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_AVX512DQ);
  63         setup_clear_cpu_cap(X86_FEATURE_AVX512BW);
  64         setup_clear_cpu_cap(X86_FEATURE_AVX512VL);
  65         setup_clear_cpu_cap(X86_FEATURE_MPX);
  66         setup_clear_cpu_cap(X86_FEATURE_XGETBV1);
  67         setup_clear_cpu_cap(X86_FEATURE_PKU);
  68 }
  69
  70 /*
  71  * Return whether the system supports a given xfeature.
  72  *
  73  * Also return the name of the (most advanced) feature that the caller requested:
  74  */
  75 int cpu_has_xfeatures(u64 xfeatures_needed, const char **feature_name)
  76 {
  77         u64 xfeatures_missing = xfeatures_needed & ~xfeatures_mask;
  78
  79         if (unlikely(feature_name)) {
  80                 long xfeature_idx, max_idx;
  81                 u64 xfeatures_print;
  82                 /*
  83                  * So we use FLS here to be able to print the most advanced
  84                  * feature that was requested but is missing. So if a driver
  85                  * asks about "XFEATURE_MASK_SSE | XFEATURE_MASK_YMM" we'll print the
  86                  * missing AVX feature - this is the most informative message
  87                  * to users:
  88                  */
  89                 if (xfeatures_missing)
  90                         xfeatures_print = xfeatures_missing;
  91                 else
  92                         xfeatures_print = xfeatures_needed;
  93
  94                 xfeature_idx = fls64(xfeatures_print)-1;
  95                 max_idx = ARRAY_SIZE(xfeature_names)-1;
  96                 xfeature_idx = min(xfeature_idx, max_idx);
  97
  98                 *feature_name = xfeature_names[xfeature_idx];
  99         }
 100
 101         if (xfeatures_missing)
 102                 return 0;
 103
 104         return 1;
 105 }
 106 EXPORT_SYMBOL_GPL(cpu_has_xfeatures);
 107
 108 /*
 109  * When executing XSAVEOPT (or other optimized XSAVE instructions), if
 110  * a processor implementation detects that an FPU state component is still
 111  * (or is again) in its initialized state, it may clear the corresponding
 112  * bit in the header.xfeatures field, and can skip the writeout of registers
 113  * to the corresponding memory layout.
 114  *
 115  * This means that when the bit is zero, the state component might still contain
 116  * some previous - non-initialized register state.
 117  *
 118  * Before writing xstate information to user-space we sanitize those components,
 119  * to always ensure that the memory layout of a feature will be in the init state
 120  * if the corresponding header bit is zero. This is to ensure that user-space doesn't
 121  * see some stale state in the memory layout during signal handling, debugging etc.
 122  */
 123 void fpstate_sanitize_xstate(struct fpu *fpu)
 124 {
 125         struct fxregs_state *fx = &fpu->state.fxsave;
 126         int feature_bit;
 127         u64 xfeatures;
 128
 129         if (!use_xsaveopt())
 130                 return;
 131
 132         xfeatures = fpu->state.xsave.header.xfeatures;
 133
 134         /*
 135          * None of the feature bits are in init state. So nothing else
 136          * to do for us, as the memory layout is up to date.
 137          */
 138         if ((xfeatures & xfeatures_mask) == xfeatures_mask)
 139                 return;
 140
 141         /*
 142          * FP is in init state
 143          */
 144         if (!(xfeatures & XFEATURE_MASK_FP)) {
 145                 fx->cwd = 0x37f;
 146                 fx->swd = 0;
 147                 fx->twd = 0;
 148                 fx->fop = 0;
 149                 fx->rip = 0;
 150                 fx->rdp = 0;
 151                 memset(&fx->st_space[0], 0, 128);
 152         }
 153
 154         /*
 155          * SSE is in init state
 156          */
 157         if (!(xfeatures & XFEATURE_MASK_SSE))
 158                 memset(&fx->xmm_space[0], 0, 256);
 159
 160         /*
 161          * First two features are FPU and SSE, which above we handled
 162          * in a special way already:
 163          */
 164         feature_bit = 0x2;
 165         xfeatures = (xfeatures_mask & ~xfeatures) >> 2;
 166
 167         /*
 168          * Update all the remaining memory layouts according to their
 169          * standard xstate layout, if their header bit is in the init
 170          * state:
 171          */
 172         while (xfeatures) {
 173                 if (xfeatures & 0x1) {
 174                         int offset = xstate_offsets[feature_bit];
 175                         int size = xstate_sizes[feature_bit];
 176
 177                         memcpy((void *)fx + offset,
 178                                (void *)&init_fpstate.xsave + offset,
 179                                size);
 180                 }
 181
 182                 xfeatures >>= 1;
 183                 feature_bit++;
 184         }
 185 }
 186
 187 /*
 188  * Enable the extended processor state save/restore feature.
