static void
putcib(Arena *arena, CIBlock *cib)
{
USED(arena);
putdblock(cib->data);
cib->data = nil;
}
/*
* write some data to the clump section at a given offset
* used to fix up corrupted arenas.
*/
uint32_t
writearena(Arena *arena, uint64_t aa, uint8_t *clbuf, uint32_t n)
{
DBlock *b;
uint64_t a;
uint32_t blocksize, off, m;
int32_t nn;
int ok;
if(n == 0)
return -1;
qlock(&arena->lock);
a = arena->size - arenadirsize(arena, arena->memstats.clumps);
if(aa >= a || aa + n > a){
qunlock(&arena->lock);
seterr(EOk, "writing beyond arena clump storage");
return -1;
}
blocksize = arena->blocksize;
a = arena->base + aa;
off = a & (blocksize - 1);
a -= off;
nn = 0;
for(;;){
b = getdblock(arena->part, a, off != 0 || off + n < blocksize ? ORDWR : OWRITE);
if(b == nil){
qunlock(&arena->lock);
return -1;
}
dirtydblock(b, DirtyArena);
m = blocksize - off;
if(m > n - nn)
m = n - nn;
memmove(&b->data[off], &clbuf[nn], m);
ok = 0;
putdblock(b);
if(ok < 0){
qunlock(&arena->lock);
return -1;
}
nn += m;
if(nn == n)
break;
off = 0;
a += blocksize;
}
qunlock(&arena->lock);
return n;
}
/*
* write the arena trailer block to the partition
*/
int
wbarena(Arena *arena)
{
DBlock *b;
int bad;
if((b = getdblock(arena->part, arena->base + arena->size, OWRITE)) == nil){
logerr(EAdmin, "can't write arena trailer: %r");
return -1;
}
dirtydblock(b, DirtyArenaTrailer);
bad = okarena(arena)<0 || packarena(arena, b->data)<0;
scorecp(b->data + arena->blocksize - VtScoreSize, arena->score);
putdblock(b);
if(bad)
return -1;
return 0;
}
/*
* read a clump of data
* n is a hint of the size of the data, not including the header
* make sure it won't run off the end, then return the number of bytes actually read
*/
uint32_t
readarena(Arena *arena, uint64_t aa, uint8_t *buf, int32_t n)
{
DBlock *b;
uint64_t a;
uint32_t blocksize, off, m;
int32_t nn;
if(n == 0)
return -1;
qlock(&arena->lock);
a = arena->size - arenadirsize(arena, arena->memstats.clumps);
qunlock(&arena->lock);
if(aa >= a){
seterr(EOk, "reading beyond arena clump storage: clumps=%d aa=%lld a=%lld -1 clumps=%lld\n", arena->memstats.clumps, aa, a, arena->size - arenadirsize(arena, arena->memstats.clumps - 1));
return -1;
}
if(aa + n > a)
n = a - aa;
blocksize = arena->blocksize;
a = arena->base + aa;
off = a & (blocksize - 1);
a -= off;
nn = 0;
for(;;){
b = getdblock(arena->part, a, OREAD);
if(b == nil)
return -1;
m = blocksize - off;
if(m > n - nn)
m = n - nn;
memmove(&buf[nn], &b->data[off], m);
putdblock(b);
nn += m;
if(nn == n)
break;
off = 0;
a += blocksize;
}
return n;
}
/*
* allocate space for the clump and write it,
* updating the arena directory
ZZZ question: should this distinguish between an arena
filling up and real errors writing the clump?
*/
uint64_t
writeaclump(Arena *arena, Clump *c, uint8_t *clbuf)
{
DBlock *b;
uint64_t a, aa;
uint32_t clump, n, nn, m, off, blocksize;
int ok;
n = c->info.size + ClumpSize + U32Size;
qlock(&arena->lock);
aa = arena->memstats.used;
if(arena->memstats.sealed
|| aa + n + U32Size + arenadirsize(arena, arena->memstats.clumps + 1) > arena->size){
if(!arena->memstats.sealed){
logerr(EOk, "seal memstats %s", arena->name);
arena->memstats.sealed = 1;
wbarena(arena);
}
qunlock(&arena->lock);
return TWID64;
}
if(packclump(c, &clbuf[0], arena->clumpmagic) < 0){
qunlock(&arena->lock);
return TWID64;
}
/*
* write the data out one block at a time
*/
blocksize = arena->blocksize;
a = arena->base + aa;
off = a & (blocksize - 1);
a -= off;
nn = 0;
for(;;){
b = getdblock(arena->part, a, off != 0 ? ORDWR : OWRITE);
if(b == nil){
qunlock(&arena->lock);
return TWID64;
}
dirtydblock(b, DirtyArena);
m = blocksize - off;
if(m > n - nn)
m = n - nn;
