fix(mqtt): bound browse observation against a chatty or malicious broker

Three review findings, all "unbounded work/memory from untrusted broker input" rather than correctness bugs - but this points at a live plant broker. 1. InferPayload decoded the WHOLE payload on MQTTnet's dispatcher thread before trimming to 64 chars. The node cap never engages here: a multi-MB blob on an already-known topic creates no node, it just updates one, so the cost repeats per message forever. Now at most PayloadInspectMaxBytes (512) are examined. A blind slice can cut a multi-byte sequence, which the strict decoder would report as "binary" - TrimToUtf8Boundary backs the window off to the last whole sequence so good UTF-8 is never mislabelled. A clipped payload is String by construction rather than inferred from a partial view. 2. The node cap bounded COUNT, not bytes. MQTT topics run to ~65KB, so 50k nodes near that ceiling is a multi-GB tree. Topics now bounded at 1024 chars and segments at 256, rejected whole (a truncated topic is a DIFFERENT topic the picker would commit as a tag address) and surfaced via an __oversized__ marker kept distinct from __truncated__ - the operator's remedy differs. 3. Control characters in a snippet would break the picker's single-line row; Trim() only strips the ends. Now scrubbed to spaces. Each guard was falsified independently by neutering it and confirming exactly one test reddens. That found a real gap: the oversized-topic test was passing via the SEGMENT bound, leaving the whole-topic bound untested - the test now uses many short segments so only the topic bound can reject it. Also records the .Driver project reference as a known, deliberate exception to the .Browser -> .Contracts pattern, with its cost (AdminUI gains Core + Polly + Serilog) and the clean fix (a leaf transport project; NOT a move into .Contracts, which is deliberately transport-free). Claude-Session: https://claude.ai/code/session_01GASWkNEi68FSCtvr6rLoEW
2026-07-24 16:30:45 -04:00
parent d421487bcd
commit 4ec01bdfc1
4 changed files with 302 additions and 22 deletions
@@ -41,9 +41,41 @@ internal sealed class MqttBrowseSession : IBrowseSession
    /// <summary>Maximum characters of a payload retained as the picker's confirm-it's-the-right-topic snippet.</summary>
    internal const int PayloadSnippetMaxChars = 64;

+    /// <summary>
+    ///     Hard cap on the payload bytes fed to the UTF-8 decode. A <c>#</c> subscription receives
+    ///     whatever the broker has — multi-MB retained blobs and JSON dumps included — and the node
+    ///     cap does not help: a large payload on an <i>already-known</i> topic creates no node, it
+    ///     just updates one, so an uncapped decode is unbounded work on MQTTnet's pump thread on
+    ///     every single message, forever. 512 is <see cref="PayloadSnippetMaxChars"/> × 4 (UTF-8's
+    ///     worst-case bytes per char, so the snippet can always be filled) doubled, leaving headroom
+    ///     for leading whitespace ahead of a scalar.
+    /// </summary>
+    internal const int PayloadInspectMaxBytes = 512;
+
+    /// <summary>
+    ///     Cap on a whole topic name. MQTT permits ~65 KB, and the node cap bounds the node
+    ///     <i>count</i> only — 50 000 nodes near that ceiling is a multi-GB tree. Topics past this
+    ///     are rejected outright rather than truncated: a truncated topic is a <i>different</i>
+    ///     topic, and the picker would commit it as a tag address.
+    /// </summary>
+    internal const int MaxTopicChars = 1024;
+
+    /// <summary>Cap on one topic segment (one tree node's label + display name).</summary>
+    internal const int MaxSegmentChars = 256;
+
    /// <summary>Sentinel node id of the "results truncated" marker, mirroring <c>CapturedTreeBrowseSession</c>.</summary>
    internal const string TruncatedNodeId = "__truncated__";

+    /// <summary>Sentinel node id of the "oversized topics rejected" marker.</summary>
+    internal const string OversizedNodeId = "__oversized__";
+
+    /// <summary>
+    ///     Throw-on-invalid so genuinely binary payloads are reported as such rather than silently
+    ///     rendered as a field of replacement characters. Cached: allocating an encoding per message
+    ///     would be per-message garbage on the dispatcher thread.
+    /// </summary>
+    private static readonly UTF8Encoding Utf8Strict = new(encoderShouldEmitUTF8Identifier: false, throwOnInvalidBytes: true);
+
    /// <summary>Bound on the courtesy DISCONNECT issued at dispose — teardown never hangs on a dead broker.</summary>
    private static readonly TimeSpan DisconnectTimeout = TimeSpan.FromSeconds(5);

