Bug with "special exceptions" when they are thrown in context of a rescue clause.
I've stumbled upon a case when ruby is supposed to throw "IOError: stream closed"(https://github.com/ruby/ruby/blob/trunk/thread.c#L4823) because there was a retained FD lock by another thread, but I'm was getting this instead of it:
RuntimeError: can't modify frozen IOError /home/vagrant/.rbenv/versions/2.3.1/lib/ruby/gems/2.3.0/gems/puma-3.6.2/lib/puma/server.rb:877:in `write' /home/vagrant/.rbenv/versions/2.3.1/lib/ruby/gems/2.3.0/gems/puma-3.6.2/lib/puma/server.rb:877:in `<<' /home/vagrant/.rbenv/versions/2.3.1/lib/ruby/gems/2.3.0/gems/puma-3.6.2/lib/puma/server.rb:877:in `stop'
I've done some digging and it appeared to a be a ruby bug with how such exceptions(so called "special exceptions") are handled. This exception is being frozen right after creation: https://github.com/ruby/ruby/blob/trunk/vm.c#L2078 but later, when it's thrown, it's being handled exactly as regular exception that is not frozen, which leads to a problem here:
https://github.com/ruby/ruby/blob/trunk/eval.c#L511 - it is an attempt to assign the "#cause" attribute on it when it all happens inside a rescue clause, since the exception itself is frozen.
I've created a script to reproduce it:
rd, wr = IO.pipe Thread.new do IO.select [rd] wr.close end begin raise 'any exception' rescue wr << 'A' end
It works with this ruby fork where I've added sleep for couple of seconds to imitate slow system call response, to keep the FD locked for a while and produce initial exception: https://github.com/ruby/ruby/blob/trunk/eval.c#L511
Updated by shyouhei (Shyouhei Urabe) about 5 years ago
- Backport changed from 2.2: UNKNOWN, 2.3: UNKNOWN, 2.4: UNKNOWN to 2.2: UNKNOWN, 2.3: REQUIRED, 2.4: REQUIRED
- Status changed from Feedback to Closed
matthewd (Matthew Draper) wrote:
Should this be closed for backporting?
Updated by MSP-Greg (Greg L) over 4 years ago
Obviously, 2.2 changes are soon to be stopped.
Recently in Puma, Travis was changed from 2.2.8 to 2.2.9, and in code using a pipe, the "can't modify frozen IOError" occurred.
This was not an issue with 2.2.8, which was used in Puma / Travis testing since 2017-09-18.
So, the question is, might this be backported to 2.2?