/*
* Replace the contents of the locked destination with the contents of the
* locked source. Destination must have one ref.
*
* Returns with the destination still with one ref and the copied chains
* with an additional lock (representing their state on the destination).
* The original chains associated with the destination are unlocked.
*/
void
hammer2_cluster_replace_locked(hammer2_cluster_t *dst, hammer2_cluster_t *src)
{
hammer2_chain_t *chain;
int i;
KKASSERT(dst->refs == 1);
dst->focus = NULL;
for (i = 0; i < src->nchains; ++i) {
chain = src->array[i];
if (chain) {
hammer2_chain_lock(chain, 0);
if (i < dst->nchains && dst->array[i])
hammer2_chain_unlock(dst->array[i]);
dst->array[i] = src->array[i];
if (dst->focus == NULL)
dst->focus = chain;
}
}
while (i < dst->nchains) {
chain = dst->array[i];
if (chain) {
hammer2_chain_unlock(chain);
dst->array[i] = NULL;
}
++i;
}
dst->nchains = src->nchains;
}
/*
* Destroy an extranious chain.
*
* Both *parentp and *chainp are locked shared.
*
* On return, *chainp will be adjusted to point to the next element in the
* iteration and locked shared.
*/
static
int
hammer2_sync_destroy(hammer2_thread_t *thr,
hammer2_chain_t **parentp, hammer2_chain_t **chainp,
hammer2_tid_t mtid, int idx)
{
hammer2_chain_t *chain;
hammer2_chain_t *parent;
hammer2_key_t key_next;
hammer2_key_t save_key;
int cache_index = -1;
chain = *chainp;
#if HAMMER2_THREAD_DEBUG
if (hammer2_debug & 1)
kprintf("destroy rec %p/%p slave %d %d.%016jx\n",
*parentp, chain,
idx, chain->bref.type, chain->bref.key);
#endif
save_key = chain->bref.key;
if (save_key != HAMMER2_KEY_MAX)
++save_key;
/*
* Try to avoid unnecessary I/O.
*
* XXX accounting not propagated up properly. We might have to do
* a RESOLVE_MAYBE here and pass 0 for the flags.
*/
hammer2_chain_unlock(chain); /* relock exclusive */
hammer2_chain_unlock(*parentp);
hammer2_chain_lock(*parentp, HAMMER2_RESOLVE_ALWAYS);
hammer2_chain_lock(chain, HAMMER2_RESOLVE_NEVER);
hammer2_chain_delete(*parentp, chain, mtid, HAMMER2_DELETE_PERMANENT);
hammer2_chain_unlock(chain);
hammer2_chain_drop(chain);
chain = NULL; /* safety */
hammer2_chain_unlock(*parentp); /* relock shared */
hammer2_chain_lock(*parentp, HAMMER2_RESOLVE_SHARED |
HAMMER2_RESOLVE_ALWAYS);
*chainp = hammer2_chain_lookup(&parent, &key_next,
save_key, HAMMER2_KEY_MAX,
&cache_index,
HAMMER2_LOOKUP_SHARED |
HAMMER2_LOOKUP_NODIRECT |
HAMMER2_LOOKUP_NODATA);
return 0;
}
hammer2_chain_t *
hammer2_inode_chain_and_parent(hammer2_inode_t *ip, int clindex,
hammer2_chain_t **parentp, int how)
{
hammer2_chain_t *chain;
hammer2_chain_t *parent;
for (;;) {
hammer2_spin_sh(&ip->cluster_spin);
if (clindex >= ip->cluster.nchains)
chain = NULL;
else
chain = ip->cluster.array[clindex].chain;
if (chain) {
hammer2_chain_ref(chain);
hammer2_spin_unsh(&ip->cluster_spin);
hammer2_chain_lock(chain, how);
} else {
hammer2_spin_unsh(&ip->cluster_spin);
}
/*
* Get parent, lock order must be (parent, chain).
*/
parent = chain->parent;
if (parent) {
hammer2_chain_ref(parent);
hammer2_chain_unlock(chain);
hammer2_chain_lock(parent, how);
hammer2_chain_lock(chain, how);
}
if (ip->cluster.array[clindex].chain == chain &&
chain->parent == parent) {
break;
}
/*
* Retry
*/
hammer2_chain_unlock(chain);
hammer2_chain_drop(chain);
if (parent) {
hammer2_chain_unlock(parent);
hammer2_chain_drop(parent);
}
}
*parentp = parent;
return chain;
}
/*
* NOTE: We don't combine the inode/chain lock because putting away an
* inode would otherwise confuse multiple lock holders of the inode.
*
* Shared locks are especially sensitive to having too many shared
* lock counts (from the same thread) on certain paths which might
* need to upgrade them. Only one count of a shared lock can be
* upgraded.
*/
hammer2_chain_t *
hammer2_inode_lock_sh(hammer2_inode_t *ip)
{
hammer2_chain_t *chain;
hammer2_inode_ref(ip);
again:
ccms_thread_lock(&ip->topo_cst, CCMS_STATE_SHARED);
chain = ip->chain;
KKASSERT(chain != NULL); /* for now */
hammer2_chain_lock(chain, HAMMER2_RESOLVE_ALWAYS |
HAMMER2_RESOLVE_SHARED);
/*
* Resolve duplication races
*/
if (hammer2_chain_refactor_test(chain, 1)) {
hammer2_chain_unlock(chain);
ccms_thread_unlock(&ip->topo_cst);
chain = hammer2_inode_lock_ex(ip);
hammer2_inode_unlock_ex(ip, chain);
goto again;
}
return (chain);
}
static int
hammer2_ioctl_pfs_snapshot(hammer2_inode_t *ip, void *data)
{
hammer2_ioc_pfs_t *pfs = data;
hammer2_dev_t *hmp;
hammer2_chain_t *chain;
hammer2_tid_t mtid;
int error;
if (pfs->name[0] == 0)
return(EINVAL);
if (pfs->name[sizeof(pfs->name)-1] != 0)
return(EINVAL);
hmp = ip->pmp->pfs_hmps[0];
if (hmp == NULL)
return (EINVAL);
hammer2_vfs_sync(ip->pmp->mp, MNT_WAIT);
hammer2_trans_init(ip->pmp, HAMMER2_TRANS_ISFLUSH);
mtid = hammer2_trans_sub(ip->pmp);
hammer2_inode_lock(ip, 0);
chain = hammer2_inode_chain(ip, 0, HAMMER2_RESOLVE_ALWAYS);
error = hammer2_chain_snapshot(chain, pfs, mtid);
hammer2_chain_unlock(chain);
hammer2_chain_drop(chain);
hammer2_inode_unlock(ip);
hammer2_trans_done(ip->pmp);
return (error);
}
/*
* NOTE: We don't combine the inode/chain lock because putting away an
* inode would otherwise confuse multiple lock holders of the inode.
*
* Shared locks are especially sensitive to having too many shared
* lock counts (from the same thread) on certain paths which might
* need to upgrade them. Only one count of a shared lock can be
* upgraded.
*/
hammer2_chain_t *
hammer2_inode_lock_sh(hammer2_inode_t *ip)
{
hammer2_chain_t *chain;
hammer2_inode_ref(ip);
for (;;) {
ccms_thread_lock(&ip->topo_cst, CCMS_STATE_SHARED);
chain = ip->chain;
KKASSERT(chain != NULL); /* for now */
hammer2_chain_lock(chain, HAMMER2_RESOLVE_ALWAYS |
HAMMER2_RESOLVE_SHARED);
/*
* Resolve duplication races, resolve hardlinks by giving
* up and cycling an exclusive lock.
