1 /*
2 * Copyright (c) 2018, 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 "classfile/javaClasses.inline.hpp"
26 #include "classfile/vmSymbols.hpp"
27 #include "code/codeCache.inline.hpp"
28 #include "code/nmethod.inline.hpp"
29 #include "code/vmreg.inline.hpp"
30 #include "compiler/oopMap.inline.hpp"
31 #include "gc/shared/continuationGCSupport.inline.hpp"
32 #include "gc/shared/gc_globals.hpp"
33 #include "gc/shared/barrierSet.hpp"
34 #include "gc/shared/memAllocator.hpp"
35 #include "gc/shared/threadLocalAllocBuffer.inline.hpp"
36 #include "interpreter/interpreter.hpp"
37 #include "jfr/jfrEvents.hpp"
38 #include "logging/log.hpp"
39 #include "logging/logStream.hpp"
40 #include "oops/access.inline.hpp"
41 #include "oops/method.inline.hpp"
42 #include "oops/oopsHierarchy.hpp"
43 #include "oops/objArrayOop.inline.hpp"
44 #include "oops/stackChunkOop.inline.hpp"
45 #include "prims/jvmtiThreadState.hpp"
46 #include "runtime/arguments.hpp"
47 #include "runtime/continuation.hpp"
48 #include "runtime/continuationEntry.inline.hpp"
49 #include "runtime/continuationHelper.inline.hpp"
50 #include "runtime/continuationJavaClasses.inline.hpp"
51 #include "runtime/continuationWrapper.inline.hpp"
52 #include "runtime/frame.inline.hpp"
53 #include "runtime/interfaceSupport.inline.hpp"
54 #include "runtime/javaThread.inline.hpp"
55 #include "runtime/jniHandles.inline.hpp"
56 #include "runtime/keepStackGCProcessed.hpp"
57 #include "runtime/objectMonitor.inline.hpp"
58 #include "runtime/orderAccess.hpp"
59 #include "runtime/prefetch.inline.hpp"
60 #include "runtime/smallRegisterMap.inline.hpp"
61 #include "runtime/sharedRuntime.hpp"
62 #include "runtime/stackChunkFrameStream.inline.hpp"
63 #include "runtime/stackFrameStream.inline.hpp"
64 #include "runtime/stackOverflow.hpp"
65 #include "runtime/stackWatermarkSet.inline.hpp"
66 #include "utilities/debug.hpp"
67 #include "utilities/exceptions.hpp"
68 #include "utilities/macros.hpp"
69 #include "utilities/vmError.hpp"
70 #if INCLUDE_ZGC
71 #include "gc/z/zStackChunkGCData.inline.hpp"
72 #endif
73
74 #include <type_traits>
75
76 /*
77 * This file contains the implementation of continuation freezing (yield) and thawing (run).
78 *
79 * This code is very latency-critical and very hot. An ordinary and well-behaved server application
80 * would likely call these operations many thousands of times per second second, on every core.
81 *
82 * Freeze might be called every time the application performs any I/O operation, every time it
83 * acquires a j.u.c. lock, every time it takes a message from a queue, and thaw can be called
84 * multiple times in each of those cases, as it is called by the return barrier, which may be
85 * invoked on method return.
86 *
87 * The amortized budget for each of those two operations is ~100-150ns. That is why, for
88 * example, every effort is made to avoid Java-VM transitions as much as possible.
89 *
90 * On the fast path, all frames are known to be compiled, and the chunk requires no barriers
91 * and so frames simply copied, and the bottom-most one is patched.
92 * On the slow path, internal pointers in interpreted frames are de/relativized to/from offsets
93 * and absolute pointers, and barriers invoked.
94 */
95
96 /************************************************
97
98 Thread-stack layout on freeze/thaw.
99 See corresponding stack-chunk layout in instanceStackChunkKlass.hpp
100
101 +----------------------------+
102 | . |
103 | . |
104 | . |
105 | carrier frames |
106 | |
107 |----------------------------|
108 | |
109 | Continuation.run |
110 | |
111 |============================|
112 | enterSpecial frame |
113 | pc |
114 | rbp |
115 | ----- |
116 ^ | int argsize | = ContinuationEntry
117 | | oopDesc* cont |
118 | | oopDesc* chunk |
119 | | ContinuationEntry* parent |
120 | | ... |
121 | |============================| <------ JavaThread::_cont_entry = entry->sp()
122 | | ? alignment word ? |
123 | |----------------------------| <--\
124 | | | |
125 | | ? caller stack args ? | | argsize (might not be 2-word aligned) words
126 Address | | | | Caller is still in the chunk.
127 | |----------------------------| |
128 | | pc (? return barrier ?) | | This pc contains the return barrier when the bottom-most frame
129 | | rbp | | isn't the last one in the continuation.
130 | | | |
131 | | frame | |
132 | | | |
133 +----------------------------| \__ Continuation frames to be frozen/thawed
134 | | /
135 | frame | |
136 | | |
137 |----------------------------| |
138 | | |
139 | frame | |
140 | | |
141 |----------------------------| <--/
142 | |
143 | doYield/safepoint stub | When preempting forcefully, we could have a safepoint stub
144 | | instead of a doYield stub
145 |============================| <- the sp passed to freeze
146 | |
147 | Native freeze/thaw frames |
148 | . |
149 | . |
150 | . |
151 +----------------------------+
152
153 ************************************************/
154
155 static const bool TEST_THAW_ONE_CHUNK_FRAME = false; // force thawing frames one-at-a-time for testing
156
157 #define CONT_JFR false // emit low-level JFR events that count slow/fast path for continuation performance debugging only
158 #if CONT_JFR
159 #define CONT_JFR_ONLY(code) code
160 #else
161 #define CONT_JFR_ONLY(code)
162 #endif
163
164 // TODO: See AbstractAssembler::generate_stack_overflow_check,
165 // Compile::bang_size_in_bytes(), m->as_SafePoint()->jvms()->interpreter_frame_size()
166 // when we stack-bang, we need to update a thread field with the lowest (farthest) bang point.
167
168 // Data invariants are defined by Continuation::debug_verify_continuation and Continuation::debug_verify_stack_chunk
169
170 // Used to just annotatate cold/hot branches
171 #define LIKELY(condition) (condition)
172 #define UNLIKELY(condition) (condition)
173
174 // debugging functions
175 #ifdef ASSERT
176 extern "C" bool dbg_is_safe(const void* p, intptr_t errvalue); // address p is readable and *(intptr_t*)p != errvalue
177
178 static void verify_continuation(oop continuation) { Continuation::debug_verify_continuation(continuation); }
179
180 static void do_deopt_after_thaw(JavaThread* thread);
181 static bool do_verify_after_thaw(JavaThread* thread, stackChunkOop chunk, outputStream* st);
182 static void log_frames(JavaThread* thread);
183 static void log_frames_after_thaw(JavaThread* thread, ContinuationWrapper& cont, intptr_t* sp, bool preempted);
184 static void print_frame_layout(const frame& f, bool callee_complete, outputStream* st = tty);
185
186 #define assert_pfl(p, ...) \
187 do { \
188 if (!(p)) { \
189 JavaThread* t = JavaThread::active(); \
190 if (t->has_last_Java_frame()) { \
191 tty->print_cr("assert(" #p ") failed:"); \
192 t->print_frame_layout(); \
193 } \
194 } \
195 vmassert(p, __VA_ARGS__); \
196 } while(0)
197
198 #else
199 static void verify_continuation(oop continuation) { }
200 #define assert_pfl(p, ...)
201 #endif
202
203 static freeze_result is_pinned0(JavaThread* thread, oop cont_scope, bool safepoint);
204 template<typename ConfigT, bool preempt> static inline freeze_result freeze_internal(JavaThread* current, intptr_t* const sp);
205
206 static inline int prepare_thaw_internal(JavaThread* thread, bool return_barrier);
207 template<typename ConfigT> static inline intptr_t* thaw_internal(JavaThread* thread, const Continuation::thaw_kind kind);
208
209
210 // Entry point to freeze. Transitions are handled manually
211 // Called from gen_continuation_yield() in sharedRuntime_<cpu>.cpp through Continuation::freeze_entry();
212 template<typename ConfigT>
213 static JRT_BLOCK_ENTRY(int, freeze(JavaThread* current, intptr_t* sp))
214 assert(sp == current->frame_anchor()->last_Java_sp(), "");
215
216 if (current->raw_cont_fastpath() > current->last_continuation()->entry_sp() || current->raw_cont_fastpath() < sp) {
217 current->set_cont_fastpath(nullptr);
218 }
219
220 return checked_cast<int>(ConfigT::freeze(current, sp));
221 JRT_END
222
223 JRT_LEAF(int, Continuation::prepare_thaw(JavaThread* thread, bool return_barrier))
224 return prepare_thaw_internal(thread, return_barrier);
225 JRT_END
226
227 template<typename ConfigT>
228 static JRT_LEAF(intptr_t*, thaw(JavaThread* thread, int kind))
229 // TODO: JRT_LEAF and NoHandleMark is problematic for JFR events.
230 // vFrameStreamCommon allocates Handles in RegisterMap for continuations.
231 // Also the preemption case with JVMTI events enabled might safepoint so
232 // undo the NoSafepointVerifier here and rely on handling by ContinuationWrapper.
233 // JRT_ENTRY instead?
234 ResetNoHandleMark rnhm;
235 DEBUG_ONLY(PauseNoSafepointVerifier pnsv(&__nsv);)
236
237 // we might modify the code cache via BarrierSetNMethod::nmethod_entry_barrier
238 MACOS_AARCH64_ONLY(ThreadWXEnable __wx(WXWrite, thread));
239 return ConfigT::thaw(thread, (Continuation::thaw_kind)kind);
240 JRT_END
241
242 JVM_ENTRY(jint, CONT_isPinned0(JNIEnv* env, jobject cont_scope)) {
243 JavaThread* thread = JavaThread::thread_from_jni_environment(env);
244 return is_pinned0(thread, JNIHandles::resolve(cont_scope), false);
245 }
246 JVM_END
247
248 ///////////
249
250 enum class oop_kind { NARROW, WIDE };
251 template <oop_kind oops, typename BarrierSetT>
252 class Config {
253 public:
254 typedef Config<oops, BarrierSetT> SelfT;
255 using OopT = std::conditional_t<oops == oop_kind::NARROW, narrowOop, oop>;
256
257 static freeze_result freeze(JavaThread* thread, intptr_t* const sp) {
258 freeze_result res = freeze_internal<SelfT, false>(thread, sp);
259 JFR_ONLY(assert((res == freeze_ok) || (res == thread->last_freeze_fail_result()), "freeze failure not set"));
260 return res;
261 }
262
263 static freeze_result freeze_preempt(JavaThread* thread, intptr_t* const sp) {
264 return freeze_internal<SelfT, true>(thread, sp);
265 }
266
267 static intptr_t* thaw(JavaThread* thread, Continuation::thaw_kind kind) {
268 return thaw_internal<SelfT>(thread, kind);
269 }
270 };
271
272 #ifdef _WINDOWS
273 static void map_stack_pages(JavaThread* thread, size_t size, address sp) {
274 address new_sp = sp - size;
275 address watermark = thread->stack_overflow_state()->shadow_zone_growth_watermark();
276
277 if (new_sp < watermark) {
278 size_t page_size = os::vm_page_size();
279 address last_touched_page = watermark - StackOverflow::stack_shadow_zone_size();
280 size_t pages_to_touch = align_up(watermark - new_sp, page_size) / page_size;
281 while (pages_to_touch-- > 0) {
282 last_touched_page -= page_size;
283 *last_touched_page = 0;
284 }
285 thread->stack_overflow_state()->set_shadow_zone_growth_watermark(new_sp);
286 }
287 }
288 #endif
289
290 static bool stack_overflow_check(JavaThread* thread, size_t size, address sp) {
291 const size_t page_size = os::vm_page_size();
292 if (size > page_size) {
293 if (sp - size < thread->stack_overflow_state()->shadow_zone_safe_limit()) {
294 return false;
295 }
296 WINDOWS_ONLY(map_stack_pages(thread, size, sp));
297 }
298 return true;
299 }
300
301 #ifdef ASSERT
302 static oop get_continuation(JavaThread* thread) {
303 assert(thread != nullptr, "");
304 assert(thread->threadObj() != nullptr, "");
305 return java_lang_Thread::continuation(thread->threadObj());
306 }
307 #endif // ASSERT
308
309 inline void clear_anchor(JavaThread* thread) {
310 thread->frame_anchor()->clear();
311 }
312
313 static void set_anchor(JavaThread* thread, intptr_t* sp, address pc) {
314 assert(pc != nullptr, "");
315
316 JavaFrameAnchor* anchor = thread->frame_anchor();
317 anchor->set_last_Java_sp(sp);
318 anchor->set_last_Java_pc(pc);
319 ContinuationHelper::set_anchor_pd(anchor, sp);
320
321 assert(thread->has_last_Java_frame(), "");
322 assert(thread->last_frame().cb() != nullptr, "");
323 }
324
325 static void set_anchor(JavaThread* thread, intptr_t* sp) {
326 address pc = ContinuationHelper::return_address_at(
327 sp - frame::sender_sp_ret_address_offset());
328 set_anchor(thread, sp, pc);
329 }
330
331 static void set_anchor_to_entry(JavaThread* thread, ContinuationEntry* entry) {
332 JavaFrameAnchor* anchor = thread->frame_anchor();
333 anchor->set_last_Java_sp(entry->entry_sp());
334 anchor->set_last_Java_pc(entry->entry_pc());
335 ContinuationHelper::set_anchor_to_entry_pd(anchor, entry);
336
337 assert(thread->has_last_Java_frame(), "");
338 assert(thread->last_frame().cb() != nullptr, "");
339 }
340
341 #if CONT_JFR
342 class FreezeThawJfrInfo : public StackObj {
343 short _e_size;
344 short _e_num_interpreted_frames;
345 public:
346
347 FreezeThawJfrInfo() : _e_size(0), _e_num_interpreted_frames(0) {}
348 inline void record_interpreted_frame() { _e_num_interpreted_frames++; }
349 inline void record_size_copied(int size) { _e_size += size << LogBytesPerWord; }
350 template<typename Event> void post_jfr_event(Event *e, oop continuation, JavaThread* jt);
351 };
352
353 template<typename Event> void FreezeThawJfrInfo::post_jfr_event(Event* e, oop continuation, JavaThread* jt) {
354 if (e->should_commit()) {
355 log_develop_trace(continuations)("JFR event: iframes: %d size: %d", _e_num_interpreted_frames, _e_size);
356 e->set_carrierThread(JFR_JVM_THREAD_ID(jt));
357 e->set_continuationClass(continuation->klass());
358 e->set_interpretedFrames(_e_num_interpreted_frames);
359 e->set_size(_e_size);
360 e->commit();
361 }
362 }
363 #endif // CONT_JFR
364
365 /////////////// FREEZE ////
366
367 class FreezeBase : public StackObj {
368 protected:
369 JavaThread* const _thread;
370 ContinuationWrapper& _cont;
371 bool _barriers; // only set when we allocate a chunk
372
373 intptr_t* _bottom_address;
374
375 // Used for preemption only
376 const bool _preempt;
377 frame _last_frame;
378
379 // Used to support freezing with held monitors
380 int _monitors_in_lockstack;
381
382 int _freeze_size; // total size of all frames plus metadata in words.
