arch/x86/kernel/fpu/xstate.c

   1 /*
   2  * xsave/xrstor support.
   3  *
   4  * Author: Suresh Siddha <suresh.b.siddha@intel.com>
   5  */
   6 #include <linux/compat.h>
   7 #include <linux/cpu.h>
   8 #include <linux/mman.h>
   9 #include <linux/pkeys.h>
  10
  11 #include <asm/fpu/api.h>
  12 #include <asm/fpu/internal.h>
  13 #include <asm/fpu/signal.h>
  14 #include <asm/fpu/regset.h>
  15 #include <asm/fpu/xstate.h>
  16
  17 #include <asm/tlbflush.h>
  18
  19 /*
  20  * Although we spell it out in here, the Processor Trace
  21  * xfeature is completely unused.  We use other mechanisms
  22  * to save/restore PT state in Linux.
  23  */
  24 static const char *xfeature_names[] =
  25 {
  26         "x87 floating point registers"  ,
  27         "SSE registers"                 ,
  28         "AVX registers"                 ,
  29         "MPX bounds registers"          ,
  30         "MPX CSR"                       ,
  31         "AVX-512 opmask"                ,
  32         "AVX-512 Hi256"                 ,
  33         "AVX-512 ZMM_Hi256"             ,
  34         "Processor Trace (unused)"      ,
  35         "Protection Keys User registers",
  36         "unknown xstate feature"        ,
  37 };
  38
  39 /*
  40  * Mask of xstate features supported by the CPU and the kernel:
  41  */
  42 u64 xfeatures_mask __read_mostly;
  43
  44 static unsigned int xstate_offsets[XFEATURE_MAX] = { [ 0 ... XFEATURE_MAX - 1] = -1};
  45 static unsigned int xstate_sizes[XFEATURE_MAX]   = { [ 0 ... XFEATURE_MAX - 1] = -1};
  46 static unsigned int xstate_comp_offsets[sizeof(xfeatures_mask)*8];
  47
  48 /*
  49  * The XSAVE area of kernel can be in standard or compacted format;
  50  * it is always in standard format for user mode. This is the user
  51  * mode standard format size used for signal and ptrace frames.
  52  */
  53 unsigned int fpu_user_xstate_size;
  54
  55 /*
  56  * Clear all of the X86_FEATURE_* bits that are unavailable
  57  * when the CPU has no XSAVE support.
  58  */
  59 void fpu__xstate_clear_all_cpu_caps(void)
  60 {
  61         setup_clear_cpu_cap(X86_FEATURE_XSAVE);
  62         setup_clear_cpu_cap(X86_FEATURE_XSAVEOPT);
  63         setup_clear_cpu_cap(X86_FEATURE_XSAVEC);
  64         setup_clear_cpu_cap(X86_FEATURE_XSAVES);
  65         setup_clear_cpu_cap(X86_FEATURE_AVX);
  66         setup_clear_cpu_cap(X86_FEATURE_AVX2);
  67         setup_clear_cpu_cap(X86_FEATURE_AVX512F);
  68         setup_clear_cpu_cap(X86_FEATURE_AVX512PF);
  69         setup_clear_cpu_cap(X86_FEATURE_AVX512ER);
  70         setup_clear_cpu_cap(X86_FEATURE_AVX512CD);
  71         setup_clear_cpu_cap(X86_FEATURE_AVX512DQ);
  72         setup_clear_cpu_cap(X86_FEATURE_AVX512BW);
  73         setup_clear_cpu_cap(X86_FEATURE_AVX512VL);
  74         setup_clear_cpu_cap(X86_FEATURE_MPX);
  75         setup_clear_cpu_cap(X86_FEATURE_XGETBV1);
  76         setup_clear_cpu_cap(X86_FEATURE_PKU);
  77         setup_clear_cpu_cap(X86_FEATURE_AVX512_4VNNIW);
  78         setup_clear_cpu_cap(X86_FEATURE_AVX512_4FMAPS);
  79 }
  80
  81 /*
  82  * Return whether the system supports a given xfeature.
  83  *
  84  * Also return the name of the (most advanced) feature that the caller requested:
  85  */
  86 int cpu_has_xfeatures(u64 xfeatures_needed, const char **feature_name)
  87 {
  88         u64 xfeatures_missing = xfeatures_needed & ~xfeatures_mask;
  89
  90         if (unlikely(feature_name)) {
  91                 long xfeature_idx, max_idx;
  92                 u64 xfeatures_print;
  93                 /*
  94                  * So we use FLS here to be able to print the most advanced
  95                  * feature that was requested but is missing. So if a driver
  96                  * asks about "XFEATURE_MASK_SSE | XFEATURE_MASK_YMM" we'll print the
  97                  * missing AVX feature - this is the most informative message
  98                  * to users:
  99                  */
 100                 if (xfeatures_missing)
 101                         xfeatures_print = xfeatures_missing;
 102                 else
 103                         xfeatures_print = xfeatures_needed;
 104
 105                 xfeature_idx = fls64(xfeatures_print)-1;
 106                 max_idx = ARRAY_SIZE(xfeature_names)-1;
 107                 xfeature_idx = min(xfeature_idx, max_idx);
 108
 109                 *feature_name = xfeature_names[xfeature_idx];
 110         }
 111
 112         if (xfeatures_missing)
 113                 return 0;
 114
 115         return 1;
 116 }
 117 EXPORT_SYMBOL_GPL(cpu_has_xfeatures);
 118
 119 static int xfeature_is_supervisor(int xfeature_nr)
 120 {
 121         /*
 122          * We currently do not support supervisor states, but if
 123          * we did, we could find out like this.
