2001-06-25 Ross Johnson <rpj@setup1.ise.canberra.edu.au>

        * create.c (pthread_create): Add priority inheritance
        attributes.
        * mutex.c (pthread_mutex_lock): Remove some overhead for
        PTHREAD_MUTEX_NORMAL mutex types. Specifically, avoid
        calling pthread_self() and pthread_equal() to check/set
        the mutex owner. Introduce a new pseudo owner for this
        type. Test results suggest increases in speed of up to
        90% for non-blocking locks.
        This is the default type of mutex used internally by other
        synchronising objects, ie. condition variables and
        read-write locks. The test rwlock7.c shows about a
        30-35% speed increase over snapshot 2001-06-06. The
        price of this is that the application developer
        must ensure correct behaviour, or explicitly set the
        mutex to a safer type such as PTHREAD_MUTEX_ERRORCHECK.
        For example, PTHREAD_MUTEX_NORMAL (or PTHREAD_MUTEX_DEFAULT)
        type mutexes will not return an error if a thread which is not
        the owner calls pthread_mutex_unlock. The call will succeed
        in unlocking the mutex if it is currently locked, but a
        subsequent unlock by the true owner will then fail with EPERM.
        This is however consistent with some other implementations.
        (pthread_mutex_unlock): Likewise.
        (pthread_mutex_trylock): Likewise.
        (pthread_mutex_destroy): Likewise.
        * attr.c (pthread_attr_init): PTHREAD_EXPLICIT_SCHED is the
        default inheritance attribute; THREAD_PRIORITY_NORMAL is
        the default priority for new threads.
        * sched.c (pthread_attr_setschedpolicy): Added routine.
        (pthread_attr_getschedpolicy): Added routine.
        (pthread_attr_setinheritsched): Added routine.
        (pthread_attr_getinheritsched): Added routine.
        * pthread.h (sched_rr_set_interval): Added as a macro;
        returns -1 with errno set to ENOSYS.

2001-06-23  Ross Johnson  <rpj@setup1.ise.canberra.edu.au>

        *sched.c (pthread_attr_setschedparam): Add priority range
        check.
        (sched_setscheduler): New function; checks for a valid
        pid and policy; checks for permission to set information
        in the target process; expects pid to be a Win32 process ID,
        not a process handle; the only scheduler policy allowed is
        SCHED_OTHER.
        (sched_getscheduler): Likewise, but checks for permission
        to query.
        * pthread.h (SCHED_*): Moved to sched.h as defined in the
        POSIX standard.
        * sched.h (SCHED_*): Moved from pthread.h.
        (pid_t): Defined if necessary.
        (sched_setscheduler): Defined.
        (sched_getscheduler): Defined.
        * pthread.def (sched_setscheduler): Exported.
        (sched_getscheduler): Likewise.

2001-06-23  Ross Johnson  <rpj@setup1.ise.canberra.edu.au>

        Contributed by -  Ralf Brese <Ralf.Brese@pdb4.siemens.de>

        * create.c (pthread_create): Set thread priority from
        thread attributes.

1 files changed, 152 insertions, 86 deletions

diff --git a/ANNOUNCE b/ANNOUNCE
index b9b4970..689b16d 100644
--- a/ANNOUNCE
+++ b/ANNOUNCE
@@ -1,4 +1,4 @@
-                 PTHREADS-WIN32 SNAPSHOT 2001-06-06

+                 PTHREADS-WIN32 SNAPSHOT 2001-??-??

                  ----------------------------------

        Web Site: http://sources.redhat.com/pthreads-win32/

       FTP Site: ftp://sources.redhat.com/pub/pthreads-win32

@@ -28,83 +28,148 @@ Change Summary (since the last snapshot)
 

 (See the ChangeLog file for details.)

 

-New:

--	Async cancellation should now work for Windows running

-	on the following processors: IX86, MIPS, ALPHA, PPC.

-	- contributors name misplaced (please contact me if it's you)

- 

--       New functions (no-ops) for source code compatibility:

-                pthread_getconcurrency()

-                  	Returns the value previously set by

-			pthread_setconcurrency(), or 0 (zero)

-			if no value was previously set.

-			The implementation does not currently use

-			this value.

-                pthread_setconcurrency()

-                        Accepts any value >= 0 but does not

-                        have any effect; returns '0', or

-                        EINVAL if the value is < 0

-                pthread_attr_getscope()

-                pthread_attr_setscope()

-                        Currently only return ENOSYS

-

--       The following mutex types and related functions are now

-        supported:

-

-                pthread_mutexattr_gettype()

-                pthread_mutexattr_settype()

-                pthread_mutexattr_setkind_np()

-                pthread_mutexattr_getkind_np()

-                PTHREAD_MUTEX_DEFAULT

-                PTHREAD_MUTEX_NORMAL or PTHREAD_MUTEX_FAST_NP

-                PTHREAD_MUTEX_ERRORCHECK or PTHREAD_MUTEX_ERRORCHECK_NP

-                PTHREAD_MUTEX_RECURSIVE or PTHREAD_MUTEX_RECURSIVE_NP

-

-	The *_NP versions are for Linux portability.

