1 /*
   2  * Copyright (c) 2007, 2018, 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/cms/cmsCardTable.hpp"
  27 #include "gc/cms/cmsHeap.hpp"
  28 #include "gc/shared/cardTableBarrierSet.hpp"
  29 #include "gc/shared/cardTableRS.hpp"
  30 #include "gc/shared/collectedHeap.hpp"
  31 #include "gc/shared/space.inline.hpp"
  32 #include "memory/allocation.inline.hpp"
  33 #include "memory/virtualspace.hpp"
  34 #include "oops/oop.inline.hpp"
  35 #include "runtime/java.hpp"
  36 #include "runtime/mutexLocker.hpp"
  37 #include "runtime/orderAccess.inline.hpp"
  38 #include "runtime/vmThread.hpp"
  39 
  40 CMSCardTable::CMSCardTable(MemRegion whole_heap, bool scanned_concurrently) :
  41     CardTableRS(whole_heap, scanned_concurrently) {
  42 }
  43 
  44 // Returns the number of chunks necessary to cover "mr".
  45 size_t CMSCardTable::chunks_to_cover(MemRegion mr) {
  46   return (size_t)(addr_to_chunk_index(mr.last()) -
  47                   addr_to_chunk_index(mr.start()) + 1);
  48 }
  49 
  50 // Returns the index of the chunk in a stride which
  51 // covers the given address.
  52 uintptr_t CMSCardTable::addr_to_chunk_index(const void* addr) {
  53   uintptr_t card = (uintptr_t) byte_for(addr);
  54   return card / ParGCCardsPerStrideChunk;
  55 }
  56 
  57 void CMSCardTable::
  58 non_clean_card_iterate_parallel_work(Space* sp, MemRegion mr,
  59                                      OopsInGenClosure* cl,
  60                                      CardTableRS* ct,
  61                                      uint n_threads) {
  62   assert(n_threads > 0, "expected n_threads > 0");
  63   assert(n_threads <= ParallelGCThreads,
  64          "n_threads: %u > ParallelGCThreads: %u", n_threads, ParallelGCThreads);
  65 
  66   // Make sure the LNC array is valid for the space.
  67   jbyte**   lowest_non_clean;
  68   uintptr_t lowest_non_clean_base_chunk_index;
  69   size_t    lowest_non_clean_chunk_size;
  70   get_LNC_array_for_space(sp, lowest_non_clean,
  71                           lowest_non_clean_base_chunk_index,
  72                           lowest_non_clean_chunk_size);
  73 
  74   uint n_strides = n_threads * ParGCStridesPerThread;
  75   SequentialSubTasksDone* pst = sp->par_seq_tasks();
  76   // Sets the condition for completion of the subtask (how many threads
  77   // need to finish in order to be done).
  78   pst->set_n_threads(n_threads);
  79   pst->set_n_tasks(n_strides);
  80 
  81   uint stride = 0;
  82   while (!pst->is_task_claimed(/* reference */ stride)) {
  83     process_stride(sp, mr, stride, n_strides,
  84                    cl, ct,
  85                    lowest_non_clean,
  86                    lowest_non_clean_base_chunk_index,
  87                    lowest_non_clean_chunk_size);
  88   }
  89   if (pst->all_tasks_completed()) {
  90     // Clear lowest_non_clean array for next time.
  91     intptr_t first_chunk_index = addr_to_chunk_index(mr.start());
  92     uintptr_t last_chunk_index  = addr_to_chunk_index(mr.last());
  93     for (uintptr_t ch = first_chunk_index; ch <= last_chunk_index; ch++) {
  94       intptr_t ind = ch - lowest_non_clean_base_chunk_index;
  95       assert(0 <= ind && ind < (intptr_t)lowest_non_clean_chunk_size,
  96              "Bounds error");
  97       lowest_non_clean[ind] = NULL;
  98     }
  99   }
 100 }
 101 
 102 void
 103 CMSCardTable::
 104 process_stride(Space* sp,
 105                MemRegion used,
 106                jint stride, int n_strides,
 107                OopsInGenClosure* cl,
 108                CardTableRS* ct,
 109                jbyte** lowest_non_clean,
 110                uintptr_t lowest_non_clean_base_chunk_index,
 111                size_t    lowest_non_clean_chunk_size) {
 112   // We go from higher to lower addresses here; it wouldn't help that much
 113   // because of the strided parallelism pattern used here.
