Introducing the BufferCache announced on the mailing list, and some small changes. L4Sys is still WIP.

git-svn-id: https://www4.informatik.uni-erlangen.de/i4svn/danceos/trunk/devel/fail@1316 8c4709b5-6ec9-48aa-a5cd-a96041d1645a

core/SAL/bochs/BochsController.cc

@@ -92,10 +92,14 @@ void BochsController::dbgEnableInstrPtrOutput(unsigned regularity, std::ostream*
 
 void BochsController::onInstrPtrChanged(address_t instrPtr, address_t address_space)
 {
-  #ifdef DEBUG
+#if 0
+	//the original code - performs magnitudes worse than
+	//the code below and is responsible for most (~87 per cent)
+	//of the slowdown of FailBochs
+#ifdef DEBUG
 	if(m_Regularity != 0 && ++m_Counter % m_Regularity == 0)
 		(*m_pDest) << "0x" << std::hex << instrPtr;
-  #endif
+#endif
 	// Check for active breakpoint-events:
 	fi::EventList::iterator it = m_EvList.begin();
 	while(it != m_EvList.end())
@@ -113,6 +117,37 @@ void BochsController::onInstrPtrChanged(address_t instrPtr, address_t address_sp
 		it++;
 	}
 	m_EvList.fireActiveEvents();
+#endif
+	//this code is highly optimised for the average case, so it me appear a bit ugly
+	bool do_fire = false;
+	unsigned i = 0;
+	fi::BufferCache<fi::BPEvent*> *buffer_cache = m_EvList.getBPBuffer();
+	while(i < buffer_cache->get_count()) {
+		fi::BPEvent *pEvBreakpt = buffer_cache->get(i);
+		if(pEvBreakpt->isMatching(instrPtr, address_space)) {
+			pEvBreakpt->setTriggerInstructionPointer(instrPtr);
+
+			//transition to STL: find the element we are working on in the Event List
+			fi::EventList::iterator it = std::find(m_EvList.begin(), m_EvList.end(), pEvBreakpt);
+			it = m_EvList.makeActive(it);
+			//find out how much elements need to be skipped to get in sync again
+			//(should be one or none, the loop is just to make sure)
+			for(unsigned j = i; j < buffer_cache->get_count(); j++) {
+				if(buffer_cache->get(j) == (*it)) {
+					i = j;
+					break;
+				}
+			}
+			// we now know we need to fire the active events - usually we do not have to
+			do_fire = true;
+			// "i" has already been set to the next element (by calling
+			// makeActive()):
+			continue; // -> skip loop  increment
+		}
+		i++;
+	}
+	if(do_fire)
+		m_EvList.fireActiveEvents();
 	// Note: SimulatorController::onBreakpointEvent will not be invoked in this
 	//       implementation.
 }

core/SAL/bochs/IOPortCom.ah

@@ -14,7 +14,7 @@
 aspect IOPortCom
 {
 // ATM only capturing bytewise output (most common, I suppose)
-	pointcut outInstruction() = "% ...::bx_cpu_c::OUT_DXAL%(...)";
+	pointcut outInstruction() = "% ...::bx_cpu_c::OUT_DXAL(...)";
 
 	advice execution (outInstruction()) : after ()
 	{
@@ -23,7 +23,7 @@ aspect IOPortCom
 		sal::simulator.onIOPortEvent(rAL, rDX, true);
 	}
 
-	pointcut inInstruction() = "% ...::bx_cpu_c::IN_ALDX%(...)";
+	pointcut inInstruction() = "% ...::bx_cpu_c::IN_ALDX(...)";
 
 	advice execution (inInstruction()) : after ()
 	{

core/controller/BufferCache.cc

@@ -0,0 +1,74 @@
+#include "BufferCache.hpp"
+#include "Event.hpp"
+
+namespace fi {
+
+template<class T> int BufferCache<T>::add(T val) {
+	size_t new_size = get_count() + 1;
+	size_t new_last_index = get_count();
+
+	int res = reallocate_buffer(new_size);
+	if (res == 0) {
+		set(new_last_index, val);
+	}
+
+	return res;
+}
+
+template<class T> int BufferCache<T>::remove(T val) {
+	bool do_remove = false;
+	for (size_t i = 0; i < get_count(); i++) {
+		if (get(i) == val) {
+			do_remove = true;
+		}
+		if (do_remove) {
+			if (i > get_count() - 1) {
+				set(i, get(i + 1));
+			}
+		}
+	}
+
+	int res = 0;
+	if (do_remove) {
+		size_t new_size = get_count() - 1;
+		res = reallocate_buffer(new_size);
+	}
+
+	return res;
+}
+
+template<class T> void BufferCache<T>::clear() {
+	set_count(0);
+	free(m_Buffer);
+	m_Buffer = NULL;
+}
+
+template<class T> int BufferCache<T>::erase(int idx) {
+	for (size_t i = idx; i < get_count() - 1; i++) {
+		set(i, get(i + 1));
+	}
+
+	size_t new_size = get_count() - 1;
+	if (reallocate_buffer(new_size) != 0)
+		return -1;
+	return idx;
+}
+
+template<class T> int BufferCache<T>::reallocate_buffer(size_t new_size) {
+	if (new_size == 0) {
+		clear();
+		return 0;
+	}
+	void *new_buffer = realloc(m_Buffer, new_size * sizeof(T));
+	if (new_buffer == NULL)
+		return 10;
+	m_Buffer = static_cast<T*>(new_buffer);
+	set_count(new_size);
+	return 0;
+}
+
+//declare whatever instances of the template
+//you are going to use here:
+template class BufferCache<BPEvent*>;
+
+} /* namespace fi */