383 int _total_align_size;
384
385 intptr_t* _cont_stack_top;
386 intptr_t* _cont_stack_bottom;
387
388 CONT_JFR_ONLY(FreezeThawJfrInfo _jfr_info;)
389
390 #ifdef ASSERT
391 intptr_t* _orig_chunk_sp;
392 int _fast_freeze_size;
393 bool _empty;
394 #endif
395
396 JvmtiSampledObjectAllocEventCollector* _jvmti_event_collector;
397
398 NOT_PRODUCT(int _frames;)
399 DEBUG_ONLY(intptr_t* _last_write;)
400
401 inline FreezeBase(JavaThread* thread, ContinuationWrapper& cont, intptr_t* sp, bool preempt);
402
403 public:
404 NOINLINE freeze_result freeze_slow();
405 void freeze_fast_existing_chunk();
406
407 CONT_JFR_ONLY(FreezeThawJfrInfo& jfr_info() { return _jfr_info; })
408 void set_jvmti_event_collector(JvmtiSampledObjectAllocEventCollector* jsoaec) { _jvmti_event_collector = jsoaec; }
409
410 inline int size_if_fast_freeze_available();
411
412 inline frame& last_frame() { return _last_frame; }
413
414 #ifdef ASSERT
415 bool check_valid_fast_path();
416 #endif
417
418 protected:
419 inline void init_rest();
420 void throw_stack_overflow_on_humongous_chunk();
421
422 // fast path
423 inline void copy_to_chunk(intptr_t* from, intptr_t* to, int size);
424 inline void unwind_frames();
425 inline void patch_stack_pd(intptr_t* frame_sp, intptr_t* heap_sp);
426
427 // slow path
428 virtual stackChunkOop allocate_chunk_slow(size_t stack_size, int argsize_md) = 0;
429
430 int cont_size() { return pointer_delta_as_int(_cont_stack_bottom, _cont_stack_top); }
431
432 private:
433 // slow path
434 frame freeze_start_frame();
435 frame freeze_start_frame_on_preempt();
436 NOINLINE freeze_result recurse_freeze(frame& f, frame& caller, int callee_argsize, bool callee_interpreted, bool top);
437 inline frame freeze_start_frame_yield_stub();
438 template<typename FKind>
439 inline freeze_result recurse_freeze_java_frame(const frame& f, frame& caller, int fsize, int argsize);
440 inline void before_freeze_java_frame(const frame& f, const frame& caller, int fsize, int argsize, bool is_bottom_frame);
441 inline void after_freeze_java_frame(const frame& hf, bool is_bottom_frame);
442 freeze_result finalize_freeze(const frame& callee, frame& caller, int argsize);
443 void patch(const frame& f, frame& hf, const frame& caller, bool is_bottom_frame);
444 NOINLINE freeze_result recurse_freeze_interpreted_frame(frame& f, frame& caller, int callee_argsize, bool callee_interpreted);
445 freeze_result recurse_freeze_compiled_frame(frame& f, frame& caller, int callee_argsize, bool callee_interpreted);
446 NOINLINE freeze_result recurse_freeze_stub_frame(frame& f, frame& caller);
447 NOINLINE freeze_result recurse_freeze_native_frame(frame& f, frame& caller);
448 NOINLINE void finish_freeze(const frame& f, const frame& top);
449
450 void freeze_lockstack(stackChunkOop chunk);
451
452 inline bool stack_overflow();
453
454 static frame sender(const frame& f) { return f.is_interpreted_frame() ? sender<ContinuationHelper::InterpretedFrame>(f)
455 : sender<ContinuationHelper::NonInterpretedUnknownFrame>(f); }
456 template<typename FKind> static inline frame sender(const frame& f);
457 template<typename FKind> frame new_heap_frame(frame& f, frame& caller);
458 inline void set_top_frame_metadata_pd(const frame& hf);
459 inline void patch_pd(frame& callee, const frame& caller);
460 void adjust_interpreted_frame_unextended_sp(frame& f);
461 static inline void prepare_freeze_interpreted_top_frame(frame& f);
462 static inline void relativize_interpreted_frame_metadata(const frame& f, const frame& hf);
463
464 protected:
465 void freeze_fast_copy(stackChunkOop chunk, int chunk_start_sp CONT_JFR_ONLY(COMMA bool chunk_is_allocated));
466 bool freeze_fast_new_chunk(stackChunkOop chunk);
467 };
468
469 template <typename ConfigT>
470 class Freeze : public FreezeBase {
471 private:
472 stackChunkOop allocate_chunk(size_t stack_size, int argsize_md);
473
474 public:
475 inline Freeze(JavaThread* thread, ContinuationWrapper& cont, intptr_t* frame_sp, bool preempt)
476 : FreezeBase(thread, cont, frame_sp, preempt) {}
477
478 freeze_result try_freeze_fast();
479
480 protected:
481 virtual stackChunkOop allocate_chunk_slow(size_t stack_size, int argsize_md) override { return allocate_chunk(stack_size, argsize_md); }
482 };
483
484 FreezeBase::FreezeBase(JavaThread* thread, ContinuationWrapper& cont, intptr_t* frame_sp, bool preempt) :
485 _thread(thread), _cont(cont), _barriers(false), _preempt(preempt), _last_frame(false /* no initialization */) {
486 DEBUG_ONLY(_jvmti_event_collector = nullptr;)
487
488 assert(_thread != nullptr, "");
489 assert(_thread->last_continuation()->entry_sp() == _cont.entrySP(), "");
490
491 DEBUG_ONLY(_cont.entry()->verify_cookie();)
492
493 assert(!Interpreter::contains(_cont.entryPC()), "");
494
495 _bottom_address = _cont.entrySP() - _cont.entry_frame_extension();
496 #ifdef _LP64
497 if (((intptr_t)_bottom_address & 0xf) != 0) {
498 _bottom_address--;
499 }
500 assert(is_aligned(_bottom_address, frame::frame_alignment), "");
501 #endif
502
503 log_develop_trace(continuations)("bottom_address: " INTPTR_FORMAT " entrySP: " INTPTR_FORMAT " argsize: " PTR_FORMAT,
504 p2i(_bottom_address), p2i(_cont.entrySP()), (_cont.entrySP() - _bottom_address) << LogBytesPerWord);
505 assert(_bottom_address != nullptr, "");
506 assert(_bottom_address <= _cont.entrySP(), "");
507 DEBUG_ONLY(_last_write = nullptr;)
508
509 assert(_cont.chunk_invariant(), "");
510 assert(!Interpreter::contains(_cont.entryPC()), "");
511 #if !defined(PPC64) || defined(ZERO)
512 static const int doYield_stub_frame_size = frame::metadata_words;
513 #else
514 static const int doYield_stub_frame_size = frame::native_abi_reg_args_size >> LogBytesPerWord;
515 #endif
516 // With preemption doYield() might not have been resolved yet
517 assert(_preempt || SharedRuntime::cont_doYield_stub()->frame_size() == doYield_stub_frame_size, "");
518
519 if (preempt) {
520 _last_frame = _thread->last_frame();
521 }
522
523 // properties of the continuation on the stack; all sizes are in words
524 _cont_stack_top = frame_sp + (!preempt ? doYield_stub_frame_size : 0); // we don't freeze the doYield stub frame
525 _cont_stack_bottom = _cont.entrySP() + (_cont.argsize() == 0 ? frame::metadata_words_at_top : 0)
526 - ContinuationHelper::frame_align_words(_cont.argsize()); // see alignment in thaw
527
528 log_develop_trace(continuations)("freeze size: %d argsize: %d top: " INTPTR_FORMAT " bottom: " INTPTR_FORMAT,
529 cont_size(), _cont.argsize(), p2i(_cont_stack_top), p2i(_cont_stack_bottom));
530 assert(cont_size() > 0, "");
531
532 if (LockingMode != LM_LIGHTWEIGHT) {
533 _monitors_in_lockstack = 0;
534 } else {
535 _monitors_in_lockstack = _thread->lock_stack().monitor_count();
536 }
537 }
538
539 void FreezeBase::init_rest() { // we want to postpone some initialization after chunk handling
540 _freeze_size = 0;
541 _total_align_size = 0;
542 NOT_PRODUCT(_frames = 0;)
543 }
544
545 void FreezeBase::freeze_lockstack(stackChunkOop chunk) {
546 assert(chunk->sp_address() - chunk->start_address() >= _monitors_in_lockstack, "no room for lockstack");
547
548 _thread->lock_stack().move_to_address((oop*)chunk->start_address());
549 chunk->set_lockstack_size(checked_cast<uint8_t>(_monitors_in_lockstack));
550 chunk->set_has_lockstack(true);
551 }
552
553 void FreezeBase::copy_to_chunk(intptr_t* from, intptr_t* to, int size) {
554 stackChunkOop chunk = _cont.tail();
555 chunk->copy_from_stack_to_chunk(from, to, size);
556 CONT_JFR_ONLY(_jfr_info.record_size_copied(size);)
557
558 #ifdef ASSERT
559 if (_last_write != nullptr) {
560 assert(_last_write == to + size, "Missed a spot: _last_write: " INTPTR_FORMAT " to+size: " INTPTR_FORMAT
561 " stack_size: %d _last_write offset: " PTR_FORMAT " to+size: " PTR_FORMAT, p2i(_last_write), p2i(to+size),
562 chunk->stack_size(), _last_write-chunk->start_address(), to+size-chunk->start_address());
563 _last_write = to;
564 }
565 #endif
566 }
567
568 static void assert_frames_in_continuation_are_safe(JavaThread* thread) {
569 #ifdef ASSERT
570 StackWatermark* watermark = StackWatermarkSet::get(thread, StackWatermarkKind::gc);
571 if (watermark == nullptr) {
572 return;
573 }
574 ContinuationEntry* ce = thread->last_continuation();
575 RegisterMap map(thread,
576 RegisterMap::UpdateMap::include,
577 RegisterMap::ProcessFrames::skip,
578 RegisterMap::WalkContinuation::skip);
579 map.set_include_argument_oops(false);
580 for (frame f = thread->last_frame(); Continuation::is_frame_in_continuation(ce, f); f = f.sender(&map)) {
581 watermark->assert_is_frame_safe(f);
582 }
583 #endif // ASSERT
584 }
585
586 #ifdef ASSERT
587 static bool monitors_on_stack(JavaThread* thread) {
588 assert_frames_in_continuation_are_safe(thread);
589 ContinuationEntry* ce = thread->last_continuation();
590 RegisterMap map(thread,
591 RegisterMap::UpdateMap::include,
592 RegisterMap::ProcessFrames::skip,
593 RegisterMap::WalkContinuation::skip);
594 map.set_include_argument_oops(false);
595 for (frame f = thread->last_frame(); Continuation::is_frame_in_continuation(ce, f); f = f.sender(&map)) {
596 if ((f.is_interpreted_frame() && ContinuationHelper::InterpretedFrame::is_owning_locks(f)) ||
597 (f.is_compiled_frame() && ContinuationHelper::CompiledFrame::is_owning_locks(map.thread(), &map, f)) ||
598 (f.is_native_frame() && ContinuationHelper::NativeFrame::is_owning_locks(map.thread(), f))) {
599 return true;
600 }
601 }
602 return false;
603 }
604 #endif // ASSERT
605
606 // Called _after_ the last possible safepoint during the freeze operation (chunk allocation)
607 void FreezeBase::unwind_frames() {
608 ContinuationEntry* entry = _cont.entry();
609 entry->flush_stack_processing(_thread);
610 assert_frames_in_continuation_are_safe(_thread);
611 assert(LockingMode != LM_LEGACY || !monitors_on_stack(_thread), "unexpected monitors on stack");
612 set_anchor_to_entry(_thread, entry);
613 }
614
615 template <typename ConfigT>
616 freeze_result Freeze<ConfigT>::try_freeze_fast() {
617 assert(_thread->thread_state() == _thread_in_vm, "");
618 assert(_thread->cont_fastpath(), "");
619
620 DEBUG_ONLY(_fast_freeze_size = size_if_fast_freeze_available();)
621 assert(_fast_freeze_size == 0, "");
622
623 stackChunkOop chunk = allocate_chunk(cont_size() + frame::metadata_words + _monitors_in_lockstack, _cont.argsize() + frame::metadata_words_at_top);
624 if (freeze_fast_new_chunk(chunk)) {
625 return freeze_ok;
626 }
627 if (_thread->has_pending_exception()) {
628 return freeze_exception;
629 }
630
631 // TODO R REMOVE when deopt change is fixed
632 assert(!_thread->cont_fastpath() || _barriers, "");
633 log_develop_trace(continuations)("-- RETRYING SLOW --");
634 return freeze_slow();
635 }
636
637 // Returns size needed if the continuation fits, otherwise 0.
638 int FreezeBase::size_if_fast_freeze_available() {
639 stackChunkOop chunk = _cont.tail();
640 if (chunk == nullptr || chunk->is_gc_mode() || chunk->requires_barriers() || chunk->has_mixed_frames()) {
641 log_develop_trace(continuations)("chunk available %s", chunk == nullptr ? "no chunk" : "chunk requires barriers");
642 return 0;
643 }
644
645 int total_size_needed = cont_size();
646 const int chunk_sp = chunk->sp();
647
648 // argsize can be nonzero if we have a caller, but the caller could be in a non-empty parent chunk,
649 // so we subtract it only if we overlap with the caller, i.e. the current chunk isn't empty.
650 // Consider leaving the chunk's argsize set when emptying it and removing the following branch,
651 // although that would require changing stackChunkOopDesc::is_empty
652 if (!chunk->is_empty()) {
653 total_size_needed -= _cont.argsize() + frame::metadata_words_at_top;
654 }
655
656 total_size_needed += _monitors_in_lockstack;
657
658 int chunk_free_room = chunk_sp - frame::metadata_words_at_bottom;
659 bool available = chunk_free_room >= total_size_needed;
660 log_develop_trace(continuations)("chunk available: %s size: %d argsize: %d top: " INTPTR_FORMAT " bottom: " INTPTR_FORMAT,
661 available ? "yes" : "no" , total_size_needed, _cont.argsize(), p2i(_cont_stack_top), p2i(_cont_stack_bottom));
662 return available ? total_size_needed : 0;
663 }
664
665 void FreezeBase::freeze_fast_existing_chunk() {
666 stackChunkOop chunk = _cont.tail();
667
668 DEBUG_ONLY(_fast_freeze_size = size_if_fast_freeze_available();)
669 assert(_fast_freeze_size > 0, "");
670
671 if (!chunk->is_empty()) { // we are copying into a non-empty chunk
672 DEBUG_ONLY(_empty = false;)
673 DEBUG_ONLY(_orig_chunk_sp = chunk->sp_address();)
674 #ifdef ASSERT
675 {
676 intptr_t* retaddr_slot = (chunk->sp_address()
677 - frame::sender_sp_ret_address_offset());
678 assert(ContinuationHelper::return_address_at(retaddr_slot) == chunk->pc(),
679 "unexpected saved return address");
680 }
681 #endif
682
683 // the chunk's sp before the freeze, adjusted to point beyond the stack-passed arguments in the topmost frame
684 // we overlap; we'll overwrite the chunk's top frame's callee arguments
685 const int chunk_start_sp = chunk->sp() + _cont.argsize() + frame::metadata_words_at_top;
686 assert(chunk_start_sp <= chunk->stack_size(), "sp not pointing into stack");
687
688 // increase max_size by what we're freezing minus the overlap
689 chunk->set_max_thawing_size(chunk->max_thawing_size() + cont_size() - _cont.argsize() - frame::metadata_words_at_top);
690
691 intptr_t* const bottom_sp = _cont_stack_bottom - _cont.argsize() - frame::metadata_words_at_top;
692 assert(bottom_sp == _bottom_address, "");
693 // Because the chunk isn't empty, we know there's a caller in the chunk, therefore the bottom-most frame
694 // should have a return barrier (installed back when we thawed it).
695 #ifdef ASSERT
696 {
697 intptr_t* retaddr_slot = (bottom_sp
698 - frame::sender_sp_ret_address_offset());
699 assert(ContinuationHelper::return_address_at(retaddr_slot)
700 == StubRoutines::cont_returnBarrier(),
701 "should be the continuation return barrier");
702 }
703 #endif
704 // We copy the fp from the chunk back to the stack because it contains some caller data,
705 // including, possibly, an oop that might have gone stale since we thawed.
706 patch_stack_pd(bottom_sp, chunk->sp_address());
707 // we don't patch the return pc at this time, so as not to make the stack unwalkable for async walks
708
709 freeze_fast_copy(chunk, chunk_start_sp CONT_JFR_ONLY(COMMA false));
710 } else { // the chunk is empty
711 const int chunk_start_sp = chunk->stack_size();
712
713 DEBUG_ONLY(_empty = true;)
714 DEBUG_ONLY(_orig_chunk_sp = chunk->start_address() + chunk_start_sp;)
715
716 chunk->set_max_thawing_size(cont_size());
717 chunk->set_bottom(chunk_start_sp - _cont.argsize() - frame::metadata_words_at_top);
718 chunk->set_sp(chunk->bottom());
719
720 freeze_fast_copy(chunk, chunk_start_sp CONT_JFR_ONLY(COMMA false));
721 }
722 }
723
724 bool FreezeBase::freeze_fast_new_chunk(stackChunkOop chunk) {
725 DEBUG_ONLY(_empty = true;)
726
727 // Install new chunk
728 _cont.set_tail(chunk);
729
730 if (UNLIKELY(chunk == nullptr || !_thread->cont_fastpath() || _barriers)) { // OOME/probably humongous
731 log_develop_trace(continuations)("Retrying slow. Barriers: %d", _barriers);
732 return false;
733 }
734
735 chunk->set_max_thawing_size(cont_size());
736
737 // in a fresh chunk, we freeze *with* the bottom-most frame's stack arguments.
738 // They'll then be stored twice: in the chunk and in the parent chunk's top frame
739 const int chunk_start_sp = cont_size() + frame::metadata_words + _monitors_in_lockstack;
740 assert(chunk_start_sp == chunk->stack_size(), "");
741
742 DEBUG_ONLY(_orig_chunk_sp = chunk->start_address() + chunk_start_sp;)
743
744 freeze_fast_copy(chunk, chunk_start_sp CONT_JFR_ONLY(COMMA true));
745
746 return true;
747 }
748
749 void FreezeBase::freeze_fast_copy(stackChunkOop chunk, int chunk_start_sp CONT_JFR_ONLY(COMMA bool chunk_is_allocated)) {
750 assert(chunk != nullptr, "");
751 assert(!chunk->has_mixed_frames(), "");
752 assert(!chunk->is_gc_mode(), "");
753 assert(!chunk->has_bitmap(), "");
754 assert(!chunk->requires_barriers(), "");
755 assert(chunk == _cont.tail(), "");
756
757 // We unwind frames after the last safepoint so that the GC will have found the oops in the frames, but before
758 // writing into the chunk. This is so that an asynchronous stack walk (not at a safepoint) that suspends us here
759 // will either see no continuation on the stack, or a consistent chunk.
760 unwind_frames();
761
762 log_develop_trace(continuations)("freeze_fast start: chunk " INTPTR_FORMAT " size: %d orig sp: %d argsize: %d",
763 p2i((oopDesc*)chunk), chunk->stack_size(), chunk_start_sp, _cont.argsize());
764 assert(chunk_start_sp <= chunk->stack_size(), "");
765 assert(chunk_start_sp >= cont_size(), "no room in the chunk");
766
767 const int chunk_new_sp = chunk_start_sp - cont_size(); // the chunk's new sp, after freeze
768 assert(!(_fast_freeze_size > 0) || (_orig_chunk_sp - (chunk->start_address() + chunk_new_sp)) == (_fast_freeze_size - _monitors_in_lockstack), "");
769
770 intptr_t* chunk_top = chunk->start_address() + chunk_new_sp;
771 #ifdef ASSERT
772 if (!_empty) {
773 intptr_t* retaddr_slot = (_orig_chunk_sp
774 - frame::sender_sp_ret_address_offset());
775 assert(ContinuationHelper::return_address_at(retaddr_slot) == chunk->pc(),
776 "unexpected saved return address");
777 }
778 #endif
779
780 log_develop_trace(continuations)("freeze_fast start: " INTPTR_FORMAT " sp: %d chunk_top: " INTPTR_FORMAT,
781 p2i(chunk->start_address()), chunk_new_sp, p2i(chunk_top));
782 intptr_t* from = _cont_stack_top - frame::metadata_words_at_bottom;
783 intptr_t* to = chunk_top - frame::metadata_words_at_bottom;
784 copy_to_chunk(from, to, cont_size() + frame::metadata_words_at_bottom);
785 // Because we're not patched yet, the chunk is now in a bad state
786
787 // patch return pc of the bottom-most frozen frame (now in the chunk)
788 // with the actual caller's return address
789 intptr_t* chunk_bottom_retaddr_slot = (chunk_top + cont_size()
790 - _cont.argsize()
791 - frame::metadata_words_at_top
792 - frame::sender_sp_ret_address_offset());
793 #ifdef ASSERT
794 if (!_empty) {
795 assert(ContinuationHelper::return_address_at(chunk_bottom_retaddr_slot)
796 == StubRoutines::cont_returnBarrier(),
797 "should be the continuation return barrier");
798 }
799 #endif
800 ContinuationHelper::patch_return_address_at(chunk_bottom_retaddr_slot,
801 chunk->pc());
802
803 // We're always writing to a young chunk, so the GC can't see it until the next safepoint.
804 chunk->set_sp(chunk_new_sp);
805
806 // set chunk->pc to the return address of the topmost frame in the chunk
807 if (_preempt) {
808 // On aarch64/riscv64, the return pc of the top frame won't necessarily be at sp[-1].
