arch/x86/kernel/fpu/xstate.c

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