Merge last written cache lsn with new main branch (#201)

Author: knizhnik

src/backend/access/gin/gininsert.c

@@ -421,8 +421,9 @@ ginbuild(Relation heap, Relation index, IndexInfo *indexInfo)
 		log_newpage_range(index, MAIN_FORKNUM,
 						  0, RelationGetNumberOfBlocks(index),
 						  true);
+		SetLastWrittenLSNForBlockRange(XactLastRecEnd, index->rd_smgr->smgr_rnode.node, MAIN_FORKNUM, 0, RelationGetNumberOfBlocks(index));
+		SetLastWrittenLSNForRelation(XactLastRecEnd, index->rd_smgr->smgr_rnode.node, MAIN_FORKNUM);
 	}
-	SetLastWrittenPageLSN(XactLastRecEnd);
 
 	smgr_end_unlogged_build(index->rd_smgr);
 

src/backend/access/gist/gistbuild.c

@@ -335,9 +335,11 @@ gistbuild(Relation heap, Relation index, IndexInfo *indexInfo)
 			log_newpage_range(index, MAIN_FORKNUM,
 							  0, RelationGetNumberOfBlocks(index),
 							  true);
+			SetLastWrittenLSNForBlockRange(XactLastRecEnd,
+							  index->rd_smgr->smgr_rnode.node,
+							  MAIN_FORKNUM, 0, RelationGetNumberOfBlocks(index));
+			SetLastWrittenLSNForRelation(XactLastRecEnd, index->rd_smgr->smgr_rnode.node, MAIN_FORKNUM);
 		}
-		SetLastWrittenPageLSN(XactLastRecEnd);
-
 		smgr_end_unlogged_build(index->rd_smgr);
 	}
 
@@ -467,7 +469,9 @@ gist_indexsortbuild(GISTBuildState *state)
 
 		lsn = log_newpage(&state->indexrel->rd_node, MAIN_FORKNUM, GIST_ROOT_BLKNO,
 					pagestate->page, true);
-		SetLastWrittenPageLSN(lsn);
+		SetLastWrittenLSNForBlock(lsn, state->indexrel->rd_smgr->smgr_rnode.node,
+								  MAIN_FORKNUM, GIST_ROOT_BLKNO);
+		SetLastWrittenLSNForRelation(lsn, state->indexrel->rd_smgr->smgr_rnode.node, MAIN_FORKNUM);
 	}
 
 	pfree(pagestate->page);

src/backend/access/spgist/spginsert.c

@@ -143,8 +143,10 @@ spgbuild(Relation heap, Relation index, IndexInfo *indexInfo)
 		log_newpage_range(index, MAIN_FORKNUM,
 						  0, RelationGetNumberOfBlocks(index),
 						  true);
+		SetLastWrittenLSNForBlockRange(XactLastRecEnd, index->rd_smgr->smgr_rnode.node,
+						  MAIN_FORKNUM, 0, RelationGetNumberOfBlocks(index));
+		SetLastWrittenLSNForRelation(XactLastRecEnd, index->rd_smgr->smgr_rnode.node, MAIN_FORKNUM);
 	}
-	SetLastWrittenPageLSN(XactLastRecEnd);
 
 	smgr_end_unlogged_build(index->rd_smgr);
 

src/backend/access/transam/xlog.c

@@ -61,6 +61,7 @@
 #include "replication/walreceiver.h"
 #include "replication/walsender.h"
 #include "storage/bufmgr.h"
+#include "storage/buf_internals.h"
 #include "storage/fd.h"
 #include "storage/ipc.h"
 #include "storage/large_object.h"
@@ -113,6 +114,7 @@ int			wal_retrieve_retry_interval = 5000;
 int			max_slot_wal_keep_size_mb = -1;
 bool		track_wal_io_timing = false;
 uint64      predefined_sysidentifier;
+int			lastWrittenLsnCacheSize;
 
 #ifdef WAL_DEBUG
 bool		XLOG_DEBUG = false;
@@ -182,6 +184,26 @@ const struct config_enum_entry recovery_target_action_options[] = {
 	{NULL, 0, false}
 };
 
