Fix -e/--exec namespace resolution on kernels without CONFIG_CHECKPOINT_RESTORE

/proc/<pid>/task/<pid>/children doesn't exist on every kernel (confirmed
missing on a real Android target), so resolve_namespace_pid() fell back to
the outer bwrap pid itself and nsenter ended up with no namespace flags at
all. Add a portable fallback that scans /proc/<n>/stat for the child whose
ppid matches, the same information pstree uses to build its tree.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Gv3s5jckJKzh6JkMoi2Akz

CLAUDE.md

@@ -422,9 +422,21 @@ Source layout (all under `src/`):
   `/proc/<outer_pid>/ns/{pid,uts,ipc,cgroup}` against this process's own showed
   them identical, while only `mnt`/`user` differed. `resolve_namespace_pid()`
   reads `/proc/<pid>/task/<pid>/children` (the direct-children list `procfs`
-  exposes) to find that real inner process and joins *its* namespaces instead —
-  falls back to `pid` itself (best-effort, not fatal) if that file can't be
-  read. For each of `{mnt→--mount, uts→--uts, ipc→--ipc, pid→--pid,
+  exposes) to find that real inner process and joins *its* namespaces instead.
+  **Real bug found via testing on a real target device, not assumed**: that
+  file requires `CONFIG_CHECKPOINT_RESTORE`, which not every kernel enables —
+  confirmed absent (not just unreadable — the file doesn't exist at all) on a
+  real Android device, where `-e/--exec` then fell back to the outer `bwrap`
+  pid itself and failed outright (`nsenter: no namespace specified`, since
+  every namespace type either matched the outer process's own or couldn't be
+  read at all). Fixed by adding `find_child_by_scanning_proc()`, a portable
+  fallback used only when the children file is missing/empty: scans
+  `/proc/<n>/stat` for any process whose ppid field equals `pid` — the same
+  information `pstree` itself reads to build its tree, which is how the actual
+  sandboxed child was located and confirmed correct on the same device via a
+  manual `nsenter -t <child_pid> -a -- /bin/sh` before the fix was written.
+  Picks the lowest matching pid if more than one child exists, for a
+  deterministic result. For each of `{mnt→--mount, uts→--uts, ipc→--ipc, pid→--pid,
   cgroup→--cgroup, user→--user}` (`net` deliberately excluded — this project
   never isolates networking, see `bwrap.cpp` below), `readlink()`s both
   `/proc/<ns_pid>/ns/<type>` and `/proc/self/ns/<type>` and only passes

src/exec_session.cpp

@@ -18,9 +18,11 @@
 
 #include <algorithm>
 #include <array>
+#include <cctype>
 #include <filesystem>
 #include <fstream>
 #include <optional>
+#include <sstream>
 #include <system_error>
 
 #include <fmt/core.h>
@@ -31,6 +33,64 @@
 
 namespace {
 
+// Fast path: /proc/<pid>/task/<pid>/children lists direct children with no
+// scanning needed. Requires CONFIG_CHECKPOINT_RESTORE, which not every kernel
+// enables -- absent (not just unreadable), confirmed by direct testing on a
+// real Android target, where the file simply doesn't exist.
+std::optional<pid_t> find_child_via_children_file(pid_t pid) {
+    std::ifstream children(fmt::format("/proc/{}/task/{}/children", pid, pid));
+    pid_t child = 0;
+    if (children >> child && child > 0) {
+        return child;
+    }
+    return std::nullopt;
+}
+
+// Portable fallback for kernels without the children file: scans /proc/<n>/stat
+// for any process whose ppid field (the first whitespace-separated field after
+// comm's closing ')' -- comm itself is parenthesized and may contain spaces or
+// parens, so it can't just be split on whitespace) equals `pid`. This is the
+// same information `pstree` itself reads to build its tree -- confirmed by
+// direct testing: `pstree -p <bwrap_pid>` found the real sandboxed child on a
+// device where the children file was missing. Picks the lowest matching pid if
+// more than one child exists, for a deterministic result.
+std::optional<pid_t> find_child_by_scanning_proc(pid_t pid) {
+    std::error_code ec;
+    std::optional<pid_t> found;
+    for (const auto& entry : std::filesystem::directory_iterator("/proc", ec)) {
+        if (ec) {
+            break;
+        }
+        const std::string name = entry.path().filename().string();
+        if (name.empty() || !std::all_of(name.begin(), name.end(),
+                                          [](unsigned char c) { return std::isdigit(c) != 0; })) {
+            continue;
+        }
+
+        std::ifstream stat_file(entry.path() / "stat");
+        std::string line;
+        if (!std::getline(stat_file, line)) {
+            continue;
+        }
+        auto close_paren = line.rfind(')');
+        if (close_paren == std::string::npos) {
+            continue;
+        }
+        std::istringstream rest(line.substr(close_paren + 1));
+        std::string state;
+        pid_t ppid = 0;
+        if (!(rest >> state >> ppid) || ppid != pid) {
+            continue;
+        }
+
+        pid_t candidate = std::stoi(name);
+        if (!found || candidate < *found) {
+            found = candidate;
+        }
+    }
+    return found;
+}
+
 // bwrap's own outer process (the one tracked in the session pid file) sets up the
 // mount/user namespaces itself, then clone()s the actual sandboxed command into
 // fresh pid/uts/ipc/cgroup/net namespaces -- clone()'s namespace-creation flags
@@ -43,10 +103,13 @@ namespace {
 // child can't be determined -- callers still get a mount/user-namespace join out
 // of that, just not the rest.
 pid_t resolve_namespace_pid(pid_t pid) {
-    std::ifstream children(fmt::format("/proc/{}/task/{}/children", pid, pid));
-    pid_t child = 0;
-    if (children >> child && child > 0) {
-        return child;
+    if (auto child = find_child_via_children_file(pid)) {
+        return *child;
+    }
+    if (auto child = find_child_by_scanning_proc(pid)) {
+        spdlog::debug("pid {}: found sandboxed child {} by scanning /proc (no .../task/.../children file)",
+                      pid, *child);
+        return *child;
     }
     spdlog::debug("could not determine pid {}'s sandboxed child process; joining its own namespaces only",
                   pid);