Skip to content

Fix deadlock#316

Merged
zenhack merged 2 commits into
capnproto:mainfrom
zenhack:fix-deadlock
Oct 1, 2022
Merged

Fix deadlock#316
zenhack merged 2 commits into
capnproto:mainfrom
zenhack:fix-deadlock

Conversation

@zenhack

@zenhack zenhack commented Oct 1, 2022

Copy link
Copy Markdown
Contributor

See the first commit message. The second commit fixes similar logic errors, though I haven't seen consequences of those errors.

There are some other tests that occasionally deadlock, but I haven't tracked down the cause yet.

It is illegal to resolve promises while holding connection locks, and I
was seeing TestHandleReturn_regression deadlock every 6th run or so;
changing this fixes it.

Note: since this makes the resolution itself and the manipulation of
the questions table non-atomic, the ordering is critical:

1. First we remove the entry from the table. This means that if the
remote vat sends us messages pertaining to that question id, they will
simply be rejected instead being acted on. This shouldn't happen
normally, but maybe a malicious vat could get us to panic or something.
2. Second, resolve the promise. After this is complete, we can be sure
the client won't be used.
3. Then, and only then, we can return the question id to the pool for
re-use.
Unlike the previous commit, I haven't observed the consequences of
these, but the same logic & ordering constraints apply.
@zenhack zenhack mentioned this pull request Oct 1, 2022
@zenhack zenhack merged commit 00c7c66 into capnproto:main Oct 1, 2022
@zenhack zenhack deleted the fix-deadlock branch October 1, 2022 06:43
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants