Error while files.add over HTTP API (http: invalid Read on closed Body)

[lidel]

This bug was originally found in ipfs/ipfs-companion#480.
Since then I was able to confirm issue originates between js-ipfs-api and HTTP API exposed by go-ipfs and created a demo app that reproduced issue on two separate machines.

Version information:

OS: two Linux boxes (one was Gentoo second was Debian)

IPFS Node:

go-ipfs version: 0.4.16-rc1-b183da3
Repo version: 7
System version: amd64/linux
Golang version: go1.10.3

HTTP API Client: js-ipfs-api 22.1.1 running in browser context

Type:

Bug

Note: This issue is about error in response produced by go-ipfs. There is a separate bug for the lack of proper handling of partial errors in js-ipfs-api at https://github.com/ipfs/js-ipfs-api/issues/797)

Description:

Sometimes (for specific payloads in browser context) the HTTP API for ipfs.files.add provided by go-ipfs returns the http: invalid Read on closed Body error.

It is especially problematic when adding multiple files at once, because error is returned in the middle of response, along with successful entries with the same payload:

How to Reproduce

The issue does not seem to appear when curl is used, it seems to be specific to browser context and the way js-ipfs-api makes requests to HTTP API.

I've found two ways to reproduce/investigate it:

A) By inspecting recorded HTTP interaction

Below is a recorded HTTP Archive (HAR) of interaction that produced the error:
Archive%2018-06-28%2015-34-47.har

Interaction was uploading these sample files with demo app (details below).

B) By Using Demo App

Turns out the .zip with demo app triggers the issue 🙃 , so its easy to reproduce:

1. Expose API of go-ipfs 0.4.15 or 0.4.16-rc1 locally on port :5001
2. Download demo app:upload-multiple-files-via-browser-ipfs-api-bug-demo.zip
   (uses ipfs-api ^22.1.1 and uploads files via ipfs.files.add with wrapWithDirectory: true)
3. Build it and start via npm install && npm start, then open form at http://localhost:3000
4. Reproduce using samples known to trigger issue at go-ipfs
   • Sample 1: upload single file: upload-multiple-files-via-browser-ipfs-api-bug-demo.zip
     • This should produce error
   • Sample 2: upload these three files at once
     • This should produce error

[Stebalien]

Looks like a go bug: golang/go#15527

As far as I can tell, work around is to avoid flushing the response till we finish reading the request.

[Stebalien]

So, I'm pretty sure that work around is incorrect (although it appears to work). However, the diagnosis is definitely correct.

[Stebalien]

This may also explain why our progress meters are broken. It may also explain a bunch of random errors and why some people have trouble adding large datasets to go-ipfs.

[Stebalien]

Ah. So, this has worked because we almost always use chunked encoding (which doesn't trigger the issue).

*edit: this worked because the client was sending "connection: close".

[Stebalien]

So, I'm not sure how to work around this. We rely on being able to stream status information while processing large requests. We usually don't notice this issue because the go-ipfs CLI usually makes a single HTTP requests and sets Connection: close however, this isn't something we're going to be able to control from the browser. Really, we don't really want to be closing connections from the browser (or other complex applications) for performance reasons. The only solutions I can see are:

1. Fork net/http and somehow try to maintain compatibility with the rest of the net/http ecosystem. I tried this. It started getting real ugly.
2. Use a different http library. I can't find any good ones that are even reasonably compatable.
3. Disable keep alive on the server (bad performance).

I get the impression it'll be a while before go fixes this bug.

[Stebalien]

@lidel try the fix/disable-keepalives branch if you need something to work with while we try to find a better fix.

[lidel]

@Stebalien thank you for looking into this. With this info I strongly suspect the bug is responsible for a bunch of random errors users of HTTP API in browser context been reporting over past two years.

I was able to confirm workaround from fix/disable-keepalives branch "solves" the issue, unfortunately most of users won't be able to run custom build.

Edited: I was too optimistic: disabling `&progress` does not work as a reliable workaround, it just changes the conditions under which issue occurs (some files started working, others stopped).

However, you mentioned streaming status information, so I looked at bidirectional aspect of `files.add` and was able to confirm a viable workaround for the client side is to simply _disable progress reporting_.

Interaction samples with/without progress reporting:

• ERROR with &progress=true: HAR (http: invalid Read on closed Body)
• OK without &progress: HAR

Not ideal (we lose progress bars etc) but at least upload will work reliably.

[Stebalien]

Beyond "not ideal", it's also not a very reliable solution. What happens if you add a Connection: close header to all requests? That may be a viable "solution" until we can fix this.

[lidel]

After playing a bit with setting Connection: close on the request it I can confirm it.. seems to work reliably.

Linking to workarounds:

• ipfs-companion: created PR #509 which fixes the header in requests made to /api/v0/add.
  • It does not modify requests sent to all APIs, only one used by ipfs-companion
  • It works only in Firefox: Chrome does not allow modification of Connection header (Unable to upload to IPFS API: passing multiple files to ipfs.files.add sometimes fails ipfs-companion#480 (comment))
• js-ipfs-api: may not be possible to fix in browsers, details in https://github.com/ipfs/js-ipfs-api/issues/797#issuecomment-401469158