1 /*
2 * Copyright (c) 2019, 2024, 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 "precompiled.hpp"
26 #include "cds/archiveBuilder.hpp"
27 #include "cds/archiveHeapLoader.inline.hpp"
28 #include "cds/archiveUtils.hpp"
29 #include "cds/cdsConfig.hpp"
30 #include "cds/classListParser.hpp"
31 #include "cds/classListWriter.hpp"
32 #include "cds/dynamicArchive.hpp"
33 #include "cds/filemap.hpp"
34 #include "cds/heapShared.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* rw_bottom = (address*)ArchiveBuilder::current()->rw_region()->base();
81 address* ro_bottom = (address*)ArchiveBuilder::current()->ro_region()->base();
82 address* cc_bottom = (address*)ArchiveBuilder::current()->cc_region()->base();
83
84 _rw_ptrmap = rw_ptrmap;
85 _ro_ptrmap = ro_ptrmap;
86 _cc_ptrmap = cc_ptrmap;
87
88 size_t rw_size = ArchiveBuilder::current()->rw_region()->used() / sizeof(address);
89 size_t ro_size = ArchiveBuilder::current()->ro_region()->used() / sizeof(address);
90 size_t cc_size = ArchiveBuilder::current()->cc_region()->used() / sizeof(address);
91 // ro_start is the first bit in _ptrmap that covers the pointer that would sit at ro_bottom.
92 // E.g., if rw_bottom = (address*)100
93 // ro_bottom = (address*)116
94 // then for 64-bit platform:
95 // ro_start = ro_bottom - rw_bottom = (116 - 100) / sizeof(address) = 2;
96 size_t ro_start = ro_bottom - rw_bottom;
97 size_t cc_start = cc_bottom - rw_bottom;
98
99 // Note: ptrmap is big enough only to cover the last pointer in cc_region or ro_region.
100 // See ArchivePtrMarker::compact()
101 if (ro_start + ro_size >_ptrmap->size()) {
102 ro_size = _ptrmap->size() - ro_start; // ro is smaller than we thought
103 cc_size = 0; // cc is empty
104 } else if (cc_size != 0 && cc_start + cc_size > _ptrmap->size()) {
105 cc_size = _ptrmap->size() - cc_start; // ro is smaller than we thought
106 }
107
108 assert(rw_size < _ptrmap->size(), "sanity");
109 assert(ro_size < _ptrmap->size(), "sanity");
110 assert(cc_size < _ptrmap->size(), "sanity");
111 assert(rw_size + ro_size + cc_size <= _ptrmap->size(), "sanity");
112
113 _rw_ptrmap->initialize(rw_size);
114 _ro_ptrmap->initialize(ro_size);
115 _cc_ptrmap->initialize(cc_size);
116
117 for (size_t i = 0; i < rw_size; i++) {
118 _rw_ptrmap->at_put(i, _ptrmap->at(i));
119 }
120 for (size_t i = 0; i < ro_size; i++) {
121 _ro_ptrmap->at_put(i, _ptrmap->at(ro_start + i));
122 }
123 for (size_t i = 0; i < cc_size; i++) {
124 _cc_ptrmap->at_put(i, _ptrmap->at(cc_start + i));
125 }
126 }
127
128 void ArchivePtrMarker::mark_pointer(address* ptr_loc) {
129 assert(_ptrmap != nullptr, "not initialized");
130 assert(!_compacted, "cannot mark anymore");
131
132 if (ptr_base() <= ptr_loc && ptr_loc < ptr_end()) {
133 address value = *ptr_loc;
134 // We don't want any pointer that points to very bottom of the archive, otherwise when
135 // MetaspaceShared::default_base_address()==0, we can't distinguish between a pointer
136 // to nothing (null) vs a pointer to an objects that happens to be at the very bottom
137 // of the archive.
