1 // SPDX-License-Identifier: GPL-2.0-only 2 /* 3 * xsave/xrstor support. 4 * 5 * Author: Suresh Siddha <suresh.b.siddha@intel.com> 6 */ 7 #include <linux/compat.h> 8 #include <linux/cpu.h> 9 #include <linux/mman.h> 10 #include <linux/pkeys.h> 11 #include <linux/seq_file.h> 12 #include <linux/proc_fs.h> 13 14 #include <asm/fpu/api.h> 15 #include <asm/fpu/internal.h> 16 #include <asm/fpu/signal.h> 17 #include <asm/fpu/regset.h> 18 #include <asm/fpu/xstate.h> 19 20 #include <asm/tlbflush.h> 21 #include <asm/cpufeature.h> 22 23 /* 24 * Although we spell it out in here, the Processor Trace 25 * xfeature is completely unused. We use other mechanisms 26 * to save/restore PT state in Linux. 27 */ 28 static const char *xfeature_names[] = 29 { 30 "x87 floating point registers" , 31 "SSE registers" , 32 "AVX registers" , 33 "MPX bounds registers" , 34 "MPX CSR" , 35 "AVX-512 opmask" , 36 "AVX-512 Hi256" , 37 "AVX-512 ZMM_Hi256" , 38 "Processor Trace (unused)" , 39 "Protection Keys User registers", 40 "unknown xstate feature" , 41 }; 42 43 static short xsave_cpuid_features[] __initdata = { 44 X86_FEATURE_FPU, 45 X86_FEATURE_XMM, 46 X86_FEATURE_AVX, 47 X86_FEATURE_MPX, 48 X86_FEATURE_MPX, 49 X86_FEATURE_AVX512F, 50 X86_FEATURE_AVX512F, 51 X86_FEATURE_AVX512F, 52 X86_FEATURE_INTEL_PT, 53 X86_FEATURE_PKU, 54 }; 55 56 /* 57 * Mask of xstate features supported by the CPU and the kernel: 58 */ 59 u64 xfeatures_mask __read_mostly; 60 61 static unsigned int xstate_offsets[XFEATURE_MAX] = { [ 0 ... XFEATURE_MAX - 1] = -1}; 62 static unsigned int xstate_sizes[XFEATURE_MAX] = { [ 0 ... XFEATURE_MAX - 1] = -1}; 63 static unsigned int xstate_comp_offsets[sizeof(xfeatures_mask)*8]; 64 65 /* 66 * The XSAVE area of kernel can be in standard or compacted format; 67 * it is always in standard format for user mode. This is the user 68 * mode standard format size used for signal and ptrace frames. 69 */ 70 unsigned int fpu_user_xstate_size; 71 72 /* 73 * Return whether the system supports a given xfeature. 74 * 75 * Also return the name of the (most advanced) feature that the caller requested: 76 */ 77 int cpu_has_xfeatures(u64 xfeatures_needed, const char **feature_name) 78 { 79 u64 xfeatures_missing = xfeatures_needed & ~xfeatures_mask; 80 81 if (unlikely(feature_name)) { 82 long xfeature_idx, max_idx; 83 u64 xfeatures_print; 84 /* 85 * So we use FLS here to be able to print the most advanced 86 * feature that was requested but is missing. So if a driver 87 * asks about "XFEATURE_MASK_SSE | XFEATURE_MASK_YMM" we'll print the 88 * missing AVX feature - this is the most informative message 89 * to users: 90 */ 91 if (xfeatures_missing) 92 xfeatures_print = xfeatures_missing; 93 else 94 xfeatures_print = xfeatures_needed; 95 96 xfeature_idx = fls64(xfeatures_print)-1; 97 max_idx = ARRAY_SIZE(xfeature_names)-1; 98 xfeature_idx = min(xfeature_idx, max_idx); 99 100 *feature_name = xfeature_names[xfeature_idx]; 101 } 102 103 if (xfeatures_missing) 104 return 0; 105 106 return 1; 107 } 108 EXPORT_SYMBOL_GPL(cpu_has_xfeatures); 109 110 static int xfeature_is_supervisor(int xfeature_nr) 111 { 112 /* 113 * We currently do not support supervisor states, but if 114 * we did, we could find out like this. 115 * 116 * SDM says: If state component 'i' is a user state component, 117 * ECX[0] return 0; if state component i is a supervisor 118 * state component, ECX[0] returns 1. 119 */ 120 u32 eax, ebx, ecx, edx; 121 122 cpuid_count(XSTATE_CPUID, xfeature_nr, &eax, &ebx, &ecx, &edx); 123 return !!(ecx & 1); 124 } 125 126 static int xfeature_is_user(int xfeature_nr) 127 { 128 return !xfeature_is_supervisor(xfeature_nr); 129 } 130 131 /* 132 * When executing XSAVEOPT (or other optimized XSAVE instructions), if 133 * a processor implementation detects that an FPU state component is still 134 * (or is again) in its initialized state, it may clear the corresponding 135 * bit in the header.xfeatures field, and can skip the writeout of registers 136 * to the corresponding memory layout. 137 * 138 * This means that when the bit is zero, the state component might still contain 139 * some previous - non-initialized register state. 140 * 141 * Before writing xstate information to user-space we sanitize those components, 142 * to always ensure that the memory layout of a feature will be in the init state 143 * if the corresponding header bit is zero. This is to ensure that user-space doesn't 144 * see some stale state in the memory layout during signal handling, debugging etc. 145 */ 146 void fpstate_sanitize_xstate(struct fpu *fpu) 147 { 148 struct fxregs_state *fx = &fpu->state.fxsave; 149 int feature_bit; 150 u64 xfeatures; 151 152 if (!use_xsaveopt()) 153 return; 154 155 xfeatures = fpu->state.xsave.header.xfeatures; 156 157 /* 158 * None of the feature bits are in init state. So nothing else 159 * to do for us, as the memory layout is up to date. 160 */ 161 if ((xfeatures & xfeatures_mask) == xfeatures_mask) 162 return; 163 164 /* 165 * FP is in init state 166 */ 167 if (!