static int
dirEntrySize(Source *s, ulong elem, ulong gen, uvlong *size)
{
Block *b;
ulong bn;
Entry e;
int epb;
epb = s->dsize/VtEntrySize;
bn = elem/epb;
elem -= bn*epb;
b = sourceBlock(s, bn, OReadOnly);
if(b == nil)
goto Err;
if(!entryUnpack(&e, b->data, elem))
goto Err;
/* hanging entries are returned as zero size */
if(!(e.flags & VtEntryActive) || e.gen != gen)
*size = 0;
else
*size = e.size;
blockPut(b);
return 1;
Err:
blockPut(b);
return 0;
}
/* assume file is locked, assume f->msource is locked */
static int
fileCheckEmpty(File *f)
{
u32int i, n;
Block *b;
MetaBlock mb;
Source *r;
r = f->msource;
n = (sourceGetSize(r)+r->dsize-1)/r->dsize;
for(i=0; i<n; i++){
b = sourceBlock(r, i, OReadOnly);
if(b == nil)
goto Err;
if(!mbUnpack(&mb, b->data, r->dsize))
goto Err;
if(mb.nindex > 0){
vtSetError(ENotEmpty);
goto Err;
}
blockPut(b);
}
return 1;
Err:
blockPut(b);
return 0;
}
Tnode*
initxsuper(void)
{
Block *b;
Tnode *t;
b = readBlock(PartSuper, 0);
if(b == nil)
return stringnode("reading super: %r");
if(!superUnpack(&super, b->data)) {
blockPut(b);
return stringnode("unpacking super: %r");
}
blockPut(b);
t = stringnode("super "
"version=%#ux "
"epoch=[%#ux,%#ux) "
"qid=%#llux "
"active=%#x "
"next=%#x "
"current=%#x "
"last=%V "
"name=%s",
super.version, super.epochLow, super.epochHigh,
super.qid, super.active, super.next, super.current,
super.last, super.name);
t->expand = xsuperexpand;
return t;
}
static int
getEntry(Source *r, Entry *e, int checkepoch)
{
u32int epoch;
Block *b;
if(r == nil){
memset(&e, 0, sizeof e);
return 1;
}
b = cacheGlobal(r->fs->cache, r->score, BtDir, r->tag, OReadOnly);
if(b == nil)
return 0;
if(!entryUnpack(e, b->data, r->offset % r->epb)){
blockPut(b);
return 0;
}
epoch = b->l.epoch;
blockPut(b);
if(checkepoch){
b = cacheGlobal(r->fs->cache, e->score, entryType(e), e->tag, OReadOnly);
if(b){
if(b->l.epoch >= epoch)
fprint(2, "warning: entry %p epoch not older %#.8ux/%d %V/%d in getEntry\n",
r, b->addr, b->l.epoch, r->score, epoch);
blockPut(b);
}
}
return 1;
}
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 void
checkEpoch(Fsck *chk, u32int epoch)
{
u32int a;
Block *b;
Entry e;
Label l;
chk->print("checking epoch %ud...\n", epoch);
for(a=0; a<chk->nblocks; a++){
if(!readLabel(chk->cache, &l, (a+chk->hint)%chk->nblocks)){
error(chk, "could not read label for addr 0x%.8#ux", a);
continue;
}
if(l.tag == RootTag && l.epoch == epoch)
break;
}
if(a == chk->nblocks){
chk->print("could not find root block for epoch %ud", epoch);
return;
}
a = (a+chk->hint)%chk->nblocks;
b = cacheLocalData(chk->cache, a, BtDir, RootTag, OReadOnly, 0);
if(b == nil){
error(chk, "could not read root block 0x%.8#ux: %R", a);
return;
}
/* no one should point at root blocks */
setBit(chk->amap, a);
setBit(chk->emap, a);
setBit(chk->xmap, a);
/*
* First entry is the rest of the file system.
* Second entry is link to previous epoch root,
* just a convenience to help the search.
