diff options
| author | Stefan Boberg <[email protected]> | 2026-05-05 15:47:48 +0200 |
|---|---|---|
| committer | GitHub Enterprise <[email protected]> | 2026-05-05 15:47:48 +0200 |
| commit | 01286c6233347d561064fc9e6cf9deaf2087ceb7 (patch) | |
| tree | bdbfdf01725baa2d2dd3d73727e6506b41421dff /src/zenserver/sessions/sessions.h | |
| parent | hub async s3 client (#1024) (diff) | |
| download | archived-zen-main.tar.xz archived-zen-main.zip | |
Branch started as a sessions-service overhaul (persistence, client liveness, UE_LOGFMT intake) and grew to pick up adjacent infrastructure work: an early-startup log backlog, a hardened `MemoryArena`, the `zen trace serve` viewer gaining a counter view + compact timeline + tabbed callsite panel, defensive fixes in the third-party `tourist` trace parser, a series of allocation reductions across the HTTP and compact-binary hot paths, and a new `zen sessions` CLI command tree.
## Sessions service
**Persistence.** Each session lives on disk under `<DataRoot>/sessions/<id>/` as `info.cb` (metadata) plus `log.bin` (length-prefixed CbObject log records). On startup the service scans that directory and loads prior sessions as ended sessions, preloading the tail of each log so historical views work after a restart. `SessionLog` is noexcept-constructed and falls back to a disabled state on disk errors, so a bad disk can't take down `RegisterSession`. `GetSession` falls back to the ended-sessions list (fixes historical log fetches over HTTP). `LoadTail` counts only successfully-parsed records.
**Pruning.** Periodic cleanup task drops ended sessions once any of three caps is exceeded: age (default 1 year), count (default 1000), or total on-disk footprint (default 50 MiB). Runs 30 s after startup, hourly thereafter. Active sessions never pruned; disk removal and directory stat happen outside the exclusive lock so a slow filesystem can't stall lookups.
**Client liveness.** Sessions carry a `ProcessHandle` for the client-reported pid, captured at registration time so Windows pid recycling can't produce false positives. A 30 s asio timer probes liveness and ends dead sessions through the normal remove path, producing a synthetic `Session ended: process exited (...)` line persisted to `log.bin`. Windows decodes common NTSTATUS exit codes to human names (Ctrl-C, access violation, stack overflow, ...); POSIX stays at plain `process exited`. Clients auto-fill `ClientPid` only for local targets (unix socket / loopback); the server defensively accepts pids only from `IsLocalMachineRequest()` peers. zenserver also reports its own pid when registering its self-session, so it shows up with a real pid in the dashboard and `zen sessions ls`.
**Synthetic end-of-session line.** `RemoveSession` takes an optional reason; before the session moves to the ended list it appends an Info-level `Session ended[: reason]` entry through the normal log path (released outside `m_Lock`). Current reasons: `client request` (HTTP DELETE), `server shutdown` (self-session), `process exited (...)` (liveness).
**UE_LOGFMT structured entries.** `POST /sessions/{id}/log` now accepts `{level, logger, format, fields}` alongside the existing `{level, logger, message}` shape. New `logtemplate.{h,cpp}` implements UE's `StructuredLog.cpp` template grammar (field paths with `.name` / `[N]`, `{{`/`}}` escapes, `$text` / `$format` / `$locformat` object conventions, bounded recursion). Renders to a displayable message at intake while persisting raw format + fields so a future UI can drill into fields without another schema bump. Hot path is zero-alloc — renders into `ExtendableStringBuilder<256>` using stack-buffered `Oid::ToString` / `IoHash::ToHexString` overloads. UI shows a `{…}` marker with the raw template + JSON-pretty fields on hover.
**Parent sessions.** `SessionInfo` gains `parent_session_id`; hub-managed storage server child processes inherit the hub's session id via `--parent-session=<id>`. `ZEN_SESSIONS_URL` env var becomes a fallback for `--sessions-url` / config when neither is provided. The in-process session log sink is disabled when a remote sessions target is configured (logs flow through `SessionsServiceClient` instead). The sessions UI groups child sessions under their parent (collapsible/expandable, sorts as a unit, supports nesting).