 189  * Called once per CPU onlining.
 190  */
 191 void fpu__init_cpu_xstate(void)
 192 {
 193         if (!boot_cpu_has(X86_FEATURE_XSAVE) || !xfeatures_mask)
 194                 return;
 195
 196         cr4_set_bits(X86_CR4_OSXSAVE);
 197         xsetbv(XCR_XFEATURE_ENABLED_MASK, xfeatures_mask);
 198 }
 199
 200 /*
 201  * Note that in the future we will likely need a pair of
 202  * functions here: one for user xstates and the other for
 203  * system xstates.  For now, they are the same.
 204  */
 205 static int xfeature_enabled(enum xfeature xfeature)
 206 {
 207         return !!(xfeatures_mask & (1UL << xfeature));
 208 }
 209
 210 /*
 211  * Record the offsets and sizes of various xstates contained
 212  * in the XSAVE state memory layout.
 213  */
 214 static void __init setup_xstate_features(void)
 215 {
 216         u32 eax, ebx, ecx, edx, i;
 217         /* start at the beginnning of the "extended state" */
 218         unsigned int last_good_offset = offsetof(struct xregs_state,
 219                                                  extended_state_area);
 220
 221         for (i = FIRST_EXTENDED_XFEATURE; i < XFEATURE_MAX; i++) {
 222                 if (!xfeature_enabled(i))
 223                         continue;
 224
 225                 cpuid_count(XSTATE_CPUID, i, &eax, &ebx, &ecx, &edx);
 226                 xstate_offsets[i] = ebx;
 227                 xstate_sizes[i] = eax;
 228                 /*
 229                  * In our xstate size checks, we assume that the
 230                  * highest-numbered xstate feature has the
 231                  * highest offset in the buffer.  Ensure it does.
 232                  */
 233                 WARN_ONCE(last_good_offset > xstate_offsets[i],
 234                         "x86/fpu: misordered xstate at %d\n", last_good_offset);
 235                 last_good_offset = xstate_offsets[i];
 236
 237                 printk(KERN_INFO "x86/fpu: xstate_offset[%d]: %4d, xstate_sizes[%d]: %4d\n", i, ebx, i, eax);
 238         }
 239 }
 240
 241 static void __init print_xstate_feature(u64 xstate_mask)
 242 {
 243         const char *feature_name;
 244
 245         if (cpu_has_xfeatures(xstate_mask, &feature_name))
 246                 pr_info("x86/fpu: Supporting XSAVE feature 0x%03Lx: '%s'\n", xstate_mask, feature_name);
 247 }
 248
 249 /*
 250  * Print out all the supported xstate features:
 251  */
 252 static void __init print_xstate_features(void)
 253 {
 254         print_xstate_feature(XFEATURE_MASK_FP);
 255         print_xstate_feature(XFEATURE_MASK_SSE);
 256         print_xstate_feature(XFEATURE_MASK_YMM);
 257         print_xstate_feature(XFEATURE_MASK_BNDREGS);
 258         print_xstate_feature(XFEATURE_MASK_BNDCSR);
 259         print_xstate_feature(XFEATURE_MASK_OPMASK);
 260         print_xstate_feature(XFEATURE_MASK_ZMM_Hi256);
 261         print_xstate_feature(XFEATURE_MASK_Hi16_ZMM);
 262         print_xstate_feature(XFEATURE_MASK_PKRU);
 263 }
 264
 265 /*
 266  * This function sets up offsets and sizes of all extended states in
 267  * xsave area. This supports both standard format and compacted format
 268  * of the xsave aread.
 269  */
 270 static void __init setup_xstate_comp(void)
 271 {
 272         unsigned int xstate_comp_sizes[sizeof(xfeatures_mask)*8];
 273         int i;
 274
 275         /*
 276          * The FP xstates and SSE xstates are legacy states. They are always
 277          * in the fixed offsets in the xsave area in either compacted form
 278          * or standard form.
