/* * Copyright 2010 OpenSourceStewardshipFoundation * * Licensed under BSD */ #include #include #include #include "VMS.h" #include "Queue_impl/BlockingQueue.h" #include "Histogram/Histogram.h" #define thdAttrs NULL //=========================================================================== void shutdownFn( void *dummy, VirtProcr *dummy2 ); SchedSlot ** create_sched_slots(); void create_masterEnv(); void create_the_coreLoop_OS_threads(); pthread_mutex_t suspendLock = PTHREAD_MUTEX_INITIALIZER; pthread_cond_t suspend_cond = PTHREAD_COND_INITIALIZER; //=========================================================================== /*Setup has two phases: * 1) Semantic layer first calls init_VMS, which creates masterEnv, and puts * the master virt procr into the work-queue, ready for first "call" * 2) Semantic layer then does its own init, which creates the seed virt * procr inside the semantic layer, ready to schedule it when * asked by the first run of the masterLoop. * *This part is bit weird because VMS really wants to be "always there", and * have applications attach and detach.. for now, this VMS is part of * the app, so the VMS system starts up as part of running the app. * *The semantic layer is isolated from the VMS internals by making the * semantic layer do setup to a state that it's ready with its * initial virt procrs, ready to schedule them to slots when the masterLoop * asks. Without this pattern, the semantic layer's setup would * have to modify slots directly to assign the initial virt-procrs, and put * them into the readyToAnimateQ itself, breaking the isolation completely. * * *The semantic layer creates the initial virt procr(s), and adds its * own environment to masterEnv, and fills in the pointers to * the requestHandler and slaveScheduler plug-in functions */ /*This allocates VMS data structures, populates the master VMSProc, * and master environment, and returns the master environment to the semantic * layer. */ void VMS__init() { create_masterEnv(); create_the_coreLoop_OS_threads(); } /*To initialize the sequential version, just don't create the threads */ void VMS__init_Seq() { create_masterEnv(); } void create_masterEnv() { MasterEnv *masterEnv; SRSWQueueStruc **readyToAnimateQs; int coreIdx; VirtProcr **masterVPs; SchedSlot ***allSchedSlots; //ptr to array of ptrs //Make the master env, which holds everything else _VMSMasterEnv = malloc( sizeof(MasterEnv) ); masterEnv = _VMSMasterEnv; //Need to set start pt here 'cause used by seed procr, which is created // before the first core loop starts up. -- not sure how yet.. // masterEnv->coreLoopStartPt = ; // masterEnv->coreLoopEndPt = ; //Make a readyToAnimateQ for each core loop readyToAnimateQs = malloc( NUM_CORES * sizeof(SRSWQueueStruc *) ); masterVPs = malloc( NUM_CORES * sizeof(VirtProcr *) ); //One array for each core, 3 in array, core's masterVP scheds all allSchedSlots = malloc( NUM_CORES * sizeof(SchedSlot *) ); for( coreIdx = 0; coreIdx < NUM_CORES; coreIdx++ ) { readyToAnimateQs[ coreIdx ] = makeSRSWQ(); //Q: should give masterVP core-specific into as its init data? masterVPs[ coreIdx ] = VMS__create_procr( &masterLoop, masterEnv ); masterVPs[ coreIdx ]->coreAnimatedBy = coreIdx; allSchedSlots[ coreIdx ] = create_sched_slots(); //makes for one core } _VMSMasterEnv->readyToAnimateQs = readyToAnimateQs; _VMSMasterEnv->masterVPs = masterVPs; _VMSMasterEnv->allSchedSlots = allSchedSlots; //Aug 19, 2010: no longer need to place initial masterVP into queue // because coreLoop now controls -- animates its masterVP when no work //==================== malloc substitute ======================== // //Testing whether malloc is using thread-local storage and therefore // causing unreliable behavior. //Just allocate a massive chunk of memory and roll own malloc/free and // make app use VMS__malloc_to, which will suspend and perform malloc // in the master, taking from this massive chunk. // initFreeList(); } /* void initMasterMalloc() { _VMSMasterEnv->mallocChunk = malloc( MASSIVE_MALLOC_SIZE ); //The free-list element is the first several locations of an // allocated chunk -- the address given to the application is pre- // pended with both the ownership structure and the free-list struc. //So, write the values of these into the first locations of // mallocChunk -- which marks it as free & puts in its size. listElem = (FreeListElem *)_VMSMasterEnv->mallocChunk; listElem->size = MASSIVE_MALLOC_SIZE - NUM_PREPEND_BYTES listElem->next = NULL; } void dissipateMasterMalloc() { //Just foo code -- to get going -- doing as if free list were link-list currElem = _VMSMasterEnv->freeList; while( currElem != NULL ) { nextElem = currElem->next; masterFree( currElem ); currElem = nextElem; } free( _VMSMasterEnv->freeList ); } */ SchedSlot ** create_sched_slots() { SchedSlot **schedSlots; int i; schedSlots = malloc( NUM_SCHED_SLOTS * sizeof(SchedSlot *) ); for( i = 0; i < NUM_SCHED_SLOTS; i++ ) { schedSlots[i] = malloc( sizeof(SchedSlot) ); //Set state to mean "handling requests done, slot needs filling" schedSlots[i]->workIsDone = FALSE; schedSlots[i]->needsProcrAssigned = TRUE; } return schedSlots; } void freeSchedSlots( SchedSlot **schedSlots ) { int i; for( i = 0; i < NUM_SCHED_SLOTS; i++ ) { free( schedSlots[i] ); } free( schedSlots ); } void create_the_coreLoop_OS_threads() { //======================================================================== // Create the Threads int coreIdx, retCode; //Need the threads to be created suspended, and wait for a signal // before proceeding -- gives time after creating to initialize other // stuff before the coreLoops set off. _VMSMasterEnv->setupComplete = 0; //Make the threads that animate the core loops for( coreIdx=0; coreIdx < NUM_CORES; coreIdx++ ) { coreLoopThdParams[coreIdx] = malloc( sizeof(ThdParams) ); coreLoopThdParams[coreIdx]->coreNum = coreIdx; retCode = pthread_create( &(coreLoopThdHandles[coreIdx]), thdAttrs, &coreLoop, (void *)(coreLoopThdParams[coreIdx]) ); if(retCode){printf("ERROR creating thread: %d\n", retCode); exit(0);} } } /*Semantic layer calls this when it want the system to start running.. * *This starts the core loops running then waits for them to exit. */ void VMS__start_the_work_then_wait_until_done() { int coreIdx; //Start the core loops running //=========================================================================== TSCount startCount, endCount; unsigned long long count = 0, freq = 0; double runTime; startCount = getTSCount(); //tell the core loop threads that setup is complete //get lock, to lock out any threads still starting up -- they'll see // that setupComplete is true before entering while loop, and so never // wait on the condition pthread_mutex_lock( &suspendLock ); _VMSMasterEnv->setupComplete = 1; pthread_mutex_unlock( &suspendLock ); pthread_cond_broadcast( &suspend_cond ); //wait for all to complete for( coreIdx=0; coreIdx < NUM_CORES; coreIdx++ ) { pthread_join( coreLoopThdHandles[coreIdx], NULL ); } //NOTE: do not clean up VMS env here -- semantic layer has to have // a chance to clean up its environment first, then do a call to free // the Master env and rest of VMS locations endCount = getTSCount(); count = endCount - startCount; runTime = (double)count / (double)TSCOUNT_FREQ; printf("\n Time startup to shutdown: %f\n", runTime); fflush( stdin ); } /*Only difference between version with an OS thread pinned to each core and * the sequential version of VMS is VMS__init_Seq, this, and coreLoop_Seq. */ void VMS__start_the_work_then_wait_until_done_Seq() { //Instead of un-suspending threads, just call the one and only // core loop (sequential version), in the main thread. coreLoop_Seq( NULL ); } /*Create stack, then create __cdecl structure on it and put initialData and * pointer to the new structure instance into the parameter positions on * the stack *Then put function pointer into nextInstrPt -- the stack is setup in std * call structure, so jumping to function ptr is same as a GCC generated * function call *No need to save registers on old stack frame, because there's no old * animator state to return to -- * */ VirtProcr * VMS__create_procr( VirtProcrFnPtr fnPtr, void *initialData ) { VirtProcr *newPr; char *stackLocs, *stackPtr; newPr = malloc( sizeof(VirtProcr) ); newPr->procrID = numProcrsCreated++; newPr->nextInstrPt = fnPtr; newPr->initialData = initialData; newPr->requests = NULL; newPr->schedSlot = NULL; // newPr->coreLoopStartPt = _VMSMasterEnv->coreLoopStartPt; //fnPtr takes two params -- void *initData & void *animProcr //alloc stack locations, make stackPtr be the highest addr minus room // for 2 params + return addr. Return addr (NULL) is in loc pointed to // by stackPtr, initData at stackPtr + 4 bytes, animatingPr just above stackLocs = malloc( VIRT_PROCR_STACK_SIZE ); if(stackLocs == 0) {perror("malloc