*/
if ((chain->flags & HAMMER2_CHAIN_DUPLICATED) == 0 &&
chain->data->ipdata.type != HAMMER2_OBJTYPE_HARDLINK) {
break;
}
hammer2_chain_unlock(chain);
ccms_thread_unlock(&ip->topo_cst);
chain = hammer2_inode_lock_ex(ip);
hammer2_inode_unlock_ex(ip, chain);
}
return (chain);
}
/*
* The caller presents a locked *chainp pointing to a HAMMER2_BREF_TYPE_INODE
* with an obj_type of HAMMER2_OBJTYPE_HARDLINK. This routine will gobble
* the *chainp and return a new locked *chainp representing the file target
* (the original *chainp will be unlocked).
*
* When a match is found the chain representing the original HARDLINK
* will be returned in *ochainp with a ref, but not locked.
*
* When no match is found *chainp is set to NULL and EIO is returned.
* (*ochainp) will still be set to the original chain with a ref but not
* locked.
*/
int
hammer2_hardlink_find(hammer2_inode_t *dip, hammer2_chain_t **chainp,
hammer2_chain_t **ochainp)
{
hammer2_chain_t *chain = *chainp;
hammer2_chain_t *parent;
hammer2_inode_t *ip;
hammer2_inode_t *pip;
hammer2_key_t key_dummy;
hammer2_key_t lhc;
int cache_index = -1;
pip = dip;
hammer2_inode_ref(pip); /* for loop */
hammer2_chain_ref(chain); /* for (*ochainp) */
*ochainp = chain;
/*
* Locate the hardlink. pip is referenced and not locked,
* ipp.
*
* chain is reused.
*/
lhc = chain->data->ipdata.inum;
hammer2_chain_unlock(chain);
chain = NULL;
while ((ip = pip) != NULL) {
parent = hammer2_inode_lock_ex(ip);
hammer2_inode_drop(ip); /* loop */
KKASSERT(parent->bref.type == HAMMER2_BREF_TYPE_INODE);
chain = hammer2_chain_lookup(&parent, &key_dummy,
lhc, lhc, &cache_index, 0);
hammer2_chain_lookup_done(parent); /* discard parent */
if (chain)
break;
pip = ip->pip; /* safe, ip held locked */
if (pip)
hammer2_inode_ref(pip); /* loop */
hammer2_inode_unlock_ex(ip, NULL);
}
/*
* chain is locked, ip is locked. Unlock ip, return the locked
* chain. *ipp is already set w/a ref count and not locked.
*
* (parent is already unlocked).
*/
if (ip)
hammer2_inode_unlock_ex(ip, NULL);
*chainp = chain;
if (chain) {
KKASSERT(chain->bref.type == HAMMER2_BREF_TYPE_INODE);
/* already locked */
return (0);
} else {
return (EIO);
}
}
void
hammer2_inode_unlock_sh(hammer2_inode_t *ip, hammer2_chain_t *chain)
{
if (chain)
hammer2_chain_unlock(chain);
ccms_thread_unlock(&ip->topo_cst);
hammer2_inode_drop(ip);
}
/*
* Locate next match or overlap under parent, replace cluster
*/
hammer2_cluster_t *
hammer2_cluster_next(hammer2_cluster_t *cparent, hammer2_cluster_t *cluster,
hammer2_key_t *key_nextp,
hammer2_key_t key_beg, hammer2_key_t key_end, int flags)
{
hammer2_chain_t *chain;
hammer2_key_t key_accum;
hammer2_key_t key_next;
int null_count;
int i;
key_accum = *key_nextp;
null_count = 0;
cluster->focus = NULL;
cparent->focus = NULL;
for (i = 0; i < cparent->nchains; ++i) {
key_next = *key_nextp;
chain = cluster->array[i];
if (chain == NULL) {
if (cparent->focus == NULL)
cparent->focus = cparent->array[i];
++null_count;
continue;
}
if (cparent->array[i] == NULL) {
if (flags & HAMMER2_LOOKUP_NOLOCK)
hammer2_chain_drop(chain);
else
hammer2_chain_unlock(chain);
++null_count;
continue;
}
chain = hammer2_chain_next(&cparent->array[i], chain,
&key_next, key_beg, key_end,
&cparent->cache_index[i], flags);
if (cparent->focus == NULL)
cparent->focus = cparent->array[i];
cluster->array[i] = chain;
if (chain == NULL) {
++null_count;
} else if (cluster->focus == NULL) {
cluster->focus = chain;
}
if (key_accum > key_next)
key_accum = key_next;
}
if (null_count == i) {
hammer2_cluster_drop(cluster);
cluster = NULL;
}
return(cluster);
}
void
hammer2_inode_unlock_ex(hammer2_inode_t *ip, hammer2_chain_t *chain)
{
/*
* XXX this will catch parent directories too which we don't
* really want.
*/
if (chain)
hammer2_chain_unlock(chain);
ccms_thread_unlock(&ip->topo_cst);
hammer2_inode_drop(ip);
}
void
hammer2_inode_unlock_ex(hammer2_inode_t *ip, hammer2_chain_t *chain)
{
/*
* XXX this will catch parent directories too which we don't
* really want.
*/
if (ip->chain && (ip->chain->flags & (HAMMER2_CHAIN_MODIFIED |
HAMMER2_CHAIN_SUBMODIFIED))) {
atomic_set_int(&ip->flags, HAMMER2_INODE_MODIFIED);
}
if (chain)
hammer2_chain_unlock(ip->hmp, chain);
ccms_thread_unlock(&ip->topo_cst);
hammer2_inode_drop(ip);
}
/*
* (Backend) Feed chain data through the cluster validator and back to
* the frontend. Chains are fed from multiple nodes concurrently
* and pipelined via per-node FIFOs in the XOP.
*
* No xop lock is needed because we are only manipulating fields under
* our direct control.
*
* Returns 0 on success and a hammer error code if sync is permanently
* lost. The caller retains a ref on the chain but by convention
* the lock is typically inherited by the xop (caller loses lock).
*
* Returns non-zero on error. In this situation the caller retains a
* ref on the chain but loses the lock (we unlock here).
*
* WARNING! The chain is moving between two different threads, it must
* be locked SHARED to retain its data mapping, not exclusive.
* When multiple operations are in progress at once, chains fed
* back to the frontend for collection can wind up being locked
* in different orders, only a shared lock can prevent a deadlock.
*
* Exclusive locks may only be used by a XOP backend node thread
* temporarily, with no direct or indirect dependencies (aka
* blocking/waiting) on other nodes.
*/
int
hammer2_xop_feed(hammer2_xop_head_t *xop, hammer2_chain_t *chain,
int clindex, int error)
{
hammer2_xop_fifo_t *fifo;
/*
* Multi-threaded entry into the XOP collector. We own the
* fifo->wi for our clindex.
*/
fifo = &xop->collect[clindex];
while (fifo->ri == fifo->wi - HAMMER2_XOPFIFO) {
tsleep_interlock(xop, 0);
if (hammer2_xop_active(xop) == 0) {
error = EINTR;
goto done;
}
if (fifo->ri == fifo->wi - HAMMER2_XOPFIFO) {
tsleep(xop, PINTERLOCKED, "h2feed", hz*60);
}
}
if (chain)
hammer2_chain_ref(chain);
fifo->errors[fifo->wi & HAMMER2_XOPFIFO_MASK] = error;
fifo->array[fifo->wi & HAMMER2_XOPFIFO_MASK] = chain;
cpu_sfence();
++fifo->wi;
atomic_add_int(&xop->check_counter, 1);
wakeup(&xop->check_counter); /* XXX optimize */
error = 0;
/*
* Cleanup. If an error occurred we eat the lock. If no error
* occurred the fifo inherits the lock and gains an additional ref.
*
* The caller's ref remains in both cases.
*/
done:
if (error && chain)
hammer2_chain_unlock(chain);
return error;
}
/*
* HAMMER2 inode locks
*
* HAMMER2 offers shared locks and exclusive locks on inodes.
