Mon, 26 Jan 2009 12:47:21 -0800
6786503: Overflow list performance can be improved
Summary: Avoid overflow list walk in CMS & ParNew when it is unnecessary. Fix a couple of correctness issues, including a C-heap leak, in ParNew at the intersection of promotion failure, work queue overflow and object array chunking. Add stress testing option and related assertion checking.
Reviewed-by: jmasa
1 /*
2 * Copyright 2000-2008 Sun Microsystems, Inc. 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 Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
20 * CA 95054 USA or visit www.sun.com if you need additional information or
21 * have any questions.
22 *
23 */
25 // This kind of "BarrierSet" allows a "CollectedHeap" to detect and
26 // enumerate ref fields that have been modified (since the last
27 // enumeration.)
29 // As it currently stands, this barrier is *imprecise*: when a ref field in
30 // an object "o" is modified, the card table entry for the card containing
31 // the head of "o" is dirtied, not necessarily the card containing the
32 // modified field itself. For object arrays, however, the barrier *is*
33 // precise; only the card containing the modified element is dirtied.
34 // Any MemRegionClosures used to scan dirty cards should take these
35 // considerations into account.
37 class Generation;
38 class OopsInGenClosure;
39 class DirtyCardToOopClosure;
41 class CardTableModRefBS: public ModRefBarrierSet {
42 // Some classes get to look at some private stuff.
43 friend class BytecodeInterpreter;
44 friend class VMStructs;
45 friend class CardTableRS;
46 friend class CheckForUnmarkedOops; // Needs access to raw card bytes.
47 #ifndef PRODUCT
48 // For debugging.
49 friend class GuaranteeNotModClosure;
50 #endif
51 protected:
53 enum CardValues {
54 clean_card = -1,
55 dirty_card = 0,
56 precleaned_card = 1,
57 claimed_card = 3,
58 last_card = 4,
59 CT_MR_BS_last_reserved = 10
60 };
62 // dirty and precleaned are equivalent wrt younger_refs_iter.
63 static bool card_is_dirty_wrt_gen_iter(jbyte cv) {
64 return cv == dirty_card || cv == precleaned_card;
65 }
67 // Returns "true" iff the value "cv" will cause the card containing it
68 // to be scanned in the current traversal. May be overridden by
69 // subtypes.
70 virtual bool card_will_be_scanned(jbyte cv) {
71 return CardTableModRefBS::card_is_dirty_wrt_gen_iter(cv);
72 }
74 // Returns "true" iff the value "cv" may have represented a dirty card at
75 // some point.
76 virtual bool card_may_have_been_dirty(jbyte cv) {
77 return card_is_dirty_wrt_gen_iter(cv);
78 }
80 // The declaration order of these const fields is important; see the
81 // constructor before changing.
82 const MemRegion _whole_heap; // the region covered by the card table
83 const size_t _guard_index; // index of very last element in the card
84 // table; it is set to a guard value
85 // (last_card) and should never be modified
86 const size_t _last_valid_index; // index of the last valid element
87 const size_t _page_size; // page size used when mapping _byte_map
88 const size_t _byte_map_size; // in bytes
89 jbyte* _byte_map; // the card marking array
91 int _cur_covered_regions;
92 // The covered regions should be in address order.
93 MemRegion* _covered;
94 // The committed regions correspond one-to-one to the covered regions.
95 // They represent the card-table memory that has been committed to service
96 // the corresponding covered region. It may be that committed region for
97 // one covered region corresponds to a larger region because of page-size
98 // roundings. Thus, a committed region for one covered region may
99 // actually extend onto the card-table space for the next covered region.
100 MemRegion* _committed;
102 // The last card is a guard card, and we commit the page for it so
103 // we can use the card for verification purposes. We make sure we never
104 // uncommit the MemRegion for that page.
105 MemRegion _guard_region;
107 protected:
108 // Initialization utilities; covered_words is the size of the covered region
109 // in, um, words.
110 inline size_t cards_required(size_t covered_words);
111 inline size_t compute_byte_map_size();
113 // Finds and return the index of the region, if any, to which the given
114 // region would be contiguous. If none exists, assign a new region and
115 // returns its index. Requires that no more than the maximum number of
116 // covered regions defined in the constructor are ever in use.
117 int find_covering_region_by_base(HeapWord* base);
119 // Same as above, but finds the region containing the given address
120 // instead of starting at a given base address.
