arch/x86/kernel/fpu/xstate.c

   1 /*
   2  * xsave/xrstor support.
   3  *
   4  * Author: Suresh Siddha <suresh.b.siddha@intel.com>
   5  */
   6 #include <linux/compat.h>
   7 #include <linux/cpu.h>
   8 #include <linux/mman.h>
   9 #include <linux/pkeys.h>
  10
  11 #include <asm/fpu/api.h>
  12 #include <asm/fpu/internal.h>
  13 #include <asm/fpu/signal.h>
  14 #include <asm/fpu/regset.h>
  15 #include <asm/fpu/xstate.h>
  16
  17 #include <asm/tlbflush.h>
  18
  19 /*
  20  * Although we spell it out in here, the Processor Trace
  21  * xfeature is completely unused.  We use other mechanisms
  22  * to save/restore PT state in Linux.
  23  */
  24 static const char *xfeature_names[] =
  25 {
  26         "x87 floating point registers"  ,
  27         "SSE registers"                 ,
  28         "AVX registers"                 ,
  29         "MPX bounds registers"          ,
  30         "MPX CSR"                       ,
  31         "AVX-512 opmask"                ,
  32         "AVX-512 Hi256"                 ,
  33         "AVX-512 ZMM_Hi256"             ,
  34         "Processor Trace (unused)"      ,
  35         "Protection Keys User registers",
  36         "unknown xstate feature"        ,
  37 };
  38
  39 /*
  40  * Mask of xstate features supported by the CPU and the kernel:
  41  */
  42 u64 xfeatures_mask __read_mostly;
  43
  44 static unsigned int xstate_offsets[XFEATURE_MAX] = { [ 0 ... XFEATURE_MAX - 1] = -1};
  45 static unsigned int xstate_sizes[XFEATURE_MAX]   = { [ 0 ... XFEATURE_MAX - 1] = -1};
  46 static unsigned int xstate_comp_offsets[sizeof(xfeatures_mask)*8];
  47
  48 /*
  49  * The XSAVE area of kernel can be in standard or compacted format;
  50  * it is always in standard format for user mode. This is the user
  51  * mode standard format size used for signal and ptrace frames.
  52  */
  53 unsigned int fpu_user_xstate_size;
  54
  55 /*
  56  * Clear all of the X86_FEATURE_* bits that are unavailable
  57  * when the CPU has no XSAVE support.
  58  */
  59 void fpu__xstate_clear_all_cpu_caps(void)
  60 {
  61         setup_clear_cpu_cap(X86_FEATURE_XSAVE);
  62         setup_clear_cpu_cap(X86_FEATURE_XSAVEOPT);
  63         setup_clear_cpu_cap(X86_FEATURE_XSAVEC);
  64         setup_clear_cpu_cap(X86_FEATURE_XSAVES);
  65         setup_clear_cpu_cap(X86_FEATURE_AVX);
  66         setup_clear_cpu_cap(X86_FEATURE_AVX2);
  67         setup_clear_cpu_cap(X86_FEATURE_AVX512F);
  68         setup_clear_cpu_cap(X86_FEATURE_AVX512IFMA);
  69         setup_clear_cpu_cap(X86_FEATURE_AVX512PF);
  70         setup_clear_cpu_cap(X86_FEATURE_AVX512ER);
  71         setup_clear_cpu_cap(X86_FEATURE_AVX512CD);
  72         setup_clear_cpu_cap(X86_FEATURE_AVX512DQ);
  73         setup_clear_cpu_cap(X86_FEATURE_AVX512BW);
  74         setup_clear_cpu_cap(X86_FEATURE_AVX512VL);
  75         setup_clear_cpu_cap(X86_FEATURE_MPX);
  76         setup_clear_cpu_cap(X86_FEATURE_XGETBV1);
  77         setup_clear_cpu_cap(X86_FEATURE_AVX512VBMI);
  78         setup_clear_cpu_cap(X86_FEATURE_PKU);
  79         setup_clear_cpu_cap(X86_FEATURE_AVX512_4VNNIW);
  80         setup_clear_cpu_cap(X86_FEATURE_AVX512_4FMAPS);
  81         setup_clear_cpu_cap(X86_FEATURE_AVX512_VPOPCNTDQ);
  82 }
  83
  84 /*
  85  * Return whether the system supports a given xfeature.
  86  *
  87  * Also return the name of the (most advanced) feature that the caller requested:
  88  */
  89 int cpu_has_xfeatures(u64 xfeatures_needed, const char **feature_name)
  90 {
  91         u64 xfeatures_missing = xfeatures_needed & ~xfeatures_mask;
  92
  93         if (unlikely(feature_name)) {
  94                 long xfeature_idx, max_idx;
  95                 u64 xfeatures_print;
  96                 /*
  97                  * So we use FLS here to be able to print the most advanced
  98                  * feature that was requested but is missing. So if a driver
  99                  * asks about "XFEATURE_MASK_SSE | XFEATURE_MASK_YMM" we'll print the
 100                  * missing AVX feature - this is the most informative message
 101                  * to users:
 102                  */
 103                 if (xfeatures_missing)
 104                         xfeatures_print = xfeatures_missing;
 105                 else
 106                         xfeatures_print = xfeatures_needed;
 107
 108                 xfeature_idx = fls64(xfeatures_print)-1;
 109                 max_idx = ARRAY_SIZE(xfeature_names)-1;
 110                 xfeature_idx = min(xfeature_idx, max_idx);
 111
 112                 *feature_name = xfeature_names[xfeature_idx];
 113         }
 114
 115         if (xfeatures_missing)
 116                 return 0;
 117
 118         return 1;
 119 }
 120 EXPORT_SYMBOL_GPL(cpu_has_xfeatures);
 121
 122 static int xfeature_is_supervisor(int xfeature_nr)
 123 {
 124         /*
 125          * We currently do not support supervisor states, but if
 126          * we did, we could find out like this.
 127          *
 128          * SDM says: If state component 'i' is a user state component,
 129          * ECX[0] return 0; if state component i is a supervisor
 130          * state component, ECX[0] returns 1.
