libseccomp: incorrect compilation of arithmetic comparisons 


When libseccomp compiles filters for 64-bit systems, it needs to split 64-bit
comparisons into 32-bit comparisons because classic BPF can't operate on 64-bit
values directly.

libseccomp offers both bitwise comparisons (NE, EQ, MASKED_EQ) and arithmetic
comparisons (LT, LE, GE, GT). Bitwise comparisons can always be implemented with
no more than two comparisons; but that doesn't work for arithmetic comparisons.

Consider the case where a filter attempts to check whether
args[0]<0x123456789abc. The cases are:

args[0].high < 0x1234:                               matches
args[0].high > 0x1234:                               no match
args[0].high == 0x1234 && args[0].low < 0x56789abc:  matches
args[0].high == 0x1234 && args[0].low >= 0x56789abc: no match

So in pseudocode, you'd want something like the following:

if args[0].high < 0x1234
  return ACCEPT
if args[0].high > 0x1234
  return REJECT
if args[0].low < 0x56789abc
  return ACCEPT
return REJECT


But actually, when libseccomp is invoked as follows:

  scmp_filter_ctx ctx = seccomp_init(SCMP_ACT_ALLOW);
  if (ctx == NULL) err(1, "seccomp_init");
  if (seccomp_rule_add(ctx, SCMP_ACT_ERRNO(EBADSLT), SCMP_SYS(mincore), 1,
                       SCMP_A0(SCMP_CMP_LT, 0x123456789abcUL)))
    err(1, "seccomp_rule_add");
  if (seccomp_load(ctx))
    err(1, "seccomp_load");

it generates the following seccomp filter:

# ./seccomp_dump 96148 simple
===== filter 0 (13 instructions) =====
0001 if arch != X86_64: [true +10, false +0] -> ret KILL
0003 if nr < 0x40000000: [true +1, false +0]
0005   if nr != 0x0000001b: [true +5, false +0] -> ret ALLOW (syscalls: <TOO MANY TO LIST>)
0007   if args[0].high < 0x00001234: [true +2, false +0] -> ret ERRNO
0009   if args[0].low >= 0x56789abc: [true +1, false +0] -> ret ALLOW (syscalls: mincore)
000a   ret ERRNO
[...]

As you can see, the case of `args[0].high > 0x1234 && args[0].low < 0x56789abc`
is handled incorrectly.



Here's a demo, tested with libseccomp from git master:
===========================================
jannh@jannh2:~/tests/libseccomp-stuff$ cat compare.c
#include <seccomp.h>
#include <err.h>
#include <stdlib.h>
#include <unistd.h>
#include <stdio.h>
#include <errno.h>
#include <sys/mman.h>

// any mincore() starting below this address should be denied with -EBADSLT
#define ADDR_LIMIT 0x123456789abcUL

static void sctest(unsigned long addr) {
  unsigned char vec;
  printf("mincore(0x%012lx, 0) = ", addr);
  int res = mincore((void*)addr, 0, &vec);
  if (res == 0) {
    printf(" 0\n");
  } else {
    printf("-%d (%m)\n", errno);
  }
}

int main(int argc, char **argv) {
  setbuf(stdout, NULL);
  printf("my pid is %d\n", (int)getpid());
  scmp_filter_ctx ctx = seccomp_init(SCMP_ACT_ALLOW);
  if (ctx == NULL) err(1, "seccomp_init");
  if (seccomp_rule_add(ctx, SCMP_ACT_ERRNO(EBADSLT), SCMP_SYS(mincore), 1,
                       SCMP_A0(SCMP_CMP_LT, ADDR_LIMIT)))
    err(1, "seccomp_rule_add");
  if (seccomp_load(ctx))
    err(1, "seccomp_load");

  sctest(0);
  sctest(0x123000000000);
  sctest(0x1230f0000000);
  sctest(0x123400000000);
  sctest(0x123450000000);
  sctest(0x123460000000);
  sctest(0x1234f0000000);
  sctest(0x123500000000);
  sctest(0x1235f0000000);
  sctest(0x123600000000);

