* cygtls.h (_threadinfo::call): Remove regparm declaration to work around

compiler bug.
* autoload.cc (TryEnterCriticalSection): Remove.
* dcrt0.cc (dll_crt0_0): Delete inappropriate setting of _my_tls.stackptr to
NULL since it has really bad consequences.  Make 'si' an automatic variable.
* cygtls.cc (_threadinfo::init_thread): Correct thinko which caused thread list
to be allocated every time.
* cygtls.h (CYGTLS_PADSIZE): Define as const int.
* sync.h: Make multiple inclusion safe.
(muto::next): Eliminate.
(muto::exiting_thread): New variable.
(muto::set_exiting_thread): New function.
(new_muto): Change to use different section for mutos since c++ give
inexplicable warning in some cases otherwise.
(new_muto1): Ditto.
* dcrt0.cc (do_exit): Call muto::set_exiting_thread here.
* sync.cc (muto_start): Eliminate.
(muto::acquire): Always give exiting thread a lock.  Never give thread a lock
if exiting.
(muto::release): Ditto for releasing.
* dtable.cc (dtable::init_lock): Unline function and define here.
* dtable.h (lock_cs): Define as a muto since critical sections seem to work
oddly on Windows Me.
(lock): Accommodate switch to muto.
(unlock): Ditto.
* exceptions.cc (setup_handler): Don't worry about acquiring mutos since that
hasn't mattered for a long time.
(signal_exit): Ditto: muto stuff will be handled automatically on exit now.
* Makefile.in (DLL_IMPORTS): Link advapi32 to ensure proper DLL initialization.
* autoload.cc (RegCloseKey): Arbitrarily choose this function as a "seed" to
pull the advapi32 link library in.  So, comment out the autoloading.
* cygtls.cc (_threadinfo::init_thread): Just clear CYGTLS_PADSIZE.
(_threadinfo::remove): Add debugging.
(_threadinfo::find_tls): Ditto.
* cygtls.h (_threadinfo::padding): Make zero length (for now?).
* dcrt0.cc (dll_crt0_0): Move more initialization here from dll_crt0_1.
(dll_crt0_1): See above.
* dtable.h (dtable::lock): Remove commented out critical section locking.
* dtable.h (dtable::init_lock): Remove commented out critical section locking.
* dtable.h (dtable::unlock): Remove commented out critical section locking.
* exceptions.cc (interruptible): bool'ize.
* init.cc (threadfunc_fe): Revert to storing threadfunc at stack bottom.
(munge_threadfunc): Ditto.  Avoid adding overhead to calibration_thread.
(prime_threads): Don't initialize tls stuff.
(dll_entry): Make minor change to initialization order.
* tlsoffsets.h: Regenerate.
* sigproc.cc (wait_sig): Add sanity check for end of process thread exit.
* select.h: Make minor formatting change.
* Makefile.in: Add still more -fomit-frame-pointer functions.
* dtable.h (dtable::lock): New function.
(dtable::unlock): New function.
(dtable::init_lock): New function.
* cygheap.h (HEAP_TLS): Declare new enum value.
(init_cygheap::threadlist): Declare new array.
(init_cygheap::sthreads): Declare new variable.
(cygheap_fdmanip::~cygheap_fdmanip): Use new dtable lock/unlock functions.
(cygheap_fdnew::cygheap_fdnew): Ditto.
(cygheap_fdget::cygheap_fdget): Ditto.
* dtable.cc (dtable_init): Initialize fdtab critical section.
(dtable::fixup_after_fork): Ditto.
(dtable::fixup_after_exec): Ditto.
(dtable::dup2): Use lock/unlock calls to protect access to fdtab.
(dtable::find_fifo): Ditto.
(dtable::fixup_before_fork): Ditto.
(dtable::fixup_before_exec): Ditto.
(dtable::set_file_pointers_for_exec): Ditto.
(dtable::vfork_child_dup): Ditto.
(dtable::vfork_parent_restore): Ditto.
* syscalls.cc (close_all_files): Ditto.
* sync.h (muto::acquired): Declare new function.
(new_muto1): Declare new macro used to specify name of muto storage.
* sync.cc (muto::acquired): Define new function.
* cygthread.cc (cygthread::stub): Remove signal chain removal call since it is
handled during initialization now.
* cygthread.cc (cygthread::simplestub): Remove signal chain removal call since
it is handled during initialization now.
* cygtls.cc (sentry): New class used for locking.  Use throughout.
