xref: /openbmc/linux/arch/x86/kernel/fpu/xstate.c (revision 0352f880)
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/bitops.h>
8 #include <linux/compat.h>
9 #include <linux/cpu.h>
10 #include <linux/mman.h>
11 #include <linux/nospec.h>
12 #include <linux/pkeys.h>
13 #include <linux/seq_file.h>
14 #include <linux/proc_fs.h>
15 #include <linux/vmalloc.h>
16 
17 #include <asm/fpu/api.h>
18 #include <asm/fpu/regset.h>
19 #include <asm/fpu/signal.h>
20 #include <asm/fpu/xcr.h>
21 
22 #include <asm/tlbflush.h>
23 #include <asm/prctl.h>
24 #include <asm/elf.h>
25 
26 #include "context.h"
27 #include "internal.h"
28 #include "legacy.h"
29 #include "xstate.h"
30 
31 #define for_each_extended_xfeature(bit, mask)				\
32 	(bit) = FIRST_EXTENDED_XFEATURE;				\
33 	for_each_set_bit_from(bit, (unsigned long *)&(mask), 8 * sizeof(mask))
34 
35 /*
36  * Although we spell it out in here, the Processor Trace
37  * xfeature is completely unused.  We use other mechanisms
38  * to save/restore PT state in Linux.
39  */
40 static const char *xfeature_names[] =
41 {
42 	"x87 floating point registers"	,
43 	"SSE registers"			,
44 	"AVX registers"			,
45 	"MPX bounds registers"		,
46 	"MPX CSR"			,
47 	"AVX-512 opmask"		,
48 	"AVX-512 Hi256"			,
49 	"AVX-512 ZMM_Hi256"		,
50 	"Processor Trace (unused)"	,
51 	"Protection Keys User registers",
52 	"PASID state",
53 	"unknown xstate feature"	,
54 	"unknown xstate feature"	,
55 	"unknown xstate feature"	,
56 	"unknown xstate feature"	,
57 	"unknown xstate feature"	,
58 	"unknown xstate feature"	,
59 	"AMX Tile config"		,
60 	"AMX Tile data"			,
61 	"unknown xstate feature"	,
62 };
63 
64 static unsigned short xsave_cpuid_features[] __initdata = {
65 	[XFEATURE_FP]				= X86_FEATURE_FPU,
66 	[XFEATURE_SSE]				= X86_FEATURE_XMM,
67 	[XFEATURE_YMM]				= X86_FEATURE_AVX,
68 	[XFEATURE_BNDREGS]			= X86_FEATURE_MPX,
69 	[XFEATURE_BNDCSR]			= X86_FEATURE_MPX,
70 	[XFEATURE_OPMASK]			= X86_FEATURE_AVX512F,
71 	[XFEATURE_ZMM_Hi256]			= X86_FEATURE_AVX512F,
72 	[XFEATURE_Hi16_ZMM]			= X86_FEATURE_AVX512F,
73 	[XFEATURE_PT_UNIMPLEMENTED_SO_FAR]	= X86_FEATURE_INTEL_PT,
74 	[XFEATURE_PKRU]				= X86_FEATURE_PKU,
75 	[XFEATURE_PASID]			= X86_FEATURE_ENQCMD,
76 	[XFEATURE_XTILE_CFG]			= X86_FEATURE_AMX_TILE,
77 	[XFEATURE_XTILE_DATA]			= X86_FEATURE_AMX_TILE,
78 };
79 
80 static unsigned int xstate_offsets[XFEATURE_MAX] __ro_after_init =
81 	{ [ 0 ... XFEATURE_MAX - 1] = -1};
82 static unsigned int xstate_sizes[XFEATURE_MAX] __ro_after_init =
83 	{ [ 0 ... XFEATURE_MAX - 1] = -1};
84 static unsigned int xstate_flags[XFEATURE_MAX] __ro_after_init;
85 
86 #define XSTATE_FLAG_SUPERVISOR	BIT(0)
87 #define XSTATE_FLAG_ALIGNED64	BIT(1)
88 
89 /*
90  * Return whether the system supports a given xfeature.
91  *
92  * Also return the name of the (most advanced) feature that the caller requested:
93  */
94 int cpu_has_xfeatures(u64 xfeatures_needed, const char **feature_name)
95 {
96 	u64 xfeatures_missing = xfeatures_needed & ~fpu_kernel_cfg.max_features;
97 
98 	if (unlikely(feature_name)) {
99 		long xfeature_idx, max_idx;
100 		u64 xfeatures_print;
101 		/*
102 		 * So we use FLS here to be able to print the most advanced
103 		 * feature that was requested but is missing. So if a driver
104 		 * asks about "XFEATURE_MASK_SSE | XFEATURE_MASK_YMM" we'll print the
105 		 * missing AVX feature - this is the most informative message
106 		 * to users:
107 		 */
108 		if (xfeatures_missing)
109 			xfeatures_print = xfeatures_missing;
110 		else
111 			xfeatures_print = xfeatures_needed;
112 
113 		xfeature_idx = fls64(xfeatures_print)-1;
114 		max_idx = ARRAY_SIZE(xfeature_names)-1;
115 		xfeature_idx = min(xfeature_idx, max_idx);
116 
117 		*feature_name = xfeature_names[xfeature_idx];
118 	}
119 
120 	if (xfeatures_missing)
121 		return 0;
122 
123 	return 1;
124 }
125 EXPORT_SYMBOL_GPL(cpu_has_xfeatures);
126 
127 static bool xfeature_is_aligned64(int xfeature_nr)
128 {
129 	return xstate_flags[xfeature_nr] & XSTATE_FLAG_ALIGNED64;
130 }
131 
132 static bool xfeature_is_supervisor(int xfeature_nr)
133 {
134 	return xstate_flags[xfeature_nr] & XSTATE_FLAG_SUPERVISOR;
135 }
136 
137 static unsigned int xfeature_get_offset(u64 xcomp_bv, int xfeature)
138 {
139 	unsigned int offs, i;
140 
141 	/*
142 	 * Non-compacted format and legacy features use the cached fixed
143 	 * offsets.
144 	 */
145 	if (!cpu_feature_enabled(X86_FEATURE_XCOMPACTED) ||
146 	    xfeature <= XFEATURE_SSE)
147 		return xstate_offsets[xfeature];
148 
149 	/*
150 	 * Compacted format offsets depend on the actual content of the
151 	 * compacted xsave area which is determined by the xcomp_bv header
152 	 * field.
153 	 */
154 	offs = FXSAVE_SIZE + XSAVE_HDR_SIZE;
155 	for_each_extended_xfeature(i, xcomp_bv) {
156 		if (xfeature_is_aligned64(i))
157 			offs = ALIGN(offs, 64);
158 		if (i == xfeature)
159 			break;
160 		offs += xstate_sizes[i];
161 	}
162 	return offs;
163 }
164 
165 /*
166  * Enable the extended processor state save/restore feature.
167  * Called once per CPU onlining.
168  */
169 void fpu__init_cpu_xstate(void)
170 {
171 	if (!boot_cpu_has(X86_FEATURE_XSAVE) || !fpu_kernel_cfg.max_features)
172 		return;
173 
174 	cr4_set_bits(X86_CR4_OSXSAVE);
175 
176 	/*
177 	 * Must happen after CR4 setup and before xsetbv() to allow KVM
178 	 * lazy passthrough.  Write independent of the dynamic state static
179 	 * key as that does not work on the boot CPU. This also ensures
180 	 * that any stale state is wiped out from XFD.
181 	 */
182 	if (cpu_feature_enabled(X86_FEATURE_XFD))
183 		wrmsrl(MSR_IA32_XFD, init_fpstate.xfd);
184 
185 	/*
186 	 * XCR_XFEATURE_ENABLED_MASK (aka. XCR0) sets user features
187 	 * managed by XSAVE{C, OPT, S} and XRSTOR{S}.  Only XSAVE user
188 	 * states can be set here.
189 	 */
190 	xsetbv(XCR_XFEATURE_ENABLED_MASK, fpu_user_cfg.max_features);
191 
192 	/*
193 	 * MSR_IA32_XSS sets supervisor states managed by XSAVES.
194 	 */
195 	if (boot_cpu_has(X86_FEATURE_XSAVES)) {
196 		wrmsrl(MSR_IA32_XSS, xfeatures_mask_supervisor() |
197 				     xfeatures_mask_independent());
198 	}
199 }
200 
201 static bool xfeature_enabled(enum xfeature xfeature)
202 {
203 	return fpu_kernel_cfg.max_features & BIT_ULL(xfeature);
204 }
205 
206 /*
207  * Record the offsets and sizes of various xstates contained
208  * in the XSAVE state memory layout.
209  */
210 static void __init setup_xstate_cache(void)
211 {
212 	u32 eax, ebx, ecx, edx, i;
213 	/* start at the beginning of the "extended state" */
214 	unsigned int last_good_offset = offsetof(struct xregs_state,
215 						 extended_state_area);
216 	/*
217 	 * The FP xstates and SSE xstates are legacy states. They are always
218 	 * in the fixed offsets in the xsave area in either compacted form
219 	 * or standard form.
