1 /*
2 * Copyright Amazon.com Inc. or its affiliates. All Rights Reserved.
3 * Copyright (c) 2025, Oracle and/or its affiliates. All rights reserved.
4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
5 *
6 * This code is free software; you can redistribute it and/or modify it
7 * under the terms of the GNU General Public License version 2 only, as
8 * published by the Free Software Foundation.
9 *
10 * This code is distributed in the hope that it will be useful, but WITHOUT
11 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
12 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
13 * version 2 for more details (a copy is included in the LICENSE file that
14 * accompanied this code).
15 *
16 * You should have received a copy of the GNU General Public License version
17 * 2 along with this work; if not, write to the Free Software Foundation,
18 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
19 *
20 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
21 * or visit www.oracle.com if you need additional information or have any
22 * questions.
23 *
24 */
25
26 #ifndef SHARE_GC_SHENANDOAH_SHENANDOAHSCANREMEMBEREDINLINE_HPP
27 #define SHARE_GC_SHENANDOAH_SHENANDOAHSCANREMEMBEREDINLINE_HPP
28
29 #include "gc/shenandoah/shenandoahScanRemembered.hpp"
30
31 #include "gc/shared/collectorCounters.hpp"
32 #include "gc/shenandoah/mode/shenandoahMode.hpp"
33 #include "gc/shenandoah/shenandoahCardStats.hpp"
34 #include "gc/shenandoah/shenandoahCardTable.hpp"
35 #include "gc/shenandoah/shenandoahHeap.hpp"
36 #include "gc/shenandoah/shenandoahHeapRegion.hpp"
37 #include "gc/shenandoah/shenandoahOldGeneration.hpp"
38 #include "logging/log.hpp"
39 #include "memory/iterator.hpp"
40 #include "oops/objArrayOop.hpp"
41 #include "oops/oop.hpp"
42
43 // Process all objects starting within count clusters beginning with first_cluster and for which the start address is
44 // less than end_of_range. For any non-array object whose header lies on a dirty card, scan the entire object,
45 // even if its end reaches beyond end_of_range. Object arrays, on the other hand, are precisely dirtied and
46 // only the portions of the array on dirty cards need to be scanned.
47 //
48 // Do not CANCEL within process_clusters. It is assumed that if a worker thread accepts responsibility for processing
49 // a chunk of work, it will finish the work it starts. Otherwise, the chunk of work will be lost in the transition to
50 // degenerated execution, leading to dangling references.
51 template <typename ClosureType>
52 void ShenandoahScanRemembered::process_clusters(size_t first_cluster, size_t count, HeapWord* end_of_range,
53 ClosureType* cl, bool use_write_table, uint worker_id) {
54
55 assert(ShenandoahHeap::heap()->old_generation()->is_parsable(), "Old generation regions must be parsable for remembered set scan");
56 // If old-gen evacuation is active, then MarkingContext for old-gen heap regions is valid. We use the MarkingContext
57 // bits to determine which objects within a DIRTY card need to be scanned. This is necessary because old-gen heap
58 // regions that are in the candidate collection set have not been coalesced and filled. Thus, these heap regions
59 // may contain zombie objects. Zombie objects are known to be dead, but have not yet been "collected". Scanning
60 // zombie objects is unsafe because the Klass pointer is not reliable, objects referenced from a zombie may have been
61 // collected (if dead), or relocated (if live), or if dead but not yet collected, we don't want to "revive" them
62 // by marking them (when marking) or evacuating them (when updating references).
63
64 // start and end addresses of range of objects to be scanned, clipped to end_of_range
65 const size_t start_card_index = first_cluster * ShenandoahCardCluster::CardsPerCluster;
66 const HeapWord* start_addr = _rs->addr_for_card_index(start_card_index);
67 // clip at end_of_range (exclusive)
68 HeapWord* end_addr = MIN2(end_of_range, (HeapWord*)start_addr + (count * ShenandoahCardCluster::CardsPerCluster
69 * CardTable::card_size_in_words()));
70 assert(start_addr < end_addr, "Empty region?");
71
72 const size_t whole_cards = (end_addr - start_addr + CardTable::card_size_in_words() - 1)/CardTable::card_size_in_words();
73 const size_t end_card_index = start_card_index + whole_cards - 1;
74 log_debug(gc, remset)("Worker %u: cluster = %zu count = %zu eor = " INTPTR_FORMAT
75 " start_addr = " INTPTR_FORMAT " end_addr = " INTPTR_FORMAT " cards = %zu",
76 worker_id, first_cluster, count, p2i(end_of_range), p2i(start_addr), p2i(end_addr), whole_cards);
77
78 // use_write_table states whether we are using the card table that is being
79 // marked by the mutators. If false, we are using a snapshot of the card table
80 // that is not subject to modifications. Even when this arg is true, and
81 // the card table is being actively marked, SATB marking ensures that we need not
82 // worry about cards marked after the processing here has passed them.