138 assert(value != (address)ptr_base(), "don't point to the bottom of the archive");
139
140 if (value != nullptr) {
141 assert(uintx(ptr_loc) % sizeof(intptr_t) == 0, "pointers must be stored in aligned addresses");
142 size_t idx = ptr_loc - ptr_base();
143 if (_ptrmap->size() <= idx) {
144 _ptrmap->resize((idx + 1) * 2);
145 }
146 assert(idx < _ptrmap->size(), "must be");
147 _ptrmap->set_bit(idx);
148 //tty->print_cr("Marking pointer [" PTR_FORMAT "] -> " PTR_FORMAT " @ " SIZE_FORMAT_W(5), p2i(ptr_loc), p2i(*ptr_loc), idx);
149 }
150 }
151 }
152
153 void ArchivePtrMarker::clear_pointer(address* ptr_loc) {
154 assert(_ptrmap != nullptr, "not initialized");
155 assert(!_compacted, "cannot clear anymore");
156
157 assert(ptr_base() <= ptr_loc && ptr_loc < ptr_end(), "must be");
158 assert(uintx(ptr_loc) % sizeof(intptr_t) == 0, "pointers must be stored in aligned addresses");
159 size_t idx = ptr_loc - ptr_base();
160 assert(idx < _ptrmap->size(), "cannot clear pointers that have not been marked");
161 _ptrmap->clear_bit(idx);
162 //tty->print_cr("Clearing pointer [" PTR_FORMAT "] -> " PTR_FORMAT " @ " SIZE_FORMAT_W(5), p2i(ptr_loc), p2i(*ptr_loc), idx);
163 }
164
165 class ArchivePtrBitmapCleaner: public BitMapClosure {
166 CHeapBitMap* _ptrmap;
167 address* _ptr_base;
168 address _relocatable_base;
169 address _relocatable_end;
170 size_t _max_non_null_offset;
171
172 public:
173 ArchivePtrBitmapCleaner(CHeapBitMap* ptrmap, address* ptr_base, address relocatable_base, address relocatable_end) :
174 _ptrmap(ptrmap), _ptr_base(ptr_base),
175 _relocatable_base(relocatable_base), _relocatable_end(relocatable_end), _max_non_null_offset(0) {}
176
177 bool do_bit(size_t offset) {
178 address* ptr_loc = _ptr_base + offset;
179 address ptr_value = *ptr_loc;
180 if (ptr_value != nullptr) {
181 assert(_relocatable_base <= ptr_value && ptr_value < _relocatable_end, "do not point to arbitrary locations!");
182 if (_max_non_null_offset < offset) {
183 _max_non_null_offset = offset;
184 }
185 } else {
186 _ptrmap->clear_bit(offset);
187 DEBUG_ONLY(log_trace(cds, reloc)("Clearing pointer [" PTR_FORMAT "] -> null @ " SIZE_FORMAT_W(9), p2i(ptr_loc), offset));
188 }
189
190 return true;
191 }
192
193 size_t max_non_null_offset() const { return _max_non_null_offset; }
194 };
195
196 void ArchivePtrMarker::compact(address relocatable_base, address relocatable_end) {
197 assert(!_compacted, "cannot compact again");
198 ArchivePtrBitmapCleaner cleaner(_ptrmap, ptr_base(), relocatable_base, relocatable_end);
199 _ptrmap->iterate(&cleaner);
200 compact(cleaner.max_non_null_offset());
201 }
202
203 void ArchivePtrMarker::compact(size_t max_non_null_offset) {
204 assert(!_compacted, "cannot compact again");
205 _ptrmap->resize(max_non_null_offset + 1);
206 _compacted = true;
207 }
208
209 char* DumpRegion::expand_top_to(char* newtop) {
210 assert(is_allocatable(), "must be initialized and not packed");
211 assert(newtop >= _top, "must not grow backwards");
212 if (newtop > _end) {
213 ArchiveBuilder::current()->report_out_of_space(_name, newtop - _top);
214 ShouldNotReachHere();
215 }
216
217 commit_to(newtop);
218 _top = newtop;
219
220 if (_max_delta > 0) {
221 uintx delta = ArchiveBuilder::current()->buffer_to_offset((address)(newtop-1));
222 if (delta > _max_delta) {
223 // This is just a sanity check and should not appear in any real world usage. This
224 // happens only if you allocate more than 2GB of shared objects and would require
225 // millions of shared classes.
