Add FINGER_BAN_ALLOWLIST to exempt trusted front-end IPs from banning

The per-IP ban tracker treats every globally-routable client equally, but an
aggregating front-end like the finger-web proxy funnels the whole internet's
federated lookups through a single IP. A burst from any one client of the proxy
(or a load test) is then attributed to the proxy's IP and, once it crosses the
failure threshold, the daemon blocks the proxy — taking out finger lookups for
everyone. Per-client abuse protection for the proxied path belongs in the proxy
(which now rate-limits per real client IP), so the daemon should trust it.

Add a FINGER_BAN_ALLOWLIST env var (comma-separated IPs). Allowlisted addresses
are marked non-trackable in the listener, so their connections are never blocked
and never recorded as offenses. Unset = unchanged behaviour.

- parse_ip_allowlist() in ban.cpp (trims entries, skips blanks) + unit tests
- listener() consults the set when computing 'trackable'
- documented in docker-compose.yml and DOCKER.md

DOCKER.md

@@ -64,6 +64,74 @@ docker run -d \
   ghcr.io/waffle2k/finger:latest
 ```
 
+## Abuse protection & client IPs (important)
+
+The daemon bans source IPs that rack up repeated failed lookups (scanners, SIP/
+HTTP probes, username guessers) -- see the "Abuse protection" section in the
+main [README.md](README.md). That protection is **per source IP**, so it only
+works if the container can see the *real* client IP.
+
+Under Docker's **default bridge networking this is not the case**: published
+ports are NAT'd so every external client arrives with the bridge gateway as its
+source (e.g. `172.20.0.1`). The daemon would see one IP for the entire internet.
+By design it treats private/RFC1918 addresses as untrackable, so rather than
+blocking everyone at once, banning simply becomes **inert** under bridge
+networking.
+
+To make abuse protection actually work in Docker, give the container the real
+client IP. In order of preference:
+
+1. **Host networking (recommended).** Add `network_mode: host` to the service
+   (and drop the `ports:` mapping -- it's ignored). The daemon then binds the
+   host's port 79 directly and sees real client IPs. This is what
+   `docker-compose.yml` in this repo uses.
+
+   Note: under host networking the container shares the host network namespace,
+   which uses the host's privileged-port rule -- so the image's non-root user
+   (UID 1000) **cannot bind port 79** and the daemon fails to listen silently.
+   Either run as root (`user: "0:0"`, as below) or `setcap
+   cap_net_bind_service=+ep` on the binary in the image to keep it non-root.
+
+   ```yaml
+   services:
+     finger:
+       image: ghcr.io/waffle2k/finger:latest
+       network_mode: host
+       user: "0:0"   # bind privileged port 79 under host networking
+       volumes:
+         - ./users:/var/finger/users
+       restart: unless-stopped
+   ```
+
+2. **macvlan network.** Give the container its own IP on the LAN. More setup,
+   but keeps the container off host networking.
+
+3. **Disable the userland proxy host-wide** (`/etc/docker/daemon.json`:
+   `{"userland-proxy": false}`, then restart dockerd). iptables DNAT then
+   preserves the source IP on published ports. This is a host-wide change that
+   restarts every container on the host -- avoid it on busy multi-service hosts.
+
+Note: bans are in-memory, so they reset when the container restarts -- the same
+trade-off as any single-process deployment.
+
+### Allowlisting a trusted front-end (`FINGER_BAN_ALLOWLIST`)
+
+Set `FINGER_BAN_ALLOWLIST` to a comma-separated list of client IPs that should
+never be tracked or banned. This is for trusted aggregating front-ends: the
+[`finger-web`](https://github.com/waffle2k/finger-web) proxy, for example,
+funnels every federated lookup through a single IP, so a burst from any one of
+*its* clients would otherwise be attributed to the proxy and ban it for
+everyone. Per-client abuse protection for that path lives in the proxy (it rate
+limits per real client IP), so the daemon should trust the proxy IP:
+
+```yaml
+environment:
+  - FINGER_BAN_ALLOWLIST=203.0.113.10,2001:db8::10
+```
+
+Addresses are matched verbatim against the connecting socket's address, so use
+canonical forms. Leave it unset for a directly-exposed daemon.
+
 ## Docker Architecture
 
 ### Multi-stage Build

Dockerfile

@@ -29,14 +29,19 @@ RUN meson compile -C builddir
 # Run tests to ensure quality
 RUN meson test -C builddir
 
-# Runtime stage - minimal Alpine Linux
-FROM alpine:latest
+# Runtime stage — match builder's glibc (Alpine/musl is incompatible
+# with our dynamically linked binary, esp. fortify _chk symbols).