*
* An inode's ip->chain pointer is resolved and stable while an inode is
* locked, and can be cleaned out at any time (become NULL) when an inode
* is not locked.
*
* The underlying chain is also locked and returned.
*
* NOTE: We don't combine the inode/chain lock because putting away an
* inode would otherwise confuse multiple lock holders of the inode.
*/
hammer2_chain_t *
hammer2_inode_lock_ex(hammer2_inode_t *ip)
{
hammer2_chain_t *chain;
hammer2_inode_ref(ip);
ccms_thread_lock(&ip->topo_cst, CCMS_STATE_EXCLUSIVE);
/*
* ip->chain fixup. Certain duplications used to move inodes
* into indirect blocks (for example) can cause ip->chain to
* become stale.
*/
again:
chain = ip->chain;
if (hammer2_chain_refactor_test(chain, 1)) {
spin_lock(&chain->core->cst.spin);
while (hammer2_chain_refactor_test(chain, 1))
chain = chain->next_parent;
if (ip->chain != chain) {
hammer2_chain_ref(chain);
spin_unlock(&chain->core->cst.spin);
hammer2_inode_repoint(ip, NULL, chain);
hammer2_chain_drop(chain);
} else {
spin_unlock(&chain->core->cst.spin);
}
}
KKASSERT(chain != NULL); /* for now */
hammer2_chain_lock(chain, HAMMER2_RESOLVE_ALWAYS);
/*
* Resolve duplication races
*/
if (hammer2_chain_refactor_test(chain, 1)) {
hammer2_chain_unlock(chain);
goto again;
}
return (chain);
}
/*
* Unlock and deref a cluster. The cluster is destroyed if this is the
* last ref.
*/
void
hammer2_cluster_unlock(hammer2_cluster_t *cluster)
{
hammer2_chain_t *chain;
int i;
KKASSERT(cluster->refs > 0);
for (i = 0; i < cluster->nchains; ++i) {
chain = cluster->array[i];
if (chain) {
hammer2_chain_unlock(chain);
if (cluster->refs == 1)
cluster->array[i] = NULL; /* safety */
}
}
if (atomic_fetchadd_int(&cluster->refs, -1) == 1) {
cluster->focus = NULL;
kfree(cluster, M_HAMMER2);
/* cluster = NULL; safety */
}
}
/*
* chain may have been moved around by the create.
*/
void
hammer2_chain_refactor(hammer2_chain_t **chainp)
{
hammer2_chain_t *chain = *chainp;
hammer2_chain_core_t *core;
core = chain->core;
while (chain->flags & HAMMER2_CHAIN_DUPLICATED) {
spin_lock(&core->cst.spin);
chain = TAILQ_NEXT(chain, core_entry);
while (chain->flags & HAMMER2_CHAIN_DUPLICATED)
chain = TAILQ_NEXT(chain, core_entry);
hammer2_chain_ref(chain);
spin_unlock(&core->cst.spin);
KKASSERT(chain->core == core);
hammer2_chain_unlock(*chainp);
hammer2_chain_lock(chain, HAMMER2_RESOLVE_ALWAYS |
HAMMER2_RESOLVE_NOREF); /* eat ref */
*chainp = chain;
}
}
/*
* When presented with a (*chainp) representing an inode of type
* OBJTYPE_HARDLINK this code will save the original inode (with a ref)
* in (*ipp), and then locate the hidden hardlink target in (dip) or
* any parent directory above (dip). The locked (*chainp) is replaced
* with a new locked (*chainp) representing the hardlink target.
*/
int
hammer2_hardlink_find(hammer2_inode_t *dip, hammer2_chain_t **chainp,
hammer2_inode_t **ipp)
{
hammer2_mount_t *hmp = dip->hmp;
hammer2_chain_t *chain = *chainp;
hammer2_chain_t *parent;
hammer2_inode_t *pip;
hammer2_key_t lhc;
*ipp = chain->u.ip;
hammer2_inode_ref(chain->u.ip);
lhc = chain->u.ip->ip_data.inum;
hammer2_inode_unlock_ex(chain->u.ip);
pip = chain->u.ip->pip;
chain = NULL;
while (pip) {
parent = &pip->chain;
KKASSERT(parent->bref.type == HAMMER2_BREF_TYPE_INODE);
hammer2_chain_lock(hmp, parent, HAMMER2_RESOLVE_ALWAYS);
chain = hammer2_chain_lookup(hmp, &parent, lhc, lhc, 0);
hammer2_chain_unlock(hmp, parent);
if (chain)
break;
pip = pip->pip;
}
*chainp = chain;
if (chain) {
KKASSERT(chain->bref.type == HAMMER2_BREF_TYPE_INODE);
/* already locked */
return (0);
} else {
return (EIO);
}
}
/*
* XXX initial NULL cluster needs reworking (pass **clusterp ?)
*
* The raw scan function is similar to lookup/next but does not seek to a key.
* Blockrefs are iterated via first_chain = (parent, NULL) and
* next_chain = (parent, chain).
*
* The passed-in parent must be locked and its data resolved. The returned
* chain will be locked. Pass chain == NULL to acquire the first sub-chain
* under parent and then iterate with the passed-in chain (which this
* function will unlock).
*/
hammer2_cluster_t *
hammer2_cluster_scan(hammer2_cluster_t *cparent, hammer2_cluster_t *cluster,
int flags)
{
hammer2_chain_t *chain;
int null_count;
int i;
null_count = 0;
for (i = 0; i < cparent->nchains; ++i) {
chain = cluster->array[i];
if (chain == NULL) {
++null_count;
continue;
}
if (cparent->array[i] == NULL) {
if (flags & HAMMER2_LOOKUP_NOLOCK)
hammer2_chain_drop(chain);
else
hammer2_chain_unlock(chain);
++null_count;
continue;
}
chain = hammer2_chain_scan(cparent->array[i], chain,
&cparent->cache_index[i], flags);
cluster->array[i] = chain;
if (chain == NULL)
++null_count;
}
if (null_count == i) {
hammer2_cluster_drop(cluster);
cluster = NULL;
}
return(cluster);
}
/*
* Lock and ref a cluster. This adds a ref to the cluster and its chains
* and then locks them.
*/
int
hammer2_cluster_lock(hammer2_cluster_t *cluster, int how)
{
hammer2_chain_t *chain;
int i;
int error;
error = 0;
atomic_add_int(&cluster->refs, 1);
for (i = 0; i < cluster->nchains; ++i) {
chain = cluster->array[i];
if (chain) {
error = hammer2_chain_lock(chain, how);
if (error) {
while (--i >= 0)
hammer2_chain_unlock(cluster->array[i]);
atomic_add_int(&cluster->refs, -1);
break;
}
}
}
return error;
}
/*
* HAMMER2 inode locks
*
* HAMMER2 offers shared locks and exclusive locks on inodes.
*
* An inode's ip->chain pointer is resolved and stable while an inode is
* locked, and can be cleaned out at any time (become NULL) when an inode
* is not locked.
*
* This function handles duplication races and hardlink replacement races
* which can cause ip's cached chain to become stale.
*
* The underlying chain is also locked and returned.
*
* NOTE: We don't combine the inode/chain lock because putting away an
* inode would otherwise confuse multiple lock holders of the inode.