121 int find_covering_region_containing(HeapWord* addr);
123 // Resize one of the regions covered by the remembered set.
124 void resize_covered_region(MemRegion new_region);
126 // Returns the leftmost end of a committed region corresponding to a
127 // covered region before covered region "ind", or else "NULL" if "ind" is
128 // the first covered region.
129 HeapWord* largest_prev_committed_end(int ind) const;
131 // Returns the part of the region mr that doesn't intersect with
132 // any committed region other than self. Used to prevent uncommitting
133 // regions that are also committed by other regions. Also protects
134 // against uncommitting the guard region.
135 MemRegion committed_unique_to_self(int self, MemRegion mr) const;
137 // Mapping from address to card marking array entry
138 jbyte* byte_for(const void* p) const {
139 assert(_whole_heap.contains(p),
140 "out of bounds access to card marking array");
141 jbyte* result = &byte_map_base[uintptr_t(p) >> card_shift];
142 assert(result >= _byte_map && result < _byte_map + _byte_map_size,
143 "out of bounds accessor for card marking array");
144 return result;
145 }
147 // The card table byte one after the card marking array
148 // entry for argument address. Typically used for higher bounds
149 // for loops iterating through the card table.
150 jbyte* byte_after(const void* p) const {
151 return byte_for(p) + 1;
152 }
154 // Iterate over the portion of the card-table which covers the given
155 // region mr in the given space and apply cl to any dirty sub-regions
156 // of mr. cl and dcto_cl must either be the same closure or cl must
157 // wrap dcto_cl. Both are required - neither may be NULL. Also, dcto_cl
158 // may be modified. Note that this function will operate in a parallel
159 // mode if worker threads are available.
160 void non_clean_card_iterate(Space* sp, MemRegion mr,
161 DirtyCardToOopClosure* dcto_cl,
162 MemRegionClosure* cl,
163 bool clear);
165 // Utility function used to implement the other versions below.
166 void non_clean_card_iterate_work(MemRegion mr, MemRegionClosure* cl,
167 bool clear);
169 void par_non_clean_card_iterate_work(Space* sp, MemRegion mr,
170 DirtyCardToOopClosure* dcto_cl,
171 MemRegionClosure* cl,
172 bool clear,
173 int n_threads);
175 // Dirty the bytes corresponding to "mr" (not all of which must be
176 // covered.)
177 void dirty_MemRegion(MemRegion mr);
179 // Clear (to clean_card) the bytes entirely contained within "mr" (not
180 // all of which must be covered.)
181 void clear_MemRegion(MemRegion mr);
183 // *** Support for parallel card scanning.
185 enum SomeConstantsForParallelism {
186 StridesPerThread = 2,
187 CardsPerStrideChunk = 256
188 };
190 // This is an array, one element per covered region of the card table.
191 // Each entry is itself an array, with one element per chunk in the
192 // covered region. Each entry of these arrays is the lowest non-clean
193 // card of the corresponding chunk containing part of an object from the
194 // previous chunk, or else NULL.
195 typedef jbyte* CardPtr;
196 typedef CardPtr* CardArr;
197 CardArr* _lowest_non_clean;
198 size_t* _lowest_non_clean_chunk_size;
199 uintptr_t* _lowest_non_clean_base_chunk_index;
200 int* _last_LNC_resizing_collection;
202 // Initializes "lowest_non_clean" to point to the array for the region
203 // covering "sp", and "lowest_non_clean_base_chunk_index" to the chunk
204 // index of the corresponding to the first element of that array.
205 // Ensures that these arrays are of sufficient size, allocating if necessary.
206 // May be called by several threads concurrently.
207 void get_LNC_array_for_space(Space* sp,
208 jbyte**& lowest_non_clean,
209 uintptr_t& lowest_non_clean_base_chunk_index,
210 size_t& lowest_non_clean_chunk_size);
212 // Returns the number of chunks necessary to cover "mr".
213 size_t chunks_to_cover(MemRegion mr) {
214 return (size_t)(addr_to_chunk_index(mr.last()) -
215 addr_to_chunk_index(mr.start()) + 1);
216 }
218 // Returns the index of the chunk in a stride which
219 // covers the given address.