stack"); exit(1);} newPr->startOfStack = stackLocs; stackPtr = ( (char *)stackLocs + VIRT_PROCR_STACK_SIZE - 0x10 ); //setup __cdecl on stack -- coreloop will switch to stackPtr before jmp *( (int *)stackPtr + 2 ) = (int) newPr; //rightmost param -- 32bit pointer *( (int *)stackPtr + 1 ) = (int) initialData; //next param to left newPr->stackPtr = stackPtr; //core loop will switch to this, then newPr->framePtr = stackPtr; //suspend loop will save new stack & frame ptr return newPr; } /*there is a label inside this function -- save the addr of this label in * the callingPr struc, as the pick-up point from which to start the next * work-unit for that procr. If turns out have to save registers, then * save them in the procr struc too. Then do assembly jump to the CoreLoop's * "done with work-unit" label. The procr struc is in the request in the * slave that animated the just-ended work-unit, so all the state is saved * there, and will get passed along, inside the request handler, to the * next work-unit for that procr. */ void VMS__suspend_procr( VirtProcr *animatingPr ) { void *jmpPt, *stackPtrAddr, *framePtrAddr, *coreLoopStackPtr; void *coreLoopFramePtr; //The request to master will cause this suspended virt procr to get // scheduled again at some future point -- to resume, core loop jumps // to the resume point (below), which causes restore of saved regs and // "return" from this call. animatingPr->nextInstrPt = &&ResumePt; //return ownership of the virt procr and sched slot to Master virt pr animatingPr->schedSlot->workIsDone = TRUE; // coreIdx = callingPr->coreAnimatedBy; stackPtrAddr = &(animatingPr->stackPtr); framePtrAddr = &(animatingPr->framePtr); jmpPt = _VMSMasterEnv->coreLoopStartPt; coreLoopFramePtr = animatingPr->coreLoopFramePtr;//need this only coreLoopStackPtr = animatingPr->coreLoopStackPtr;//safety //Save the virt procr's stack and frame ptrs, asm volatile("movl %0, %%eax; \ movl %%esp, (%%eax); \ movl %1, %%eax; \ movl %%ebp, (%%eax) "\ /* outputs */ : "=g" (stackPtrAddr), "=g" (framePtrAddr) \ /* inputs */ : \ /* clobber */ : "%eax" \ ); //=========================== Measurement stuff ======================== #ifdef MEAS__TIME_STAMP_SUSP //record time stamp: compare to time-stamp recorded below saveLowTimeStampCountInto( animatingPr->preSuspTSCLow ); #endif //======================================================================= //restore coreloop's frame ptr, then jump back to "start" of core loop //Note, GCC compiles to assembly that saves esp and ebp in the stack // frame -- so have to explicitly do assembly that saves to memory asm volatile("movl %0, %%eax; \ movl %1, %%esp; \ movl %2, %%ebp; \ jmp %%eax " \ /* outputs */ : \ /* inputs */ : "m" (jmpPt), "m"(coreLoopStackPtr), "m"(coreLoopFramePtr)\ /* clobber */ : "memory", "%eax", "%ebx", "%ecx", "%edx", "%edi","%esi" \ ); //list everything as clobbered to force GCC to save all // live vars that are in regs on stack before this // assembly, so that stack pointer is correct, before jmp ResumePt: #ifdef MEAS__TIME_STAMP_SUSP //NOTE: only take low part of count -- do sanity check when take diff saveLowTimeStampCountInto( animatingPr->postSuspTSCLow ); #endif return; } /* *This adds a request to dissipate, then suspends the processor so that the * request handler will receive the request. The request handler is what * does the work of freeing memory and removing the processor from the * semantic environment's data structures. *The request handler also is what figures out when to shutdown the VMS * system -- which causes all the core loop threads to die, and returns from * the call that started up VMS to perform the work. * *This form is a bit misleading to understand if one is trying to figure out * how VMS works -- it looks like a normal function call, but inside it * sends a request to the request handler and suspends the processor, which * jumps out of the VMS__dissipate_procr function, and out of all nestings * above it, transferring the work of dissipating to the request handler, * which then does the actual work -- causing the processor that animated * the call of this function to disappear and the "hanging" state of this * function to just poof into thin air -- the virtual processor's trace * never returns from this call, but instead the virtual processor's trace * gets suspended in this call and all the virt processor's state disap- * pears -- making that suspend the last thing in