 279          */
 280         xstate_comp_offsets[0] = 0;
 281         xstate_comp_offsets[1] = offsetof(struct fxregs_state, xmm_space);
 282
 283         if (!boot_cpu_has(X86_FEATURE_XSAVES)) {
 284                 for (i = FIRST_EXTENDED_XFEATURE; i < XFEATURE_MAX; i++) {
 285                         if (xfeature_enabled(i)) {
 286                                 xstate_comp_offsets[i] = xstate_offsets[i];
 287                                 xstate_comp_sizes[i] = xstate_sizes[i];
 288                         }
 289                 }
 290                 return;
 291         }
 292
 293         xstate_comp_offsets[FIRST_EXTENDED_XFEATURE] =
 294                 FXSAVE_SIZE + XSAVE_HDR_SIZE;
 295
 296         for (i = FIRST_EXTENDED_XFEATURE; i < XFEATURE_MAX; i++) {
 297                 if (xfeature_enabled(i))
 298                         xstate_comp_sizes[i] = xstate_sizes[i];
 299                 else
 300                         xstate_comp_sizes[i] = 0;
 301
 302                 if (i > FIRST_EXTENDED_XFEATURE)
 303                         xstate_comp_offsets[i] = xstate_comp_offsets[i-1]
 304                                         + xstate_comp_sizes[i-1];
 305
 306         }
 307 }
 308
 309 /*
 310  * setup the xstate image representing the init state
 311  */
 312 static void __init setup_init_fpu_buf(void)
 313 {
 314         static int on_boot_cpu __initdata = 1;
 315
 316         WARN_ON_FPU(!on_boot_cpu);
 317         on_boot_cpu = 0;
 318
 319         if (!boot_cpu_has(X86_FEATURE_XSAVE))
 320                 return;
 321
 322         setup_xstate_features();
 323         print_xstate_features();
 324
 325         if (boot_cpu_has(X86_FEATURE_XSAVES)) {
 326                 init_fpstate.xsave.header.xcomp_bv = (u64)1 << 63 | xfeatures_mask;
 327                 init_fpstate.xsave.header.xfeatures = xfeatures_mask;
 328         }
 329
 330         /*
 331          * Init all the features state with header_bv being 0x0
 332          */
 333         copy_kernel_to_xregs_booting(&init_fpstate.xsave);
 334
 335         /*
 336          * Dump the init state again. This is to identify the init state
 337          * of any feature which is not represented by all zero's.
 338          */
 339         copy_xregs_to_kernel_booting(&init_fpstate.xsave);
 340 }
 341
 342 static int xfeature_is_supervisor(int xfeature_nr)
 343 {
 344         /*
 345          * We currently do not support supervisor states, but if
 346          * we did, we could find out like this.
 347          *
 348          * SDM says: If state component i is a user state component,
 349          * ECX[0] return 0; if state component i is a supervisor
 350          * state component, ECX[0] returns 1.
 351         u32 eax, ebx, ecx, edx;
 352         cpuid_count(XSTATE_CPUID, xfeature_nr, &eax, &ebx, &ecx, &edx;
 353         return !!(ecx & 1);
 354         */
 355         return 0;
 356 }
 357 /*
 358 static int xfeature_is_user(int xfeature_nr)
 359 {
 360         return !xfeature_is_supervisor(xfeature_nr);
 361 }
 362 */
 363
 364 /*
 365  * This check is important because it is easy to get XSTATE_*
 366  * confused with XSTATE_BIT_*.
 367  */
 368 #define CHECK_XFEATURE(nr) do {         \
 369         WARN_ON(nr < FIRST_EXTENDED_XFEATURE);  \
 370         WARN_ON(nr >= XFEATURE_MAX);    \
 371 } while (0)
 372
 373 /*
 374  * We could cache this like xstate_size[], but we only use
 375  * it here, so it would be a waste of space.
