void
archFree(Arch *a)
{
/* kill slave */
vtLock(a->lk);
a->die = vtRendezAlloc(a->lk);
vtWakeup(a->starve);
while(a->ref > 1)
vtSleep(a->die);
vtUnlock(a->lk);
vtRendezFree(a->starve);
vtRendezFree(a->die);
vtLockFree(a->lk);
vtMemFree(a);
}
void
superWrite(Block* b, Super* super, int forceWrite)
{
superPack(super, b->data);
blockDirty(b);
if(forceWrite){
while(!blockWrite(b, Waitlock)){
/* this should no longer happen */
fprint(2, "%s: could not write super block; "
"waiting 10 seconds\n", argv0);
sleep(10*1000);
}
while(b->iostate != BioClean && b->iostate != BioDirty){
assert(b->iostate == BioWriting);
vtSleep(b->ioready);
}
/*
* it's okay that b might still be dirty.
* that means it got written out but with an old root pointer,
* but the other fields went out, and those are the ones
* we really care about. (specifically, epochHigh; see fsSnapshot).
*/
}
}
static void
archThread(void *v)
{
Arch *a = v;
Block *b;
Param p;
Super super;
int ret;
u32int addr;
uchar rbuf[VtRootSize];
VtRoot root;
vtThreadSetName("arch");
for(;;){
/* look for work */
vtLock(a->fs->elk);
b = superGet(a->c, &super);
if(b == nil){
vtUnlock(a->fs->elk);
fprint(2, "archThread: superGet: %R\n");
sleep(60*1000);
continue;
}
addr = super.next;
if(addr != NilBlock && super.current == NilBlock){
super.current = addr;
super.next = NilBlock;
superPack(&super, b->data);
blockDirty(b);
}else
addr = super.current;
blockPut(b);
vtUnlock(a->fs->elk);
if(addr == NilBlock){
/* wait for work */
vtLock(a->lk);
vtSleep(a->starve);
if(a->die != nil)
goto Done;
vtUnlock(a->lk);
continue;
}
sleep(10*1000); /* window of opportunity to provoke races */
/* do work */
memset(&p, 0, sizeof p);
p.blockSize = a->blockSize;
p.dsize = 3*VtEntrySize; /* root has three Entries */
p.c = a->c;
p.a = a;
ret = archWalk(&p, addr, BtDir, RootTag);
switch(ret){
default:
abort();
case ArchFailure:
fprint(2, "archiveBlock %#ux: %R\n", addr);
sleep(60*1000);
continue;
case ArchSuccess:
case ArchFaked:
break;
}
if(0) fprint(2, "archiveSnapshot 0x%#ux: maxdepth %ud nfixed %ud"
" send %ud nfailsend %ud nvisit %ud"
" nreclaim %ud nfake %ud nreal %ud\n",
addr, p.maxdepth, p.nfixed,
p.nsend, p.nfailsend, p.nvisit,
p.nreclaim, p.nfake, p.nreal);
if(0) fprint(2, "archiveBlock %V (%ud)\n", p.score, p.blockSize);
/* tie up vac root */
memset(&root, 0, sizeof root);
root.version = VtRootVersion;
strecpy(root.type, root.type+sizeof root.type, "vac");
strecpy(root.name, root.name+sizeof root.name, "fossil");
memmove(root.score, p.score, VtScoreSize);
memmove(root.prev, super.last, VtScoreSize);
root.blockSize = a->blockSize;
vtRootPack(&root, rbuf);
if(!vtWrite(a->z, p.score, VtRootType, rbuf, VtRootSize)
|| !vtSha1Check(p.score, rbuf, VtRootSize)){
fprint(2, "vtWriteBlock %#ux: %R\n", addr);
sleep(60*1000);
continue;
}
/* record success */
vtLock(a->fs->elk);
b = superGet(a->c, &super);
if(b == nil){
vtUnlock(a->fs->elk);
fprint(2, "archThread: superGet: %R\n");
sleep(60*1000);
continue;
}
super.current = NilBlock;
memmove(super.last, p.score, VtScoreSize);
superPack(&super, b->data);
blockDirty(b);
blockPut(b);
vtUnlock(a->fs->elk);
consPrint("archive vac:%V\n", p.score);
}
Done:
a->ref--;
vtWakeup(a->die);
vtUnlock(a->lk);
}
static int
archWalk(Param *p, u32int addr, uchar type, u32int tag)
{
int ret, i, x, psize, dsize;
uchar *data, score[VtScoreSize];
Block *b;
Label l;
Entry *e;
WalkPtr w;
p->nvisit++;
b = cacheLocalData(p->c, addr, type, tag, OReadWrite,0);
if(b == nil){
fprint(2, "archive(%ud, %#ux): cannot find block: %R\n", p->snapEpoch, addr);