+FROM ubuntu:24.04
 
-# Install runtime dependencies (if any)
-RUN apk add --no-cache \
-    libstdc++ \
-    && addgroup -g 1000 finger \
-    && adduser -D -s /bin/sh -u 1000 -G finger finger
+ENV DEBIAN_FRONTEND=noninteractive
+
+RUN apt-get update && apt-get install -y --no-install-recommends \
+    libstdc++6 \
+    netcat-openbsd \
+    && rm -rf /var/lib/apt/lists/* \
+    && (userdel -r ubuntu 2>/dev/null || true) \
+    && groupadd -g 1000 finger \
+    && useradd -m -u 1000 -g finger -s /bin/sh finger
 
 # Copy the compiled binary from builder stage
 COPY --from=builder /app/builddir/finger /usr/local/bin/finger
@@ -56,7 +61,7 @@ EXPOSE 79
 
 # Add health check
 HEALTHCHECK --interval=30s --timeout=3s --start-period=5s --retries=3 \
-    CMD nc -z localhost 79 || exit 1
+    CMD nc -w 1 127.0.0.1 79 < /dev/null || exit 1
 
 # Set metadata labels
 LABEL org.opencontainers.image.title="finger"

FREEBSD.md

@@ -51,7 +51,7 @@ Create `/usr/local/etc/rc.d/fingerd`:
 name="fingerd"
 rcvar="fingerd_enable"
 command="/usr/sbin/daemon"
-command_args="-f -p /var/run/fingerd.pid /usr/local/bin/finger"
+command_args="-f -p /var/run/fingerd.pid -o /var/log/fingerd.log /usr/local/bin/finger"
 pidfile="/var/run/fingerd.pid"
 # procname must be the full path so rc.subr can match it against ps output
 procname="/usr/local/bin/finger"

README.md

@@ -59,3 +59,13 @@ and execute `docker compose up -d`
 
 # Setting your status
 within the `./users` directory, create a file named after the user you wish to have a response. That's it!
+
+# Abuse protection
+Most traffic on port 79 is not finger at all -- HTTP and SIP probes, TLS
+handshakes, and username-guessing scanners. None of these resolve to a plan
+file, so the daemon treats any request that fails to read a plan as an
+"offense" and timestamps it against the source IP. When an IP records more than
+3 failures within a rolling 24-hour window, its connections are dropped
+(without being read or answered) until those failures age back out of the
+window. Legitimate lookups that hit a real plan never count against an IP. All
+state is in-memory; thresholds live in `BanTracker::Config` (`ban.hpp`).

ban.cpp

@@ -0,0 +1,105 @@
+#include "ban.hpp"
+
+#include <cstdint>
+#include <string>
+
+bool is_bannable_address(const boost::asio::ip::address &addr) {
+  if (addr.is_loopback() || addr.is_unspecified() || addr.is_multicast()) {
+    return false;
+  }
+
+  if (addr.is_v4()) {
+    const std::uint32_t a = addr.to_v4().to_uint();
+    if ((a & 0xFF000000u) == 0x0A000000u) return false; // 10.0.0.0/8
+    if ((a & 0xFFF00000u) == 0xAC100000u) return false; // 172.16.0.0/12
+    if ((a & 0xFFFF0000u) == 0xC0A80000u) return false; // 192.168.0.0/16
+    if ((a & 0xFFFF0000u) == 0xA9FE0000u) return false; // 169.254.0.0/16 link-local
+    if ((a & 0xFFC00000u) == 0x64400000u) return false; // 100.64.0.0/10 CGNAT / Tailscale
+    return true;
+  }
+
+  // IPv6: drop link-local (fe80::/10) and unique-local (fc00::/7).
+  const auto v6 = addr.to_v6();
+  if (v6.is_link_local()) {
+    return false;
+  }
+  if ((v6.to_bytes()[0] & 0xFEu) == 0xFCu) {
+    return false;
+  }
+  return true;
+}
+
+std::unordered_set<std::string> parse_ip_allowlist(std::string_view csv) {
+  std::unordered_set<std::string> out;
+  std::size_t start = 0;
+  while (start <= csv.size()) {
+    const std::size_t comma = csv.find(',', start);
+    const std::size_t end =
+        (comma == std::string_view::npos) ? csv.size() : comma;
+    std::string_view tok = csv.substr(start, end - start);
+    const std::size_t a = tok.find_first_not_of(" \t\r\n");
+    if (a != std::string_view::npos) {
+      const std::size_t b = tok.find_last_not_of(" \t\r\n");
+      out.emplace(tok.substr(a, b - a + 1));
+    }
+    if (comma == std::string_view::npos) {
+      break;
+    }
+    start = comma + 1;
+  }
+  return out;
+}
+
+namespace {
+// Count timestamps that fall within (now - window, now]. The deque is kept in
+// ascending order, so the in-window entries are always a suffix.
