total size of enabled xstates in XCR0/xfeatures_mask.
616  *
617  * Note the SDM's wording here.  "sub-function 0" only enumerates
618  * the size of the *user* states.  If we use it to size a buffer
619  * that we use 'XSAVES' on, we could potentially overflow the
620  * buffer because 'XSAVES' saves system states too.
621  *
622  * Note that we do not currently set any bits on IA32_XSS so
623  * 'XCR0 | IA32_XSS == XCR0' for now.
624  */
625 static unsigned int __init get_xsaves_size(void)
626 {
627 	unsigned int eax, ebx, ecx, edx;
628 	/*
629 	 * - CPUID function 0DH, sub-function 1:
630 	 *    EBX enumerates the size (in bytes) required by
631 	 *    the XSAVES instruction for an XSAVE area
632 	 *    containing all the state components
633 	 *    corresponding to bits currently set in
634 	 *    XCR0 | IA32_XSS.
635 	 */
636 	cpuid_count(XSTATE_CPUID, 1, &eax, &ebx, &ecx, &edx);
637 	return ebx;
638 }
639 
640 static unsigned int __init get_xsave_size(void)
641 {
642 	unsigned int eax, ebx, ecx, edx;
643 	/*
644 	 * - CPUID function 0DH, sub-function 0:
645 	 *    EBX enumerates the size (in bytes) required by
646 	 *    the XSAVE instruction for an XSAVE area
647 	 *    containing all the *user* state components
648 	 *    corresponding to bits currently set in XCR0.
649 	 */
650 	cpuid_count(XSTATE_CPUID, 0, &eax, &ebx, &ecx, &edx);
651 	return ebx;
652 }
653 
654 /*
655  * Will the runtime-enumerated 'xstate_size' fit in the init
656  * task's statically-allocated buffer?
657  */
658 static bool is_supported_xstate_size(unsigned int test_xstate_size)
659 {
660 	if (test_xstate_size <= sizeof(union fpregs_state))
661 		return true;
662 
663 	pr_warn("x86/fpu: xstate buffer too small (%zu < %d), disabling xsave\n",
664 			sizeof(union fpregs_state), test_xstate_size);
665 	return false;
666 }
667 
668 static int __init init_xstate_size(void)
669 {
670 	/* Recompute the context size for enabled features: */
671 	unsigned int possible_xstate_size;
672 	unsigned int xsave_size;
673 
674 	xsave_size = get_xsave_size();
675 
676 	if (boot_cpu_has(X86_FEATURE_XSAVES))
677 		possible_xstate_size = get_xsaves_size();
678 	else
679 		possible_xstate_size = xsave_size;
680 
681 	/* Ensure we have the space to store all enabled: */
682 	if (!is_supported_xstate_size(possible_xstate_size))
683 		return -EINVAL;
684 
685 	/*
686 	 * The size is OK, we are definitely going to use xsave,
687 	 * make it known to the world that we need more space.
688 	 */
689 	fpu_kernel_xstate_size = possible_xstate_size;
690 	do_extra_xstate_size_checks();
691 
692 	/*
693 	 * User space is always in standard format.
694 	 */
695 	fpu_user_xstate_size = xsave_size;
696 	return 0;
697 }
698 
699 /*
700  * We enabled the XSAVE hardware, but something went wrong and
701  * we can not use it.  Disable it.
702  */
703 static void fpu__init_disable_system_xstate(void)
704 {
705 	xfeatures_mask = 0;
706 	cr4_clear_bits(X86_CR4_OSXSAVE);
707 	setup_clear_cpu_cap(X86_FEATURE_XSAVE);
708 }
709 
710 /*
711  * Enable and initialize the xsave feature.
712  * Called once per system bootup.