if(strcmp(vtGetError(), ELabelMismatch) == 0){
/* might as well plod on so we write _something_ to Venti */
memmove(p->score, vtZeroScore, VtScoreSize);
return ArchFaked;
}
return ArchFailure;
}
if(DEBUG) fprint(2, "%*sarchive(%ud, %#ux): block label %L\n",
p->depth*2, "", p->snapEpoch, b->addr, &b->l);
p->depth++;
if(p->depth > p->maxdepth)
p->maxdepth = p->depth;
data = b->data;
if((b->l.state&BsVenti) == 0){
initWalk(&w, b, b->l.type==BtDir ? p->dsize : p->psize);
for(i=0; nextWalk(&w, score, &type, &tag, &e); i++){
if(e){
if(!(e->flags&VtEntryActive))
continue;
if((e->snap && !e->archive)
|| (e->flags&VtEntryNoArchive)){
if(0) fprint(2, "snap; faking %#ux\n", b->addr);
if(data == b->data){
data = copyBlock(b, p->blockSize);
if(data == nil){
ret = ArchFailure;
goto Out;
}
w.data = data;
}
memmove(e->score, vtZeroScore, VtScoreSize);
e->depth = 0;
e->size = 0;
e->tag = 0;
e->flags &= ~VtEntryLocal;
entryPack(e, data, w.n-1);
continue;
}
}
addr = globalToLocal(score);
if(addr == NilBlock)
continue;
dsize = p->dsize;
psize = p->psize;
if(e){
p->dsize= e->dsize;
p->psize = e->psize;
}
vtUnlock(b->lk);
x = archWalk(p, addr, type, tag);
vtLock(b->lk);
if(e){
p->dsize = dsize;
p->psize = psize;
}
while(b->iostate != BioClean && b->iostate != BioDirty)
vtSleep(b->ioready);
switch(x){
case ArchFailure:
fprint(2, "archWalk %#ux failed; ptr is in %#ux offset %d\n",
addr, b->addr, i);
ret = ArchFailure;
goto Out;
case ArchFaked:
/*
* When we're writing the entry for an archive directory
* (like /archive/2003/1215) then even if we've faked
* any data, record the score unconditionally.
* This way, we will always record the Venti score here.
* Otherwise, temporary data or corrupted file system
* would cause us to keep holding onto the on-disk
* copy of the archive.
*/
if(e==nil || !e->archive)
if(data == b->data){
if(0) fprint(2, "faked %#ux, faking %#ux (%V)\n", addr, b->addr, p->score);
data = copyBlock(b, p->blockSize);
if(data == nil){
ret = ArchFailure;
goto Out;
}
w.data = data;
}
/* fall through */
if(0) fprint(2, "falling\n");
case ArchSuccess:
if(e){
memmove(e->score, p->score, VtScoreSize);
e->flags &= ~VtEntryLocal;
entryPack(e, data, w.n-1);
}else
memmove(data+(w.n-1)*VtScoreSize, p->score, VtScoreSize);
if(data == b->data){
blockDirty(b);
/*
* If b is in the active tree, then we need to note that we've
* just removed addr from the active tree (replacing it with the
* copy we just stored to Venti). If addr is in other snapshots,
* this will close addr but not free it, since it has a non-empty
* epoch range.
*
* If b is in the active tree but has been copied (this can happen
* if we get killed at just the right moment), then we will
* mistakenly leak its kids.
*
* The children of an archive directory (e.g., /archive/2004/0604)
* are not treated as in the active tree.
*/
if((b->l.state&BsCopied)==0 && (e==nil || e->snap==0))
blockRemoveLink(b, addr, p->l.type, p->l.tag, 0);
}
break;
}
}
if(!ventiSend(p->a, b, data)){
p->nfailsend++;
ret = ArchFailure;
goto Out;
}
p->nsend++;
if(data != b->data)
p->nfake++;
if(data == b->data){ /* not faking it, so update state */
p->nreal++;
l = b->l;
l.state |= BsVenti;
if(!blockSetLabel(b, &l, 0)){
ret = ArchFailure;
goto Out;
}
}
}
shaBlock(p->score, b, data, p->blockSize);
if(0) fprint(2, "ventisend %V %p %p %p\n", p->score, data, b->data, w.data);
ret = data!=b->data ? ArchFaked : ArchSuccess;
p->l = b->l;
Out:
if(data != b->data)
vtMemFree(data);
p->depth--;
blockPut(b);
return ret;
}
