1 /* 2 * Copyright (c) 2019, 2025, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "cds/archiveBuilder.hpp" 26 #include "cds/archiveHeapLoader.inline.hpp" 27 #include "cds/archiveUtils.hpp" 28 #include "cds/cdsConfig.hpp" 29 #include "cds/classListParser.hpp" 30 #include "cds/classListWriter.hpp" 31 #include "cds/dynamicArchive.hpp" 32 #include "cds/filemap.hpp" 33 #include "cds/heapShared.hpp" 34 #include "cds/lambdaProxyClassDictionary.hpp" 35 #include "cds/metaspaceShared.hpp" 36 #include "classfile/classLoader.hpp" 37 #include "classfile/systemDictionaryShared.hpp" 38 #include "classfile/vmClasses.hpp" 39 #include "interpreter/bootstrapInfo.hpp" 40 #include "memory/metaspaceUtils.hpp" 41 #include "memory/resourceArea.hpp" 42 #include "oops/compressedOops.inline.hpp" 43 #include "oops/klass.inline.hpp" 44 #include "runtime/arguments.hpp" 45 #include "utilities/bitMap.inline.hpp" 46 #include "utilities/debug.hpp" 47 #include "utilities/formatBuffer.hpp" 48 #include "utilities/globalDefinitions.hpp" 49 #include "utilities/spinYield.hpp" 50 51 CHeapBitMap* ArchivePtrMarker::_ptrmap = nullptr; 52 CHeapBitMap* ArchivePtrMarker::_rw_ptrmap = nullptr; 53 CHeapBitMap* ArchivePtrMarker::_ro_ptrmap = nullptr; 54 CHeapBitMap* ArchivePtrMarker::_cc_ptrmap = nullptr; 55 VirtualSpace* ArchivePtrMarker::_vs; 56 57 bool ArchivePtrMarker::_compacted; 58 59 void ArchivePtrMarker::initialize(CHeapBitMap* ptrmap, VirtualSpace* vs) { 60 assert(_ptrmap == nullptr, "initialize only once"); 61 assert(_rw_ptrmap == nullptr, "initialize only once"); 62 assert(_ro_ptrmap == nullptr, "initialize only once"); 63 assert(_cc_ptrmap == nullptr, "initialize only once"); 64 _vs = vs; 65 _compacted = false; 66 _ptrmap = ptrmap; 67 68 // Use this as initial guesstimate. We should need less space in the 69 // archive, but if we're wrong the bitmap will be expanded automatically. 70 size_t estimated_archive_size = MetaspaceGC::capacity_until_GC(); 71 // But set it smaller in debug builds so we always test the expansion code. 72 // (Default archive is about 12MB). 73 DEBUG_ONLY(estimated_archive_size = 6 * M); 74 75 // We need one bit per pointer in the archive. 76 _ptrmap->initialize(estimated_archive_size / sizeof(intptr_t)); 77 } 78 79 void ArchivePtrMarker::initialize_rw_ro_cc_maps(CHeapBitMap* rw_ptrmap, CHeapBitMap* ro_ptrmap, CHeapBitMap* cc_ptrmap) { 80 address* buff_bottom = (address*)ArchiveBuilder::current()->buffer_bottom(); 81 address* rw_bottom = (address*)ArchiveBuilder::current()->rw_region()->base(); 82 address* ro_bottom = (address*)ArchiveBuilder::current()->ro_region()->base(); 83 address* cc_bottom = (address*)ArchiveBuilder::current()->cc_region()->base(); 84 85 // The bit in _ptrmap that cover the very first word in the rw/ro/cc regions. 86 size_t rw_start = rw_bottom - buff_bottom; 87 size_t ro_start = ro_bottom - buff_bottom; 88 size_t cc_start = cc_bottom - buff_bottom; 89 90 // The number of bits used by the rw/ro ptrmaps. We might have lots of zero 91 // bits at the bottom and top of rw/ro ptrmaps, but these zeros will be 92 // removed by FileMapInfo::write_bitmap_region(). 