 124          *
 125          * SDM says: If state component 'i' is a user state component,
 126          * ECX[0] return 0; if state component i is a supervisor
 127          * state component, ECX[0] returns 1.
 128          */
 129         u32 eax, ebx, ecx, edx;
 130
 131         cpuid_count(XSTATE_CPUID, xfeature_nr, &eax, &ebx, &ecx, &edx);
 132         return !!(ecx & 1);
 133 }
 134
 135 static int xfeature_is_user(int xfeature_nr)
 136 {
 137         return !xfeature_is_supervisor(xfeature_nr);
 138 }
 139
 140 /*
 141  * When executing XSAVEOPT (or other optimized XSAVE instructions), if
 142  * a processor implementation detects that an FPU state component is still
 143  * (or is again) in its initialized state, it may clear the corresponding
 144  * bit in the header.xfeatures field, and can skip the writeout of registers
 145  * to the corresponding memory layout.
 146  *
 147  * This means that when the bit is zero, the state component might still contain
 148  * some previous - non-initialized register state.
 149  *
 150  * Before writing xstate information to user-space we sanitize those components,
 151  * to always ensure that the memory layout of a feature will be in the init state
 152  * if the corresponding header bit is zero. This is to ensure that user-space doesn't
 153  * see some stale state in the memory layout during signal handling, debugging etc.
 154  */
 155 void fpstate_sanitize_xstate(struct fpu *fpu)
 156 {
 157         struct fxregs_state *fx = &fpu->state.fxsave;
 158         int feature_bit;
 159         u64 xfeatures;
 160
 161         if (!use_xsaveopt())
 162                 return;
 163
 164         xfeatures = fpu->state.xsave.header.xfeatures;
 165
 166         /*
 167          * None of the feature bits are in init state. So nothing else
 168          * to do for us, as the memory layout is up to date.
 169          */
 170         if ((xfeatures & xfeatures_mask) == xfeatures_mask)
 171                 return;
 172
 173         /*
 174          * FP is in init state
 175          */
 176         if (!(xfeatures & XFEATURE_MASK_FP)) {
 177                 fx->cwd = 0x37f;
 178                 fx->swd = 0;
 179                 fx->twd = 0;
 180                 fx->fop = 0;
 181                 fx->rip = 0;
 182                 fx->rdp = 0;
 183                 memset(&fx->st_space[0], 0, 128);
 184         }
 185
 186         /*
 187          * SSE is in init state
 188          */
 189         if (!(xfeatures & XFEATURE_MASK_SSE))
 190                 memset(&fx->xmm_space[0], 0, 256);
 191
 192         /*
 193          * First two features are FPU and SSE, which above we handled
 194          * in a special way already:
 195          */
 196         feature_bit = 0x2;
 197         xfeatures = (xfeatures_mask & ~xfeatures) >> 2;
 198
 199         /*
 200          * Update all the remaining memory layouts according to their
 201          * standard xstate layout, if their header bit is in the init
 202          * state:
 203          */
 204         while (xfeatures) {
 205                 if (xfeatures & 0x1) {
 206                         int offset = xstate_comp_offsets[feature_bit];
 207                         int size = xstate_sizes[feature_bit];
 208
 209                         memcpy((void *)fx + offset,
 210                                (void *)&init_fpstate.xsave + offset,
 211                                size);
 212                 }
 213
 214                 xfeatures >>= 1;
 215                 feature_bit++;
 216         }
 217 }
 218
 219 /*
 220  * Enable the extended processor state save/restore feature.
 221  * Called once per CPU onlining.
 222  */
 223 void fpu__init_cpu_xstate(void)
 224 {
 225         if (!boot_cpu_has(X86_FEATURE_XSAVE) || !xfeatures_mask)
 226                 return;
 227         /*
 228          * Make it clear that XSAVES supervisor states are not yet
 229          * implemented should anyone expect it to work by changing
 230          * bits in XFEATURE_MASK_* macros and XCR0.
 231          */
 232         WARN_ONCE((xfeatures_mask & XFEATURE_MASK_SUPERVISOR),
 233                 "x86/fpu: XSAVES supervisor states are not yet implemented.\n");
 234
 235         xfeatures_mask &= ~XFEATURE_MASK_SUPERVISOR;
 236
 237         cr4_set_bits(X86_CR4_OSXSAVE);
 238         xsetbv(XCR_XFEATURE_ENABLED_MASK, xfeatures_mask);
 239 }
 240
 241 /*
 242  * Note that in the future we will likely need a pair of
 243  * functions here: one for user xstates and the other for
 244  * system xstates.  For now, they are the same.
 245  */
 246 static int xfeature_enabled(enum xfeature xfeature)
 247 {
 248         return !!(xfeatures_mask & (1UL << xfeature));
 249 }
 250
 251 /*
 252  * Record the offsets and sizes of various xstates contained
 253  * in the XSAVE state memory layout.
 254  */
 255 static void __init setup_xstate_features(void)
 256 {
 257         u32 eax, ebx, ecx, edx, i;
 258         /* start at the beginnning of the "extended state" */
 259         unsigned int last_good_offset = offsetof(struct xregs_state,
 260                                                  extended_state_area);
 261         /*
 262          * The FP xstates and SSE xstates are legacy states. They are always
 263          * in the fixed offsets in the xsave area in either compacted form
 264          * or standard form.
