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/cmsHeap.hpp"

    27 #include "gc/shared/cardTableModRefBS.hpp"

    28 #include "gc/shared/cardTableRS.hpp"

    29 #include "gc/shared/collectedHeap.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 CardTableRS::

    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 CardTableRS::

    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 void

   165 CardTableRS::

   166 process_chunk_boundaries(Space* sp,

   167                          DirtyCardToOopClosure* dcto_cl,

   168                          MemRegion chunk_mr,

   169                          MemRegion used,

   170                          jbyte** lowest_non_clean,

   171                          uintptr_t lowest_non_clean_base_chunk_index,

   172                          size_t    lowest_non_clean_chunk_size)

   173 {

   174   // We must worry about non-array objects that cross chunk boundaries,

   175   // because such objects are both precisely and imprecisely marked:

   176   // .. if the head of such an object is dirty, the entire object

   177   //    needs to be scanned, under the interpretation that this

   178   //    was an imprecise mark

   179   // .. if the head of such an object is not dirty, we can assume

   180   //    precise marking and it's efficient to scan just the dirty

   181   //    cards.

   182   // In either case, each scanned reference must be scanned precisely

   183   // once so as to avoid cloning of a young referent. For efficiency,

   184   // our closures depend on this property and do not protect against

   185   // double scans.

   186 

   187   uintptr_t start_chunk_index = addr_to_chunk_index(chunk_mr.start());

   188   assert(start_chunk_index >= lowest_non_clean_base_chunk_index, "Bounds error.");

   189   uintptr_t cur_chunk_index   = start_chunk_index - lowest_non_clean_base_chunk_index;

   190 

   191   // First, set "our" lowest_non_clean entry, which would be

   192   // used by the thread scanning an adjoining left chunk with

   193   // a non-array object straddling the mutual boundary.

   194   // Find the object that spans our boundary, if one exists.

   195   // first_block is the block possibly straddling our left boundary.

   196   HeapWord* first_block = sp->block_start(chunk_mr.start());

   197   assert((chunk_mr.start() != used.start()) || (first_block == chunk_mr.start()),

   198          "First chunk should always have a co-initial block");

   199   // Does the block straddle the chunk's left boundary, and is it

   200   // a non-array object?

   201   if (first_block < chunk_mr.start()        // first block straddles left bdry

   202       && sp->block_is_obj(first_block)      // first block is an object

   203       && !(oop(first_block)->is_objArray()  // first block is not an array (arrays are precisely dirtied)

   204            || oop(first_block)->is_typeArray())) {

   205     // Find our least non-clean card, so that a left neighbor

   206     // does not scan an object straddling the mutual boundary

   207     // too far to the right, and attempt to scan a portion of

   208     // that object twice.

   209     jbyte* first_dirty_card = NULL;

   210     jbyte* last_card_of_first_obj =

   211         byte_for(first_block + sp->block_size(first_block) - 1);

   212     jbyte* first_card_of_cur_chunk = byte_for(chunk_mr.start());

   213     jbyte* last_card_of_cur_chunk = byte_for(chunk_mr.last());

   214     jbyte* last_card_to_check =

   215       (jbyte*) MIN2((intptr_t) last_card_of_cur_chunk,

   216                     (intptr_t) last_card_of_first_obj);

   217     // Note that this does not need to go beyond our last card

   218     // if our first object completely straddles this chunk.

   219     for (jbyte* cur = first_card_of_cur_chunk;

   220          cur <= last_card_to_check; cur++) {

   221       jbyte val = *cur;

   222       if (card_will_be_scanned(val)) {

   223         first_dirty_card = cur; break;

   224       } else {

   225         assert(!card_may_have_been_dirty(val), "Error");

   226       }

   227     }

   228     if (first_dirty_card != NULL) {

   229       assert(cur_chunk_index < lowest_non_clean_chunk_size, "Bounds error.");

   230       assert(lowest_non_clean[cur_chunk_index] == NULL,

   231              "Write exactly once : value should be stable hereafter for this round");

   232       lowest_non_clean[cur_chunk_index] = first_dirty_card;

   233     }

   234   } else {

   235     // In this case we can help our neighbor by just asking them

   236     // to stop at our first card (even though it may not be dirty).

   237     assert(lowest_non_clean[cur_chunk_index] == NULL, "Write once : value should be stable hereafter");

   238     jbyte* first_card_of_cur_chunk = byte_for(chunk_mr.start());

   239     lowest_non_clean[cur_chunk_index] = first_card_of_cur_chunk;

   240   }

   241 

   242   // Next, set our own max_to_do, which will strictly/exclusively bound

   243   // the highest address that we will scan past the right end of our chunk.

   244   HeapWord* max_to_do = NULL;

   245   if (chunk_mr.end() < used.end()) {

   246     // This is not the last chunk in the used region.

   247     // What is our last block? We check the first block of

   248     // the next (right) chunk rather than strictly check our last block

   249     // because it's potentially more efficient to do so.

