xref: /openbmc/linux/arch/x86/kernel/fpu/xstate.c (revision 11a163f2)
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 	"PASID state",
41 	"unknown xstate feature"	,
42 };
43 
44 static short xsave_cpuid_features[] __initdata = {
45 	X86_FEATURE_FPU,
46 	X86_FEATURE_XMM,
47 	X86_FEATURE_AVX,
48 	X86_FEATURE_MPX,
49 	X86_FEATURE_MPX,
50 	X86_FEATURE_AVX512F,
51 	X86_FEATURE_AVX512F,
52 	X86_FEATURE_AVX512F,
53 	X86_FEATURE_INTEL_PT,
54 	X86_FEATURE_PKU,
55 	X86_FEATURE_ENQCMD,
56 };
57 
58 /*
59  * This represents the full set of bits that should ever be set in a kernel
60  * XSAVE buffer, both supervisor and user xstates.
61  */
62 u64 xfeatures_mask_all __read_mostly;
63 
64 static unsigned int xstate_offsets[XFEATURE_MAX] = { [ 0 ... XFEATURE_MAX - 1] = -1};
65 static unsigned int xstate_sizes[XFEATURE_MAX]   = { [ 0 ... XFEATURE_MAX - 1] = -1};
66 static unsigned int xstate_comp_offsets[XFEATURE_MAX] = { [ 0 ... XFEATURE_MAX - 1] = -1};
67 static unsigned int xstate_supervisor_only_offsets[XFEATURE_MAX] = { [ 0 ... XFEATURE_MAX - 1] = -1};
68 
69 /*
70  * The XSAVE area of kernel can be in standard or compacted format;
71  * it is always in standard format for user mode. This is the user
72  * mode standard format size used for signal and ptrace frames.
73  */
74 unsigned int fpu_user_xstate_size;
75 
76 /*
77  * Return whether the system supports a given xfeature.
78  *
79  * Also return the name of the (most advanced) feature that the caller requested:
80  */
81 int cpu_has_xfeatures(u64 xfeatures_needed, const char **feature_name)
82 {
83 	u64 xfeatures_missing = xfeatures_needed & ~xfeatures_mask_all;
84 
85 	if (unlikely(feature_name)) {
86 		long xfeature_idx, max_idx;
87 		u64 xfeatures_print;
88 		/*
89 		 * So we use FLS here to be able to print the most advanced
90 		 * feature that was requested but is missing. So if a driver
91 		 * asks about "XFEATURE_MASK_SSE | XFEATURE_MASK_YMM" we'll print the
92 		 * missing AVX feature - this is the most informative message
93 		 * to users:
94 		 */
95 		if (xfeatures_missing)
96 			xfeatures_print = xfeatures_missing;
97 		else
98 			xfeatures_print = xfeatures_needed;
99 
100 		xfeature_idx = fls64(xfeatures_print)-1;
101 		max_idx = ARRAY_SIZE(xfeature_names)-1;
102 		xfeature_idx = min(xfeature_idx, max_idx);
103 
104 		*feature_name = xfeature_names[xfeature_idx];
105 	}
106 
107 	if (xfeatures_missing)
108 		return 0;
109 
110 	return 1;
111 }
112 EXPORT_SYMBOL_GPL(cpu_has_xfeatures);
113 
114 static bool xfeature_is_supervisor(int xfeature_nr)
115 {
116 	/*
117 	 * Extended State Enumeration Sub-leaves (EAX = 0DH, ECX = n, n > 1)
118 	 * returns ECX[0] set to (1) for a supervisor state, and cleared (0)
119 	 * for a user state.
120 	 */
121 	u32 eax, ebx, ecx, edx;
122 
123 	cpuid_count(XSTATE_CPUID, xfeature_nr, &eax, &ebx, &ecx, &edx);
124 	return ecx & 1;
125 }
126 
127 /*
128  * When executing XSAVEOPT (or other optimized XSAVE instructions), if
129  * a processor implementation detects that an FPU state component is still
130  * (or is again) in its initialized state, it may clear the corresponding
131  * bit in the header.xfeatures field, and can skip the writeout of registers
132  * to the corresponding memory layout.
133  *
134  * This means that when the bit is zero, the state component might still contain
135  * some previous - non-initialized register state.
136  *
137  * Before writing xstate information to user-space we sanitize those components,
138  * to always ensure that the memory layout of a feature will be in the init state
139  * if the corresponding header bit is zero. This is to ensure that user-space doesn't
140  * see some stale state in the memory layout during signal handling, debugging etc.
141  */
142 void fpstate_sanitize_xstate(struct fpu *fpu)
143 {
144 	struct fxregs_state *fx = &fpu->state.fxsave;
145 	int feature_bit;
146 	u64 xfeatures;
147 
148 	if (!use_xsaveopt())
149 		return;
150 
151 	xfeatures = fpu->state.xsave.header.xfeatures;
152 
153 	/*
154 	 * None of the feature bits are in init state. So nothing else
155 	 * to do for us, as the memory layout is up to date.
156 	 */
157 	if ((xfeatures & xfeatures_mask_all) == xfeatures_mask_all)
158 		return;
159 
160 	/*
161 	 * FP is in init state
162 	 */
163 	if (!(xfeatures & XFEATURE_MASK_FP)) {
164 		fx->cwd = 0x37f;
165 		fx->swd = 0;
166 		fx->twd = 0;
167 		fx->fop = 0;
168 		fx->rip = 0;
169 		fx->rdp = 0;
170 		memset(&fx->st_space[0], 0, 128);
171 	}
172 
173 	/*
174 	 * SSE is in init state
175 	 */
176 	if (!(xfeatures & XFEATURE_MASK_SSE))
177 		memset(&fx->xmm_space[0], 0, 256);
178 
179 	/*
180 	 * First two features are FPU and SSE, which above we handled
181 	 * in a special way already:
182 	 */
183 	feature_bit = 0x2;
184 	xfeatures = (xfeatures_mask_user() & ~xfeatures) >> 2;
185 
186 	/*
187 	 * Update all the remaining memory layouts according to their
188 	 * standard xstate layout, if their header bit is in the init
189 	 * state:
190 	 */
191 	while (xfeatures) {
192 		if (xfeatures & 0x1) {
193 			int offset = xstate_comp_offsets[feature_bit];
194 			int size = xstate_sizes[feature_bit];
195 
196 			memcpy((void *)fx + offset,
197 			       (void *)&init_fpstate.xsave + offset,
198 			       size);
199 		}
200 
201 		xfeatures >>= 1;
202 		feature_bit++;
203 	}
204 }
205 
206 /*
207  * Enable the extended processor state save/restore feature.
208  * Called once per CPU onlining.
209  */
210 void fpu__init_cpu_xstate(void)
211 {
212 	u64 unsup_bits;
213 
214 	if (!boot_cpu_has(X86_FEATURE_XSAVE) || !xfeatures_mask_all)
215 		return;
216 	/*
217 	 * Unsupported supervisor xstates should not be found in
218 	 * the xfeatures mask.
