1 #ifdef USE_PRAGMA_IDENT_HDR
2 #pragma ident "@(#)cardTableModRefBS.hpp 1.53 07/10/04 10:49:32 JVM"
3 #endif
4 /*
5 * Copyright 2000-2006 Sun Microsystems, Inc. All Rights Reserved.
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
7 *
8 * This code is free software; you can redistribute it and/or modify it
9 * under the terms of the GNU General Public License version 2 only, as
10 * published by the Free Software Foundation.
11 *
12 * This code is distributed in the hope that it will be useful, but WITHOUT
13 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
14 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
15 * version 2 for more details (a copy is included in the LICENSE file that
16 * accompanied this code).
17 *
18 * You should have received a copy of the GNU General Public License version
19 * 2 along with this work; if not, write to the Free Software Foundation,
20 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
21 *
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
24 * have any questions.
25 *
26 */
27
28 // This kind of "BarrierSet" allows a "CollectedHeap" to detect and
29 // enumerate ref fields that have been modified (since the last
30 // enumeration.)
31
32 // As it currently stands, this barrier is *imprecise*: when a ref field in
33 // an object "o" is modified, the card table entry for the card containing
34 // the head of "o" is dirtied, not necessarily the card containing the
35 // modified field itself. For object arrays, however, the barrier *is*
36 // precise; only the card containing the modified element is dirtied.
37 // Any MemRegionClosures used to scan dirty cards should take these
38 // considerations into account.
39
40 class Generation;
41 class OopsInGenClosure;
42 class DirtyCardToOopClosure;
43
44 class CardTableModRefBS: public ModRefBarrierSet {
45 // Some classes get to look at some private stuff.
46 friend class BytecodeInterpreter;
47 friend class VMStructs;
48 friend class CardTableRS;
49 friend class CheckForUnmarkedOops; // Needs access to raw card bytes.
50 #ifndef PRODUCT
51 // For debugging.
52 friend class GuaranteeNotModClosure;
53 #endif
54 protected:
55
56 enum CardValues {
57 clean_card = -1,
58 dirty_card = 0,
59 precleaned_card = 1,
60 last_card = 4,
61 CT_MR_BS_last_reserved = 10
62 };
63
64 // dirty and precleaned are equivalent wrt younger_refs_iter.
65 static bool card_is_dirty_wrt_gen_iter(jbyte cv) {
66 return cv == dirty_card || cv == precleaned_card;
67 }
68
69 // Returns "true" iff the value "cv" will cause the card containing it
70 // to be scanned in the current traversal. May be overridden by
71 // subtypes.
72 virtual bool card_will_be_scanned(jbyte cv) {
73 return CardTableModRefBS::card_is_dirty_wrt_gen_iter(cv);
74 }
75
76 // Returns "true" iff the value "cv" may have represented a dirty card at
77 // some point.
78 virtual bool card_may_have_been_dirty(jbyte cv) {
79 return card_is_dirty_wrt_gen_iter(cv);
80 }
81
82 // The declaration order of these const fields is important; see the
83 // constructor before changing.
84 const MemRegion _whole_heap; // the region covered by the card table
85 const size_t _guard_index; // index of very last element in the card
86 // table; it is set to a guard value
87 // (last_card) and should never be modified
88 const size_t _last_valid_index; // index of the last valid element
89 const size_t _page_size; // page size used when mapping _byte_map
90 const size_t _byte_map_size; // in bytes
91 jbyte* _byte_map; // the card marking array
92
93 int _cur_covered_regions;
94 // The covered regions should be in address order.
95 MemRegion* _covered;
96 // The committed regions correspond one-to-one to the covered regions.
97 // They represent the card-table memory that has been committed to service
98 // the corresponding covered region. It may be that committed region for
99 // one covered region corresponds to a larger region because of page-size
100 // roundings. Thus, a committed region for one covered region may
101 // actually extend onto the card-table space for the next covered region.
102 MemRegion* _committed;
103
104 // The last card is a guard card, and we commit the page for it so
105 // we can use the card for verification purposes. We make sure we never
106 // uncommit the MemRegion for that page.
107 MemRegion _guard_region;
108
109 protected:
110 // Initialization utilities; covered_words is the size of the covered region
111 // in, um, words.