809 // Also, on x64, if the top frame is the native wrapper frame, sp[-1] will not
810 // be the pc we used when creating the oopmap. Get the top's frame last pc from
811 // the anchor instead.
812 address last_pc = _last_frame.pc();
813 ContinuationHelper::patch_return_address_at(chunk_top - frame::sender_sp_ret_address_offset(), last_pc);
814 chunk->set_pc(last_pc);
815 } else {
816 chunk->set_pc(ContinuationHelper::return_address_at(
817 _cont_stack_top - frame::sender_sp_ret_address_offset()));
818 }
819
820 if (_monitors_in_lockstack > 0) {
821 freeze_lockstack(chunk);
822 }
823
824 _cont.write();
825
826 log_develop_trace(continuations)("FREEZE CHUNK #" INTPTR_FORMAT " (young)", _cont.hash());
827 LogTarget(Trace, continuations) lt;
828 if (lt.develop_is_enabled()) {
829 LogStream ls(lt);
830 chunk->print_on(true, &ls);
831 }
832
833 // Verification
834 assert(_cont.chunk_invariant(), "");
835 chunk->verify();
836
837 #if CONT_JFR
838 EventContinuationFreezeFast e;
839 if (e.should_commit()) {
840 e.set_id(cast_from_oop<u8>(chunk));
841 DEBUG_ONLY(e.set_allocate(chunk_is_allocated);)
842 e.set_size(cont_size() << LogBytesPerWord);
843 e.commit();
844 }
845 #endif
846 }
847
848 NOINLINE freeze_result FreezeBase::freeze_slow() {
849 #ifdef ASSERT
850 ResourceMark rm;
851 #endif
852
853 log_develop_trace(continuations)("freeze_slow #" INTPTR_FORMAT, _cont.hash());
854 assert(_thread->thread_state() == _thread_in_vm || _thread->thread_state() == _thread_blocked, "");
855
856 #if CONT_JFR
857 EventContinuationFreezeSlow e;
858 if (e.should_commit()) {
859 e.set_id(cast_from_oop<u8>(_cont.continuation()));
860 e.commit();
861 }
862 #endif
863
864 init_rest();
865
866 HandleMark hm(Thread::current());
867
868 frame f = freeze_start_frame();
869
870 LogTarget(Debug, continuations) lt;
871 if (lt.develop_is_enabled()) {
872 LogStream ls(lt);
873 f.print_on(&ls);
874 }
875
876 frame caller; // the frozen caller in the chunk
877 freeze_result res = recurse_freeze(f, caller, 0, false, true);
878
879 if (res == freeze_ok) {
880 finish_freeze(f, caller);
881 _cont.write();
882 }
883
884 return res;
885 }
886
887 frame FreezeBase::freeze_start_frame() {
888 if (LIKELY(!_preempt)) {
889 return freeze_start_frame_yield_stub();
890 } else {
891 return freeze_start_frame_on_preempt();
892 }
893 }
894
895 frame FreezeBase::freeze_start_frame_yield_stub() {
896 frame f = _thread->last_frame();
897 assert(SharedRuntime::cont_doYield_stub()->contains(f.pc()), "must be");
898 f = sender<ContinuationHelper::NonInterpretedUnknownFrame>(f);
899 assert(Continuation::is_frame_in_continuation(_thread->last_continuation(), f), "");
900 return f;
901 }
902
903 frame FreezeBase::freeze_start_frame_on_preempt() {
904 assert(_last_frame.sp() == _thread->last_frame().sp(), "_last_frame should be already initialized");
905 assert(Continuation::is_frame_in_continuation(_thread->last_continuation(), _last_frame), "");
906 return _last_frame;
907 }
908
909 // The parameter callee_argsize includes metadata that has to be part of caller/callee overlap.
910 NOINLINE freeze_result FreezeBase::recurse_freeze(frame& f, frame& caller, int callee_argsize, bool callee_interpreted, bool top) {
911 assert(f.unextended_sp() < _bottom_address, ""); // see recurse_freeze_java_frame
912 assert(f.is_interpreted_frame() || ((top && _preempt) == ContinuationHelper::Frame::is_stub(f.cb()))
913 || ((top && _preempt) == f.is_native_frame()), "");
914
915 if (stack_overflow()) {
916 return freeze_exception;
917 }
918
919 if (f.is_compiled_frame()) {
920 if (UNLIKELY(f.oop_map() == nullptr)) {
921 // special native frame
922 return freeze_pinned_native;
923 }
924 return recurse_freeze_compiled_frame(f, caller, callee_argsize, callee_interpreted);
925 } else if (f.is_interpreted_frame()) {
926 assert(!f.interpreter_frame_method()->is_native() || (top && _preempt), "");
927 return recurse_freeze_interpreted_frame(f, caller, callee_argsize, callee_interpreted);
928 } else if (top && _preempt) {
929 assert(f.is_native_frame() || f.is_runtime_frame(), "");
930 return f.is_native_frame() ? recurse_freeze_native_frame(f, caller) : recurse_freeze_stub_frame(f, caller);
931 } else {
932 // Frame can't be frozen. Most likely the call_stub or upcall_stub
933 // which indicates there are further natives frames up the stack.
934 return freeze_pinned_native;
935 }
936 }
937
938 // The parameter callee_argsize includes metadata that has to be part of caller/callee overlap.
939 // See also StackChunkFrameStream<frame_kind>::frame_size()
940 template<typename FKind>
941 inline freeze_result FreezeBase::recurse_freeze_java_frame(const frame& f, frame& caller, int fsize, int argsize) {
942 assert(FKind::is_instance(f), "");
943
944 assert(fsize > 0, "");
945 assert(argsize >= 0, "");
946 _freeze_size += fsize;
947 NOT_PRODUCT(_frames++;)
948
949 assert(FKind::frame_bottom(f) <= _bottom_address, "");
950
951 // We don't use FKind::frame_bottom(f) == _bottom_address because on x64 there's sometimes an extra word between
952 // enterSpecial and an interpreted frame
953 if (FKind::frame_bottom(f) >= _bottom_address - 1) {
954 return finalize_freeze(f, caller, argsize); // recursion end
955 } else {
956 frame senderf = sender<FKind>(f);
957 assert(FKind::interpreted || senderf.sp() == senderf.unextended_sp(), "");
958 freeze_result result = recurse_freeze(senderf, caller, argsize, FKind::interpreted, false); // recursive call
959 return result;
960 }
961 }
962
963 inline void FreezeBase::before_freeze_java_frame(const frame& f, const frame& caller, int fsize, int argsize, bool is_bottom_frame) {
964 LogTarget(Trace, continuations) lt;
965 if (lt.develop_is_enabled()) {
966 LogStream ls(lt);
967 ls.print_cr("======== FREEZING FRAME interpreted: %d bottom: %d", f.is_interpreted_frame(), is_bottom_frame);
968 ls.print_cr("fsize: %d argsize: %d", fsize, argsize);
969 f.print_value_on(&ls);
970 }
971 assert(caller.is_interpreted_frame() == Interpreter::contains(caller.pc()), "");
972 }
973
974 inline void FreezeBase::after_freeze_java_frame(const frame& hf, bool is_bottom_frame) {
975 LogTarget(Trace, continuations) lt;
976 if (lt.develop_is_enabled()) {
977 LogStream ls(lt);
978 DEBUG_ONLY(hf.print_value_on(&ls);)
979 assert(hf.is_heap_frame(), "should be");
980 DEBUG_ONLY(print_frame_layout(hf, false, &ls);)
981 if (is_bottom_frame) {
982 ls.print_cr("bottom h-frame:");
983 hf.print_on(&ls);
984 }
985 }
986 }
987
988 // The parameter argsize_md includes metadata that has to be part of caller/callee overlap.
989 // See also StackChunkFrameStream<frame_kind>::frame_size()
990 freeze_result FreezeBase::finalize_freeze(const frame& callee, frame& caller, int argsize_md) {
991 int argsize = argsize_md - frame::metadata_words_at_top;
992 assert(callee.is_interpreted_frame()
993 || ContinuationHelper::Frame::is_stub(callee.cb())
994 || callee.cb()->as_nmethod()->is_osr_method()
995 || argsize == _cont.argsize(), "argsize: %d cont.argsize: %d", argsize, _cont.argsize());
996 log_develop_trace(continuations)("bottom: " INTPTR_FORMAT " count %d size: %d argsize: %d",
997 p2i(_bottom_address), _frames, _freeze_size << LogBytesPerWord, argsize);
998
999 LogTarget(Trace, continuations) lt;
1000
1001 #ifdef ASSERT
1002 bool empty = _cont.is_empty();
1003 log_develop_trace(continuations)("empty: %d", empty);
1004 #endif
1005
1006 stackChunkOop chunk = _cont.tail();
1007
1008 assert(chunk == nullptr || (chunk->max_thawing_size() == 0) == chunk->is_empty(), "");
1009
1010 _freeze_size += frame::metadata_words; // for top frame's metadata
1011
1012 int overlap = 0; // the args overlap the caller -- if there is one in this chunk and is of the same kind
1013 int unextended_sp = -1;
1014 if (chunk != nullptr) {
1015 if (!chunk->is_empty()) {
1016 StackChunkFrameStream<ChunkFrames::Mixed> last(chunk);
1017 unextended_sp = chunk->to_offset(StackChunkFrameStream<ChunkFrames::Mixed>(chunk).unextended_sp());
1018 bool top_interpreted = Interpreter::contains(chunk->pc());
1019 if (callee.is_interpreted_frame() == top_interpreted) {
1020 overlap = argsize_md;
1021 }
1022 } else {
1023 unextended_sp = chunk->stack_size() - frame::metadata_words_at_top;
1024 }
1025 }
1026
1027 log_develop_trace(continuations)("finalize _size: %d overlap: %d unextended_sp: %d", _freeze_size, overlap, unextended_sp);
1028
1029 _freeze_size -= overlap;
1030 assert(_freeze_size >= 0, "");
1031
1032 assert(chunk == nullptr || chunk->is_empty()
1033 || unextended_sp == chunk->to_offset(StackChunkFrameStream<ChunkFrames::Mixed>(chunk).unextended_sp()), "");
1034 assert(chunk != nullptr || unextended_sp < _freeze_size, "");
1035
1036 _freeze_size += _monitors_in_lockstack;
1037
1038 // _barriers can be set to true by an allocation in freeze_fast, in which case the chunk is available
1039 bool allocated_old_in_freeze_fast = _barriers;
1040 assert(!allocated_old_in_freeze_fast || (unextended_sp >= _freeze_size && chunk->is_empty()),
1041 "Chunk allocated in freeze_fast is of insufficient size "
1042 "unextended_sp: %d size: %d is_empty: %d", unextended_sp, _freeze_size, chunk->is_empty());
1043 assert(!allocated_old_in_freeze_fast || (!UseZGC && !UseG1GC), "Unexpected allocation");
1044
1045 DEBUG_ONLY(bool empty_chunk = true);
1046 if (unextended_sp < _freeze_size || chunk->is_gc_mode() || (!allocated_old_in_freeze_fast && chunk->requires_barriers())) {
1047 // ALLOCATE NEW CHUNK
1048
1049 if (lt.develop_is_enabled()) {
1050 LogStream ls(lt);
1051 if (chunk == nullptr) {
1052 ls.print_cr("no chunk");
1053 } else {
1054 ls.print_cr("chunk barriers: %d _size: %d free size: %d",
1055 chunk->requires_barriers(), _freeze_size, chunk->sp() - frame::metadata_words);
1056 chunk->print_on(&ls);
1057 }
1058 }
1059
1060 _freeze_size += overlap; // we're allocating a new chunk, so no overlap
1061 // overlap = 0;
1062
1063 chunk = allocate_chunk_slow(_freeze_size, argsize_md);
1064 if (chunk == nullptr) {
1065 return freeze_exception;
1066 }
1067
1068 // Install new chunk
1069 _cont.set_tail(chunk);
1070 assert(chunk->is_empty(), "");
1071 } else {
1072 // REUSE EXISTING CHUNK
1073 log_develop_trace(continuations)("Reusing chunk mixed: %d empty: %d", chunk->has_mixed_frames(), chunk->is_empty());
1074 if (chunk->is_empty()) {
1075 int sp = chunk->stack_size() - argsize_md;
1076 chunk->set_sp(sp);
1077 chunk->set_bottom(sp);
1078 _freeze_size += overlap;
1079 assert(chunk->max_thawing_size() == 0, "");
1080 } DEBUG_ONLY(else empty_chunk = false;)
1081 }
1082 assert(!chunk->is_gc_mode(), "");
1083 assert(!chunk->has_bitmap(), "");
1084 chunk->set_has_mixed_frames(true);
1085
1086 assert(chunk->requires_barriers() == _barriers, "");
1087 assert(!_barriers || chunk->is_empty(), "");
1088
1089 assert(!chunk->is_empty() || StackChunkFrameStream<ChunkFrames::Mixed>(chunk).is_done(), "");
1090 assert(!chunk->is_empty() || StackChunkFrameStream<ChunkFrames::Mixed>(chunk).to_frame().is_empty(), "");
1091
1092 if (_preempt) {
1093 frame f = _thread->last_frame();
1094 if (f.is_interpreted_frame()) {
1095 // Some platforms do not save the last_sp in the top interpreter frame on VM calls.
1096 // We need it so that on resume we can restore the sp to the right place, since
1097 // thawing might add an alignment word to the expression stack (see finish_thaw()).
1098 // We do it now that we know freezing will be successful.
1099 prepare_freeze_interpreted_top_frame(f);
1100 }
1101 }
1102
1103 // We unwind frames after the last safepoint so that the GC will have found the oops in the frames, but before
1104 // writing into the chunk. This is so that an asynchronous stack walk (not at a safepoint) that suspends us here
1105 // will either see no continuation or a consistent chunk.
1106 unwind_frames();
1107
1108 chunk->set_max_thawing_size(chunk->max_thawing_size() + _freeze_size - _monitors_in_lockstack - frame::metadata_words);
1109
1110 if (lt.develop_is_enabled()) {
1111 LogStream ls(lt);
1112 ls.print_cr("top chunk:");
1113 chunk->print_on(&ls);
1114 }
1115
1116 if (_monitors_in_lockstack > 0) {
1117 freeze_lockstack(chunk);
1118 }
1119
1120 // The topmost existing frame in the chunk; or an empty frame if the chunk is empty
1121 caller = StackChunkFrameStream<ChunkFrames::Mixed>(chunk).to_frame();
1122
1123 DEBUG_ONLY(_last_write = caller.unextended_sp() + (empty_chunk ? argsize_md : overlap);)
1124
1125 assert(chunk->is_in_chunk(_last_write - _freeze_size),
1126 "last_write-size: " INTPTR_FORMAT " start: " INTPTR_FORMAT, p2i(_last_write-_freeze_size), p2i(chunk->start_address()));
1127 #ifdef ASSERT
1128 if (lt.develop_is_enabled()) {
1129 LogStream ls(lt);
1130 ls.print_cr("top hframe before (freeze):");
1131 assert(caller.is_heap_frame(), "should be");
1132 caller.print_on(&ls);
1133 }
1134
1135 assert(!empty || Continuation::is_continuation_entry_frame(callee, nullptr), "");
1136
1137 frame entry = sender(callee);
1138
1139 assert((!empty && Continuation::is_return_barrier_entry(entry.pc())) || (empty && Continuation::is_continuation_enterSpecial(entry)), "");
1140 assert(callee.is_interpreted_frame() || entry.sp() == entry.unextended_sp(), "");
1141 #endif
1142
1143 return freeze_ok_bottom;
1144 }
1145
1146 // After freezing a frame we need to possibly adjust some values related to the caller frame.
1147 void FreezeBase::patch(const frame& f, frame& hf, const frame& caller, bool is_bottom_frame) {
1148 if (is_bottom_frame) {
1149 // If we're the bottom frame, we need to replace the return barrier with the real
1150 // caller's pc.
1151 address last_pc = caller.pc();
1152 assert((last_pc == nullptr) == _cont.tail()->is_empty(), "");
1153 ContinuationHelper::Frame::patch_pc(caller, last_pc);
1154 } else {
1155 assert(!caller.is_empty(), "");
1156 }
1157
1158 patch_pd(hf, caller);
1159
1160 if (f.is_interpreted_frame()) {
1161 assert(hf.is_heap_frame(), "should be");
1162 ContinuationHelper::InterpretedFrame::patch_sender_sp(hf, caller);
1163 }
1164
1165 #ifdef ASSERT
1166 if (hf.is_compiled_frame()) {
1167 if (f.is_deoptimized_frame()) { // TODO DEOPT: long term solution: unroll on freeze and patch pc
1168 log_develop_trace(continuations)("Freezing deoptimized frame");
1169 assert(f.cb()->as_nmethod()->is_deopt_pc(f.raw_pc()), "");
1170 assert(f.cb()->as_nmethod()->is_deopt_pc(ContinuationHelper::Frame::real_pc(f)), "");
1171 }
1172 }
1173 #endif
1174 }
1175
1176 #ifdef ASSERT
1177 static void verify_frame_top(const frame& f, intptr_t* top) {
1178 ResourceMark rm;
1179 InterpreterOopMap mask;
1180 f.interpreted_frame_oop_map(&mask);
1181 assert(top <= ContinuationHelper::InterpretedFrame::frame_top(f, &mask),
1182 "frame_top: " INTPTR_FORMAT " Interpreted::frame_top: " INTPTR_FORMAT,
1183 p2i(top), p2i(ContinuationHelper::InterpretedFrame::frame_top(f, &mask)));
1184 }
1185 #endif // ASSERT
1186
1187 // The parameter callee_argsize includes metadata that has to be part of caller/callee overlap.
