Two optimizations to the pager: (1) Write dirty pages back to the database

file in order and (2) Keep a separate list of in-memory pages that are in
the checkpoint journal in order to speed a checkpoint commit. (CVS 783)

FossilOrigin-Name: a6ef6657a4377684dc2fce7be2bbf009fd2d2f37

manifest

@@ -1,5 +1,5 @@
-C Try\sto\sbetter\sdetect\swhen\sthe\slibrary\sis\scompiled\sfor\slarge\sfile\ssupport\s(LFS)\nbut\sthe\ssupport\sis\snot\savailable\sin\sthe\shost\sOS\skernel.\s(CVS\s782)
+C Two\soptimizations\sto\sthe\spager:\s(1)\sWrite\sdirty\spages\sback\sto\sthe\sdatabase\nfile\sin\sorder\sand\s(2)\sKeep\sa\sseparate\slist\sof\sin-memory\spages\sthat\sare\sin\nthe\scheckpoint\sjournal\sin\sorder\sto\sspeed\sa\scheckpoint\scommit.\s(CVS\s783)
-D 2002-11-09T00:33:16
+D 2002-11-10T23:32:57
 F Makefile.in d6c9a85c2a5e696843201d090dcf8bf2f8716f2a
 F Makefile.linux-gcc b86a99c493a5bfb402d1d9178dcdc4bd4b32f906
 F README f1de682fbbd94899d50aca13d387d1b3fd3be2dd
@@ -32,7 +32,7 @@ F src/main.c f04f93b8928d6d85976e5137fea146a46de1fd6e
 F src/md5.c fe4f9c9c6f71dfc26af8da63e4d04489b1430565
 F src/os.c caf5a34b35a2d99a58457517261c879ac29b0a05
 F src/os.h 1caaea972d1c0401cfe6300aba51fb0f6fe435c9
-F src/pager.c 292853d08658df23f1044fba1a793a210475964e
+F src/pager.c 27d9b94641e968bafabfd2119d9ba4304ccb69a4
 F src/pager.h 6991c9c2dc5e4c7f2df4d4ba47d1c6458f763a32
 F src/parse.y 469c9636ff713e63c00234662209f11668671ae9
 F src/printf.c 5c50fc1da75c8f5bf432b1ad17d91d6653acd167
@@ -149,7 +149,7 @@ F www/speed.tcl a20a792738475b68756ea7a19321600f23d1d803
 F www/sqlite.tcl ae3dcfb077e53833b59d4fcc94d8a12c50a44098
 F www/tclsqlite.tcl 1db15abeb446aad0caf0b95b8b9579720e4ea331
 F www/vdbe.tcl 2013852c27a02a091d39a766bc87cff329f21218
-P 2008b56fe11e49d52e28f47d14ccd70504e6c094
+P a29d60ecc5ee3f535142a81f56eecbef7875ef22
-R 4ac4711a79ff9ee10d8741795898da69
+R 52d1e18697b72d70d44fdac37bed7445
 U drh
-Z c8abc0129b7b7c57014c7437e465381e
+Z 7623b80b72f443af5ce24d861a13e847

manifest.uuid

@@ -1 +1 @@
-a29d60ecc5ee3f535142a81f56eecbef7875ef22
+a6ef6657a4377684dc2fce7be2bbf009fd2d2f37

src/pager.c

@@ -18,7 +18,7 @@
 ** file simultaneously, or one process from reading the database while
 ** another is writing.
 **
-** @(#) $Id: pager.c,v 1.56 2002/11/09 00:33:16 drh Exp $
+** @(#) $Id: pager.c,v 1.57 2002/11/10 23:32:57 drh Exp $
 */
 #include "os.h"         /* Must be first to enable large file support */
 #include "sqliteInt.h"
@@ -74,6 +74,8 @@ struct PgHdr {
   int nRef;                      /* Number of users of this page */
   PgHdr *pNextFree, *pPrevFree;  /* Freelist of pages where nRef==0 */
   PgHdr *pNextAll, *pPrevAll;    /* A list of all pages */
+  PgHdr *pNextCkpt, *pPrevCkpt;  /* List of pages in the checkpoint journal */
+  PgHdr *pSort;                  /* Next in list of pages to be written */
   u8 inJournal;                  /* TRUE if has been written to journal */
   u8 inCkpt;                     /* TRUE if written to the checkpoint journal */
   u8 dirty;                      /* TRUE if we need to write back changes */
@@ -130,6 +132,7 @@ struct Pager {
   u8 *aInCkpt;                /* One bit for each page in the database */
   PgHdr *pFirst, *pLast;      /* List of free pages */
   PgHdr *pAll;                /* List of all pages */
+  PgHdr *pCkpt;               /* List of pages in the checkpoint journal */
   PgHdr *aHash[N_PG_HASH];    /* Hash table to map page number of PgHdr */
 };
 