(_threadinfo::reset_exception): Don't pop stack.
(_threadinfo::find_tls): Move from exceptions.cc.
(_threadinfo::init_thread): Initialize array of threads rather than linked
list.  Take second argument indicating thread function for this thread.
(_threadinfo::remove): Search thread array rather than linked list.  Use sentry
to lock.  Only unlock if we got the lock.
(_threadinfo::find_tls): Ditto for first two.
(handle_threadlist_exception): Handle exceptions when manipulating the thread
list in case of premature thread termination.
(_threadinfo::init_threadlist_exceptions): Ditto.
* cygtls.h (TLS_STACK_SIZE): Decrease size.
(_threadinfo::padding): Add element to avoid overwriting lower part of stack.
(_threadinfo::remove): Add a "wait" argument to control how long we wait for a
lock before removing.
* exceptions.cc (init_exception_handler): Make global.  Take argument to
control exception handler being set.
(ctrl_c_handler): Wait forever when removing self from signal chain.
(_threadinfo::find_tls): Move to cygtls.cc.
(sig_handle): Reorganize detection for thread-specific signals.
* heap.cc (heap_init): Rework slightly.  Make fatal error more verbose.  Remove
malloc initialization since it can't happen during dll attach.
* init.cc (search_for): Move address to search for on stack here.
(threadfunc_ix): Ditto for stack offset.  Make shared so that stack walk
potentially only has to be done once when cygwin processes are running.
(threadfunc_fe): Use standard tls to store thread function (may change back
later).
(calibration_thread): New function.  Potentially called to find threadfunc_ix.
(munge_threadfunc): Search for "search_for" value on stack.  Output warning
when thread func not found on stack.  Use standard tls to store thread
function.
(prime_threads): New function.  Called to prime thread front end.
(dll_entry): Call dll_crt0_0 here when DLL_PROCESS_ATTACH.  Call prime_threads
here.  Try to remove thread from signal list here.
* sigproc.cc (wait_sig): Initialize threadlist exception stuff here.
* thread.cc (pthread::exit): Pass argument to signal list remove function.
* thread.h: Remove obsolete *ResourceLock defines.
* tlsoffsets.h: Regenerate.
* winsup.h (spf): Define temporary debug macro to be deleted later.
* dcrt0.cc (dll_crt0_0): New function, called during DLL initialization.
Mainly consists of code pulled from dll_crt0_1.
(dll_crt0_1): See above.
(_dll_crt0): Wait for initial calibration thread to complete, if appropriate.
Move some stuff to dll_crt0_0.
(initialize_main_tls): Accommodate argument change to
_thread_info::init_thread.
* fork.cc (fork_child): Ditto.
(sync_with_child): Fix debug message.
* external.cc (cygwin_internal): Remove special considerations for
uninitialized dll since initialization happens during dll attach now.
* dlfcn.cc (dlopen): Remove obsolete *ResourceLock calls.
(dlclose): Ditto.
* cygheap.h (init_cygheap::close_ctty): Declare new function.
* cygheap.cc (init_cygheap::close_ctty): Define new function.
* syscalls.cc (close_all_files): Use close_ctty.
(setsid): Ditto.
* cygthread.cc (cygthread::stub): Remove exception initialization.
* cygthread.cc (cygthread::stub): Remove exception initialization.
(cygthread::simplestub): Ditto.
* thread.cc (pthread::thread_init_wrapper): Ditto.
* cygtls.cc (_last_thread): Make static.
(_threadinfo::call2): Initialize exception handler here.
(_threadinfo::find_tls): Move here.
* exceptions.cc (_threadinfo::find_tls): Move.
* dcrt0.cc (__api_fatal): Add prefix info to message here rather than including
it in every call to function.
* winsup.h (api_fatal): Accommodate above change.
* debug.cc (add_handle): Don't do anything if cygheap not around.
(mark_closed): Ditto.
* dll_init.cc (dll_list::detach): Fix debug output.
* fork.cc (sync_with_child): Ditto.
(vfork): Improve debug output.
* heap.cc (heap_init): Ditto.
* exceptions.cc (try_to_debug): Clarify message when debugger attaches.