**Platform reporting.** `SessionInfo` gains `Platform`, flowed end-to-end: client auto-fills via `GetRuntimePlatformName()`, server persists in `info.cb` (`plat`) and emits on GET. UI renders as a SimpleIcons-style inline SVG (windows / macOS / iOS / linux / wine / android / playstation / xbox / nintendo) with case-insensitive alias resolution (Win32/Win64, PS4/PS5, XSX/XSS, NintendoSwitch, iPhone/iPad, Darwin/OSX). Unknown values fall back to text; sorting runs on the underlying string.
**WebSocket log streaming.** Sessions UI moves from 2 s polling to a WebSocket push model. New `WsSubscriber` has a stable id + helper methods. UI caps the log-line DOM at 5 000 entries with a shared cursor-regression helper, factored out of two call sites. Per-broadcast allocations trimmed on the push path; fixed a stack overrun in the WS log broadcast hex-id buffer.
**Log memory.** `LogEntry::Level` is now `logging::LogLevel` (1 byte) instead of `std::string` (~32 B) — saves ~310 KB per full 10 k-entry deque and eliminates a per-message allocation in the in-proc sink. On-disk format writes an int32 and accepts either int or legacy string on read. `LogEntry` strings now live in a `MemoryArena`; logger names are interned across the deque. `SessionLog::Append` and `WriteSessionInfoFile` drop their `UniqueBuffer` round-trip and write `CbObject::GetView()` straight through `BasicFile` / `SafeWriteFile`. Multi-entry `POST /log` batched under one lock + one push.
**In-proc log timestamps.** `InProcSessionLogSink::TimePointToDateTime` previously preserved only whole seconds, so every in-proc entry rendered at `.000` ms in the dashboard and `zen sessions tail`. It now adds the sub-second part (nanoseconds → 100 ns ticks) to keep ms precision end-to-end.
**UI.** Side "Session Details" panel is gone — its info is inline in the table (appname, mode, platform, id, timestamps, this/log pills, active dot). Bottom panel is a tabbed `Log | Metadata` view with a right-side "Session Information" panel beside metadata; log-only controls (filter, newest-first, follow, log-level filter, expand/collapse) hide when Metadata is active, polling keeps running across tab switches. Wide-mode toggle fills the viewport edge-to-edge. Log lines show the logger category; timestamps render in 24 h with zero-padded fields regardless of locale. Sessions list defaults to All / 10 per page / created-desc, gains click-to-sort headers on the full dataset, a header filter box, and a pager aligned to the table's right edge. Duplicate auto-injected `<h1>Sessions</h1>` removed.
## `zen sessions` CLI
New command tree on the `zen` client for inspecting the sessions service from the terminal:
- **`zen sessions ls`** — lists sessions (active first, ended next; newest-first within each group) with id, status, app/mode, pid, created, duration, and log count. Supports `--status active|ended|all` (default `all`).
- **`zen sessions status`** — prints the sessions service summary: self id, active / ended counts, and the read/write/delete/list/request/bad-request counters from `/stats/sessions`.
- **`zen sessions tail [session]`** — tails a session's log. With no argument it tails zenserver's own session (resolved via `/sessions/list`'s `self_id`); an explicit 24-hex id targets any session, including ended ones (historical replay). `--lines N` (default 50, 0 = all buffered) trims the initial dump client-side. `--follow` prefers a WebSocket push subscription on `/sessions/ws` for sub-second latency; on upgrade failure (older server, blocked port, unix-socket transport) it falls back to HTTP cursor polling at `--interval-ms` (default 500), with sleeps chunked to 50 ms so Ctrl-C reacts quickly. Output matches `zen::logging::FullFormatter` (`[YY-MM-DD HH:MM:SS.mmm] [lvl] [logger] message`); on a TTY the level is colored and the logger is bold, with continuation lines indented under the message column using the *visible* prefix width. 404 surfaces as `(session ended)` and connection errors as `(server gone)` — both clean exits, so stopping the server mid-tail no longer prints a stack trace.
- **`zen sessions ui`** — opens `<host>/dashboard/?page=sessions` in the user's default browser. Rejects unix-socket hosts.
A small `ZenServiceClient::IsUnixSocket()` helper now wraps the unix-socket check used by `ui`, `sessions tail` (WS path), and `sessions ui`.
## Logging
`BacklogSink` captures early-startup log entries in a fixed-capacity ring so late-attached sinks (session sink, file sink) can replay them. Detaches from the broadcast list when disabled; backed by destructor-only cleanup (no `unique_ptr` indirection per entry). Tuned defaults so the backlog covers typical bring-up without unbounded growth.
