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.hpp"

    38 #include "runtime/vmThread.hpp"

    39 

    40 CMSCardTable::CMSCardTable(MemRegion whole_heap) :

    41     CardTableRS(whole_heap, CMSPrecleaningEnabled /* 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   CardValue** 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->try_claim_task(/* 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                CardValue** 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   CardValue* start_card  = byte_for(used.start());

   118   CardValue* 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   CardValue* chunk_card_start;

   122 

   123   if ((uintptr_t)stride >= start_chunk_stride_num) {

   124     chunk_card_start = (start_card +

   125                         (stride - start_chunk_stride_num) * ParGCCardsPerStrideChunk);

   126   } else {

   127     // Go ahead to the next chunk group boundary, then to the requested stride.

   128     chunk_card_start = (start_card +

   129                         (n_strides - start_chunk_stride_num + stride) * ParGCCardsPerStrideChunk);

   130   }

   131 

   132   while (chunk_card_start < end_card) {

   133     // Even though we go from lower to higher addresses below, the

   134     // strided parallelism can interleave the actual processing of the

   135     // dirty pages in various ways. For a specific chunk within this

   136     // stride, we take care to avoid double scanning or missing a card

   137     // by suitably initializing the "min_done" field in process_chunk_boundaries()

   138     // below, together with the dirty region extension accomplished in

   139     // DirtyCardToOopClosure::do_MemRegion().

   140     CardValue* chunk_card_end = chunk_card_start + ParGCCardsPerStrideChunk;

   141     // Invariant: chunk_mr should be fully contained within the "used" region.

   142     MemRegion chunk_mr = MemRegion(addr_for(chunk_card_start),

   143                                    chunk_card_end >= end_card ?

   144                                    used.end() : addr_for(chunk_card_end));

   145     assert(chunk_mr.word_size() > 0, "[chunk_card_start > used_end)");

   146     assert(used.contains(chunk_mr), "chunk_mr should be subset of used");

   147 

   148     // This function is used by the parallel card table iteration.

   149     const bool parallel = true;

   150 

   151     DirtyCardToOopClosure* dcto_cl = sp->new_dcto_cl(cl, precision(),

   152                                                      cl->gen_boundary(),

   153                                                      parallel);

   154     ClearNoncleanCardWrapper clear_cl(dcto_cl, ct, parallel);

   155 

   156 

   157     // Process the chunk.

   158     process_chunk_boundaries(sp,

   159                              dcto_cl,

   160                              chunk_mr,

   161                              used,

   162                              lowest_non_clean,

   163                              lowest_non_clean_base_chunk_index,

   164                              lowest_non_clean_chunk_size);

   165 

   166     // We want the LNC array updates above in process_chunk_boundaries

   167     // to be visible before any of the card table value changes as a

   168     // result of the dirty card iteration below.

   169     OrderAccess::storestore();

   170 

   171     // We want to clear the cards: clear_cl here does the work of finding

   172     // contiguous dirty ranges of cards to process and clear.

   173     clear_cl.do_MemRegion(chunk_mr);

   174 

   175     // Find the next chunk of the stride.

   176     chunk_card_start += ParGCCardsPerStrideChunk * n_strides;

   177   }

   178 }

   179 

   180 void

   181 CMSCardTable::

   182 process_chunk_boundaries(Space* sp,

   183                          DirtyCardToOopClosure* dcto_cl,

   184                          MemRegion chunk_mr,

   185                          MemRegion used,

   186                          CardValue** lowest_non_clean,

   187                          uintptr_t lowest_non_clean_base_chunk_index,

   188                          size_t    lowest_non_clean_chunk_size)

   189 {

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

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

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

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

   194   //    was an imprecise mark

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

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

   197   //    cards.

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

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

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

   201   // double scans.

   202 

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

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

   205   uintptr_t cur_chunk_index   = start_chunk_index - lowest_non_clean_base_chunk_index;

   206 

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

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

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

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

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

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

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

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

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

   216   // a non-array object?

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

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

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

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

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

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

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

   224     // that object twice.

   225     CardValue* first_dirty_card = NULL;

   226     CardValue* last_card_of_first_obj =

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

   228     CardValue* first_card_of_cur_chunk = byte_for(chunk_mr.start());

   229     CardValue* last_card_of_cur_chunk = byte_for(chunk_mr.last());

   230     CardValue* last_card_to_check = MIN2(last_card_of_cur_chunk, last_card_of_first_obj);

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

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

   233     for (CardValue* cur = first_card_of_cur_chunk;

   234          cur <= last_card_to_check; cur++) {

   235       CardValue val = *cur;

   236       if (card_will_be_scanned(val)) {

   237         first_dirty_card = cur;

   238         break;

   239       } else {

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

   241       }

   242     }

   243     if (first_dirty_card != NULL) {

   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     }

   249   } else {

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

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

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

   253     CardValue* first_card_of_cur_chunk = byte_for(chunk_mr.start());

   254     lowest_non_clean[cur_chunk_index] = first_card_of_cur_chunk;

   255   }

   256 

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

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

   259   HeapWord* max_to_do = NULL;

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

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

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

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

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

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

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

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

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

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

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

   271       max_to_do = chunk_mr.end();

   272     } else {

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

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

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

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

   277       // the chunk.

   278       CardValue* const last_obj_card = byte_for(last_block);

   279       const CardValue val = *last_obj_card;

   280       if (!card_will_be_scanned(val)) {

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

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

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

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

   285         max_to_do = chunk_mr.end();

   286       } else {

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

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

   289         // bound our processing.