83 const CardValue* const ctbm = _rs->get_card_table_byte_map(use_write_table);
84
85 // If old gen evacuation is active, ctx will hold the completed marking of
86 // old generation objects. We'll only scan objects that are marked live by
87 // the old generation marking. These include objects allocated since the
88 // start of old generation marking (being those above TAMS).
89 const ShenandoahHeap* heap = ShenandoahHeap::heap();
90 const ShenandoahMarkingContext* ctx = heap->old_generation()->is_mark_complete() ?
91 heap->marking_context() : nullptr;
92
93 // The region we will scan is the half-open interval [start_addr, end_addr),
94 // and lies entirely within a single region.
95 const ShenandoahHeapRegion* region = ShenandoahHeap::heap()->heap_region_containing(start_addr);
96 assert(region->contains(end_addr - 1), "Slice shouldn't cross regions");
97
98 // This code may have implicit assumptions of examining only old gen regions.
99 assert(region->is_old(), "We only expect to be processing old regions");
100 assert(!region->is_humongous(), "Humongous regions can be processed more efficiently;"
101 "see process_humongous_clusters()");
102 // tams and ctx below are for old generation marking. As such, young gen roots must
103 // consider everything above tams, since it doesn't represent a TAMS for young gen's
104 // SATB marking.
105 const HeapWord* tams = (ctx == nullptr ? region->bottom() : ctx->top_at_mark_start(region));
106
107 NOT_PRODUCT(ShenandoahCardStats stats(whole_cards, card_stats(worker_id));)
108
109 // In the case of imprecise marking, we remember the lowest address
110 // scanned in a range of dirty cards, as we work our way left from the
111 // highest end_addr. This serves as another upper bound on the address we will
112 // scan as we move left over each contiguous range of dirty cards.
113 HeapWord* upper_bound = nullptr;
114
115 // Starting at the right end of the address range, walk backwards accumulating
116 // a maximal dirty range of cards, then process those cards.
117 ssize_t cur_index = (ssize_t) end_card_index;
118 assert(cur_index >= 0, "Overflow");
119 assert(((ssize_t)start_card_index) >= 0, "Overflow");
120 while (cur_index >= (ssize_t)start_card_index) {
121
122 // We'll continue the search starting with the card for the upper bound
123 // address identified by the last dirty range that we processed, if any,
124 // skipping any cards at higher addresses.
125 if (upper_bound != nullptr) {
126 ssize_t right_index = _rs->card_index_for_addr(upper_bound);
127 assert(right_index >= 0, "Overflow");
128 cur_index = MIN2(cur_index, right_index);
129 assert(upper_bound < end_addr, "Program logic");
130 end_addr = upper_bound; // lower end_addr
131 upper_bound = nullptr; // and clear upper_bound
132 if (end_addr <= start_addr) {
133 assert(right_index <= (ssize_t)start_card_index, "Program logic");
134 // We are done with our cluster
135 return;
136 }
137 }
138
139 if (ctbm[cur_index] == CardTable::dirty_card_val()) {
140 // ==== BEGIN DIRTY card range processing ====
141
142 const size_t dirty_r = cur_index; // record right end of dirty range (inclusive)
143 while (--cur_index >= (ssize_t)start_card_index && ctbm[cur_index] == CardTable::dirty_card_val()) {
144 // walk back over contiguous dirty cards to find left end of dirty range (inclusive)
145 }
146 // [dirty_l, dirty_r] is a "maximal" closed interval range of dirty card indices:
147 // it may not be maximal if we are using the write_table, because of concurrent
148 // mutations dirtying the card-table. It may also not be maximal if an upper bound
149 // was established by the scan of the previous chunk.
