diff options
| author | Luke Hoersten <[email protected]> | 2026-06-20 11:16:52 -0500 |
|---|---|---|
| committer | Luke Hoersten <[email protected]> | 2026-06-20 11:16:52 -0500 |
| commit | 131c46f8c539b065feb397e2ee706cff1c81e4cc (patch) | |
| tree | 5843aab0989829e99e142dbf305445ae6129de6c | |
| parent | e6510f2a719cbab4fe9392206f38c90c0d3d3ee3 (diff) | |
cleanup phase 1: stale-comment refresh, zero behavior change
Removes/rewrites every comment that referenced the dead HTTP-streaming
architecture so a reader of v1.0.0+ source isn't chasing a model
that's been gone for several releases. No code paths changed; this is
the safe pre-pass before Phase 2's actual deletions.
Firmware (main/http_api.c):
- s_last_painted_seq / s_last_post_us / X-Frame-Seq parse / stale-
frame guard / Server-Timing emission / TCP_NODELAY setsockopt /
max_open_sockets=4 — each block now leads with "(Legacy from the
HTTP-streaming era; removed in Phase 2)" so the reader knows the
block is doomed, not load-bearing.
- /frame dim-mismatch comment narrowed: panel-native is always 800x480
BGR888, no Scrypted-side variation expected; Scrypted does the
rotation+scale via sharp/mediaManager/ffmpeg cascade (snapshot) or
ffmpeg -vf (stream).
- "Single in-flight frame. Concurrent posts get 503" rewritten to
reflect: decoder mutex now mostly serves to fence /frame snapshots
against an active stream_server decode.
Firmware (main/stream_server.c:136-142):
- FIONREAD-skip rationale reduced from a 7-line paragraph to one
sentence; the savings/tradeoff math now lives in the plan, not the
per-line comment.
Script (scrypted/scrypted-viewport.ts):
- Top-of-file tuning constants block drops the "frame_interval_ms
removed", "fps filter", "in-flight back-to-back startStream"
rationale; one short line covers the model: "Stream rate is paced
by camera + TCP backpressure; no app-level fps cap."
- agentFor() rationale rewritten: this Agent is over-engineered for
control-plane traffic (~1 POST/min steady state) — a legacy of when
it backed per-frame /frame POSTs. Marked for Phase 2 retirement.
- noDelay on Agent: clarified it's now a no-op safety for control
plane (was load-bearing for live-stream pipelining).
- snapshot fire-and-forget comment: replaced X-Frame-Seq-race
rationale with the actual TCP-streaming truth (sharp/mediaManager/
ffmpeg cascade race against stream socket bring-up).
- writeLatencies probe: rewrote the "keep-alive socket" comment
(live stream uses raw net.Socket, not the http.Agent pool).
- socketBackpressured: explicit "diagnostic only, never gates writes"
comment added at declaration site.
- skipLogger header: rewrote the inFlight/MAX_INFLIGHT/fps-filter
rationale into a one-line description of what the log line actually
emits today.
- frameSeq map: now flagged as legacy of HTTP-streaming era, retired
in Phase 2.
