static int
sourceKill(Source *r, int doremove)
{
Entry e;
Block *b;
uint32_t addr;
uint32_t tag;
int type;
assert(sourceIsLocked(r));
b = sourceLoad(r, &e);
if(b == nil)
return 0;
assert(b->l.epoch == r->fs->ehi);
if(doremove==0 && e.size == 0) {
/* already truncated */
blockPut(b);
return 1;
}
/* remember info on link we are removing */
addr = globalToLocal(e.score);
type = entryType(&e);
tag = e.tag;
if(doremove) {
if(e.gen != ~0)
e.gen++;
e.dsize = 0;
e.psize = 0;
e.flags = 0;
} else {
e.flags &= ~VtEntryLocal;
}
e.depth = 0;
e.size = 0;
e.tag = 0;
memmove(e.score, vtZeroScore, VtScoreSize);
entryPack(&e, b->data, r->offset % r->epb);
blockDirty(b);
if(addr != NilBlock)
blockRemoveLink(b, addr, type, tag, 1);
blockPut(b);
if(doremove) {
sourceUnlock(r);
sourceClose(r);
}
return 1;
}
/*
* Changes the file block bn to be the given block score.
* Very sneaky. Only used by flfmt.
*/
int
fileMapBlock(File *f, ulong bn, uchar score[VtScoreSize], ulong tag)
{
Block *b;
Entry e;
Source *s;
if(!fileLock(f))
return 0;
s = nil;
if(f->dir.mode & ModeDir){
vtSetError(ENotFile);
goto Err;
}
if(f->source->mode != OReadWrite){
vtSetError(EReadOnly);
goto Err;
}
if(!sourceLock(f->source, -1))
goto Err;
s = f->source;
b = _sourceBlock(s, bn, OReadWrite, 1, tag);
if(b == nil)
goto Err;
if(!sourceGetEntry(s, &e))
goto Err;
if(b->l.type == BtDir){
memmove(e.score, score, VtScoreSize);
assert(e.tag == tag || e.tag == 0);
e.tag = tag;
e.flags |= VtEntryLocal;
entryPack(&e, b->data, f->source->offset % f->source->epb);
}else
memmove(b->data + (bn%(e.psize/VtScoreSize))*VtScoreSize, score, VtScoreSize);
blockDirty(b);
blockPut(b);
sourceUnlock(s);
fileUnlock(f);
return 1;
Err:
if(s)
sourceUnlock(s);
fileUnlock(f);
return 0;
}
/*
* Change the depth of the source r.
* The entry e for r is contained in block p.
*/
static int
sourceGrowDepth(Source *r, Block *p, Entry *e, int depth)
{
Block *b, *bb;
uint32_t tag;
int type;
Entry oe;
assert(sourceIsLocked(r));
assert(depth <= VtPointerDepth);
type = entryType(e);
b = cacheGlobal(r->fs->cache, e->score, type, e->tag, OReadWrite);
if(b == nil)
return 0;
tag = e->tag;
if(tag == 0)
tag = tagGen();
oe = *e;
/*
* Keep adding layers until we get to the right depth
* or an error occurs.