93 size_t rw_size = ArchiveBuilder::current()->rw_region()->used() / sizeof(address); 94 size_t ro_size = ArchiveBuilder::current()->ro_region()->used() / sizeof(address); 95 size_t cc_size = ArchiveBuilder::current()->cc_region()->used() / sizeof(address); 96 97 // The last (exclusive) bit in _ptrmap that covers the rw/ro regions. 98 // Note: _ptrmap is dynamically expanded only when an actual pointer is written, so 99 // it may not be as large as we want. 100 size_t rw_end = MIN2<size_t>(rw_start + rw_size, _ptrmap->size()); 101 size_t ro_end = MIN2<size_t>(ro_start + ro_size, _ptrmap->size()); 102 size_t cc_end = MIN2<size_t>(cc_start + cc_size, _ptrmap->size()); 103 104 rw_ptrmap->initialize(rw_size); 105 ro_ptrmap->initialize(ro_size); 106 cc_ptrmap->initialize(cc_size); 107 108 for (size_t rw_bit = rw_start; rw_bit < rw_end; rw_bit++) { 109 rw_ptrmap->at_put(rw_bit - rw_start, _ptrmap->at(rw_bit)); 110 } 111 112 for(size_t ro_bit = ro_start; ro_bit < ro_end; ro_bit++) { 113 ro_ptrmap->at_put(ro_bit - ro_start, _ptrmap->at(ro_bit)); 114 } 115 116 for (size_t cc_bit = cc_start; cc_bit < cc_end; cc_bit++) { 117 cc_ptrmap->at_put(cc_bit - cc_start, _ptrmap->at(cc_bit)); 118 } 119 120 _rw_ptrmap = rw_ptrmap; 121 _ro_ptrmap = ro_ptrmap; 122 _cc_ptrmap = cc_ptrmap; 123 } 124 125 void ArchivePtrMarker::mark_pointer(address* ptr_loc) { 126 assert(_ptrmap != nullptr, "not initialized"); 127 assert(!_compacted, "cannot mark anymore"); 128 129 if (ptr_base() <= ptr_loc && ptr_loc < ptr_end()) { 130 address value = *ptr_loc; 131 // We don't want any pointer that points to very bottom of the archive, otherwise when 132 // MetaspaceShared::default_base_address()==0, we can't distinguish between a pointer 133 // to nothing (null) vs a pointer to an objects that happens to be at the very bottom 134 // of the archive. 135 assert(value != (address)ptr_base(), "don't point to the bottom of the archive"); 136 137 if (value != nullptr) { 138 assert(uintx(ptr_loc) % sizeof(intptr_t) == 0, "pointers must be stored in aligned addresses"); 139 size_t idx = ptr_loc - ptr_base(); 140 if (_ptrmap->size() <= idx) { 141 _ptrmap->resize((idx + 1) * 2); 142 } 143 assert(idx < _ptrmap->size(), "must be"); 144 _ptrmap->set_bit(idx); 145 //tty->print_cr("Marking pointer [" PTR_FORMAT "] -> " PTR_FORMAT " @ %5zu", p2i(ptr_loc), p2i(*ptr_loc), idx); 146 } 147 } 148 } 149 150 void ArchivePtrMarker::clear_pointer(address* ptr_loc) { 151 assert(_ptrmap != nullptr, "not initialized"); 152 assert(!_compacted, "cannot clear anymore"); 153 154 assert(ptr_base() <= ptr_loc && ptr_loc < ptr_end(), "must be"); 155 assert(uintx(ptr_loc) % sizeof(intptr_t) == 0, "pointers must be stored in aligned addresses"); 156 size_t idx = ptr_loc - ptr_base(); 157 assert(idx < _ptrmap->size(), "cannot clear pointers that have not been marked"); 158 _ptrmap->clear_bit(idx); 159 //tty->print_cr("Clearing pointer [" PTR_FORMAT "] -> " PTR_FORMAT " @ %5zu", p2i(ptr_loc), p2i(*ptr_loc), idx); 160 } 161 162 class ArchivePtrBitmapCleaner: public BitMapClosure { 163 CHeapBitMap* _ptrmap; 164 address* _ptr_base; 165 address _relocatable_base; 166 address _relocatable_end; 167 size_t _max_non_null_offset; 168 169 public: 170 ArchivePtrBitmapCleaner(CHeapBitMap* ptrmap, address* ptr_base, address