11 #include "hashtable.h"
13 #include "snapshotimp.h"
17 #define FAILURE(mesg) { model_print("failed in the API: %s with errno relative message: %s\n", mesg, strerror(errno)); exit(EXIT_FAILURE); }
20 #define SSDEBUG model_print
22 #define SSDEBUG(...) do { } while (0)
25 /* extern declaration definition */
26 struct SnapShot *snapshotrecord = NULL;
28 /** PageAlignedAdressUpdate return a page aligned address for the
29 * address being added as a side effect the numBytes are also changed.
31 static void * PageAlignAddressUpward(void *addr)
33 return (void *)((((uintptr_t)addr) + PAGESIZE - 1) & ~(PAGESIZE - 1));
36 #if !USE_MPROTECT_SNAPSHOT
38 * These variables are necessary because the stack is shared region and
39 * there exists a race between all processes executing the same function.
40 * To avoid the problem above, we require variables allocated in 'safe' regions.
41 * The bug was actually observed with the forkID, these variables below are
42 * used to indicate the various contexts to which to switch to.
44 * @savedSnapshotContext: contains the point to which takesnapshot() call should switch to.
45 * @savedUserSnapshotContext: contains the point to which the process whose snapshotid is equal to the rollbackid should switch to
46 * @snapshotid: it is a running counter for the various forked processes snapshotid. it is incremented and set in a persistently shared record
48 static ucontext_t savedSnapshotContext;
49 static ucontext_t savedUserSnapshotContext;
50 static snapshot_id snapshotid = 0;
52 #else /* USE_MPROTECT_SNAPSHOT */
54 /** ReturnPageAlignedAddress returns a page aligned address for the
55 * address being added as a side effect the numBytes are also changed.
57 static void * ReturnPageAlignedAddress(void *addr)
59 return (void *)(((uintptr_t)addr) & ~(PAGESIZE - 1));
62 /** The initSnapShotRecord method initialized the snapshotting data
63 * structures for the mprotect based snapshot.
65 static void initSnapShotRecord(unsigned int numbackingpages, unsigned int numsnapshots, unsigned int nummemoryregions)
67 snapshotrecord = (struct SnapShot *)model_malloc(sizeof(struct SnapShot));
68 snapshotrecord->regionsToSnapShot = (struct MemoryRegion *)model_malloc(sizeof(struct MemoryRegion) * nummemoryregions);
69 snapshotrecord->backingStoreBasePtr = (void *)model_malloc(sizeof(snapshot_page_t) * (numbackingpages + 1));
70 //Page align the backingstorepages
71 snapshotrecord->backingStore = (snapshot_page_t *)PageAlignAddressUpward(snapshotrecord->backingStoreBasePtr);
72 snapshotrecord->backingRecords = (struct BackingPageRecord *)model_malloc(sizeof(struct BackingPageRecord) * numbackingpages);
73 snapshotrecord->snapShots = (struct SnapShotRecord *)model_malloc(sizeof(struct SnapShotRecord) * numsnapshots);
74 snapshotrecord->lastSnapShot = 0;
75 snapshotrecord->lastBackingPage = 0;
76 snapshotrecord->lastRegion = 0;
77 snapshotrecord->maxRegions = nummemoryregions;
78 snapshotrecord->maxBackingPages = numbackingpages;
79 snapshotrecord->maxSnapShots = numsnapshots;
82 /** HandlePF is the page fault handler for mprotect based snapshotting
85 static void HandlePF(int sig, siginfo_t *si, void *unused)
87 if (si->si_code == SEGV_MAPERR) {
88 model_print("Real Fault at %p\n", si->si_addr);
90 model_print("For debugging, place breakpoint at: %s:%d\n",
94 void* addr = ReturnPageAlignedAddress(si->si_addr);
96 unsigned int backingpage = snapshotrecord->lastBackingPage++; //Could run out of pages...
97 if (backingpage == snapshotrecord->maxBackingPages) {
98 model_print("Out of backing pages at %p\n", si->si_addr);
103 memcpy(&(snapshotrecord->backingStore[backingpage]), addr, sizeof(snapshot_page_t));
104 //remember where to copy page back to
105 snapshotrecord->backingRecords[backingpage].basePtrOfPage = addr;
106 //set protection to read/write
107 if (mprotect(addr, sizeof(snapshot_page_t), PROT_READ | PROT_WRITE)) {
109 // Handle error by quitting?
112 #endif /* USE_MPROTECT_SNAPSHOT */
114 #if !USE_MPROTECT_SNAPSHOT
115 void createSharedMemory()
117 //step 1. create shared memory.
