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
  22 /* #define SECCOMP_DEBUG 1 */
  23
  24 #ifdef CONFIG_SECCOMP_FILTER
  25 #include <asm/syscall.h>
  26 #include <linux/filter.h>
  27 #include <linux/ptrace.h>
  28 #include <linux/security.h>
  29 #include <linux/slab.h>
  30 #include <linux/tracehook.h>
  31 #include <linux/uaccess.h>
  32
  33 /**
  34  * struct seccomp_filter - container for seccomp BPF programs
  35  *
  36  * @usage: reference count to manage the object lifetime.
  37  *         get/put helpers should be used when accessing an instance
  38  *         outside of a lifetime-guarded section.  In general, this
  39  *         is only needed for handling filters shared across tasks.
  40  * @prev: points to a previously installed, or inherited, filter
  41  * @len: the number of instructions in the program
  42  * @insnsi: the BPF program instructions to evaluate
  43  *
  44  * seccomp_filter objects are organized in a tree linked via the @prev
  45  * pointer.  For any task, it appears to be a singly-linked list starting
  46  * with current->seccomp.filter, the most recently attached or inherited filter.
  47  * However, multiple filters may share a @prev node, by way of fork(), which
  48  * results in a unidirectional tree existing in memory.  This is similar to
  49  * how namespaces work.
  50  *
  51  * seccomp_filter objects should never be modified after being attached
  52  * to a task_struct (other than @usage).
  53  */
  54 struct seccomp_filter {
  55         atomic_t usage;
  56         struct seccomp_filter *prev;
  57         struct sk_filter *prog;
  58 };
  59
  60 /* Limit any path through the tree to 256KB worth of instructions. */
  61 #define MAX_INSNS_PER_PATH ((1 << 18) / sizeof(struct sock_filter))
  62
  63 /*
  64  * Endianness is explicitly ignored and left for BPF program authors to manage
  65  * as per the specific architecture.
  66  */
  67 static void populate_seccomp_data(struct seccomp_data *sd)
  68 {
  69         struct task_struct *task = current;
  70         struct pt_regs *regs = task_pt_regs(task);
  71         unsigned long args[6];
  72
  73         sd->nr = syscall_get_nr(task, regs);
  74         sd->arch = syscall_get_arch();
  75         syscall_get_arguments(task, regs, 0, 6, args);
  76         sd->args[0] = args[0];
  77         sd->args[1] = args[1];
  78         sd->args[2] = args[2];
  79         sd->args[3] = args[3];
  80         sd->args[4] = args[4];
  81         sd->args[5] = args[5];
  82         sd->instruction_pointer = KSTK_EIP(task);
  83 }
  84
  85 /**
  86  *      seccomp_check_filter - verify seccomp filter code
  87  *      @filter: filter to verify
  88  *      @flen: length of filter
  89  *
  90  * Takes a previously checked filter (by sk_chk_filter) and
  91  * redirects all filter code that loads struct sk_buff data
  92  * and related data through seccomp_bpf_load.  It also
  93  * enforces length and alignment checking of those loads.
  94  *
  95  * Returns 0 if the rule set is legal or -EINVAL if not.