220 	 */
221 	xstate_offsets[XFEATURE_FP]	= 0;
222 	xstate_sizes[XFEATURE_FP]	= offsetof(struct fxregs_state,
223 						   xmm_space);
224 
225 	xstate_offsets[XFEATURE_SSE]	= xstate_sizes[XFEATURE_FP];
226 	xstate_sizes[XFEATURE_SSE]	= sizeof_field(struct fxregs_state,
227 						       xmm_space);
228 
229 	for_each_extended_xfeature(i, fpu_kernel_cfg.max_features) {
230 		cpuid_count(XSTATE_CPUID, i, &eax, &ebx, &ecx, &edx);
231 
232 		xstate_sizes[i] = eax;
233 		xstate_flags[i] = ecx;
234 
235 		/*
236 		 * If an xfeature is supervisor state, the offset in EBX is
237 		 * invalid, leave it to -1.
238 		 */
239 		if (xfeature_is_supervisor(i))
240 			continue;
241 
242 		xstate_offsets[i] = ebx;
243 
244 		/*
245 		 * In our xstate size checks, we assume that the highest-numbered
246 		 * xstate feature has the highest offset in the buffer.  Ensure
247 		 * it does.
248 		 */
249 		WARN_ONCE(last_good_offset > xstate_offsets[i],
250 			  "x86/fpu: misordered xstate at %d\n", last_good_offset);
251 
252 		last_good_offset = xstate_offsets[i];
253 	}
254 }
255 
256 static void __init print_xstate_feature(u64 xstate_mask)
257 {
258 	const char *feature_name;
259 
260 	if (cpu_has_xfeatures(xstate_mask, &feature_name))
261 		pr_info("x86/fpu: Supporting XSAVE feature 0x%03Lx: '%s'\n", xstate_mask, feature_name);
262 }
263 
264 /*
265  * Print out all the supported xstate features:
266  */
267 static void __init print_xstate_features(void)
268 {
269 	print_xstate_feature(XFEATURE_MASK_FP);
270 	print_xstate_feature(XFEATURE_MASK_SSE);
271 	print_xstate_feature(XFEATURE_MASK_YMM);
272 	print_xstate_feature(XFEATURE_MASK_BNDREGS);
273 	print_xstate_feature(XFEATURE_MASK_BNDCSR);
274 	print_xstate_feature(XFEATURE_MASK_OPMASK);
275 	print_xstate_feature(XFEATURE_MASK_ZMM_Hi256);
276 	print_xstate_feature(XFEATURE_MASK_Hi16_ZMM);
277 	print_xstate_feature(XFEATURE_MASK_PKRU);
278 	print_xstate_feature(XFEATURE_MASK_PASID);
279 	print_xstate_feature(XFEATURE_MASK_XTILE_CFG);
280 	print_xstate_feature(XFEATURE_MASK_XTILE_DATA);
281 }
282 
283 /*
284  * This check is important because it is easy to get XSTATE_*
285  * confused with XSTATE_BIT_*.
286  */
287 #define CHECK_XFEATURE(nr) do {		\
288 	WARN_ON(nr < FIRST_EXTENDED_XFEATURE);	\
289 	WARN_ON(nr >= XFEATURE_MAX);	\
290 } while (0)
291 
292 /*
293  * Print out xstate component offsets and sizes
294  */
295 static void __init print_xstate_offset_size(void)
296 {
297 	int i;
298 
299 	for_each_extended_xfeature(i, fpu_kernel_cfg.max_features) {
300 		pr_info("x86/fpu: xstate_offset[%d]: %4d, xstate_sizes[%d]: %4d\n",
301 			i, xfeature_get_offset(fpu_kernel_cfg.max_features, i),
302 			i, xstate_sizes[i]);
303 	}
304 }
305 
306 /*
307  * This function is called only during boot time when x86 caps are not set
308  * up and alternative can not be used yet.
309  */
310 static __init void os_xrstor_booting(struct xregs_state *xstate)
311 {
312 	u64 mask = fpu_kernel_cfg.max_features & XFEATURE_MASK_FPSTATE;
313 	u32 lmask = mask;
314 	u32 hmask = mask >> 32;
315 	int err;
316 
317 	if (cpu_feature_enabled(X86_FEATURE_XSAVES))
318 		XSTATE_OP(XRSTORS, xstate, lmask, hmask, err);
319 	else
320 		XSTATE_OP(XRSTOR, xstate, lmask, hmask, err);
321 
322 	/*
323 	 * We should never fault when copying from a kernel buffer, and the FPU
324 	 * state we set at boot time should be valid.
325 	 */
326 	WARN_ON_FPU(err);
327 }
328 
329 /*
330  * All supported features have either init state all zeros or are
331  * handled in setup_init_fpu() individually. This is an explicit
332  * feature list and does not use XFEATURE_MASK*SUPPORTED to catch
333  * newly added supported features at build time and make people
334  * actually look at the init state for the new feature.
335  */
336 #define XFEATURES_INIT_FPSTATE_HANDLED		\
337 	(XFEATURE_MASK_FP |			\
338 	 XFEATURE_MASK_SSE |			\
339 	 XFEATURE_MASK_YMM |			\
340 	 XFEATURE_MASK_OPMASK |			\
341 	 XFEATURE_MASK_ZMM_Hi256 |		\
342 	 XFEATURE_MASK_Hi16_ZMM	 |		\
343 	 XFEATURE_MASK_PKRU |			\
344 	 XFEATURE_MASK_BNDREGS |		\
345 	 XFEATURE_MASK_BNDCSR |			\
346 	 XFEATURE_MASK_PASID |			\
347 	 XFEATURE_MASK_XTILE)
348 
349 /*
350  * setup the xstate image representing the init state
351  */
352 static void __init setup_init_fpu_buf(void)
353 {
354 	BUILD_BUG_ON((XFEATURE_MASK_USER_SUPPORTED |
355 		      XFEATURE_MASK_SUPERVISOR_SUPPORTED) !=
356 		     XFEATURES_INIT_FPSTATE_HANDLED);
357 
358 	if (!boot_cpu_has(X86_FEATURE_XSAVE))
359 		return;
360 
361 	print_xstate_features();
362 
363 	xstate_init_xcomp_bv(&init_fpstate.regs.xsave, init_fpstate.xfeatures);
364 
365 	/*
366 	 * Init all the features state with header.xfeatures being 0x0
367 	 */
368 	os_xrstor_booting(&init_fpstate.regs.xsave);
369 
370 	/*
371 	 * All components are now in init state. Read the state back so
372 	 * that init_fpstate contains all non-zero init state. This only
373 	 * works with XSAVE, but not with XSAVEOPT and XSAVEC/S because
374 	 * those use the init optimization which skips writing data for
375 	 * components in init state.
376 	 *
377 	 * XSAVE could be used, but that would require to reshuffle the
378 	 * data when XSAVEC/S is available because XSAVEC/S uses xstate
379 	 * compaction. But doing so is a pointless exercise because most
380 	 * components have an all zeros init state except for the legacy
381 	 * ones (FP and SSE). Those can be saved with FXSAVE into the
382 	 * legacy area. Adding new features requires to ensure that init
383 	 * state is all zeroes or if not to add the necessary handling
384 	 * here.
385 	 */
386 	fxsave(&init_fpstate.regs.fxsave);
387 }
388 
389 int xfeature_size(int xfeature_nr)
390 {
391 	u32 eax, ebx, ecx, edx;
392 
393 	CHECK_XFEATURE(xfeature_nr);
394 	cpuid_count(XSTATE_CPUID, xfeature_nr, &eax, &ebx, &ecx, &edx);
395 	return eax;
396 }
397 
398 /* Validate an xstate header supplied by userspace (ptrace or sigreturn) */
399 static int validate_user_xstate_header(const struct xstate_header *hdr,
400 				       struct fpstate *fpstate)
401 {
402 	/* No unknown or supervisor features may be set */
403 	if (hdr->xfeatures & ~fpstate->user_xfeatures)
404 		return -EINVAL;
405 
406 	/* Userspace must use the uncompacted format */
407 	if (hdr->xcomp_bv)
408 		return -EINVAL;
409 
410 	/*
411 	 * If 'reserved' is shrunken to add a new field, make sure to validate
412 	 * that new field here!
413 	 */
414 	BUILD_BUG_ON(sizeof(hdr->reserved) != 48);
415 
416 	/* No reserved bits may be set */
417 	if (memchr_inv(hdr->reserved, 0, sizeof(hdr->reserved)))
418 		return -EINVAL;
419 
420 	return 0;
421 }
422 
423 static void __init __xstate_dump_leaves(void)
424 {
425 	int i;
426 	u32 eax, ebx, ecx, edx;
427 	static int should_dump = 1;
428 
429 	if (!should_dump)
430 		return;
431 	should_dump = 0;
432 	/*
433 	 * Dump out a few leaves past the ones that we support
434 	 * just in case there are some goodies up there
435 	 */
436 	for (i = 0; i < XFEATURE_MAX + 10; i++) {
437 		cpuid_count(XSTATE_CPUID, i, &eax, &ebx, &ecx, &edx);
438 		pr_warn("CPUID[%02x, %02x]: eax=%08x ebx=%08x ecx=%08x edx=%08x\n",
439 			XSTATE_CPUID, i, eax, ebx, ecx, edx);
440 	}
441 }
442 
443 #define XSTATE_WARN_ON(x) do {							\
444 	if (WARN_ONCE(x, "XSAVE consistency problem, dumping leaves")) {	\
445 		__xstate_dump_leaves();						\
446 	}									\
447 } while (0)
448 
449 #define XCHECK_SZ(sz, nr, nr_macro, __struct) do {			\
450 	if ((nr == nr_macro) &&						\
451 	    WARN_ONCE(sz != sizeof(__struct),				\
452 		"%s: struct is %zu bytes, cpu state %d bytes\n",	\
453 		__stringify(nr_macro), sizeof(__struct), sz)) {		\
454 		__xstate_dump_leaves();					\
455 	}								\
456 } while (0)
457 
458 /**
459  * check_xtile_data_against_struct - Check tile data state size.
