venti: fix sync deadlock, add /proc stub

src/cmd/venti/srv/arena.c

@@ -293,13 +293,12 @@ ZZZ question: should this distinguish between an arena
 filling up and real errors writing the clump?
  */
 u64int
-writeaclump(Arena *arena, Clump *c, u8int *clbuf, u64int start, u64int *pa)
+writeaclump(Arena *arena, Clump *c, u8int *clbuf)
 {
 	DBlock *b;
 	u64int a, aa;
 	u32int clump, n, nn, m, off, blocksize;
 	int ok;
-	AState as;
 
 	n = c->info.size + ClumpSize + U32Size;
 	qlock(&arena->lock);
@@ -309,10 +308,6 @@ writeaclump(Arena *arena, Clump *c, u8int *clbuf, u64int start, u64int *pa)
 		if(!arena->memstats.sealed){
 			logerr(EOk, "seal memstats %s", arena->name);
 			arena->memstats.sealed = 1;
-			as.arena = arena;
-			as.aa = start+aa;
-			as.stats = arena->memstats;
-			setdcachestate(&as);
 		}
 		qunlock(&arena->lock);
 		return TWID64;
@@ -390,14 +385,6 @@ NoCIG:
 	writeclumpinfo(arena, clump, &c->info);
 	wbarena(arena);
 
-	/* set up for call to setdcachestate */
-	as.arena = arena;
-	as.aa = start+arena->memstats.used;
-	as.stats = arena->memstats;
-
-	/* update this before calling setdcachestate so it cannot be behind dcache.diskstate */
-	*pa = start+aa;
-	setdcachestate(&as);
 	qunlock(&arena->lock);
 
 	return aa;

src/cmd/venti/srv/buildindex.c

@@ -36,7 +36,7 @@ static void	arenapartproc(void*);
 void
 usage(void)
 {
-	fprint(2, "usage: buildindex [-b] [-i isect]... [-M imem] venti.conf\n");
+	fprint(2, "usage: buildindex [-bd] [-i isect]... [-M imem] venti.conf\n");
 	threadexitsall("usage");
 }
 
@@ -54,13 +54,13 @@ threadmain(int argc, char *argv[])
 	case 'b':
 		bloom = 1;
 		break;
+	case 'd':	/* debugging - make sure to run all 3 passes */
+		dumb = 1;
+		break;
 	case 'i':
 		isect = vtrealloc(isect, (nisect+1)*sizeof(isect[0]));
 		isect[nisect++] = EARGF(usage());
 		break;
-	case 'd':	/* debugging - make sure to run all 3 passes */
-		dumb = 1;
-		break;
 	case 'M':
 		imem = unittoull(EARGF(usage()));
 		break;
@@ -222,22 +222,28 @@ arenapartproc(void *v)
 		if(a->memstats.clumps)
 			fprint(2, "%T arena %s: %d entries\n", 
 				a->name, a->memstats.clumps);
-		addr = ix->amap[i].start;
+		/*
-		for(clump=0; clump<a->memstats.clumps; clump+=n){
+		 * Running the loop backwards accesses the 
+		 * clump info blocks forwards, since they are
+		 * stored in reverse order at the end of the arena.
+		 * This speeds things slightly.
+		 */
+		addr = ix->amap[i].start + a->memstats.used;
+		for(clump=a->memstats.clumps; clump > 0; clump-=n){
 			n = ClumpChunks;
-			if(n > a->memstats.clumps - clump)
+			if(n > clump)
-				n = a->memstats.clumps - clump;
+				n = clump;
-			if(readclumpinfos(a, clump, cis, n) != n){
+			if(readclumpinfos(a, clump-n, cis, n) != n){
 				fprint(2, "%T arena %s: directory read: %r\n", a->name);
 				errors = 1;
 				break;
 			}
-			for(j=0; j<n; j++){
+			for(j=n-1; j>=0; j--){
 				ci = &cis[j];
 				ie.ia.type = ci->type;
 				ie.ia.size = ci->uncsize;
+				addr -= ci->size + ClumpSize;
 				ie.ia.addr = addr;
-				addr += ci->size + ClumpSize;
 				ie.ia.blocks = (ci->size + ClumpSize + (1<<ABlockLog)-1) >> ABlockLog;
 				scorecp(ie.score, ci->score);
 				if(ci->type == VtCorruptType)
@@ -253,6 +259,8 @@ arenapartproc(void *v)
 				}
 			}
 		}
+		if(addr != ix->amap[i].start)
+			fprint(2, "%T arena %s: clump miscalculation %lld != %lld\n", a->name, addr, ix->amap[i].start);
 	}
 	add(&arenaentries, tot);
 	add(&skipentries, nskip);