150 const size_t dirty_l = cur_index + 1; // record left end of dirty range (inclusive)
151 // Check that we identified a boundary on our left
152 assert(ctbm[dirty_l] == CardTable::dirty_card_val(), "First card in range should be dirty");
153 assert(dirty_l == start_card_index || use_write_table
154 || ctbm[dirty_l - 1] == CardTable::clean_card_val(),
155 "Interval isn't maximal on the left");
156 assert(dirty_r >= dirty_l, "Error");
157 assert(ctbm[dirty_r] == CardTable::dirty_card_val(), "Last card in range should be dirty");
158 // Record alternations, dirty run length, and dirty card count
159 NOT_PRODUCT(stats.record_dirty_run(dirty_r - dirty_l + 1);)
160
161 // Find first object that starts this range:
162 // [left, right) is a maximal right-open interval of dirty cards
163 HeapWord* left = _rs->addr_for_card_index(dirty_l); // inclusive
164 HeapWord* right = _rs->addr_for_card_index(dirty_r + 1); // exclusive
165 // Clip right to end_addr established above (still exclusive)
166 right = MIN2(right, end_addr);
167 assert(right <= region->top() && end_addr <= region->top(), "Busted bounds");
168 const MemRegion mr(left, right);
169
170 // NOTE: We'll not call block_start() repeatedly
171 // on a very large object if its head card is dirty. If not,
172 // (i.e. the head card is clean) we'll call it each time we
173 // process a new dirty range on the object. This is always
174 // the case for large object arrays, which are typically more
175 // common.
176 HeapWord* p = _scc->block_start(dirty_l);
177 oop obj = cast_to_oop(p);
178
179 // PREFIX: The object that straddles into this range of dirty cards
180 // from the left may be subject to special treatment unless
181 // it is an object array.
182 if (p < left && !obj->is_refArray()) {
183 // The mutator (both compiler and interpreter, but not JNI?)
184 // typically dirty imprecisely (i.e. only the head of an object),
185 // but GC closures typically dirty the object precisely. (It would
186 // be nice to have everything be precise for maximum efficiency.)
187 //
188 // To handle this, we check the head card of the object here and,
189 // if dirty, (arrange to) scan the object in its entirety. If we
190 // find the head card clean, we'll scan only the portion of the
191 // object lying in the dirty card range below, assuming this was
192 // the result of precise marking by GC closures.
193
194 // index of the "head card" for p
195 const size_t hc_index = _rs->card_index_for_addr(p);
196 if (ctbm[hc_index] == CardTable::dirty_card_val()) {
197 // Scan or skip the object, depending on location of its
198 // head card, and remember that we'll have processed all
199 // the objects back up to p, which is thus an upper bound
200 // for the next iteration of a dirty card loop.
201 upper_bound = p; // remember upper bound for next chunk
202 if (p < start_addr) {
203 // if object starts in a previous slice, it'll be handled
204 // in its entirety by the thread processing that slice; we can
205 // skip over it and avoid an unnecessary extra scan.
206 assert(obj == cast_to_oop(p), "Inconsistency detected");
207 p += obj->size();
208 } else {
209 // the object starts in our slice, we scan it in its entirety
210 assert(obj == cast_to_oop(p), "Inconsistency detected");
211 if (ctx == nullptr || ctx->is_marked(obj)) {
212 // Scan the object in its entirety
213 p += obj->oop_iterate_size(cl);
214 } else {
215 assert(p < tams, "Error 1 in ctx/marking/tams logic");
216 // Skip over any intermediate dead objects
217 p = ctx->get_next_marked_addr(p, tams);
218 assert(p <= tams, "Error 2 in ctx/marking/tams logic");
219 }
220 }
221 assert(p > left, "Should have processed into interior of dirty range");
222 }
223 }
224
225 size_t i = 0;
226 HeapWord* last_p = nullptr;
227
228 // BODY: Deal with (other) objects in this dirty card range
229 while (p < right) {
230 obj = cast_to_oop(p);
231 // walk right scanning eligible objects
232 if (ctx == nullptr || ctx->is_marked(obj)) {
233 // we need to remember the last object ptr we scanned, in case we need to
234 // complete a partial suffix scan after mr, see below
235 last_p = p;
236 // apply the closure to the oops in the portion of
237 // the object within mr.
