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  * @insns: 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_S_LD_W_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_S_LD_W_LEN:
 113                         ftest->code = BPF_LD | BPF_IMM;
 114                         ftest->k = sizeof(struct seccomp_data);
 115                         continue;
 116                 case BPF_S_LDX_W_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_S_RET_K:
 122                 case BPF_S_RET_A:
 123                 case BPF_S_ALU_ADD_K:
 124                 case BPF_S_ALU_ADD_X:
 125                 case BPF_S_ALU_SUB_K:
 126                 case BPF_S_ALU_SUB_X:
 127                 case BPF_S_ALU_MUL_K:
 128                 case BPF_S_ALU_MUL_X:
 129                 case BPF_S_ALU_DIV_X:
 130                 case BPF_S_ALU_AND_K:
 131                 case BPF_S_ALU_AND_X:
 132                 case BPF_S_ALU_OR_K:
 133                 case BPF_S_ALU_OR_X:
 134                 case BPF_S_ALU_XOR_K:
 135                 case BPF_S_ALU_XOR_X:
 136                 case BPF_S_ALU_LSH_K:
 137                 case BPF_S_ALU_LSH_X:
 138                 case BPF_S_ALU_RSH_K:
 139                 case BPF_S_ALU_RSH_X:
 140                 case BPF_S_ALU_NEG:
 141                 case BPF_S_LD_IMM:
 142                 case BPF_S_LDX_IMM:
 143                 case BPF_S_MISC_TAX:
 144                 case BPF_S_MISC_TXA:
 145                 case BPF_S_ALU_DIV_K:
 146                 case BPF_S_LD_MEM:
 147                 case BPF_S_LDX_MEM:
 148                 case BPF_S_ST:
 149                 case BPF_S_STX:
 150                 case BPF_S_JMP_JA:
 151                 case BPF_S_JMP_JEQ_K:
 152                 case BPF_S_JMP_JEQ_X:
 153                 case BPF_S_JMP_JGE_K:
 154                 case BPF_S_JMP_JGE_X:
 155                 case BPF_S_JMP_JGT_K:
 156                 case BPF_S_JMP_JGT_X:
 157                 case BPF_S_JMP_JSET_K:
 158                 case BPF_S_JMP_JSET_X:
 159                         sk_decode_filter(ftest, ftest);
 160                         continue;
 161                 default:
 162                         return -EINVAL;
 163                 }
 164         }
 165         return 0;
 166 }
 167
 168 /**
 169  * seccomp_run_filters - evaluates all seccomp filters against @syscall
 170  * @syscall: number of the current system call
 171  *
 172  * Returns valid seccomp BPF response codes.
 173  */
 174 static u32 seccomp_run_filters(int syscall)
 175 {
 176         struct seccomp_filter *f;
 177         struct seccomp_data sd;
 178         u32 ret = SECCOMP_RET_ALLOW;
 179
 180         /* Ensure unexpected behavior doesn't result in failing open. */
 181         if (WARN_ON(current->seccomp.filter == NULL))
 182                 return SECCOMP_RET_KILL;
 183
 184         populate_seccomp_data(&sd);
 185
 186         /*
 187          * All filters in the list are evaluated and the lowest BPF return
 188          * value always takes priority (ignoring the DATA).
 189          */
 190         for (f = current->seccomp.filter; f; f = f->prev) {
 191                 u32 cur_ret = SK_RUN_FILTER(f->prog, (void *)&sd);
 192
 193                 if ((cur_ret & SECCOMP_RET_ACTION) < (ret & SECCOMP_RET_ACTION))
 194                         ret = cur_ret;
 195         }
 196         return ret;
 197 }
 198
 199 /**
 200  * seccomp_attach_filter: Attaches a seccomp filter to current.
 201  * @fprog: BPF program to install
 202  *
 203  * Returns 0 on success or an errno on failure.