460  *
461  * Calculate the state size by multiplying the single tile size which is
462  * recorded in a C struct, and the number of tiles that the CPU informs.
463  * Compare the provided size with the calculation.
464  *
465  * @size:	The tile data state size
466  *
467  * Returns:	0 on success, -EINVAL on mismatch.
468  */
469 static int __init check_xtile_data_against_struct(int size)
470 {
471 	u32 max_palid, palid, state_size;
472 	u32 eax, ebx, ecx, edx;
473 	u16 max_tile;
474 
475 	/*
476 	 * Check the maximum palette id:
477 	 *   eax: the highest numbered palette subleaf.
478 	 */
479 	cpuid_count(TILE_CPUID, 0, &max_palid, &ebx, &ecx, &edx);
480 
481 	/*
482 	 * Cross-check each tile size and find the maximum number of
483 	 * supported tiles.
484 	 */
485 	for (palid = 1, max_tile = 0; palid <= max_palid; palid++) {
486 		u16 tile_size, max;
487 
488 		/*
489 		 * Check the tile size info:
490 		 *   eax[31:16]:  bytes per title
491 		 *   ebx[31:16]:  the max names (or max number of tiles)
492 		 */
493 		cpuid_count(TILE_CPUID, palid, &eax, &ebx, &edx, &edx);
494 		tile_size = eax >> 16;
495 		max = ebx >> 16;
496 
497 		if (tile_size != sizeof(struct xtile_data)) {
498 			pr_err("%s: struct is %zu bytes, cpu xtile %d bytes\n",
499 			       __stringify(XFEATURE_XTILE_DATA),
500 			       sizeof(struct xtile_data), tile_size);
501 			__xstate_dump_leaves();
502 			return -EINVAL;
503 		}
504 
505 		if (max > max_tile)
506 			max_tile = max;
507 	}
508 
509 	state_size = sizeof(struct xtile_data) * max_tile;
510 	if (size != state_size) {
511 		pr_err("%s: calculated size is %u bytes, cpu state %d bytes\n",
512 		       __stringify(XFEATURE_XTILE_DATA), state_size, size);
513 		__xstate_dump_leaves();
514 		return -EINVAL;
515 	}
516 	return 0;
517 }
518 
519 /*
520  * We have a C struct for each 'xstate'.  We need to ensure
521  * that our software representation matches what the CPU
522  * tells us about the state's size.
523  */
524 static bool __init check_xstate_against_struct(int nr)
525 {
526 	/*
527 	 * Ask the CPU for the size of the state.
528 	 */
529 	int sz = xfeature_size(nr);
530 	/*
531 	 * Match each CPU state with the corresponding software
532 	 * structure.
533 	 */
534 	XCHECK_SZ(sz, nr, XFEATURE_YMM,       struct ymmh_struct);
535 	XCHECK_SZ(sz, nr, XFEATURE_BNDREGS,   struct mpx_bndreg_state);
536 	XCHECK_SZ(sz, nr, XFEATURE_BNDCSR,    struct mpx_bndcsr_state);
537 	XCHECK_SZ(sz, nr, XFEATURE_OPMASK,    struct avx_512_opmask_state);
538 	XCHECK_SZ(sz, nr, XFEATURE_ZMM_Hi256, struct avx_512_zmm_uppers_state);
539 	XCHECK_SZ(sz, nr, XFEATURE_Hi16_ZMM,  struct avx_512_hi16_state);
540 	XCHECK_SZ(sz, nr, XFEATURE_PKRU,      struct pkru_state);
541 	XCHECK_SZ(sz, nr, XFEATURE_PASID,     struct ia32_pasid_state);
542 	XCHECK_SZ(sz, nr, XFEATURE_XTILE_CFG, struct xtile_cfg);
543 
544 	/* The tile data size varies between implementations. */
545 	if (nr == XFEATURE_XTILE_DATA)
546 		check_xtile_data_against_struct(sz);
547 
548 	/*
549 	 * Make *SURE* to add any feature numbers in below if
550 	 * there are "holes" in the xsave state component
551 	 * numbers.
552 	 */
553 	if ((nr < XFEATURE_YMM) ||
554 	    (nr >= XFEATURE_MAX) ||
555 	    (nr == XFEATURE_PT_UNIMPLEMENTED_SO_FAR) ||
556 	    ((nr >= XFEATURE_RSRVD_COMP_11) && (nr <= XFEATURE_RSRVD_COMP_16))) {
557 		WARN_ONCE(1, "no structure for xstate: %d\n", nr);
558 		XSTATE_WARN_ON(1);
559 		return false;
560 	}
561 	return true;
562 }
563 
564 static unsigned int xstate_calculate_size(u64 xfeatures, bool compacted)
565 {
566 	unsigned int topmost = fls64(xfeatures) -  1;
567 	unsigned int offset = xstate_offsets[topmost];
568 
569 	if (topmost <= XFEATURE_SSE)
570 		return sizeof(struct xregs_state);
571 
572 	if (compacted)
573 		offset = xfeature_get_offset(xfeatures, topmost);
574 	return offset + xstate_sizes[topmost];
575 }
576 
577 /*
578  * This essentially double-checks what the cpu told us about
579  * how large the XSAVE buffer needs to be.  We are recalculating
580  * it to be safe.
581  *
582  * Independent XSAVE features allocate their own buffers and are not
583  * covered by these checks. Only the size of the buffer for task->fpu
584  * is checked here.
585  */
586 static bool __init paranoid_xstate_size_valid(unsigned int kernel_size)
587 {
588 	bool compacted = cpu_feature_enabled(X86_FEATURE_XCOMPACTED);
589 	bool xsaves = cpu_feature_enabled(X86_FEATURE_XSAVES);
590 	unsigned int size = FXSAVE_SIZE + XSAVE_HDR_SIZE;
591 	int i;
592 
593 	for_each_extended_xfeature(i, fpu_kernel_cfg.max_features) {
594 		if (!check_xstate_against_struct(i))
595 			return false;
596 		/*
597 		 * Supervisor state components can be managed only by
598 		 * XSAVES.
599 		 */
600 		if (!xsaves && xfeature_is_supervisor(i)) {
601 			XSTATE_WARN_ON(1);
602 			return false;
603 		}
604 	}
605 	size = xstate_calculate_size(fpu_kernel_cfg.max_features, compacted);
606 	XSTATE_WARN_ON(size != kernel_size);
607 	return size == kernel_size;
608 }
609 
610 /*
611  * Get total size of enabled xstates in XCR0 | IA32_XSS.
612  *
613  * Note the SDM's wording here.  "sub-function 0" only enumerates
614  * the size of the *user* states.  If we use it to size a buffer
615  * that we use 'XSAVES' on, we could potentially overflow the
616  * buffer because 'XSAVES' saves system states too.
617  *
618  * This also takes compaction into account. So this works for
619  * XSAVEC as well.
620  */
621 static unsigned int __init get_compacted_size(void)
622 {
623 	unsigned int eax, ebx, ecx, edx;
624 	/*
625 	 * - CPUID function 0DH, sub-function 1:
626 	 *    EBX enumerates the size (in bytes) required by
627 	 *    the XSAVES instruction for an XSAVE area
628 	 *    containing all the state components
629 	 *    corresponding to bits currently set in
630 	 *    XCR0 | IA32_XSS.
631 	 *
632 	 * When XSAVES is not available but XSAVEC is (virt), then there
633 	 * are no supervisor states, but XSAVEC still uses compacted
634 	 * format.
635 	 */
636 	cpuid_count(XSTATE_CPUID, 1, &eax, &ebx, &ecx, &edx);
637 	return ebx;
638 }
639 
640 /*
641  * Get the total size of the enabled xstates without the independent supervisor
642  * features.
643  */
644 static unsigned int __init get_xsave_compacted_size(void)
645 {
646 	u64 mask = xfeatures_mask_independent();
647 	unsigned int size;
648 
649 	if (!mask)
650 		return get_compacted_size();
651 
652 	/* Disable independent features. */
653 	wrmsrl(MSR_IA32_XSS, xfeatures_mask_supervisor());
654 
655 	/*
656 	 * Ask the hardware what size is required of the buffer.
657 	 * This is the size required for the task->fpu buffer.
658 	 */
659 	size = get_compacted_size();
660 
661 	/* Re-enable independent features so XSAVES will work on them again. */
662 	wrmsrl(MSR_IA32_XSS, xfeatures_mask_supervisor() | mask);
663 
664 	return size;
665 }
666 
667 static unsigned int __init get_xsave_size_user(void)
668 {
669 	unsigned int eax, ebx, ecx, edx;
670 	/*
671 	 * - CPUID function 0DH, sub-function 0:
672 	 *    EBX enumerates the size (in bytes) required by
673 	 *    the XSAVE instruction for an XSAVE area
674 	 *    containing all the *user* state components
675 	 *    corresponding to bits currently set in XCR0.
