SEGV calling ca_context_create from file scope constructor

[Jeff Hill]

From Margaret Votava,

We are trying to understand a problem with a DESY/DOOCS utility. It generates a segmentation fault within ca_task_initialize (stack trace below). The

epicsMutex ca_client_context::defaultServiceInstallMutex;

constructor is not being called in all OS platform/Compiler combinations.

In particular, a failing combination is
Linux 2.6.8-22.0.2/gcc v3.4.3 (red hat)
Linux 2.6.1/gcc v3.2.3 (red hat)

but works in Solaris with Solaris compiler.

We believe that this is a linker problem.

Has anyone seen before or have any clues of how to fix?

Thank you kindly,
Margaret

### epics code ###

(gdb) where
#0 epicsMutexLock (pmutexNode=0x0)
    at ../../../src/libCom/osi/epicsMutex.cpp:118
#1 0x006add15 in epicsMutex::lock (this=0x0)
    at ../../../src/libCom/osi/epicsMutex.cpp:211
#2 0x00ca19cb in ca_client_context (this=0x8d8ce08,
    enablePreemptiveCallback=false) at ../../../include/epicsGuard.h:68 #3 0x00c88305 in ca_context_create (
    premptiveCallbackSelect=ca_disable_preemptive_callback)
    at ../access.cpp:204
#4 0x00c8835e in ca_task_initialize () at ../access.cpp:180

### doocs code ####

#5 0x00a3e9f5 in epics_init ()
    at
/home/votava/smtf/doocs/libs/clientlib/source/clientlib/eq_svr.cc:1928
#6 0x00a464b5 in EqCall (this=0x8d8cc88)
    at
/home/votava/smtf/doocs/libs/clientlib/source/clientlib/eq_client.cc:369
#7 0x00e05837 in __static_initialization_and_destruction_0
(__initialize_p=1,
    __priority=65535)
    at
/home/votava/smtf/doocs/libs/dddlib/source/dddlib/panels/device_attr_stubs.cc:26
#8 0x00e05917 in global constructors keyed to ea ()
    at
/home/votava/smtf/doocs/libs/dddlib/source/dddlib/panels/device_attr_stubs.cc:400
#9 0x00e36819 in __do_global_ctors_aux ()
   from
/home/votava/smtf/doocs/ups/../ups/../libs/dddlib/ups/../ups/../Linux/lib/libddd.so.2.2.5
#10 0x00d45add in _init ()
   from
/home/votava/smtf/doocs/ups/../ups/../libs/dddlib/ups/../ups/../Linux/lib/libddd.so.2.2.5
#11 0x00271318 in _dl_init_internal () from /lib/ld-linux.so.2 #12 0x002657ff in _dl_start_user () from /lib/ld-linux.so.2

Additional information:

The issue you have discovered occurs only when calling ca_context_create (a.k.a. ca_task_initialize) from your C++ file scope object's constructor.

The static variable "ca_client_context::defaultServiceInstallMutex" is also file scope, and this implies that it's constructor along with the constructor of your device_attr_stubs both run before main() is called.

Which file scope constructor runs first is very C++ compiler implementation dependent.

Let me know if the attached patch resolves your troubles.

Jeff

cvs diff -u -wb -i -- ca_client_context.cpp oldAccess.h (in directory D:\users\hill\R3.14.dll_hell_fix\epics\base\src\ca\)
Index: ca_client_context.cpp
===================================================================
RCS file: /net/phoebus/epicsmgr/cvsroot/epics/base/src/ca/ca_client_context.cpp,v
retrieving revision 1.9.2.18
diff -c -u -w -b -i -r1.9.2.18 ca_client_context.cpp
cvs diff: conflicting specifications of output style
--- ca_client_context.cpp 13 Apr 2005 17:28:13 -0000 1.9.2.18
+++ ca_client_context.cpp 17 Feb 2006 22:22:11 -0000
@@ -47,6 +47,7 @@
 {
     epicsThreadPrivateDelete ( caClientCallbackThreadId );
     caClientCallbackThreadId = 0;
+ delete ca_client_context::pDefaultServiceInstallMutex;
 }

 // runs once only for each process
@@ -54,13 +55,14 @@
 {
     caClientCallbackThreadId = epicsThreadPrivateCreate ();
     assert ( caClientCallbackThreadId );
+ ca_client_context::pDefaultServiceInstallMutex = new epicsMutex;
     epicsAtExit ( cacExitHandler,0 );
 }

 extern epicsThreadPrivateId caClientContextId;

 cacService * ca_client_context::pDefaultService = 0;
-epicsMutex ca_client_context::defaultServiceInstallMutex;
+epicsMutex * ca_client_context::pDefaultServiceInstallMutex;

 ca_client_context::ca_client_context ( bool enablePreemptiveCallback ) :
     ca_exception_func ( 0 ), ca_exception_arg ( 0 ),
@@ -77,9 +79,7 @@

     epicsThreadOnce ( & cacOnce, cacOnceFunc, 0 );