(xfeatures & XFEATURE_MASK_FP)) { 168 fx->cwd = 0x37f; 169 fx->swd = 0; 170 fx->twd = 0; 171 fx->fop = 0; 172 fx->rip = 0; 173 fx->rdp = 0; 174 memset(&fx->st_space[0], 0, 128); 175 } 176 177 /* 178 * SSE is in init state 179 */ 180 if (!(xfeatures & XFEATURE_MASK_SSE)) 181 memset(&fx->xmm_space[0], 0, 256); 182 183 /* 184 * First two features are FPU and SSE, which above we handled 185 * in a special way already: 186 */ 187 feature_bit = 0x2; 188 xfeatures = (xfeatures_mask & ~xfeatures) >> 2; 189 190 /* 191 * Update all the remaining memory layouts according to their 192 * standard xstate layout, if their header bit is in the init 193 * state: 194 */ 195 while (xfeatures) { 196 if (xfeatures & 0x1) { 197 int offset = xstate_comp_offsets[feature_bit]; 198 int size = xstate_sizes[feature_bit]; 199 200 memcpy((void *)fx + offset, 201 (void *)&init_fpstate.xsave + offset, 202 size); 203 } 204 205 xfeatures >>= 1; 206 feature_bit++; 207 } 208 } 209 210 /* 211 * Enable the extended processor state save/restore feature. 212 * Called once per CPU onlining. 213 */ 214 void fpu__init_cpu_xstate(void) 215 { 216 if (!boot_cpu_has(X86_FEATURE_XSAVE) || !xfeatures_mask) 217 return; 218 /* 219 * Make it clear that XSAVES supervisor states are not yet 220 * implemented should anyone expect it to work by changing 221 * bits in XFEATURE_MASK_* macros and XCR0. 222 */ 223 WARN_ONCE((xfeatures_mask & XFEATURE_MASK_SUPERVISOR), 224 "x86/fpu: XSAVES supervisor states are not yet implemented.\n"); 225 226 xfeatures_mask &= ~XFEATURE_MASK_SUPERVISOR; 227 228 cr4_set_bits(X86_CR4_OSXSAVE); 229 xsetbv(XCR_XFEATURE_ENABLED_MASK, xfeatures_mask); 230 } 231 232 /* 233 * Note that in the future we will likely need a pair of 234 * functions here: one for user xstates and the other for 235 * system xstates. For now, they are the same. 236 */ 237 static int xfeature_enabled(enum xfeature xfeature) 238 { 239 return !!(xfeatures_mask & (1UL << xfeature)); 240 } 241 242 /* 243 * Record the offsets and sizes of various xstates contained 244 * in the XSAVE state memory layout. 245 */ 246 static void __init setup_xstate_features(void) 247 { 248 u32 eax, ebx, ecx, edx, i; 249 /* start at the beginnning of the "extended state" */ 250 unsigned int last_good_offset = offsetof(struct xregs_state, 251 extended_state_area); 252 /* 253 * The FP xstates and SSE xstates are legacy states. They are always 254 * in the fixed offsets in the xsave area in either compacted form 255 * or standard form. 256 */ 257 xstate_offsets[0] = 0; 258 xstate_sizes[0] = offsetof(struct fxregs_state, xmm_space); 259 xstate_offsets[1] = xstate_sizes[0]; 260 xstate_sizes[1] = FIELD_SIZEOF(struct fxregs_state, xmm_space); 261 262 for (i = FIRST_EXTENDED_XFEATURE; i < XFEATURE_MAX; i++) { 263 if (!xfeature_enabled(i)) 264 continue; 265 266 cpuid_count(XSTATE_CPUID, i, &eax, &ebx, &ecx, &edx); 267 268 /* 269 * If an xfeature is supervisor state, the offset 270 * in EBX is invalid. We leave it to -1. 271 */ 272 if (xfeature_is_user(i)) 273 xstate_offsets[i] = ebx; 274 275 xstate_sizes[i] = eax; 276 /* 277 * In our xstate size checks, we assume that the 278 * highest-numbered xstate feature has the 279 * highest offset in the buffer. Ensure it does. 280 */ 281 WARN_ONCE(last_good_offset > xstate_offsets[i], 282 "x86/fpu: misordered xstate at %d\n", last_good_offset); 283 last_good_offset = xstate_offsets[i]; 284 } 285 } 286 287 static void __init print_xstate_feature(u64 xstate_mask) 288 { 289 const char *feature_name; 290 291 if (cpu_has_xfeatures(xstate_mask, &feature_name)) 292 pr_info("x86/fpu: Supporting XSAVE feature 0x%03Lx: '%s'\n", xstate_mask, feature_name); 293 } 294 295 /* 296 * Print out all the supported xstate features: 297 */ 298 static void __init print_xstate_features(void) 299 { 300 print_xstate_feature(XFEATURE_MASK_FP); 301 print_xstate_feature(XFEATURE_MASK_SSE); 302 print_xstate_feature(XFEATURE_MASK_YMM); 303 print_xstate_feature(XFEATURE_MASK_BNDREGS); 304 print_xstate_feature(XFEATURE_MASK_BNDCSR); 305 print_xstate_feature(XFEATURE_MASK_OPMASK); 306 print_xstate_feature(XFEATURE_MASK_ZMM_Hi256); 307 print_xstate_feature(XFEATURE_MASK_Hi16_ZMM); 308 print_xstate_feature(XFEATURE_MASK_PKRU); 309 } 310 311 /* 312 * This check is important because it is easy to get XSTATE_* 313 * confused with XSTATE_BIT_*. 