*/
hammer2_chain_t *
hammer2_inode_lock_ex(hammer2_inode_t *ip)
{
hammer2_chain_t *chain;
hammer2_chain_t *ochain;
hammer2_chain_core_t *core;
int error;
hammer2_inode_ref(ip);
ccms_thread_lock(&ip->topo_cst, CCMS_STATE_EXCLUSIVE);
chain = ip->chain;
core = chain->core;
for (;;) {
if (chain->flags & HAMMER2_CHAIN_DUPLICATED) {
spin_lock(&core->cst.spin);
while (chain->flags & HAMMER2_CHAIN_DUPLICATED)
chain = TAILQ_NEXT(chain, core_entry);
hammer2_chain_ref(chain);
spin_unlock(&core->cst.spin);
hammer2_inode_repoint(ip, NULL, chain);
hammer2_chain_drop(chain);
}
hammer2_chain_lock(chain, HAMMER2_RESOLVE_ALWAYS);
if ((chain->flags & HAMMER2_CHAIN_DUPLICATED) == 0)
break;
hammer2_chain_unlock(chain);
}
if (chain->data->ipdata.type == HAMMER2_OBJTYPE_HARDLINK &&
(chain->flags & HAMMER2_CHAIN_DELETED) == 0) {
error = hammer2_hardlink_find(ip->pip, &chain, &ochain);
hammer2_chain_drop(ochain);
KKASSERT(error == 0);
/* XXX error handling */
}
return (chain);
}
/*
* chain may have been moved around by the create.
*/
static
void
hammer2_chain_refactor(hammer2_chain_t **chainp)
{
hammer2_chain_t *chain = *chainp;
hammer2_chain_core_t *core;
core = chain->core;
spin_lock(&core->cst.spin);
while (hammer2_chain_refactor_test(chain, 1)) {
chain = chain->next_parent;
while (hammer2_chain_refactor_test(chain, 1))
chain = chain->next_parent;
hammer2_chain_ref(chain);
spin_unlock(&core->cst.spin);
hammer2_chain_unlock(*chainp);
hammer2_chain_lock(chain, HAMMER2_RESOLVE_ALWAYS |
HAMMER2_RESOLVE_NOREF); /* eat ref */
*chainp = chain;
spin_lock(&core->cst.spin);
}
spin_unlock(&core->cst.spin);
}
/*
* Given an exclusively locked inode and chain we consolidate its chain
* for hardlink creation, adding (nlinks) to the file's link count and
* potentially relocating the inode to a directory common to ip->pip and tdip.
*
* Replaces (*chainp) if consolidation occurred, unlocking the old chain
* and returning a new locked chain.
*
* NOTE! This function will also replace ip->chain.
*/
int
hammer2_hardlink_consolidate(hammer2_trans_t *trans,
hammer2_inode_t *ip, hammer2_chain_t **chainp,
hammer2_inode_t *cdip, hammer2_chain_t **cdchainp,
int nlinks)
{
hammer2_inode_data_t *ipdata;
hammer2_chain_t *chain;
hammer2_chain_t *nchain;
int error;
chain = *chainp;
if (nlinks == 0 && /* no hardlink needed */
(chain->data->ipdata.name_key & HAMMER2_DIRHASH_VISIBLE)) {
return (0);
}
if (hammer2_hardlink_enable < 0) { /* fake hardlinks */
return (0);
}
if (hammer2_hardlink_enable == 0) { /* disallow hardlinks */
hammer2_chain_unlock(chain);
*chainp = NULL;
return (ENOTSUP);
}
/*
* If no change in the hardlink's target directory is required and
* this is already a hardlink target, all we need to do is adjust
* the link count.
*/
if (cdip == ip->pip &&
(chain->data->ipdata.name_key & HAMMER2_DIRHASH_VISIBLE) == 0) {
if (nlinks) {
hammer2_chain_modify(trans, &chain, 0);
chain->data->ipdata.nlinks += nlinks;
}
error = 0;
goto done;
}
/*
* chain is the real inode. If it's visible we have to convert it
* to a hardlink pointer. If it is not visible then it is already
* a hardlink target and only needs to be deleted.
*/
KKASSERT((chain->flags & HAMMER2_CHAIN_DELETED) == 0);
KKASSERT(chain->data->ipdata.type != HAMMER2_OBJTYPE_HARDLINK);
if (chain->data->ipdata.name_key & HAMMER2_DIRHASH_VISIBLE) {
/*
* We are going to duplicate chain later, causing its
* media block to be shifted to the duplicate. Even though
* we are delete-duplicating nchain here it might decide not
* to reallocate the block. Set FORCECOW to force it to.
*/
nchain = chain;
hammer2_chain_lock(nchain, HAMMER2_RESOLVE_ALWAYS);
atomic_set_int(&nchain->flags, HAMMER2_CHAIN_FORCECOW);
hammer2_chain_delete_duplicate(trans, &nchain,
HAMMER2_DELDUP_RECORE);
KKASSERT((chain->flags & HAMMER2_CHAIN_DUPLICATED) == 0);
ipdata = &nchain->data->ipdata;
ipdata->target_type = ipdata->type;
ipdata->type = HAMMER2_OBJTYPE_HARDLINK;
ipdata->uflags = 0;
ipdata->rmajor = 0;
ipdata->rminor = 0;
ipdata->ctime = 0;
ipdata->mtime = 0;
ipdata->atime = 0;
ipdata->btime = 0;
bzero(&ipdata->uid, sizeof(ipdata->uid));
bzero(&ipdata->gid, sizeof(ipdata->gid));
ipdata->op_flags = HAMMER2_OPFLAG_DIRECTDATA;
ipdata->cap_flags = 0;
ipdata->mode = 0;
ipdata->size = 0;
ipdata->nlinks = 1;
ipdata->iparent = 0; /* XXX */
ipdata->pfs_type = 0;
ipdata->pfs_inum = 0;
bzero(&ipdata->pfs_clid, sizeof(ipdata->pfs_clid));
bzero(&ipdata->pfs_fsid, sizeof(ipdata->pfs_fsid));
ipdata->data_quota = 0;
ipdata->data_count = 0;
ipdata->inode_quota = 0;
ipdata->inode_count = 0;
ipdata->attr_tid = 0;
ipdata->dirent_tid = 0;
bzero(&ipdata->u, sizeof(ipdata->u));
/* XXX transaction ids */
} else {
hammer2_chain_delete(trans, chain, 0);
nchain = NULL;
}
/*
* chain represents the hardlink target and is now flagged deleted.
* duplicate it to the parent directory and adjust nlinks.
*
* WARNING! The shiftup() call can cause nchain to be moved into
* an indirect block, and our nchain will wind up pointing
* to the older/original version.
*/
KKASSERT(chain->flags & HAMMER2_CHAIN_DELETED);
hammer2_hardlink_shiftup(trans, &chain, cdip, cdchainp, nlinks, &error);
if (error == 0)
hammer2_inode_repoint(ip, cdip, chain);
/*
* Unlock the original chain last as the lock blocked races against
* the creation of the new hardlink target.
*/
if (nchain)
hammer2_chain_unlock(nchain);
done:
/*
* Cleanup, chain/nchain already dealt with.
*/
*chainp = chain;
hammer2_inode_drop(cdip);
return (error);
}
/*
* Unlink the file from the specified directory inode. The directory inode
* does not need to be locked.
*
* isdir determines whether a directory/non-directory check should be made.
* No check is made if isdir is set to -1.
*
* isopen specifies whether special unlink-with-open-descriptor handling
* must be performed. If set to -1 the caller is deleting a PFS and we
* check whether the chain is mounted or not (chain->pmp != NULL). 1 is
* implied if it is mounted.
*
* If isopen is 1 and nlinks drops to 0 this function must move the chain
* to a special hidden directory until last-close occurs on the file.
*
* NOTE! The underlying file can still be active with open descriptors
* or if the chain is being manually held (e.g. for rename).
*
* The caller is responsible for fixing up ip->chain if e.g. a
* rename occurs (see chain_duplicate()).