   250     HeapWord* const last_block = sp->block_start(chunk_mr.end());

   251     assert(last_block <= chunk_mr.end(), "In case this property changes.");

   252     if ((last_block == chunk_mr.end())     // our last block does not straddle boundary

   253         || !sp->block_is_obj(last_block)   // last_block isn't an object

   254         || oop(last_block)->is_objArray()  // last_block is an array (precisely marked)

   255         || oop(last_block)->is_typeArray()) {

   256       max_to_do = chunk_mr.end();

   257     } else {

   258       assert(last_block < chunk_mr.end(), "Tautology");

   259       // It is a non-array object that straddles the right boundary of this chunk.

   260       // last_obj_card is the card corresponding to the start of the last object

   261       // in the chunk.  Note that the last object may not start in

   262       // the chunk.

   263       jbyte* const last_obj_card = byte_for(last_block);

   264       const jbyte val = *last_obj_card;

   265       if (!card_will_be_scanned(val)) {

   266         assert(!card_may_have_been_dirty(val), "Error");

   267         // The card containing the head is not dirty.  Any marks on

   268         // subsequent cards still in this chunk must have been made

   269         // precisely; we can cap processing at the end of our chunk.

   270         max_to_do = chunk_mr.end();

   271       } else {

   272         // The last object must be considered dirty, and extends onto the

   273         // following chunk.  Look for a dirty card in that chunk that will

   274         // bound our processing.

   275         jbyte* limit_card = NULL;

   276         const size_t last_block_size = sp->block_size(last_block);

   277         jbyte* const last_card_of_last_obj =

   278           byte_for(last_block + last_block_size - 1);

   279         jbyte* const first_card_of_next_chunk = byte_for(chunk_mr.end());

   280         // This search potentially goes a long distance looking

   281         // for the next card that will be scanned, terminating

   282         // at the end of the last_block, if no earlier dirty card

   283         // is found.

   284         assert(byte_for(chunk_mr.end()) - byte_for(chunk_mr.start()) == ParGCCardsPerStrideChunk,

   285                "last card of next chunk may be wrong");

   286         for (jbyte* cur = first_card_of_next_chunk;

   287              cur <= last_card_of_last_obj; cur++) {

   288           const jbyte val = *cur;

   289           if (card_will_be_scanned(val)) {

   290             limit_card = cur; break;

   291           } else {

   292             assert(!card_may_have_been_dirty(val), "Error: card can't be skipped");

   293           }

   294         }

   295         if (limit_card != NULL) {

   296           max_to_do = addr_for(limit_card);

   297           assert(limit_card != NULL && max_to_do != NULL, "Error");

   298         } else {

   299           // The following is a pessimistic value, because it's possible

   300           // that a dirty card on a subsequent chunk has been cleared by

   301           // the time we get to look at it; we'll correct for that further below,

   302           // using the LNC array which records the least non-clean card

   303           // before cards were cleared in a particular chunk.

   304           limit_card = last_card_of_last_obj;

   305           max_to_do = last_block + last_block_size;

   306           assert(limit_card != NULL && max_to_do != NULL, "Error");

   307         }

   308         assert(0 < cur_chunk_index+1 && cur_chunk_index+1 < lowest_non_clean_chunk_size,

   309                "Bounds error.");

   310         // It is possible that a dirty card for the last object may have been

   311         // cleared before we had a chance to examine it. In that case, the value

   312         // will have been logged in the LNC for that chunk.

   313         // We need to examine as many chunks to the right as this object

   314         // covers. However, we need to bound this checking to the largest

   315         // entry in the LNC array: this is because the heap may expand

   316         // after the LNC array has been created but before we reach this point,

   317         // and the last block in our chunk may have been expanded to include

   318         // the expansion delta (and possibly subsequently allocated from, so

   319         // it wouldn't be sufficient to check whether that last block was

   320         // or was not an object at this point).

   321         uintptr_t last_chunk_index_to_check = addr_to_chunk_index(last_block + last_block_size - 1)

   322                                               - lowest_non_clean_base_chunk_index;

   323         const uintptr_t last_chunk_index    = addr_to_chunk_index(used.last())

   324                                               - lowest_non_clean_base_chunk_index;

   325         if (last_chunk_index_to_check > last_chunk_index) {

   326           assert(last_block + last_block_size > used.end(),

   327                  "Inconsistency detected: last_block [" PTR_FORMAT "," PTR_FORMAT "]"

   328                  " does not exceed used.end() = " PTR_FORMAT ","

   329                  " yet last_chunk_index_to_check " INTPTR_FORMAT

   330                  " exceeds last_chunk_index " INTPTR_FORMAT,

   331                  p2i(last_block), p2i(last_block + last_block_size),

   332                  p2i(used.end()),

   333                  last_chunk_index_to_check, last_chunk_index);

   334           assert(sp->used_region().end() > used.end(),

   335                  "Expansion did not happen: "