 114 
 115   // Find the first card address of the first chunk in the stride that is
 116   // at least "bottom" of the used region.
 117   jbyte*    start_card  = byte_for(used.start());
 118   jbyte*    end_card    = byte_after(used.last());
 119   uintptr_t start_chunk = addr_to_chunk_index(used.start());
 120   uintptr_t start_chunk_stride_num = start_chunk % n_strides;
 121   jbyte* chunk_card_start;
 122 
 123   if ((uintptr_t)stride >= start_chunk_stride_num) {
 124     chunk_card_start = (jbyte*)(start_card +
 125                                 (stride - start_chunk_stride_num) *
 126                                 ParGCCardsPerStrideChunk);
 127   } else {
 128     // Go ahead to the next chunk group boundary, then to the requested stride.
 129     chunk_card_start = (jbyte*)(start_card +
 130                                 (n_strides - start_chunk_stride_num + stride) *
 131                                 ParGCCardsPerStrideChunk);
 132   }
 133 
 134   while (chunk_card_start < end_card) {
 135     // Even though we go from lower to higher addresses below, the
 136     // strided parallelism can interleave the actual processing of the
 137     // dirty pages in various ways. For a specific chunk within this
 138     // stride, we take care to avoid double scanning or missing a card
 139     // by suitably initializing the "min_done" field in process_chunk_boundaries()
 140     // below, together with the dirty region extension accomplished in
 141     // DirtyCardToOopClosure::do_MemRegion().
 142     jbyte*    chunk_card_end = chunk_card_start + ParGCCardsPerStrideChunk;
 143     // Invariant: chunk_mr should be fully contained within the "used" region.
 144     MemRegion chunk_mr       = MemRegion(addr_for(chunk_card_start),
 145                                          chunk_card_end >= end_card ?
 146                                            used.end() : addr_for(chunk_card_end));
 147     assert(chunk_mr.word_size() > 0, "[chunk_card_start > used_end)");
 148     assert(used.contains(chunk_mr), "chunk_mr should be subset of used");
 149 
 150     // This function is used by the parallel card table iteration.
 151     const bool parallel = true;
 152 
 153     DirtyCardToOopClosure* dcto_cl = sp->new_dcto_cl(cl, precision(),
 154                                                      cl->gen_boundary(),
 155                                                      parallel);
 156     ClearNoncleanCardWrapper clear_cl(dcto_cl, ct, parallel);
 157 
 158 
 159     // Process the chunk.
 160     process_chunk_boundaries(sp,
 161                              dcto_cl,
 162                              chunk_mr,
 163                              used,
 164                              lowest_non_clean,
 165                              lowest_non_clean_base_chunk_index,
 166                              lowest_non_clean_chunk_size);
 167 
 168     // We want the LNC array updates above in process_chunk_boundaries
 169     // to be visible before any of the card table value changes as a
 170     // result of the dirty card iteration below.
 171     OrderAccess::storestore();
 172 
 173     // We want to clear the cards: clear_cl here does the work of finding
 174     // contiguous dirty ranges of cards to process and clear.
 175     clear_cl.do_MemRegion(chunk_mr);
 176 
 177     // Find the next chunk of the stride.