*/
int
hammer2_unlink_file(hammer2_trans_t *trans, hammer2_inode_t *dip,
const uint8_t *name, size_t name_len,
int isdir, int *hlinkp, struct nchandle *nch)
{
hammer2_inode_data_t *ipdata;
hammer2_chain_t *parent;
hammer2_chain_t *ochain;
hammer2_chain_t *chain;
hammer2_chain_t *dparent;
hammer2_chain_t *dchain;
hammer2_key_t key_dummy;
hammer2_key_t key_next;
hammer2_key_t lhc;
int error;
int cache_index = -1;
uint8_t type;
error = 0;
ochain = NULL;
lhc = hammer2_dirhash(name, name_len);
/*
* Search for the filename in the directory
*/
if (hlinkp)
*hlinkp = 0;
parent = hammer2_inode_lock_ex(dip);
chain = hammer2_chain_lookup(&parent, &key_next,
lhc, lhc + HAMMER2_DIRHASH_LOMASK,
&cache_index, 0);
while (chain) {
if (chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
name_len == chain->data->ipdata.name_len &&
bcmp(name, chain->data->ipdata.filename, name_len) == 0) {
break;
}
chain = hammer2_chain_next(&parent, chain, &key_next,
key_next,
lhc + HAMMER2_DIRHASH_LOMASK,
&cache_index, 0);
}
hammer2_inode_unlock_ex(dip, NULL); /* retain parent */
/*
* Not found or wrong type (isdir < 0 disables the type check).
* If a hardlink pointer, type checks use the hardlink target.
*/
if (chain == NULL) {
error = ENOENT;
goto done;
}
if ((type = chain->data->ipdata.type) == HAMMER2_OBJTYPE_HARDLINK) {
if (hlinkp)
*hlinkp = 1;
type = chain->data->ipdata.target_type;
}
if (type == HAMMER2_OBJTYPE_DIRECTORY && isdir == 0) {
error = ENOTDIR;
goto done;
}
if (type != HAMMER2_OBJTYPE_DIRECTORY && isdir >= 1) {
error = EISDIR;
goto done;
}
/*
* Hardlink must be resolved. We can't hold the parent locked
* while we do this or we could deadlock.
*
* On success chain will be adjusted to point at the hardlink target
* and ochain will point to the hardlink pointer in the original
* directory. Otherwise chain remains pointing to the original.
*/
if (chain->data->ipdata.type == HAMMER2_OBJTYPE_HARDLINK) {
hammer2_chain_unlock(parent);
parent = NULL;
error = hammer2_hardlink_find(dip, &chain, &ochain);
}
/*
* If this is a directory the directory must be empty. However, if
* isdir < 0 we are doing a rename and the directory does not have
* to be empty, and if isdir > 1 we are deleting a PFS/snapshot
* and the directory does not have to be empty.
*
* NOTE: We check the full key range here which covers both visible
* and invisible entries. Theoretically there should be no
* invisible (hardlink target) entries if there are no visible
* entries.
*/
if (type == HAMMER2_OBJTYPE_DIRECTORY && isdir == 1) {
dparent = hammer2_chain_lookup_init(chain, 0);
dchain = hammer2_chain_lookup(&dparent, &key_dummy,
0, (hammer2_key_t)-1,
&cache_index,
HAMMER2_LOOKUP_NODATA);
if (dchain) {
hammer2_chain_unlock(dchain);
hammer2_chain_lookup_done(dparent);
error = ENOTEMPTY;
goto done;
}
hammer2_chain_lookup_done(dparent);
dparent = NULL;
/* dchain NULL */
}
/*
* Ok, we can now unlink the chain. We always decrement nlinks even
* if the entry can be deleted in case someone has the file open and
* does an fstat().
*
* The chain itself will no longer be in the on-media topology but
* can still be flushed to the media (e.g. if an open descriptor
* remains). When the last vnode/ip ref goes away the chain will
* be marked unmodified, avoiding any further (now unnecesary) I/O.
*
* A non-NULL ochain indicates a hardlink.
*/
if (ochain) {
/*
* Delete the original hardlink pointer unconditionally.
* (any open descriptors will migrate to the hardlink
* target and have no affect on this operation).
*
* NOTE: parent from above is NULL when ochain != NULL
* so we can reuse it.
*/
hammer2_chain_lock(ochain, HAMMER2_RESOLVE_ALWAYS);
hammer2_chain_delete(trans, ochain, 0);
hammer2_chain_unlock(ochain);
}
/*
* Decrement nlinks on the hardlink target (or original file if
* there it was not hardlinked). Delete the target when nlinks
* reaches 0 with special handling if (isopen) is set.
*
* NOTE! In DragonFly the vnops function calls cache_unlink() after
* calling us here to clean out the namecache association,
* (which does not represent a ref for the open-test), and to
* force finalization of the vnode if/when the last ref gets
* dropped.
*
* NOTE! Files are unlinked by rename and then relinked. nch will be
* passed as NULL in this situation. hammer2_inode_connect()
* will bump nlinks.
*/
KKASSERT(chain != NULL);
hammer2_chain_modify(trans, &chain, 0);
ipdata = &chain->data->ipdata;
--ipdata->nlinks;
if ((int64_t)ipdata->nlinks < 0) /* XXX debugging */
ipdata->nlinks = 0;
if (ipdata->nlinks == 0) {
if ((chain->flags & HAMMER2_CHAIN_PFSROOT) && chain->pmp) {
error = EINVAL;
kprintf("hammer2: PFS \"%s\" cannot be deleted "
"while still mounted\n",
ipdata->filename);
goto done;
}
if (nch && cache_isopen(nch)) {
kprintf("WARNING: unlinking open file\n");
atomic_set_int(&chain->flags, HAMMER2_CHAIN_UNLINKED);
hammer2_inode_move_to_hidden(trans, &chain,
ipdata->inum);
} else {
hammer2_chain_delete(trans, chain, 0);
}
}
error = 0;
done:
if (chain)
hammer2_chain_unlock(chain);
if (parent)
hammer2_chain_lookup_done(parent);
if (ochain)
hammer2_chain_drop(ochain);
return error;
}
/*
* Create a new PFS under the super-root
*/
static int
hammer2_ioctl_pfs_create(hammer2_inode_t *ip, void *data)
{
hammer2_inode_data_t *nipdata;
hammer2_chain_t *nchain;
hammer2_dev_t *hmp;
hammer2_ioc_pfs_t *pfs;
hammer2_inode_t *nip;
hammer2_tid_t mtid;
int error;
hmp = ip->pmp->pfs_hmps[0];
if (hmp == NULL)
return (EINVAL);
pfs = data;
nip = NULL;
if (pfs->name[0] == 0)
return(EINVAL);
pfs->name[sizeof(pfs->name) - 1] = 0; /* ensure 0-termination */
if (hammer2_ioctl_pfs_lookup(ip, pfs) == 0)
return(EEXIST);
hammer2_trans_init(hmp->spmp, 0);
mtid = hammer2_trans_sub(hmp->spmp);
nip = hammer2_inode_create(hmp->spmp->iroot, NULL, NULL,
pfs->name, strlen(pfs->name), 0,
1, HAMMER2_OBJTYPE_DIRECTORY, 0,
HAMMER2_INSERT_PFSROOT, &error);
if (error == 0) {
hammer2_inode_modify(nip);
nchain = hammer2_inode_chain(nip, 0, HAMMER2_RESOLVE_ALWAYS);
hammer2_chain_modify(nchain, mtid, 0);
nipdata = &nchain->data->ipdata;
nip->meta.pfs_type = pfs->pfs_type;
nip->meta.pfs_subtype = pfs->pfs_subtype;
nip->meta.pfs_clid = pfs->pfs_clid;
nip->meta.pfs_fsid = pfs->pfs_fsid;
nip->meta.op_flags |= HAMMER2_OPFLAG_PFSROOT;
/*
* Set default compression and check algorithm. This
* can be changed later.
*
* Do not allow compression on PFS's with the special name
* "boot", the boot loader can't decompress (yet).
