52 // In 53 // 1) ParallelTaskTerminator::offer_termination() where _n_threads 54 // must be set to the correct value so that count of workers that 55 // have offered termination will exactly match the number 56 // working on the task. Tasks such as those derived from GCTask 57 // use ParallelTaskTerminator's. Tasks that want load balancing 58 // by work stealing use this method to gauge completion. 59 // 2) SubTasksDone has a variable _n_threads that is used in 60 // all_tasks_completed() to determine completion. all_tasks_complete() 61 // counts the number of tasks that have been done and then reset 62 // the SubTasksDone so that it can be used again. When the number of 63 // tasks is set to the number of GC workers, then _n_threads must 64 // be set to the number of active GC workers. G1CollectedHeap, 65 // HRInto_G1RemSet, GenCollectedHeap and SharedHeap have SubTasksDone. 66 // This seems too many. 67 // 3) SequentialSubTasksDone has an _n_threads that is used in 68 // a way similar to SubTasksDone and has the same dependency on the 69 // number of active GC workers. CompactibleFreeListSpace and Space 70 // have SequentialSubTasksDone's. 71 // Example of using SubTasksDone and SequentialSubTasksDone 72 // G1CollectedHeap::g1_process_strong_roots() calls 73 // process_strong_roots(false, // no scoping; this is parallel code 74 // is_scavenging, so, 75 // &buf_scan_non_heap_roots, 76 // &eager_scan_code_roots); 77 // which delegates to SharedHeap::process_strong_roots() and uses 78 // SubTasksDone* _process_strong_tasks to claim tasks. 79 // process_strong_roots() calls 80 // rem_set()->younger_refs_iterate() 81 // to scan the card table and which eventually calls down into 82 // CardTableModRefBS::par_non_clean_card_iterate_work(). This method 83 // uses SequentialSubTasksDone* _pst to claim tasks. 84 // Both SubTasksDone and SequentialSubTasksDone call their method 85 // all_tasks_completed() to count the number of GC workers that have 86 // finished their work. That logic is "when all the workers are 87 // finished the tasks are finished". 88 // 89 // The pattern that appears in the code is to set _n_threads 90 // to a value > 1 before a task that you would like executed in parallel 91 // and then to set it to 0 after that task has completed. A value of 92 // 0 is a "special" value in set_n_threads() which translates to 93 // setting _n_threads to 1. 94 // 95 // Some code uses _n_termination to decide if work should be done in 96 // parallel. The notorious possibly_parallel_oops_do() in threads.cpp 97 // is an example of such code. Look for variable "is_par" for other 98 // examples. 99 // 165 166 // Iterate over all spaces in use in the heap, in an undefined order. 167 virtual void space_iterate(SpaceClosure* cl) = 0; 168 169 // A SharedHeap will contain some number of spaces. This finds the 170 // space whose reserved area contains the given address, or else returns 171 // NULL. 172 virtual Space* space_containing(const void* addr) const = 0; 173 174 bool no_gc_in_progress() { return !is_gc_active(); } 175 176 // Some collectors will perform "process_strong_roots" in parallel. 177 // Such a call will involve claiming some fine-grained tasks, such as 178 // scanning of threads. To make this process simpler, we provide the 179 // "strong_roots_parity()" method. Collectors that start parallel tasks 180 // whose threads invoke "process_strong_roots" must 181 // call "change_strong_roots_parity" in sequential code starting such a 182 // task. (This also means that a parallel thread may only call 183 // process_strong_roots once.) 184 // 185 // For calls to process_strong_roots by sequential code, the parity is 186 // updated automatically. 187 // 188 // The idea is that objects representing fine-grained tasks, such as 189 // threads, will contain a "parity" field. A task will is claimed in the 190 // current "process_strong_roots" call only if its parity field is the 191 // same as the "strong_roots_parity"; task claiming is accomplished by 192 // updating the parity field to the strong_roots_parity with a CAS. 193 // 194 // If the client meats this spec, then strong_roots_parity() will have 195 // the following properties: 196 // a) to return a different value than was returned before the last 197 // call to change_strong_roots_parity, and 198 // c) to never return a distinguished value (zero) with which such 199 // task-claiming variables may be initialized, to indicate "never 200 // claimed". 201 private: 202 void change_strong_roots_parity(); 203 public: 204 int strong_roots_parity() { return _strong_roots_parity; } 205 206 // Call these in sequential code around process_strong_roots. 207 // strong_roots_prologue calls change_strong_roots_parity, if 208 // parallel tasks are enabled. 