1188 // See also StackChunkFrameStream<frame_kind>::frame_size()
1189 NOINLINE freeze_result FreezeBase::recurse_freeze_interpreted_frame(frame& f, frame& caller,
1190 int callee_argsize /* incl. metadata */,
1191 bool callee_interpreted) {
1192 adjust_interpreted_frame_unextended_sp(f);
1193
1194 // The frame's top never includes the stack arguments to the callee
1195 intptr_t* const stack_frame_top = ContinuationHelper::InterpretedFrame::frame_top(f, callee_argsize, callee_interpreted);
1196 intptr_t* const stack_frame_bottom = ContinuationHelper::InterpretedFrame::frame_bottom(f);
1197 const int fsize = pointer_delta_as_int(stack_frame_bottom, stack_frame_top);
1198
1199 DEBUG_ONLY(verify_frame_top(f, stack_frame_top));
1200
1201 Method* frame_method = ContinuationHelper::Frame::frame_method(f);
1202 // including metadata between f and its args
1203 const int argsize = ContinuationHelper::InterpretedFrame::stack_argsize(f) + frame::metadata_words_at_top;
1204
1205 log_develop_trace(continuations)("recurse_freeze_interpreted_frame %s _size: %d fsize: %d argsize: %d callee_interpreted: %d",
1206 frame_method->name_and_sig_as_C_string(), _freeze_size, fsize, argsize, callee_interpreted);
1207 // we'd rather not yield inside methods annotated with @JvmtiMountTransition
1208 assert(!ContinuationHelper::Frame::frame_method(f)->jvmti_mount_transition(), "");
1209
1210 freeze_result result = recurse_freeze_java_frame<ContinuationHelper::InterpretedFrame>(f, caller, fsize, argsize);
1211 if (UNLIKELY(result > freeze_ok_bottom)) {
1212 return result;
1213 }
1214
1215 bool is_bottom_frame = result == freeze_ok_bottom;
1216 assert(!caller.is_empty() || is_bottom_frame, "");
1217
1218 DEBUG_ONLY(before_freeze_java_frame(f, caller, fsize, 0, is_bottom_frame);)
1219
1220 frame hf = new_heap_frame<ContinuationHelper::InterpretedFrame>(f, caller);
1221 _total_align_size += frame::align_wiggle; // add alignment room for internal interpreted frame alignment on AArch64/PPC64
1222
1223 intptr_t* heap_frame_top = ContinuationHelper::InterpretedFrame::frame_top(hf, callee_argsize, callee_interpreted);
1224 intptr_t* heap_frame_bottom = ContinuationHelper::InterpretedFrame::frame_bottom(hf);
1225 assert(heap_frame_bottom == heap_frame_top + fsize, "");
1226
1227 // Some architectures (like AArch64/PPC64/RISC-V) add padding between the locals and the fixed_frame to keep the fp 16-byte-aligned.
1228 // On those architectures we freeze the padding in order to keep the same fp-relative offsets in the fixed_frame.
1229 copy_to_chunk(stack_frame_top, heap_frame_top, fsize);
1230 assert(!is_bottom_frame || !caller.is_interpreted_frame() || (heap_frame_top + fsize) == (caller.unextended_sp() + argsize), "");
1231
1232 relativize_interpreted_frame_metadata(f, hf);
1233
1234 patch(f, hf, caller, is_bottom_frame);
1235
1236 CONT_JFR_ONLY(_jfr_info.record_interpreted_frame();)
1237 DEBUG_ONLY(after_freeze_java_frame(hf, is_bottom_frame);)
1238 caller = hf;
1239
1240 // Mark frame_method's GC epoch for class redefinition on_stack calculation.
1241 frame_method->record_gc_epoch();
1242
1243 return freeze_ok;
1244 }
1245
1246 // The parameter callee_argsize includes metadata that has to be part of caller/callee overlap.
1247 // See also StackChunkFrameStream<frame_kind>::frame_size()
1248 freeze_result FreezeBase::recurse_freeze_compiled_frame(frame& f, frame& caller,
1249 int callee_argsize /* incl. metadata */,
1250 bool callee_interpreted) {
1251 // The frame's top never includes the stack arguments to the callee
1252 intptr_t* const stack_frame_top = ContinuationHelper::CompiledFrame::frame_top(f, callee_argsize, callee_interpreted);
1253 intptr_t* const stack_frame_bottom = ContinuationHelper::CompiledFrame::frame_bottom(f);
1254 // including metadata between f and its stackargs
1255 const int argsize = ContinuationHelper::CompiledFrame::stack_argsize(f) + frame::metadata_words_at_top;
1256 const int fsize = pointer_delta_as_int(stack_frame_bottom + argsize, stack_frame_top);
1257
1258 log_develop_trace(continuations)("recurse_freeze_compiled_frame %s _size: %d fsize: %d argsize: %d",
1259 ContinuationHelper::Frame::frame_method(f) != nullptr ?
1260 ContinuationHelper::Frame::frame_method(f)->name_and_sig_as_C_string() : "",
1261 _freeze_size, fsize, argsize);
1262 // we'd rather not yield inside methods annotated with @JvmtiMountTransition
1263 assert(!ContinuationHelper::Frame::frame_method(f)->jvmti_mount_transition(), "");
1264
1265 freeze_result result = recurse_freeze_java_frame<ContinuationHelper::CompiledFrame>(f, caller, fsize, argsize);
1266 if (UNLIKELY(result > freeze_ok_bottom)) {
1267 return result;
1268 }
1269
1270 bool is_bottom_frame = result == freeze_ok_bottom;
1271 assert(!caller.is_empty() || is_bottom_frame, "");
1272
1273 DEBUG_ONLY(before_freeze_java_frame(f, caller, fsize, argsize, is_bottom_frame);)
1274
1275 frame hf = new_heap_frame<ContinuationHelper::CompiledFrame>(f, caller);
1276
1277 intptr_t* heap_frame_top = ContinuationHelper::CompiledFrame::frame_top(hf, callee_argsize, callee_interpreted);
1278
1279 copy_to_chunk(stack_frame_top, heap_frame_top, fsize);
1280 assert(!is_bottom_frame || !caller.is_compiled_frame() || (heap_frame_top + fsize) == (caller.unextended_sp() + argsize), "");
1281
1282 if (caller.is_interpreted_frame()) {
1283 // When thawing the frame we might need to add alignment (see Thaw::align)
1284 _total_align_size += frame::align_wiggle;
1285 }
1286
1287 patch(f, hf, caller, is_bottom_frame);
1288
1289 assert(is_bottom_frame || Interpreter::contains(ContinuationHelper::CompiledFrame::real_pc(caller)) == caller.is_interpreted_frame(), "");
1290
1291 DEBUG_ONLY(after_freeze_java_frame(hf, is_bottom_frame);)
1292 caller = hf;
1293 return freeze_ok;
1294 }
1295
1296 NOINLINE freeze_result FreezeBase::recurse_freeze_stub_frame(frame& f, frame& caller) {
1297 DEBUG_ONLY(frame fsender = sender(f);)
1298 assert(fsender.is_compiled_frame(), "sender should be compiled frame");
1299
1300 intptr_t* const stack_frame_top = ContinuationHelper::StubFrame::frame_top(f);
1301 const int fsize = f.cb()->frame_size();
1302
1303 log_develop_trace(continuations)("recurse_freeze_stub_frame %s _size: %d fsize: %d :: " INTPTR_FORMAT " - " INTPTR_FORMAT,
1304 f.cb()->name(), _freeze_size, fsize, p2i(stack_frame_top), p2i(stack_frame_top+fsize));
1305
1306 freeze_result result = recurse_freeze_java_frame<ContinuationHelper::StubFrame>(f, caller, fsize, 0);
1307 if (UNLIKELY(result > freeze_ok_bottom)) {
1308 return result;
1309 }
1310
1311 assert(result == freeze_ok, "should have caller");
1312 DEBUG_ONLY(before_freeze_java_frame(f, caller, fsize, 0, false /*is_bottom_frame*/);)
1313
1314 frame hf = new_heap_frame<ContinuationHelper::StubFrame>(f, caller);
1315 intptr_t* heap_frame_top = ContinuationHelper::StubFrame::frame_top(hf);
1316
1317 copy_to_chunk(stack_frame_top, heap_frame_top, fsize);
1318
1319 patch(f, hf, caller, false /*is_bottom_frame*/);
1320
1321 DEBUG_ONLY(after_freeze_java_frame(hf, false /*is_bottom_frame*/);)
1322
1323 caller = hf;
1324 return freeze_ok;
1325 }
1326
1327 NOINLINE freeze_result FreezeBase::recurse_freeze_native_frame(frame& f, frame& caller) {
1328 if (!f.cb()->as_nmethod()->method()->is_object_wait0()) {
1329 assert(f.cb()->as_nmethod()->method()->is_synchronized(), "");
1330 // Synchronized native method case. Unlike the interpreter native wrapper, the compiled
1331 // native wrapper tries to acquire the monitor after marshalling the arguments from the
1332 // caller into the native convention. This is so that we have a valid oopMap in case of
1333 // having to block in the slow path. But that would require freezing those registers too
1334 // and then fixing them back on thaw in case of oops. To avoid complicating things and
1335 // given that this would be a rare case anyways just pin the vthread to the carrier.
1336 return freeze_pinned_native;
1337 }
1338
1339 intptr_t* const stack_frame_top = ContinuationHelper::NativeFrame::frame_top(f);
1340 // There are no stackargs but argsize must include the metadata
1341 const int argsize = frame::metadata_words_at_top;
1342 const int fsize = f.cb()->frame_size() + argsize;
1343
1344 log_develop_trace(continuations)("recurse_freeze_native_frame %s _size: %d fsize: %d :: " INTPTR_FORMAT " - " INTPTR_FORMAT,
1345 f.cb()->name(), _freeze_size, fsize, p2i(stack_frame_top), p2i(stack_frame_top+fsize));
1346
1347 freeze_result result = recurse_freeze_java_frame<ContinuationHelper::NativeFrame>(f, caller, fsize, argsize);
1348 if (UNLIKELY(result > freeze_ok_bottom)) {
1349 return result;
1350 }
1351
1352 assert(result == freeze_ok, "should have caller frame");
1353 DEBUG_ONLY(before_freeze_java_frame(f, caller, fsize, argsize, false /* is_bottom_frame */);)
1354
1355 frame hf = new_heap_frame<ContinuationHelper::NativeFrame>(f, caller);
1356 intptr_t* heap_frame_top = ContinuationHelper::NativeFrame::frame_top(hf);
1357
1358 copy_to_chunk(stack_frame_top, heap_frame_top, fsize);
1359
1360 if (caller.is_interpreted_frame()) {
1361 // When thawing the frame we might need to add alignment (see Thaw::align)
1362 _total_align_size += frame::align_wiggle;
1363 }
1364
1365 patch(f, hf, caller, false /* is_bottom_frame */);
1366
1367 DEBUG_ONLY(after_freeze_java_frame(hf, false /* is_bottom_frame */);)
1368
1369 caller = hf;
1370 return freeze_ok;
1371 }
1372
1373 NOINLINE void FreezeBase::finish_freeze(const frame& f, const frame& top) {
1374 stackChunkOop chunk = _cont.tail();
1375
1376 LogTarget(Trace, continuations) lt;
1377 if (lt.develop_is_enabled()) {
1378 LogStream ls(lt);
1379 assert(top.is_heap_frame(), "should be");
1380 top.print_on(&ls);
1381 }
1382
1383 set_top_frame_metadata_pd(top);
1384
1385 chunk->set_sp(chunk->to_offset(top.sp()));
1386 chunk->set_pc(top.pc());
1387
1388 chunk->set_max_thawing_size(chunk->max_thawing_size() + _total_align_size);
1389
1390 assert(chunk->sp_address() - chunk->start_address() >= _monitors_in_lockstack, "clash with lockstack");
1391
1392 // At this point the chunk is consistent
1393
1394 if (UNLIKELY(_barriers)) {
1395 log_develop_trace(continuations)("do barriers on old chunk");
1396 // Serial and Parallel GC can allocate objects directly into the old generation.
1397 // Then we want to relativize the derived pointers eagerly so that
1398 // old chunks are all in GC mode.
1399 assert(!UseG1GC, "G1 can not deal with allocating outside of eden");
1400 assert(!UseZGC, "ZGC can not deal with allocating chunks visible to marking");
1401 if (UseShenandoahGC) {
1402 _cont.tail()->relativize_derived_pointers_concurrently();
1403 } else {
1404 ContinuationGCSupport::transform_stack_chunk(_cont.tail());
1405 }
1406 // For objects in the old generation we must maintain the remembered set
1407 _cont.tail()->do_barriers<stackChunkOopDesc::BarrierType::Store>();
1408 }
1409
1410 log_develop_trace(continuations)("finish_freeze: has_mixed_frames: %d", chunk->has_mixed_frames());
1411 if (lt.develop_is_enabled()) {
1412 LogStream ls(lt);
1413 chunk->print_on(true, &ls);
1414 }
1415
1416 if (lt.develop_is_enabled()) {
1417 LogStream ls(lt);
1418 ls.print_cr("top hframe after (freeze):");
1419 assert(_cont.last_frame().is_heap_frame(), "should be");
1420 _cont.last_frame().print_on(&ls);
1421 DEBUG_ONLY(print_frame_layout(top, false, &ls);)
1422 }
1423
1424 assert(_cont.chunk_invariant(), "");
1425 }
1426
1427 inline bool FreezeBase::stack_overflow() { // detect stack overflow in recursive native code
1428 JavaThread* t = !_preempt ? _thread : JavaThread::current();
1429 assert(t == JavaThread::current(), "");
1430 if (os::current_stack_pointer() < t->stack_overflow_state()->shadow_zone_safe_limit()) {
1431 if (!_preempt) {
1432 ContinuationWrapper::SafepointOp so(t, _cont); // could also call _cont.done() instead
1433 Exceptions::_throw_msg(t, __FILE__, __LINE__, vmSymbols::java_lang_StackOverflowError(), "Stack overflow while freezing");
1434 }
1435 return true;
1436 }
1437 return false;
1438 }
1439
1440 class StackChunkAllocator : public MemAllocator {
1441 const size_t _stack_size;
1442 int _argsize_md;
1443 ContinuationWrapper& _continuation_wrapper;
1444 JvmtiSampledObjectAllocEventCollector* const _jvmti_event_collector;
1445 mutable bool _took_slow_path;
1446
1447 // Does the minimal amount of initialization needed for a TLAB allocation.
1448 // We don't need to do a full initialization, as such an allocation need not be immediately walkable.
1449 virtual oop initialize(HeapWord* mem) const override {
1450 assert(_stack_size > 0, "");
1451 assert(_stack_size <= max_jint, "");
1452 assert(_word_size > _stack_size, "");
1453
1454 // zero out fields (but not the stack)
1455 const size_t hs = oopDesc::header_size();
1456 if (oopDesc::has_klass_gap()) {
1457 oopDesc::set_klass_gap(mem, 0);
1458 }
1459 Copy::fill_to_aligned_words(mem + hs, vmClasses::StackChunk_klass()->size_helper() - hs);
1460
1461 int bottom = (int)_stack_size - _argsize_md;
1462
1463 jdk_internal_vm_StackChunk::set_size(mem, (int)_stack_size);
1464 jdk_internal_vm_StackChunk::set_bottom(mem, bottom);
1465 jdk_internal_vm_StackChunk::set_sp(mem, bottom);
1466
1467 return finish(mem);
1468 }
1469
1470 stackChunkOop allocate_fast() const {
1471 if (!UseTLAB) {
1472 return nullptr;
1473 }
1474
1475 HeapWord* const mem = MemAllocator::mem_allocate_inside_tlab_fast();
1476 if (mem == nullptr) {
1477 return nullptr;
1478 }
1479
1480 oop obj = initialize(mem);
1481 return stackChunkOopDesc::cast(obj);
1482 }
1483
1484 public:
1485 StackChunkAllocator(Klass* klass,
1486 size_t word_size,
1487 Thread* thread,
1488 size_t stack_size,
1489 int argsize_md,
1490 ContinuationWrapper& continuation_wrapper,
1491 JvmtiSampledObjectAllocEventCollector* jvmti_event_collector)
1492 : MemAllocator(klass, word_size, thread),
1493 _stack_size(stack_size),
1494 _argsize_md(argsize_md),
1495 _continuation_wrapper(continuation_wrapper),
1496 _jvmti_event_collector(jvmti_event_collector),
1497 _took_slow_path(false) {}
1498
1499 // Provides it's own, specialized allocation which skips instrumentation
1500 // if the memory can be allocated without going to a slow-path.
1501 stackChunkOop allocate() const {
1502 // First try to allocate without any slow-paths or instrumentation.
1503 stackChunkOop obj = allocate_fast();
1504 if (obj != nullptr) {
1505 return obj;
1506 }
1507
1508 // Now try full-blown allocation with all expensive operations,
1509 // including potentially safepoint operations.
1510 _took_slow_path = true;
1511
1512 // Protect unhandled Loom oops
1513 ContinuationWrapper::SafepointOp so(_thread, _continuation_wrapper);
1514
1515 // Can safepoint
1516 _jvmti_event_collector->start();
1517
1518 // Can safepoint
1519 return stackChunkOopDesc::cast(MemAllocator::allocate());
1520 }
1521
1522 bool took_slow_path() const {
1523 return _took_slow_path;
1524 }
1525 };
1526
1527 template <typename ConfigT>
1528 stackChunkOop Freeze<ConfigT>::allocate_chunk(size_t stack_size, int argsize_md) {
1529 log_develop_trace(continuations)("allocate_chunk allocating new chunk");
1530
1531 InstanceStackChunkKlass* klass = InstanceStackChunkKlass::cast(vmClasses::StackChunk_klass());
1532 size_t size_in_words = klass->instance_size(stack_size);
1533
1534 if (CollectedHeap::stack_chunk_max_size() > 0 && size_in_words >= CollectedHeap::stack_chunk_max_size()) {
1535 if (!_preempt) {
1536 throw_stack_overflow_on_humongous_chunk();
1537 }
1538 return nullptr;
1539 }
1540
1541 JavaThread* current = _preempt ? JavaThread::current() : _thread;
1542 assert(current == JavaThread::current(), "should be current");
1543
1544 // Allocate the chunk.
1545 //
1546 // This might safepoint while allocating, but all safepointing due to
1547 // instrumentation have been deferred. This property is important for
1548 // some GCs, as this ensures that the allocated object is in the young
1549 // generation / newly allocated memory.
