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;
}
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;
}
Fs *
fsOpen(char *file, VtSession *z, int32_t ncache, int mode)
{
int fd, m;
uint8_t oscore[VtScoreSize];
Block *b, *bs;
Disk *disk;
Fs *fs;
Super super;
switch(mode){
default:
vtSetError(EBadMode);
return nil;
case OReadOnly:
m = OREAD;
break;
case OReadWrite:
m = ORDWR;
break;
}
fd = open(file, m);
if(fd < 0){
vtSetError("open %s: %r", file);
return nil;
}
bwatchInit();
disk = diskAlloc(fd);
if(disk == nil){
vtSetError("diskAlloc: %R");
close(fd);
return nil;
}
fs = vtMemAllocZ(sizeof(Fs));
fs->mode = mode;
fs->name = vtStrDup(file);
fs->blockSize = diskBlockSize(disk);
fs->elk = vtLockAlloc();
fs->cache = cacheAlloc(disk, z, ncache, mode);
if(mode == OReadWrite && z)
fs->arch = archInit(fs->cache, disk, fs, z);
fs->z = z;
b = cacheLocal(fs->cache, PartSuper, 0, mode);
if(b == nil)
goto Err;
if(!superUnpack(&super, b->data)){
blockPut(b);
vtSetError("bad super block");
goto Err;
}
blockPut(b);
fs->ehi = super.epochHigh;
fs->elo = super.epochLow;
//fprint(2, "%s: fs->ehi %d fs->elo %d active=%d\n", argv0, fs->ehi, fs->elo, super.active);
fs->source = sourceRoot(fs, super.active, mode);
if(fs->source == nil){
/*
* Perhaps it failed because the block is copy-on-write.
* Do the copy and try again.
*/
if(mode == OReadOnly || strcmp(vtGetError(), EBadRoot) != 0)
goto Err;
b = cacheLocalData(fs->cache, super.active, BtDir, RootTag,
OReadWrite, 0);
if(b == nil){
vtSetError("cacheLocalData: %R");
goto Err;
}
if(b->l.epoch == fs->ehi){
blockPut(b);
vtSetError("bad root source block");
goto Err;
}
b = blockCopy(b, RootTag, fs->ehi, fs->elo);
if(b == nil)
goto Err;
localToGlobal(super.active, oscore);
super.active = b->addr;
bs = cacheLocal(fs->cache, PartSuper, 0, OReadWrite);
if(bs == nil){
blockPut(b);
vtSetError("cacheLocal: %R");
goto Err;
}
superPack(&super, bs->data);
blockDependency(bs, b, 0, oscore, nil);
blockPut(b);
blockDirty(bs);
blockRemoveLink(bs, globalToLocal(oscore), BtDir, RootTag, 0);
blockPut(bs);
fs->source = sourceRoot(fs, super.active, mode);
if(fs->source == nil){
vtSetError("sourceRoot: %R");
goto Err;
}
}
//fprint(2, "%s: got fs source\n", argv0);
vtRLock(fs->elk);
fs->file = fileRoot(fs->source);
fs->source->file = fs->file; /* point back */
vtRUnlock(fs->elk);
if(fs->file == nil){
vtSetError("fileRoot: %R");
goto Err;
}
//fprint(2, "%s: got file root\n", argv0);
if(mode == OReadWrite){
fs->metaFlush = periodicAlloc(fsMetaFlush, fs, 1000);
fs->snap = snapInit(fs);
}
return fs;
Err:
fprint(2, "%s: fsOpen error\n", argv0);
fsClose(fs);
return nil;
}
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;
}
static int
sourceShrinkSize(Source *r, Entry *e, uint64_t size)
{
int i, type, ppb;
uint64_t ptrsz;
uint32_t addr;
uint8_t score[VtScoreSize];
Block *b;
type = entryType(e);
b = cacheGlobal(r->fs->cache, e->score, type, e->tag, OReadWrite);
if(b == nil)
return 0;
ptrsz = e->dsize;
ppb = e->psize/VtScoreSize;
for(i=0; i+1<e->depth; i++)
ptrsz *= ppb;
while(type&BtLevelMask) {
if(b->addr == NilBlock || b->l.epoch != r->fs->ehi) {
/* not worth copying the block just so we can zero some of it */
blockPut(b);
return 0;
}
/*
* invariant: each pointer in the tree rooted at b accounts for ptrsz bytes
*/
/* zero the pointers to unnecessary blocks */
i = (size+ptrsz-1)/ptrsz;
for(; i<ppb; i++) {
addr = globalToLocal(b->data+i*VtScoreSize);
memmove(b->data+i*VtScoreSize, vtZeroScore, VtScoreSize);
blockDirty(b);
if(addr != NilBlock)
blockRemoveLink(b, addr, type-1, e->tag, 1);
}
/* recurse (go around again) on the partially necessary block */
i = size/ptrsz;
size = size%ptrsz;
if(size == 0) {
blockPut(b);
return 1;
}
ptrsz /= ppb;
type--;
memmove(score, b->data+i*VtScoreSize, VtScoreSize);
blockPut(b);
b = cacheGlobal(r->fs->cache, score, type, e->tag, OReadWrite);
if(b == nil)
return 0;
}
if(b->addr == NilBlock || b->l.epoch != r->fs->ehi) {
blockPut(b);
return 0;
}
/*
* No one ever truncates BtDir blocks.
*/
if(type == BtData && e->dsize > size) {
memset(b->data+size, 0, e->dsize-size);
blockDirty(b);
}
blockPut(b);
return 1;
}