219 	 */
220 	unsup_bits = xfeatures_mask_all & XFEATURE_MASK_SUPERVISOR_UNSUPPORTED;
221 	WARN_ONCE(unsup_bits, "x86/fpu: Found unsupported supervisor xstates: 0x%llx\n",
222 		  unsup_bits);
223 
224 	xfeatures_mask_all &= ~XFEATURE_MASK_SUPERVISOR_UNSUPPORTED;
225 
226 	cr4_set_bits(X86_CR4_OSXSAVE);
227 
228 	/*
229 	 * XCR_XFEATURE_ENABLED_MASK (aka. XCR0) sets user features
230 	 * managed by XSAVE{C, OPT, S} and XRSTOR{S}.  Only XSAVE user
231 	 * states can be set here.
232 	 */
233 	xsetbv(XCR_XFEATURE_ENABLED_MASK, xfeatures_mask_user());
234 
235 	/*
236 	 * MSR_IA32_XSS sets supervisor states managed by XSAVES.
237 	 */
238 	if (boot_cpu_has(X86_FEATURE_XSAVES)) {
239 		wrmsrl(MSR_IA32_XSS, xfeatures_mask_supervisor() |
240 				     xfeatures_mask_dynamic());
241 	}
242 }
243 
244 static bool xfeature_enabled(enum xfeature xfeature)
245 {
246 	return xfeatures_mask_all & BIT_ULL(xfeature);
247 }
248 
249 /*
250  * Record the offsets and sizes of various xstates contained
251  * in the XSAVE state memory layout.
252  */
253 static void __init setup_xstate_features(void)
254 {
255 	u32 eax, ebx, ecx, edx, i;
256 	/* start at the beginnning of the "extended state" */
257 	unsigned int last_good_offset = offsetof(struct xregs_state,
258 						 extended_state_area);
259 	/*
260 	 * The FP xstates and SSE xstates are legacy states. They are always
261 	 * in the fixed offsets in the xsave area in either compacted form
262 	 * or standard form.
263 	 */
264 	xstate_offsets[XFEATURE_FP]	= 0;
265 	xstate_sizes[XFEATURE_FP]	= offsetof(struct fxregs_state,
266 						   xmm_space);
267 
268 	xstate_offsets[XFEATURE_SSE]	= xstate_sizes[XFEATURE_FP];
269 	xstate_sizes[XFEATURE_SSE]	= sizeof_field(struct fxregs_state,
270 						       xmm_space);
271 
272 	for (i = FIRST_EXTENDED_XFEATURE; i < XFEATURE_MAX; i++) {
273 		if (!xfeature_enabled(i))
274 			continue;
275 
276 		cpuid_count(XSTATE_CPUID, i, &eax, &ebx, &ecx, &edx);
277 
278 		xstate_sizes[i] = eax;
279 
280 		/*
281 		 * If an xfeature is supervisor state, the offset in EBX is
282 		 * invalid, leave it to -1.
283 		 */
284 		if (xfeature_is_supervisor(i))
285 			continue;
286 
287 		xstate_offsets[i] = ebx;
288 
289 		/*
290 		 * In our xstate size checks, we assume that the highest-numbered
291 		 * xstate feature has the highest offset in the buffer.  Ensure
292 		 * it does.
293 		 */
294 		WARN_ONCE(last_good_offset > xstate_offsets[i],
295 			  "x86/fpu: misordered xstate at %d\n", last_good_offset);
296 
297 		last_good_offset = xstate_offsets[i];
298 	}
299 }
300 
301 static void __init print_xstate_feature(u64 xstate_mask)
302 {
303 	const char *feature_name;
304 
305 	if (cpu_has_xfeatures(xstate_mask, &feature_name))
306 		pr_info("x86/fpu: Supporting XSAVE feature 0x%03Lx: '%s'\n", xstate_mask, feature_name);
307 }
308 
309 /*
310  * Print out all the supported xstate features:
311  */
312 static void __init print_xstate_features(void)
313 {
314 	print_xstate_feature(XFEATURE_MASK_FP);
315 	print_xstate_feature(XFEATURE_MASK_SSE);
316 	print_xstate_feature(XFEATURE_MASK_YMM);
317 	print_xstate_feature(XFEATURE_MASK_BNDREGS);
318 	print_xstate_feature(XFEATURE_MASK_BNDCSR);
319 	print_xstate_feature(XFEATURE_MASK_OPMASK);
320 	print_xstate_feature(XFEATURE_MASK_ZMM_Hi256);
321 	print_xstate_feature(XFEATURE_MASK_Hi16_ZMM);
322 	print_xstate_feature(XFEATURE_MASK_PKRU);
323 	print_xstate_feature(XFEATURE_MASK_PASID);
324 }
325 
326 /*
327  * This check is important because it is easy to get XSTATE_*
328  * confused with XSTATE_BIT_*.
329  */
330 #define CHECK_XFEATURE(nr) do {		\
331 	WARN_ON(nr < FIRST_EXTENDED_XFEATURE);	\
332 	WARN_ON(nr >= XFEATURE_MAX);	\
333 } while (0)
334 
335 /*
336  * We could cache this like xstate_size[], but we only use
337  * it here, so it would be a waste of space.
338  */
339 static int xfeature_is_aligned(int xfeature_nr)
340 {
341 	u32 eax, ebx, ecx, edx;
342 
343 	CHECK_XFEATURE(xfeature_nr);
344 
345 	if (!xfeature_enabled(xfeature_nr)) {
346 		WARN_ONCE(1, "Checking alignment of disabled xfeature %d\n",
347 			  xfeature_nr);
348 		return 0;
349 	}
350 
351 	cpuid_count(XSTATE_CPUID, xfeature_nr, &eax, &ebx, &ecx, &edx);
352 	/*
353 	 * The value returned by ECX[1] indicates the alignment
354 	 * of state component 'i' when the compacted format
355 	 * of the extended region of an XSAVE area is used:
356 	 */
357 	return !!(ecx & 2);
358 }
359 
360 /*
361  * This function sets up offsets and sizes of all extended states in
362  * xsave area. This supports both standard format and compacted format
363  * of the xsave area.
364  */
365 static void __init setup_xstate_comp_offsets(void)
366 {
367 	unsigned int next_offset;
368 	int i;
369 
370 	/*
371 	 * The FP xstates and SSE xstates are legacy states. They are always
372 	 * in the fixed offsets in the xsave area in either compacted form
373 	 * or standard form.
374 	 */
375 	xstate_comp_offsets[XFEATURE_FP] = 0;
376 	xstate_comp_offsets[XFEATURE_SSE] = offsetof(struct fxregs_state,
377 						     xmm_space);
378 
379 	if (!boot_cpu_has(X86_FEATURE_XSAVES)) {
380 		for (i = FIRST_EXTENDED_XFEATURE; i < XFEATURE_MAX; i++) {
381 			if (xfeature_enabled(i))
382 				xstate_comp_offsets[i] = xstate_offsets[i];
383 		}
384 		return;
385 	}
386 
387 	next_offset = FXSAVE_SIZE + XSAVE_HDR_SIZE;
388 
389 	for (i = FIRST_EXTENDED_XFEATURE; i < XFEATURE_MAX; i++) {
390 		if (!xfeature_enabled(i))
391 			continue;
392 
393 		if (xfeature_is_aligned(i))
394 			next_offset = ALIGN(next_offset, 64);
395 
396 		xstate_comp_offsets[i] = next_offset;
397 		next_offset += xstate_sizes[i];
398 	}
399 }
400 
401 /*
402  * Setup offsets of a supervisor-state-only XSAVES buffer:
403  *
404  * The offsets stored in xstate_comp_offsets[] only work for one specific
405  * value of the Requested Feature BitMap (RFBM).  In cases where a different
406  * RFBM value is used, a different set of offsets is required.  This set of
407  * offsets is for when RFBM=xfeatures_mask_supervisor().