the virt procr's trace. */ void VMS__dissipate_procr( VirtProcr *procrToDissipate ) { VMSReqst *req; req = malloc( sizeof(VMSReqst) ); // req->virtProcrFrom = callingPr; req->reqType = dissipate; req->nextReqst = procrToDissipate->requests; procrToDissipate->requests = req; VMS__suspend_procr( procrToDissipate ); } /*This inserts the semantic-layer's request data into standard VMS carrier */ inline void VMS__add_sem_request( void *semReqData, VirtProcr *callingPr ) { VMSReqst *req; req = malloc( sizeof(VMSReqst) ); // req->virtProcrFrom = callingPr; req->reqType = semantic; req->semReqData = semReqData; req->nextReqst = callingPr->requests; callingPr->requests = req; } /*Use this to get first request before starting request handler's loop */ VMSReqst * VMS__take_top_request_from( VirtProcr *procrWithReq ) { VMSReqst *req; req = procrWithReq->requests; if( req == NULL ) return req; procrWithReq->requests = procrWithReq->requests->nextReqst; return req; } /*A subtle bug due to freeing then accessing "next" after freed caused this * form of call to be put in -- so call this at end of request handler loop * that iterates through the requests. */ VMSReqst * VMS__free_top_and_give_next_request_from( VirtProcr *procrWithReq ) { VMSReqst *req; req = procrWithReq->requests; if( req == NULL ) return NULL; procrWithReq->requests = procrWithReq->requests->nextReqst; VMS__free_request( req ); return procrWithReq->requests; } //TODO: add a semantic-layer supplied "freer" for the semantic-data portion // of a request -- IE call with both a virt procr and a fn-ptr to request // freer (also maybe put sem request freer as a field in virt procr?) //MeasVMS relies right now on this only freeing VMS layer of request -- the // semantic portion of request is alloc'd and freed by request handler void VMS__free_request( VMSReqst *req ) { free( req ); } inline int VMS__isSemanticReqst( VMSReqst *req ) { return ( req->reqType == semantic ); } inline void * VMS__take_sem_reqst_from( VMSReqst *req ) { return req->semReqData; } inline int VMS__isDissipateReqst( VMSReqst *req ) { return ( req->reqType == dissipate ); } inline int VMS__isCreateReqst( VMSReqst *req ) { return ( req->reqType == regCreated ); } void VMS__send_req_to_register_new_procr(VirtProcr *newPr, VirtProcr *reqstingPr) { VMSReqst *req; req = malloc( sizeof(VMSReqst) ); req->reqType = regCreated; req->semReqData = newPr; req->nextReqst = reqstingPr->requests; reqstingPr->requests = req; VMS__suspend_procr( reqstingPr ); } /*This must be called by the request handler plugin -- it cannot be called * from the semantic library "dissipate processor" function -- instead, the * semantic layer has to generate a request for the plug-in to call this * function. *The reason is that this frees the virtual processor's stack -- which is * still in use inside semantic library calls! * *This frees or recycles all the state owned by and comprising the VMS * portion of the animating virtual procr. The request handler must first * free any semantic data created for the processor that didn't use the * VMS_malloc mechanism. Then it calls this, which first asks the malloc * system to disown any state that did use VMS_malloc, and then frees the * statck and the processor-struct itself. *If the dissipated processor is the sole (remaining) owner of VMS__malloc'd * state, then that state gets freed (or sent to recycling) as a side-effect * of dis-owning it. */ void VMS__handle_dissipate_reqst( VirtProcr *animatingPr ) { //dis-own all locations owned by this processor, causing to be freed // any locations that it is (was) sole owner of //TODO: implement VMS__malloc system, including "give up ownership" //The dissipate request might still be attached, so remove and free it VMS__free_top_and_give_next_request_from( animatingPr ); //NOTE: initialData was given to the processor, so should either have // been alloc'd with VMS__malloc, or freed by the level above animPr. //So, all that's left to free here is the stack and the VirtProcr struc // itself free( animatingPr->startOfStack ); free( animatingPr ); } //TODO: re-architect so that have clean separation between request handler // and master loop, for dissipate, create, shutdown, and other non-semantic // requests. Issue is chain: one removes requests from AppVP, one dispatches // on type of request, and one handles each type.. but some types require // action from both request handler