209 class StrongRootsScope : public MarkingCodeBlobClosure::MarkScope { 210 public: 211 StrongRootsScope(SharedHeap* outer, bool activate = true); 212 ~StrongRootsScope(); 213 }; 214 friend class StrongRootsScope; 215 216 enum ScanningOption { 217 SO_None = 0x0, 218 SO_AllClasses = 0x1, 219 SO_SystemClasses = 0x2, 220 SO_Strings = 0x4, 221 SO_AllCodeCache = 0x8, 222 SO_ScavengeCodeCache = 0x10 223 }; 224 225 FlexibleWorkGang* workers() const { return _workers; } 226 227 // Invoke the "do_oop" method the closure "roots" on all root locations. 228 // The "so" argument determines which roots the closure is applied to: 229 // "SO_None" does none; 230 // "SO_AllClasses" applies the closure to all entries in the SystemDictionary; 231 // "SO_SystemClasses" to all the "system" classes and loaders; 232 // "SO_Strings" applies the closure to all entries in StringTable; 233 // "SO_AllCodeCache" applies the closure to all elements of the CodeCache. 234 // "SO_ScavengeCodeCache" applies the closure to elements on the scavenge root list in the CodeCache. 235 void process_strong_roots(bool activate_scope, 236 ScanningOption so, 237 OopClosure* roots, 238 KlassClosure* klass_closure); 239 240 // Apply "root_closure" to the JNI weak roots.. 241 void process_weak_roots(OopClosure* root_closure); 242 243 // The functions below are helper functions that a subclass of 244 // "SharedHeap" can use in the implementation of its virtual 245 // functions. 246 247 public: 248 249 // Do anything common to GC's. 250 virtual void gc_prologue(bool full) = 0; 251 virtual void gc_epilogue(bool full) = 0; 252 253 // Sets the number of parallel threads that will be doing tasks 254 // (such as process strong roots) subsequently. 255 virtual void set_par_threads(uint t); 256 257 int n_termination(); 258 void set_n_termination(int t); 259 260 // 261 // New methods from CollectedHeap 262 // 263 264 // Some utilities. 265 void print_size_transition(outputStream* out, 266 size_t bytes_before, 267 size_t bytes_after, 268 size_t capacity); 269 }; 270 271 inline SharedHeap::ScanningOption operator|(SharedHeap::ScanningOption so0, SharedHeap::ScanningOption so1) { 272 return static_cast<SharedHeap::ScanningOption>(static_cast<int>(so0) | static_cast<int>(so1)); 273 } 274 | 52 // In 53 // 1) ParallelTaskTerminator::offer_termination() where _n_threads 54 // must be set to the correct value so that count of workers that 55 // have offered termination will exactly match the number 56 // working on the task. Tasks such as those derived from GCTask 57 // use ParallelTaskTerminator's. Tasks that want load balancing 58 // by work stealing use this method to gauge completion. 59 // 2) SubTasksDone has a variable _n_threads that is used in 60 // all_tasks_completed() to determine completion. all_tasks_complete() 61 // counts the number of tasks that have been done and then reset 62 // the SubTasksDone so that it can be used again. When the number of 63 // tasks is set to the number of GC workers, then _n_threads must 64 // be set to the number of active GC workers. G1CollectedHeap, 65 // HRInto_G1RemSet, GenCollectedHeap and SharedHeap have SubTasksDone. 66 // This seems too many. 67 // 3) SequentialSubTasksDone has an _n_threads that is used in 68 // a way similar to SubTasksDone and has the same dependency on the 69 // number of active GC workers. CompactibleFreeListSpace and Space 70 // have SequentialSubTasksDone's. 71 // Example of using SubTasksDone and SequentialSubTasksDone 72 // G1CollectedHeap::g1_process_roots() 73 // to SharedHeap::process_roots() and uses 74 // SubTasksDone* _process_strong_tasks to claim tasks. 75 // process_roots() calls 76 // rem_set()->younger_refs_iterate() 77 // to scan the card table and which eventually calls down into 78 // CardTableModRefBS::par_non_clean_card_iterate_work(). This method 79 // uses SequentialSubTasksDone* _pst to claim tasks. 80 // Both SubTasksDone and SequentialSubTasksDone call their method 81 // all_tasks_completed() to count the number of GC workers that have 82 // finished their work. That logic is "when all the workers are 83 // finished the tasks are finished". 84 // 85 // The pattern that appears in the code is to set _n_threads 86 // to a value > 1 before a task that you would like executed in parallel 87 // and then to set it to 0 after that task has completed. A value of 88 // 0 is a "special" value in set_n_threads() which translates to 89 // setting _n_threads to 1. 90 // 91 // Some code uses _n_termination to decide if work should be done in 92 // parallel. The notorious possibly_parallel_oops_do() in threads.cpp 93 // is an example of such code. Look for variable "is_par" for other 94 // examples. 95 // 161 162 // Iterate over all spaces in use in the heap, in an undefined order. 163 virtual void space_iterate(SpaceClosure* cl) = 0; 164 165 // A SharedHeap will contain some number of spaces. This finds the 166 // space whose reserved area contains the given address, or else returns 167 // NULL. 168 virtual Space* space_containing(const void* addr) const = 0; 169 170 bool no_gc_in_progress() { return !is_gc_active(); } 171 172 // Some collectors will perform "process_strong_roots" in parallel. 