*/
if(!entryUnpack(&e, b->data, 0)){
error(chk, "could not unpack root block 0x%.8#ux: %R", a);
blockPut(b);
return;
}
walkEpoch(chk, b, e.score, BtDir, e.tag, epoch);
if(entryUnpack(&e, b->data, 1))
chk->hint = globalToLocal(e.score);
blockPut(b);
}
/*
* 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 int
fileMetaRemove(File *f, char *uid)
{
Block *b;
MetaBlock mb;
MetaEntry me;
int i;
File *up;
up = f->up;
fileWAccess(up, uid);
fileMetaLock(f);
sourceLock(up->msource, OReadWrite);
b = sourceBlock(up->msource, f->boff, OReadWrite);
if(b == nil)
goto Err;
if(!mbUnpack(&mb, b->data, up->msource->dsize))
{
fprint(2, "U\n");
goto Err;
}
if(!mbSearch(&mb, f->dir.elem, &i, &me))
{
fprint(2, "S\n");
goto Err;
}
mbDelete(&mb, i);
mbPack(&mb);
sourceUnlock(up->msource);
blockDirty(b);
blockPut(b);
f->removed = 1;
f->boff = NilBlock;
f->dirty = 0;
fileMetaUnlock(f);
return 1;
Err:
sourceUnlock(up->msource);
blockPut(b);
fileMetaUnlock(f);
return 0;
}
static int
deeFill(DirEntryEnum *dee)
{
int i, n;
Source *meta, *source;
MetaBlock mb;
MetaEntry me;
File *f;
Block *b;
DirEntry *de;
/* clean up first */
for(i=dee->i; i<dee->n; i++)
deCleanup(dee->buf+i);
vtMemFree(dee->buf);
dee->buf = nil;
dee->i = 0;
dee->n = 0;
f = dee->file;
source = f->source;
meta = f->msource;
b = sourceBlock(meta, dee->boff, OReadOnly);
if(b == nil)
goto Err;
if(!mbUnpack(&mb, b->data, meta->dsize))
goto Err;
n = mb.nindex;
dee->buf = vtMemAlloc(n * sizeof(DirEntry));
for(i=0; i<n; i++){
de = dee->buf + i;
meUnpack(&me, &mb, i);
if(!deUnpack(de, &me))
goto Err;
dee->n++;
if(!(de->mode & ModeDir))
if(!dirEntrySize(source, de->entry, de->gen, &de->size))
goto Err;
}
dee->boff++;
blockPut(b);
return 1;
Err:
blockPut(b);
return 0;
}
static void
scanSource(Fsck *chk, char *name, Source *r)
{
u32int a, nb, o;
Block *b;
Entry e;
if(!chk->useventi && globalToLocal(r->score)==NilBlock)
return;
if(!sourceGetEntry(r, &e)){
error(chk, "could not get entry for %s", name);
return;
}
a = globalToLocal(e.score);
if(!chk->useventi && a==NilBlock)
return;
if(getBit(chk->smap, a))
return;
setBit(chk->smap, a);
nb = (sourceGetSize(r) + r->dsize-1) / r->dsize;
for(o = 0; o < nb; o++){
b = sourceBlock(r, o, OReadOnly);
if(b == nil){
error(chk, "could not read block in data file %s", name);
continue;
}
if(b->addr != NilBlock && getBit(chk->errmap, b->addr)){
warn(chk, "previously reported error in block %ux is in file %s",
b->addr, name);
}
blockPut(b);
}
}
int
fsEpochLow(Fs *fs, uint32_t low)
{
Block *bs;
Super super;
vtLock(fs->elk);
if(low > fs->ehi){
vtSetError("bad low epoch (must be <= %ud)", fs->ehi);
vtUnlock(fs->elk);
return 0;
}
if((bs = superGet(fs->cache, &super)) == nil){
vtUnlock(fs->elk);
return 0;
}
super.epochLow = low;
fs->elo = low;
superWrite(bs, &super, 1);
blockPut(bs);
vtUnlock(fs->elk);
return 1;
}
/*
* Lock two (usually sibling) sources. This needs special care
* because the Entries for both sources might be in the same block.
* We also try to lock blocks in left-to-right order within the tree.
*/
int
sourceLock2(Source *r, Source *rr, int mode)
{
Block *b, *bb;
if(rr == nil)
return sourceLock(r, mode);
if(mode == -1)
mode = r->mode;
if(r->parent==rr->parent && r->offset/r->epb == rr->offset/rr->epb) {
b = sourceLoadBlock(r, mode);
if(b == nil)
return 0;
if(memcmp(r->score, rr->score, VtScoreSize) != 0) {
memmove(rr->score, b->score, VtScoreSize);
rr->scoreEpoch = b->l.epoch;
rr->tag = b->l.tag;
rr->epoch = rr->fs->ehi;
}
blockDupLock(b);
bb = b;
} else if(r->parent==rr->parent || r->offset > rr->offset) {
bb = sourceLoadBlock(rr, mode);
b = sourceLoadBlock(r, mode);
} else {
b = sourceLoadBlock(r, mode);
bb = sourceLoadBlock(rr, mode);
}
if(b == nil || bb == nil) {
if(b)
blockPut(b);
if(bb)
blockPut(bb);
return 0;
}
/*
* The fact that we are holding b and bb serves
* as the lock entitling us to write to r->b and rr->b.