 131          */
 132         u32 eax, ebx, ecx, edx;
 133
 134         cpuid_count(XSTATE_CPUID, xfeature_nr, &eax, &ebx, &ecx, &edx);
 135         return !!(ecx & 1);
 136 }
 137
 138 static int xfeature_is_user(int xfeature_nr)
 139 {
 140         return !xfeature_is_supervisor(xfeature_nr);
 141 }
 142
 143 /*
 144  * When executing XSAVEOPT (or other optimized XSAVE instructions), if
 145  * a processor implementation detects that an FPU state component is still
 146  * (or is again) in its initialized state, it may clear the corresponding
 147  * bit in the header.xfeatures field, and can skip the writeout of registers
 148  * to the corresponding memory layout.
 149  *
 150  * This means that when the bit is zero, the state component might still contain
 151  * some previous - non-initialized register state.
 152  *
 153  * Before writing xstate information to user-space we sanitize those components,
 154  * to always ensure that the memory layout of a feature will be in the init state
 155  * if the corresponding header bit is zero. This is to ensure that user-space doesn't
 156  * see some stale state in the memory layout during signal handling, debugging etc.
 157  */
 158 void fpstate_sanitize_xstate(struct fpu *fpu)
 159 {
 160         struct fxregs_state *fx = &fpu->state.fxsave;
 161         int feature_bit;
 162         u64 xfeatures;
 163
 164         if (!use_xsaveopt())
 165                 return;
 166
 167         xfeatures = fpu->state.xsave.header.xfeatures;
 168
 169         /*
 170          * None of the feature bits are in init state. So nothing else
 171          * to do for us, as the memory layout is up to date.
 172          */
 173         if ((xfeatures & xfeatures_mask) == xfeatures_mask)
 174                 return;
 175
 176         /*
 177          * FP is in init state
 178          */
 179         if (!(xfeatures & XFEATURE_MASK_FP)) {
 180                 fx->cwd = 0x37f;
 181                 fx->swd = 0;
 182                 fx->twd = 0;
 183                 fx->fop = 0;
 184                 fx->rip = 0;
 185                 fx->rdp = 0;
 186                 memset(&fx->st_space[0], 0, 128);
 187         }
 188
 189         /*
 190          * SSE is in init state
 191          */
 192         if (!(xfeatures & XFEATURE_MASK_SSE))
 193                 memset(&fx->xmm_space[0], 0, 256);
 194
 195         /*
 196          * First two features are FPU and SSE, which above we handled
 197          * in a special way already:
 198          */
 199         feature_bit = 0x2;
 200         xfeatures = (xfeatures_mask & ~xfeatures) >> 2;
 201
 202         /*
 203          * Update all the remaining memory layouts according to their
 204          * standard xstate layout, if their header bit is in the init
 205          * state:
 206          */
 207         while (xfeatures) {
 208                 if (xfeatures & 0x1) {
 209                         int offset = xstate_comp_offsets[feature_bit];
 210                         int size = xstate_sizes[feature_bit];
 211
 212                         memcpy((void *)fx + offset,
 213                                (void *)&init_fpstate.xsave + offset,
 214                                size);
 215                 }
 216
 217                 xfeatures >>= 1;
 218                 feature_bit++;
 219         }
 220 }
 221
 222 /*
 223  * Enable the extended processor state save/restore feature.
 224  * Called once per CPU onlining.
 225  */
 226 void fpu__init_cpu_xstate(void)
 227 {
 228         if (!boot_cpu_has(X86_FEATURE_XSAVE) || !xfeatures_mask)
 229                 return;
 230         /*
 231          * Make it clear that XSAVES supervisor states are not yet
 232          * implemented should anyone expect it to work by changing
 233          * bits in XFEATURE_MASK_* macros and XCR0.
 234          */
 235         WARN_ONCE((xfeatures_mask & XFEATURE_MASK_SUPERVISOR),
 236                 "x86/fpu: XSAVES supervisor states are not yet implemented.\n");
 237
 238         xfeatures_mask &= ~XFEATURE_MASK_SUPERVISOR;
 239
 240         cr4_set_bits(X86_CR4_OSXSAVE);
 241         xsetbv(XCR_XFEATURE_ENABLED_MASK, xfeatures_mask);
 242 }
 243
 244 /*
 245  * Note that in the future we will likely need a pair of
 246  * functions here: one for user xstates and the other for
 247  * system xstates.  For now, they are the same.
 248  */
 249 static int xfeature_enabled(enum xfeature xfeature)
 250 {
 251         return !!(xfeatures_mask & (1UL << xfeature));
 252 }
 253
 254 /*
 255  * Record the offsets and sizes of various xstates contained
 256  * in the XSAVE state memory layout.
 257  */
 258 static void __init setup_xstate_features(void)
 259 {
 260         u32 eax, ebx, ecx, edx, i;
 261         /* start at the beginnning of the "extended state" */
 262         unsigned int last_good_offset = offsetof(struct xregs_state,
 263                                                  extended_state_area);
 264         /*
 265          * The FP xstates and SSE xstates are legacy states. They are always
 266          * in the fixed offsets in the xsave area in either compacted form
 267          * or standard form.
 268          */
 269         xstate_offsets[0] = 0;
 270         xstate_sizes[0] = offsetof(struct fxregs_state, xmm_space);
 271         xstate_offsets[1] = xstate_sizes[0];
 272         xstate_sizes[1] = FIELD_SIZEOF(struct fxregs_state, xmm_space);
 273
 274         for (i = FIRST_EXTENDED_XFEATURE; i < XFEATURE_MAX; i++) {
 275                 if (!xfeature_enabled(i))
 276                         continue;
 277
 278                 cpuid_count(XSTATE_CPUID, i, &eax, &ebx, &ecx, &edx);
 279
 280                 /*
 281                  * If an xfeature is supervisor state, the offset
 282                  * in EBX is invalid. We leave it to -1.
 283                  */
 284                 if (xfeature_is_user(i))
 285                         xstate_offsets[i] = ebx;
 286
 287                 xstate_sizes[i] = eax;
 288                 /*
 289                  * In our xstate size checks, we assume that the
 290                  * highest-numbered xstate feature has the
 291                  * highest offset in the buffer.  Ensure it does.
