kernel/seccomp.c

   1 /*
   2  * linux/kernel/seccomp.c
   3  *
   4  * Copyright 2004-2005  Andrea Arcangeli <andrea@cpushare.com>
   5  *
   6  * Copyright (C) 2012 Google, Inc.
   7  * Will Drewry <wad@chromium.org>
   8  *
   9  * This defines a simple but solid secure-computing facility.
  10  *
  11  * Mode 1 uses a fixed list of allowed system calls.
  12  * Mode 2 allows user-defined system call filters in the form
  13  *        of Berkeley Packet Filters/Linux Socket Filters.
  14  */
  15
  16 #include <linux/refcount.h>
  17 #include <linux/audit.h>
  18 #include <linux/compat.h>
  19 #include <linux/coredump.h>
  20 #include <linux/sched.h>
  21 #include <linux/sched/task_stack.h>
  22 #include <linux/seccomp.h>
  23 #include <linux/slab.h>
  24 #include <linux/syscalls.h>
  25
  26 #ifdef CONFIG_HAVE_ARCH_SECCOMP_FILTER
  27 #include <asm/syscall.h>
  28 #endif
  29
  30 #ifdef CONFIG_SECCOMP_FILTER
  31 #include <linux/filter.h>
  32 #include <linux/pid.h>
  33 #include <linux/ptrace.h>
  34 #include <linux/security.h>
  35 #include <linux/tracehook.h>
  36 #include <linux/uaccess.h>
  37
  38 /**
  39  * struct seccomp_filter - container for seccomp BPF programs
  40  *
  41  * @usage: reference count to manage the object lifetime.
  42  *         get/put helpers should be used when accessing an instance
  43  *         outside of a lifetime-guarded section.  In general, this
  44  *         is only needed for handling filters shared across tasks.
  45  * @prev: points to a previously installed, or inherited, filter
  46  * @prog: the BPF program to evaluate
  47  *
  48  * seccomp_filter objects are organized in a tree linked via the @prev
  49  * pointer.  For any task, it appears to be a singly-linked list starting
  50  * with current->seccomp.filter, the most recently attached or inherited filter.
  51  * However, multiple filters may share a @prev node, by way of fork(), which
  52  * results in a unidirectional tree existing in memory.  This is similar to
  53  * how namespaces work.
  54  *
  55  * seccomp_filter objects should never be modified after being attached
  56  * to a task_struct (other than @usage).
  57  */
  58 struct seccomp_filter {
  59         refcount_t usage;
  60         struct seccomp_filter *prev;
  61         struct bpf_prog *prog;
  62 };
  63
  64 /* Limit any path through the tree to 256KB worth of instructions. */
  65 #define MAX_INSNS_PER_PATH ((1 << 18) / sizeof(struct sock_filter))
  66
  67 /*
  68  * Endianness is explicitly ignored and left for BPF program authors to manage
  69  * as per the specific architecture.
  70  */
  71 static void populate_seccomp_data(struct seccomp_data *sd)
  72 {
  73         struct task_struct *task = current;
  74         struct pt_regs *regs = task_pt_regs(task);
  75         unsigned long args[6];
  76
  77         sd->nr = syscall_get_nr(task, regs);
  78         sd->arch = syscall_get_arch();
  79         syscall_get_arguments(task, regs, 0, 6, args);
  80         sd->args[0] = args[0];
  81         sd->args[1] = args[1];
  82         sd->args[2] = args[2];
  83         sd->args[3] = args[3];
  84         sd->args[4] = args[4];
  85         sd->args[5] = args[5];
  86         sd->instruction_pointer = KSTK_EIP(task);
  87 }
  88
  89 /**
  90  *      seccomp_check_filter - verify seccomp filter code
  91  *      @filter: filter to verify
  92  *      @flen: length of filter
  93  *
  94  * Takes a previously checked filter (by bpf_check_classic) and
  95  * redirects all filter code that loads struct sk_buff data
  96  * and related data through seccomp_bpf_load.  It also
  97  * enforces length and alignment checking of those loads.
  98  *
  99  * Returns 0 if the rule set is legal or -EINVAL if not.