112 inline size_t cards_required(size_t covered_words);
113 inline size_t compute_byte_map_size();
114
115 // Finds and return the index of the region, if any, to which the given
116 // region would be contiguous. If none exists, assign a new region and
117 // returns its index. Requires that no more than the maximum number of
118 // covered regions defined in the constructor are ever in use.
119 int find_covering_region_by_base(HeapWord* base);
120
121 // Same as above, but finds the region containing the given address
122 // instead of starting at a given base address.
123 int find_covering_region_containing(HeapWord* addr);
124
125 // Resize one of the regions covered by the remembered set.
126 void resize_covered_region(MemRegion new_region);
127
128 // Returns the leftmost end of a committed region corresponding to a
129 // covered region before covered region "ind", or else "NULL" if "ind" is
130 // the first covered region.
131 HeapWord* largest_prev_committed_end(int ind) const;
132
133 // Returns the part of the region mr that doesn't intersect with
134 // any committed region other than self. Used to prevent uncommitting
135 // regions that are also committed by other regions. Also protects
136 // against uncommitting the guard region.
137 MemRegion committed_unique_to_self(int self, MemRegion mr) const;
138
139 // Mapping from address to card marking array entry
140 jbyte* byte_for(const void* p) const {
141 assert(_whole_heap.contains(p),
142 "out of bounds access to card marking array");
143 jbyte* result = &byte_map_base[uintptr_t(p) >> card_shift];
144 assert(result >= _byte_map && result < _byte_map + _byte_map_size,
145 "out of bounds accessor for card marking array");
146 return result;
147 }
148
149 // The card table byte one after the card marking array
150 // entry for argument address. Typically used for higher bounds
151 // for loops iterating through the card table.
152 jbyte* byte_after(const void* p) const {
153 return byte_for(p) + 1;
154 }
155
156 // Mapping from card marking array entry to address of first word
157 HeapWord* addr_for(const jbyte* p) const {
158 assert(p >= _byte_map && p < _byte_map + _byte_map_size,
159 "out of bounds access to card marking array");
160 size_t delta = pointer_delta(p, byte_map_base, sizeof(jbyte));
161 HeapWord* result = (HeapWord*) (delta << card_shift);
162 assert(_whole_heap.contains(result),
163 "out of bounds accessor from card marking array");
164 return result;
165 }
166
167 // Iterate over the portion of the card-table which covers the given
168 // region mr in the given space and apply cl to any dirty sub-regions
169 // of mr. cl and dcto_cl must either be the same closure or cl must
170 // wrap dcto_cl. Both are required - neither may be NULL. Also, dcto_cl
171 // may be modified. Note that this function will operate in a parallel
172 // mode if worker threads are available.
173 void non_clean_card_iterate(Space* sp, MemRegion mr,
174 DirtyCardToOopClosure* dcto_cl,
175 MemRegionClosure* cl,
176 bool clear);
177
178 // Utility function used to implement the other versions below.
179 void non_clean_card_iterate_work(MemRegion mr, MemRegionClosure* cl,
180 bool clear);
181
182 void par_non_clean_card_iterate_work(Space* sp, MemRegion mr,
183 DirtyCardToOopClosure* dcto_cl,
184 MemRegionClosure* cl,
185 bool clear,
186 int n_threads);
187
188 // Dirty the bytes corresponding to "mr" (not all of which must be
189 // covered.)
190 void dirty_MemRegion(MemRegion mr);
191
192 // Clear (to clean_card) the bytes entirely contained within "mr" (not
193 // all of which must be covered.)
194 void clear_MemRegion(MemRegion mr);
195
196 // *** Support for parallel card scanning.
197
198 enum SomeConstantsForParallelism {
199 StridesPerThread = 2,
200 CardsPerStrideChunk = 256
201 };
202
203 // This is an array, one element per covered region of the card table.
204 // Each entry is itself an array, with one element per chunk in the
205 // covered region. Each entry of these arrays is the lowest non-clean
206 // card of the corresponding chunk containing part of an object from the
207 // previous chunk, or else NULL.