@@ -253,6 +256,45 @@ static int pager_errcode(Pager *pPager){
   return rc;
 }
 
+/*
+** Add or remove a page from the list of all pages that are in the
+** checkpoint journal.
+**
+** The Pager keeps a separate list of pages that are currently in
+** the checkpoint journal.  This helps the sqlitepager_ckpt_commit()
+** routine run MUCH faster for the common case where there are many
+** pages in memory but only a few are in the checkpoint journal.
+*/
+static void page_add_to_ckpt_list(PgHdr *pPg){
+  Pager *pPager = pPg->pPager;
+  if( pPg->inCkpt ) return;
+  assert( pPg->pPrevCkpt==0 && pPg->pNextCkpt==0 );
+  pPg->pPrevCkpt = 0;
+  if( pPager->pCkpt ){
+    pPager->pCkpt->pPrevCkpt = pPg;
+  }
+  pPg->pNextCkpt = pPager->pCkpt;
+  pPager->pCkpt = pPg;
+  pPg->inCkpt = 1;
+}
+static void page_remove_from_ckpt_list(PgHdr *pPg){
+  if( !pPg->inCkpt ) return;
+  if( pPg->pPrevCkpt ){
+    assert( pPg->pPrevCkpt->pNextCkpt==pPg );
+    pPg->pPrevCkpt->pNextCkpt = pPg->pNextCkpt;
+  }else{
+    assert( pPg->pPager->pCkpt==pPg );
+    pPg->pPager->pCkpt = pPg->pNextCkpt;
+  }
+  if( pPg->pNextCkpt ){
+    assert( pPg->pNextCkpt->pPrevCkpt==pPg );
+    pPg->pNextCkpt->pPrevCkpt = pPg->pPrevCkpt;
+  }
+  pPg->pNextCkpt = 0;
+  pPg->pPrevCkpt = 0;
+  pPg->inCkpt = 0;
+}
+
 /*
 ** Find a page in the hash table given its page number.  Return
 ** a pointer to the page or NULL if not found.
@@ -752,6 +794,41 @@ int sqlitepager_ref(void *pData){
   return SQLITE_OK;
 }
 
+/*
+** The parameters are pointers to the head of two sorted lists
+** of page headers.  Merge these two lists together and return
+** a single sorted list.  This routine forms the core of the 
+** merge-sort algorithm that sorts dirty pages into accending
+** order prior to writing them back to the disk.
+**
+** In the case of a tie, left sorts in front of right.
+**
+** Headers are sorted in order of ascending page number.
+*/
+static PgHdr *page_merge(PgHdr *pLeft, PgHdr *pRight){
+  PgHdr sHead;
+  PgHdr *pTail;
+  pTail = &sHead;
+  pTail->pSort = 0;
+  while( pLeft && pRight ){
+    if( pLeft->pgno<=pRight->pgno ){
+      pTail->pSort = pLeft;
+      pLeft = pLeft->pSort;
+    }else{
+      pTail->pSort = pRight;
+      pRight = pRight->pSort;
+    }
+    pTail = pTail->pSort;
+  }
+  if( pLeft ){
+    pTail->pSort = pLeft;
+  }else if( pRight ){
+    pTail->pSort = pRight;
+  }
+  return sHead.pSort;
+}
+
+
 /*
 ** Sync the journal and then write all free dirty pages to the database
 ** file.
@@ -768,10 +845,24 @@ int sqlitepager_ref(void *pData){
 ** If we are writing to temporary database, there is no need to preserve
 ** the integrity of the journal file, so we can save time and skip the
 ** fsync().
+**
+** This routine goes to the extra trouble of sorting all the dirty
+** pages by their page number prior to writing them.  Tests show that
+** writing pages in order by page number gives a modest speed improvement
+** under Linux.  
 */
 static int syncAllPages(Pager *pPager){
   PgHdr *pPg;
+  PgHdr *pToWrite;
+# define NSORT 28
+  Pgno lastPgno;
+  int i;
+  PgHdr *apSorter[NSORT];
   int rc = SQLITE_OK;
+
+  /* Sync the journal before modifying the main database
+  ** (assuming there is a journal and it needs to be synced.)
+  */
   if( pPager->needSync ){
     if( !pPager->tempFile ){
       rc = sqliteOsSync(&pPager->jfd);
@@ -779,13 +870,57 @@ static int syncAllPages(Pager *pPager){
     }
     pPager->needSync = 0;
   }
+
+  /* Create a list of all dirty pages
+  */
+  pToWrite = 0;
   for(pPg=pPager->pFirst; pPg; pPg=pPg->pNextFree){
     if( pPg->dirty ){
+      pPg->pSort = pToWrite;
+      pToWrite = pPg;
+    }
+  }
+
+  /* Sort the list of dirty pages into accending order by
+  ** page number
+  */
+  for(i=0; i<NSORT; i++){
+    apSorter[i] = 0;
+  }
+  while( pToWrite ){
+    pPg = pToWrite;
+    pToWrite = pPg->pSort;
+    pPg->pSort = 0;
+    for(i=0; i<NSORT-1; i++){
+      if( apSorter[i]==0 ){
+        apSorter[i] = pPg;
+        break;
+      }else{
+        pPg = page_merge(apSorter[i], pPg);
+        apSorter[i] = 0;
+      }
+    }
+    if( i>=NSORT-1 ){
+      apSorter[NSORT-1] = page_merge(apSorter[NSORT-1],pPg);
+    }
+  }
+  pToWrite = 0;
+  for(i=0; i<NSORT; i++){
+    pToWrite = page_merge(apSorter[i], pToWrite);
+  }
+
+  /* Write all dirty pages back to the database and mark
+  ** them all clean.
+  */
+  lastPgno = 0;
+  for(pPg=pToWrite; pPg; pPg=pPg->pSort){
+    if( lastPgno==0 || pPg->pgno!=lastPgno-1 ){
       sqliteOsSeek(&pPager->fd, (pPg->pgno-1)*SQLITE_PAGE_SIZE);
+    }
     rc = sqliteOsWrite(&pPager->fd, PGHDR_TO_DATA(pPg), SQLITE_PAGE_SIZE);
     if( rc!=SQLITE_OK ) break;
     pPg->dirty = 0;
-    }
+    lastPgno = pPg->pgno;
   }
   return rc;
 }
@@ -984,10 +1119,11 @@ int sqlitepager_get(Pager *pPager, Pgno pgno, void **ppPage){
     }else{
       pPg->inJournal = 0;
     }
-    if( pPager->aInCkpt && (int)pgno<=pPager->ckptSize ){
+    if( pPager->aInCkpt && (int)pgno<=pPager->ckptSize
-      pPg->inCkpt = (pPager->aInCkpt[pgno/8] & (1<<(pgno&7)))!=0;
+             && (pPager->aInCkpt[pgno/8] & (1<<(pgno&7)))!=0 ){
+      page_add_to_ckpt_list(pPg);
     }else{
-      pPg->inCkpt = 0;
+      page_remove_from_ckpt_list(pPg);
     }
     pPg->dirty = 0;
     pPg->nRef = 1;
@@ -1248,7 +1384,7 @@ int sqlitepager_write(void *pData){
     pPg->inJournal = 1;
     if( pPager->ckptInUse ){
       pPager->aInCkpt[pPg->pgno/8] |= 1<<(pPg->pgno&7);
-      pPg->inCkpt = 1;
+      page_add_to_ckpt_list(pPg);
     }
   }
 
