Project

General

Profile

Actions

Backport #5700

closed

fork {} segfaults during VM cleanup when run inside Fiber

Added by normalperson (Eric Wong) almost 13 years ago. Updated almost 13 years ago.

Status:
Closed
Assignee:
-
[ruby-core:41456]

Description

The issue is very easy to reproduce:

Fiber.new do
  p Process.waitpid2(fork {})
end.resume

Parent process successfully exits, child process segfaults while
it is exiting.

Backtrace is attached (gdb_bt.txt)

MALLOC_CHECK_=3 with GNU libc malloc() implementation detects an
attempt to free invalid pointer (attachment: malloc_check_3.txt)

I can also reproduce this with 1.9.2-p290 and 1.9.3-p0 as well as
latest trunk, so it has been around a while and a fix needs to be
backported.


Files

gdb_bt.txt (2.57 KB) gdb_bt.txt GDB backtrace normalperson (Eric Wong), 12/03/2011 10:37 AM
malloc_check_3.txt (6.12 KB) malloc_check_3.txt output with MALLOC_CHECK_=3 when run with GNU libc normalperson (Eric Wong), 12/03/2011 10:37 AM
fiber_fork.rb (54 Bytes) fiber_fork.rb script to reproduce error normalperson (Eric Wong), 12/03/2011 10:37 AM
bug5700.patch (2.06 KB) bug5700.patch nagachika (Tomoyuki Chikanaga), 02/10/2012 12:30 PM

Updated by normalperson (Eric Wong) almost 13 years ago

Eric Wong wrote:

http://redmine.ruby-lang.org/issues/5700

Hello, I would like to see this fixed for the next 1.9.3 release.

I'm pretty sure this is an easy fix for somebody already familiar with
the Fiber/continuation code (unfortunately I am not familiar with it).

Updated by nagachika (Tomoyuki Chikanaga) almost 13 years ago

Hmm, it don't seems easy problem for me... Apparently I'm not so familiar with fiber.

Anyway I've written a patch for this issue. I know it's somewhat ad-hoc, but it works.
ko1 and nobu how do you think?

Updated by normalperson (Eric Wong) almost 13 years ago

Tomoyuki Chikanaga wrote:

Anyway I've written a patch for this issue. I know it's somewhat ad-hoc, but it works.
ko1 and nobu how do you think?

Thanks for looking into this. This fix works for me, too :>

Updated by nobu (Nobuyoshi Nakada) almost 13 years ago

One particular case is OK, two cases would be still OK, but three cases are not exceptional already.
We'll need some standard way to tell if a given object can be discarded immediately or not, I guess.

Anyway, it would be enough for the time being.

Updated by nagachika (Tomoyuki Chikanaga) almost 13 years ago

I agree. In fact we don't have to protect all Fibers but only a root Fiber of main Thread.

But can I check-in previous patch for a temporary workaround? The next release of 1.9.3 is almost coming.

Actions #6

Updated by nagachika (Tomoyuki Chikanaga) almost 13 years ago

  • Tracker changed from Bug to Backport
  • Project changed from Ruby master to Backport193
  • Category deleted (core)
  • Target version deleted (2.0.0)

I've committed at r34629, r34630. Please backport them.

Actions #7

Updated by naruse (Yui NARUSE) almost 13 years ago

  • Status changed from Open to Closed
  • % Done changed from 0 to 100

This issue was solved with changeset r34637.
Eric, thank you for reporting this issue.
Your contribution to Ruby is greatly appreciated.
May Ruby be with you.


merge revision(s) 34629,34630:

* cont.c (rb_fiber_reset_root_local_storage): add a new function to
  restore rb_thread_t::local_storage.

* cont.c (rb_obj_is_fiber): add a new function to tell finalizer to
  prevent fibers from destroy.

* gc.c (rb_objspace_call_finalizer): don't sweep fibers at finalizing
  objspace.

* internal.h (rb_fiber_reset_root_local_storage, rb_obj_is_fiber):
  add prototypes.

* vm.c (ruby_vm_destruct): reset main thread's local_storage before
  free main thread. rb_thread_t::local_storage is replaced by fiber's
  local storage when forked from fiber, and it should be already freed
  when the fiber was destroyed.

* test/ruby/test_fiber.rb (test_fork_from_fiber): add test for fork
  from fiber.
  when the fiber was destroyed. [ruby-core:41456] [Bug #5700]
Actions

Also available in: Atom PDF

Like0
Like0Like0Like0Like0Like0Like0Like0