linux/tools/testing/selftests/x86/sigreturn.c
<<
>>
Prefs
   1// SPDX-License-Identifier: GPL-2.0-only
   2/*
   3 * sigreturn.c - tests for x86 sigreturn(2) and exit-to-userspace
   4 * Copyright (c) 2014-2015 Andrew Lutomirski
   5 *
   6 * This is a series of tests that exercises the sigreturn(2) syscall and
   7 * the IRET / SYSRET paths in the kernel.
   8 *
   9 * For now, this focuses on the effects of unusual CS and SS values,
  10 * and it has a bunch of tests to make sure that ESP/RSP is restored
  11 * properly.
  12 *
  13 * The basic idea behind these tests is to raise(SIGUSR1) to create a
  14 * sigcontext frame, plug in the values to be tested, and then return,
  15 * which implicitly invokes sigreturn(2) and programs the user context
  16 * as desired.
  17 *
  18 * For tests for which we expect sigreturn and the subsequent return to
  19 * user mode to succeed, we return to a short trampoline that generates
  20 * SIGTRAP so that the meat of the tests can be ordinary C code in a
  21 * SIGTRAP handler.
  22 *
  23 * The inner workings of each test is documented below.
  24 *
  25 * Do not run on outdated, unpatched kernels at risk of nasty crashes.
  26 */
  27
  28#define _GNU_SOURCE
  29
  30#include <sys/time.h>
  31#include <time.h>
  32#include <stdlib.h>
  33#include <sys/syscall.h>
  34#include <unistd.h>
  35#include <stdio.h>
  36#include <string.h>
  37#include <inttypes.h>
  38#include <sys/mman.h>
  39#include <sys/signal.h>
  40#include <sys/ucontext.h>
  41#include <asm/ldt.h>
  42#include <err.h>
  43#include <setjmp.h>
  44#include <stddef.h>
  45#include <stdbool.h>
  46#include <sys/ptrace.h>
  47#include <sys/user.h>
  48
  49/* Pull in AR_xyz defines. */
  50typedef unsigned int u32;
  51typedef unsigned short u16;
  52#include "../../../../arch/x86/include/asm/desc_defs.h"
  53
  54/*
  55 * Copied from asm/ucontext.h, as asm/ucontext.h conflicts badly with the glibc
  56 * headers.
  57 */
  58#ifdef __x86_64__
  59/*
  60 * UC_SIGCONTEXT_SS will be set when delivering 64-bit or x32 signals on
  61 * kernels that save SS in the sigcontext.  All kernels that set
  62 * UC_SIGCONTEXT_SS will correctly restore at least the low 32 bits of esp
  63 * regardless of SS (i.e. they implement espfix).
  64 *
  65 * Kernels that set UC_SIGCONTEXT_SS will also set UC_STRICT_RESTORE_SS
  66 * when delivering a signal that came from 64-bit code.
  67 *
  68 * Sigreturn restores SS as follows:
  69 *
  70 * if (saved SS is valid || UC_STRICT_RESTORE_SS is set ||
  71 *     saved CS is not 64-bit)
  72 *         new SS = saved SS  (will fail IRET and signal if invalid)
  73 * else
  74 *         new SS = a flat 32-bit data segment
  75 */
  76#define UC_SIGCONTEXT_SS       0x2
  77#define UC_STRICT_RESTORE_SS   0x4
  78#endif
  79
  80/*
  81 * In principle, this test can run on Linux emulation layers (e.g.
  82 * Illumos "LX branded zones").  Solaris-based kernels reserve LDT
  83 * entries 0-5 for their own internal purposes, so start our LDT
  84 * allocations above that reservation.  (The tests don't pass on LX
  85 * branded zones, but at least this lets them run.)
  86 */
  87#define LDT_OFFSET 6
  88
  89/* An aligned stack accessible through some of our segments. */
  90static unsigned char stack16[65536] __attribute__((aligned(4096)));
  91
  92/*
  93 * An aligned int3 instruction used as a trampoline.  Some of the tests
  94 * want to fish out their ss values, so this trampoline copies ss to eax
  95 * before the int3.
  96 */
  97asm (".pushsection .text\n\t"
  98     ".type int3, @function\n\t"
  99     ".align 4096\n\t"
 100     "int3:\n\t"
 101     "mov %ss,%ecx\n\t"
 102     "int3\n\t"
 103     ".size int3, . - int3\n\t"
 104     ".align 4096, 0xcc\n\t"
 105     ".popsection");
 106extern char int3[4096];
 107
 108/*
 109 * At startup, we prepapre:
 110 *
 111 * - ldt_nonexistent_sel: An LDT entry that doesn't exist (all-zero
 112 *   descriptor or out of bounds).