relocatable_base, address relocatable_end) : 171 _ptrmap(ptrmap), _ptr_base(ptr_base), 172 _relocatable_base(relocatable_base), _relocatable_end(relocatable_end), _max_non_null_offset(0) {} 173 174 bool do_bit(size_t offset) { 175 address* ptr_loc = _ptr_base + offset; 176 address ptr_value = *ptr_loc; 177 if (ptr_value != nullptr) { 178 assert(_relocatable_base <= ptr_value && ptr_value < _relocatable_end, "do not point to arbitrary locations!"); 179 if (_max_non_null_offset < offset) { 180 _max_non_null_offset = offset; 181 } 182 } else { 183 _ptrmap->clear_bit(offset); 184 DEBUG_ONLY(log_trace(cds, reloc)("Clearing pointer [" PTR_FORMAT "] -> null @ %9zu", p2i(ptr_loc), offset)); 185 } 186 187 return true; 188 } 189 190 size_t max_non_null_offset() const { return _max_non_null_offset; } 191 }; 192 193 void ArchivePtrMarker::compact(address relocatable_base, address relocatable_end) { 194 assert(!_compacted, "cannot compact again"); 195 ArchivePtrBitmapCleaner cleaner(_ptrmap, ptr_base(), relocatable_base, relocatable_end); 196 _ptrmap->iterate(&cleaner); 197 compact(cleaner.max_non_null_offset()); 198 } 199 200 void ArchivePtrMarker::compact(size_t max_non_null_offset) { 201 assert(!_compacted, "cannot compact again"); 202 _ptrmap->resize(max_non_null_offset + 1); 203 _compacted = true; 204 } 205 206 char* DumpRegion::expand_top_to(char* newtop) { 207 assert(is_allocatable(), "must be initialized and not packed"); 208 assert(newtop >= _top, "must not grow backwards"); 209 if (newtop > _end) { 210 ArchiveBuilder::current()->report_out_of_space(_name, newtop - _top); 211 ShouldNotReachHere(); 212 } 213 214 commit_to(newtop); 215 _top = newtop; 216 217 if (_max_delta > 0) { 218 uintx delta = ArchiveBuilder::current()->buffer_to_offset((address)(newtop-1)); 219 if (delta > _max_delta) { 220 // This is just a sanity check and should not appear in any real world usage. This 221 // happens only if you allocate more than 2GB of shared objects and would require 222 // millions of shared classes. 223 log_error(cds)("Out of memory in the CDS archive: Please reduce the number of shared classes."); 224 MetaspaceShared::unrecoverable_writing_error(); 225 } 226 } 227 228 return _top; 229 } 230 231 void DumpRegion::commit_to(char* newtop) { 232 assert(CDSConfig::is_dumping_archive(), "sanity"); 233 char* base = _rs->base(); 234 size_t need_committed_size = newtop - base; 235 size_t has_committed_size = _vs->committed_size(); 236 if (need_committed_size < has_committed_size) { 237 return; 238 } 239 240 size_t min_bytes = need_committed_size - has_committed_size; 241 size_t preferred_bytes = 1 * M; 242 size_t uncommitted = _vs->reserved_size() - has_committed_size; 243 244 size_t commit = MAX2(min_bytes, preferred_bytes); 245 commit = MIN2(commit, uncommitted); 246 assert(commit <= uncommitted, "sanity"); 247 248 if (!