   336                  "[" PTR_FORMAT "," PTR_FORMAT ") -> [" PTR_FORMAT "," PTR_FORMAT ")",

   337                  p2i(sp->used_region().start()), p2i(sp->used_region().end()),

   338                  p2i(used.start()), p2i(used.end()));

   339           last_chunk_index_to_check = last_chunk_index;

   340         }

   341         for (uintptr_t lnc_index = cur_chunk_index + 1;

   342              lnc_index <= last_chunk_index_to_check;

   343              lnc_index++) {

   344           jbyte* lnc_card = lowest_non_clean[lnc_index];

   345           if (lnc_card != NULL) {

   346             // we can stop at the first non-NULL entry we find

   347             if (lnc_card <= limit_card) {

   348               limit_card = lnc_card;

   349               max_to_do = addr_for(limit_card);

   350               assert(limit_card != NULL && max_to_do != NULL, "Error");

   351             }

   352             // In any case, we break now

   353             break;

   354           }  // else continue to look for a non-NULL entry if any

   355         }

   356         assert(limit_card != NULL && max_to_do != NULL, "Error");

   357       }

   358       assert(max_to_do != NULL, "OOPS 1 !");

   359     }

   360     assert(max_to_do != NULL, "OOPS 2!");

   361   } else {

   362     max_to_do = used.end();

   363   }

   364   assert(max_to_do != NULL, "OOPS 3!");

   365   // Now we can set the closure we're using so it doesn't to beyond

   366   // max_to_do.

   367   dcto_cl->set_min_done(max_to_do);

   368 #ifndef PRODUCT

   369   dcto_cl->set_last_bottom(max_to_do);

   370 #endif

   371 }

   372 

   373 void

   374 CardTableRS::

   375 get_LNC_array_for_space(Space* sp,

   376                         jbyte**& lowest_non_clean,

   377                         uintptr_t& lowest_non_clean_base_chunk_index,

   378                         size_t& lowest_non_clean_chunk_size) {

   379 

   380   int       i        = find_covering_region_containing(sp->bottom());

   381   MemRegion covered  = _covered[i];

   382   size_t    n_chunks = chunks_to_cover(covered);

   383 

   384   // Only the first thread to obtain the lock will resize the

   385   // LNC array for the covered region.  Any later expansion can't affect

   386   // the used_at_save_marks region.

   387   // (I observed a bug in which the first thread to execute this would

   388   // resize, and then it would cause "expand_and_allocate" that would

   389   // increase the number of chunks in the covered region.  Then a second

   390   // thread would come and execute this, see that the size didn't match,

   391   // and free and allocate again.  So the first thread would be using a

   392   // freed "_lowest_non_clean" array.)

   393 

   394   // Do a dirty read here. If we pass the conditional then take the rare

   395   // event lock and do the read again in case some other thread had already

   396   // succeeded and done the resize.

   397   int cur_collection = CMSHeap::heap()->total_collections();

   398   // Updated _last_LNC_resizing_collection[i] must not be visible before

   399   // _lowest_non_clean and friends are visible. Therefore use acquire/release

   400   // to guarantee this on non TSO architecures.

   401   if (OrderAccess::load_acquire(&_last_LNC_resizing_collection[i]) != cur_collection) {

   402     MutexLocker x(ParGCRareEvent_lock);

   403     // This load_acquire is here for clarity only. The MutexLocker already fences.

   404     if (OrderAccess::load_acquire(&_last_LNC_resizing_collection[i]) != cur_collection) {

   405       if (_lowest_non_clean[i] == NULL ||

   406           n_chunks != _lowest_non_clean_chunk_size[i]) {

   407 

   408         // Should we delete the old?

   409         if (_lowest_non_clean[i] != NULL) {

   410           assert(n_chunks != _lowest_non_clean_chunk_size[i],

   411                  "logical consequence");

   412           FREE_C_HEAP_ARRAY(CardPtr, _lowest_non_clean[i]);

   413           _lowest_non_clean[i] = NULL;

   414         }

   415         // Now allocate a new one if necessary.

   416         if (_lowest_non_clean[i] == NULL) {

   417           _lowest_non_clean[i]                  = NEW_C_HEAP_ARRAY(CardPtr, n_chunks, mtGC);

   418           _lowest_non_clean_chunk_size[i]       = n_chunks;

   419           _lowest_non_clean_base_chunk_index[i] = addr_to_chunk_index(covered.start());

   420           for (int j = 0; j < (int)n_chunks; j++)

   421             _lowest_non_clean[i][j] = NULL;

   422         }

   423       }

   424       // Make sure this gets visible only after _lowest_non_clean* was initialized

   425       OrderAccess::release_store(&_last_LNC_resizing_collection[i], cur_collection);

   426     }

   427   }

   428   // In any case, now do the initialization.

   429   lowest_non_clean                  = _lowest_non_clean[i];

   430   lowest_non_clean_base_chunk_index = _lowest_non_clean_base_chunk_index[i];

   431   lowest_non_clean_chunk_size       = _lowest_non_clean_chunk_size[i];

   432 }