 113 * - code16_sel: A 16-bit LDT code segment pointing to int3.
 114 * - data16_sel: A 16-bit LDT data segment pointing to stack16.
 115 * - npcode32_sel: A 32-bit not-present LDT code segment pointing to int3.
 116 * - npdata32_sel: A 32-bit not-present LDT data segment pointing to stack16.
 117 * - gdt_data16_idx: A 16-bit GDT data segment pointing to stack16.
 118 * - gdt_npdata32_idx: A 32-bit not-present GDT data segment pointing to
 119 *   stack16.
 120 *
 121 * For no particularly good reason, xyz_sel is a selector value with the
 122 * RPL and LDT bits filled in, whereas xyz_idx is just an index into the
 123 * descriptor table.  These variables will be zero if their respective
 124 * segments could not be allocated.
 125 */
 126static unsigned short ldt_nonexistent_sel;
 127static unsigned short code16_sel, data16_sel, npcode32_sel, npdata32_sel;
 128
 129static unsigned short gdt_data16_idx, gdt_npdata32_idx;
 130
 131static unsigned short GDT3(int idx)
 132{
 133        return (idx << 3) | 3;
 134}
 135
 136static unsigned short LDT3(int idx)
 137{
 138        return (idx << 3) | 7;
 139}
 140
 141/* Our sigaltstack scratch space. */
 142static char altstack_data[SIGSTKSZ];
 143
 144static void sethandler(int sig, void (*handler)(int, siginfo_t *, void *),
 145                       int flags)
 146{
 147        struct sigaction sa;
 148        memset(&sa, 0, sizeof(sa));
 149        sa.sa_sigaction = handler;
 150        sa.sa_flags = SA_SIGINFO | flags;
 151        sigemptyset(&sa.sa_mask);
 152        if (sigaction(sig, &sa, 0))
 153                err(1, "sigaction");
 154}
 155
 156static void clearhandler(int sig)
 157{
 158        struct sigaction sa;
 159        memset(&sa, 0, sizeof(sa));
 160        sa.sa_handler = SIG_DFL;
 161        sigemptyset(&sa.sa_mask);
 162        if (sigaction(sig, &sa, 0))
 163                err(1, "sigaction");
 164}
 165
 166static void add_ldt(const struct user_desc *desc, unsigned short *var,
 167                    const char *name)
 168{
 169        if (syscall(SYS_modify_ldt, 1, desc, sizeof(*desc)) == 0) {
 170                *var = LDT3(desc->entry_number);
 171        } else {
 172                printf("[NOTE]\tFailed to create %s segment\n", name);
 173                *var = 0;
 174        }
 175}
 176
 177static void setup_ldt(void)
 178{
 179        if ((unsigned long)stack16 > (1ULL << 32) - sizeof(stack16))
 180                errx(1, "stack16 is too high\n");
 181        if ((unsigned long)int3 > (1ULL << 32) - sizeof(int3))
 182                errx(1, "int3 is too high\n");
 183
 184        ldt_nonexistent_sel = LDT3(LDT_OFFSET + 2);
 185
 186        const struct user_desc code16_desc = {
 187                .entry_number    = LDT_OFFSET + 0,
 188                .base_addr       = (unsigned long)int3,
 189                .limit           = 4095,
 190                .seg_32bit       = 0,
 191                .contents        = 2, /* Code, not conforming */
 192                .read_exec_only  = 0,
 193                .limit_in_pages  = 0,
 194                .seg_not_present = 0,
 195                .useable         = 0
 196        };
 197        add_ldt(&code16_desc, &code16_sel, "code16");
 198
 199        const struct user_desc data16_desc = {
 200                .entry_number    = LDT_OFFSET + 1,
 201                .base_addr       = (unsigned long)stack16,
 202                .limit           = 0xffff,
 203                .seg_32bit       = 0,
 204                .contents        = 0, /* Data, grow-up */
 205                .read_exec_only  = 0,
 206                .limit_in_pages  = 0,
 207                .seg_not_present = 0,
 208                .useable         = 0
 209        };
 210        add_ldt(&data16_desc, &data16_sel, "data16");
 211
 212        const struct user_desc npcode32_desc = {
 213                .entry_number    = LDT_OFFSET + 3,
 214                .base_addr       = (unsigned long)int3,
 215                .limit           = 4095,
 216                .seg_32bit       = 1,
 217                .contents        = 2, /* Code, not conforming */
 218                .read_exec_only  = 0,
 219                .limit_in_pages  = 0,
 220                .seg_not_present = 1,
 221                .useable         = 0