1550 StackChunkAllocator allocator(klass, size_in_words, current, stack_size, argsize_md, _cont, _jvmti_event_collector);
1551 stackChunkOop chunk = allocator.allocate();
1552
1553 if (chunk == nullptr) {
1554 return nullptr; // OOME
1555 }
1556
1557 // assert that chunk is properly initialized
1558 assert(chunk->stack_size() == (int)stack_size, "");
1559 assert(chunk->size() >= stack_size, "chunk->size(): %zu size: %zu", chunk->size(), stack_size);
1560 assert(chunk->sp() == chunk->bottom(), "");
1561 assert((intptr_t)chunk->start_address() % 8 == 0, "");
1562 assert(chunk->max_thawing_size() == 0, "");
1563 assert(chunk->pc() == nullptr, "");
1564 assert(chunk->is_empty(), "");
1565 assert(chunk->flags() == 0, "");
1566 assert(chunk->is_gc_mode() == false, "");
1567 assert(chunk->lockstack_size() == 0, "");
1568
1569 // fields are uninitialized
1570 chunk->set_parent_access<IS_DEST_UNINITIALIZED>(_cont.last_nonempty_chunk());
1571 chunk->set_cont_access<IS_DEST_UNINITIALIZED>(_cont.continuation());
1572
1573 #if INCLUDE_ZGC
1574 if (UseZGC) {
1575 ZStackChunkGCData::initialize(chunk);
1576 assert(!chunk->requires_barriers(), "ZGC always allocates in the young generation");
1577 _barriers = false;
1578 } else
1579 #endif
1580 #if INCLUDE_SHENANDOAHGC
1581 if (UseShenandoahGC) {
1582 _barriers = chunk->requires_barriers();
1583 } else
1584 #endif
1585 {
1586 if (!allocator.took_slow_path()) {
1587 // Guaranteed to be in young gen / newly allocated memory
1588 assert(!chunk->requires_barriers(), "Unfamiliar GC requires barriers on TLAB allocation");
1589 _barriers = false;
1590 } else {
1591 // Some GCs could put direct allocations in old gen for slow-path
1592 // allocations; need to explicitly check if that was the case.
1593 _barriers = chunk->requires_barriers();
1594 }
1595 }
1596
1597 if (_barriers) {
1598 log_develop_trace(continuations)("allocation requires barriers");
1599 }
1600
1601 assert(chunk->parent() == nullptr || chunk->parent()->is_stackChunk(), "");
1602
1603 return chunk;
1604 }
1605
1606 void FreezeBase::throw_stack_overflow_on_humongous_chunk() {
1607 ContinuationWrapper::SafepointOp so(_thread, _cont); // could also call _cont.done() instead
1608 Exceptions::_throw_msg(_thread, __FILE__, __LINE__, vmSymbols::java_lang_StackOverflowError(), "Humongous stack chunk");
1609 }
1610
1611 #if INCLUDE_JVMTI
1612 static int num_java_frames(ContinuationWrapper& cont) {
1613 ResourceMark rm; // used for scope traversal in num_java_frames(nmethod*, address)
1614 int count = 0;
1615 for (stackChunkOop chunk = cont.tail(); chunk != nullptr; chunk = chunk->parent()) {
1616 count += chunk->num_java_frames();
1617 }
1618 return count;
1619 }
1620
1621 static void invalidate_jvmti_stack(JavaThread* thread) {
1622 if (thread->is_interp_only_mode()) {
1623 JvmtiThreadState *state = thread->jvmti_thread_state();
1624 if (state != nullptr)
1625 state->invalidate_cur_stack_depth();
1626 }
1627 }
1628
1629 static void jvmti_yield_cleanup(JavaThread* thread, ContinuationWrapper& cont) {
1630 if (JvmtiExport::can_post_frame_pop()) {
1631 int num_frames = num_java_frames(cont);
1632
1633 ContinuationWrapper::SafepointOp so(Thread::current(), cont);
1634 JvmtiExport::continuation_yield_cleanup(JavaThread::current(), num_frames);
1635 }
1636 invalidate_jvmti_stack(thread);
1637 }
1638
1639 static void jvmti_mount_end(JavaThread* current, ContinuationWrapper& cont, frame top) {
1640 assert(current->vthread() != nullptr, "must be");
1641
1642 HandleMarkCleaner hm(current);
1643 Handle vth(current, current->vthread());
1644
1645 ContinuationWrapper::SafepointOp so(current, cont);
1646
1647 // Since we might safepoint set the anchor so that the stack can be walked.
1648 set_anchor(current, top.sp());
1649
1650 JRT_BLOCK
1651 JvmtiVTMSTransitionDisabler::VTMS_vthread_mount((jthread)vth.raw_value(), false);
1652
1653 if (current->pending_contended_entered_event()) {
1654 JvmtiExport::post_monitor_contended_entered(current, current->contended_entered_monitor());
1655 current->set_contended_entered_monitor(nullptr);
1656 }
1657 JRT_BLOCK_END
1658
1659 clear_anchor(current);
1660 }
1661 #endif // INCLUDE_JVMTI
1662
1663 #ifdef ASSERT
1664 // There are no interpreted frames if we're not called from the interpreter and we haven't ancountered an i2c
1665 // adapter or called Deoptimization::unpack_frames. As for native frames, upcalls from JNI also go through the
1666 // interpreter (see JavaCalls::call_helper), while the UpcallLinker explicitly sets cont_fastpath.
1667 bool FreezeBase::check_valid_fast_path() {
1668 ContinuationEntry* ce = _thread->last_continuation();
1669 RegisterMap map(_thread,
1670 RegisterMap::UpdateMap::skip,
1671 RegisterMap::ProcessFrames::skip,
1672 RegisterMap::WalkContinuation::skip);
1673 map.set_include_argument_oops(false);
1674 bool is_top_frame = true;
1675 for (frame f = freeze_start_frame(); Continuation::is_frame_in_continuation(ce, f); f = f.sender(&map), is_top_frame = false) {
1676 if (!((f.is_compiled_frame() && !f.is_deoptimized_frame()) || (is_top_frame && (f.is_runtime_frame() || f.is_native_frame())))) {
1677 return false;
1678 }
1679 }
1680 return true;
1681 }
1682 #endif // ASSERT
1683
1684 static inline freeze_result freeze_epilog(ContinuationWrapper& cont) {
1685 verify_continuation(cont.continuation());
1686 assert(!cont.is_empty(), "");
1687
1688 log_develop_debug(continuations)("=== End of freeze cont ### #" INTPTR_FORMAT, cont.hash());
1689 return freeze_ok;
1690 }
1691
1692 static freeze_result freeze_epilog(JavaThread* thread, ContinuationWrapper& cont, freeze_result res) {
1693 if (UNLIKELY(res != freeze_ok)) {
1694 JFR_ONLY(thread->set_last_freeze_fail_result(res);)
1695 verify_continuation(cont.continuation());
1696 log_develop_trace(continuations)("=== end of freeze (fail %d)", res);
1697 return res;
1698 }
1699
1700 JVMTI_ONLY(jvmti_yield_cleanup(thread, cont)); // can safepoint
1701 return freeze_epilog(cont);
1702 }
1703
1704 static freeze_result preempt_epilog(ContinuationWrapper& cont, freeze_result res, frame& old_last_frame) {
1705 if (UNLIKELY(res != freeze_ok)) {
1706 verify_continuation(cont.continuation());
1707 log_develop_trace(continuations)("=== end of freeze (fail %d)", res);
1708 return res;
1709 }
1710
1711 patch_return_pc_with_preempt_stub(old_last_frame);
1712 cont.tail()->set_preempted(true);
1713
1714 return freeze_epilog(cont);
1715 }
1716
1717 template<typename ConfigT, bool preempt>
1718 static inline freeze_result freeze_internal(JavaThread* current, intptr_t* const sp) {
1719 assert(!current->has_pending_exception(), "");
1720
1721 #ifdef ASSERT
1722 log_trace(continuations)("~~~~ freeze sp: " INTPTR_FORMAT "JavaThread: " INTPTR_FORMAT, p2i(current->last_continuation()->entry_sp()), p2i(current));
1723 log_frames(current);
1724 #endif
1725
1726 CONT_JFR_ONLY(EventContinuationFreeze event;)
1727
1728 ContinuationEntry* entry = current->last_continuation();
1729
1730 oop oopCont = entry->cont_oop(current);
1731 assert(oopCont == current->last_continuation()->cont_oop(current), "");
1732 assert(ContinuationEntry::assert_entry_frame_laid_out(current), "");
1733
1734 verify_continuation(oopCont);
1735 ContinuationWrapper cont(current, oopCont);
1736 log_develop_debug(continuations)("FREEZE #" INTPTR_FORMAT " " INTPTR_FORMAT, cont.hash(), p2i((oopDesc*)oopCont));
1737
1738 assert(entry->is_virtual_thread() == (entry->scope(current) == java_lang_VirtualThread::vthread_scope()), "");
1739
1740 assert(LockingMode == LM_LEGACY || (current->held_monitor_count() == 0 && current->jni_monitor_count() == 0),
1741 "Held monitor count should only be used for LM_LEGACY: " INT64_FORMAT " JNI: " INT64_FORMAT, (int64_t)current->held_monitor_count(), (int64_t)current->jni_monitor_count());
1742
1743 if (entry->is_pinned() || current->held_monitor_count() > 0) {
1744 log_develop_debug(continuations)("PINNED due to critical section/hold monitor");
1745 verify_continuation(cont.continuation());
1746 freeze_result res = entry->is_pinned() ? freeze_pinned_cs : freeze_pinned_monitor;
1747 if (!preempt) {
1748 JFR_ONLY(current->set_last_freeze_fail_result(res);)
1749 }
1750 log_develop_trace(continuations)("=== end of freeze (fail %d)", res);
1751 // Avoid Thread.yield() loops without safepoint polls.
1752 if (SafepointMechanism::should_process(current) && !preempt) {
1753 cont.done(); // allow safepoint
1754 ThreadInVMfromJava tivmfj(current);
1755 }
1756 return res;
1757 }
1758
1759 Freeze<ConfigT> freeze(current, cont, sp, preempt);
1760
1761 assert(!current->cont_fastpath() || freeze.check_valid_fast_path(), "");
1762 bool fast = UseContinuationFastPath && current->cont_fastpath();
1763 if (fast && freeze.size_if_fast_freeze_available() > 0) {
1764 freeze.freeze_fast_existing_chunk();
1765 CONT_JFR_ONLY(freeze.jfr_info().post_jfr_event(&event, oopCont, current);)
1766 return !preempt ? freeze_epilog(cont) : preempt_epilog(cont, freeze_ok, freeze.last_frame());
1767 }
1768
1769 if (preempt) {
1770 JvmtiSampledObjectAllocEventCollector jsoaec(false);
1771 freeze.set_jvmti_event_collector(&jsoaec);
1772
1773 freeze_result res = fast ? freeze.try_freeze_fast() : freeze.freeze_slow();
1774
1775 CONT_JFR_ONLY(freeze.jfr_info().post_jfr_event(&event, oopCont, current);)
1776 preempt_epilog(cont, res, freeze.last_frame());
1777 return res;
1778 }
1779
1780 log_develop_trace(continuations)("chunk unavailable; transitioning to VM");
1781 assert(current == JavaThread::current(), "must be current thread");
1782 JRT_BLOCK
1783 // delays a possible JvmtiSampledObjectAllocEventCollector in alloc_chunk
1784 JvmtiSampledObjectAllocEventCollector jsoaec(false);
1785 freeze.set_jvmti_event_collector(&jsoaec);
1786
1787 freeze_result res = fast ? freeze.try_freeze_fast() : freeze.freeze_slow();
1788
1789 CONT_JFR_ONLY(freeze.jfr_info().post_jfr_event(&event, oopCont, current);)
1790 freeze_epilog(current, cont, res);
1791 cont.done(); // allow safepoint in the transition back to Java
1792 return res;
1793 JRT_BLOCK_END
1794 }
1795
1796 static freeze_result is_pinned0(JavaThread* thread, oop cont_scope, bool safepoint) {
1797 ContinuationEntry* entry = thread->last_continuation();
1798 if (entry == nullptr) {
1799 return freeze_ok;
1800 }
1801 if (entry->is_pinned()) {
1802 return freeze_pinned_cs;
1803 } else if (thread->held_monitor_count() > 0) {
1804 return freeze_pinned_monitor;
1805 }
1806
1807 RegisterMap map(thread,
1808 RegisterMap::UpdateMap::include,
1809 RegisterMap::ProcessFrames::skip,
1810 RegisterMap::WalkContinuation::skip);
1811 map.set_include_argument_oops(false);
1812 frame f = thread->last_frame();
1813
1814 if (!safepoint) {
1815 f = f.sender(&map); // this is the yield frame
1816 } else { // safepoint yield
1817 #if (defined(X86) || defined(AARCH64) || defined(RISCV64)) && !defined(ZERO)
1818 f.set_fp(f.real_fp()); // Instead of this, maybe in ContinuationWrapper::set_last_frame always use the real_fp?
1819 #else
1820 Unimplemented();
1821 #endif
1822 if (!Interpreter::contains(f.pc())) {
1823 assert(ContinuationHelper::Frame::is_stub(f.cb()), "must be");
1824 assert(f.oop_map() != nullptr, "must be");
1825 f.oop_map()->update_register_map(&f, &map); // we have callee-save registers in this case
1826 }
1827 }
1828
1829 while (true) {
1830 if ((f.is_interpreted_frame() && f.interpreter_frame_method()->is_native()) || f.is_native_frame()) {
1831 return freeze_pinned_native;
1832 }
1833
1834 f = f.sender(&map);
1835 if (!Continuation::is_frame_in_continuation(entry, f)) {
1836 oop scope = jdk_internal_vm_Continuation::scope(entry->cont_oop(thread));
1837 if (scope == cont_scope) {
1838 break;
1839 }
1840 intx monitor_count = entry->parent_held_monitor_count();
1841 entry = entry->parent();
1842 if (entry == nullptr) {
1843 break;
1844 }
1845 if (entry->is_pinned()) {
1846 return freeze_pinned_cs;
1847 } else if (monitor_count > 0) {
1848 return freeze_pinned_monitor;
1849 }
1850 }
1851 }
1852 return freeze_ok;
1853 }
1854
1855 /////////////// THAW ////
1856
1857 static int thaw_size(stackChunkOop chunk) {
1858 int size = chunk->max_thawing_size();
1859 size += frame::metadata_words; // For the top pc+fp in push_return_frame or top = stack_sp - frame::metadata_words in thaw_fast
1860 size += 2*frame::align_wiggle; // in case of alignments at the top and bottom
1861 return size;
1862 }
1863
1864 // make room on the stack for thaw
1865 // returns the size in bytes, or 0 on failure
1866 static inline int prepare_thaw_internal(JavaThread* thread, bool return_barrier) {
1867 log_develop_trace(continuations)("~~~~ prepare_thaw return_barrier: %d", return_barrier);
1868
1869 assert(thread == JavaThread::current(), "");
1870
1871 ContinuationEntry* ce = thread->last_continuation();
1872 assert(ce != nullptr, "");
1873 oop continuation = ce->cont_oop(thread);
1874 assert(continuation == get_continuation(thread), "");
1875 verify_continuation(continuation);
1876
1877 stackChunkOop chunk = jdk_internal_vm_Continuation::tail(continuation);
1878 assert(chunk != nullptr, "");
1879
1880 // The tail can be empty because it might still be available for another freeze.
1881 // However, here we want to thaw, so we get rid of it (it will be GCed).
1882 if (UNLIKELY(chunk->is_empty())) {
1883 chunk = chunk->parent();
1884 assert(chunk != nullptr, "");
1885 assert(!chunk->is_empty(), "");
1886 jdk_internal_vm_Continuation::set_tail(continuation, chunk);
1887 }
1888
1889 // Verification
1890 chunk->verify();
1891 assert(chunk->max_thawing_size() > 0, "chunk invariant violated; expected to not be empty");
1892
1893 // Only make space for the last chunk because we only thaw from the last chunk
1894 int size = thaw_size(chunk) << LogBytesPerWord;
1895
1896 const address bottom = (address)thread->last_continuation()->entry_sp();
1897 // 300 is an estimate for stack size taken for this native code, in addition to StackShadowPages
1898 // for the Java frames in the check below.
1899 if (!stack_overflow_check(thread, size + 300, bottom)) {
1900 return 0;
1901 }
1902
1903 log_develop_trace(continuations)("prepare_thaw bottom: " INTPTR_FORMAT " top: " INTPTR_FORMAT " size: %d",
1904 p2i(bottom), p2i(bottom - size), size);
1905 return size;
1906 }
1907
1908 class ThawBase : public StackObj {
1909 protected:
1910 JavaThread* _thread;
1911 ContinuationWrapper& _cont;
1912 CONT_JFR_ONLY(FreezeThawJfrInfo _jfr_info;)
1913
1914 intptr_t* _fastpath;
1915 bool _barriers;
1916 bool _preempted_case;
1917 intptr_t* _top_unextended_sp_before_thaw;
1918 int _align_size;
1919 DEBUG_ONLY(intptr_t* _top_stack_address);
1920
1921 StackChunkFrameStream<ChunkFrames::Mixed> _stream;
1922
1923 NOT_PRODUCT(int _frames;)
1924
1925 protected:
1926 ThawBase(JavaThread* thread, ContinuationWrapper& cont) :
1927 _thread(thread), _cont(cont),
1928 _fastpath(nullptr) {
1929 DEBUG_ONLY(_top_unextended_sp_before_thaw = nullptr;)
1930 assert (cont.tail() != nullptr, "no last chunk");
1931 DEBUG_ONLY(_top_stack_address = _cont.entrySP() - thaw_size(cont.tail());)
1932 }
1933
1934 void clear_chunk(stackChunkOop chunk);
1935 template<bool check_stub>
1936 int remove_top_compiled_frame_from_chunk(stackChunkOop chunk, int &argsize);
1937 void copy_from_chunk(intptr_t* from, intptr_t* to, int size);
1938
1939 void thaw_lockstack(stackChunkOop chunk);
1940
1941 // fast path
1942 inline void prefetch_chunk_pd(void* start, int size_words);
1943 void patch_return(intptr_t* sp, bool is_last);
1944
1945 intptr_t* handle_preempted_continuation(intptr_t* sp, Continuation::preempt_kind preempt_kind, bool fast_case);
1946 inline intptr_t* push_cleanup_continuation();
1947 void throw_interrupted_exception(JavaThread* current, frame& top);
1948
1949 void recurse_thaw(const frame& heap_frame, frame& caller, int num_frames, bool top_on_preempt_case);
1950 void finish_thaw(frame& f);
1951
1952 private:
1953 template<typename FKind> bool recurse_thaw_java_frame(frame& caller, int num_frames);
1954 void finalize_thaw(frame& entry, int argsize);
1955
1956 inline bool seen_by_gc();
1957
1958 inline void before_thaw_java_frame(const frame& hf, const frame& caller, bool bottom, int num_frame);
1959 inline void after_thaw_java_frame(const frame& f, bool bottom);
1960 inline void patch(frame& f, const frame& caller, bool bottom);
1961 void clear_bitmap_bits(address start, address end);
1962
1963 NOINLINE void recurse_thaw_interpreted_frame(const frame& hf, frame& caller, int num_frames);
1964 void recurse_thaw_compiled_frame(const frame& hf, frame& caller, int num_frames, bool stub_caller);
1965 void recurse_thaw_stub_frame(const frame& hf, frame& caller, int num_frames);
1966 void recurse_thaw_native_frame(const frame& hf, frame& caller, int num_frames);
1967
1968 void push_return_frame(frame& f);
1969 inline frame new_entry_frame();
1970 template<typename FKind> frame new_stack_frame(const frame& hf, frame& caller, bool bottom);
1971 inline void patch_pd(frame& f, const frame& sender);
1972 inline void patch_pd(frame& f, intptr_t* caller_sp);
1973 inline intptr_t* align(const frame& hf, intptr_t* frame_sp, frame& caller, bool bottom);
1974
1975 void maybe_set_fastpath(intptr_t* sp) { if (sp > _fastpath) _fastpath = sp; }
1976
1977 static inline void derelativize_interpreted_frame_metadata(const frame& hf, const frame& f);
1978
1979 public:
1980 CONT_JFR_ONLY(FreezeThawJfrInfo& jfr_info() { return _jfr_info; })
1981 };
1982
1983 template <typename ConfigT>
1984 class Thaw : public ThawBase {
1985 public:
1986 Thaw(JavaThread* thread, ContinuationWrapper& cont) : ThawBase(thread, cont) {}
1987
1988 inline bool can_thaw_fast(stackChunkOop chunk) {
1989 return !_barriers
1990 && _thread->cont_fastpath_thread_state()
1991 && !chunk->has_thaw_slowpath_condition()
1992 && !PreserveFramePointer;
1993 }
1994
1995 inline intptr_t* thaw(Continuation::thaw_kind kind);
1996 template<bool check_stub = false>
1997 NOINLINE intptr_t* thaw_fast(stackChunkOop chunk);
1998 NOINLINE intptr_t* thaw_slow(stackChunkOop chunk, Continuation::thaw_kind kind);
1999 inline void patch_caller_links(intptr_t* sp, intptr_t* bottom);
2000 };
2001
2002 template <typename ConfigT>
2003 inline intptr_t* Thaw<ConfigT>::thaw(Continuation::thaw_kind kind) {
2004 verify_continuation(_cont.continuation());
2005 assert(!jdk_internal_vm_Continuation::done(_cont.continuation()), "");
2006 assert(!_cont.is_empty(), "");
2007
2008 stackChunkOop chunk = _cont.tail();
2009 assert(chunk != nullptr, "guaranteed by prepare_thaw");
2010 assert(!chunk->is_empty(), "guaranteed by prepare_thaw");
2011
2012 _barriers = chunk->requires_barriers();
2013 return (LIKELY(can_thaw_fast(chunk))) ? thaw_fast(chunk)
2014 : thaw_slow(chunk, kind);
2015 }
2016
2017 class ReconstructedStack : public StackObj {
2018 intptr_t* _base; // _cont.entrySP(); // top of the entry frame
2019 int _thaw_size;
2020 int _argsize;
2021 public:
2022 ReconstructedStack(intptr_t* base, int thaw_size, int argsize)
2023 : _base(base), _thaw_size(thaw_size - (argsize == 0 ? frame::metadata_words_at_top : 0)), _argsize(argsize) {
2024 // The only possible source of misalignment is stack-passed arguments b/c compiled frames are 16-byte aligned.