src/cmd/venti/srv/checkarenas.c

@@ -24,7 +24,7 @@ checkarena(Arena *arena, int scan, int fix)
 
 	err = 0;
 	for(;;){
-		e = syncarena(arena, 0, 1000, 0, fix);
+		e = syncarena(arena, 1000, 0, fix);
 		err |= e;
 		if(!(e & SyncHeader))
 			break;

src/cmd/venti/srv/clump.c

@@ -62,19 +62,17 @@ storeclump(Index *ix, ZBlock *zb, u8int *sc, int type, u32int creator, IAddr *ia
 	memset(cb->data+ClumpSize+dsize, 0, 4);
 	cl.info.size = dsize;
 
-	ia->addr = 0;
+	a = writeiclump(ix, &cl, cb->data);
-	ia->type = type;
-	ia->size = size;
-	ia->blocks = (dsize + ClumpSize + (1 << ABlockLog) - 1) >> ABlockLog;
-
-	a = writeiclump(ix, &cl, cb->data, &ia->addr);
-
 	trace(TraceLump, "storeclump exit %lld", a);
-
 	freezblock(cb);
 	if(a == TWID64)
 		return -1;
 
+	ia->addr = a;
+	ia->type = type;
+	ia->size = size;
+	ia->blocks = (dsize + ClumpSize + (1 << ABlockLog) - 1) >> ABlockLog;
+
 /*
 	qlock(&stats.lock);
 	stats.clumpwrites++;

src/cmd/venti/srv/dat.h

@@ -466,6 +466,8 @@ struct Index
 	AMap		*smap;			/* mapping of buckets to index sections */
 	int		narenas;
 	AMap		*amap;			/* mapping from index addesses to arenas */
+	
+	QLock	writing;
 };
 
 /*

src/cmd/venti/srv/dcache.c

@@ -55,9 +55,6 @@ struct DCache
 	u8int		*mem;			/* memory for all block descriptors */
 	int		ndirty;			/* number of dirty blocks */
 	int		maxdirty;		/* max. number of dirty blocks */
-
-	AState	diskstate;
-	AState	state;
 };
 
 typedef struct Ra Ra;
@@ -123,26 +120,6 @@ initdcache(u32int mem)
 	vtproc(delaykickroundproc, &dcache.round);
 }
 
-void
-setdcachestate(AState *a)
-{
-	trace(TraceBlock, "setdcachestate %s 0x%llux clumps %d", a->arena ? a->arena->name : nil, a->aa, a->stats.clumps);
-	qlock(&dcache.lock);
-	dcache.state = *a;
-	qunlock(&dcache.lock);
-}
-
-AState
-diskstate(void)
-{
-	AState a;
-
-	qlock(&dcache.lock);
-	a = dcache.diskstate;
-	qunlock(&dcache.lock);
-	return a;
-}
-
 static u32int
 pbhash(u64int addr)
 {
@@ -637,7 +614,6 @@ flushproc(void *v)
 	int i, j, n;
 	ulong t0;
 	DBlock *b, **write;
-	AState as;
 
 	USED(v);
 	threadsetname("flushproc");
@@ -648,10 +624,6 @@ flushproc(void *v)
 		t0 = nsec()/1000;
 		trace(TraceProc, "build t=%lud", (ulong)(nsec()/1000)-t0);
 
-		qlock(&dcache.lock);
-		as = dcache.state;
-		qunlock(&dcache.lock);
-
 		write = dcache.write;
 		n = 0;
 		for(i=0; i<dcache.nblocks; i++){
@@ -688,7 +660,6 @@ flushproc(void *v)
 		 */
 		trace(TraceProc, "undirty.%d t=%lud", j, (ulong)(nsec()/1000)-t0);
 		qlock(&dcache.lock);
-		dcache.diskstate = as;
 		for(i=0; i<n; i++){
 			b = write[i];
 			--dcache.ndirty;

src/cmd/venti/srv/fixarenas.c

@@ -661,7 +661,7 @@ isonearena(void)
 	return u32(pagein(0, Block)) == ArenaHeadMagic;
 }
 