 100  */
 101 static int seccomp_check_filter(struct sock_filter *filter, unsigned int flen)
 102 {
 103         int pc;
 104         for (pc = 0; pc < flen; pc++) {
 105                 struct sock_filter *ftest = &filter[pc];
 106                 u16 code = ftest->code;
 107                 u32 k = ftest->k;
 108
 109                 switch (code) {
 110                 case BPF_LD | BPF_W | BPF_ABS:
 111                         ftest->code = BPF_LDX | BPF_W | BPF_ABS;
 112                         /* 32-bit aligned and not out of bounds. */
 113                         if (k >= sizeof(struct seccomp_data) || k & 3)
 114                                 return -EINVAL;
 115                         continue;
 116                 case BPF_LD | BPF_W | BPF_LEN:
 117                         ftest->code = BPF_LD | BPF_IMM;
 118                         ftest->k = sizeof(struct seccomp_data);
 119                         continue;
 120                 case BPF_LDX | BPF_W | BPF_LEN:
 121                         ftest->code = BPF_LDX | BPF_IMM;
 122                         ftest->k = sizeof(struct seccomp_data);
 123                         continue;
 124                 /* Explicitly include allowed calls. */
 125                 case BPF_RET | BPF_K:
 126                 case BPF_RET | BPF_A:
 127                 case BPF_ALU | BPF_ADD | BPF_K:
 128                 case BPF_ALU | BPF_ADD | BPF_X:
 129                 case BPF_ALU | BPF_SUB | BPF_K:
 130                 case BPF_ALU | BPF_SUB | BPF_X:
 131                 case BPF_ALU | BPF_MUL | BPF_K:
 132                 case BPF_ALU | BPF_MUL | BPF_X:
 133                 case BPF_ALU | BPF_DIV | BPF_K:
 134                 case BPF_ALU | BPF_DIV | BPF_X:
 135                 case BPF_ALU | BPF_AND | BPF_K:
 136                 case BPF_ALU | BPF_AND | BPF_X:
 137                 case BPF_ALU | BPF_OR | BPF_K:
 138                 case BPF_ALU | BPF_OR | BPF_X:
 139                 case BPF_ALU | BPF_XOR | BPF_K:
 140                 case BPF_ALU | BPF_XOR | BPF_X:
 141                 case BPF_ALU | BPF_LSH | BPF_K:
 142                 case BPF_ALU | BPF_LSH | BPF_X:
 143                 case BPF_ALU | BPF_RSH | BPF_K:
 144                 case BPF_ALU | BPF_RSH | BPF_X:
 145                 case BPF_ALU | BPF_NEG:
 146                 case BPF_LD | BPF_IMM:
 147                 case BPF_LDX | BPF_IMM:
 148                 case BPF_MISC | BPF_TAX:
 149                 case BPF_MISC | BPF_TXA:
 150                 case BPF_LD | BPF_MEM:
 151                 case BPF_LDX | BPF_MEM:
 152                 case BPF_ST:
 153                 case BPF_STX:
 154                 case BPF_JMP | BPF_JA:
 155                 case BPF_JMP | BPF_JEQ | BPF_K:
 156                 case BPF_JMP | BPF_JEQ | BPF_X:
 157                 case BPF_JMP | BPF_JGE | BPF_K:
 158                 case BPF_JMP | BPF_JGE | BPF_X:
 159                 case BPF_JMP | BPF_JGT | BPF_K:
 160                 case BPF_JMP | BPF_JGT | BPF_X:
 161                 case BPF_JMP | BPF_JSET | BPF_K:
 162                 case BPF_JMP | BPF_JSET | BPF_X:
 163                         continue;
 164                 default:
 165                         return -EINVAL;
 166                 }
 167         }
 168         return 0;
 169 }
 170
 171 /**
 172  * seccomp_run_filters - evaluates all seccomp filters against @sd
 173  * @sd: optional seccomp data to be passed to filters
 174  * @match: stores struct seccomp_filter that resulted in the return value,
 175  *         unless filter returned SECCOMP_RET_ALLOW, in which case it will
 176  *         be unchanged.
 177  *
 178  * Returns valid seccomp BPF response codes.
 179  */
 180 static u32 seccomp_run_filters(const struct seccomp_data *sd,
 181                                struct seccomp_filter **match)
 182 {
 183         struct seccomp_data sd_local;
 184         u32 ret = SECCOMP_RET_ALLOW;
 185         /* Make sure cross-thread synced filter points somewhere sane. */
 186         struct seccomp_filter *f =
 187                         lockless_dereference(current->seccomp.filter);
 188
 189         /* Ensure unexpected behavior doesn't result in failing open. */
 190         if (unlikely(WARN_ON(f == NULL)))
 191                 return SECCOMP_RET_KILL;
 192
 193         if (!sd) {
 194                 populate_seccomp_data(&sd_local);
 195                 sd = &sd_local;
 196         }
 197
 198         /*
 199          * All filters in the list are evaluated and the lowest BPF return
 200          * value always takes priority (ignoring the DATA).