*/
r->b = b;
rr->b = bb;
return 1;
}
/*
* the file is locked already
* f->msource is unlocked
*/
static File *
dirLookup(File *f, char *elem)
{
int i;
MetaBlock mb;
MetaEntry me;
Block *b;
Source *meta;
File *ff;
u32int bo, nb;
meta = f->msource;
b = nil;
if(!sourceLock(meta, -1))
return nil;
nb = (sourceGetSize(meta)+meta->dsize-1)/meta->dsize;
for(bo=0; bo<nb; bo++){
b = sourceBlock(meta, bo, OReadOnly);
if(b == nil)
goto Err;
if(!mbUnpack(&mb, b->data, meta->dsize))
goto Err;
if(mbSearch(&mb, elem, &i, &me)){
ff = fileAlloc(f->fs);
if(!deUnpack(&ff->dir, &me)){
fileFree(ff);
goto Err;
}
sourceUnlock(meta);
blockPut(b);
ff->boff = bo;
ff->mode = f->mode;
ff->issnapshot = f->issnapshot;
return ff;
}
blockPut(b);
b = nil;
}
vtSetError(ENoFile);
/* fall through */
Err:
sourceUnlock(meta);
blockPut(b);
return nil;
}
void
sourceUnlock(Source *r)
{
Block *b;
if(r->b == nil) {
fprint(2, "sourceUnlock: already unlocked\n");
abort();
}
b = r->b;
r->b = nil;
blockPut(b);
}
int
sourceGetEntry(Source *r, Entry *e)
{
Block *b;
assert(sourceIsLocked(r));
b = sourceLoad(r, e);
if(b == nil)
return 0;
blockPut(b);
return 1;
}
Source *
sourceRoot(Fs *fs, uint32_t addr, int mode)
{
Source *r;
Block *b;
b = cacheLocalData(fs->cache, addr, BtDir, RootTag, mode, 0);
if(b == nil)
return nil;
if(mode == OReadWrite && b->l.epoch != fs->ehi) {
consPrint("sourceRoot: fs->ehi = %ud, b->l = %L\n",
fs->ehi, &b->l);
blockPut(b);
vtSetError(EBadRoot);
return nil;
}
r = sourceAlloc(fs, b, nil, 0, mode, 0);
blockPut(b);
return r;
}
/*
* 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;
}
Block*
readBlock(int part, u32int addr)
{
u32int start, end;
u64int offset;
int n, nn;
Block *b;
uchar *buf;
start = partStart(part);
end = partEnd(part);
if(addr >= end-start) {
werrstr("bad addr 0x%.8ux; wanted 0x%.8ux - 0x%.8ux", addr, start, end);
return nil;
}
b = allocBlock();
b->addr = addr;
buf = b->data;
offset = ((u64int)(addr+start))*h.blockSize;
n = h.blockSize;
while(n > 0) {
nn = pread(fd, buf, n, offset);
if(nn < 0) {
blockPut(b);
return nil;
}
if(nn == 0) {
werrstr("short read");
blockPut(b);
return nil;
}
n -= nn;
offset += nn;
buf += nn;
}
return b;
}
Block*
dataBlock(uchar score[VtScoreSize], uint type, uint tag)
{
Block *b, *bl;
int lpb;
Label l;
u32int addr;
addr = globalToLocal(score);
if(addr == NilBlock)
return ventiBlock(score, type);
lpb = h.blockSize/LabelSize;
bl = readBlock(PartLabel, addr/lpb);
if(bl == nil)
return nil;
if(!labelUnpack(&l, bl->data, addr%lpb)) {
werrstr("%r");
blockPut(bl);
return nil;
}
blockPut(bl);
if(l.type != type) {
werrstr("type mismatch; got %d (%s) wanted %d (%s)",
l.type, btStr(l.type), type, btStr(type));
return nil;
}
if(tag && l.tag != tag) {
werrstr("tag mismatch; got 0x%.8ux wanted 0x%.8ux",
l.tag, tag);
return nil;
}
b = readBlock(PartData, addr);
if(b == nil)
return nil;
b->l = l;
return b;
}
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;
}
uint64_t
sourceGetSize(Source *r)
{
Entry e;
Block *b;
assert(sourceIsLocked(r));
b = sourceLoad(r, &e);
if(b == nil)
return 0;
blockPut(b);
return e.size;
}
/*
* 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;
}
/*
* We've just walked the whole write buffer. Notice blocks that
* aren't marked available but that we didn't visit. They are lost.