-static int tabsizes[] = { 16*1024, 64*1024, 512*1024, };
+static int tabsizes[] = { 16*1024, 64*1024, 512*1024, 768*1024, };
 /*
  * Poke around on the disk to guess what the ArenaPart numbers are.
  */
@@ -807,8 +807,9 @@ guessgeometry(void)
 	 * Fmtarenas used to use 64k tab, now uses 512k tab.
 	 */
 	if(ap.arenabase == 0){
+		print("trying standard arena bases...\n");
 		for(i=0; i<nelem(tabsizes); i++){
-			ap.arenabase = ROUNDUP(PartBlank+HeadSize, ap.blocksize);
+			ap.arenabase = ROUNDUP(PartBlank+HeadSize+tabsizes[i], ap.blocksize);
 			p = pagein(ap.arenabase, Block);
 			if(u32(p) == ArenaHeadMagic)
 				break;
@@ -1554,7 +1555,7 @@ guessarena(vlong offset0, int anum, ArenaHead *head, Arena *arena,
 	bcit = cibuf;
 	ecit = cibuf+ncibuf;
 	
-	smart = 1;
+	smart = 0;	/* Somehow the smart code doesn't do corrupt clumps right. */
 Again:
 	nbad = 0;
 	ci = bci;

src/cmd/venti/srv/fns.h

@@ -29,7 +29,6 @@ void		delaykickroundproc(void*);
 void		dirtydblock(DBlock*, int);
 void		diskaccess(int);
 void		disksched(void);
-AState	diskstate(void);
 void		*emalloc(ulong);
 void		emptydcache(void);
 void		emptyicache(void);
@@ -60,6 +59,7 @@ vlong	hargint(HConnect*, char*, vlong);
 int		hdebug(HConnect*);
 int		hdisk(HConnect*);
 int		hnotfound(HConnect*);
+int		hproc(HConnect*);
 int		hsethtml(HConnect*);
 int		hsettext(HConnect*);
 int		httpdinit(char *address, char *webroot);
@@ -68,6 +68,7 @@ IEntry*	icachedirty(u32int, u32int, u64int);
 ulong	icachedirtyfrac(void);
 void		icacheclean(IEntry*);
 int		icachelookup(u8int *score, int type, IAddr *ia);
+AState	icachestate(void);
 int		ientrycmp(const void *vie1, const void *vie2);
 char		*ifileline(IFile *f);
 int		ifilename(IFile *f, char *dst);
@@ -91,7 +92,7 @@ Part*		initpart(char *name, int mode);
 void		initround(Round*, char*, int);
 int		initventi(char *config, Config *conf);
 void		insertlump(Lump *lump, Packet *p);
-int		insertscore(u8int *score, IAddr *ia, int state);
+int		insertscore(u8int *score, IAddr *ia, int state, AState *as);
 void		kickdcache(void);
 void		kickicache(void);
 void		kickround(Round*, int wait);
@@ -156,7 +157,6 @@ int		runconfig(char *config, Config*);
 int		scorecmp(u8int *, u8int *);
 void		scoremem(u8int *score, u8int *buf, int size);
 void		setatailstate(AState*);
-void		setdcachestate(AState*);
 void		seterr(int severity, char *fmt, ...);
 void		setstat(int, long);
 void		settrace(char *type);
@@ -170,9 +170,8 @@ int		strscore(char *s, u8int *score);
 int		stru32int(char *s, u32int *r);
 int		stru64int(char *s, u64int *r);
 void		sumarena(Arena *arena);
-int		syncarena(Arena *arena, u64int start, u32int n, int zok, int fix);
+int		syncarena(Arena *arena, u32int n, int zok, int fix);
-int		syncarenaindex(Index *ix, Arena *arena, u32int clump, u64int a, int fix, int *pflush, int check);
+int		syncindex(Index *ix);
-int		syncindex(Index *ix, int fix, int mustflushicache, int check);
 void		trace(char *type, char*, ...);
 void		traceinit(void);
 int		u64log2(u64int v);
@@ -201,12 +200,12 @@ void		wbbloomhead(Bloom*);
 int		wbisect(ISect *is);
 int		wbindex(Index *ix);
 int		whackblock(u8int *dst, u8int *src, int ssize);
-u64int		writeaclump(Arena *a, Clump *c, u8int *clbuf, u64int, u64int*);
+u64int		writeaclump(Arena *a, Clump *c, u8int *clbuf);
 u32int		writearena(Arena *arena, u64int aa, u8int *clbuf, u32int n);
 int		writebloom(Bloom*);
 int		writeclumpinfo(Arena *arean, int clump, ClumpInfo *ci);
 int		writepng(Hio*, Memimage*);
-u64int		writeiclump(Index *ix, Clump *c, u8int *clbuf, u64int*);
+u64int		writeiclump(Index *ix, Clump *c, u8int *clbuf);
 int		writelump(Packet *p, u8int *score, int type, u32int creator, uint ms);
 int		writepart(Part *part, u64int addr, u8int *buf, u32int n);
 int		writeqlump(Lump *u, Packet *p, int creator, uint ms);