*/
nip->meta.comp_algo =
HAMMER2_ENC_ALGO(HAMMER2_COMP_NEWFS_DEFAULT);
nip->meta.check_algo =
HAMMER2_ENC_ALGO( HAMMER2_CHECK_ISCSI32);
if (strcasecmp(pfs->name, "boot") == 0) {
nip->meta.comp_algo =
HAMMER2_ENC_ALGO(HAMMER2_COMP_AUTOZERO);
}
#if 0
hammer2_blockref_t bref;
/* XXX new PFS needs to be rescanned / added */
bref = nchain->bref;
kprintf("ADD LOCAL PFS (IOCTL): %s\n", nipdata->filename);
hammer2_pfsalloc(nchain, nipdata, bref.modify_tid);
#endif
/* XXX rescan */
hammer2_chain_unlock(nchain);
hammer2_chain_drop(nchain);
/*
* Super-root isn't mounted, fsync it
*/
hammer2_inode_ref(nip);
hammer2_inode_unlock(nip);
hammer2_inode_fsync(nip);
hammer2_inode_drop(nip);
}
hammer2_trans_done(hmp->spmp);
return (error);
}
/*
* Synchronize the inode's frontend state with the chain state prior
* to any explicit flush of the inode or any strategy write call.
*
* Called with a locked inode.
*/
void
hammer2_inode_fsync(hammer2_trans_t *trans, hammer2_inode_t *ip,
hammer2_chain_t **chainp)
{
hammer2_inode_data_t *ipdata;
hammer2_chain_t *parent;
hammer2_chain_t *chain;
hammer2_key_t lbase;
hammer2_key_t key_next;
int cache_index;
ipdata = &ip->chain->data->ipdata;
if (ip->flags & HAMMER2_INODE_MTIME) {
ipdata = hammer2_chain_modify_ip(trans, ip, chainp, 0);
atomic_clear_int(&ip->flags, HAMMER2_INODE_MTIME);
ipdata->mtime = ip->mtime;
}
if ((ip->flags & HAMMER2_INODE_RESIZED) && ip->size < ipdata->size) {
ipdata = hammer2_chain_modify_ip(trans, ip, chainp, 0);
ipdata->size = ip->size;
atomic_clear_int(&ip->flags, HAMMER2_INODE_RESIZED);
/*
* We must delete any chains beyond the EOF. The chain
* straddling the EOF will be pending in the bioq.
*/
lbase = (ipdata->size + HAMMER2_PBUFMASK64) &
~HAMMER2_PBUFMASK64;
parent = hammer2_chain_lookup_init(ip->chain, 0);
chain = hammer2_chain_lookup(&parent, &key_next,
lbase, (hammer2_key_t)-1,
&cache_index,
HAMMER2_LOOKUP_NODATA);
while (chain) {
/*
* Degenerate embedded case, nothing to loop on
*/
if (chain->bref.type == HAMMER2_BREF_TYPE_INODE) {
hammer2_chain_unlock(chain);
break;
}
if (chain->bref.type == HAMMER2_BREF_TYPE_DATA) {
hammer2_chain_delete(trans, chain, 0);
}
chain = hammer2_chain_next(&parent, chain, &key_next,
key_next, (hammer2_key_t)-1,
&cache_index,
HAMMER2_LOOKUP_NODATA);
}
hammer2_chain_lookup_done(parent);
} else
if ((ip->flags & HAMMER2_INODE_RESIZED) && ip->size > ipdata->size) {
ipdata = hammer2_chain_modify_ip(trans, ip, chainp, 0);
ipdata->size = ip->size;
atomic_clear_int(&ip->flags, HAMMER2_INODE_RESIZED);
/*
* When resizing larger we may not have any direct-data
* available.
*/
if ((ipdata->op_flags & HAMMER2_OPFLAG_DIRECTDATA) &&
ip->size > HAMMER2_EMBEDDED_BYTES) {
ipdata->op_flags &= ~HAMMER2_OPFLAG_DIRECTDATA;
bzero(&ipdata->u.blockset, sizeof(ipdata->u.blockset));
}
}
}
/*
* Create a new inode in the specified directory using the vattr to
* figure out the type of inode.
*
* If no error occurs the new inode with its chain locked is returned in
* *nipp, otherwise an error is returned and *nipp is set to NULL.
*
* If vap and/or cred are NULL the related fields are not set and the
* inode type defaults to a directory. This is used when creating PFSs
* under the super-root, so the inode number is set to 1 in this case.
*
* dip is not locked on entry.
*/
hammer2_inode_t *
hammer2_inode_create(hammer2_trans_t *trans, hammer2_inode_t *dip,
struct vattr *vap, struct ucred *cred,
const uint8_t *name, size_t name_len,
hammer2_chain_t **chainp, int *errorp)
{
hammer2_inode_data_t *dipdata;
hammer2_inode_data_t *nipdata;
hammer2_chain_t *chain;
hammer2_chain_t *parent;
hammer2_inode_t *nip;
hammer2_key_t key_dummy;
hammer2_key_t lhc;
int error;
uid_t xuid;
uuid_t dip_uid;
uuid_t dip_gid;
uint32_t dip_mode;
uint8_t dip_algo;
int cache_index = -1;
lhc = hammer2_dirhash(name, name_len);
*errorp = 0;
/*
* Locate the inode or indirect block to create the new
* entry in. At the same time check for key collisions
* and iterate until we don't get one.
*
* NOTE: hidden inodes do not have iterators.
*/
retry:
parent = hammer2_inode_lock_ex(dip);
dipdata = &dip->chain->data->ipdata;
dip_uid = dipdata->uid;
dip_gid = dipdata->gid;
dip_mode = dipdata->mode;
dip_algo = dipdata->comp_algo;
error = 0;
while (error == 0) {
chain = hammer2_chain_lookup(&parent, &key_dummy,
lhc, lhc, &cache_index, 0);
if (chain == NULL)
break;
if ((lhc & HAMMER2_DIRHASH_VISIBLE) == 0)
error = ENOSPC;
if ((lhc & HAMMER2_DIRHASH_LOMASK) == HAMMER2_DIRHASH_LOMASK)
error = ENOSPC;
hammer2_chain_unlock(chain);
chain = NULL;
++lhc;
}
if (error == 0) {
error = hammer2_chain_create(trans, &parent, &chain,
lhc, 0,
HAMMER2_BREF_TYPE_INODE,
HAMMER2_INODE_BYTES);
}
/*
* Cleanup and handle retries.
*/
if (error == EAGAIN) {
hammer2_chain_ref(parent);
hammer2_inode_unlock_ex(dip, parent);
hammer2_chain_wait(parent);
hammer2_chain_drop(parent);
goto retry;
}
hammer2_inode_unlock_ex(dip, parent);
if (error) {
KKASSERT(chain == NULL);
*errorp = error;
return (NULL);
}
/*
* Set up the new inode.
*
* NOTE: *_get() integrates chain's lock into the inode lock.
*
* NOTE: Only one new inode can currently be created per
* transaction. If the need arises we can adjust
* hammer2_trans_init() to allow more.
*
* NOTE: nipdata will have chain's blockset data.
