From d4d89c88c5e26be70c976a756e874af65ad6ec55 Mon Sep 17 00:00:00 2001 From: Heikki Linnakangas Date: Thu, 8 Sep 2016 14:31:31 +0300 Subject: [PATCH 1/3] Don't bother to pre-read tuples into SortTuple slots during merge. That only seems to add overhead. We're doing the same number of READTUP() calls either way, but we're spreading the memory usage over a larger area if we try to pre-read, so it doesn't seem worth it. The pre-reading can be helpful, to trigger the OS readahead of the underlying tape, because it will make the read pattern appear more sequential. But we'll fix that in the next patch, by teaching logtape.c to read in larger chunks. --- src/backend/utils/sort/tuplesort.c | 889 ++++++++++--------------------------- 1 file changed, 223 insertions(+), 666 deletions(-) diff --git a/src/backend/utils/sort/tuplesort.c b/src/backend/utils/sort/tuplesort.c index c8fbcf8..b9fb99c 100644 --- a/src/backend/utils/sort/tuplesort.c +++ b/src/backend/utils/sort/tuplesort.c @@ -162,7 +162,7 @@ bool optimize_bounded_sort = true; * The objects we actually sort are SortTuple structs. These contain * a pointer to the tuple proper (might be a MinimalTuple or IndexTuple), * which is a separate palloc chunk --- we assume it is just one chunk and - * can be freed by a simple pfree() (except during final on-the-fly merge, + * can be freed by a simple pfree() (except during merge, * when memory is used in batch). SortTuples also contain the tuple's * first key column in Datum/nullflag format, and an index integer. * @@ -203,6 +203,20 @@ typedef struct int tupindex; /* see notes above */ } SortTuple; +/* + * During merge, we use a pre-allocated set of fixed-size buffers to store + * tuples in. To avoid palloc/pfree overhead. + * + * 'nextfree' is valid when this chunk is in the free list. When in use, the + * buffer holds a tuple. + */ +#define MERGETUPLEBUFFER_SIZE 1024 + +typedef union MergeTupleBuffer +{ + union MergeTupleBuffer *nextfree; + char buffer[MERGETUPLEBUFFER_SIZE]; +} MergeTupleBuffer; /* * Possible states of a Tuplesort object. These denote the states that @@ -307,14 +321,6 @@ struct Tuplesortstate int tapenum, unsigned int len); /* - * Function to move a caller tuple. This is usually implemented as a - * memmove() shim, but function may also perform additional fix-up of - * caller tuple where needed. Batch memory support requires the movement - * of caller tuples from one location in memory to another. - */ - void (*movetup) (void *dest, void *src, unsigned int len); - - /* * This array holds the tuples now in sort memory. If we are in state * INITIAL, the tuples are in no particular order; if we are in state * SORTEDINMEM, the tuples are in final sorted order; in states BUILDRUNS @@ -332,12 +338,40 @@ struct Tuplesortstate /* * Memory for tuples is sometimes allocated in batch, rather than * incrementally. This implies that incremental memory accounting has - * been abandoned. Currently, this only happens for the final on-the-fly - * merge step. Large batch allocations can store tuples (e.g. - * IndexTuples) without palloc() fragmentation and other overhead. + * been abandoned. Currently, this happens when we start merging. + * Large batch allocations can store tuples (e.g. IndexTuples) without + * palloc() fragmentation and other overhead. + * + * For the batch memory, we use one large allocation, divided into + * MERGETUPLEBUFFER_SIZE chunks. The allocation is sized to hold + * one chunk per tape, plus one additional chunk. We need that many + * chunks to hold all the tuples kept in the heap during merge, plus + * the one we have last returned from the sort. + * + * Initially, all the chunks are kept in a linked list, in freeBufferHead. + * When a tuple is read from a tape, it is put to the next available + * chunk, if it fits. If the tuple is larger than MERGETUPLEBUFFER_SIZE, + * it is palloc'd instead. + * + * When we're done processing a tuple, we return the chunk back to the + * free list, or pfree() if it was palloc'd. We know that a tuple was + * allocated from the batch memory arena, if its pointer value is between + * mergeTupleBuffersBegin and -End. */ bool batchUsed; + char *batchMemoryBegin; /* beginning of batch memory arena */ + char *batchMemoryEnd; /* end of batch memory arena */ + MergeTupleBuffer *freeBufferHead; /* head of free list */ + + /* + * When we return a tuple to the caller that came from a tape (that is, + * in TSS_SORTEDONTAPE or TSS_FINALMERGE modes), we remember the tuple + * in 'readlasttuple', so that we can recycle the memory on next + * gettuple call. + */ + void *readlasttuple; + /* * While building initial runs, this indicates if the replacement * selection strategy is in use. When it isn't, then a simple hybrid @@ -358,42 +392,11 @@ struct Tuplesortstate */ /* - * These variables are only used during merge passes. mergeactive[i] is + * This variable is only used during merge passes. mergeactive[i] is * true if we are reading an input run from (actual) tape number i and - * have not yet exhausted that run. mergenext[i] is the memtuples index - * of the next pre-read tuple (next to be loaded into the heap) for tape - * i, or 0 if we are out of pre-read tuples. mergelast[i] similarly - * points to the last pre-read tuple from each tape. mergeavailslots[i] - * is the number of unused memtuples[] slots reserved for tape i, and - * mergeavailmem[i] is the amount of unused space allocated for tape i. - * mergefreelist and mergefirstfree keep track of unused locations in the - * memtuples[] array. The memtuples[].tupindex fields link together - * pre-read tuples for each tape as well as recycled locations in - * mergefreelist. It