 201          */
 202         for (; f; f = f->prev) {
 203                 u32 cur_ret = BPF_PROG_RUN(f->prog, sd);
 204
 205                 if ((cur_ret & SECCOMP_RET_ACTION) < (ret & SECCOMP_RET_ACTION)) {
 206                         ret = cur_ret;
 207                         *match = f;
 208                 }
 209         }
 210         return ret;
 211 }
 212 #endif /* CONFIG_SECCOMP_FILTER */
 213
 214 static inline bool seccomp_may_assign_mode(unsigned long seccomp_mode)
 215 {
 216         assert_spin_locked(&current->sighand->siglock);
 217
 218         if (current->seccomp.mode && current->seccomp.mode != seccomp_mode)
 219                 return false;
 220
 221         return true;
 222 }
 223
 224 static inline void seccomp_assign_mode(struct task_struct *task,
 225                                        unsigned long seccomp_mode)
 226 {
 227         assert_spin_locked(&task->sighand->siglock);
 228
 229         task->seccomp.mode = seccomp_mode;
 230         /*
 231          * Make sure TIF_SECCOMP cannot be set before the mode (and
 232          * filter) is set.
 233          */
 234         smp_mb__before_atomic();
 235         set_tsk_thread_flag(task, TIF_SECCOMP);
 236 }
 237
 238 #ifdef CONFIG_SECCOMP_FILTER
 239 /* Returns 1 if the parent is an ancestor of the child. */
 240 static int is_ancestor(struct seccomp_filter *parent,
 241                        struct seccomp_filter *child)
 242 {
 243         /* NULL is the root ancestor. */
 244         if (parent == NULL)
 245                 return 1;
 246         for (; child; child = child->prev)
 247                 if (child == parent)
 248                         return 1;
 249         return 0;
 250 }
 251
 252 /**
 253  * seccomp_can_sync_threads: checks if all threads can be synchronized
 254  *
 255  * Expects sighand and cred_guard_mutex locks to be held.
 256  *
 257  * Returns 0 on success, -ve on error, or the pid of a thread which was
 258  * either not in the correct seccomp mode or it did not have an ancestral
 259  * seccomp filter.
 260  */
 261 static inline pid_t seccomp_can_sync_threads(void)
 262 {
 263         struct task_struct *thread, *caller;
 264
 265         BUG_ON(!mutex_is_locked(&current->signal->cred_guard_mutex));
 266         assert_spin_locked(&current->sighand->siglock);
 267
 268         /* Validate all threads being eligible for synchronization. */
 269         caller = current;
 270         for_each_thread(caller, thread) {
 271                 pid_t failed;
 272
 273                 /* Skip current, since it is initiating the sync. */
 274                 if (thread == caller)
 275                         continue;
 276
 277                 if (thread->seccomp.mode == SECCOMP_MODE_DISABLED ||
 278                     (thread->seccomp.mode == SECCOMP_MODE_FILTER &&
 279                      is_ancestor(thread->seccomp.filter,
 280                                  caller->seccomp.filter)))
 281                         continue;
 282
 283                 /* Return the first thread that cannot be synchronized. */
 284                 failed = task_pid_vnr(thread);
 285                 /* If the pid cannot be resolved, then return -ESRCH */
 286                 if (unlikely(WARN_ON(failed == 0)))
 287                         failed = -ESRCH;
 288                 return failed;
 289         }
 290
 291         return 0;
 292 }
 293
 294 /**
 295  * seccomp_sync_threads: sets all threads to use current's filter
 296  *
 297  * Expects sighand and cred_guard_mutex locks to be held, and for
 298  * seccomp_can_sync_threads() to have returned success already
 299  * without dropping the locks.
 300  *
 301  */
 302 static inline void seccomp_sync_threads(void)
 303 {
 304         struct task_struct *thread, *caller;
 305
 306         BUG_ON(!mutex_is_locked(&current->signal->cred_guard_mutex));
 307         assert_spin_locked(&current->sighand->siglock);
 308
 309         /* Synchronize all threads. */
 310         caller = current;
 311         for_each_thread(caller, thread) {
 312                 /* Skip current, since it needs no changes. */
 313                 if (thread == caller)
 314                         continue;
 315
 316                 /* Get a task reference for the new leaf node. */
 317                 get_seccomp_filter(caller);
 318                 /*
 319                  * Drop the task reference to the shared ancestor since
 320                  * current's path will hold a reference.  (This also
 321                  * allows a put before the assignment.)
 322                  */
 323                 put_seccomp_filter(thread);
 324                 smp_store_release(&thread->seccomp.filter,
 325                                   caller->seccomp.filter);
 326
 327                 /*
 328                  * Don't let an unprivileged task work around
 329                  * the no_new_privs restriction by creating
 330                  * a thread that sets it up, enters seccomp,
 331                  * then dies.