 265          */
 266         xstate_offsets[0] = 0;
 267         xstate_sizes[0] = offsetof(struct fxregs_state, xmm_space);
 268         xstate_offsets[1] = xstate_sizes[0];
 269         xstate_sizes[1] = FIELD_SIZEOF(struct fxregs_state, xmm_space);
 270
 271         for (i = FIRST_EXTENDED_XFEATURE; i < XFEATURE_MAX; i++) {
 272                 if (!xfeature_enabled(i))
 273                         continue;
 274
 275                 cpuid_count(XSTATE_CPUID, i, &eax, &ebx, &ecx, &edx);
 276
 277                 /*
 278                  * If an xfeature is supervisor state, the offset
 279                  * in EBX is invalid. We leave it to -1.
 280                  */
 281                 if (xfeature_is_user(i))
 282                         xstate_offsets[i] = ebx;
 283
 284                 xstate_sizes[i] = eax;
 285                 /*
 286                  * In our xstate size checks, we assume that the
 287                  * highest-numbered xstate feature has the
 288                  * highest offset in the buffer.  Ensure it does.
 289                  */
 290                 WARN_ONCE(last_good_offset > xstate_offsets[i],
 291                         "x86/fpu: misordered xstate at %d\n", last_good_offset);
 292                 last_good_offset = xstate_offsets[i];
 293         }
 294 }
 295
 296 static void __init print_xstate_feature(u64 xstate_mask)
 297 {
 298         const char *feature_name;
 299
 300         if (cpu_has_xfeatures(xstate_mask, &feature_name))
 301                 pr_info("x86/fpu: Supporting XSAVE feature 0x%03Lx: '%s'\n", xstate_mask, feature_name);
 302 }
 303
 304 /*
 305  * Print out all the supported xstate features:
 306  */
 307 static void __init print_xstate_features(void)
 308 {
 309         print_xstate_feature(XFEATURE_MASK_FP);
 310         print_xstate_feature(XFEATURE_MASK_SSE);
 311         print_xstate_feature(XFEATURE_MASK_YMM);
 312         print_xstate_feature(XFEATURE_MASK_BNDREGS);
 313         print_xstate_feature(XFEATURE_MASK_BNDCSR);
 314         print_xstate_feature(XFEATURE_MASK_OPMASK);
 315         print_xstate_feature(XFEATURE_MASK_ZMM_Hi256);
 316         print_xstate_feature(XFEATURE_MASK_Hi16_ZMM);
 317         print_xstate_feature(XFEATURE_MASK_PKRU);
 318 }
 319
 320 /*
 321  * This check is important because it is easy to get XSTATE_*
 322  * confused with XSTATE_BIT_*.
 323  */
 324 #define CHECK_XFEATURE(nr) do {         \
 325         WARN_ON(nr < FIRST_EXTENDED_XFEATURE);  \
 326         WARN_ON(nr >= XFEATURE_MAX);    \
 327 } while (0)
 328
 329 /*
 330  * We could cache this like xstate_size[], but we only use
 331  * it here, so it would be a waste of space.
 332  */
 333 static int xfeature_is_aligned(int xfeature_nr)
 334 {
 335         u32 eax, ebx, ecx, edx;
 336
 337         CHECK_XFEATURE(xfeature_nr);
 338         cpuid_count(XSTATE_CPUID, xfeature_nr, &eax, &ebx, &ecx, &edx);
 339         /*
 340          * The value returned by ECX[1] indicates the alignment
 341          * of state component 'i' when the compacted format
 342          * of the extended region of an XSAVE area is used:
 343          */
 344         return !!(ecx & 2);
 345 }
 346
 347 /*
 348  * This function sets up offsets and sizes of all extended states in
 349  * xsave area. This supports both standard format and compacted format
 350  * of the xsave aread.
 351  */
 352 static void __init setup_xstate_comp(void)
 353 {
 354         unsigned int xstate_comp_sizes[sizeof(xfeatures_mask)*8];
 355         int i;
 356
 357         /*
 358          * The FP xstates and SSE xstates are legacy states. They are always
 359          * in the fixed offsets in the xsave area in either compacted form
 360          * or standard form.
