Fix for bug 71718. Make Mac NSPR work on dual CPU Mac OS X macines by using MP critical regions to fix threading synchronization problems. r=gordon, wtc.

git-svn-id: svn://10.0.0.236/branches/NSPRPUB_PRE_4_2_CLIENT_BRANCH@103266 18797224-902f-48f8-a5cc-f745e15eee43
This commit is contained in:
sfraser%netscape.com
2001-09-20 03:42:34 +00:00
parent 80d937bbfa
commit 00479da414
6 changed files with 269 additions and 46 deletions

View File

@@ -138,6 +138,23 @@ struct _MDFileDesc {
** Interrupts Related definitions
*/
#define _MD_GET_INTSOFF() (_pr_intsOff)
#define _MD_INTSOFF(_is) \
PR_BEGIN_MACRO \
ENTER_CRITICAL_REGION(); \
(_is) = _PR_MD_GET_INTSOFF(); \
_PR_MD_SET_INTSOFF(1); \
LEAVE_CRITICAL_REGION(); \
PR_END_MACRO
#if TARGET_CARBON
extern void _MD_SetIntsOff(PRInt32 ints);
#define _MD_SET_INTSOFF(_val) _MD_SetIntsOff(_val)
#else /* not TARGET_CARBON */
#define _MD_SET_INTSOFF(_val) (_pr_intsOff = _val)
#endif /* TARGET_CARBON */
#define _MD_START_INTERRUPTS _MD_StartInterrupts
#define _MD_STOP_INTERRUPTS _MD_StopInterrupts
#define _MD_BLOCK_CLOCK_INTERRUPTS()
@@ -238,6 +255,8 @@ extern PRStatus _MD_InitThread(PRThread *thread);
/*
** Initialize the thread context preparing it to execute _main.
** *sp = 0 zeros out the sp for the first stack frame so that
** stack walking code can find the top of the stack.
*/
#if defined(powerc) || defined(__powerc)
#define _MD_INIT_CONTEXT(_thread, _sp, _main, _status) \
@@ -248,6 +267,7 @@ extern PRStatus _MD_InitThread(PRThread *thread);
*((PRBool *)_status) = PR_TRUE; \
(void) setjmp(jb); \
sp = INIT_STACKPTR(_sp); \
*sp = 0; \
(_MD_GET_SP(_thread)) = (long) sp; \
tvect = (unsigned long *)_main; \
(_MD_GET_PC(_thread)) = (int) *tvect; \
@@ -627,4 +647,36 @@ extern PRStatus _MD_CloseFileMap(struct PRFileMap *fmap);
extern void SetLogFileTypeCreator(const char *logFile);
extern int _MD_mac_get_nonblocking_connect_error(PRInt32 osfd);
/*
* Critical section support
*/
#define MAC_CRITICAL_REGIONS TARGET_CARBON
#if MAC_CRITICAL_REGIONS
extern void InitCriticalRegion();
extern void TermCriticalRegion();
extern void EnterCritialRegion();
extern void LeaveCritialRegion();
#define INIT_CRITICAL_REGION() InitCriticalRegion()
#define TERM_CRITICAL_REGION() TermCriticalRegion()
#define ENTER_CRITICAL_REGION() EnterCritialRegion()
#define LEAVE_CRITICAL_REGION() LeaveCritialRegion()
#else
#define INIT_CRITICAL_REGION()
#define TERM_CRITICAL_REGION()
#define ENTER_CRITICAL_REGION()
#define LEAVE_CRITICAL_REGION()
#endif
#endif /* prmacos_h___ */

View File

@@ -328,8 +328,10 @@ NSPR_API(PRInt32) _pr_intsOff;
#define _MD_LAST_THREAD() (_pr_lastThread)
#define _MD_SET_LAST_THREAD(t) (_pr_lastThread = t)
#ifndef XP_MAC
#define _MD_GET_INTSOFF() (_pr_intsOff)
#define _MD_SET_INTSOFF(_val) (_pr_intsOff = _val)
#endif
/* The unbalanced curly braces in these two macros are intentional */
@@ -374,12 +376,20 @@ extern PRInt32 _native_threads_only;
#else
#ifdef XP_MAC
#define _PR_INTSOFF(_is) _MD_INTSOFF(_is)
#else /* XP_MAC */
#define _PR_INTSOFF(_is) \
PR_BEGIN_MACRO \
(_is) = _PR_MD_GET_INTSOFF(); \
_PR_MD_SET_INTSOFF(1); \
PR_END_MACRO
#endif /* XP_MAC */
#define _PR_FAST_INTSON(_is) \
PR_BEGIN_MACRO \
_PR_MD_SET_INTSOFF(_is); \