 222        };
 223        add_ldt(&npcode32_desc, &npcode32_sel, "npcode32");
 224
 225        const struct user_desc npdata32_desc = {
 226                .entry_number    = LDT_OFFSET + 4,
 227                .base_addr       = (unsigned long)stack16,
 228                .limit           = 0xffff,
 229                .seg_32bit       = 1,
 230                .contents        = 0, /* Data, grow-up */
 231                .read_exec_only  = 0,
 232                .limit_in_pages  = 0,
 233                .seg_not_present = 1,
 234                .useable         = 0
 235        };
 236        add_ldt(&npdata32_desc, &npdata32_sel, "npdata32");
 237
 238        struct user_desc gdt_data16_desc = {
 239                .entry_number    = -1,
 240                .base_addr       = (unsigned long)stack16,
 241                .limit           = 0xffff,
 242                .seg_32bit       = 0,
 243                .contents        = 0, /* Data, grow-up */
 244                .read_exec_only  = 0,
 245                .limit_in_pages  = 0,
 246                .seg_not_present = 0,
 247                .useable         = 0
 248        };
 249
 250        if (syscall(SYS_set_thread_area, &gdt_data16_desc) == 0) {
 251                /*
 252                 * This probably indicates vulnerability to CVE-2014-8133.
 253                 * Merely getting here isn't definitive, though, and we'll
 254                 * diagnose the problem for real later on.
 255                 */
 256                printf("[WARN]\tset_thread_area allocated data16 at index %d\n",
 257                       gdt_data16_desc.entry_number);
 258                gdt_data16_idx = gdt_data16_desc.entry_number;
 259        } else {
 260                printf("[OK]\tset_thread_area refused 16-bit data\n");
 261        }
 262
 263        struct user_desc gdt_npdata32_desc = {
 264                .entry_number    = -1,
 265                .base_addr       = (unsigned long)stack16,
 266                .limit           = 0xffff,
 267                .seg_32bit       = 1,
 268                .contents        = 0, /* Data, grow-up */
 269                .read_exec_only  = 0,
 270                .limit_in_pages  = 0,
 271                .seg_not_present = 1,
 272                .useable         = 0
 273        };
 274
 275        if (syscall(SYS_set_thread_area, &gdt_npdata32_desc) == 0) {
 276                /*
 277                 * As a hardening measure, newer kernels don't allow this.
 278                 */
 279                printf("[WARN]\tset_thread_area allocated npdata32 at index %d\n",
 280                       gdt_npdata32_desc.entry_number);
 281                gdt_npdata32_idx = gdt_npdata32_desc.entry_number;
 282        } else {
 283                printf("[OK]\tset_thread_area refused 16-bit data\n");
 284        }
 285}
 286
 287/* State used by our signal handlers. */
 288static gregset_t initial_regs, requested_regs, resulting_regs;
 289
 290/* Instructions for the SIGUSR1 handler. */
 291static volatile unsigned short sig_cs, sig_ss;
 292static volatile sig_atomic_t sig_trapped, sig_err, sig_trapno;
 293#ifdef __x86_64__
 294static volatile sig_atomic_t sig_corrupt_final_ss;
 295#endif
 296
 297/* Abstractions for some 32-bit vs 64-bit differences. */
 298#ifdef __x86_64__
 299# define REG_IP REG_RIP
 300# define REG_SP REG_RSP
 301# define REG_CX REG_RCX
 302
 303struct selectors {
 304        unsigned short cs, gs, fs, ss;
 305};
 306
 307static unsigned short *ssptr(ucontext_t *ctx)
 308{
 309        struct selectors *sels = (void *)&ctx->uc_mcontext.gregs[REG_CSGSFS];
 310        return &sels->ss;
 311}
 312
 313static unsigned short *csptr(ucontext_t *ctx)
 314{
 315        struct selectors *sels = (void *)&ctx->uc_mcontext.gregs[REG_CSGSFS];
 316        return &sels->cs;
 317}
 318#else
 319# define REG_IP REG_EIP
 320# define REG_SP REG_ESP
 321# define REG_CX REG_ECX
 322
 323static greg_t *ssptr(ucontext_t *ctx)
 324{
 325        return &ctx->uc_mcontext.gregs[REG_SS];
 326}
 327
 328static greg_t *csptr(ucontext_t *ctx)
 329{
 330        return &ctx->uc_mcontext.gregs[REG_CS];
 331}
 332#endif
 333
 334/*
 335 * Checks a given selector for its code bitness or returns -1 if it's not
 336 * a usable code segment selector.
