1 /* 2 * Copyright (c) 2007, 2015, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "gc/shared/cardTableModRefBS.hpp" 27 #include "gc/shared/cardTableRS.hpp" 28 #include "gc/shared/collectedHeap.hpp" 29 #include "gc/shared/genCollectedHeap.hpp" 30 #include "gc/shared/space.inline.hpp" 31 #include "memory/allocation.inline.hpp" 32 #include "memory/virtualspace.hpp" 33 #include "oops/oop.inline.hpp" 34 #include "runtime/java.hpp" 35 #include "runtime/mutexLocker.hpp" 36 #include "runtime/orderAccess.inline.hpp" 37 #include "runtime/vmThread.hpp" 38 39 void CardTableModRefBSForCTRS:: 40 non_clean_card_iterate_parallel_work(Space* sp, MemRegion mr, 41 OopsInGenClosure* cl, 42 CardTableRS* ct, 43 uint n_threads) { 44 assert(n_threads > 0, "expected n_threads > 0"); 45 assert(n_threads <= ParallelGCThreads, 46 "n_threads: %u > ParallelGCThreads: %u", n_threads, ParallelGCThreads); 47 48 // Make sure the LNC array is valid for the space. 49 jbyte** lowest_non_clean; 50 uintptr_t lowest_non_clean_base_chunk_index; 51 size_t lowest_non_clean_chunk_size; 52 get_LNC_array_for_space(sp, lowest_non_clean, 53 lowest_non_clean_base_chunk_index, 54 lowest_non_clean_chunk_size); 55 56 uint n_strides = n_threads * ParGCStridesPerThread; 57 SequentialSubTasksDone* pst = sp->par_seq_tasks(); 58 // Sets the condition for completion of the subtask (how many threads 59 // need to finish in order to be done). 60 pst->set_n_threads(n_threads); 61 pst->set_n_tasks(n_strides); 62 63 uint stride = 0; 64 while (!pst->is_task_claimed(/* reference */ stride)) { 65 process_stride(sp, mr, stride, n_strides, 66 cl, ct, 67 lowest_non_clean, 68 lowest_non_clean_base_chunk_index, 69 lowest_non_clean_chunk_size); 70 } 71 if (pst->all_tasks_completed()) { 72 // Clear lowest_non_clean array for next time. 73 intptr_t first_chunk_index = addr_to_chunk_index(mr.start()); 74 uintptr_t last_chunk_index = addr_to_chunk_index(mr.last()); 75 for (uintptr_t ch = first_chunk_index; ch <= last_chunk_index; ch++) { 76 intptr_t ind = ch - lowest_non_clean_base_chunk_index; 77 assert(0 <= ind && ind < (intptr_t)lowest_non_clean_chunk_size, 78 "Bounds error"); 79 lowest_non_clean[ind] = NULL; 80 } 81 } 82 } 83 84 void 85 CardTableModRefBSForCTRS:: 86 process_stride(Space* sp, 87 MemRegion used, 88 jint stride, int n_strides, 89 OopsInGenClosure* cl, 90 CardTableRS* ct, 91 jbyte** lowest_non_clean, 92 uintptr_t lowest_non_clean_base_chunk_index, 93 size_t lowest_non_clean_chunk_size) { 94 // We go from higher to lower addresses here; it wouldn't help that much 95 // because of the strided parallelism pattern used here. 96 97 // Find the first card address of the first chunk in the stride that is 98 // at least "bottom" of the used region. 99 jbyte* start_card = byte_for(used.start()); 100 jbyte* end_card = byte_after(used.last()); 101 uintptr_t start_chunk = addr_to_chunk_index(used.start()); 102 uintptr_t start_chunk_stride_num = start_chunk % n_strides; 103 jbyte* chunk_card_start; 104 105 if ((uintptr_t)stride >= start_chunk_stride_num) { 106 chunk_card_start = (jbyte*)(start_card + 107 (stride - start_chunk_stride_num) * 108 ParGCCardsPerStrideChunk); 109 } else { 110 // Go ahead to the next chunk group boundary, then to the requested stride. 