226 log_error(cds)("Out of memory in the CDS archive: Please reduce the number of shared classes.");
227 MetaspaceShared::unrecoverable_writing_error();
228 }
229 }
230
231 return _top;
232 }
233
234 void DumpRegion::commit_to(char* newtop) {
235 assert(CDSConfig::is_dumping_archive(), "sanity");
236 char* base = _rs->base();
237 size_t need_committed_size = newtop - base;
238 size_t has_committed_size = _vs->committed_size();
239 if (need_committed_size < has_committed_size) {
240 return;
241 }
242
243 size_t min_bytes = need_committed_size - has_committed_size;
244 size_t preferred_bytes = 1 * M;
245 size_t uncommitted = _vs->reserved_size() - has_committed_size;
246
247 size_t commit = MAX2(min_bytes, preferred_bytes);
248 commit = MIN2(commit, uncommitted);
249 assert(commit <= uncommitted, "sanity");
250
251 if (!_vs->expand_by(commit, false)) {
252 log_error(cds)("Failed to expand shared space to " SIZE_FORMAT " bytes",
253 need_committed_size);
254 MetaspaceShared::unrecoverable_writing_error();
255 }
256
257 const char* which;
258 if (_rs->base() == (char*)MetaspaceShared::symbol_rs_base()) {
259 which = "symbol";
260 } else {
261 which = "shared";
262 }
263 log_debug(cds)("Expanding %s spaces by " SIZE_FORMAT_W(7) " bytes [total " SIZE_FORMAT_W(9) " bytes ending at %p]",
264 which, commit, _vs->actual_committed_size(), _vs->high());
265 }
266
267 char* DumpRegion::allocate(size_t num_bytes, size_t alignment) {
268 // Always align to at least minimum alignment
269 alignment = MAX2(SharedSpaceObjectAlignment, alignment);
270 char* p = (char*)align_up(_top, alignment);
271 char* newtop = p + align_up(num_bytes, (size_t)SharedSpaceObjectAlignment);
272 expand_top_to(newtop);
273 memset(p, 0, newtop - p);
274 return p;
275 }
276
277 void DumpRegion::append_intptr_t(intptr_t n, bool need_to_mark) {
278 assert(is_aligned(_top, sizeof(intptr_t)), "bad alignment");
279 intptr_t *p = (intptr_t*)_top;
280 char* newtop = _top + sizeof(intptr_t);
281 expand_top_to(newtop);
282 *p = n;
283 if (need_to_mark) {
284 ArchivePtrMarker::mark_pointer(p);
285 }
286 }
287
288 void DumpRegion::print(size_t total_bytes) const {
289 char* base = used() > 0 ? ArchiveBuilder::current()->to_requested(_base) : nullptr;
290 log_debug(cds)("%s space: " SIZE_FORMAT_W(9) " [ %4.1f%% of total] out of " SIZE_FORMAT_W(9) " bytes [%5.1f%% used] at " INTPTR_FORMAT,
291 _name, used(), percent_of(used(), total_bytes), reserved(), percent_of(used(), reserved()),
292 p2i(base));
293 }
294
295 void DumpRegion::print_out_of_space_msg(const char* failing_region, size_t needed_bytes) {
296 log_error(cds)("[%-8s] " PTR_FORMAT " - " PTR_FORMAT " capacity =%9d, allocated =%9d",
297 _name, p2i(_base), p2i(_top), int(_end - _base), int(_top - _base));
298 if (strcmp(_name, failing_region) == 0) {
299 log_error(cds)(" required = %d", int(needed_bytes));
300 }
301 }
302
303 void DumpRegion::init(ReservedSpace* rs, VirtualSpace* vs) {
304 _rs = rs;
305 _vs = vs;
306 // Start with 0 committed bytes. The memory will be committed as needed.
307 if (!_vs->initialize(*_rs, 0)) {
308 fatal("Unable to allocate memory for shared space");
309 }
310 _base = _top = _rs->base();
311 _end = _rs->end();
312 }
313
314 void DumpRegion::pack(DumpRegion* next) {
315 if (!is_packed()) {
316 _end = (char*)align_up(_top, MetaspaceShared::core_region_alignment());
317 _is_packed = true;
318 }
319 if (next != nullptr) {
320 next->_rs = _rs;
321 next->_vs = _vs;
322 next->_base = next->_top = this->_end;
323 next->_end = _rs->end();
324 }
325 }
326
327 void WriteClosure::do_ptr(void** p) {
328 // Write ptr into the archive; ptr can be:
329 // (a) null -> written as 0
330 // (b) a "buffered" address -> written as is
331 // (c) a "source" address -> convert to "buffered" and write
332 // The common case is (c). E.g., when writing the vmClasses into the archive.
