Repository navigation
Open
Description
I asked Claude for a code review of this repo.
First, according to claude, this repo is mainly relevant for sbt 1.x
- sbt 1.13.x uses it for every server and client connection. JNI is the default on Apple Silicon, JNA everywhere else.
- sbt 2.x uses the JDK's own Unix domain sockets
sbt 2.x still bundles ipcsocket in sbtn for Windows, as the named pipe client for servers older than 2.1.0.
Both sbt 1.x and 2.x use ipcsocket from the 1.x branch; develop never shipped. Fixes need to be backported to 1.x where they apply.
PRs
| PR | Fix | Priority | Status |
|---|---|---|---|
| #96 | ForwardingSocketChannel ignored ByteBuffer position/limit (crash at end of stream, endless write loop) |
– (not in 1.x) |
merged |
| #97 | JNI: socket path longer than the system limit overwrote memory and crashed the JVM; array bounds | High | merged |
| #98 | Native library loaded from a predictable /tmp directory another local user could control |
Medium, security | merged |
| #99 | Unix socket closed itself at end of stream, so a response after shutdownOutput() was lost |
Medium, behavior change | merged |
| #100 | Closing a Windows named pipe socket leaked the pipe, two event handles and a TCP socket | High, behavior change | merged |
| #102 | JNA: a timed-out named pipe read kept writing into freed memory and lost data; a pending write could read freed memory | Medium (on 1.x only the write part applies) |
open |
| #103 | Closing sockets and server sockets: wake blocked threads, isBound/isClosed, close() during accept(), Windows close hang |
High | open |
| later | JNI named pipe read/write kept going after errors: hang when the peer closes, read timeout fails concurrent writes | Medium | needs #102 |
| later | JNA reported wrong Windows error codes | Low | needs #102 |
| later | Another program could take over the named pipe's name ("pipe squatting") | Medium, security | needs #103 |
| later | CI: run JNI tests with -Xcheck:jni, fail on warnings from ipcsocket's JNI code |
– | needs the JNI PR |
Impact on sbt 1.13.x
- sbt's server reads every connection with a 5 s timeout, but ipcsocket
1.xignores read timeouts, so the timeout bugs fixed by Fix use of freed memory when a named pipe read times out (JNA) #102 and the JNI PR don't affect sbt. - sbt sets
sbt.ipcsocket.tmpdirto a directory as predictable as ipcsocket's default; Don't load the native library from a directory another user controls #98 doesn't check it, so this is fixed in sbt ([2.x] fix: Don't extract native libraries to a directory another user controls sbt#10009, [1.x] bport: Don't extract native libraries to a directory another user controls sbt#10013). - On Apple Silicon (JNI), the thin client and BSP can't connect when the socket path has non-ASCII characters (Fix Unix socket paths and array bounds in the JNI code #97). This comes from reading the code; it hasn't been reproduced.
- sbt's
BootServerSocketskipsclose()on Windows because it is "very slow"; that's the close hang fixed in Fix closing sockets and server sockets #103, so sbt can drop the workaround. - With Keep a Unix socket open after reading end of stream #99, sbt should check for -1 in
ReadJsonFromInputStream's body loop: a client disconnecting mid-message would otherwise make it spin.
Not fixed
- On Windows, a closed peer surfaces as an
IOExceptionfromread, not -1. Left alone because callers may depend on it. - Low priority: JNI string leaks, a fd leak when
bindfails in theUnixDomainServerSocketconstructor, no arm64 JNI library on Windows, a missing NULL check onCreateEvent, inefficientreadLine/readAll. - The bundled native binaries predate the recent C fixes; they need a rebuild before the next release.
Metadata
Metadata
Assignees
Labels
No labels
Activity
eed3si9n commentedon Oct 28, 2026
That's right. Since JDK 17 ships with Unix domain sockets, I've ripped out ipcsocket from sbt's develop branch (sbt 2.1.0) from the most part. Since we build sbtn off of the develop branch, and is backported to 1.13.x and 2.0.x, I've kept named pipe support so new sbtn can continue to talk to old sbt servers.
eed3si9n commentedon Oct 29, 2026
FYI - ForwardingSocketChannel was an idea I had before sbt 2.x requiring JDK 17, and thought it could multiplex between sockets and SocketChannel. It shipped as 2.0.0-M2, but I couldn't get it to stabilize so the develop branch never shipped. sbt 1.x and 2.x are both using ipcsocket 1.x branch.
lrytz commentedon Oct 29, 2026
OK, thanks for the heads-up. I'll do backport PRs to 1.x for the ones that apply.