 332                  */
 333                 if (task_no_new_privs(caller))
 334                         task_set_no_new_privs(thread);
 335
 336                 /*
 337                  * Opt the other thread into seccomp if needed.
 338                  * As threads are considered to be trust-realm
 339                  * equivalent (see ptrace_may_access), it is safe to
 340                  * allow one thread to transition the other.
 341                  */
 342                 if (thread->seccomp.mode == SECCOMP_MODE_DISABLED)
 343                         seccomp_assign_mode(thread, SECCOMP_MODE_FILTER);
 344         }
 345 }
 346
 347 /**
 348  * seccomp_prepare_filter: Prepares a seccomp filter for use.
 349  * @fprog: BPF program to install
 350  *
 351  * Returns filter on success or an ERR_PTR on failure.
 352  */
 353 static struct seccomp_filter *seccomp_prepare_filter(struct sock_fprog *fprog)
 354 {
 355         struct seccomp_filter *sfilter;
 356         int ret;
 357         const bool save_orig = IS_ENABLED(CONFIG_CHECKPOINT_RESTORE);
 358
 359         if (fprog->len == 0 || fprog->len > BPF_MAXINSNS)
 360                 return ERR_PTR(-EINVAL);
 361
 362         BUG_ON(INT_MAX / fprog->len < sizeof(struct sock_filter));
 363
 364         /*
 365          * Installing a seccomp filter requires that the task has
 366          * CAP_SYS_ADMIN in its namespace or be running with no_new_privs.
 367          * This avoids scenarios where unprivileged tasks can affect the
 368          * behavior of privileged children.
 369          */
 370         if (!task_no_new_privs(current) &&
 371             security_capable_noaudit(current_cred(), current_user_ns(),
 372                                      CAP_SYS_ADMIN) != 0)
 373                 return ERR_PTR(-EACCES);
 374
 375         /* Allocate a new seccomp_filter */
 376         sfilter = kzalloc(sizeof(*sfilter), GFP_KERNEL | __GFP_NOWARN);
 377         if (!sfilter)
 378                 return ERR_PTR(-ENOMEM);
 379
 380         ret = bpf_prog_create_from_user(&sfilter->prog, fprog,
 381                                         seccomp_check_filter, save_orig);
 382         if (ret < 0) {
 383                 kfree(sfilter);
 384                 return ERR_PTR(ret);
 385         }
 386
 387         refcount_set(&sfilter->usage, 1);
 388
 389         return sfilter;
 390 }
 391
 392 /**
 393  * seccomp_prepare_user_filter - prepares a user-supplied sock_fprog
 394  * @user_filter: pointer to the user data containing a sock_fprog.
 395  *
 396  * Returns 0 on success and non-zero otherwise.
 397  */
 398 static struct seccomp_filter *
 399 seccomp_prepare_user_filter(const char __user *user_filter)
 400 {
 401         struct sock_fprog fprog;
 402         struct seccomp_filter *filter = ERR_PTR(-EFAULT);
 403
 404 #ifdef CONFIG_COMPAT
 405         if (in_compat_syscall()) {
 406                 struct compat_sock_fprog fprog32;
 407                 if (copy_from_user(&fprog32, user_filter, sizeof(fprog32)))
 408                         goto out;
 409                 fprog.len = fprog32.len;
 410                 fprog.filter = compat_ptr(fprog32.filter);
 411         } else /* falls through to the if below. */
 412 #endif
 413         if (copy_from_user(&fprog, user_filter, sizeof(fprog)))
 414                 goto out;
 415         filter = seccomp_prepare_filter(&fprog);
 416 out:
 417         return filter;
 418 }
 419
 420 /**
 421  * seccomp_attach_filter: validate and attach filter
 422  * @flags:  flags to change filter behavior
 423  * @filter: seccomp filter to add to the current process
 424  *
 425  * Caller must be holding current->sighand->siglock lock.
 426  *
 427  * Returns 0 on success, -ve on error.
 428  */
 429 static long seccomp_attach_filter(unsigned int flags,
 430                                   struct seccomp_filter *filter)
 431 {
 432         unsigned long total_insns;
 433         struct seccomp_filter *walker;
 434
 435         assert_spin_locked(&current->sighand->siglock);
 436
 437         /* Validate resulting filter length. */
 438         total_insns = filter->prog->len;
 439         for (walker = current->seccomp.filter; walker; walker = walker->prev)
 440                 total_insns += walker->prog->len + 4;  /* 4 instr penalty */
 441         if (total_insns > MAX_INSNS_PER_PATH)
 442                 return -ENOMEM;
 443
 444         /* If thread sync has been requested, check that it is possible. */
 445         if (flags & SECCOMP_FILTER_FLAG_TSYNC) {
 446                 int ret;
 447
 448                 ret = seccomp_can_sync_threads();
 449                 if (ret)
 450                         return ret;
 451         }
 452
 453         /*
 454          * If there is an existing filter, make it the prev and don't drop its
 455          * task reference.