+
+typedef struct LastWrittenLsnCacheEntry
+{
+	BufferTag	key;
+	XLogRecPtr	lsn;
+	/* double linked list for LRU replacement algorithm */
+	dlist_node	lru_node;
+} LastWrittenLsnCacheEntry;
+
+
+/*
+ * Cache of last written LSN for each relation chunk (hash bucket).
+ * Also to provide request LSN for smgrnblocks, smgrexists there is pseudokey=InvalidBlockId which stores LSN of last
+ * relation metadata update.
+ * Size of the cache is limited by GUC variable lastWrittenLsnCacheSize ("lsn_cache_size"),
+ * pages are replaced using LRU algorithm, based on L2-list.
+ * Access to this cache is protected by 'LastWrittenLsnLock'.
+ */
+static HTAB *lastWrittenLsnCache;
+
 /*
  * Statistics for current checkpoint are collected in this global struct.
  * Because only the checkpointer or a stand-alone backend can perform
@@ -749,7 +771,17 @@ typedef struct XLogCtlData
 	 * XLOG_FPW_CHANGE record that instructs full_page_writes is disabled.
 	 */
 	XLogRecPtr	lastFpwDisableRecPtr;
-	XLogRecPtr  lastWrittenPageLSN;
+
+	/*
+	 * Maximal last written LSN for pages not present in lastWrittenLsnCache
+	 */
+	XLogRecPtr  maxLastWrittenLsn;
+
+	/*
+	 * Double linked list to implement LRU replacement policy for last written LSN cache.
+	 * Access to this list as well as to last written LSN cache is protected by 'LastWrittenLsnLock'.
+	 */
+	dlist_head lastWrittenLsnLRU;
 
 	/* neon: copy of startup's RedoStartLSN for walproposer's use */
 	XLogRecPtr	RedoStartLSN;
@@ -772,6 +804,8 @@ static WALInsertLockPadded *WALInsertLocks = NULL;
  */
 static ControlFileData *ControlFile = NULL;
 
+#define LAST_WRITTEN_LSN_CACHE_BUCKET 1024 /* blocks = 8Mb */
+
 /*
  * Calculate the amount of space left on the page after 'endptr'. Beware
  * multiple evaluation!
@@ -5141,11 +5175,8 @@ LocalProcessControlFile(bool reset)
 	ReadControlFile();
 }
 
-/*
- * Initialization of shared memory for XLOG
- */
-Size
-XLOGShmemSize(void)
+static Size
+XLOGCtlShmemSize(void)
 {
 	Size		size;
 
@@ -5185,6 +5216,16 @@ XLOGShmemSize(void)
 	return size;
 }
 
+/*
+ * Initialization of shared memory for XLOG
+ */
+Size
+XLOGShmemSize(void)
+{
+	return XLOGCtlShmemSize() +
+		hash_estimate_size(lastWrittenLsnCacheSize, sizeof(LastWrittenLsnCacheEntry));
+}
+
 void
 XLOGShmemInit(void)
 {
@@ -5214,6 +5255,15 @@ XLOGShmemInit(void)
 	XLogCtl = (XLogCtlData *)
 		ShmemInitStruct("XLOG Ctl", XLOGShmemSize(), &foundXLog);
 
+	{
+		static HASHCTL info;
+		info.keysize = sizeof(BufferTag);
+		info.entrysize = sizeof(LastWrittenLsnCacheEntry);
+		lastWrittenLsnCache = ShmemInitHash("last_written_lsn_cache",
+											lastWrittenLsnCacheSize, lastWrittenLsnCacheSize,
+											&info,
+											HASH_ELEM | HASH_BLOBS);
+	}
 	localControlFile = ControlFile;
 	ControlFile = (ControlFileData *)
 		ShmemInitStruct("Control File", sizeof(ControlFileData), &foundCFile);
@@ -8117,7 +8167,8 @@ StartupXLOG(void)
 
 	XLogCtl->LogwrtRqst.Write = EndOfLog;
 	XLogCtl->LogwrtRqst.Flush = EndOfLog;
-	XLogCtl->lastWrittenPageLSN = EndOfLog;
+	XLogCtl->maxLastWrittenLsn = EndOfLog;
+	dlist_init(&XLogCtl->lastWrittenLsnLRU);
 
 	LocalSetXLogInsertAllowed();
 