@@ -60,6 +92,7 @@ internal sealed class MqttBrowseSession : IBrowseSession
    private int _disposed;
    private long _lastUsedTicksUtc = DateTime.UtcNow.Ticks;
    private int _nodeCount;
+    private int _oversized;
    private int _publishCount;
    private int _truncated;

@@ -95,6 +128,14 @@ internal sealed class MqttBrowseSession : IBrowseSession
    /// <summary>True once the node cap stopped recording; surfaced as a marker node from <see cref="RootAsync"/>.</summary>
    internal bool Truncated => Volatile.Read(ref _truncated) != 0;

+    /// <summary>
+    ///     True once a topic was rejected for breaching <see cref="MaxTopicChars"/> /
+    ///     <see cref="MaxSegmentChars"/>; surfaced as its own marker node from
+    ///     <see cref="RootAsync"/>. Kept distinct from <see cref="Truncated"/> because the operator's
+    ///     remedy differs — narrow the prefix, versus this broker publishes absurd topic names.
+    /// </summary>
+    internal bool OversizedRejected => Volatile.Read(ref _oversized) != 0;
+
    private bool Disposed => Volatile.Read(ref _disposed) != 0;

    /// <summary>
@@ -123,6 +164,15 @@ internal sealed class MqttBrowseSession : IBrowseSession
                HasChildrenHint: false));
        }

+        if (OversizedRejected)
+        {
+            nodes.Add(new BrowseNode(
+                OversizedNodeId,
+                "⚠ Some topics were too long to show — use manual entry for those",
+                BrowseNodeKind.Folder,
+                HasChildrenHint: false));
+        }
+
        Touch();
        return Task.FromResult<IReadOnlyList<BrowseNode>>(nodes);
    }
@@ -209,7 +259,24 @@ internal sealed class MqttBrowseSession : IBrowseSession
    {
        if (Disposed || string.IsNullOrWhiteSpace(topic)) return;

+        if (topic.Length > MaxTopicChars)
+        {
+            Volatile.Write(ref _oversized, 1);
+            return;
+        }
+
        var segments = topic.Split('/');
+
+        // Checked as a pre-pass so a rejected topic leaves no half-built path behind.
+        foreach (var segment in segments)
+        {
+            if (segment.Length > MaxSegmentChars)
+            {
+                Volatile.Write(ref _oversized, 1);
+                return;
+            }
+        }
+
        var level = _roots;
        TopicNode? node = null;
        var path = string.Empty;
@@ -270,6 +337,10 @@ internal sealed class MqttBrowseSession : IBrowseSession
    /// <summary>
    ///     Best-effort type inference from a payload body. Explicit authoring stays the authority
    ///     (design §4) — this only decorates the picker. Non-UTF-8 bodies are reported as opaque.
+    ///     <para>
+    ///         Runs on MQTTnet's dispatcher thread, so the work is bounded before any decode: at most
+    ///         <see cref="PayloadInspectMaxBytes"/> bytes are ever examined, however large the message.
+    ///     </para>
    /// </summary>
    /// <param name="payload">The raw message body.</param>
    /// <returns>The inferred type and a bounded display snippet.</returns>
@@ -277,37 +348,102 @@ internal sealed class MqttBrowseSession : IBrowseSession
    {
        if (payload.IsEmpty) return new ObservedValue(DriverDataType.String, "(empty)");

+        var totalBytes = payload.Length;
+        var clipped = totalBytes > PayloadInspectMaxBytes;
+        // A blind slice can cut a multi-byte sequence in half, which the strict decoder would report
+        // as "binary" — mislabelling a perfectly good UTF-8 payload. Back the window off to the last
+        // whole sequence instead.
+        var window = clipped ? TrimToUtf8Boundary(payload[..PayloadInspectMaxBytes]) : payload;
+
        string text;
        try
        {
-            // Throw-on-invalid so genuinely binary payloads are reported as such rather than
-            // silently rendered as a field of replacement characters.
-            text = new UTF8Encoding(encoderShouldEmitUTF8Identifier: false, throwOnInvalidBytes: true)
-                .GetString(payload);
+            text = Utf8Strict.GetString(window);
        }
        catch (DecoderFallbackException)
        {
-            return new ObservedValue(DriverDataType.String, $"({payload.Length} bytes, binary)");
+            return new ObservedValue(DriverDataType.String, $"({totalBytes} bytes, binary)");
        }

        var trimmed = text.Trim();
-        var snippet = trimmed.Length > PayloadSnippetMaxChars
-            ? string.Concat(trimmed.AsSpan(0, PayloadSnippetMaxChars), "…")
-            : trimmed;

