diff --git a/src/backend/access/common/heaptuple.c b/src/backend/access/common/heaptuple.c index e39b977..acf88de 100644 --- a/src/backend/access/common/heaptuple.c +++ b/src/backend/access/common/heaptuple.c @@ -60,7 +60,11 @@ #include "access/sysattr.h" #include "access/tuptoaster.h" #include "executor/tuptable.h" +#include "utils/datum.h" +#include "utils/pg_lzcompress.h" +/* guc variable for EWT compression ratio*/ +int wal_update_compression_ratio = 25; /* Does att's datatype allow packing into the 1-byte-header varlena format? */ #define ATT_IS_PACKABLE(att) \ @@ -617,6 +621,49 @@ heap_copytuple_with_tuple(HeapTuple src, HeapTuple dest) memcpy((char *) dest->t_data, (char *) src->t_data, src->t_len); } +/* ---------------- + * heap_delta_encode + * + * Calculate the delta between two tuples, using pglz. The result is + * stored in *encdata. *encdata must point to a PGLZ_header buffer, with at + * least PGLZ_MAX_OUTPUT(newtup->t_len) bytes. + * ---------------- + */ +bool +heap_delta_encode(TupleDesc tupleDesc, HeapTuple oldtup, HeapTuple newtup, + char *encdata, uint32 *enclen) +{ + PGLZ_Strategy strategy; + + strategy = *PGLZ_strategy_default; + strategy.min_comp_rate = wal_update_compression_ratio; + + return pglz_delta_encode( + (char *) newtup->t_data + offsetof(HeapTupleHeaderData, t_bits), + newtup->t_len - offsetof(HeapTupleHeaderData, t_bits), + (char *) oldtup->t_data + offsetof(HeapTupleHeaderData, t_bits), + oldtup->t_len - offsetof(HeapTupleHeaderData, t_bits), + encdata, enclen, &strategy + ); +} + +/* ---------------- + * heap_delta_decode + * + * Decode a tuple using delta-encoded WAL tuple and old tuple version. + * ---------------- + */ +void +heap_delta_decode(char *encdata, uint32 enclen, HeapTuple oldtup, HeapTuple newtup) +{ + return pglz_delta_decode(encdata, enclen, + (char *) newtup->t_data + offsetof(HeapTupleHeaderData, t_bits), + MaxHeapTupleSize - offsetof(HeapTupleHeaderData, t_bits), + &newtup->t_len, + (char *) oldtup->t_data + offsetof(HeapTupleHeaderData, t_bits), + oldtup->t_len - offsetof(HeapTupleHeaderData, t_bits)); +} + /* * heap_form_tuple * construct a tuple from the given values[] and isnull[] arrays, diff --git a/src/backend/access/heap/heapam.c b/src/backend/access/heap/heapam.c index fe56318..24c117c 100644 --- a/src/backend/access/heap/heapam.c +++ b/src/backend/access/heap/heapam.c @@ -70,10 +70,12 @@ #include "utils/snapmgr.h" #include "utils/syscache.h" #include "utils/tqual.h" +#include "utils/pg_lzcompress.h" /* GUC variable */ bool synchronize_seqscans = true; +extern int wal_update_compression_ratio; static HeapScanDesc heap_beginscan_internal(Relation relation, @@ -5815,6 +5817,12 @@ log_heap_update(Relation reln, Buffer oldbuf, XLogRecPtr recptr; XLogRecData rdata[4]; Page page = BufferGetPage(newbuf); + char *newtupdata; + int newtuplen; + bool compressed = false; + + /* Structure which holds EWT */ + char buf[MaxHeapTupleSize]; /* Caller should not call me on a non-WAL-logged relation */ Assert(RelationNeedsWAL(reln)); @@ -5824,15 +5832,47 @@ log_heap_update(Relation reln, Buffer oldbuf, else info = XLOG_HEAP_UPDATE; + newtupdata = ((char *) newtup->t_data) + offsetof(HeapTupleHeaderData, t_bits); + newtuplen = newtup->t_len - offsetof(HeapTupleHeaderData, t_bits); + + /* + * EWT can be generated for all new tuple versions created by Update + * operation. Currently we do it when both the old and new tuple versions + * are on same page, because during recovery if the page containing old + * tuple is corrupt, it should not cascade that corruption to other pages. + * Under the general assumption that for long runs most updates tend to + * create new tuple version on same page, there should not be significant + * impact on WAL reduction or performance. + * + * We should not generate EWT when we need to backup the whole bolck in + * WAL as in that case there is no saving by reduced WAL size. + */ + if (wal_update_compression_ratio != 0 && (oldbuf == newbuf) && !XLogCheckBufferNeedsBackup(newbuf)) + { + uint32 enclen; + /* Delta-encode the new tuple using the old tuple */ + if (heap_delta_encode(reln->rd_att, oldtup, newtup, buf, &enclen)) + { + compressed = true; + newtupdata = buf; + newtuplen = enclen; + } + } + + xlrec.flags = 0; xlrec.target.node = reln->rd_node; xlrec.target.tid = oldtup->t_self; xlrec.old_xmax = HeapTupleHeaderGetRawXmax(oldtup->t_data); xlrec.old_infobits_set = compute_infobits(oldtup->t_data->t_infomask, oldtup->t_data->t_infomask2); xlrec.new_xmax = HeapTupleHeaderGetRawXmax(newtup->t_data); - xlrec.all_visible_cleared = all_visible_cleared; + if (all_visible_cleared) + xlrec.flags |= XL_HEAP_UPDATE_ALL_VISIBLE_CLEARED; xlrec.newtid = newtup->t_self; - xlrec.new_all_visible_cleared = new_all_visible_cleared; + if (new_all_visible_cleared) + xlrec.flags |= XL_HEAP_UPDATE_NEW_ALL_VISIBLE_CLEARED; + if (compressed) + xlrec.flags |= XL_HEAP_UPDATE_DELTA_ENCODED; rdata[0].data = (char *) &xlrec; rdata[0].len = SizeOfHeapUpdate; @@ -5859,9 +5899,12 @@ log_heap_update(Relation reln, Buffer oldbuf, rdata[2].buffer_std = true; rdata[2].next = &(rdata[3]); - /* PG73FORMAT: write bitmap [+ padding] [+ oid] + data */ - rdata[3].data = (char *) newtup->t_data + offsetof(HeapTupleHeaderData, t_bits); - rdata[3].len = newtup->t_len - offsetof(HeapTupleHeaderData, t_bits); + /* + * PG73FORMAT: write bitmap [+ padding] [+ oid] + data follows ......... + * OR PG93FORMAT [If encoded]: LZ header + Encoded data follows + */ + rdata[3].data = newtupdata; + rdata[3].len = newtuplen; rdata[3].buffer = newbuf; rdata[3].buffer_std = true; rdata[3].next = NULL; @@ -6671,7 +6714,10 @@ heap_xlog_update(XLogRecPtr lsn, XLogRecord *record, bool hot_update) Page page; OffsetNumber offnum; ItemId lp = NULL; + HeapTupleData newtup; + HeapTupleData oldtup; HeapTupleHeader htup; + HeapTupleHeader oldtupdata = NULL; struct { HeapTupleHeaderData hdr; @@ -6686,7 +6732,7 @@ heap_xlog_update(XLogRecPtr lsn, XLogRecord *record, bool hot_update) * The visibility map may need to be fixed even if the heap page is * already up-to-date. */ - if (xlrec->all_visible_cleared) + if (xlrec->flags & XL_HEAP_UPDATE_ALL_VISIBLE_CLEARED) { Relation reln = CreateFakeRelcacheEntry(xlrec->target.node); BlockNumber block = ItemPointerGetBlockNumber(&xlrec->target.tid); @@ -6746,7 +6792,7 @@ heap_xlog_update(XLogRecPtr lsn, XLogRecord *record, bool hot_update) if (PageGetMaxOffsetNumber(page) < offnum || !ItemIdIsNormal(lp)) elog(PANIC, "heap_update_redo: invalid lp"); - htup = (HeapTupleHeader) PageGetItem(page, lp); + oldtupdata = htup = (HeapTupleHeader) PageGetItem(page, lp); htup->t_infomask &= ~(HEAP_XMAX_BITS | HEAP_MOVED); htup->t_infomask2 &= ~HEAP_KEYS_UPDATED; @@ -6764,7 +6810,7 @@ heap_xlog_update(XLogRecPtr lsn, XLogRecord *record, bool hot_update) /* Mark the page as a candidate for pruning */ PageSetPrunable(page, record->xl_xid); - if (xlrec->all_visible_cleared) + if (xlrec->flags & XL_HEAP_UPDATE_ALL_VISIBLE_CLEARED) PageClearAllVisible(page); /* @@ -6788,7 +6834,7 @@ newt:; * The visibility map may need to be fixed even if the heap page is * already up-to-date. */ - if (xlrec->new_all_visible_cleared) + if (xlrec->flags & XL_HEAP_UPDATE_NEW_ALL_VISIBLE_CLEARED) { Relation reln = CreateFakeRelcacheEntry(xlrec->target.node); BlockNumber block = ItemPointerGetBlockNumber(&xlrec->newtid); @@ -6851,10 +6897,31 @@ newsame:; SizeOfHeapHeader); htup = &tbuf.hdr; MemSet((char *) htup, 0, sizeof(HeapTupleHeaderData)); - /* PG73FORMAT: get bitmap [+ padding] [+ oid] + data */ - memcpy((char *) htup + offsetof(HeapTupleHeaderData, t_bits), - (char *) xlrec + hsize, - newlen); + + /* + * If the record is EWT then decode it. + */ + if (xlrec->flags & XL_HEAP_UPDATE_DELTA_ENCODED) + { + /* + * PG93FORMAT: Header + Control byte + history reference (2 - 3)bytes + * + New data (1 byte length + variable data)+ ... + */ + oldtup.t_data = oldtupdata; + oldtup.t_len = ItemIdGetLength(lp); + newtup.t_data = htup; + + heap_delta_decode((char *) xlrec + hsize, newlen, &oldtup, &newtup); + newlen = newtup.t_len; + } + else + { + /* PG73FORMAT: get bitmap [+ padding] [+ oid] + data */ + memcpy((char *) htup + offsetof(HeapTupleHeaderData, t_bits), + (char *) xlrec + hsize, + newlen); + } + newlen += offsetof(HeapTupleHeaderData, t_bits); htup->t_infomask2 = xlhdr.t_infomask2; htup->t_infomask = xlhdr.t_infomask; @@ -6870,7 +6937,7 @@ newsame:; if (offnum == InvalidOffsetNumber) elog(PANIC, "heap_update_redo: failed to add tuple"); - if (xlrec->new_all_visible_cleared) + if (xlrec->flags & XL_HEAP_UPDATE_NEW_ALL_VISIBLE_CLEARED) PageClearAllVisible(page); freespace = PageGetHeapFreeSpace(page); /* needed to update FSM below */ diff --git a/src/backend/access/transam/xlog.c b/src/backend/access/transam/xlog.c index 07c68ad..c3a94a2 100644 --- a/src/backend/access/transam/xlog.c +++ b/src/backend/access/transam/xlog.c @@ -1249,6 +1249,28 @@ begin:; } /* + * Determine whether the buffer referenced has to be backed up. Since