220 uintptr_t addr_to_chunk_index(const void* addr) {
221 uintptr_t card = (uintptr_t) byte_for(addr);
222 return card / CardsPerStrideChunk;
223 }
225 // Apply cl, which must either itself apply dcto_cl or be dcto_cl,
226 // to the cards in the stride (of n_strides) within the given space.
227 void process_stride(Space* sp,
228 MemRegion used,
229 jint stride, int n_strides,
230 DirtyCardToOopClosure* dcto_cl,
231 MemRegionClosure* cl,
232 bool clear,
233 jbyte** lowest_non_clean,
234 uintptr_t lowest_non_clean_base_chunk_index,
235 size_t lowest_non_clean_chunk_size);
237 // Makes sure that chunk boundaries are handled appropriately, by
238 // adjusting the min_done of dcto_cl, and by using a special card-table
239 // value to indicate how min_done should be set.
240 void process_chunk_boundaries(Space* sp,
241 DirtyCardToOopClosure* dcto_cl,
242 MemRegion chunk_mr,
243 MemRegion used,
244 jbyte** lowest_non_clean,
245 uintptr_t lowest_non_clean_base_chunk_index,
246 size_t lowest_non_clean_chunk_size);
248 public:
249 // Constants
250 enum SomePublicConstants {
251 card_shift = 9,
252 card_size = 1 << card_shift,
253 card_size_in_words = card_size / sizeof(HeapWord)
254 };
256 static int clean_card_val() { return clean_card; }
257 static int dirty_card_val() { return dirty_card; }
258 static int claimed_card_val() { return claimed_card; }
259 static int precleaned_card_val() { return precleaned_card; }
261 // For RTTI simulation.
262 bool is_a(BarrierSet::Name bsn) {
263 return bsn == BarrierSet::CardTableModRef || ModRefBarrierSet::is_a(bsn);
264 }
266 CardTableModRefBS(MemRegion whole_heap, int max_covered_regions);
268 // *** Barrier set functions.
270 bool has_write_ref_pre_barrier() { return false; }
272 inline bool write_ref_needs_barrier(void* field, oop new_val) {
273 // Note that this assumes the perm gen is the highest generation
274 // in the address space
275 return new_val != NULL && !new_val->is_perm();
276 }
278 // Record a reference update. Note that these versions are precise!
279 // The scanning code has to handle the fact that the write barrier may be
280 // either precise or imprecise. We make non-virtual inline variants of
281 // these functions here for performance.
282 protected:
283 void write_ref_field_work(oop obj, size_t offset, oop newVal);
284 void write_ref_field_work(void* field, oop newVal);
285 public:
287 bool has_write_ref_array_opt() { return true; }
288 bool has_write_region_opt() { return true; }
290 inline void inline_write_region(MemRegion mr) {
291 dirty_MemRegion(mr);
292 }
293 protected:
294 void write_region_work(MemRegion mr) {
295 inline_write_region(mr);
296 }
297 public:
299 inline void inline_write_ref_array(MemRegion mr) {
300 dirty_MemRegion(mr);
301 }
302 protected:
303 void write_ref_array_work(MemRegion mr) {
304 inline_write_ref_array(mr);
305 }
306 public:
308 bool is_aligned(HeapWord* addr) {
309 return is_card_aligned(addr);
310 }
312 // *** Card-table-barrier-specific things.
314 inline void inline_write_ref_field_pre(void* field, oop newVal) {}
316 inline void inline_write_ref_field(void* field, oop newVal) {
317 jbyte* byte = byte_for(field);
318 *byte = dirty_card;
319 }
321 // These are used by G1, when it uses the card table as a temporary data
322 // structure for card claiming.
323 bool is_card_dirty(size_t card_index) {
324 return _byte_map[card_index] == dirty_card_val();
325 }
327 void mark_card_dirty(size_t card_index) {
328 _byte_map[card_index] = dirty_card_val();
329 }
331 bool is_card_claimed(size_t card_index) {
332 return _byte_map[card_index] == claimed_card_val();
333 }
335 bool claim_card(size_t card_index);
337 bool is_card_clean(size_t card_index) {
338 return _byte_map[card_index] == clean_card_val();
339 }
341 // Card marking array base (adjusted for heap low boundary)
342 // This would be the 0th element of _byte_map, if the heap started at 0x0.
343 // But since the heap starts at some higher address, this points to somewhere
344 // before the beginning of the actual _byte_map.
345 jbyte* byte_map_base;
347 // Return true if "p" is at the start of a card.