-

-        In this implementation PTHREAD_MUTEX_DEFAULT is

-        mapped to PTHREAD_MUTEX_NORMAL for compatibility

-        with major Unix vendors. This is also the default

-	if no type is specified.

-

-	Please note that the default behaviour in previous

-	versions of the library was equivalent to

-	PTHREAD_MUTEX_RECURSIVE. This will be a problem

-	for some applications unless they are updated to

-	explicitly set the type of any mutexes that are, or

-	could be, locked recursively.

-

-        PTHREAD_MUTEX_NORMAL will cause thread deadlock

-        if the owner of a mutex tries to relock it without

-        first unlocking it. It has slightly less overhead

-	that other types.

-        - Thomas Pfaff <tpfaff@gmx.net>

-

--       Pthreads-win32 mutexes are now based on Win32

-        critical sections for all Windows versions. The

-	implementation no longer depends on

-	TryEnterCriticalSection.

-        - Thomas Pfaff <tpfaff@gmx.net>

-

--	New implementation of condition variables which attempts to

-	correct problems of spurious wakeups, unfairness, and

-	broadcast deadlock. See README.CV for details of the

-	discussion.

-	  - Alexander Terekhov <TEREKHOV@de.ibm.com>

-	  - Louis Thomas <lthomas@arbitrade.com>

-

-Bugs fixed:

--       Pthread_mutex_trylock() now properly returns EBUSY

-        even when the current thread owns the mutex.

-        Consequently, pthread_mutex_destroy() will no longer

-        destroy a locked mutex (it will return EBUSY).

-        - Thomas Pfaff <tpfaff@gmx.net>

-

--	Errno warnings and building with /MD;

-	calling conventions (stdcall v cdecl);

-	thread termination problems;

-	- Milan Gardian <Milan.Gardian@LEIBINGER.com>

+-----------------------

+Additions to Scheduling

+-----------------------

+New routines:

+	pthread_attr_setinheritsched()

+	pthread_attr_getinheritsched()

+	pthread_attr_setschedpolicy()

+	pthread_attr_getschedpolicy()

+	sched_setscheduler()

+	sched_getscheduler()

+	sched_rr_get_interval()

+Now defined:

+	_POSIX_THREAD_PRIORITY_SCHEDULING

+

+These routines complete the set required for defining

+_POSIX_THREAD_PRIORITY_SCHEDULING.

+

+

+sched_setscheduler

+sched_getscheduler

+------------------

+These routines will require patching to work with UWIN

+or any other system that provides it's own pid_t.

+Pthread.h conditionally defines pid_t as a DWORD, which

+is the type returned by GetCurrentProcessId().

+

+The only supported policy is SCHED_OTHER, however, in

+order to provide semantic compatibility these routines

+verify the following:

+- that the process identified by pid exists;

+- that permission is granted to set or query the policy;

+The appropriate error is returned if either of these fail.

+On success, both routines return SCHED_OTHER.

+

+

+sched_rr_get_interval

+---------------------

+Always returns -1 and sets errno to ENOTSUP.

+

+

+pthread_attr_setschedpolicy

+pthread_attr_getschedpolicy

+---------------------------

+The only valid supported policy is SCHED_OTHER.

+Attempting to set other policies results in an ENOTSUP

+error.

+

+

+pthread_attr_setinheritsched

+pthread_attr_getinheritsched

+----------------------------

+The two possible values that can be set are

+PTHREAD_INHERIT_SCHED and PTHREAD_EXPLICIT_SCHED.

+

+Neither the POSIX standard nor the Single Unix Spec

+specifies which should be the default value.

+Consequently, implementations use different defaults,

+eg (from a scan of the web):

+

+PTHREAD_INHERIT_SCHED default:

+	HP, MKS Toolkit, QNX, AIX (?)

+

+PTHREAD_EXPLICIT_SCHED default:

+	IBM OS/400, Linux, Sun

+

+All Win32 threads are created with THREAD_PRIORITY_NORMAL.

+They do not inherit the priority of the parent thread or the

+process. This behaviour is equivalent to the following

+Pthreads defaults:

+

+	Inheritance:	PTHREAD_EXPLICIT_SCHED

+	Priority:	THREAD_PRIORITY_NORMAL

+

+These are also the defaults in pthreads-win32, and now

+reinforced by changes to the library which now actually

+use these values and routines. This choice maintains the

+notion that, from the Pthread point-of-view, Win32

+threads (those not created via pthread_create()) are

+treated as detached POSIX threads with default attribute

+values.