we don't + * yet have the insert lock, fullPageWrites and forcePageWrites could change + * later, but will not cause any problem because this function is used only to + * identify whether EWT is required for WAL update. + */ +bool +XLogCheckBufferNeedsBackup(Buffer buffer) +{ + bool doPageWrites; + Page page; + + page = BufferGetPage(buffer); + + doPageWrites = XLogCtl->Insert.fullPageWrites || XLogCtl->Insert.forcePageWrites; + + if (doPageWrites && PageGetLSN(page) <= RedoRecPtr) + return true; /* buffer requires backup */ + + return false; /* buffer does not need to be backed up */ +} + +/* * Determine whether the buffer referenced by an XLogRecData item has to * be backed up, and if so fill a BkpBlock struct for it. In any case * save the buffer's LSN at *lsn. diff --git a/src/backend/utils/adt/pg_lzcompress.c b/src/backend/utils/adt/pg_lzcompress.c index 66c64c1..35e8206 100644 --- a/src/backend/utils/adt/pg_lzcompress.c +++ b/src/backend/utils/adt/pg_lzcompress.c @@ -112,7 +112,7 @@ * of identical bytes like trailing spaces) and for bigger ones * our 4K maximum look-back distance is too small. * - * The compressor creates a table for 8192 lists of positions. + * The compressor creates a table for lists of positions. * For each input position (except the last 3), a hash key is * built from the 4 next input bytes and the position remembered * in the appropriate list. Thus, the table points to linked @@ -120,7 +120,10 @@ * matching strings. This is done on the fly while the input * is compressed into the output area. Table entries are only * kept for the last 4096 input positions, since we cannot use - * back-pointers larger than that anyway. + * back-pointers larger than that anyway. The size of the hash + * table depends on the size of the input - a larger table has + * a larger startup cost, as it needs to be initialized to zero, + * but reduces the number of hash collisions on long inputs. * * For each byte in the input, it's hash key (built from this * byte and the next 3) is used to find the appropriate list @@ -180,8 +183,7 @@ * Local definitions * ---------- */ -#define PGLZ_HISTORY_LISTS 8192 /* must be power of 2 */ -#define PGLZ_HISTORY_MASK (PGLZ_HISTORY_LISTS - 1) +#define PGLZ_MAX_HISTORY_LISTS 8192 /* must be power of 2 */ #define PGLZ_HISTORY_SIZE 4096 #define PGLZ_MAX_MATCH 273 @@ -198,9 +200,9 @@ */ typedef struct PGLZ_HistEntry { - struct PGLZ_HistEntry *next; /* links for my hash key's list */ - struct PGLZ_HistEntry *prev; - int hindex; /* my current hash key */ + int16 next; /* links for my hash key's list */ + int16 prev; + uint32 hindex; /* my current hash key */ const char *pos; /* my input position */ } PGLZ_HistEntry; @@ -241,9 +243,11 @@ const PGLZ_Strategy *const PGLZ_strategy_always = &strategy_always_data; * Statically allocated work arrays for history * ---------- */ -static PGLZ_HistEntry *hist_start[PGLZ_HISTORY_LISTS]; -static PGLZ_HistEntry hist_entries[PGLZ_HISTORY_SIZE]; +static int16 hist_start[PGLZ_MAX_HISTORY_LISTS]; +static PGLZ_HistEntry hist_entries[PGLZ_HISTORY_SIZE + 1]; +/* Element 0 in hist_entries is unused, and means 'invalid'. */ +#define INVALID_ENTRY 0 /* ---------- * pglz_hist_idx - @@ -257,12 +261,40 @@ static PGLZ_HistEntry hist_entries[PGLZ_HISTORY_SIZE]; * hash keys more. * ---------- */ -#define pglz_hist_idx(_s,_e) ( \ +#define pglz_hist_idx(_s,_e, _mask) ( \ ((((_e) - (_s)) < 4) ? (int) (_s)[0] : \ - (((_s)[0] << 9) ^ ((_s)[1] << 6) ^ \ - ((_s)[2] << 3) ^ (_s)[3])) & (PGLZ_HISTORY_MASK) \ + (((_s)[0] << 6) ^ ((_s)[1] << 4) ^ \ + ((_s)[2] << 2) ^ (_s)[3])) & (_mask) \ ) +/* + * pglz_hash_init and pglz_hash_roll can be use to calculate the hash in + * a rolling fashion. First, call pglz_hash_init, with a pointer to the first + * byte. Then call pglz_hash_roll for every subsequent byte. After each + * pglz_hash_roll() call, hindex holds the (masked) hash of the current byte. + * + * a,b,c,d are local variables these macros use to store state. These macros + * don't check for end-of-buffer like pglz_hist_idx() does, so these cannot be + * used on the last 3 bytes of input. + */ +#define pglz_hash_init(_p,hindex,a,b,c,d) \ + do { \ + a = 0; \ + b = _p[0]; \ + c = _p[1]; \ + d = _p[2]; \ + hindex = (b << 4) ^ (c << 2) ^ d; \ + } while (0) + +#define pglz_hash_roll(_p,hindex,a,b,c,d,_mask) \ + do { \ + /* subtract old a */ \ + hindex ^= a; \ + /* shift and add byte */ \ + a = b; b = c; c = d; d = _p[3]; \ + hindex = ((hindex << 2) ^ d) & (_mask); \ + } while (0) + /* ---------- * pglz_hist_add - @@ -276,32 +308,49 @@ static PGLZ_HistEntry hist_entries[PGLZ_HISTORY_SIZE]; * _hn and _recycle are modified in the macro. * ---------- */ -#define pglz_hist_add(_hs,_he,_hn,_recycle,_s,_e) \ +#define pglz_hist_add(_hs,_he,_hn,_recycle,_s,_e, _hindex) \ do { \ - int __hindex = pglz_hist_idx((_s),(_e)); \ - PGLZ_HistEntry **__myhsp = &(_hs)[__hindex]; \ + int16 *__myhsp = &(_hs)[_hindex]; \ PGLZ_HistEntry *__myhe = &(_he)[_hn]; \ if (_recycle) { \ - if (__myhe->prev == NULL) \ + if (__myhe->prev == INVALID_ENTRY) \ (_hs)[__myhe->hindex] = __myhe->next; \ else \ - __myhe->prev->next = __myhe->next; \ - if (__myhe->next != NULL) \ - __myhe->next->prev = __myhe->prev; \ + (_he)[__myhe->prev].next = __myhe->next; \ + if (__myhe->next != INVALID_ENTRY) \ + (_he)[__myhe->next].prev = __myhe->prev; \ } \ __myhe->next = *__myhsp; \ - __myhe->prev = NULL; \ - __myhe->hindex = __hindex; \ + __myhe->prev = INVALID_ENTRY; \ + __myhe->hindex = _hindex; \ __myhe->pos = (_s); \ - if (*__myhsp != NULL) \ - (*__myhsp)->prev = __myhe; \ - *__myhsp = __myhe; \ - if (++(_hn) >= PGLZ_HISTORY_SIZE) { \ - (_hn) = 0; \ + if (*__myhsp != INVALID_ENTRY) \ + (_he)[(*__myhsp)].prev = _hn; \ + *__myhsp = _hn; \ + if (++(_hn) >= PGLZ_HISTORY_SIZE + 1) { \ + (_hn) = 1; \ (_recycle) = true; \ } \ } while (0) +/* + * An version of pglz_hist_add() that doesn't do recycling. Can be used if + * you know the input fits in PGLZ_HISTORY_SIZE. + */ +#define pglz_hist_add_no_recycle(_hs,_he,_hn,_s,_e, _hindex) \ +do { \ + int16 *__myhsp = &(_hs)[_hindex]; \ + PGLZ_HistEntry *__myhe = &(_he)[_hn]; \ + __myhe->next = *__myhsp; \ + __myhe->prev = INVALID_ENTRY; \ + __myhe->hindex = _hindex; \ + __myhe->pos = (_s); \ + if (*__myhsp != INVALID_ENTRY) \ + (_he)[(*__myhsp)].prev = _hn; \ + *__myhsp = _hn; \ + ++(_hn); \ +} while (0) + /* ---------- * pglz_out_ctrl - @@ -372,28 +421,42 @@ do { \ * ---------- */ static inline int -pglz_find_match(PGLZ_HistEntry **hstart, const char *input, const char *end, - int *lenp, int *offp, int good_match, int good_drop) +pglz_find_match(int16 *hstart, const char *input, const char *end, + int *lenp, int *offp, int good_match, int good_drop, + const char *hend, int hindex) { - PGLZ_HistEntry *hent; + int16 hentno; int32 len = 0; int32 off = 0; /* * Traverse the linked history list until a good enough match is found. */ - hent = hstart[pglz_hist_idx(input, end)]; - while (hent) + hentno = hstart[hindex]; + while (hentno != INVALID_ENTRY) { + PGLZ_HistEntry *hent = &hist_entries[hentno]; const char *ip = input; const char *hp = hent->pos; int32 thisoff; int32 thislen; + int32 maxlen; + + maxlen = PGLZ_MAX_MATCH; + if (end - input < maxlen) + maxlen = end - input; + if (hend && (hend - hp < maxlen)) + maxlen = hend - hp; + /* * Stop if the offset does not fit into our tag anymore. */ - thisoff = ip - hp; + if (!hend) + thisoff = ip - hp; + else + thisoff = hend - hp; + if (thisoff >= 0x0fff) break; @@ -413,7 +476,7 @@ pglz_find_match(PGLZ_HistEntry **hstart, const char *input, const char *end, thislen = len; ip += len; hp += len; - while (ip < end && *ip == *hp && thislen < PGLZ_MAX_MATCH) + while (*ip == *hp && thislen < maxlen) { thislen++; ip++; @@ -423,7 +486,7 @@ pglz_find_match(PGLZ_HistEntry **hstart, const char *input, const char *end, } else { - while (ip < end && *ip == *hp && thislen < PGLZ_MAX_MATCH) + while (*ip == *hp && thislen < maxlen) { thislen++; ip++; @@ -443,13 +506,13 @@ pglz_find_match(PGLZ_HistEntry **hstart, const char *input, const char *end, /* * Advance to the next history entry */ - hent = hent->next; + hentno = hent->next; /* * Be happy with lesser good matches the more entries we visited. But * no point in doing calculation if we're at end of list. */ - if (hent) + if (hentno != INVALID_ENTRY) { if (len >= good_match) break; @@ -471,6 +534,29 @@ pglz_find_match(PGLZ_HistEntry **hstart, const char *input, const char *end, return 0; } +static int +choose_hash_size(int slen) +{ + int hashsz; + + /* + * Experiments suggest that these hash sizes work pretty well. A large + * hash table minimizes collision, but has a higher startup cost. For + * a small input, the startup cost dominates. The table size must be + * a power of two. + */ + if (slen < 128) + hashsz = 512; + else if (slen < 256) + hashsz = 1024; + else if (slen < 512) + hashsz = 2048; + else if (slen < 1024) + hashsz = 4096; + else + hashsz = 8192; + return hashsz; +} /* ---------- * pglz_compress - @@ -484,7 +570,7 @@ pglz_compress(const char *source, int32 slen, PGLZ_Header *dest, { unsigned char *bp = ((unsigned char *) dest) + sizeof(PGLZ_Header); unsigned char *bstart = bp; - int hist_next = 0; + int hist_next = 1; bool hist_recycle = false; const char *dp = source; const char *dend = source + slen; @@ -500,6 +586,8 @@ pglz_compress(const char *source, int32 slen, PGLZ_Header *dest, int32 result_size; int32 result_max; int32 need_rate; + int hashsz; + int mask; /* * Our fallback strategy is the default. @@ -555,17 +643,21 @@ pglz_compress(const char *source, int32 slen, PGLZ_Header *dest, else result_max = (slen * (100 - need_rate)) / 100; + hashsz = choose_hash_size(slen); + mask = hashsz - 1; + /* * Initialize the history lists to empty. We do not need to zero the * hist_entries[] array; its entries are initialized as they are used. */ - memset(hist_start, 0, sizeof(hist_start)); + memset(hist_start, 0, hashsz * sizeof(int16)); /* * Compress the source directly into the output buffer. */ while (dp < dend) { + uint32 hindex; /* * If we already exceeded the maximum result size, fail. * @@ -588,8 +680,9 @@ pglz_compress(const char *source, int32 slen, PGLZ_Header *dest, /* * Try to find a match in the history */ + hindex = pglz_hist_idx(dp, dend, mask); if (pglz_find_match(hist_start, dp, dend, &match_len, - &match_off, good_match, good_drop)) + &match_off, good_match, good_drop, NULL, hindex)) { /* * Create the tag and add history entries for all matched @@ -598,9 +691,10 @@ pglz_compress(const char *source, int32 slen, PGLZ_Header *dest, pglz_out_tag(ctrlp, ctrlb, ctrl, bp, match_len, match_off); while (match_len--) { + hindex = pglz_hist_idx(dp, dend, mask); pglz_hist_add(hist_start, hist_entries, hist_next, hist_recycle, - dp, dend); + dp, dend, hindex); dp++; /* Do not do this ++ in the line above! */ /* The macro would do it four times - Jan. */ } @@ -614,7 +708,7 @@ pglz_compress(const char *source, int32 slen, PGLZ_Header *dest, pglz_out_literal(ctrlp, ctrlb, ctrl, bp, *dp); pglz_hist_add(hist_start, hist_entries, hist_next, hist_recycle, - dp, dend); + dp, dend, hindex); dp++; /* Do not do this ++ in the line above! */ /* The macro would do it four times - Jan. */ } @@ -637,6 +731,198 @@ pglz_compress(const char *source, int32 slen, PGLZ_Header *dest, return true; } +/* + * Delta encoding. + * + * The 'source' is encoded using the same pglz algorithm used for compression. + * The difference with pglz_compress is that the back-references refer to + * the 'history', instead of earlier offsets in 'source'. + * + * The encoded result is written to *dest, and its length is returned in + * *finallen. + */ +bool +pglz_delta_encode(const char *source, int32 slen, + const char *history, int32 hlen, + char *dest, uint32 *finallen, + const PGLZ_Strategy *strategy) +{ + unsigned char *bp = ((unsigned char *) dest); + unsigned char *bstart = bp; + const char *dp = source; + const char *dend = source + slen; + const char *hp = history; + const char *hend = history + hlen; + unsigned char ctrl_dummy = 0; + unsigned char *ctrlp = &ctrl_dummy; + unsigned char ctrlb = 0; + unsigned char ctrl = 0; + bool found_match = false; + int32 match_len = 0; + int32 match_off; + int32 result_size; + int32 result_max; + int32 good_match; + int32 good_drop; + int32 need_rate; + int hist_next = 0; + int hashsz; + int mask; + int32 a,b,c,d; + int32 hindex; + + /* + * Tuples of length greater than PGLZ_HISTORY_SIZE are not allowed for + * delta encode as this is the maximum size of history offset. + */ + if (hlen >= PGLZ_HISTORY_SIZE || hlen < 4) + return false; + + /* + * Our fallback strategy is the default. + */ + if (strategy == NULL) + strategy = PGLZ_strategy_default; + + /* + * If the strategy forbids compression (at all or if source chunk size out + * of range), fail. + */ + if (strategy->match_size_good <= 0 || + slen < strategy->min_input_size || + slen > strategy->max_input_size) + return false; + + need_rate = strategy->min_comp_rate; + if (need_rate < 0) + need_rate = 0; + else if (need_rate > 99) + need_rate = 99; + + /* + * Limit the match parameters to the supported range. + */ + good_match = strategy->match_size_good; + if (good_match > PGLZ_MAX_MATCH) + good_match = PGLZ_MAX_MATCH; + else if (good_match < 17) + good_match = 17; + + good_drop = strategy->match_size_drop; + if (good_drop < 0) + good_drop = 0; + else if (good_drop > 100) + good_drop = 100; + + /* + * Compute the maximum result size allowed by the strategy, namely