and master loop -- maybe just give the // request handler calls like: VMS__handle_X_request_type void endOSThreadFn( void *initData, VirtProcr *animatingPr ); /*This is called by the semantic layer's request handler when it decides its * time to shut down the VMS system. Calling this causes the core loop OS * threads to exit, which unblocks the entry-point function that started up * VMS, and allows it to grab the result and return to the original single- * threaded application. * *The _VMSMasterEnv is needed by this shut down function, so the create-seed- * and-wait function has to free a bunch of stuff after it detects the * threads have all died: the masterEnv, the thread-related locations, * masterVP any AppVPs that might still be allocated and sitting in the * semantic environment, or have been orphaned in the _VMSWorkQ. * *NOTE: the semantic plug-in is expected to use VMS__malloc_to to get all the * locations it needs, and give ownership to masterVP. Then, they will be * automatically freed when the masterVP is dissipated. (This happens after * the core loop threads have all exited) * *In here,create one core-loop shut-down processor for each core loop and put * them all directly into the readyToAnimateQ. *Note, this function can ONLY be called after the semantic environment no * longer cares if AppVPs get animated after the point this is called. In * other words, this can be used as an abort, or else it should only be * called when all AppVPs have finished dissipate requests -- only at that * point is it sure that all results have completed. */ void VMS__handle_shutdown_reqst( void *dummy, VirtProcr *animatingPr ) { int coreIdx; VirtProcr *shutDownPr; //create the shutdown processors, one for each core loop -- put them // directly into the Q -- each core will die when gets one for( coreIdx=0; coreIdx < NUM_CORES; coreIdx++ ) { shutDownPr = VMS__create_procr( &endOSThreadFn, NULL ); writeSRSWQ( shutDownPr, _VMSMasterEnv->readyToAnimateQs[coreIdx] ); } } /*Am trying to be cute, avoiding IF statement in coreLoop that checks for * a special shutdown procr. Ended up with extra-complex shutdown sequence. *This function has the sole purpose of setting the stack and framePtr * to the coreLoop's stack and framePtr.. it does that then jumps to the * core loop's shutdown point -- might be able to just call Pthread_exit * from here, but am going back to the pthread's stack and setting everything * up just as if it never jumped out, before calling pthread_exit. *The end-point of core loop will free the stack and so forth of the * processor that animates this function, (this fn is transfering the * animator of the AppVP that is in turn animating this function over * to core loop function -- note that this slices out a level of virtual * processors). */ void endOSThreadFn( void *initData, VirtProcr *animatingPr ) { void *jmpPt, *coreLoopStackPtr, *coreLoopFramePtr; jmpPt = _VMSMasterEnv->coreLoopEndPt; coreLoopStackPtr = animatingPr->coreLoopStackPtr; coreLoopFramePtr = animatingPr->coreLoopFramePtr; asm volatile("movl %0, %%eax; \ movl %1, %%esp; \ movl %2, %%ebp; \ jmp %%eax " \ /* outputs */ : \ /* inputs */ : "m" (jmpPt), "m"(coreLoopStackPtr), "m"(coreLoopFramePtr)\ /* clobber */ : "memory", "%eax", "%ebx", "%ecx", "%edx", "%edi","%esi" \ ); } /*This is called after the threads have shut down and control has returned * to the semantic layer, in the entry point function in the main thread. * It has to free anything allocated during VMS_init, and any other alloc'd * locations that might be left over. */ void VMS__cleanup_after_shutdown() { SRSWQueueStruc **readyToAnimateQs; int coreIdx; VirtProcr **masterVPs; SchedSlot ***allSchedSlots; //ptr to array of ptrs readyToAnimateQs = _VMSMasterEnv->readyToAnimateQs; masterVPs = _VMSMasterEnv->masterVPs; allSchedSlots = _VMSMasterEnv->allSchedSlots; for( coreIdx = 0; coreIdx < NUM_CORES; coreIdx++ ) { freeSRSWQ( readyToAnimateQs[ coreIdx ] ); VMS__handle_dissipate_reqst( masterVPs[ coreIdx ] ); freeSchedSlots( allSchedSlots[ coreIdx ] ); } free( _VMSMasterEnv->readyToAnimateQs ); free( _VMSMasterEnv->masterVPs ); free( _VMSMasterEnv->allSchedSlots ); free( _VMSMasterEnv ); } //=========================================================================== inline TSCount getTSCount() { unsigned int low, high; TSCount out; saveTimeStampCountInto( low, high ); out = high; out = (out << 32) + low; return out; }