is OK to use 0 as a null link in these lists, because - * memtuples[0] is part of the merge heap and is never a pre-read tuple. + * have not yet exhausted that run. */ bool *mergeactive; /* active input run source? */ - int *mergenext; /* first preread tuple for each source */ - int *mergelast; /* last preread tuple for each source */ - int *mergeavailslots; /* slots left for prereading each tape */ - int64 *mergeavailmem; /* availMem for prereading each tape */ - int mergefreelist; /* head of freelist of recycled slots */ - int mergefirstfree; /* first slot never used in this merge */ - - /* - * Per-tape batch state, when final on-the-fly merge consumes memory from - * just a few large allocations. - * - * Aside from the general benefits of performing fewer individual retail - * palloc() calls, this also helps make merging more cache efficient, - * since each tape's tuples must naturally be accessed sequentially (in - * sorted order). - */ - int64 spacePerTape; /* Space (memory) for tuples (not slots) */ - char **mergetuples; /* Each tape's memory allocation */ - char **mergecurrent; /* Current offset into each tape's memory */ - char **mergetail; /* Last item's start point for each tape */ - char **mergeoverflow; /* Retail palloc() "overflow" for each tape */ /* * Variables for Algorithm D. Note that destTape is a "logical" tape @@ -481,11 +484,33 @@ struct Tuplesortstate #endif }; +/* + * Is the given tuple allocated from the batch memory arena? + */ +#define IS_MERGETUPLE_BUFFER(state, tuple) \ + ((char *) tuple >= state->batchMemoryBegin && \ + (char *) tuple < state->batchMemoryEnd) + +/* + * Return the given tuple to the batch memory free list, or free it + * if it was palloc'd. + */ +#define RELEASE_MERGETUPLE_BUFFER(state, tuple) \ + do { \ + MergeTupleBuffer *buf = (MergeTupleBuffer *) tuple; \ + \ + if (IS_MERGETUPLE_BUFFER(state, tuple)) \ + { \ + buf->nextfree = state->freeBufferHead; \ + state->freeBufferHead = buf; \ + } else \ + pfree(tuple); \ + } while(0) + #define COMPARETUP(state,a,b) ((*(state)->comparetup) (a, b, state)) #define COPYTUP(state,stup,tup) ((*(state)->copytup) (state, stup, tup)) #define WRITETUP(state,tape,stup) ((*(state)->writetup) (state, tape, stup)) #define READTUP(state,stup,tape,len) ((*(state)->readtup) (state, stup, tape, len)) -#define MOVETUP(dest,src,len) ((*(state)->movetup) (dest, src, len)) #define LACKMEM(state) ((state)->availMem < 0 && !(state)->batchUsed) #define USEMEM(state,amt) ((state)->availMem -= (amt)) #define FREEMEM(state,amt) ((state)->availMem += (amt)) @@ -553,16 +578,8 @@ static void inittapes(Tuplesortstate *state); static void selectnewtape(Tuplesortstate *state); static void mergeruns(Tuplesortstate *state); static void mergeonerun(Tuplesortstate *state); -static void beginmerge(Tuplesortstate *state, bool finalMergeBatch); -static void batchmemtuples(Tuplesortstate *state); -static void mergebatch(Tuplesortstate *state, int64 spacePerTape); -static void mergebatchone(Tuplesortstate *state, int srcTape, - SortTuple *stup, bool *should_free); -static void mergebatchfreetape(Tuplesortstate *state, int srcTape, - SortTuple *rtup, bool *should_free); -static void *mergebatchalloc(Tuplesortstate *state, int tapenum, Size tuplen); -static void mergepreread(Tuplesortstate *state); -static void mergeprereadone(Tuplesortstate *state, int srcTape); +static void beginmerge(Tuplesortstate *state); +static bool mergereadnext(Tuplesortstate *state, int srcTape, SortTuple *stup); static void dumptuples(Tuplesortstate *state, bool alltuples); static void dumpbatch(Tuplesortstate *state, bool alltuples); static void make_bounded_heap(Tuplesortstate *state); @@ -574,7 +591,7 @@ static void tuplesort_heap_siftup(Tuplesortstate *state, bool checkIndex); static void reversedirection(Tuplesortstate *state); static unsigned int getlen(Tuplesortstate *state, int tapenum, bool eofOK); static void markrunend(Tuplesortstate *state, int tapenum); -static void *readtup_alloc(Tuplesortstate *state, int tapenum, Size tuplen); +static void *readtup_alloc(Tuplesortstate *state, Size tuplen); static int comparetup_heap(const SortTuple *a, const SortTuple *b, Tuplesortstate *state); static void copytup_heap(Tuplesortstate *state, SortTuple *stup, void *tup); @@ -582,7 +599,6 @@ static void writetup_heap(Tuplesortstate *state, int tapenum, SortTuple *stup); static void readtup_heap(Tuplesortstate *state, SortTuple *stup, int tapenum, unsigned int len); -static void movetup_heap(void *dest, void *src, unsigned int len); static int comparetup_cluster(const SortTuple *a, const SortTuple *b, Tuplesortstate *state); static void copytup_cluster(Tuplesortstate *state, SortTuple *stup, void *tup); @@ -590,7 +606,6 @@ static void writetup_cluster(Tuplesortstate *state, int tapenum, SortTuple *stup); static void readtup_cluster(Tuplesortstate *state, SortTuple *stup, int tapenum, unsigned int len); -static void movetup_cluster(void *dest, void *src, unsigned int len); static int comparetup_index_btree(const SortTuple *a, const SortTuple *b, Tuplesortstate *state); static int comparetup_index_hash(const SortTuple *a, const SortTuple *b, @@ -600,7 +615,6 @@ static void writetup_index(Tuplesortstate *state, int tapenum, SortTuple *stup); static void readtup_index(Tuplesortstate *state, SortTuple *stup, int tapenum, unsigned int len); -static void movetup_index(void *dest, void *src, unsigned int len); static int comparetup_datum(const SortTuple *a, const SortTuple *b, Tuplesortstate *state); static void copytup_datum(Tuplesortstate *state, SortTuple *stup, void *tup); @@ -608,7 +622,6 @@ static void writetup_datum(Tuplesortstate *state, int tapenum, SortTuple *stup); static void readtup_datum(Tuplesortstate *state, SortTuple *stup, int tapenum, unsigned int len); -static void movetup_datum(void *dest, void *src, unsigned int len); static void free_sort_tuple(Tuplesortstate *state, SortTuple *stup); /* @@ -760,7 +773,6 @@ tuplesort_begin_heap(TupleDesc tupDesc, state->copytup = copytup_heap; state->writetup = writetup_heap; state->readtup = readtup_heap; - state->movetup = movetup_heap; state->tupDesc = tupDesc; /* assume we need not copy tupDesc */ state->abbrevNext = 10; @@ -833,7 +845,6 @@ tuplesort_begin_cluster(TupleDesc tupDesc, state->copytup = copytup_cluster; state->writetup = writetup_cluster; state->readtup = readtup_cluster; - state->movetup = movetup_cluster; state->abbrevNext = 10; state->indexInfo = BuildIndexInfo(indexRel); @@ -925,7 +936,6 @@ tuplesort_begin_index_btree(Relation heapRel, state->copytup = copytup_index; state->writetup = writetup_index; state->readtup = readtup_index; - state->movetup = movetup_index; state->abbrevNext = 10; state->heapRel = heapRel; @@ -993,7 +1003,6 @@ tuplesort_begin_index_hash(Relation heapRel, state->copytup = copytup_index; state->writetup = writetup_index; state->readtup = readtup_index; - state->movetup = movetup_index; state->heapRel = heapRel; state->indexRel = indexRel; @@ -1036,7 +1045,6 @@ tuplesort_begin_datum(Oid datumType, Oid sortOperator, Oid sortCollation, state->copytup = copytup_datum; state->writetup = writetup_datum; state->readtup = readtup_datum; - state->movetup = movetup_datum; state->abbrevNext = 10; state->datumType = datumType; @@ -1881,14 +1889,33 @@ tuplesort_gettuple_common(Tuplesortstate *state, bool forward, case TSS_SORTEDONTAPE: Assert(forward || state->randomAccess); Assert(!state->batchUsed); - *should_free = true; + + /* + * The buffer holding the tuple that we returned in previous + * gettuple call can now be reused. + */ + if (state->readlasttuple) + { + RELEASE_MERGETUPLE_BUFFER(state, state->readlasttuple); + state->readlasttuple = NULL; + } + if (forward) { if (state->eof_reached) return false; + if ((tuplen = getlen(state, state->result_tape, true)) != 0) { READTUP(state, stup, state->result_tape, tuplen); + + /* + * Remember the tuple we return, so that we can recycle its + * memory on next call. (This can be NULL, in the Datum case). + */ + state->readlasttuple = stup->tuple; + + *should_free = false; return true; } else @@ -1962,6 +1989,14 @@ tuplesort_gettuple_common(Tuplesortstate *state, bool forward, tuplen)) elog(ERROR, "bogus tuple length in backward scan"); READTUP(state, stup, state->result_tape, tuplen); + + /* + * Remember the tuple we return, so that we can recycle its + * memory on next call. (This can be NULL, in the Datum case). + */ + state->readlasttuple = stup->tuple; + + *should_free = false; return true; case TSS_FINALMERGE: @@ -1971,13 +2006,22 @@ tuplesort_gettuple_common(Tuplesortstate *state, bool forward, *should_free = false; /* + * The buffer holding the tuple that we returned in previous + * gettuple call can now be reused. + */ + if (state->readlasttuple) + { + RELEASE_MERGETUPLE_BUFFER(state, state->readlasttuple); + state->readlasttuple = NULL; + } + + /* * This code should match the inner loop of mergeonerun(). */ if (state->memtupcount > 0) { int srcTape = state->memtuples[0].tupindex; - int tupIndex; - SortTuple *newtup; + SortTuple newtup; /* * Returned tuple is still counted in our memory space most of @@ -1988,42 +2032,22 @@ tuplesort_gettuple_common(Tuplesortstate *state, bool forward, */ *stup = state->memtuples[0]; tuplesort_heap_siftup(state, false); - if ((tupIndex = state->mergenext[srcTape]) == 0) - { - /* - * out of preloaded data on this tape, try to read more - * - * Unlike mergeonerun(), we only preload from the single - * tape that's run dry, though not before preparing its - * batch memory for a new round of sequential consumption. - * See mergepreread() comments. - */ - if (state->batchUsed) - mergebatchone(state, srcTape, stup, should_free); - mergeprereadone(state, srcTape); + /* + * Remember the tuple we return, so that we can recycle its + * memory on next call. (This can be NULL, in the Datum case). + */ + state->readlasttuple = stup->tuple; - /* - * if still no data, we've reached end of run on this tape - */ - if ((tupIndex = state->mergenext[srcTape]) == 0) - { - /* Free tape's buffer, avoiding dangling pointer */ - if (state->batchUsed) - mergebatchfreetape(state, srcTape, stup, should_free); - return true; - } + /* pull next tuple from tape, insert in heap */ + if (!mergereadnext(state, srcTape, &newtup)) + { + /* we've reached end of run on this tape */ + return true; } - /* pull next preread tuple from list, insert in heap */ - newtup = &state->memtuples[tupIndex]; - state->mergenext[srcTape] = newtup->tupindex; - if (state->mergenext[srcTape] == 0) - state->mergelast[srcTape] = 0; - tuplesort_heap_insert(state, newtup, srcTape, false); - /* put the now-unused memtuples entry on the freelist */ - newtup->tupindex = state->mergefreelist; - state->mergefreelist = tupIndex; - state->mergeavailslots[srcTape]++; + + tuplesort_heap_insert(state, &newtup, srcTape, false); + return true; } return false; @@ -2325,7 +2349,8 @@ inittapes(Tuplesortstate *state) #endif /* - * Decrease availMem to reflect the space needed for tape buffers; but + * Decrease availMem to reflect the space needed for tape buffers, when + * writing the initial runs; but * don't decrease it to the point that we have no room for tuples. (That * case is only likely to occur if sorting pass-by-value Datums; in all * other scenarios the memtuples[] array is unlikely