        // Ordered narrowest-first. "1" infers Int64 rather than Boolean: an integer reading is the
        // commoner meaning and, unlike the reverse, does not silently collapse a range to two states.
-        var dataType = trimmed switch
+        // A clipped payload is never a scalar, and inferring from a partial view could read a
+        // whitespace-padded prefix as a number it is not — so it is String by construction.
+        var dataType = clipped
+            ? DriverDataType.String
+            : trimmed switch
+            {
+                _ when bool.TryParse(trimmed, out _) => DriverDataType.Boolean,
+                _ when long.TryParse(trimmed, NumberStyles.Integer, CultureInfo.InvariantCulture, out _)
+                    => DriverDataType.Int64,
+                _ when double.TryParse(trimmed, NumberStyles.Float, CultureInfo.InvariantCulture, out _)
+                    => DriverDataType.Float64,
+                _ => DriverDataType.String,
+            };
+
+        return new ObservedValue(dataType, Sanitize(trimmed, clipped));
+    }
+
+    /// <summary>
+    ///     Renders decoded payload text as a single-line, length-bounded snippet. Control characters
+    ///     (newlines, tabs, ANSI escapes) become spaces — <see cref="string.Trim()"/> only strips them
+    ///     at the ends, and an embedded one would break the picker's single-line attribute row.
+    /// </summary>
+    /// <param name="text">The trimmed decoded text.</param>
+    /// <param name="clipped">Whether bytes beyond the inspection window were discarded.</param>
+    /// <returns>The display snippet.</returns>
+    private static string Sanitize(string text, bool clipped)
+    {
+        var overLength = text.Length > PayloadSnippetMaxChars;
+        var span = overLength ? text.AsSpan(0, PayloadSnippetMaxChars) : text.AsSpan();
+
+        // clipped, not just overLength: a huge payload whose first 512 bytes are mostly whitespace
+        // yields a short snippet that would otherwise claim to be the whole message.
+        return overLength || clipped ? Scrub(span) + "…" : Scrub(span);
+    }
+
+    /// <summary>Replaces control characters with spaces and trims the trailing edge.</summary>
+    /// <param name="span">The characters to scrub.</param>
+    /// <returns>The scrubbed text.</returns>
+    private static string Scrub(ReadOnlySpan<char> span)
+    {
+        Span<char> buffer = span.Length <= 128 ? stackalloc char[span.Length] : new char[span.Length];
+        for (var i = 0; i < span.Length; i++) buffer[i] = char.IsControl(span[i]) ? ' ' : span[i];
+        return new string(buffer).TrimEnd();
+    }
+
+    /// <summary>
+    ///     Shortens a byte window so it ends on a whole UTF-8 sequence. Walks back over continuation
+    ///     bytes (<c>10xxxxxx</c>) to the lead byte and drops it when the sequence it starts runs past
+    ///     the window. A stray continuation byte is left in place — that is genuine corruption, and the
+    ///     strict decoder should say so.
+    /// </summary>
+    /// <param name="bytes">The blindly-sliced window.</param>
+    /// <returns>The window, shortened to a sequence boundary.</returns>
+    private static ReadOnlySpan<byte> TrimToUtf8Boundary(ReadOnlySpan<byte> bytes)
+    {
+        var i = bytes.Length - 1;
+        var walked = 0;
+        while (i >= 0 && (bytes[i] & 0xC0) == 0x80 && walked < 3)
        {
-            _ when bool.TryParse(trimmed, out _) => DriverDataType.Boolean,
-            _ when long.TryParse(trimmed, NumberStyles.Integer, CultureInfo.InvariantCulture, out _)
-                => DriverDataType.Int64,
-            _ when double.TryParse(trimmed, NumberStyles.Float, CultureInfo.InvariantCulture, out _)
-                => DriverDataType.Float64,
-            _ => DriverDataType.String,
+            i--;
+            walked++;
+        }
+
+        if (i < 0) return ReadOnlySpan<byte>.Empty; // nothing but continuation bytes
+
+        var needed = bytes[i] switch
+        {
+            < 0x80 => 1,   // ASCII
+            >= 0xF0 => 4,
+            >= 0xE0 => 3,
+            >= 0xC0 => 2,
+            _ => 1,        // stray continuation byte — let the decoder reject it
        };