 292                  */
 293                 WARN_ONCE(last_good_offset > xstate_offsets[i],
 294                         "x86/fpu: misordered xstate at %d\n", last_good_offset);
 295                 last_good_offset = xstate_offsets[i];
 296         }
 297 }
 298
 299 static void __init print_xstate_feature(u64 xstate_mask)
 300 {
 301         const char *feature_name;
 302
 303         if (cpu_has_xfeatures(xstate_mask, &feature_name))
 304                 pr_info("x86/fpu: Supporting XSAVE feature 0x%03Lx: '%s'\n", xstate_mask, feature_name);
 305 }
 306
 307 /*
 308  * Print out all the supported xstate features:
 309  */
 310 static void __init print_xstate_features(void)
 311 {
 312         print_xstate_feature(XFEATURE_MASK_FP);
 313         print_xstate_feature(XFEATURE_MASK_SSE);
 314         print_xstate_feature(XFEATURE_MASK_YMM);
 315         print_xstate_feature(XFEATURE_MASK_BNDREGS);
 316         print_xstate_feature(XFEATURE_MASK_BNDCSR);
 317         print_xstate_feature(XFEATURE_MASK_OPMASK);
 318         print_xstate_feature(XFEATURE_MASK_ZMM_Hi256);
 319         print_xstate_feature(XFEATURE_MASK_Hi16_ZMM);
 320         print_xstate_feature(XFEATURE_MASK_PKRU);
 321 }
 322
 323 /*
 324  * This check is important because it is easy to get XSTATE_*
 325  * confused with XSTATE_BIT_*.
 326  */
 327 #define CHECK_XFEATURE(nr) do {         \
 328         WARN_ON(nr < FIRST_EXTENDED_XFEATURE);  \
 329         WARN_ON(nr >= XFEATURE_MAX);    \
 330 } while (0)
 331
 332 /*
 333  * We could cache this like xstate_size[], but we only use
 334  * it here, so it would be a waste of space.
 335  */
 336 static int xfeature_is_aligned(int xfeature_nr)
 337 {
 338         u32 eax, ebx, ecx, edx;
 339
 340         CHECK_XFEATURE(xfeature_nr);
 341         cpuid_count(XSTATE_CPUID, xfeature_nr, &eax, &ebx, &ecx, &edx);
 342         /*
 343          * The value returned by ECX[1] indicates the alignment
 344          * of state component 'i' when the compacted format
 345          * of the extended region of an XSAVE area is used:
 346          */
 347         return !!(ecx & 2);
 348 }
 349
 350 /*
 351  * This function sets up offsets and sizes of all extended states in
 352  * xsave area. This supports both standard format and compacted format
 353  * of the xsave aread.
 354  */
 355 static void __init setup_xstate_comp(void)
 356 {
 357         unsigned int xstate_comp_sizes[sizeof(xfeatures_mask)*8];
 358         int i;
 359
 360         /*
 361          * The FP xstates and SSE xstates are legacy states. They are always
 362          * in the fixed offsets in the xsave area in either compacted form
 363          * or standard form.
 364          */
 365         xstate_comp_offsets[0] = 0;
 366         xstate_comp_offsets[1] = offsetof(struct fxregs_state, xmm_space);
 367
 368         if (!boot_cpu_has(X86_FEATURE_XSAVES)) {
 369                 for (i = FIRST_EXTENDED_XFEATURE; i < XFEATURE_MAX; i++) {
 370                         if (xfeature_enabled(i)) {
 371                                 xstate_comp_offsets[i] = xstate_offsets[i];
 372                                 xstate_comp_sizes[i] = xstate_sizes[i];
 373                         }
 374                 }
 375                 return;
 376         }
 377
 378         xstate_comp_offsets[FIRST_EXTENDED_XFEATURE] =
 379                 FXSAVE_SIZE + XSAVE_HDR_SIZE;
 380
 381         for (i = FIRST_EXTENDED_XFEATURE; i < XFEATURE_MAX; i++) {
 382                 if (xfeature_enabled(i))
 383                         xstate_comp_sizes[i] = xstate_sizes[i];
 384                 else
 385                         xstate_comp_sizes[i] = 0;
 386
 387                 if (i > FIRST_EXTENDED_XFEATURE) {
 388                         xstate_comp_offsets[i] = xstate_comp_offsets[i-1]
 389                                         + xstate_comp_sizes[i-1];
 390
 391                         if (xfeature_is_aligned(i))
 392                                 xstate_comp_offsets[i] =
 393                                         ALIGN(xstate_comp_offsets[i], 64);
 394                 }
 395         }
 396 }
 397
 398 /*
 399  * Print out xstate component offsets and sizes
 400  */
 401 static void __init print_xstate_offset_size(void)
 402 {
 403         int i;
 404
 405         for (i = FIRST_EXTENDED_XFEATURE; i < XFEATURE_MAX; i++) {
 406                 if (!xfeature_enabled(i))
 407                         continue;
 408                 pr_info("x86/fpu: xstate_offset[%d]: %4d, xstate_sizes[%d]: %4d\n",
 409                          i, xstate_comp_offsets[i], i, xstate_sizes[i]);
 410         }
 411 }
 412
 413 /*
 414  * setup the xstate image representing the init state
 415  */
 416 static void __init setup_init_fpu_buf(void)
 417 {
 418         static int on_boot_cpu __initdata = 1;
 419
 420         WARN_ON_FPU(!on_boot_cpu);
 421         on_boot_cpu = 0;
 422
 423         if (!boot_cpu_has(X86_FEATURE_XSAVE))
 424                 return;
 425
 426         setup_xstate_features();
 427         print_xstate_features();
 428
 429         if (boot_cpu_has(X86_FEATURE_XSAVES))
 430                 init_fpstate.xsave.header.xcomp_bv = (u64)1 << 63 | xfeatures_mask;
 431
 432         /*
 433          * Init all the features state with header.xfeatures being 0x0
 434          */
 435         copy_kernel_to_xregs_booting(&init_fpstate.xsave);
 436
 437         /*
 438          * Dump the init state again. This is to identify the init state
 439          * of any feature which is not represented by all zero's.
 440          */
 441         copy_xregs_to_kernel_booting(&init_fpstate.xsave);
 442 }
 443
 444 static int xfeature_uncompacted_offset(int xfeature_nr)
 445 {
 446         u32 eax, ebx, ecx, edx;
 447
 448         /*
 449          * Only XSAVES supports supervisor states and it uses compacted
 450          * format. Checking a supervisor state's uncompacted offset is
 451          * an error.