core/controller/BufferCache.hpp

@@ -0,0 +1,82 @@
+#ifndef __BUFFERCACHE_HPP__
+#define __BUFFERCACHE_HPP__
+
+#include <stdlib.h>
+
+namespace fi {
+
+/**
+ * \class BufferCache
+ *
+ * \brief A simple dynamic array
+ *
+ * This class is intended to serve as a kind of cache for the entirely STL-based,
+ * untyped and therefore quite slow event handling mechanism of Fail*.
+ * To keep the code easily readable, the buffer management methods
+ * are less performant than the could be (remove() and erase() have linear complexity).
+ */
+template<class T> class BufferCache {
+public:
+	BufferCache()
+		: m_Buffer(NULL), m_Buffer_count(0) {}
+	~BufferCache() {}
+	/**
+	 * Add an element to the array. The object pointed to remains untouched.
+	 * @param val the element to add
+	 * @returns 0 if successful, an error code otherwise (ATM only 10 if malloc() fails)
+	 */
+	int add(T val);
+	/**
+	 * Remove an element from the array. The object pointed to remains untouched.
+	 * @param val the element to remove
+	 * @returns 0 if successful, an error code otherwise (ATM only 10 if malloc() fails)
+	 */
+	int remove(T val);
+	/**
+	 * Remove an element at a specific position. The object pointed to remains untouched.
+	 * @param val the element to remove
+	 * @returns a pointer to the given element's successor if successful, -1 otherwise
+	 */
+	int erase(int i);
+	/**
+	 * Clears the array, removing all elements. The objects pointed to remain untouched.
+	 */
+	void clear();
+	/**
+	 * Retrieve an element from the array. Should be inlined.
+	 * @param idx the position to retrieve the element from
+	 * @returns the element at the given position
+	 */
+	inline T get(size_t idx) { return m_Buffer[idx]; }
+	/**
+	 * Set an element at a given position. Should be inlined.
+	 * @param idx the position to change an element at
+	 * @param val the new value of the given element
+	 */
+	inline void set(size_t idx, T val) { m_Buffer[idx] = val; }
+	/**
+	 * Retrieves the current length of the array. Should be inlined.
+	 * @returns the array length
+	 */
+	inline size_t get_count() { return m_Buffer_count; }
+protected:
+	/**
+	 * Changes the current length of the array. Should be inlined.
+	 * @param new_count the new array length
+	 */
+	inline void set_count(size_t new_count) { m_Buffer_count = new_count; }
+	/**
+	 * Reallocates the buffer. This implementation is extremely primitive,
+	 * but since the amount of entries is small,
+	 * this will not be significant, hopefully. Should be inlined.
+	 * @param new_size the new number of elements in the array
+	 * @returns 0 if successful, an error code otherwise (ATM only 10 if malloc() fails)
+	 */
+	inline int reallocate_buffer(size_t new_size);
+private:
+	T *m_Buffer;
+	size_t m_Buffer_count;
+};
+
+} /* namespace fi */
+#endif /* BUFFERCACHE_H_ */

core/controller/CMakeLists.txt

@@ -1,4 +1,5 @@
 set(SRCS
+	BufferCache.cc
 	CampaignManager.cc
 	CoroutineManager.cc
 	Event.cc