314 */ 315 #define CHECK_XFEATURE(nr) do { \ 316 WARN_ON(nr < FIRST_EXTENDED_XFEATURE); \ 317 WARN_ON(nr >= XFEATURE_MAX); \ 318 } while (0) 319 320 /* 321 * We could cache this like xstate_size[], but we only use 322 * it here, so it would be a waste of space. 323 */ 324 static int xfeature_is_aligned(int xfeature_nr) 325 { 326 u32 eax, ebx, ecx, edx; 327 328 CHECK_XFEATURE(xfeature_nr); 329 cpuid_count(XSTATE_CPUID, xfeature_nr, &eax, &ebx, &ecx, &edx); 330 /* 331 * The value returned by ECX[1] indicates the alignment 332 * of state component 'i' when the compacted format 333 * of the extended region of an XSAVE area is used: 334 */ 335 return !!(ecx & 2); 336 } 337 338 /* 339 * This function sets up offsets and sizes of all extended states in 340 * xsave area. This supports both standard format and compacted format 341 * of the xsave aread. 342 */ 343 static void __init setup_xstate_comp(void) 344 { 345 unsigned int xstate_comp_sizes[sizeof(xfeatures_mask)*8]; 346 int i; 347 348 /* 349 * The FP xstates and SSE xstates are legacy states. They are always 350 * in the fixed offsets in the xsave area in either compacted form 351 * or standard form. 352 */ 353 xstate_comp_offsets[0] = 0; 354 xstate_comp_offsets[1] = offsetof(struct fxregs_state, xmm_space); 355 356 if (!boot_cpu_has(X86_FEATURE_XSAVES)) { 357 for (i = FIRST_EXTENDED_XFEATURE; i < XFEATURE_MAX; i++) { 358 if (xfeature_enabled(i)) { 359 xstate_comp_offsets[i] = xstate_offsets[i]; 360 xstate_comp_sizes[i] = xstate_sizes[i]; 361 } 362 } 363 return; 364 } 365 366 xstate_comp_offsets[FIRST_EXTENDED_XFEATURE] = 367 FXSAVE_SIZE + XSAVE_HDR_SIZE; 368 369 for (i = FIRST_EXTENDED_XFEATURE; i < XFEATURE_MAX; i++) { 370 if (xfeature_enabled(i)) 371 xstate_comp_sizes[i] = xstate_sizes[i]; 372 else 373 xstate_comp_sizes[i] = 0; 374 375 if (i > FIRST_EXTENDED_XFEATURE) { 376 xstate_comp_offsets[i] = xstate_comp_offsets[i-1] 377 + xstate_comp_sizes[i-1]; 378 379 if (xfeature_is_aligned(i)) 380 xstate_comp_offsets[i] = 381 ALIGN(xstate_comp_offsets[i], 64); 382 } 383 } 384 } 385 386 /* 387 * Print out xstate component offsets and sizes 388 */ 389 static void __init print_xstate_offset_size(void) 390 { 391 int i; 392 393 for (i = FIRST_EXTENDED_XFEATURE; i < XFEATURE_MAX; i++) { 394 if (!xfeature_enabled(i)) 395 continue; 396 pr_info("x86/fpu: xstate_offset[%d]: %4d, xstate_sizes[%d]: %4d\n", 397 i, xstate_comp_offsets[i], i, xstate_sizes[i]); 398 } 399 } 400 401 /* 402 * setup the xstate image representing the init state 403 */ 404 static void __init setup_init_fpu_buf(void) 405 { 406 static int on_boot_cpu __initdata = 1; 407 408 WARN_ON_FPU(!on_boot_cpu); 409 on_boot_cpu = 0; 410 411 if (!boot_cpu_has(X86_FEATURE_XSAVE)) 412 return; 413 414 setup_xstate_features(); 415 print_xstate_features(); 416 417 if (boot_cpu_has(X86_FEATURE_XSAVES)) 418 init_fpstate.xsave.header.xcomp_bv = (u64)1 << 63 | xfeatures_mask; 419 420 /* 421 * Init all the features state with header.xfeatures being 0x0 422 */ 423 copy_kernel_to_xregs_booting(&init_fpstate.xsave); 424 425 /* 426 * Dump the init state again. This is to identify the init state 427 * of any feature which is not represented by all zero's. 428 */ 429 copy_xregs_to_kernel_booting(&init_fpstate.xsave); 430 } 431 432 static int xfeature_uncompacted_offset(int xfeature_nr) 433 { 434 u32 eax, ebx, ecx, edx; 435 436 /* 437 * Only XSAVES supports supervisor states and it uses compacted 438 * format. Checking a supervisor state's uncompacted offset is 439 * an error. 440 */ 441 if (XFEATURE_MASK_SUPERVISOR & BIT_ULL(xfeature_nr)) { 442 WARN_ONCE(1, "No fixed offset for xstate %d\n", xfeature_nr); 443 return -1; 444 } 445 446 CHECK_XFEATURE(xfeature_nr); 447 cpuid_count(XSTATE_CPUID, xfeature_nr, &eax, &ebx, &ecx, &edx); 448 return ebx; 449 } 450 451 static int xfeature_size(int xfeature_nr) 452 { 453 u32 eax, ebx, ecx, edx; 454 455 CHECK_XFEATURE(xfeature_nr); 456 cpuid_count(XSTATE_CPUID, xfeature_nr, &eax, &ebx, &ecx, &edx); 457 return eax; 458 } 459 460 /* 461 * 'XSAVES' implies two different things: 462 * 1. saving of supervisor/system state 463 * 2. using the compacted format 464 * 465 * Use this function when dealing with the compacted format so 466 * that it is obvious which aspect of 'XSAVES' is being handled 467 * by the calling code. 468 */ 469 int using_compacted_format(void) 470 { 471 return boot_cpu_has(X86_FEATURE_XSAVES); 472 } 473 474 /* Validate an xstate header supplied by userspace (ptrace or sigreturn) */ 475 int validate_xstate_header(const struct xstate_header *hdr) 476 { 477 /* No unknown or supervisor features may be set */ 478 if (hdr->xfeatures & (~xfeatures_mask | XFEATURE_MASK_SUPERVISOR)) 479 return -EINVAL; 480 481 /* Userspace must use the uncompacted