 178     chunk_card_start += ParGCCardsPerStrideChunk * n_strides;
 179   }
 180 }
 181 
 182 void
 183 CMSCardTable::
 184 process_chunk_boundaries(Space* sp,
 185                          DirtyCardToOopClosure* dcto_cl,
 186                          MemRegion chunk_mr,
 187                          MemRegion used,
 188                          jbyte** lowest_non_clean,
 189                          uintptr_t lowest_non_clean_base_chunk_index,
 190                          size_t    lowest_non_clean_chunk_size)
 191 {
 192   // We must worry about non-array objects that cross chunk boundaries,
 193   // because such objects are both precisely and imprecisely marked:
 194   // .. if the head of such an object is dirty, the entire object
 195   //    needs to be scanned, under the interpretation that this
 196   //    was an imprecise mark
 197   // .. if the head of such an object is not dirty, we can assume
 198   //    precise marking and it's efficient to scan just the dirty
 199   //    cards.
 200   // In either case, each scanned reference must be scanned precisely
 201   // once so as to avoid cloning of a young referent. For efficiency,
 202   // our closures depend on this property and do not protect against
 203   // double scans.
 204 
 205   uintptr_t start_chunk_index = addr_to_chunk_index(chunk_mr.start());
 206   assert(start_chunk_index >= lowest_non_clean_base_chunk_index, "Bounds error.");
 207   uintptr_t cur_chunk_index   = start_chunk_index - lowest_non_clean_base_chunk_index;
 208 
 209   // First, set "our" lowest_non_clean entry, which would be
 210   // used by the thread scanning an adjoining left chunk with
 211   // a non-array object straddling the mutual boundary.
 212   // Find the object that spans our boundary, if one exists.
 213   // first_block is the block possibly straddling our left boundary.
 214   HeapWord* first_block = sp->block_start(chunk_mr.start());
 215   assert((chunk_mr.start() != used.start()) || (first_block == chunk_mr.start()),
 216          "First chunk should always have a co-initial block");
 217   // Does the block straddle the chunk's left boundary, and is it
 218   // a non-array object?
 219   if (first_block < chunk_mr.start()        // first block straddles left bdry
 220       && sp->block_is_obj(first_block)      // first block is an object
 221       && !(oop(first_block)->is_objArray()  // first block is not an array (arrays are precisely dirtied)
 222            || oop(first_block)->is_typeArray())) {
 223     // Find our least non-clean card, so that a left neighbor
 224     // does not scan an object straddling the mutual boundary
 225     // too far to the right, and attempt to scan a portion of
 226     // that object twice.
 227     jbyte* first_dirty_card = NULL;
 228     jbyte* last_card_of_first_obj =
 229         byte_for(first_block + sp->block_size(first_block) - 1);
 230     jbyte* first_card_of_cur_chunk = byte_for(chunk_mr.start());
 231     jbyte* last_card_of_cur_chunk = byte_for(chunk_mr.last());
 232     jbyte* last_card_to_check =
 233       (jbyte*) MIN2((intptr_t) last_card_of_cur_chunk,
 234                     (intptr_t) last_card_of_first_obj);
 235     // Note that this does not need to go beyond our last card
 236     // if our first object completely straddles this chunk.
 237     for (jbyte* cur = first_card_of_cur_chunk;
 238          cur <= last_card_to_check; cur++) {
 239       jbyte val = *cur;
 240       if (card_will_be_scanned(val)) {
 241         first_dirty_card = cur; break;
 242       } else {
 243         assert(!card_may_have_been_dirty(val), "Error");
 244       }
 245     }
 246     if (first_dirty_card != NULL) {
 247       assert(cur_chunk_index < lowest_non_clean_chunk_size, "Bounds error.");
 248       assert(lowest_non_clean[cur_chunk_index] == NULL,
 249              "Write exactly once : value should be stable hereafter for this round");
 250       lowest_non_clean[cur_chunk_index] = first_dirty_card;
 251     }
 252   } else {
 253     // In this case we can help our neighbor by just asking them
 254     // to stop at our first card (even though it may not be dirty).