111 chunk_card_start = (jbyte*)(start_card + 112 (n_strides - start_chunk_stride_num + stride) * 113 ParGCCardsPerStrideChunk); 114 } 115 116 while (chunk_card_start < end_card) { 117 // Even though we go from lower to higher addresses below, the 118 // strided parallelism can interleave the actual processing of the 119 // dirty pages in various ways. For a specific chunk within this 120 // stride, we take care to avoid double scanning or missing a card 121 // by suitably initializing the "min_done" field in process_chunk_boundaries() 122 // below, together with the dirty region extension accomplished in 123 // DirtyCardToOopClosure::do_MemRegion(). 124 jbyte* chunk_card_end = chunk_card_start + ParGCCardsPerStrideChunk; 125 // Invariant: chunk_mr should be fully contained within the "used" region. 126 MemRegion chunk_mr = MemRegion(addr_for(chunk_card_start), 127 chunk_card_end >= end_card ? 128 used.end() : addr_for(chunk_card_end)); 129 assert(chunk_mr.word_size() > 0, "[chunk_card_start > used_end)"); 130 assert(used.contains(chunk_mr), "chunk_mr should be subset of used"); 131 132 // This function is used by the parallel card table iteration. 133 const bool parallel = true; 134 135 DirtyCardToOopClosure* dcto_cl = sp->new_dcto_cl(cl, precision(), 136 cl->gen_boundary(), 137 parallel); 138 ClearNoncleanCardWrapper clear_cl(dcto_cl, ct, parallel); 139 140 141 // Process the chunk. 142 process_chunk_boundaries(sp, 143 dcto_cl, 144 chunk_mr, 145 used, 146 lowest_non_clean, 147 lowest_non_clean_base_chunk_index, 148 lowest_non_clean_chunk_size); 149 150 // We want the LNC array updates above in process_chunk_boundaries 151 // to be visible before any of the card table value changes as a 152 // result of the dirty card iteration below. 153 OrderAccess::storestore(); 154 155 // We want to clear the cards: clear_cl here does the work of finding 156 // contiguous dirty ranges of cards to process and clear. 157 clear_cl.do_MemRegion(chunk_mr); 158 159 // Find the next chunk of the stride. 160 chunk_card_start += ParGCCardsPerStrideChunk * n_strides; 161 } 162 } 163 164 165 // If you want a talkative process_chunk_boundaries, 166 // then #define NOISY(x) x 167 #ifdef NOISY 168 #error "Encountered a global preprocessor flag, NOISY, which might clash with local definition to follow" 169 #else 170 #define NOISY(x) 171 #endif 172 173 void 174 CardTableModRefBSForCTRS:: 175 process_chunk_boundaries(Space* sp, 176 DirtyCardToOopClosure* dcto_cl, 177 MemRegion chunk_mr, 178 MemRegion used, 179 jbyte** lowest_non_clean, 180 uintptr_t lowest_non_clean_base_chunk_index, 181 size_t lowest_non_clean_chunk_size) 182 { 183 // We must worry about non-array objects that cross chunk boundaries, 184 // because such objects are both precisely and imprecisely marked: 185 // .. if the head of such an object is dirty, the entire object 186 // needs to be scanned, under the interpretation that this 187 // was an imprecise mark 188 // .. if the head of such an object is not dirty, we can assume 189 // precise marking and it's efficient to scan just the dirty 190 // cards. 191 // In either case, each scanned reference must be scanned precisely 192 // once so as to avoid cloning of a young referent. For efficiency, 193 // our closures depend on this property and do not protect against 194 // double scans. 195 196 uintptr_t start_chunk_index = addr_to_chunk_index(chunk_mr.start()); 197 assert(start_chunk_index >= lowest_non_clean_base_chunk_index, "Bounds error."); 198 uintptr_t cur_chunk_index = start_chunk_index - lowest_non_clean_base_chunk_index; 199 200 NOISY(tty->print_cr("===========================================================================");) 201 NOISY(tty->print_cr(" process_chunk_boundary: Called with [" PTR_FORMAT "," PTR_FORMAT ")", 202 chunk_mr.start(), chunk_mr.end());) 203 204 // First, set "our" lowest_non_clean entry, which would be 205 // used by the thread scanning an adjoining left chunk with 206 // a non-array object straddling the mutual boundary. 