713  */
714 void __init fpu__init_system_xstate(void)
715 {
716 	unsigned int eax, ebx, ecx, edx;
717 	static int on_boot_cpu __initdata = 1;
718 	int err;
719 	int i;
720 
721 	WARN_ON_FPU(!on_boot_cpu);
722 	on_boot_cpu = 0;
723 
724 	if (!boot_cpu_has(X86_FEATURE_FPU)) {
725 		pr_info("x86/fpu: No FPU detected\n");
726 		return;
727 	}
728 
729 	if (!boot_cpu_has(X86_FEATURE_XSAVE)) {
730 		pr_info("x86/fpu: x87 FPU will use %s\n",
731 			boot_cpu_has(X86_FEATURE_FXSR) ? "FXSAVE" : "FSAVE");
732 		return;
733 	}
734 
735 	if (boot_cpu_data.cpuid_level < XSTATE_CPUID) {
736 		WARN_ON_FPU(1);
737 		return;
738 	}
739 
740 	cpuid_count(XSTATE_CPUID, 0, &eax, &ebx, &ecx, &edx);
741 	xfeatures_mask = eax + ((u64)edx << 32);
742 
743 	if ((xfeatures_mask & XFEATURE_MASK_FPSSE) != XFEATURE_MASK_FPSSE) {
744 		/*
745 		 * This indicates that something really unexpected happened
746 		 * with the enumeration.  Disable XSAVE and try to continue
747 		 * booting without it.  This is too early to BUG().
748 		 */
749 		pr_err("x86/fpu: FP/SSE not present amongst the CPU's xstate features: 0x%llx.\n", xfeatures_mask);
750 		goto out_disable;
751 	}
752 
753 	/*
754 	 * Clear XSAVE features that are disabled in the normal CPUID.
755 	 */
756 	for (i = 0; i < ARRAY_SIZE(xsave_cpuid_features); i++) {
757 		if (!boot_cpu_has(xsave_cpuid_features[i]))
758 			xfeatures_mask &= ~BIT(i);
759 	}
760 
761 	xfeatures_mask &= fpu__get_supported_xfeatures_mask();
762 
763 	/* Enable xstate instructions to be able to continue with initialization: */
764 	fpu__init_cpu_xstate();
765 	err = init_xstate_size();
766 	if (err)
767 		goto out_disable;
768 
769 	/*
770 	 * Update info used for ptrace frames; use standard-format size and no
771 	 * supervisor xstates:
772 	 */
773 	update_regset_xstate_info(fpu_user_xstate_size,	xfeatures_mask & ~XFEATURE_MASK_SUPERVISOR);
774 
775 	fpu__init_prepare_fx_sw_frame();
776 	setup_init_fpu_buf();
777 	setup_xstate_comp();
778 	print_xstate_offset_size();
779 
780 	pr_info("x86/fpu: Enabled xstate features 0x%llx, context size is %d bytes, using '%s' format.\n",
781 		xfeatures_mask,
782 		fpu_kernel_xstate_size,
783 		boot_cpu_has(X86_FEATURE_XSAVES) ? "compacted" : "standard");
784 	return;
785 
786 out_disable:
787 	/* something went wrong, try to boot without any XSAVE support */
788 	fpu__init_disable_system_xstate();
789 }
790 
791 /*
792  * Restore minimal FPU state after suspend:
793  */
794 void fpu__resume_cpu(void)
795 {
796 	/*
797 	 * Restore XCR0 on xsave capable CPUs:
798 	 */
799 	if (boot_cpu_has(X86_FEATURE_XSAVE))
800 		xsetbv(XCR_XFEATURE_ENABLED_MASK, xfeatures_mask);
801 }
802 
803 /*
804  * Given an xstate feature nr, calculate where in the xsave
805  * buffer the state is.  Callers should ensure that the buffer
806  * is valid.
807  */
808 static void *__raw_xsave_addr(struct xregs_state *xsave, int xfeature_nr)
809 {
810 	if (!xfeature_enabled(xfeature_nr)) {
811 		WARN_ON_FPU(1);
812 		return NULL;
813 	}
814 
815 	return (void *)xsave + xstate_comp_offsets[xfeature_nr];
816 }
817 /*
818  * Given the xsave area and a state inside, this function returns the
819  * address of the state.
820  *
821  * This is the API that is called to get xstate address in either
822  * standard format or compacted format of xsave area.
823  *
824  * Note that if there is no data for the field in the xsave buffer
825  * this will return NULL.
826  *
827  * Inputs:
828  *	xstate: the thread's storage area for all FPU data
829  *	xfeature_nr: state which is defined in xsave.h (e.g. XFEATURE_FP,
830  *	XFEATURE_SSE, etc...)
831  * Output:
832  *	address of the state in the xsave area, or NULL if the
833  *	field is not present in the xsave buffer.
834  */
835 void *get_xsave_addr(struct xregs_state *xsave, int xfeature_nr)
836 {
837 	/*
838 	 * Do we even *have* xsave state?
