explanation: we use pthread_once to define a mutex type descriptor,
used to define some of our mutexes as recursive mutexes. Now,
pthread_once relies on a counter stored in a given location;
we used a non-exported global var for this counter.
Unfortunately this ties the mutex initialisation to the static
initialisation of the compilation unit holding this counter variable.
Theoretically it would be possible (we never observed such an incident)
that, during static initialisation, a singleton was brought up,
which requires a class-scoped lock, implemented as recursive mutex.
And it would be possible for this singleton locking to happen prior
to initialisation of the mentioned counter variable.
As a fix, I've moved the counter varialbe into a function scoped
static variable, since that is guaranteed by the C++ runtime system
to be initialised at first usage of the function, irrespective of the
initialisation order of the enclosing compilation units
Note: this drops some backwards compatibility. We're targeting now
roughly the range between Ubuntu-Precise (LTS) and Debian/testing,
with Debian/stable as the reference system.
The naming scheme for Boost-Libraries was adjusted with Boost-1.42
for Unix-Platforms. Now the '-mt' suffix isn't included any more, but
the libraries available through the usual packaging mechanisms can be
assumed to be thread safe.
See also http://issues.lumiera.org/ticket/759
this is really creepy: the same(!) instance of the singleton factory
sees different addresses of the class static variable, depending on
the compilation unit.
Please note that the type of the concrete factory function is *erased*
when exiting the constructor function of ConfigurableHolder
basically this reproduces the problem in a simplified setup.
Especially note that we're going through a single instance of the factory,
yet still this single instance 'sees' two different locations of the
class static variable
...but still dynamically linking against the core lib
But the actual template instantiations happen now within the two
compilation units, which are linked statically.
When looking into the symbol table, we can see that the static
field is emitted two times
readelf -W target/clang-static-init -s | c++filt |less
Observations:
- the initial observation was that we get two instances of the config rules service
- obviously this it is *not* the initialisation of a static variable accessed from
multiple compilation units. But the access from two compilation units is crucial
- we can exclude the effect of all other initialisation. It *is* in SingletonSubclass
- we can exclude the effect of dynamic linking. Even two translation units
linked statically exhibit the same problem
rebuild this test case in the research area, to be able to verify with various compilers
clang-3.2 requires a clarification here (while previous versions
of clang and GCC automatically resolved the ambiguity by assuming
use of a nested, dependent template).
Clang seems to evaluate the terms of a function call in another order
than GCC -- this uncovered re-entrance errors in some metaprogramming tests,
where we re-used a global formatter object in recursive instantiations.
Clang is more insistent when it comes to enforcing 'protected' visibility.
Since in this case the basic design can be considered sane and optimal, the
only (and obvious) solution is to nest the PIMPL into a default base class
for implementation; this mirrors the structure of the interface.
Compilation with Clang 3.0 (which is available in Debian/stable) fails,
mostly due to some scoping and naming inconsistencies which weren't detected
by GCC. At some instances, Clang seems to have problems to figure out a
perfectly valid type definition; these can be resolved by more explicit
typing (which is preferrable anyway)
by providing a custom copy function one can adjust otherwise non-copyable
elements. This should be used cautionary because dereferencing elements may
poison the cache and thus have some considerable performance impact
(profile this)
using our util::_Fmt front-end helps to reduce the code size,
since all usages rely on a single inclusion of boost::format
including boost::format via header can cause quite some code bloat
NOTE: partial solution, still some further includes to reorganise
initial draft of an RfC to discuss and define the
requirements for other parts of the application to relie on
note: this commit fixes a merge error; the RfC was lost
while combining documentation and code branches