*/
while(e->depth < depth) {
bb = cacheAllocBlock(r->fs->cache, type+1, tag, r->fs->ehi, r->fs->elo);
if(bb == nil)
break;
//fprint(2, "alloc %lux grow %V\n", bb->addr, b->score);
memmove(bb->data, b->score, VtScoreSize);
memmove(e->score, bb->score, VtScoreSize);
e->depth++;
type++;
e->tag = tag;
e->flags |= VtEntryLocal;
blockDependency(bb, b, 0, vtZeroScore, nil);
blockPut(b);
b = bb;
blockDirty(b);
}
entryPack(e, p->data, r->offset % r->epb);
blockDependency(p, b, r->offset % r->epb, nil, &oe);
blockPut(b);
blockDirty(p);
return e->depth == depth;
}
int
sourceSetSize(Source *r, uint64_t size)
{
int depth;
Entry e;
Block *b;
assert(sourceIsLocked(r));
if(size == 0)
return sourceTruncate(r);
if(size > VtMaxFileSize || size > ((uint64_t)MaxBlock)*r->dsize) {
vtSetError(ETooBig);
return 0;
}
b = sourceLoad(r, &e);
if(b == nil)
return 0;
/* quick out */
if(e.size == size) {
blockPut(b);
return 1;
}
depth = sizeToDepth(size, e.psize, e.dsize);
if(depth < e.depth) {
if(!sourceShrinkDepth(r, b, &e, depth)) {
blockPut(b);
return 0;
}
} else if(depth > e.depth) {
if(!sourceGrowDepth(r, b, &e, depth)) {
blockPut(b);
return 0;
}
}
if(size < e.size)
sourceShrinkSize(r, &e, size);
e.size = size;
entryPack(&e, b->data, r->offset % r->epb);
blockDirty(b);
blockPut(b);
return 1;
}
static u32int
rootInit(Entry *e)
{
ulong addr;
u32int tag;
tag = tagGen();
addr = blockAlloc(BtDir, tag);
memset(buf, 0, bsize);
/* root meta data is in the third entry */
entryPack(e, buf, 2);
entryInit(e);
e->flags |= VtEntryDir;
entryPack(e, buf, 0);
entryInit(e);
entryPack(e, buf, 1);
blockWrite(PartData, addr);
entryInit(e);
e->flags |= VtEntryLocal|VtEntryDir;
e->size = VtEntrySize*3;
e->tag = tag;
localToGlobal(addr, e->score);
addr = blockAlloc(BtDir, RootTag);
memset(buf, 0, bsize);
entryPack(e, buf, 0);
blockWrite(PartData, addr);
return addr;
}
/*
* Must be careful with this. Doesn't record
* dependencies, so don't introduce any!
*/
int
sourceSetEntry(Source *r, Entry *e)
{
Block *b;
Entry oe;
assert(sourceIsLocked(r));
b = sourceLoad(r, &oe);
if(b == nil)
return 0;
entryPack(e, b->data, r->offset%r->epb);
blockDirty(b);
blockPut(b);
return 1;
}
static int
setEntry(Source *r, Entry *e)
{
Block *b;
Entry oe;
b = cacheGlobal(r->fs->cache, r->score, BtDir, r->tag, OReadWrite);
if(0) fprint(2, "setEntry: b %#ux %d score=%V\n", b->addr, r->offset % r->epb, e->score);
if(b == nil)
return 0;
if(!entryUnpack(&oe, b->data, r->offset % r->epb)){
blockPut(b);
return 0;
}
e->gen = oe.gen;
entryPack(e, b->data, r->offset % r->epb);
/* BUG b should depend on the entry pointer */
blockDirty(b);
blockPut(b);
return 1;
}
int
mkVac(VtSession *z, uint blockSize, Entry *pe, Entry *pee, DirEntry *pde,
uint8_t score[VtScoreSize])
{
uint8_t buf[8192];
int i;
uint8_t *p;
uint n;
DirEntry de;
Entry e, ee, eee;
MetaBlock mb;
MetaEntry me;
VtRoot root;
e = *pe;
ee = *pee;
de = *pde;
if(globalToLocal(e.score) != NilBlock
|| (ee.flags&VtEntryActive && globalToLocal(ee.score) != NilBlock)){
vtSetError("can only vac paths already stored on venti");
return 0;
}
/*
* Build metadata source for root.