View File

@@ -91,20 +91,21 @@ typedef struct ExtendedParamBlock ExtendedParamBlock;
static void AsyncIOCompletion (ExtendedParamBlock *pbAsyncPtr)
{
_PRCPU *cpu = _PR_MD_CURRENT_CPU();
PRThread *thread = pbAsyncPtr->thread;
PRThread *thread = pbAsyncPtr->thread;
PRIntn is;
if (_PR_MD_GET_INTSOFF()) {
cpu->u.missed[cpu->where] |= _PR_MISSED_IO;
thread->md.missedIONotify = PR_TRUE;
return;
cpu->u.missed[cpu->where] |= _PR_MISSED_IO;
return;
}
_PR_MD_SET_INTSOFF(1);
thread->md.osErrCode = noErr;
DoneWaitingOnThisThread(thread);
_PR_INTSOFF(is);
_PR_MD_SET_INTSOFF(0);
thread->md.osErrCode = noErr;
DoneWaitingOnThisThread(thread);
_PR_FAST_INTSON(is);
}
void _MD_SetError(OSErr oserror)

View File

@@ -171,8 +171,8 @@ static pascal void DNSNotifierRoutine(void * contextPtr, OTEventCode otEvent, O
switch (otEvent) {
case T_DNRSTRINGTOADDRCOMPLETE:
if (_PR_MD_GET_INTSOFF()) {
cpu->u.missed[cpu->where] |= _PR_MISSED_IO;
dnsContext.thread->md.missedIONotify = PR_TRUE;
cpu->u.missed[cpu->where] |= _PR_MISSED_IO;
return;
}
DoneWaitingOnThisThread(dnsContext.thread);
@@ -187,8 +187,8 @@ static pascal void DNSNotifierRoutine(void * contextPtr, OTEventCode otEvent, O
dnsContext.serviceRef = nil;
if (_PR_MD_GET_INTSOFF()) {
cpu->u.missed[cpu->where] |= _PR_MISSED_IO;
dnsContext.thread->md.missedIONotify = PR_TRUE;
cpu->u.missed[cpu->where] |= _PR_MISSED_IO;
return;
}
DoneWaitingOnThisThread(dnsContext.thread);
@@ -294,8 +294,8 @@ WakeUpNotifiedThread(PRThread *thread, OTResult result)
if (thread) {
thread->md.osErrCode = result;
if (_PR_MD_GET_INTSOFF()) {
cpu->u.missed[cpu->where] |= _PR_MISSED_IO;
thread->md.missedIONotify = PR_TRUE;
cpu->u.missed[cpu->where] |= _PR_MISSED_IO;
return;
}
DoneWaitingOnThisThread(thread);
@@ -1171,8 +1171,8 @@ static pascal void RawEndpointNotifierRoutine(void * contextPtr, OTEventCode co
if (thread) {
thread->md.osErrCode = result;
if (_PR_MD_GET_INTSOFF()) {
cpu->u.missed[cpu->where] |= _PR_MISSED_IO;
thread->md.asyncNotifyPending = PR_TRUE;
cpu->u.missed[cpu->where] |= _PR_MISSED_IO;
return;
}
DoneWaitingOnThisThread(thread);