 376  */
 377 static int xfeature_is_aligned(int xfeature_nr)
 378 {
 379         u32 eax, ebx, ecx, edx;
 380
 381         CHECK_XFEATURE(xfeature_nr);
 382         cpuid_count(XSTATE_CPUID, xfeature_nr, &eax, &ebx, &ecx, &edx);
 383         /*
 384          * The value returned by ECX[1] indicates the alignment
 385          * of state component i when the compacted format
 386          * of the extended region of an XSAVE area is used
 387          */
 388         return !!(ecx & 2);
 389 }
 390
 391 static int xfeature_uncompacted_offset(int xfeature_nr)
 392 {
 393         u32 eax, ebx, ecx, edx;
 394
 395         CHECK_XFEATURE(xfeature_nr);
 396         cpuid_count(XSTATE_CPUID, xfeature_nr, &eax, &ebx, &ecx, &edx);
 397         return ebx;
 398 }
 399
 400 static int xfeature_size(int xfeature_nr)
 401 {
 402         u32 eax, ebx, ecx, edx;
 403
 404         CHECK_XFEATURE(xfeature_nr);
 405         cpuid_count(XSTATE_CPUID, xfeature_nr, &eax, &ebx, &ecx, &edx);
 406         return eax;
 407 }
 408
 409 /*
 410  * 'XSAVES' implies two different things:
 411  * 1. saving of supervisor/system state
 412  * 2. using the compacted format
 413  *
 414  * Use this function when dealing with the compacted format so
 415  * that it is obvious which aspect of 'XSAVES' is being handled
 416  * by the calling code.
 417  */
 418 static int using_compacted_format(void)
 419 {
 420         return boot_cpu_has(X86_FEATURE_XSAVES);
 421 }
 422
 423 static void __xstate_dump_leaves(void)
 424 {
 425         int i;
 426         u32 eax, ebx, ecx, edx;
 427         static int should_dump = 1;
 428
 429         if (!should_dump)
 430                 return;
 431         should_dump = 0;
 432         /*
 433          * Dump out a few leaves past the ones that we support
 434          * just in case there are some goodies up there
 435          */
 436         for (i = 0; i < XFEATURE_MAX + 10; i++) {
 437                 cpuid_count(XSTATE_CPUID, i, &eax, &ebx, &ecx, &edx);
 438                 pr_warn("CPUID[%02x, %02x]: eax=%08x ebx=%08x ecx=%08x edx=%08x\n",
 439                         XSTATE_CPUID, i, eax, ebx, ecx, edx);
 440         }
 441 }
 442
 443 #define XSTATE_WARN_ON(x) do {                                                  \
 444         if (WARN_ONCE(x, "XSAVE consistency problem, dumping leaves")) {        \
 445                 __xstate_dump_leaves();                                         \
 446         }                                                                       \
 447 } while (0)
 448
 449 #define XCHECK_SZ(sz, nr, nr_macro, __struct) do {                      \
 450         if ((nr == nr_macro) &&                                         \
 451             WARN_ONCE(sz != sizeof(__struct),                           \
 452                 "%s: struct is %zu bytes, cpu state %d bytes\n",        \
 453                 __stringify(nr_macro), sizeof(__struct), sz)) {         \
 454                 __xstate_dump_leaves();                                 \
 455         }                                                               \
 456 } while (0)
 457
 458 /*
 459  * We have a C struct for each 'xstate'.  We need to ensure
 460  * that our software representation matches what the CPU
 461  * tells us about the state's size.
 462  */
 463 static void check_xstate_against_struct(int nr)
 464 {
 465         /*
 466          * Ask the CPU for the size of the state.
 467          */
 468         int sz = xfeature_size(nr);
 469         /*
 470          * Match each CPU state with the corresponding software
 471          * structure.
 472          */
 473         XCHECK_SZ(sz, nr, XFEATURE_YMM,       struct ymmh_struct);
 474         XCHECK_SZ(sz, nr, XFEATURE_BNDREGS,   struct mpx_bndreg_state);
 475         XCHECK_SZ(sz, nr, XFEATURE_BNDCSR,    struct mpx_bndcsr_state);
 476         XCHECK_SZ(sz, nr, XFEATURE_OPMASK,    struct avx_512_opmask_state);
 477         XCHECK_SZ(sz, nr, XFEATURE_ZMM_Hi256, struct avx_512_zmm_uppers_state);
 478         XCHECK_SZ(sz, nr, XFEATURE_Hi16_ZMM,  struct avx_512_hi16_state);
 479         XCHECK_SZ(sz, nr, XFEATURE_PKRU,      struct pkru_state);
 480
 481         /*
 482          * Make *SURE* to add any feature numbers in below if
 483          * there are "holes" in the xsave state component
 484          * numbers.
 485          */
 486         if ((nr < XFEATURE_YMM) ||
 487             (nr >= XFEATURE_MAX) ||
 488             (nr == XFEATURE_PT_UNIMPLEMENTED_SO_FAR)) {
 489                 WARN_ONCE(1, "no structure for xstate: %d\n", nr);
 490                 XSTATE_WARN_ON(1);
 491         }
 492 }
 493
 494 /*
 495  * This essentially double-checks what the cpu told us about
 496  * how large the XSAVE buffer needs to be.  We are recalculating
 497  * it to be safe.