 452          */
 453         if (XFEATURE_MASK_SUPERVISOR & (1 << xfeature_nr)) {
 454                 WARN_ONCE(1, "No fixed offset for xstate %d\n", xfeature_nr);
 455                 return -1;
 456         }
 457
 458         CHECK_XFEATURE(xfeature_nr);
 459         cpuid_count(XSTATE_CPUID, xfeature_nr, &eax, &ebx, &ecx, &edx);
 460         return ebx;
 461 }
 462
 463 static int xfeature_size(int xfeature_nr)
 464 {
 465         u32 eax, ebx, ecx, edx;
 466
 467         CHECK_XFEATURE(xfeature_nr);
 468         cpuid_count(XSTATE_CPUID, xfeature_nr, &eax, &ebx, &ecx, &edx);
 469         return eax;
 470 }
 471
 472 /*
 473  * 'XSAVES' implies two different things:
 474  * 1. saving of supervisor/system state
 475  * 2. using the compacted format
 476  *
 477  * Use this function when dealing with the compacted format so
 478  * that it is obvious which aspect of 'XSAVES' is being handled
 479  * by the calling code.
 480  */
 481 int using_compacted_format(void)
 482 {
 483         return boot_cpu_has(X86_FEATURE_XSAVES);
 484 }
 485
 486 /* Validate an xstate header supplied by userspace (ptrace or sigreturn) */
 487 int validate_xstate_header(const struct xstate_header *hdr)
 488 {
 489         /* No unknown or supervisor features may be set */
 490         if (hdr->xfeatures & (~xfeatures_mask | XFEATURE_MASK_SUPERVISOR))
 491                 return -EINVAL;
 492
 493         /* Userspace must use the uncompacted format */
 494         if (hdr->xcomp_bv)
 495                 return -EINVAL;
 496
 497         /*
 498          * If 'reserved' is shrunken to add a new field, make sure to validate
 499          * that new field here!
 500          */
 501         BUILD_BUG_ON(sizeof(hdr->reserved) != 48);
 502
 503         /* No reserved bits may be set */
 504         if (memchr_inv(hdr->reserved, 0, sizeof(hdr->reserved)))
 505                 return -EINVAL;
 506
 507         return 0;
 508 }
 509
 510 static void __xstate_dump_leaves(void)
 511 {
 512         int i;
 513         u32 eax, ebx, ecx, edx;
 514         static int should_dump = 1;
 515
 516         if (!should_dump)
 517                 return;
 518         should_dump = 0;
 519         /*
 520          * Dump out a few leaves past the ones that we support
 521          * just in case there are some goodies up there
 522          */
 523         for (i = 0; i < XFEATURE_MAX + 10; i++) {
 524                 cpuid_count(XSTATE_CPUID, i, &eax, &ebx, &ecx, &edx);
 525                 pr_warn("CPUID[%02x, %02x]: eax=%08x ebx=%08x ecx=%08x edx=%08x\n",
 526                         XSTATE_CPUID, i, eax, ebx, ecx, edx);
 527         }
 528 }
 529
 530 #define XSTATE_WARN_ON(x) do {                                                  \
 531         if (WARN_ONCE(x, "XSAVE consistency problem, dumping leaves")) {        \
 532                 __xstate_dump_leaves();                                         \
 533         }                                                                       \
 534 } while (0)
 535
 536 #define XCHECK_SZ(sz, nr, nr_macro, __struct) do {                      \
 537         if ((nr == nr_macro) &&                                         \
 538             WARN_ONCE(sz != sizeof(__struct),                           \
 539                 "%s: struct is %zu bytes, cpu state %d bytes\n",        \
 540                 __stringify(nr_macro), sizeof(__struct), sz)) {         \
 541                 __xstate_dump_leaves();                                 \
 542         }                                                               \
 543 } while (0)
 544
 545 /*
 546  * We have a C struct for each 'xstate'.  We need to ensure
 547  * that our software representation matches what the CPU
 548  * tells us about the state's size.
 549  */
 550 static void check_xstate_against_struct(int nr)
 551 {
 552         /*
 553          * Ask the CPU for the size of the state.
 554          */
 555         int sz = xfeature_size(nr);
 556         /*
 557          * Match each CPU state with the corresponding software
 558          * structure.
 559          */
 560         XCHECK_SZ(sz, nr, XFEATURE_YMM,       struct ymmh_struct);
 561         XCHECK_SZ(sz, nr, XFEATURE_BNDREGS,   struct mpx_bndreg_state);
 562         XCHECK_SZ(sz, nr, XFEATURE_BNDCSR,    struct mpx_bndcsr_state);
 563         XCHECK_SZ(sz, nr, XFEATURE_OPMASK,    struct avx_512_opmask_state);
 564         XCHECK_SZ(sz, nr, XFEATURE_ZMM_Hi256, struct avx_512_zmm_uppers_state);
 565         XCHECK_SZ(sz, nr, XFEATURE_Hi16_ZMM,  struct avx_512_hi16_state);
 566         XCHECK_SZ(sz, nr, XFEATURE_PKRU,      struct pkru_state);
 567
 568         /*
 569          * Make *SURE* to add any feature numbers in below if
 570          * there are "holes" in the xsave state component
 571          * numbers.
 572          */
 573         if ((nr < XFEATURE_YMM) ||
 574             (nr >= XFEATURE_MAX) ||
 575             (nr == XFEATURE_PT_UNIMPLEMENTED_SO_FAR)) {
 576                 WARN_ONCE(1, "no structure for xstate: %d\n", nr);
 577                 XSTATE_WARN_ON(1);
 578         }
 579 }
 580
 581 /*
 582  * This essentially double-checks what the cpu told us about
 583  * how large the XSAVE buffer needs to be.  We are recalculating
 584  * it to be safe.
 585  */
 586 static void do_extra_xstate_size_checks(void)
 587 {
 588         int paranoid_xstate_size = FXSAVE_SIZE + XSAVE_HDR_SIZE;
 589         int i;
 590
 591         for (i = FIRST_EXTENDED_XFEATURE; i < XFEATURE_MAX; i++) {
 592                 if (!xfeature_enabled(i))
 593                         continue;
 594
 595                 check_xstate_against_struct(i);
 596                 /*
 597                  * Supervisor state components can be managed only by
 598                  * XSAVES, which is compacted-format only.