2025 assert(argsize != 0 || (_base - _thaw_size) == ContinuationHelper::frame_align_pointer(_base - _thaw_size), "");
2026 // We're at most one alignment word away from entrySP
2027 assert(_base - 1 <= top() + total_size() + frame::metadata_words_at_bottom, "missed entry frame");
2028 }
2029
2030 int entry_frame_extension() const { return _argsize + (_argsize > 0 ? frame::metadata_words_at_top : 0); }
2031
2032 // top and bottom stack pointers
2033 intptr_t* sp() const { return ContinuationHelper::frame_align_pointer(_base - _thaw_size); }
2034 intptr_t* bottom_sp() const { return ContinuationHelper::frame_align_pointer(_base - entry_frame_extension()); }
2035
2036 // several operations operate on the totality of the stack being reconstructed,
2037 // including the metadata words
2038 intptr_t* top() const { return sp() - frame::metadata_words_at_bottom; }
2039 int total_size() const { return _thaw_size + frame::metadata_words_at_bottom; }
2040 };
2041
2042 inline void ThawBase::clear_chunk(stackChunkOop chunk) {
2043 chunk->set_sp(chunk->bottom());
2044 chunk->set_max_thawing_size(0);
2045 }
2046
2047 template<bool check_stub>
2048 int ThawBase::remove_top_compiled_frame_from_chunk(stackChunkOop chunk, int &argsize) {
2049 bool empty = false;
2050 StackChunkFrameStream<ChunkFrames::CompiledOnly> f(chunk);
2051 DEBUG_ONLY(intptr_t* const chunk_sp = chunk->start_address() + chunk->sp();)
2052 assert(chunk_sp == f.sp(), "");
2053 assert(chunk_sp == f.unextended_sp(), "");
2054
2055 int frame_size = f.cb()->frame_size();
2056 argsize = f.stack_argsize();
2057
2058 assert(!f.is_stub() || check_stub, "");
2059 if (check_stub && f.is_stub()) {
2060 // If we don't thaw the top compiled frame too, after restoring the saved
2061 // registers back in Java, we would hit the return barrier to thaw one more
2062 // frame effectively overwriting the restored registers during that call.
2063 f.next(SmallRegisterMap::instance(), true /* stop */);
2064 assert(!f.is_done(), "");
2065
2066 f.get_cb();
2067 assert(f.is_compiled(), "");
2068 frame_size += f.cb()->frame_size();
2069 argsize = f.stack_argsize();
2070
2071 if (f.cb()->as_nmethod()->is_marked_for_deoptimization()) {
2072 // The caller of the runtime stub when the continuation is preempted is not at a
2073 // Java call instruction, and so cannot rely on nmethod patching for deopt.
2074 log_develop_trace(continuations)("Deoptimizing runtime stub caller");
2075 f.to_frame().deoptimize(nullptr); // the null thread simply avoids the assertion in deoptimize which we're not set up for
2076 }
2077 }
2078
2079 f.next(SmallRegisterMap::instance(), true /* stop */);
2080 empty = f.is_done();
2081 assert(!empty || argsize == chunk->argsize(), "");
2082
2083 if (empty) {
2084 clear_chunk(chunk);
2085 } else {
2086 chunk->set_sp(chunk->sp() + frame_size);
2087 chunk->set_max_thawing_size(chunk->max_thawing_size() - frame_size);
2088 // We set chunk->pc to the return pc into the next frame
2089 chunk->set_pc(f.pc());
2090 #ifdef ASSERT
2091 {
2092 intptr_t* retaddr_slot = (chunk_sp
2093 + frame_size
2094 - frame::sender_sp_ret_address_offset());
2095 assert(f.pc() == ContinuationHelper::return_address_at(retaddr_slot),
2096 "unexpected pc");
2097 }
2098 #endif
2099 }
2100 assert(empty == chunk->is_empty(), "");
2101 // returns the size required to store the frame on stack, and because it is a
2102 // compiled frame, it must include a copy of the arguments passed by the caller
2103 return frame_size + argsize + frame::metadata_words_at_top;
2104 }
2105
2106 void ThawBase::thaw_lockstack(stackChunkOop chunk) {
2107 int lockStackSize = chunk->lockstack_size();
2108 assert(lockStackSize > 0 && lockStackSize <= LockStack::CAPACITY, "");
2109
2110 oop tmp_lockstack[LockStack::CAPACITY];
2111 chunk->transfer_lockstack(tmp_lockstack, _barriers);
2112 _thread->lock_stack().move_from_address(tmp_lockstack, lockStackSize);
2113
2114 chunk->set_lockstack_size(0);
2115 chunk->set_has_lockstack(false);
2116 }
2117
2118 void ThawBase::copy_from_chunk(intptr_t* from, intptr_t* to, int size) {
2119 assert(to >= _top_stack_address, "overwrote past thawing space"
2120 " to: " INTPTR_FORMAT " top_address: " INTPTR_FORMAT, p2i(to), p2i(_top_stack_address));
2121 assert(to + size <= _cont.entrySP(), "overwrote past thawing space");
2122 _cont.tail()->copy_from_chunk_to_stack(from, to, size);
2123 CONT_JFR_ONLY(_jfr_info.record_size_copied(size);)
2124 }
2125
2126 void ThawBase::patch_return(intptr_t* sp, bool is_last) {
2127 log_develop_trace(continuations)("thaw_fast patching -- sp: " INTPTR_FORMAT, p2i(sp));
2128
2129 address pc = !is_last ? StubRoutines::cont_returnBarrier() : _cont.entryPC();
2130 ContinuationHelper::patch_return_address_at(
2131 sp - frame::sender_sp_ret_address_offset(),
2132 pc);
2133 }
2134
2135 template <typename ConfigT>
2136 template<bool check_stub>
2137 NOINLINE intptr_t* Thaw<ConfigT>::thaw_fast(stackChunkOop chunk) {
2138 assert(chunk == _cont.tail(), "");
2139 assert(!chunk->has_mixed_frames(), "");
2140 assert(!chunk->requires_barriers(), "");
2141 assert(!chunk->has_bitmap(), "");
2142 assert(!_thread->is_interp_only_mode(), "");
2143
2144 LogTarget(Trace, continuations) lt;
2145 if (lt.develop_is_enabled()) {
2146 LogStream ls(lt);
2147 ls.print_cr("thaw_fast");
2148 chunk->print_on(true, &ls);
2149 }
2150
2151 // Below this heuristic, we thaw the whole chunk, above it we thaw just one frame.
2152 static const int threshold = 500; // words
2153
2154 const int full_chunk_size = chunk->stack_size() - chunk->sp(); // this initial size could be reduced if it's a partial thaw
2155 int argsize, thaw_size;
2156
2157 intptr_t* const chunk_sp = chunk->start_address() + chunk->sp();
2158
2159 bool partial, empty;
2160 if (LIKELY(!TEST_THAW_ONE_CHUNK_FRAME && (full_chunk_size < threshold))) {
2161 prefetch_chunk_pd(chunk->start_address(), full_chunk_size); // prefetch anticipating memcpy starting at highest address
2162
2163 partial = false;
2164 argsize = chunk->argsize(); // must be called *before* clearing the chunk
2165 clear_chunk(chunk);
2166 thaw_size = full_chunk_size;
2167 empty = true;
2168 } else { // thaw a single frame
2169 partial = true;
2170 thaw_size = remove_top_compiled_frame_from_chunk<check_stub>(chunk, argsize);
2171 empty = chunk->is_empty();
2172 }
2173
2174 // Are we thawing the last frame(s) in the continuation
2175 const bool is_last = empty && chunk->parent() == nullptr;
2176 assert(!is_last || argsize == 0, "");
2177
2178 log_develop_trace(continuations)("thaw_fast partial: %d is_last: %d empty: %d size: %d argsize: %d entrySP: " PTR_FORMAT,
2179 partial, is_last, empty, thaw_size, argsize, p2i(_cont.entrySP()));
2180
2181 ReconstructedStack rs(_cont.entrySP(), thaw_size, argsize);
2182
2183 // also copy metadata words at frame bottom
2184 copy_from_chunk(chunk_sp - frame::metadata_words_at_bottom, rs.top(), rs.total_size());
2185
2186 // update the ContinuationEntry
2187 _cont.set_argsize(argsize);
2188 log_develop_trace(continuations)("setting entry argsize: %d", _cont.argsize());
2189 assert(rs.bottom_sp() == _cont.entry()->bottom_sender_sp(), "");
2190
2191 // install the return barrier if not last frame, or the entry's pc if last
2192 patch_return(rs.bottom_sp(), is_last);
2193
2194 // insert the back links from callee to caller frames
2195 patch_caller_links(rs.top(), rs.top() + rs.total_size());
2196
2197 assert(is_last == _cont.is_empty(), "");
2198 assert(_cont.chunk_invariant(), "");
2199
2200 #if CONT_JFR
2201 EventContinuationThawFast e;
2202 if (e.should_commit()) {
2203 e.set_id(cast_from_oop<u8>(chunk));
2204 e.set_size(thaw_size << LogBytesPerWord);
2205 e.set_full(!partial);
2206 e.commit();
2207 }
2208 #endif
2209
2210 #ifdef ASSERT
2211 set_anchor(_thread, rs.sp());
2212 log_frames(_thread);
2213 if (LoomDeoptAfterThaw) {
2214 do_deopt_after_thaw(_thread);
2215 }
2216 clear_anchor(_thread);
2217 #endif
2218
2219 return rs.sp();
2220 }
2221
2222 inline bool ThawBase::seen_by_gc() {
2223 return _barriers || _cont.tail()->is_gc_mode();
2224 }
2225
2226 static inline void relativize_chunk_concurrently(stackChunkOop chunk) {
2227 #if INCLUDE_ZGC || INCLUDE_SHENANDOAHGC
2228 if (UseZGC || UseShenandoahGC) {
2229 chunk->relativize_derived_pointers_concurrently();
2230 }
2231 #endif
2232 }
2233
2234 template <typename ConfigT>
2235 NOINLINE intptr_t* Thaw<ConfigT>::thaw_slow(stackChunkOop chunk, Continuation::thaw_kind kind) {
2236 Continuation::preempt_kind preempt_kind;
2237 bool retry_fast_path = false;
2238
2239 _preempted_case = chunk->preempted();
2240 if (_preempted_case) {
2241 ObjectWaiter* waiter = java_lang_VirtualThread::objectWaiter(_thread->vthread());
2242 if (waiter != nullptr) {
2243 // Mounted again after preemption. Resume the pending monitor operation,
2244 // which will be either a monitorenter or Object.wait() call.
2245 ObjectMonitor* mon = waiter->monitor();
2246 preempt_kind = waiter->is_wait() ? Continuation::freeze_on_wait : Continuation::freeze_on_monitorenter;
2247
2248 bool mon_acquired = mon->resume_operation(_thread, waiter, _cont);
2249 assert(!mon_acquired || mon->has_owner(_thread), "invariant");
2250 if (!mon_acquired) {
2251 // Failed to acquire monitor. Return to enterSpecial to unmount again.
2252 return push_cleanup_continuation();
2253 }
2254 chunk = _cont.tail(); // reload oop in case of safepoint in resume_operation (if posting JVMTI events).
2255 } else {
2256 // Preemption cancelled in moniterenter case. We actually acquired
2257 // the monitor after freezing all frames so nothing to do.
2258 preempt_kind = Continuation::freeze_on_monitorenter;
2259 }
2260 // Call this first to avoid racing with GC threads later when modifying the chunk flags.
2261 relativize_chunk_concurrently(chunk);
2262 chunk->set_preempted(false);
2263 retry_fast_path = true;
2264 } else {
2265 relativize_chunk_concurrently(chunk);
2266 }
2267
2268 // On first thaw after freeze restore oops to the lockstack if any.
2269 assert(chunk->lockstack_size() == 0 || kind == Continuation::thaw_top, "");
2270 if (kind == Continuation::thaw_top && chunk->lockstack_size() > 0) {
2271 thaw_lockstack(chunk);
2272 retry_fast_path = true;
2273 }
2274
2275 // Retry the fast path now that we possibly cleared the FLAG_HAS_LOCKSTACK
2276 // and FLAG_PREEMPTED flags from the stackChunk.