676 	 */
677 	cpuid_count(XSTATE_CPUID, 0, &eax, &ebx, &ecx, &edx);
678 	return ebx;
679 }
680 
681 static int __init init_xstate_size(void)
682 {
683 	/* Recompute the context size for enabled features: */
684 	unsigned int user_size, kernel_size, kernel_default_size;
685 	bool compacted = cpu_feature_enabled(X86_FEATURE_XCOMPACTED);
686 
687 	/* Uncompacted user space size */
688 	user_size = get_xsave_size_user();
689 
690 	/*
691 	 * XSAVES kernel size includes supervisor states and uses compacted
692 	 * format. XSAVEC uses compacted format, but does not save
693 	 * supervisor states.
694 	 *
695 	 * XSAVE[OPT] do not support supervisor states so kernel and user
696 	 * size is identical.
697 	 */
698 	if (compacted)
699 		kernel_size = get_xsave_compacted_size();
700 	else
701 		kernel_size = user_size;
702 
703 	kernel_default_size =
704 		xstate_calculate_size(fpu_kernel_cfg.default_features, compacted);
705 
706 	if (!paranoid_xstate_size_valid(kernel_size))
707 		return -EINVAL;
708 
709 	fpu_kernel_cfg.max_size = kernel_size;
710 	fpu_user_cfg.max_size = user_size;
711 
712 	fpu_kernel_cfg.default_size = kernel_default_size;
713 	fpu_user_cfg.default_size =
714 		xstate_calculate_size(fpu_user_cfg.default_features, false);
715 
716 	return 0;
717 }
718 
719 /*
720  * We enabled the XSAVE hardware, but something went wrong and
721  * we can not use it.  Disable it.
722  */
723 static void __init fpu__init_disable_system_xstate(unsigned int legacy_size)
724 {
725 	fpu_kernel_cfg.max_features = 0;
726 	cr4_clear_bits(X86_CR4_OSXSAVE);
727 	setup_clear_cpu_cap(X86_FEATURE_XSAVE);
728 
729 	/* Restore the legacy size.*/
730 	fpu_kernel_cfg.max_size = legacy_size;
731 	fpu_kernel_cfg.default_size = legacy_size;
732 	fpu_user_cfg.max_size = legacy_size;
733 	fpu_user_cfg.default_size = legacy_size;
734 
735 	/*
736 	 * Prevent enabling the static branch which enables writes to the
737 	 * XFD MSR.
738 	 */
739 	init_fpstate.xfd = 0;
740 
741 	fpstate_reset(&current->thread.fpu);
742 }
743 
744 /*
745  * Enable and initialize the xsave feature.
746  * Called once per system bootup.
747  */
748 void __init fpu__init_system_xstate(unsigned int legacy_size)
749 {
750 	unsigned int eax, ebx, ecx, edx;
751 	u64 xfeatures;
752 	int err;
753 	int i;
754 
755 	if (!boot_cpu_has(X86_FEATURE_FPU)) {
756 		pr_info("x86/fpu: No FPU detected\n");
757 		return;
758 	}
759 
760 	if (!boot_cpu_has(X86_FEATURE_XSAVE)) {
761 		pr_info("x86/fpu: x87 FPU will use %s\n",
762 			boot_cpu_has(X86_FEATURE_FXSR) ? "FXSAVE" : "FSAVE");
763 		return;
764 	}
765 
766 	if (boot_cpu_data.cpuid_level < XSTATE_CPUID) {
767 		WARN_ON_FPU(1);
768 		return;
769 	}
770 
771 	/*
772 	 * Find user xstates supported by the processor.
773 	 */
774 	cpuid_count(XSTATE_CPUID, 0, &eax, &ebx, &ecx, &edx);
775 	fpu_kernel_cfg.max_features = eax + ((u64)edx << 32);
776 
777 	/*
778 	 * Find supervisor xstates supported by the processor.
779 	 */
780 	cpuid_count(XSTATE_CPUID, 1, &eax, &ebx, &ecx, &edx);
781 	fpu_kernel_cfg.max_features |= ecx + ((u64)edx << 32);
782 
783 	if ((fpu_kernel_cfg.max_features & XFEATURE_MASK_FPSSE) != XFEATURE_MASK_FPSSE) {
784 		/*
785 		 * This indicates that something really unexpected happened
786 		 * with the enumeration.  Disable XSAVE and try to continue
787 		 * booting without it.  This is too early to BUG().
788 		 */
789 		pr_err("x86/fpu: FP/SSE not present amongst the CPU's xstate features: 0x%llx.\n",
790 		       fpu_kernel_cfg.max_features);
791 		goto out_disable;
792 	}
793 
794 	/*
795 	 * Clear XSAVE features that are disabled in the normal CPUID.
796 	 */
797 	for (i = 0; i < ARRAY_SIZE(xsave_cpuid_features); i++) {
798 		unsigned short cid = xsave_cpuid_features[i];
799 
800 		/* Careful: X86_FEATURE_FPU is 0! */
801 		if ((i != XFEATURE_FP && !cid) || !boot_cpu_has(cid))
802 			fpu_kernel_cfg.max_features &= ~BIT_ULL(i);
803 	}
804 
805 	if (!cpu_feature_enabled(X86_FEATURE_XFD))
806 		fpu_kernel_cfg.max_features &= ~XFEATURE_MASK_USER_DYNAMIC;
807 
808 	if (!cpu_feature_enabled(X86_FEATURE_XSAVES))
809 		fpu_kernel_cfg.max_features &= XFEATURE_MASK_USER_SUPPORTED;
810 	else
811 		fpu_kernel_cfg.max_features &= XFEATURE_MASK_USER_SUPPORTED |
812 					XFEATURE_MASK_SUPERVISOR_SUPPORTED;
813 
814 	fpu_user_cfg.max_features = fpu_kernel_cfg.max_features;
815 	fpu_user_cfg.max_features &= XFEATURE_MASK_USER_SUPPORTED;
816 
817 	/* Clean out dynamic features from default */
818 	fpu_kernel_cfg.default_features = fpu_kernel_cfg.max_features;
819 	fpu_kernel_cfg.default_features &= ~XFEATURE_MASK_USER_DYNAMIC;
820 
821 	fpu_user_cfg.default_features = fpu_user_cfg.max_features;
822 	fpu_user_cfg.default_features &= ~XFEATURE_MASK_USER_DYNAMIC;
823 
824 	/* Store it for paranoia check at the end */
825 	xfeatures = fpu_kernel_cfg.max_features;
826 
827 	/*
828 	 * Initialize the default XFD state in initfp_state and enable the
829 	 * dynamic sizing mechanism if dynamic states are available.  The
830 	 * static key cannot be enabled here because this runs before
831 	 * jump_label_init(). This is delayed to an initcall.
832 	 */
833 	init_fpstate.xfd = fpu_user_cfg.max_features & XFEATURE_MASK_USER_DYNAMIC;
834 
835 	/* Set up compaction feature bit */
836 	if (cpu_feature_enabled(X86_FEATURE_XSAVEC) ||
837 	    cpu_feature_enabled(X86_FEATURE_XSAVES))
838 		setup_force_cpu_cap(X86_FEATURE_XCOMPACTED);
839 
840 	/* Enable xstate instructions to be able to continue with initialization: */
841 	fpu__init_cpu_xstate();
842 
843 	/* Cache size, offset and flags for initialization */
844 	setup_xstate_cache();
845 
846 	err = init_xstate_size();
847 	if (err)
848 		goto out_disable;
849 
850 	/* Reset the state for the current task */
851 	fpstate_reset(&current->thread.fpu);
852 
853 	/*
854 	 * Update info used for ptrace frames; use standard-format size and no
855 	 * supervisor xstates:
856 	 */
857 	update_regset_xstate_info(fpu_user_cfg.max_size,
858 				  fpu_user_cfg.max_features);
859 
860 	/*
861 	 * init_fpstate excludes dynamic states as they are large but init
862 	 * state is zero.
863 	 */
864 	init_fpstate.size		= fpu_kernel_cfg.default_size;
865 	init_fpstate.xfeatures		= fpu_kernel_cfg.default_features;
866 
867 	if (init_fpstate.size > sizeof(init_fpstate.regs)) {
868 		pr_warn("x86/fpu: init_fpstate buffer too small (%zu < %d), disabling XSAVE\n",
869 			sizeof(init_fpstate.regs), init_fpstate.size);
870 		goto out_disable;
871 	}
872 
873 	setup_init_fpu_buf();
874 
875 	/*
876 	 * Paranoia check whether something in the setup modified the
877 	 * xfeatures mask.
