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