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 */
cpu_has_xfeatures(u64 xfeatures_needed,const char ** feature_name)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
xfeature_is_aligned64(int xfeature_nr)127 static bool xfeature_is_aligned64(int xfeature_nr)
128 {
129 return xstate_flags[xfeature_nr] & XSTATE_FLAG_ALIGNED64;
130 }
131
xfeature_is_supervisor(int xfeature_nr)132 static bool xfeature_is_supervisor(int xfeature_nr)
133 {
134 return xstate_flags[xfeature_nr] & XSTATE_FLAG_SUPERVISOR;
135 }
136
xfeature_get_offset(u64 xcomp_bv,int xfeature)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 */
fpu__init_cpu_xstate(void)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
xfeature_enabled(enum xfeature xfeature)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 */
setup_xstate_cache(void)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
print_xstate_feature(u64 xstate_mask)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 */
print_xstate_features(void)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 */
print_xstate_offset_size(void)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 */
os_xrstor_booting(struct xregs_state * xstate)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 */
setup_init_fpu_buf(void)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
xfeature_size(int xfeature_nr)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) */
validate_user_xstate_header(const struct xstate_header * hdr,struct fpstate * fpstate)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
__xstate_dump_leaves(void)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, fmt, ...) do { \
444 if (WARN_ONCE(x, "XSAVE consistency problem: " fmt, ##__VA_ARGS__)) { \
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 */
check_xtile_data_against_struct(int size)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 */
check_xstate_against_struct(int nr)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 XSTATE_WARN_ON(1, "No structure for xstate: %d\n", nr);
558 return false;
559 }
560 return true;
561 }
562
xstate_calculate_size(u64 xfeatures,bool compacted)563 static unsigned int xstate_calculate_size(u64 xfeatures, bool compacted)
564 {
565 unsigned int topmost = fls64(xfeatures) - 1;
566 unsigned int offset = xstate_offsets[topmost];
567
568 if (topmost <= XFEATURE_SSE)
569 return sizeof(struct xregs_state);
570
571 if (compacted)
572 offset = xfeature_get_offset(xfeatures, topmost);
573 return offset + xstate_sizes[topmost];
574 }
575
576 /*
577 * This essentially double-checks what the cpu told us about
578 * how large the XSAVE buffer needs to be. We are recalculating
579 * it to be safe.
580 *
581 * Independent XSAVE features allocate their own buffers and are not
582 * covered by these checks. Only the size of the buffer for task->fpu
583 * is checked here.
584 */
paranoid_xstate_size_valid(unsigned int kernel_size)585 static bool __init paranoid_xstate_size_valid(unsigned int kernel_size)
586 {
587 bool compacted = cpu_feature_enabled(X86_FEATURE_XCOMPACTED);
588 bool xsaves = cpu_feature_enabled(X86_FEATURE_XSAVES);
589 unsigned int size = FXSAVE_SIZE + XSAVE_HDR_SIZE;
590 int i;
591
592 for_each_extended_xfeature(i, fpu_kernel_cfg.max_features) {
593 if (!check_xstate_against_struct(i))
594 return false;
595 /*
596 * Supervisor state components can be managed only by
597 * XSAVES.
598 */
599 if (!xsaves && xfeature_is_supervisor(i)) {
600 XSTATE_WARN_ON(1, "Got supervisor feature %d, but XSAVES not advertised\n", i);
601 return false;
602 }
603 }
604 size = xstate_calculate_size(fpu_kernel_cfg.max_features, compacted);
605 XSTATE_WARN_ON(size != kernel_size,
606 "size %u != kernel_size %u\n", 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 */
get_compacted_size(void)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 */
get_xsave_compacted_size(void)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
get_xsave_size_user(void)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
init_xstate_size(void)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 */
fpu__init_disable_system_xstate(unsigned int legacy_size)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(¤t->thread.fpu);
742 }
743
744 /*
745 * Enable and initialize the xsave feature.
746 * Called once per system bootup.