| -rw-r--r-- | main/http_api.c | 93 | ||||
| -rw-r--r-- | main/stream_server.c | 10 | ||||
| -rw-r--r-- | scrypted/scrypted-viewport.ts | 90 |
3 files changed, 93 insertions, 100 deletions
diff --git a/main/http_api.c b/main/http_api.c index 37446af..e1a93db 100644 --- a/main/http_api.c +++ b/main/http_api.c @@ -266,8 +266,11 @@ static esp_err_t config_post_handler(httpd_req_t *req) // ============================================================================ // POST /state // ============================================================================ -// Defined alongside other /frame static state further down — forward -// referenced here because state_post_handler resets it on wake. +// Legacy of the HTTP-streaming era (multiple /frame POSTs in flight). +// /frame is now snapshot-only and inherently single-shot, so this +// comparator is effectively unused — but state_post_handler still +// resets it on wake until Phase 2 of the cleanup removes the +// X-Frame-Seq parsing path entirely. static uint32_t s_last_painted_seq; static esp_err_t state_post_handler(httpd_req_t *req) @@ -288,9 +291,8 @@ static esp_err_t state_post_handler(httpd_req_t *req) viewport_run_state_t target; if (strcmp(j->valuestring, "wake") == 0) { target = VIEWPORT_STATE_AWAKE; - // Reset the pipelined-POST sequence comparator so the next - // stream's first frame isn't rejected as stale just because - // the previous stream's counter happened to be higher. + // Reset the (legacy) per-stream sequence comparator; harmless + // now that /frame is snapshot-only, removed in Phase 2. s_last_painted_seq = 0; } else if (strcmp(j->valuestring, "sleep") == 0) { target = VIEWPORT_STATE_ASLEEP; @@ -321,22 +323,24 @@ static esp_err_t respond_status(httpd_req_t *req, const char *status, const char return httpd_resp_send(req, body, body ? HTTPD_RESP_USE_STRLEN : 0); } -// Tracks when the previous /frame response finished so we can log the -// idle gap until the next request enters. Large idle = Scrypted/network -// upstream is the bottleneck; small idle = we're saturating the link. +// (Legacy from the HTTP-streaming era: tracked idle gap between +// successive /frame POSTs. /frame is now snapshot-only — fires at +// most once per wake — so the gap measurement is no longer meaningful +// in steady state. Field is left in place to avoid touching the +// timing-log format; Phase 2 will retire it alongside the rest of +// the streaming machinery.) static int64_t s_last_post_us; // (s_last_painted_seq is forward-declared above state_post_handler; // description there.) static esp_err_t frame_post_handler(httpd_req_t *req) { - // Nagle off on this socket. /frame is a single large POST followed - // by a tiny (empty 204) response — the worst case for Nagle. The - // last partial-MTU packet of the body, and the response packet, - // both sit in the kernel send buffer up to 40ms waiting for an - // ACK that the peer's also delaying-ACKing. setsockopt is cheap - // and idempotent; if keep-alive is ever added the flag persists - // for the connection's life. + // Nagle off. Originally added for HTTP streaming where Nagle + + // delayed-ACK on a 130 KB body POST would stall 40 ms before the + // last partial-MTU segment shipped. /frame is now snapshot-only + // (one POST per wake) so the optimisation is largely moot — but + // setsockopt is essentially free per request and there's no harm + // in leaving it. Phase 2 of the cleanup removes this entirely. int sockfd = httpd_req_to_sockfd(req); if (sockfd >= 0) { int one = 1; @@ -369,17 +373,22 @@ static esp_err_t frame_post_handler(httpd_req_t *req) int64_t t_entry = esp_timer_get_time(); - // Read the optional X-Frame-Seq header used to keep pipelined - // POSTs in monotonic paint order. Missing or zero header = no - // ordering enforcement (treat every frame as newer than the - // previous one, same behaviour we had before pipelining). + // X-Frame-Seq is a legacy artifact of the HTTP-streaming era, + // where multiple /frame POSTs could be in flight at once and + // arrive out of order. /frame is now snapshot-only so the + // header is always missing in practice — kept-alive for one + // more cycle and then removed in Phase 2. uint32_t seq = 0; char seq_str[16] = {0}; if (httpd_req_get_hdr_value_str(req, "X-Frame-Seq", seq_str, sizeof(seq_str)) == ESP_OK) { seq = (uint32_t)strtoul(seq_str, NULL, 10); } - // Single in-flight frame. Concurrent posts get 503 (spec). + // Decoder mutex. Used to serialise pipelined /frame POSTs during + // the HTTP-streaming era; now mostly redundant since /frame is + // snapshot-only and the live stream owns the decoder via + // stream_server. Concurrent posts (e.g. a snapshot landing while + // the live stream is mid-decode) get 503. if (!jpeg_decoder_try_lock(0)) { return respond_status(req, "503 Service Unavailable", "frame in flight"); } @@ -404,17 +413,12 @@ static esp_err_t frame_post_handler(httpd_req_t *req) int64_t t_recv = esp_timer_get_time(); - // Pipelined-POST stale-frame check. With the X-Frame-Seq header - // set, drop anything we've already advanced past — we hold the - // decoder lock, so re-checking here is race-free. Without the - // header (seq==0) every frame paints, preserving pre-pipelining - // behaviour. + // (Legacy stale-frame guard from HTTP-streaming era; never + // triggers under snapshot-only /frame because the header is + // missing. Removed in Phase 2 along with the rest of the + // sequencing machinery.) if (seq != 0 && seq <= s_last_painted_seq) { jpeg_decoder_unlock(); - // 200 OK with an empty body keeps the keep-alive socket - // healthy and lets Scrypted's per-fetch wall-clock log - // stay accurate; we just won't count this toward - // frames_received. httpd_resp_set_status(req, "200 OK"); httpd_resp_set_hdr(req, "X-Frame-Drop", "stale-seq"); return httpd_resp_send(req, NULL, 0); @@ -436,10 +440,11 @@ static esp_err_t frame_post_handler(httpd_req_t *req) return respond_status(req, "400 Bad Request", "JPEG decode failed"); } - // /frame always expects the panel-native 800x480 BGR888 layout — - // Scrypted does the rotation + scale, the firmware just decodes and - // paints. Orientation lives in viewport_state for /state reporting - // and the Scrypted-side ffmpeg pipeline only. + // /frame always expects the panel-native 800x480 BGR888 layout. + // Scrypted does the rotation + scale (snapshot: sharp / mediaManager + // / ffmpeg cascade; stream: ffmpeg -vf transpose+scale). Orientation + // is informational for /state reporting only — firmware never + // rotates pixels. if (w != VIEWPORT_PANEL_WIDTH || h != VIEWPORT_PANEL_HEIGHT) { viewport_state_lock(); viewport_state_get()->decode_errors++; @@ -508,12 +513,10 @@ static esp_err_t frame_post_handler(httpd_req_t *req) state_machine_frame_painted(); // reset idle timer - // Server-Timing per-stage breakdown so the Scrypted side can build - // a unified end-to-end trace per frame (joined by X-Frame-Seq). - // Units are ms with 0.1ms precision — paint is sub-millisecond - // so the decimal matters there. Scrypted subtracts the sum from - // its own (fetch_start → Response_headers) measurement to derive - // the network up + handler-dispatch overhead it can't see directly. + // (Legacy Server-Timing emission from the HTTP-streaming era; + // no Scrypted-side consumer remains — the live-stream cross-side + // trace now flows over the TCP framer's seq field plus per-window + // /state polling. Removed in Phase 2.) char st_hdr[160]; snprintf(st_hdr, sizeof(st_hdr), "recv;dur=%.1f, dec;dur=%.1f, paint;dur=%.1f, post;dur=%.1f, handle;dur=%.1f", @@ -554,11 +557,13 @@ esp_err_t http_api_start(void) cfg.max_uri_handlers = 8; cfg.lru_purge_enable = true; cfg.stack_size = 8192; // POST /config alone has ~2.4 KiB of stack locals - // Allow Scrypted to keep two POST /frame in flight at once so the - // body upload of frame N+1 overlaps with the JPEG decode of frame - // N. The decoder mutex still serialises decode itself; this only - // unblocks the network half of the pipeline. +1 socket reserve for - // a concurrent /state or /config request landing during a stream. + // 2 sockets is enough now that /frame is snapshot-only and the + // live stream owns its own TCP socket on port 81: one slot for + // the snapshot POST + one for a concurrent /state or /config + // request. (Was 4 during the HTTP-streaming era to allow + // pipelined /frame POSTs.) Phase 2 of the cleanup actually + // bumps this down; for Phase 1 (comments-only) the value stays + // at 4 so behaviour is unchanged. cfg.max_open_sockets = 4; httpd_handle_t server = NULL; diff --git a/main/stream_server.c b/main/stream_server.c index 