Spurious "Connection died, tried 5 times before giving up" using HTTP/2

[alexcrichton]

I did this

Currently the Cargo package manager for Rust uses Curl to download files over HTTP, namely dependency libraries. Recently Cargo, on Rust's nightly channel, updated from curl 7.76.0 to curl 7.77.0. Since updating users are reporting an increase in errors seen while downloading.

I've personally been seeing this in a number of CI builds that use Cargo with curl 7.77.0, and the errors as-reported by Cargo is:

error: failed to download from `https://crates.io/api/v1/crates/wat/1.0.38/download`

Caused by:
  [55] Failed sending data to the peer (Connection died, tried 5 times before giving up)

I'm able to reproduce this locally on a machine where it just downloads dependencies in a loop, and frequently it will hit this error while downloading. Since this was an apparent regression in Curl I did a bisection of commits between 7.76.0 and 7.77.0 where on each commit I built that version of Curl into Cargo and then downloaded all of Cargo's own dependencies 100 times, where if any download failed that mean the commit was "bad". Bisection resulted in pointing to this commit -- 252790c -- where download would fail spuriously after that but would succeed at least 100 times before that.

Unfortunately reproduction is unlikely to be super easy unless you want to dive into Cargo (which I'm assuming you'd rather not). Cargo uses Curl as a library, so Cargo's usage is pretty far removed from the curl command line tool for example. In theory Cargo is doing pretty a pretty bland request of "please download all these files over HTTP/2" with some reasonable timeout and retry limits configured. Cargo enables HTTP/2 and pipewait with a maximum of 2 connections per-host at this time.

To hopefully help in assisting with debugging this I captured debug logs from curl during a failed transfer. These debug logs were captured at 252790c with DEBUG_HTTP2 enabled as well.

I'm happy to help test out patches and/or get more debugging information if necessary. I can post reproduction steps as well but it requires using Rust to build Cargo and having a nontrivial setup to thread through a custom Curl dependency into Cargo.

I expected the following

Given the frequency of these errors it seems unlikely that this is caused by an actual spurious error, so the expectation is that something is retried internally as necessary or requeued as necessary without bubbling up an error.

curl/libcurl version

This relates to curl 7.77.0, specifically an apparent regression introduced with 252790c

operating system

The system I tested this on was:

Linux 4.19.0-9-arm64 #1 SMP Debian 4.19.118-2+deb10u1 (2020-06-07) aarch64 GNU/Linux

but this seems to appear frequently with GitHub Actions runners as well, and I've seen it locally on my macbook too.

[alexcrichton]

Some further debugging and I think I've got a way to reproduce this a bit easier. Given this urls.txt file I can reproduce the issue with the curl command line executable:

curl --remote-name-all -Ss $(cat /urls.txt) -L -Z --parallel-max 100 --http2-prior-knowledge

I've re-bisected using this command, and bisection again points at 252790c as the commit between 7.76 and 7.77 where errors started happening.

For reference I used this Dockerfile for bisection and for nghttp2 it was using nghttp2/nghttp2@31253f4 (current HEAD at the time of me running this). I was also building locally with ./configure --disable-shared --enable-debug --with-openssl --with-nghttp2.

The custom build also yields:

# curl --version --verbose
curl 7.77.0-DEV (aarch64-unknown-linux-gnu) libcurl/7.77.0-DEV OpenSSL/1.1.1f zlib/1.2.11 nghttp2/1.40.0
Release-Date: [unreleased]
Protocols: dict file ftp ftps gopher gophers http https imap imaps mqtt pop3 pop3s rtsp smb smbs smtp smtps telnet tftp
Features: alt-svc AsynchDNS Debug HSTS HTTP2 HTTPS-proxy IPv6 Largefile libz NTLM NTLM_WB SSL TLS-SRP TrackMemory UnixSockets

and these are the verbose logs of a failed transfer

[alexcrichton]

Er sorry, I forgot to enable http2 debug logging for those logs. This is a failure log with DEBUG_HTTP2 defined in lib/http2.c.

Additionally I'm not sure if this is related to #5611, it may be the same issue?

[bagder]

curl --remote-name-all -Ss $(cat /urls.txt) -L -

Does this command line reliably reproduce the problem for you? I've run it many times now on my dev box (using current curl built from git) but it always succeeds. Does the problem still remain for you with curl 7.78.0 ?

[alexcrichton]

I indeed cannot reproduce on the current master branch, but I believe the reason for that is that nghttp2 is no longer detected after b4b34db, so http2 isn't used (and there's no bug unless http2 is used).

Inside of config.log I see:

/usr/bin/ld: cannot find -l-lnghttp2

Is this a local configuration error for me? When you build locally do you have http2 enabled with nghttp2?

[alexcrichton]

For reference, though, on fbf2659 (current master) if I revert b4b34db I'm able to still reproduce the issue on my end. Curl still occasionally prints out:

$ curl --remote-name-all -Ss $(cat /urls.txt) -L -Z --parallel-max 100 --http2-prior-knowledge
curl: (55) Connection died, tried 5 times before giving up

[bagder]

How silly, so okey #7515 should fix that diversion...

[jay]

I'm able to reproduce this in Windows but rarely and I'm unable to get it down to fewer URLs to eliminate the noise. I had to put the URL list in a config style file (eg lines prefixed with url= like url = http://theurl) and passed them via config (curl ... -K urls.txt) due to shell differences. I looked through a wireshark decrypted http2 capture and I see all redirects and nothing unusual. I should note I did it manually and may have missed something, since I don't have an automated way to scan the packet capture for unusual circumstance; it's more of I know it when I see it.

[alexcrichton]

For your reproduction on Windows, were following redirects enabled? It's expected that the URL https://crates.io/api/v1/crates/quote/1.0.9/download redirects to https://static.crates.io/crates/quote/quote-1.0.9.crate, and at least when I was testing locally I had to ensure redirects were followed to get it to more reliably fail.