static void
icacheinsert(u8int score[VtScoreSize], IAddr *ia, int state)
{
IEntry *ie;
if((ie = poplast(&icache.free)) == nil && (ie = evictlru()) == nil){
addstat(StatIcacheStall, 1);
while((ie = poplast(&icache.free)) == nil && (ie = evictlru()) == nil){
// Could safely return here if state == IEClean.
// But if state == IEDirty, have to wait to make
// sure we don't lose an index write.
// Let's wait all the time.
flushdcache();
kickicache();
rsleep(&icache.full);
}
addstat(StatIcacheStall, -1);
}
memmove(ie->score, score, VtScoreSize);
ie->state = state;
ie->ia = *ia;
if(state == IEClean){
addstat(StatIcachePrefetch, 1);
pushfirst(&icache.clean, ie);
}else{
addstat(StatIcacheWrite, 1);
assert(state == IEDirty);
icache.ndirty++;
setstat(StatIcacheDirty, icache.ndirty);
delaykickicache();
pushfirst(&icache.dirty, ie);
}
ihashinsert(icache.hash, ie);
}
int
syncindex(Index *ix)
{
Arena *arena;
int i, e, e1, ok;
ok = 0;
for(i = 0; i < ix->narenas; i++){
trace(TraceProc, "syncindex start %d", i);
arena = ix->arenas[i];
e = syncarena(arena, TWID32, 1, 1);
e1 = e;
e1 &= ~(SyncHeader|SyncCIZero|SyncCIErr);
if(e & SyncHeader)
fprint(2, "arena %s: header is out-of-date\n", arena->name);
if(e1){
fprint(2, "arena %s: %x\n", arena->name, e1);
ok = -1;
continue;
}
flushdcache();
if(arena->memstats.clumps == arena->diskstats.clumps)
continue;
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;
}
static int
hdcacheflush(HConnect *c)
{
Hio *hout;
int r;
r = hsettext(c);
if(r < 0)
return r;
hout = &c->hout;
flushdcache();
hprint(hout, "flushed dcache\n");
hflush(hout);
return 0;
}
static int
syncarenaindex(Arena *arena, uint64_t a0)
{
int ok;
uint32_t clump;
uint64_t a;
ClumpInfo ci;
IAddr ia;
AState as;
if(arena->diskstats.clumps == arena->memstats.clumps)
return 0;
memset(&as, 0, sizeof as);
as.arena = arena;
as.stats = arena->diskstats;
ok = 0;
a = a0 + arena->diskstats.used;
for(clump=arena->diskstats.clumps; clump < arena->memstats.clumps; clump++){
if(readclumpinfo(arena, clump, &ci) < 0){
fprint(2, "%s: clump %d: cannot read clumpinfo\n",
arena->name, clump);
ok = -1;
break;
}
ia.type = ci.type;
ia.size = ci.uncsize;
ia.addr = a;
ia.blocks = (ClumpSize + ci.size + (1 << ABlockLog) - 1) >> ABlockLog;
a += ClumpSize + ci.size;
as.stats.used += ClumpSize + ci.size;
as.stats.uncsize += ia.size;
as.stats.clumps++;
if(ci.uncsize > ci.size)
as.stats.cclumps++;
as.aa = a;
insertscore(ci.score, &ia, IEDirty, &as);
}
flushdcache();
return ok;
}
/*
* Return a singly-linked list of dirty index entries.
* with 32-bit hash numbers between lo and hi
* and address < limit.
