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