2277 if (retry_fast_path && can_thaw_fast(chunk)) {
2278 intptr_t* sp = thaw_fast<true>(chunk);
2279 if (_preempted_case) {
2280 return handle_preempted_continuation(sp, preempt_kind, true /* fast_case */);
2281 }
2282 return sp;
2283 }
2284
2285 LogTarget(Trace, continuations) lt;
2286 if (lt.develop_is_enabled()) {
2287 LogStream ls(lt);
2288 ls.print_cr("thaw slow return_barrier: %d " INTPTR_FORMAT, kind, p2i(chunk));
2289 chunk->print_on(true, &ls);
2290 }
2291
2292 #if CONT_JFR
2293 EventContinuationThawSlow e;
2294 if (e.should_commit()) {
2295 e.set_id(cast_from_oop<u8>(_cont.continuation()));
2296 e.commit();
2297 }
2298 #endif
2299
2300 DEBUG_ONLY(_frames = 0;)
2301 _align_size = 0;
2302 int num_frames = kind == Continuation::thaw_top ? 2 : 1;
2303
2304 _stream = StackChunkFrameStream<ChunkFrames::Mixed>(chunk);
2305 _top_unextended_sp_before_thaw = _stream.unextended_sp();
2306
2307 frame heap_frame = _stream.to_frame();
2308 if (lt.develop_is_enabled()) {
2309 LogStream ls(lt);
2310 ls.print_cr("top hframe before (thaw):");
2311 assert(heap_frame.is_heap_frame(), "should have created a relative frame");
2312 heap_frame.print_value_on(&ls);
2313 }
2314
2315 frame caller; // the thawed caller on the stack
2316 recurse_thaw(heap_frame, caller, num_frames, _preempted_case);
2317 finish_thaw(caller); // caller is now the topmost thawed frame
2318 _cont.write();
2319
2320 assert(_cont.chunk_invariant(), "");
2321
2322 JVMTI_ONLY(invalidate_jvmti_stack(_thread));
2323
2324 _thread->set_cont_fastpath(_fastpath);
2325
2326 intptr_t* sp = caller.sp();
2327
2328 if (_preempted_case) {
2329 return handle_preempted_continuation(sp, preempt_kind, false /* fast_case */);
2330 }
2331 return sp;
2332 }
2333
2334 void ThawBase::recurse_thaw(const frame& heap_frame, frame& caller, int num_frames, bool top_on_preempt_case) {
2335 log_develop_debug(continuations)("thaw num_frames: %d", num_frames);
2336 assert(!_cont.is_empty(), "no more frames");
2337 assert(num_frames > 0, "");
2338 assert(!heap_frame.is_empty(), "");
2339
2340 if (top_on_preempt_case && (heap_frame.is_native_frame() || heap_frame.is_runtime_frame())) {
2341 heap_frame.is_native_frame() ? recurse_thaw_native_frame(heap_frame, caller, 2) : recurse_thaw_stub_frame(heap_frame, caller, 2);
2342 } else if (!heap_frame.is_interpreted_frame()) {
2343 recurse_thaw_compiled_frame(heap_frame, caller, num_frames, false);
2344 } else {
2345 recurse_thaw_interpreted_frame(heap_frame, caller, num_frames);
2346 }
2347 }
2348
2349 template<typename FKind>
2350 bool ThawBase::recurse_thaw_java_frame(frame& caller, int num_frames) {
2351 assert(num_frames > 0, "");
2352
2353 DEBUG_ONLY(_frames++;)
2354
2355 int argsize = _stream.stack_argsize();
2356
2357 _stream.next(SmallRegisterMap::instance());
2358 assert(_stream.to_frame().is_empty() == _stream.is_done(), "");
2359
2360 // we never leave a compiled caller of an interpreted frame as the top frame in the chunk
2361 // as it makes detecting that situation and adjusting unextended_sp tricky
2362 if (num_frames == 1 && !_stream.is_done() && FKind::interpreted && _stream.is_compiled()) {
2363 log_develop_trace(continuations)("thawing extra compiled frame to not leave a compiled interpreted-caller at top");
2364 num_frames++;
2365 }
2366
2367 if (num_frames == 1 || _stream.is_done()) { // end recursion
2368 finalize_thaw(caller, FKind::interpreted ? 0 : argsize);
2369 return true; // bottom
2370 } else { // recurse
2371 recurse_thaw(_stream.to_frame(), caller, num_frames - 1, false /* top_on_preempt_case */);
2372 return false;
2373 }
2374 }
2375
2376 void ThawBase::finalize_thaw(frame& entry, int argsize) {
2377 stackChunkOop chunk = _cont.tail();
2378
2379 if (!_stream.is_done()) {
2380 assert(_stream.sp() >= chunk->sp_address(), "");
2381 chunk->set_sp(chunk->to_offset(_stream.sp()));
2382 chunk->set_pc(_stream.pc());
2383 } else {
2384 chunk->set_sp(chunk->bottom());
2385 chunk->set_pc(nullptr);
2386 }
2387 assert(_stream.is_done() == chunk->is_empty(), "");
2388
2389 int total_thawed = pointer_delta_as_int(_stream.unextended_sp(), _top_unextended_sp_before_thaw);
2390 chunk->set_max_thawing_size(chunk->max_thawing_size() - total_thawed);
2391
2392 _cont.set_argsize(argsize);
2393 entry = new_entry_frame();
2394
2395 assert(entry.sp() == _cont.entrySP(), "");
2396 assert(Continuation::is_continuation_enterSpecial(entry), "");
2397 assert(_cont.is_entry_frame(entry), "");
2398 }
2399
2400 inline void ThawBase::before_thaw_java_frame(const frame& hf, const frame& caller, bool bottom, int num_frame) {
2401 LogTarget(Trace, continuations) lt;
2402 if (lt.develop_is_enabled()) {
2403 LogStream ls(lt);
2404 ls.print_cr("======== THAWING FRAME: %d", num_frame);
2405 assert(hf.is_heap_frame(), "should be");
2406 hf.print_value_on(&ls);
2407 }
2408 assert(bottom == _cont.is_entry_frame(caller), "bottom: %d is_entry_frame: %d", bottom, _cont.is_entry_frame(hf));
2409 }
2410
2411 inline void ThawBase::after_thaw_java_frame(const frame& f, bool bottom) {
2412 #ifdef ASSERT
2413 LogTarget(Trace, continuations) lt;
2414 if (lt.develop_is_enabled()) {
2415 LogStream ls(lt);
2416 ls.print_cr("thawed frame:");
2417 print_frame_layout(f, false, &ls); // f.print_on(&ls);
2418 }
2419 #endif
2420 }
2421
2422 inline void ThawBase::patch(frame& f, const frame& caller, bool bottom) {
2423 assert(!bottom || caller.fp() == _cont.entryFP(), "");
2424 if (bottom) {
2425 ContinuationHelper::Frame::patch_pc(caller, _cont.is_empty() ? caller.pc()
2426 : StubRoutines::cont_returnBarrier());
2427 } else {
2428 // caller might have been deoptimized during thaw but we've overwritten the return address when copying f from the heap.
2429 // If the caller is not deoptimized, pc is unchanged.
2430 ContinuationHelper::Frame::patch_pc(caller, caller.raw_pc());
2431 }
2432
2433 patch_pd(f, caller);
2434
2435 if (f.is_interpreted_frame()) {
2436 ContinuationHelper::InterpretedFrame::patch_sender_sp(f, caller);
2437 }
2438
2439 assert(!bottom || !_cont.is_empty() || Continuation::is_continuation_entry_frame(f, nullptr), "");
2440 assert(!bottom || (_cont.is_empty() != Continuation::is_cont_barrier_frame(f)), "");
2441 }
2442
2443 void ThawBase::clear_bitmap_bits(address start, address end) {
2444 assert(is_aligned(start, wordSize), "should be aligned: " PTR_FORMAT, p2i(start));
2445 assert(is_aligned(end, VMRegImpl::stack_slot_size), "should be aligned: " PTR_FORMAT, p2i(end));
2446
2447 // we need to clear the bits that correspond to arguments as they reside in the caller frame
2448 // or they will keep objects that are otherwise unreachable alive.
2449
2450 // Align `end` if UseCompressedOops is not set to avoid UB when calculating the bit index, since
2451 // `end` could be at an odd number of stack slots from `start`, i.e might not be oop aligned.
2452 // If that's the case the bit range corresponding to the last stack slot should not have bits set
2453 // anyways and we assert that before returning.
2454 address effective_end = UseCompressedOops ? end : align_down(end, wordSize);
2455 log_develop_trace(continuations)("clearing bitmap for " INTPTR_FORMAT " - " INTPTR_FORMAT, p2i(start), p2i(effective_end));
2456 stackChunkOop chunk = _cont.tail();
2457 chunk->bitmap().clear_range(chunk->bit_index_for(start), chunk->bit_index_for(effective_end));
2458 assert(effective_end == end || !chunk->bitmap().at(chunk->bit_index_for(effective_end)), "bit should not be set");
2459 }
2460
2461 intptr_t* ThawBase::handle_preempted_continuation(intptr_t* sp, Continuation::preempt_kind preempt_kind, bool fast_case) {
2462 assert(preempt_kind == Continuation::freeze_on_wait || preempt_kind == Continuation::freeze_on_monitorenter, "");
2463 frame top(sp);
2464 assert(top.pc() == *(address*)(sp - frame::sender_sp_ret_address_offset()), "");
2465
2466 #if INCLUDE_JVMTI
2467 // Finish the VTMS transition.
2468 assert(_thread->is_in_VTMS_transition(), "must be");
2469 bool is_vthread = Continuation::continuation_scope(_cont.continuation()) == java_lang_VirtualThread::vthread_scope();
2470 if (is_vthread) {
2471 if (JvmtiVTMSTransitionDisabler::VTMS_notify_jvmti_events()) {
2472 jvmti_mount_end(_thread, _cont, top);
2473 } else {
2474 _thread->set_is_in_VTMS_transition(false);
2475 java_lang_Thread::set_is_in_VTMS_transition(_thread->vthread(), false);
2476 }
2477 }
2478 #endif
2479
2480 if (fast_case) {
2481 // If we thawed in the slow path the runtime stub/native wrapper frame already
2482 // has the correct fp (see ThawBase::new_stack_frame). On the fast path though,
2483 // we copied the original fp at the time of freeze which now will have to be fixed.
2484 assert(top.is_runtime_frame() || top.is_native_frame(), "");
2485 int fsize = top.cb()->frame_size();
2486 patch_pd(top, sp + fsize);
2487 }
2488
2489 if (preempt_kind == Continuation::freeze_on_wait) {
2490 // Check now if we need to throw IE exception.
2491 if (_thread->pending_interrupted_exception()) {
2492 throw_interrupted_exception(_thread, top);
2493 _thread->set_pending_interrupted_exception(false);
2494 }
2495 } else if (top.is_runtime_frame()) {
2496 // The continuation might now run on a different platform thread than the previous time so
2497 // we need to adjust the current thread saved in the stub frame before restoring registers.
2498 JavaThread** thread_addr = frame::saved_thread_address(top);
2499 if (thread_addr != nullptr) *thread_addr = _thread;
2500 }
2501 return sp;
2502 }
2503
2504 void ThawBase::throw_interrupted_exception(JavaThread* current, frame& top) {
2505 ContinuationWrapper::SafepointOp so(current, _cont);
2506 // Since we might safepoint set the anchor so that the stack can be walked.
2507 set_anchor(current, top.sp());
2508 JRT_BLOCK
2509 THROW(vmSymbols::java_lang_InterruptedException());
2510 JRT_BLOCK_END
2511 clear_anchor(current);
2512 }
2513
2514 NOINLINE void ThawBase::recurse_thaw_interpreted_frame(const frame& hf, frame& caller, int num_frames) {
2515 assert(hf.is_interpreted_frame(), "");
2516
2517 if (UNLIKELY(seen_by_gc())) {
2518 _cont.tail()->do_barriers<stackChunkOopDesc::BarrierType::Store>(_stream, SmallRegisterMap::instance());
2519 }
2520
2521 const bool is_bottom_frame = recurse_thaw_java_frame<ContinuationHelper::InterpretedFrame>(caller, num_frames);
2522
2523 DEBUG_ONLY(before_thaw_java_frame(hf, caller, is_bottom_frame, num_frames);)
2524
2525 _align_size += frame::align_wiggle; // possible added alignment for internal interpreted frame alignment om AArch64
2526
2527 frame f = new_stack_frame<ContinuationHelper::InterpretedFrame>(hf, caller, is_bottom_frame);
2528
2529 intptr_t* const stack_frame_top = f.sp() + frame::metadata_words_at_top;
2530 intptr_t* const stack_frame_bottom = ContinuationHelper::InterpretedFrame::frame_bottom(f);
2531 intptr_t* const heap_frame_top = hf.unextended_sp() + frame::metadata_words_at_top;
2532 intptr_t* const heap_frame_bottom = ContinuationHelper::InterpretedFrame::frame_bottom(hf);
2533
2534 assert(hf.is_heap_frame(), "should be");
2535 assert(!f.is_heap_frame(), "should not be");
2536
2537 const int fsize = pointer_delta_as_int(heap_frame_bottom, heap_frame_top);
2538 assert((stack_frame_bottom == stack_frame_top + fsize), "");
2539
2540 // Some architectures (like AArch64/PPC64/RISC-V) add padding between the locals and the fixed_frame to keep the fp 16-byte-aligned.
2541 // On those architectures we freeze the padding in order to keep the same fp-relative offsets in the fixed_frame.
2542 copy_from_chunk(heap_frame_top, stack_frame_top, fsize);
2543
2544 // Make sure the relativized locals is already set.
2545 assert(f.interpreter_frame_local_at(0) == stack_frame_bottom - 1, "invalid frame bottom");
2546
2547 derelativize_interpreted_frame_metadata(hf, f);
2548 patch(f, caller, is_bottom_frame);
2549
2550 assert(f.is_interpreted_frame_valid(_cont.thread()), "invalid thawed frame");
2551 assert(stack_frame_bottom <= ContinuationHelper::Frame::frame_top(caller), "");
2552
2553 CONT_JFR_ONLY(_jfr_info.record_interpreted_frame();)
2554
2555 maybe_set_fastpath(f.sp());
2556
2557 Method* m = hf.interpreter_frame_method();
2558 assert(!m->is_native() || !is_bottom_frame, "should be top frame of thaw_top case; missing caller frame");
2559 const int locals = m->max_locals();
2560
2561 if (!is_bottom_frame) {
2562 // can only fix caller once this frame is thawed (due to callee saved regs)
2563 _cont.tail()->fix_thawed_frame(caller, SmallRegisterMap::instance());
2564 } else if (_cont.tail()->has_bitmap() && locals > 0) {
2565 assert(hf.is_heap_frame(), "should be");
2566 address start = (address)(heap_frame_bottom - locals);
2567 address end = (address)heap_frame_bottom;
2568 clear_bitmap_bits(start, end);
2569 }
2570
2571 DEBUG_ONLY(after_thaw_java_frame(f, is_bottom_frame);)
2572 caller = f;
2573 }
2574
2575 void ThawBase::recurse_thaw_compiled_frame(const frame& hf, frame& caller, int num_frames, bool stub_caller) {
2576 assert(hf.is_compiled_frame(), "");
2577 assert(_preempted_case || !stub_caller, "stub caller not at preemption");
2578
2579 if (!stub_caller && UNLIKELY(seen_by_gc())) { // recurse_thaw_stub_frame already invoked our barriers with a full regmap
2580 _cont.tail()->do_barriers<stackChunkOopDesc::BarrierType::Store>(_stream, SmallRegisterMap::instance());
2581 }
2582
2583 const bool is_bottom_frame = recurse_thaw_java_frame<ContinuationHelper::CompiledFrame>(caller, num_frames);
2584
2585 DEBUG_ONLY(before_thaw_java_frame(hf, caller, is_bottom_frame, num_frames);)
2586
2587 assert(caller.sp() == caller.unextended_sp(), "");
2588
2589 if ((!is_bottom_frame && caller.is_interpreted_frame()) || (is_bottom_frame && Interpreter::contains(_cont.tail()->pc()))) {
2590 _align_size += frame::align_wiggle; // we add one whether or not we've aligned because we add it in recurse_freeze_compiled_frame
2591 }
2592
2593 // new_stack_frame must construct the resulting frame using hf.pc() rather than hf.raw_pc() because the frame is not
2594 // yet laid out in the stack, and so the original_pc is not stored in it.
2595 // As a result, f.is_deoptimized_frame() is always false and we must test hf to know if the frame is deoptimized.
2596 frame f = new_stack_frame<ContinuationHelper::CompiledFrame>(hf, caller, is_bottom_frame);
2597 intptr_t* const stack_frame_top = f.sp();
2598 intptr_t* const heap_frame_top = hf.unextended_sp();
2599
2600 const int added_argsize = (is_bottom_frame || caller.is_interpreted_frame()) ? hf.compiled_frame_stack_argsize() : 0;
2601 int fsize = ContinuationHelper::CompiledFrame::size(hf) + added_argsize;
2602 assert(fsize <= (int)(caller.unextended_sp() - f.unextended_sp()), "");
2603
2604 intptr_t* from = heap_frame_top - frame::metadata_words_at_bottom;
2605 intptr_t* to = stack_frame_top - frame::metadata_words_at_bottom;
2606 // copy metadata, except the metadata at the top of the (unextended) entry frame
2607 int sz = fsize + frame::metadata_words_at_bottom + (is_bottom_frame && added_argsize == 0 ? 0 : frame::metadata_words_at_top);
2608
2609 // If we're the bottom-most thawed frame, we're writing to within one word from entrySP
2610 // (we might have one padding word for alignment)
2611 assert(!is_bottom_frame || (_cont.entrySP() - 1 <= to + sz && to + sz <= _cont.entrySP()), "");
2612 assert(!is_bottom_frame || hf.compiled_frame_stack_argsize() != 0 || (to + sz && to + sz == _cont.entrySP()), "");
2613
2614 copy_from_chunk(from, to, sz); // copying good oops because we invoked barriers above
2615
2616 patch(f, caller, is_bottom_frame);
2617
2618 // f.is_deoptimized_frame() is always false and we must test hf.is_deoptimized_frame() (see comment above)
2619 assert(!f.is_deoptimized_frame(), "");
2620 if (hf.is_deoptimized_frame()) {
2621 maybe_set_fastpath(f.sp());
2622 } else if (_thread->is_interp_only_mode()
2623 || (stub_caller && f.cb()->as_nmethod()->is_marked_for_deoptimization())) {
2624 // The caller of the safepoint stub when the continuation is preempted is not at a call instruction, and so
2625 // cannot rely on nmethod patching for deopt.
2626 assert(_thread->is_interp_only_mode() || stub_caller, "expected a stub-caller");
2627
2628 log_develop_trace(continuations)("Deoptimizing thawed frame");
2629 DEBUG_ONLY(ContinuationHelper::Frame::patch_pc(f, nullptr));
2630
2631 f.deoptimize(nullptr); // the null thread simply avoids the assertion in deoptimize which we're not set up for
2632 assert(f.is_deoptimized_frame(), "");
2633 assert(ContinuationHelper::Frame::is_deopt_return(f.raw_pc(), f), "");
2634 maybe_set_fastpath(f.sp());
2635 }
2636
2637 if (!is_bottom_frame) {
2638 // can only fix caller once this frame is thawed (due to callee saved regs); this happens on the stack
2639 _cont.tail()->fix_thawed_frame(caller, SmallRegisterMap::instance());
2640 } else if (_cont.tail()->has_bitmap() && added_argsize > 0) {
2641 address start = (address)(heap_frame_top + ContinuationHelper::CompiledFrame::size(hf) + frame::metadata_words_at_top);
2642 int stack_args_slots = f.cb()->as_nmethod()->num_stack_arg_slots(false /* rounded */);
2643 int argsize_in_bytes = stack_args_slots * VMRegImpl::stack_slot_size;
2644 clear_bitmap_bits(start, start + argsize_in_bytes);
2645 }
2646
2647 DEBUG_ONLY(after_thaw_java_frame(f, is_bottom_frame);)
2648 caller = f;
2649 }
2650
2651 void ThawBase::recurse_thaw_stub_frame(const frame& hf, frame& caller, int num_frames) {
2652 DEBUG_ONLY(_frames++;)
2653
2654 if (UNLIKELY(seen_by_gc())) {
2655 // Process the stub's caller here since we might need the full map.