## `zen trace serve` viewer
- Compact timeline mode for high-density views.
- New `TRACE_INT_VALUE` / `TRACE_FLOAT_VALUE` counter trace points + a counters page in the viewer.
- Callsite tables collapsed into a single tabbed panel.
- Lossless `Oid <-> Guid` bridge for trace session ids; trace `SessionId` plumbed through.
- `tourist` parser hardening: bounds-check `BufferStream::read`, validate `Type::info_size` before `patch()`, convert `parse_important_aux` to a loop (avoids deep recursion), widen `ParserPool` index to `uint32`, bounds-check field offsets in the dispatcher, pin `Types::parse` buffer up-front.
## `MemoryArena`
Configurable chunk size, inline chunk list, oversize requests routed to truly-dedicated chunks (no slack waste, no fragmentation when one allocation is much larger than the chunk).
## Allocation cleanups across hot paths
- `zenhttp::HttpRequestRouter::HandleRequest` and `FormatPackageMessageInternal`: drop heap allocations.
- Compact-binary validation: `eastl::fixed_vector` + `eastl::sort`; eliminate `std::vector` churn.
- `zenserverprocess`: trim transient allocations in spawn paths.
- Sessions HTTP intake / broadcast: drop transient `std::string` allocs.
Diffstat (limited to 'src/zenserver/sessions/sessions.h')
| -rw-r--r-- | src/zenserver/sessions/sessions.h | 193 |
1 files changed, 174 insertions, 19 deletions
diff --git a/src/zenserver/sessions/sessions.h b/src/zenserver/sessions/sessions.h index a84ca6506..a722704e0 100644 --- a/src/zenserver/sessions/sessions.h +++ b/src/zenserver/sessions/sessions.h @@ -4,6 +4,8 @@ #include <zencore/compactbinary.h> #include <zencore/logbase.h> +#include <zencore/memory/memoryarena.h> +#include <zencore/process.h> #include <zencore/thread.h> #include <zencore/uid.h> @@ -11,7 +13,10 @@ ZEN_THIRD_PARTY_INCLUDES_START #include <EASTL/deque.h> #include <tsl/robin_map.h> ZEN_THIRD_PARTY_INCLUDES_END +#include <filesystem> +#include <functional> #include <optional> +#include <span> #include <string> #include <vector> @@ -34,25 +39,62 @@ public: Oid Id; std::string AppName; std::string Mode; - Oid JobId; - CbObject Metadata; - DateTime CreatedAt; - DateTime UpdatedAt; - DateTime EndedAt{0}; + std::string Platform; // Reported by the client, e.g. "windows", "linux", "macos" + uint32_t ClientPid = 0; // Non-zero = local PID to probe for liveness. 0 = don't track. + Oid ParentSessionId; + // Optional task/action identifier used to associate this session with a + // specific unit of work. Distinct from ParentSessionId, which records + // process/session ancestry. + Oid JobId; + CbObject Metadata; + DateTime CreatedAt; + DateTime UpdatedAt; + DateTime EndedAt{0}; }; + /// Stored form of a log entry. The string fields are arena-borrowed + /// `const char*` — they live in the owning Session's MemoryArena and + /// are valid only for that Session's lifetime. Default copy is + /// intentionally shallow (string pointers are shared with the source); + /// callers must not let copies outlive the originating Session. + /// + /// Build entries via `LogEntryInput` and route them through + /// `Session::AppendLog` / `AppendLogBatch`, which intern logger names + /// and arena-allocate the other strings before storing. struct LogEntry { - DateTime Timestamp; - std::string Level; - std::string Message; - CbObject Data; + DateTime Timestamp{0}; + // Sentinel: Off means "no level set" (e.g. plain-text POSTed entries + // where the client didn't include a level). Real log entries use + // Trace..Critical, so Off is free to reuse as "omit on serialize". + logging::LogLevel Level = logging::Off; + // Arena pointers (null-terminated). Empty string is the default + // — never null, so callers don't need to guard. + const char* LoggerName = ""; // Interned: one canonical copy per unique name across the session. + const char* Message = ""; // For structured entries: the rendered form (populated at intake). + const char* Format = ""; // UE_LOGFMT template; "" for plain entries. + CbObject Fields; // Present only when Format is non-empty. + }; + + /// Input form used to build an entry on the way into a Session. The + /// string_view fields are caller-borrowed; AppendLog interns/copies + /// them into the Session's arena before any LogEntry is