+int count_in_window(const std::deque<BanTracker::clock::time_point> &ts,
+                    BanTracker::clock::time_point now,
+                    BanTracker::clock::duration window) {
+  const auto cutoff = now - window;
+  int count = 0;
+  for (auto it = ts.rbegin(); it != ts.rend() && *it > cutoff; ++it) {
+    ++count;
+  }
+  return count;
+}
+} // namespace
+
+bool BanTracker::is_blocked(const std::string &ip, clock::time_point now) const {
+  auto it = offenders_.find(ip);
+  if (it == offenders_.end()) {
+    return false;
+  }
+  return count_in_window(it->second, now, cfg_.window) > cfg_.threshold;
+}
+
+BanTracker::OffenseResult
+BanTracker::record_offense(const std::string &ip, clock::time_point now) {
+  auto &ts = offenders_[ip];
+  const auto cutoff = now - cfg_.window;
+
+  // Drop this IP's timestamps that have aged out of the window.
+  while (!ts.empty() && ts.front() <= cutoff) {
+    ts.pop_front();
+  }
+
+  ts.push_back(now);
+
+  const int count = static_cast<int>(ts.size());
+  return {count, count > cfg_.threshold};
+}
+
+void BanTracker::sweep(clock::time_point now) {
+  const auto cutoff = now - cfg_.window;
+  for (auto it = offenders_.begin(); it != offenders_.end();) {
+    auto &ts = it->second;
+    while (!ts.empty() && ts.front() <= cutoff) {
+      ts.pop_front();
+    }
+    if (ts.empty()) {
+      it = offenders_.erase(it);
+    } else {
+      ++it;
+    }
+  }
+}

ban.hpp

@@ -0,0 +1,90 @@
+#pragma once
+
+#include <boost/asio/ip/address.hpp>
+#include <chrono>
+#include <cstddef>
+#include <deque>
+#include <string>
+#include <string_view>
+#include <unordered_map>
+#include <unordered_set>
+
+// BanTracker records the timestamps of "offenses" -- requests that are
+// obviously not finger queries -- per client IP, over a rolling time window.
+// When an IP has more than `threshold` offenses still inside the window, it is
+// blocked and its connections are dropped. Offense timestamps older than the
+// window are pruned, so a blocked IP automatically frees itself once its old
+// offenses age out.
+//
+// All state is in-memory: the daemon runs a single io_context thread, so every
+// call happens on the same thread and no locking is required. Time is passed
+// in as a steady_clock time_point rather than read internally, so the logic is
+// deterministic and unit-testable.
+class BanTracker {
+public:
+  using clock = std::chrono::steady_clock;
+
+  struct Config {
+    int threshold = 3;                              // block when offenses exceed this
+    clock::duration window = std::chrono::hours(24); // rolling window length
+  };
+
+  struct OffenseResult {
+    int count;    // offenses within the window, including this one
+    bool blocked; // true if the IP is now blocked (count > threshold)
+  };
+
+  BanTracker() = default;
+  explicit BanTracker(Config cfg) : cfg_(cfg) {}
+
+  // True if ip currently has more than `threshold` offenses inside the rolling
+  // window. Does not mutate state.
+  bool is_blocked(const std::string &ip, clock::time_point now) const;
+
+  // Record one offense from ip at `now`. Prunes that IP's expired timestamps,
+  // appends this one, and reports the in-window count and whether it is now
+  // blocked.
+  OffenseResult record_offense(const std::string &ip, clock::time_point now);
+
+  // Drop timestamps older than the window across all IPs, removing any IP left
+  // with no offenses. Safe to call periodically to keep the map bounded.
+  void sweep(clock::time_point now);
+
+  // Number of tracked IPs (for introspection and tests).
+  std::size_t tracked() const { return offenders_.size(); }
+
+  const Config &config() const { return cfg_; }
+
+private:
+  Config cfg_{};
+  // Per-IP offense timestamps, kept in ascending order (steady_clock is
+  // monotonic, so appends are always newest-last).