core/controller/EventList.cc

@@ -12,6 +12,10 @@ EventId EventList::add(BaseEvent* ev, ExperimentFlow* pExp)
 	// a zero counter does not make sense
 	assert(ev->getCounter() != 0);
 	ev->setParent(pExp); // event is linked to experiment flow
+
+	BPEvent *bp_ev;
+	if((bp_ev = dynamic_cast<BPEvent*>(ev)) != NULL)
+		m_Bp_cache.add(bp_ev);
 	m_BufferList.push_back(ev);
 	return (ev->getId());
 }
@@ -27,6 +31,7 @@ void EventList::remove(BaseEvent* ev)
 			sal::simulator.onEventDeletion(*it);
 		for (firelist_t::iterator it = m_FireList.begin(); it != m_FireList.end(); it++)
 			sal::simulator.onEventDeletion(*it);
+		m_Bp_cache.clear();
 		m_BufferList.clear();
 		// all remaining active events must not fire anymore
 		m_DeleteList.insert(m_DeleteList.end(), m_FireList.begin(), m_FireList.end());
@@ -36,6 +41,10 @@ void EventList::remove(BaseEvent* ev)
 	//   * if ev in m_FireList, copy to m_DeleteList
 	} else {
 		sal::simulator.onEventDeletion(ev);
+
+		BPEvent *bp_ev;
+		if((bp_ev = dynamic_cast<BPEvent*>(ev)) != NULL)
+			m_Bp_cache.remove(bp_ev);
 		m_BufferList.remove(ev);
 		firelist_t::const_iterator it =
 			std::find(m_FireList.begin(), m_FireList.end(), ev);
@@ -61,6 +70,10 @@ EventList::iterator EventList::m_remove(iterator it, bool skip_deletelist)
 		sal::simulator.onEventDeletion(*it);
 		m_DeleteList.push_back(*it);
 	}
+
+	BPEvent *bp_ev;
+	if((bp_ev = dynamic_cast<BPEvent*>(*it)) != NULL)
+		m_Bp_cache.remove(bp_ev);
 	return (m_BufferList.erase(it));
 }
 
@@ -80,6 +93,7 @@ void EventList::remove(ExperimentFlow* flow)
 		for (bufferlist_t::iterator it = m_BufferList.begin();
 		     it != m_BufferList.end(); it++)
 			sal::simulator.onEventDeletion(*it); // invoke event handler
+		m_Bp_cache.clear();
 		m_BufferList.clear();
 	} else { // remove all events corresponding to a specific experiment ("flow"):
 		for (bufferlist_t::iterator it = m_BufferList.begin();

core/controller/EventList.hpp

@@ -7,6 +7,7 @@
 #include <algorithm>
 
 #include "Event.hpp"
+#include "BufferCache.hpp"
 
 namespace fi
 {
@@ -53,6 +54,7 @@ class EventList
 		deletelist_t m_DeleteList; //!< the deleted events (used temporarily)
 		// TODO: Hashing?
 		BaseEvent* m_pFired; //!< the recently fired Event-object
+		BufferCache<BPEvent*> m_Bp_cache;
 	public:
 		/**
 		 * The iterator of this class used to loop through the list of
@@ -180,6 +182,11 @@ class EventList
 		 * TODO: Improve naming (instead of "fireActiveEvents")?
 		 */
 		void fireActiveEvents();
+		/**
+		 * Retrieves the BPEvent buffer cache.
+		 * @returns the buffer cache
+		 */
+		inline BufferCache<BPEvent*> *getBPBuffer() { return &m_Bp_cache; }
 };
 
 }; // end-of-namespace: fi

core/experiments/l4sys/campaign.cc

@@ -33,7 +33,7 @@ bool L4SysCampaign::run()
 	int count = 0;
 	//iterate over one register
 	for (int bit_offset = 0; bit_offset < 1; ++bit_offset) {
-		for (int instr_offset = 0; instr_offset < COOL_ECC_NUMINSTR; ++instr_offset) {
+		for (int instr_offset = 0; instr_offset < L4SYS_NUMINSTR; ++instr_offset) {
 			L4SysExperimentData *d = new L4SysExperimentData;
 			d->msg.set_instr_offset(instr_offset);
 			d->msg.set_bit_offset(bit_offset);

core/experiments/l4sys/experiment.cc

@@ -1,4 +1,5 @@
 #include <iostream>
+#include <map>
 #include <sys/types.h>
 #include <sys/stat.h>
 #include <unistd.h>
@@ -33,7 +34,7 @@ using std::endl;
 char const * const state_folder = "l4sys.state";
 char const * const instr_list_fn = "ip.list";
 char const * const golden_run_fn = "golden.out";
-sal::address_t const aspace = 0x1fd44000;
+sal::address_t const aspace = 0x01e00000;
 std::string output;
 std::vector<sal::address_t> instr_list;
 std::string golden_run;
@@ -80,14 +81,12 @@ bool L4SysExperiment::run() {
 
 	log << "startup" << endl;
 