built. Use + /// this struct rather than constructing LogEntry directly so the + /// arena ownership invariant stays one-sided. + struct LogEntryInput + { + DateTime Timestamp{0}; + logging::LogLevel Level = logging::Off; + std::string_view LoggerName; + std::string_view Message; + std::string_view Format; + CbObject Fields; }; class Session : public TRefCounted<Session> { public: - Session(const SessionInfo& Info); + Session(const SessionInfo& Info, Ref<SessionLog> Log = {}, ProcessHandle ClientProcess = {}); ~Session(); Session(Session&&) = delete; @@ -67,12 +109,29 @@ public: void SetEndedAt(DateTime When) { m_Info.EndedAt = When; } - void AppendLog(LogEntry Entry); + /// Appends an entry to the in-memory deque and to the persisted + /// log. Returns the new cursor value (m_TotalAppended post- + /// increment). Logger name is interned, message and format are + /// arena-allocated — the input's string_views may safely be + /// caller-stack-bound. + uint64_t AppendLog(LogEntryInput Input); + + /// Append-many counterpart that takes the deque lock exactly once + /// for the whole batch. Use this when an inbound HTTP POST carries + /// multiple entries — single-lock semantics keep entries from one + /// caller contiguous on the wire even when other appends race in, + /// and the WS-push observer can fire just once for the whole batch. + /// Returns the new cursor (the value at the tail of the batch). + uint64_t AppendLogBatch(std::span<LogEntryInput> Inputs); + std::vector<LogEntry> GetLogEntries(uint32_t Limit = 0, uint32_t Offset = 0) const; uint64_t GetLogCount() const; /// Returns entries appended after the given cursor and the new cursor value. /// A cursor of 0 returns all entries currently in the deque. + /// The returned LogEntries borrow strings from this Session's + /// arena — callers must hold a Ref<Session> for as long as they + /// keep the result. struct CursorResult { std::vector<LogEntry> Entries; @@ -81,26 +140,118 @@ public: }; CursorResult GetLogEntriesAfter(uint64_t AfterCursor) const; + // Seed this session with pre-existing log entries (e.g. loaded from disk + // on startup). Sets the total-appended counter to reflect what was on + // disk so cursors remain meaningful for historical sessions. The inputs + // are interned/arena-allocated into this session. + void PreloadEntries(std::span<const LogEntryInput> Tail, uint64_t TotalCount); + + /// Process handle used for client-liveness checks. Acquired at + /// registration time (while the pid is known to refer to the reporting + /// process) and held for the session's lifetime; on Windows this is a + /// real HANDLE tied to the specific process instance and is immune to + /// pid reuse. Invalid (IsValid() == false) for remote sessions or when + /// OpenProcess() failed. Set once at construction — no synchronization + /// needed for readers. + const ProcessHandle& GetClientProcess() const { return m_ClientProcess; } + ProcessHandle& GetClientProcess() { return m_ClientProcess; } + + static constexpr uint32_t MaxLogEntries = 10000; + private: + // Intern a logger name into m_LogArena and return the canonical + // pointer for that name. Subsequent calls with the same string + // return the same pointer. Caller must hold m_LogLock exclusive. + const char* InternLoggerNameLocked(std::string_view Name); + + // Allocate a copy of Str into m_LogArena and return a null- + // terminated pointer. No deduplication. Caller must hold m_LogLock + // exclusive. Empty input returns "" (no allocation). + const char* AllocateLogStringLocked(std::string_view Str); + SessionInfo m_Info; Ref<SessionLog> m_Log; + ProcessHandle m_ClientProcess; mutable RwLock m_LogLock; eastl::deque<LogEntry> m_LogEntries; uint64_t m_TotalAppended = 0; // monotonically increasing counter - - static constexpr uint32_t MaxLogEntries = 10000; + // String storage for the in-memory deque. LoggerName is interned + // (one canonical copy per unique name); Message and Format are + // duplicated per entry. Both die with the Session — so the + // LogEntry pointers do too. Sized to fit a typical session's + // strings in one chunk; spills to additional chunks otherwise. + MemoryArena m_LogArena{4096}; + tsl::robin_map<std::string_view, const