747 */
fpu__init_system_xstate(unsigned int legacy_size)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(¤t->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 */
fpu__resume_cpu(void)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 */
__raw_xsave_addr(struct xregs_state * xsave,int xfeature_nr)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 */
get_xsave_addr(struct xregs_state * xsave,int xfeature_nr)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 */
arch_set_user_pkey_access(struct task_struct * tsk,int pkey,unsigned long init_val)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
copy_feature(bool from_xstate,struct membuf * to,void * xstate,void * init_xstate,unsigned int size)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 */
__copy_xstate_to_uabi_buf(struct membuf to,struct fpstate * fpstate,u32 pkru_val,enum xstate_copy_mode copy_mode)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 */
copy_xstate_to_uabi_buf(struct membuf to,struct task_struct * tsk,enum xstate_copy_mode copy_mode)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
copy_from_buffer(void * dst,unsigned int offset,unsigned int size,const void * kbuf,const void __user * ubuf)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 /**
1204 * copy_uabi_to_xstate - Copy a UABI format buffer to the kernel xstate
1205 * @fpstate: The fpstate buffer to copy to
1206 * @kbuf: The UABI format buffer, if it comes from the kernel
1207 * @ubuf: The UABI format buffer, if it comes from userspace
1208 * @pkru: The location to write the PKRU value to
1209 *
1210 * Converts from the UABI format into the kernel internal hardware
1211 * dependent format.
1212 *
1213 * This function ultimately has three different callers with distinct PKRU
1214 * behavior.
1215 * 1. When called from sigreturn the PKRU register will be restored from
1216 * @fpstate via an XRSTOR. Correctly copying the UABI format buffer to
1217 * @fpstate is sufficient to cover this case, but the caller will also
1218 * pass a pointer to the thread_struct's pkru field in @pkru and updating
1219 * it is harmless.
1220 * 2. When called from ptrace the PKRU register will be restored from the
1221 * thread_struct's pkru field. A pointer to that is passed in @pkru.
1222 * The kernel will restore it manually, so the XRSTOR behavior that resets
1223 * the PKRU register to the hardware init value (0) if the corresponding
1224 * xfeatures bit is not set is emulated here.
1225 * 3. When called from KVM the PKRU register will be restored from the vcpu's
1226 * pkru field. A pointer to that is passed in @pkru. KVM hasn't used
1227 * XRSTOR and hasn't had the PKRU resetting behavior described above. To
1228 * preserve that KVM behavior, it passes NULL for @pkru if the xfeatures
1229 * bit is not set.
1230 */
copy_uabi_to_xstate(struct fpstate * fpstate,const void * kbuf,const void __user * ubuf,u32 * pkru)1231 static int copy_uabi_to_xstate(struct fpstate *fpstate, const void *kbuf,
1232 const void __user *ubuf, u32 *pkru)
1233 {
1234 struct xregs_state *xsave = &fpstate->regs.xsave;
1235 unsigned int offset, size;
1236 struct xstate_header hdr;
1237 u64 mask;
1238 int i;
1239
1240 offset = offsetof(struct xregs_state, header);
1241 if (copy_from_buffer(&hdr, offset, sizeof(hdr), kbuf, ubuf))
1242 return -EFAULT;
1243
1244 if (validate_user_xstate_header(&hdr, fpstate))
1245 return -EINVAL;
1246