to occupy more than @@ -2350,14 +2375,6 @@ inittapes(Tuplesortstate *state) state->tapeset = LogicalTapeSetCreate(maxTapes); state->mergeactive = (bool *) palloc0(maxTapes * sizeof(bool)); - state->mergenext = (int *) palloc0(maxTapes * sizeof(int)); - state->mergelast = (int *) palloc0(maxTapes * sizeof(int)); - state->mergeavailslots = (int *) palloc0(maxTapes * sizeof(int)); - state->mergeavailmem = (int64 *) palloc0(maxTapes * sizeof(int64)); - state->mergetuples = (char **) palloc0(maxTapes * sizeof(char *)); - state->mergecurrent = (char **) palloc0(maxTapes * sizeof(char *)); - state->mergetail = (char **) palloc0(maxTapes * sizeof(char *)); - state->mergeoverflow = (char **) palloc0(maxTapes * sizeof(char *)); state->tp_fib = (int *) palloc0(maxTapes * sizeof(int)); state->tp_runs = (int *) palloc0(maxTapes * sizeof(int)); state->tp_dummy = (int *) palloc0(maxTapes * sizeof(int)); @@ -2468,6 +2485,8 @@ mergeruns(Tuplesortstate *state) svTape, svRuns, svDummy; + char *p; + int i; Assert(state->status == TSS_BUILDRUNS); Assert(state->memtupcount == 0); @@ -2504,6 +2523,36 @@ mergeruns(Tuplesortstate *state) return; } + /* + * We no longer need a large memtuples array, only one slot per tape. Shrink + * it, to make the memory available for other use. We only need one slot per + * tape. + */ + pfree(state->memtuples); + FREEMEM(state, state->memtupsize * sizeof(SortTuple)); + state->memtupsize = state->maxTapes; + state->memtuples = (SortTuple *) palloc(state->maxTapes * sizeof(SortTuple)); + USEMEM(state, state->memtupsize * sizeof(SortTuple)); + + /* + * From this point on, we no longer use the USEMEM()/LACKMEM() mechanism to + * track memory usage. + */ + state->batchUsed = true; + + /* Initialize the merge tuple buffer arena. */ + state->batchMemoryBegin = palloc((state->maxTapes + 1) * MERGETUPLEBUFFER_SIZE); + state->batchMemoryEnd = state->batchMemoryBegin + (state->maxTapes + 1) * MERGETUPLEBUFFER_SIZE; + state->freeBufferHead = (MergeTupleBuffer *) state->batchMemoryBegin; + + p = state->batchMemoryBegin; + for (i = 0; i < state->maxTapes; i++) + { + ((MergeTupleBuffer *) p)->nextfree = (MergeTupleBuffer *) (p + MERGETUPLEBUFFER_SIZE); + p += MERGETUPLEBUFFER_SIZE; + } + ((MergeTupleBuffer *) p)->nextfree = NULL; + /* End of step D2: rewind all output tapes to prepare for merging */ for (tapenum = 0; tapenum < state->tapeRange; tapenum++) LogicalTapeRewind(state->tapeset, tapenum, false); @@ -2534,7 +2583,7 @@ mergeruns(Tuplesortstate *state) /* Tell logtape.c we won't be writing anymore */ LogicalTapeSetForgetFreeSpace(state->tapeset); /* Initialize for the final merge pass */ - beginmerge(state, state->tuples); + beginmerge(state); state->status = TSS_FINALMERGE; return; } @@ -2617,16 +2666,12 @@ mergeonerun(Tuplesortstate *state) { int destTape = state->tp_tapenum[state->tapeRange]; int srcTape; - int tupIndex; - SortTuple *tup; - int64 priorAvail, - spaceFreed; /* * Start the merge by loading one tuple from each active source tape into * the heap. We can also decrease the input run/dummy run counts. */ - beginmerge(state, false); + beginmerge(state); /* * Execute merge by repeatedly extracting lowest tuple in heap, writing it @@ -2635,33 +2680,25 @@ mergeonerun(Tuplesortstate *state) */ while (state->memtupcount > 0) { + SortTuple stup; + /* write the tuple to destTape */ - priorAvail = state->availMem; srcTape = state->memtuples[0].tupindex; WRITETUP(state, destTape, &state->memtuples[0]); - /* writetup adjusted total free space, now fix per-tape space */ - spaceFreed = state->availMem - priorAvail; - state->mergeavailmem[srcTape] += spaceFreed; + + /* Recycle the buffer we just wrote out, for the next read */ + RELEASE_MERGETUPLE_BUFFER(state, state->memtuples[0].tuple); + /* compact the heap */ tuplesort_heap_siftup(state, false); - if ((tupIndex = state->mergenext[srcTape]) == 0) + + /* pull next tuple from tape, insert in heap */ + if (!mergereadnext(state, srcTape, &stup)) { - /* out of preloaded data on this tape, try to read more */ - mergepreread(state); - /* if still no data, we've reached end of run on this tape */ - if ((tupIndex = state->mergenext[srcTape]) == 0) - continue; + /* we've reached end of run on this tape */ + continue; } - /* pull next preread tuple from list, insert in heap */ - tup = &state->memtuples[tupIndex]; - state->mergenext[srcTape] = tup->tupindex; - if (state->mergenext[srcTape] == 0) - state->mergelast[srcTape] = 0; - tuplesort_heap_insert(state, tup, srcTape, false); - /* put the now-unused memtuples entry on the freelist */ - tup->tupindex = state->mergefreelist; - state->mergefreelist = tupIndex; - state->mergeavailslots[srcTape]++; + tuplesort_heap_insert(state, &stup, srcTape, false); } /* @@ -2694,18 +2731,13 @@ mergeonerun(Tuplesortstate *state) * which tapes contain active input runs in mergeactive[]. Then, load * as many tuples as we can from each active input tape, and finally * fill the merge heap with the first tuple from each active tape. - * - * finalMergeBatch indicates if this is the beginning of a final on-the-fly - * merge where a batched allocation of tuple memory is required. */ static void -beginmerge(Tuplesortstate *state, bool finalMergeBatch) +beginmerge(Tuplesortstate *state) { int activeTapes; int tapenum; int srcTape; - int slotsPerTape; - int64 spacePerTape; /* Heap should be empty here */ Assert(state->memtupcount == 0); @@ -2729,497 +2761,48 @@ beginmerge(Tuplesortstate *state, bool finalMergeBatch) } state->activeTapes = activeTapes; - /* Clear merge-pass state variables */ - memset(state->mergenext, 0, - state->maxTapes * sizeof(*state->mergenext)); - memset(state->mergelast, 