878 	 */
879 	if (xfeatures != fpu_kernel_cfg.max_features) {
880 		pr_err("x86/fpu: xfeatures modified from 0x%016llx to 0x%016llx during init, disabling XSAVE\n",
881 		       xfeatures, fpu_kernel_cfg.max_features);
882 		goto out_disable;
883 	}
884 
885 	print_xstate_offset_size();
886 	pr_info("x86/fpu: Enabled xstate features 0x%llx, context size is %d bytes, using '%s' format.\n",
887 		fpu_kernel_cfg.max_features,
888 		fpu_kernel_cfg.max_size,
889 		boot_cpu_has(X86_FEATURE_XCOMPACTED) ? "compacted" : "standard");
890 	return;
891 
892 out_disable:
893 	/* something went wrong, try to boot without any XSAVE support */
894 	fpu__init_disable_system_xstate(legacy_size);
895 }
896 
897 /*
898  * Restore minimal FPU state after suspend:
899  */
900 void fpu__resume_cpu(void)
901 {
902 	/*
903 	 * Restore XCR0 on xsave capable CPUs:
904 	 */
905 	if (cpu_feature_enabled(X86_FEATURE_XSAVE))
906 		xsetbv(XCR_XFEATURE_ENABLED_MASK, fpu_user_cfg.max_features);
907 
908 	/*
909 	 * Restore IA32_XSS. The same CPUID bit enumerates support
910 	 * of XSAVES and MSR_IA32_XSS.
911 	 */
912 	if (cpu_feature_enabled(X86_FEATURE_XSAVES)) {
913 		wrmsrl(MSR_IA32_XSS, xfeatures_mask_supervisor()  |
914 				     xfeatures_mask_independent());
915 	}
916 
917 	if (fpu_state_size_dynamic())
918 		wrmsrl(MSR_IA32_XFD, current->thread.fpu.fpstate->xfd);
919 }
920 
921 /*
922  * Given an xstate feature nr, calculate where in the xsave
923  * buffer the state is.  Callers should ensure that the buffer
924  * is valid.
925  */
926 static void *__raw_xsave_addr(struct xregs_state *xsave, int xfeature_nr)
927 {
928 	u64 xcomp_bv = xsave->header.xcomp_bv;
929 
930 	if (WARN_ON_ONCE(!xfeature_enabled(xfeature_nr)))
931 		return NULL;
932 
933 	if (cpu_feature_enabled(X86_FEATURE_XCOMPACTED)) {
934 		if (WARN_ON_ONCE(!(xcomp_bv & BIT_ULL(xfeature_nr))))
935 			return NULL;
936 	}
937 
938 	return (void *)xsave + xfeature_get_offset(xcomp_bv, xfeature_nr);
939 }
940 
941 /*
942  * Given the xsave area and a state inside, this function returns the
943  * address of the state.
944  *
945  * This is the API that is called to get xstate address in either
946  * standard format or compacted format of xsave area.
947  *
948  * Note that if there is no data for the field in the xsave buffer
949  * this will return NULL.
950  *
951  * Inputs:
952  *	xstate: the thread's storage area for all FPU data
953  *	xfeature_nr: state which is defined in xsave.h (e.g. XFEATURE_FP,
954  *	XFEATURE_SSE, etc...)
955  * Output:
956  *	address of the state in the xsave area, or NULL if the
957  *	field is not present in the xsave buffer.
958  */
959 void *get_xsave_addr(struct xregs_state *xsave, int xfeature_nr)
960 {
961 	/*
962 	 * Do we even *have* xsave state?
963 	 */
964 	if (!boot_cpu_has(X86_FEATURE_XSAVE))
965 		return NULL;
966 
967 	/*
968 	 * We should not ever be requesting features that we
969 	 * have not enabled.
970 	 */
971 	if (WARN_ON_ONCE(!xfeature_enabled(xfeature_nr)))
972 		return NULL;
973 
974 	/*
975 	 * This assumes the last 'xsave*' instruction to
976 	 * have requested that 'xfeature_nr' be saved.
977 	 * If it did not, we might be seeing and old value
978 	 * of the field in the buffer.
979 	 *
980 	 * This can happen because the last 'xsave' did not
981 	 * request that this feature be saved (unlikely)
982 	 * or because the "init optimization" caused it
983 	 * to not be saved.
984 	 */
985 	if (!(xsave->header.xfeatures & BIT_ULL(xfeature_nr)))
986 		return NULL;
987 
988 	return __raw_xsave_addr(xsave, xfeature_nr);
989 }
990 
991 #ifdef CONFIG_ARCH_HAS_PKEYS
992 
993 /*
994  * This will go out and modify PKRU register to set the access
995  * rights for @pkey to @init_val.
996  */
997 int arch_set_user_pkey_access(struct task_struct *tsk, int pkey,
998 			      unsigned long init_val)
999 {
1000 	u32 old_pkru, new_pkru_bits = 0;
1001 	int pkey_shift;
1002 
1003 	/*
1004 	 * This check implies XSAVE support.  OSPKE only gets
1005 	 * set if we enable XSAVE and we enable PKU in XCR0.
1006 	 */
1007 	if (!cpu_feature_enabled(X86_FEATURE_OSPKE))
1008 		return -EINVAL;
1009 
1010 	/*
1011 	 * This code should only be called with valid 'pkey'
1012 	 * values originating from in-kernel users.  Complain
1013 	 * if a bad value is observed.
1014 	 */
1015 	if (WARN_ON_ONCE(pkey >= arch_max_pkey()))
1016 		return -EINVAL;
1017 
1018 	/* Set the bits we need in PKRU:  */
1019 	if (init_val & PKEY_DISABLE_ACCESS)
1020 		new_pkru_bits |= PKRU_AD_BIT;
1021 	if (init_val & PKEY_DISABLE_WRITE)
1022 		new_pkru_bits |= PKRU_WD_BIT;
1023 
1024 	/* Shift the bits in to the correct place in PKRU for pkey: */
1025 	pkey_shift = pkey * PKRU_BITS_PER_PKEY;
1026 	new_pkru_bits <<= pkey_shift;
1027 
1028 	/* Get old PKRU and mask off any old bits in place: */
1029 	old_pkru = read_pkru();
1030 	old_pkru &= ~((PKRU_AD_BIT|PKRU_WD_BIT) << pkey_shift);
1031 
1032 	/* Write old part along with new part: */
1033 	write_pkru(old_pkru | new_pkru_bits);
1034 
1035 	return 0;
1036 }
1037 #endif /* ! CONFIG_ARCH_HAS_PKEYS */
1038 
1039 static void copy_feature(bool from_xstate, struct membuf *to, void *xstate,
1040 			 void *init_xstate, unsigned int size)
1041 {
1042 	membuf_write(to, from_xstate ? xstate : init_xstate, size);
1043 }
1044 
1045 /**
1046  * __copy_xstate_to_uabi_buf - Copy kernel saved xstate to a UABI buffer
1047  * @to:		membuf descriptor
1048  * @fpstate:	The fpstate buffer from which to copy
1049  * @pkru_val:	The PKRU value to store in the PKRU component
1050  * @copy_mode:	The requested copy mode
1051  *
1052  * Converts from kernel XSAVE or XSAVES compacted format to UABI conforming
1053  * format, i.e. from the kernel internal hardware dependent storage format
1054  * to the requested @mode. UABI XSTATE is always uncompacted!
1055  *
1056  * It supports partial copy but @to.pos always starts from zero.
1057  */
1058 void __copy_xstate_to_uabi_buf(struct membuf to, struct fpstate *fpstate,
1059 			       u32 pkru_val, enum xstate_copy_mode copy_mode)
1060 {
1061 	const unsigned int off_mxcsr = offsetof(struct fxregs_state, mxcsr);
1062 	struct xregs_state *xinit = &init_fpstate.regs.xsave;
1063 	struct xregs_state *xsave = &fpstate->regs.xsave;
1064 	struct xstate_header header;
1065 	unsigned int zerofrom;
1066 	u64 mask;
1067 	int i;
1068 
1069 	memset(&header, 0, sizeof(header));
1070 	header.xfeatures = xsave->header.xfeatures;
1071 
1072 	/* Mask out the feature bits depending on copy mode */
1073 	switch (copy_mode) {
1074 	case XSTATE_COPY_FP:
1075 		header.xfeatures &= XFEATURE_MASK_FP;
1076 		break;
1077 
1078 	case XSTATE_COPY_FX:
1079 		header.xfeatures &= XFEATURE_MASK_FP | XFEATURE_MASK_SSE;
1080 		break;
1081 
1082 	case XSTATE_COPY_XSAVE:
1083 		header.xfeatures &= fpstate->user_xfeatures;
1084 		break;
1085 	}
1086 
1087 	/* Copy FP state up to MXCSR */
1088 	copy_feature(header.xfeatures & XFEATURE_MASK_FP, &to, &xsave->i387,
1089 		     &xinit->i387, off_mxcsr);
1090 
1091 	/* Copy MXCSR when SSE or YMM are set in the feature mask */
1092 	copy_feature(header.xfeatures & (XFEATURE_MASK_SSE | XFEATURE_MASK_YMM),
1093 		     &to, &xsave->i387.mxcsr, &xinit->i387.mxcsr,
1094 		     MXCSR_AND_FLAGS_SIZE);
1095 
1096 	/* Copy the remaining FP state */
1097 	copy_feature(header.xfeatures & XFEATURE_MASK_FP,
1098 		     &to, &xsave->i387.st_space, &xinit->i387.st_space,
1099 		     sizeof(xsave->i387.st_space));
1100 
1101 	/* Copy the SSE state - shared with YMM, but independently managed */
1102 	copy_feature(header.xfeatures & XFEATURE_MASK_SSE,
1103 		     &to, &xsave->i387.xmm_space, &xinit->i387.xmm_space,
1104 		     sizeof(xsave->i387.xmm_space));
1105 
1106 	if (copy_mode != XSTATE_COPY_XSAVE)
1107 		goto out;
1108 
1109 	/* Zero the padding area */
1110 	membuf_zero(&to, sizeof(xsave->i387.padding));
1111 
1112 	/* Copy xsave->i387.sw_reserved */
1113 	membuf_write(&to, xstate_fx_sw_bytes, sizeof(xsave->i387.sw_reserved));
1114 
1115 	/* Copy the user space relevant state of @xsave->header */
1116 	membuf_write(&to, &header, sizeof(header));
1117 
1118 	zerofrom = offsetof(struct xregs_state, extended_state_area);
1119 
1120 	/*
1121 	 * The ptrace buffer is in non-compacted XSAVE format.  In
1122 	 * non-compacted format disabled features still occupy state space,
1123 	 * but there is no state to copy from in the compacted
1124 	 * init_fpstate. The gap tracking will zero these states.
