Farid Zaripov wrote:
The __rw_get_moneypunct () function in punct.cpp is not thread safe
for the same
facets. If this function is invoked with the same pfacet parameter in
dirrerent threads
simultaneously, then some one thread successfully create __rw_setlocale
object
but other threads will blocked by global mutex in __rw_setlocale ctor.
Then when
the mutex will be released after the updating pfacet->_C_data(), the
next thread will
do the same initialization as the first thread.
I see what you're saying: the other thread will end up overwriting
the data written by the first thread. The newly data should be the
same, so I wonder if this just a benign race condition or if it's
more sinister than that?
Either way, I agree that it should be fixed. I took the liberty
to add more detail to your comment in the patch below to make
it even more clear what's going on (I had some difficulty
following the logic at first so I think it might be helpful to
others).
ChangeLog:
Can you reference the Jira bug here please?
* punct.cpp (__rw_get_moneypunct): After creating __rw_setlocale
object check
if the locale data is initialized by another thread to prevent
multiple initialization\
of the locale data in facet.
----------
Index: punct.cpp
===================================================================
--- punct.cpp (revision 580097)
+++ punct.cpp (working copy)
@@ -308,6 +308,10 @@
// set all categories -- need LC_NUMERIC and LC_CTYPE
const __rw_setlocale clocale (locname, _RWSTD_LC_ALL);
+ if (pfacet->_C_data ())
+ // call self recursively on already initialized in another
thread `impdata'
+ return __rw_get_moneypunct (pfacet, flags);
+
I suggest:
+ if (pfacet->_C_data ()) {
+ // check to see if another thread may have set _C_data()
+ // while we were waiting for the lock in __rw_setlocale
+ // ctor above and, if so, call self recursively on the
+ // already initialized `impdata'
+ return __rw_get_moneypunct (pfacet, flags);
+ }
Thanks
Martin