src/cmd/venti/srv/hproc.c

@@ -0,0 +1,17 @@
+#include "stdinc.h"
+#include "dat.h"
+#include "fns.h"
+#include "xml.h"
+
+int
+hproc(HConnect *c)
+{
+	int r;
+	
+	if((r = hsettext(c)) < 0)
+		return r;
+	hprint(&c->hout, "/proc only implemented on Plan 9\n");
+	hflush(&c->hout);
+	return 0;
+}
+

src/cmd/venti/srv/httpd.c

@@ -69,6 +69,7 @@ httpdinit(char *address, char *dir)
 	httpdobj("/emptydcache", hdcacheempty);
 	httpdobj("/disk", hdisk);
 	httpdobj("/debug", hdebug);
+	httpdobj("/proc/", hproc);
 
 	if(vtproc(listenproc, address) < 0)
 		return -1;
@@ -565,11 +566,11 @@ darena(Hio *hout, Arena *arena)
 	if(scorecmp(zeroscore, arena->score) != 0)
 		hprint(hout, "\tscore=%V\n", arena->score);
 
-	hprint(hout, "\tmem: clumps=%d compressed clumps=%d data=%,lld compressed data=%,lld storage=%,lld\n",
+	hprint(hout, "\twritten: clumps=%d compressed clumps=%d data=%,lld compressed data=%,lld storage=%,lld\n",
 		arena->memstats.clumps, arena->memstats.cclumps, arena->memstats.uncsize,
 		arena->memstats.used - arena->memstats.clumps * ClumpSize,
 		arena->memstats.used + arena->memstats.clumps * ClumpInfoSize);
-	hprint(hout, "\tdisk: clumps=%d compressed clumps=%d data=%,lld compressed data=%,lld storage=%,lld\n",
+	hprint(hout, "\tindexed: clumps=%d compressed clumps=%d data=%,lld compressed data=%,lld storage=%,lld\n",
 		arena->diskstats.clumps, arena->diskstats.cclumps, arena->diskstats.uncsize,
 		arena->diskstats.used - arena->diskstats.clumps * ClumpSize,
 		arena->diskstats.used + arena->diskstats.clumps * ClumpInfoSize);

src/cmd/venti/srv/icache.c

@@ -20,6 +20,7 @@ struct ICache
 	IEntry	dirty;
 	u32int	maxdirty;
 	u32int	ndirty;
+	AState	as;
 
 	ISum	**sum;
 	int		nsum;
@@ -398,7 +399,7 @@ icachelookup(u8int score[VtScoreSize], int type, IAddr *ia)
 }
 
 int
-insertscore(u8int score[VtScoreSize], IAddr *ia, int state)
+insertscore(u8int score[VtScoreSize], IAddr *ia, int state, AState *as)
 {
 	ISum *toload;
 
@@ -409,6 +410,13 @@ insertscore(u8int score[VtScoreSize], IAddr *ia, int state)
 	else{
 		assert(state == IEDirty);
 		toload = nil;
+		if(as == nil)
+			fprint(2, "%T insertscore IEDirty without as; called from %lux\n", getcallerpc(&score));
+		else{
+			if(icache.as.aa > as->aa)
+				fprint(2, "%T insertscore: aa moving backward: %#llux -> %#llux\n", icache.as.aa, as->aa);
+			icache.as = *as;
+		}
 	}
 	qunlock(&icache.lock);
 	if(toload){
@@ -443,7 +451,7 @@ lookupscore_untimed(u8int score[VtScoreSize], int type, IAddr *ia)
 	if(loadientry(mainindex, score, type, &d) < 0)
 		return -1;
 	
-	insertscore(score, &d.ia, IEClean);
+	insertscore(score, &d.ia, IEClean, nil);
 	*ia = d.ia;
 	return 0;
 }
@@ -507,6 +515,16 @@ icachedirty(u32int lo, u32int hi, u64int limit)
 	return dirty;
 }
 
+AState
+icachestate(void)
+{
+	AState as;
+
+	qlock(&icache.lock);
+	as = icache.as;
+	qunlock(&icache.lock);
+	return as;
+}
 
