Skip to content

HTTPS clone URL

Subversion checkout URL

You can clone with HTTPS or Subversion.

Download ZIP
Commits on Oct 5, 2012
  1. Dirkjan Bussink

    Make sure that we setup a barrier before installing the jitted code

    dbussink authored
    Also make sure we remove the setup call frame at the proper moment.
  2. Dirkjan Bussink
  3. Dirkjan Bussink
  4. Dirkjan Bussink
  5. Dirkjan Bussink

    Fix race condition when concurrently reading and modifying header

    dbussink authored
    This fixes a potential number of race conditions. The most problematic
    one is the following. If we are in the middle of inflating a header and
    concurrently reading another flag there's a potential problem. If we
    check the header status first and it's not inflated but then read the
    flags after it's inflated, the flags contain bogus information.
    
    We fix this by making sure we always first copy the header, then check
    inflation status and return this copy if it is not inflated. This means
    that the code can see slightly outdated information, but it can already
    handle that properly because this type of race could happen in other
    places too.
    
    We also change every modification of changing the flags to be atomic.
    This is necessary to prevent certain information from being lost. If we
    for example freeze an object while concurrently inflating the header, it
    could be that we set the frozen bit in the header after we already
    copied the flags into the inflated header. In that case setting the
    frozen status would be lost.
    
    This is not a huge problem, but the side effect is more problematic.
    Since we reuse the header space as the pointer to the inflated header,
    modifying a bit there would mean the inflated header would point at a
    totally different inflated header. This would mean the object would
    suddenly get a very different status.
    
    These changes are related to issue #1940. In this issue the behavior
    exposed itself as follows. The code resulted in a concurrent scenario
    where an object header was inflated while concurrently retrieving an
    instance variable. Because of this, the returned obj_type was the value
    of the bits inside the inflated header pointer:
    
    $2 = {
      <rubinius::ObjectHeader> = {
        header = {
          f = {
            inflated = 1,
            meaning = 0,
            obj_type = 217,
            zone = rubinius::UnspecifiedZone,
            age = 2,
            Forwarded = 1,
            Remember = 1,
            Marked = 1,
            InImmix = 0,
            Pinned = 0,
            Frozen = 0,
            Tainted = 1,
            Untrusted = 0,
            LockContended = 0,
            unused = 0,
            aux_word = 1
          },
          flags64 = 4312680137,
          all_flags = 0x1010e46c9
        },
        klass_ = 0x104a3ff90,
        ivars_ = 0x1a,
        __body__ = {0x0}
      }, <No data fields>}
    
    The value here is obj_type 217, which is actually an invalid value. What
    happens because of this is that the code in Object#get_ivar would see
    that obj_type. This resulted in it not recognizing it as a packed object
    and this not running the proper code to retrieve the instance variable.
    
    This resulted in a nil being returned as the instance variable instead
    of the actual instance variable and then resulted in a NoMethodError.
  6. Dirkjan Bussink
  7. Dirkjan Bussink
Commits on Oct 4, 2012
  1. Dirkjan Bussink

    Store the thread name in the managed thread

    dbussink authored
    This makes debugging somewhat easier when we deadlock because we can see
    which thread is actually problematic if we call methods on other running
    threads such as getting backtraces.
  2. Dirkjan Bussink

    Fix deadlock in starting a Thread

    dbussink authored
    When we create a new Thread object, it is allocated with a the
    initializion lock locked. This means that any other method can needs to
    grab this lock has to wait for the thread to be actually started.
    
    If a method such as getting a backtrace of the thread is called before
    the thread is scheduled to be started, it can deadlock because it can't
    grab the lock and if there is simultanous GC triggered.
  3. Dirkjan Bussink
Commits on Oct 3, 2012
  1. Brian Shirai
  2. Brian Shirai

    Cleaned up at_exit spec.

    brixen authored
  3. Brian Shirai
  4. Brian Shirai

    Merge pull request #1898 from voxik/ignore-block-device-test

    brixen authored
    Ignore block device test.
  5. Brian Shirai
  6. Brian Shirai
  7. Brian Shirai

    Merge pull request #1936 from ryoqun/raise-nil-false

    brixen authored
    Correctly raise TypeError even if nil or false is given
  8. Brian Shirai

    Merge pull request #1937 from voxik/install-manpages

    brixen authored
    Install man pages.
  9. Brian Shirai
  10. Brian Shirai
  11. Ryo Onodera
  12. Ryo Onodera
  13. Brian Shirai

    Changed configure --libdir to --appdir. Closes #1919.

    brixen authored
    The --libdir option is retained for specifying where to put the Rubinius
    shared library. We don't build a shared or static library yet but we will
    when we have an embedding API.
    
    The --appdir option specifies where the application runtime and library
    files will be installed. The use of --libdir for this was misleading.
  14. Brian Shirai
  15. Brian Shirai
  16. Brian Shirai
  17. Brian Shirai

    Silence iconv warning.

    brixen authored
Commits on Oct 2, 2012
  1. Dirkjan Bussink

    Trigger a wakeup after setting up the listening socket for the agent

    dbussink authored
    This fixes the problem on OS X where opening an agent console would just
    hang. Related to #1933.
  2. Install man pages.

    authored
  3. Dirkjan Bussink

    Teach the JIT about fixnum_mul

    dbussink authored
    We have to work around a bug in LLVM here which makes smul.with.overflow
    not work correctly. It will not properly detect overflows if we use the
    i63 / i31 integer types as described here:
    
    http://llvm.org/bugs/show_bug.cgi?id=13991
    
    We work around it by do multiplication with overflow detection for i64
    and i32 which does work and manually do a bounds check for Fixnum::MIN
    and Fixnum::MAX.
  4. Dirkjan Bussink
  5. Dirkjan Bussink

    Use better descriptive name for comparison operations

    dbussink authored
    float_compare and fixnum_compare are actually the comparators for the
    Float#<=> and Fixnum#<=> behavior, so this name makes it less ambiguous
    what these wrappers are for.
Commits on Oct 1, 2012
  1. Dirkjan Bussink
  2. Dirkjan Bussink

    Add missing file

    dbussink authored
  3. Dirkjan Bussink
Something went wrong with that request. Please try again.