 599                  */
 600                 if (!using_compacted_format())
 601                         XSTATE_WARN_ON(xfeature_is_supervisor(i));
 602
 603                 /* Align from the end of the previous feature */
 604                 if (xfeature_is_aligned(i))
 605                         paranoid_xstate_size = ALIGN(paranoid_xstate_size, 64);
 606                 /*
 607                  * The offset of a given state in the non-compacted
 608                  * format is given to us in a CPUID leaf.  We check
 609                  * them for being ordered (increasing offsets) in
 610                  * setup_xstate_features().
 611                  */
 612                 if (!using_compacted_format())
 613                         paranoid_xstate_size = xfeature_uncompacted_offset(i);
 614                 /*
 615                  * The compacted-format offset always depends on where
 616                  * the previous state ended.
 617                  */
 618                 paranoid_xstate_size += xfeature_size(i);
 619         }
 620         XSTATE_WARN_ON(paranoid_xstate_size != fpu_kernel_xstate_size);
 621 }
 622
 623
 624 /*
 625  * Get total size of enabled xstates in XCR0/xfeatures_mask.
 626  *
 627  * Note the SDM's wording here.  "sub-function 0" only enumerates
 628  * the size of the *user* states.  If we use it to size a buffer
 629  * that we use 'XSAVES' on, we could potentially overflow the
 630  * buffer because 'XSAVES' saves system states too.
 631  *
 632  * Note that we do not currently set any bits on IA32_XSS so
 633  * 'XCR0 | IA32_XSS == XCR0' for now.
 634  */
 635 static unsigned int __init get_xsaves_size(void)
 636 {
 637         unsigned int eax, ebx, ecx, edx;
 638         /*
 639          * - CPUID function 0DH, sub-function 1:
 640          *    EBX enumerates the size (in bytes) required by
 641          *    the XSAVES instruction for an XSAVE area
 642          *    containing all the state components
 643          *    corresponding to bits currently set in
 644          *    XCR0 | IA32_XSS.
 645          */
 646         cpuid_count(XSTATE_CPUID, 1, &eax, &ebx, &ecx, &edx);
 647         return ebx;
 648 }
 649
 650 static unsigned int __init get_xsave_size(void)
 651 {
 652         unsigned int eax, ebx, ecx, edx;
 653         /*
 654          * - CPUID function 0DH, sub-function 0:
 655          *    EBX enumerates the size (in bytes) required by
 656          *    the XSAVE instruction for an XSAVE area
 657          *    containing all the *user* state components
 658          *    corresponding to bits currently set in XCR0.
 659          */
 660         cpuid_count(XSTATE_CPUID, 0, &eax, &ebx, &ecx, &edx);
 661         return ebx;
 662 }
 663
 664 /*
 665  * Will the runtime-enumerated 'xstate_size' fit in the init
 666  * task's statically-allocated buffer?
 667  */
 668 static bool is_supported_xstate_size(unsigned int test_xstate_size)
 669 {
 670         if (test_xstate_size <= sizeof(union fpregs_state))
 671                 return true;
 672
 673         pr_warn("x86/fpu: xstate buffer too small (%zu < %d), disabling xsave\n",
 674                         sizeof(union fpregs_state), test_xstate_size);
 675         return false;
 676 }
 677
 678 static int init_xstate_size(void)
 679 {
 680         /* Recompute the context size for enabled features: */
 681         unsigned int possible_xstate_size;
 682         unsigned int xsave_size;
 683
 684         xsave_size = get_xsave_size();
 685
 686         if (boot_cpu_has(X86_FEATURE_XSAVES))
 687                 possible_xstate_size = get_xsaves_size();
 688         else
 689                 possible_xstate_size = xsave_size;
 690
 691         /* Ensure we have the space to store all enabled: */
 692         if (!is_supported_xstate_size(possible_xstate_size))
 693                 return -EINVAL;
 694
 695         /*
 696          * The size is OK, we are definitely going to use xsave,
 697          * make it known to the world that we need more space.
 698          */
 699         fpu_kernel_xstate_size = possible_xstate_size;
 700         do_extra_xstate_size_checks();
 701
 702         /*
 703          * User space is always in standard format.
 704          */
 705         fpu_user_xstate_size = xsave_size;
 706         return 0;
 707 }
 708
 709 /*
 710  * We enabled the XSAVE hardware, but something went wrong and
 711  * we can not use it.  Disable it.
 712  */
 713 static void fpu__init_disable_system_xstate(void)
 714 {
 715         xfeatures_mask = 0;
 716         cr4_clear_bits(X86_CR4_OSXSAVE);
 717         fpu__xstate_clear_all_cpu_caps();
 718 }
 719
 720 /*
 721  * Enable and initialize the xsave feature.
 722  * Called once per system bootup.
 723  */
 724 void __init fpu__init_system_xstate(void)
 725 {
 726         unsigned int eax, ebx, ecx, edx;
 727         static int on_boot_cpu __initdata = 1;
 728         int err;
 729
 730         WARN_ON_FPU(!on_boot_cpu);
 731         on_boot_cpu = 0;
 732
 733         if (!boot_cpu_has(X86_FEATURE_FPU)) {
 734                 pr_info("x86/fpu: No FPU detected\n");
 735                 return;
 736         }
 737
 738         if (!boot_cpu_has(X86_FEATURE_XSAVE)) {
 739                 pr_info("x86/fpu: x87 FPU will use %s\n",
 740                         boot_cpu_has(X86_FEATURE_FXSR) ? "FXSAVE" : "FSAVE");
 741                 return;
 742         }
 743
 744         if (boot_cpu_data.cpuid_level < XSTATE_CPUID) {
 745                 WARN_ON_FPU(1);
 746                 return;
 747         }
 748
 749         cpuid_count(XSTATE_CPUID, 0, &eax, &ebx, &ecx, &edx);
 750         xfeatures_mask = eax + ((u64)edx << 32);
 751
 752         if ((xfeatures_mask & XFEATURE_MASK_FPSSE) != XFEATURE_MASK_FPSSE) {
 753                 /*
 754                  * This indicates that something really unexpected happened
 755                  * with the enumeration.  Disable XSAVE and try to continue
 756                  * booting without it.  This is too early to BUG().