 361          */
 362         xstate_comp_offsets[0] = 0;
 363         xstate_comp_offsets[1] = offsetof(struct fxregs_state, xmm_space);
 364
 365         if (!boot_cpu_has(X86_FEATURE_XSAVES)) {
 366                 for (i = FIRST_EXTENDED_XFEATURE; i < XFEATURE_MAX; i++) {
 367                         if (xfeature_enabled(i)) {
 368                                 xstate_comp_offsets[i] = xstate_offsets[i];
 369                                 xstate_comp_sizes[i] = xstate_sizes[i];
 370                         }
 371                 }
 372                 return;
 373         }
 374
 375         xstate_comp_offsets[FIRST_EXTENDED_XFEATURE] =
 376                 FXSAVE_SIZE + XSAVE_HDR_SIZE;
 377
 378         for (i = FIRST_EXTENDED_XFEATURE; i < XFEATURE_MAX; i++) {
 379                 if (xfeature_enabled(i))
 380                         xstate_comp_sizes[i] = xstate_sizes[i];
 381                 else
 382                         xstate_comp_sizes[i] = 0;
 383
 384                 if (i > FIRST_EXTENDED_XFEATURE) {
 385                         xstate_comp_offsets[i] = xstate_comp_offsets[i-1]
 386                                         + xstate_comp_sizes[i-1];
 387
 388                         if (xfeature_is_aligned(i))
 389                                 xstate_comp_offsets[i] =
 390                                         ALIGN(xstate_comp_offsets[i], 64);
 391                 }
 392         }
 393 }
 394
 395 /*
 396  * Print out xstate component offsets and sizes
 397  */
 398 static void __init print_xstate_offset_size(void)
 399 {
 400         int i;
 401
 402         for (i = FIRST_EXTENDED_XFEATURE; i < XFEATURE_MAX; i++) {
 403                 if (!xfeature_enabled(i))
 404                         continue;
 405                 pr_info("x86/fpu: xstate_offset[%d]: %4d, xstate_sizes[%d]: %4d\n",
 406                          i, xstate_comp_offsets[i], i, xstate_sizes[i]);
 407         }
 408 }
 409
 410 /*
 411  * setup the xstate image representing the init state
 412  */
 413 static void __init setup_init_fpu_buf(void)
 414 {
 415         static int on_boot_cpu __initdata = 1;
 416
 417         WARN_ON_FPU(!on_boot_cpu);
 418         on_boot_cpu = 0;
 419
 420         if (!boot_cpu_has(X86_FEATURE_XSAVE))
 421                 return;
 422
 423         setup_xstate_features();
 424         print_xstate_features();
 425
 426         if (boot_cpu_has(X86_FEATURE_XSAVES))
 427                 init_fpstate.xsave.header.xcomp_bv = (u64)1 << 63 | xfeatures_mask;
 428
 429         /*
 430          * Init all the features state with header.xfeatures being 0x0
 431          */
 432         copy_kernel_to_xregs_booting(&init_fpstate.xsave);
 433
 434         /*
 435          * Dump the init state again. This is to identify the init state
 436          * of any feature which is not represented by all zero's.
 437          */
 438         copy_xregs_to_kernel_booting(&init_fpstate.xsave);
 439 }
 440
 441 static int xfeature_uncompacted_offset(int xfeature_nr)
 442 {
 443         u32 eax, ebx, ecx, edx;
 444
 445         /*
 446          * Only XSAVES supports supervisor states and it uses compacted
 447          * format. Checking a supervisor state's uncompacted offset is
 448          * an error.
 449          */
 450         if (XFEATURE_MASK_SUPERVISOR & (1 << xfeature_nr)) {
 451                 WARN_ONCE(1, "No fixed offset for xstate %d\n", xfeature_nr);
 452                 return -1;
 453         }
 454
 455         CHECK_XFEATURE(xfeature_nr);
 456         cpuid_count(XSTATE_CPUID, xfeature_nr, &eax, &ebx, &ecx, &edx);
 457         return ebx;
 458 }
 459
 460 static int xfeature_size(int xfeature_nr)
 461 {
 462         u32 eax, ebx, ecx, edx;
 463
 464         CHECK_XFEATURE(xfeature_nr);
 465         cpuid_count(XSTATE_CPUID, xfeature_nr, &eax, &ebx, &ecx, &edx);
 466         return eax;
 467 }
 468
 469 /*
 470  * 'XSAVES' implies two different things:
 471  * 1. saving of supervisor/system state
 472  * 2. using the compacted format
 473  *
 474  * Use this function when dealing with the compacted format so
 475  * that it is obvious which aspect of 'XSAVES' is being handled
 476  * by the calling code.
 477  */
 478 int using_compacted_format(void)
 479 {
 480         return boot_cpu_has(X86_FEATURE_XSAVES);
 481 }
 482
 483 static void __xstate_dump_leaves(void)
 484 {
 485         int i;
 486         u32 eax, ebx, ecx, edx;
 487         static int should_dump = 1;
 488
 489         if (!should_dump)
 490                 return;
 491         should_dump = 0;
 492         /*
 493          * Dump out a few leaves past the ones that we support
 494          * just in case there are some goodies up there
 495          */
 496         for (i = 0; i < XFEATURE_MAX + 10; i++) {
 497                 cpuid_count(XSTATE_CPUID, i, &eax, &ebx, &ecx, &edx);
 498                 pr_warn("CPUID[%02x, %02x]: eax=%08x ebx=%08x ecx=%08x edx=%08x\n",
 499                         XSTATE_CPUID, i, eax, ebx, ecx, edx);
 500         }
 501 }
 502
 503 #define XSTATE_WARN_ON(x) do {                                                  \
 504         if (WARN_ONCE(x, "XSAVE consistency problem, dumping leaves")) {        \
 505                 __xstate_dump_leaves();                                         \
 506         }                                                                       \
 507 } while (0)
 508
 509 #define XCHECK_SZ(sz, nr, nr_macro, __struct) do {                      \
 510         if ((nr == nr_macro) &&                                         \
 511             WARN_ONCE(sz != sizeof(__struct),                           \
 512                 "%s: struct is %zu bytes, cpu state %d bytes\n",        \
 513                 __stringify(nr_macro), sizeof(__struct), sz)) {         \
 514                 __xstate_dump_leaves();                                 \
 515         }                                                               \
 516 } while (0)
 517
 518 /*
 519  * We have a C struct for each 'xstate'.  We need to ensure
 520  * that our software representation matches what the CPU
 521  * tells us about the state's size.
 522  */
 523 static void check_xstate_against_struct(int nr)
 524 {
 525         /*
 526          * Ask the CPU for the size of the state.