348 bool is_card_aligned(HeapWord* p) {
349 jbyte* pcard = byte_for(p);
350 return (addr_for(pcard) == p);
351 }
353 // The kinds of precision a CardTableModRefBS may offer.
354 enum PrecisionStyle {
355 Precise,
356 ObjHeadPreciseArray
357 };
359 // Tells what style of precision this card table offers.
360 PrecisionStyle precision() {
361 return ObjHeadPreciseArray; // Only one supported for now.
362 }
364 // ModRefBS functions.
365 virtual void invalidate(MemRegion mr, bool whole_heap = false);
366 void clear(MemRegion mr);
367 void dirty(MemRegion mr);
368 void mod_oop_in_space_iterate(Space* sp, OopClosure* cl,
369 bool clear = false,
370 bool before_save_marks = false);
372 // *** Card-table-RemSet-specific things.
374 // Invoke "cl.do_MemRegion" on a set of MemRegions that collectively
375 // includes all the modified cards (expressing each card as a
376 // MemRegion). Thus, several modified cards may be lumped into one
377 // region. The regions are non-overlapping, and are visited in
378 // *decreasing* address order. (This order aids with imprecise card
379 // marking, where a dirty card may cause scanning, and summarization
380 // marking, of objects that extend onto subsequent cards.)
381 // If "clear" is true, the card is (conceptually) marked unmodified before
382 // applying the closure.
383 void mod_card_iterate(MemRegionClosure* cl, bool clear = false) {
384 non_clean_card_iterate_work(_whole_heap, cl, clear);
385 }
387 // Like the "mod_cards_iterate" above, except only invokes the closure
388 // for cards within the MemRegion "mr" (which is required to be
389 // card-aligned and sized.)
390 void mod_card_iterate(MemRegion mr, MemRegionClosure* cl,
391 bool clear = false) {
392 non_clean_card_iterate_work(mr, cl, clear);
393 }
395 static uintx ct_max_alignment_constraint();
397 // Apply closure "cl" to the dirty cards containing some part of
398 // MemRegion "mr".
399 void dirty_card_iterate(MemRegion mr, MemRegionClosure* cl);
401 // Return the MemRegion corresponding to the first maximal run
402 // of dirty cards lying completely within MemRegion mr.
403 // If reset is "true", then sets those card table entries to the given
404 // value.
405 MemRegion dirty_card_range_after_reset(MemRegion mr, bool reset,
406 int reset_val);
408 // Set all the dirty cards in the given region to precleaned state.
409 void preclean_dirty_cards(MemRegion mr);
411 // Provide read-only access to the card table array.
412 const jbyte* byte_for_const(const void* p) const {
413 return byte_for(p);
414 }
415 const jbyte* byte_after_const(const void* p) const {
416 return byte_after(p);
417 }
419 // Mapping from card marking array entry to address of first word
420 HeapWord* addr_for(const jbyte* p) const {
421 assert(p >= _byte_map && p < _byte_map + _byte_map_size,
422 "out of bounds access to card marking array");
423 size_t delta = pointer_delta(p, byte_map_base, sizeof(jbyte));
424 HeapWord* result = (HeapWord*) (delta << card_shift);
425 assert(_whole_heap.contains(result),
426 "out of bounds accessor from card marking array");
427 return result;
428 }
430 // Mapping from address to card marking array index.
431 size_t index_for(void* p) {
432 assert(_whole_heap.contains(p),
433 "out of bounds access to card marking array");
434 return byte_for(p) - _byte_map;
435 }
437 void verify();
438 void verify_guard();
440 void verify_clean_region(MemRegion mr) PRODUCT_RETURN;
442 static size_t par_chunk_heapword_alignment() {
443 return CardsPerStrideChunk * card_size_in_words;
444 }
446 };
448 class CardTableRS;
450 // A specialization for the CardTableRS gen rem set.
451 class CardTableModRefBSForCTRS: public CardTableModRefBS {
452 CardTableRS* _rs;
453 protected:
454 bool card_will_be_scanned(jbyte cv);
455 bool card_may_have_been_dirty(jbyte cv);
456 public:
457 CardTableModRefBSForCTRS(MemRegion whole_heap,
458 int max_covered_regions) :
459 CardTableModRefBS(whole_heap, max_covered_regions) {}
461 void set_CTRS(CardTableRS* rs) { _rs = rs; }
462 };