[ruby-core:70923] [Ruby trunk - Bug #10686] Memory leaking from torture test of symbol GC

Issue #10686 has been updated by Rei Odaira.

File test_symbol.patch added

I would like to propose two changes in `test/ruby/test_symbol.rb (TestSymbol#test_symbol_fstr_leak)`.

First, this test checks only the increase in the virtual size, but it should check the rss as well. On Mac OS X, the memory leak of the symbols increases the rss, not the virtual size. The following example demonstrates the increase in the rss when the fix in symbol.c is not applied. Therefore, `test_symbol_fstr_leak` should specify "rss: true" as an argument to `assert_no_memory_leak()`.

~~~
$ ruby --disable=gems -I test/lib -r memory_status -e 'p Memory::Status.new; 200_000.times { |i| i.to_s.to_sym }; GC.start; p Memory::Status.new'
#<struct Memory::Status size=2505801728, rss=4321280>
#<struct Memory::Status size=2563211264, rss=47579136>
~~~

The second change I am proposing is to add a warm-up phase before measuring the actual memory size increase. On a certain AIX machine, `test_symbol_fstr_leak` causes 2.6x increase in the virtual size, even after applying the fix in symbol.c.

~~~
$ ruby --disable=gems -I test/lib -r memory_status -e 'p Memory::Status.new; 200_000.times { |i| i.to_s.to_sym }; GC.start; p Memory::Status.new'
#<struct Memory::Status size=1646592, rss=4091904>
#<struct Memory::Status size=4263936, rss=6725632>
~~~

However, I don't think this is a leak, because if you sequentially run a similar test twice and only measure the second run, neither the virtual size or the rss increases so much, as follows.

~~~
$ ruby --disable=gems -I test/lib -r memory_status -e '200_000.times { |i| i.to_s.to_sym }; GC.start; p Memory::Status.new; 200_000.times { |i| (i+200_000).to_s.to_sym }; GC.start; p Memory::Status.new'
#<struct Memory::Status size=4276224, rss=6758400>
#<struct Memory::Status size=4321280, rss=6803456>
~~~

Because the warm-up phase inevitably increases memory usage, we should measure the steady state after the warm-up phase when detecting a memory leak.

The attached patch includes these two changes. I confirmed my patch on Mac OS X, POWER Linux, and AIX.  Without the fix in symbol.c, `test_symbol_fstr_leak` correctly fails on every environment, and with the fix in symbol.c, it succeeds.

----------------------------------------
Bug #10686: Memory leaking from torture test of symbol GC
https://bugs.ruby-lang.org/issues/10686#change-54284

* Author: Charles Nutter
* Status: Closed
* Priority: Normal
* Assignee: 
* ruby -v: 2.2p0
* Backport: 2.0.0: DONTNEED, 2.1: DONTNEED, 2.2: DONE
----------------------------------------
The following code appears to grow without bounds when running on MRI 2.2p0 (and grows *very* fast...hold on to your RAM):

```ruby
x = 0; loop { (x += 1).to_s.to_sym }
```

I asked ko1 about this on Twitter and he said it appears to be leaking strings somewhere.

---Files--------------------------------
test_symbol.patch (610 Bytes)


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