1247 /* Validate MXCSR when any of the related features is in use */
1248 mask = XFEATURE_MASK_FP | XFEATURE_MASK_SSE | XFEATURE_MASK_YMM;
1249 if (hdr.xfeatures & mask) {
1250 u32 mxcsr[2];
1251
1252 offset = offsetof(struct fxregs_state, mxcsr);
1253 if (copy_from_buffer(mxcsr, offset, sizeof(mxcsr), kbuf, ubuf))
1254 return -EFAULT;
1255
1256 /* Reserved bits in MXCSR must be zero. */
1257 if (mxcsr[0] & ~mxcsr_feature_mask)
1258 return -EINVAL;
1259
1260 /* SSE and YMM require MXCSR even when FP is not in use. */
1261 if (!(hdr.xfeatures & XFEATURE_MASK_FP)) {
1262 xsave->i387.mxcsr = mxcsr[0];
1263 xsave->i387.mxcsr_mask = mxcsr[1];
1264 }
1265 }
1266
1267 for (i = 0; i < XFEATURE_MAX; i++) {
1268 mask = BIT_ULL(i);
1269
1270 if (hdr.xfeatures & mask) {
1271 void *dst = __raw_xsave_addr(xsave, i);
1272
1273 offset = xstate_offsets[i];
1274 size = xstate_sizes[i];
1275
1276 if (copy_from_buffer(dst, offset, size, kbuf, ubuf))
1277 return -EFAULT;
1278 }
1279 }
1280
1281 if (hdr.xfeatures & XFEATURE_MASK_PKRU) {
1282 struct pkru_state *xpkru;
1283
1284 xpkru = __raw_xsave_addr(xsave, XFEATURE_PKRU);
1285 *pkru = xpkru->pkru;
1286 } else {
1287 /*
1288 * KVM may pass NULL here to indicate that it does not need
1289 * PKRU updated.
1290 */
1291 if (pkru)
1292 *pkru = 0;
1293 }
1294
1295 /*
1296 * The state that came in from userspace was user-state only.
1297 * Mask all the user states out of 'xfeatures':
1298 */
1299 xsave->header.xfeatures &= XFEATURE_MASK_SUPERVISOR_ALL;
1300
1301 /*
1302 * Add back in the features that came in from userspace:
1303 */
1304 xsave->header.xfeatures |= hdr.xfeatures;
1305
1306 return 0;
1307 }
1308
1309 /*
1310 * Convert from a ptrace standard-format kernel buffer to kernel XSAVE[S]
1311 * format and copy to the target thread. Used by ptrace and KVM.
1312 */
copy_uabi_from_kernel_to_xstate(struct fpstate * fpstate,const void * kbuf,u32 * pkru)1313 int copy_uabi_from_kernel_to_xstate(struct fpstate *fpstate, const void *kbuf, u32 *pkru)
1314 {
1315 return copy_uabi_to_xstate(fpstate, kbuf, NULL, pkru);
1316 }
1317
1318 /*
1319 * Convert from a sigreturn standard-format user-space buffer to kernel
1320 * XSAVE[S] format and copy to the target thread. This is called from the
1321 * sigreturn() and rt_sigreturn() system calls.
1322 */
copy_sigframe_from_user_to_xstate(struct task_struct * tsk,const void __user * ubuf)1323 int copy_sigframe_from_user_to_xstate(struct task_struct *tsk,
1324 const void __user *ubuf)
1325 {
1326 return copy_uabi_to_xstate(tsk->thread.fpu.fpstate, NULL, ubuf, &tsk->thread.pkru);
1327 }
1328
validate_independent_components(u64 mask)1329 static bool validate_independent_components(u64 mask)
1330 {
1331 u64 xchk;
1332
1333 if (WARN_ON_FPU(!cpu_feature_enabled(X86_FEATURE_XSAVES)))
1334 return false;
1335
1336 xchk = ~xfeatures_mask_independent();
1337
1338 if (WARN_ON_ONCE(!mask || mask & xchk))
1339 return false;
1340
1341 return true;
1342 }
1343
1344 /**
1345 * xsaves - Save selected components to a kernel xstate buffer
1346 * @xstate: Pointer to the buffer
1347 * @mask: Feature mask to select the components to save
1348 *
1349 * The @xstate buffer must be 64 byte aligned and correctly initialized as
1350 * XSAVES does not write the full xstate header. Before first use the
1351 * buffer should be zeroed otherwise a consecutive XRSTORS from that buffer
1352 * can #GP.
1353 *
1354 * The feature mask must be a subset of the independent features.