format */ 482 if (hdr->xcomp_bv) 483 return -EINVAL; 484 485 /* 486 * If 'reserved' is shrunken to add a new field, make sure to validate 487 * that new field here! 488 */ 489 BUILD_BUG_ON(sizeof(hdr->reserved) != 48); 490 491 /* No reserved bits may be set */ 492 if (memchr_inv(hdr->reserved, 0, sizeof(hdr->reserved))) 493 return -EINVAL; 494 495 return 0; 496 } 497 498 static void __xstate_dump_leaves(void) 499 { 500 int i; 501 u32 eax, ebx, ecx, edx; 502 static int should_dump = 1; 503 504 if (!should_dump) 505 return; 506 should_dump = 0; 507 /* 508 * Dump out a few leaves past the ones that we support 509 * just in case there are some goodies up there 510 */ 511 for (i = 0; i < XFEATURE_MAX + 10; i++) { 512 cpuid_count(XSTATE_CPUID, i, &eax, &ebx, &ecx, &edx); 513 pr_warn("CPUID[%02x, %02x]: eax=%08x ebx=%08x ecx=%08x edx=%08x\n", 514 XSTATE_CPUID, i, eax, ebx, ecx, edx); 515 } 516 } 517 518 #define XSTATE_WARN_ON(x) do { \ 519 if (WARN_ONCE(x, "XSAVE consistency problem, dumping leaves")) { \ 520 __xstate_dump_leaves(); \ 521 } \ 522 } while (0) 523 524 #define XCHECK_SZ(sz, nr, nr_macro, __struct) do { \ 525 if ((nr == nr_macro) && \ 526 WARN_ONCE(sz != sizeof(__struct), \ 527 "%s: struct is %zu bytes, cpu state %d bytes\n", \ 528 __stringify(nr_macro), sizeof(__struct), sz)) { \ 529 __xstate_dump_leaves(); \ 530 } \ 531 } while (0) 532 533 /* 534 * We have a C struct for each 'xstate'. We need to ensure 535 * that our software representation matches what the CPU 536 * tells us about the state's size. 537 */ 538 static void check_xstate_against_struct(int nr) 539 { 540 /* 541 * Ask the CPU for the size of the state. 542 */ 543 int sz = xfeature_size(nr); 544 /* 545 * Match each CPU state with the corresponding software 546 * structure. 547 */ 548 XCHECK_SZ(sz, nr, XFEATURE_YMM, struct ymmh_struct); 549 XCHECK_SZ(sz, nr, XFEATURE_BNDREGS, struct mpx_bndreg_state); 550 XCHECK_SZ(sz, nr, XFEATURE_BNDCSR, struct mpx_bndcsr_state); 551 XCHECK_SZ(sz, nr, XFEATURE_OPMASK, struct avx_512_opmask_state); 552 XCHECK_SZ(sz, nr, XFEATURE_ZMM_Hi256, struct avx_512_zmm_uppers_state); 553 XCHECK_SZ(sz, nr, XFEATURE_Hi16_ZMM, struct avx_512_hi16_state); 554 XCHECK_SZ(sz, nr, XFEATURE_PKRU, struct pkru_state); 555 556 /* 557 * Make *SURE* to add any feature numbers in below if 558 * there are "holes" in the xsave state component 559 * numbers. 560 */ 561 if ((nr < XFEATURE_YMM) || 562 (nr >= XFEATURE_MAX) || 563 (nr == XFEATURE_PT_UNIMPLEMENTED_SO_FAR)) { 564 WARN_ONCE(1, "no structure for xstate: %d\n", nr); 565 XSTATE_WARN_ON(1); 566 } 567 } 568 569 /* 570 * This essentially double-checks what the cpu told us about 571 * how large the XSAVE buffer needs to be. We are recalculating 572 * it to be safe. 573 */ 574 static void do_extra_xstate_size_checks(void) 575 { 576 int paranoid_xstate_size = FXSAVE_SIZE + XSAVE_HDR_SIZE; 577 int i; 578 579 for (i = FIRST_EXTENDED_XFEATURE; i < XFEATURE_MAX; i++) { 580 if (!xfeature_enabled(i)) 581 continue; 582 583 check_xstate_against_struct(i); 584 /* 585 * Supervisor state components can be managed only by 586 * XSAVES, which is compacted-format only. 587 */ 588 if (!using_compacted_format()) 589 XSTATE_WARN_ON(xfeature_is_supervisor(i)); 590 591 /* Align from the end of the previous feature */ 592 if (xfeature_is_aligned(i)) 593 paranoid_xstate_size = ALIGN(paranoid_xstate_size, 64); 594 /* 595 * The offset of a given state in the non-compacted 596 * format is given to us in a CPUID leaf. We check 597 * them for being ordered (increasing offsets) in 598 * setup_xstate_features(). 599 */ 600 if (!using_compacted_format()) 601 paranoid_xstate_size = xfeature_uncompacted_offset(i); 602 /* 603 * The compacted-format offset always depends on where 604 * the previous state ended. 605 */ 606 paranoid_xstate_size += xfeature_size(i); 607 } 608 XSTATE_WARN_ON(paranoid_xstate_size != fpu_kernel_xstate_size); 609 } 610 611 612 /* 613 * Get total size of enabled xstates in XCR0/xfeatures_mask. 614 * 615 * Note the SDM's wording here. "sub-function 0" only enumerates 616 * the size of the *user* states. If we use it to size a buffer 617 * that we use 'XSAVES' on, we could potentially overflow the 618 * buffer because 'XSAVES' saves system states too. 619 * 620 * Note that we do not currently set any bits on IA32_XSS so 621 * 'XCR0 | IA32_XSS == XCR0' for now. 622 */ 623 static unsigned int __init get_xsaves_size(void) 624 { 625 unsigned int eax, ebx, ecx, edx; 626 /* 627 * - CPUID function 0DH, sub-function 1: 628 * EBX enumerates the size (in bytes) required by 629 * the XSAVES instruction for an XSAVE area 630 * containing all the state components 631 * corresponding to bits currently set in 632 * XCR0 | IA32_XSS. 633 */ 634 cpuid_count(XSTATE_CPUID, 1, &eax, &ebx, &ecx, &edx); 635 return ebx; 636 } 637 638 static unsigned int __init get_xsave_size(void) 639 { 640 unsigned int eax, ebx, ecx, edx; 641 /* 642 * - CPUID function 0DH, sub-function 0: 643 * EBX enumerates the size (in bytes) required by 644 * the XSAVE instruction for an XSAVE area 645 * containing all the *user* state components 646 * corresponding to bits currently set in XCR0. 