+  std::unordered_map<std::string, std::deque<clock::time_point>> offenders_;
+};
+
+// Whether a client address is meaningful to track and ban. Only globally
+// routable unicast addresses qualify. Loopback, RFC1918 private, CGNAT
+// (100.64/10), link-local, IPv6 unique-local, and multicast addresses all
+// return false.
+//
+// This matters because the daemon can only ban what it can see: behind Docker's
+// default bridge networking every external client is SNAT'd to the bridge
+// gateway (a 172.16/12 address), so banning per source IP would collapse all
+// clients into one and block everyone. Skipping non-global addresses makes
+// banning correct where the real client IP is visible (e.g. the FreeBSD jail,
+// where pf rdr preserves it) and inert where it is not (Docker bridge), with no
+// deployment-specific configuration.
+bool is_bannable_address(const boost::asio::ip::address &addr);
+
+// Parse a comma-separated list of IP addresses (the value of the
+// FINGER_BAN_ALLOWLIST env var) into a set of address strings. Whitespace
+// around each entry is trimmed and empty entries are skipped. The strings are
+// matched verbatim against boost::asio's address().to_string() output, so use
+// canonical forms (e.g. "147.182.255.203", "2a01:4f8:190:7447::2").
+//
+// Allowlisting exists for trusted aggregating front-ends — notably the
+// finger-web proxy, which funnels every federated lookup through one IP. Without
+// it, a burst from any single client of the proxy is attributed to the proxy's
+// IP and bans the proxy for everyone; per-client abuse protection for that path
+// lives in the proxy instead.
+std::unordered_set<std::string> parse_ip_allowlist(std::string_view csv);

docker-compose.yml

@@ -3,19 +3,42 @@ version: '3.8'
 services:
   finger:
     build: .
-    ports:
-      - "79:79"
+    # IMPORTANT: host networking is what lets the daemon's abuse protection
+    # work. Under Docker's default bridge networking every external client is
+    # SNAT'd to the bridge gateway (a 172.16/12 address), so the daemon sees a
+    # single source IP for everyone -- the per-IP ban logic can't tell clients
+    # apart and (by design) treats that private address as untrackable, leaving
+    # banning inert. Host networking exposes the real client IP, so repeat
+    # offenders actually get blocked.
+    network_mode: host
+    # Under host networking the container shares the host net namespace, which
+    # uses the host's privileged-port rule -- so the image's non-root user
+    # (UID 1000) cannot bind port 79 and the daemon fails to listen silently.
+    # Run as root to bind it. (Alternative: setcap cap_net_bind_service on the
+    # binary in the image to keep it non-root.)
+    user: "0:0"
+    # FINGER_BAN_ALLOWLIST: comma-separated client IPs that are never tracked or
+    # banned. Use it for trusted aggregating front-ends — e.g. the finger-web
+    # proxy, which funnels every federated lookup through one IP; without an
+    # allowlist a burst from any single client of the proxy is attributed to the
+    # proxy and bans it for everyone (per-client abuse protection for that path
+    # lives in the proxy). Leave unset for a directly-exposed daemon.
+    # environment:
+    #   - FINGER_BAN_ALLOWLIST=203.0.113.10,2001:db8::10
     volumes:
       - ./users:/var/finger/users
     restart: unless-stopped
     healthcheck:
-      test: ["CMD", "nc", "-z", "localhost", "79"]
+      test: ["CMD-SHELL", "nc -w 1 127.0.0.1 79 < /dev/null || exit 1"]
       interval: 30s
       timeout: 10s
       retries: 3
       start_period: 40s
 
-  # Example using published image instead of building locally
+  # Bridge-networking alternative (quick local testing only). NOTE: with this
+  # mode the daemon only ever sees the bridge gateway IP, so abuse protection
+  # is effectively disabled. Prefer host networking above for any public-facing
+  # deployment.
   # finger:
   #   image: ghcr.io/waffle2k/finger:latest
   #   ports:

handler.cpp

@@ -1,4 +1,6 @@
 #include "handler.hpp"
+#include <algorithm>
+#include <cctype>
 #include <filesystem>
 #include <fstream>
 #include <string_view>
@@ -65,8 +67,16 @@ std::string process(const std::string &username, const IFilesystemWrapper &fs,
     return std::string("InvalidInput: ") + e.what() + std::string("\r\n");
   }
 
+  // Plan-file lookup is case-insensitive: normalise the requested name to
+  // lower-case so e.g. "Pete" resolves the on-disk "pete" plan. Plan filenames
+  // are always lower-case; the original spelling is still echoed back below
+  // when no plan exists.
