Bug #14353: $SAFE should stay at least thread-local for compatibility

• Status: Closed
• Assignee: matz (Yukihiro Matsumoto)
• Target version: -
• Backport: 2.3: UNKNOWN, 2.4: UNKNOWN, 2.5: UNKNOWN

[Eregon (Benoit Daloze)]

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:

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):

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 (Koichi Sasada) Can we switch it back to thread-local for compatibility, avoiding headaches and keeping it usable with multiple threads?

[Eregon (Benoit Daloze)]

• Assignee set to ko1 (Koichi Sasada)

[Eregon (Benoit Daloze)]

• Target version set to 2.6

[Eregon (Benoit Daloze)]

• Related to Feature #14250: Make `$SAFE` process global state and allow to set 0 again added

[shevegen (Robert A. Heiler)]

I have seen $VERBOSE used in ways very similar as you described though and have used
it myself; most commonly I used it e. g. when I would use syck for yaml files rather
than psych, in order to suppress warnings I'd get during loading. It never felt very
elegant though to use the re-assignment trick. Can't think of a more elegant
solution either.

[ko1 (Koichi Sasada)]

• Assignee changed from ko1 (Koichi Sasada) to matz (Yukihiro Matsumoto)

Of course, I proposed $SAFE as Fiber local.

However, Matz said process global is enough.
akr also said nobody use $SAFE correctly, so that such incompatibility is not a matter.

I'm neutral. Matz, please decide it.

[Eregon (Benoit Daloze)]

Of course, I proposed $SAFE as Fiber local.

:)

If $SAFE had no effect at all, then I think it would be fine to make $SAFE a normal process-wide global variable.
But this is not the case, $SAFE still causes SecurityError.
Therefore any library/code setting $SAFE will potentially break other threads, which is quite incompatible.

I think a more compatible way to remove $SAFE is:

• Warn that $SAFE is deprecated.
• Make it Fiber-local or Thread-local. (because very few people expect it frame-local and it incurs a cost)
• Warn that $SAFE is has no longer any effect (or maybe even raise when $SAFE is set, like for levels 2-4) and remove $SAFE checks, so the value of $SAFE has no effect.
• Remove the variable entirely or let it be process global.

Something like

ruby -e 'Thread.new{ $SAFE=1 }.join; Thread.new{ load "test.rb".taint }.join'

fails, prints no warning and could happen in any big application if $SAFE is set to 1 in some Thread.
This doesn't seem worth breaking. Even more so without a warning.

[naruse (Yui NARUSE)]

• Target version deleted (2.6)

[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

[matz (Yukihiro Matsumoto)]

I understand the compatibility situation. But I don't think it's worth effort for development. I vote for making $SAFE no effect (in 2.7).

Matz.

[k0kubun (Takashi Kokubun)]

• Is duplicate of Feature #5455: $SAFE should be removed added

[jeremyevans0 (Jeremy Evans)]

• Status changed from Open to Closed

$SAFE is being deprecated in 2.7 and will revert to normal global variable in 3.0, so this can be closed.