 337 */
 338int cs_bitness(unsigned short cs)
 339{
 340        uint32_t valid = 0, ar;
 341        asm ("lar %[cs], %[ar]\n\t"
 342             "jnz 1f\n\t"
 343             "mov $1, %[valid]\n\t"
 344             "1:"
 345             : [ar] "=r" (ar), [valid] "+rm" (valid)
 346             : [cs] "r" (cs));
 347
 348        if (!valid)
 349                return -1;
 350
 351        bool db = (ar & (1 << 22));
 352        bool l = (ar & (1 << 21));
 353
 354        if (!(ar & (1<<11)))
 355            return -1;  /* Not code. */
 356
 357        if (l && !db)
 358                return 64;
 359        else if (!l && db)
 360                return 32;
 361        else if (!l && !db)
 362                return 16;
 363        else
 364                return -1;      /* Unknown bitness. */
 365}
 366
 367/*
 368 * Checks a given selector for its code bitness or returns -1 if it's not
 369 * a usable code segment selector.
 370 */
 371bool is_valid_ss(unsigned short cs)
 372{
 373        uint32_t valid = 0, ar;
 374        asm ("lar %[cs], %[ar]\n\t"
 375             "jnz 1f\n\t"
 376             "mov $1, %[valid]\n\t"
 377             "1:"
 378             : [ar] "=r" (ar), [valid] "+rm" (valid)
 379             : [cs] "r" (cs));
 380
 381        if (!valid)
 382                return false;
 383
 384        if ((ar & AR_TYPE_MASK) != AR_TYPE_RWDATA &&
 385            (ar & AR_TYPE_MASK) != AR_TYPE_RWDATA_EXPDOWN)
 386                return false;
 387
 388        return (ar & AR_P);
 389}
 390
 391/* Number of errors in the current test case. */
 392static volatile sig_atomic_t nerrs;
 393
 394static void validate_signal_ss(int sig, ucontext_t *ctx)
 395{
 396#ifdef __x86_64__
 397        bool was_64bit = (cs_bitness(*csptr(ctx)) == 64);
 398
 399        if (!(ctx->uc_flags & UC_SIGCONTEXT_SS)) {
 400                printf("[FAIL]\tUC_SIGCONTEXT_SS was not set\n");
 401                nerrs++;
 402
 403                /*
 404                 * This happens on Linux 4.1.  The rest will fail, too, so
 405                 * return now to reduce the noise.
 406                 */
 407                return;
 408        }
 409
 410        /* UC_STRICT_RESTORE_SS is set iff we came from 64-bit mode. */
 411        if (!!(ctx->uc_flags & UC_STRICT_RESTORE_SS) != was_64bit) {
 412                printf("[FAIL]\tUC_STRICT_RESTORE_SS was wrong in signal %d\n",
 413                       sig);
 414                nerrs++;
 415        }
 416
 417        if (is_valid_ss(*ssptr(ctx))) {
 418                /*
 419                 * DOSEMU was written before 64-bit sigcontext had SS, and
 420                 * it tries to figure out the signal source SS by looking at
 421                 * the physical register.  Make sure that keeps working.
 422                 */
 423                unsigned short hw_ss;
 424                asm ("mov %%ss, %0" : "=rm" (hw_ss));
 425                if (hw_ss != *ssptr(ctx)) {
 426                        printf("[FAIL]\tHW SS didn't match saved SS\n");
 427                        nerrs++;
 428                }
 429        }
 430#endif
 431}
 432
 433/*
 434 * SIGUSR1 handler.  Sets CS and SS as requested and points IP to the
 435 * int3 trampoline.  Sets SP to a large known value so that we can see
 436 * whether the value round-trips back to user mode correctly.
 437 */
 438static void sigusr1(int sig, siginfo_t *info, void *ctx_void)
 439{
 440        ucontext_t *ctx = (ucontext_t*)ctx_void;
 441
 442        validate_signal_ss(sig, ctx);
 443
 444        memcpy(&initial_regs, &ctx->uc_mcontext.gregs, sizeof(gregset_t));
 445
 446        *csptr(ctx) = sig_cs;
 447        *ssptr(ctx) = sig_ss;
 448
 449        ctx->uc_mcontext.gregs[REG_IP] =
 450                sig_cs == code16_sel ? 0 : (unsigned long)&int3;
 451        ctx->uc_mcontext.gregs[REG_SP] = (unsigned long)0x8badf00d5aadc0deULL;
 452        ctx->uc_mcontext.gregs[REG_CX] = 0;
 453
 454#ifdef __i386__
 455        /*
 456         * Make sure the kernel doesn't inadvertently use DS or ES-relative
 457         * accesses in a region where user DS or ES is loaded.
 458         *
 459         * Skip this for 64-bit builds because long mode doesn't care about
 460         * DS and ES and skipping it increases test coverage a little bit,
 461         * since 64-bit kernels can still run the 32-bit build.
 462         */
 463        ctx->uc_mcontext.gregs[REG_DS] = 0;
 464        ctx->uc_mcontext.gregs[REG_ES] = 0;
 465#endif
 466
 467        memcpy(&requested_regs, &ctx->uc_mcontext.gregs, sizeof(gregset_t));
 468        requested_regs[REG_CX] = *ssptr(ctx);   /* The asm code does this. */
 469
 470        return;
 471}
 472
 473/*
 474 * Called after a successful sigreturn (via int3) or from a failed
 475 * sigreturn (directly by kernel).  Restores our state so that the
 476 * original raise(SIGUSR1) returns.
 477 */
 478static void sigtrap(int sig, siginfo_t *info, void *ctx_void)
 479{
 480        ucontext_t *ctx = (ucontext_t*)ctx_void;
 481
 482        validate_signal_ss(sig, ctx);
 483
 484        sig_err = ctx->uc_mcontext.gregs[REG_ERR];
 485        sig_trapno = ctx->uc_mcontext.gregs[REG_TRAPNO];
 486
 487        unsigned short ss;
 488        asm ("mov %%ss,%0" : "=r" (ss));
 489
 490        greg_t asm_ss = ctx->uc_mcontext.gregs[REG_CX];
 491        if (asm_ss != sig_ss && sig == SIGTRAP) {
 492                /* Sanity check failure. */
 493                printf("[FAIL]\tSIGTRAP: ss = %hx, frame ss = %hx, ax = %llx\n",
 494                       ss, *ssptr(ctx), (unsigned long long)asm_ss);
 495                nerrs++;
 496        }
 497
 498        memcpy(&resulting_regs, &ctx->uc_mcontext.gregs, sizeof(gregset_t));
 499        memcpy(&ctx->uc_mcontext.gregs, &initial_regs, sizeof(gregset_t));
 500
 501#ifdef __x86_64__
 502        if (sig_corrupt_final_ss) {
 503                if (ctx->uc_flags & UC_STRICT_RESTORE_SS) {
 504                        printf("[FAIL]\tUC_STRICT_RESTORE_SS was set inappropriately\n");
 505                        nerrs++;
 506                } else {
 507                        /*
 508                         * DOSEMU transitions from 32-bit to 64-bit mode by
 509                         * adjusting sigcontext, and it requires that this work
 510                         * even if the saved SS is bogus.
 511                         */
 512                        printf("\tCorrupting SS on return to 64-bit mode\n");
 513                        *ssptr(ctx) = 0;
 514                }
 515        }
 516#endif
 517
 518        sig_trapped = sig;
 519}
 520
 521#ifdef __x86_64__
 522/* Tests recovery if !UC_STRICT_RESTORE_SS */
 523static void sigusr2(int sig, siginfo_t *info, void *ctx_void)
 524{
 525        ucontext_t *ctx = (ucontext_t*)ctx_void;
 526
 527        if (!(ctx->uc_flags & UC_STRICT_RESTORE_SS)) {
 528                printf("[FAIL]\traise(2) didn't set UC_STRICT_RESTORE_SS\n");
 529                nerrs++;
 530                return;  /* We can't do the rest. */
 531        }
 532
 533        ctx->uc_flags &= ~UC_STRICT_RESTORE_SS;
 534        *ssptr(ctx) = 0;
 535
 536        /* Return.  The kernel should recover without sending another signal. */
 537}
 538
 539static int test_nonstrict_ss(void)
 540{
 541        clearhandler(SIGUSR1);
 542        clearhandler(SIGTRAP);
 543        clearhandler(SIGSEGV);
 544        clearhandler(SIGILL);
 545        sethandler(SIGUSR2, sigusr2, 0);
 546
 547        nerrs = 0;
 548
 549        printf("[RUN]\tClear UC_STRICT_RESTORE_SS and corrupt SS\n");