 456          */
 457         filter->prev = current->seccomp.filter;
 458         current->seccomp.filter = filter;
 459
 460         /* Now that the new filter is in place, synchronize to all threads. */
 461         if (flags & SECCOMP_FILTER_FLAG_TSYNC)
 462                 seccomp_sync_threads();
 463
 464         return 0;
 465 }
 466
 467 /* get_seccomp_filter - increments the reference count of the filter on @tsk */
 468 void get_seccomp_filter(struct task_struct *tsk)
 469 {
 470         struct seccomp_filter *orig = tsk->seccomp.filter;
 471         if (!orig)
 472                 return;
 473         /* Reference count is bounded by the number of total processes. */
 474         refcount_inc(&orig->usage);
 475 }
 476
 477 static inline void seccomp_filter_free(struct seccomp_filter *filter)
 478 {
 479         if (filter) {
 480                 bpf_prog_destroy(filter->prog);
 481                 kfree(filter);
 482         }
 483 }
 484
 485 /* put_seccomp_filter - decrements the ref count of tsk->seccomp.filter */
 486 void put_seccomp_filter(struct task_struct *tsk)
 487 {
 488         struct seccomp_filter *orig = tsk->seccomp.filter;
 489         /* Clean up single-reference branches iteratively. */
 490         while (orig && refcount_dec_and_test(&orig->usage)) {
 491                 struct seccomp_filter *freeme = orig;
 492                 orig = orig->prev;
 493                 seccomp_filter_free(freeme);
 494         }
 495 }
 496
 497 static void seccomp_init_siginfo(siginfo_t *info, int syscall, int reason)
 498 {
 499         memset(info, 0, sizeof(*info));
 500         info->si_signo = SIGSYS;
 501         info->si_code = SYS_SECCOMP;
 502         info->si_call_addr = (void __user *)KSTK_EIP(current);
 503         info->si_errno = reason;
 504         info->si_arch = syscall_get_arch();
 505         info->si_syscall = syscall;
 506 }
 507
 508 /**
 509  * seccomp_send_sigsys - signals the task to allow in-process syscall emulation
 510  * @syscall: syscall number to send to userland
 511  * @reason: filter-supplied reason code to send to userland (via si_errno)
 512  *
 513  * Forces a SIGSYS with a code of SYS_SECCOMP and related sigsys info.
 514  */
 515 static void seccomp_send_sigsys(int syscall, int reason)
 516 {
 517         struct siginfo info;
 518         seccomp_init_siginfo(&info, syscall, reason);
 519         force_sig_info(SIGSYS, &info, current);
 520 }
 521 #endif  /* CONFIG_SECCOMP_FILTER */
 522
 523 /*
 524  * Secure computing mode 1 allows only read/write/exit/sigreturn.
 525  * To be fully secure this must be combined with rlimit
 526  * to limit the stack allocations too.
 527  */
 528 static const int mode1_syscalls[] = {
 529         __NR_seccomp_read, __NR_seccomp_write, __NR_seccomp_exit, __NR_seccomp_sigreturn,
 530         0, /* null terminated */
 531 };
 532
 533 static void __secure_computing_strict(int this_syscall)
 534 {
 535         const int *syscall_whitelist = mode1_syscalls;
 536 #ifdef CONFIG_COMPAT
 537         if (in_compat_syscall())
 538                 syscall_whitelist = get_compat_mode1_syscalls();
 539 #endif
 540         do {
 541                 if (*syscall_whitelist == this_syscall)
 542                         return;
 543         } while (*++syscall_whitelist);
 544
 545 #ifdef SECCOMP_DEBUG
 546         dump_stack();
 547 #endif
 548         audit_seccomp(this_syscall, SIGKILL, SECCOMP_RET_KILL);
 549         do_exit(SIGKILL);
 550 }
 551
 552 #ifndef CONFIG_HAVE_ARCH_SECCOMP_FILTER
 553 void secure_computing_strict(int this_syscall)
 554 {
 555         int mode = current->seccomp.mode;
 556
 557         if (IS_ENABLED(CONFIG_CHECKPOINT_RESTORE) &&
 558             unlikely(current->ptrace & PT_SUSPEND_SECCOMP))
 559                 return;
 560
 561         if (mode == SECCOMP_MODE_DISABLED)
 562                 return;
 563         else if (mode == SECCOMP_MODE_STRICT)
 564                 __secure_computing_strict(this_syscall);
 565         else
 566                 BUG();
 567 }
 568 #else
 569
 570 #ifdef CONFIG_SECCOMP_FILTER
 571 static int __seccomp_filter(int this_syscall, const struct seccomp_data *sd,
 572                             const bool recheck_after_trace)
 573 {
 574         u32 filter_ret, action;
 575         struct seccomp_filter *match = NULL;
 576         int data;
 577
 578         /*
 579          * Make sure that any changes to mode from another thread have
 580          * been seen after TIF_SECCOMP was seen.