*/
IEntry*
icachedirty(u32int lo, u32int hi, u64int limit)
{
u32int h;
IEntry *ie, *dirty;
dirty = nil;
trace(TraceProc, "icachedirty enter");
qlock(&icache.lock);
for(ie = icache.dirty.next; ie != &icache.dirty; ie=ie->next){
if(ie->state == IEDirty && ie->ia.addr <= limit){
h = hashbits(ie->score, 32);
if(lo <= h && h <= hi){
ie->nextdirty = dirty;
dirty = ie;
}
}
}
qunlock(&icache.lock);
trace(TraceProc, "icachedirty exit");
if(dirty == nil)
flushdcache();
return dirty;
}
void
threadmain(int argc, char *argv[])
{
int vers;
ArenaPart *ap;
Part *part;
Arena *arena;
uint64_t addr, limit, asize, apsize;
char *file, *name, aname[ANameSize];
int i, n, blocksize, tabsize, zero;
ventifmtinstall();
statsinit();
blocksize = 8 * 1024;
asize = 512 * 1024 *1024;
tabsize = 512 * 1024; /* BUG: should be determine from number of arenas */
zero = -1;
vers = ArenaVersion5;
ARGBEGIN{
case 'D':
settrace(EARGF(usage()));
break;
case 'a':
asize = unittoull(EARGF(usage()));
if(asize == TWID64)
usage();
break;
case 'b':
blocksize = unittoull(EARGF(usage()));
if(blocksize == ~0)
usage();
if(blocksize > MaxDiskBlock){
fprint(2, "block size too large, max %d\n", MaxDiskBlock);
threadexitsall("usage");
}
break;
case '4':
vers = ArenaVersion4;
break;
case 'Z':
zero = 0;
break;
default:
usage();
break;
}ARGEND
if(zero == -1){
if(vers == ArenaVersion4)
zero = 1;
else
zero = 0;
}
if(argc != 2)
usage();
name = argv[0];
file = argv[1];
if(nameok(name) < 0)
sysfatal("illegal name template %s", name);
part = initpart(file, ORDWR|ODIRECT);
if(part == nil)
sysfatal("can't open partition %s: %r", file);
if(zero)
zeropart(part, blocksize);
maxblocksize = blocksize;
initdcache(20*blocksize);
ap = newarenapart(part, blocksize, tabsize);
if(ap == nil)
sysfatal("can't initialize arena: %r");
apsize = ap->size - ap->arenabase;
n = apsize / asize;
if(apsize - (n * asize) >= MinArenaSize)
n++;
fprint(2, "fmtarenas %s: %,d arenas, %,lld bytes storage, %,d bytes for index map\n",
file, n, apsize, ap->tabsize);
ap->narenas = n;
ap->map = MKNZ(AMap, n);
ap->arenas = MKNZ(Arena*, n);
addr = ap->arenabase;
for(i = 0; i < n; i++){
limit = addr + asize;
if(limit >= ap->size || ap->size - limit < MinArenaSize){
limit = ap->size;
if(limit - addr < MinArenaSize)
sysfatal("bad arena set math: runt arena at %lld,%lld %lld", addr, limit, ap->size);
}
snprint(aname, ANameSize, "%s%d", name, i);
if(0) fprint(2, "adding arena %s at [%lld,%lld)\n", aname, addr, limit);
arena = newarena(part, vers, aname, addr, limit - addr, blocksize);
if(!arena)
fprint(2, "can't make new arena %s: %r", aname);
freearena(arena);
ap->map[i].start = addr;
ap->map[i].stop = limit;
namecp(ap->map[i].name, aname);
addr = limit;
}
if(wbarenapart(ap) < 0)
fprint(2, "can't write back arena partition header for %s: %r\n", file);
flushdcache();
threadexitsall(0);
}
int
writeqlump(Lump *u, Packet *p, int creator, uint ms)
{
ZBlock *flat;
Packet *old;
IAddr ia;
int ok;
if(lookupscore(u->score, u->type, &ia) == 0){
if(verifywrites == 0){
/* assume the data is here! */
packetfree(p);
ms = msec() - ms;
addstat2(StatRpcWriteOld, 1, StatRpcWriteOldTime, ms);
return 0;
}
/*
* if the read fails,
* assume it was corrupted data and store the block again
*/
old = readilump(u, &ia, u->score);
if(old != nil){
ok = 0;
if(packetcmp(p, old) != 0){
uchar nscore[VtScoreSize];
packetsha1(old, nscore);
if(scorecmp(u->score, nscore) != 0)
seterr(EStrange, "readilump returned bad data %V not %V", nscore, u->score);
else
seterr(EStrange, "score collision %V", u->score);
ok = -1;
}
packetfree(p);
packetfree(old);
ms = msec() - ms;