 757                  */
 758                 pr_err("x86/fpu: FP/SSE not present amongst the CPU's xstate features: 0x%llx.\n", xfeatures_mask);
 759                 goto out_disable;
 760         }
 761
 762         xfeatures_mask &= fpu__get_supported_xfeatures_mask();
 763
 764         /* Enable xstate instructions to be able to continue with initialization: */
 765         fpu__init_cpu_xstate();
 766         err = init_xstate_size();
 767         if (err)
 768                 goto out_disable;
 769
 770         /*
 771          * Update info used for ptrace frames; use standard-format size and no
 772          * supervisor xstates:
 773          */
 774         update_regset_xstate_info(fpu_user_xstate_size, xfeatures_mask & ~XFEATURE_MASK_SUPERVISOR);
 775
 776         fpu__init_prepare_fx_sw_frame();
 777         setup_init_fpu_buf();
 778         setup_xstate_comp();
 779         print_xstate_offset_size();
 780
 781         pr_info("x86/fpu: Enabled xstate features 0x%llx, context size is %d bytes, using '%s' format.\n",
 782                 xfeatures_mask,
 783                 fpu_kernel_xstate_size,
 784                 boot_cpu_has(X86_FEATURE_XSAVES) ? "compacted" : "standard");
 785         return;
 786
 787 out_disable:
 788         /* something went wrong, try to boot without any XSAVE support */
 789         fpu__init_disable_system_xstate();
 790 }
 791
 792 /*
 793  * Restore minimal FPU state after suspend:
 794  */
 795 void fpu__resume_cpu(void)
 796 {
 797         /*
 798          * Restore XCR0 on xsave capable CPUs:
 799          */
 800         if (boot_cpu_has(X86_FEATURE_XSAVE))
 801                 xsetbv(XCR_XFEATURE_ENABLED_MASK, xfeatures_mask);
 802 }
 803
 804 /*
 805  * Given an xstate feature mask, calculate where in the xsave
 806  * buffer the state is.  Callers should ensure that the buffer
 807  * is valid.
 808  *
 809  * Note: does not work for compacted buffers.
 810  */
 811 void *__raw_xsave_addr(struct xregs_state *xsave, int xstate_feature_mask)
 812 {
 813         int feature_nr = fls64(xstate_feature_mask) - 1;
 814
 815         if (!xfeature_enabled(feature_nr)) {
 816                 WARN_ON_FPU(1);
 817                 return NULL;
 818         }
 819
 820         return (void *)xsave + xstate_comp_offsets[feature_nr];
 821 }
 822 /*
 823  * Given the xsave area and a state inside, this function returns the
 824  * address of the state.
 825  *
 826  * This is the API that is called to get xstate address in either
 827  * standard format or compacted format of xsave area.
 828  *
 829  * Note that if there is no data for the field in the xsave buffer
 830  * this will return NULL.
 831  *
 832  * Inputs:
 833  *      xstate: the thread's storage area for all FPU data
 834  *      xstate_feature: state which is defined in xsave.h (e.g.
 835  *      XFEATURE_MASK_FP, XFEATURE_MASK_SSE, etc...)
 836  * Output:
 837  *      address of the state in the xsave area, or NULL if the
 838  *      field is not present in the xsave buffer.
 839  */
 840 void *get_xsave_addr(struct xregs_state *xsave, int xstate_feature)
 841 {
 842         /*
 843          * Do we even *have* xsave state?
 844          */
 845         if (!boot_cpu_has(X86_FEATURE_XSAVE))
 846                 return NULL;
 847
 848         /*
 849          * We should not ever be requesting features that we
 850          * have not enabled.  Remember that pcntxt_mask is
 851          * what we write to the XCR0 register.
 852          */
 853         WARN_ONCE(!(xfeatures_mask & xstate_feature),
 854                   "get of unsupported state");
 855         /*
 856          * This assumes the last 'xsave*' instruction to
 857          * have requested that 'xstate_feature' be saved.
 858          * If it did not, we might be seeing and old value
 859          * of the field in the buffer.
 860          *
 861          * This can happen because the last 'xsave' did not
 862          * request that this feature be saved (unlikely)
 863          * or because the "init optimization" caused it
 864          * to not be saved.
 865          */
 866         if (!(xsave->header.xfeatures & xstate_feature))
 867                 return NULL;
 868
 869         return __raw_xsave_addr(xsave, xstate_feature);
 870 }
 871 EXPORT_SYMBOL_GPL(get_xsave_addr);
 872
 873 /*
 874  * This wraps up the common operations that need to occur when retrieving
 875  * data from xsave state.  It first ensures that the current task was
 876  * using the FPU and retrieves the data in to a buffer.  It then calculates
 877  * the offset of the requested field in the buffer.
 878  *
 879  * This function is safe to call whether the FPU is in use or not.
 880  *
 881  * Note that this only works on the current task.
 882  *
 883  * Inputs:
 884  *      @xsave_state: state which is defined in xsave.h (e.g. XFEATURE_MASK_FP,
 885  *      XFEATURE_MASK_SSE, etc...)
 886  * Output:
 887  *      address of the state in the xsave area or NULL if the state
 888  *      is not present or is in its 'init state'.
 889  */
 890 const void *get_xsave_field_ptr(int xsave_state)
 891 {
 892         struct fpu *fpu = &current->thread.fpu;
 893
 894         if (!fpu->initialized)
 895                 return NULL;
 896         /*
 897          * fpu__save() takes the CPU's xstate registers
 898          * and saves them off to the 'fpu memory buffer.
 899          */
 900         fpu__save(fpu);
 901
 902         return get_xsave_addr(&fpu->state.xsave, xsave_state);
 903 }
 904
 905 #ifdef CONFIG_ARCH_HAS_PKEYS
 906
 907 #define NR_VALID_PKRU_BITS (CONFIG_NR_PROTECTION_KEYS * 2)
 908 #define PKRU_VALID_MASK (NR_VALID_PKRU_BITS - 1)
 909 /*
 910  * This will go out and modify PKRU register to set the access
 911  * rights for @pkey to @init_val.