408  */
409 static void __init setup_supervisor_only_offsets(void)
410 {
411 	unsigned int next_offset;
412 	int i;
413 
414 	next_offset = FXSAVE_SIZE + XSAVE_HDR_SIZE;
415 
416 	for (i = FIRST_EXTENDED_XFEATURE; i < XFEATURE_MAX; i++) {
417 		if (!xfeature_enabled(i) || !xfeature_is_supervisor(i))
418 			continue;
419 
420 		if (xfeature_is_aligned(i))
421 			next_offset = ALIGN(next_offset, 64);
422 
423 		xstate_supervisor_only_offsets[i] = next_offset;
424 		next_offset += xstate_sizes[i];
425 	}
426 }
427 
428 /*
429  * Print out xstate component offsets and sizes
430  */
431 static void __init print_xstate_offset_size(void)
432 {
433 	int i;
434 
435 	for (i = FIRST_EXTENDED_XFEATURE; i < XFEATURE_MAX; i++) {
436 		if (!xfeature_enabled(i))
437 			continue;
438 		pr_info("x86/fpu: xstate_offset[%d]: %4d, xstate_sizes[%d]: %4d\n",
439 			 i, xstate_comp_offsets[i], i, xstate_sizes[i]);
440 	}
441 }
442 
443 /*
444  * setup the xstate image representing the init state
445  */
446 static void __init setup_init_fpu_buf(void)
447 {
448 	static int on_boot_cpu __initdata = 1;
449 
450 	WARN_ON_FPU(!on_boot_cpu);
451 	on_boot_cpu = 0;
452 
453 	if (!boot_cpu_has(X86_FEATURE_XSAVE))
454 		return;
455 
456 	setup_xstate_features();
457 	print_xstate_features();
458 
459 	if (boot_cpu_has(X86_FEATURE_XSAVES))
460 		init_fpstate.xsave.header.xcomp_bv = XCOMP_BV_COMPACTED_FORMAT |
461 						     xfeatures_mask_all;
462 
463 	/*
464 	 * Init all the features state with header.xfeatures being 0x0
465 	 */
466 	copy_kernel_to_xregs_booting(&init_fpstate.xsave);
467 
468 	/*
469 	 * Dump the init state again. This is to identify the init state
470 	 * of any feature which is not represented by all zero's.
471 	 */
472 	copy_xregs_to_kernel_booting(&init_fpstate.xsave);
473 }
474 
475 static int xfeature_uncompacted_offset(int xfeature_nr)
476 {
477 	u32 eax, ebx, ecx, edx;
478 
479 	/*
480 	 * Only XSAVES supports supervisor states and it uses compacted
481 	 * format. Checking a supervisor state's uncompacted offset is
482 	 * an error.
483 	 */
484 	if (XFEATURE_MASK_SUPERVISOR_ALL & BIT_ULL(xfeature_nr)) {
485 		WARN_ONCE(1, "No fixed offset for xstate %d\n", xfeature_nr);
486 		return -1;
487 	}
488 
489 	CHECK_XFEATURE(xfeature_nr);
490 	cpuid_count(XSTATE_CPUID, xfeature_nr, &eax, &ebx, &ecx, &edx);
491 	return ebx;
492 }
493 
494 int xfeature_size(int xfeature_nr)
495 {
496 	u32 eax, ebx, ecx, edx;
497 
498 	CHECK_XFEATURE(xfeature_nr);
499 	cpuid_count(XSTATE_CPUID, xfeature_nr, &eax, &ebx, &ecx, &edx);
500 	return eax;
501 }
502 
503 /*
504  * 'XSAVES' implies two different things:
505  * 1. saving of supervisor/system state
506  * 2. using the compacted format
507  *
508  * Use this function when dealing with the compacted format so
509  * that it is obvious which aspect of 'XSAVES' is being handled
510  * by the calling code.
511  */
512 int using_compacted_format(void)
513 {
514 	return boot_cpu_has(X86_FEATURE_XSAVES);
515 }
516 
517 /* Validate an xstate header supplied by userspace (ptrace or sigreturn) */
518 int validate_user_xstate_header(const struct xstate_header *hdr)
519 {
520 	/* No unknown or supervisor features may be set */
521 	if (hdr->xfeatures & ~xfeatures_mask_user())
522 		return -EINVAL;
523 
524 	/* Userspace must use the uncompacted format */
525 	if (hdr->xcomp_bv)
526 		return -EINVAL;
527 
528 	/*
529 	 * If 'reserved' is shrunken to add a new field, make sure to validate
530 	 * that new field here!
531 	 */
532 	BUILD_BUG_ON(sizeof(hdr->reserved) != 48);
533 
534 	/* No reserved bits may be set */
535 	if (memchr_inv(hdr->reserved, 0, sizeof(hdr->reserved)))
536 		return -EINVAL;
537 
538 	return 0;
539 }
540 
541 static void __xstate_dump_leaves(void)
542 {
543 	int i;
544 	u32 eax, ebx, ecx, edx;
545 	static int should_dump = 1;
546 
547 	if (!should_dump)
548 		return;
549 	should_dump = 0;
550 	/*
551 	 * Dump out a few leaves past the ones that we support
552 	 * just in case there are some goodies up there
553 	 */
554 	for (i = 0; i < XFEATURE_MAX + 10; i++) {
555 		cpuid_count(XSTATE_CPUID, i, &eax, &ebx, &ecx, &edx);
556 		pr_warn("CPUID[%02x, %02x]: eax=%08x ebx=%08x ecx=%08x edx=%08x\n",
557 			XSTATE_CPUID, i, eax, ebx, ecx, edx);
558 	}
559 }
560 
561 #define XSTATE_WARN_ON(x) do {							\
562 	if (WARN_ONCE(x, "XSAVE consistency problem, dumping leaves")) {	\
563 		__xstate_dump_leaves();						\
564 	}									\
565 } while (0)
566 
567 #define XCHECK_SZ(sz, nr, nr_macro, __struct) do {			\
568 	if ((nr == nr_macro) &&						\
569 	    WARN_ONCE(sz != sizeof(__struct),				\
570 		"%s: struct is %zu bytes, cpu state %d bytes\n",	\
571 		__stringify(nr_macro), sizeof(__struct), sz)) {		\
572 		__xstate_dump_leaves();					\
573 	}								\
574 } while (0)
575 
576 /*
577  * We have a C struct for each 'xstate'.  We need to ensure
578  * that our software representation matches what the CPU
579  * tells us about the state's size.
580  */
581 static void check_xstate_against_struct(int nr)
582 {
583 	/*
584 	 * Ask the CPU for the size of the state.