-	//FIXME: this is a race condition:
-	//only one L4SysExperiment instance should execute this block
-	//at a time
+#ifdef PREPARE_EXPERIMENT
 
 	struct stat teststruct;
 	// STEP 1: run until interesting function starts, and save state
 	if (stat(state_folder, &teststruct) == -1) {
-		bp.setWatchInstructionPointer(COOL_ECC_FUNC_ENTRY);
+		bp.setWatchInstructionPointer(L4SYS_FUNC_ENTRY);
 		sal::simulator.addEventAndWait(&bp);
 
 		log << "test function entry reached, saving state" << endl;
@@ -112,7 +111,7 @@ bool L4SysExperiment::run() {
 		std::ofstream instr_list_file(instr_list_fn);
 		instr_list_file << std::hex;
 		bp.setWatchInstructionPointer(fi::ANY_ADDR);
-		while (bp.getTriggerInstructionPointer() != COOL_ECC_CALCDONE) {
+		while (bp.getTriggerInstructionPointer() != L4SYS_FUNC_EXIT) {
 			sal::simulator.addEventAndWait(&bp);
 			//short sanity check
 			sal::address_t curr_instr = bp.getTriggerInstructionPointer();
@@ -146,7 +145,7 @@ bool L4SysExperiment::run() {
 		sal::simulator.addSuppressedInterrupt(32);
 
 		std::ofstream golden_run_file(golden_run_fn);
-		bp.setWatchInstructionPointer(COOL_ECC_CALCDONE);
+		bp.setWatchInstructionPointer(L4SYS_FUNC_EXIT);
 		bp.setCounter(times_run);
 		sal::simulator.addEvent(&bp);
 		fi::BaseEvent* ev = waitIOOrOther(true);
@@ -183,10 +182,10 @@ bool L4SysExperiment::run() {
 		output.reserve(flen);
 	}
 
-	//end of critical section
+#endif
 
 	// STEP 4: The actual experiment.
-	for (int i = 0; i < COOL_ECC_NUMINSTR; i++) {
+	for (int i = 0; i < 1/*L4SYS_NUMINSTR*/; i++) {
 		log << "restoring state" << endl;
 		sal::simulator.restore(state_folder);
 
@@ -195,7 +194,7 @@ bool L4SysExperiment::run() {
 		if (!m_jc.getParam(param)) {
 			log << "Dying." << endl;
 			// communicate that we were told to die
-			sal::simulator.terminate(1); // "return (false);" ?
+			sal::simulator.terminate(1);
 		}
 		int id = param.getWorkloadID();
 		int instr_offset = param.msg.instr_offset();
@@ -239,10 +238,10 @@ bool L4SysExperiment::run() {
 		}
 
 		// aftermath
-		fi::BPSingleEvent ev_done(COOL_ECC_CALCDONE, aspace);
+		fi::BPSingleEvent ev_done(L4SYS_FUNC_EXIT, aspace);
 		ev_done.setCounter(times_run);
 		sal::simulator.addEvent(&ev_done);
-		const unsigned instr_run = times_run * COOL_ECC_NUMINSTR;
+		const unsigned instr_run = times_run * L4SYS_NUMINSTR;
 		fi::BPSingleEvent ev_timeout(fi::ANY_ADDR, aspace);
 		ev_timeout.setCounter(instr_run + 3000);
 		sal::simulator.addEvent(&ev_timeout);
@@ -264,7 +263,7 @@ bool L4SysExperiment::run() {
 		if (ev == &ev_done) {
 			if (strcmp(output.c_str(), golden_run.c_str()) == 0) {
 				log << std::dec << "Result DONE" << endl;
-				param.msg.set_resulttype(param.msg.CALCDONE);
+				param.msg.set_resulttype(param.msg.DONE);
 			} else {
 				log << std::dec << "Result WRONG" << endl;
 				param.msg.set_resulttype(param.msg.WRONG);

core/experiments/l4sys/experimentInfo.hpp

@@ -3,9 +3,10 @@
 
 // FIXME autogenerate this
 
-#define COOL_ECC_FUNC_ENTRY		0x1007cd0
-#define COOL_ECC_CALCDONE		0x1007d3a
-#define COOL_ECC_NUMINSTR		3166
-#define HEADLESS_EXPERIMENT
+#define L4SYS_FUNC_ENTRY		0x1007cd0
+#define L4SYS_FUNC_EXIT			0x1007d3a
+#define L4SYS_NUMINSTR			3184
+//#define HEADLESS_EXPERIMENT
+#define PREPARE_EXPERIMENT
 
 #endif

core/experiments/l4sys/l4sys.proto

@@ -6,7 +6,7 @@ message L4SysProtoMsg {
 	// results
 	// make these optional to reduce overhead for server->client communication
 	enum ResultType {
-		CALCDONE = 1;
+		DONE = 1;
 		TIMEOUT = 2;
 		TRAP = 3;
 		INTR = 4;

core/jobserver/JobServer.cc

@@ -45,7 +45,7 @@ ExperimentData *JobServer::getDone()
 	 && m_doneJobs.Size() == 0) {
 		// FRICKEL workaround
 		sleep(1);
-		ExperimentData *exp;
+		ExperimentData *exp = NULL;
 		if (m_doneJobs.Dequeue_nb(exp)) {
 			return exp;
 		}