 498  */
 499 static void do_extra_xstate_size_checks(void)
 500 {
 501         int paranoid_xstate_size = FXSAVE_SIZE + XSAVE_HDR_SIZE;
 502         int i;
 503
 504         for (i = FIRST_EXTENDED_XFEATURE; i < XFEATURE_MAX; i++) {
 505                 if (!xfeature_enabled(i))
 506                         continue;
 507
 508                 check_xstate_against_struct(i);
 509                 /*
 510                  * Supervisor state components can be managed only by
 511                  * XSAVES, which is compacted-format only.
 512                  */
 513                 if (!using_compacted_format())
 514                         XSTATE_WARN_ON(xfeature_is_supervisor(i));
 515
 516                 /* Align from the end of the previous feature */
 517                 if (xfeature_is_aligned(i))
 518                         paranoid_xstate_size = ALIGN(paranoid_xstate_size, 64);
 519                 /*
 520                  * The offset of a given state in the non-compacted
 521                  * format is given to us in a CPUID leaf.  We check
 522                  * them for being ordered (increasing offsets) in
 523                  * setup_xstate_features().
 524                  */
 525                 if (!using_compacted_format())
 526                         paranoid_xstate_size = xfeature_uncompacted_offset(i);
 527                 /*
 528                  * The compacted-format offset always depends on where
 529                  * the previous state ended.
 530                  */
 531                 paranoid_xstate_size += xfeature_size(i);
 532         }
 533         XSTATE_WARN_ON(paranoid_xstate_size != xstate_size);
 534 }
 535
 536 /*
 537  * Calculate total size of enabled xstates in XCR0/xfeatures_mask.
 538  *
 539  * Note the SDM's wording here.  "sub-function 0" only enumerates
 540  * the size of the *user* states.  If we use it to size a buffer
 541  * that we use 'XSAVES' on, we could potentially overflow the
 542  * buffer because 'XSAVES' saves system states too.
 543  *
 544  * Note that we do not currently set any bits on IA32_XSS so
 545  * 'XCR0 | IA32_XSS == XCR0' for now.
 546  */
 547 static unsigned int __init calculate_xstate_size(void)
 548 {
 549         unsigned int eax, ebx, ecx, edx;
 550         unsigned int calculated_xstate_size;
 551
 552         if (!boot_cpu_has(X86_FEATURE_XSAVES)) {
 553                 /*
 554                  * - CPUID function 0DH, sub-function 0:
 555                  *    EBX enumerates the size (in bytes) required by
 556                  *    the XSAVE instruction for an XSAVE area
 557                  *    containing all the *user* state components
 558                  *    corresponding to bits currently set in XCR0.
 559                  */
 560                 cpuid_count(XSTATE_CPUID, 0, &eax, &ebx, &ecx, &edx);
 561                 calculated_xstate_size = ebx;
 562         } else {
 563                 /*
 564                  * - CPUID function 0DH, sub-function 1:
 565                  *    EBX enumerates the size (in bytes) required by
 566                  *    the XSAVES instruction for an XSAVE area
 567                  *    containing all the state components
 568                  *    corresponding to bits currently set in
 569                  *    XCR0 | IA32_XSS.
 570                  */
 571                 cpuid_count(XSTATE_CPUID, 1, &eax, &ebx, &ecx, &edx);
 572                 calculated_xstate_size = ebx;
 573         }
 574         return calculated_xstate_size;
 575 }
 576
 577 /*
 578  * Will the runtime-enumerated 'xstate_size' fit in the init
 579  * task's statically-allocated buffer?
 580  */
 581 static bool is_supported_xstate_size(unsigned int test_xstate_size)
 582 {
 583         if (test_xstate_size <= sizeof(union fpregs_state))
 584                 return true;
 585
 586         pr_warn("x86/fpu: xstate buffer too small (%zu < %d), disabling xsave\n",
 587                         sizeof(union fpregs_state), test_xstate_size);
 588         return false;
 589 }
 590
 591 static int init_xstate_size(void)
 592 {
 593         /* Recompute the context size for enabled features: */
 594         unsigned int possible_xstate_size = calculate_xstate_size();
 595
 596         /* Ensure we have the space to store all enabled: */
 597         if (!is_supported_xstate_size(possible_xstate_size))
 598                 return -EINVAL;
 599
 600         /*
 601          * The size is OK, we are definitely going to use xsave,
 602          * make it known to the world that we need more space.