 /*
  * The singly-linked non-circular list of index entries ie

src/cmd/venti/srv/icachewrite.c

@@ -12,7 +12,7 @@ static void icachewritecoord(void*);
 static IEntry *iesort(IEntry*);
 
 int icachesleeptime = 1000;	/* milliseconds */
-int minicachesleeptime = 50;
+int minicachesleeptime = 0;
 
 enum
 {
@@ -242,18 +242,20 @@ icachewritecoord(void *v)
 	threadsetname("icachewritecoord");
 
 	ix = mainindex;
-	iwrite.as = diskstate();
+	iwrite.as = icachestate();
 
 	for(;;){
 		trace(TraceProc, "icachewritecoord sleep");
 		waitforkick(&iwrite.round);
 		trace(TraceWork, "start");
-		as = diskstate();
+		as = icachestate();
 		if(as.arena==iwrite.as.arena && as.aa==iwrite.as.aa){
 			/* will not be able to do anything more than last flush - kick disk */
+			fprint(2, "icache: nothing to do - kick dcache\n");
 			trace(TraceProc, "icachewritecoord kick dcache");
 			kickdcache();
 			trace(TraceProc, "icachewritecoord kicked dcache");
+			goto SkipWork;	/* won't do anything; don't bother rewriting bloom filter */
 		}
 		iwrite.as = as;
 
@@ -271,9 +273,11 @@ icachewritecoord(void *v)
 				err |= recvul(ix->bloom->writedonechan);
 
 			trace(TraceProc, "icachewritecoord donewrite err=%d", err);
-			if(err == 0)
+			if(err == 0){
 				setatailstate(&iwrite.as);
+			}
 		}
+	SkipWork:
 		icacheclean(nil);	/* wake up anyone waiting */
 		trace(TraceWork, "finish");
 		addstat(StatIcacheFlush, 1);

src/cmd/venti/srv/index.c

@@ -541,20 +541,33 @@ ZZZ question: should this distinguish between an arena
 filling up and real errors writing the clump?
  */
 u64int
-writeiclump(Index *ix, Clump *c, u8int *clbuf, u64int *pa)
+writeiclump(Index *ix, Clump *c, u8int *clbuf)
 {
 	u64int a;
 	int i;
+	IAddr ia;
+	AState as;
 
 	trace(TraceLump, "writeiclump enter");
+	qlock(&ix->writing);
 	for(i = ix->mapalloc; i < ix->narenas; i++){
-		a = writeaclump(ix->arenas[i], c, clbuf, ix->amap[i].start, pa);
+		a = writeaclump(ix->arenas[i], c, clbuf);
 		if(a != TWID64){
-			ix->mapalloc = i;	/* assuming write is atomic, race is okay */
+			ix->mapalloc = i;
+			ia.addr = ix->amap[i].start + a;
+			ia.type = c->info.type;
+			ia.size = c->info.uncsize;
+			ia.blocks = (c->info.size + ClumpSize + (1<<ABlockLog) - 1) >> ABlockLog;
+			as.arena = ix->arenas[i];
+			as.aa = ia.addr;
+			as.stats = as.arena->memstats;
+			insertscore(c->info.score, &ia, IEDirty, &as);
+			qunlock(&ix->writing);
 			trace(TraceLump, "writeiclump exit");
-			return a;
+			return ia.addr;
 		}
 	}
+	qunlock(&ix->writing);
 
 	seterr(EAdmin, "no space left in arenas");
 	trace(TraceLump, "writeiclump failed");

src/cmd/venti/srv/lump.c

@@ -173,8 +173,6 @@ writeqlump(Lump *u, Packet *p, int creator, uint ms)
 	flat = packet2zblock(p, packetsize(p));
 	ok = storeclump(mainindex, flat, u->score, u->type, creator, &ia);
 	freezblock(flat);
-	if(ok == 0)
-		ok = insertscore(u->score, &ia, IEDirty);
 	if(ok == 0)
 		insertlump(u, p);
 	else

src/cmd/venti/srv/mkfile

@@ -13,6 +13,7 @@ LIBOFILES=\
 	dump.$O\
 	graph.$O\
 	hdisk.$O\
+	hproc.$O\
 	httpd.$O\
 	icache.$O\
 	icachewrite.$O\