585 	 */
586 	int sz = xfeature_size(nr);
587 	/*
588 	 * Match each CPU state with the corresponding software
589 	 * structure.
590 	 */
591 	XCHECK_SZ(sz, nr, XFEATURE_YMM,       struct ymmh_struct);
592 	XCHECK_SZ(sz, nr, XFEATURE_BNDREGS,   struct mpx_bndreg_state);
593 	XCHECK_SZ(sz, nr, XFEATURE_BNDCSR,    struct mpx_bndcsr_state);
594 	XCHECK_SZ(sz, nr, XFEATURE_OPMASK,    struct avx_512_opmask_state);
595 	XCHECK_SZ(sz, nr, XFEATURE_ZMM_Hi256, struct avx_512_zmm_uppers_state);
596 	XCHECK_SZ(sz, nr, XFEATURE_Hi16_ZMM,  struct avx_512_hi16_state);
597 	XCHECK_SZ(sz, nr, XFEATURE_PKRU,      struct pkru_state);
598 	XCHECK_SZ(sz, nr, XFEATURE_PASID,     struct ia32_pasid_state);
599 
600 	/*
601 	 * Make *SURE* to add any feature numbers in below if
602 	 * there are "holes" in the xsave state component
603 	 * numbers.
604 	 */
605 	if ((nr < XFEATURE_YMM) ||
606 	    (nr >= XFEATURE_MAX) ||
607 	    (nr == XFEATURE_PT_UNIMPLEMENTED_SO_FAR) ||
608 	    ((nr >= XFEATURE_RSRVD_COMP_11) && (nr <= XFEATURE_LBR))) {
609 		WARN_ONCE(1, "no structure for xstate: %d\n", nr);
610 		XSTATE_WARN_ON(1);
611 	}
612 }
613 
614 /*
615  * This essentially double-checks what the cpu told us about
616  * how large the XSAVE buffer needs to be.  We are recalculating
617  * it to be safe.
618  *
619  * Dynamic XSAVE features allocate their own buffers and are not
620  * covered by these checks. Only the size of the buffer for task->fpu
621  * is checked here.
622  */
623 static void do_extra_xstate_size_checks(void)
624 {
625 	int paranoid_xstate_size = FXSAVE_SIZE + XSAVE_HDR_SIZE;
626 	int i;
627 
628 	for (i = FIRST_EXTENDED_XFEATURE; i < XFEATURE_MAX; i++) {
629 		if (!xfeature_enabled(i))
630 			continue;
631 
632 		check_xstate_against_struct(i);
633 		/*
634 		 * Supervisor state components can be managed only by
635 		 * XSAVES, which is compacted-format only.
636 		 */
637 		if (!using_compacted_format())
638 			XSTATE_WARN_ON(xfeature_is_supervisor(i));
639 
640 		/* Align from the end of the previous feature */
641 		if (xfeature_is_aligned(i))
642 			paranoid_xstate_size = ALIGN(paranoid_xstate_size, 64);
643 		/*
644 		 * The offset of a given state in the non-compacted
645 		 * format is given to us in a CPUID leaf.  We check
646 		 * them for being ordered (increasing offsets) in
647 		 * setup_xstate_features().
648 		 */
649 		if (!using_compacted_format())
650 			paranoid_xstate_size = xfeature_uncompacted_offset(i);
651 		/*
652 		 * The compacted-format offset always depends on where
653 		 * the previous state ended.
654 		 */
655 		paranoid_xstate_size += xfeature_size(i);
656 	}
657 	XSTATE_WARN_ON(paranoid_xstate_size != fpu_kernel_xstate_size);
658 }
659 
660 
661 /*
662  * Get total size of enabled xstates in XCR0 | IA32_XSS.
663  *
664  * Note the SDM's wording here.  "sub-function 0" only enumerates
665  * the size of the *user* states.  If we use it to size a buffer
666  * that we use 'XSAVES' on, we could potentially overflow the
667  * buffer because 'XSAVES' saves system states too.
668  */
669 static unsigned int __init get_xsaves_size(void)
670 {
671 	unsigned int eax, ebx, ecx, edx;
672 	/*
673 	 * - CPUID function 0DH, sub-function 1:
674 	 *    EBX enumerates the size (in bytes) required by
675 	 *    the XSAVES instruction for an XSAVE area
676 	 *    containing all the state components
677 	 *    corresponding to bits currently set in
678 	 *    XCR0 | IA32_XSS.
679 	 */
680 	cpuid_count(XSTATE_CPUID, 1, &eax, &ebx, &ecx, &edx);
681 	return ebx;
682 }
683 
684 /*
685  * Get the total size of the enabled xstates without the dynamic supervisor
686  * features.
687  */
688 static unsigned int __init get_xsaves_size_no_dynamic(void)
689 {
690 	u64 mask = xfeatures_mask_dynamic();
691 	unsigned int size;
692 
693 	if (!mask)
694 		return get_xsaves_size();
695 
696 	/* Disable dynamic features. */
697 	wrmsrl(MSR_IA32_XSS, xfeatures_mask_supervisor());
698 
699 	/*
700 	 * Ask the hardware what size is required of the buffer.
701 	 * This is the size required for the task->fpu buffer.
702 	 */
703 	size = get_xsaves_size();
704 
705 	/* Re-enable dynamic features so XSAVES will work on them again. */
706 	wrmsrl(MSR_IA32_XSS, xfeatures_mask_supervisor() | mask);
707 
708 	return size;
709 }
710 
711 static unsigned int __init get_xsave_size(void)
712 {
713 	unsigned int eax, ebx, ecx, edx;
714 	/*
715 	 * - CPUID function 0DH, sub-function 0:
716 	 *    EBX enumerates the size (in bytes) required by
717 	 *    the XSAVE instruction for an XSAVE area
718 	 *    containing all the *user* state components
719 	 *    corresponding to bits currently set in XCR0.
720 	 */
721 	cpuid_count(XSTATE_CPUID, 0, &eax, &ebx, &ecx, &edx);
722 	return ebx;
723 }
724 
725 /*
726  * Will the runtime-enumerated 'xstate_size' fit in the init
727  * task's statically-allocated buffer?
728  */
729 static bool is_supported_xstate_size(unsigned int test_xstate_size)
730 {
731 	if (test_xstate_size <= sizeof(union fpregs_state))
732 		return true;
733 
734 	pr_warn("x86/fpu: xstate buffer too small (%zu < %d), disabling xsave\n",
735 			sizeof(union fpregs_state), test_xstate_size);
736 	return false;
737 }
738 
739 static int __init init_xstate_size(void)
740 {
741 	/* Recompute the context size for enabled features: */
742 	unsigned int possible_xstate_size;
743 	unsigned int xsave_size;
744 
745 	xsave_size = get_xsave_size();
746 
747 	if (boot_cpu_has(X86_FEATURE_XSAVES))
748 		possible_xstate_size = get_xsaves_size_no_dynamic();
749 	else
750 		possible_xstate_size = xsave_size;
751 
752 	/* Ensure we have the space to store all enabled: */
753 	if (!is_supported_xstate_size(possible_xstate_size))
754 		return -EINVAL;
755 
756 	/*
757 	 * The size is OK, we are definitely going to use xsave,
758 	 * make it known to the world that we need more space.