 550        raise(SIGUSR2);
 551        if (!nerrs)
 552                printf("[OK]\tIt worked\n");
 553
 554        return nerrs;
 555}
 556#endif
 557
 558/* Finds a usable code segment of the requested bitness. */
 559int find_cs(int bitness)
 560{
 561        unsigned short my_cs;
 562
 563        asm ("mov %%cs,%0" :  "=r" (my_cs));
 564
 565        if (cs_bitness(my_cs) == bitness)
 566                return my_cs;
 567        if (cs_bitness(my_cs + (2 << 3)) == bitness)
 568                return my_cs + (2 << 3);
 569        if (my_cs > (2<<3) && cs_bitness(my_cs - (2 << 3)) == bitness)
 570            return my_cs - (2 << 3);
 571        if (cs_bitness(code16_sel) == bitness)
 572                return code16_sel;
 573
 574        printf("[WARN]\tCould not find %d-bit CS\n", bitness);
 575        return -1;
 576}
 577
 578static int test_valid_sigreturn(int cs_bits, bool use_16bit_ss, int force_ss)
 579{
 580        int cs = find_cs(cs_bits);
 581        if (cs == -1) {
 582                printf("[SKIP]\tCode segment unavailable for %d-bit CS, %d-bit SS\n",
 583                       cs_bits, use_16bit_ss ? 16 : 32);
 584                return 0;
 585        }
 586
 587        if (force_ss != -1) {
 588                sig_ss = force_ss;
 589        } else {
 590                if (use_16bit_ss) {
 591                        if (!data16_sel) {
 592                                printf("[SKIP]\tData segment unavailable for %d-bit CS, 16-bit SS\n",
 593                                       cs_bits);
 594                                return 0;
 595                        }
 596                        sig_ss = data16_sel;
 597                } else {
 598                        asm volatile ("mov %%ss,%0" : "=r" (sig_ss));
 599                }
 600        }
 601
 602        sig_cs = cs;
 603
 604        printf("[RUN]\tValid sigreturn: %d-bit CS (%hx), %d-bit SS (%hx%s)\n",
 605               cs_bits, sig_cs, use_16bit_ss ? 16 : 32, sig_ss,
 606               (sig_ss & 4) ? "" : ", GDT");
 607
 608        raise(SIGUSR1);
 609
 610        nerrs = 0;
 611
 612        /*
 613         * Check that each register had an acceptable value when the
 614         * int3 trampoline was invoked.
 615         */
 616        for (int i = 0; i < NGREG; i++) {
 617                greg_t req = requested_regs[i], res = resulting_regs[i];
 618
 619                if (i == REG_TRAPNO || i == REG_IP)
 620                        continue;       /* don't care */
 621
 622                if (i == REG_SP) {
 623                        /*
 624                         * If we were using a 16-bit stack segment, then
 625                         * the kernel is a bit stuck: IRET only restores
 626                         * the low 16 bits of ESP/RSP if SS is 16-bit.
 627                         * The kernel uses a hack to restore bits 31:16,
 628                         * but that hack doesn't help with bits 63:32.
 629                         * On Intel CPUs, bits 63:32 end up zeroed, and, on
 630                         * AMD CPUs, they leak the high bits of the kernel
 631                         * espfix64 stack pointer.  There's very little that
 632                         * the kernel can do about it.
 633                         *
 634                         * Similarly, if we are returning to a 32-bit context,
 635                         * the CPU will often lose the high 32 bits of RSP.