 912  */
 913 int arch_set_user_pkey_access(struct task_struct *tsk, int pkey,
 914                 unsigned long init_val)
 915 {
 916         u32 old_pkru;
 917         int pkey_shift = (pkey * PKRU_BITS_PER_PKEY);
 918         u32 new_pkru_bits = 0;
 919
 920         /*
 921          * This check implies XSAVE support.  OSPKE only gets
 922          * set if we enable XSAVE and we enable PKU in XCR0.
 923          */
 924         if (!boot_cpu_has(X86_FEATURE_OSPKE))
 925                 return -EINVAL;
 926
 927         /* Set the bits we need in PKRU:  */
 928         if (init_val & PKEY_DISABLE_ACCESS)
 929                 new_pkru_bits |= PKRU_AD_BIT;
 930         if (init_val & PKEY_DISABLE_WRITE)
 931                 new_pkru_bits |= PKRU_WD_BIT;
 932
 933         /* Shift the bits in to the correct place in PKRU for pkey: */
 934         new_pkru_bits <<= pkey_shift;
 935
 936         /* Get old PKRU and mask off any old bits in place: */
 937         old_pkru = read_pkru();
 938         old_pkru &= ~((PKRU_AD_BIT|PKRU_WD_BIT) << pkey_shift);
 939
 940         /* Write old part along with new part: */
 941         write_pkru(old_pkru | new_pkru_bits);
 942
 943         return 0;
 944 }
 945 #endif /* ! CONFIG_ARCH_HAS_PKEYS */
 946
 947 /*
 948  * Weird legacy quirk: SSE and YMM states store information in the
 949  * MXCSR and MXCSR_FLAGS fields of the FP area. That means if the FP
 950  * area is marked as unused in the xfeatures header, we need to copy
 951  * MXCSR and MXCSR_FLAGS if either SSE or YMM are in use.
 952  */
 953 static inline bool xfeatures_mxcsr_quirk(u64 xfeatures)
 954 {
 955         if (!(xfeatures & (XFEATURE_MASK_SSE|XFEATURE_MASK_YMM)))
 956                 return false;
 957
 958         if (xfeatures & XFEATURE_MASK_FP)
 959                 return false;
 960
 961         return true;
 962 }
 963
 964 /*
 965  * This is similar to user_regset_copyout(), but will not add offset to
 966  * the source data pointer or increment pos, count, kbuf, and ubuf.
 967  */
 968 static inline void
 969 __copy_xstate_to_kernel(void *kbuf, const void *data,
 970                         unsigned int offset, unsigned int size, unsigned int size_total)
 971 {
 972         if (offset < size_total) {
 973                 unsigned int copy = min(size, size_total - offset);
 974
 975                 memcpy(kbuf + offset, data, copy);
 976         }
 977 }
 978
 979 /*
 980  * Convert from kernel XSAVES compacted format to standard format and copy
 981  * to a kernel-space ptrace buffer.
 982  *
 983  * It supports partial copy but pos always starts from zero. This is called
 984  * from xstateregs_get() and there we check the CPU has XSAVES.
 985  */
 986 int copy_xstate_to_kernel(void *kbuf, struct xregs_state *xsave, unsigned int offset_start, unsigned int size_total)
 987 {
 988         unsigned int offset, size;
 989         struct xstate_header header;
 990         int i;
 991
 992         /*
 993          * Currently copy_regset_to_user() starts from pos 0:
 994          */
 995         if (unlikely(offset_start != 0))
 996                 return -EFAULT;
 997
 998         /*
 999          * The destination is a ptrace buffer; we put in only user xstates:
1000          */
1001         memset(&header, 0, sizeof(header));
1002         header.xfeatures = xsave->header.xfeatures;
1003         header.xfeatures &= ~XFEATURE_MASK_SUPERVISOR;
1004
1005         /*
1006          * Copy xregs_state->header:
1007          */
1008         offset = offsetof(struct xregs_state, header);
1009         size = sizeof(header);
1010
1011         __copy_xstate_to_kernel(kbuf, &header, offset, size, size_total);
1012
1013         for (i = 0; i < XFEATURE_MAX; i++) {
1014                 /*
1015                  * Copy only in-use xstates:
1016                  */
1017                 if ((header.xfeatures >> i) & 1) {
1018                         void *src = __raw_xsave_addr(xsave, 1 << i);
1019
1020                         offset = xstate_offsets[i];
1021                         size = xstate_sizes[i];
1022
1023                         /* The next component has to fit fully into the output buffer: */
1024                         if (offset + size > size_total)
1025                                 break;
1026
1027                         __copy_xstate_to_kernel(kbuf, src, offset, size, size_total);
1028                 }
1029
1030         }
1031
1032         if (xfeatures_mxcsr_quirk(header.xfeatures)) {
1033                 offset = offsetof(struct fxregs_state, mxcsr);
1034                 size = MXCSR_AND_FLAGS_SIZE;
1035                 __copy_xstate_to_kernel(kbuf, &xsave->i387.mxcsr, offset, size, size_total);
1036         }
1037
1038         /*
1039          * Fill xsave->i387.sw_reserved value for ptrace frame:
1040          */
1041         offset = offsetof(struct fxregs_state, sw_reserved);
1042         size = sizeof(xstate_fx_sw_bytes);
1043
1044         __copy_xstate_to_kernel(kbuf, xstate_fx_sw_bytes, offset, size, size_total);
1045
1046         return 0;
1047 }
1048
1049 static inline int
1050 __copy_xstate_to_user(void __user *ubuf, const void *data, unsigned int offset, unsigned int size, unsigned int size_total)
1051 {
1052         if (!size)
1053                 return 0;
1054
1055         if (offset < size_total) {
1056                 unsigned int copy = min(size, size_total - offset);
1057
1058                 if (__copy_to_user(ubuf + offset, data, copy))
1059                         return -EFAULT;
1060         }
1061         return 0;
1062 }
1063
1064 /*
1065  * Convert from kernel XSAVES compacted format to standard format and copy
1066  * to a user-space buffer. It supports partial copy but pos always starts from
1067  * zero. This is called from xstateregs_get() and there we check the CPU
1068  * has XSAVES.
