/*
* 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;
}
/*
* caller must lock f->source and f->msource
* caller must NOT lock the source and msource
* referenced by dir.
*/
static u32int
fileMetaAlloc(File *f, DirEntry *dir, u32int start)
{
u32int nb, bo;
Block *b, *bb;
MetaBlock mb;
int nn;
uchar *p;
int i, n, epb;
MetaEntry me;
Source *s, *ms;
s = f->source;
ms = f->msource;
n = deSize(dir);
nb = (sourceGetSize(ms)+ms->dsize-1)/ms->dsize;
b = nil;
if(start > nb)
start = nb;
for(bo=start; bo<nb; bo++){
b = sourceBlock(ms, bo, OReadWrite);
if(b == nil)
goto Err;
if(!mbUnpack(&mb, b->data, ms->dsize))
goto Err;
nn = (mb.maxsize*FullPercentage/100) - mb.size + mb.free;
if(n <= nn && mb.nindex < mb.maxindex)
break;
blockPut(b);
b = nil;
}
/* add block to meta file */
if(b == nil){
b = sourceBlock(ms, bo, OReadWrite);
if(b == nil)
goto Err;
sourceSetSize(ms, (nb+1)*ms->dsize);
mbInit(&mb, b->data, ms->dsize, ms->dsize/BytesPerEntry);
}
p = mbAlloc(&mb, n);
if(p == nil){
/* mbAlloc might have changed block */
mbPack(&mb);
blockDirty(b);
vtSetError(EBadMeta);
goto Err;
}
mbSearch(&mb, dir->elem, &i, &me);
assert(me.p == nil);
me.p = p;
me.size = n;
dePack(dir, &me);
mbInsert(&mb, i, &me);
mbPack(&mb);
/* meta block depends on super block for qid ... */
bb = cacheLocal(b->c, PartSuper, 0, OReadOnly);
blockDependency(b, bb, -1, nil, nil);
blockPut(bb);
/* ... and one or two dir entries */
epb = s->dsize/VtEntrySize;
bb = sourceBlock(s, dir->entry/epb, OReadOnly);
blockDependency(b, bb, -1, nil, nil);
blockPut(bb);
if(dir->mode & ModeDir){
bb = sourceBlock(s, dir->mentry/epb, OReadOnly);
blockDependency(b, bb, -1, nil, nil);
blockPut(bb);
}
blockDirty(b);
blockPut(b);
return bo;
Err:
blockPut(b);
return NilBlock;
}
/* assumes metaLock is held */
static int
fileMetaFlush2(File *f, char *oelem)
{
File *fp;
Block *b, *bb;
MetaBlock mb;
MetaEntry me, me2;
int i, n;
u32int boff;
if(!f->dirty)
return 0;
if(oelem == nil)
oelem = f->dir.elem;
//print("fileMetaFlush %s->%s\n", oelem, f->dir.elem);
fp = f->up;
if(!sourceLock(fp->msource, -1))
return -1;
/* can happen if source is clri'ed out from under us */
if(f->boff == NilBlock)
goto Err1;
b = sourceBlock(fp->msource, f->boff, OReadWrite);
if(b == nil)
goto Err1;
if(!mbUnpack(&mb, b->data, fp->msource->dsize))
goto Err;
if(!mbSearch(&mb, oelem, &i, &me))
goto Err;
n = deSize(&f->dir);
if(0)fprint(2, "old size %d new size %d\n", me.size, n);
if(mbResize(&mb, &me, n)){
/* fits in the block */
mbDelete(&mb, i);
if(strcmp(f->dir.elem, oelem) != 0)
mbSearch(&mb, f->dir.elem, &i, &me2);
dePack(&f->dir, &me);
mbInsert(&mb, i, &me);
mbPack(&mb);
blockDirty(b);
blockPut(b);
sourceUnlock(fp->msource);
f->dirty = 0;
return 1;
}
/*
* moving entry to another block
* it is feasible for the fs to crash leaving two copies
* of the directory entry. This is just too much work to
* fix. Given that entries are only allocated in a block that
* is less than PercentageFull, most modifications of meta data
* will fit within the block. i.e. this code should almost
* never be executed.
*/
boff = fileMetaAlloc(fp, &f->dir, f->boff+1);
if(boff == NilBlock){
/* mbResize might have modified block */
mbPack(&mb);
blockDirty(b);
goto Err;
}
fprint(2, "fileMetaFlush moving entry from %ud -> %ud\n", f->boff, boff);
f->boff = boff;
/* make sure deletion goes to disk after new entry */
bb = sourceBlock(fp->msource, f->boff, OReadWrite);
mbDelete(&mb, i);
mbPack(&mb);
blockDependency(b, bb, -1, nil, nil);
blockPut(bb);
blockDirty(b);
blockPut(b);
sourceUnlock(fp->msource);
f->dirty = 0;
return 1;
Err:
blockPut(b);
Err1:
sourceUnlock(fp->msource);
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
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;
}