_vs->expand_by(commit, false)) { 249 log_error(cds)("Failed to expand shared space to %zu bytes", 250 need_committed_size); 251 MetaspaceShared::unrecoverable_writing_error(); 252 } 253 254 const char* which; 255 if (_rs->base() == (char*)MetaspaceShared::symbol_rs_base()) { 256 which = "symbol"; 257 } else { 258 which = "shared"; 259 } 260 log_debug(cds)("Expanding %s spaces by %7zu bytes [total %9zu bytes ending at %p]", 261 which, commit, _vs->actual_committed_size(), _vs->high()); 262 } 263 264 char* DumpRegion::allocate(size_t num_bytes, size_t alignment) { 265 // Always align to at least minimum alignment 266 alignment = MAX2(SharedSpaceObjectAlignment, alignment); 267 char* p = (char*)align_up(_top, alignment); 268 char* newtop = p + align_up(num_bytes, (size_t)SharedSpaceObjectAlignment); 269 expand_top_to(newtop); 270 memset(p, 0, newtop - p); 271 return p; 272 } 273 274 void DumpRegion::append_intptr_t(intptr_t n, bool need_to_mark) { 275 assert(is_aligned(_top, sizeof(intptr_t)), "bad alignment"); 276 intptr_t *p = (intptr_t*)_top; 277 char* newtop = _top + sizeof(intptr_t); 278 expand_top_to(newtop); 279 *p = n; 280 if (need_to_mark) { 281 ArchivePtrMarker::mark_pointer(p); 282 } 283 } 284 285 void DumpRegion::print(size_t total_bytes) const { 286 char* base = used() > 0 ? ArchiveBuilder::current()->to_requested(_base) : nullptr; 287 log_debug(cds)("%s space: %9zu [ %4.1f%% of total] out of %9zu bytes [%5.1f%% used] at " INTPTR_FORMAT, 288 _name, used(), percent_of(used(), total_bytes), reserved(), percent_of(used(), reserved()), 289 p2i(base)); 290 } 291 292 void DumpRegion::print_out_of_space_msg(const char* failing_region, size_t needed_bytes) { 293 log_error(cds)("[%-8s] " PTR_FORMAT " - " PTR_FORMAT " capacity =%9d, allocated =%9d", 294 _name, p2i(_base), p2i(_top), int(_end - _base), int(_top - _base)); 295 if (strcmp(_name, failing_region) == 0) { 296 log_error(cds)(" required = %d", int(needed_bytes)); 297 } 298 } 299 300 void DumpRegion::init(ReservedSpace* rs, VirtualSpace* vs) { 301 _rs = rs; 302 _vs = vs; 303 // Start with 0 committed bytes. The memory will be committed as needed. 304 if (!_vs->initialize(*_rs, 0)) { 305 fatal("Unable to allocate memory for shared space"); 306 } 307 _base = _top = _rs->base(); 308 _end = _rs->end(); 309 } 310 311 void DumpRegion::pack(DumpRegion* next) { 312 if (!is_packed()) { 313 _end = (char*)align_up(_top, MetaspaceShared::core_region_alignment()); 314 _is_packed = true; 315 } 316 if (next != nullptr) { 317 next->_rs = _rs; 318 next->_vs = _vs; 319 next->_base = next->_top = this->_end; 320 next->_end = _rs->end(); 321 } 322 } 323 324 void WriteClosure::do_ptr(void** p) { 325 // Write ptr into the archive; ptr can be: 326 // (a) null -> written as 0 327 // (b) a "buffered" address -> written as is 328 // (c) a "source" address -> convert to "buffered" and write 329 // The common case is (c). E.g., when writing the vmClasses into the archive. 330 // We have (b) only when we don't have a corresponding source object. E.g., 331 // the archived c++ vtable entries. 332 address ptr = *(address*)p; 333 if (ptr != nullptr && !ArchiveBuilder::current()->is_in_buffer_space(ptr)) { 334 ptr = ArchiveBuilder::current()->get_buffered_addr(ptr); 335 } 336 // null pointers do not need to be converted to offsets 337 if (ptr != nullptr) { 338 ptr = (address)ArchiveBuilder::current()->buffer_to_offset(ptr); 339 } 340 _dump_region->append_intptr_t((intptr_t)ptr, false); 341 } 342 343 void ReadClosure::do_ptr(void** p) { 344 assert(*p == nullptr, "initializing