 636                         */
 637
 638                        if (res == req)
 639                                continue;
 640
 641                        if (cs_bits != 64 && ((res ^ req) & 0xFFFFFFFF) == 0) {
 642                                printf("[NOTE]\tSP: %llx -> %llx\n",
 643                                       (unsigned long long)req,
 644                                       (unsigned long long)res);
 645                                continue;
 646                        }
 647
 648                        printf("[FAIL]\tSP mismatch: requested 0x%llx; got 0x%llx\n",
 649                               (unsigned long long)requested_regs[i],
 650                               (unsigned long long)resulting_regs[i]);
 651                        nerrs++;
 652                        continue;
 653                }
 654
 655                bool ignore_reg = false;
 656#if __i386__
 657                if (i == REG_UESP)
 658                        ignore_reg = true;
 659#else
 660                if (i == REG_CSGSFS) {
 661                        struct selectors *req_sels =
 662                                (void *)&requested_regs[REG_CSGSFS];
 663                        struct selectors *res_sels =
 664                                (void *)&resulting_regs[REG_CSGSFS];
 665                        if (req_sels->cs != res_sels->cs) {
 666                                printf("[FAIL]\tCS mismatch: requested 0x%hx; got 0x%hx\n",
 667                                       req_sels->cs, res_sels->cs);
 668                                nerrs++;
 669                        }
 670
 671                        if (req_sels->ss != res_sels->ss) {
 672                                printf("[FAIL]\tSS mismatch: requested 0x%hx; got 0x%hx\n",
 673                                       req_sels->ss, res_sels->ss);
 674                                nerrs++;
 675                        }
 676
 677                        continue;
 678                }
 679#endif
 680
 681                /* Sanity check on the kernel */
 682                if (i == REG_CX && req != res) {
 683                        printf("[FAIL]\tCX (saved SP) mismatch: requested 0x%llx; got 0x%llx\n",
 684                               (unsigned long long)req,
 685                               (unsigned long long)res);
 686                        nerrs++;
 687                        continue;
 688                }
 689
 690                if (req != res && !ignore_reg) {
 691                        printf("[FAIL]\tReg %d mismatch: requested 0x%llx; got 0x%llx\n",
 692                               i, (unsigned long long)req,
 693                               (unsigned long long)res);
 694                        nerrs++;
 695                }
 696        }
 697
 698        if (nerrs == 0)
 699                printf("[OK]\tall registers okay\n");
 700
 701        return nerrs;
 702}
 703
 704static int test_bad_iret(int cs_bits, unsigned short ss, int force_cs)
 705{
 706        int cs = force_cs == -1 ? find_cs(cs_bits) : force_cs;
 707        if (cs == -1)
 708                return 0;
 709
 710        sig_cs = cs;
 711        sig_ss = ss;
 712
 713        printf("[RUN]\t%d-bit CS (%hx), bogus SS (%hx)\n",
 714               cs_bits, sig_cs, sig_ss);
 715
 716        sig_trapped = 0;
 717        raise(SIGUSR1);
 718        if (sig_trapped) {
 719                char errdesc[32] = "";
 720                if (sig_err) {
 721                        const char *src = (sig_err & 1) ? " EXT" : "";
 722                        const char *table;
 723                        if ((sig_err & 0x6) == 0x0)
 724                                table = "GDT";
 725                        else if ((sig_err & 0x6) == 0x4)
 726                                table = "LDT";
 727                        else if ((sig_err & 0x6) == 0x2)
 728                                table = "IDT";
 729                        else
 730                                table = "???";
 731
 732                        sprintf(errdesc, "%s%s index %d, ",
 733                                table, src, sig_err >> 3);
 734                }
 735
 736                char trapname[32];
 737                if (sig_trapno == 13)
 738                        strcpy(trapname, "GP");
 739                else if (sig_trapno == 11)
 740                        strcpy(trapname, "NP");
 741                else if (sig_trapno == 12)
 742                        strcpy(trapname, "SS");
 743                else if (sig_trapno == 32)
 744                        strcpy(trapname, "IRET");  /* X86_TRAP_IRET */
 745                else
 746                        sprintf(trapname, "%d", sig_trapno);
 747
 748                printf("[OK]\tGot #%s(0x%lx) (i.e. %s%s)\n",
 749                       trapname, (unsigned long)sig_err,
 750                       errdesc, strsignal(sig_trapped));
 751                return 0;
 752        } else {
 753                /*
 754                 * This also implicitly tests UC_STRICT_RESTORE_SS:
 755                 * We check that these signals set UC_STRICT_RESTORE_SS and,
 756                 * if UC_STRICT_RESTORE_SS doesn't cause strict behavior,
 757                 * then we won't get SIGSEGV.
 758                 */
 759                printf("[FAIL]\tDid not get SIGSEGV\n");
 760                return 1;
 761        }
 762}
 763
 764int main()
 765{
 766        int total_nerrs = 0;
 767        unsigned short my_cs, my_ss;
 768
 769        asm volatile ("mov %%cs,%0" : "=r" (my_cs));
 770        asm volatile ("mov %%ss,%0" : "=r" (my_ss));
 771        setup_ldt();
 772
 773        stack_t stack = {
 774                .ss_sp = altstack_data,
 775                .ss_size = SIGSTKSZ,
 776        };
 777        if (sigaltstack(&stack, NULL) != 0)
 778                err(1, "sigaltstack");
 779
 780        sethandler(SIGUSR1, sigusr1, 0);
 781        sethandler(SIGTRAP, sigtrap, SA_ONSTACK);
 782
 783        /* Easy cases: return to a 32-bit SS in each possible CS bitness. */
 784        total_nerrs += test_valid_sigreturn(64, false, -1);
 785        total_nerrs += test_valid_sigreturn(32, false, -1);
 786        total_nerrs += test_valid_sigreturn(16, false, -1);
 787
 788        /*
 789         * Test easy espfix cases: return to a 16-bit LDT SS in each possible
 790         * CS bitness.  NB: with a long mode CS, the SS bitness is irrelevant.