208 typedef jbyte* CardPtr;
209 typedef CardPtr* CardArr;
210 CardArr* _lowest_non_clean;
211 size_t* _lowest_non_clean_chunk_size;
212 uintptr_t* _lowest_non_clean_base_chunk_index;
213 int* _last_LNC_resizing_collection;
214
215 // Initializes "lowest_non_clean" to point to the array for the region
216 // covering "sp", and "lowest_non_clean_base_chunk_index" to the chunk
217 // index of the corresponding to the first element of that array.
218 // Ensures that these arrays are of sufficient size, allocating if necessary.
219 // May be called by several threads concurrently.
220 void get_LNC_array_for_space(Space* sp,
221 jbyte**& lowest_non_clean,
222 uintptr_t& lowest_non_clean_base_chunk_index,
223 size_t& lowest_non_clean_chunk_size);
224
225 // Returns the number of chunks necessary to cover "mr".
226 size_t chunks_to_cover(MemRegion mr) {
227 return (size_t)(addr_to_chunk_index(mr.last()) -
228 addr_to_chunk_index(mr.start()) + 1);
229 }
230
231 // Returns the index of the chunk in a stride which
232 // covers the given address.
233 uintptr_t addr_to_chunk_index(const void* addr) {
234 uintptr_t card = (uintptr_t) byte_for(addr);
235 return card / CardsPerStrideChunk;
236 }
237
238 // Apply cl, which must either itself apply dcto_cl or be dcto_cl,
239 // to the cards in the stride (of n_strides) within the given space.
240 void process_stride(Space* sp,
241 MemRegion used,
242 jint stride, int n_strides,
243 DirtyCardToOopClosure* dcto_cl,
244 MemRegionClosure* cl,
245 bool clear,
246 jbyte** lowest_non_clean,
247 uintptr_t lowest_non_clean_base_chunk_index,
248 size_t lowest_non_clean_chunk_size);
249
250 // Makes sure that chunk boundaries are handled appropriately, by
251 // adjusting the min_done of dcto_cl, and by using a special card-table
252 // value to indicate how min_done should be set.
253 void process_chunk_boundaries(Space* sp,
254 DirtyCardToOopClosure* dcto_cl,
255 MemRegion chunk_mr,
256 MemRegion used,
257 jbyte** lowest_non_clean,
258 uintptr_t lowest_non_clean_base_chunk_index,
259 size_t lowest_non_clean_chunk_size);
260
261 public:
262 // Constants
263 enum SomePublicConstants {
264 card_shift = 9,
265 card_size = 1 << card_shift,
266 card_size_in_words = card_size / sizeof(HeapWord)
267 };
268
269 // For RTTI simulation.
270 BarrierSet::Name kind() { return BarrierSet::CardTableModRef; }
271 bool is_a(BarrierSet::Name bsn) {
272 return bsn == BarrierSet::CardTableModRef || bsn == BarrierSet::ModRef;
273 }
274
275 CardTableModRefBS(MemRegion whole_heap, int max_covered_regions);
276
277 // *** Barrier set functions.
278
279 inline bool write_ref_needs_barrier(oop* field, oop new_val) {
280 // Note that this assumes the perm gen is the highest generation
281 // in the address space
282 return new_val != NULL && !new_val->is_perm();
283 }
284
285 // Record a reference update. Note that these versions are precise!
286 // The scanning code has to handle the fact that the write barrier may be
287 // either precise or imprecise. We make non-virtual inline variants of
288 // these functions here for performance.
289 protected:
290 void write_ref_field_work(oop obj, size_t offset, oop newVal);
291 void write_ref_field_work(oop* field, oop newVal);
292 public:
293
294 bool has_write_ref_array_opt() { return true; }
295 bool has_write_region_opt() { return true; }
296
297 inline void inline_write_region(MemRegion mr) {
298 dirty_MemRegion(mr);
299 }
300 protected:
301 void write_region_work(MemRegion mr) {
302 inline_write_region(mr);
303 }
304 public:
305
306 inline void inline_write_ref_array(MemRegion mr) {
307 dirty_MemRegion(mr);
308 }
309 protected:
310 void write_ref_array_work(MemRegion mr) {
311 inline_write_ref_array(mr);
312 }
313 public:
314
315 bool is_aligned(HeapWord* addr) {
316 return is_card_aligned(addr);
317 }
318
319 // *** Card-table-barrier-specific things.