*/
chain->data->ipdata.inum = trans->inode_tid;
nip = hammer2_inode_get(dip->pmp, dip, chain);
nipdata = &chain->data->ipdata;
if (vap) {
KKASSERT(trans->inodes_created == 0);
nipdata->type = hammer2_get_obj_type(vap->va_type);
nipdata->inum = trans->inode_tid;
++trans->inodes_created;
switch (nipdata->type) {
case HAMMER2_OBJTYPE_CDEV:
case HAMMER2_OBJTYPE_BDEV:
nipdata->rmajor = vap->va_rmajor;
nipdata->rminor = vap->va_rminor;
break;
default:
break;
}
} else {
nipdata->type = HAMMER2_OBJTYPE_DIRECTORY;
nipdata->inum = 1;
}
/* Inherit parent's inode compression mode. */
nip->comp_heuristic = 0;
nipdata->comp_algo = dip_algo;
nipdata->version = HAMMER2_INODE_VERSION_ONE;
hammer2_update_time(&nipdata->ctime);
nipdata->mtime = nipdata->ctime;
if (vap)
nipdata->mode = vap->va_mode;
nipdata->nlinks = 1;
if (vap) {
if (dip && dip->pmp) {
xuid = hammer2_to_unix_xid(&dip_uid);
xuid = vop_helper_create_uid(dip->pmp->mp,
dip_mode,
xuid,
cred,
&vap->va_mode);
} else {
/* super-root has no dip and/or pmp */
xuid = 0;
}
if (vap->va_vaflags & VA_UID_UUID_VALID)
nipdata->uid = vap->va_uid_uuid;
else if (vap->va_uid != (uid_t)VNOVAL)
hammer2_guid_to_uuid(&nipdata->uid, vap->va_uid);
else
hammer2_guid_to_uuid(&nipdata->uid, xuid);
if (vap->va_vaflags & VA_GID_UUID_VALID)
nipdata->gid = vap->va_gid_uuid;
else if (vap->va_gid != (gid_t)VNOVAL)
hammer2_guid_to_uuid(&nipdata->gid, vap->va_gid);
else if (dip)
nipdata->gid = dip_gid;
}
/*
* Regular files and softlinks allow a small amount of data to be
* directly embedded in the inode. This flag will be cleared if
* the size is extended past the embedded limit.
*/
if (nipdata->type == HAMMER2_OBJTYPE_REGFILE ||
nipdata->type == HAMMER2_OBJTYPE_SOFTLINK) {
nipdata->op_flags |= HAMMER2_OPFLAG_DIRECTDATA;
}
KKASSERT(name_len < HAMMER2_INODE_MAXNAME);
bcopy(name, nipdata->filename, name_len);
nipdata->name_key = lhc;
nipdata->name_len = name_len;
*chainp = chain;
return (nip);
}
/*
* cparent is locked exclusively, with an extra ref, cluster is not locked.
* Replace element [i] in the cluster.
*/
static
int
hammer2_sync_replace(hammer2_thread_t *thr,
hammer2_chain_t *parent, hammer2_chain_t *chain,
hammer2_tid_t mtid, int idx,
hammer2_chain_t *focus)
{
int nradix;
uint8_t otype;
#if HAMMER2_THREAD_DEBUG
if (hammer2_debug & 1)
kprintf("replace rec %p slave %d %d.%016jx mod=%016jx\n",
chain,
idx,
focus->bref.type, focus->bref.key, mtid);
#endif
hammer2_chain_unlock(chain);
hammer2_chain_lock(chain, HAMMER2_RESOLVE_ALWAYS);
if (chain->bytes != focus->bytes) {
/* XXX what if compressed? */
nradix = hammer2_getradix(chain->bytes);
hammer2_chain_resize(NULL, parent, chain,
mtid, nradix, 0);
}
hammer2_chain_modify(chain, mtid, 0);
otype = chain->bref.type;
chain->bref.type = focus->bref.type;
chain->bref.methods = focus->bref.methods;
chain->bref.keybits = focus->bref.keybits;
chain->bref.vradix = focus->bref.vradix;
/* mirror_tid updated by flush */
KKASSERT(chain->bref.modify_tid == mtid);
chain->bref.flags = focus->bref.flags;
/* key already present */
/* check code will be recalculated */
chain->error = 0;
/*
* Copy data body.
*/
switch(chain->bref.type) {
case HAMMER2_BREF_TYPE_INODE:
if ((focus->data->ipdata.meta.op_flags &
HAMMER2_OPFLAG_DIRECTDATA) == 0) {
/*
* If DIRECTDATA is transitioning to 0 or the old
* chain is not an inode we have to initialize
* the block table.
*/
if (otype != HAMMER2_BREF_TYPE_INODE ||
(chain->data->ipdata.meta.op_flags &
HAMMER2_OPFLAG_DIRECTDATA)) {
kprintf("chain inode trans away from dd\n");
bzero(&chain->data->ipdata.u,
sizeof(chain->data->ipdata.u));
}
bcopy(focus->data, chain->data,
offsetof(hammer2_inode_data_t, u));
/* XXX setcheck on inode should not be needed */
hammer2_chain_setcheck(chain, chain->data);
break;
}
/* fall through */
case HAMMER2_BREF_TYPE_DATA:
bcopy(focus->data, chain->data, chain->bytes);
hammer2_chain_setcheck(chain, chain->data);
break;
default:
KKASSERT(0);
break;
}
hammer2_chain_unlock(chain);
hammer2_chain_lock(chain, HAMMER2_RESOLVE_SHARED |
HAMMER2_RESOLVE_MAYBE);
return 0;
}
/*
* Shift *chainp up to the specified directory, change the filename
* to "0xINODENUMBER", and adjust the key. The chain becomes the
* invisible hardlink target.
*
* The original *chainp has already been marked deleted.
*/
static
void
hammer2_hardlink_shiftup(hammer2_trans_t *trans, hammer2_chain_t **chainp,
hammer2_inode_t *dip, hammer2_chain_t **dchainp,
int nlinks, int *errorp)
{
hammer2_inode_data_t *nipdata;
hammer2_chain_t *chain;
hammer2_chain_t *xchain;
hammer2_key_t key_dummy;
hammer2_key_t lhc;
hammer2_blockref_t bref;
int cache_index = -1;
chain = *chainp;
lhc = chain->data->ipdata.inum;
KKASSERT((lhc & HAMMER2_DIRHASH_VISIBLE) == 0);
/*
* Locate the inode or indirect block to create the new
* entry in. lhc represents the inode number so there is
* no collision iteration.
*
* There should be no key collisions with invisible inode keys.
*
* WARNING! Must use inode_lock_ex() on dip to handle a stale
* dip->chain cache.
*/
retry:
*errorp = 0;
xchain = hammer2_chain_lookup(dchainp, &key_dummy,
lhc, lhc, &cache_index, 0);
if (xchain) {
kprintf("X3 chain %p dip %p dchain %p dip->chain %p\n",
xchain, dip, *dchainp, dip->chain);
hammer2_chain_unlock(xchain);
xchain = NULL;
*errorp = ENOSPC;
#if 0
Debugger("X3");
#endif
}
/*
* Create entry in common parent directory using the seek position
* calculated above.
*
* We must refactor chain because it might have been shifted into
* an indirect chain by the create.
*/
if (*errorp == 0) {
KKASSERT(xchain == NULL);
#if 0
*errorp = hammer2_chain_create(trans, dchainp, &xchain,
lhc, 0,
HAMMER2_BREF_TYPE_INODE,/* n/a */
HAMMER2_INODE_BYTES); /* n/a */
#endif
/*XXX this somehow isn't working on chain XXX*/
/*KKASSERT(xxx)*/
}
/*
* Cleanup and handle retries.
*/
if (*errorp == EAGAIN) {
kprintf("R");
hammer2_chain_wait(*dchainp);
hammer2_chain_drop(*dchainp);
goto retry;
}
/*
* Handle the error case
*/
if (*errorp) {
panic("error2");
KKASSERT(xchain == NULL);
return;
}
/*
* Use xchain as a placeholder for (lhc). Duplicate chain to the
* same target bref as xchain and then delete xchain. The duplication
* occurs after xchain in flush order even though xchain is deleted
* after the duplication. XXX
*
* WARNING! Duplications (to a different parent) can cause indirect
* blocks to be inserted, refactor xchain.
*/
bref = chain->bref;
bref.key = lhc; /* invisible dir entry key */
bref.keybits = 0;
hammer2_chain_duplicate(trans, dchainp, &chain, &bref, 0, 2);
/*
* chain is now 'live' again.. adjust the filename.