 255     assert(lowest_non_clean[cur_chunk_index] == NULL, "Write once : value should be stable hereafter");
 256     jbyte* first_card_of_cur_chunk = byte_for(chunk_mr.start());
 257     lowest_non_clean[cur_chunk_index] = first_card_of_cur_chunk;
 258   }
 259 
 260   // Next, set our own max_to_do, which will strictly/exclusively bound
 261   // the highest address that we will scan past the right end of our chunk.
 262   HeapWord* max_to_do = NULL;
 263   if (chunk_mr.end() < used.end()) {
 264     // This is not the last chunk in the used region.
 265     // What is our last block? We check the first block of
 266     // the next (right) chunk rather than strictly check our last block
 267     // because it's potentially more efficient to do so.
 268     HeapWord* const last_block = sp->block_start(chunk_mr.end());
 269     assert(last_block <= chunk_mr.end(), "In case this property changes.");
 270     if ((last_block == chunk_mr.end())     // our last block does not straddle boundary
 271         || !sp->block_is_obj(last_block)   // last_block isn't an object
 272         || oop(last_block)->is_objArray()  // last_block is an array (precisely marked)
 273         || oop(last_block)->is_typeArray()) {
 274       max_to_do = chunk_mr.end();
 275     } else {
 276       assert(last_block < chunk_mr.end(), "Tautology");
 277       // It is a non-array object that straddles the right boundary of this chunk.
 278       // last_obj_card is the card corresponding to the start of the last object
 279       // in the chunk.  Note that the last object may not start in
 280       // the chunk.
 281       jbyte* const last_obj_card = byte_for(last_block);
 282       const jbyte val = *last_obj_card;
 283       if (!card_will_be_scanned(val)) {
 284         assert(!card_may_have_been_dirty(val), "Error");
 285         // The card containing the head is not dirty.  Any marks on
 286         // subsequent cards still in this chunk must have been made
 287         // precisely; we can cap processing at the end of our chunk.
 288         max_to_do = chunk_mr.end();
 289       } else {
 290         // The last object must be considered dirty, and extends onto the
 291         // following chunk.  Look for a dirty card in that chunk that will
 292         // bound our processing.
 293         jbyte* limit_card = NULL;
 294         const size_t last_block_size = sp->block_size(last_block);
 295         jbyte* const last_card_of_last_obj =
 296           byte_for(last_block + last_block_size - 1);
 297         jbyte* const first_card_of_next_chunk = byte_for(chunk_mr.end());
 298         // This search potentially goes a long distance looking
 299         // for the next card that will be scanned, terminating
 300         // at the end of the last_block, if no earlier dirty card
 301         // is found.
 302         assert(byte_for(chunk_mr.end()) - byte_for(chunk_mr.start()) == ParGCCardsPerStrideChunk,
 303                "last card of next chunk may be wrong");
 304         for (jbyte* cur = first_card_of_next_chunk;
 305              cur <= last_card_of_last_obj; cur++) {
 306           const jbyte val = *cur;
 307           if (card_will_be_scanned(val)) {
 308             limit_card = cur; break;
 309           } else {
 310             assert(!card_may_have_been_dirty(val), "Error: card can't be skipped");
 311           }
 312         }
 313         if (limit_card != NULL) {
 314           max_to_do = addr_for(limit_card);
 315           assert(limit_card != NULL && max_to_do != NULL, "Error");
 316         } else {
 317           // The following is a pessimistic value, because it's possible
 318           // that a dirty card on a subsequent chunk has been cleared by
 319           // the time we get to look at it; we'll correct for that further below,
 320           // using the LNC array which records the least non-clean card
 321           // before cards were cleared in a particular chunk.
 322           limit_card = last_card_of_last_obj;
 323           max_to_do = last_block + last_block_size;
 324           assert(limit_card != NULL && max_to_do != NULL, "Error");
 325         }
 326         assert(0 < cur_chunk_index+1 && cur_chunk_index+1 < lowest_non_clean_chunk_size,
 327                "Bounds error.");
 328         // It is possible that a dirty card for the last object may have been
 329         // cleared before we had a chance to examine it. In that case, the value
 330         // will have been logged in the LNC for that chunk.