839 	 */
840 	if (!boot_cpu_has(X86_FEATURE_XSAVE))
841 		return NULL;
842 
843 	/*
844 	 * We should not ever be requesting features that we
845 	 * have not enabled.  Remember that xfeatures_mask is
846 	 * what we write to the XCR0 register.
847 	 */
848 	WARN_ONCE(!(xfeatures_mask & BIT_ULL(xfeature_nr)),
849 		  "get of unsupported state");
850 	/*
851 	 * This assumes the last 'xsave*' instruction to
852 	 * have requested that 'xfeature_nr' be saved.
853 	 * If it did not, we might be seeing and old value
854 	 * of the field in the buffer.
855 	 *
856 	 * This can happen because the last 'xsave' did not
857 	 * request that this feature be saved (unlikely)
858 	 * or because the "init optimization" caused it
859 	 * to not be saved.
860 	 */
861 	if (!(xsave->header.xfeatures & BIT_ULL(xfeature_nr)))
862 		return NULL;
863 
864 	return __raw_xsave_addr(xsave, xfeature_nr);
865 }
866 EXPORT_SYMBOL_GPL(get_xsave_addr);
867 
868 /*
869  * This wraps up the common operations that need to occur when retrieving
870  * data from xsave state.  It first ensures that the current task was
871  * using the FPU and retrieves the data in to a buffer.  It then calculates
872  * the offset of the requested field in the buffer.
873  *
874  * This function is safe to call whether the FPU is in use or not.
875  *
876  * Note that this only works on the current task.
877  *
878  * Inputs:
879  *	@xfeature_nr: state which is defined in xsave.h (e.g. XFEATURE_FP,
880  *	XFEATURE_SSE, etc...)
881  * Output:
882  *	address of the state in the xsave area or NULL if the state
883  *	is not present or is in its 'init state'.
884  */
885 const void *get_xsave_field_ptr(int xfeature_nr)
886 {
887 	struct fpu *fpu = &current->thread.fpu;
888 
889 	/*
890 	 * fpu__save() takes the CPU's xstate registers
891 	 * and saves them off to the 'fpu memory buffer.
892 	 */
893 	fpu__save(fpu);
894 
895 	return get_xsave_addr(&fpu->state.xsave, xfeature_nr);
896 }
897 
898 #ifdef CONFIG_ARCH_HAS_PKEYS
899 
900 #define NR_VALID_PKRU_BITS (CONFIG_NR_PROTECTION_KEYS * 2)
901 #define PKRU_VALID_MASK (NR_VALID_PKRU_BITS - 1)
902 /*
903  * This will go out and modify PKRU register to set the access
904  * rights for @pkey to @init_val.
905  */
906 int arch_set_user_pkey_access(struct task_struct *tsk, int pkey,
907 		unsigned long init_val)
908 {
909 	u32 old_pkru;
910 	int pkey_shift = (pkey * PKRU_BITS_PER_PKEY);
911 	u32 new_pkru_bits = 0;
912 
913 	/*
914 	 * This check implies XSAVE support.  OSPKE only gets
915 	 * set if we enable XSAVE and we enable PKU in XCR0.
916 	 */
917 	if (!boot_cpu_has(X86_FEATURE_OSPKE))
918 		return -EINVAL;
919 
920 	/* Set the bits we need in PKRU:  */
921 	if (init_val & PKEY_DISABLE_ACCESS)
922 		new_pkru_bits |= PKRU_AD_BIT;
923 	if (init_val & PKEY_DISABLE_WRITE)
924 		new_pkru_bits |= PKRU_WD_BIT;
925 
926 	/* Shift the bits in to the correct place in PKRU for pkey: */
927 	new_pkru_bits <<= pkey_shift;
928 
929 	/* Get old PKRU and mask off any old bits in place: */
930 	old_pkru = read_pkru();
931 	old_pkru &= ~((PKRU_AD_BIT|PKRU_WD_BIT) << pkey_shift);
932 
933 	/* Write old part along with new part: */
934 	write_pkru(old_pkru | new_pkru_bits);
935 
936 	return 0;
937 }
938 #endif /* ! CONFIG_ARCH_HAS_PKEYS */
939 
940 /*
941  * Weird legacy quirk: SSE and YMM states store information in the
942  * MXCSR and MXCSR_FLAGS fields of the FP area. That means if the FP
943  * area is marked as unused in the xfeatures header, we need to copy
944  * MXCSR and MXCSR_FLAGS if either SSE or YMM are in use.