 527          */
 528         int sz = xfeature_size(nr);
 529         /*
 530          * Match each CPU state with the corresponding software
 531          * structure.
 532          */
 533         XCHECK_SZ(sz, nr, XFEATURE_YMM,       struct ymmh_struct);
 534         XCHECK_SZ(sz, nr, XFEATURE_BNDREGS,   struct mpx_bndreg_state);
 535         XCHECK_SZ(sz, nr, XFEATURE_BNDCSR,    struct mpx_bndcsr_state);
 536         XCHECK_SZ(sz, nr, XFEATURE_OPMASK,    struct avx_512_opmask_state);
 537         XCHECK_SZ(sz, nr, XFEATURE_ZMM_Hi256, struct avx_512_zmm_uppers_state);
 538         XCHECK_SZ(sz, nr, XFEATURE_Hi16_ZMM,  struct avx_512_hi16_state);
 539         XCHECK_SZ(sz, nr, XFEATURE_PKRU,      struct pkru_state);
 540
 541         /*
 542          * Make *SURE* to add any feature numbers in below if
 543          * there are "holes" in the xsave state component
 544          * numbers.
 545          */
 546         if ((nr < XFEATURE_YMM) ||
 547             (nr >= XFEATURE_MAX) ||
 548             (nr == XFEATURE_PT_UNIMPLEMENTED_SO_FAR)) {
 549                 WARN_ONCE(1, "no structure for xstate: %d\n", nr);
 550                 XSTATE_WARN_ON(1);
 551         }
 552 }
 553
 554 /*
 555  * This essentially double-checks what the cpu told us about
 556  * how large the XSAVE buffer needs to be.  We are recalculating
 557  * it to be safe.
 558  */
 559 static void do_extra_xstate_size_checks(void)
 560 {
 561         int paranoid_xstate_size = FXSAVE_SIZE + XSAVE_HDR_SIZE;
 562         int i;
 563
 564         for (i = FIRST_EXTENDED_XFEATURE; i < XFEATURE_MAX; i++) {
 565                 if (!xfeature_enabled(i))
 566                         continue;
 567
 568                 check_xstate_against_struct(i);
 569                 /*
 570                  * Supervisor state components can be managed only by
 571                  * XSAVES, which is compacted-format only.
 572                  */
 573                 if (!using_compacted_format())
 574                         XSTATE_WARN_ON(xfeature_is_supervisor(i));
 575
 576                 /* Align from the end of the previous feature */
 577                 if (xfeature_is_aligned(i))
 578                         paranoid_xstate_size = ALIGN(paranoid_xstate_size, 64);
 579                 /*
 580                  * The offset of a given state in the non-compacted
 581                  * format is given to us in a CPUID leaf.  We check
 582                  * them for being ordered (increasing offsets) in
 583                  * setup_xstate_features().
 584                  */
 585                 if (!using_compacted_format())
 586                         paranoid_xstate_size = xfeature_uncompacted_offset(i);
 587                 /*
 588                  * The compacted-format offset always depends on where
 589                  * the previous state ended.
 590                  */
 591                 paranoid_xstate_size += xfeature_size(i);
 592         }
 593         XSTATE_WARN_ON(paranoid_xstate_size != fpu_kernel_xstate_size);
 594 }
 595
 596
 597 /*
 598  * Get total size of enabled xstates in XCR0/xfeatures_mask.
 599  *
 600  * Note the SDM's wording here.  "sub-function 0" only enumerates
 601  * the size of the *user* states.  If we use it to size a buffer
 602  * that we use 'XSAVES' on, we could potentially overflow the
 603  * buffer because 'XSAVES' saves system states too.
 604  *
 605  * Note that we do not currently set any bits on IA32_XSS so
 606  * 'XCR0 | IA32_XSS == XCR0' for now.
 607  */
 608 static unsigned int __init get_xsaves_size(void)
 609 {
 610         unsigned int eax, ebx, ecx, edx;
 611         /*
 612          * - CPUID function 0DH, sub-function 1:
 613          *    EBX enumerates the size (in bytes) required by
 614          *    the XSAVES instruction for an XSAVE area
 615          *    containing all the state components
 616          *    corresponding to bits currently set in
 617          *    XCR0 | IA32_XSS.
 618          */
 619         cpuid_count(XSTATE_CPUID, 1, &eax, &ebx, &ecx, &edx);
 620         return ebx;
 621 }
 622
 623 static unsigned int __init get_xsave_size(void)
 624 {
 625         unsigned int eax, ebx, ecx, edx;
 626         /*
 627          * - CPUID function 0DH, sub-function 0:
 628          *    EBX enumerates the size (in bytes) required by
 629          *    the XSAVE instruction for an XSAVE area
 630          *    containing all the *user* state components
 631          *    corresponding to bits currently set in XCR0.
 632          */
 633         cpuid_count(XSTATE_CPUID, 0, &eax, &ebx, &ecx, &edx);
 634         return ebx;
 635 }
 636
 637 /*
 638  * Will the runtime-enumerated 'xstate_size' fit in the init
 639  * task's statically-allocated buffer?