238 p += obj->oop_iterate_size(cl, mr);
239 NOT_PRODUCT(i++);
240 } else {
241 // forget the last object pointer we remembered
242 last_p = nullptr;
243 assert(p < tams, "Tams and above are implicitly marked in ctx");
244 // object under tams isn't marked: skip to next live object
245 p = ctx->get_next_marked_addr(p, tams);
246 assert(p <= tams, "Error 3 in ctx/marking/tams logic");
247 }
248 }
249
250 // SUFFIX: Fix up a possible incomplete scan at right end of window
251 // by scanning the portion of a non-objArray that wasn't done.
252 if (p > right && last_p != nullptr) {
253 assert(last_p < right, "Error");
254 // check if last_p suffix needs scanning
255 const oop last_obj = cast_to_oop(last_p);
256 if (!last_obj->is_refArray()) {
257 // scan the remaining suffix of the object
258 const MemRegion last_mr(right, p);
259 assert(p == last_p + last_obj->size(), "Would miss portion of last_obj");
260 last_obj->oop_iterate(cl, last_mr);
261 log_develop_debug(gc, remset)("Fixed up non-objArray suffix scan in [" INTPTR_FORMAT ", " INTPTR_FORMAT ")",
262 p2i(last_mr.start()), p2i(last_mr.end()));
263 } else {
264 log_develop_debug(gc, remset)("Skipped suffix scan of objArray in [" INTPTR_FORMAT ", " INTPTR_FORMAT ")",
265 p2i(right), p2i(p));
266 }
267 }
268 NOT_PRODUCT(stats.record_scan_obj_cnt(i);)
269
270 // ==== END DIRTY card range processing ====
271 } else {
272 // ==== BEGIN CLEAN card range processing ====
273
274 // If we are using the write table (during update refs, e.g.), a mutator may dirty
275 // a card at any time. This is fine for the algorithm below because it is only
276 // counting contiguous runs of clean cards (and only for non-product builds).
277 assert(use_write_table || ctbm[cur_index] == CardTable::clean_card_val(), "Error");
278
279 // walk back over contiguous clean cards
280 size_t i = 0;
281 while (--cur_index >= (ssize_t)start_card_index && ctbm[cur_index] == CardTable::clean_card_val()) {
282 NOT_PRODUCT(i++);
283 }
284 // Record alternations, clean run length, and clean card count
285 NOT_PRODUCT(stats.record_clean_run(i);)
286
287 // ==== END CLEAN card range processing ====
288 }
289 }
290 }
291
292 // Given that this range of clusters is known to span a humongous object spanned by region r, scan the
293 // portion of the humongous object that corresponds to the specified range.
294 template <typename ClosureType>
295 inline void
296 ShenandoahScanRemembered::process_humongous_clusters(ShenandoahHeapRegion* r, size_t first_cluster, size_t count,
297 HeapWord *end_of_range, ClosureType *cl, bool use_write_table) {
298 ShenandoahHeapRegion* start_region = r->humongous_start_region();
299 HeapWord* p = start_region->bottom();
300 oop obj = cast_to_oop(p);
301 assert(r->is_humongous(), "Only process humongous regions here");
302 assert(start_region->is_humongous_start(), "Should be start of humongous region");
303 assert(p + obj->size() >= end_of_range, "Humongous object ends before range ends");
304
305 size_t first_card_index = first_cluster * ShenandoahCardCluster::CardsPerCluster;
306 HeapWord* first_cluster_addr = _rs->addr_for_card_index(first_card_index);
307 size_t spanned_words = count * ShenandoahCardCluster::CardsPerCluster * CardTable::card_size_in_words();
308 start_region->oop_iterate_humongous_slice_dirty(cl, first_cluster_addr, spanned_words, use_write_table);
309 }
310
311
312 // This method takes a region & determines the end of the region that the worker can scan.
