@mark1russell7/machines 0.3.23 → 0.3.24

package/README.md

@@ -25,7 +25,7 @@ The algebra reads these documents. It gives the closed form of each timer, the f
 
 | Export | Identifier | System and version | Layer | Sources, by rung |
 | --- | --- | --- | --- | --- |
-| `LINUX_TCP` | `linux-tcp` | TCP of Linux 6.6, for kernels from 6.6 to 6.9 | kernel | 4 normative, 7 implemented |
+| `LINUX_TCP` | `linux-tcp` | TCP of Linux 6.6, for kernels from 6.6 to 6.9 | kernel | 4 normative, 8 implemented |
 | `NF_CONNTRACK_TCP` | `nf-conntrack-tcp` | The TCP state machine of conntrack on Linux 6.6.87.2 | middlebox | 4 implemented |
 | `REDIS_CLIENT` | `redis-client` | One client connection of the Redis 8.10.2 server | server | 3 implemented, 1 documented |
 | `IOREDIS_STATUS` | `ioredis-status` | The connection status of ioredis 6.0.0, standalone | driver | 5 implemented |
@@ -109,6 +109,12 @@ A change to the version of a system changes its sources too. Each fact of the do
 
 In this repository, a snapshot of the predictions of the cells (`packages/cells/src/__snapshots__/predictions.json`) shows each prediction that a change to a document moves.
 
+### 0.3.24
+
+- **The SYN timer of Linux TCP** (`syn_retransmit`). It is a backoff timer: each firing sends the SYN again. The waits 0 to `tcp_syn_linear_timeouts` are 1 s, and then each wait doubles up to `TCP_RTO_MAX`. The source is `tcp_retransmit_timer` in `net/ipv4/tcp_timer.c` of 6.6, and 6.17 has the same waits. At HZ 1000 a wait of 1 s is up to 64 ms late. The models are `linux.syn-wait` and `linux.syn-wait-lateness` of the algebra 0.7.3.
+  - `recovery` over this timer gives the first SYN after a stall. Kind run 40123580623 measured it: a SYN to a pod that was gone stalled until IPVS dropped the pod.
+  - One prediction of the snapshot moves, the kind cell T07.defaults.12.
+
 ### 0.3.18
 
 - **Caddy 2.11.7 as a proxy of gRPC** (`caddy-http2`), from the source at its tag and of golang.org/x/net v0.59.0. With no stream open for 5 min, Caddy sends GOAWAY NO_ERROR, and it closes the connection 1 s later. A PING of the client does not reset that timer. A failure of the upstream before the headers of a response gives 502, thus UNAVAILABLE. Caddy retries no call at the defaults, and at SIGTERM it waits for each stream.

package/build/index.js

@@ -3478,6 +3478,10 @@ var linux_tcp_default = {
 		{
 			"ref": "net/ipv4/tcp_input.c:791-834 tcp_event_data_recv(): the first data starts quick-ACK mode; :326-333 tcp_in_quickack_mode(); :5552-5576 __tcp_ack_snd_check(): an ACK at once or delayed; net/ipv4/tcp_output.c:161-176 tcp_event_data_sent(): ping-pong mode",
 			"rung": "implemented"
+		},
+		{
+			"ref": "net/ipv4/tcp_timer.c:604 and :624-631 tcp_retransmit_timer(): each firing sends the SYN again, and in SYN_SENT the RTO doubles only after tcp_syn_linear_timeouts firings; the same at 6.17 (:448 total_rto, :671-678)",
+			"rung": "implemented"
 		}
 	],
 	states: {
@@ -3541,6 +3545,10 @@ var linux_tcp_default = {
 			"note": "the first retransmission the RTO timer makes, which sets retrans_stamp; the tail loss probe one RTO earlier does not"
 		},
 		"timer.syn.exhausted": { "kind": "timer" },
+		"timer.syn": {
+			"kind": "timer",
+			"note": "the SYN timer fires and the SYN goes out again"
+		},
 		"timer.synack.exhausted": { "kind": "timer" },
 		"timer.rto": {
 			"kind": "timer",
@@ -3661,6 +3669,32 @@ var linux_tcp_default = {
 			"fires": "timer.syn.exhausted",
 			"note": "6.6: linear 1 s waits for the first tcp_syn_linear_timeouts retransmissions, then doubling; 131 s at the defaults. Each wait is a timer of the wheel, rounded up at its level: up to 6.08 s late in all at HZ 1000, because the wait of 64 s can be 4.096 s late"
 		},
+		"syn_retransmit": {
+			"duration": {
+				"model": "linux.syn-wait",
+				"knobs": ["fired", "tcp_syn_linear_timeouts"],
+				"unit": "s"
+			},
+			"lateness": {
+				"model": "linux.syn-wait-lateness",
+				"knobs": [
+					"fired",
+					"tcp_syn_linear_timeouts",
+					"hz"
+				],
+				"unit": "s"
+			},
+			"early": {
+				"model": "linux.jiffy",
+				"knobs": ["hz"],
+				"unit": "s"
+			},
+			"arms": ["enter:SYN_SENT"],
+			"cancels": ["rx.synack", "rx.rst"],
+			"fires": "timer.syn",
+			"backoff": { "resets": ["enter:SYN_SENT"] },
+			"note": "each firing sends the SYN again: the waits 0 to tcp_syn_linear_timeouts keep the initial RTO of 1 s (TCP_TIMEOUT_INIT), then each firing doubles the wait, up to TCP_RTO_MAX. Each wait is a timer of the wheel: a wait of 1 s is up to 64 ms late at HZ 1000. syn_giveup is what gives up on it. Through a stall of the path, the first SYN after its end gets through (recovery)"
+		},
 		"retransmit": {
 			"duration": {
 				"model": "linux.rto-wait",