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