333 // We have (b) only when we don't have a corresponding source object. E.g.,
334 // the archived c++ vtable entries.
335 address ptr = *(address*)p;
336 if (ptr != nullptr && !ArchiveBuilder::current()->is_in_buffer_space(ptr)) {
337 ptr = ArchiveBuilder::current()->get_buffered_addr(ptr);
338 }
339 // null pointers do not need to be converted to offsets
340 if (ptr != nullptr) {
341 ptr = (address)ArchiveBuilder::current()->buffer_to_offset(ptr);
342 }
343 _dump_region->append_intptr_t((intptr_t)ptr, false);
344 }
345
346 void ReadClosure::do_ptr(void** p) {
347 assert(*p == nullptr, "initializing previous initialized pointer.");
348 intptr_t obj = nextPtr();
349 assert(obj >= 0, "sanity.");
350 *p = (obj != 0) ? (void*)(_base_address + obj) : (void*)obj;
351 }
352
353 void ReadClosure::do_u4(u4* p) {
354 intptr_t obj = nextPtr();
355 *p = (u4)(uintx(obj));
356 }
357
358 void ReadClosure::do_int(int* p) {
359 intptr_t obj = nextPtr();
360 *p = (int)(intx(obj));
361 }
362
363 void ReadClosure::do_bool(bool* p) {
364 intptr_t obj = nextPtr();
365 *p = (bool)(uintx(obj));
366 }
367
368 void ReadClosure::do_tag(int tag) {
369 int old_tag;
370 old_tag = (int)(intptr_t)nextPtr();
371 // do_int(&old_tag);
372 assert(tag == old_tag, "tag doesn't match (%d, expected %d)", old_tag, tag);
373 FileMapInfo::assert_mark(tag == old_tag);
374 }
375
376 void ArchiveUtils::log_to_classlist(BootstrapInfo* bootstrap_specifier, TRAPS) {
377 if (ClassListWriter::is_enabled()) {
378 if (SystemDictionaryShared::is_supported_invokedynamic(bootstrap_specifier)) {
379 const constantPoolHandle& pool = bootstrap_specifier->pool();
380 if (SystemDictionaryShared::is_builtin_loader(pool->pool_holder()->class_loader_data())) {
381 // Currently lambda proxy classes are supported only for the built-in loaders.
382 ResourceMark rm(THREAD);
383 int pool_index = bootstrap_specifier->bss_index();
384 ClassListWriter w;
385 w.stream()->print("%s %s", ClassListParser::lambda_proxy_tag(), pool->pool_holder()->name()->as_C_string());
386 CDSIndyInfo cii;
387 ClassListParser::populate_cds_indy_info(pool, pool_index, &cii, CHECK);
388 GrowableArray<const char*>* indy_items = cii.items();
389 for (int i = 0; i < indy_items->length(); i++) {
390 w.stream()->print(" %s", indy_items->at(i));
391 }
392 w.stream()->cr();
393 }
394 }
395 }
396 }
397
398
399 // "boot", "platform", "app" or nullptr
400 const char* ArchiveUtils::builtin_loader_name_or_null(oop loader) {
401 if (loader == nullptr) {
402 return "boot";
403 } else if (loader == SystemDictionary::java_platform_loader()) {
404 return "platform";
405 } else if (loader == SystemDictionary::java_system_loader()) {
406 return "app";
407 } else {
408 return nullptr;
409 }
410 }
411
412 // "boot", "platform", "app". Asserts if not a built-in-loader
413 const char* ArchiveUtils::builtin_loader_name(oop loader) {
414 const char* name = builtin_loader_name_or_null(loader);
415 assert(name != nullptr, "must be a built-in loader");
416 return name;
417 }
418
419 bool ArchiveUtils::builtin_loader_from_type(const char* loader_type, oop* value_ret) {
420 if (strcmp(loader_type, "boot") == 0) {
421 *value_ret = nullptr;
422 return true;
423 } else if (strcmp(loader_type, "platform") == 0) {
424 *value_ret = SystemDictionary::java_platform_loader();
425 return true;