118 void *memMapBase = mmap(0, SHARED_MEMORY_DEFAULT + STACK_SIZE_DEFAULT, PROT_READ | PROT_WRITE, MAP_SHARED | MAP_ANON, -1, 0);
119 if (MAP_FAILED == memMapBase)
122 //Setup snapshot record at top of free region
123 snapshotrecord = (struct SnapShot *)memMapBase;
124 snapshotrecord->mSharedMemoryBase = (void *)((uintptr_t)memMapBase + sizeof(struct SnapShot));
125 snapshotrecord->mStackBase = (void *)((uintptr_t)memMapBase + SHARED_MEMORY_DEFAULT);
126 snapshotrecord->mStackSize = STACK_SIZE_DEFAULT;
127 snapshotrecord->mIDToRollback = -1;
128 snapshotrecord->currSnapShotID = 0;
133 /** The initSnapshotLibrary function initializes the snapshot library.
134 * @param entryPoint the function that should run the program.
136 #if USE_MPROTECT_SNAPSHOT
138 void initSnapshotLibrary(unsigned int numbackingpages,
139 unsigned int numsnapshots, unsigned int nummemoryregions,
140 unsigned int numheappages, VoidFuncPtr entryPoint)
142 /* Setup a stack for our signal handler.... */
144 ss.ss_sp = PageAlignAddressUpward(model_malloc(SIGSTACKSIZE + PAGESIZE - 1));
145 ss.ss_size = SIGSTACKSIZE;
147 sigaltstack(&ss, NULL);
150 sa.sa_flags = SA_SIGINFO | SA_NODEFER | SA_RESTART | SA_ONSTACK;
151 sigemptyset(&sa.sa_mask);
152 sa.sa_sigaction = HandlePF;
154 if (sigaction(SIGBUS, &sa, NULL) == -1) {
155 model_print("SIGACTION CANNOT BE INSTALLED\n");
159 if (sigaction(SIGSEGV, &sa, NULL) == -1) {
160 model_print("SIGACTION CANNOT BE INSTALLED\n");
164 initSnapShotRecord(numbackingpages, numsnapshots, nummemoryregions);
166 // EVIL HACK: We need to make sure that calls into the HandlePF method don't cause dynamic links
167 // The problem is that we end up protecting state in the dynamic linker...
168 // Solution is to call our signal handler before we start protecting stuff...
171 memset(&si, 0, sizeof(si));
172 si.si_addr = ss.ss_sp;
173 HandlePF(SIGSEGV, &si, NULL);
174 snapshotrecord->lastBackingPage--; //remove the fake page we copied
176 void *basemySpace = model_malloc((numheappages + 1) * PAGESIZE);
177 void *pagealignedbase = PageAlignAddressUpward(basemySpace);
178 user_snapshot_space = create_mspace_with_base(pagealignedbase, numheappages * PAGESIZE, 1);
179 addMemoryRegionToSnapShot(pagealignedbase, numheappages);
181 void *base_model_snapshot_space = model_malloc((numheappages + 1) * PAGESIZE);
182 pagealignedbase = PageAlignAddressUpward(base_model_snapshot_space);
183 model_snapshot_space = create_mspace_with_base(pagealignedbase, numheappages * PAGESIZE, 1);
184 addMemoryRegionToSnapShot(pagealignedbase, numheappages);
189 void initSnapshotLibrary(unsigned int numbackingpages,
190 unsigned int numsnapshots, unsigned int nummemoryregions,
191 unsigned int numheappages, VoidFuncPtr entryPoint)
194 createSharedMemory();
196 void *base_model_snapshot_space = malloc((numheappages + 1) * PAGESIZE);
197 void *pagealignedbase = PageAlignAddressUpward(base_model_snapshot_space);
198 model_snapshot_space = create_mspace_with_base(pagealignedbase, numheappages * PAGESIZE, 1);
200 //step 2 setup the stack context.
201 ucontext_t newContext;
202 getcontext(&newContext);
203 newContext.uc_stack.ss_sp = snapshotrecord->mStackBase;
204 newContext.uc_stack.ss_size = STACK_SIZE_DEFAULT;
205 makecontext(&newContext, entryPoint, 0);
206 /* switch to a new entryPoint context, on a new stack */
207 swapcontext(&savedSnapshotContext, &newContext);
209 /* switch back here when takesnapshot is called */
211 snapshotid = snapshotrecord->currSnapShotID;
212 /* This bool indicates that the current process's snapshotid is same
213 as the id to which the rollback needs to occur */
215 bool rollback = false;
217 snapshotrecord->currSnapShotID = snapshotid + 1;
221 /* If the rollback bool is set, switch to the context we need to
222 return to during a rollback. */
224 setcontext(&(snapshotrecord->mContextToRollback));
226 /*Child process which is forked as a result of takesnapshot
227 call should switch back to the takesnapshot context*/
228 setcontext(&savedUserSnapshotContext);
234 SSDEBUG("The process id of child is %d and the process id of this process is %d and snapshot id is %d\n",
235 forkedID, getpid(), snapshotid);
238 retVal = waitpid(forkedID, &status, 0);
239 } while (-1 == retVal && errno == EINTR);
241 if (snapshotrecord->mIDToRollback != snapshotid) {
250 /** The addMemoryRegionToSnapShot function assumes that addr is page aligned.