  while (1) pause();
}
jannh@jannh2:~/tests/libseccomp-stuff$ gcc -o compare compare.c -Wall -I/h/git/foreign/libseccomp/include/ -L/h/git/foreign/libseccomp/src/.libs -lseccomp -Wl,-rpath /h/git/foreign/libseccomp/src/.libs/
jannh@jannh2:~/tests/libseccomp-stuff$ ./compare 
my pid is 104373
mincore(0x000000000000, 0) = -57 (Invalid slot)
mincore(0x123000000000, 0) = -57 (Invalid slot)
mincore(0x1230f0000000, 0) = -57 (Invalid slot)
mincore(0x123400000000, 0) = -57 (Invalid slot)
mincore(0x123450000000, 0) = -57 (Invalid slot)
mincore(0x123460000000, 0) =  0
mincore(0x1234f0000000, 0) =  0
mincore(0x123500000000, 0) = -57 (Invalid slot)
mincore(0x1235f0000000, 0) =  0
mincore(0x123600000000, 0) = -57 (Invalid slot)
===========================================


This probably isn't terribly interesting for most users of libseccomp, but the
Tor daemon
(https://gitweb.torproject.org/tor.git/tree/src/lib/sandbox/sandbox.c) does use
arithmetic comparisons to prevent writes to a certain memory region:

===========================================
  /*
   * Allow mprotect with PROT_READ|PROT_WRITE because openssl uses it, but
   * never over the memory region used by the protected strings.
   *
   * PROT_READ|PROT_WRITE was originally fully allowed in sb_mprotect(), but
   * had to be removed due to limitation of libseccomp regarding intervals.
   *
   * There is a restriction on how much you can mprotect with R|W up to the
   * size of the canary.
   */
  ret = seccomp_rule_add_3(ctx, SCMP_ACT_ALLOW, SCMP_SYS(mprotect),
      SCMP_CMP(0, SCMP_CMP_LT, (intptr_t) pr_mem_base),
      SCMP_CMP(1, SCMP_CMP_LE, MALLOC_MP_LIM),
      SCMP_CMP(2, SCMP_CMP_EQ, PROT_READ|PROT_WRITE));
[...]
  ret = seccomp_rule_add_3(ctx, SCMP_ACT_ALLOW, SCMP_SYS(mprotect),
      SCMP_CMP(0, SCMP_CMP_GT, (intptr_t) pr_mem_base + pr_mem_size +
          MALLOC_MP_LIM),
      SCMP_CMP(1, SCMP_CMP_LE, MALLOC_MP_LIM),
      SCMP_CMP(2, SCMP_CMP_EQ, PROT_READ|PROT_WRITE));
[...]
===========================================

systemd also has some code that uses arithmetic comparisons in
https://github.com/systemd/systemd/blob/master/src/shared/seccomp-util.c ,
specifically for two purposes:

 - If you whitelist a range of address families for socket() using
   RestrictAddressFamilies, anything outside that range gets blocked with
   SCMP_CMP_LT/SCMP_CMP_GT.
 - If you restrict the use of scheduling classes, anything above the permitted
   class is blocked via SCMP_CMP_GT.

(Both of these, by the way, are for syscalls that silently discard the upper 32
bits of their arguments.)

The start of the second seccomp filter generated for a systemd unit with
"RestrictAddressFamilies=AF_INET AF_INET6" is:

===== filter 1 (57 instructions) =====
0001 if arch != X86_64: [true +54, false +0] -> ret ALLOW (syscalls: <TOO MANY TO LIST>)
0003 if nr < 0x40000000: [true +1, false +0]
0005   if nr != 0x00000029: [true +50, false +0] -> ret ALLOW (syscalls: <TOO MANY TO LIST>)
0007   if args[0].high != 0x00000000: [true +42, false +0]
0033     if args[0].high < 0x00000000: [true +3, false +0] -> ret ERRNO
0035     if args[0].low > 0x0000000a: [true +1, false +0] -> ret ERRNO
0036     if args[0].low >= 0x00000002: [true +1, false +0] -> ret ALLOW (syscalls: socket)
0037     ret ERRNO

So this filter will e.g. permit socket() calls in the range from 0x100000002 to
0x10000000a (and the kernel will ignore the high bit, meaning that in effect,
this filter grants access to families like AF_AX25); but as far as I can tell,
the other filter installed by systemd prevents this.


In the open-source users of libseccomp that I have been able to find on
codesearch.debian.net, this issue doesn't seem to have significant
impact; but someone might rely on this behavior, so I've decided to treat this
as a security bug.