1 files changed, 51 insertions, 7 deletions

diff --git a/winsup/cygwin/init.cc b/winsup/cygwin/init.cc
index f6a29f2d6..a85a8a270 100644
--- a/winsup/cygwin/init.cc
+++ b/winsup/cygwin/init.cc
@@ -16,51 +16,95 @@ details. */
 #include "cygtls.h"
 
 int NO_COPY dynamically_loaded;
+static char *search_for = (char *) cygthread::stub;
+static unsigned threadfunc_ix __attribute__((section ("cygwin_dll_common"), shared)) = 0;
+DWORD tls_func;
+
+HANDLE sync_startup;
+
+#define OLDFUNC_OFFSET -1
 
 static void WINAPI
 threadfunc_fe (VOID *arg)
 {
-  _threadinfo::call ((DWORD (*)  (void *, void *)) (((char **) _tlsbase)[-1]), arg);
+  _threadinfo::call ((DWORD (*)  (void *, void *)) (((char **) _tlsbase)[OLDFUNC_OFFSET]), arg);
+  // void *threadfunc = (void *) TlsGetValue (tls_func);
+  // TlsFree (tls_func);
+  // _threadinfo::call ((DWORD (*)  (void *, void *)) (threadfunc), arg);
+}
+
+static DWORD WINAPI
+calibration_thread (VOID *arg)
+{
+  ExitThread (0);
 }
 
 static void
 munge_threadfunc (HANDLE cygwin_hmodule)
 {
-  char **ebp = (char **) __builtin_frame_address (0); 
-  static unsigned threadfunc_ix;
+  char **ebp = (char **) __builtin_frame_address (0);
   if (!threadfunc_ix)
     {
       for (char **peb = ebp; peb < (char **) _tlsbase; peb++)
-	if (*peb == (char *) cygthread::stub)
+	if (*peb == search_for)
 	  {
 	    threadfunc_ix = peb - ebp;
 	    goto foundit;
 	  }
+#ifdef DEBUGGING
+      system_printf ("non-fatal warning: unknown thread! search_for %p, cygthread::stub %p, calibration_thread %p, possible func offset %p",
+		     search_for, cygthread::stub, calibration_thread, ebp[137]);
+#endif
+      try_to_debug ();
       return;
     }
 
 foundit:
   char *threadfunc = ebp[threadfunc_ix];
-  ebp[threadfunc_ix] = (char *) threadfunc_fe;
-  ((char **) _tlsbase)[-1] = threadfunc;
+  if (threadfunc == (char *) calibration_thread)
+    /* no need for the overhead */;
+  else
+    {
+      ebp[threadfunc_ix] = (char *) threadfunc_fe;
+      ((char **) _tlsbase)[OLDFUNC_OFFSET] = threadfunc;
+      // TlsSetValue (tls_func, (void *) threadfunc);
+    }
 }
 
+void
+prime_threads ()
+{
+  // tls_func = TlsAlloc ();
+  if (!threadfunc_ix)
+    {
+      DWORD id;
+      search_for = (char *) calibration_thread;
+      sync_startup = CreateThread (NULL, 0, calibration_thread, 0, 0, &id);
+    }
+}
+
+extern void __stdcall dll_crt0_0 ();
+
 extern "C" int WINAPI
 dll_entry (HANDLE h, DWORD reason, void *static_load)
 {
   switch (reason)
     {
     case DLL_PROCESS_ATTACH:
+      prime_threads ();
       dynamically_loaded = (static_load == NULL);
       // __cygwin_user_data.impure_ptr = &_my_tls.local_clib;
-      _my_tls.stackptr = NULL;
+      dll_crt0_0 ();
+      // small_printf ("%u, %p, %p\n", cygwin_pid (GetCurrentProcessId ()), _tlstop, _tlsbase);
       break;
     case DLL_PROCESS_DETACH:
       break;
     case DLL_THREAD_ATTACH:
       munge_threadfunc (h);
+      // small_printf ("%u, %p, %p\n", cygwin_pid (GetCurrentProcessId ()), _tlstop, _tlsbase);
       break;
     case DLL_THREAD_DETACH:
+      _my_tls.remove (0);
       break;
     }
   return 1;