1125 	 */
1126 	mask = fpstate->user_xfeatures;
1127 
1128 	/*
1129 	 * Dynamic features are not present in init_fpstate. When they are
1130 	 * in an all zeros init state, remove those from 'mask' to zero
1131 	 * those features in the user buffer instead of retrieving them
1132 	 * from init_fpstate.
1133 	 */
1134 	if (fpu_state_size_dynamic())
1135 		mask &= (header.xfeatures | xinit->header.xcomp_bv);
1136 
1137 	for_each_extended_xfeature(i, mask) {
1138 		/*
1139 		 * If there was a feature or alignment gap, zero the space
1140 		 * in the destination buffer.
1141 		 */
1142 		if (zerofrom < xstate_offsets[i])
1143 			membuf_zero(&to, xstate_offsets[i] - zerofrom);
1144 
1145 		if (i == XFEATURE_PKRU) {
1146 			struct pkru_state pkru = {0};
1147 			/*
1148 			 * PKRU is not necessarily up to date in the
1149 			 * XSAVE buffer. Use the provided value.
1150 			 */
1151 			pkru.pkru = pkru_val;
1152 			membuf_write(&to, &pkru, sizeof(pkru));
1153 		} else {
1154 			copy_feature(header.xfeatures & BIT_ULL(i), &to,
1155 				     __raw_xsave_addr(xsave, i),
1156 				     __raw_xsave_addr(xinit, i),
1157 				     xstate_sizes[i]);
1158 		}
1159 		/*
1160 		 * Keep track of the last copied state in the non-compacted
1161 		 * target buffer for gap zeroing.
1162 		 */
1163 		zerofrom = xstate_offsets[i] + xstate_sizes[i];
1164 	}
1165 
1166 out:
1167 	if (to.left)
1168 		membuf_zero(&to, to.left);
1169 }
1170 
1171 /**
1172  * copy_xstate_to_uabi_buf - Copy kernel saved xstate to a UABI buffer
1173  * @to:		membuf descriptor
1174  * @tsk:	The task from which to copy the saved xstate
1175  * @copy_mode:	The requested copy mode
1176  *
1177  * Converts from kernel XSAVE or XSAVES compacted format to UABI conforming
1178  * format, i.e. from the kernel internal hardware dependent storage format
1179  * to the requested @mode. UABI XSTATE is always uncompacted!
1180  *
1181  * It supports partial copy but @to.pos always starts from zero.
1182  */
1183 void copy_xstate_to_uabi_buf(struct membuf to, struct task_struct *tsk,
1184 			     enum xstate_copy_mode copy_mode)
1185 {
1186 	__copy_xstate_to_uabi_buf(to, tsk->thread.fpu.fpstate,
1187 				  tsk->thread.pkru, copy_mode);
1188 }
1189 
1190 static int copy_from_buffer(void *dst, unsigned int offset, unsigned int size,
1191 			    const void *kbuf, const void __user *ubuf)
1192 {
1193 	if (kbuf) {
1194 		memcpy(dst, kbuf + offset, size);
1195 	} else {
1196 		if (copy_from_user(dst, ubuf + offset, size))
1197 			return -EFAULT;
1198 	}
1199 	return 0;
1200 }
1201 
1202 
1203 static int copy_uabi_to_xstate(struct fpstate *fpstate, const void *kbuf,
1204 			       const void __user *ubuf)
1205 {
1206 	struct xregs_state *xsave = &fpstate->regs.xsave;
1207 	unsigned int offset, size;
1208 	struct xstate_header hdr;
1209 	u64 mask;
1210 	int i;
1211 
1212 	offset = offsetof(struct xregs_state, header);
1213 	if (copy_from_buffer(&hdr, offset, sizeof(hdr), kbuf, ubuf))
1214 		return -EFAULT;
1215 
1216 	if (validate_user_xstate_header(&hdr, fpstate))
1217 		return -EINVAL;
1218 
1219 	/* Validate MXCSR when any of the related features is in use */
1220 	mask = XFEATURE_MASK_FP | XFEATURE_MASK_SSE | XFEATURE_MASK_YMM;
1221 	if (hdr.xfeatures & mask) {
1222 		u32 mxcsr[2];
1223 
1224 		offset = offsetof(struct fxregs_state, mxcsr);
1225 		if (copy_from_buffer(mxcsr, offset, sizeof(mxcsr), kbuf, ubuf))
1226 			return -EFAULT;
1227 
1228 		/* Reserved bits in MXCSR must be zero. */
1229 		if (mxcsr[0] & ~mxcsr_feature_mask)
1230 			return -EINVAL;
1231 
1232 		/* SSE and YMM require MXCSR even when FP is not in use. */
1233 		if (!(hdr.xfeatures & XFEATURE_MASK_FP)) {
1234 			xsave->i387.mxcsr = mxcsr[0];
1235 			xsave->i387.mxcsr_mask = mxcsr[1];
1236 		}
1237 	}
1238 
1239 	for (i = 0; i < XFEATURE_MAX; i++) {
1240 		mask = BIT_ULL(i);
1241 
1242 		if (hdr.xfeatures & mask) {
1243 			void *dst = __raw_xsave_addr(xsave, i);
1244 
1245 			offset = xstate_offsets[i];
1246 			size = xstate_sizes[i];
1247 
1248 			if (copy_from_buffer(dst, offset, size, kbuf, ubuf))
1249 				return -EFAULT;
1250 		}
1251 	}
1252 
1253 	/*
1254 	 * The state that came in from userspace was user-state only.
1255 	 * Mask all the user states out of 'xfeatures':
1256 	 */
1257 	xsave->header.xfeatures &= XFEATURE_MASK_SUPERVISOR_ALL;
1258 
1259 	/*
1260 	 * Add back in the features that came in from userspace:
1261 	 */
1262 	xsave->header.xfeatures |= hdr.xfeatures;
1263 
1264 	return 0;
1265 }
1266 
1267 /*
1268  * Convert from a ptrace standard-format kernel buffer to kernel XSAVE[S]
1269  * format and copy to the target thread. Used by ptrace and KVM.
1270  */
1271 int copy_uabi_from_kernel_to_xstate(struct fpstate *fpstate, const void *kbuf)
1272 {
1273 	return copy_uabi_to_xstate(fpstate, kbuf, NULL);
1274 }
1275 
1276 /*
1277  * Convert from a sigreturn standard-format user-space buffer to kernel
1278  * XSAVE[S] format and copy to the target thread. This is called from the
1279  * sigreturn() and rt_sigreturn() system calls.
1280  */
1281 int copy_sigframe_from_user_to_xstate(struct fpstate *fpstate,
1282 				      const void __user *ubuf)
1283 {
1284 	return copy_uabi_to_xstate(fpstate, NULL, ubuf);
1285 }
1286 
1287 static bool validate_independent_components(u64 mask)
1288 {
1289 	u64 xchk;
1290 
1291 	if (WARN_ON_FPU(!cpu_feature_enabled(X86_FEATURE_XSAVES)))
1292 		return false;
1293 
1294 	xchk = ~xfeatures_mask_independent();
1295 
1296 	if (WARN_ON_ONCE(!mask || mask & xchk))
1297 		return false;
1298 
1299 	return true;
1300 }
1301 
1302 /**
1303  * xsaves - Save selected components to a kernel xstate buffer
1304  * @xstate:	Pointer to the buffer
1305  * @mask:	Feature mask to select the components to save
1306  *
1307  * The @xstate buffer must be 64 byte aligned and correctly initialized as
1308  * XSAVES does not write the full xstate header. Before first use the
1309  * buffer should be zeroed otherwise a consecutive XRSTORS from that buffer
1310  * can #GP.
1311  *
1312  * The feature mask must be a subset of the independent features.
1313  */
1314 void xsaves(struct xregs_state *xstate, u64 mask)
1315 {
1316 	int err;
1317 
1318 	if (!validate_independent_components(mask))
1319 		return;
1320 
1321 	XSTATE_OP(XSAVES, xstate, (u32)mask, (u32)(mask >> 32), err);
1322 	WARN_ON_ONCE(err);
1323 }
1324 
1325 /**
1326  * xrstors - Restore selected components from a kernel xstate buffer
1327  * @xstate:	Pointer to the buffer
1328  * @mask:	Feature mask to select the components to restore
1329  *
1330  * The @xstate buffer must be 64 byte aligned and correctly initialized
1331  * otherwise XRSTORS from that buffer can #GP.
