x/crypto/ssh: Channel.Read incorrectly returns write error instead of io.EOF

[hwh33]

What version of Go are you using (go version)?

go version go1.16.3 darwin/amd64

Does this issue reproduce with the latest release?

Yes.

What operating system and processor architecture are you using (go env)?

go env Output

GO111MODULE=""
GOARCH="amd64"
GOBIN=""
GOCACHE="/Users/hwh33/Library/Caches/go-build"
GOENV="/Users/hwh33/Library/Application Support/go/env"
GOEXE=""
GOFLAGS=""
GOHOSTARCH="amd64"
GOHOSTOS="darwin"
GOINSECURE=""
GOMODCACHE="/Users/hwh33/Code/golang/pkg/mod"
GONOPROXY="github.com/getlantern/*"
GONOSUMDB="github.com/getlantern/*"
GOOS="darwin"
GOPATH="/Users/hwh33/Code/golang"
GOPRIVATE="github.com/getlantern/*"
GOPROXY="https://proxy.golang.org,direct"
GOROOT="/usr/local/go"
GOSUMDB="sum.golang.org"
GOTMPDIR=""
GOTOOLDIR="/usr/local/go/pkg/tool/darwin_amd64"
GOVCS=""
GOVERSION="go1.16.3"
GCCGO="gccgo"
AR="ar"
CC="clang"
CXX="clang++"
CGO_ENABLED="1"
GOMOD="/Users/hwh33/Code/golang/src/github.com/getlantern/ossh/go.mod"
CGO_CFLAGS="-g -O2"
CGO_CPPFLAGS=""
CGO_CXXFLAGS="-g -O2"
CGO_FFLAGS="-g -O2"
CGO_LDFLAGS="-g -O2"
PKG_CONFIG="pkg-config"
GOGCCFLAGS="-fPIC -arch x86_64 -m64 -pthread -fno-caret-diagnostics -Qunused-arguments -fmessage-length=0 -fdebug-prefix-map=/var/folders/cm/89l8fqlj6c5dc09fc9mb14dh0000gn/T/go-build3841693344=/tmp/go-build -gno-record-gcc-switches -fno-common"

What did you do?

• Set up a pair of ssh.Conns (client and server).
• Open a channel between the connections.
• Write a bunch of data to the client side of the channel.
• Close the client connection.
• Read from the server channel.

Example:
https://play.golang.org/p/JAd3orfmiye

(The Go Playground sometimes sees the TCP dial/listen as a deadlock. It will often work though.)

What did you expect to see?

The data from the peer, then io.EOF.

What did you see instead?

Partial data, and a write-related error. Usually write: broken pipe or write: protocol wrong type for socket

More details

This only happens if the peer has closed the ssh.Conn (or the underlying transport). I believe the bug is rooted in an incorrect assumption here. The comment indicates that the channel should be tolerant of errors adjusting the window when the peer has closed the connection. The assumption is that such errors will be exclusively io.EOF. Perhaps something has changed, but this is not currently the case.

I'm actually not sure what the fix should be here as I'm not sure what error could be expected. I don't think net.Conn.Write makes any kind of guarantees about behavior when the peer has closed the connection. Somehow though, the buffered data should be returned to the reader, followed by io.EOF.

[hwh33]

It seems like there's another issue beyond the block of code I linked to above (this one). Even if you ignore the error returned by c.adjustWindow there, the Channel.Read call still sometimes returns an early io.EOF.

[cagedmantis]

/cc @FiloSottile

[willmo]

https://go-review.googlesource.com/c/crypto/+/459915 would make encountering this less likely (as a side effect), since c.adjustWindow would write less often.

I thought about making another CL to simply ignore all errors from c.adjustWindow or its call to c.sendMessage, but I'm not sure if that's really a sensible thing to do. But returning all non-EOF errors seems wrong too.

Even if you ignore the error returned by c.adjustWindow there, the Channel.Read call still sometimes returns an early io.EOF.

I don't think it's strictly guaranteed that the data in your Channel.Write made it out to the underlying net.Conn before you closed it, even though you waited for Channel.Write to return. For example, writes are buffered during periodic rekeying. Other than that, though, I don't know why this would happen with your playground example, since the server should read and buffer the channel data before seeing the connection close, and buffer.Read doesn't return EOF until all buffered data is consumed.