313 template <typename ClosureType>
314 inline void
315 ShenandoahScanRemembered::process_region_slice(ShenandoahHeapRegion *region, size_t start_offset, size_t clusters,
316 HeapWord *end_of_range, ClosureType *cl, bool use_write_table,
317 uint worker_id) {
318
319 // This is called only for young gen collection, when we scan old gen regions
320 assert(region->is_old(), "Expecting an old region");
321 HeapWord *start_of_range = region->bottom() + start_offset;
322 size_t start_cluster_no = cluster_for_addr(start_of_range);
323 assert(addr_for_cluster(start_cluster_no) == start_of_range, "process_region_slice range must align on cluster boundary");
324
325 // region->end() represents the end of memory spanned by this region, but not all of this
326 // memory is eligible to be scanned because some of this memory has not yet been allocated.
327 //
328 // region->top() represents the end of allocated memory within this region. Any addresses
329 // beyond region->top() should not be scanned as that memory does not hold valid objects.
330
331 if (use_write_table) {
332 // This is update-refs servicing.
333 if (end_of_range > region->get_update_watermark()) {
334 end_of_range = region->get_update_watermark();
335 }
336 } else {
337 // This is concurrent mark servicing. Note that TAMS for this region is TAMS at start of old-gen
338 // collection. Here, we need to scan up to TAMS for most recently initiated young-gen collection.
339 // Since all LABs are retired at init mark, and since replacement LABs are allocated lazily, and since no
340 // promotions occur until evacuation phase, TAMS for most recent young-gen is same as top().
341 if (end_of_range > region->top()) {
342 end_of_range = region->top();
343 }
344 }
345
346 log_debug(gc)("Remembered set scan processing Region %zu, from " PTR_FORMAT " to " PTR_FORMAT ", using %s table",
347 region->index(), p2i(start_of_range), p2i(end_of_range),
348 use_write_table? "read/write (updating)": "read (marking)");
349
350 // Note that end_of_range may point to the middle of a cluster because we limit scanning to
351 // region->top() or region->get_update_watermark(). We avoid processing past end_of_range.
352 // Objects that start between start_of_range and end_of_range, including humongous objects, will
353 // be fully processed by process_clusters. In no case should we need to scan past end_of_range.
354 if (start_of_range < end_of_range) {
355 if (region->is_humongous()) {
356 ShenandoahHeapRegion* start_region = region->humongous_start_region();
357 process_humongous_clusters(start_region, start_cluster_no, clusters, end_of_range, cl, use_write_table);
358 } else {
359 process_clusters(start_cluster_no, clusters, end_of_range, cl, use_write_table, worker_id);
360 }
361 }
362 }
363
364 inline bool ShenandoahRegionChunkIterator::has_next() const {
365 return _index < _total_chunks;
366 }
367
368 inline bool ShenandoahRegionChunkIterator::next(struct ShenandoahRegionChunk *assignment) {
369 if (_index >= _total_chunks) {
370 return false;
371 }
372 size_t new_index = Atomic::add(&_index, (size_t) 1, memory_order_relaxed);
373 if (new_index > _total_chunks) {
374 // First worker that hits new_index == _total_chunks continues, other
375 // contending workers return false.
376 return false;
377 }
378 // convert to zero-based indexing
379 new_index--;
380 assert(new_index < _total_chunks, "Error");
381
382 // Find the group number for the assigned chunk index
383 size_t group_no;
384 for (group_no = 0; new_index >= _group_entries[group_no]; group_no++)
385 ;
386 assert(group_no < _num_groups, "Cannot have group no greater or equal to _num_groups");
387
388 // All size computations measured in HeapWord
389 size_t region_size_words = ShenandoahHeapRegion::region_size_words();
390 size_t group_region_index = _region_index[group_no];
391 size_t group_region_offset = _group_offset[group_no];
392
393 size_t index_within_group = (group_no == 0)? new_index: new_index - _group_entries[group_no - 1];
394 size_t group_chunk_size = _group_chunk_size[group_no];
395 size_t offset_of_this_chunk = group_region_offset + index_within_group * group_chunk_size;
396 size_t regions_spanned_by_chunk_offset = offset_of_this_chunk / region_size_words;
397 size_t offset_within_region = offset_of_this_chunk % region_size_words;
398
399 size_t region_index = group_region_index + regions_spanned_by_chunk_offset;
400
401 assignment->_r = _heap->get_region(region_index);
402 assignment->_chunk_offset = offset_within_region;
403 assignment->_chunk_size = group_chunk_size;
404 return true;
405 }
406
407 #endif // SHARE_GC_SHENANDOAH_SHENANDOAHSCANREMEMBEREDINLINE_HPP