 640  */
 641 static bool is_supported_xstate_size(unsigned int test_xstate_size)
 642 {
 643         if (test_xstate_size <= sizeof(union fpregs_state))
 644                 return true;
 645
 646         pr_warn("x86/fpu: xstate buffer too small (%zu < %d), disabling xsave\n",
 647                         sizeof(union fpregs_state), test_xstate_size);
 648         return false;
 649 }
 650
 651 static int init_xstate_size(void)
 652 {
 653         /* Recompute the context size for enabled features: */
 654         unsigned int possible_xstate_size;
 655         unsigned int xsave_size;
 656
 657         xsave_size = get_xsave_size();
 658
 659         if (boot_cpu_has(X86_FEATURE_XSAVES))
 660                 possible_xstate_size = get_xsaves_size();
 661         else
 662                 possible_xstate_size = xsave_size;
 663
 664         /* Ensure we have the space to store all enabled: */
 665         if (!is_supported_xstate_size(possible_xstate_size))
 666                 return -EINVAL;
 667
 668         /*
 669          * The size is OK, we are definitely going to use xsave,
 670          * make it known to the world that we need more space.
 671          */
 672         fpu_kernel_xstate_size = possible_xstate_size;
 673         do_extra_xstate_size_checks();
 674
 675         /*
 676          * User space is always in standard format.
 677          */
 678         fpu_user_xstate_size = xsave_size;
 679         return 0;
 680 }
 681
 682 /*
 683  * We enabled the XSAVE hardware, but something went wrong and
 684  * we can not use it.  Disable it.
 685  */
 686 static void fpu__init_disable_system_xstate(void)
 687 {
 688         xfeatures_mask = 0;
 689         cr4_clear_bits(X86_CR4_OSXSAVE);
 690         fpu__xstate_clear_all_cpu_caps();
 691 }
 692
 693 /*
 694  * Enable and initialize the xsave feature.
 695  * Called once per system bootup.
 696  */
 697 void __init fpu__init_system_xstate(void)
 698 {
 699         unsigned int eax, ebx, ecx, edx;
 700         static int on_boot_cpu __initdata = 1;
 701         int err;
 702
 703         WARN_ON_FPU(!on_boot_cpu);
 704         on_boot_cpu = 0;
 705
 706         if (!boot_cpu_has(X86_FEATURE_XSAVE)) {
 707                 pr_info("x86/fpu: Legacy x87 FPU detected.\n");
 708                 return;
 709         }
 710
 711         if (boot_cpu_data.cpuid_level < XSTATE_CPUID) {
 712                 WARN_ON_FPU(1);
 713                 return;
 714         }
 715
 716         cpuid_count(XSTATE_CPUID, 0, &eax, &ebx, &ecx, &edx);
 717         xfeatures_mask = eax + ((u64)edx << 32);
 718
 719         if ((xfeatures_mask & XFEATURE_MASK_FPSSE) != XFEATURE_MASK_FPSSE) {
 720                 /*
 721                  * This indicates that something really unexpected happened
 722                  * with the enumeration.  Disable XSAVE and try to continue
 723                  * booting without it.  This is too early to BUG().
 724                  */
 725                 pr_err("x86/fpu: FP/SSE not present amongst the CPU's xstate features: 0x%llx.\n", xfeatures_mask);
 726                 goto out_disable;
 727         }
 728
 729         xfeatures_mask &= fpu__get_supported_xfeatures_mask();
 730
 731         /* Enable xstate instructions to be able to continue with initialization: */
 732         fpu__init_cpu_xstate();
 733         err = init_xstate_size();
 734         if (err)
 735                 goto out_disable;
 736
 737         /*
 738          * Update info used for ptrace frames; use standard-format size and no
 739          * supervisor xstates:
 740          */
 741         update_regset_xstate_info(fpu_user_xstate_size, xfeatures_mask & ~XFEATURE_MASK_SUPERVISOR);
 742
 743         fpu__init_prepare_fx_sw_frame();
 744         setup_init_fpu_buf();
 745         setup_xstate_comp();
 746         print_xstate_offset_size();
 747
 748         pr_info("x86/fpu: Enabled xstate features 0x%llx, context size is %d bytes, using '%s' format.\n",
 749                 xfeatures_mask,
 750                 fpu_kernel_xstate_size,
 751                 boot_cpu_has(X86_FEATURE_XSAVES) ? "compacted" : "standard");
 752         return;
 753
 754 out_disable:
 755         /* something went wrong, try to boot without any XSAVE support */
 756         fpu__init_disable_system_xstate();
 757 }
 758
 759 /*
 760  * Restore minimal FPU state after suspend:
 761  */
 762 void fpu__resume_cpu(void)
 763 {
 764         /*
 765          * Restore XCR0 on xsave capable CPUs:
 766          */
 767         if (boot_cpu_has(X86_FEATURE_XSAVE))
 768                 xsetbv(XCR_XFEATURE_ENABLED_MASK, xfeatures_mask);
 769 }
 770
 771 /*
 772  * Given an xstate feature mask, calculate where in the xsave
 773  * buffer the state is.  Callers should ensure that the buffer
 774  * is valid.
 775  *
 776  * Note: does not work for compacted buffers.
 777  */
 778 void *__raw_xsave_addr(struct xregs_state *xsave, int xstate_feature_mask)
 779 {
 780         int feature_nr = fls64(xstate_feature_mask) - 1;
 781
 782         if (!xfeature_enabled(feature_nr)) {
 783                 WARN_ON_FPU(1);
 784                 return NULL;
 785         }
 786
 787         return (void *)xsave + xstate_comp_offsets[feature_nr];
 788 }
 789 /*
 790  * Given the xsave area and a state inside, this function returns the
 791  * address of the state.
 792  *
 793  * This is the API that is called to get xstate address in either
 794  * standard format or compacted format of xsave area.