426 } else if (strcmp(loader_type, "app") == 0) {
427 *value_ret = SystemDictionary::java_system_loader();
428 return true;
429 } else {
430 DEBUG_ONLY(*value_ret = cast_to_oop((void*)badOopVal));
431 return false;
432 }
433 }
434
435 oop ArchiveUtils::builtin_loader_from_type(int loader_type) {
436 if (loader_type == ClassLoader::BOOT_LOADER) {
437 return nullptr;
438 } else if (loader_type == ClassLoader::PLATFORM_LOADER) {
439 return SystemDictionary::java_platform_loader();
440 } else if (loader_type == ClassLoader::APP_LOADER) {
441 return SystemDictionary::java_system_loader();
442 } else {
443 ShouldNotReachHere();
444 return nullptr;
445 }
446 }
447
448 bool ArchiveUtils::has_aot_initialized_mirror(InstanceKlass* src_ik) {
449 if (SystemDictionaryShared::is_excluded_class(src_ik)) {
450 assert(!ArchiveBuilder::current()->has_been_buffered(src_ik), "sanity");
451 return false;
452 }
453 return ArchiveBuilder::current()->get_buffered_addr(src_ik)->has_aot_initialized_mirror();
454 }
455
456 size_t HeapRootSegments::size_in_bytes(size_t seg_idx) {
457 assert(seg_idx < _count, "In range");
458 return objArrayOopDesc::object_size(size_in_elems(seg_idx)) * HeapWordSize;
459 }
460
461 int HeapRootSegments::size_in_elems(size_t seg_idx) {
462 assert(seg_idx < _count, "In range");
463 if (seg_idx != _count - 1) {
464 return _max_size_in_elems;
465 } else {
466 // Last slice, leftover
467 return _roots_count % _max_size_in_elems;
468 }
469 }
470
471 size_t HeapRootSegments::segment_offset(size_t seg_idx) {
472 assert(seg_idx < _count, "In range");
473 return _base_offset + seg_idx * _max_size_in_bytes;
474 }
475
476 ArchiveWorkers::ArchiveWorkers() :
477 _end_semaphore(0),
478 _num_workers(max_workers()),
479 _started_workers(0),
480 _finish_tokens(0),
481 _state(UNUSED),
482 _task(nullptr) {}
483
484 ArchiveWorkers::~ArchiveWorkers() {
485 assert(Atomic::load(&_state) != WORKING, "Should not be working");
486 }
487
488 int ArchiveWorkers::max_workers() {
489 // The pool is used for short-lived bursty tasks. We do not want to spend
490 // too much time creating and waking up threads unnecessarily. Plus, we do
491 // not want to overwhelm large machines. This is why we want to be very
492 // conservative about the number of workers actually needed.
493 return MAX2(0, log2i_graceful(os::active_processor_count()));
494 }
495
496 bool ArchiveWorkers::is_parallel() {
497 return _num_workers > 0;
498 }
499
500 void ArchiveWorkers::start_worker_if_needed() {
501 while (true) {
502 int cur = Atomic::load(&_started_workers);
503 if (cur >= _num_workers) {
504 return;
505 }
506 if (Atomic::cmpxchg(&_started_workers, cur, cur + 1, memory_order_relaxed) == cur) {
507 new ArchiveWorkerThread(this);
508 return;
509 }
510 }
511 }
512
513 void ArchiveWorkers::run_task(ArchiveWorkerTask* task) {
514 assert(Atomic::load(&_state) == UNUSED, "Should be unused yet");
515 assert(Atomic::load(&_task) == nullptr, "Should not have running tasks");
516 Atomic::store(&_state, WORKING);
517
518 if (is_parallel()) {
519 run_task_multi(task);
520 } else {
521 run_task_single(task);
522 }
523
524 assert(Atomic::load(&_state) == WORKING, "Should be working");
525 Atomic::store(&_state, SHUTDOWN);
526 }
527
528 void ArchiveWorkers::run_task_single(ArchiveWorkerTask* task) {
529 // Single thread needs no chunking.