 603          */
 604         xstate_size = possible_xstate_size;
 605         do_extra_xstate_size_checks();
 606         return 0;
 607 }
 608
 609 /*
 610  * We enabled the XSAVE hardware, but something went wrong and
 611  * we can not use it.  Disable it.
 612  */
 613 static void fpu__init_disable_system_xstate(void)
 614 {
 615         xfeatures_mask = 0;
 616         cr4_clear_bits(X86_CR4_OSXSAVE);
 617         fpu__xstate_clear_all_cpu_caps();
 618 }
 619
 620 /*
 621  * Enable and initialize the xsave feature.
 622  * Called once per system bootup.
 623  */
 624 void __init fpu__init_system_xstate(void)
 625 {
 626         unsigned int eax, ebx, ecx, edx;
 627         static int on_boot_cpu __initdata = 1;
 628         int err;
 629
 630         WARN_ON_FPU(!on_boot_cpu);
 631         on_boot_cpu = 0;
 632
 633         if (!boot_cpu_has(X86_FEATURE_XSAVE)) {
 634                 pr_info("x86/fpu: Legacy x87 FPU detected.\n");
 635                 return;
 636         }
 637
 638         if (boot_cpu_data.cpuid_level < XSTATE_CPUID) {
 639                 WARN_ON_FPU(1);
 640                 return;
 641         }
 642
 643         cpuid_count(XSTATE_CPUID, 0, &eax, &ebx, &ecx, &edx);
 644         xfeatures_mask = eax + ((u64)edx << 32);
 645
 646         if ((xfeatures_mask & XFEATURE_MASK_FPSSE) != XFEATURE_MASK_FPSSE) {
 647                 pr_err("x86/fpu: FP/SSE not present amongst the CPU's xstate features: 0x%llx.\n", xfeatures_mask);
 648                 BUG();
 649         }
 650
 651         xfeatures_mask &= fpu__get_supported_xfeatures_mask();
 652
 653         /* Enable xstate instructions to be able to continue with initialization: */
 654         fpu__init_cpu_xstate();
 655         err = init_xstate_size();
 656         if (err) {
 657                 /* something went wrong, boot without any XSAVE support */
 658                 fpu__init_disable_system_xstate();
 659                 return;
 660         }
 661
 662         update_regset_xstate_info(xstate_size, xfeatures_mask);
 663         fpu__init_prepare_fx_sw_frame();
 664         setup_init_fpu_buf();
 665         setup_xstate_comp();
 666
 667         pr_info("x86/fpu: Enabled xstate features 0x%llx, context size is %d bytes, using '%s' format.\n",
 668                 xfeatures_mask,
 669                 xstate_size,
 670                 boot_cpu_has(X86_FEATURE_XSAVES) ? "compacted" : "standard");
 671 }
 672
 673 /*
 674  * Restore minimal FPU state after suspend:
 675  */
 676 void fpu__resume_cpu(void)
 677 {
 678         /*
 679          * Restore XCR0 on xsave capable CPUs:
 680          */
 681         if (boot_cpu_has(X86_FEATURE_XSAVE))
 682                 xsetbv(XCR_XFEATURE_ENABLED_MASK, xfeatures_mask);
 683 }
 684
 685 /*
 686  * Given an xstate feature mask, calculate where in the xsave
 687  * buffer the state is.  Callers should ensure that the buffer
 688  * is valid.
 689  *
 690  * Note: does not work for compacted buffers.
 691  */
 692 void *__raw_xsave_addr(struct xregs_state *xsave, int xstate_feature_mask)
 693 {
 694         int feature_nr = fls64(xstate_feature_mask) - 1;
 695
 696         return (void *)xsave + xstate_comp_offsets[feature_nr];
 697 }
 698 /*
 699  * Given the xsave area and a state inside, this function returns the
 700  * address of the state.
 701  *
 702  * This is the API that is called to get xstate address in either
 703  * standard format or compacted format of xsave area.
 704  *
 705  * Note that if there is no data for the field in the xsave buffer
 706  * this will return NULL.
 707  *
 708  * Inputs:
 709  *      xstate: the thread's storage area for all FPU data
 710  *      xstate_feature: state which is defined in xsave.h (e.g.
 711  *      XFEATURE_MASK_FP, XFEATURE_MASK_SSE, etc...)
 712  * Output:
 713  *      address of the state in the xsave area, or NULL if the
 714  *      field is not present in the xsave buffer.