@@ -8893,29 +8944,141 @@ GetInsertRecPtr(void)
 }
 
 /*
- * GetLastWrittenPageLSN -- Returns maximal LSN of written page
+ * GetLastWrittenLSN -- Returns maximal LSN of written page.
+ * It returns an upper bound for the last written LSN of a given page,
+ * either from a cached last written LSN or a global maximum last written LSN.
+ * If rnode is InvalidOid then we calculate maximum among all cached LSN and maxLastWrittenLsn.
+ * If cache is large enough ,iterting through all hash items may be rather expensive.
+ * But GetLastWrittenLSN(InvalidOid) is used only by zenith_dbsize which is not performance critical.
  */
 XLogRecPtr
-GetLastWrittenPageLSN(void)
+GetLastWrittenLSN(RelFileNode rnode, ForkNumber forknum, BlockNumber blkno)
 {
 	XLogRecPtr lsn;
-	SpinLockAcquire(&XLogCtl->info_lck);
-	lsn = XLogCtl->lastWrittenPageLSN;
-	SpinLockRelease(&XLogCtl->info_lck);
+	LastWrittenLsnCacheEntry* entry;
+
+	LWLockAcquire(LastWrittenLsnLock, LW_SHARED);
+
+	/* Maximal last written LSN among all non-cached pages */
+	lsn = XLogCtl->maxLastWrittenLsn;
+
+	if (rnode.relNode != InvalidOid)
+	{
+		BufferTag key;
+		key.rnode = rnode;
+		key.forkNum = forknum;
+		key.blockNum = blkno / LAST_WRITTEN_LSN_CACHE_BUCKET;
+		entry = hash_search(lastWrittenLsnCache, &key, HASH_FIND, NULL);
+		if (entry != NULL)
+			lsn = entry->lsn;
+	}
+	else
+	{
+		HASH_SEQ_STATUS seq;
+		/* Find maximum of all cached LSNs */
+		hash_seq_init(&seq, lastWrittenLsnCache);
+		while ((entry = (LastWrittenLsnCacheEntry *) hash_seq_search(&seq)) != NULL)
+		{
+			if (entry->lsn > lsn)
+				lsn = entry->lsn;
+		}
+	}
+	LWLockRelease(LastWrittenLsnLock);
 
