bpf: don't let ldimm64 leak map addresses on unprivileged
authorDaniel Borkmann <daniel@iogearbox.net>
Sun, 7 May 2017 22:04:09 +0000 (00:04 +0200)
committerDavid S. Miller <davem@davemloft.net>
Mon, 8 May 2017 19:06:46 +0000 (15:06 -0400)
The patch fixes two things at once:

1) It checks the env->allow_ptr_leaks and only prints the map address to
   the log if we have the privileges to do so, otherwise it just dumps 0
   as we would when kptr_restrict is enabled on %pK. Given the latter is
   off by default and not every distro sets it, I don't want to rely on
   this, hence the 0 by default for unprivileged.

2) Printing of ldimm64 in the verifier log is currently broken in that
   we don't print the full immediate, but only the 32 bit part of the
   first insn part for ldimm64. Thus, fix this up as well; it's okay to
   access, since we verified all ldimm64 earlier already (including just
   constants) through replace_map_fd_with_map_ptr().

Fixes: 1be7f75d1668 ("bpf: enable non-root eBPF programs")
Fixes: cbd357008604 ("bpf: verifier (add ability to receive verification log)")
Reported-by: Jann Horn <jannh@google.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
kernel/bpf/verifier.c

index c2ff608c1984ed3f7e5031d5f9fbcfe479f7fc67..c5b56c92f8e255d1b13634ad07c460969484f2f4 100644 (file)
@@ -298,7 +298,8 @@ static const char *const bpf_jmp_string[16] = {
        [BPF_EXIT >> 4] = "exit",
 };
 
-static void print_bpf_insn(struct bpf_insn *insn)
+static void print_bpf_insn(const struct bpf_verifier_env *env,
+                          const struct bpf_insn *insn)
 {
        u8 class = BPF_CLASS(insn->code);
 
@@ -362,9 +363,19 @@ static void print_bpf_insn(struct bpf_insn *insn)
                                insn->code,
                                bpf_ldst_string[BPF_SIZE(insn->code) >> 3],
                                insn->src_reg, insn->imm);
-               } else if (BPF_MODE(insn->code) == BPF_IMM) {
-                       verbose("(%02x) r%d = 0x%x\n",
-                               insn->code, insn->dst_reg, insn->imm);
+               } else if (BPF_MODE(insn->code) == BPF_IMM &&
+                          BPF_SIZE(insn->code) == BPF_DW) {
+                       /* At this point, we already made sure that the second
+                        * part of the ldimm64 insn is accessible.
+                        */
+                       u64 imm = ((u64)(insn + 1)->imm << 32) | (u32)insn->imm;
+                       bool map_ptr = insn->src_reg == BPF_PSEUDO_MAP_FD;
+
+                       if (map_ptr && !env->allow_ptr_leaks)
+                               imm = 0;
+
+                       verbose("(%02x) r%d = 0x%llx\n", insn->code,
+                               insn->dst_reg, (unsigned long long)imm);
                } else {
                        verbose("BUG_ld_%02x\n", insn->code);
                        return;
@@ -2853,7 +2864,7 @@ static int do_check(struct bpf_verifier_env *env)
 
                if (log_level) {
                        verbose("%d: ", insn_idx);
-                       print_bpf_insn(insn);
+                       print_bpf_insn(env, insn);
                }
 
                err = ext_analyzer_insn_hook(env, insn_idx, prev_insn_idx);