 581          */
 582         rmb();
 583
 584         filter_ret = seccomp_run_filters(sd, &match);
 585         data = filter_ret & SECCOMP_RET_DATA;
 586         action = filter_ret & SECCOMP_RET_ACTION;
 587
 588         switch (action) {
 589         case SECCOMP_RET_ERRNO:
 590                 /* Set low-order bits as an errno, capped at MAX_ERRNO. */
 591                 if (data > MAX_ERRNO)
 592                         data = MAX_ERRNO;
 593                 syscall_set_return_value(current, task_pt_regs(current),
 594                                          -data, 0);
 595                 goto skip;
 596
 597         case SECCOMP_RET_TRAP:
 598                 /* Show the handler the original registers. */
 599                 syscall_rollback(current, task_pt_regs(current));
 600                 /* Let the filter pass back 16 bits of data. */
 601                 seccomp_send_sigsys(this_syscall, data);
 602                 goto skip;
 603
 604         case SECCOMP_RET_TRACE:
 605                 /* We've been put in this state by the ptracer already. */
 606                 if (recheck_after_trace)
 607                         return 0;
 608
 609                 /* ENOSYS these calls if there is no tracer attached. */
 610                 if (!ptrace_event_enabled(current, PTRACE_EVENT_SECCOMP)) {
 611                         syscall_set_return_value(current,
 612                                                  task_pt_regs(current),
 613                                                  -ENOSYS, 0);
 614                         goto skip;
 615                 }
 616
 617                 /* Allow the BPF to provide the event message */
 618                 ptrace_event(PTRACE_EVENT_SECCOMP, data);
 619                 /*
 620                  * The delivery of a fatal signal during event
 621                  * notification may silently skip tracer notification,
 622                  * which could leave us with a potentially unmodified
 623                  * syscall that the tracer would have liked to have
 624                  * changed. Since the process is about to die, we just
 625                  * force the syscall to be skipped and let the signal
 626                  * kill the process and correctly handle any tracer exit
 627                  * notifications.
 628                  */
 629                 if (fatal_signal_pending(current))
 630                         goto skip;
 631                 /* Check if the tracer forced the syscall to be skipped. */
 632                 this_syscall = syscall_get_nr(current, task_pt_regs(current));
 633                 if (this_syscall < 0)
 634                         goto skip;
 635
 636                 /*
 637                  * Recheck the syscall, since it may have changed. This
 638                  * intentionally uses a NULL struct seccomp_data to force
 639                  * a reload of all registers. This does not goto skip since
 640                  * a skip would have already been reported.
 641                  */
 642                 if (__seccomp_filter(this_syscall, NULL, true))
 643                         return -1;
 644
 645                 return 0;
 646
 647         case SECCOMP_RET_ALLOW:
 648                 /*
 649                  * Note that the "match" filter will always be NULL for
 650                  * this action since SECCOMP_RET_ALLOW is the starting
 651                  * state in seccomp_run_filters().