 204  */
 205 static long seccomp_attach_filter(struct sock_fprog *fprog)
 206 {
 207         struct seccomp_filter *filter;
 208         unsigned long fp_size = fprog->len * sizeof(struct sock_filter);
 209         unsigned long total_insns = fprog->len;
 210         struct sock_filter *fp;
 211         int new_len;
 212         long ret;
 213
 214         if (fprog->len == 0 || fprog->len > BPF_MAXINSNS)
 215                 return -EINVAL;
 216
 217         for (filter = current->seccomp.filter; filter; filter = filter->prev)
 218                 total_insns += filter->prog->len + 4;  /* include a 4 instr penalty */
 219         if (total_insns > MAX_INSNS_PER_PATH)
 220                 return -ENOMEM;
 221
 222         /*
 223          * Installing a seccomp filter requires that the task have
 224          * CAP_SYS_ADMIN in its namespace or be running with no_new_privs.
 225          * This avoids scenarios where unprivileged tasks can affect the
 226          * behavior of privileged children.
 227          */
 228         if (!current->no_new_privs &&
 229             security_capable_noaudit(current_cred(), current_user_ns(),
 230                                      CAP_SYS_ADMIN) != 0)
 231                 return -EACCES;
 232
 233         fp = kzalloc(fp_size, GFP_KERNEL|__GFP_NOWARN);
 234         if (!fp)
 235                 return -ENOMEM;
 236
 237         /* Copy the instructions from fprog. */
 238         ret = -EFAULT;
 239         if (copy_from_user(fp, fprog->filter, fp_size))
 240                 goto free_prog;
 241
 242         /* Check and rewrite the fprog via the skb checker */
 243         ret = sk_chk_filter(fp, fprog->len);
 244         if (ret)
 245                 goto free_prog;
 246
 247         /* Check and rewrite the fprog for seccomp use */
 248         ret = seccomp_check_filter(fp, fprog->len);
 249         if (ret)
 250                 goto free_prog;
 251
 252         /* Convert 'sock_filter' insns to 'sock_filter_int' insns */
 253         ret = sk_convert_filter(fp, fprog->len, NULL, &new_len);
 254         if (ret)
 255                 goto free_prog;
 256
 257         /* Allocate a new seccomp_filter */
 258         ret = -ENOMEM;
 259         filter = kzalloc(sizeof(struct seccomp_filter),
 260                          GFP_KERNEL|__GFP_NOWARN);
 261         if (!filter)
 262                 goto free_prog;
 263
 264         filter->prog = kzalloc(sk_filter_size(new_len),
 265                                GFP_KERNEL|__GFP_NOWARN);
 266         if (!filter->prog)
 267                 goto free_filter;
 268
 269         ret = sk_convert_filter(fp, fprog->len, filter->prog->insnsi, &new_len);
 270         if (ret)
 271                 goto free_filter_prog;
 272         kfree(fp);
 273
 274         atomic_set(&filter->usage, 1);
 275         filter->prog->len = new_len;
 276         filter->prog->bpf_func = (void *)sk_run_filter_int_seccomp;
 277
 278         /* JIT internal BPF into native HW instructions */
 279         bpf_int_jit_compile(filter->prog);
 280
 281         /*
 282          * If there is an existing filter, make it the prev and don't drop its
 283          * task reference.
 284          */
 285         filter->prev = current->seccomp.filter;
 286         current->seccomp.filter = filter;
 287         return 0;
 288
 289 free_filter_prog:
 290         kfree(filter->prog);
 291 free_filter:
 292         kfree(filter);
 293 free_prog:
 294         kfree(fp);
 295         return ret;
 296 }
 297
 298 /**
 299  * seccomp_attach_user_filter - attaches a user-supplied sock_fprog
 300  * @user_filter: pointer to the user data containing a sock_fprog.
 301  *
 302  * Returns 0 on success and non-zero otherwise.
