X-Git-Url: http://plrg.eecs.uci.edu/git/?p=c11tester.git;a=blobdiff_plain;f=snapshot.cc;h=0d38edc1c1cbc7ee0b34d393aeee50baa2e05ee8;hp=e006189c44ba1741208183ea8cbddd222b381dfd;hb=251ac4b4bf3a9f2d3cfacc1e6618200ca1c431ac;hpb=b5000a06086de6ea8799168d463f018cab785830 diff --git a/snapshot.cc b/snapshot.cc index e006189c..0d38edc1 100644 --- a/snapshot.cc +++ b/snapshot.cc @@ -3,339 +3,218 @@ #include #include #include -#include "hashtable.h" -#include -#include -#include "snapshot.h" -#include "snapshotimp.h" -#include "mymemory.h" -#include -#include -#include -#include +#include #include #include -#include +#include "hashtable.h" +#include "snapshot.h" +#include "mymemory.h" #include "common.h" +#include "context.h" +#include "model.h" -#define FAILURE(mesg) { printf("failed in the API: %s with errno relative message: %s\n", mesg, strerror( errno ) ); exit(EXIT_FAILURE); } -#ifdef CONFIG_SSDEBUG -#define SSDEBUG printf -#else -#define SSDEBUG(...) do { } while (0) -#endif +#define SHARED_MEMORY_DEFAULT (200 * ((size_t)1 << 20)) // 100mb for the shared memory +#define STACK_SIZE_DEFAULT (((size_t)1 << 20) * 20) // 20 mb out of the above 100 mb for my stack -/* extern declaration definition */ -struct SnapShot * snapshotrecord = NULL; +struct fork_snapshotter { + /** @brief Pointer to the shared (non-snapshot) memory heap base + * (NOTE: this has size SHARED_MEMORY_DEFAULT - sizeof(*fork_snap)) */ + void *mSharedMemoryBase; -/** PageAlignedAdressUpdate return a page aligned address for the - * address being added as a side effect the numBytes are also changed. - */ -static void * PageAlignAddressUpward(void * addr) { - return (void *)((((uintptr_t)addr)+PAGESIZE-1)&~(PAGESIZE-1)); -} + /** @brief Pointer to the shared (non-snapshot) stack region */ + void *mStackBase; -#if !USE_MPROTECT_SNAPSHOT -/** @statics -* These variables are necessary because the stack is shared region and -* there exists a race between all processes executing the same function. -* To avoid the problem above, we require variables allocated in 'safe' regions. -* The bug was actually observed with the forkID, these variables below are -* used to indicate the various contexts to which to switch to. -* -* @savedSnapshotContext: contains the point to which takesnapshot() call should switch to. -* @savedUserSnapshotContext: contains the point to which the process whose snapshotid is equal to the rollbackid should switch to -* @snapshotid: it is a running counter for the various forked processes snapshotid. it is incremented and set in a persistently shared record -*/ -static ucontext_t savedSnapshotContext; -static ucontext_t savedUserSnapshotContext; -static snapshot_id snapshotid = 0; + /** @brief Size of the shared stack */ + size_t mStackSize; -#else /* USE_MPROTECT_SNAPSHOT */ + /** + * @brief Stores the ID that we are attempting to roll back to + * + * Used in inter-process communication so that each process can + * determine whether or not to take over execution (w/ matching ID) or + * exit (we're rolling back even further). Dubiously marked 'volatile' + * to prevent compiler optimizations from messing with the + * inter-process behavior. + */ + volatile snapshot_id mIDToRollback; -/** ReturnPageAlignedAddress returns a page aligned address for the - * address being added as a side effect the numBytes are also changed. - */ -static void * ReturnPageAlignedAddress(void * addr) { - return (void *)(((uintptr_t)addr)&~(PAGESIZE-1)); -} -/** The initSnapShotRecord method initialized the snapshotting data - * structures for the mprotect based snapshot. - */ -static void initSnapShotRecord(unsigned int numbackingpages, unsigned int numsnapshots, unsigned int nummemoryregions) { - snapshotrecord=( struct SnapShot * )model_malloc(sizeof(struct