 652                  */
 653                 return 0;
 654
 655         case SECCOMP_RET_KILL:
 656         default:
 657                 audit_seccomp(this_syscall, SIGSYS, action);
 658                 /* Dump core only if this is the last remaining thread. */
 659                 if (get_nr_threads(current) == 1) {
 660                         siginfo_t info;
 661
 662                         /* Show the original registers in the dump. */
 663                         syscall_rollback(current, task_pt_regs(current));
 664                         /* Trigger a manual coredump since do_exit skips it. */
 665                         seccomp_init_siginfo(&info, this_syscall, data);
 666                         do_coredump(&info);
 667                 }
 668                 do_exit(SIGSYS);
 669         }
 670
 671         unreachable();
 672
 673 skip:
 674         audit_seccomp(this_syscall, 0, action);
 675         return -1;
 676 }
 677 #else
 678 static int __seccomp_filter(int this_syscall, const struct seccomp_data *sd,
 679                             const bool recheck_after_trace)
 680 {
 681         BUG();
 682 }
 683 #endif
 684
 685 int __secure_computing(const struct seccomp_data *sd)
 686 {
 687         int mode = current->seccomp.mode;
 688         int this_syscall;
 689
 690         if (IS_ENABLED(CONFIG_CHECKPOINT_RESTORE) &&
 691             unlikely(current->ptrace & PT_SUSPEND_SECCOMP))
 692                 return 0;
 693
 694         this_syscall = sd ? sd->nr :
 695                 syscall_get_nr(current, task_pt_regs(current));
 696
 697         switch (mode) {
 698         case SECCOMP_MODE_STRICT:
 699                 __secure_computing_strict(this_syscall);  /* may call do_exit */
 700                 return 0;
 701         case SECCOMP_MODE_FILTER:
 702                 return __seccomp_filter(this_syscall, sd, false);
 703         default:
 704                 BUG();
 705         }
 706 }
 707 #endif /* CONFIG_HAVE_ARCH_SECCOMP_FILTER */
 708
 709 long prctl_get_seccomp(void)
 710 {
 711         return current->seccomp.mode;
 712 }
 713
 714 /**
 715  * seccomp_set_mode_strict: internal function for setting strict seccomp
 716  *
 717  * Once current->seccomp.mode is non-zero, it may not be changed.
 718  *
 719  * Returns 0 on success or -EINVAL on failure.
 720  */
 721 static long seccomp_set_mode_strict(void)
 722 {
 723         const unsigned long seccomp_mode = SECCOMP_MODE_STRICT;
 724         long ret = -EINVAL;
 725
 726         spin_lock_irq(&current->sighand->siglock);
 727
 728         if (!seccomp_may_assign_mode(seccomp_mode))
 729                 goto out;
 730
 731 #ifdef TIF_NOTSC
 732         disable_TSC();
 733 #endif
 734         seccomp_assign_mode(current, seccomp_mode);
 735         ret = 0;
 736
 737 out:
 738         spin_unlock_irq(&current->sighand->siglock);
 739
 740         return ret;
 741 }
 742
 743 #ifdef CONFIG_SECCOMP_FILTER
 744 /**
 745  * seccomp_set_mode_filter: internal function for setting seccomp filter
 746  * @flags:  flags to change filter behavior
 747  * @filter: struct sock_fprog containing filter
 748  *
 749  * This function may be called repeatedly to install additional filters.
 750  * Every filter successfully installed will be evaluated (in reverse order)
 751  * for each system call the task makes.
 752  *
 753  * Once current->seccomp.mode is non-zero, it may not be changed.
 754  *
 755  * Returns 0 on success or -EINVAL on failure.
 756  */
 757 static long seccomp_set_mode_filter(unsigned int flags,
 758                                     const char __user *filter)
 759 {
 760         const unsigned long seccomp_mode = SECCOMP_MODE_FILTER;
 761         struct seccomp_filter *prepared = NULL;
 762         long ret = -EINVAL;
 763
 764         /* Validate flags. */
 765         if (flags & ~SECCOMP_FILTER_FLAG_MASK)
 766                 return -EINVAL;
 767
 768         /* Prepare the new filter before holding any locks. */
 769         prepared = seccomp_prepare_user_filter(filter);
 770         if (IS_ERR(prepared))
 771                 return PTR_ERR(prepared);
 772
 773         /*
 774          * Make sure we cannot change seccomp or nnp state via TSYNC
 775          * while another thread is in the middle of calling exec.