2656 RegisterMap map(nullptr,
2657 RegisterMap::UpdateMap::include,
2658 RegisterMap::ProcessFrames::skip,
2659 RegisterMap::WalkContinuation::skip);
2660 map.set_include_argument_oops(false);
2661 _stream.next(&map);
2662 assert(!_stream.is_done(), "");
2663 _cont.tail()->do_barriers<stackChunkOopDesc::BarrierType::Store>(_stream, &map);
2664 } else {
2665 _stream.next(SmallRegisterMap::instance());
2666 assert(!_stream.is_done(), "");
2667 }
2668
2669 recurse_thaw_compiled_frame(_stream.to_frame(), caller, num_frames, true);
2670
2671 assert(caller.is_compiled_frame(), "");
2672 assert(caller.sp() == caller.unextended_sp(), "");
2673
2674 DEBUG_ONLY(before_thaw_java_frame(hf, caller, false /*is_bottom_frame*/, num_frames);)
2675
2676 frame f = new_stack_frame<ContinuationHelper::StubFrame>(hf, caller, false);
2677 intptr_t* stack_frame_top = f.sp();
2678 intptr_t* heap_frame_top = hf.sp();
2679 int fsize = ContinuationHelper::StubFrame::size(hf);
2680
2681 copy_from_chunk(heap_frame_top - frame::metadata_words, stack_frame_top - frame::metadata_words,
2682 fsize + frame::metadata_words);
2683
2684 patch(f, caller, false /*is_bottom_frame*/);
2685
2686 // can only fix caller once this frame is thawed (due to callee saved regs)
2687 RegisterMap map(nullptr,
2688 RegisterMap::UpdateMap::include,
2689 RegisterMap::ProcessFrames::skip,
2690 RegisterMap::WalkContinuation::skip);
2691 map.set_include_argument_oops(false);
2692 f.oop_map()->update_register_map(&f, &map);
2693 ContinuationHelper::update_register_map_with_callee(caller, &map);
2694 _cont.tail()->fix_thawed_frame(caller, &map);
2695
2696 DEBUG_ONLY(after_thaw_java_frame(f, false /*is_bottom_frame*/);)
2697 caller = f;
2698 }
2699
2700 void ThawBase::recurse_thaw_native_frame(const frame& hf, frame& caller, int num_frames) {
2701 assert(hf.is_native_frame(), "");
2702 assert(_preempted_case && hf.cb()->as_nmethod()->method()->is_object_wait0(), "");
2703
2704 if (UNLIKELY(seen_by_gc())) { // recurse_thaw_stub_frame already invoked our barriers with a full regmap
2705 _cont.tail()->do_barriers<stackChunkOopDesc::BarrierType::Store>(_stream, SmallRegisterMap::instance());
2706 }
2707
2708 const bool is_bottom_frame = recurse_thaw_java_frame<ContinuationHelper::NativeFrame>(caller, num_frames);
2709 assert(!is_bottom_frame, "");
2710
2711 DEBUG_ONLY(before_thaw_java_frame(hf, caller, is_bottom_frame, num_frames);)
2712
2713 assert(caller.sp() == caller.unextended_sp(), "");
2714
2715 if (caller.is_interpreted_frame()) {
2716 _align_size += frame::align_wiggle; // we add one whether or not we've aligned because we add it in recurse_freeze_native_frame
2717 }
2718
2719 // new_stack_frame must construct the resulting frame using hf.pc() rather than hf.raw_pc() because the frame is not
2720 // yet laid out in the stack, and so the original_pc is not stored in it.
2721 // As a result, f.is_deoptimized_frame() is always false and we must test hf to know if the frame is deoptimized.
2722 frame f = new_stack_frame<ContinuationHelper::NativeFrame>(hf, caller, false /* bottom */);
2723 intptr_t* const stack_frame_top = f.sp();
2724 intptr_t* const heap_frame_top = hf.unextended_sp();
2725
2726 int fsize = ContinuationHelper::NativeFrame::size(hf);
2727 assert(fsize <= (int)(caller.unextended_sp() - f.unextended_sp()), "");
2728
2729 intptr_t* from = heap_frame_top - frame::metadata_words_at_bottom;
2730 intptr_t* to = stack_frame_top - frame::metadata_words_at_bottom;
2731 int sz = fsize + frame::metadata_words_at_bottom;
2732
2733 copy_from_chunk(from, to, sz); // copying good oops because we invoked barriers above
2734
2735 patch(f, caller, false /* bottom */);
2736
2737 // f.is_deoptimized_frame() is always false and we must test hf.is_deoptimized_frame() (see comment above)
2738 assert(!f.is_deoptimized_frame(), "");
2739 assert(!hf.is_deoptimized_frame(), "");
2740 assert(!f.cb()->as_nmethod()->is_marked_for_deoptimization(), "");
2741
2742 // can only fix caller once this frame is thawed (due to callee saved regs); this happens on the stack
2743 _cont.tail()->fix_thawed_frame(caller, SmallRegisterMap::instance());
2744
2745 DEBUG_ONLY(after_thaw_java_frame(f, false /* bottom */);)
2746 caller = f;
2747 }
2748
2749 void ThawBase::finish_thaw(frame& f) {
2750 stackChunkOop chunk = _cont.tail();
2751
2752 if (chunk->is_empty()) {
2753 // Only remove chunk from list if it can't be reused for another freeze
2754 if (seen_by_gc()) {
2755 _cont.set_tail(chunk->parent());
2756 } else {
2757 chunk->set_has_mixed_frames(false);
2758 }
2759 chunk->set_max_thawing_size(0);
2760 } else {
2761 chunk->set_max_thawing_size(chunk->max_thawing_size() - _align_size);
2762 }
2763 assert(chunk->is_empty() == (chunk->max_thawing_size() == 0), "");
2764
2765 if (!is_aligned(f.sp(), frame::frame_alignment)) {
2766 assert(f.is_interpreted_frame(), "");
2767 f.set_sp(align_down(f.sp(), frame::frame_alignment));
2768 }
2769 push_return_frame(f);
2770 chunk->fix_thawed_frame(f, SmallRegisterMap::instance()); // can only fix caller after push_return_frame (due to callee saved regs)
2771
2772 assert(_cont.is_empty() == _cont.last_frame().is_empty(), "");
2773
2774 log_develop_trace(continuations)("thawed %d frames", _frames);
2775
2776 LogTarget(Trace, continuations) lt;
2777 if (lt.develop_is_enabled()) {
2778 LogStream ls(lt);
2779 ls.print_cr("top hframe after (thaw):");
2780 _cont.last_frame().print_value_on(&ls);
2781 }
2782 }
2783
2784 void ThawBase::push_return_frame(frame& f) { // see generate_cont_thaw
2785 assert(!f.is_compiled_frame() || f.is_deoptimized_frame() == f.cb()->as_nmethod()->is_deopt_pc(f.raw_pc()), "");
2786 assert(!f.is_compiled_frame() || f.is_deoptimized_frame() == (f.pc() != f.raw_pc()), "");
2787
2788 LogTarget(Trace, continuations) lt;
2789 if (lt.develop_is_enabled()) {
2790 LogStream ls(lt);
2791 ls.print_cr("push_return_frame");
2792 f.print_value_on(&ls);
2793 }
2794
2795 assert(f.sp() - frame::metadata_words_at_bottom >= _top_stack_address, "overwrote past thawing space"
2796 " to: " INTPTR_FORMAT " top_address: " INTPTR_FORMAT, p2i(f.sp() - frame::metadata_words), p2i(_top_stack_address));
2797 ContinuationHelper::Frame::patch_pc(f, f.raw_pc()); // in case we want to deopt the frame in a full transition, this is checked.
2798 ContinuationHelper::push_pd(f);
2799
2800 assert(ContinuationHelper::Frame::assert_frame_laid_out(f), "");
2801 }
2802
2803 // returns new top sp
2804 // called after preparations (stack overflow check and making room)
2805 template<typename ConfigT>
2806 static inline intptr_t* thaw_internal(JavaThread* thread, const Continuation::thaw_kind kind) {
2807 assert(thread == JavaThread::current(), "Must be current thread");
2808
2809 CONT_JFR_ONLY(EventContinuationThaw event;)
2810
2811 log_develop_trace(continuations)("~~~~ thaw kind: %d sp: " INTPTR_FORMAT, kind, p2i(thread->last_continuation()->entry_sp()));
2812
2813 ContinuationEntry* entry = thread->last_continuation();
2814 assert(entry != nullptr, "");
2815 oop oopCont = entry->cont_oop(thread);
2816
2817 assert(!jdk_internal_vm_Continuation::done(oopCont), "");
2818 assert(oopCont == get_continuation(thread), "");
2819 verify_continuation(oopCont);
2820
2821 assert(entry->is_virtual_thread() == (entry->scope(thread) == java_lang_VirtualThread::vthread_scope()), "");
2822
2823 ContinuationWrapper cont(thread, oopCont);
2824 log_develop_debug(continuations)("THAW #" INTPTR_FORMAT " " INTPTR_FORMAT, cont.hash(), p2i((oopDesc*)oopCont));
2825
2826 #ifdef ASSERT
2827 set_anchor_to_entry(thread, cont.entry());
2828 log_frames(thread);
2829 clear_anchor(thread);
2830 #endif
2831
2832 DEBUG_ONLY(bool preempted = cont.tail()->preempted();)
2833 Thaw<ConfigT> thw(thread, cont);
2834 intptr_t* const sp = thw.thaw(kind);
2835 assert(is_aligned(sp, frame::frame_alignment), "");
2836 DEBUG_ONLY(log_frames_after_thaw(thread, cont, sp, preempted);)
2837
2838 CONT_JFR_ONLY(thw.jfr_info().post_jfr_event(&event, cont.continuation(), thread);)
2839
2840 verify_continuation(cont.continuation());
2841 log_develop_debug(continuations)("=== End of thaw #" INTPTR_FORMAT, cont.hash());
2842
2843 return sp;
2844 }
2845
2846 #ifdef ASSERT
2847 static void do_deopt_after_thaw(JavaThread* thread) {
2848 int i = 0;
2849 StackFrameStream fst(thread, true, false);
2850 fst.register_map()->set_include_argument_oops(false);
2851 ContinuationHelper::update_register_map_with_callee(*fst.current(), fst.register_map());
2852 for (; !fst.is_done(); fst.next()) {
2853 if (fst.current()->cb()->is_nmethod()) {
2854 nmethod* nm = fst.current()->cb()->as_nmethod();
2855 if (!nm->method()->is_continuation_native_intrinsic()) {
2856 nm->make_deoptimized();
2857 }
2858 }
2859 }
2860 }
2861
2862 class ThawVerifyOopsClosure: public OopClosure {
2863 intptr_t* _p;
2864 outputStream* _st;
2865 bool is_good_oop(oop o) {
2866 return dbg_is_safe(o, -1) && dbg_is_safe(o->klass(), -1) && oopDesc::is_oop(o) && o->klass()->is_klass();
2867 }
2868 public:
2869 ThawVerifyOopsClosure(outputStream* st) : _p(nullptr), _st(st) {}
2870 intptr_t* p() { return _p; }
2871 void reset() { _p = nullptr; }
2872
2873 virtual void do_oop(oop* p) {
2874 oop o = *p;
2875 if (o == nullptr || is_good_oop(o)) {
2876 return;
2877 }
2878 _p = (intptr_t*)p;
2879 _st->print_cr("*** non-oop " PTR_FORMAT " found at " PTR_FORMAT, p2i(*p), p2i(p));
2880 }
2881 virtual void do_oop(narrowOop* p) {
2882 oop o = RawAccess<>::oop_load(p);
2883 if (o == nullptr || is_good_oop(o)) {
2884 return;
2885 }
2886 _p = (intptr_t*)p;
2887 _st->print_cr("*** (narrow) non-oop %x found at " PTR_FORMAT, (int)(*p), p2i(p));
2888 }
2889 };
2890
2891 static bool do_verify_after_thaw(JavaThread* thread, stackChunkOop chunk, outputStream* st) {
2892 assert(thread->has_last_Java_frame(), "");
2893
2894 ResourceMark rm;
2895 ThawVerifyOopsClosure cl(st);
2896 NMethodToOopClosure cf(&cl, false);
2897
2898 StackFrameStream fst(thread, true, false);
2899 fst.register_map()->set_include_argument_oops(false);
2900 ContinuationHelper::update_register_map_with_callee(*fst.current(), fst.register_map());
2901 for (; !fst.is_done() && !Continuation::is_continuation_enterSpecial(*fst.current()); fst.next()) {
2902 if (fst.current()->cb()->is_nmethod() && fst.current()->cb()->as_nmethod()->is_marked_for_deoptimization()) {
2903 st->print_cr(">>> do_verify_after_thaw deopt");
2904 fst.current()->deoptimize(nullptr);
2905 fst.current()->print_on(st);
2906 }
2907
2908 fst.current()->oops_do(&cl, &cf, fst.register_map());
2909 if (cl.p() != nullptr) {
2910 frame fr = *fst.current();
2911 st->print_cr("Failed for frame barriers: %d",chunk->requires_barriers());
2912 fr.print_on(st);
2913 if (!fr.is_interpreted_frame()) {
2914 st->print_cr("size: %d argsize: %d",
2915 ContinuationHelper::NonInterpretedUnknownFrame::size(fr),
2916 ContinuationHelper::NonInterpretedUnknownFrame::stack_argsize(fr));
2917 }
2918 VMReg reg = fst.register_map()->find_register_spilled_here(cl.p(), fst.current()->sp());
2919 if (reg != nullptr) {
2920 st->print_cr("Reg %s %d", reg->name(), reg->is_stack() ? (int)reg->reg2stack() : -99);
2921 }
2922 cl.reset();
2923 DEBUG_ONLY(thread->print_frame_layout();)
2924 if (chunk != nullptr) {
2925 chunk->print_on(true, st);
2926 }
2927 return false;
2928 }
2929 }
2930 return true;
2931 }
2932
2933 static void log_frames(JavaThread* thread) {
2934 const static int show_entry_callers = 3;
2935 LogTarget(Trace, continuations) lt;
2936 if (!lt.develop_is_enabled()) {
2937 return;
2938 }
2939 LogStream ls(lt);
2940
2941 ls.print_cr("------- frames --------- for thread " INTPTR_FORMAT, p2i(thread));
2942 if (!thread->has_last_Java_frame()) {
2943 ls.print_cr("NO ANCHOR!");
2944 }
2945
2946 RegisterMap map(thread,
2947 RegisterMap::UpdateMap::include,
2948 RegisterMap::ProcessFrames::include,
2949 RegisterMap::WalkContinuation::skip);
2950 map.set_include_argument_oops(false);
2951
2952 if (false) {
2953 for (frame f = thread->last_frame(); !f.is_entry_frame(); f = f.sender(&map)) {
2954 f.print_on(&ls);
2955 }
2956 } else {
2957 map.set_skip_missing(true);
2958 ResetNoHandleMark rnhm;
2959 ResourceMark rm;
2960 HandleMark hm(Thread::current());
2961 FrameValues values;
2962
2963 int i = 0;
2964 int post_entry = -1;
2965 for (frame f = thread->last_frame(); !f.is_first_frame(); f = f.sender(&map), i++) {
2966 f.describe(values, i, &map, i == 0);
2967 if (post_entry >= 0 || Continuation::is_continuation_enterSpecial(f))
2968 post_entry++;
2969 if (post_entry >= show_entry_callers)
2970 break;
2971 }
2972 values.print_on(thread, &ls);
2973 }
2974
2975 ls.print_cr("======= end frames =========");
2976 }
2977
2978 static void log_frames_after_thaw(JavaThread* thread, ContinuationWrapper& cont, intptr_t* sp, bool preempted) {
2979 intptr_t* sp0 = sp;
2980 address pc0 = *(address*)(sp - frame::sender_sp_ret_address_offset());
2981
2982 if (preempted && sp0 == cont.entrySP()) {
2983 // Still preempted (monitor not acquired) so no frames were thawed.
2984 assert(cont.tail()->preempted(), "");
2985 set_anchor(thread, cont.entrySP(), cont.entryPC());
2986 } else {
2987 set_anchor(thread, sp0);
2988 }
2989
2990 log_frames(thread);
2991 if (LoomVerifyAfterThaw) {
2992 assert(do_verify_after_thaw(thread, cont.tail(), tty), "");
2993 }
2994 assert(ContinuationEntry::assert_entry_frame_laid_out(thread), "");
2995 clear_anchor(thread);
2996
2997 LogTarget(Trace, continuations) lt;
2998 if (lt.develop_is_enabled()) {
2999 LogStream ls(lt);
3000 ls.print_cr("Jumping to frame (thaw):");
3001 frame(sp).print_value_on(&ls);
3002 }
3003 }
3004 #endif // ASSERT
3005
3006 #include CPU_HEADER_INLINE(continuationFreezeThaw)
3007
3008 #ifdef ASSERT
3009 static void print_frame_layout(const frame& f, bool callee_complete, outputStream* st) {
3010 ResourceMark rm;
3011 FrameValues values;
3012 assert(f.get_cb() != nullptr, "");
3013 RegisterMap map(f.is_heap_frame() ?
3014 nullptr :
3015 JavaThread::current(),
3016 RegisterMap::UpdateMap::include,
3017 RegisterMap::ProcessFrames::skip,
3018 RegisterMap::WalkContinuation::skip);
3019 map.set_include_argument_oops(false);
3020 map.set_skip_missing(true);
3021 if (callee_complete) {
3022 frame::update_map_with_saved_link(&map, ContinuationHelper::Frame::callee_link_address(f));
3023 }
3024 const_cast<frame&>(f).describe(values, 0, &map, true);
3025 values.print_on(static_cast<JavaThread*>(nullptr), st);
3026 }
3027 #endif
3028
3029 static address thaw_entry = nullptr;
3030 static address freeze_entry = nullptr;
3031 static address freeze_preempt_entry = nullptr;
3032
3033 address Continuation::thaw_entry() {
3034 return ::thaw_entry;
3035 }
3036
3037 address Continuation::freeze_entry() {
3038 return ::freeze_entry;
3039 }
3040
3041 address Continuation::freeze_preempt_entry() {
3042 return ::freeze_preempt_entry;
3043 }
3044
3045 class ConfigResolve {
3046 public:
3047 static void resolve() { resolve_compressed(); }
3048
3049 static void resolve_compressed() {
3050 UseCompressedOops ? resolve_gc<true>()
3051 : resolve_gc<false>();
3052 }
3053
3054 private:
3055 template <bool use_compressed>
3056 static void resolve_gc() {
3057 BarrierSet* bs = BarrierSet::barrier_set();
3058 assert(bs != nullptr, "freeze/thaw invoked before BarrierSet is set");
3059 switch (bs->kind()) {
3060 #define BARRIER_SET_RESOLVE_BARRIER_CLOSURE(bs_name) \
3061 case BarrierSet::bs_name: { \
3062 resolve<use_compressed, typename BarrierSet::GetType<BarrierSet::bs_name>::type>(); \
3063 } \
3064 break;
3065 FOR_EACH_CONCRETE_BARRIER_SET_DO(BARRIER_SET_RESOLVE_BARRIER_CLOSURE)
3066 #undef BARRIER_SET_RESOLVE_BARRIER_CLOSURE
3067
3068 default:
3069 fatal("BarrierSet resolving not implemented");
3070 };
3071 }
3072
3073 template <bool use_compressed, typename BarrierSetT>
3074 static void resolve() {
3075 typedef Config<use_compressed ? oop_kind::NARROW : oop_kind::WIDE, BarrierSetT> SelectedConfigT;
3076
3077 freeze_entry = (address)freeze<SelectedConfigT>;
3078 freeze_preempt_entry = (address)SelectedConfigT::freeze_preempt;
3079
3080 // If we wanted, we could templatize by kind and have three different thaw entries
3081 thaw_entry = (address)thaw<SelectedConfigT>;
3082 }
3083 };
3084
3085 void Continuation::init() {
3086 ConfigResolve::resolve();
3087 }