207 // Find the object that spans our boundary, if one exists. 208 // first_block is the block possibly straddling our left boundary. 209 HeapWord* first_block = sp->block_start(chunk_mr.start()); 210 assert((chunk_mr.start() != used.start()) || (first_block == chunk_mr.start()), 211 "First chunk should always have a co-initial block"); 212 // Does the block straddle the chunk's left boundary, and is it 213 // a non-array object? 214 if (first_block < chunk_mr.start() // first block straddles left bdry 215 && sp->block_is_obj(first_block) // first block is an object 216 && !(oop(first_block)->is_objArray() // first block is not an array (arrays are precisely dirtied) 217 || oop(first_block)->is_typeArray())) { 218 // Find our least non-clean card, so that a left neighbor 219 // does not scan an object straddling the mutual boundary 220 // too far to the right, and attempt to scan a portion of 221 // that object twice. 222 jbyte* first_dirty_card = NULL; 223 jbyte* last_card_of_first_obj = 224 byte_for(first_block + sp->block_size(first_block) - 1); 225 jbyte* first_card_of_cur_chunk = byte_for(chunk_mr.start()); 226 jbyte* last_card_of_cur_chunk = byte_for(chunk_mr.last()); 227 jbyte* last_card_to_check = 228 (jbyte*) MIN2((intptr_t) last_card_of_cur_chunk, 229 (intptr_t) last_card_of_first_obj); 230 // Note that this does not need to go beyond our last card 231 // if our first object completely straddles this chunk. 232 for (jbyte* cur = first_card_of_cur_chunk; 233 cur <= last_card_to_check; cur++) { 234 jbyte val = *cur; 235 if (card_will_be_scanned(val)) { 236 first_dirty_card = cur; break; 237 } else { 238 assert(!card_may_have_been_dirty(val), "Error"); 239 } 240 } 241 if (first_dirty_card != NULL) { 242 NOISY(tty->print_cr(" LNC: Found a dirty card at " PTR_FORMAT " in current chunk", 243 first_dirty_card);) 244 assert(cur_chunk_index < lowest_non_clean_chunk_size, "Bounds error."); 245 assert(lowest_non_clean[cur_chunk_index] == NULL, 246 "Write exactly once : value should be stable hereafter for this round"); 247 lowest_non_clean[cur_chunk_index] = first_dirty_card; 248 } NOISY(else { 249 tty->print_cr(" LNC: Found no dirty card in current chunk; leaving LNC entry NULL"); 250 // In the future, we could have this thread look for a non-NULL value to copy from its 251 // right neighbor (up to the end of the first object). 252 if (last_card_of_cur_chunk < last_card_of_first_obj) { 253 tty->print_cr(" LNC: BEWARE!!! first obj straddles past right end of chunk:\n" 254 " might be efficient to get value from right neighbor?"); 255 } 256 }) 257 } else { 258 // In this case we can help our neighbor by just asking them 259 // to stop at our first card (even though it may not be dirty). 260 NOISY(tty->print_cr(" LNC: first block is not a non-array object; setting LNC to first card of current chunk");) 261 assert(lowest_non_clean[cur_chunk_index] == NULL, "Write once : value should be stable hereafter"); 262 jbyte* first_card_of_cur_chunk = byte_for(chunk_mr.start()); 263 lowest_non_clean[cur_chunk_index] = first_card_of_cur_chunk; 264 } 265 NOISY(tty->print_cr(" process_chunk_boundary: lowest_non_clean[" INTPTR_FORMAT "] = " PTR_FORMAT 266 " which corresponds to the heap address " PTR_FORMAT, 267 cur_chunk_index, lowest_non_clean[cur_chunk_index], 268 (lowest_non_clean[cur_chunk_index] != NULL) 269 ? addr_for(lowest_non_clean[cur_chunk_index]) 270 : NULL);) 271 NOISY(tty->print_cr("---------------------------------------------------------------------------");) 272 273 // Next, set our own max_to_do, which will strictly/exclusively bound 274 // the highest address that we will scan past the right end of our chunk. 