-        return new ObservedValue(dataType, snippet);
+        return bytes.Length - i >= needed ? bytes : bytes[..i];
    }

    /// <summary>Projects one level of the tree into browse nodes, ordered so the picker is stable.</summary>
@@ -19,6 +19,14 @@ namespace ZB.MOM.WW.OtOpcUa.Driver.Mqtt.Browser;
 ///         publishes: an operator opening a picker must not be able to inject a message into a
 ///         running plant.
 ///     </para>
+///     <para>
+///         <b>Layering note.</b> Unlike every other bespoke browser here, this project references the
+///         runtime <c>.Driver</c> project (to reuse <c>MqttConnection.BuildClientOptions</c> rather
+///         than fork the TLS / CA-pin path). That is a known, deliberate exception with a documented
+///         cost and a documented clean fix — see the comment on that <c>ProjectReference</c> in
+///         <c>ZB.MOM.WW.OtOpcUa.Driver.Mqtt.Browser.csproj</c> before copying this project as a
+///         template.
+///     </para>
 /// </summary>
 public sealed class MqttDriverBrowser : IDriverBrowser
 {
@@ -14,12 +14,32 @@
        <ProjectReference Include="..\..\Core\ZB.MOM.WW.OtOpcUa.Commons\ZB.MOM.WW.OtOpcUa.Commons.csproj"/>
        <ProjectReference Include="..\ZB.MOM.WW.OtOpcUa.Driver.Mqtt.Contracts\ZB.MOM.WW.OtOpcUa.Driver.Mqtt.Contracts.csproj"/>
        <!--
-            Deviation from the plan's ref list (which named .Contracts only): the browse connect
-            reuses MqttConnection.BuildClientOptions rather than reimplementing it. That method
-            owns the TLS posture, the CA-pin chain validator and the credential handling; a second
-            copy here would be a security divergence waiting to happen (browse silently not
-            honouring a pinned CA while the runtime driver does). It is pure — no I/O, no network —
-            so the dependency costs nothing at browse time.
+            ┌──────────────────────────────────────────────────────────────────────────────────┐
+            │ KNOWN, DELIBERATE EXCEPTION to the .Browser → .Contracts pattern.                 │
+            │ DO NOT COPY THIS LINE INTO A NEW DRIVER'S BROWSER WITHOUT READING THIS.           │
+            └──────────────────────────────────────────────────────────────────────────────────┘
+
+            Every other bespoke browser in this repo (OpcUaClient, Galaxy) references only its own
+            .Contracts project plus its transport package. This one also references the runtime
+            .Driver project. That is a real widening, not a free one, and it was accepted knowingly:
+
+            WHY: the browse CONNECT reuses MqttConnection.BuildClientOptions, which owns the TLS
+              posture, the PEM CA-pin chain validator, the credential handling and the protocol
+              version mapping (~100 lines). A second copy here would be a silent security
+              divergence — browse quietly not honouring a pinned CA while the runtime driver does.
+              The method is pure (no I/O, no network), so the dependency costs nothing at run time.
+
+            WHAT IT COSTS: the AdminUI's project graph does NOT otherwise include Core.csproj
+              (only Host.csproj and the runtime Driver.* projects do). Referencing this browser
+              from the AdminUI therefore pulls in Core + Polly.Core + Serilog at build/deploy time.
+
+            THE CLEAN FIX (deferred, not blocked): a small leaf project — say
+              ZB.MOM.WW.OtOpcUa.Driver.Mqtt.Transport (net10, refs .Contracts + MQTTnet) — holding
+              BuildClientOptions/ConfigureTls/the CA-pin helpers, referenced by BOTH .Driver and
+              .Browser. Note the obvious-looking alternative does NOT work: those members cannot
+              move into .Contracts, because .Contracts is deliberately transport-free (Task 1
+              removed its MQTTnet reference to keep it so) and BuildClientOptions returns
+              MqttClientOptions, an MQTTnet type.
        -->
        <ProjectReference Include="..\ZB.MOM.WW.OtOpcUa.Driver.Mqtt\ZB.MOM.WW.OtOpcUa.Driver.Mqtt.csproj"/>
    </ItemGroup>