From: knu@...
Date: 2014-10-31T13:22:03+00:00
Subject: [ruby-core:66035] [ruby-trunk - Bug #9494] [Closed] Race condition in autoload of Digest::SHA256, etc.

Issue #9494 has been updated by Akinori MUSHA.

Status changed from Assigned to Closed
% Done changed from 0 to 100

Applied in changeset r48213.

----------
Make Digest() thread-safe.

* ext/digest/lib/digest.rb (Digest()): This function should now be
  thread-safe.  If you have a problem with regard to on-demand
  loading under a multi-threaded environment, preload "digest/*"
  modules on boot or use this method instead of directly
  referencing Digest::*. [Bug #9494]
  cf. https://github.com/aws/aws-sdk-ruby/issues/525

----------------------------------------
Bug #9494: Race condition in autoload of Digest::SHA256, etc.
https://bugs.ruby-lang.org/issues/9494#change-49754

* Author: Conrad Irwin
* Status: Closed
* Priority: Normal
* Assignee: Akinori MUSHA
* Category: ext
* Target version: 
* ruby -v: 2.0.0p247
* Backport: 1.9.3: UNKNOWN, 2.0.0: UNKNOWN, 2.1: UNKNOWN
----------------------------------------
At the moment the digest extension uses const_missing to implement autoload. Unfortunately there's a race condition: another thread can run after the constant is defined, but before it is initialized. (http://git.io/YW3bbA)

There's a minimal test case here: http://git.io/5umBqQ ��� this was happening to us regularly in production because each sidekiq worker would race to create a hash.

The work around we have deployed is to require digest/sha2.so pre-emptively. And the attached patch does that.

An alternative solution would be to use `autoload`, but that is deprecated; or to change the way meta-data is stored for Digest classes so that it can be accessed immediately. There's no way to use a mutex inside const_missing to fix this, because the bug is that the constant becomes not-missing before it is fully defined.

---Files--------------------------------
require-digest-upfront.patch (1006 Bytes)


-- 
https://bugs.ruby-lang.org/