   290         CardValue* limit_card = NULL;

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

   292         CardValue* const last_card_of_last_obj =

   293           byte_for(last_block + last_block_size - 1);

   294         CardValue* const first_card_of_next_chunk = byte_for(chunk_mr.end());

   295         // This search potentially goes a long distance looking

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

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

   298         // is found.

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

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

   301         for (CardValue* cur = first_card_of_next_chunk;

   302              cur <= last_card_of_last_obj; cur++) {

   303           const CardValue val = *cur;

   304           if (card_will_be_scanned(val)) {

   305             limit_card = cur; break;

   306           } else {

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

   308           }

   309         }

   310         if (limit_card != NULL) {

   311           max_to_do = addr_for(limit_card);

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

   313         } else {

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

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

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

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

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

   319           limit_card = last_card_of_last_obj;

   320           max_to_do = last_block + last_block_size;

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

   322         }

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

   324                "Bounds error.");

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

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

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

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

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

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

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

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

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

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

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

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

   337                                               - lowest_non_clean_base_chunk_index;

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

   339                                               - lowest_non_clean_base_chunk_index;

   340         if (last_chunk_index_to_check > last_chunk_index) {

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

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

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

   344                  " yet last_chunk_index_to_check " INTPTR_FORMAT

   345                  " exceeds last_chunk_index " INTPTR_FORMAT,

   346                  p2i(last_block), p2i(last_block + last_block_size),

   347                  p2i(used.end()),

   348                  last_chunk_index_to_check, last_chunk_index);

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

   350                  "Expansion did not happen: "

   351                  "[" PTR_FORMAT "," PTR_FORMAT ") -> [" PTR_FORMAT "," PTR_FORMAT ")",

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

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

   354           last_chunk_index_to_check = last_chunk_index;

   355         }

   356         for (uintptr_t lnc_index = cur_chunk_index + 1;

   357              lnc_index <= last_chunk_index_to_check;

   358              lnc_index++) {

   359           CardValue* lnc_card = lowest_non_clean[lnc_index];

   360           if (lnc_card != NULL) {

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

   362             if (lnc_card <= limit_card) {

   363               limit_card = lnc_card;

   364               max_to_do = addr_for(limit_card);

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

   366             }

   367             // In any case, we break now

   368             break;

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

   370         }

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

   372       }

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

   374     }

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

   376   } else {

   377     max_to_do = used.end();

   378   }

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

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

   381   // max_to_do.

   382   dcto_cl->set_min_done(max_to_do);

   383 #ifndef PRODUCT

   384   dcto_cl->set_last_bottom(max_to_do);

   385 #endif

   386 }

   387 

   388 void

   389 CMSCardTable::

   390 get_LNC_array_for_space(Space* sp,

   391                         CardValue**& lowest_non_clean,

   392                         uintptr_t& lowest_non_clean_base_chunk_index,

   393                         size_t& lowest_non_clean_chunk_size) {

   394 

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

   396   MemRegion covered  = _covered[i];

   397   size_t    n_chunks = chunks_to_cover(covered);

   398 

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

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

   401   // the used_at_save_marks region.

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

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

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

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

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

   407   // freed "_lowest_non_clean" array.)

   408 

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

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

   411   // succeeded and done the resize.

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

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

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

   415   // to guarantee this on non TSO architecures.

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

   417     MutexLocker x(ParGCRareEvent_lock);

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

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

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

   421           n_chunks != _lowest_non_clean_chunk_size[i]) {

   422 

   423         // Should we delete the old?

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

   425           assert(n_chunks != _lowest_non_clean_chunk_size[i],

   426                  "logical consequence");

   427           FREE_C_HEAP_ARRAY(CardPtr, _lowest_non_clean[i]);

   428           _lowest_non_clean[i] = NULL;

   429         }

   430         // Now allocate a new one if necessary.

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

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

   433           _lowest_non_clean_chunk_size[i]       = n_chunks;

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

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

   436             _lowest_non_clean[i][j] = NULL;

   437         }

   438       }

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

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

   441     }

   442   }

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

   444   lowest_non_clean                  = _lowest_non_clean[i];

   445   lowest_non_clean_base_chunk_index = _lowest_non_clean_base_chunk_index[i];

   446   lowest_non_clean_chunk_size       = _lowest_non_clean_chunk_size[i];

   447 }

   448 

   449 #ifdef ASSERT

   450 void CMSCardTable::verify_used_region_at_save_marks(Space* sp) const {

   451   MemRegion ur    = sp->used_region();

   452   MemRegion urasm = sp->used_region_at_save_marks();

   453 

   454   if (!ur.contains(urasm)) {

   455     log_warning(gc)("CMS+ParNew: Did you forget to call save_marks()? "

   456                     "[" PTR_FORMAT ", " PTR_FORMAT ") is not contained in "

   457                     "[" PTR_FORMAT ", " PTR_FORMAT ")",

   458                     p2i(urasm.start()), p2i(urasm.end()), p2i(ur.start()), p2i(ur.end()));

   459     MemRegion ur2 = sp->used_region();

   460     MemRegion urasm2 = sp->used_region_at_save_marks();

   461     if (!ur.equals(ur2)) {

   462       log_warning(gc)("CMS+ParNew: Flickering used_region()!!");

   463     }

   464     if (!urasm.equals(urasm2)) {

   465       log_warning(gc)("CMS+ParNew: Flickering used_region_at_save_marks()!!");

   466     }

   467     ShouldNotReachHere();

   468   }

   469 }

   470 #endif // ASSERT