the + * input size minus the minimum wanted compression rate. This had better + * be <= slen, else we might overrun the provided output buffer. + */ + if (slen > (INT_MAX / 100)) + { + /* Approximate to avoid overflow */ + result_max = (slen / 100) * (100 - need_rate); + } + else + result_max = (slen * (100 - need_rate)) / 100; + + hashsz = choose_hash_size(hlen); + mask = hashsz - 1; + + /* + * Initialize the history lists to empty. We do not need to zero the + * hist_entries[] array; its entries are initialized as they are used. + */ + memset(hist_start, 0, hashsz * sizeof(int16)); + + pglz_hash_init(hp, hindex, a,b,c,d); + while (hp < hend - 4) + { + /* + * TODO: It would be nice to behave like the history and the source + * strings were concatenated, so that you could compress using the + * new data, too. + */ + pglz_hash_roll(hp, hindex, a,b,c,d, mask); + pglz_hist_add_no_recycle(hist_start, hist_entries, + hist_next, + hp, hend, hindex); + hp++; /* Do not do this ++ in the line above! */ + } + + /* + * Loop through the input. + */ + match_off = 0; + pglz_hash_init(dp, hindex,a,b,c,d); + while (dp < dend - 4) + { + /* + * If we already exceeded the maximum result size, fail. + * + * We check once per loop; since the loop body could emit as many as 4 + * bytes (a control byte and 3-byte tag), PGLZ_MAX_OUTPUT() had better + * allow 4 slop bytes. + */ + if (bp - bstart >= result_max) + return false; + + /* + * Try to find a match in the history + */ + pglz_hash_roll(dp,hindex,a,b,c,d,mask); + if (pglz_find_match(hist_start, dp, dend, &match_len, + &match_off, good_match, good_drop, hend, hindex)) + { + /* + * Create the tag and add history entries for all matched + * characters. + */ + pglz_out_tag(ctrlp, ctrlb, ctrl, bp, match_len, match_off); + dp += match_len; + found_match = true; + pglz_hash_init(dp, hindex,a,b,c,d); + } + else + { + /* + * No match found. Copy one literal byte. + */ + pglz_out_literal(ctrlp, ctrlb, ctrl, bp, *dp); + dp++; /* Do not do this ++ in the line above! */ + /* The macro would do it four times - Jan. */ + } + } + + if (!found_match) + return false; + + /* Handle the last few bytes as literals */ + while (dp < dend) + { + pglz_out_literal(ctrlp, ctrlb, ctrl, bp, *dp); + dp++; /* Do not do this ++ in the line above! */ + } + + /* + * Write out the last control byte and check that we haven't overrun the + * output size allowed by the strategy. + */ + *ctrlp = ctrlb; + result_size = bp - bstart; + +#ifdef DELTA_DEBUG + elog(LOG, "old %d new %d compressed %d", hlen, slen, result_size); +#endif + + /* + * Success - need only fill in the actual length of the compressed datum. + */ + *finallen = result_size; + + return true; +} /* ---------- * pglz_decompress - @@ -740,3 +1026,107 @@ pglz_decompress(const PGLZ_Header *source, char *dest) * That's it. */ } + +/* ---------- + * pglz_delta_decode + * + * Decompresses source into dest. + * To decompress, it uses history if provided. + * ---------- + */ +void +pglz_delta_decode(const char *source, uint32 srclen, + char *dest, uint32 destlen, uint32 *finallen, + const char *history, uint32 histlen) +{ + const unsigned char *sp; + const unsigned char *srcend; + unsigned char *dp; + unsigned char *destend; + const char *hend; + + sp = ((const unsigned char *) source); + srcend = ((const unsigned char *) source) + srclen; + dp = (unsigned char *) dest; + destend = dp + destlen; + hend = history + histlen; + + while (sp < srcend && dp < destend) + { + /* + * Read one control byte and process the next 8 items (or as many as + * remain in the compressed input). + */ + unsigned char ctrl = *sp++; + int ctrlc; + + for (ctrlc = 0; ctrlc < 8 && sp < srcend; ctrlc++) + { + if (ctrl & 1) + { + /* + * Otherwise it contains the match length minus 3 and the + * upper 4 bits of the offset. The next following byte + * contains the lower 8 bits of the offset. If the length is + * coded as 18, another extension tag byte tells how much + * longer the match really was (0-255). + */ + int32 len; + int32 off; + + len = (sp[0] & 0x0f) + 3; + off = ((sp[0] & 0xf0) << 4) | sp[1]; + sp += 2; + if (len == 18) + len += *sp++; + + /* + * Check for output buffer overrun, to ensure we don't clobber + * memory in case of corrupt input. Note: we must advance dp + * here to ensure the error is detected below the loop. We + * don't simply put the elog inside the loop since that will + * probably interfere with optimization. + */ + if (dp + len > destend) + { + dp += len; + break; + } + + /* + * Now