char*> m_InternedLoggerNames; }; - SessionsService(); + /// Construct a SessionsService. If StorageRoot is non-empty, session + /// metadata and logs are persisted under that directory (one subdirectory + /// per session id) and previously-persisted sessions are loaded as ended. + explicit SessionsService(std::filesystem::path StorageRoot = {}); ~SessionsService(); - bool RegisterSession(const Oid& SessionId, std::string AppName, std::string Mode, const Oid& JobId, CbObjectView Metadata); - bool UpdateSession(const Oid& SessionId, CbObjectView Metadata); - Ref<Session> GetSession(const Oid& SessionId) const; + bool RegisterSession(const Oid& SessionId, + std::string AppName, + std::string Mode, + std::string Platform, + uint32_t ClientPid, + const Oid& ParentSessionId, + const Oid& JobId, + CbObjectView Metadata); + bool UpdateSession(const Oid& SessionId, CbObjectView Metadata); + Ref<Session> GetSession(const Oid& SessionId) const; std::vector<Ref<Session>> GetSessions() const; std::vector<Ref<Session>> GetEndedSessions() const; - bool RemoveSession(const Oid& SessionId); - uint64_t GetSessionCount() const; + /// Ends a session. If Reason is non-empty, a synthetic log line is + /// appended to the session log before it's moved to ended so the + /// historical log has a clear closing event. + bool RemoveSession(const Oid& SessionId, std::string_view Reason = {}); + uint64_t GetSessionCount() const; + + /// Appends a log entry to `SessionId` and, if the session exists, + /// invokes the log-appended callback with the new cursor so downstream + /// push subscribers (e.g. the HTTP WS broadcast) can deliver the delta + /// without polling. Returns the new cursor, or 0 if the session is + /// unknown. Fires the callback AFTER any internal locks are released + /// so the callback can safely call back into this service. + uint64_t AppendLog(const Oid& SessionId, LogEntryInput Input); + + /// Batch counterpart of AppendLog. Atomic with respect to other + /// appends to the same session — entries land contiguously on the + /// wire and persist in order — and fires exactly one push-callback + /// for the whole batch. Empty batches and unknown sessions are + /// no-ops returning 0. + uint64_t AppendLogBatch(const Oid& SessionId, std::span<LogEntryInput> Inputs); + + /// Observer fired after an entry is appended to any session. Replaces + /// any previously set callback. Pass {} to clear. Only one listener is + /// supported — the single consumer today is the HTTP WebSocket push. + using LogAppendedCallback = std::function<void(const Oid& SessionId, uint64_t NewCursor)>; + void SetLogAppendedCallback(LogAppendedCallback Callback); + + /// Drop ended sessions that are too old, that push us over the count + /// limit, or that push the on-disk footprint over the byte budget, and + /// delete their persisted directories. Active sessions are never + /// pruned. Returns the number removed by each criterion. + struct PruneResult + { + size_t ExpiredByAge = 0; + size_t ExpiredByCount = 0; + size_t ExpiredByStorage = 0; + }; + PruneResult PruneExpired(TimeSpan MaxAge, size_t MaxCount, uint64_t MaxStorageBytes); + + /// End any active session whose tracked client process is no longer + /// running. Sessions with an invalid ProcessHandle (remote, or + /// OpenProcess failed at registration) are skipped. Returns the number + /// of sessions ended by this pass. + size_t CheckProcessLiveness(); + + // Tuning defaults. Expressed in whole days / bytes so they're easy to + // override from a future command-line flag without touching internals. + static constexpr int kDefaultMaxSessionAgeDays = 365; + static constexpr size_t kDefaultMaxSessionCount = 1000; + static constexpr uint64_t kDefaultMaxStorageBytes = 50ull * 1024 * 1024; // 50 MiB private: LoggerRef& Log() { return m_Log; } @@ -110,6 +261,10 @@ private: tsl::robin_map<Oid, Ref<Session>, Oid::Hasher> m_Sessions; std::vector<Ref<Session>> m_EndedSessions; std::unique_ptr<SessionLogStore> m_SessionLogs; + // Set once at wiring-time (single consumer), never reassigned while + // hot, so no dedicated lock — just a plain member. Copy-on-call + // guards against the theoretical re-register race below. + LogAppendedCallback m_LogAppendedCallback; }; } // namespace zen |