647 */ 648 cpuid_count(XSTATE_CPUID, 0, &eax, &ebx, &ecx, &edx); 649 return ebx; 650 } 651 652 /* 653 * Will the runtime-enumerated 'xstate_size' fit in the init 654 * task's statically-allocated buffer? 655 */ 656 static bool is_supported_xstate_size(unsigned int test_xstate_size) 657 { 658 if (test_xstate_size <= sizeof(union fpregs_state)) 659 return true; 660 661 pr_warn("x86/fpu: xstate buffer too small (%zu < %d), disabling xsave\n", 662 sizeof(union fpregs_state), test_xstate_size); 663 return false; 664 } 665 666 static int __init init_xstate_size(void) 667 { 668 /* Recompute the context size for enabled features: */ 669 unsigned int possible_xstate_size; 670 unsigned int xsave_size; 671 672 xsave_size = get_xsave_size(); 673 674 if (boot_cpu_has(X86_FEATURE_XSAVES)) 675 possible_xstate_size = get_xsaves_size(); 676 else 677 possible_xstate_size = xsave_size; 678 679 /* Ensure we have the space to store all enabled: */ 680 if (!is_supported_xstate_size(possible_xstate_size)) 681 return -EINVAL; 682 683 /* 684 * The size is OK, we are definitely going to use xsave, 685 * make it known to the world that we need more space. 686 */ 687 fpu_kernel_xstate_size = possible_xstate_size; 688 do_extra_xstate_size_checks(); 689 690 /* 691 * User space is always in standard format. 692 */ 693 fpu_user_xstate_size = xsave_size; 694 return 0; 695 } 696 697 /* 698 * We enabled the XSAVE hardware, but something went wrong and 699 * we can not use it. Disable it. 700 */ 701 static void fpu__init_disable_system_xstate(void) 702 { 703 xfeatures_mask = 0; 704 cr4_clear_bits(X86_CR4_OSXSAVE); 705 setup_clear_cpu_cap(X86_FEATURE_XSAVE); 706 } 707 708 /* 709 * Enable and initialize the xsave feature. 710 * Called once per system bootup. 711 */ 712 void __init fpu__init_system_xstate(void) 713 { 714 unsigned int eax, ebx, ecx, edx; 715 static int on_boot_cpu __initdata = 1; 716 int err; 717 int i; 718 719 WARN_ON_FPU(!on_boot_cpu); 720 on_boot_cpu = 0; 721 722 if (!boot_cpu_has(X86_FEATURE_FPU)) { 723 pr_info("x86/fpu: No FPU detected\n"); 724 return; 725 } 726 727 if (!boot_cpu_has(X86_FEATURE_XSAVE)) { 728 pr_info("x86/fpu: x87 FPU will use %s\n", 729 boot_cpu_has(X86_FEATURE_FXSR) ? "FXSAVE" : "FSAVE"); 730 return; 731 } 732 733 if (boot_cpu_data.cpuid_level < XSTATE_CPUID) { 734 WARN_ON_FPU(1); 735 return; 736 } 737 738 cpuid_count(XSTATE_CPUID, 0, &eax, &ebx, &ecx, &edx); 739 xfeatures_mask = eax + ((u64)edx << 32); 740 741 if ((xfeatures_mask & XFEATURE_MASK_FPSSE) != XFEATURE_MASK_FPSSE) { 742 /* 743 * This indicates that something really unexpected happened 744 * with the enumeration. Disable XSAVE and try to continue 745 * booting without it. This is too early to BUG(). 746 */ 747 pr_err("x86/fpu: FP/SSE not present amongst the CPU's xstate features: 0x%llx.\n", xfeatures_mask); 748 goto out_disable; 749 } 750 751 /* 752 * Clear XSAVE features that are disabled in the normal CPUID. 753 */ 754 for (i = 0; i < ARRAY_SIZE(xsave_cpuid_features); i++) { 755 if (!boot_cpu_has(xsave_cpuid_features[i])) 756 xfeatures_mask &= ~BIT(i); 757 } 758 759 xfeatures_mask &= fpu__get_supported_xfeatures_mask(); 760 761 /* Enable xstate instructions to be able to continue with initialization: */ 762 fpu__init_cpu_xstate(); 763 err = init_xstate_size(); 764 if (err) 765 goto out_disable; 766 767 /* 768 * Update info used for ptrace frames; use standard-format size and no 769 * supervisor xstates: 770 */ 771 update_regset_xstate_info(fpu_user_xstate_size, xfeatures_mask & ~XFEATURE_MASK_SUPERVISOR); 772 773 fpu__init_prepare_fx_sw_frame(); 774 setup_init_fpu_buf(); 775 setup_xstate_comp(); 776 print_xstate_offset_size(); 777 778 pr_info("x86/fpu: Enabled xstate features 0x%llx, context size is %d bytes, using '%s' format.\n", 779 xfeatures_mask, 780 fpu_kernel_xstate_size, 781 boot_cpu_has(X86_FEATURE_XSAVES) ? "compacted" : "standard"); 782 return; 783 784 out_disable: 785 /* something went wrong, try to boot without any XSAVE support */ 786 fpu__init_disable_system_xstate(); 787 } 788 789 /* 790 * Restore minimal FPU state after suspend: 791 */ 792 void fpu__resume_cpu(void) 793 { 794 /* 795 * Restore XCR0 on xsave capable CPUs: 796 */ 797 if (boot_cpu_has(X86_FEATURE_XSAVE)) 798 xsetbv(XCR_XFEATURE_ENABLED_MASK, xfeatures_mask); 799 } 800 801 /* 802 * Given an xstate feature nr, calculate where in the xsave 803 * buffer the state is. Callers should ensure that the buffer 804 * is valid. 