945  */
946 static inline bool xfeatures_mxcsr_quirk(u64 xfeatures)
947 {
948 	if (!(xfeatures & (XFEATURE_MASK_SSE|XFEATURE_MASK_YMM)))
949 		return false;
950 
951 	if (xfeatures & XFEATURE_MASK_FP)
952 		return false;
953 
954 	return true;
955 }
956 
957 /*
958  * This is similar to user_regset_copyout(), but will not add offset to
959  * the source data pointer or increment pos, count, kbuf, and ubuf.
960  */
961 static inline void
962 __copy_xstate_to_kernel(void *kbuf, const void *data,
963 			unsigned int offset, unsigned int size, unsigned int size_total)
964 {
965 	if (offset < size_total) {
966 		unsigned int copy = min(size, size_total - offset);
967 
968 		memcpy(kbuf + offset, data, copy);
969 	}
970 }
971 
972 /*
973  * Convert from kernel XSAVES compacted format to standard format and copy
974  * to a kernel-space ptrace buffer.
975  *
976  * It supports partial copy but pos always starts from zero. This is called
977  * from xstateregs_get() and there we check the CPU has XSAVES.
978  */
979 int copy_xstate_to_kernel(void *kbuf, struct xregs_state *xsave, unsigned int offset_start, unsigned int size_total)
980 {
981 	unsigned int offset, size;
982 	struct xstate_header header;
983 	int i;
984 
985 	/*
986 	 * Currently copy_regset_to_user() starts from pos 0:
987 	 */
988 	if (unlikely(offset_start != 0))
989 		return -EFAULT;
990 
991 	/*
992 	 * The destination is a ptrace buffer; we put in only user xstates:
993 	 */
994 	memset(&header, 0, sizeof(header));
995 	header.xfeatures = xsave->header.xfeatures;
996 	header.xfeatures &= ~XFEATURE_MASK_SUPERVISOR;
997 
998 	/*
999 	 * Copy xregs_state->header:
1000 	 */
1001 	offset = offsetof(struct xregs_state, header);
1002 	size = sizeof(header);
1003 
1004 	__copy_xstate_to_kernel(kbuf, &header, offset, size, size_total);
1005 
1006 	for (i = 0; i < XFEATURE_MAX; i++) {
1007 		/*
1008 		 * Copy only in-use xstates:
1009 		 */
1010 		if ((header.xfeatures >> i) & 1) {
1011 			void *src = __raw_xsave_addr(xsave, i);
1012 
1013 			offset = xstate_offsets[i];
1014 			size = xstate_sizes[i];
1015 
1016 			/* The next component has to fit fully into the output buffer: */
1017 			if (offset + size > size_total)
1018 				break;
1019 
1020 			__copy_xstate_to_kernel(kbuf, src, offset, size, size_total);
1021 		}
1022 
1023 	}
1024 
1025 	if (xfeatures_mxcsr_quirk(header.xfeatures)) {
1026 		offset = offsetof(struct fxregs_state, mxcsr);
1027 		size = MXCSR_AND_FLAGS_SIZE;
1028 		__copy_xstate_to_kernel(kbuf, &xsave->i387.mxcsr, offset, size, size_total);
1029 	}
1030 
1031 	/*
1032 	 * Fill xsave->i387.sw_reserved value for ptrace frame:
1033 	 */
1034 	offset = offsetof(struct fxregs_state, sw_reserved);
1035 	size = sizeof(xstate_fx_sw_bytes);
1036 
1037 	__copy_xstate_to_kernel(kbuf, xstate_fx_sw_bytes, offset, size, size_total);
1038 
1039 	return 0;
1040 }
1041 
1042 static inline int
1043 __copy_xstate_to_user(void __user *ubuf, const void *data, unsigned int offset, unsigned int size, unsigned int size_total)
1044 {
1045 	if (!size)
1046 		return 0;
1047 
1048 	if (offset < size_total) {
1049 		unsigned int copy = min(size, size_total - offset);
1050 
1051 		if (__copy_to_user(ubuf + offset, data, copy))
1052 			return -EFAULT;
1053 	}
1054 	return 0;
1055 }
1056 
1057 /*
1058  * Convert from kernel XSAVES compacted format to standard format and copy
1059  * to a user-space buffer. It supports partial copy but pos always starts from
1060  * zero. This is called from xstateregs_get() and there we check the CPU
1061  * has XSAVES.