275 HeapWord* max_to_do = NULL; 276 if (chunk_mr.end() < used.end()) { 277 // This is not the last chunk in the used region. 278 // What is our last block? We check the first block of 279 // the next (right) chunk rather than strictly check our last block 280 // because it's potentially more efficient to do so. 281 HeapWord* const last_block = sp->block_start(chunk_mr.end()); 282 assert(last_block <= chunk_mr.end(), "In case this property changes."); 283 if ((last_block == chunk_mr.end()) // our last block does not straddle boundary 284 || !sp->block_is_obj(last_block) // last_block isn't an object 285 || oop(last_block)->is_objArray() // last_block is an array (precisely marked) 286 || oop(last_block)->is_typeArray()) { 287 max_to_do = chunk_mr.end(); 288 NOISY(tty->print_cr(" process_chunk_boundary: Last block on this card is not a non-array object;\n" 289 " max_to_do left at " PTR_FORMAT, max_to_do);) 290 } else { 291 assert(last_block < chunk_mr.end(), "Tautology"); 292 // It is a non-array object that straddles the right boundary of this chunk. 293 // last_obj_card is the card corresponding to the start of the last object 294 // in the chunk. Note that the last object may not start in 295 // the chunk. 296 jbyte* const last_obj_card = byte_for(last_block); 297 const jbyte val = *last_obj_card; 298 if (!card_will_be_scanned(val)) { 299 assert(!card_may_have_been_dirty(val), "Error"); 300 // The card containing the head is not dirty. Any marks on 301 // subsequent cards still in this chunk must have been made 302 // precisely; we can cap processing at the end of our chunk. 303 max_to_do = chunk_mr.end(); 304 NOISY(tty->print_cr(" process_chunk_boundary: Head of last object on this card is not dirty;\n" 305 " max_to_do left at " PTR_FORMAT, 306 max_to_do);) 307 } else { 308 // The last object must be considered dirty, and extends onto the 309 // following chunk. Look for a dirty card in that chunk that will 310 // bound our processing. 311 jbyte* limit_card = NULL; 312 const size_t last_block_size = sp->block_size(last_block); 313 jbyte* const last_card_of_last_obj = 314 byte_for(last_block + last_block_size - 1); 315 jbyte* const first_card_of_next_chunk = byte_for(chunk_mr.end()); 316 // This search potentially goes a long distance looking 317 // for the next card that will be scanned, terminating 318 // at the end of the last_block, if no earlier dirty card 319 // is found. 320 assert(byte_for(chunk_mr.end()) - byte_for(chunk_mr.start()) == ParGCCardsPerStrideChunk, 321 "last card of next chunk may be wrong"); 322 for (jbyte* cur = first_card_of_next_chunk; 323 cur <= last_card_of_last_obj; cur++) { 324 const jbyte val = *cur; 325 if (card_will_be_scanned(val)) { 326 NOISY(tty->print_cr(" Found a non-clean card " PTR_FORMAT " with value 0x%x", 327 cur, (int)val);) 328 limit_card = cur; break; 329 } else { 330 assert(!card_may_have_been_dirty(val), "Error: card can't be skipped"); 331 } 332 } 333 if (limit_card != NULL) { 334 max_to_do = addr_for(limit_card); 335 assert(limit_card != NULL && max_to_do != NULL, "Error"); 336 NOISY(tty->print_cr(" process_chunk_boundary: Found a dirty card at " PTR_FORMAT 337 " max_to_do set at " PTR_FORMAT " which is before end of last block in chunk: " 338 PTR_FORMAT " + " PTR_FORMAT " = " PTR_FORMAT, 339 limit_card, max_to_do, last_block, last_block_size, (last_block+last_block_size));) 340 } else { 341 // The following is a pessimistic value, because it's possible 342 // that a dirty card on a subsequent chunk has been cleared by 343 // the time we get to look at it; we'll correct for that further below, 344 // using the LNC array which records the least non-clean card 345 // before cards were cleared in a particular chunk. 346 limit_card = last_card_of_last_obj; 347 max_to_do = last_block + last_block_size; 348 assert(limit_card != NULL && max_to_do != NULL, "Error"); 349 NOISY(tty->print_cr(" process_chunk_boundary: Found no dirty card before end of last block in chunk\n" 350 " Setting limit_card to " PTR_FORMAT 351 " and max_to_do " PTR_FORMAT " + " PTR_FORMAT " = " PTR_FORMAT, 352 limit_card, last_block, last_block_size, max_to_do);) 353 } 354 assert(0 < cur_chunk_index+1 && cur_chunk_index+1 < lowest_non_clean_chunk_size, 355 "Bounds error."); 356 // It is possible that a dirty card for the last object may have been 357 // cleared before we had a chance to examine it. In that case, the value 358 // will have been logged in the LNC for that chunk. 359 // We need to examine as many chunks to the right as this object 360 // covers. However, we need to bound this checking to the largest 361 // entry in the LNC array: this is because the heap may expand 362 // after the LNC array has been created but before we reach this point, 363 // and the last block in our chunk may have been expanded to include 364 // the expansion delta (and possibly subsequently allocated from, so 365 // it wouldn't be sufficient to check whether that last block was 366 // or was not an object at this point). 367 uintptr_t last_chunk_index_to_check = addr_to_chunk_index(last_block + last_block_size - 1) 368 - lowest_non_clean_base_chunk_index; 369 const uintptr_t last_chunk_index = addr_to_chunk_index(used.last()) 370 - lowest_non_clean_base_chunk_index; 371 if (last_chunk_index_to_check > last_chunk_index) { 372 assert(last_block + last_block_size > used.end(), 373 "Inconsistency detected: last_block [" PTR_FORMAT "," PTR_FORMAT "]" 374 " does not exceed used.end() = " PTR_FORMAT "," 375 " yet last_chunk_index_to_check " INTPTR_FORMAT 376 " exceeds last_chunk_index " INTPTR_FORMAT, 377 p2i(last_block), p2i(last_block + last_block_size), 378 p2i(used.end()), 379 last_chunk_index_to_check, last_chunk_index); 380 assert(sp->used_region().end() > used.end(), 381 "Expansion did not happen: " 382 "[" PTR_FORMAT "," PTR_FORMAT ") -> [" PTR_FORMAT "," PTR_FORMAT ")", 383 p2i(sp->used_region().start()), p2i(sp->used_region().end()), 384 p2i(used.start()), p2i(used.end())); 385 NOISY(tty->print_cr(" process_chunk_boundary: heap expanded; explicitly bounding last_chunk");) 386 last_chunk_index_to_check = last_chunk_index; 387 } 388 for (uintptr_t lnc_index = cur_chunk_index + 1; 389 lnc_index <= last_chunk_index_to_check; 390 lnc_index++) { 391 jbyte* lnc_card = lowest_non_clean[lnc_index]; 392 if (lnc_card != NULL) { 393 // we can stop at the first non-NULL entry we find 394 if (lnc_card <= limit_card) { 395 NOISY(tty->print_cr(" process_chunk_boundary: LNC card " PTR_FORMAT " is lower than limit_card " PTR_FORMAT, 396 " max_to_do will be lowered to " PTR_FORMAT " from " PTR_FORMAT, 397 lnc_card, limit_card, addr_for(lnc_card), max_to_do);) 398 limit_card = lnc_card; 399 max_to_do = addr_for(limit_card); 400 assert(limit_card != NULL && max_to_do != NULL, "Error"); 401 } 402 // In any case, we break now 403 break; 404 } // else continue to look for a non-NULL entry if any 405 } 406 assert(limit_card != NULL && max_to_do != NULL, "Error"); 407 } 408 assert(max_to_do != NULL, "OOPS 1 !"); 409 } 410 assert(max_to_do != NULL, "OOPS 2!"); 411 } else { 412 max_to_do = used.end(); 413 NOISY(tty->print_cr(" process_chunk_boundary: Last chunk of this space;\n" 414 " max_to_do left at " PTR_FORMAT, 415 max_to_do);) 416 } 417 assert(max_to_do != NULL, "OOPS 3!"); 418 // Now we can set the closure we're using so it doesn't to beyond 419 // max_to_do. 420 dcto_cl->set_min_done(max_to_do); 421 #ifndef PRODUCT 422 dcto_cl->set_last_bottom(max_to_do); 423 #endif 424 NOISY(tty->print_cr("===========================================================================\n");) 425 } 426 427 #undef NOISY 428 429 void 430 CardTableModRefBSForCTRS:: 431 get_LNC_array_for_space(Space* sp, 432 jbyte**& lowest_non_clean, 433 uintptr_t& lowest_non_clean_base_chunk_index, 434 size_t& lowest_non_clean_chunk_size) { 435 436 int i = find_covering_region_containing(sp->bottom()); 437 MemRegion covered = _covered[i]; 438 size_t n_chunks = chunks_to_cover(covered); 439 440 // Only the first thread to obtain the lock will resize the 441 // LNC array for the covered region. Any later expansion can't affect 442 // the used_at_save_marks region. 443 // (I observed a bug in which the first thread to execute this would 444 // resize, and then it would cause "expand_and_allocate" that would 445 // increase the number of chunks in the covered region. Then a second 446 // thread would come and execute this, see that the size didn't match, 447 // and free and allocate again. So the first thread would be using a 448 // freed "_lowest_non_clean" array.) 449 450 // Do a dirty read here. If we pass the conditional then take the rare 451 // event lock and do the read again in case some other thread had already 452 // succeeded and done the resize. 453 int cur_collection = GenCollectedHeap::heap()->total_collections(); 454 if (_last_LNC_resizing_collection[i] != cur_collection) { 455 MutexLocker x(ParGCRareEvent_lock); 456 if (_last_LNC_resizing_collection[i] != cur_collection) { 457 if (_lowest_non_clean[i] == NULL || 458 n_chunks != _lowest_non_clean_chunk_size[i]) { 459 460 // Should we delete the old? 461 if (_lowest_non_clean[i] != NULL) { 462 assert(n_chunks != _lowest_non_clean_chunk_size[i], 463 "logical consequence"); 464 FREE_C_HEAP_ARRAY(CardPtr, _lowest_non_clean[i]); 465 _lowest_non_clean[i] = NULL; 466 } 467 // Now allocate a new one if necessary. 468 if (_lowest_non_clean[i] == NULL) { 469 _lowest_non_clean[i] = NEW_C_HEAP_ARRAY(CardPtr, n_chunks, mtGC); 470 _lowest_non_clean_chunk_size[i] = n_chunks; 471 _lowest_non_clean_base_chunk_index[i] = addr_to_chunk_index(covered.start()); 472 for (int j = 0; j < (int)n_chunks; j++) 473 _lowest_non_clean[i][j] = NULL; 474 } 475 } 476 _last_LNC_resizing_collection[i] = cur_collection; 477 } 478 } 479 // In any case, now do the initialization. 480 lowest_non_clean = _lowest_non_clean[i]; 481 lowest_non_clean_base_chunk_index = _lowest_non_clean_base_chunk_index[i]; 482 lowest_non_clean_chunk_size = _lowest_non_clean_chunk_size[i]; 483 }