805 */ 806 static void *__raw_xsave_addr(struct xregs_state *xsave, int xfeature_nr) 807 { 808 if (!xfeature_enabled(xfeature_nr)) { 809 WARN_ON_FPU(1); 810 return NULL; 811 } 812 813 return (void *)xsave + xstate_comp_offsets[xfeature_nr]; 814 } 815 /* 816 * Given the xsave area and a state inside, this function returns the 817 * address of the state. 818 * 819 * This is the API that is called to get xstate address in either 820 * standard format or compacted format of xsave area. 821 * 822 * Note that if there is no data for the field in the xsave buffer 823 * this will return NULL. 824 * 825 * Inputs: 826 * xstate: the thread's storage area for all FPU data 827 * xfeature_nr: state which is defined in xsave.h (e.g. XFEATURE_FP, 828 * XFEATURE_SSE, etc...) 829 * Output: 830 * address of the state in the xsave area, or NULL if the 831 * field is not present in the xsave buffer. 832 */ 833 void *get_xsave_addr(struct xregs_state *xsave, int xfeature_nr) 834 { 835 /* 836 * Do we even *have* xsave state? 837 */ 838 if (!boot_cpu_has(X86_FEATURE_XSAVE)) 839 return NULL; 840 841 /* 842 * We should not ever be requesting features that we 843 * have not enabled. Remember that pcntxt_mask is 844 * what we write to the XCR0 register. 845 */ 846 WARN_ONCE(!(xfeatures_mask & BIT_ULL(xfeature_nr)), 847 "get of unsupported state"); 848 /* 849 * This assumes the last 'xsave*' instruction to 850 * have requested that 'xfeature_nr' be saved. 851 * If it did not, we might be seeing and old value 852 * of the field in the buffer. 853 * 854 * This can happen because the last 'xsave' did not 855 * request that this feature be saved (unlikely) 856 * or because the "init optimization" caused it 857 * to not be saved. 858 */ 859 if (!(xsave->header.xfeatures & BIT_ULL(xfeature_nr))) 860 return NULL; 861 862 return __raw_xsave_addr(xsave, xfeature_nr); 863 } 864 EXPORT_SYMBOL_GPL(get_xsave_addr); 865 866 /* 867 * This wraps up the common operations that need to occur when retrieving 868 * data from xsave state. It first ensures that the current task was 869 * using the FPU and retrieves the data in to a buffer. It then calculates 870 * the offset of the requested field in the buffer. 871 * 872 * This function is safe to call whether the FPU is in use or not. 873 * 874 * Note that this only works on the current task. 875 * 876 * Inputs: 877 * @xfeature_nr: state which is defined in xsave.h (e.g. XFEATURE_FP, 878 * XFEATURE_SSE, etc...) 879 * Output: 880 * address of the state in the xsave area or NULL if the state 881 * is not present or is in its 'init state'. 882 */ 883 const void *get_xsave_field_ptr(int xfeature_nr) 884 { 885 struct fpu *fpu = ¤t->thread.fpu; 886 887 /* 888 * fpu__save() takes the CPU's xstate registers 889 * and saves them off to the 'fpu memory buffer. 890 */ 891 fpu__save(fpu); 892 893 return get_xsave_addr(&fpu->state.xsave, xfeature_nr); 894 } 895 896 #ifdef CONFIG_ARCH_HAS_PKEYS 897 898 #define NR_VALID_PKRU_BITS (CONFIG_NR_PROTECTION_KEYS * 2) 899 #define PKRU_VALID_MASK (NR_VALID_PKRU_BITS - 1) 900 /* 901 * This will go out and modify PKRU register to set the access 902 * rights for @pkey to @init_val. 903 */ 904 int arch_set_user_pkey_access(struct task_struct *tsk, int pkey, 905 unsigned long init_val) 906 { 907 u32 old_pkru; 908 int pkey_shift = (pkey * PKRU_BITS_PER_PKEY); 909 u32 new_pkru_bits = 0; 910 911 /* 912 * This check implies XSAVE support. OSPKE only gets 913 * set if we enable XSAVE and we enable PKU in XCR0. 914 */ 915 if (!boot_cpu_has(X86_FEATURE_OSPKE)) 916 return -EINVAL; 917 918 /* Set the bits we need in PKRU: */ 919 if (init_val & PKEY_DISABLE_ACCESS) 920 new_pkru_bits |= PKRU_AD_BIT; 921 if (init_val & PKEY_DISABLE_WRITE) 922 new_pkru_bits |= PKRU_WD_BIT; 923 924 /* Shift the bits in to the correct place in PKRU for pkey: */ 925 new_pkru_bits <<= pkey_shift; 926 927 /* Get old PKRU and mask off any old bits in place: */ 928 old_pkru = read_pkru(); 929 old_pkru &= ~((PKRU_AD_BIT|PKRU_WD_BIT) << pkey_shift); 930 931 /* Write old part along with new part: */ 932 write_pkru(old_pkru | new_pkru_bits); 933 934 return 0; 935 } 936 #endif /* ! CONFIG_ARCH_HAS_PKEYS */ 937 938 /* 939 * Weird legacy quirk: SSE and YMM states store information in the 940 * MXCSR and MXCSR_FLAGS fields of the FP area. That means if the FP 941 * area is marked as unused in the xfeatures header, we need to copy 942 * MXCSR and MXCSR_FLAGS if either SSE or YMM are in use. 943 */ 944 static inline bool xfeatures_mxcsr_quirk(u64 xfeatures) 945 { 946 if (!