+  std::string lookup = username;
+  std::transform(lookup.begin(), lookup.end(), lookup.begin(),
+                 [](unsigned char c) { return std::tolower(c); });
+
   // Attempt to open the plan file (if any) and return the contents as a string
-  std::filesystem::path planPath = basepath / username;
+  std::filesystem::path planPath = basepath / lookup;
 
   // Check if the plan file exists using the filesystem wrapper
   if (!fs.exists(planPath)) {

main.cpp

@@ -1,15 +1,21 @@
 #include <iostream>
 
+#include <boost/asio/as_tuple.hpp>
 #include <boost/asio/co_spawn.hpp>
 #include <boost/asio/deferred.hpp>
 #include <boost/asio/detached.hpp>
 #include <boost/asio/io_context.hpp>
 #include <boost/asio/ip/tcp.hpp>
 #include <boost/asio/signal_set.hpp>
+#include <boost/asio/steady_timer.hpp>
 #include <boost/asio/write.hpp>
+#include <chrono>
 #include <cstdio>
-#include <syslog.h>
+#include <cstdlib>
+#include <string>
+#include <unordered_set>
 
+#include "ban.hpp"
 #include "handler.hpp"
 
 using boost::asio::awaitable;
@@ -23,35 +29,74 @@ awaitable<std::string> dofinger(const std::string &username) {
   co_return process(username);
 }
 
-awaitable<void> echo(tcp::socket socket, std::string client_addr) {
+awaitable<void> echo(tcp::socket socket, std::string client_addr, bool trackable,
+                     BanTracker &bans) {
   try {
+    auto now = std::chrono::steady_clock::now();
+
+    // An IP that has racked up too many failed lookups (scanners, username
+    // guessers, non-finger junk) is dropped without being read or answered.
+    // Only globally-routable addresses are tracked: behind Docker's bridge
+    // every client is SNAT'd to the gateway, so banning there would block
+    // everyone at once (see is_bannable_address()).
+    if (trackable && bans.is_blocked(client_addr, now)) {
+      std::printf("finger drop from %s: blocked\n", client_addr.c_str());
+      co_return;
+    }
+
     char data[1024];
-    auto bytes_read =
-        co_await socket.async_read_some(boost::asio::buffer(data), deferred);
+    auto [read_ec, bytes_read] = co_await socket.async_read_some(
+        boost::asio::buffer(data), boost::asio::as_tuple(deferred));
+    if (read_ec) {
+      // Client hung up before sending a request: health checks (which connect
+      // and immediately close), port scanners, and reset connections all land
+      // here. This is normal -- don't log it as an exception.
+      co_return;
+    }
     std::string username(data, bytes_read);
     // Remove trailing \r\n characters
     while (!username.empty() &&
            (username.back() == '\r' || username.back() == '\n')) {
       username.pop_back();
     }
-    syslog(LOG_INFO, "finger request from %s for user '%s'",
+    std::printf("finger request from %s for user '%s'\n",
                 client_addr.c_str(), username.c_str());
     auto response = co_await dofinger(username);
-    if (response.compare(std::string(username)) == 0) {
-      // No plan found
-      co_await async_write(
-          socket, boost::asio::buffer(std::string("No plan found\r\n")),
-          deferred);
+
+    // A "failure" is simply any request that does not resolve to a readable
+    // plan file: an unknown user, rejected input, or non-finger junk. Each
+    // failure is timestamped against the client IP; once an IP exceeds the
+    // threshold within the rolling window, the is_blocked() check above starts
+    // dropping its connections. This also frustrates username guessing.
+    bool plan_served =
+        response != username && response.rfind("InvalidInput:", 0) != 0;
+    if (!plan_served) {
+      if (trackable) {
+        auto res = bans.record_offense(client_addr, now);
+        std::printf("finger miss from %s for '%s' (%d failures in window)%s\n",
+                    client_addr.c_str(), username.c_str(), res.count,
+                    res.blocked ? " -- now blocked" : "");
+      } else {
+        std::printf("finger miss from %s for '%s' (not tracked)\n",
+                    client_addr.c_str(), username.c_str());
+      }
+      // Best-effort reply; ignore write errors (the client may have already
+      // gone away).