0, - state->maxTapes * sizeof(*state->mergelast)); - state->mergefreelist = 0; /* nothing in the freelist */ - state->mergefirstfree = activeTapes; /* 1st slot avail for preread */ - - if (finalMergeBatch) - { - /* Free outright buffers for tape never actually allocated */ - FREEMEM(state, (state->maxTapes - activeTapes) * TAPE_BUFFER_OVERHEAD); - - /* - * Grow memtuples one last time, since the palloc() overhead no longer - * incurred can make a big difference - */ - batchmemtuples(state); - } - /* * Initialize space allocation to let each active input tape have an equal * share of preread space. */ Assert(activeTapes > 0); - slotsPerTape = (state->memtupsize - state->mergefirstfree) / activeTapes; - Assert(slotsPerTape > 0); - spacePerTape = MAXALIGN_DOWN(state->availMem / activeTapes); - for (srcTape = 0; srcTape < state->maxTapes; srcTape++) - { - if (state->mergeactive[srcTape]) - { - state->mergeavailslots[srcTape] = slotsPerTape; - state->mergeavailmem[srcTape] = spacePerTape; - } - } - - /* - * Preallocate tuple batch memory for each tape. This is the memory used - * for tuples themselves (not SortTuples), so it's never used by - * pass-by-value datum sorts. Memory allocation is performed here at most - * once per sort, just in advance of the final on-the-fly merge step. - */ - if (finalMergeBatch) - mergebatch(state, spacePerTape); - - /* - * Preread as many tuples as possible (and at least one) from each active - * tape - */ - mergepreread(state); /* Load the merge heap with the first tuple from each input tape */ for (srcTape = 0; srcTape < state->maxTapes; srcTape++) { - int tupIndex = state->mergenext[srcTape]; - SortTuple *tup; - - if (tupIndex) - { - tup = &state->memtuples[tupIndex]; - state->mergenext[srcTape] = tup->tupindex; - if (state->mergenext[srcTape] == 0) - state->mergelast[srcTape] = 0; - tuplesort_heap_insert(state, tup, srcTape, false); - /* put the now-unused memtuples entry on the freelist */ - tup->tupindex = state->mergefreelist; - state->mergefreelist = tupIndex; - state->mergeavailslots[srcTape]++; - -#ifdef TRACE_SORT - if (trace_sort && finalMergeBatch) - { - int64 perTapeKB = (spacePerTape + 1023) / 1024; - int64 usedSpaceKB; - int usedSlots; + SortTuple tup; - /* - * Report how effective batchmemtuples() was in balancing the - * number of slots against the need for memory for the - * underlying tuples (e.g. IndexTuples). The big preread of - * all tapes when switching to FINALMERGE state should be - * fairly representative of memory utilization during the - * final merge step, and in any case is the only point at - * which all tapes are guaranteed to have depleted either - * their batch memory allowance or slot allowance. Ideally, - * both will be completely depleted for every tape by now. - */ - usedSpaceKB = (state->mergecurrent[srcTape] - - state->mergetuples[srcTape] + 1023) / 1024; - usedSlots = slotsPerTape - state->mergeavailslots[srcTape]; - - elog(LOG, "tape %d initially used " INT64_FORMAT " KB of " - INT64_FORMAT " KB batch (%2.3f) and %d out of %d slots " - "(%2.3f)", srcTape, - usedSpaceKB, perTapeKB, - (double) usedSpaceKB / (double) perTapeKB, - usedSlots, slotsPerTape, - (double) usedSlots / (double) slotsPerTape); - } -#endif - } + if (mergereadnext(state, srcTape, &tup)) + tuplesort_heap_insert(state, &tup, srcTape, false); } } /* - * batchmemtuples - grow memtuples without palloc overhead + * mergereadnext - load tuple from one merge input tape * - * When called, availMem should be approximately the amount of memory we'd - * require to allocate memtupsize - memtupcount tuples (not SortTuples/slots) - * that were allocated with palloc() overhead, and in doing so use up all - * allocated slots. However, though slots and tuple memory is in balance - * following the last grow_memtuples() call, that's predicated on the observed - * average tuple size for the "final" grow_memtuples() call, which includes - * palloc overhead. During the final merge pass, where we will arrange to - * squeeze out the palloc overhead, we might need more slots in the memtuples - * array. - * - * To make that happen, arrange for the amount of remaining memory to be - * exactly equal to the palloc overhead multiplied by the current size of - * the memtuples array, force the grow_memtuples flag back to true (it's - * probably but not necessarily false on entry to this routine), and then - * call grow_memtuples. This simulates loading enough tuples to fill the - * whole memtuples array and then having some space left over because of the - * elided palloc overhead. We expect that grow_memtuples() will conclude that - * it can't double the size of the memtuples array but that it can increase - * it by some percentage; but if it does decide to double it, that just means - * that we've never managed to use many slots in the memtuples array, in which - * case doubling it shouldn't hurt anything anyway. - */ -static void -batchmemtuples(Tuplesortstate *state) -{ - int64 refund; - int64 availMemLessRefund; - int memtupsize = state->memtupsize; - - /* For simplicity, assume no memtuples are actually currently counted */ - Assert(state->memtupcount == 0); - - /* - * Refund STANDARDCHUNKHEADERSIZE per tuple. - * - * This sometimes fails to make memory use perfectly balanced, but it - * should never make the situation worse. Note that Assert-enabled builds - * get a larger refund, due to a varying STANDARDCHUNKHEADERSIZE. - */ - refund = memtupsize * STANDARDCHUNKHEADERSIZE; - availMemLessRefund = state->availMem - refund; - - /* - * To establish balanced memory use after refunding palloc overhead, - * temporarily have our accounting