1355 */
xsaves(struct xregs_state * xstate,u64 mask)1356 void xsaves(struct xregs_state *xstate, u64 mask)
1357 {
1358 int err;
1359
1360 if (!validate_independent_components(mask))
1361 return;
1362
1363 XSTATE_OP(XSAVES, xstate, (u32)mask, (u32)(mask >> 32), err);
1364 WARN_ON_ONCE(err);
1365 }
1366
1367 /**
1368 * xrstors - Restore selected components from a kernel xstate buffer
1369 * @xstate: Pointer to the buffer
1370 * @mask: Feature mask to select the components to restore
1371 *
1372 * The @xstate buffer must be 64 byte aligned and correctly initialized
1373 * otherwise XRSTORS from that buffer can #GP.
1374 *
1375 * Proper usage is to restore the state which was saved with
1376 * xsaves() into @xstate.
1377 *
1378 * The feature mask must be a subset of the independent features.
1379 */
xrstors(struct xregs_state * xstate,u64 mask)1380 void xrstors(struct xregs_state *xstate, u64 mask)
1381 {
1382 int err;
1383
1384 if (!validate_independent_components(mask))
1385 return;
1386
1387 XSTATE_OP(XRSTORS, xstate, (u32)mask, (u32)(mask >> 32), err);
1388 WARN_ON_ONCE(err);
1389 }
1390
1391 #if IS_ENABLED(CONFIG_KVM)
fpstate_clear_xstate_component(struct fpstate * fps,unsigned int xfeature)1392 void fpstate_clear_xstate_component(struct fpstate *fps, unsigned int xfeature)
1393 {
1394 void *addr = get_xsave_addr(&fps->regs.xsave, xfeature);
1395
1396 if (addr)
1397 memset(addr, 0, xstate_sizes[xfeature]);
1398 }
1399 EXPORT_SYMBOL_GPL(fpstate_clear_xstate_component);
1400 #endif
1401
1402 #ifdef CONFIG_X86_64
1403
1404 #ifdef CONFIG_X86_DEBUG_FPU
1405 /*
1406 * Ensure that a subsequent XSAVE* or XRSTOR* instruction with RFBM=@mask
1407 * can safely operate on the @fpstate buffer.
1408 */
xstate_op_valid(struct fpstate * fpstate,u64 mask,bool rstor)1409 static bool xstate_op_valid(struct fpstate *fpstate, u64 mask, bool rstor)
1410 {
1411 u64 xfd = __this_cpu_read(xfd_state);
1412
1413 if (fpstate->xfd == xfd)
1414 return true;
1415
1416 /*
1417 * The XFD MSR does not match fpstate->xfd. That's invalid when
1418 * the passed in fpstate is current's fpstate.
1419 */
1420 if (fpstate->xfd == current->thread.fpu.fpstate->xfd)
1421 return false;
1422
1423 /*
1424 * XRSTOR(S) from init_fpstate are always correct as it will just
1425 * bring all components into init state and not read from the
1426 * buffer. XSAVE(S) raises #PF after init.
1427 */
1428 if (fpstate == &init_fpstate)
1429 return rstor;
1430
1431 /*
1432 * XSAVE(S): clone(), fpu_swap_kvm_fpu()
1433 * XRSTORS(S): fpu_swap_kvm_fpu()
1434 */
1435
1436 /*
1437 * No XSAVE/XRSTOR instructions (except XSAVE itself) touch
1438 * the buffer area for XFD-disabled state components.
1439 */
1440 mask &= ~xfd;
1441
1442 /*
1443 * Remove features which are valid in fpstate. They
1444 * have space allocated in fpstate.
1445 */
1446 mask &= ~fpstate->xfeatures;
1447
1448 /*
1449 * Any remaining state components in 'mask' might be written
1450 * by XSAVE/XRSTOR. Fail validation it found.