 795  *
 796  * Note that if there is no data for the field in the xsave buffer
 797  * this will return NULL.
 798  *
 799  * Inputs:
 800  *      xstate: the thread's storage area for all FPU data
 801  *      xstate_feature: state which is defined in xsave.h (e.g.
 802  *      XFEATURE_MASK_FP, XFEATURE_MASK_SSE, etc...)
 803  * Output:
 804  *      address of the state in the xsave area, or NULL if the
 805  *      field is not present in the xsave buffer.
 806  */
 807 void *get_xsave_addr(struct xregs_state *xsave, int xstate_feature)
 808 {
 809         /*
 810          * Do we even *have* xsave state?
 811          */
 812         if (!boot_cpu_has(X86_FEATURE_XSAVE))
 813                 return NULL;
 814
 815         /*
 816          * We should not ever be requesting features that we
 817          * have not enabled.  Remember that pcntxt_mask is
 818          * what we write to the XCR0 register.
 819          */
 820         WARN_ONCE(!(xfeatures_mask & xstate_feature),
 821                   "get of unsupported state");
 822         /*
 823          * This assumes the last 'xsave*' instruction to
 824          * have requested that 'xstate_feature' be saved.
 825          * If it did not, we might be seeing and old value
 826          * of the field in the buffer.
 827          *
 828          * This can happen because the last 'xsave' did not
 829          * request that this feature be saved (unlikely)
 830          * or because the "init optimization" caused it
 831          * to not be saved.
 832          */
 833         if (!(xsave->header.xfeatures & xstate_feature))
 834                 return NULL;
 835
 836         return __raw_xsave_addr(xsave, xstate_feature);
 837 }
 838 EXPORT_SYMBOL_GPL(get_xsave_addr);
 839
 840 /*
 841  * This wraps up the common operations that need to occur when retrieving
 842  * data from xsave state.  It first ensures that the current task was
 843  * using the FPU and retrieves the data in to a buffer.  It then calculates
 844  * the offset of the requested field in the buffer.
 845  *
 846  * This function is safe to call whether the FPU is in use or not.
 847  *
 848  * Note that this only works on the current task.
 849  *
 850  * Inputs:
 851  *      @xsave_state: state which is defined in xsave.h (e.g. XFEATURE_MASK_FP,
 852  *      XFEATURE_MASK_SSE, etc...)
 853  * Output:
 854  *      address of the state in the xsave area or NULL if the state
 855  *      is not present or is in its 'init state'.
 856  */
 857 const void *get_xsave_field_ptr(int xsave_state)
 858 {
 859         struct fpu *fpu = &current->thread.fpu;
 860
 861         if (!fpu->fpstate_active)
 862                 return NULL;
 863         /*
 864          * fpu__save() takes the CPU's xstate registers
 865          * and saves them off to the 'fpu memory buffer.
 866          */
 867         fpu__save(fpu);
 868
 869         return get_xsave_addr(&fpu->state.xsave, xsave_state);
 870 }
 871
 872 #ifdef CONFIG_ARCH_HAS_PKEYS
 873
 874 #define NR_VALID_PKRU_BITS (CONFIG_NR_PROTECTION_KEYS * 2)
 875 #define PKRU_VALID_MASK (NR_VALID_PKRU_BITS - 1)
 876 /*
 877  * This will go out and modify PKRU register to set the access
 878  * rights for @pkey to @init_val.
 879  */
 880 int arch_set_user_pkey_access(struct task_struct *tsk, int pkey,
 881                 unsigned long init_val)
 882 {
 883         u32 old_pkru;
 884         int pkey_shift = (pkey * PKRU_BITS_PER_PKEY);
 885         u32 new_pkru_bits = 0;
 886
 887         /*
 888          * This check implies XSAVE support.  OSPKE only gets
 889          * set if we enable XSAVE and we enable PKU in XCR0.
 890          */
 891         if (!boot_cpu_has(X86_FEATURE_OSPKE))
 892                 return -EINVAL;
 893         /*
 894          * For most XSAVE components, this would be an arduous task:
 895          * brining fpstate up to date with fpregs, updating fpstate,
 896          * then re-populating fpregs.  But, for components that are
 897          * never lazily managed, we can just access the fpregs
 898          * directly.  PKRU is never managed lazily, so we can just
 899          * manipulate it directly.  Make sure it stays that way.
 900          */
 901         WARN_ON_ONCE(!use_eager_fpu());
 902
 903         /* Set the bits we need in PKRU:  */
 904         if (init_val & PKEY_DISABLE_ACCESS)
 905                 new_pkru_bits |= PKRU_AD_BIT;
 906         if (init_val & PKEY_DISABLE_WRITE)
 907                 new_pkru_bits |= PKRU_WD_BIT;
 908
 909         /* Shift the bits in to the correct place in PKRU for pkey: */
 910         new_pkru_bits <<= pkey_shift;
 911
 912         /* Get old PKRU and mask off any old bits in place: */
 913         old_pkru = read_pkru();
 914         old_pkru &= ~((PKRU_AD_BIT|PKRU_WD_BIT) << pkey_shift);
 915
 916         /* Write old part along with new part: */
 917         write_pkru(old_pkru | new_pkru_bits);
 918
 919         return 0;
 920 }
 921 #endif /* ! CONFIG_ARCH_HAS_PKEYS */
 922
 923 /*
 924  * This is similar to user_regset_copyout(), but will not add offset to
 925  * the source data pointer or increment pos, count, kbuf, and ubuf.