  96  */
  97 static int seccomp_check_filter(struct sock_filter *filter, unsigned int flen)
  98 {
  99         int pc;
 100         for (pc = 0; pc < flen; pc++) {
 101                 struct sock_filter *ftest = &filter[pc];
 102                 u16 code = ftest->code;
 103                 u32 k = ftest->k;
 104
 105                 switch (code) {
 106                 case BPF_LD | BPF_W | BPF_ABS:
 107                         ftest->code = BPF_LDX | BPF_W | BPF_ABS;
 108                         /* 32-bit aligned and not out of bounds. */
 109                         if (k >= sizeof(struct seccomp_data) || k & 3)
 110                                 return -EINVAL;
 111                         continue;
 112                 case BPF_LD | BPF_W | BPF_LEN:
 113                         ftest->code = BPF_LD | BPF_IMM;
 114                         ftest->k = sizeof(struct seccomp_data);
 115                         continue;
 116                 case BPF_LDX | BPF_W | BPF_LEN:
 117                         ftest->code = BPF_LDX | BPF_IMM;
 118                         ftest->k = sizeof(struct seccomp_data);
 119                         continue;
 120                 /* Explicitly include allowed calls. */
 121                 case BPF_RET | BPF_K:
 122                 case BPF_RET | BPF_A:
 123                 case BPF_ALU | BPF_ADD | BPF_K:
 124                 case BPF_ALU | BPF_ADD | BPF_X:
 125                 case BPF_ALU | BPF_SUB | BPF_K:
 126                 case BPF_ALU | BPF_SUB | BPF_X:
 127                 case BPF_ALU | BPF_MUL | BPF_K:
 128                 case BPF_ALU | BPF_MUL | BPF_X:
 129                 case BPF_ALU | BPF_DIV | BPF_K:
 130                 case BPF_ALU | BPF_DIV | BPF_X:
 131                 case BPF_ALU | BPF_AND | BPF_K:
 132                 case BPF_ALU | BPF_AND | BPF_X:
 133                 case BPF_ALU | BPF_OR | BPF_K:
 134                 case BPF_ALU | BPF_OR | BPF_X:
 135                 case BPF_ALU | BPF_XOR | BPF_K:
 136                 case BPF_ALU | BPF_XOR | BPF_X:
 137                 case BPF_ALU | BPF_LSH | BPF_K:
 138                 case BPF_ALU | BPF_LSH | BPF_X:
 139                 case BPF_ALU | BPF_RSH | BPF_K:
 140                 case BPF_ALU | BPF_RSH | BPF_X:
 141                 case BPF_ALU | BPF_NEG:
 142                 case BPF_LD | BPF_IMM:
 143                 case BPF_LDX | BPF_IMM:
 144                 case BPF_MISC | BPF_TAX:
 145                 case BPF_MISC | BPF_TXA:
 146                 case BPF_LD | BPF_MEM:
 147                 case BPF_LDX | BPF_MEM:
 148                 case BPF_ST:
 149                 case BPF_STX:
 150                 case BPF_JMP | BPF_JA:
 151                 case BPF_JMP | BPF_JEQ | BPF_K:
 152                 case BPF_JMP | BPF_JEQ | BPF_X:
 153                 case BPF_JMP | BPF_JGE | BPF_K:
 154                 case BPF_JMP | BPF_JGE | BPF_X:
 155                 case BPF_JMP | BPF_JGT | BPF_K:
 156                 case BPF_JMP | BPF_JGT | BPF_X:
 157                 case BPF_JMP | BPF_JSET | BPF_K:
 158                 case BPF_JMP | BPF_JSET | BPF_X:
 159                         continue;
 160                 default:
 161                         return -EINVAL;
 162                 }
 163         }
 164         return 0;
 165 }
 166
 167 /**
 168  * seccomp_run_filters - evaluates all seccomp filters against @syscall
 169  * @syscall: number of the current system call
 170  *
 171  * Returns valid seccomp BPF response codes.
 172  */
 173 static u32 seccomp_run_filters(int syscall)
 174 {
 175         struct seccomp_filter *f;
 176         struct seccomp_data sd;
 177         u32 ret = SECCOMP_RET_ALLOW;
 178
 179         /* Ensure unexpected behavior doesn't result in failing open. */
 180         if (WARN_ON(current->seccomp.filter == NULL))
 181                 return SECCOMP_RET_KILL;
 182
 183         populate_seccomp_data(&sd);
 184
 185         /*
 186          * All filters in the list are evaluated and the lowest BPF return
 187          * value always takes priority (ignoring the DATA).
 188          */
 189         for (f = current->seccomp.filter; f; f = f->prev) {
 190                 u32 cur_ret = SK_RUN_FILTER(f->prog, (void *)&sd);
 191
 192                 if ((cur_ret & SECCOMP_RET_ACTION) < (ret & SECCOMP_RET_ACTION))
 193                         ret = cur_ret;
 194         }
 195         return ret;
 196 }
 197
 198 /**
 199  * seccomp_attach_filter: Attaches a seccomp filter to current.
