x/tools/gopls: crashes due to apparent data race (gopls v16+ on linux)

[findleyr]

While we've found and fixed many gopls crashes thanks to telemetry, as we gain more users it appears we're accumulating a long tail of issues that "can't happen" based on easily verifiable local invariants.

In some cases, these were related to misreporting of stacks by the runtime (#70637), but others cannot be explained by a misreported stack, and indicate some form of memory corruption or (most likely) data race.

This issue is an umbrella for those bugs. @adonovan and I will consolidate them here.

[gabyhelp]

Related Issues

• x/tools/gopls/internal/lsp/cache: data race in overlayFS methods #58151 (closed)

_{(Emoji vote if this was helpful or unhelpful; more detailed feedback welcome in this discussion.)}

[adonovan]

Fun fact: with one exception, all of the attached "can't happen" telemetry issues are from gopls/v0.17 on linux/amd64, even though linux accounts for only 22% of gopls users. I'm going to tentatively call this a linux-specific bug introduced after gopls 15. (Recall that 16 had broken telemetry.)

This is the current set of all reports:

golang.org/x/tools/[email protected] go1.23.2 linux/amd64 vscode (2)
golang.org/x/tools/[email protected] go1.23.2 linux/amd64 vscode (2)
golang.org/x/tools/[email protected] go1.23.4 linux/amd64 vscode (1)
golang.org/x/tools/[email protected] go1.23.4 linux/amd64 vscode (1)
golang.org/x/tools/[email protected] go1.23.4 linux/amd64 vscode (1)
golang.org/x/tools/[email protected] go1.23.4 linux/amd64 vscode (1)
golang.org/x/tools/[email protected] go1.23.4 linux/amd64 vscode (1)

It's conceivable that all come from a single VS Code user with faulty hardware, though unlikely since there are 3 distinct gopls/Go version combinations here. (Maybe they were frantically upgrading after crashes?)

The exception is #71067, from gopls/v0.15 on darwin/arm64, and which could be explained by a bug in IExportShallow. I will split it out.

[adonovan]

@findleyr Perhaps you could run gopls built with -race for a while? I've been doing it on my mac, but no problems so far.

[adonovan]

The latest associated issue (#72931) is something new:

• golang.org/x/tools/[email protected] go1.23.6 linux/amd64 vscode (1)