/* * Copyright 2010 OpenSourceCodeStewardshipFoundation * * Licensed under BSD */ #include #include #include #include "VMS.h" #include "Queue_impl/BlockingQueue.h" //=========================================================================== void shutdownFn( void *dummy, VirtProcr *dummy2 ); void create_sched_slots( MasterEnv *masterEnv ); //=========================================================================== /*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 workQ 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() { MasterEnv *masterEnv; CASQueueStruc *workQ; //Make the central work-queue _VMSWorkQ = makeCASQ(); workQ = _VMSWorkQ; _VMSMasterEnv = malloc( sizeof(MasterEnv) ); masterEnv = _VMSMasterEnv; //create the master virtual processor masterEnv->masterVirtPr = VMS__create_procr( &masterLoop, masterEnv ); create_sched_slots( masterEnv ); //Set slot 0 to be the master virt procr & set flags just in case masterEnv->schedSlots[0]->needsProcrAssigned = FALSE; //says don't touch masterEnv->schedSlots[0]->workIsDone = FALSE; //says don't touch masterEnv->schedSlots[0]->procrAssignedToSlot = masterEnv->masterVirtPr; masterEnv->masterVirtPr->schedSlot = masterEnv->schedSlots[0]; masterEnv->stillRunning = FALSE; //First core loop to start up gets this, which will schedule seed Pr //TODO: debug: check address of masterVirtPr writeCASQ( masterEnv->masterVirtPr, workQ ); numProcrsCreated = 1; //======================================================================== // Create the Threads int coreIdx; //Make params given to the win threads that animate the core loops for( coreIdx=0; coreIdx < NUM_CORES; coreIdx++ ) { coreLoopThdParams[coreIdx] = malloc( sizeof(ThdParams) ); coreLoopThdParams[coreIdx]->coreNum = coreIdx; //make the core loop threads, born in suspended state coreLoopThdHandles[ coreIdx ] = CreateThread ( NULL, // Security attributes 0, // Stack size coreLoop, coreLoopThdParams[coreIdx], CREATE_SUSPENDED, &(coreLoopThdIds[coreIdx]) ); } } void create_sched_slots( MasterEnv *masterEnv ) { SchedSlot **schedSlots, **filledSlots; int i; schedSlots = malloc( NUM_SCHED_SLOTS * sizeof(SchedSlot *) ); filledSlots = malloc( NUM_SCHED_SLOTS * sizeof(SchedSlot *) ); masterEnv->schedSlots = schedSlots; masterEnv->filledSlots = filledSlots; 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; } } /*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 //=========================================================================== LARGE_INTEGER stPerfCount, endPerfCount, countFreq; unsigned long long count = 0, freq = 0; double runTime; QueryPerformanceCounter( &stPerfCount ); //start them running for( coreIdx=0; coreIdx < NUM_CORES; coreIdx++ ) { //Create the threads ResumeThread( coreLoopThdHandles[coreIdx] ); //starts thread } //wait for all to complete for( coreIdx=0; coreIdx < NUM_CORES; coreIdx++ ) { WaitForSingleObject(coreLoopThdHandles[coreIdx], INFINITE); } //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 QueryPerformanceCounter( &endPerfCount ); count = endPerfCount.QuadPart - stPerfCount.QuadPart; QueryPerformanceFrequency( &countFreq ); freq = countFreq.QuadPart; runTime = (double)count / (double)freq; printf("\n Time startup to shutdown: %f\n", runTime); fflush( stdin ); } /*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; //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 ); 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 *callingPr ) { 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. callingPr->nextInstrPt = &&ResumePt; //return ownership of the virt procr and sched slot to Master virt pr callingPr->schedSlot->workIsDone = TRUE; // coreIdx = callingPr->coreAnimatedBy; stackPtrAddr = &(callingPr->stackPtr); framePtrAddr = &(callingPr->framePtr); jmpPt = callingPr->coreLoopStartPt; coreLoopFramePtr = callingPr->coreLoopFramePtr;//need this only coreLoopStackPtr = callingPr->coreLoopStackPtr;//shouldn't need -- safety //Save the virt procr's stack and frame ptrs, 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 %%esp, (%%eax); \ movl %1, %%eax; \ movl %%ebp, (%%eax); \ movl %2, %%eax; \ movl %3, %%esp; \ movl %4, %%ebp; \ jmp %%eax " \ /* outputs */ : "=g" (stackPtrAddr), "=g" (framePtrAddr) \ /* inputs */ : "g" (jmpPt), "g"(coreLoopStackPtr), "g"(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: return; } /*This is equivalent to "jump back to core loop" -- it's mainly only used * just after adding dissipate request to a processor -- so the semantic * layer is the only place it will be seen and/or used. * *It does almost the same thing as suspend, except don't need to save the * stack nor set the nextInstrPt * *As of June 30, 2010 just implementing as a