src/cmd/venti/srv/syncarena.c

@@ -25,7 +25,7 @@ clumpinfocmp(ClumpInfo *c, ClumpInfo *d)
  * returns 0 if ok, flags if error occurred
  */
 int
-syncarena(Arena *arena, u64int start, u32int n, int zok, int fix)
+syncarena(Arena *arena, u32int n, int zok, int fix)
 {
 	ZBlock *lump;
 	Clump cl;
@@ -53,7 +53,7 @@ syncarena(Arena *arena, u64int start, u32int n, int zok, int fix)
 			fprint(2, "%s: illegal clump magic number=%#8.8ux at clump=%d\n", arena->name, magic, clump);
 			/* err |= SyncDataErr; */
 			if(fix && writeclumpmagic(arena, aa, ClumpFreeMagic) < 0){
-				fprint(2, "can't write corrected clump free magic: %r");
+				fprint(2, "%s: can't write corrected clump free magic: %r", arena->name);
 				err |= SyncFixErr;
 			}
 			break;
@@ -136,9 +136,8 @@ syncarena(Arena *arena, u64int start, u32int n, int zok, int fix)
 	|| cclumps != arena->memstats.cclumps
 	|| uncsize != arena->memstats.uncsize){
 		err |= SyncHeader;
-		fprint(2, "arena %s: start=%lld fix=%d flush=%d %lld->%lld %ud->%ud %ud->%ud %lld->%lld\n",
+		fprint(2, "arena %s: fix=%d flush=%d %lld->%lld %ud->%ud %ud->%ud %lld->%lld\n",
 			arena->name,
-			start,
 			fix,
 			flush,
 			used, arena->memstats.used,

src/cmd/venti/srv/syncindex.c

@@ -6,7 +6,7 @@ static	int	verbose;
 void
 usage(void)
 {
-	fprint(2, "usage: syncindex [-fv] [-B blockcachesize] config\n");
+	fprint(2, "usage: syncindex [-v] [-B blockcachesize] config\n");
 	threadexitsall("usage");
 }
 
@@ -16,9 +16,7 @@ void
 threadmain(int argc, char *argv[])
 {
 	u32int bcmem, icmem;
-	int fix;
 
-	fix = 0;
 	bcmem = 0;
 	icmem = 0;
 	ARGBEGIN{
@@ -28,9 +26,6 @@ threadmain(int argc, char *argv[])
 	case 'I':
 		icmem = unittoull(EARGF(usage()));
 		break;
-	case 'f':
-		fix++;
-		break;
 	case 'v':
 		verbose++;
 		break;
@@ -39,9 +34,6 @@ threadmain(int argc, char *argv[])
 		break;
 	}ARGEND
 
-	if(!fix)
-		readonly = 1;
-
 	if(argc != 1)
 		usage();
 
@@ -63,8 +55,10 @@ threadmain(int argc, char *argv[])
 
 	if(verbose)
 		printindex(2, mainindex);
-	if(syncindex(mainindex, fix, 1, 0) < 0)
+	if(syncindex(mainindex) < 0)
 		sysfatal("failed to sync index=%s: %r\n", mainindex->name);
+	flushicache();
+	flushdcache();
 
 	threadexitsall(0);
 }

src/cmd/venti/srv/syncindex0.c

@@ -2,184 +2,92 @@
 #include "dat.h"
 #include "fns.h"
 
-enum
+static int
+syncarenaindex(Arena *arena, u64int a0)
 {
-	ClumpChunks	= 32*1024
+	int ok;
-};
+	u32int clump;
-
+	u64int a;
-static int missing, wrong;
+	ClumpInfo ci;
-
-/*
- * shell sort is plenty good enough
- * because we're going to do a bunch of disk i/o's
- */
-static void
-sortclumpinfo(ClumpInfo *ci, int *s, int n)
-{
-	int i, j, m, t;
-
-	for(m = (n + 3) / 5; m > 0; m = (m + 1) / 3){
-		for(i = n - m; i-- > 0;){
-			for(j = i + m; j < n; j += m){
-				if(memcmp(ci[s[j - m]].score, ci[s[j]].score, VtScoreSize) <= 0)
-					break;
-				t = s[j];
-				s[j] = s[j - m];
-				s[j - m] = t;
-			}
-		}
-	}
-}
-
-int
-syncarenaindex(Index *ix, Arena *arena, u32int clump, u64int a, int fix, int *pflush, int check)
-{
-	Packet *pack;
-	IEntry ie;
 	IAddr ia;
-	ClumpInfo *ci, *cis;
+	AState as;
-	u64int *addrs;
-	int i, n, ok, *s, flush;
 	