 200  * @fprog: BPF program to install
 201  *
 202  * Returns 0 on success or an errno on failure.
 203  */
 204 static long seccomp_attach_filter(struct sock_fprog *fprog)
 205 {
 206         struct seccomp_filter *filter;
 207         unsigned long fp_size = fprog->len * sizeof(struct sock_filter);
 208         unsigned long total_insns = fprog->len;
 209         struct sock_filter *fp;
 210         int new_len;
 211         long ret;
 212
 213         if (fprog->len == 0 || fprog->len > BPF_MAXINSNS)
 214                 return -EINVAL;
 215
 216         for (filter = current->seccomp.filter; filter; filter = filter->prev)
 217                 total_insns += filter->prog->len + 4;  /* include a 4 instr penalty */
 218         if (total_insns > MAX_INSNS_PER_PATH)
 219                 return -ENOMEM;
 220
 221         /*
 222          * Installing a seccomp filter requires that the task has
 223          * CAP_SYS_ADMIN in its namespace or be running with no_new_privs.
 224          * This avoids scenarios where unprivileged tasks can affect the
 225          * behavior of privileged children.
 226          */
 227         if (!current->no_new_privs &&
 228             security_capable_noaudit(current_cred(), current_user_ns(),
 229                                      CAP_SYS_ADMIN) != 0)
 230                 return -EACCES;
 231
 232         fp = kzalloc(fp_size, GFP_KERNEL|__GFP_NOWARN);
 233         if (!fp)
 234                 return -ENOMEM;
 235
 236         /* Copy the instructions from fprog. */
 237         ret = -EFAULT;
 238         if (copy_from_user(fp, fprog->filter, fp_size))
 239                 goto free_prog;
 240
 241         /* Check and rewrite the fprog via the skb checker */
 242         ret = sk_chk_filter(fp, fprog->len);
 243         if (ret)
 244                 goto free_prog;
 245
 246         /* Check and rewrite the fprog for seccomp use */
 247         ret = seccomp_check_filter(fp, fprog->len);
 248         if (ret)
 249                 goto free_prog;
 250
 251         /* Convert 'sock_filter' insns to 'sock_filter_int' insns */
 252         ret = sk_convert_filter(fp, fprog->len, NULL, &new_len);
 253         if (ret)
 254                 goto free_prog;
 255
 256         /* Allocate a new seccomp_filter */
 257         ret = -ENOMEM;
 258         filter = kzalloc(sizeof(struct seccomp_filter),
 259                          GFP_KERNEL|__GFP_NOWARN);
 260         if (!filter)
 261                 goto free_prog;
 262
 263         filter->prog = kzalloc(sk_filter_size(new_len),
 264                                GFP_KERNEL|__GFP_NOWARN);
 265         if (!filter->prog)
 266                 goto free_filter;
 267
 268         ret = sk_convert_filter(fp, fprog->len, filter->prog->insnsi, &new_len);
 269         if (ret)
 270                 goto free_filter_prog;
 271         kfree(fp);
 272
 273         atomic_set(&filter->usage, 1);
 274         filter->prog->len = new_len;
 275
 276         sk_filter_select_runtime(filter->prog);
 277
 278         /*
 279          * If there is an existing filter, make it the prev and don't drop its
 280          * task reference.
 281          */
 282         filter->prev = current->seccomp.filter;
 283         current->seccomp.filter = filter;
 284         return 0;
 285
 286 free_filter_prog:
 287         kfree(filter->prog);
 288 free_filter:
 289         kfree(filter);
 290 free_prog:
 291         kfree(fp);
 292         return ret;
 293 }
 294
 295 /**
 296  * seccomp_attach_user_filter - attaches a user-supplied sock_fprog
 297  * @user_filter: pointer to the user data containing a sock_fprog.
 298  *
 299  * Returns 0 on success and non-zero otherwise.