we copy the bytes specified by the tag from history + * to OUTPUT. + */ + memcpy(dp, hend - off, len); + dp += len; + } + else + { + /* + * An unset control bit means LITERAL BYTE. So we just + * copy one from INPUT to OUTPUT. + */ + if (dp >= destend) /* check for buffer overrun */ + break; /* do not clobber memory */ + + *dp++ = *sp++; + } + + /* + * Advance the control bit + */ + ctrl >>= 1; + } + } + + /* + * Check we decompressed the right amount. + */ + if (sp != srcend) + elog(PANIC, "compressed data is corrupt"); + + /* + * That's it. + */ + *finallen = ((char *) dp - dest); +} diff --git a/src/backend/utils/misc/guc.c b/src/backend/utils/misc/guc.c index 22ba35f..6ff6b23 100644 --- a/src/backend/utils/misc/guc.c +++ b/src/backend/utils/misc/guc.c @@ -124,6 +124,7 @@ extern char *default_tablespace; extern char *temp_tablespaces; extern bool ignore_checksum_failure; extern bool synchronize_seqscans; +extern int wal_update_compression_ratio; extern int ssl_renegotiation_limit; extern char *SSLCipherSuites; @@ -2410,6 +2411,17 @@ static struct config_int ConfigureNamesInt[] = NULL, NULL, NULL }, + { + /* Not for general use */ + {"wal_update_compression_ratio", PGC_USERSET, DEVELOPER_OPTIONS, + gettext_noop("Sets the compression ratio of delta record for wal update"), + NULL, + }, + &wal_update_compression_ratio, + 25, 0, 100, + NULL, NULL, NULL + }, + /* End-of-list marker */ { {NULL, 0, 0, NULL, NULL}, NULL, 0, 0, 0, NULL, NULL, NULL diff --git a/src/include/access/heapam_xlog.h b/src/include/access/heapam_xlog.h index e58eae5..386277d 100644 --- a/src/include/access/heapam_xlog.h +++ b/src/include/access/heapam_xlog.h @@ -147,13 +147,22 @@ typedef struct xl_heap_update TransactionId old_xmax; /* xmax of the old tuple */ TransactionId new_xmax; /* xmax of the new tuple */ ItemPointerData newtid; /* new inserted tuple id */ - uint8 old_infobits_set; /* infomask bits to set on old tuple */ - bool all_visible_cleared; /* PD_ALL_VISIBLE was cleared */ - bool new_all_visible_cleared; /* same for the page of newtid */ + uint8 old_infobits_set; /* infomask bits to set on old tuple */ + uint8 flags; /* flag bits, see below */ /* NEW TUPLE xl_heap_header AND TUPLE DATA FOLLOWS AT END OF STRUCT */ } xl_heap_update; -#define SizeOfHeapUpdate (offsetof(xl_heap_update, new_all_visible_cleared) + sizeof(bool)) +#define XL_HEAP_UPDATE_ALL_VISIBLE_CLEARED 0x01 /* Indicates as old + * page's all visible + * bit is cleared */ +#define XL_HEAP_UPDATE_NEW_ALL_VISIBLE_CLEARED 0x02 /* Indicates as new + * page's all visible + * bit is cleared */ +#define XL_HEAP_UPDATE_DELTA_ENCODED 0x04 /* Indicates as the + * update operation is + * delta encoded */ + +#define SizeOfHeapUpdate (offsetof(xl_heap_update, flags) + sizeof(uint8)) /* * This is what we need to know about vacuum page cleanup/redirect diff --git a/src/include/access/htup_details.h b/src/include/access/htup_details.h index cd01ecd..1ef550b 100644 --- a/src/include/access/htup_details.h +++ b/src/include/access/htup_details.h @@ -687,6 +687,11 @@ extern HeapTuple heap_modify_tuple(HeapTuple tuple, extern void heap_deform_tuple(HeapTuple tuple, TupleDesc tupleDesc, Datum *values, bool *isnull); +extern bool heap_delta_encode(TupleDesc tupleDesc, HeapTuple oldtup, + HeapTuple newtup, char *encdata, uint32 *enclen); +extern void heap_delta_decode (char *encdata, uint32 enclen, HeapTuple oldtup, + HeapTuple newtup); + /* these three are deprecated versions of the three above: */ extern HeapTuple heap_formtuple(TupleDesc tupleDescriptor, Datum *values, char *nulls); diff --git a/src/include/access/xlog.h b/src/include/access/xlog.h index f8f06c1..56efcac 100644 --- a/src/include/access/xlog.h +++ b/src/include/access/xlog.h @@ -261,6 +261,7 @@ typedef struct CheckpointStatsData extern CheckpointStatsData CheckpointStats; extern XLogRecPtr XLogInsert(RmgrId rmid, uint8 info, XLogRecData *rdata); +extern bool XLogCheckBufferNeedsBackup(Buffer buffer); extern void XLogFlush(XLogRecPtr RecPtr); extern bool XLogBackgroundFlush(void); extern bool XLogNeedsFlush(XLogRecPtr RecPtr); diff --git a/src/include/utils/pg_lzcompress.h b/src/include/utils/pg_lzcompress.h index 4af24a3..5add61a 100644 --- a/src/include/utils/pg_lzcompress.h +++ b/src/include/utils/pg_lzcompress.h @@ -107,6 +107,12 @@ extern const PGLZ_Strategy *const PGLZ_strategy_always; */ extern bool pglz_compress(const char *source, int32 slen, PGLZ_Header *dest, const PGLZ_Strategy *strategy); +extern bool