*/
static void
checkLeak(Fsck *chk)
{
u32int a, nfree, nlost;
Block *b;
Label l;
nfree = 0;
nlost = 0;
for(a = 0; a < chk->nblocks; a++){
if(!readLabel(chk->cache, &l, a)){
error(chk, "could not read label: addr 0x%ux %d %d: %R",
a, l.type, l.state);
continue;
}
if(getBit(chk->amap, a))
continue;
if(l.state == BsFree || l.epochClose <= chk->fs->elo ||
l.epochClose == l.epoch){
nfree++;
setBit(chk->amap, a);
continue;
}
if(l.state&BsClosed)
continue;
nlost++;
// warn(chk, "unreachable block: addr 0x%ux type %d tag 0x%ux "
// "state %s epoch %ud close %ud", a, l.type, l.tag,
// bsStr(l.state), l.epoch, l.epochClose);
b = cacheLocal(chk->cache, PartData, a, OReadOnly);
if(b == nil){
error(chk, "could not read block 0x%#.8ux", a);
continue;
}
chk->close(chk, b, 0);
chk->nclose++;
setBit(chk->amap, a);
blockPut(b);
}
chk->print("fsys blocks: total=%ud used=%ud(%.1f%%) free=%ud(%.1f%%) lost=%ud(%.1f%%)\n",
chk->nblocks,
chk->nblocks - nfree-nlost,
100.*(chk->nblocks - nfree - nlost)/chk->nblocks,
nfree, 100.*nfree/chk->nblocks,
nlost, 100.*nlost/chk->nblocks);
}
int
saveQid(Fs *fs)
{
Block *b;
Super super;
uint64_t qidMax;
if((b = superGet(fs->cache, &super)) == nil)
return 0;
qidMax = super.qid;
blockPut(b);
if(!fileSetQidSpace(fs->file, 0, qidMax))
return 0;
return 1;
}
Block*
superGet(Cache *c, Super* super)
{
Block *b;
if((b = cacheLocal(c, PartSuper, 0, OReadWrite)) == nil){
fprint(2, "%s: superGet: cacheLocal failed: %R\n", argv0);
return nil;
}
if(!superUnpack(super, b->data)){
fprint(2, "%s: superGet: superUnpack failed: %R\n", argv0);
blockPut(b);
return nil;
}
return b;
}
int
fsNextQid(Fs *fs, uint64_t *qid)
{
Block *b;
Super super;
if((b = superGet(fs->cache, &super)) == nil)
return 0;
*qid = super.qid++;
/*
* It's okay if the super block doesn't go to disk immediately,
* since fileMetaAlloc will record a dependency between the
* block holding this qid and the super block. See file.c:/^fileMetaAlloc.
*/
superWrite(b, &super, 0);
blockPut(b);
return 1;
}
void
fsCheck(Fsck *chk)
{
Block *b;
Super super;
checkInit(chk);
b = superGet(chk->cache, &super);
if(b == nil){
chk->print("could not load super block: %R");
return;
}
blockPut(b);
chk->hint = super.active;
checkEpochs(chk);
chk->smap = vtMemAllocZ(chk->nblocks/8+1);
checkDirs(chk);
vtMemFree(chk->smap);
}
Source *
sourceOpen(Source *r, uint32_t offset, int mode, int issnapshot)
{
uint32_t bn;
Block *b;
assert(sourceIsLocked(r));
if(r->mode == OReadWrite)
assert(r->epoch == r->b->l.epoch);
if(!r->dir) {
vtSetError(ENotDir);
return nil;
}
bn = offset/(r->dsize/VtEntrySize);
b = sourceBlock(r, bn, mode);
if(b == nil)
return nil;
r = sourceAlloc(r->fs, b, r, offset, mode, issnapshot);
blockPut(b);
return r;
}
Block*
ventiBlock(uchar score[VtScoreSize], uint type)
{
int n;
Block *b;
b = allocBlock();
memmove(b->score, score, VtScoreSize);
b->addr = NilBlock;
n = vtread(z, b->score, vtType[type], b->data, h.blockSize);
if(n < 0) {
fprint(2, "vtread returns %d: %r\n", n);
blockPut(b);
return nil;
}
vtzeroextend(vtType[type], b->data, n, h.blockSize);
b->l.type = type;
b->l.state = 0;
b->l.tag = 0;
b->l.epoch = 0;
return b;
}