320
321 inline void inline_write_ref_field(oop* field, oop newVal) {
322 jbyte* byte = byte_for(field);
323 *byte = dirty_card;
324 }
325
326 // Card marking array base (adjusted for heap low boundary)
327 // This would be the 0th element of _byte_map, if the heap started at 0x0.
328 // But since the heap starts at some higher address, this points to somewhere
329 // before the beginning of the actual _byte_map.
330 jbyte* byte_map_base;
331
332 // Return true if "p" is at the start of a card.
333 bool is_card_aligned(HeapWord* p) {
334 jbyte* pcard = byte_for(p);
335 return (addr_for(pcard) == p);
336 }
337
338 // The kinds of precision a CardTableModRefBS may offer.
339 enum PrecisionStyle {
340 Precise,
341 ObjHeadPreciseArray
342 };
343
344 // Tells what style of precision this card table offers.
345 PrecisionStyle precision() {
346 return ObjHeadPreciseArray; // Only one supported for now.
347 }
348
349 // ModRefBS functions.
350 void invalidate(MemRegion mr);
351 void clear(MemRegion mr);
352 void mod_oop_in_space_iterate(Space* sp, OopClosure* cl,
353 bool clear = false,
354 bool before_save_marks = false);
355
356 // *** Card-table-RemSet-specific things.
357
358 // Invoke "cl.do_MemRegion" on a set of MemRegions that collectively
359 // includes all the modified cards (expressing each card as a
360 // MemRegion). Thus, several modified cards may be lumped into one
361 // region. The regions are non-overlapping, and are visited in
362 // *decreasing* address order. (This order aids with imprecise card
363 // marking, where a dirty card may cause scanning, and summarization
364 // marking, of objects that extend onto subsequent cards.)
365 // If "clear" is true, the card is (conceptually) marked unmodified before
366 // applying the closure.
367 void mod_card_iterate(MemRegionClosure* cl, bool clear = false) {
368 non_clean_card_iterate_work(_whole_heap, cl, clear);
369 }
370
371 // Like the "mod_cards_iterate" above, except only invokes the closure
372 // for cards within the MemRegion "mr" (which is required to be
373 // card-aligned and sized.)
374 void mod_card_iterate(MemRegion mr, MemRegionClosure* cl,
375 bool clear = false) {
376 non_clean_card_iterate_work(mr, cl, clear);
377 }
378
379 static uintx ct_max_alignment_constraint();
380
381 // Apply closure cl to the dirty cards lying completely
382 // within MemRegion mr, setting the cards to precleaned.
383 void dirty_card_iterate(MemRegion mr, MemRegionClosure* cl);
384
385 // Return the MemRegion corresponding to the first maximal run
386 // of dirty cards lying completely within MemRegion mr, after
387 // marking those cards precleaned.
388 MemRegion dirty_card_range_after_preclean(MemRegion mr);
389
390 // Set all the dirty cards in the given region to precleaned state.
391 void preclean_dirty_cards(MemRegion mr);
392
393 // Mapping from address to card marking array index.
394 int index_for(void* p) {
395 assert(_whole_heap.contains(p),
396 "out of bounds access to card marking array");
397 return byte_for(p) - _byte_map;
398 }
399
400 void verify();
401 void verify_guard();
402
403 void verify_clean_region(MemRegion mr) PRODUCT_RETURN;
404
405 static size_t par_chunk_heapword_alignment() {
406 return CardsPerStrideChunk * card_size_in_words;
407 }
408 };
409
410 class CardTableRS;
411
412 // A specialization for the CardTableRS gen rem set.
413 class CardTableModRefBSForCTRS: public CardTableModRefBS {
414 CardTableRS* _rs;
415 protected:
416 bool card_will_be_scanned(jbyte cv);
417 bool card_may_have_been_dirty(jbyte cv);
418 public:
419 CardTableModRefBSForCTRS(MemRegion whole_heap,
420 int max_covered_regions) :
421 CardTableModRefBS(whole_heap, max_covered_regions) {}
422
423 void set_CTRS(CardTableRS* rs) { _rs = rs; }
424 };