1069  */
1070 int copy_xstate_to_user(void __user *ubuf, struct xregs_state *xsave, unsigned int offset_start, unsigned int size_total)
1071 {
1072         unsigned int offset, size;
1073         int ret, i;
1074         struct xstate_header header;
1075
1076         /*
1077          * Currently copy_regset_to_user() starts from pos 0:
1078          */
1079         if (unlikely(offset_start != 0))
1080                 return -EFAULT;
1081
1082         /*
1083          * The destination is a ptrace buffer; we put in only user xstates:
1084          */
1085         memset(&header, 0, sizeof(header));
1086         header.xfeatures = xsave->header.xfeatures;
1087         header.xfeatures &= ~XFEATURE_MASK_SUPERVISOR;
1088
1089         /*
1090          * Copy xregs_state->header:
1091          */
1092         offset = offsetof(struct xregs_state, header);
1093         size = sizeof(header);
1094
1095         ret = __copy_xstate_to_user(ubuf, &header, offset, size, size_total);
1096         if (ret)
1097                 return ret;
1098
1099         for (i = 0; i < XFEATURE_MAX; i++) {
1100                 /*
1101                  * Copy only in-use xstates:
1102                  */
1103                 if ((header.xfeatures >> i) & 1) {
1104                         void *src = __raw_xsave_addr(xsave, 1 << i);
1105
1106                         offset = xstate_offsets[i];
1107                         size = xstate_sizes[i];
1108
1109                         /* The next component has to fit fully into the output buffer: */
1110                         if (offset + size > size_total)
1111                                 break;
1112
1113                         ret = __copy_xstate_to_user(ubuf, src, offset, size, size_total);
1114                         if (ret)
1115                                 return ret;
1116                 }
1117
1118         }
1119
1120         if (xfeatures_mxcsr_quirk(header.xfeatures)) {
1121                 offset = offsetof(struct fxregs_state, mxcsr);
1122                 size = MXCSR_AND_FLAGS_SIZE;
1123                 __copy_xstate_to_user(ubuf, &xsave->i387.mxcsr, offset, size, size_total);
1124         }
1125
1126         /*
1127          * Fill xsave->i387.sw_reserved value for ptrace frame:
1128          */
1129         offset = offsetof(struct fxregs_state, sw_reserved);
1130         size = sizeof(xstate_fx_sw_bytes);
1131
1132         ret = __copy_xstate_to_user(ubuf, xstate_fx_sw_bytes, offset, size, size_total);
1133         if (ret)
1134                 return ret;
1135
1136         return 0;
1137 }
1138
1139 /*
1140  * Convert from a ptrace standard-format kernel buffer to kernel XSAVES format
1141  * and copy to the target thread. This is called from xstateregs_set().
1142  */
1143 int copy_kernel_to_xstate(struct xregs_state *xsave, const void *kbuf)
1144 {
1145         unsigned int offset, size;
1146         int i;
1147         struct xstate_header hdr;
1148
1149         offset = offsetof(struct xregs_state, header);
1150         size = sizeof(hdr);
1151
1152         memcpy(&hdr, kbuf + offset, size);
1153
1154         if (validate_xstate_header(&hdr))
1155                 return -EINVAL;
1156
1157         for (i = 0; i < XFEATURE_MAX; i++) {
1158                 u64 mask = ((u64)1 << i);
1159
1160                 if (hdr.xfeatures & mask) {
1161                         void *dst = __raw_xsave_addr(xsave, 1 << i);
1162
1163                         offset = xstate_offsets[i];
1164                         size = xstate_sizes[i];
1165
1166                         memcpy(dst, kbuf + offset, size);
1167                 }
1168         }
1169
1170         if (xfeatures_mxcsr_quirk(hdr.xfeatures)) {
1171                 offset = offsetof(struct fxregs_state, mxcsr);
1172                 size = MXCSR_AND_FLAGS_SIZE;
1173                 memcpy(&xsave->i387.mxcsr, kbuf + offset, size);
1174         }
1175
1176         /*
1177          * The state that came in from userspace was user-state only.
1178          * Mask all the user states out of 'xfeatures':
1179          */
1180         xsave->header.xfeatures &= XFEATURE_MASK_SUPERVISOR;
1181
1182         /*
1183          * Add back in the features that came in from userspace:
1184          */
1185         xsave->header.xfeatures |= hdr.xfeatures;
1186
1187         return 0;
1188 }
1189
1190 /*
1191  * Convert from a ptrace or sigreturn standard-format user-space buffer to
1192  * kernel XSAVES format and copy to the target thread. This is called from
1193  * xstateregs_set(), as well as potentially from the sigreturn() and
1194  * rt_sigreturn() system calls.
1195  */
1196 int copy_user_to_xstate(struct xregs_state *xsave, const void __user *ubuf)
1197 {
1198         unsigned int offset, size;
1199         int i;
1200         struct xstate_header hdr;
1201
1202         offset = offsetof(struct xregs_state, header);
1203         size = sizeof(hdr);
1204
1205         if (__copy_from_user(&hdr, ubuf + offset, size))
1206                 return -EFAULT;
1207
1208         if (validate_xstate_header(&hdr))
1209                 return -EINVAL;
1210
1211         for (i = 0; i < XFEATURE_MAX; i++) {
1212                 u64 mask = ((u64)1 << i);
1213
1214                 if (hdr.xfeatures & mask) {
1215                         void *dst = __raw_xsave_addr(xsave, 1 << i);
1216
1217                         offset = xstate_offsets[i];
1218                         size = xstate_sizes[i];
1219
1220                         if (__copy_from_user(dst, ubuf + offset, size))
1221                                 return -EFAULT;
1222                 }
1223         }
1224
1225         if (xfeatures_mxcsr_quirk(hdr.xfeatures)) {
1226                 offset = offsetof(struct fxregs_state, mxcsr);
1227                 size = MXCSR_AND_FLAGS_SIZE;
1228                 if (__copy_from_user(&xsave->i387.mxcsr, ubuf + offset, size))
1229                         return -EFAULT;
1230         }
1231
1232         /*
1233          * The state that came in from userspace was user-state only.
1234          * Mask all the user states out of 'xfeatures':
1235          */
1236         xsave->header.xfeatures &= XFEATURE_MASK_SUPERVISOR;
1237
1238         /*
1239          * Add back in the features that came in from userspace:
1240          */
1241         xsave->header.xfeatures |= hdr.xfeatures;
1242
1243         return 0;
1244 }