+      co_await async_write(socket,
+                           boost::asio::buffer(std::string("No plan found\r\n")),
+                           boost::asio::as_tuple(deferred));
       co_return;
     }
-    co_await async_write(socket, boost::asio::buffer(response), deferred);
+    co_await async_write(socket, boost::asio::buffer(response),
+                         boost::asio::as_tuple(deferred));
     co_return;
   } catch (std::exception &e) {
-    syslog(LOG_ERR, "echo exception: %s", e.what());
+    std::printf("echo exception: %s\n", e.what());
   }
 }
 
-awaitable<void> listener() {
+awaitable<void> listener(BanTracker &bans,
+                         const std::unordered_set<std::string> &allowlist) {
   auto executor = co_await this_coro::executor;
   tcp::acceptor acceptor(executor, {tcp::v4(), 79});
   for (;;) {
@@ -60,24 +105,50 @@ awaitable<void> listener() {
     auto endpoint = socket.remote_endpoint(ec);
     std::string client_addr =
         ec ? std::string("unknown") : endpoint.address().to_string();
-    co_spawn(executor, echo(std::move(socket), std::move(client_addr)),
+    // Allowlisted IPs (trusted aggregating front-ends like the finger-web
+    // proxy) are never tracked, so their bursts neither block them nor count
+    // as offenses.
+    bool trackable = !ec && is_bannable_address(endpoint.address()) &&
+                     allowlist.find(client_addr) == allowlist.end();
+    co_spawn(executor,
+             echo(std::move(socket), std::move(client_addr), trackable, bans),
              detached);
   }
 }
 
+// Periodically prune offense records that have aged out of the window so the
+// tracker's memory stays bounded even for IPs that never reconnect.
+awaitable<void> sweeper(BanTracker &bans) {
+  boost::asio::steady_timer timer(co_await this_coro::executor);
+  for (;;) {
+    timer.expires_after(std::chrono::minutes(10));
+    co_await timer.async_wait(deferred);
+    bans.sweep(std::chrono::steady_clock::now());
+  }
+}
+
 int main() {
-  openlog("fingerd", LOG_PID, LOG_DAEMON);
+  // Line-buffer stdout so docker logs / tail -f see entries in real time.
+  std::setvbuf(stdout, nullptr, _IOLBF, 0);
   try {
     boost::asio::io_context io_context(1);
+    BanTracker bans;
+
+    const char *allow_env = std::getenv("FINGER_BAN_ALLOWLIST");
+    const std::unordered_set<std::string> allowlist =
+        parse_ip_allowlist(allow_env ? allow_env : "");
+    for (const auto &ip : allowlist) {
+      std::printf("ban allowlist: %s (never tracked or blocked)\n", ip.c_str());
+    }
 
     boost::asio::signal_set signals(io_context, SIGINT, SIGTERM);
     signals.async_wait([&](auto, auto) { io_context.stop(); });
 
-    co_spawn(io_context, listener(), detached);
+    co_spawn(io_context, listener(bans, allowlist), detached);
+    co_spawn(io_context, sweeper(bans), detached);
 
     io_context.run();
   } catch (std::exception &e) {
-    syslog(LOG_ERR, "fatal exception: %s", e.what());
+    std::printf("fatal exception: %s\n", e.what());
   }
-  closelog();
 }

meson.build

@@ -14,7 +14,7 @@ gtest_dep = dependency('gtest', main : true, required : true)
 gmock_dep = dependency('gmock', main : true, required : true)
 
 executable('finger',
-  'main.cpp','handler.cpp',
+  'main.cpp','handler.cpp','ban.cpp',
   dependencies : [boost_dep, threads_dep],
   install : true)
 
@@ -33,7 +33,13 @@ test_real_fs_exe = executable('test_handler_real_filesystem',
   'test_handler_real_filesystem.cpp', 'handler.cpp',
   dependencies : [boost_dep, threads_dep, gtest_dep, gmock_dep])
 
+# Ban tracker test executable
+test_ban_exe = executable('test_ban',
+  'test_ban.cpp', 'ban.cpp',
+  dependencies : [boost_dep, threads_dep, gtest_dep, gmock_dep])
+
 # Register the tests
 test('handler_tests', test_exe)
 test('handler_mock_tests', test_mock_exe)
 test('handler_real_filesystem_tests', test_real_fs_exe)
+test('ban_tests', test_ban_exe)

test_ban.cpp

@@ -0,0 +1,152 @@
+#include "ban.hpp"
+#include <boost/asio/ip/address.hpp>
+#include <gtest/gtest.h>
+
+using namespace std::chrono_literals;
+using clock_t_ = BanTracker::clock;
+
+// Work well away from the steady_clock epoch so that subtracting the window
+// never underflows and default-constructed time_points are unambiguous.