(xfeatures & (XFEATURE_MASK_SSE|XFEATURE_MASK_YMM))) 947 return false; 948 949 if (xfeatures & XFEATURE_MASK_FP) 950 return false; 951 952 return true; 953 } 954 955 /* 956 * This is similar to user_regset_copyout(), but will not add offset to 957 * the source data pointer or increment pos, count, kbuf, and ubuf. 958 */ 959 static inline void 960 __copy_xstate_to_kernel(void *kbuf, const void *data, 961 unsigned int offset, unsigned int size, unsigned int size_total) 962 { 963 if (offset < size_total) { 964 unsigned int copy = min(size, size_total - offset); 965 966 memcpy(kbuf + offset, data, copy); 967 } 968 } 969 970 /* 971 * Convert from kernel XSAVES compacted format to standard format and copy 972 * to a kernel-space ptrace buffer. 973 * 974 * It supports partial copy but pos always starts from zero. This is called 975 * from xstateregs_get() and there we check the CPU has XSAVES. 976 */ 977 int copy_xstate_to_kernel(void *kbuf, struct xregs_state *xsave, unsigned int offset_start, unsigned int size_total) 978 { 979 unsigned int offset, size; 980 struct xstate_header header; 981 int i; 982 983 /* 984 * Currently copy_regset_to_user() starts from pos 0: 985 */ 986 if (unlikely(offset_start != 0)) 987 return -EFAULT; 988 989 /* 990 * The destination is a ptrace buffer; we put in only user xstates: 991 */ 992 memset(&header, 0, sizeof(header)); 993 header.xfeatures = xsave->header.xfeatures; 994 header.xfeatures &= ~XFEATURE_MASK_SUPERVISOR; 995 996 /* 997 * Copy xregs_state->header: 998 */ 999 offset = offsetof(struct xregs_state, header); 1000 size = sizeof(header); 1001 1002 __copy_xstate_to_kernel(kbuf, &header, offset, size, size_total); 1003 1004 for (i = 0; i < XFEATURE_MAX; i++) { 1005 /* 1006 * Copy only in-use xstates: 1007 */ 1008 if ((header.xfeatures >> i) & 1) { 1009 void *src = __raw_xsave_addr(xsave, i); 1010 1011 offset = xstate_offsets[i]; 1012 size = xstate_sizes[i]; 1013 1014 /* The next component has to fit fully into the output buffer: */ 1015 if (offset + size > size_total) 1016 break; 1017 1018 __copy_xstate_to_kernel(kbuf, src, offset, size, size_total); 1019 } 1020 1021 } 1022 1023 if (xfeatures_mxcsr_quirk(header.xfeatures)) { 1024 offset = offsetof(struct fxregs_state, mxcsr); 1025 size = MXCSR_AND_FLAGS_SIZE; 1026 __copy_xstate_to_kernel(kbuf, &xsave->i387.mxcsr, offset, size, size_total); 1027 } 1028 1029 /* 1030 * Fill xsave->i387.sw_reserved value for ptrace frame: 1031 */ 1032 offset = offsetof(struct fxregs_state, sw_reserved); 1033 size = sizeof(xstate_fx_sw_bytes); 1034 1035 __copy_xstate_to_kernel(kbuf, xstate_fx_sw_bytes, offset, size, size_total); 1036 1037 return 0; 1038 } 1039 1040 static inline int 1041 __copy_xstate_to_user(void __user *ubuf, const void *data, unsigned int offset, unsigned int size, unsigned int size_total) 1042 { 1043 if (!size) 1044 return 0; 1045 1046 if (offset < size_total) { 1047 unsigned int copy = min(size, size_total - offset); 1048 1049 if (__copy_to_user(ubuf + offset, data, copy)) 1050 return -EFAULT; 1051 } 1052 return 0; 1053 } 1054 1055 /* 1056 * Convert from kernel XSAVES compacted format to standard format and copy 1057 * to a user-space buffer. It supports partial copy but pos always starts from 1058 * zero. This is called from xstateregs_get() and there we check the CPU 1059 * has XSAVES. 1060 */ 1061 int copy_xstate_to_user(void __user *ubuf, struct xregs_state *xsave, unsigned int offset_start, unsigned int size_total) 1062 { 1063 unsigned int offset, size; 1064 int ret, i; 1065 struct xstate_header header; 1066 1067 /* 1068 * Currently copy_regset_to_user() starts from pos 0: 1069 */ 1070 if (unlikely(offset_start != 0)) 1071 return -EFAULT; 1072 1073 /* 1074 * The destination is a ptrace buffer; we put in only user xstates: 1075 */ 1076 memset(&header, 0, sizeof(header)); 1077 header.xfeatures = xsave->header.xfeatures; 1078 header.xfeatures &= ~XFEATURE_MASK_SUPERVISOR; 1079 1080 /* 1081 * Copy xregs_state->header: 1082 */ 1083 offset = offsetof(struct xregs_state, header); 1084 size = sizeof(header); 1085 1086 ret = __copy_xstate_to_user(ubuf, &header, offset, size, size_total); 1087 if (ret) 1088 return ret; 1089 1090 for (i = 0; i < XFEATURE_MAX; i++) { 1091 /* 1092 * Copy only in-use xstates: 1093 */ 1094 if ((header.xfeatures >> i) & 1) { 1095 void *src = __raw_xsave_addr(xsave, i); 1096 1097 offset = xstate_offsets[i]; 1098 size = xstate_sizes[i]; 1099 1100 /* The next component has to fit fully into the output buffer: */ 1101 if (offset + size > size_total) 1102 break; 1103 1104 ret = __copy_xstate_to_user(ubuf, src, offset, size, size_total); 1105 if (ret) 1106 return ret; 1107 } 1108 1109 } 