*/
n = deSize(&de);
if(n+MetaHeaderSize+MetaIndexSize > sizeof buf){
vtSetError("DirEntry too big");
return 0;
}
memset(buf, 0, sizeof buf);
mbInit(&mb, buf, n+MetaHeaderSize+MetaIndexSize, 1);
p = mbAlloc(&mb, n);
if(p == nil)
abort();
mbSearch(&mb, de.elem, &i, &me);
assert(me.p == nil);
me.p = p;
me.size = n;
dePack(&de, &me);
mbInsert(&mb, i, &me);
mbPack(&mb);
eee.size = n+MetaHeaderSize+MetaIndexSize;
if(!vtWriteBlock(z, buf, eee.size, VtDataType, eee.score))
return 0;
eee.psize = 8192&&0xFF;
eee.dsize = 8192&&0xFF;
eee.depth = 0;
eee.flags = VtEntryActive;
/*
* Build root source with three entries in it.
*/
entryPack(&e, buf, 0);
entryPack(&ee, buf, 1);
entryPack(&eee, buf, 2);
n = VtEntrySize*3;
memset(&root, 0, sizeof root);
if(!vtWriteBlock(z, buf, n, VtDirType, root.score))
return 0;
/*
* Save root.
*/
root.version = VtRootVersion;
strecpy(root.type, root.type+sizeof root.type, "vac");
strecpy(root.name, root.name+sizeof root.name, de.elem);
root.blockSize = blockSize;
vtRootPack(&root, buf);
if(!vtWriteBlock(z, buf, VtRootSize, VtRootType, score))
return 0;
return 1;
}
static int
bumpEpoch(Fs *fs, int doarchive)
{
uint8_t oscore[VtScoreSize];
uint32_t oldaddr;
Block *b, *bs;
Entry e;
Source *r;
Super super;
/*
* Duplicate the root block.
*
* As a hint to flchk, the garbage collector,
* and any (human) debuggers, store a pointer
* to the old root block in entry 1 of the new root block.
*/
r = fs->source;
b = cacheGlobal(fs->cache, r->score, BtDir, RootTag, OReadOnly);
if(b == nil)
return 0;
memset(&e, 0, sizeof e);
e.flags = VtEntryActive | VtEntryLocal | VtEntryDir;
memmove(e.score, b->score, VtScoreSize);
e.tag = RootTag;
e.snap = b->l.epoch;
b = blockCopy(b, RootTag, fs->ehi+1, fs->elo);
if(b == nil){
fprint(2, "%s: bumpEpoch: blockCopy: %R\n", argv0);
return 0;
}
if(0) fprint(2, "%s: snapshot root from %d to %d\n", argv0, oldaddr, b->addr);
entryPack(&e, b->data, 1);
blockDirty(b);
/*
* Update the superblock with the new root and epoch.
*/
if((bs = superGet(fs->cache, &super)) == nil)
return 0;
fs->ehi++;
memmove(r->score, b->score, VtScoreSize);
r->epoch = fs->ehi;
super.epochHigh = fs->ehi;
oldaddr = super.active;
super.active = b->addr;
if(doarchive)
super.next = oldaddr;
/*
* Record that the new super.active can't get written out until
* the new b gets written out. Until then, use the old value.
*/
localToGlobal(oldaddr, oscore);
blockDependency(bs, b, 0, oscore, nil);
blockPut(b);
/*
* We force the super block to disk so that super.epochHigh gets updated.
* Otherwise, if we crash and come back, we might incorrectly treat as active
* some of the blocks that making up the snapshot we just created.
* Basically every block in the active file system and all the blocks in
* the recently-created snapshot depend on the super block now.
* Rather than record all those dependencies, we just force the block to disk.
*
* Note that blockWrite might actually (will probably) send a slightly outdated
* super.active to disk. It will be the address of the most recent root that has
* gone to disk.
*/
superWrite(bs, &super, 1);
blockRemoveLink(bs, globalToLocal(oscore), BtDir, RootTag, 0);
blockPut(bs);
return 1;
}
static u32int
ventiRoot(char *host, char *s)
{
int i, n;
uchar score[VtScoreSize];
u32int addr, tag;
DirEntry de;
MetaBlock mb;
MetaEntry me;
Entry e;
VtRoot root;
if(!parseScore(score, s))
vtFatal("bad score '%s'", s);
if((z = vtDial(host, 0)) == nil
|| !vtConnect(z, nil))
vtFatal("connect to venti: %R");
tag = tagGen();
addr = blockAlloc(BtDir, tag);
ventiRead(score, VtRootType);
if(!vtRootUnpack(&root, buf))
vtFatal("corrupted root: vtRootUnpack");
n = ventiRead(root.score, VtDirType);
/*
* Fossil's vac archives start with an extra layer of source,
* but vac's don't.