previous initialized pointer."); 345 intptr_t obj = nextPtr(); 346 assert(obj >= 0, "sanity."); 347 *p = (obj != 0) ? (void*)(_base_address + obj) : (void*)obj; 348 } 349 350 void ReadClosure::do_u4(u4* p) { 351 intptr_t obj = nextPtr(); 352 *p = (u4)(uintx(obj)); 353 } 354 355 void ReadClosure::do_int(int* p) { 356 intptr_t obj = nextPtr(); 357 *p = (int)(intx(obj)); 358 } 359 360 void ReadClosure::do_bool(bool* p) { 361 intptr_t obj = nextPtr(); 362 *p = (bool)(uintx(obj)); 363 } 364 365 void ReadClosure::do_tag(int tag) { 366 int old_tag; 367 old_tag = (int)(intptr_t)nextPtr(); 368 // do_int(&old_tag); 369 assert(tag == old_tag, "tag doesn't match (%d, expected %d)", old_tag, tag); 370 FileMapInfo::assert_mark(tag == old_tag); 371 } 372 373 void ArchiveUtils::log_to_classlist(BootstrapInfo* bootstrap_specifier, TRAPS) { 374 if (ClassListWriter::is_enabled()) { 375 if (LambdaProxyClassDictionary::is_supported_invokedynamic(bootstrap_specifier)) { 376 const constantPoolHandle& pool = bootstrap_specifier->pool(); 377 if (SystemDictionaryShared::is_builtin_loader(pool->pool_holder()->class_loader_data())) { 378 // Currently lambda proxy classes are supported only for the built-in loaders. 379 ResourceMark rm(THREAD); 380 int pool_index = bootstrap_specifier->bss_index(); 381 ClassListWriter w; 382 w.stream()->print("%s %s", ClassListParser::lambda_proxy_tag(), pool->pool_holder()->name()->as_C_string()); 383 CDSIndyInfo cii; 384 ClassListParser::populate_cds_indy_info(pool, pool_index, &cii, CHECK); 385 GrowableArray<const char*>* indy_items = cii.items(); 386 for (int i = 0; i < indy_items->length(); i++) { 387 w.stream()->print(" %s", indy_items->at(i)); 388 } 389 w.stream()->cr(); 390 } 391 } 392 } 393 } 394 395 396 // "boot", "platform", "app" or nullptr 397 const char* ArchiveUtils::builtin_loader_name_or_null(oop loader) { 398 if (loader == nullptr) { 399 return "boot"; 400 } else if (loader == SystemDictionary::java_platform_loader()) { 401 return "platform"; 402 } else if (loader == SystemDictionary::java_system_loader()) { 403 return "app"; 404 } else { 405 return nullptr; 406 } 407 } 408 409 // "boot", "platform", "app". Asserts if not a built-in-loader 410 const char* ArchiveUtils::builtin_loader_name(oop loader) { 411 const char* name = builtin_loader_name_or_null(loader); 412 assert(name != nullptr, "must be a built-in loader"); 413 return name; 414 } 415 416 bool ArchiveUtils::builtin_loader_from_type(const char* loader_type, oop* value_ret) { 417 if (strcmp(loader_type, "boot") == 0) { 418 *value_ret = nullptr; 419 return true; 420 } else if (strcmp(loader_type, "platform") == 0) { 421 *value_ret = SystemDictionary::java_platform_loader(); 422 return true; 423 } else if (strcmp(loader_type, "app") == 0) { 424 *value_ret = SystemDictionary::java_system_loader(); 425 return true; 426 } else { 427 DEBUG_ONLY(*value_ret = cast_to_oop((void*)badOopVal)); 428 return false; 429 } 430 } 431 432 oop ArchiveUtils::builtin_loader_from_type(int loader_type) { 433 if (loader_type == ClassLoader::BOOT_LOADER) { 434 return nullptr; 435 } else if (loader_type == ClassLoader::PLATFORM_LOADER) { 436 return SystemDictionary::java_platform_loader(); 437 } else if (loader_type == ClassLoader::APP_LOADER) { 438 return SystemDictionary::java_system_loader(); 439 } else { 440 ShouldNotReachHere(); 441 return nullptr; 442 } 443 } 444 445 bool ArchiveUtils::has_aot_initialized_mirror(InstanceKlass* src_ik) { 446 if (SystemDictionaryShared::is_excluded_class(src_ik)) { 447 assert(!ArchiveBuilder::current()->has_been_buffered(src_ik), "sanity"); 448 return false; 449 } 450 return ArchiveBuilder::current()->get_buffered_addr(src_ik)->has_aot_initialized_mirror(); 451 } 452 453 size_t HeapRootSegments::size_in_bytes(size_t seg_idx) { 454 assert(seg_idx < _count, "In range"); 455 return objArrayOopDesc::object_size(size_in_elems(seg_idx)) * HeapWordSize; 456 } 457 458 int HeapRootSegments::size_in_elems(size_t seg_idx) { 459 assert(seg_idx < _count, "In range"); 460 if (seg_idx != _count - 1) { 461 return _max_size_in_elems; 462 } else { 463 // Last slice, leftover 464 return _roots_count % _max_size_in_elems; 465 } 466 } 467 468 size_t HeapRootSegments::segment_offset(size_t seg_idx) { 469 assert(seg_idx < _count, "In range"); 470 return _base_offset + seg_idx * _max_size_in_bytes; 471 } 472 473 ArchiveWorkers::ArchiveWorkers() : 474 _end_semaphore(0), 475 _num_workers(max_workers()), 476 _started_workers(0), 477 _finish_tokens(0), 478 _state(UNUSED), 479 _task(nullptr) {} 480 481 ArchiveWorkers::~ArchiveWorkers() { 482 assert(Atomic::load(&_state) != WORKING, "Should not be working"); 483 } 484 485 int ArchiveWorkers::max_workers() { 486 // The pool is used for short-lived bursty tasks. We do not want to spend 487 // too much time creating and waking up threads unnecessarily. Plus, we do 488 // not want to overwhelm large machines. This is why we want to be very 489 // conservative about the number of workers actually needed. 490 return MAX2(0, log2i_graceful(os::active_processor_count())); 491 } 492 493 bool ArchiveWorkers::is_parallel() { 494 return _num_workers > 0; 495 } 496 497 void ArchiveWorkers::start_worker_if_needed() { 498 while (true) { 499 int cur = Atomic::load(&_started_workers); 500 if (cur >= _num_workers) { 501 return; 502 } 503 if (Atomic::cmpxchg(&_started_workers, cur, cur + 1, memory_order_relaxed) == cur) { 504 new ArchiveWorkerThread(this); 505 return; 506 } 507 } 508 } 509 510 void ArchiveWorkers::run_task(ArchiveWorkerTask* task) { 511 assert(Atomic::load(&_state) == UNUSED, "Should be unused yet"); 512 assert(Atomic::load(&_task) == nullptr, "Should not have running tasks"); 513 Atomic::store(&_state, WORKING); 514 515 if (is_parallel()) { 516 run_task_multi(task); 517 } else { 518 run_task_single(task); 519 } 520 521 assert(Atomic::load(&_state) == WORKING, "Should be working"); 522 Atomic::store(&_state, SHUTDOWN); 523 } 524 525 void ArchiveWorkers::run_task_single(ArchiveWorkerTask* task) { 526 // Single thread needs no chunking. 527 task->configure_max_chunks(1); 528 529 // Execute the task ourselves, as there are no workers. 530 task->work(0, 1); 531 } 532 533 void ArchiveWorkers::run_task_multi(ArchiveWorkerTask* task) { 534 // Multiple threads can work with multiple chunks. 535 task->configure_max_chunks(_num_workers * CHUNKS_PER_WORKER); 536 537 // Set up the run and publish the task. Issue one additional finish token 538 // to cover the semaphore shutdown path, see below. 