1451 */
1452 return !mask;
1453 }
1454
xfd_validate_state(struct fpstate * fpstate,u64 mask,bool rstor)1455 void xfd_validate_state(struct fpstate *fpstate, u64 mask, bool rstor)
1456 {
1457 WARN_ON_ONCE(!xstate_op_valid(fpstate, mask, rstor));
1458 }
1459 #endif /* CONFIG_X86_DEBUG_FPU */
1460
xfd_update_static_branch(void)1461 static int __init xfd_update_static_branch(void)
1462 {
1463 /*
1464 * If init_fpstate.xfd has bits set then dynamic features are
1465 * available and the dynamic sizing must be enabled.
1466 */
1467 if (init_fpstate.xfd)
1468 static_branch_enable(&__fpu_state_size_dynamic);
1469 return 0;
1470 }
arch_initcall(xfd_update_static_branch)1471 arch_initcall(xfd_update_static_branch)
1472
1473 void fpstate_free(struct fpu *fpu)
1474 {
1475 if (fpu->fpstate && fpu->fpstate != &fpu->__fpstate)
1476 vfree(fpu->fpstate);
1477 }
1478
1479 /**
1480 * fpstate_realloc - Reallocate struct fpstate for the requested new features
1481 *
1482 * @xfeatures: A bitmap of xstate features which extend the enabled features
1483 * of that task
1484 * @ksize: The required size for the kernel buffer
1485 * @usize: The required size for user space buffers
1486 * @guest_fpu: Pointer to a guest FPU container. NULL for host allocations
1487 *
1488 * Note vs. vmalloc(): If the task with a vzalloc()-allocated buffer
1489 * terminates quickly, vfree()-induced IPIs may be a concern, but tasks
1490 * with large states are likely to live longer.
1491 *
1492 * Returns: 0 on success, -ENOMEM on allocation error.
1493 */
fpstate_realloc(u64 xfeatures,unsigned int ksize,unsigned int usize,struct fpu_guest * guest_fpu)1494 static int fpstate_realloc(u64 xfeatures, unsigned int ksize,
1495 unsigned int usize, struct fpu_guest *guest_fpu)
1496 {
1497 struct fpu *fpu = ¤t->thread.fpu;
1498 struct fpstate *curfps, *newfps = NULL;
1499 unsigned int fpsize;
1500 bool in_use;
1501
1502 fpsize = ksize + ALIGN(offsetof(struct fpstate, regs), 64);
1503
1504 newfps = vzalloc(fpsize);
1505 if (!newfps)
1506 return -ENOMEM;
1507 newfps->size = ksize;
1508 newfps->user_size = usize;
1509 newfps->is_valloc = true;
1510
1511 /*
1512 * When a guest FPU is supplied, use @guest_fpu->fpstate
1513 * as reference independent whether it is in use or not.
1514 */
1515 curfps = guest_fpu ? guest_fpu->fpstate : fpu->fpstate;
1516
1517 /* Determine whether @curfps is the active fpstate */
1518 in_use = fpu->fpstate == curfps;
1519
1520 if (guest_fpu) {
1521 newfps->is_guest = true;
1522 newfps->is_confidential = curfps->is_confidential;
1523 newfps->in_use = curfps->in_use;
1524 guest_fpu->xfeatures |= xfeatures;
1525 guest_fpu->uabi_size = usize;
1526 }
1527
1528 fpregs_lock();
1529 /*
1530 * If @curfps is in use, ensure that the current state is in the
1531 * registers before swapping fpstate as that might invalidate it
1532 * due to layout changes.