-	trace(TraceProc, "syncarenaindex enter");
+	if(arena->diskstats.clumps == arena->memstats.clumps)
+		return 0;
+	
+	memset(&as, 0, sizeof as);
+	as.arena = arena;
+	as.stats = arena->diskstats;
 
-	flush = 0;
-	cis = MKN(ClumpInfo, ClumpChunks);
-	addrs = MKN(u64int, ClumpChunks);
-	s = MKN(int, ClumpChunks);
 	ok = 0;
-	for(; clump < arena->memstats.clumps; clump += n){
+	a = a0 + arena->diskstats.used;
-		n = ClumpChunks;
+	for(clump=arena->diskstats.clumps; clump < arena->memstats.clumps; clump++){
-		if(n > arena->memstats.clumps - clump)
+		if(readclumpinfo(arena, clump, &ci) < 0){
-			n = arena->memstats.clumps - clump;
+			fprint(2, "%s: clump %d: cannot read clumpinfo\n",
-		n = readclumpinfos(arena, clump, cis, n);
+				arena->name, clump);
-		if(n <= 0){
-			fprint(2, "arena directory read failed\n");
 			ok = -1;
 			break;
 		}
 
-		for(i = 0; i < n; i++){
+		ia.type = ci.type;
-			addrs[i] = a;
+		ia.size = ci.uncsize;
-			a += cis[i].size + ClumpSize;
+		ia.addr = a;
-			s[i] = i;
+		ia.blocks = (ClumpSize + ci.size + (1 << ABlockLog) - 1) >> ABlockLog;
-		}
+		a += ClumpSize + ci.size;
 
-		sortclumpinfo(cis, s, n);
+		as.stats.used += ClumpSize + ci.size;
-
+		as.stats.uncsize += ia.size;
-		for(i = 0; i < n; i++){
+		as.stats.clumps++;
-			ci = &cis[s[i]];
+		if(ci.uncsize > ci.size)
-			ia.type = ci->type;
+			as.stats.cclumps++;
-			ia.size = ci->uncsize;
+		as.aa = a;
-			ia.addr = addrs[s[i]];
+		insertscore(ci.score, &ia, IEDirty, &as);
-			ia.blocks = (ci->size + ClumpSize + (1 << ABlockLog) - 1) >> ABlockLog;
-
-			if(!check)
-				goto Add;
-			if(loadientry(ix, ci->score, ci->type, &ie) < 0){
-				trace(TraceProc, "syncarenaindex missing block %V.%d", ci->score, ci->type);
-				missing++;
-			if(0)	fprint(2, "missing block type=%d score=%V\n", ci->type, ci->score);
-			}else if(iaddrcmp(&ia, &ie.ia) != 0){
-				trace(TraceProc, "syncarenaindex mismatched entry");
-				fprint(2, "\nmismatched index entry and clump at %d\n", clump + i);
-				fprint(2, "\tclump: type=%d size=%d blocks=%d addr=%lld\n", ia.type, ia.size, ia.blocks, ia.addr);
-				fprint(2, "\tindex: type=%d size=%d block=%d addr=%lld\n", ie.ia.type, ie.ia.size, ie.ia.blocks, ie.ia.addr);
-				pack = readlump(ie.score, ie.ia.type, ie.ia.size, nil);
-				packetfree(pack);
-				if(pack != nil){
-					fprint(2, "duplicated lump\n");
-					continue;
-				}
-				wrong++;
-			}else
-				continue;
-		Add:
-			if(!fix){
-				ok = -1;
-				continue;
-			}
-			flush = 1;
-			trace(TraceProc, "syncarenaindex insert %V", ci->score);
-			insertscore(ci->score, &ia, IEDirty);
-		}
-
-		if(0 && clump / 1000 != (clump + n) / 1000)
-			fprint(2, ".");
-	}
-	free(cis);
-	free(addrs);
-	free(s);
-	if(flush){
-		flushdcache();
-		*pflush = 1;
 	}
+	flushdcache();
 	return ok;
 }
 
 int
-syncindex(Index *ix, int fix, int mustflush, int check)
+syncindex(Index *ix)
 {
 	Arena *arena;
-	AState as;
+	int i, e, e1, ok;
-	u64int a;
-	int i, e, e1, ok, ok1, flush;
 