759 	 */
760 	fpu_kernel_xstate_size = possible_xstate_size;
761 	do_extra_xstate_size_checks();
762 
763 	/*
764 	 * User space is always in standard format.
765 	 */
766 	fpu_user_xstate_size = xsave_size;
767 	return 0;
768 }
769 
770 /*
771  * We enabled the XSAVE hardware, but something went wrong and
772  * we can not use it.  Disable it.
773  */
774 static void fpu__init_disable_system_xstate(void)
775 {
776 	xfeatures_mask_all = 0;
777 	cr4_clear_bits(X86_CR4_OSXSAVE);
778 	setup_clear_cpu_cap(X86_FEATURE_XSAVE);
779 }
780 
781 /*
782  * Enable and initialize the xsave feature.
783  * Called once per system bootup.
784  */
785 void __init fpu__init_system_xstate(void)
786 {
787 	unsigned int eax, ebx, ecx, edx;
788 	static int on_boot_cpu __initdata = 1;
789 	int err;
790 	int i;
791 
792 	WARN_ON_FPU(!on_boot_cpu);
793 	on_boot_cpu = 0;
794 
795 	if (!boot_cpu_has(X86_FEATURE_FPU)) {
796 		pr_info("x86/fpu: No FPU detected\n");
797 		return;
798 	}
799 
800 	if (!boot_cpu_has(X86_FEATURE_XSAVE)) {
801 		pr_info("x86/fpu: x87 FPU will use %s\n",
802 			boot_cpu_has(X86_FEATURE_FXSR) ? "FXSAVE" : "FSAVE");
803 		return;
804 	}
805 
806 	if (boot_cpu_data.cpuid_level < XSTATE_CPUID) {
807 		WARN_ON_FPU(1);
808 		return;
809 	}
810 
811 	/*
812 	 * Find user xstates supported by the processor.
813 	 */
814 	cpuid_count(XSTATE_CPUID, 0, &eax, &ebx, &ecx, &edx);
815 	xfeatures_mask_all = eax + ((u64)edx << 32);
816 
817 	/*
818 	 * Find supervisor xstates supported by the processor.
819 	 */
820 	cpuid_count(XSTATE_CPUID, 1, &eax, &ebx, &ecx, &edx);
821 	xfeatures_mask_all |= ecx + ((u64)edx << 32);
822 
823 	if ((xfeatures_mask_user() & XFEATURE_MASK_FPSSE) != XFEATURE_MASK_FPSSE) {
824 		/*
825 		 * This indicates that something really unexpected happened
826 		 * with the enumeration.  Disable XSAVE and try to continue
827 		 * booting without it.  This is too early to BUG().
828 		 */
829 		pr_err("x86/fpu: FP/SSE not present amongst the CPU's xstate features: 0x%llx.\n",
830 		       xfeatures_mask_all);
831 		goto out_disable;
832 	}
833 
834 	/*
835 	 * Clear XSAVE features that are disabled in the normal CPUID.
836 	 */
837 	for (i = 0; i < ARRAY_SIZE(xsave_cpuid_features); i++) {
838 		if (!boot_cpu_has(xsave_cpuid_features[i]))
839 			xfeatures_mask_all &= ~BIT_ULL(i);
840 	}
841 
842 	xfeatures_mask_all &= fpu__get_supported_xfeatures_mask();
843 
844 	/* Enable xstate instructions to be able to continue with initialization: */
845 	fpu__init_cpu_xstate();
846 	err = init_xstate_size();
847 	if (err)
848 		goto out_disable;
849 
850 	/*
851 	 * Update info used for ptrace frames; use standard-format size and no
852 	 * supervisor xstates:
853 	 */
854 	update_regset_xstate_info(fpu_user_xstate_size, xfeatures_mask_user());
855 
856 	fpu__init_prepare_fx_sw_frame();
857 	setup_init_fpu_buf();
858 	setup_xstate_comp_offsets();
859 	setup_supervisor_only_offsets();
860 	print_xstate_offset_size();
861 
862 	pr_info("x86/fpu: Enabled xstate features 0x%llx, context size is %d bytes, using '%s' format.\n",
863 		xfeatures_mask_all,
864 		fpu_kernel_xstate_size,
865 		boot_cpu_has(X86_FEATURE_XSAVES) ? "compacted" : "standard");
866 	return;
867 
868 out_disable:
869 	/* something went wrong, try to boot without any XSAVE support */
870 	fpu__init_disable_system_xstate();
871 }
872 
873 /*
874  * Restore minimal FPU state after suspend:
875  */
876 void fpu__resume_cpu(void)
877 {
878 	/*
879 	 * Restore XCR0 on xsave capable CPUs:
880 	 */
881 	if (boot_cpu_has(X86_FEATURE_XSAVE))
882 		xsetbv(XCR_XFEATURE_ENABLED_MASK, xfeatures_mask_user());
883 
884 	/*
885 	 * Restore IA32_XSS. The same CPUID bit enumerates support
886 	 * of XSAVES and MSR_IA32_XSS.
887 	 */
888 	if (boot_cpu_has(X86_FEATURE_XSAVES)) {
889 		wrmsrl(MSR_IA32_XSS, xfeatures_mask_supervisor()  |
890 				     xfeatures_mask_dynamic());
891 	}
892 }
893 
894 /*
895  * Given an xstate feature nr, calculate where in the xsave
896  * buffer the state is.  Callers should ensure that the buffer
897  * is valid.
898  */
899 static void *__raw_xsave_addr(struct xregs_state *xsave, int xfeature_nr)
900 {
901 	if (!xfeature_enabled(xfeature_nr)) {
902 		WARN_ON_FPU(1);
903 		return NULL;
904 	}
905 
906 	return (void *)xsave + xstate_comp_offsets[xfeature_nr];
907 }
908 /*
909  * Given the xsave area and a state inside, this function returns the
910  * address of the state.
911  *
912  * This is the API that is called to get xstate address in either
913  * standard format or compacted format of xsave area.
914  *
915  * Note that if there is no data for the field in the xsave buffer
916  * this will return NULL.
917  *
918  * Inputs:
919  *	xstate: the thread's storage area for all FPU data
920  *	xfeature_nr: state which is defined in xsave.h (e.g. XFEATURE_FP,
921  *	XFEATURE_SSE, etc...)
922  * Output:
923  *	address of the state in the xsave area, or NULL if the
924  *	field is not present in the xsave buffer.
925  */
926 void *get_xsave_addr(struct xregs_state *xsave, int xfeature_nr)
927 {
928 	/*
929 	 * Do we even *have* xsave state?
930 	 */
931 	if (!boot_cpu_has(X86_FEATURE_XSAVE))
932 		return NULL;
933 
934 	/*
935 	 * We should not ever be requesting features that we
936 	 * have not enabled.
937 	 */
938 	WARN_ONCE(!(xfeatures_mask_all & BIT_ULL(xfeature_nr)),
939 		  "get of unsupported state");
940 	/*
941 	 * This assumes the last 'xsave*' instruction to
942 	 * have requested that 'xfeature_nr' be saved.