 776          */
 777         if (flags & SECCOMP_FILTER_FLAG_TSYNC &&
 778             mutex_lock_killable(&current->signal->cred_guard_mutex))
 779                 goto out_free;
 780
 781         spin_lock_irq(&current->sighand->siglock);
 782
 783         if (!seccomp_may_assign_mode(seccomp_mode))
 784                 goto out;
 785
 786         ret = seccomp_attach_filter(flags, prepared);
 787         if (ret)
 788                 goto out;
 789         /* Do not free the successfully attached filter. */
 790         prepared = NULL;
 791
 792         seccomp_assign_mode(current, seccomp_mode);
 793 out:
 794         spin_unlock_irq(&current->sighand->siglock);
 795         if (flags & SECCOMP_FILTER_FLAG_TSYNC)
 796                 mutex_unlock(&current->signal->cred_guard_mutex);
 797 out_free:
 798         seccomp_filter_free(prepared);
 799         return ret;
 800 }
 801 #else
 802 static inline long seccomp_set_mode_filter(unsigned int flags,
 803                                            const char __user *filter)
 804 {
 805         return -EINVAL;
 806 }
 807 #endif
 808
 809 /* Common entry point for both prctl and syscall. */
 810 static long do_seccomp(unsigned int op, unsigned int flags,
 811                        const char __user *uargs)
 812 {
 813         switch (op) {
 814         case SECCOMP_SET_MODE_STRICT:
 815                 if (flags != 0 || uargs != NULL)
 816                         return -EINVAL;
 817                 return seccomp_set_mode_strict();
 818         case SECCOMP_SET_MODE_FILTER:
 819                 return seccomp_set_mode_filter(flags, uargs);
 820         default:
 821                 return -EINVAL;
 822         }
 823 }
 824
 825 SYSCALL_DEFINE3(seccomp, unsigned int, op, unsigned int, flags,
 826                          const char __user *, uargs)
 827 {
 828         return do_seccomp(op, flags, uargs);
 829 }
 830
 831 /**
 832  * prctl_set_seccomp: configures current->seccomp.mode
 833  * @seccomp_mode: requested mode to use
 834  * @filter: optional struct sock_fprog for use with SECCOMP_MODE_FILTER
 835  *
 836  * Returns 0 on success or -EINVAL on failure.
 837  */
 838 long prctl_set_seccomp(unsigned long seccomp_mode, char __user *filter)
 839 {
 840         unsigned int op;
 841         char __user *uargs;
 842
 843         switch (seccomp_mode) {
 844         case SECCOMP_MODE_STRICT:
 845                 op = SECCOMP_SET_MODE_STRICT;
 846                 /*
 847                  * Setting strict mode through prctl always ignored filter,
 848                  * so make sure it is always NULL here to pass the internal
 849                  * check in do_seccomp().
 850                  */
 851                 uargs = NULL;
 852                 break;
 853         case SECCOMP_MODE_FILTER:
 854                 op = SECCOMP_SET_MODE_FILTER;
 855                 uargs = filter;
 856                 break;
 857         default:
 858                 return -EINVAL;
 859         }
 860
 861         /* prctl interface doesn't have flags, so they are always zero. */
 862         return do_seccomp(op, 0, uargs);
 863 }
 864
 865 #if defined(CONFIG_SECCOMP_FILTER) && defined(CONFIG_CHECKPOINT_RESTORE)
 866 long seccomp_get_filter(struct task_struct *task, unsigned long filter_off,
 867                         void __user *data)
 868 {
 869         struct seccomp_filter *filter;
 870         struct sock_fprog_kern *fprog;
 871         long ret;
 872         unsigned long count = 0;
 873
 874         if (!capable(CAP_SYS_ADMIN) ||
 875             current->seccomp.mode != SECCOMP_MODE_DISABLED) {
 876                 return -EACCES;
 877         }
 878
 879         spin_lock_irq(&task->sighand->siglock);
 880         if (task->seccomp.mode != SECCOMP_MODE_FILTER) {
 881                 ret = -EINVAL;
 882                 goto out;
 883         }
 884
 885         filter = task->seccomp.filter;
 886         while (filter) {
 887                 filter = filter->prev;
 888                 count++;
 889         }
 890
 891         if (filter_off >= count) {
 892                 ret = -ENOENT;
 893                 goto out;
 894         }
 895         count -= filter_off;
 896
 897         filter = task->seccomp.filter;
 898         while (filter && count > 1) {
 899                 filter = filter->prev;
 900                 count--;
 901         }
 902
 903         if (WARN_ON(count != 1 || !filter)) {
 904                 /* The filter tree shouldn't shrink while we're using it. */
 905                 ret = -ENOENT;
 906                 goto out;
 907         }
 908
 909         fprog = filter->prog->orig_prog;
 910         if (!fprog) {
 911                 /* This must be a new non-cBPF filter, since we save
 912                  * every cBPF filter's orig_prog above when
 913                  * CONFIG_CHECKPOINT_RESTORE is enabled.
 914                  */
 915                 ret = -EMEDIUMTYPE;
 916                 goto out;
 917         }
 918
 919         ret = fprog->len;
 920         if (!data)
 921                 goto out;
 922
 923         get_seccomp_filter(task);
 924         spin_unlock_irq(&task->sighand->siglock);
 925
 926         if (copy_to_user(data, fprog->filter, bpf_classic_proglen(fprog)))
 927                 ret = -EFAULT;
 928
 929         put_seccomp_filter(task);
 930         return ret;
 931
 932 out:
 933         spin_unlock_irq(&task->sighand->siglock);
 934         return ret;
 935 }
 936 #endif