indicate that we've allocated all - * memory we're allowed to less that refund, and call grow_memtuples() to - * have it increase the number of slots. - */ - state->growmemtuples = true; - USEMEM(state, availMemLessRefund); - (void) grow_memtuples(state); - /* Should not matter, but be tidy */ - FREEMEM(state, availMemLessRefund); - state->growmemtuples = false; - -#ifdef TRACE_SORT - if (trace_sort) - { - Size OldKb = (memtupsize * sizeof(SortTuple) + 1023) / 1024; - Size NewKb = (state->memtupsize * sizeof(SortTuple) + 1023) / 1024; - - elog(LOG, "grew memtuples %1.2fx from %d (%zu KB) to %d (%zu KB) for final merge", - (double) NewKb / (double) OldKb, - memtupsize, OldKb, - state->memtupsize, NewKb); - } -#endif -} - -/* - * mergebatch - initialize tuple memory in batch - * - * This allows sequential access to sorted tuples buffered in memory from - * tapes/runs on disk during a final on-the-fly merge step. Note that the - * memory is not used for SortTuples, but for the underlying tuples (e.g. - * MinimalTuples). - * - * Note that when batch memory is used, there is a simple division of space - * into large buffers (one per active tape). The conventional incremental - * memory accounting (calling USEMEM() and FREEMEM()) is abandoned. Instead, - * when each tape's memory budget is exceeded, a retail palloc() "overflow" is - * performed, which is then immediately detected in a way that is analogous to - * LACKMEM(). This keeps each tape's use of memory fair, which is always a - * goal. - */ -static void -mergebatch(Tuplesortstate *state, int64 spacePerTape) -{ - int srcTape; - - Assert(state->activeTapes > 0); - Assert(state->tuples); - - /* - * For the purposes of tuplesort's memory accounting, the batch allocation - * is special, and regular memory accounting through USEMEM() calls is - * abandoned (see mergeprereadone()). - */ - for (srcTape = 0; srcTape < state->maxTapes; srcTape++) - { - char *mergetuples; - - if (!state->mergeactive[srcTape]) - continue; - - /* Allocate buffer for each active tape */ - mergetuples = MemoryContextAllocHuge(state->tuplecontext, - spacePerTape); - - /* Initialize state for tape */ - state->mergetuples[srcTape] = mergetuples; - state->mergecurrent[srcTape] = mergetuples; - state->mergetail[srcTape] = mergetuples; - state->mergeoverflow[srcTape] = NULL; - } - - state->batchUsed = true; - state->spacePerTape = spacePerTape; -} - -/* - * mergebatchone - prepare batch memory for one merge input tape - * - * This is called following the exhaustion of preread tuples for one input - * tape. All that actually occurs is that the state for the source tape is - * reset to indicate that all memory may be reused. - * - * This routine must deal with fixing up the tuple that is about to be returned - * to the client, due to "overflow" allocations. - */ -static void -mergebatchone(Tuplesortstate *state, int srcTape, SortTuple *rtup, - bool *should_free) -{ - Assert(state->batchUsed); - - /* - * Tuple about to be returned to caller ("stup") is final preread tuple - * from tape, just removed from the top of the heap. Special steps around - * memory management must be performed for that tuple, to make sure it - * isn't overwritten early. - */ - if (!state->mergeoverflow[srcTape]) - { - Size tupLen; - - /* - * Mark tuple buffer range for reuse, but be careful to move final, - * tail tuple to start of space for next run so that it's available to - * caller when stup is returned, and remains available at least until - * the next tuple is requested. - */ - tupLen = state->mergecurrent[srcTape] - state->mergetail[srcTape]; - state->mergecurrent[srcTape] = state->mergetuples[srcTape]; - MOVETUP(state->mergecurrent[srcTape], state->mergetail[srcTape], - tupLen); - - /* Make SortTuple at top of the merge heap point to new tuple */ - rtup->tuple = (void *) state->mergecurrent[srcTape]; - - state->mergetail[srcTape] = state->mergecurrent[srcTape]; - state->mergecurrent[srcTape] += tupLen; - } - else - { - /* - * Handle an "overflow" retail palloc. - * - * This is needed when we run out of tuple memory for the tape. - */ - state->mergecurrent[srcTape] = state->mergetuples[srcTape]; - state->mergetail[srcTape] = state->mergetuples[srcTape]; - - if (rtup->tuple) - { - Assert(rtup->tuple == (void *) state->mergeoverflow[srcTape]); - /* Caller should free palloc'd tuple */ - *should_free = true; - } - state->mergeoverflow[srcTape] = NULL; - } -} - -/* - * mergebatchfreetape - handle final clean-up for batch memory once tape is - * about to become exhausted - * - * All tuples are returned from tape, but a single final tuple, *rtup, is to be - * passed back to caller. Free tape's batch allocation buffer while ensuring - * that the final tuple is managed appropriately. - */ -static void -mergebatchfreetape(Tuplesortstate *state, int srcTape, SortTuple *rtup, - bool *should_free) -{ - Assert(state->batchUsed); - Assert(state->status == TSS_FINALMERGE); - - /* - * Tuple may or may not already be an overflow allocation from - * mergebatchone() - */ - if (!