1062  */
1063 int copy_xstate_to_user(void __user *ubuf, struct xregs_state *xsave, unsigned int offset_start, unsigned int size_total)
1064 {
1065 	unsigned int offset, size;
1066 	int ret, i;
1067 	struct xstate_header header;
1068 
1069 	/*
1070 	 * Currently copy_regset_to_user() starts from pos 0:
1071 	 */
1072 	if (unlikely(offset_start != 0))
1073 		return -EFAULT;
1074 
1075 	/*
1076 	 * The destination is a ptrace buffer; we put in only user xstates:
1077 	 */
1078 	memset(&header, 0, sizeof(header));
1079 	header.xfeatures = xsave->header.xfeatures;
1080 	header.xfeatures &= ~XFEATURE_MASK_SUPERVISOR;
1081 
1082 	/*
1083 	 * Copy xregs_state->header:
1084 	 */
1085 	offset = offsetof(struct xregs_state, header);
1086 	size = sizeof(header);
1087 
1088 	ret = __copy_xstate_to_user(ubuf, &header, offset, size, size_total);
1089 	if (ret)
1090 		return ret;
1091 
1092 	for (i = 0; i < XFEATURE_MAX; i++) {
1093 		/*
1094 		 * Copy only in-use xstates:
1095 		 */
1096 		if ((header.xfeatures >> i) & 1) {
1097 			void *src = __raw_xsave_addr(xsave, i);
1098 
1099 			offset = xstate_offsets[i];
1100 			size = xstate_sizes[i];
1101 
1102 			/* The next component has to fit fully into the output buffer: */
1103 			if (offset + size > size_total)
1104 				break;
1105 
1106 			ret = __copy_xstate_to_user(ubuf, src, offset, size, size_total);
1107 			if (ret)
1108 				return ret;
1109 		}
1110 
1111 	}
1112 
1113 	if (xfeatures_mxcsr_quirk(header.xfeatures)) {
1114 		offset = offsetof(struct fxregs_state, mxcsr);
1115 		size = MXCSR_AND_FLAGS_SIZE;
1116 		__copy_xstate_to_user(ubuf, &xsave->i387.mxcsr, offset, size, size_total);
1117 	}
1118 
1119 	/*
1120 	 * Fill xsave->i387.sw_reserved value for ptrace frame:
1121 	 */
1122 	offset = offsetof(struct fxregs_state, sw_reserved);
1123 	size = sizeof(xstate_fx_sw_bytes);
1124 
1125 	ret = __copy_xstate_to_user(ubuf, xstate_fx_sw_bytes, offset, size, size_total);
1126 	if (ret)
1127 		return ret;
1128 
1129 	return 0;
1130 }
1131 
1132 /*
1133  * Convert from a ptrace standard-format kernel buffer to kernel XSAVES format
1134  * and copy to the target thread. This is called from xstateregs_set().
1135  */
1136 int copy_kernel_to_xstate(struct xregs_state *xsave, const void *kbuf)
1137 {
1138 	unsigned int offset, size;
1139 	int i;
1140 	struct xstate_header hdr;
1141 
1142 	offset = offsetof(struct xregs_state, header);
1143 	size = sizeof(hdr);
1144 
1145 	memcpy(&hdr, kbuf + offset, size);
1146 
1147 	if (validate_xstate_header(&hdr))
1148 		return -EINVAL;
1149 
1150 	for (i = 0; i < XFEATURE_MAX; i++) {
1151 		u64 mask = ((u64)1 << i);
1152 
1153 		if (hdr.xfeatures & mask) {
1154 			void *dst = __raw_xsave_addr(xsave, i);
1155 
1156 			offset = xstate_offsets[i];
1157 			size = xstate_sizes[i];
1158 
1159 			memcpy(dst, kbuf + offset, size);
1160 		}
1161 	}
1162 
1163 	if (xfeatures_mxcsr_quirk(hdr.xfeatures)) {
1164 		offset = offsetof(struct fxregs_state, mxcsr);
1165 		size = MXCSR_AND_FLAGS_SIZE;
1166 		memcpy(&xsave->i387.mxcsr, kbuf + offset, size);
1167 	}
1168 
1169 	/*
1170 	 * The state that came in from userspace was user-state only.