 	ok = 0;
-	flush = 0;
 	for(i = 0; i < ix->narenas; i++){
 		trace(TraceProc, "syncindex start %d", i);
 		arena = ix->arenas[i];
-		/*
+		e = syncarena(arena, TWID32, 1, 1);
-		 * Syncarena will scan through the arena looking for blocks
-		 * that have been forgotten.  It will update arena->memstats.used,
-		 * so save the currenct copy as the place to start the 
-		 * syncarenaindex scan.
-		 */
-		a = arena->memstats.used;
-		e = syncarena(arena, ix->amap[i].start, TWID32, fix, fix);
 		e1 = e;
-		if(fix)
+		e1 &= ~(SyncHeader|SyncCIZero|SyncCIErr);
-			e1 &= ~(SyncHeader|SyncCIZero|SyncCIErr);
+		if(e & SyncHeader)
-		if(e1 == SyncHeader)
 			fprint(2, "arena %s: header is out-of-date\n", arena->name);
-		if(e1)
+		if(e1){
+			fprint(2, "arena %s: %x\n", arena->name, e1);
 			ok = -1;
-		else{
+			continue;
-			/*
-			 * use diskstats not memstats here, because diskstats
-			 * is what has been indexed; memstats is what has 
-			 * made it to disk (confusing names).
-			 */
-			ok1 = syncarenaindex(ix, arena,
-					arena->diskstats.clumps,
-					ix->amap[i].start + arena->diskstats.used,
-					fix, &flush, check);
-			if(ok1 < 0)
-				fprint(2, "syncarenaindex: %r\n");
-			if(fix && ok1==0 && (e & SyncHeader) && wbarena(arena) < 0)
-				fprint(2, "arena=%s header write failed: %r\n", arena->name);
-			ok |= ok1;
-
-			as.arena = arena;
-			as.aa = ix->amap[i].start + arena->memstats.used;
-			as.stats = arena->memstats;
-			setdcachestate(&as);
 		}
-	}
-	if(missing || wrong)
-		fprint(2, "syncindex: %d missing entries, %d wrong entries (flush=%d)\n", missing, wrong, flush);
-	if(fix && wbindex(ix) < 0){
-		fprint(2, "can't write back index header for %s: %r\n", ix->name);
-		return -1;
-	}
-	if(fix && flush){
 		flushdcache();
-		if(mustflush){
+		
-			flushicache();
+		if(arena->memstats.clumps == arena->diskstats.clumps)
-			flushdcache();
+			continue;
-		}else
+		
-			kickicache();
+		fprint(2, "%T %s: indexing %d clumps...\n",
+			arena->name,
+			arena->memstats.clumps - arena->diskstats.clumps);
+
+		if(syncarenaindex(arena, ix->amap[i].start) < 0){
+			fprint(2, "arena %s: syncarenaindex: %r\n", arena->name);
+			ok = -1;
+			continue;
+		}
+		if(wbarena(arena) < 0){
+			fprint(2, "arena %s: wbarena: %r\n", arena->name);
+			ok = -1;
+			continue;
+		}
+		flushdcache();
+		delaykickicache();
 	}
 	return ok;
 }

src/cmd/venti/srv/venti.c

@@ -161,7 +161,7 @@ threadmain(int argc, char *argv[])
 		startbloomproc(mainindex->bloom);
 
 	fprint(2, "sync...");
-	if(!readonly && syncindex(mainindex, 1, 0, 0) < 0)
+	if(!readonly && syncindex(mainindex) < 0)
 		sysfatal("can't sync server: %r");
 
 	if(!readonly && queuewrites){

src/cmd/venti/srv/wrarena.c

@@ -133,7 +133,6 @@ threadmain(int argc, char *argv[])
 	Arena *arena;
 	u64int offset, aoffset;
 	Part *part;
-	Dir *d;
 	uchar buf[8192];
 	ArenaHead head;
 	ZClump zerocl;
@@ -178,9 +177,6 @@ threadmain(int argc, char *argv[])
 
 	statsinit();
 
-	if((d = dirstat(file)) == nil)
-		sysfatal("can't stat file %s: %r", file);
-
 	part = initpart(file, OREAD);
 	if(part == nil)
 		sysfatal("can't open file %s: %r", file);
@@ -190,9 +186,9 @@ threadmain(int argc, char *argv[])
 	if(unpackarenahead(&head, buf) < 0)
 		sysfatal("corrupted arena header: %r");
 
-	if(aoffset+head.size > d->length)
+	if(aoffset+head.size > part->size)
 		sysfatal("arena is truncated: want %llud bytes have %llud\n",
-			head.size, d->length);
+			head.size, part->size);
 
 	partblocksize(part, head.blocksize);
 	initdcache(8 * MaxDiskBlock);