943 	 * If it did not, we might be seeing and old value
944 	 * of the field in the buffer.
945 	 *
946 	 * This can happen because the last 'xsave' did not
947 	 * request that this feature be saved (unlikely)
948 	 * or because the "init optimization" caused it
949 	 * to not be saved.
950 	 */
951 	if (!(xsave->header.xfeatures & BIT_ULL(xfeature_nr)))
952 		return NULL;
953 
954 	return __raw_xsave_addr(xsave, xfeature_nr);
955 }
956 EXPORT_SYMBOL_GPL(get_xsave_addr);
957 
958 /*
959  * This wraps up the common operations that need to occur when retrieving
960  * data from xsave state.  It first ensures that the current task was
961  * using the FPU and retrieves the data in to a buffer.  It then calculates
962  * the offset of the requested field in the buffer.
963  *
964  * This function is safe to call whether the FPU is in use or not.
965  *
966  * Note that this only works on the current task.
967  *
968  * Inputs:
969  *	@xfeature_nr: state which is defined in xsave.h (e.g. XFEATURE_FP,
970  *	XFEATURE_SSE, etc...)
971  * Output:
972  *	address of the state in the xsave area or NULL if the state
973  *	is not present or is in its 'init state'.
974  */
975 const void *get_xsave_field_ptr(int xfeature_nr)
976 {
977 	struct fpu *fpu = &current->thread.fpu;
978 
979 	/*
980 	 * fpu__save() takes the CPU's xstate registers
981 	 * and saves them off to the 'fpu memory buffer.
982 	 */
983 	fpu__save(fpu);
984 
985 	return get_xsave_addr(&fpu->state.xsave, xfeature_nr);
986 }
987 
988 #ifdef CONFIG_ARCH_HAS_PKEYS
989 
990 /*
991  * This will go out and modify PKRU register to set the access
992  * rights for @pkey to @init_val.
993  */
994 int arch_set_user_pkey_access(struct task_struct *tsk, int pkey,
995 		unsigned long init_val)
996 {
997 	u32 old_pkru;
998 	int pkey_shift = (pkey * PKRU_BITS_PER_PKEY);
999 	u32 new_pkru_bits = 0;
1000 
1001 	/*
1002 	 * This check implies XSAVE support.  OSPKE only gets
1003 	 * set if we enable XSAVE and we enable PKU in XCR0.
1004 	 */
1005 	if (!boot_cpu_has(X86_FEATURE_OSPKE))
1006 		return -EINVAL;
1007 
1008 	/*
1009 	 * This code should only be called with valid 'pkey'
1010 	 * values originating from in-kernel users.  Complain
1011 	 * if a bad value is observed.
1012 	 */
1013 	WARN_ON_ONCE(pkey >= arch_max_pkey());
1014 
1015 	/* Set the bits we need in PKRU:  */
1016 	if (init_val & PKEY_DISABLE_ACCESS)
1017 		new_pkru_bits |= PKRU_AD_BIT;
1018 	if (init_val & PKEY_DISABLE_WRITE)
1019 		new_pkru_bits |= PKRU_WD_BIT;
1020 
1021 	/* Shift the bits in to the correct place in PKRU for pkey: */
1022 	new_pkru_bits <<= pkey_shift;
1023 
1024 	/* Get old PKRU and mask off any old bits in place: */
1025 	old_pkru = read_pkru();
1026 	old_pkru &= ~((PKRU_AD_BIT|PKRU_WD_BIT) << pkey_shift);
1027 
1028 	/* Write old part along with new part: */
1029 	write_pkru(old_pkru | new_pkru_bits);
1030 
1031 	return 0;
1032 }
1033 #endif /* ! CONFIG_ARCH_HAS_PKEYS */
1034 
1035 /*
1036  * Weird legacy quirk: SSE and YMM states store information in the
1037  * MXCSR and MXCSR_FLAGS fields of the FP area. That means if the FP
1038  * area is marked as unused in the xfeatures header, we need to copy
1039  * MXCSR and MXCSR_FLAGS if either SSE or YMM are in use.
1040  */
1041 static inline bool xfeatures_mxcsr_quirk(u64 xfeatures)
1042 {
1043 	if (!(xfeatures & (XFEATURE_MASK_SSE|XFEATURE_MASK_YMM)))
1044 		return false;
1045 
1046 	if (xfeatures & XFEATURE_MASK_FP)
1047 		return false;
1048 
1049 	return true;
1050 }
1051 
1052 static void fill_gap(struct membuf *to, unsigned *last, unsigned offset)
1053 {
1054 	if (*last >= offset)
1055 		return;
1056 	membuf_write(to, (void *)&init_fpstate.xsave + *last, offset - *last);
1057 	*last = offset;
1058 }
1059 
1060 static void copy_part(struct membuf *to, unsigned *last, unsigned offset,
1061 		      unsigned size, void *from)
1062 {
1063 	fill_gap(to, last, offset);
1064 	membuf_write(to, from, size);
1065 	*last = offset + size;
1066 }
1067 
1068 /*
1069  * Convert from kernel XSAVES compacted format to standard format and copy
1070  * to a kernel-space ptrace buffer.
1071  *
1072  * It supports partial copy but pos always starts from zero. This is called
1073  * from xstateregs_get() and there we check the CPU has XSAVES.
1074  */
1075 void copy_xstate_to_kernel(struct membuf to, struct xregs_state *xsave)
1076 {
1077 	struct xstate_header header;
1078 	const unsigned off_mxcsr = offsetof(struct fxregs_state, mxcsr);
1079 	unsigned size = to.left;
1080 	unsigned last = 0;
1081 	int i;
1082 
1083 	/*
1084 	 * The destination is a ptrace buffer; we put in only user xstates:
1085 	 */
1086 	memset(&header, 0, sizeof(header));
1087 	header.xfeatures = xsave->header.xfeatures;
1088 	header.xfeatures &= xfeatures_mask_user();
1089 
1090 	if (header.xfeatures & XFEATURE_MASK_FP)
1091 		copy_part(&to, &last, 0, off_mxcsr, &xsave->i387);
1092 	if (header.xfeatures & (XFEATURE_MASK_SSE | XFEATURE_MASK_YMM))
1093 		copy_part(&to, &last, off_mxcsr,
1094 			  MXCSR_AND_FLAGS_SIZE, &xsave->i387.mxcsr);
1095 	if (header.xfeatures & XFEATURE_MASK_FP)
1096 		copy_part(&to, &last, offsetof(struct fxregs_state, st_space),
1097 			  128, &xsave->i387.st_space);
1098 	if (header.xfeatures & XFEATURE_MASK_SSE)
1099 		copy_part(&to, &last, xstate_offsets[XFEATURE_SSE],
1100 			  256, &xsave->i387.xmm_space);
1101 	/*
1102 	 * Fill xsave->i387.sw_reserved value for ptrace frame:
1103 	 */
1104 	copy_part(&to, &last, offsetof(struct fxregs_state, sw_reserved),
1105 		  48, xstate_fx_sw_bytes);
1106 	/*
1107 	 * Copy xregs_state->header:
1108 	 */
1109 	copy_part(&to, &last, offsetof(struct xregs_state, header),
1110 		  sizeof(header), &header);
1111 
1112 	for (i = FIRST_EXTENDED_XFEATURE; i < XFEATURE_MAX; i++) {
1113 		/*
1114 		 * Copy only in-use xstates:
1115 		 */
1116 		if ((header.xfeatures >> i) & 1) {
1117 			void *src = __raw_xsave_addr(xsave, i);
1118 
1119 			copy_part(&to, &last, xstate_offsets[i],
1120 				  xstate_sizes[i], src);
1121 		}
1122 
1123 	}
1124 	fill_gap(&to, &last, size);
1125 }
1126 
1127 /*
1128  * Convert from a ptrace standard-format kernel buffer to kernel XSAVES format
1129  * and copy to the target thread. This is called from xstateregs_set().