1533 */
1534 if (in_use && test_thread_flag(TIF_NEED_FPU_LOAD))
1535 fpregs_restore_userregs();
1536
1537 newfps->xfeatures = curfps->xfeatures | xfeatures;
1538
1539 if (!guest_fpu)
1540 newfps->user_xfeatures = curfps->user_xfeatures | xfeatures;
1541
1542 newfps->xfd = curfps->xfd & ~xfeatures;
1543
1544 /* Do the final updates within the locked region */
1545 xstate_init_xcomp_bv(&newfps->regs.xsave, newfps->xfeatures);
1546
1547 if (guest_fpu) {
1548 guest_fpu->fpstate = newfps;
1549 /* If curfps is active, update the FPU fpstate pointer */
1550 if (in_use)
1551 fpu->fpstate = newfps;
1552 } else {
1553 fpu->fpstate = newfps;
1554 }
1555
1556 if (in_use)
1557 xfd_update_state(fpu->fpstate);
1558 fpregs_unlock();
1559
1560 /* Only free valloc'ed state */
1561 if (curfps && curfps->is_valloc)
1562 vfree(curfps);
1563
1564 return 0;
1565 }
1566
validate_sigaltstack(unsigned int usize)1567 static int validate_sigaltstack(unsigned int usize)
1568 {
1569 struct task_struct *thread, *leader = current->group_leader;
1570 unsigned long framesize = get_sigframe_size();
1571
1572 lockdep_assert_held(¤t->sighand->siglock);
1573
1574 /* get_sigframe_size() is based on fpu_user_cfg.max_size */
1575 framesize -= fpu_user_cfg.max_size;
1576 framesize += usize;
1577 for_each_thread(leader, thread) {
1578 if (thread->sas_ss_size && thread->sas_ss_size < framesize)
1579 return -ENOSPC;
1580 }
1581 return 0;
1582 }
1583
__xstate_request_perm(u64 permitted,u64 requested,bool guest)1584 static int __xstate_request_perm(u64 permitted, u64 requested, bool guest)
1585 {
1586 /*
1587 * This deliberately does not exclude !XSAVES as we still might
1588 * decide to optionally context switch XCR0 or talk the silicon
1589 * vendors into extending XFD for the pre AMX states, especially
1590 * AVX512.
1591 */
1592 bool compacted = cpu_feature_enabled(X86_FEATURE_XCOMPACTED);
1593 struct fpu *fpu = ¤t->group_leader->thread.fpu;
1594 struct fpu_state_perm *perm;
1595 unsigned int ksize, usize;
1596 u64 mask;
1597 int ret = 0;
1598
1599 /* Check whether fully enabled */
1600 if ((permitted & requested) == requested)
1601 return 0;
1602
1603 /* Calculate the resulting kernel state size */
1604 mask = permitted | requested;
1605 /* Take supervisor states into account on the host */
1606 if (!guest)
1607 mask |= xfeatures_mask_supervisor();
1608 ksize = xstate_calculate_size(mask, compacted);
1609
1610 /* Calculate the resulting user state size */
1611 mask &= XFEATURE_MASK_USER_SUPPORTED;
1612 usize = xstate_calculate_size(mask, false);
1613
1614 if (!guest) {
1615 ret = validate_sigaltstack(usize);
1616 if (ret)
1617 return ret;
1618 }
1619
1620 perm = guest ? &fpu->guest_perm : &fpu->perm;
1621 /* Pairs with the READ_ONCE() in xstate_get_group_perm() */
1622 WRITE_ONCE(perm->__state_perm, mask);
1623 /* Protected by sighand lock */
1624 perm->__state_size = ksize;
1625 perm->__user_state_size = usize;
1626 return ret;
1627 }
1628
1629 /*
1630 * Permissions array to map facilities with more than one component
1631 */
1632 static const u64 xstate_prctl_req[XFEATURE_MAX] = {
1633 [XFEATURE_XTILE_DATA] = XFEATURE_MASK_XTILE_DATA,
1634 };
1635
xstate_request_perm(unsigned long idx,bool guest)1636 static int xstate_request_perm(unsigned long idx, bool guest)
1637 {
1638 u64 permitted, requested;
1639 int ret;
1640
1641 if (idx >= XFEATURE_MAX)
1642 return -EINVAL;
1643
1644 /*
1645 * Look up the facility mask which can require more than
1646 * one xstate component.