addstat2(StatRpcWriteOld, 1, StatRpcWriteOldTime, ms);
return ok;
}
logerr(EAdmin, "writelump: read %V failed, rewriting: %r\n", u->score);
}
flat = packet2zblock(p, packetsize(p));
ok = storeclump(mainindex, flat, u->score, u->type, creator, &ia);
freezblock(flat);
if(ok == 0)
insertlump(u, p);
else
packetfree(p);
if(syncwrites){
flushdcache();
flushicache();
flushdcache();
}
ms = msec() - ms;
addstat2(StatRpcWriteNew, 1, StatRpcWriteNewTime, ms);
return ok;
}
void
sumarena(Arena *arena)
{
ZBlock *b;
DigestState s;
uint64_t a, e;
uint32_t bs;
int t;
uint8_t score[VtScoreSize];
bs = MaxIoSize;
if(bs < arena->blocksize)
bs = arena->blocksize;
/*
* read & sum all blocks except the last one
*/
flushdcache();
memset(&s, 0, sizeof s);
b = alloczblock(bs, 0, arena->part->blocksize);
e = arena->base + arena->size;
for(a = arena->base - arena->blocksize; a + arena->blocksize <= e; a += bs){
disksched();
while((t=arenasumsleeptime) == SleepForever){
sleep(1000);
disksched();
}
sleep(t);
if(a + bs > e)
bs = arena->blocksize;
if(readpart(arena->part, a, b->data, bs) < 0)
goto ReadErr;
addstat(StatSumRead, 1);
addstat(StatSumReadBytes, bs);
sha1(b->data, bs, nil, &s);
}
/*
* the last one is special, since it may already have the checksum included
*/
bs = arena->blocksize;
if(readpart(arena->part, e, b->data, bs) < 0){
ReadErr:
logerr(EOk, "sumarena can't sum %s, read at %lld failed: %r", arena->name, a);
freezblock(b);
return;
}
addstat(StatSumRead, 1);
addstat(StatSumReadBytes, bs);
sha1(b->data, bs-VtScoreSize, nil, &s);
sha1(zeroscore, VtScoreSize, nil, &s);
sha1(nil, 0, score, &s);
/*
* check for no checksum or the same
*/
if(scorecmp(score, &b->data[bs - VtScoreSize]) != 0
&& scorecmp(zeroscore, &b->data[bs - VtScoreSize]) != 0)
logerr(EOk, "overwriting mismatched checksums for arena=%s, found=%V calculated=%V",
arena->name, &b->data[bs - VtScoreSize], score);
freezblock(b);
qlock(&arena->lock);
scorecp(arena->score, score);
wbarena(arena);
qunlock(&arena->lock);
}
static void
ventiserver(void *v)
{
Packet *p;
VtReq *r;
char err[ERRMAX];
uint ms;
int cached, ok;
USED(v);
threadsetname("ventiserver");
trace(TraceWork, "start");
while((r = vtgetreq(ventisrv)) != nil){
trace(TraceWork, "finish");
trace(TraceWork, "start request %F", &r->tx);
trace(TraceRpc, "<- %F", &r->tx);
r->rx.msgtype = r->tx.msgtype+1;
addstat(StatRpcTotal, 1);
if(0) print("req (arenas[0]=%p sects[0]=%p) %F\n",
mainindex->arenas[0], mainindex->sects[0], &r->tx);
switch(r->tx.msgtype){
default:
vtrerror(r, "unknown request");
break;
case VtTread:
ms = msec();
r->rx.data = readlump(r->tx.score, r->tx.blocktype, r->tx.count, &cached);
ms = msec() - ms;
addstat2(StatRpcRead, 1, StatRpcReadTime, ms);
if(r->rx.data == nil){
addstat(StatRpcReadFail, 1);
rerrstr(err, sizeof err);
vtrerror(r, err);
}else{
addstat(StatRpcReadBytes, packetsize(r->rx.data));
addstat(StatRpcReadOk, 1);
if(cached)
addstat2(StatRpcReadCached, 1, StatRpcReadCachedTime, ms);
else
addstat2(StatRpcReadUncached, 1, StatRpcReadUncachedTime, ms);
}
break;
case VtTwrite:
if(readonly){
vtrerror(r, "read only");
break;
}
p = r->tx.data;
r->tx.data = nil;
addstat(StatRpcWriteBytes, packetsize(p));
ms = msec();
ok = writelump(p, r->rx.score, r->tx.blocktype, 0, ms);
ms = msec() - ms;
addstat2(StatRpcWrite, 1, StatRpcWriteTime, ms);
if(ok < 0){
addstat(StatRpcWriteFail, 1);
rerrstr(err, sizeof err);
vtrerror(r, err);
}
break;
case VtTsync:
flushqueue();
flushdcache();
break;
}
trace(TraceRpc, "-> %F", &r->rx);
vtrespond(r);
trace(TraceWork, "start");
}
flushdcache();
flushicache();
threadexitsall(0);
}