 791         *
 792         * This catches the original missing-espfix-on-64-bit-kernels issue
 793         * as well as CVE-2014-8134.
 794         */
 795        total_nerrs += test_valid_sigreturn(64, true, -1);
 796        total_nerrs += test_valid_sigreturn(32, true, -1);
 797        total_nerrs += test_valid_sigreturn(16, true, -1);
 798
 799        if (gdt_data16_idx) {
 800                /*
 801                 * For performance reasons, Linux skips espfix if SS points
 802                 * to the GDT.  If we were able to allocate a 16-bit SS in
 803                 * the GDT, see if it leaks parts of the kernel stack pointer.
 804                 *
 805                 * This tests for CVE-2014-8133.
 806                 */
 807                total_nerrs += test_valid_sigreturn(64, true,
 808                                                    GDT3(gdt_data16_idx));
 809                total_nerrs += test_valid_sigreturn(32, true,
 810                                                    GDT3(gdt_data16_idx));
 811                total_nerrs += test_valid_sigreturn(16, true,
 812                                                    GDT3(gdt_data16_idx));
 813        }
 814
 815#ifdef __x86_64__
 816        /* Nasty ABI case: check SS corruption handling. */
 817        sig_corrupt_final_ss = 1;
 818        total_nerrs += test_valid_sigreturn(32, false, -1);
 819        total_nerrs += test_valid_sigreturn(32, true, -1);
 820        sig_corrupt_final_ss = 0;
 821#endif
 822
 823        /*
 824         * We're done testing valid sigreturn cases.  Now we test states
 825         * for which sigreturn itself will succeed but the subsequent
 826         * entry to user mode will fail.
 827         *
 828         * Depending on the failure mode and the kernel bitness, these
 829         * entry failures can generate SIGSEGV, SIGBUS, or SIGILL.
 830         */
 831        clearhandler(SIGTRAP);
 832        sethandler(SIGSEGV, sigtrap, SA_ONSTACK);
 833        sethandler(SIGBUS, sigtrap, SA_ONSTACK);
 834        sethandler(SIGILL, sigtrap, SA_ONSTACK);  /* 32-bit kernels do this */
 835
 836        /* Easy failures: invalid SS, resulting in #GP(0) */
 837        test_bad_iret(64, ldt_nonexistent_sel, -1);
 838        test_bad_iret(32, ldt_nonexistent_sel, -1);
 839        test_bad_iret(16, ldt_nonexistent_sel, -1);
 840
 841        /* These fail because SS isn't a data segment, resulting in #GP(SS) */
 842        test_bad_iret(64, my_cs, -1);
 843        test_bad_iret(32, my_cs, -1);
 844        test_bad_iret(16, my_cs, -1);
 845
 846        /* Try to return to a not-present code segment, triggering #NP(SS). */
 847        test_bad_iret(32, my_ss, npcode32_sel);
 848
 849        /*
 850         * Try to return to a not-present but otherwise valid data segment.
 851         * This will cause IRET to fail with #SS on the espfix stack.  This
 852         * exercises CVE-2014-9322.
 853         *
 854         * Note that, if espfix is enabled, 64-bit Linux will lose track
 855         * of the actual cause of failure and report #GP(0) instead.
 856         * This would be very difficult for Linux to avoid, because
 857         * espfix64 causes IRET failures to be promoted to #DF, so the
 858         * original exception frame is never pushed onto the stack.
 859         */
 860        test_bad_iret(32, npdata32_sel, -1);
 861
 862        /*
 863         * Try to return to a not-present but otherwise valid data
 864         * segment without invoking espfix.  Newer kernels don't allow
 865         * this to happen in the first place.  On older kernels, though,
 866         * this can trigger CVE-2014-9322.
 867         */
 868        if (gdt_npdata32_idx)
 869                test_bad_iret(32, GDT3(gdt_npdata32_idx), -1);
 870
 871#ifdef __x86_64__
 872        total_nerrs += test_nonstrict_ss();
 873#endif
 874
 875        return total_nerrs ? 1 : 0;
 876}
 877