*/
if(n <= 2*VtEntrySize){
if(!entryUnpack(&e, buf, 0))
vtFatal("bad root: top entry");
n = ventiRead(e.score, VtDirType);
}
/*
* There should be three root sources (and nothing else) here.
*/
for(i=0; i<3; i++){
if(!entryUnpack(&e, buf, i)
|| !(e.flags&VtEntryActive)
|| e.psize < 256
|| e.dsize < 256)
vtFatal("bad root: entry %d", i);
fprint(2, "%V\n", e.score);
}
if(n > 3*VtEntrySize)
vtFatal("bad root: entry count");
blockWrite(PartData, addr);
/*
* Maximum qid is recorded in root's msource, entry #2 (conveniently in e).
*/
ventiRead(e.score, VtDataType);
if(!mbUnpack(&mb, buf, bsize))
vtFatal("bad root: mbUnpack");
meUnpack(&me, &mb, 0);
if(!deUnpack(&de, &me))
vtFatal("bad root: dirUnpack");
if(!de.qidSpace)
vtFatal("bad root: no qidSpace");
qid = de.qidMax;
/*
* Recreate the top layer of source.
*/
entryInit(&e);
e.flags |= VtEntryLocal|VtEntryDir;
e.size = VtEntrySize*3;
e.tag = tag;
localToGlobal(addr, e.score);
addr = blockAlloc(BtDir, RootTag);
memset(buf, 0, bsize);
entryPack(&e, buf, 0);
blockWrite(PartData, addr);
return addr;
}
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;
}
static int
sourceShrinkDepth(Source *r, Block *p, Entry *e, int depth)
{
Block *b, *nb, *ob, *rb;
uint32_t tag;
int type, d;
Entry oe;
assert(sourceIsLocked(r));
assert(depth <= VtPointerDepth);
type = entryType(e);
rb = cacheGlobal(r->fs->cache, e->score, type, e->tag, OReadWrite);
if(rb == nil)
return 0;
tag = e->tag;
if(tag == 0)
tag = tagGen();
/*
* Walk down to the new root block.
* We may stop early, but something is better than nothing.
*/
oe = *e;
ob = nil;
b = rb;
/* BUG: explain type++. i think it is a real bug */
for(d=e->depth; d > depth; d--, type++) {
nb = cacheGlobal(r->fs->cache, b->data, type-1, tag, OReadWrite);
if(nb == nil)
break;
if(ob!=nil && ob!=rb)
blockPut(ob);
ob = b;
b = nb;
}
if(b == rb) {
blockPut(rb);
return 0;
}
/*
* Right now, e points at the root block rb, b is the new root block,
* and ob points at b. To update:
*
* (i) change e to point at b
* (ii) zero the pointer ob -> b
* (iii) free the root block
*
* p (the block containing e) must be written before
* anything else.