1130  */
1131 int copy_kernel_to_xstate(struct xregs_state *xsave, const void *kbuf)
1132 {
1133 	unsigned int offset, size;
1134 	int i;
1135 	struct xstate_header hdr;
1136 
1137 	offset = offsetof(struct xregs_state, header);
1138 	size = sizeof(hdr);
1139 
1140 	memcpy(&hdr, kbuf + offset, size);
1141 
1142 	if (validate_user_xstate_header(&hdr))
1143 		return -EINVAL;
1144 
1145 	for (i = 0; i < XFEATURE_MAX; i++) {
1146 		u64 mask = ((u64)1 << i);
1147 
1148 		if (hdr.xfeatures & mask) {
1149 			void *dst = __raw_xsave_addr(xsave, i);
1150 
1151 			offset = xstate_offsets[i];
1152 			size = xstate_sizes[i];
1153 
1154 			memcpy(dst, kbuf + offset, size);
1155 		}
1156 	}
1157 
1158 	if (xfeatures_mxcsr_quirk(hdr.xfeatures)) {
1159 		offset = offsetof(struct fxregs_state, mxcsr);
1160 		size = MXCSR_AND_FLAGS_SIZE;
1161 		memcpy(&xsave->i387.mxcsr, kbuf + offset, size);
1162 	}
1163 
1164 	/*
1165 	 * The state that came in from userspace was user-state only.
1166 	 * Mask all the user states out of 'xfeatures':
1167 	 */
1168 	xsave->header.xfeatures &= XFEATURE_MASK_SUPERVISOR_ALL;
1169 
1170 	/*
1171 	 * Add back in the features that came in from userspace:
1172 	 */
1173 	xsave->header.xfeatures |= hdr.xfeatures;
1174 
1175 	return 0;
1176 }
1177 
1178 /*
1179  * Convert from a ptrace or sigreturn standard-format user-space buffer to
1180  * kernel XSAVES format and copy to the target thread. This is called from
1181  * xstateregs_set(), as well as potentially from the sigreturn() and
1182  * rt_sigreturn() system calls.
1183  */
1184 int copy_user_to_xstate(struct xregs_state *xsave, const void __user *ubuf)
1185 {
1186 	unsigned int offset, size;
1187 	int i;
1188 	struct xstate_header hdr;
1189 
1190 	offset = offsetof(struct xregs_state, header);
1191 	size = sizeof(hdr);
1192 
1193 	if (__copy_from_user(&hdr, ubuf + offset, size))
1194 		return -EFAULT;
1195 
1196 	if (validate_user_xstate_header(&hdr))
1197 		return -EINVAL;
1198 
1199 	for (i = 0; i < XFEATURE_MAX; i++) {
1200 		u64 mask = ((u64)1 << i);
1201 
1202 		if (hdr.xfeatures & mask) {
1203 			void *dst = __raw_xsave_addr(xsave, i);
1204 
1205 			offset = xstate_offsets[i];
1206 			size = xstate_sizes[i];
1207 
1208 			if (__copy_from_user(dst, ubuf + offset, size))
1209 				return -EFAULT;
1210 		}
1211 	}
1212 
1213 	if (xfeatures_mxcsr_quirk(hdr.xfeatures)) {
1214 		offset = offsetof(struct fxregs_state, mxcsr);
1215 		size = MXCSR_AND_FLAGS_SIZE;
1216 		if (__copy_from_user(&xsave->i387.mxcsr, ubuf + offset, size))
1217 			return -EFAULT;
1218 	}
1219 
1220 	/*
1221 	 * The state that came in from userspace was user-state only.
1222 	 * Mask all the user states out of 'xfeatures':
1223 	 */
1224 	xsave->header.xfeatures &= XFEATURE_MASK_SUPERVISOR_ALL;
1225 
1226 	/*
1227 	 * Add back in the features that came in from userspace:
1228 	 */
1229 	xsave->header.xfeatures |= hdr.xfeatures;
1230 
1231 	return 0;
1232 }
1233 
1234 /*
1235  * Save only supervisor states to the kernel buffer.  This blows away all
1236  * old states, and is intended to be used only in __fpu__restore_sig(), where
1237  * user states are restored from the user buffer.
1238  */
1239 void copy_supervisor_to_kernel(struct xregs_state *xstate)
1240 {
1241 	struct xstate_header *header;
1242 	u64 max_bit, min_bit;
1243 	u32 lmask, hmask;
1244 	int err, i;
1245 
1246 	if (WARN_ON(!boot_cpu_has(X86_FEATURE_XSAVES)))
1247 		return;
1248 
1249 	if (!xfeatures_mask_supervisor())
1250 		return;
1251 
1252 	max_bit = __fls(xfeatures_mask_supervisor());
1253 	min_bit = __ffs(xfeatures_mask_supervisor());
1254 
1255 	lmask = xfeatures_mask_supervisor();
1256 	hmask = xfeatures_mask_supervisor() >> 32;
1257 	XSTATE_OP(XSAVES, xstate, lmask, hmask, err);
1258 
1259 	/* We should never fault when copying to a kernel buffer: */
1260 	if (WARN_ON_FPU(err))
1261 		return;
1262 
1263 	/*
1264 	 * At this point, the buffer has only supervisor states and must be
1265 	 * converted back to normal kernel format.
1266 	 */
1267 	header = &xstate->header;
1268 	header->xcomp_bv |= xfeatures_mask_all;
1269 
1270 	/*
1271 	 * This only moves states up in the buffer.  Start with
1272 	 * the last state and move backwards so that states are
1273 	 * not overwritten until after they are moved.  Note:
1274 	 * memmove() allows overlapping src/dst buffers.
1275 	 */
1276 	for (i = max_bit; i >= min_bit; i--) {
1277 		u8 *xbuf = (u8 *)xstate;
1278 
1279 		if (!((header->xfeatures >> i) & 1))
1280 			continue;
1281 
1282 		/* Move xfeature 'i' into its normal location */
1283 		memmove(xbuf + xstate_comp_offsets[i],
1284 			xbuf + xstate_supervisor_only_offsets[i],
1285 			xstate_sizes[i]);
1286 	}
1287 }
1288 
1289 /**
1290  * copy_dynamic_supervisor_to_kernel() - Save dynamic supervisor states to
1291  *                                       an xsave area
1292  * @xstate: A pointer to an xsave area
1293  * @mask: Represent the dynamic supervisor features saved into the xsave area
1294  *
1295  * Only the dynamic supervisor states sets in the mask are saved into the xsave
1296  * area (See the comment in XFEATURE_MASK_DYNAMIC for the details of dynamic
1297  * supervisor feature). Besides the dynamic supervisor states, the legacy
1298  * region and XSAVE header are also saved into the xsave area. The supervisor
1299  * features in the XFEATURE_MASK_SUPERVISOR_SUPPORTED and
1300  * XFEATURE_MASK_SUPERVISOR_UNSUPPORTED are not saved.
