[ruby-core:91222] [Ruby trunk Bug#14353] $SAFE should stay at least thread-local for compatibility

Issue #14353 has been updated by vo.x (Vit Ondruch).


Just FTR, as per this discussion:

https://lists.fedoraproject.org/archives/list/[e-mail]/message/V234THAZ2ETTFVLJZXJYPH2QO5W7A3KL/

The global $SAFE breaks gettext testsuite:

https://github.com/ruby-gettext/gettext/commit/49b9f4ca66583395ddfa91503afd7593f069de18

As well as there are issues in postgresql-plruby:

https://github.com/devrimgunduz/postgresql-plruby/blob/master/extconf.rb#L21

----------------------------------------
Bug #14353: $SAFE should stay at least thread-local for compatibility
https://bugs.ruby-lang.org/issues/14353#change-76462

* Author: Eregon (Benoit Daloze)
* Status: Open
* Priority: Normal
* Assignee: matz (Yukihiro Matsumoto)
* Target version: 
* ruby -v: 
* Backport: 2.3: UNKNOWN, 2.4: UNKNOWN, 2.5: UNKNOWN
----------------------------------------
In #14250 $SAFE changed from a frame+thread-local variable to a process-wide global variable.

This feels wrong and breaks the most common usage of $SAFE in tests:

~~~ ruby
Thread.new {
  $SAFE = 1
  sth that should be checked to work under $SAFE==1
}.join
~~~

It is very clear this is incompatible given how many files (33!) had to be changed in r61510.
And it has wide ranging confusing side-effects, one example: https://travis-ci.org/ruby/spec/jobs/328524568

I agree frame-local is too much for $SAFE.

But removing thread-local seems to only introduce large incompatibilities.

It also makes it impossible to use it in a thread-safe way.
The common pattern (not necessarily for $SAFE, more often for $VERBOSE):

~~~ ruby
begin
  old = $SAFE
  $SAFE = 1
  something under SAFE==1
ensure
  $SAFE = old
end
~~~

is unsafe if two threads run it concurrently (The last thread executing `$SAFE = old` might restore to 1 even though it should be 0).

(Actually I believe most built-in variables (e.g. $VERBOSE) should be thread-local and not process-wide due to this)

Since $SAFE is being deprecated and removed, I don't see any reason to make it more incompatible than needed.

@ko1 Can we switch it back to thread-local for compatibility, avoiding headaches and keeping it usable with multiple threads?



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

Unsubscribe: <mailto:[e-mail]?subject=unsubscribe>
<http://lists.ruby-lang.org/cgi-bin/mailman/options/ruby-core>