cmd/pprof: invokeVisualizer is racy

[dsnet]

In invokeVisualizer, we defer deleting the temporary file when invoking the visualizer until the end of the pprof program. However, the invokeVisualizer does not wait for the viewer to finish when not running in interactive mode.

Thus, the following race occurs in non-interactive mode:

• pprof generates some report
• pprof invokes the browser to show report; Chrome starts running
• Without waiting for browser to end, we return from invokeVisualizer and the pprof program begins terminations.
• The DeferDelete is run, deleting the report file
• Chrome has finished starting up, but cannot find the file. Oops.

The fix is probably to wait for the visualizer to end regardless of "interactivity".

[minux]

Even if it waits for the visualizer command to exit, the problem still exists because some visualizer exits before the UI is closed (e.g. chrome when a chrome window is already open.)

[dgryski]

/cc @rauls5382

[dsnet]

I noticed that there is already a 1 second wait in the logic I pointed to. I must have been really unlucky and Chrome somehow took longer than 1 second to load up. I'm not sure if there's anything we could do here then. Increasing the wait time seems like a bad idea (how high is high enough?).

[rauls5382]

Note the logic is to do wait for the visualizer in non-interactive mode. On interactive mode (pprof shell) we do not wait to allow you to let you go back to the command line and spawn other visualizers in parallel.

Having said that, there is a fundamental race condition: we to write the files, run the viewer, wait for the viewer read the files, and then delete them. The problem happens when the visualizer is a command that signals another command (eg chrome), so the visualizer exits even though the files haven't been read.

I could agree with special-casing the delay for some platforms.