1301  *
1302  * The xsave area must be 64-bytes aligned.
1303  */
1304 void copy_dynamic_supervisor_to_kernel(struct xregs_state *xstate, u64 mask)
1305 {
1306 	u64 dynamic_mask = xfeatures_mask_dynamic() & mask;
1307 	u32 lmask, hmask;
1308 	int err;
1309 
1310 	if (WARN_ON_FPU(!boot_cpu_has(X86_FEATURE_XSAVES)))
1311 		return;
1312 
1313 	if (WARN_ON_FPU(!dynamic_mask))
1314 		return;
1315 
1316 	lmask = dynamic_mask;
1317 	hmask = dynamic_mask >> 32;
1318 
1319 	XSTATE_OP(XSAVES, xstate, lmask, hmask, err);
1320 
1321 	/* Should never fault when copying to a kernel buffer */
1322 	WARN_ON_FPU(err);
1323 }
1324 
1325 /**
1326  * copy_kernel_to_dynamic_supervisor() - Restore dynamic supervisor states from
1327  *                                       an xsave area
1328  * @xstate: A pointer to an xsave area
1329  * @mask: Represent the dynamic supervisor features restored from the xsave area
1330  *
1331  * Only the dynamic supervisor states sets in the mask are restored from the
1332  * xsave area (See the comment in XFEATURE_MASK_DYNAMIC for the details of
1333  * dynamic supervisor feature). Besides the dynamic supervisor states, the
1334  * legacy region and XSAVE header are also restored from the xsave area. The
1335  * supervisor features in the XFEATURE_MASK_SUPERVISOR_SUPPORTED and
1336  * XFEATURE_MASK_SUPERVISOR_UNSUPPORTED are not restored.
1337  *
1338  * The xsave area must be 64-bytes aligned.
1339  */
1340 void copy_kernel_to_dynamic_supervisor(struct xregs_state *xstate, u64 mask)
1341 {
1342 	u64 dynamic_mask = xfeatures_mask_dynamic() & mask;
1343 	u32 lmask, hmask;
1344 	int err;
1345 
1346 	if (WARN_ON_FPU(!boot_cpu_has(X86_FEATURE_XSAVES)))
1347 		return;
1348 
1349 	if (WARN_ON_FPU(!dynamic_mask))
1350 		return;
1351 
1352 	lmask = dynamic_mask;
1353 	hmask = dynamic_mask >> 32;
1354 
1355 	XSTATE_OP(XRSTORS, xstate, lmask, hmask, err);
1356 
1357 	/* Should never fault when copying from a kernel buffer */
1358 	WARN_ON_FPU(err);
1359 }
1360 
1361 #ifdef CONFIG_PROC_PID_ARCH_STATUS
1362 /*
1363  * Report the amount of time elapsed in millisecond since last AVX512
1364  * use in the task.
1365  */
1366 static void avx512_status(struct seq_file *m, struct task_struct *task)
1367 {
1368 	unsigned long timestamp = READ_ONCE(task->thread.fpu.avx512_timestamp);
1369 	long delta;
1370 
1371 	if (!timestamp) {
1372 		/*
1373 		 * Report -1 if no AVX512 usage
1374 		 */
1375 		delta = -1;
1376 	} else {
1377 		delta = (long)(jiffies - timestamp);
1378 		/*
1379 		 * Cap to LONG_MAX if time difference > LONG_MAX
1380 		 */
1381 		if (delta < 0)
1382 			delta = LONG_MAX;
1383 		delta = jiffies_to_msecs(delta);
1384 	}
1385 
1386 	seq_put_decimal_ll(m, "AVX512_elapsed_ms:\t", delta);
1387 	seq_putc(m, '\n');
1388 }
1389 
1390 /*
1391  * Report architecture specific information
1392  */
1393 int proc_pid_arch_status(struct seq_file *m, struct pid_namespace *ns,
1394 			struct pid *pid, struct task_struct *task)
1395 {
1396 	/*
1397 	 * Report AVX512 state if the processor and build option supported.
1398 	 */
1399 	if (cpu_feature_enabled(X86_FEATURE_AVX512F))
1400 		avx512_status(m, task);
1401 
1402 	return 0;
1403 }
1404 #endif /* CONFIG_PROC_PID_ARCH_STATUS */
1405 
1406 #ifdef CONFIG_IOMMU_SUPPORT
1407 void update_pasid(void)
1408 {
1409 	u64 pasid_state;
1410 	u32 pasid;
1411 
1412 	if (!cpu_feature_enabled(X86_FEATURE_ENQCMD))
1413 		return;
1414 
1415 	if (!current->mm)
1416 		return;
1417 
1418 	pasid = READ_ONCE(current->mm->pasid);
1419 	/* Set the valid bit in the PASID MSR/state only for valid pasid. */
1420 	pasid_state = pasid == PASID_DISABLED ?
1421 		      pasid : pasid | MSR_IA32_PASID_VALID;
1422 
1423 	/*
1424 	 * No need to hold fregs_lock() since the task's fpstate won't
1425 	 * be changed by others (e.g. ptrace) while the task is being
1426 	 * switched to or is in IPI.
1427 	 */
1428 	if (!test_thread_flag(TIF_NEED_FPU_LOAD)) {
1429 		/* The MSR is active and can be directly updated. */
1430 		wrmsrl(MSR_IA32_PASID, pasid_state);
1431 	} else {
1432 		struct fpu *fpu = &current->thread.fpu;
1433 		struct ia32_pasid_state *ppasid_state;
1434 		struct xregs_state *xsave;
1435 
1436 		/*
1437 		 * The CPU's xstate registers are not currently active. Just
1438 		 * update the PASID state in the memory buffer here. The
1439 		 * PASID MSR will be loaded when returning to user mode.
1440 		 */
1441 		xsave = &fpu->state.xsave;
1442 		xsave->header.xfeatures |= XFEATURE_MASK_PASID;
1443 		ppasid_state = get_xsave_addr(xsave, XFEATURE_PASID);
1444 		/*
1445 		 * Since XFEATURE_MASK_PASID is set in xfeatures, ppasid_state
1446 		 * won't be NULL and no need to check its value.
1447 		 *
1448 		 * Only update the task's PASID state when it's different
1449 		 * from the mm's pasid.
1450 		 */
1451 		if (ppasid_state->pasid != pasid_state) {
1452 			/*
1453 			 * Invalid fpregs so that state restoring will pick up
1454 			 * the PASID state.
1455 			 */
1456 			__fpu_invalidate_fpregs_state(fpu);
1457 			ppasid_state->pasid = pasid_state;
1458 		}
1459 	}
1460 }
1461 #endif /* CONFIG_IOMMU_SUPPORT */
1462