pglz_delta_encode(const char *source, int32 slen, + const char *history, int32 hlen, + char *dest, uint32 *finallen, const PGLZ_Strategy *strategy); extern void pglz_decompress(const PGLZ_Header *source, char *dest); +extern void pglz_delta_decode(const char *source, uint32 srclen, + char *dest, uint32 destlen, uint32 *finallen, + const char *history, uint32 histlen); #endif /* _PG_LZCOMPRESS_H_ */ diff --git a/src/test/regress/expected/update.out b/src/test/regress/expected/update.out index 71b856f..af46df2 100644 --- a/src/test/regress/expected/update.out +++ b/src/test/regress/expected/update.out @@ -97,3 +97,73 @@ SELECT a, b, char_length(c) FROM update_test; (2 rows) DROP TABLE update_test; +-- +-- Test to update continuos and non continuos columns +-- +DROP TABLE IF EXISTS update_test; +NOTICE: table "update_test" does not exist, skipping +CREATE TABLE update_test ( + bser bigserial, + bln boolean, + ename VARCHAR(25), + perf_f float(8), + grade CHAR, + dept CHAR(5) NOT NULL, + dob DATE, + idnum INT, + addr VARCHAR(30) NOT NULL, + destn CHAR(6), + Gend CHAR, + samba BIGINT, + hgt float, + ctime TIME +); +INSERT INTO update_test VALUES ( + nextval('update_test_bser_seq'::regclass), + TRUE, + 'Test', + 7.169, + 'B', + 'CSD', + '2000-01-01', + 520, + 'road2, + streeeeet2, + city2', + 'dcy2', + 'M', + 12000, + 50.4, + '00:00:00.0' +); +SELECT * from update_test; + bser | bln | ename | perf_f | grade | dept | dob | idnum | addr | destn | gend | samba | hgt | ctime +------+-----+-------+--------+-------+-------+------------+-------+-----------------------------+--------+------+-------+------+---------- + 1 | t | Test | 7.169 | B | CSD | 01-01-2000 | 520 | road2, +| dcy2 | M | 12000 | 50.4 | 00:00:00 + | | | | | | | | streeeeet2,+| | | | | + | | | | | | | | city2 | | | | | +(1 row) + +-- update first column +UPDATE update_test SET bser = bser - 1 + 1; +-- update middle column +UPDATE update_test SET perf_f = 8.9; +-- update last column +UPDATE update_test SET ctime = '00:00:00.1'; +-- update 3 continuos columns +UPDATE update_test SET destn = 'dcy2', samba = 0 WHERE Gend = 'M' and dept = 'CSD'; +-- update two non continuos columns +UPDATE update_test SET destn = 'moved', samba = 0; +UPDATE update_test SET bln = FALSE, hgt = 10.1; +-- update causing some column alignment difference +UPDATE update_test SET ename = 'Tes'; +UPDATE update_test SET dept = 'Test'; +SELECT * from update_test; + bser | bln | ename | perf_f | grade | dept | dob | idnum | addr | destn | gend | samba | hgt | ctime +------+-----+-------+--------+-------+-------+------------+-------+-----------------------------+--------+------+-------+------+------------ + 1 | f | Tes | 8.9 | B | Test | 01-01-2000 | 520 | road2, +| moved | M | 0 | 10.1 | 00:00:00.1 + | | | | | | | | streeeeet2,+| | | | | + | | | | | | | | city2 | | | | | +(1 row) + +DROP TABLE update_test; diff --git a/src/test/regress/sql/update.sql b/src/test/regress/sql/update.sql index a8a028f..1806992 100644 --- a/src/test/regress/sql/update.sql +++ b/src/test/regress/sql/update.sql @@ -59,3 +59,70 @@ UPDATE update_test SET c = repeat('x', 10000) WHERE c = 'car'; SELECT a, b, char_length(c) FROM update_test; DROP TABLE update_test; + + +-- +-- Test to update continuos and non continuos columns +-- + +DROP TABLE IF EXISTS update_test; +CREATE TABLE update_test ( + bser bigserial, + bln boolean, + ename VARCHAR(25), + perf_f float(8), + grade CHAR, + dept CHAR(5) NOT NULL, + dob DATE, + idnum INT, + addr VARCHAR(30) NOT NULL, + destn CHAR(6), + Gend CHAR, + samba BIGINT, + hgt float, + ctime TIME +); + +INSERT INTO update_test VALUES ( + nextval('update_test_bser_seq'::regclass), + TRUE, + 'Test', + 7.169, + 'B', + 'CSD', + '2000-01-01', + 520, + 'road2, + streeeeet2, + city2', + 'dcy2', + 'M', + 12000, + 50.4, + '00:00:00.0' +); + +SELECT * from update_test; + +-- update first column +UPDATE update_test SET bser = bser - 1 + 1; + +-- update middle column +UPDATE update_test SET perf_f = 8.9; + +-- update last column +UPDATE update_test SET ctime = '00:00:00.1'; + +-- update 3 continuos columns +UPDATE update_test SET destn = 'dcy2', samba = 0 WHERE Gend = 'M' and dept = 'CSD'; + +-- update two non continuos columns +UPDATE update_test SET destn = 'moved', samba = 0; +UPDATE update_test SET bln = FALSE, hgt = 10.1; + +-- update causing some column alignment difference +UPDATE update_test SET ename = 'Tes'; +UPDATE update_test SET dept = 'Test'; + +SELECT * from update_test; +DROP TABLE update_test;