1647 */
1648 idx = array_index_nospec(idx, ARRAY_SIZE(xstate_prctl_req));
1649 requested = xstate_prctl_req[idx];
1650 if (!requested)
1651 return -EOPNOTSUPP;
1652
1653 if ((fpu_user_cfg.max_features & requested) != requested)
1654 return -EOPNOTSUPP;
1655
1656 /* Lockless quick check */
1657 permitted = xstate_get_group_perm(guest);
1658 if ((permitted & requested) == requested)
1659 return 0;
1660
1661 /* Protect against concurrent modifications */
1662 spin_lock_irq(¤t->sighand->siglock);
1663 permitted = xstate_get_group_perm(guest);
1664
1665 /* First vCPU allocation locks the permissions. */
1666 if (guest && (permitted & FPU_GUEST_PERM_LOCKED))
1667 ret = -EBUSY;
1668 else
1669 ret = __xstate_request_perm(permitted, requested, guest);
1670 spin_unlock_irq(¤t->sighand->siglock);
1671 return ret;
1672 }
1673
__xfd_enable_feature(u64 xfd_err,struct fpu_guest * guest_fpu)1674 int __xfd_enable_feature(u64 xfd_err, struct fpu_guest *guest_fpu)
1675 {
1676 u64 xfd_event = xfd_err & XFEATURE_MASK_USER_DYNAMIC;
1677 struct fpu_state_perm *perm;
1678 unsigned int ksize, usize;
1679 struct fpu *fpu;
1680
1681 if (!xfd_event) {
1682 if (!guest_fpu)
1683 pr_err_once("XFD: Invalid xfd error: %016llx\n", xfd_err);
1684 return 0;
1685 }
1686
1687 /* Protect against concurrent modifications */
1688 spin_lock_irq(¤t->sighand->siglock);
1689
1690 /* If not permitted let it die */
1691 if ((xstate_get_group_perm(!!guest_fpu) & xfd_event) != xfd_event) {
1692 spin_unlock_irq(¤t->sighand->siglock);
1693 return -EPERM;
1694 }
1695
1696 fpu = ¤t->group_leader->thread.fpu;
1697 perm = guest_fpu ? &fpu->guest_perm : &fpu->perm;
1698 ksize = perm->__state_size;
1699 usize = perm->__user_state_size;
1700
1701 /*
1702 * The feature is permitted. State size is sufficient. Dropping
1703 * the lock is safe here even if more features are added from
1704 * another task, the retrieved buffer sizes are valid for the
1705 * currently requested feature(s).
1706 */
1707 spin_unlock_irq(¤t->sighand->siglock);
1708
1709 /*
1710 * Try to allocate a new fpstate. If that fails there is no way
1711 * out.
1712 */
1713 if (fpstate_realloc(xfd_event, ksize, usize, guest_fpu))
1714 return -EFAULT;
1715 return 0;
1716 }
1717
xfd_enable_feature(u64 xfd_err)1718 int xfd_enable_feature(u64 xfd_err)
1719 {
1720 return __xfd_enable_feature(xfd_err, NULL);
1721 }
1722
1723 #else /* CONFIG_X86_64 */
xstate_request_perm(unsigned long idx,bool guest)1724 static inline int xstate_request_perm(unsigned long idx, bool guest)
1725 {
1726 return -EPERM;
1727 }
1728 #endif /* !CONFIG_X86_64 */
1729
xstate_get_guest_group_perm(void)1730 u64 xstate_get_guest_group_perm(void)
1731 {
1732 return xstate_get_group_perm(true);
1733 }
1734 EXPORT_SYMBOL_GPL(xstate_get_guest_group_perm);
1735
1736 /**
1737 * fpu_xstate_prctl - xstate permission operations
1738 * @tsk: Redundant pointer to current
1739 * @option: A subfunction of arch_prctl()
1740 * @arg2: option argument
1741 * Return: 0 if successful; otherwise, an error code
1742 *
1743 * Option arguments:
1744 *
1745 * ARCH_GET_XCOMP_SUPP: Pointer to user space u64 to store the info
1746 * ARCH_GET_XCOMP_PERM: Pointer to user space u64 to store the info
1747 * ARCH_REQ_XCOMP_PERM: Facility number requested
1748 *
1749 * For facilities which require more than one XSTATE component, the request
1750 * must be the highest state component number related to that facility,
1751 * e.g. for AMX which requires XFEATURE_XTILE_CFG(17) and
1752 * XFEATURE_XTILE_DATA(18) this would be XFEATURE_XTILE_DATA(18).