View File

@@ -36,11 +36,13 @@
#include <string.h>
#include <Types.h>
#include <MacTypes.h>
#include <Timer.h>
#include <OSUtils.h>
#include <LowMem.h>
#include <Multiprocessing.h>
#include <Gestalt.h>
TimerUPP gTimerCallbackUPP = NULL;
@@ -168,24 +170,26 @@ _PRInterruptTable _pr_interruptTable[] = {
pascal void TimerCallback(TMTaskPtr tmTaskPtr)
{
_PRCPU *cpu = _PR_MD_CURRENT_CPU();
PRIntn is;
if (_PR_MD_GET_INTSOFF()) {
cpu->u.missed[cpu->where] |= _PR_MISSED_CLOCK;
PrimeTime((QElemPtr)tmTaskPtr, kMacTimerInMiliSecs);
return;
PrimeTime((QElemPtr)tmTaskPtr, kMacTimerInMiliSecs);
return;
}
_PR_MD_SET_INTSOFF(1);
// And tell nspr that a clock interrupt occured.
_PR_ClockInterrupt();
if ((_PR_RUNQREADYMASK(cpu)) >> ((_PR_MD_CURRENT_THREAD()->priority)))
_PR_SET_RESCHED_FLAG();
_PR_MD_SET_INTSOFF(0);
_PR_INTSOFF(is);
// Reset the clock timer so that we fire again.
PrimeTime((QElemPtr)tmTaskPtr, kMacTimerInMiliSecs);
// And tell nspr that a clock interrupt occured.
_PR_ClockInterrupt();
if ((_PR_RUNQREADYMASK(cpu)) >> ((_PR_MD_CURRENT_THREAD()->priority)))
_PR_SET_RESCHED_FLAG();
_PR_FAST_INTSON(is);
// Reset the clock timer so that we fire again.
PrimeTime((QElemPtr)tmTaskPtr, kMacTimerInMiliSecs);
}
@@ -219,25 +223,22 @@ void _MD_StopInterrupts(void)
void _MD_PauseCPU(PRIntervalTime timeout)
{
#pragma unused (timeout)
/* unsigned long finalTicks; */
EventRecord theEvent;
if (timeout != PR_INTERVAL_NO_WAIT)
{
EventRecord theEvent;
if (timeout != PR_INTERVAL_NO_WAIT) {
/* Delay(1,&finalTicks); */
/*
** Calling WaitNextEvent() here is suboptimal. This routine should
** pause the process until IO or the timeout occur, yielding time to
** other processes on operating systems that require this (Mac OS classic).
** WaitNextEvent() may incur too much latency, and has other problems,
** such as the potential to drop suspend/resume events, and to handle
** AppleEvents at a time at which we're not prepared to handle them.
*/
(void) WaitNextEvent(nullEvent, &theEvent, 1, NULL);
/*
** Rather than calling Delay() which basically just wedges the processor
** we'll instead call WaitNextEvent() with a mask that ignores all events
** which gives other apps a chance to get time rather than just locking up
** the machine when we're waiting for a long time (or in an infinite loop,
** whichever comes first)
*/
(void)WaitNextEvent(nullEvent, &theEvent, 1, NULL);
(void) _MD_IOInterrupt();
}
(void) _MD_IOInterrupt();
}
}
@@ -276,6 +277,11 @@ void WaitOnThisThread(PRThread *thread, PRIntervalTime timeout)
PRIntervalTime timein = PR_IntervalNow();
PRStatus status = PR_SUCCESS;
// Turn interrupts off to avoid a race over lock ownership with the callback
// (which can fire at any time). Interrupts may stay off until we leave
// this function, or another NSPR thread turns them back on. They certainly
// stay off until PR_WaitCondVar() relinquishes the asyncIOLock lock, which
// is what we care about.
_PR_INTSOFF(is);
PR_Lock(thread->md.asyncIOLock);
if (timeout == PR_INTERVAL_NO_TIMEOUT) {
@@ -302,9 +308,24 @@ void DoneWaitingOnThisThread(PRThread *thread)
{
intn is;
PR_ASSERT(thread->md.asyncIOLock->owner == NULL);
// DoneWaitingOnThisThread() is called from OT notifiers and async file I/O
// callbacks that can run at "interrupt" time (Classic Mac OS) or on pthreads
// that may run concurrently with the main threads (Mac OS X). They can thus
// be called when any NSPR thread is running, or even while NSPR is in a
// thread context switch. It is therefore vital that we can guarantee to
// be able to get the asyncIOLock without blocking (thus avoiding code
// that makes assumptions about the current NSPR thread etc). To achieve
// this, we use NSPR interrrupts as a semaphore on the lock; all code
// that grabs the lock also disables interrupts for the time the lock
// is held. Callers of DoneWaitingOnThisThread() thus have to check whether
// interrupts are already off, and, if so, simply set the missed_IO flag on
// the CPU rather than calling this function.
_PR_INTSOFF(is);
PR_Lock(thread->md.asyncIOLock);
thread->io_pending = PR_FALSE;
thread->io_pending = PR_FALSE;
/* let the waiting thread know that async IO completed */
PR_NotifyCondVar(thread->md.asyncIOCVar);
PR_Unlock(thread->md.asyncIOLock);
@@ -319,6 +340,7 @@ PR_IMPLEMENT(void) PR_Mac_WaitForAsyncNotify(PRIntervalTime timeout)
PRStatus status = PR_SUCCESS;
PRThread *thread = _PR_MD_CURRENT_THREAD();
// See commments in WaitOnThisThread()
_PR_INTSOFF(is);
PR_Lock(thread->md.asyncIOLock);
if (timeout == PR_INTERVAL_NO_TIMEOUT) {
@@ -344,11 +366,14 @@ void AsyncNotify(PRThread *thread)