1332  *
1333  * Proper usage is to restore the state which was saved with
1334  * xsaves() into @xstate.
1335  *
1336  * The feature mask must be a subset of the independent features.
1337  */
1338 void xrstors(struct xregs_state *xstate, u64 mask)
1339 {
1340 	int err;
1341 
1342 	if (!validate_independent_components(mask))
1343 		return;
1344 
1345 	XSTATE_OP(XRSTORS, xstate, (u32)mask, (u32)(mask >> 32), err);
1346 	WARN_ON_ONCE(err);
1347 }
1348 
1349 #if IS_ENABLED(CONFIG_KVM)
1350 void fpstate_clear_xstate_component(struct fpstate *fps, unsigned int xfeature)
1351 {
1352 	void *addr = get_xsave_addr(&fps->regs.xsave, xfeature);
1353 
1354 	if (addr)
1355 		memset(addr, 0, xstate_sizes[xfeature]);
1356 }
1357 EXPORT_SYMBOL_GPL(fpstate_clear_xstate_component);
1358 #endif
1359 
1360 #ifdef CONFIG_X86_64
1361 
1362 #ifdef CONFIG_X86_DEBUG_FPU
1363 /*
1364  * Ensure that a subsequent XSAVE* or XRSTOR* instruction with RFBM=@mask
1365  * can safely operate on the @fpstate buffer.
1366  */
1367 static bool xstate_op_valid(struct fpstate *fpstate, u64 mask, bool rstor)
1368 {
1369 	u64 xfd = __this_cpu_read(xfd_state);
1370 
1371 	if (fpstate->xfd == xfd)
1372 		return true;
1373 
1374 	 /*
1375 	  * The XFD MSR does not match fpstate->xfd. That's invalid when
1376 	  * the passed in fpstate is current's fpstate.
1377 	  */
1378 	if (fpstate->xfd == current->thread.fpu.fpstate->xfd)
1379 		return false;
1380 
1381 	/*
1382 	 * XRSTOR(S) from init_fpstate are always correct as it will just
1383 	 * bring all components into init state and not read from the
1384 	 * buffer. XSAVE(S) raises #PF after init.
1385 	 */
1386 	if (fpstate == &init_fpstate)
1387 		return rstor;
1388 
1389 	/*
1390 	 * XSAVE(S): clone(), fpu_swap_kvm_fpu()
1391 	 * XRSTORS(S): fpu_swap_kvm_fpu()
1392 	 */
1393 
1394 	/*
1395 	 * No XSAVE/XRSTOR instructions (except XSAVE itself) touch
1396 	 * the buffer area for XFD-disabled state components.
1397 	 */
1398 	mask &= ~xfd;
1399 
1400 	/*
1401 	 * Remove features which are valid in fpstate. They
1402 	 * have space allocated in fpstate.
1403 	 */
1404 	mask &= ~fpstate->xfeatures;
1405 
1406 	/*
1407 	 * Any remaining state components in 'mask' might be written
1408 	 * by XSAVE/XRSTOR. Fail validation it found.
1409 	 */
1410 	return !mask;
1411 }
1412 
1413 void xfd_validate_state(struct fpstate *fpstate, u64 mask, bool rstor)
1414 {
1415 	WARN_ON_ONCE(!xstate_op_valid(fpstate, mask, rstor));
1416 }
1417 #endif /* CONFIG_X86_DEBUG_FPU */
1418 
1419 static int __init xfd_update_static_branch(void)
1420 {
1421 	/*
1422 	 * If init_fpstate.xfd has bits set then dynamic features are
1423 	 * available and the dynamic sizing must be enabled.
1424 	 */
1425 	if (init_fpstate.xfd)
1426 		static_branch_enable(&__fpu_state_size_dynamic);
1427 	return 0;
1428 }
1429 arch_initcall(xfd_update_static_branch)
1430 
1431 void fpstate_free(struct fpu *fpu)
1432 {
1433 	if (fpu->fpstate && fpu->fpstate != &fpu->__fpstate)
1434 		vfree(fpu->fpstate);
1435 }
1436 
1437 /**
1438  * fpstate_realloc - Reallocate struct fpstate for the requested new features
1439  *
1440  * @xfeatures:	A bitmap of xstate features which extend the enabled features
1441  *		of that task
1442  * @ksize:	The required size for the kernel buffer
1443  * @usize:	The required size for user space buffers
1444  * @guest_fpu:	Pointer to a guest FPU container. NULL for host allocations
1445  *
1446  * Note vs. vmalloc(): If the task with a vzalloc()-allocated buffer
1447  * terminates quickly, vfree()-induced IPIs may be a concern, but tasks
1448  * with large states are likely to live longer.
1449  *
1450  * Returns: 0 on success, -ENOMEM on allocation error.
1451  */
1452 static int fpstate_realloc(u64 xfeatures, unsigned int ksize,
1453 			   unsigned int usize, struct fpu_guest *guest_fpu)
1454 {
1455 	struct fpu *fpu = &current->thread.fpu;
1456 	struct fpstate *curfps, *newfps = NULL;
1457 	unsigned int fpsize;
1458 	bool in_use;
1459 
1460 	fpsize = ksize + ALIGN(offsetof(struct fpstate, regs), 64);
1461 
1462 	newfps = vzalloc(fpsize);
1463 	if (!newfps)
1464 		return -ENOMEM;
1465 	newfps->size = ksize;
1466 	newfps->user_size = usize;
1467 	newfps->is_valloc = true;
1468 
1469 	/*
1470 	 * When a guest FPU is supplied, use @guest_fpu->fpstate
1471 	 * as reference independent whether it is in use or not.
1472 	 */
1473 	curfps = guest_fpu ? guest_fpu->fpstate : fpu->fpstate;
1474 
1475 	/* Determine whether @curfps is the active fpstate */
1476 	in_use = fpu->fpstate == curfps;
1477 
1478 	if (guest_fpu) {
1479 		newfps->is_guest = true;
1480 		newfps->is_confidential = curfps->is_confidential;
1481 		newfps->in_use = curfps->in_use;
1482 		guest_fpu->xfeatures |= xfeatures;
1483 		guest_fpu->uabi_size = usize;
1484 	}
1485 
1486 	fpregs_lock();
1487 	/*
1488 	 * If @curfps is in use, ensure that the current state is in the
1489 	 * registers before swapping fpstate as that might invalidate it
1490 	 * due to layout changes.
1491 	 */
1492 	if (in_use && test_thread_flag(TIF_NEED_FPU_LOAD))
1493 		fpregs_restore_userregs();
1494 
1495 	newfps->xfeatures = curfps->xfeatures | xfeatures;
1496 
1497 	if (!guest_fpu)
1498 		newfps->user_xfeatures = curfps->user_xfeatures | xfeatures;
1499 
1500 	newfps->xfd = curfps->xfd & ~xfeatures;
1501 
1502 	/* Do the final updates within the locked region */
1503 	xstate_init_xcomp_bv(&newfps->regs.xsave, newfps->xfeatures);
1504 
1505 	if (guest_fpu) {
1506 		guest_fpu->fpstate = newfps;
1507 		/* If curfps is active, update the FPU fpstate pointer */
1508 		if (in_use)
1509 			fpu->fpstate = newfps;
1510 	} else {
1511 		fpu->fpstate = newfps;
1512 	}
1513 
1514 	if (in_use)
1515 		xfd_update_state(fpu->fpstate);
1516 	fpregs_unlock();
1517 
1518 	/* Only free valloc'ed state */
1519 	if (curfps && curfps->is_valloc)
1520 		vfree(curfps);
1521 
1522 	return 0;
1523 }
1524 
1525 static int validate_sigaltstack(unsigned int usize)
1526 {
1527 	struct task_struct *thread, *leader = current->group_leader;
1528 	unsigned long framesize = get_sigframe_size();
1529 
1530 	lockdep_assert_held(&current->sighand->siglock);
1531 
1532 	/* get_sigframe_size() is based on fpu_user_cfg.max_size */
1533 	framesize -= fpu_user_cfg.max_size;
1534 	framesize += usize;
1535 	for_each_thread(leader, thread) {
1536 		if (thread->sas_ss_size && thread->sas_ss_size < framesize)
1537 			return -ENOSPC;
1538 	}
1539 	return 0;
1540 }
1541 
1542 static int __xstate_request_perm(u64 permitted, u64 requested, bool guest)
1543 {
1544 	/*
1545 	 * This deliberately does not exclude !XSAVES as we still might
1546 	 * decide to optionally context switch XCR0 or talk the silicon
1547 	 * vendors into extending XFD for the pre AMX states, especially
1548 	 * AVX512.