1753 */
fpu_xstate_prctl(int option,unsigned long arg2)1754 long fpu_xstate_prctl(int option, unsigned long arg2)
1755 {
1756 u64 __user *uptr = (u64 __user *)arg2;
1757 u64 permitted, supported;
1758 unsigned long idx = arg2;
1759 bool guest = false;
1760
1761 switch (option) {
1762 case ARCH_GET_XCOMP_SUPP:
1763 supported = fpu_user_cfg.max_features | fpu_user_cfg.legacy_features;
1764 return put_user(supported, uptr);
1765
1766 case ARCH_GET_XCOMP_PERM:
1767 /*
1768 * Lockless snapshot as it can also change right after the
1769 * dropping the lock.
1770 */
1771 permitted = xstate_get_host_group_perm();
1772 permitted &= XFEATURE_MASK_USER_SUPPORTED;
1773 return put_user(permitted, uptr);
1774
1775 case ARCH_GET_XCOMP_GUEST_PERM:
1776 permitted = xstate_get_guest_group_perm();
1777 permitted &= XFEATURE_MASK_USER_SUPPORTED;
1778 return put_user(permitted, uptr);
1779
1780 case ARCH_REQ_XCOMP_GUEST_PERM:
1781 guest = true;
1782 fallthrough;
1783
1784 case ARCH_REQ_XCOMP_PERM:
1785 if (!IS_ENABLED(CONFIG_X86_64))
1786 return -EOPNOTSUPP;
1787
1788 return xstate_request_perm(idx, guest);
1789
1790 default:
1791 return -EINVAL;
1792 }
1793 }
1794
1795 #ifdef CONFIG_PROC_PID_ARCH_STATUS
1796 /*
1797 * Report the amount of time elapsed in millisecond since last AVX512
1798 * use in the task.
1799 */
avx512_status(struct seq_file * m,struct task_struct * task)1800 static void avx512_status(struct seq_file *m, struct task_struct *task)
1801 {
1802 unsigned long timestamp = READ_ONCE(task->thread.fpu.avx512_timestamp);
1803 long delta;
1804
1805 if (!timestamp) {
1806 /*
1807 * Report -1 if no AVX512 usage
1808 */
1809 delta = -1;
1810 } else {
1811 delta = (long)(jiffies - timestamp);
1812 /*
1813 * Cap to LONG_MAX if time difference > LONG_MAX
1814 */
1815 if (delta < 0)
1816 delta = LONG_MAX;
1817 delta = jiffies_to_msecs(delta);
1818 }
1819
1820 seq_put_decimal_ll(m, "AVX512_elapsed_ms:\t", delta);
1821 seq_putc(m, '\n');
1822 }
1823
1824 /*
1825 * Report architecture specific information
1826 */
proc_pid_arch_status(struct seq_file * m,struct pid_namespace * ns,struct pid * pid,struct task_struct * task)1827 int proc_pid_arch_status(struct seq_file *m, struct pid_namespace *ns,
1828 struct pid *pid, struct task_struct *task)
1829 {
1830 /*
1831 * Report AVX512 state if the processor and build option supported.
1832 */
1833 if (cpu_feature_enabled(X86_FEATURE_AVX512F))
1834 avx512_status(m, task);
1835
1836 return 0;
1837 }
1838 #endif /* CONFIG_PROC_PID_ARCH_STATUS */
1839