+static const clock_t_::time_point kBase = clock_t_::time_point{} + 1000h;
+
+TEST(BanTracker, UnknownIpIsNotBlocked) {
+  BanTracker bt;
+  EXPECT_FALSE(bt.is_blocked("1.2.3.4", kBase));
+}
+
+TEST(BanTracker, BlocksOnlyAfterMoreThanThreshold) {
+  BanTracker bt; // default threshold = 3, so block on the 4th failure
+  EXPECT_FALSE(bt.record_offense("1.2.3.4", kBase).blocked); // 1
+  EXPECT_FALSE(bt.record_offense("1.2.3.4", kBase).blocked); // 2
+  EXPECT_FALSE(bt.record_offense("1.2.3.4", kBase).blocked); // 3
+  EXPECT_FALSE(bt.is_blocked("1.2.3.4", kBase));
+  auto r = bt.record_offense("1.2.3.4", kBase); // 4
+  EXPECT_TRUE(r.blocked);
+  EXPECT_EQ(r.count, 4);
+  EXPECT_TRUE(bt.is_blocked("1.2.3.4", kBase));
+}
+
+TEST(BanTracker, TracksEachIpIndependently) {
+  BanTracker bt;
+  for (int i = 0; i < 4; ++i) {
+    bt.record_offense("1.1.1.1", kBase);
+  }
+  EXPECT_TRUE(bt.is_blocked("1.1.1.1", kBase));
+  EXPECT_FALSE(bt.is_blocked("2.2.2.2", kBase));
+}
+
+TEST(BanTracker, OffensesAgeOutOfRollingWindow) {
+  BanTracker bt;
+  // Four failures spread over a couple of hours -> blocked.
+  for (int i = 0; i < 4; ++i) {
+    bt.record_offense("1.2.3.4", kBase + i * 1h);
+  }
+  EXPECT_TRUE(bt.is_blocked("1.2.3.4", kBase + 3h));
+
+  // 24h after the first failure, that one drops out of the window: only 3
+  // remain, so the IP is no longer blocked.
+  EXPECT_FALSE(bt.is_blocked("1.2.3.4", kBase + 24h + 1min));
+}
+
+TEST(BanTracker, WindowBoundaryIsExclusiveAtCutoff) {
+  BanTracker bt;
+  // Exactly window-old timestamps are pruned (cutoff is inclusive of <=).
+  bt.record_offense("1.2.3.4", kBase);
+  auto r = bt.record_offense("1.2.3.4", kBase + 24h);
+  EXPECT_EQ(r.count, 1); // the kBase entry was pruned before appending
+}
+
+TEST(BanTracker, SweepRemovesFullyExpiredIp) {
+  BanTracker bt;
+  for (int i = 0; i < 4; ++i) {
+    bt.record_offense("1.2.3.4", kBase);
+  }
+  EXPECT_EQ(bt.tracked(), 1u);
+  bt.sweep(kBase + 24h + 1min); // all offenses aged out
+  EXPECT_EQ(bt.tracked(), 0u);
+}
+
+TEST(BanTracker, SweepKeepsStillActiveIp) {
+  BanTracker bt;
+  for (int i = 0; i < 4; ++i) {
+    bt.record_offense("1.2.3.4", kBase);
+  }
+  bt.sweep(kBase + 1h); // still inside the window
+  EXPECT_EQ(bt.tracked(), 1u);
+  EXPECT_TRUE(bt.is_blocked("1.2.3.4", kBase + 1h));
+}
+
+TEST(BanTracker, RespectsCustomConfig) {
+  BanTracker bt(BanTracker::Config{/*threshold=*/1, /*window=*/1h});
+  EXPECT_FALSE(bt.record_offense("9.9.9.9", kBase).blocked); // 1, not > 1
+  EXPECT_TRUE(bt.record_offense("9.9.9.9", kBase).blocked);  // 2 > 1
+  EXPECT_TRUE(bt.is_blocked("9.9.9.9", kBase));
+  EXPECT_FALSE(bt.is_blocked("9.9.9.9", kBase + 1h + 1min)); // window elapsed
+}
+
+static bool bannable(const char *ip) {
+  return is_bannable_address(boost::asio::ip::make_address(ip));
+}
+
+TEST(BannableAddress, GlobalIpv4IsBannable) {
+  EXPECT_TRUE(bannable("8.8.8.8"));
+  EXPECT_TRUE(bannable("192.184.167.198")); // a real scanner seen in the logs
+  EXPECT_TRUE(bannable("1.2.3.4"));
+}
+
+TEST(BannableAddress, PrivateAndLocalIpv4AreNotBannable) {