539 Atomic::store(&_finish_tokens, _num_workers + 1); 540 Atomic::release_store(&_task, task); 541 542 // Kick off pool startup by starting a single worker, and proceed 543 // immediately to executing the task locally. 544 start_worker_if_needed(); 545 546 // Execute the task ourselves, while workers are catching up. 547 // This allows us to hide parts of task handoff latency. 548 task->run(); 549 550 // Done executing task locally, wait for any remaining workers to complete. 551 // Once all workers report, we can proceed to termination. To do this safely, 552 // we need to make sure every worker has left. A spin-wait alone would suffice, 553 // but we do not want to burn cycles on it. A semaphore alone would not be safe, 554 // since workers can still be inside it as we proceed from wait here. So we block 555 // on semaphore first, and then spin-wait for all workers to terminate. 556 _end_semaphore.wait(); 557 SpinYield spin; 558 while (Atomic::load(&_finish_tokens) != 0) { 559 spin.wait(); 560 } 561 562 OrderAccess::fence(); 563 564 assert(Atomic::load(&_finish_tokens) == 0, "All tokens are consumed"); 565 } 566 567 void ArchiveWorkers::run_as_worker() { 568 assert(is_parallel(), "Should be in parallel mode"); 569 570 ArchiveWorkerTask* task = Atomic::load_acquire(&_task); 571 task->run(); 572 573 // All work done in threads should be visible to caller. 574 OrderAccess::fence(); 575 576 // Signal the pool the work is complete, and we are exiting. 577 // Worker cannot do anything else with the pool after this. 578 if (Atomic::sub(&_finish_tokens, 1, memory_order_relaxed) == 1) { 579 // Last worker leaving. Notify the pool it can unblock to spin-wait. 580 // Then consume the last token and leave. 581 _end_semaphore.signal(); 582 int last = Atomic::sub(&_finish_tokens, 1, memory_order_relaxed); 583 assert(last == 0, "Should be"); 584 } 585 } 586 587 void ArchiveWorkerTask::run() { 588 while (true) { 589 int chunk = Atomic::load(&_chunk); 590 if (chunk >= _max_chunks) { 591 return; 592 } 593 if (Atomic::cmpxchg(&_chunk, chunk, chunk + 1, memory_order_relaxed) == chunk) { 594 assert(0 <= chunk && chunk < _max_chunks, "Sanity"); 595 work(chunk, _max_chunks); 596 } 597 } 598 } 599 600 void ArchiveWorkerTask::configure_max_chunks(int max_chunks) { 601 if (_max_chunks == 0) { 602 _max_chunks = max_chunks; 603 } 604 } 605 606 ArchiveWorkerThread::ArchiveWorkerThread(ArchiveWorkers* pool) : NamedThread(), _pool(pool) { 607 set_name("ArchiveWorkerThread"); 608 if (os::create_thread(this, os::os_thread)) { 609 os::start_thread(this); 610 } else { 611 vm_exit_during_initialization("Unable to create archive worker", 612 os::native_thread_creation_failed_msg()); 613 } 614 } 615 616 void ArchiveWorkerThread::run() { 617 // Avalanche startup: each worker starts two others. 618 _pool->start_worker_if_needed(); 619 _pool->start_worker_if_needed(); 620 621 // Set ourselves up. 622 os::set_priority(this, NearMaxPriority); 623 624 // Work. 625 _pool->run_as_worker(); 626 } 627 628 void ArchiveWorkerThread::post_run() { 629 this->NamedThread::post_run(); 630 delete this; 631 }