 331         // We need to examine as many chunks to the right as this object
 332         // covers. However, we need to bound this checking to the largest
 333         // entry in the LNC array: this is because the heap may expand
 334         // after the LNC array has been created but before we reach this point,
 335         // and the last block in our chunk may have been expanded to include
 336         // the expansion delta (and possibly subsequently allocated from, so
 337         // it wouldn't be sufficient to check whether that last block was
 338         // or was not an object at this point).
 339         uintptr_t last_chunk_index_to_check = addr_to_chunk_index(last_block + last_block_size - 1)
 340                                               - lowest_non_clean_base_chunk_index;
 341         const uintptr_t last_chunk_index    = addr_to_chunk_index(used.last())
 342                                               - lowest_non_clean_base_chunk_index;
 343         if (last_chunk_index_to_check > last_chunk_index) {
 344           assert(last_block + last_block_size > used.end(),
 345                  "Inconsistency detected: last_block [" PTR_FORMAT "," PTR_FORMAT "]"
 346                  " does not exceed used.end() = " PTR_FORMAT ","
 347                  " yet last_chunk_index_to_check " INTPTR_FORMAT
 348                  " exceeds last_chunk_index " INTPTR_FORMAT,
 349                  p2i(last_block), p2i(last_block + last_block_size),
 350                  p2i(used.end()),
 351                  last_chunk_index_to_check, last_chunk_index);
 352           assert(sp->used_region().end() > used.end(),
 353                  "Expansion did not happen: "
 354                  "[" PTR_FORMAT "," PTR_FORMAT ") -> [" PTR_FORMAT "," PTR_FORMAT ")",
 355                  p2i(sp->used_region().start()), p2i(sp->used_region().end()),
 356                  p2i(used.start()), p2i(used.end()));
 357           last_chunk_index_to_check = last_chunk_index;
 358         }
 359         for (uintptr_t lnc_index = cur_chunk_index + 1;
 360              lnc_index <= last_chunk_index_to_check;
 361              lnc_index++) {
 362           jbyte* lnc_card = lowest_non_clean[lnc_index];
 363           if (lnc_card != NULL) {
 364             // we can stop at the first non-NULL entry we find
 365             if (lnc_card <= limit_card) {
 366               limit_card = lnc_card;
 367               max_to_do = addr_for(limit_card);
 368               assert(limit_card != NULL && max_to_do != NULL, "Error");
 369             }
 370             // In any case, we break now
 371             break;
 372           }  // else continue to look for a non-NULL entry if any
 373         }
 374         assert(limit_card != NULL && max_to_do != NULL, "Error");
 375       }
 376       assert(max_to_do != NULL, "OOPS 1 !");
 377     }
 378     assert(max_to_do != NULL, "OOPS 2!");
 379   } else {
 380     max_to_do = used.end();
 381   }
 382   assert(max_to_do != NULL, "OOPS 3!");
 383   // Now we can set the closure we're using so it doesn't to beyond
 384   // max_to_do.
 385   dcto_cl->set_min_done(max_to_do);
 386 #ifndef PRODUCT
 387   dcto_cl->set_last_bottom(max_to_do);
 388 #endif
 389 }
 390 
 391 void
 392 CMSCardTable::
 393 get_LNC_array_for_space(Space* sp,
 394                         jbyte**& lowest_non_clean,
 395                         uintptr_t& lowest_non_clean_base_chunk_index,
 396                         size_t& lowest_non_clean_chunk_size) {
 397 
 398   int       i        = find_covering_region_containing(sp->bottom());
 399   MemRegion covered  = _covered[i];
 400   size_t    n_chunks = chunks_to_cover(covered);
 401 
 402   // Only the first thread to obtain the lock will resize the
 403   // LNC array for the covered region.  Any later expansion can't affect
 404   // the used_at_save_marks region.