 715  */
 716 void *get_xsave_addr(struct xregs_state *xsave, int xstate_feature)
 717 {
 718         /*
 719          * Do we even *have* xsave state?
 720          */
 721         if (!boot_cpu_has(X86_FEATURE_XSAVE))
 722                 return NULL;
 723
 724         /*
 725          * We should not ever be requesting features that we
 726          * have not enabled.  Remember that pcntxt_mask is
 727          * what we write to the XCR0 register.
 728          */
 729         WARN_ONCE(!(xfeatures_mask & xstate_feature),
 730                   "get of unsupported state");
 731         /*
 732          * This assumes the last 'xsave*' instruction to
 733          * have requested that 'xstate_feature' be saved.
 734          * If it did not, we might be seeing and old value
 735          * of the field in the buffer.
 736          *
 737          * This can happen because the last 'xsave' did not
 738          * request that this feature be saved (unlikely)
 739          * or because the "init optimization" caused it
 740          * to not be saved.
 741          */
 742         if (!(xsave->header.xfeatures & xstate_feature))
 743                 return NULL;
 744
 745         return __raw_xsave_addr(xsave, xstate_feature);
 746 }
 747 EXPORT_SYMBOL_GPL(get_xsave_addr);
 748
 749 /*
 750  * This wraps up the common operations that need to occur when retrieving
 751  * data from xsave state.  It first ensures that the current task was
 752  * using the FPU and retrieves the data in to a buffer.  It then calculates
 753  * the offset of the requested field in the buffer.
 754  *
 755  * This function is safe to call whether the FPU is in use or not.
 756  *
 757  * Note that this only works on the current task.
 758  *
 759  * Inputs:
 760  *      @xsave_state: state which is defined in xsave.h (e.g. XFEATURE_MASK_FP,
 761  *      XFEATURE_MASK_SSE, etc...)
 762  * Output:
 763  *      address of the state in the xsave area or NULL if the state
 764  *      is not present or is in its 'init state'.
 765  */
 766 const void *get_xsave_field_ptr(int xsave_state)
 767 {
 768         struct fpu *fpu = &current->thread.fpu;
 769
 770         if (!fpu->fpstate_active)
 771                 return NULL;
 772         /*
 773          * fpu__save() takes the CPU's xstate registers
 774          * and saves them off to the 'fpu memory buffer.
 775          */
 776         fpu__save(fpu);
 777
 778         return get_xsave_addr(&fpu->state.xsave, xsave_state);
 779 }
 780
 781
 782 /*
 783  * Set xfeatures (aka XSTATE_BV) bit for a feature that we want
 784  * to take out of its "init state".  This will ensure that an
 785  * XRSTOR actually restores the state.
 786  */
 787 static void fpu__xfeature_set_non_init(struct xregs_state *xsave,
 788                 int xstate_feature_mask)
 789 {
 790         xsave->header.xfeatures |= xstate_feature_mask;
 791 }
 792
 793 /*
 794  * This function is safe to call whether the FPU is in use or not.
 795  *
 796  * Note that this only works on the current task.
 797  *
 798  * Inputs:
 799  *      @xsave_state: state which is defined in xsave.h (e.g. XFEATURE_MASK_FP,
 800  *      XFEATURE_MASK_SSE, etc...)
 801  *      @xsave_state_ptr: a pointer to a copy of the state that you would
 802  *      like written in to the current task's FPU xsave state.  This pointer
 803  *      must not be located in the current tasks's xsave area.
 804  * Output:
 805  *      address of the state in the xsave area or NULL if the state
 806  *      is not present or is in its 'init state'.
 807  */
 808 static void fpu__xfeature_set_state(int xstate_feature_mask,
 809                 void *xstate_feature_src, size_t len)
 810 {
 811         struct xregs_state *xsave = &current->thread.fpu.state.xsave;
 812         struct fpu *fpu = &current->thread.fpu;
 813         void *dst;
 814
 815         if (!boot_cpu_has(X86_FEATURE_XSAVE)) {
 816                 WARN_ONCE(1, "%s() attempted with no xsave support", __func__);
 817                 return;
 818         }
 819
 820         /*
 821          * Tell the FPU code that we need the FPU state to be in
 822          * 'fpu' (not in the registers), and that we need it to
 823          * be stable while we write to it.