 300  */
 301 static long seccomp_attach_user_filter(char __user *user_filter)
 302 {
 303         struct sock_fprog fprog;
 304         long ret = -EFAULT;
 305
 306 #ifdef CONFIG_COMPAT
 307         if (is_compat_task()) {
 308                 struct compat_sock_fprog fprog32;
 309                 if (copy_from_user(&fprog32, user_filter, sizeof(fprog32)))
 310                         goto out;
 311                 fprog.len = fprog32.len;
 312                 fprog.filter = compat_ptr(fprog32.filter);
 313         } else /* falls through to the if below. */
 314 #endif
 315         if (copy_from_user(&fprog, user_filter, sizeof(fprog)))
 316                 goto out;
 317         ret = seccomp_attach_filter(&fprog);
 318 out:
 319         return ret;
 320 }
 321
 322 /* get_seccomp_filter - increments the reference count of the filter on @tsk */
 323 void get_seccomp_filter(struct task_struct *tsk)
 324 {
 325         struct seccomp_filter *orig = tsk->seccomp.filter;
 326         if (!orig)
 327                 return;
 328         /* Reference count is bounded by the number of total processes. */
 329         atomic_inc(&orig->usage);
 330 }
 331
 332 /* put_seccomp_filter - decrements the ref count of tsk->seccomp.filter */
 333 void put_seccomp_filter(struct task_struct *tsk)
 334 {
 335         struct seccomp_filter *orig = tsk->seccomp.filter;
 336         /* Clean up single-reference branches iteratively. */
 337         while (orig && atomic_dec_and_test(&orig->usage)) {
 338                 struct seccomp_filter *freeme = orig;
 339                 orig = orig->prev;
 340                 sk_filter_free(freeme->prog);
 341                 kfree(freeme);
 342         }
 343 }
 344
 345 /**
 346  * seccomp_send_sigsys - signals the task to allow in-process syscall emulation
 347  * @syscall: syscall number to send to userland
 348  * @reason: filter-supplied reason code to send to userland (via si_errno)
 349  *
 350  * Forces a SIGSYS with a code of SYS_SECCOMP and related sigsys info.
 351  */
 352 static void seccomp_send_sigsys(int syscall, int reason)
 353 {
 354         struct siginfo info;
 355         memset(&info, 0, sizeof(info));
 356         info.si_signo = SIGSYS;
 357         info.si_code = SYS_SECCOMP;
 358         info.si_call_addr = (void __user *)KSTK_EIP(current);
 359         info.si_errno = reason;
 360         info.si_arch = syscall_get_arch();
 361         info.si_syscall = syscall;
 362         force_sig_info(SIGSYS, &info, current);
 363 }
 364 #endif  /* CONFIG_SECCOMP_FILTER */
 365
 366 /*
 367  * Secure computing mode 1 allows only read/write/exit/sigreturn.
 368  * To be fully secure this must be combined with rlimit
 369  * to limit the stack allocations too.
 370  */
 371 static int mode1_syscalls[] = {
 372         __NR_seccomp_read, __NR_seccomp_write, __NR_seccomp_exit, __NR_seccomp_sigreturn,
 373         0, /* null terminated */
 374 };
 375
 376 #ifdef CONFIG_COMPAT
 377 static int mode1_syscalls_32[] = {
 378         __NR_seccomp_read_32, __NR_seccomp_write_32, __NR_seccomp_exit_32, __NR_seccomp_sigreturn_32,
 379         0, /* null terminated */
 380 };
 381 #endif
 382
 383 int __secure_computing(int this_syscall)
 384 {
 385         int mode = current->seccomp.mode;
 386         int exit_sig = 0;
 387         int *syscall;
 388         u32 ret;
 389
 390         switch (mode) {
 391         case SECCOMP_MODE_STRICT:
 392                 syscall = mode1_syscalls;
 393 #ifdef CONFIG_COMPAT
 394                 if (is_compat_task())
 395                         syscall = mode1_syscalls_32;
 396 #endif
 397                 do {
 398                         if (*syscall == this_syscall)
 399                                 return 0;
 400                 } while (*++syscall);
 401                 exit_sig = SIGKILL;
 402                 ret = SECCOMP_RET_KILL;
 403                 break;
 404 #ifdef CONFIG_SECCOMP_FILTER
 405         case SECCOMP_MODE_FILTER: {
 406                 int data;
 407                 struct pt_regs *regs = task_pt_regs(current);
 408                 ret = seccomp_run_filters(this_syscall);
 409                 data = ret & SECCOMP_RET_DATA;
 410                 ret &= SECCOMP_RET_ACTION;