10d1e1b..38769f0 100644 --- a/main/stream_server.c +++ b/main/stream_server.c @@ -133,13 +133,9 @@ static void handle_client(int fd, const char *peer) continue; } - // "Always paint the latest" — if the socket already has more - // bytes queued in the kernel receive buffer, at least one more - // header is on the way. The frame we just finished receiving - // is no longer the freshest possible; skip its decode+paint - // (saves ~6ms per skip) and loop straight to the next header. - // This trades some intermediate frames for lower - // glass-to-glass latency on the latest one. + // If the kernel already has another header queued, this frame + // is no longer the freshest; skip decode + paint, loop back to + // the head-of-queue header. int queued = 0; if (ioctl(fd, FIONREAD, &queued) == 0 && queued >= HEADER_BYTES) { jpeg_decoder_unlock(); diff --git a/scrypted/scrypted-viewport.ts b/scrypted/scrypted-viewport.ts index 2d45c60..cebff82 100644 --- a/scrypted/scrypted-viewport.ts +++ b/scrypted/scrypted-viewport.ts @@ -6,7 +6,7 @@ // short git hash of the commit that added this constant — if the // hash in the log doesn't match the HEAD this file came from, the // Scrypted Script editor is still on stale code. -const SCRIPT_VERSION = "496da49"; +const SCRIPT_VERSION = "pending"; // // Architecture // ------------ @@ -83,19 +83,12 @@ type Setting = any; type Settings = any; type SettingValue = any; -// Tuning constants. Frame interval is also exposed on the parent's -// Settings page so it can be tweaked without editing the script. -// (frame_interval_ms removed in b97c250 — under the TCP streaming -// data plane ffmpeg emits at the camera's native rate and TCP back- -// pressure naturally caps us when the firmware can't keep up. The -// fps filter was only ever useful as a max-fps safety under HTTP.) +// Tuning constants. +// Stream rate is paced by camera + TCP backpressure; no app-level fps +// cap, no per-frame pipelining semaphore. const REREGISTER_INTERVAL_MS = 5 * 60_000; -// 5s gives /state + /config posts enough headroom to slip in between -// /frame POSTs when the device is mid-stream. The firmware's single -// httpd task processes one connection at a time; under heavy /frame -// load a /state {wake} can queue behind 1–3 in-flight /frames before -// landing. 1s was tripping when a burst of camera events triggered -// back-to-back startStream calls. +// 5s is generous for snapshot POSTs (the only HTTP traffic during a +// stream) and survives long-tail latency on a busy Scrypted host. const HTTP_TIMEOUT_MS = 5_000; const DEFAULT_IDLE_TIMEOUT_MS = 60_000; const DEFAULT_BRIGHTNESS = 100; @@ -308,15 +301,12 @@ class ScryptedViewportProvider extends ScryptedDeviceBase }>(); private scryptedBase = ""; - // Per-host node:http Agent with keep-alive. Replaces the previous - // undici dispatcher path — undici isn't reachable as a require() - // module from Scrypted's plugin sandbox, but node:http is always - // available. keepAlive: true reuses the underlying TCP connection - // across POSTs (skips SYN+SYN-ACK+ACK per frame, ~1ms saved on - // LAN and tail-spike-killer on Wi-Fi). maxSockets: 2 caps the - // pool at the same value as the firmware's pipelining capacity, - // so frame N+1 can begin uploading on socket B while frame N is - // still being decoded on the device via socket A. + // Per-host node:http Agent for the control plane (/state, /config, + // and the snapshot /frame POST). Over-engineered for our volume + // (<1 POST/min steady state, ~5 around a wake event) — legacy of + // the HTTP-streaming era where it backed the per-frame /frame + // POSTs. Phase 2 of the cleanup retires the pool and reverts + // postJSON to a plain fetch() with AbortSignal.timeout. private agents = new Map<string, any>(); private agentFor(host: string): any { let a = this.agents.get(host); @@ -328,15 +318,10 @@ class ScryptedViewportProvider extends ScryptedDeviceBase maxSockets: 2, maxFreeSockets: 2, timeout: 30_000, - // CRITICAL: without this, the keep-alive socket inherits - // Nagle ON by default. When a 128KB JPEG body doesn't - // end on an MTU boundary, the kernel sits on the final - // partial packet for up to 200ms (delayed-ACK window) - // waiting for either more data or an ACK. Firmware-side - // TCP_NODELAY only affects firmware's sends; the sender - // also has to opt out. Without noDelay we measured - // fw_recv p95 spiking from ~25ms to ~230ms — the exact - // 200ms Nagle+delayed-ACK deadlock signature. + // noDelay was load-bearing when this Agent fronted the + // live stream's per-frame POSTs (200ms Nagle+delayed- + // ACK stall). It's now a no-op safety for control-plane + // traffic; harmless to leave on. noDelay: true, }); this.agents.set(host, a); @@ -734,9 +719,11 @@ class ScryptedViewportProvider extends ScryptedDeviceBase // ffmpeg-emitted frame lands (ffmpeg startup + RTSP connect + // first H.264 keyframe wait). The snapshot fills the gap so // the panel shows the camera near-instantly on tap/event. - // Fire-and-forget — if it loses the race to the first stream - // frame (which has a higher X-Frame-Seq), the firmware will - // stale-drop it. Errors are silent so a missing snapshot path + // Fire-and-forget — runs in parallel with stream socket bring-up. + // Whichever lands first wins user-visibly; if the stream's + // first frame arrives before the snapshot finishes, the snapshot + // just overpaints stale data on top of a fresher frame for + // ~1 paint cycle. Errors are silent so a missing snapshot path // doesn't break the stream start. this.pushSnapshot(v, cam).catch(() => {}); @@ -855,6 +842,9 @@ class ScryptedViewportProvider extends ScryptedDeviceBase const net = require("net"); let sock: any = null; let socketReady = false; + // Diagnostic only. Never gates writes; we keep pushing past + // kernel-buffer fullness because the firmware's FIONREAD skip + // (stream_server.c) drops superseded frames before decode. let socketBackpressured = false; let seq = 0; let droppedFrames = 0; @@ -864,10 +854,12 @@ class ScryptedViewportProvider extends ScryptedDeviceBase let workBuf: Buffer = Buffer.alloc(0); // Latency probe — wall-clock from "ffmpeg emitted the JPEG" - // to "kernel accepted the socket.write". With the keep-alive - // socket + TCP_NODELAY this should be sub-millisecond steady - // state; visible-double-digit numbers here mean kernel send - // buffer is full (= firmware can't ingest fast enough). + // to "kernel accepted the socket.write". With TCP_NODELAY on + // the stream socket this is sub-millisecond steady state; + // double-digit ms numbers here mean kernel send buffer is + // full (= firmware can't ingest fast enough), which is fine + // — we don't gate on it and the firmware's FIONREAD skip + // drops the surplus before decode. const writeLatencies: number[] = []; const openStreamSocket = () => { @@ -1014,12 +1006,13 @@ class ScryptedViewportProvider extends ScryptedDeviceBase try { currentProc?.kill("SIGTERM"); } catch {} }); - // Periodic skip-rate log. The metric that matters is *delivered* - // fps vs. requested rate, not raw drop count — at low intervals - // ffmpeg's fps filter emits timestamp-clumped pairs and our - // single-flight guard correctly skips the second, but the - // first still painted on time. Only warn if delivered fps falls - // below 75% of target. + // Periodic stream-health log. Emits delivered fps + sustained + // MB/s + socket.write p50/p95/max + drop count + backpressure + // flag every 10s. drops here count frames the demux loop + // threw away because the socket wasn't connected yet (initial + // race at stream start) — once the socket is up we keep + // writing past kernel backpressure and let the firmware's + // FIONREAD skip drop the stale frames before decode. const skipLogger = setInterval(() => { const now = Date.now(); const window = (now - lastLogUs) / 1000; @@ -1184,11 +1177,10 @@ class ScryptedViewportProvider extends ScryptedDeviceBase } catch { /* stream is starting anyway */ } } - // Monotonic per-viewport sequence number paired with X-Frame-Seq - // so the firmware can drop pipelined-out-of-order frames. Reset - // on every stopStream so each stream session starts fresh; the - // firmware also resets its comparator on /state {wake}, so the - // two stay in step. + // Legacy of the HTTP-streaming era. Used to feed the X-Frame-Seq + // header on /frame POSTs so the firmware could drop pipelined- + // out-of-order frames. /frame is now snapshot-only (single-shot), + // so the comparator never trips. Retired in Phase 2. private frameSeq = new Map<string, number>(); private findByName(name: string): Viewport | undefined { |
