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