 405   // (I observed a bug in which the first thread to execute this would
 406   // resize, and then it would cause "expand_and_allocate" that would
 407   // increase the number of chunks in the covered region.  Then a second
 408   // thread would come and execute this, see that the size didn't match,
 409   // and free and allocate again.  So the first thread would be using a
 410   // freed "_lowest_non_clean" array.)
 411 
 412   // Do a dirty read here. If we pass the conditional then take the rare
 413   // event lock and do the read again in case some other thread had already
 414   // succeeded and done the resize.
 415   int cur_collection = CMSHeap::heap()->total_collections();
 416   // Updated _last_LNC_resizing_collection[i] must not be visible before
 417   // _lowest_non_clean and friends are visible. Therefore use acquire/release
 418   // to guarantee this on non TSO architecures.
 419   if (OrderAccess::load_acquire(&_last_LNC_resizing_collection[i]) != cur_collection) {
 420     MutexLocker x(ParGCRareEvent_lock);
 421     // This load_acquire is here for clarity only. The MutexLocker already fences.
 422     if (OrderAccess::load_acquire(&_last_LNC_resizing_collection[i]) != cur_collection) {
 423       if (_lowest_non_clean[i] == NULL ||
 424           n_chunks != _lowest_non_clean_chunk_size[i]) {
 425 
 426         // Should we delete the old?
 427         if (_lowest_non_clean[i] != NULL) {
 428           assert(n_chunks != _lowest_non_clean_chunk_size[i],
 429                  "logical consequence");
 430           FREE_C_HEAP_ARRAY(CardPtr, _lowest_non_clean[i]);
 431           _lowest_non_clean[i] = NULL;
 432         }
 433         // Now allocate a new one if necessary.
 434         if (_lowest_non_clean[i] == NULL) {
 435           _lowest_non_clean[i]                  = NEW_C_HEAP_ARRAY(CardPtr, n_chunks, mtGC);
 436           _lowest_non_clean_chunk_size[i]       = n_chunks;
 437           _lowest_non_clean_base_chunk_index[i] = addr_to_chunk_index(covered.start());
 438           for (int j = 0; j < (int)n_chunks; j++)
 439             _lowest_non_clean[i][j] = NULL;
 440         }
 441       }
 442       // Make sure this gets visible only after _lowest_non_clean* was initialized
 443       OrderAccess::release_store(&_last_LNC_resizing_collection[i], cur_collection);
 444     }
 445   }
 446   // In any case, now do the initialization.
 447   lowest_non_clean                  = _lowest_non_clean[i];
 448   lowest_non_clean_base_chunk_index = _lowest_non_clean_base_chunk_index[i];
 449   lowest_non_clean_chunk_size       = _lowest_non_clean_chunk_size[i];
 450 }
 451 
 452 #ifdef ASSERT
 453 void CMSCardTable::verify_used_region_at_save_marks(Space* sp) const {
 454   MemRegion ur    = sp->used_region();
 455   MemRegion urasm = sp->used_region_at_save_marks();
 456 
 457   if (!ur.contains(urasm)) {
 458     log_warning(gc)("CMS+ParNew: Did you forget to call save_marks()? "
 459                     "[" PTR_FORMAT ", " PTR_FORMAT ") is not contained in "
 460                     "[" PTR_FORMAT ", " PTR_FORMAT ")",
 461                     p2i(urasm.start()), p2i(urasm.end()), p2i(ur.start()), p2i(ur.end()));
 462     MemRegion ur2 = sp->used_region();
 463     MemRegion urasm2 = sp->used_region_at_save_marks();
 464     if (!ur.equals(ur2)) {
 465       log_warning(gc)("CMS+ParNew: Flickering used_region()!!");
 466     }
 467     if (!urasm.equals(urasm2)) {
 468       log_warning(gc)("CMS+ParNew: Flickering used_region_at_save_marks()!!");
 469     }
 470     ShouldNotReachHere();
 471   }
 472 }
 473 #endif // ASSERT
--- EOF ---