252 void addMemoryRegionToSnapShot(void *addr, unsigned int numPages)
254 #if USE_MPROTECT_SNAPSHOT
255 unsigned int memoryregion = snapshotrecord->lastRegion++;
256 if (memoryregion == snapshotrecord->maxRegions) {
257 model_print("Exceeded supported number of memory regions!\n");
261 snapshotrecord->regionsToSnapShot[memoryregion].basePtr = addr;
262 snapshotrecord->regionsToSnapShot[memoryregion].sizeInPages = numPages;
263 #endif //NOT REQUIRED IN THE CASE OF FORK BASED SNAPSHOTS.
266 /** The takeSnapshot function takes a snapshot.
267 * @return The snapshot identifier.
269 snapshot_id takeSnapshot()
271 #if USE_MPROTECT_SNAPSHOT
272 for (unsigned int region = 0; region < snapshotrecord->lastRegion; region++) {
273 if (mprotect(snapshotrecord->regionsToSnapShot[region].basePtr, snapshotrecord->regionsToSnapShot[region].sizeInPages * sizeof(snapshot_page_t), PROT_READ) == -1) {
275 model_print("Failed to mprotect inside of takeSnapShot\n");
279 unsigned int snapshot = snapshotrecord->lastSnapShot++;
280 if (snapshot == snapshotrecord->maxSnapShots) {
281 model_print("Out of snapshots\n");
284 snapshotrecord->snapShots[snapshot].firstBackingPage = snapshotrecord->lastBackingPage;
288 swapcontext(&savedUserSnapshotContext, &savedSnapshotContext);
289 SSDEBUG("TAKESNAPSHOT RETURN\n");
294 /** The rollBack function rollback to the given snapshot identifier.
295 * @param theID is the snapshot identifier to rollback to.
297 void rollBack(snapshot_id theID)
299 #if USE_MPROTECT_SNAPSHOT == 2
300 if (snapshotrecord->lastSnapShot == (theID + 1)) {
301 for (unsigned int page = snapshotrecord->snapShots[theID].firstBackingPage; page < snapshotrecord->lastBackingPage; page++) {
302 memcpy(snapshotrecord->backingRecords[page].basePtrOfPage, &snapshotrecord->backingStore[page], sizeof(snapshot_page_t));
308 #if USE_MPROTECT_SNAPSHOT
309 HashTable< void *, bool, uintptr_t, 4, model_malloc, model_calloc, model_free> duplicateMap;
310 for (unsigned int region = 0; region < snapshotrecord->lastRegion; region++) {
311 if (mprotect(snapshotrecord->regionsToSnapShot[region].basePtr, snapshotrecord->regionsToSnapShot[region].sizeInPages * sizeof(snapshot_page_t), PROT_READ | PROT_WRITE) == -1) {
313 model_print("Failed to mprotect inside of takeSnapShot\n");
317 for (unsigned int page = snapshotrecord->snapShots[theID].firstBackingPage; page < snapshotrecord->lastBackingPage; page++) {
318 if (!duplicateMap.contains(snapshotrecord->backingRecords[page].basePtrOfPage)) {
319 duplicateMap.put(snapshotrecord->backingRecords[page].basePtrOfPage, true);
320 memcpy(snapshotrecord->backingRecords[page].basePtrOfPage, &snapshotrecord->backingStore[page], sizeof(snapshot_page_t));
323 snapshotrecord->lastSnapShot = theID;
324 snapshotrecord->lastBackingPage = snapshotrecord->snapShots[theID].firstBackingPage;
325 takeSnapshot(); //Make sure current snapshot is still good...All later ones are cleared
327 snapshotrecord->mIDToRollback = theID;
328 volatile int sTemp = 0;
329 getcontext(&snapshotrecord->mContextToRollback);
331 * This is used to quit the process on rollback, so that the process
332 * which needs to rollback can quit allowing the process whose
333 * snapshotid matches the rollbackid to switch to this context and
338 SSDEBUG("Invoked rollback\n");
342 * This fix obviates the need for a finalize call. hence less dependences for model-checker....
344 snapshotrecord->mIDToRollback = -1;