SnapShot)); - snapshotrecord->regionsToSnapShot=( struct MemoryRegion * )model_malloc(sizeof(struct MemoryRegion)*nummemoryregions); - snapshotrecord->backingStoreBasePtr= ( struct SnapShotPage * )model_malloc( sizeof( struct SnapShotPage ) * (numbackingpages + 1) ); - //Page align the backingstorepages - snapshotrecord->backingStore=( struct SnapShotPage * )PageAlignAddressUpward(snapshotrecord->backingStoreBasePtr); - snapshotrecord->backingRecords=( struct BackingPageRecord * )model_malloc(sizeof(struct BackingPageRecord)*numbackingpages); - snapshotrecord->snapShots= ( struct SnapShotRecord * )model_malloc(sizeof(struct SnapShotRecord)*numsnapshots); - snapshotrecord->lastSnapShot=0; - snapshotrecord->lastBackingPage=0; - snapshotrecord->lastRegion=0; - snapshotrecord->maxRegions=nummemoryregions; - snapshotrecord->maxBackingPages=numbackingpages; - snapshotrecord->maxSnapShots=numsnapshots; -} -/** HandlePF is the page fault handler for mprotect based snapshotting - * algorithm. + /** @brief Inter-process tracking of the next snapshot ID */ + snapshot_id currSnapShotID; +}; + +static struct fork_snapshotter *fork_snap = NULL; +ucontext_t shared_ctxt; + +/** @statics + * These variables are necessary because the stack is shared region and + * there exists a race between all processes executing the same function. + * To avoid the problem above, we require variables allocated in 'safe' regions. + * The bug was actually observed with the forkID, these variables below are + * used to indicate the various contexts to which to switch to. + * + * @private_ctxt: the context which is internal to the current process. Used + * for running the internal snapshot/rollback loop. + * @exit_ctxt: a special context used just for exiting from a process (so we + * can use swapcontext() instead of setcontext() + hacks) + * @snapshotid: it is a running counter for the various forked processes + * snapshotid. it is incremented and set in a persistently shared record */ -static void HandlePF( int sig, siginfo_t *si, void * unused){ - if( si->si_code == SEGV_MAPERR ){ - printf("Real Fault at %p\n", si->si_addr); - print_trace(); - exit( EXIT_FAILURE ); - } - void* addr = ReturnPageAlignedAddress(si->si_addr); +static ucontext_t private_ctxt; +static snapshot_id snapshotid = 0; - unsigned int backingpage=snapshotrecord->lastBackingPage++; //Could run out of pages... - if (backingpage==snapshotrecord->maxBackingPages) { - printf("Out of backing pages at %p\n", si->si_addr); - exit( EXIT_FAILURE ); - } +/** + * @brief Create a new context, with a given stack and entry function + * @param ctxt The context structure to fill + * @param stack The stack to run the new context in + * @param stacksize The size of the stack + * @param func The entry point function for the context + */ +static void create_context(ucontext_t *ctxt, void *stack, size_t stacksize, + void (*func)(void)) +{ + getcontext(ctxt); + ctxt->uc_stack.ss_sp = stack; + ctxt->uc_stack.ss_size = stacksize; + ctxt->uc_link = NULL; + makecontext(ctxt, func, 0); +} - //copy page - memcpy(&(snapshotrecord->backingStore[backingpage]), addr, sizeof(struct SnapShotPage)); - //remember where to copy page back to - snapshotrecord->backingRecords[backingpage].basePtrOfPage=addr; - //set protection to read/write - if (mprotect( addr, sizeof(struct SnapShotPage), PROT_READ | PROT_WRITE )) { - perror("mprotect"); - // Handle error by quitting? - } +/** @brief An empty function, used for an "empty" context which just exits a + * process */ +static void fork_exit() +{ + _Exit(EXIT_SUCCESS); } -#endif /* USE_MPROTECT_SNAPSHOT */ -#if !USE_MPROTECT_SNAPSHOT -void createSharedMemory(){ +static void createSharedMemory() +{ //step 1. create shared memory. - void * memMapBase = mmap( 0, SHARED_MEMORY_DEFAULT + STACK_SIZE_DEFAULT, PROT_READ | PROT_WRITE, MAP_SHARED|MAP_ANON, -1, 0 ); - if( MAP_FAILED == memMapBase ) - FAILURE("mmap"); + void *memMapBase = mmap(0, SHARED_MEMORY_DEFAULT + STACK_SIZE_DEFAULT, PROT_READ | PROT_WRITE, MAP_SHARED | MAP_ANON, -1, 0); + if (memMapBase == MAP_FAILED) { + perror("mmap"); + exit(EXIT_FAILURE); + } //Setup snapshot record at top of free region - snapshotrecord = ( struct SnapShot * )memMapBase; - snapshotrecord->mSharedMemoryBase = (void *)((uintptr_t)memMapBase + sizeof(struct SnapShot)); - snapshotrecord->mStackBase = (void *)((uintptr_t)memMapBase + SHARED_MEMORY_DEFAULT); - snapshotrecord->mStackSize = STACK_SIZE_DEFAULT; - snapshotrecord->mIDToRollback = -1; - snapshotrecord->currSnapShotID = 0; + fork_snap = (struct fork_snapshotter *)memMapBase; + fork_snap->mSharedMemoryBase = (void *)((uintptr_t)memMapBase + sizeof(*fork_snap)); + fork_snap->mStackBase = (void *)((uintptr_t)memMapBase + SHARED_MEMORY_DEFAULT); + fork_snap->mStackSize = STACK_SIZE_DEFAULT; + fork_snap->mIDToRollback = -1; + fork_snap->currSnapShotID = 0; } -#endif - -/** The initSnapshotLibrary function initializes the snapshot library. - * @param entryPoint the function that should run the program. +/** + * Create a new mspace pointer for the non-snapshotting (i.e., inter-process + * shared) memory region. Only for fork-based snapshotting. + * + * @return The shared memory mspace */ -#if USE_MPROTECT_SNAPSHOT - -void initSnapshotLibrary(unsigned int numbackingpages, - unsigned int numsnapshots, unsigned int nummemoryregions, - unsigned int numheappages, VoidFuncPtr entryPoint) { - /* Setup a stack for our signal handler.... */ - stack_t ss; - ss.ss_sp = PageAlignAddressUpward(model_malloc(SIGSTACKSIZE+PAGESIZE-1)); - ss.ss_size = SIGSTACKSIZE; - ss.ss_flags = 0; - sigaltstack(&ss, NULL); - - struct sigaction sa; - sa.sa_flags = SA_SIGINFO | SA_NODEFER | SA_RESTART | SA_ONSTACK; - sigemptyset( &sa.sa_mask ); - sa.sa_sigaction = HandlePF; -#ifdef MAC - if( sigaction( SIGBUS, &sa, NULL ) == -1 ){ - printf("SIGACTION CANNOT BE INSTALLED\n"); - exit(EXIT_FAILURE); - } -#endif - if( sigaction( SIGSEGV, &sa, NULL ) == -1 ){ - printf("SIGACTION CANNOT BE INSTALLED\n"); - exit(EXIT_FAILURE); - } - - initSnapShotRecord(numbackingpages, numsnapshots, nummemoryregions); - - // EVIL HACK: We need to make sure that calls into the HandlePF method don't cause dynamic links - // The problem is that we end up protecting state in the dynamic linker... - // Solution is to call our signal handler before we start protecting stuff... - - siginfo_t si; - memset(&si, 0, sizeof(si)); - si.si_addr=ss.ss_sp; - HandlePF(SIGSEGV, &si, NULL); - snapshotrecord->lastBackingPage--; //remove the fake page we copied - - void *basemySpace = model_malloc((numheappages+1)*PAGESIZE); - void * pagealignedbase=PageAlignAddressUpward(basemySpace); - user_snapshot_space = create_mspace_with_base(pagealignedbase, numheappages * PAGESIZE, 1); - addMemoryRegionToSnapShot(pagealignedbase, numheappages); - - void *base_model_snapshot_space = model_malloc((numheappages + 1) * PAGESIZE); - pagealignedbase = PageAlignAddressUpward(base_model_snapshot_space); - model_snapshot_space = create_mspace_with_base(pagealignedbase, numheappages * PAGESIZE, 1); - addMemoryRegionToSnapShot(pagealignedbase, numheappages); - - entryPoint(); +mspace create_shared_mspace() +{ + if (!fork_snap) + createSharedMemory(); + return create_mspace_with_base((void *)(fork_snap->mSharedMemoryBase), SHARED_MEMORY_DEFAULT - sizeof(*fork_snap), 1); } -#else -void initSnapshotLibrary(unsigned int numbackingpages, - unsigned int numsnapshots, unsigned int nummemoryregions, - unsigned int numheappages, VoidFuncPtr entryPoint) { - if (!snapshotrecord) + +static void fork_snapshot_init(unsigned int numbackingpages, + unsigned int numsnapshots, unsigned int nummemoryregions, + unsigned int numheappages) +{ + if (!fork_snap) createSharedMemory(); - void *base_model_snapshot_space = malloc((numheappages + 1) * PAGESIZE); - void *pagealignedbase = PageAlignAddressUpward(base_model_snapshot_space); - model_snapshot_space = create_mspace_with_base(pagealignedbase, numheappages * PAGESIZE, 1); + model_snapshot_space = create_mspace(numheappages * PAGESIZE, 1); +} - //step 2 setup the stack context. - ucontext_t newContext; - getcontext( &newContext ); - newContext.uc_stack.ss_sp = snapshotrecord->mStackBase; - newContext.uc_stack.ss_size = STACK_SIZE_DEFAULT; - makecontext( &newContext, entryPoint, 0 ); - /* switch to a new entryPoint context, on a new stack */ - swapcontext(&savedSnapshotContext, &newContext); +volatile int modellock = 0; +static void fork_loop() { /* switch back here when takesnapshot is called */ - pid_t forkedID = 0; - snapshotid = snapshotrecord->currSnapShotID; - /* This bool indicates that the current process's snapshotid is same - as the id to which the rollback needs to occur */ - - bool rollback = false; - while( true ){ - snapshotrecord->currSnapShotID=snapshotid+1; - forkedID = fork(); - - if( 0 == forkedID ){ - /* If the rollback bool is set, switch to the context we need to - return to during a rollback. */ - if( rollback) { - setcontext( &( snapshotrecord->mContextToRollback ) ); - } else { - /*Child process which is forked as a result of takesnapshot - call should switch back to the takesnapshot context*/ - setcontext( &savedUserSnapshotContext ); - } - } else { - int status; - int retVal; + snapshotid = fork_snap->currSnapShotID; + if (model->params.nofork) { + setcontext(&shared_ctxt); + _Exit(EXIT_SUCCESS); + } - SSDEBUG("The process id of child is %d and the process id of this process is %d and snapshot id is %d\n", - forkedID, getpid(), snapshotid ); + while (true) { + pid_t forkedID; + fork_snap->currSnapShotID = snapshotid + 1; - do { - retVal=waitpid( forkedID, &status, 0 ); - } while( -1 == retVal && errno == EINTR ); + modellock = 1; + forkedID = fork(); + modellock = 0; - if( snapshotrecord->mIDToRollback != snapshotid ) { - exit(EXIT_SUCCESS); + if (0 == forkedID) { + setcontext(&shared_ctxt); + } else { + DEBUG("parent PID: %d, child PID: %d, snapshot ID: %d\n", + getpid(), forkedID, snapshotid); + + while (waitpid(forkedID, NULL, 0) < 0) { + /* waitpid() may be interrupted */ + if (errno != EINTR) { + perror("waitpid"); + exit(EXIT_FAILURE); + } } - rollback = true; + + if (fork_snap->mIDToRollback != snapshotid) + _Exit(EXIT_SUCCESS); } } } -#endif -/** The addMemoryRegionToSnapShot function assumes that addr is page aligned. +static void fork_startExecution() { + /* switch to a new entryPoint context, on a new stack */ + create_context(&private_ctxt, snapshot_calloc(STACK_SIZE_DEFAULT, 1), STACK_SIZE_DEFAULT, fork_loop); +} + +static snapshot_id fork_take_snapshot() { + model_swapcontext(&shared_ctxt, &private_ctxt); + DEBUG("TAKESNAPSHOT RETURN\n"); + fork_snap->mIDToRollback = -1; + return snapshotid; +} + +static void fork_roll_back(snapshot_id theID) +{ + DEBUG("Rollback\n"); + fork_snap->mIDToRollback = theID; + fork_exit(); +} + +/** + * @brief Initializes the snapshot system + * @param entryPoint the function that should run the program. */ -void addMemoryRegionToSnapShot( void * addr, unsigned int numPages) { -#if USE_MPROTECT_SNAPSHOT - unsigned int memoryregion=snapshotrecord->lastRegion++; - if (memoryregion==snapshotrecord->maxRegions) { - printf("Exceeded supported number of memory regions!