memmove(&b->data[off], &clbuf[nn], m);
ok = 0;
putdblock(b);
if(ok < 0){
qunlock(&arena->lock);
return TWID64;
}
nn += m;
if(nn == n)
break;
off = 0;
a += blocksize;
}
arena->memstats.used += c->info.size + ClumpSize;
arena->memstats.uncsize += c->info.uncsize;
if(c->info.size < c->info.uncsize)
arena->memstats.cclumps++;
clump = arena->memstats.clumps;
if(clump % ArenaCIGSize == 0){
if(arena->cig == nil){
loadcig(arena);
if(arena->cig == nil)
goto NoCIG;
}
/* add aa as start of next cig */
if(clump/ArenaCIGSize != arena->ncig){
fprint(2, "bad arena cig computation %s: writing clump %d but %d cigs\n",
arena->name, clump, arena->ncig);
arena->ncig = -1;
vtfree(arena->cig);
arena->cig = nil;
goto NoCIG;
}
arena->cig = vtrealloc(arena->cig, (arena->ncig+1)*sizeof arena->cig[0]);
arena->cig[arena->ncig++].offset = aa;
}
NoCIG:
arena->memstats.clumps++;
if(arena->memstats.clumps == 0)
sysfatal("clumps wrapped");
arena->wtime = now();
if(arena->ctime == 0)
arena->ctime = arena->wtime;
writeclumpinfo(arena, clump, &c->info);
wbarena(arena);
qunlock(&arena->lock);
return aa;
}
static int
icachewritesect(Index *ix, ISect *is, u8int *buf)
{
int err, i, werr, h, bsize, t;
u32int lo, hi;
u64int addr, naddr;
uint nbuf, off;
DBlock *b;
IBucket ib;
IEntry *ie, *iedirty, **l, *chunk;
lo = is->start * ix->div;
if(TWID32/ix->div < is->stop)
hi = TWID32;
else
hi = is->stop * ix->div - 1;
trace(TraceProc, "icachewritesect enter %ud %ud %llud",
lo, hi, iwrite.as.aa);
iedirty = icachedirty(lo, hi, iwrite.as.aa);
iedirty = iesort(iedirty);
bsize = 1 << is->blocklog;
err = 0;
while(iedirty){
disksched();
while((t = icachesleeptime) == SleepForever){
sleep(1000);
disksched();
}
if(t < minicachesleeptime)
t = minicachesleeptime;
if(t > 0)
sleep(t);
trace(TraceProc, "icachewritesect nextchunk");
chunk = nextchunk(ix, is, &iedirty, &addr, &nbuf);
trace(TraceProc, "icachewritesect readpart 0x%llux+0x%ux",
addr, nbuf);
if(readpart(is->part, addr, buf, nbuf) < 0){
fprint(2, "%s: part %s addr 0x%llux: icachewritesect "
"readpart: %r\n", argv0, is->part->name, addr);
err = -1;
continue;
}
trace(TraceProc, "icachewritesect updatebuf");
addstat(StatIsectReadBytes, nbuf);
addstat(StatIsectRead, 1);
for(l=&chunk; (ie=*l)!=nil; l=&ie->nextdirty){
again:
naddr = ie2diskaddr(ix, is, ie);
off = naddr - addr;
if(off+bsize > nbuf){
fprint(2, "%s: whoops! addr=0x%llux nbuf=%ud "
"addr+nbuf=0x%llux naddr=0x%llux\n",
argv0, addr, nbuf, addr+nbuf, naddr);
assert(off+bsize <= nbuf);
}
unpackibucket(&ib, buf+off, is->bucketmagic);
if(okibucket(&ib, is) < 0){
fprint(2, "%s: bad bucket XXX\n", argv0);
goto skipit;
}
trace(TraceProc, "icachewritesect add %V at 0x%llux",
ie->score, naddr);
h = bucklook(ie->score, ie->ia.type, ib.data, ib.n);
if(h & 1){
h ^= 1;
packientry(ie, &ib.data[h]);
}else if(ib.n < is->buckmax){
memmove(&ib.data[h + IEntrySize], &ib.data[h],
ib.n*IEntrySize - h);
ib.n++;
packientry(ie, &ib.data[h]);
}else{
fprint(2, "%s: bucket overflow XXX\n", argv0);
skipit:
err = -1;
*l = ie->nextdirty;
ie = *l;
if(ie)
goto again;
else
break;
}
packibucket(&ib, buf+off, is->bucketmagic);
}
diskaccess(1);
trace(TraceProc, "icachewritesect writepart", addr, nbuf);
werr = 0;
if(writepart(is->part, addr, buf, nbuf) < 0 || flushpart(is->part) < 0)
werr = -1;
for(i=0; i<nbuf; i+=bsize){
if((b = _getdblock(is->part, addr+i, ORDWR, 0)) != nil){
memmove(b->data, buf+i, bsize);
putdblock(b);
}
}
if(werr < 0){
fprint(2, "%s: part %s addr 0x%llux: icachewritesect "
"writepart: %r\n", argv0, is->part->name, addr);
err = -1;
continue;
}
addstat(StatIsectWriteBytes, nbuf);
addstat(StatIsectWrite, 1);
icacheclean(chunk);
}
trace(TraceProc, "icachewritesect done");
return err;
}