{
intn is;
PR_ASSERT(thread->md.asyncIOLock->owner == NULL);
// See commments in DoneWaitingOnThisThread()
_PR_INTSOFF(is);
PR_Lock(thread->md.asyncIOLock);
thread->md.asyncNotifyPending = PR_TRUE;
thread->md.asyncNotifyPending = PR_TRUE;
/* let the waiting thread know that async IO completed */
PR_NotifyCondVar(thread->md.asyncIOCVar); // let thread know that async IO completed
PR_NotifyCondVar(thread->md.asyncIOCVar);
PR_Unlock(thread->md.asyncIOLock);
_PR_FAST_INTSON(is);
}
@@ -359,8 +384,8 @@ PR_IMPLEMENT(void) PR_Mac_PostAsyncNotify(PRThread *thread)
_PRCPU * cpu = _PR_MD_CURRENT_CPU();
if (_PR_MD_GET_INTSOFF()) {
cpu->u.missed[cpu->where] |= _PR_MISSED_IO;
thread->md.missedAsyncNotify = PR_TRUE;
cpu->u.missed[cpu->where] |= _PR_MISSED_IO;
} else {
AsyncNotify(thread);
}
@@ -407,3 +432,135 @@ PRStatus _MD_KillProcess(PRProcess *process)
PR_SetError(PR_NOT_IMPLEMENTED_ERROR, unimpErr);
return PR_FAILURE;
}
//##############################################################################
//##############################################################################
#pragma mark -
#pragma mark INTERRUPT SUPPORT
#if TARGET_CARBON
/*
This critical region support is required for Mac NSPR to work correctly on dual CPU
machines on Mac OS X. This note explains why.
NSPR uses a timer task, and has callbacks for async file I/O and Open Transport
whose runtime behaviour differs depending on environment. On "Classic" Mac OS
these run at "interrupt" time (OS-level interrupts, that is, not NSPR interrupts),
and can thus preempt other code, but they always run to completion.
On Mac OS X, these are all emulated using MP tasks, which sit atop pthreads. Thus,
they can be preempted at any time (and not necessarily run to completion), and can
also run *concurrently* with eachother, and with application code, on multiple
CPU machines. Note that all NSPR threads are emulated, and all run on the main
application MP task.
We thus have to use MP critical sections to protect data that is shared between
the various callbacks and the main MP thread. It so happens that NSPR has this
concept of software interrupts, and making interrupt-off times be critical
sections works.
*/
/*
Whether to use critical regions. True if running on Mac OS X and later
*/
PRBool gUseCriticalRegions;
/*
Count of the number of times we've entered the critical region.
We need this because ENTER_CRITICAL_REGION() will *not* block when
called from different NSPR threads (which all run on one MP thread),
and we need to ensure that when code turns interrupts back on (by
settings _pr_intsOff to 0) we exit the critical section enough times
to leave it.
*/
PRInt32 gCriticalRegionEntryCount;
void _MD_SetIntsOff(PRInt32 ints)
{
ENTER_CRITICAL_REGION();
gCriticalRegionEntryCount ++;
_pr_intsOff = ints;
if (!ints)
{
PRInt32 i = gCriticalRegionEntryCount;
gCriticalRegionEntryCount = 0;
for ( ;i > 0; i --) {
LEAVE_CRITICAL_REGION();
}
}
}
#endif /* TARGET_CARBON */
//##############################################################################
//##############################################################################
#pragma mark -
#pragma mark CRITICAL REGION SUPPORT
#if MAC_CRITICAL_REGIONS
MPCriticalRegionID gCriticalRegion;
void InitCriticalRegion()
{
long systemVersion;
OSStatus err;
// we only need to do critical region stuff on Mac OS X
err = Gestalt(gestaltSystemVersion, &systemVersion);
gUseCriticalRegions = (err == noErr) && (systemVersion >= 0x00001000);
if (!gUseCriticalRegions) return;
err = MPCreateCriticalRegion(&gCriticalRegion);
PR_ASSERT(err == noErr);
}
void TermCriticalRegion()
{
OSStatus err;
if (!gUseCriticalRegions) return;
err = MPDeleteCriticalRegion(gCriticalRegion);
PR_ASSERT(err == noErr);
}
void EnterCritialRegion()
{
OSStatus err;
if (!gUseCriticalRegions) return;
PR_ASSERT(gCriticalRegion != kInvalidID);
err = MPEnterCriticalRegion(gCriticalRegion, kDurationForever /* 10000 * kDurationMillisecond */ );
PR_ASSERT(err == noErr);
}
void LeaveCritialRegion()
{
OSStatus err;
if (!gUseCriticalRegions) return;
PR_ASSERT(gCriticalRegion != kInvalidID);
err = MPExitCriticalRegion(gCriticalRegion);
PR_ASSERT(err == noErr);
}
#endif // MAC_CRITICAL_REGIONS

View File

@@ -288,6 +288,8 @@ void _MD_EarlyInit()
{
Handle environmentVariables;
INIT_CRITICAL_REGION();
#if !defined(MAC_NSPR_STANDALONE)
// MacintoshInitializeMemory(); Moved to mdmacmem.c: AllocateRawMemory(Size blockSize)
#else
@@ -374,6 +376,7 @@ void CleanupTermProc(void)
_MD_StopInterrupts(); // deactive Time Manager task
CLOSE_OPEN_TRANSPORT();
TERM_CRITICAL_REGION();
__NSTerminate();
}