\n"); - exit(EXIT_FAILURE); - } +void snapshot_system_init(unsigned int numbackingpages, + unsigned int numsnapshots, unsigned int nummemoryregions, + unsigned int numheappages) +{ + fork_snapshot_init(numbackingpages, numsnapshots, nummemoryregions, numheappages); +} - snapshotrecord->regionsToSnapShot[ memoryregion ].basePtr=addr; - snapshotrecord->regionsToSnapShot[ memoryregion ].sizeInPages=numPages; -#endif //NOT REQUIRED IN THE CASE OF FORK BASED SNAPSHOTS. +void startExecution() { + fork_startExecution(); } -/** The takeSnapshot function takes a snapshot. +/** Takes a snapshot of memory. * @return The snapshot identifier. */ -snapshot_id takeSnapshot( ){ -#if USE_MPROTECT_SNAPSHOT - for(unsigned int region=0; regionlastRegion;region++) { - if( mprotect(snapshotrecord->regionsToSnapShot[region].basePtr, snapshotrecord->regionsToSnapShot[region].sizeInPages*sizeof(struct SnapShotPage), PROT_READ ) == -1 ){ - perror("mprotect"); - printf("Failed to mprotect inside of takeSnapShot\n"); - exit(EXIT_FAILURE); - } - } - unsigned int snapshot=snapshotrecord->lastSnapShot++; - if (snapshot==snapshotrecord->maxSnapShots) { - printf("Out of snapshots\n"); - exit(EXIT_FAILURE); - } - snapshotrecord->snapShots[snapshot].firstBackingPage=snapshotrecord->lastBackingPage; - - return snapshot; -#else - swapcontext( &savedUserSnapshotContext, &savedSnapshotContext ); - SSDEBUG("TAKESNAPSHOT RETURN\n"); - return snapshotid; -#endif +snapshot_id take_snapshot() +{ + return fork_take_snapshot(); } -/** The rollBack function rollback to the given snapshot identifier. +/** Rolls the memory state back to the given snapshot identifier. * @param theID is the snapshot identifier to rollback to. */ -void rollBack( snapshot_id theID ){ -#if USE_MPROTECT_SNAPSHOT==2 - if (snapshotrecord->lastSnapShot==(theID+1)) { - for(unsigned int page=snapshotrecord->snapShots[theID].firstBackingPage; pagelastBackingPage; page++) { - memcpy(snapshotrecord->backingRecords[page].basePtrOfPage, &snapshotrecord->backingStore[page], sizeof(struct SnapShotPage)); - } - return; - } -#endif - -#if USE_MPROTECT_SNAPSHOT - HashTable< void *, bool, uintptr_t, 4, model_malloc, model_calloc, model_free> duplicateMap; - for(unsigned int region=0; regionlastRegion;region++) { - if( mprotect(snapshotrecord->regionsToSnapShot[region].basePtr, snapshotrecord->regionsToSnapShot[region].sizeInPages*sizeof(struct SnapShotPage), PROT_READ | PROT_WRITE ) == -1 ){ - perror("mprotect"); - printf("Failed to mprotect inside of takeSnapShot\n"); - exit(EXIT_FAILURE); - } - } - for(unsigned int page=snapshotrecord->snapShots[theID].firstBackingPage; pagelastBackingPage; page++) { - if( !duplicateMap.contains(snapshotrecord->backingRecords[page].basePtrOfPage )) { - duplicateMap.put(snapshotrecord->backingRecords[page].basePtrOfPage, true); - memcpy(snapshotrecord->backingRecords[page].basePtrOfPage, &snapshotrecord->backingStore[page], sizeof(struct SnapShotPage)); - } - } - snapshotrecord->lastSnapShot=theID; - snapshotrecord->lastBackingPage=snapshotrecord->snapShots[theID].firstBackingPage; - takeSnapshot(); //Make sure current snapshot is still good...All later ones are cleared -#else - snapshotrecord->mIDToRollback = theID; - volatile int sTemp = 0; - getcontext( &snapshotrecord->mContextToRollback ); - /* - * This is used to quit the process on rollback, so that the process - * which needs to rollback can quit allowing the process whose - * snapshotid matches the rollbackid to switch to this context and - * continue.... - */ - if( !sTemp ){ - sTemp = 1; - SSDEBUG("Invoked rollback\n"); - exit(EXIT_SUCCESS); - } - /* - * This fix obviates the need for a finalize call. hence less dependences for model-checker.... - * - */ - snapshotrecord->mIDToRollback = -1; -#endif +void snapshot_roll_back(snapshot_id theID) +{ + fork_roll_back(theID); } -