call to suspend -- just sugar */ void VMS__return_from_fn( VirtProcr *animatingPr ) { VMS__suspend_procr( animatingPr ); } /*Not sure yet the form going to put "dissipate" in, so this is the third * possibility -- the semantic layer can just make a macro that looks like * a call to its name, then expands to a call to this. * *As of June 30, 2010 this looks like the top choice.. * *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; } //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 (or maybe put request freer as a field in virt procr?) void VMS__remove_and_free_top_request( VirtProcr *procrWithReq ) { VMSReqst *req; req = procrWithReq->requests; procrWithReq->requests = procrWithReq->requests->nextReqst; free( req ); } //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?) void VMS__free_request( VMSReqst *req ) { free( req ); } 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; } 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_register_new_procr_request(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 ); } /*The semantic layer figures out when the work is done ( perhaps by a call * in the application to "work all done", or perhaps all the virtual * processors have dissipated.. a.s.o. ) * *The semantic layer is responsible for making sure all work has fully * completed before using this to shutdown the VMS system. * *After the semantic layer has determined it wants to shut down, the * next time the Master Loop calls the scheduler plug-in, the scheduler * then calls this function and returns the virtual processor it gets back. * *When the shut-down processor runs, it first frees all locations malloc'd to * the VMS system (that wasn't * specified as return-locations). Then it creates one core-loop shut-down * processor for each core loop and puts them all into the workQ. When a * core loop animates a core loop shut-down processor, it causes exit-thread * to run, and when all core loop threads have exited, then the "wait for * work to finish" in the main thread is woken, and the function-call that * started all the work returns. * *The function animated by this processor performs the shut-down work. */ VirtProcr * VMS__create_the_shutdown_procr() { return VMS__create_procr( &shutdownFn, NULL ); } /*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__free_procr_locs( 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__remove_and_free_top_request( animatingPr ); free( animatingPr->startOfStack ); //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 ); } /*This is the function run by the special "shut-down" processor * *The _VMSMasterEnv is needed by this shut down function, so the "wait" * function run in the main loop has to free it, and the thread-related * locations (coreLoopThdParams a.s.o.). *However, the semantic environment and all data malloc'd to VMS can be * freed here. * *NOTE: the semantic plug-in is expected to use VMS__malloc to get all the * locations it needs -- they will be automatically freed by the standard * "free all owned locations" * *Free any locations malloc'd to the VMS system (that weren't * specified as return-locations). *Then create one core-loop shut-down processor for each core loop and puts * them all into the workQ. */ void shutdownFn( void *dummy, VirtProcr *animatingPr ) { int coreIdx; VirtProcr *shutDownPr; CASQueueStruc *workQ = _VMSWorkQ; //free all the locations owned within the VMS system //TODO: write VMS__malloc and free.. -- take the DKU malloc as starting pt //make the core loop shut-down processors and put them into the workQ for( coreIdx=0; coreIdx < NUM_CORES; coreIdx++ ) { shutDownPr = VMS__create_procr( NULL, NULL ); shutDownPr->nextInstrPt = _VMSMasterEnv->coreLoopShutDownPt; writeCASQ( shutDownPr, workQ ); } //This is an issue: the animating processor of this function may not // get its request handled before all the cores have shutdown. //TODO: after all the threads stop, clean out the MasterEnv, the // SemanticEnv, and the workQ before returning. VMS__dissipate_procr( animatingPr ); //will never come back from this } /*This has to free anything allocated during VMS_init, and any other alloc'd * locations that might be left over. */ void VMS__shutdown() { int i; free( _VMSWorkQ ); free( _VMSMasterEnv->filledSlots ); for( i = 0; i < NUM_SCHED_SLOTS; i++ ) { free( _VMSMasterEnv->schedSlots[i] ); } free( _VMSMasterEnv->schedSlots); VMS__free_procr_locs( _VMSMasterEnv->masterVirtPr ); free( _VMSMasterEnv ); } //=========================================================================== inline TSCount getTSCount() { unsigned int low, high; TSCount out; saveTimeStampCountInto( low, high ); out = high; out = (out << 32) + low; return out; }