- {
- // this wont consistently work if called from file scope constructor
- epicsGuard < epicsMutex > guard ( ca_client_context::defaultServiceInstallMutex );
+ epicsGuard < epicsMutex > guard ( *ca_client_context::pDefaultServiceInstallMutex );
         if ( ca_client_context::pDefaultService ) {
             this->pServiceContext.reset (
                 & ca_client_context::pDefaultService->contextCreate (
@@ -88,7 +88,6 @@
         else {
             this->pServiceContext.reset ( new cac ( this->mutex, this->cbMutex, *this ) );
         }
- }

     this->sock = epicsSocketCreate ( AF_INET, SOCK_DGRAM, IPPROTO_UDP );
     if ( this->sock == INVALID_SOCKET ) {
@@ -734,8 +733,9 @@

 void ca_client_context::installDefaultService ( cacService & service )
 {
- // this wont consistently work if called from file scope constructor
- epicsGuard < epicsMutex > guard ( ca_client_context::defaultServiceInstallMutex );
+ epicsThreadOnce ( & cacOnce, cacOnceFunc, 0 );
+
+ epicsGuard < epicsMutex > guard ( *ca_client_context::pDefaultServiceInstallMutex );
     if ( ca_client_context::pDefaultService ) {
         throw std::logic_error
             ( "CA in-memory service already installed and can't be replaced");
Index: oldAccess.h
===================================================================
RCS file: /net/phoebus/epicsmgr/cvsroot/epics/base/src/ca/oldAccess.h,v
retrieving revision 1.47.2.14
diff -c -u -w -b -i -r1.47.2.14 oldAccess.h
cvs diff: conflicting specifications of output style
--- oldAccess.h 19 Oct 2004 20:48:51 -0000 1.47.2.14
+++ oldAccess.h 17 Feb 2006 22:07:21 -0000
@@ -286,6 +286,9 @@
     void operator delete ( void * );
 };

+extern "C" void cacOnceFunc ( void * );
+extern "C" void cacExitHandler ( void *);
+
 struct ca_client_context : public cacContextNotify
 {
 public:
@@ -412,8 +415,10 @@
  ca_client_context ( const ca_client_context & );
  ca_client_context & operator = ( const ca_client_context & );

+ friend void cacOnceFunc ( void * );
+ friend void cacExitHandler ( void *);
     static cacService * pDefaultService;
- static epicsMutex defaultServiceInstallMutex;
+ static epicsMutex * pDefaultServiceInstallMutex;
 };

 int fetchClientContext ( ca_client_context * * ppcac );

Original Mantis Bug: mantis-239
    http://www.aps.anl.gov/epics/mantis/view_bug_page.php?f_id=239

[Jeff Hill (johill-lanl)]

I susequently installed this patch to reduce the scope of the previously mentioned patch.

Saved settings for D:\users\hill\R3.14.dll_hell_fix\epics\base successfully...
cvs diff -u -wb -i -- ca_client_context.cpp (in directory D:\users\hill\R3.14.dll_hell_fix\epics\base\src\ca\)
Index: ca_client_context.cpp
===================================================================
RCS file: /net/phoebus/epicsmgr/cvsroot/epics/base/src/ca/ca_client_context.cpp,v
retrieving revision 1.9.2.19
diff -c -u -w -b -i -r1.9.2.19 ca_client_context.cpp
cvs diff: conflicting specifications of output style
--- ca_client_context.cpp 17 Feb 2006 22:31:42 -0000 1.9.2.19
+++ ca_client_context.cpp 17 Feb 2006 22:38:36 -0000
@@ -78,7 +78,7 @@
     }

     epicsThreadOnce ( & cacOnce, cacOnceFunc, 0 );
-
+ {
     epicsGuard < epicsMutex > guard ( *ca_client_context::pDefaultServiceInstallMutex );
     if ( ca_client_context::pDefaultService ) {
         this->pServiceContext.reset (
@@ -88,6 +88,7 @@
     else {
         this->pServiceContext.reset ( new cac ( this->mutex, this->cbMutex, *this ) );
     }
+ }

     this->sock = epicsSocketCreate ( AF_INET, SOCK_DGRAM, IPPROTO_UDP );
     if ( this->sock == INVALID_SOCKET ) {

***** CVS exited normally with code 1 *****

[Andrew Johnson (anj)]

Patch included in R3.14.9.

[Andrew Johnson (anj)]

R3.14.9 Released.