1549 	 */
1550 	bool compacted = cpu_feature_enabled(X86_FEATURE_XCOMPACTED);
1551 	struct fpu *fpu = &current->group_leader->thread.fpu;
1552 	struct fpu_state_perm *perm;
1553 	unsigned int ksize, usize;
1554 	u64 mask;
1555 	int ret = 0;
1556 
1557 	/* Check whether fully enabled */
1558 	if ((permitted & requested) == requested)
1559 		return 0;
1560 
1561 	/* Calculate the resulting kernel state size */
1562 	mask = permitted | requested;
1563 	/* Take supervisor states into account on the host */
1564 	if (!guest)
1565 		mask |= xfeatures_mask_supervisor();
1566 	ksize = xstate_calculate_size(mask, compacted);
1567 
1568 	/* Calculate the resulting user state size */
1569 	mask &= XFEATURE_MASK_USER_SUPPORTED;
1570 	usize = xstate_calculate_size(mask, false);
1571 
1572 	if (!guest) {
1573 		ret = validate_sigaltstack(usize);
1574 		if (ret)
1575 			return ret;
1576 	}
1577 
1578 	perm = guest ? &fpu->guest_perm : &fpu->perm;
1579 	/* Pairs with the READ_ONCE() in xstate_get_group_perm() */
1580 	WRITE_ONCE(perm->__state_perm, mask);
1581 	/* Protected by sighand lock */
1582 	perm->__state_size = ksize;
1583 	perm->__user_state_size = usize;
1584 	return ret;
1585 }
1586 
1587 /*
1588  * Permissions array to map facilities with more than one component
1589  */
1590 static const u64 xstate_prctl_req[XFEATURE_MAX] = {
1591 	[XFEATURE_XTILE_DATA] = XFEATURE_MASK_XTILE_DATA,
1592 };
1593 
1594 static int xstate_request_perm(unsigned long idx, bool guest)
1595 {
1596 	u64 permitted, requested;
1597 	int ret;
1598 
1599 	if (idx >= XFEATURE_MAX)
1600 		return -EINVAL;
1601 
1602 	/*
1603 	 * Look up the facility mask which can require more than
1604 	 * one xstate component.
1605 	 */
1606 	idx = array_index_nospec(idx, ARRAY_SIZE(xstate_prctl_req));
1607 	requested = xstate_prctl_req[idx];
1608 	if (!requested)
1609 		return -EOPNOTSUPP;
1610 
1611 	if ((fpu_user_cfg.max_features & requested) != requested)
1612 		return -EOPNOTSUPP;
1613 
1614 	/* Lockless quick check */
1615 	permitted = xstate_get_group_perm(guest);
1616 	if ((permitted & requested) == requested)
1617 		return 0;
1618 
1619 	/* Protect against concurrent modifications */
1620 	spin_lock_irq(&current->sighand->siglock);
1621 	permitted = xstate_get_group_perm(guest);
1622 
1623 	/* First vCPU allocation locks the permissions. */
1624 	if (guest && (permitted & FPU_GUEST_PERM_LOCKED))
1625 		ret = -EBUSY;
1626 	else
1627 		ret = __xstate_request_perm(permitted, requested, guest);
1628 	spin_unlock_irq(&current->sighand->siglock);
1629 	return ret;
1630 }
1631 
1632 int __xfd_enable_feature(u64 xfd_err, struct fpu_guest *guest_fpu)
1633 {
1634 	u64 xfd_event = xfd_err & XFEATURE_MASK_USER_DYNAMIC;
1635 	struct fpu_state_perm *perm;
1636 	unsigned int ksize, usize;
1637 	struct fpu *fpu;
1638 
1639 	if (!xfd_event) {
1640 		if (!guest_fpu)
1641 			pr_err_once("XFD: Invalid xfd error: %016llx\n", xfd_err);
1642 		return 0;
1643 	}
1644 
1645 	/* Protect against concurrent modifications */
1646 	spin_lock_irq(&current->sighand->siglock);
1647 
1648 	/* If not permitted let it die */
1649 	if ((xstate_get_group_perm(!!guest_fpu) & xfd_event) != xfd_event) {
1650 		spin_unlock_irq(&current->sighand->siglock);
1651 		return -EPERM;
1652 	}
1653 
1654 	fpu = &current->group_leader->thread.fpu;
1655 	perm = guest_fpu ? &fpu->guest_perm : &fpu->perm;
1656 	ksize = perm->__state_size;
1657 	usize = perm->__user_state_size;
1658 
1659 	/*
1660 	 * The feature is permitted. State size is sufficient.  Dropping
1661 	 * the lock is safe here even if more features are added from
1662 	 * another task, the retrieved buffer sizes are valid for the
1663 	 * currently requested feature(s).
1664 	 */
1665 	spin_unlock_irq(&current->sighand->siglock);
1666 
1667 	/*
1668 	 * Try to allocate a new fpstate. If that fails there is no way
1669 	 * out.
1670 	 */
1671 	if (fpstate_realloc(xfd_event, ksize, usize, guest_fpu))
1672 		return -EFAULT;
1673 	return 0;
1674 }
1675 
1676 int xfd_enable_feature(u64 xfd_err)
1677 {
1678 	return __xfd_enable_feature(xfd_err, NULL);
1679 }
1680 
1681 #else /* CONFIG_X86_64 */
1682 static inline int xstate_request_perm(unsigned long idx, bool guest)
1683 {
1684 	return -EPERM;
1685 }
1686 #endif  /* !CONFIG_X86_64 */
1687 
1688 u64 xstate_get_guest_group_perm(void)
1689 {
1690 	return xstate_get_group_perm(true);
1691 }
1692 EXPORT_SYMBOL_GPL(xstate_get_guest_group_perm);
1693 
1694 /**
1695  * fpu_xstate_prctl - xstate permission operations
1696  * @tsk:	Redundant pointer to current
1697  * @option:	A subfunction of arch_prctl()
1698  * @arg2:	option argument
1699  * Return:	0 if successful; otherwise, an error code
1700  *
1701  * Option arguments:
1702  *
1703  * ARCH_GET_XCOMP_SUPP: Pointer to user space u64 to store the info
1704  * ARCH_GET_XCOMP_PERM: Pointer to user space u64 to store the info
1705  * ARCH_REQ_XCOMP_PERM: Facility number requested
1706  *
1707  * For facilities which require more than one XSTATE component, the request
1708  * must be the highest state component number related to that facility,
1709  * e.g. for AMX which requires XFEATURE_XTILE_CFG(17) and
1710  * XFEATURE_XTILE_DATA(18) this would be XFEATURE_XTILE_DATA(18).
1711  */
1712 long fpu_xstate_prctl(int option, unsigned long arg2)
1713 {
1714 	u64 __user *uptr = (u64 __user *)arg2;
1715 	u64 permitted, supported;
1716 	unsigned long idx = arg2;
1717 	bool guest = false;
1718 
1719 	switch (option) {
1720 	case ARCH_GET_XCOMP_SUPP:
1721 		supported = fpu_user_cfg.max_features |	fpu_user_cfg.legacy_features;
1722 		return put_user(supported, uptr);
1723 
1724 	case ARCH_GET_XCOMP_PERM:
1725 		/*
1726 		 * Lockless snapshot as it can also change right after the
1727 		 * dropping the lock.
1728 		 */
1729 		permitted = xstate_get_host_group_perm();
1730 		permitted &= XFEATURE_MASK_USER_SUPPORTED;
1731 		return put_user(permitted, uptr);
1732 
1733 	case ARCH_GET_XCOMP_GUEST_PERM:
1734 		permitted = xstate_get_guest_group_perm();
1735 		permitted &= XFEATURE_MASK_USER_SUPPORTED;
1736 		return put_user(permitted, uptr);
1737 
1738 	case ARCH_REQ_XCOMP_GUEST_PERM:
1739 		guest = true;
1740 		fallthrough;
1741 
1742 	case ARCH_REQ_XCOMP_PERM:
1743 		if (!IS_ENABLED(CONFIG_X86_64))
1744 			return -EOPNOTSUPP;
1745 
1746 		return xstate_request_perm(idx, guest);
1747 
1748 	default:
1749 		return -EINVAL;
1750 	}
1751 }
1752 
1753 #ifdef CONFIG_PROC_PID_ARCH_STATUS
1754 /*
1755  * Report the amount of time elapsed in millisecond since last AVX512
1756  * use in the task.
1757  */
1758 static void avx512_status(struct seq_file *m, struct task_struct *task)
1759 {
1760 	unsigned long timestamp = READ_ONCE(task->thread.fpu.avx512_timestamp);
1761 	long delta;
1762 
1763 	if (!timestamp) {
1764 		/*
1765 		 * Report -1 if no AVX512 usage
1766 		 */
1767 		delta = -1;
1768 	} else {
1769 		delta = (long)(jiffies - timestamp);
1770 		/*
1771 		 * Cap to LONG_MAX if time difference > LONG_MAX
1772 		 */
1773 		if (delta < 0)
1774 			delta = LONG_MAX;
1775 		delta = jiffies_to_msecs(delta);
1776 	}
1777 
1778 	seq_put_decimal_ll(m, "AVX512_elapsed_ms:\t", delta);
1779 	seq_putc(m, '\n');
1780 }
1781 
1782 /*
1783  * Report architecture specific information
1784  */
1785 int proc_pid_arch_status(struct seq_file *m, struct pid_namespace *ns,
1786 			struct pid *pid, struct task_struct *task)
1787 {
1788 	/*
1789 	 * Report AVX512 state if the processor and build option supported.
1790 	 */
1791 	if (cpu_feature_enabled(X86_FEATURE_AVX512F))
1792 		avx512_status(m, task);
1793 
1794 	return 0;
1795 }
1796 #endif /* CONFIG_PROC_PID_ARCH_STATUS */
1797