1488990 - Do we handle content processes failing to advance to PROCESS_CONNECTED?

Status: NEW

(Core :: DOM: Content Processes, enhancement, P2)

Description

[Jed Davis [:jld] ⟨⏰|UTC-7⟩ ⟦he/him⟧]

GeckoChildProcessHost has several ideas of when the process is “launched”.  The PROCESS_CREATED is reached when there's an OS-level identifier for the process (pid, handle); PROCESS_CONNECTED is later, when the IPC hello message is received.

On Unix the channel is a socketpair, so if the child fails to start or crashes before sending the hello message, the parent will get EOF and go to the PROCESS_ERROR state and fire error callbacks and so on.  On Windows it's a named pipe that the process has to connect back to, and this is where there might be a problem: if the parent doesn't wait for PROCESS_CONNECTED with a timeout, would we ever detect the process failing to start?

GMP processes use GeckoChildProcessHost::SyncLaunch; NPAPI and GPU processes use WaitUntilConnected, and VR processes use the OnChannelConnected callback.  Content processes, however, use LaunchAndWaitForProcessHandle and don't appear to have any kind of timeout for reaching PROCESS_CONNECTED.

So it's possible that if a content process fails to start on Windows, the ContentParent might get stuck instead of reporting an error.  I don't have a Windows dev env at the moment so I haven't tried testing this yet; I found it while reading the code as part of bug 1446161.

Comment 1

[Jed Davis [:jld] ⟨⏰|UTC-7⟩ ⟦he/him⟧]

On second thought this probably belongs to DOM: Content Processes, because it's more or less specific to ContentParent.

Comment 2

[Jed Davis [:jld] ⟨⏰|UTC-7⟩ ⟦he/him⟧]

Bug 1471124 suggests that the answer is “no”: if I'm interpreting the data in that bug correctly, a content process is hanging indefinitely after it's PROCESS_CREATED and before it's even execed (so definitely isn't getting to PROCESS_CONNECTED) and one of the observed results was that every nth tab would fail to display, for n = the number of content processes. Which implies that a non-PROCESS_CONNECTED content process not only isn't ever timed out, but is repeatedly sent requests to create new PBrowsers while in that state.