 411                 switch (ret) {
 412                 case SECCOMP_RET_ERRNO:
 413                         /* Set the low-order 16-bits as a errno. */
 414                         syscall_set_return_value(current, regs,
 415                                                  -data, 0);
 416                         goto skip;
 417                 case SECCOMP_RET_TRAP:
 418                         /* Show the handler the original registers. */
 419                         syscall_rollback(current, regs);
 420                         /* Let the filter pass back 16 bits of data. */
 421                         seccomp_send_sigsys(this_syscall, data);
 422                         goto skip;
 423                 case SECCOMP_RET_TRACE:
 424                         /* Skip these calls if there is no tracer. */
 425                         if (!ptrace_event_enabled(current, PTRACE_EVENT_SECCOMP)) {
 426                                 syscall_set_return_value(current, regs,
 427                                                          -ENOSYS, 0);
 428                                 goto skip;
 429                         }
 430                         /* Allow the BPF to provide the event message */
 431                         ptrace_event(PTRACE_EVENT_SECCOMP, data);
 432                         /*
 433                          * The delivery of a fatal signal during event
 434                          * notification may silently skip tracer notification.
 435                          * Terminating the task now avoids executing a system
 436                          * call that may not be intended.
 437                          */
 438                         if (fatal_signal_pending(current))
 439                                 break;
 440                         if (syscall_get_nr(current, regs) < 0)
 441                                 goto skip;  /* Explicit request to skip. */
 442
 443                         return 0;
 444                 case SECCOMP_RET_ALLOW:
 445                         return 0;
 446                 case SECCOMP_RET_KILL:
 447                 default:
 448                         break;
 449                 }
 450                 exit_sig = SIGSYS;
 451                 break;
 452         }
 453 #endif
 454         default:
 455                 BUG();
 456         }
 457
 458 #ifdef SECCOMP_DEBUG
 459         dump_stack();
 460 #endif
 461         audit_seccomp(this_syscall, exit_sig, ret);
 462         do_exit(exit_sig);
 463 #ifdef CONFIG_SECCOMP_FILTER
 464 skip:
 465         audit_seccomp(this_syscall, exit_sig, ret);
 466 #endif
 467         return -1;
 468 }
 469
 470 long prctl_get_seccomp(void)
 471 {
 472         return current->seccomp.mode;
 473 }
 474
 475 /**
 476  * prctl_set_seccomp: configures current->seccomp.mode
 477  * @seccomp_mode: requested mode to use
 478  * @filter: optional struct sock_fprog for use with SECCOMP_MODE_FILTER
 479  *
 480  * This function may be called repeatedly with a @seccomp_mode of
 481  * SECCOMP_MODE_FILTER to install additional filters.  Every filter
 482  * successfully installed will be evaluated (in reverse order) for each system
 483  * call the task makes.
 484  *
 485  * Once current->seccomp.mode is non-zero, it may not be changed.
 486  *
 487  * Returns 0 on success or -EINVAL on failure.
 488  */
 489 long prctl_set_seccomp(unsigned long seccomp_mode, char __user *filter)
 490 {
 491         long ret = -EINVAL;
 492
 493         if (current->seccomp.mode &&
 494             current->seccomp.mode != seccomp_mode)
 495                 goto out;
 496
 497         switch (seccomp_mode) {
 498         case SECCOMP_MODE_STRICT:
 499                 ret = 0;
 500 #ifdef TIF_NOTSC
 501                 disable_TSC();
 502 #endif
 503                 break;
 504 #ifdef CONFIG_SECCOMP_FILTER
 505         case SECCOMP_MODE_FILTER:
 506                 ret = seccomp_attach_user_filter(filter);
 507                 if (ret)
 508                         goto out;
 509                 break;
 510 #endif
 511         default:
 512                 goto out;
 513         }
 514
 515         current->seccomp.mode = seccomp_mode;
 516         set_thread_flag(TIF_SECCOMP);
 517 out:
 518         return ret;
 519 }