 303  */
 304 static long seccomp_attach_user_filter(char __user *user_filter)
 305 {
 306         struct sock_fprog fprog;
 307         long ret = -EFAULT;
 308
 309 #ifdef CONFIG_COMPAT
 310         if (is_compat_task()) {
 311                 struct compat_sock_fprog fprog32;
 312                 if (copy_from_user(&fprog32, user_filter, sizeof(fprog32)))
 313                         goto out;
 314                 fprog.len = fprog32.len;
 315                 fprog.filter = compat_ptr(fprog32.filter);
 316         } else /* falls through to the if below. */
 317 #endif
 318         if (copy_from_user(&fprog, user_filter, sizeof(fprog)))
 319                 goto out;
 320         ret = seccomp_attach_filter(&fprog);
 321 out:
 322         return ret;
 323 }
 324
 325 /* get_seccomp_filter - increments the reference count of the filter on @tsk */
 326 void get_seccomp_filter(struct task_struct *tsk)
 327 {
 328         struct seccomp_filter *orig = tsk->seccomp.filter;
 329         if (!orig)
 330                 return;
 331         /* Reference count is bounded by the number of total processes. */
 332         atomic_inc(&orig->usage);
 333 }
 334
 335 /* put_seccomp_filter - decrements the ref count of tsk->seccomp.filter */
 336 void put_seccomp_filter(struct task_struct *tsk)
 337 {
 338         struct seccomp_filter *orig = tsk->seccomp.filter;
 339         /* Clean up single-reference branches iteratively. */
 340         while (orig && atomic_dec_and_test(&orig->usage)) {
 341                 struct seccomp_filter *freeme = orig;
 342                 orig = orig->prev;
 343                 bpf_jit_free(freeme->prog);
 344                 kfree(freeme);
 345         }
 346 }
 347
 348 /**
 349  * seccomp_send_sigsys - signals the task to allow in-process syscall emulation
 350  * @syscall: syscall number to send to userland
 351  * @reason: filter-supplied reason code to send to userland (via si_errno)
 352  *
 353  * Forces a SIGSYS with a code of SYS_SECCOMP and related sigsys info.
 354  */
 355 static void seccomp_send_sigsys(int syscall, int reason)
 356 {
 357         struct siginfo info;
 358         memset(&info, 0, sizeof(info));
 359         info.si_signo = SIGSYS;
 360         info.si_code = SYS_SECCOMP;
 361         info.si_call_addr = (void __user *)KSTK_EIP(current);
 362         info.si_errno = reason;
 363         info.si_arch = syscall_get_arch();
 364         info.si_syscall = syscall;
 365         force_sig_info(SIGSYS, &info, current);
 366 }
 367 #endif  /* CONFIG_SECCOMP_FILTER */
 368
 369 /*
 370  * Secure computing mode 1 allows only read/write/exit/sigreturn.
 371  * To be fully secure this must be combined with rlimit
 372  * to limit the stack allocations too.
 373  */
 374 static int mode1_syscalls[] = {
 375         __NR_seccomp_read, __NR_seccomp_write, __NR_seccomp_exit, __NR_seccomp_sigreturn,
 376         0, /* null terminated */
 377 };
 378
 379 #ifdef CONFIG_COMPAT
 380 static int mode1_syscalls_32[] = {
 381         __NR_seccomp_read_32, __NR_seccomp_write_32, __NR_seccomp_exit_32, __NR_seccomp_sigreturn_32,
 382         0, /* null terminated */
 383 };
 384 #endif
 385
 386 int __secure_computing(int this_syscall)
 387 {
 388         int mode = current->seccomp.mode;
 389         int exit_sig = 0;
 390         int *syscall;
 391         u32 ret;
 392
 393         switch (mode) {
 394         case SECCOMP_MODE_STRICT:
 395                 syscall = mode1_syscalls;
 396 #ifdef CONFIG_COMPAT
 397                 if (is_compat_task())
 398                         syscall = mode1_syscalls_32;
 399 #endif
 400                 do {
 401                         if (*syscall == this_syscall)
 402                                 return 0;
 403                 } while (*++syscall);
 404                 exit_sig = SIGKILL;
 405                 ret = SECCOMP_RET_KILL;
 406                 break;
 407 #ifdef CONFIG_SECCOMP_FILTER
 408         case SECCOMP_MODE_FILTER: {
 409                 int data;
 410                 struct pt_regs *regs = task_pt_regs(current);
 411                 ret = seccomp_run_filters(this_syscall);
 412                 data = ret & SECCOMP_RET_DATA;
 413                 ret &= SECCOMP_RET_ACTION;