 926  */
 927 static inline int xstate_copyout(unsigned int pos, unsigned int count,
 928                                  void *kbuf, void __user *ubuf,
 929                                  const void *data, const int start_pos,
 930                                  const int end_pos)
 931 {
 932         if ((count == 0) || (pos < start_pos))
 933                 return 0;
 934
 935         if (end_pos < 0 || pos < end_pos) {
 936                 unsigned int copy = (end_pos < 0 ? count : min(count, end_pos - pos));
 937
 938                 if (kbuf) {
 939                         memcpy(kbuf + pos, data, copy);
 940                 } else {
 941                         if (__copy_to_user(ubuf + pos, data, copy))
 942                                 return -EFAULT;
 943                 }
 944         }
 945         return 0;
 946 }
 947
 948 /*
 949  * Convert from kernel XSAVES compacted format to standard format and copy
 950  * to a ptrace buffer. It supports partial copy but pos always starts from
 951  * zero. This is called from xstateregs_get() and there we check the CPU
 952  * has XSAVES.
 953  */
 954 int copyout_from_xsaves(unsigned int pos, unsigned int count, void *kbuf,
 955                         void __user *ubuf, struct xregs_state *xsave)
 956 {
 957         unsigned int offset, size;
 958         int ret, i;
 959         struct xstate_header header;
 960
 961         /*
 962          * Currently copy_regset_to_user() starts from pos 0:
 963          */
 964         if (unlikely(pos != 0))
 965                 return -EFAULT;
 966
 967         /*
 968          * The destination is a ptrace buffer; we put in only user xstates:
 969          */
 970         memset(&header, 0, sizeof(header));
 971         header.xfeatures = xsave->header.xfeatures;
 972         header.xfeatures &= ~XFEATURE_MASK_SUPERVISOR;
 973
 974         /*
 975          * Copy xregs_state->header:
 976          */
 977         offset = offsetof(struct xregs_state, header);
 978         size = sizeof(header);
 979
 980         ret = xstate_copyout(offset, size, kbuf, ubuf, &header, 0, count);
 981
 982         if (ret)
 983                 return ret;
 984
 985         for (i = 0; i < XFEATURE_MAX; i++) {
 986                 /*
 987                  * Copy only in-use xstates:
 988                  */
 989                 if ((header.xfeatures >> i) & 1) {
 990                         void *src = __raw_xsave_addr(xsave, 1 << i);
 991
 992                         offset = xstate_offsets[i];
 993                         size = xstate_sizes[i];
 994
 995                         ret = xstate_copyout(offset, size, kbuf, ubuf, src, 0, count);
 996
 997                         if (ret)
 998                                 return ret;
 999
1000                         if (offset + size >= count)
1001                                 break;
1002                 }
1003
1004         }
1005
1006         /*
1007          * Fill xsave->i387.sw_reserved value for ptrace frame:
1008          */
1009         offset = offsetof(struct fxregs_state, sw_reserved);
1010         size = sizeof(xstate_fx_sw_bytes);
1011
1012         ret = xstate_copyout(offset, size, kbuf, ubuf, xstate_fx_sw_bytes, 0, count);
1013
1014         if (ret)
1015                 return ret;
1016
1017         return 0;
1018 }
1019
1020 /*
1021  * Convert from a ptrace standard-format buffer to kernel XSAVES format
1022  * and copy to the target thread. This is called from xstateregs_set() and
1023  * there we check the CPU has XSAVES and a whole standard-sized buffer
1024  * exists.
1025  */
1026 int copyin_to_xsaves(const void *kbuf, const void __user *ubuf,
1027                      struct xregs_state *xsave)
1028 {
1029         unsigned int offset, size;
1030         int i;
1031         u64 xfeatures;
1032         u64 allowed_features;
1033
1034         offset = offsetof(struct xregs_state, header);
1035         size = sizeof(xfeatures);
1036
1037         if (kbuf) {
1038                 memcpy(&xfeatures, kbuf + offset, size);
1039         } else {
1040                 if (__copy_from_user(&xfeatures, ubuf + offset, size))
1041                         return -EFAULT;
1042         }
1043
1044         /*
1045          * Reject if the user sets any disabled or supervisor features:
1046          */
1047         allowed_features = xfeatures_mask & ~XFEATURE_MASK_SUPERVISOR;
1048
1049         if (xfeatures & ~allowed_features)
1050                 return -EINVAL;
1051
1052         for (i = 0; i < XFEATURE_MAX; i++) {
1053                 u64 mask = ((u64)1 << i);
1054
1055                 if (xfeatures & mask) {
1056                         void *dst = __raw_xsave_addr(xsave, 1 << i);
1057
1058                         offset = xstate_offsets[i];
1059                         size = xstate_sizes[i];
1060
1061                         if (kbuf) {
1062                                 memcpy(dst, kbuf + offset, size);
1063                         } else {
1064                                 if (__copy_from_user(dst, ubuf + offset, size))
1065                                         return -EFAULT;
1066                         }
1067                 }
1068         }
1069
1070         /*
1071          * The state that came in from userspace was user-state only.
1072          * Mask all the user states out of 'xfeatures':
1073          */
1074         xsave->header.xfeatures &= XFEATURE_MASK_SUPERVISOR;
1075
1076         /*
1077          * Add back in the features that came in from userspace:
1078          */
1079         xsave->header.xfeatures |= xfeatures;
1080
1081         return 0;
1082 }