Bug 27947 – A bug in SQLiteTransaction can lead to a failed ASSERT

• Component: New Bugs
• Status: RESOLVED FIXED

[Dumitru Daniliuc]

SQLiteTransaction::begin() sets m_db.m_transactionInProgress to true even if m_db.executeCommand("BEGIN;") fails. This can lead to an ASSERT failing incorrectly in SQLTransaction::openTransaction() (steps 1 + 2).

[Dumitru Daniliuc]

Created attachment 33997 [details]
patch

[Darin Adler]

Comment on attachment 33997 [details]
patch

Seems fine. Is there any way to make a regression test for this? Under what circumstances will the BEGIN command fail?

r=me

[Dimitri Glazkov (Google)]

Should we need another assert after BEGIN. This shouldn't be happening normally, right?

It seems like we should be doing something more than just reporting !inProgress in these cases.

[Dumitru Daniliuc]

Created attachment 34000 [details]
test case

(In reply to comment #2)
> (From update of attachment 33997 [details])
> Seems fine. Is there any way to make a regression test for this? Under what
> circumstances will the BEGIN command fail?
> 
> r=me

BEGIN seems to fail if the same DB thread tries to run 2 transactions in 2 different Database objects that point to the same DB file. I attached the example we used.

I don't have a detailed explanation for what's going on yet -- we think there's a race condition here and I'm still debugging it (details + patch to come in a future bug). However, always setting m_db.m_transactionInProgress to true seems wrong and leads to crashes in Chromium's and Safari's debug builds (you might have to click the "Test" button a couple of times). So I thought it wouldn't hurt to submit a quick patch for this before we're done investigating the bigger issue.

[Dimitri Glazkov (Google)]

Landed as http://trac.webkit.org/changeset/46736.

Please create a bug for tracking the origin of the issue.