 414                 switch (ret) {
 415                 case SECCOMP_RET_ERRNO:
 416                         /* Set the low-order 16-bits as a errno. */
 417                         syscall_set_return_value(current, regs,
 418                                                  -data, 0);
 419                         goto skip;
 420                 case SECCOMP_RET_TRAP:
 421                         /* Show the handler the original registers. */
 422                         syscall_rollback(current, regs);
 423                         /* Let the filter pass back 16 bits of data. */
 424                         seccomp_send_sigsys(this_syscall, data);
 425                         goto skip;
 426                 case SECCOMP_RET_TRACE:
 427                         /* Skip these calls if there is no tracer. */
 428                         if (!ptrace_event_enabled(current, PTRACE_EVENT_SECCOMP)) {
 429                                 syscall_set_return_value(current, regs,
 430                                                          -ENOSYS, 0);
 431                                 goto skip;
 432                         }
 433                         /* Allow the BPF to provide the event message */
 434                         ptrace_event(PTRACE_EVENT_SECCOMP, data);
 435                         /*
 436                          * The delivery of a fatal signal during event
 437                          * notification may silently skip tracer notification.
 438                          * Terminating the task now avoids executing a system
 439                          * call that may not be intended.
 440                          */
 441                         if (fatal_signal_pending(current))
 442                                 break;
 443                         if (syscall_get_nr(current, regs) < 0)
 444                                 goto skip;  /* Explicit request to skip. */
 445
 446                         return 0;
 447                 case SECCOMP_RET_ALLOW:
 448                         return 0;
 449                 case SECCOMP_RET_KILL:
 450                 default:
 451                         break;
 452                 }
 453                 exit_sig = SIGSYS;
 454                 break;
 455         }
 456 #endif
 457         default:
 458                 BUG();
 459         }
 460
 461 #ifdef SECCOMP_DEBUG
 462         dump_stack();
 463 #endif
 464         audit_seccomp(this_syscall, exit_sig, ret);
 465         do_exit(exit_sig);
 466 #ifdef CONFIG_SECCOMP_FILTER
 467 skip:
 468         audit_seccomp(this_syscall, exit_sig, ret);
 469 #endif
 470         return -1;
 471 }
 472
 473 long prctl_get_seccomp(void)
 474 {
 475         return current->seccomp.mode;
 476 }
 477
 478 /**
 479  * prctl_set_seccomp: configures current->seccomp.mode
 480  * @seccomp_mode: requested mode to use
 481  * @filter: optional struct sock_fprog for use with SECCOMP_MODE_FILTER
 482  *
 483  * This function may be called repeatedly with a @seccomp_mode of
 484  * SECCOMP_MODE_FILTER to install additional filters.  Every filter
 485  * successfully installed will be evaluated (in reverse order) for each system
 486  * call the task makes.
 487  *
 488  * Once current->seccomp.mode is non-zero, it may not be changed.
 489  *
 490  * Returns 0 on success or -EINVAL on failure.
 491  */
 492 long prctl_set_seccomp(unsigned long seccomp_mode, char __user *filter)
 493 {
 494         long ret = -EINVAL;
 495
 496         if (current->seccomp.mode &&
 497             current->seccomp.mode != seccomp_mode)
 498                 goto out;
 499
 500         switch (seccomp_mode) {
 501         case SECCOMP_MODE_STRICT:
 502                 ret = 0;
 503 #ifdef TIF_NOTSC
 504                 disable_TSC();
 505 #endif
 506                 break;
 507 #ifdef CONFIG_SECCOMP_FILTER
 508         case SECCOMP_MODE_FILTER:
 509                 ret = seccomp_attach_user_filter(filter);
 510                 if (ret)
 511                         goto out;
 512                 break;
 513 #endif
 514         default:
 515                 goto out;
 516         }
 517
 518         current->seccomp.mode = seccomp_mode;
 519         set_thread_flag(TIF_SECCOMP);
 520 out:
 521         return ret;
 522 }