@@ -1268,7 +1404,7 @@ int sqlitepager_write(void *pData){
     }
     assert( pPager->aInCkpt!=0 );
     pPager->aInCkpt[pPg->pgno/8] |= 1<<(pPg->pgno&7);
-    pPg->inCkpt = 1;
+    page_add_to_ckpt_list(pPg);
   }
 
   /* Update the database size and return.
@@ -1352,14 +1488,14 @@ void sqlitepager_dont_rollback(void *pData){
     pPg->inJournal = 1;
     if( pPager->ckptInUse ){
       pPager->aInCkpt[pPg->pgno/8] |= 1<<(pPg->pgno&7);
-      pPg->inCkpt = 1;
+      page_add_to_ckpt_list(pPg);
     }
   }
   if( pPager->ckptInUse && !pPg->inCkpt && (int)pPg->pgno<=pPager->ckptSize ){
     assert( pPg->inJournal || (int)pPg->pgno>pPager->origDbSize );
     assert( pPager->aInCkpt!=0 );
     pPager->aInCkpt[pPg->pgno/8] |= 1<<(pPg->pgno&7);
-    pPg->inCkpt = 1;
+    page_add_to_ckpt_list(pPg);
   }
 }
 
@@ -1521,15 +1657,19 @@ ckpt_begin_failed:
 */
 int sqlitepager_ckpt_commit(Pager *pPager){
   if( pPager->ckptInUse ){
-    PgHdr *pPg;
+    PgHdr *pPg, *pNext;
     sqliteOsSeek(&pPager->cpfd, 0);
     sqliteOsTruncate(&pPager->cpfd, 0);
     pPager->ckptInUse = 0;
     sqliteFree( pPager->aInCkpt );
     pPager->aInCkpt = 0;
-    for(pPg=pPager->pAll; pPg; pPg=pPg->pNextAll){
+    for(pPg=pPager->pCkpt; pPg; pPg=pNext){
+      pNext = pPg->pNextCkpt;
+      assert( pPg->inCkpt );
       pPg->inCkpt = 0;
+      pPg->pPrevCkpt = pPg->pNextCkpt = 0;
     }
+    pPager->pCkpt = 0;
   }
   return SQLITE_OK;
 }