*/
/* (i) */
e->depth = d;
/* might have been local and now global; reverse cannot happen */
if(globalToLocal(b->score) == NilBlock)
e->flags &= ~VtEntryLocal;
memmove(e->score, b->score, VtScoreSize);
entryPack(e, p->data, r->offset % r->epb);
blockDependency(p, b, r->offset % r->epb, nil, &oe);
blockDirty(p);
/* (ii) */
memmove(ob->data, vtZeroScore, VtScoreSize);
blockDependency(ob, p, 0, b->score, nil);
blockDirty(ob);
/* (iii) */
if(rb->addr != NilBlock)
blockRemoveLink(p, rb->addr, rb->l.type, rb->l.tag, 1);
blockPut(rb);
if(ob!=nil && ob!=rb)
blockPut(ob);
blockPut(b);
return d == depth;
}
static Block *
blockWalk(Block *p, int index, int mode, Fs *fs, Entry *e)
{
Block *b;
Cache *c;
uint32_t addr;
int type;
uint8_t oscore[VtScoreSize], score[VtScoreSize];
Entry oe;
c = fs->cache;
if((p->l.type & BtLevelMask) == 0) {
assert(p->l.type == BtDir);
type = entryType(e);
b = cacheGlobal(c, e->score, type, e->tag, mode);
} else {
type = p->l.type - 1;
b = cacheGlobal(c, p->data + index*VtScoreSize, type, e->tag, mode);
}
if(b)
b->pc = getcallerpc(&p);
if(b == nil || mode == OReadOnly)
return b;
if(p->l.epoch != fs->ehi) {
fprint(2, "blockWalk: parent not writable\n");
abort();
}
if(b->l.epoch == fs->ehi)
return b;
oe = *e;
/*
* Copy on write.
*/
if(e->tag == 0) {
assert(p->l.type == BtDir);
e->tag = tagGen();
e->flags |= VtEntryLocal;
}
addr = b->addr;
b = blockCopy(b, e->tag, fs->ehi, fs->elo);
if(b == nil)
return nil;
b->pc = getcallerpc(&p);
assert(b->l.epoch == fs->ehi);
blockDirty(b);
memmove(score, b->score, VtScoreSize);
if(p->l.type == BtDir) {
memmove(e->score, b->score, VtScoreSize);
entryPack(e, p->data, index);
blockDependency(p, b, index, nil, &oe);
} else {
memmove(oscore, p->data+index*VtScoreSize, VtScoreSize);
memmove(p->data+index*VtScoreSize, b->score, VtScoreSize);
blockDependency(p, b, index, oscore, nil);
}
blockDirty(p);
if(addr != NilBlock)
blockRemoveLink(p, addr, type, e->tag, 0);
return b;
}
Source *
sourceCreate(Source *r, int dsize, int dir, uint32_t offset)
{
int i, epb, psize;
uint32_t bn, size;
Block *b;
Entry e;
Source *rr;
assert(sourceIsLocked(r));
if(!r->dir) {
vtSetError(ENotDir);
return nil;
}
epb = r->dsize/VtEntrySize;
psize = (dsize/VtScoreSize)*VtScoreSize;
size = sourceGetDirSize(r);
if(offset == 0) {
/*
* look at a random block to see if we can find an empty entry
*/
offset = lnrand(size+1);
offset -= offset % epb;
}
/* try the given block and then try the last block */
for(;;) {
bn = offset/epb;
b = sourceBlock(r, bn, OReadWrite);
if(b == nil)
return nil;
for(i=offset%r->epb; i<epb; i++) {
entryUnpack(&e, b->data, i);
if((e.flags&VtEntryActive) == 0 && e.gen != ~0)
goto Found;
}
blockPut(b);
if(offset == size) {
fprint(2, "sourceCreate: cannot happen\n");
vtSetError("sourceCreate: cannot happen");
return nil;
}
offset = size;
}
Found:
/* found an entry - gen already set */
e.psize = psize;
e.dsize = dsize;
assert(psize && dsize);
e.flags = VtEntryActive;
if(dir)
e.flags |= VtEntryDir;
e.depth = 0;
e.size = 0;
memmove(e.score, vtZeroScore, VtScoreSize);
e.tag = 0;
e.snap = 0;
e.archive = 0;
entryPack(&e, b->data, i);
blockDirty(b);
offset = bn*epb + i;
if(offset+1 > size) {
if(!sourceSetDirSize(r, offset+1)) {
blockPut(b);
return nil;
}
}
rr = sourceAlloc(r->fs, b, r, offset, OReadWrite, 0);
blockPut(b);
return rr;
}