1171 	 * Mask all the user states out of 'xfeatures':
1172 	 */
1173 	xsave->header.xfeatures &= XFEATURE_MASK_SUPERVISOR;
1174 
1175 	/*
1176 	 * Add back in the features that came in from userspace:
1177 	 */
1178 	xsave->header.xfeatures |= hdr.xfeatures;
1179 
1180 	return 0;
1181 }
1182 
1183 /*
1184  * Convert from a ptrace or sigreturn standard-format user-space buffer to
1185  * kernel XSAVES format and copy to the target thread. This is called from
1186  * xstateregs_set(), as well as potentially from the sigreturn() and
1187  * rt_sigreturn() system calls.
1188  */
1189 int copy_user_to_xstate(struct xregs_state *xsave, const void __user *ubuf)
1190 {
1191 	unsigned int offset, size;
1192 	int i;
1193 	struct xstate_header hdr;
1194 
1195 	offset = offsetof(struct xregs_state, header);
1196 	size = sizeof(hdr);
1197 
1198 	if (__copy_from_user(&hdr, ubuf + offset, size))
1199 		return -EFAULT;
1200 
1201 	if (validate_xstate_header(&hdr))
1202 		return -EINVAL;
1203 
1204 	for (i = 0; i < XFEATURE_MAX; i++) {
1205 		u64 mask = ((u64)1 << i);
1206 
1207 		if (hdr.xfeatures & mask) {
1208 			void *dst = __raw_xsave_addr(xsave, i);
1209 
1210 			offset = xstate_offsets[i];
1211 			size = xstate_sizes[i];
1212 
1213 			if (__copy_from_user(dst, ubuf + offset, size))
1214 				return -EFAULT;
1215 		}
1216 	}
1217 
1218 	if (xfeatures_mxcsr_quirk(hdr.xfeatures)) {
1219 		offset = offsetof(struct fxregs_state, mxcsr);
1220 		size = MXCSR_AND_FLAGS_SIZE;
1221 		if (__copy_from_user(&xsave->i387.mxcsr, ubuf + offset, size))
1222 			return -EFAULT;
1223 	}
1224 
1225 	/*
1226 	 * The state that came in from userspace was user-state only.
1227 	 * Mask all the user states out of 'xfeatures':
1228 	 */
1229 	xsave->header.xfeatures &= XFEATURE_MASK_SUPERVISOR;
1230 
1231 	/*
1232 	 * Add back in the features that came in from userspace:
1233 	 */
1234 	xsave->header.xfeatures |= hdr.xfeatures;
1235 
1236 	return 0;
1237 }
1238 
1239 #ifdef CONFIG_PROC_PID_ARCH_STATUS
1240 /*
1241  * Report the amount of time elapsed in millisecond since last AVX512
1242  * use in the task.
1243  */
1244 static void avx512_status(struct seq_file *m, struct task_struct *task)
1245 {
1246 	unsigned long timestamp = READ_ONCE(task->thread.fpu.avx512_timestamp);
1247 	long delta;
1248 
1249 	if (!timestamp) {
1250 		/*
1251 		 * Report -1 if no AVX512 usage
1252 		 */
1253 		delta = -1;
1254 	} else {
1255 		delta = (long)(jiffies - timestamp);
1256 		/*
1257 		 * Cap to LONG_MAX if time difference > LONG_MAX
1258 		 */
1259 		if (delta < 0)
1260 			delta = LONG_MAX;
1261 		delta = jiffies_to_msecs(delta);
1262 	}
1263 
1264 	seq_put_decimal_ll(m, "AVX512_elapsed_ms:\t", delta);
1265 	seq_putc(m, '\n');
1266 }
1267 
1268 /*
1269  * Report architecture specific information
1270  */
1271 int proc_pid_arch_status(struct seq_file *m, struct pid_namespace *ns,
1272 			struct pid *pid, struct task_struct *task)
1273 {
1274 	/*
1275 	 * Report AVX512 state if the processor and build option supported.
1276 	 */
1277 	if (cpu_feature_enabled(X86_FEATURE_AVX512F))
1278 		avx512_status(m, task);
1279 
1280 	return 0;
1281 }
1282 #endif /* CONFIG_PROC_PID_ARCH_STATUS */
1283