173 // Such a call will involve claiming some fine-grained tasks, such as 174 // scanning of threads. To make this process simpler, we provide the 175 // "strong_roots_parity()" method. Collectors that start parallel tasks 176 // whose threads invoke "process_strong_roots" must 177 // call "change_strong_roots_parity" in sequential code starting such a 178 // task. (This also means that a parallel thread may only call 179 // process_strong_roots once.) 180 // 181 // For calls to process_roots by sequential code, the parity is 182 // updated automatically. 183 // 184 // The idea is that objects representing fine-grained tasks, such as 185 // threads, will contain a "parity" field. A task will is claimed in the 186 // current "process_roots" call only if its parity field is the 187 // same as the "strong_roots_parity"; task claiming is accomplished by 188 // updating the parity field to the strong_roots_parity with a CAS. 189 // 190 // If the client meats this spec, then strong_roots_parity() will have 191 // the following properties: 192 // a) to return a different value than was returned before the last 193 // call to change_strong_roots_parity, and 194 // c) to never return a distinguished value (zero) with which such 195 // task-claiming variables may be initialized, to indicate "never 196 // claimed". 197 public: 198 int strong_roots_parity() { return _strong_roots_parity; } 199 200 // Call these in sequential code around process_roots. 201 // strong_roots_prologue calls change_strong_roots_parity, if 202 // parallel tasks are enabled. 203 class StrongRootsScope : public MarkingCodeBlobClosure::MarkScope { 204 // Used to implement the Thread work barrier. 205 static Monitor* _lock; 206 207 SharedHeap* _sh; 208 volatile jint _n_workers_done_with_threads; 209 210 public: 211 StrongRootsScope(SharedHeap* heap, bool activate = true); 212 ~StrongRootsScope(); 213 214 // Mark that this thread is done with the Threads work. 215 void mark_worker_done_with_threads(uint n_workers); 216 // Wait until all n_workers are done with the Threads work. 217 void wait_until_all_workers_done_with_threads(uint n_workers); 218 }; 219 friend class StrongRootsScope; 220 221 // The current active StrongRootScope 222 StrongRootsScope* _strong_roots_scope; 223 224 StrongRootsScope* active_strong_roots_scope() const; 225 226 private: 227 void register_strong_roots_scope(StrongRootsScope* scope); 228 void unregister_strong_roots_scope(StrongRootsScope* scope); 229 void change_strong_roots_parity(); 230 231 public: 232 enum ScanningOption { 233 SO_None = 0x0, 234 SO_AllCodeCache = 0x8, 235 SO_ScavengeCodeCache = 0x10 236 }; 237 238 FlexibleWorkGang* workers() const { return _workers; } 239 240 // Invoke the "do_oop" method the closure "roots" on all root locations. 241 // The "so" argument determines which roots the closure is applied to: 242 // "SO_None" does none; 243 // "SO_AllCodeCache" applies the closure to all elements of the CodeCache. 244 // "SO_ScavengeCodeCache" applies the closure to elements on the scavenge root list in the CodeCache. 245 void process_roots(bool activate_scope, 246 ScanningOption so, 247 OopClosure* strong_roots, 248 OopClosure* weak_roots, 249 CLDClosure* strong_cld_closure, 250 CLDClosure* weak_cld_closure, 251 CodeBlobClosure* code_roots); 252 void process_all_roots(bool activate_scope, 253 ScanningOption so, 254 OopClosure* roots, 255 CLDClosure* cld_closure, 256 CodeBlobClosure* code_roots); 257 void process_strong_roots(bool activate_scope, 258 ScanningOption so, 259 OopClosure* roots, 260 CLDClosure* cld_closure, 261 CodeBlobClosure* code_roots); 262 263 264 // Apply "root_closure" to the JNI weak roots.. 265 void process_weak_roots(OopClosure* root_closure); 266 267 // The functions below are helper functions that a subclass of 268 // "SharedHeap" can use in the implementation of its virtual 269 // functions. 270 271 public: 272 273 // Do anything common to GC's. 274 virtual void gc_prologue(bool full) = 0; 275 virtual void gc_epilogue(bool full) = 0; 276 277 // Sets the number of parallel threads that will be doing tasks 278 // (such as process roots) subsequently. 279 virtual void set_par_threads(uint t); 280 281 int n_termination(); 282 void set_n_termination(int t); 283 284 // 285 // New methods from CollectedHeap 286 // 287 288 // Some utilities. 289 void print_size_transition(outputStream* out, 290 size_t bytes_before, 291 size_t bytes_after, 292 size_t capacity); 293 }; 294 295 inline SharedHeap::ScanningOption operator|(SharedHeap::ScanningOption so0, SharedHeap::ScanningOption so1) { 296 return static_cast<SharedHeap::ScanningOption>(static_cast<int>(so0) | static_cast<int>(so1)); 297 } 298 |