 824          */
 825         fpu__current_fpstate_write_begin();
 826
 827         /*
 828          * This method *WILL* *NOT* work for compact-format
 829          * buffers.  If the 'xstate_feature_mask' is unset in
 830          * xcomp_bv then we may need to move other feature state
 831          * "up" in the buffer.
 832          */
 833         if (xsave->header.xcomp_bv & xstate_feature_mask) {
 834                 WARN_ON_ONCE(1);
 835                 goto out;
 836         }
 837
 838         /* find the location in the xsave buffer of the desired state */
 839         dst = __raw_xsave_addr(&fpu->state.xsave, xstate_feature_mask);
 840
 841         /*
 842          * Make sure that the pointer being passed in did not
 843          * come from the xsave buffer itself.
 844          */
 845         WARN_ONCE(xstate_feature_src == dst, "set from xsave buffer itself");
 846
 847         /* put the caller-provided data in the location */
 848         memcpy(dst, xstate_feature_src, len);
 849
 850         /*
 851          * Mark the xfeature so that the CPU knows there is state
 852          * in the buffer now.
 853          */
 854         fpu__xfeature_set_non_init(xsave, xstate_feature_mask);
 855 out:
 856         /*
 857          * We are done writing to the 'fpu'.  Reenable preeption
 858          * and (possibly) move the fpstate back in to the fpregs.
 859          */
 860         fpu__current_fpstate_write_end();
 861 }
 862
 863 #define NR_VALID_PKRU_BITS (CONFIG_NR_PROTECTION_KEYS * 2)
 864 #define PKRU_VALID_MASK (NR_VALID_PKRU_BITS - 1)
 865
 866 /*
 867  * This will go out and modify the XSAVE buffer so that PKRU is
 868  * set to a particular state for access to 'pkey'.
 869  *
 870  * PKRU state does affect kernel access to user memory.  We do
 871  * not modfiy PKRU *itself* here, only the XSAVE state that will
 872  * be restored in to PKRU when we return back to userspace.
 873  */
 874 int arch_set_user_pkey_access(struct task_struct *tsk, int pkey,
 875                 unsigned long init_val)
 876 {
 877         struct xregs_state *xsave = &tsk->thread.fpu.state.xsave;
 878         struct pkru_state *old_pkru_state;
 879         struct pkru_state new_pkru_state;
 880         int pkey_shift = (pkey * PKRU_BITS_PER_PKEY);
 881         u32 new_pkru_bits = 0;
 882
 883         /*
 884          * This check implies XSAVE support.  OSPKE only gets
 885          * set if we enable XSAVE and we enable PKU in XCR0.
 886          */
 887         if (!boot_cpu_has(X86_FEATURE_OSPKE))
 888                 return -EINVAL;
 889
 890         /* Set the bits we need in PKRU  */
 891         if (init_val & PKEY_DISABLE_ACCESS)
 892                 new_pkru_bits |= PKRU_AD_BIT;
 893         if (init_val & PKEY_DISABLE_WRITE)
 894                 new_pkru_bits |= PKRU_WD_BIT;
 895
 896         /* Shift the bits in to the correct place in PKRU for pkey. */
 897         new_pkru_bits <<= pkey_shift;
 898
 899         /* Locate old copy of the state in the xsave buffer */
 900         old_pkru_state = get_xsave_addr(xsave, XFEATURE_MASK_PKRU);
 901
 902         /*
 903          * When state is not in the buffer, it is in the init
 904          * state, set it manually.  Otherwise, copy out the old
 905          * state.
 906          */
 907         if (!old_pkru_state)
 908                 new_pkru_state.pkru = 0;
 909         else
 910                 new_pkru_state.pkru = old_pkru_state->pkru;
 911
 912         /* mask off any old bits in place */
 913         new_pkru_state.pkru &= ~((PKRU_AD_BIT|PKRU_WD_BIT) << pkey_shift);
 914         /* Set the newly-requested bits */
 915         new_pkru_state.pkru |= new_pkru_bits;
 916
 917         /*
 918          * We could theoretically live without zeroing pkru.pad.
 919          * The current XSAVE feature state definition says that
 920          * only bytes 0->3 are used.  But we do not want to
 921          * chance leaking kernel stack out to userspace in case a
 922          * memcpy() of the whole xsave buffer was done.
 923          *
 924          * They're in the same cacheline anyway.
 925          */
 926         new_pkru_state.pad = 0;
 927
 928         fpu__xfeature_set_state(XFEATURE_MASK_PKRU, &new_pkru_state,
 929                         sizeof(new_pkru_state));
 930
 931         return 0;
 932 }