1110 1111 if (xfeatures_mxcsr_quirk(header.xfeatures)) { 1112 offset = offsetof(struct fxregs_state, mxcsr); 1113 size = MXCSR_AND_FLAGS_SIZE; 1114 __copy_xstate_to_user(ubuf, &xsave->i387.mxcsr, offset, size, size_total); 1115 } 1116 1117 /* 1118 * Fill xsave->i387.sw_reserved value for ptrace frame: 1119 */ 1120 offset = offsetof(struct fxregs_state, sw_reserved); 1121 size = sizeof(xstate_fx_sw_bytes); 1122 1123 ret = __copy_xstate_to_user(ubuf, xstate_fx_sw_bytes, offset, size, size_total); 1124 if (ret) 1125 return ret; 1126 1127 return 0; 1128 } 1129 1130 /* 1131 * Convert from a ptrace standard-format kernel buffer to kernel XSAVES format 1132 * and copy to the target thread. This is called from xstateregs_set(). 1133 */ 1134 int copy_kernel_to_xstate(struct xregs_state *xsave, const void *kbuf) 1135 { 1136 unsigned int offset, size; 1137 int i; 1138 struct xstate_header hdr; 1139 1140 offset = offsetof(struct xregs_state, header); 1141 size = sizeof(hdr); 1142 1143 memcpy(&hdr, kbuf + offset, size); 1144 1145 if (validate_xstate_header(&hdr)) 1146 return -EINVAL; 1147 1148 for (i = 0; i < XFEATURE_MAX; i++) { 1149 u64 mask = ((u64)1 << i); 1150 1151 if (hdr.xfeatures & mask) { 1152 void *dst = __raw_xsave_addr(xsave, i); 1153 1154 offset = xstate_offsets[i]; 1155 size = xstate_sizes[i]; 1156 1157 memcpy(dst, kbuf + offset, size); 1158 } 1159 } 1160 1161 if (xfeatures_mxcsr_quirk(hdr.xfeatures)) { 1162 offset = offsetof(struct fxregs_state, mxcsr); 1163 size = MXCSR_AND_FLAGS_SIZE; 1164 memcpy(&xsave->i387.mxcsr, kbuf + offset, size); 1165 } 1166 1167 /* 1168 * The state that came in from userspace was user-state only. 1169 * Mask all the user states out of 'xfeatures': 1170 */ 1171 xsave->header.xfeatures &= XFEATURE_MASK_SUPERVISOR; 1172 1173 /* 1174 * Add back in the features that came in from userspace: 1175 */ 1176 xsave->header.xfeatures |= hdr.xfeatures; 1177 1178 return 0; 1179 } 1180 1181 /* 1182 * Convert from a ptrace or sigreturn standard-format user-space buffer to 1183 * kernel XSAVES format and copy to the target thread. This is called from 1184 * xstateregs_set(), as well as potentially from the sigreturn() and 1185 * rt_sigreturn() system calls. 1186 */ 1187 int copy_user_to_xstate(struct xregs_state *xsave, const void __user *ubuf) 1188 { 1189 unsigned int offset, size; 1190 int i; 1191 struct xstate_header hdr; 1192 1193 offset = offsetof(struct xregs_state, header); 1194 size = sizeof(hdr); 1195 1196 if (__copy_from_user(&hdr, ubuf + offset, size)) 1197 return -EFAULT; 1198 1199 if (validate_xstate_header(&hdr)) 1200 return -EINVAL; 1201 1202 for (i = 0; i < XFEATURE_MAX; i++) { 1203 u64 mask = ((u64)1 << i); 1204 1205 if (hdr.xfeatures & mask) { 1206 void *dst = __raw_xsave_addr(xsave, i); 1207 1208 offset = xstate_offsets[i]; 1209 size = xstate_sizes[i]; 1210 1211 if (__copy_from_user(dst, ubuf + offset, size)) 1212 return -EFAULT; 1213 } 1214 } 1215 1216 if (xfeatures_mxcsr_quirk(hdr.xfeatures)) { 1217 offset = offsetof(struct fxregs_state, mxcsr); 1218 size = MXCSR_AND_FLAGS_SIZE; 1219 if (__copy_from_user(&xsave->i387.mxcsr, ubuf + offset, size)) 1220 return -EFAULT; 1221 } 1222 1223 /* 1224 * The state that came in from userspace was user-state only. 1225 * Mask all the user states out of 'xfeatures': 1226 */ 1227 xsave->header.xfeatures &= XFEATURE_MASK_SUPERVISOR; 1228 1229 /* 1230 * Add back in the features that came in from userspace: 1231 */ 1232 xsave->header.xfeatures |= hdr.xfeatures; 1233 1234 return 0; 1235 } 1236 1237 #ifdef CONFIG_PROC_PID_ARCH_STATUS 1238 /* 1239 * Report the amount of time elapsed in millisecond since last AVX512 1240 * use in the task. 1241 */ 1242 static void avx512_status(struct seq_file *m, struct task_struct *task) 1243 { 1244 unsigned long timestamp = READ_ONCE(task->thread.fpu.avx512_timestamp); 1245 long delta; 1246 1247 if (!timestamp) { 1248 /* 1249 * Report -1 if no AVX512 usage 1250 */ 1251 delta = -1; 1252 } else { 1253 delta = (long)(jiffies - timestamp); 1254 /* 1255 * Cap to LONG_MAX if time difference > LONG_MAX 1256 */ 1257 if (delta < 0) 1258 delta = LONG_MAX; 1259 delta = jiffies_to_msecs(delta); 1260 } 1261 1262 seq_put_decimal_ll(m, "AVX512_elapsed_ms:\t", delta); 1263 seq_putc(m, '\n'); 1264 } 1265 1266 /* 1267 * Report architecture specific information 1268 */ 1269 int proc_pid_arch_status(struct seq_file *m, struct pid_namespace *ns, 1270 struct pid *pid, struct task_struct *task) 1271 { 1272 /* 1273 * Report AVX512 state if the processor and build option supported. 1274 */ 1275 if (cpu_feature_enabled(X86_FEATURE_AVX512F)) 1276 avx512_status(m, task); 1277 1278 return 0; 1279 } 1280 #endif /* CONFIG_PROC_PID_ARCH_STATUS */ 1281