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