*should_free && rtup->tuple) - { - /* - * Final tuple still in tape's batch allocation. - * - * Return palloc()'d copy to caller, and have it freed in a similar - * manner to overflow allocation. Otherwise, we'd free batch memory - * and pass back a pointer to garbage. Note that we deliberately - * allocate this in the parent tuplesort context, to be on the safe - * side. - */ - Size tuplen; - void *oldTuple = rtup->tuple; - - tuplen = state->mergecurrent[srcTape] - state->mergetail[srcTape]; - rtup->tuple = MemoryContextAlloc(state->sortcontext, tuplen); - MOVETUP(rtup->tuple, oldTuple, tuplen); - *should_free = true; - } - - /* Free spacePerTape-sized buffer */ - pfree(state->mergetuples[srcTape]); -} - -/* - * mergebatchalloc - allocate memory for one tuple using a batch memory - * "logical allocation". - * - * This is used for the final on-the-fly merge phase only. READTUP() routines - * receive memory from here in place of palloc() and USEMEM() calls. - * - * Tuple tapenum is passed, ensuring each tape's tuples are stored in sorted, - * contiguous order (while allowing safe reuse of memory made available to - * each tape). This maximizes locality of access as tuples are returned by - * final merge. - * - * Caller must not subsequently attempt to free memory returned here. In - * general, only mergebatch* functions know about how memory returned from - * here should be freed, and this function's caller must ensure that batch - * memory management code will definitely have the opportunity to do the right - * thing during the final on-the-fly merge. - */ -static void * -mergebatchalloc(Tuplesortstate *state, int tapenum, Size tuplen) -{ - Size reserve_tuplen = MAXALIGN(tuplen); - char *ret; - - /* Should overflow at most once before mergebatchone() call: */ - Assert(state->mergeoverflow[tapenum] == NULL); - Assert(state->batchUsed); - - /* It should be possible to use precisely spacePerTape memory at once */ - if (state->mergecurrent[tapenum] + reserve_tuplen <= - state->mergetuples[tapenum] + state->spacePerTape) - { - /* - * Usual case -- caller is returned pointer into its tape's buffer, - * and an offset from that point is recorded as where tape has - * consumed up to for current round of preloading. - */ - ret = state->mergetail[tapenum] = state->mergecurrent[tapenum]; - state->mergecurrent[tapenum] += reserve_tuplen; - } - else - { - /* - * Allocate memory, and record as tape's overflow allocation. This - * will be detected quickly, in a similar fashion to a LACKMEM() - * condition, and should not happen again before a new round of - * preloading for caller's tape. Note that we deliberately allocate - * this in the parent tuplesort context, to be on the safe side. - * - * Sometimes, this does not happen because merging runs out of slots - * before running out of memory. - */ - ret = state->mergeoverflow[tapenum] = - MemoryContextAlloc(state->sortcontext, tuplen); - } - - return ret; -} - -/* - * mergepreread - load tuples from merge input tapes - * - * This routine exists to improve sequentiality of reads during a merge pass, - * as explained in the header comments of this file. Load tuples from each - * active source tape until the tape's run is exhausted or it has used up - * its fair share of available memory. In any case, we guarantee that there - * is at least one preread tuple available from each unexhausted input tape. - * - * We invoke this routine at the start of a merge pass for initial load, - * and then whenever any tape's preread data runs out. Note that we load - * as much data as possible from all tapes, not just the one that ran out. - * This is because logtape.c works best with a usage pattern that alternates - * between reading a lot of data and writing a lot of data, so whenever we - * are forced to read, we should fill working memory completely. - * - * In FINALMERGE state, we *don't* use this routine, but instead just preread - * from the single tape that ran dry. There's no read/write alternation in - * that state and so no point in scanning through all the tapes to fix one. - * (Moreover, there may be quite a lot of inactive tapes in that state, since - * we might have had many fewer runs than tapes. In a regular tape-to-tape - * merge we can expect most of the tapes to be active. Plus, only - * FINALMERGE state has to consider memory management for a batch - * allocation.) - */ -static void -mergepreread(Tuplesortstate *state) -{ - int srcTape; - - for (srcTape = 0; srcTape < state->maxTapes; srcTape++) - mergeprereadone(state, srcTape); -} - -/* - * mergeprereadone - load tuples from one merge input tape + * Returns false on EOF. * * Read tuples from the specified tape until it has used up its free memory * or array slots; but ensure that we have at least one tuple, if any are * to be had. */ -static void -mergeprereadone(Tuplesortstate *state, int srcTape) +static bool +mergereadnext(Tuplesortstate *state, int srcTape, SortTuple *stup) { unsigned int tuplen; - SortTuple stup; - int tupIndex; - int64 priorAvail, - spaceUsed; if (!state->mergeactive[srcTape]) - return; /* tape's run is already exhausted */ + return false; /* tape's run is already exhausted */ - /* - * Manage per-tape availMem. Only actually matters when batch memory not - * in use. - */ - priorAvail = state->availMem; - state->availMem = state->mergeavailmem[srcTape]; - - /* - * When batch memory is used if final on-the-fly merge, only mergeoverflow - * test is relevant; otherwise, only LACKMEM() test is relevant. - */ - while ((state->mergeavailslots[srcTape] > 0 && - state->mergeoverflow[srcTape] == NULL && !LACKMEM(state)) || - state->mergenext[srcTape] == 0) + /* read next tuple, if any */ + if ((tuplen = getlen(state, srcTape, true)) == 0) { - /* read next tuple, if any */ - if ((tuplen = getlen(state, srcTape, true)) == 0) - { - state->mergeactive[srcTape] = false; - break; - } - READTUP(state, &stup, srcTape, tuplen); - /* find a free slot in memtuples[] for it */ - tupIndex = state->mergefreelist; - if (tupIndex) - state->mergefreelist = state->memtuples[tupIndex].tupindex; - else - { - tupIndex = state->mergefirstfree++; - Assert(tupIndex < state->memtupsize); - } - state->mergeavailslots[srcTape]--; - /* store tuple, append to list for its tape */ - stup.tupindex = 0; - state->memtuples[tupIndex] = stup; - if (state->mergelast[srcTape]) - state->memtuples[state->mergelast[srcTape]].tupindex = tupIndex; - else - state->mergenext[srcTape] = tupIndex; - state->mergelast[srcTape] = tupIndex; + state->mergeactive[srcTape] = false; + return false; } - /* update per-tape and global availmem counts */ - spaceUsed = state->mergeavailmem[srcTape] - state->availMem; - state->mergeavailmem[srcTape] = state->availMem; - state->availMem = priorAvail - spaceUsed; + READTUP(state, stup, srcTape, tuplen); + + return true; } /* @@ -3857,27 +3440,24 @@ markrunend(Tuplesortstate *state, int tapenum) * routines. */ static void * -readtup_alloc(Tuplesortstate *state, int tapenum, Size tuplen) +readtup_alloc(Tuplesortstate *state, Size tuplen) { - if (state->batchUsed) - { - /* - * No USEMEM() call, because during final on-the-fly merge accounting - * is based on tape-private state. ("Overflow" allocations are - * detected as an indication that a new round or preloading is - * required. Preloading marks existing contents of tape's batch buffer - * for reuse.) - */ - return mergebatchalloc(state, tapenum, tuplen); - } + MergeTupleBuffer *buf; + + /* + * We pre-allocate enough buffers in the arena that we should never run out. + */ + Assert(state->freeBufferHead); + + if (tuplen > MERGETUPLEBUFFER_SIZE || !state->freeBufferHead) + return MemoryContextAlloc(state->sortcontext, tuplen); else { - char *ret; + buf = state->freeBufferHead; + /* Reuse this buffer */ + state->freeBufferHead = buf->nextfree; - /* Batch allocation yet to be performed */ - ret = MemoryContextAlloc(state->tuplecontext, tuplen); - USEMEM(state, GetMemoryChunkSpace(ret)); - return ret; + return buf; } } @@ -4046,8 +3626,11 @@ writetup_heap(Tuplesortstate *state, int tapenum, SortTuple *stup) LogicalTapeWrite(state->tapeset, tapenum, (void *) &tuplen, sizeof(tuplen)); - FREEMEM(state, GetMemoryChunkSpace(tuple)); - heap_free_minimal_tuple(tuple); + if (!state->batchUsed) + { + FREEMEM(state, GetMemoryChunkSpace(tuple)); + heap_free_minimal_tuple(tuple); + } } static void @@ -4056,7 +3639,7 @@ readtup_heap(Tuplesortstate *state, SortTuple *stup, { unsigned int tupbodylen = len - sizeof(int); unsigned int tuplen = tupbodylen + MINIMAL_TUPLE_DATA_OFFSET; - MinimalTuple tuple = (MinimalTuple) readtup_alloc(state, tapenum, tuplen); + MinimalTuple tuple = (MinimalTuple) readtup_alloc(state, tuplen); char *tupbody = (char *) tuple + MINIMAL_TUPLE_DATA_OFFSET; HeapTupleData htup; @@ -4077,12 +3660,6 @@ readtup_heap(Tuplesortstate *state, SortTuple *stup, &stup->isnull1); } -static void -movetup_heap(void *dest, void *src, unsigned int len) -{ - memmove(dest, src, len); -} - /* * Routines specialized for the CLUSTER case (HeapTuple data, with * comparisons per a btree index definition) @@ -4289,8 +3866,11 @@ writetup_cluster(Tuplesortstate *state, int tapenum, SortTuple *stup) LogicalTapeWrite(state->tapeset, tapenum, &tuplen, sizeof(tuplen)); - FREEMEM(state, GetMemoryChunkSpace(tuple)); - heap_freetuple(tuple); + if (!state->batchUsed) + { + FREEMEM(state, GetMemoryChunkSpace(tuple)); + heap_freetuple(tuple); + } } static void @@ -4299,7 +3879,6 @@ readtup_cluster(Tuplesortstate *state, SortTuple *stup, { unsigned int t_len = tuplen - sizeof(ItemPointerData) - sizeof(int); HeapTuple tuple = (HeapTuple) readtup_alloc(state, - tapenum, t_len + HEAPTUPLESIZE); /* Reconstruct the HeapTupleData header */ @@ -4324,19 +3903,6 @@ readtup_cluster(Tuplesortstate *state, SortTuple *stup, &stup->isnull1); } -static void -movetup_cluster(void *dest, void *src, unsigned int len) -{ - HeapTuple tuple; - - memmove(dest, src, len); - - /* Repoint the HeapTupleData header */ - tuple = (HeapTuple) dest; - tuple->t_data = (HeapTupleHeader) ((char *) tuple + HEAPTUPLESIZE); -} - - /* * Routines specialized for IndexTuple case * @@ -4604,8 +4170,11 @@ writetup_index(Tuplesortstate *state, int tapenum, SortTuple *stup) LogicalTapeWrite(state->tapeset, tapenum, (void *) &tuplen, sizeof(tuplen)); - FREEMEM(state, GetMemoryChunkSpace(tuple)); - pfree(tuple); + if (!state->batchUsed) + { + FREEMEM(state, GetMemoryChunkSpace(tuple)); + pfree(tuple); + } } static void @@ -4613,7 +4182,7 @@ readtup_index(Tuplesortstate *state, SortTuple *stup, int tapenum, unsigned int len) { unsigned int tuplen = len - sizeof(unsigned int); - IndexTuple tuple = (IndexTuple) readtup_alloc(state, tapenum, tuplen); + IndexTuple tuple = (IndexTuple) readtup_alloc(state, tuplen); LogicalTapeReadExact(state->tapeset, tapenum, tuple, tuplen); @@ -4628,12 +4197,6 @@ readtup_index(Tuplesortstate *state, SortTuple *stup, &stup->isnull1); } -static void -movetup_index(void *dest, void *src, unsigned int len) -{ - memmove(dest, src, len); -} - /* * Routines specialized for DatumTuple case */ @@ -4700,7 +4263,7 @@ writetup_datum(Tuplesortstate *state, int tapenum, SortTuple *stup) LogicalTapeWrite(state->tapeset, tapenum, (void *) &writtenlen, sizeof(writtenlen)); - if (stup->tuple) + if (!state->batchUsed && stup->tuple) { FREEMEM(state, GetMemoryChunkSpace(stup->tuple)); pfree(stup->tuple); @@ -4730,7 +4293,7 @@ readtup_datum(Tuplesortstate *state, SortTuple *stup, } else { - void *raddr = readtup_alloc(state, tapenum, tuplen); + void *raddr = readtup_alloc(state, tuplen); LogicalTapeReadExact(state->tapeset, tapenum, raddr, tuplen); @@ -4744,12 +4307,6 @@ readtup_datum(Tuplesortstate *state, SortTuple *stup, &tuplen, sizeof(tuplen)); } -static void -movetup_datum(void *dest, void *src, unsigned int len) -{ - memmove(dest, src, len); -} - /* * Convenience routine to free a tuple previously loaded into sort memory */ -- 2.9.3