Bug #15661

Disallow concurrent Dir.chdir with block

Added by headius (Charles Nutter) over 1 year ago. Updated about 13 hours ago.

Target version:


Dir.chdir with a block should disallow concurrent use, since it will almost never produce the results a user expects.

In calls for Dir.chdir to be made thread-safe were rejected because the underlying native call is process-global. This is reasonable because there's no way to easily make the native chdir be thread-local (at least not without larger changes to CRuby itself).

However the block form of chdir is explicitly bounded, and clearly indicates that the dir should be changed only for the given block. I believe Dir.chdir should prevent multiple threads from attempting to do this bounded chdir at the same time.

Currently, if two threads attempt to do a Dir.chdir with a block, one of them will see a warning: "conflicting chdir during another chdir block". This warning is presumably intended to inform the user that they may see unpredictable results, but I can think of no cases where you would ever see predictable results.

In other words, there's no reason to allow a user to do concurrent Dir.chdir with a block because they will always be at risk of unpredictable results, and I believe this only leads to hard-to-diagnose bugs.

The warning should be a hard error.

The warning should also be turned into an error if a non-block Dir.chdir call happens while a block Dir.chdir is in operation. The important thing is to protect the integrity of the current directory during the lifetime of that block invocation.

In CRuby terms, the patch would be to check for chdir_blocking > 0 and then simply error out, unless it's happening on the same thread.

Concurrent non-block Dir.chdir would be unaffected. This only protects the block form from having the current dir change while it is executing.


Updated by shevegen (Robert A. Heiler) over 1 year ago

I sort of agree with headius. I also can not really think of a possible and
deliberate use case for concurrent Dir.chdir with a block. Perhaps something
is missing (someone has a use case?) but otherwise I sort of agree with him

Updated by nobu (Nobuyoshi Nakada) over 1 year ago

  • Description updated (diff)

Why not blocking until another block terminates?

Updated by headius (Charles Nutter) over 1 year ago

nobu (Nobuyoshi Nakada) I actually considered that my first preference, except for the introduction of a new global lock.

I worry about suddenly introducing deadlocks into someone's code, though that code would have to be doing some locking as well as a concurrent chdir somewhere (both locks need to be contended). Perhaps that code deserves what it got, but then again maybe it's better to just have a hard error anyway. I'm willing to discuss it more.

Updated by matz (Yukihiro Matsumoto) 1 day ago

It's hard to estimate how big the incompatibility is. But I'd like to try it.
I hope we don't see any big problem.


Also available in: Atom PDF