Autocannon sends two extra bytes at the end of the body, breaking the protocol

I’ve tested uWS HTTP server and passed little body in every http request. It produced strange results. I asked (https://github.com/uNetworking/uWebSockets.js/issues/228) creator of uWS why it happens and he researched what "autocannon sends two extra bytes at the end of the body, breaking the protocol. uWebSockets does not close the connection immediately, as it has a different parsing strategy. "

Can you fix it, please?

$ autocannon -c 100 -p 100 -d 10 -m POST -b '{"a":1}' http://0.0.0.0:8080
Running 10s test @ http://0.0.0.0:8080
100 connections with 100 pipelining factor

┌─────────┬──────┬──────┬───────┬───────┬──────────┬──────────┬────────────┐
│ Stat    │ 2.5% │ 50%  │ 97.5% │ 99%   │ Avg      │ Stdev    │ Max        │
├─────────┼──────┼──────┼───────┼───────┼──────────┼──────────┼────────────┤
│ Latency │ 0 ms │ 0 ms │ 56 ms │ 82 ms │ 13.01 ms │ 34.33 ms │ 3618.88 ms │
└─────────┴──────┴──────┴───────┴───────┴──────────┴──────────┴────────────┘
┌───────────┬─────┬──────┬────────┬────────┬────────┬─────────┬─────────┐
│ Stat      │ 1%  │ 2.5% │ 50%    │ 97.5%  │ Avg    │ Stdev   │ Min     │
├───────────┼─────┼──────┼────────┼────────┼────────┼─────────┼─────────┤
│ Req/Sec   │ 0   │ 0    │ 1253   │ 3683   │ 1655.6 │ 1489.25 │ 26      │
├───────────┼─────┼──────┼────────┼────────┼────────┼─────────┼─────────┤
│ Bytes/Sec │ 0 B │ 0 B  │ 231 kB │ 678 kB │ 305 kB │ 274 kB  │ 4.78 kB │
└───────────┴─────┴──────┴────────┴────────┴────────┴─────────┴─────────┘

Req/Bytes counts sampled once per second.

17k requests in 10.15s, 3.05 MB read
7k errors (0 timeouts)

[ghost]

You can clearly see in Wireshark that autocannon sends invalid HTTP in the case of post. It sends Content-length: 7 followed by 7 bytes of content followed by 2 invalid bytes, breaking the protocol, causing the connection to eventually fail

It works just fine in wrk, https://github.com/wg/wrk as that client actually follows HTTP.

The protocol is like this:

header: value rn header: value rn rnrn content

but autocannon sends

header: value rn header: value rn rnrn content rn <- those shouldn’t be there, they break the protocol, botching the entire benchmark as things deteriorate to HTTP1.0 kind of connection close

[GlenTiki]

Oh, it looks like you’re correct. Thanks for investigating. We add those /r/n for simpler logic for pipelining reasons, I think. I don’t think we realised it was invalid. I’ll put together a fix.

Thanks again!

On Wed 20 Nov 2019, 3:34 p.m. Matteo Collina, [e-mail] wrote:

I’m happy to take a PR, I’m currently swamped. Il 20 nov 2019, 16:32 +0100, Alex Hultman [e-mail], ha scritto:

— You are receiving this because you commented. Reply to this email directly, view it on GitHub, or unsubscribe.

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/mcollina/autocannon/issues/229?email_source=notifications&email_token=ABUWMBZWJPHG3A6U6LZNDEDQUVKGXA5CNFSM4JPCTER2YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEESMGZA#issuecomment-556057444, or unsubscribe https://github.com/notifications/unsubscribe-auth/ABUWMBZADJES6DM2N2JUORLQUVKGXANCNFSM4JPCTERQ .