+

+

+------------------

+Changes to Mutexes

+------------------

+Snapshot-2001-06-06 included Thomas Pfaff's enhancements

+to the mutex routines to improve speed. The improvements

+are most apparent on Win9x class systems where pthreads-win32

+previously used Win32 mutexes rather than critical

+sections as the underlying mechanism. The enhancements

+also resulted in speed improvements in other primitives

+which use mutexes internally, such as condition variables

+and read-write locks. Thomas also added mutex

+types to the library as described in the Single Unix

+Specification documentation, and as provided with

+the majority of major Unix and Linux Pthreads

+implementations.

+

+Changes have been made to further improve the speed of the

+default PTHREAD_MUTEX_NORMAL type (and therefore also

+PTHREAD_MUTEX_DEFAULT which is equivalent in pthreads-win32).

+

+Specifically, the library no longer sets or checks the real

+mutex owner when locking, unlocking, trylocking, or

+destroying PTHREAD_MUTEX_NORMAL mutexes. This saves

+significant overhead and results in measured speed increases

+of around 90 percent for non-blocking lock operations, and a

+slight improvement for blocking lock operations. Since the

+type of mutex used internally is PTHREAD_MUTEX_DEFAULT, this

+also results in additional speed improvements to CVs and R/W

+lock operations. Subjective observation shows an

+improvement of approximately 30-35% in R/W locks

+(from tests/rwlock7.c). This is compared to the

+already improved snapshot-2001-06-06.

+

+The price paid for this improvement is that some checking

+for errors is not done for these mutex types. The onus

+is placed on the developer to check application code

+for logical errors, or use a safer mutex type such as

+PTHREAD_MUTEX_ERRORCHECK. For example, it is now

+possible for a non-owner thread to unlock or destroy

+a mutex of these types. However, because the owner

+is not simply left as NULL but set to a special anonymous

+owner value when locked and then reset to NULL when

+unlocked or destroyed, an error will ultimately eventuate

+when the owner thread subsequently attempts to unlock or

+destroy the mutex.

+

+These mutex changes appear to be consistent with both

+the behaviour exhibited by other implementations and their

+documentation, including the Open Group documentation.

+

+

+-------

+Bug fix

+-------

+Pthread_create now sets the priority of the new thread

+from the value set in the thread attribute. 

+- from Ralf.Brese@pdb4.siemens.de.

+

 

+---------------------------

 Known bugs in this snapshot

 ---------------------------

 

@@ -173,24 +238,22 @@ reliably.
 Level of standards conformance

 ------------------------------

 

-The following POSIX 1003.1c 1995 options are defined:

+The following POSIX 1003.1c 1995 and POSIX 1003.1b options are defined:

 

       _POSIX_THREADS

       _POSIX_THREAD_SAFE_FUNCTIONS

       _POSIX_THREAD_ATTR_STACKSIZE

+      _POSIX_THREAD_PRIORITY_SCHEDULING

+      _POSIX_SEMAPHORES

 

 

 The following POSIX 1003.1c 1995 options are not defined:

 

       _POSIX_THREAD_ATTR_STACKADDR

-      _POSIX_THREAD_PRIORITY_SCHEDULING

       _POSIX_THREAD_PRIO_INHERIT

       _POSIX_THREAD_PRIO_PROTECT

       _POSIX_THREAD_PROCESS_SHARED

 

-The following POSIX 1003.1b option is defined:

-

-      _POSIX_SEMAPHORES

 

 The following functions are implemented:

 

@@ -291,6 +354,10 @@ The following functions are implemented:
       ---------------------------

       pthread_attr_getschedparam  

       pthread_attr_setschedparam  

+      pthread_attr_getinheritsched

+      pthread_attr_setinheritsched

+      pthread_attr_getschedpolicy (only supports SCHED_OTHER)

+      pthread_attr_setschedpolicy (only supports SCHED_OTHER)

       pthread_getschedparam

       pthread_setschedparam

       pthread_getconcurrency

@@ -299,6 +366,9 @@ The following functions are implemented:
       pthread_attr_setscope  (returns an error ENOSYS)

       sched_get_priority_max (POSIX 1b)

       sched_get_priority_min (POSIX 1b)

+      sched_rr_set_interval  (POSIX 1b  - returns an error ENOSYS)

+      sched_setscheduler     (POSIX 1b  - only supports SCHED_OTHER)

+      sched_getscheduler     (POSIX 1b  - only supports SCHED_OTHER)

       sched_yield            (POSIX 1b)

 

       ---------------------------

@@ -344,10 +414,6 @@ The following functions are not implemented:
       ---------------------------

       RealTime Scheduling

       ---------------------------

-      pthread_attr_getinheritsched

-      pthread_attr_getschedpolicy

-      pthread_attr_setinheritsched

-      pthread_attr_setschedpolicy

       pthread_mutex_getprioceiling

       pthread_mutex_setprioceiling

       pthread_mutex_attr_getprioceiling