530 task->configure_max_chunks(1);
531
532 // Execute the task ourselves, as there are no workers.
533 task->work(0, 1);
534 }
535
536 void ArchiveWorkers::run_task_multi(ArchiveWorkerTask* task) {
537 // Multiple threads can work with multiple chunks.
538 task->configure_max_chunks(_num_workers * CHUNKS_PER_WORKER);
539
540 // Set up the run and publish the task. Issue one additional finish token
541 // to cover the semaphore shutdown path, see below.
542 Atomic::store(&_finish_tokens, _num_workers + 1);
543 Atomic::release_store(&_task, task);
544
545 // Kick off pool startup by starting a single worker, and proceed
546 // immediately to executing the task locally.
547 start_worker_if_needed();
548
549 // Execute the task ourselves, while workers are catching up.
550 // This allows us to hide parts of task handoff latency.
551 task->run();
552
553 // Done executing task locally, wait for any remaining workers to complete.
554 // Once all workers report, we can proceed to termination. To do this safely,
555 // we need to make sure every worker has left. A spin-wait alone would suffice,
556 // but we do not want to burn cycles on it. A semaphore alone would not be safe,
557 // since workers can still be inside it as we proceed from wait here. So we block
558 // on semaphore first, and then spin-wait for all workers to terminate.
559 _end_semaphore.wait();
560 SpinYield spin;
561 while (Atomic::load(&_finish_tokens) != 0) {
562 spin.wait();
563 }
564
565 OrderAccess::fence();
566
567 assert(Atomic::load(&_finish_tokens) == 0, "All tokens are consumed");
568 }
569
570 void ArchiveWorkers::run_as_worker() {
571 assert(is_parallel(), "Should be in parallel mode");
572
573 ArchiveWorkerTask* task = Atomic::load_acquire(&_task);
574 task->run();
575
576 // All work done in threads should be visible to caller.
577 OrderAccess::fence();
578
579 // Signal the pool the work is complete, and we are exiting.
580 // Worker cannot do anything else with the pool after this.
581 if (Atomic::sub(&_finish_tokens, 1, memory_order_relaxed) == 1) {
582 // Last worker leaving. Notify the pool it can unblock to spin-wait.
583 // Then consume the last token and leave.
584 _end_semaphore.signal();
585 int last = Atomic::sub(&_finish_tokens, 1, memory_order_relaxed);
586 assert(last == 0, "Should be");
587 }
588 }
589
590 void ArchiveWorkerTask::run() {
591 while (true) {
592 int chunk = Atomic::load(&_chunk);
593 if (chunk >= _max_chunks) {
594 return;
595 }
596 if (Atomic::cmpxchg(&_chunk, chunk, chunk + 1, memory_order_relaxed) == chunk) {
597 assert(0 <= chunk && chunk < _max_chunks, "Sanity");
598 work(chunk, _max_chunks);
599 }
600 }
601 }
602
603 void ArchiveWorkerTask::configure_max_chunks(int max_chunks) {
604 if (_max_chunks == 0) {
605 _max_chunks = max_chunks;
606 }
607 }
608
609 ArchiveWorkerThread::ArchiveWorkerThread(ArchiveWorkers* pool) : NamedThread(), _pool(pool) {
610 set_name("ArchiveWorkerThread");
611 if (os::create_thread(this, os::os_thread)) {
612 os::start_thread(this);
613 } else {
614 vm_exit_during_initialization("Unable to create archive worker",
615 os::native_thread_creation_failed_msg());
616 }
617 }
618
619 void ArchiveWorkerThread::run() {
620 // Avalanche startup: each worker starts two others.
621 _pool->start_worker_if_needed();
622 _pool->start_worker_if_needed();
623
624 // Set ourselves up.
625 os::set_priority(this, NearMaxPriority);
626
627 // Work.
628 _pool->run_as_worker();
629 }
630
631 void ArchiveWorkerThread::post_run() {
632 this->NamedThread::post_run();
633 delete this;
634 }