+  EXPECT_FALSE(bannable("127.0.0.1"));   // loopback
+  EXPECT_FALSE(bannable("10.1.2.3"));    // 10/8
+  EXPECT_FALSE(bannable("172.20.0.1"));  // Docker bridge gateway (172.16/12)
+  EXPECT_FALSE(bannable("172.31.255.1"));
+  EXPECT_FALSE(bannable("192.168.1.104")); // the finger jail's own LAN IP
+  EXPECT_FALSE(bannable("169.254.10.1")); // link-local
+  EXPECT_FALSE(bannable("224.0.0.1"));    // multicast
+}
+
+TEST(BannableAddress, CgnatRangeIsNotBannable) {
+  EXPECT_FALSE(bannable("100.64.0.1"));   // bottom of 100.64/10 (CGNAT/Tailscale)
+  EXPECT_FALSE(bannable("100.127.255.1")); // top of the range
+  EXPECT_TRUE(bannable("100.63.255.1"));   // just below the range -> public
+  EXPECT_TRUE(bannable("100.128.0.1"));    // just above the range -> public
+}
+
+TEST(BannableAddress, Ipv6Classification) {
+  EXPECT_TRUE(bannable("2001:4860:4860::8888")); // global
+  EXPECT_FALSE(bannable("::1"));                 // loopback
+  EXPECT_FALSE(bannable("fe80::1"));             // link-local
+  EXPECT_FALSE(bannable("fc00::1"));             // unique-local
+  EXPECT_FALSE(bannable("fd12:3456::1"));        // unique-local
+}
+
+TEST(IpAllowlist, ParsesCommaSeparatedTrimmedEntries) {
+  auto a = parse_ip_allowlist("147.182.255.203, 10.0.0.1 ,\t2a01:4f8:190:7447::2");
+  EXPECT_EQ(a.size(), 3u);
+  EXPECT_TRUE(a.count("147.182.255.203"));
+  EXPECT_TRUE(a.count("10.0.0.1"));
+  EXPECT_TRUE(a.count("2a01:4f8:190:7447::2"));
+}
+
+TEST(IpAllowlist, SingleEntryNoCommas) {
+  auto a = parse_ip_allowlist("147.182.255.203");
+  EXPECT_EQ(a.size(), 1u);
+  EXPECT_TRUE(a.count("147.182.255.203"));
+}
+
+TEST(IpAllowlist, EmptyAndBlankYieldEmptySet) {
+  EXPECT_TRUE(parse_ip_allowlist("").empty());
+  EXPECT_TRUE(parse_ip_allowlist("   ").empty());
+  EXPECT_TRUE(parse_ip_allowlist(",, ,\t,").empty()); // only separators/blanks
+}
+
+TEST(IpAllowlist, IgnoresEmptyEntriesBetweenCommas) {
+  auto a = parse_ip_allowlist("8.8.8.8,,9.9.9.9,");
+  EXPECT_EQ(a.size(), 2u);
+  EXPECT_TRUE(a.count("8.8.8.8"));
+  EXPECT_TRUE(a.count("9.9.9.9"));
+}
+
+int main(int argc, char **argv) {
+  ::testing::InitGoogleTest(&argc, argv);
+  return RUN_ALL_TESTS();
+}

test_handler_mock.cpp

@@ -64,6 +64,21 @@ TEST_F(ProcessMockTest, ProcessWithEmptyFile) {
   EXPECT_EQ(result, "emptyfileuser");
 }
 
+// Username lookup is case-insensitive: a mixed-case request is lowercased
+// before the plan-file path is built, so "Pete" reads .../pete.
+TEST_F(ProcessMockTest, ProcessLowercasesUsernameForLookup) {
+  using ::testing::Return;
+  const std::filesystem::path base{"/var/finger/users/"};
+
+  EXPECT_CALL(*mock_filesystem, exists(base / "pete"))
+      .WillOnce(Return(true));
+  EXPECT_CALL(*mock_filesystem, read_file(base / "pete"))
+      .WillOnce(Return("Just another hacker.\r\n"));
+
+  std::string result = process("Pete", *mock_filesystem, base);
+  EXPECT_EQ(result, "Just another hacker.\r\n");
+}
+
 // Test showing multiple expectations
 TEST_F(ProcessMockTest, MultipleFileOperations) {
   using ::testing::_;