httpd

An HTTP/1.1 server written in sysl: sockets, timers and tasks from kairos's libuv driver, and parsing from llhttp. One task accepts; each connection is one task that reads, parses, answers and decides when the connection ends.

It does keep-alive and pipelining. llhttp is fed whatever each read returned, so a request split across reads is assembled across them, and several pipelined requests in one read are answered in order, in one write.

Running it

sysl build .
./httpd 8080 static

httpd [port] [root]: the port defaults to 8080 and the root to static. SIGINT or SIGTERM shuts it down gracefully.

Routes

route	answer
GET /	Hello, world!
GET /static/<file>	the file from the root directory, read on libuv's thread pool; 404 if it is missing, or if the name holds a / or starts with .
POST /echo	the request body
GET /sleep?ms=N	slept, after an await lp.sleep(N ms); N is 0 to 60000

A wrong method gets 405 and any other path gets 404.

What it refuses, and how it ends

case	answer
a request not complete within the request timeout (10 s), counted from the end of the previous answer	408, then close
a body over the limit (1 MiB), declared or chunked	413, then close
a head over 8 KiB	431, then close
a malformed request	400, with llhttp's reason as the body, after the good requests before it
a connection over the cap (10,000)	503, then close
SIGINT / SIGTERM	the listener and every idle connection close at once; a connection that is answering finishes its answer, says Connection: close, and then closes. The loop ends when the last one has

A client that disappears mid-answer only ends its own task: the failed write returns, and the task's defer takes the connection out of the server's table and closes it.

Testing

sysl test .

Twenty tests. Each one starts the server in process on a port the kernel chooses, and drives it with clients that are kairos tasks on the same loop. The clients write raw bytes and compare the exact bytes that come back. The suite covers every route, chunked bodies, keep-alive against Connection: close, HTTP/1.0, pipelined requests split at every byte boundary, a request sent a byte at a time, and every case in the table above. One more test is marked ignore. It needs a sysl fix, because the task lp.timeout gives up on is released too late.

Against C

The baseline is a minimal C server built on the same libuv and the same llhttp, compiled with clang -O2. It serves GET / and /sleep only. The sysl server is built at optimization = "2". The load generator is oha (-z 8s), with keep-alive and everything on one machine: an Apple M5 Pro (18 cores, 64 GB) on macOS 26.6.2, measured 2026-10-13 with sysl 0.1.0-alpha.11, kairos 0.1.1, llhttp 0.2.0 and libuv 1.x from Homebrew.

route	connections	sysl req/s	C req/s	sysl / C	sysl p50	C p50	sysl p99	C p99
GET /	1	31,243	31,619	0.99	0.029 ms	0.029 ms	0.078 ms	0.073 ms
GET /	64	154,085	180,105	0.86	0.40 ms	0.35 ms	0.62 ms	0.57 ms
GET /	512	149,976	199,305	0.75	3.46 ms	2.51 ms	4.37 ms	4.87 ms
/sleep?ms=10	1	90	90	1.00	11.12 ms	11.12 ms	11.22 ms	11.21 ms
/sleep?ms=10	64	5,501	5,741	0.96	11.64 ms	11.41 ms	11.96 ms	11.89 ms
/sleep?ms=10	512	35,735	47,516	0.75	14.32 ms	10.73 ms	16.21 ms	12.03 ms

At 512 connections a profile of GET / is mostly write and read system calls. What sysl adds on top is mostly malloc and free: a fresh slice for every read (zeroed and then copied), and the strings that build each answer. Over two 60-second runs of GET / at 512 connections (about 17 million requests), resident memory stayed flat during each run: 61.9 MB, then 69.5 MB. Most of that is libuv's 64 KiB read buffer for each open connection.

Connections leak, but requests do not. Every connection leaves 17 counted allocations behind (SYSL_COUNT_ALLOCS=1, 100 and 1,000 connections). Fourteen of them are llhttp's Requests: the reader holds its parser, and the parser's callback captures the reader, so the two form a cycle that nothing ever frees.

License

ISC