 	return lsn;
 }
 
 /*
- * SetLastWrittenPageLSN -- Set maximal LSN of written page
+ * SetLastWrittenLSNForBlockRange -- Set maximal LSN of written page range.
+ * We maintain cache of last written LSNs with limited size and LRU replacement
+ * policy. To reduce cache size we store max LSN not for each page, but for
+ * bucket (1024 blocks). This cache allows to use old LSN when
+ * requesting pages of unchanged or appended relations.
+ *
+ * rnode.relNode can be InvalidOid, in this case maxLastWrittenLsn is updated.
+ * SetLastWrittenLsn with dummy rnode is used by createdb and dbase_redo functions.
  */
 void
-SetLastWrittenPageLSN(XLogRecPtr lsn)
+SetLastWrittenLSNForBlockRange(XLogRecPtr lsn, RelFileNode rnode, ForkNumber forknum, BlockNumber from, BlockNumber till)
 {
-	SpinLockAcquire(&XLogCtl->info_lck);
-	if (lsn > XLogCtl->lastWrittenPageLSN)
-		XLogCtl->lastWrittenPageLSN = lsn;
-	SpinLockRelease(&XLogCtl->info_lck);
+	if (lsn == InvalidXLogRecPtr)
+		return;
+
+	LWLockAcquire(LastWrittenLsnLock, LW_EXCLUSIVE);
+	if (rnode.relNode == InvalidOid)
+	{
+		if (lsn > XLogCtl->maxLastWrittenLsn)
+			XLogCtl->maxLastWrittenLsn = lsn;
+	}
+	else
+	{
+		LastWrittenLsnCacheEntry* entry;
+		BufferTag key;
+		bool found;
+		BlockNumber bucket;
+
+		key.rnode = rnode;
+		key.forkNum = forknum;
+		for (bucket = from / LAST_WRITTEN_LSN_CACHE_BUCKET;
+			 bucket <= till / LAST_WRITTEN_LSN_CACHE_BUCKET;
+			 bucket++)
+		{
+			key.blockNum = bucket;
+			entry = hash_search(lastWrittenLsnCache, &key, HASH_ENTER, &found);
+			if (found)
+			{
+				if (lsn > entry->lsn)
+					entry->lsn = lsn;
+				/* Unlink from LRU list */
+				dlist_delete(&entry->lru_node);
+			}
+			else
+			{
+				entry->lsn = lsn;
+				if (hash_get_num_entries(lastWrittenLsnCache) > lastWrittenLsnCacheSize)
+				{
+					/* Replace least recently used entry */
+					LastWrittenLsnCacheEntry* victim = dlist_container(LastWrittenLsnCacheEntry, lru_node, dlist_pop_head_node(&XLogCtl->lastWrittenLsnLRU));
+					/* Adjust max LSN for not cached relations/chunks if needed */
+					if (victim->lsn > XLogCtl->maxLastWrittenLsn)
+						XLogCtl->maxLastWrittenLsn = victim->lsn;
+
+					hash_search(lastWrittenLsnCache, victim, HASH_REMOVE, NULL);
+				}
+			}
+			/* Link to the end of LRU list */
+			dlist_push_tail(&XLogCtl->lastWrittenLsnLRU, &entry->lru_node);
+		}
+	}
+	LWLockRelease(LastWrittenLsnLock);
+}
+
+/*
+ * SetLastWrittenLSNForBlock -- Set maximal LSN for block
+ */
+void
+SetLastWrittenLSNForBlock(XLogRecPtr lsn, RelFileNode rnode, ForkNumber forknum, BlockNumber blkno)
+{
+	SetLastWrittenLSNForBlockRange(lsn, rnode, forknum, blkno, blkno);
+}
+
+/*
+ * SetLastWrittenLSNForRelation -- Set maximal LSN for relation metadata
+ */
+void
+SetLastWrittenLSNForRelation(XLogRecPtr lsn, RelFileNode rnode, ForkNumber forknum)
+{
+	SetLastWrittenLSNForBlock(lsn, rnode, forknum, REL_METADATA_PSEUDO_BLOCKNO);
+}
+
+/*
+ * SetLastWrittenLSNForDatabase -- Set maximal LSN for the whole database
+ */
+void
+SetLastWrittenLSNForDatabase(XLogRecPtr lsn)
+{
+	RelFileNode dummyNode = {InvalidOid, InvalidOid, InvalidOid};
+	SetLastWrittenLSNForBlock(lsn, dummyNode, MAIN_FORKNUM, 0);
 }
 
 /*

src/backend/commands/dbcommands.c

@@ -685,7 +685,7 @@ createdb(ParseState *pstate, const CreatedbStmt *stmt)
 
 				lsn = XLogInsert(RM_DBASE_ID,
 								 XLOG_DBASE_CREATE | XLR_SPECIAL_REL_UPDATE);
-				SetLastWrittenPageLSN(lsn);
+				SetLastWrittenLSNForDatabase(lsn);
 			}
 		}
 		table_endscan(scan);
@@ -2363,8 +2363,7 @@ dbase_redo(XLogReaderState *record)
 		 */
 		{
 			XLogRecPtr	lsn = record->EndRecPtr;
-
-			SetLastWrittenPageLSN(lsn);
+			SetLastWrittenLSNForDatabase(lsn);
 		}
 	}
 	else if (info == XLOG_DBASE_DROP)

src/backend/storage/lmgr/lwlocknames.txt

@@ -53,3 +53,4 @@ XactTruncationLock					44
 # 45 was XactTruncationLock until removal of BackendRandomLock
 WrapLimitsVacuumLock				46
 NotifyQueueTailLock					47
+LastWrittenLsnLock					48

src/backend/utils/misc/guc.c

@@ -2357,6 +2357,16 @@ static struct config_int ConfigureNamesInt[] =
 		NULL, NULL, NULL
 	},
 
+	{
+		{"lsn_cache_size", PGC_POSTMASTER, UNGROUPED,
+			gettext_noop("Size of las written LSN cache used by Neon."),
+			NULL
+		},
+		&lastWrittenLsnCacheSize,
+		1024, 10, 1000000, /* 1024 is enough to hold 10GB database with 8Mb bucket */
+		NULL, NULL, NULL
+	},
+
 	{
 		{"temp_buffers", PGC_USERSET, RESOURCES_MEM,
 			gettext_noop("Sets the maximum number of temporary buffers used by each session."),