*
* Directory entries are inodes but this is a hidden hardlink
* target. The name isn't used but to ease debugging give it
* a name after its inode number.
*/
hammer2_chain_modify(trans, &chain, 0);
nipdata = &chain->data->ipdata;
ksnprintf(nipdata->filename, sizeof(nipdata->filename),
"0x%016jx", (intmax_t)nipdata->inum);
nipdata->name_len = strlen(nipdata->filename);
nipdata->name_key = lhc;
nipdata->nlinks += nlinks;
*chainp = chain;
}
/*
* Used to scan and retrieve PFS information. PFS's are directories under
* the super-root.
*
* To scan PFSs pass name_key=0. The function will scan for the next
* PFS and set all fields, as well as set name_next to the next key.
* When no PFSs remain, name_next is set to (hammer2_key_t)-1.
*
* To retrieve a particular PFS by key, specify the key but note that
* the ioctl will return the lowest key >= specified_key, so the caller
* must verify the key.
*
* To retrieve the PFS associated with the file descriptor, pass
* name_key set to (hammer2_key_t)-1.
*/
static int
hammer2_ioctl_pfs_get(hammer2_inode_t *ip, void *data)
{
const hammer2_inode_data_t *ripdata;
hammer2_dev_t *hmp;
hammer2_ioc_pfs_t *pfs;
hammer2_chain_t *parent;
hammer2_chain_t *chain;
hammer2_key_t key_next;
hammer2_key_t save_key;
int cache_index = -1;
int error;
hmp = ip->pmp->pfs_hmps[0];
if (hmp == NULL)
return (EINVAL);
pfs = data;
save_key = pfs->name_key;
error = 0;
/*
* Setup
*/
if (save_key == (hammer2_key_t)-1) {
hammer2_inode_lock(ip->pmp->iroot, 0);
parent = NULL;
chain = hammer2_inode_chain(hmp->spmp->iroot, 0,
HAMMER2_RESOLVE_ALWAYS |
HAMMER2_RESOLVE_SHARED);
} else {
hammer2_inode_lock(hmp->spmp->iroot, 0);
parent = hammer2_inode_chain(hmp->spmp->iroot, 0,
HAMMER2_RESOLVE_ALWAYS |
HAMMER2_RESOLVE_SHARED);
chain = hammer2_chain_lookup(&parent, &key_next,
pfs->name_key, HAMMER2_KEY_MAX,
&cache_index,
HAMMER2_LOOKUP_SHARED);
}
/*
* Locate next PFS
*/
while (chain) {
if (chain->bref.type == HAMMER2_BREF_TYPE_INODE)
break;
if (parent == NULL) {
hammer2_chain_unlock(chain);
hammer2_chain_drop(chain);
chain = NULL;
break;
}
chain = hammer2_chain_next(&parent, chain, &key_next,
key_next, HAMMER2_KEY_MAX,
&cache_index,
HAMMER2_LOOKUP_SHARED);
}
/*
* Load the data being returned by the ioctl.
*/
if (chain) {
ripdata = &chain->data->ipdata;
pfs->name_key = ripdata->meta.name_key;
pfs->pfs_type = ripdata->meta.pfs_type;
pfs->pfs_subtype = ripdata->meta.pfs_subtype;
pfs->pfs_clid = ripdata->meta.pfs_clid;
pfs->pfs_fsid = ripdata->meta.pfs_fsid;
KKASSERT(ripdata->meta.name_len < sizeof(pfs->name));
bcopy(ripdata->filename, pfs->name, ripdata->meta.name_len);
pfs->name[ripdata->meta.name_len] = 0;
ripdata = NULL; /* safety */
/*
* Calculate name_next, if any.
*/
if (parent == NULL) {
pfs->name_next = (hammer2_key_t)-1;
} else {
chain = hammer2_chain_next(&parent, chain, &key_next,
key_next, HAMMER2_KEY_MAX,
&cache_index,
HAMMER2_LOOKUP_SHARED);
if (chain)
pfs->name_next = chain->bref.key;
else
pfs->name_next = (hammer2_key_t)-1;
}
} else {
pfs->name_next = (hammer2_key_t)-1;
error = ENOENT;
}
/*
* Cleanup
*/
if (chain) {
hammer2_chain_unlock(chain);
hammer2_chain_drop(chain);
}
if (parent) {
hammer2_chain_unlock(parent);
hammer2_chain_drop(parent);
}
if (save_key == (hammer2_key_t)-1) {
hammer2_inode_unlock(ip->pmp->iroot);
} else {
hammer2_inode_unlock(hmp->spmp->iroot);
}
return (error);
}
/*
* Connect the target inode represented by (*chainp) to the media topology
* at (dip, name, len). The caller can pass a rough *chainp, this function
* will issue lookup()s to position the parent chain properly for the
* chain insertion.
*
* If hlink is TRUE this function creates an OBJTYPE_HARDLINK directory
* entry instead of connecting (*chainp).
*
* If hlink is FALSE this function uses chain_duplicate() to make a copy
* if (*chainp) in the directory entry. (*chainp) is likely to be deleted
* by the caller in this case (e.g. rename).
*/
int
hammer2_inode_connect(hammer2_trans_t *trans,
hammer2_chain_t **chainp, int hlink,
hammer2_inode_t *dip, hammer2_chain_t **dchainp,
const uint8_t *name, size_t name_len,
hammer2_key_t lhc)
{
hammer2_inode_data_t *ipdata;
hammer2_chain_t *nchain;
hammer2_chain_t *ochain;
hammer2_key_t key_dummy;
int cache_index = -1;
int error;
/*
* Since ochain is either disconnected from the topology or represents
* a hardlink terminus which is always a parent of or equal to dip,
* we should be able to safely lock dip->chain for our setup.
*
* WARNING! Must use inode_lock_ex() on dip to handle a stale
* dip->chain cache.
*/
ochain = *chainp;
/*
* If name is non-NULL we calculate lhc, else we use the passed-in
* lhc.
*/
if (name) {
lhc = hammer2_dirhash(name, name_len);
/*
* Locate the inode or indirect block to create the new
* entry in. At the same time check for key collisions
* and iterate until we don't get one.
*/
error = 0;
while (error == 0) {
nchain = hammer2_chain_lookup(dchainp, &key_dummy,
lhc, lhc,
&cache_index, 0);
if (nchain == NULL)
break;
if ((lhc & HAMMER2_DIRHASH_LOMASK) ==
HAMMER2_DIRHASH_LOMASK) {
error = ENOSPC;
}
hammer2_chain_unlock(nchain);
nchain = NULL;
++lhc;
}
} else {
/*
* Reconnect to specific key (used when moving
* unlinked-but-open files into the hidden directory).
*/
nchain = hammer2_chain_lookup(dchainp, &key_dummy,
lhc, lhc, &cache_index, 0);
KKASSERT(nchain == NULL);
}
if (error == 0) {
if (hlink) {
/*
* Hardlink pointer needed, create totally fresh
* directory entry.
*
* We must refactor ochain because it might have
* been shifted into an indirect chain by the
* create.
*/
KKASSERT(nchain == NULL);
error = hammer2_chain_create(trans, dchainp, &nchain,
lhc, 0,
HAMMER2_BREF_TYPE_INODE,
HAMMER2_INODE_BYTES);
hammer2_chain_refactor(&ochain);
} else {
/*
* Reconnect the original chain and rename. Use
* chain_duplicate(). The caller will likely delete
* or has already deleted the original chain in
* this case.
*
* NOTE: chain_duplicate() generates a new chain
* with CHAIN_DELETED cleared (ochain typically
* has it set from the file unlink).
*
* WARNING! Can cause held-over chains to require a
* refactor. Fortunately we have none (our
* locked chains are passed into and
* modified by the call).
*/
nchain = ochain;
ochain = NULL;
hammer2_chain_duplicate(trans, NULL, &nchain, NULL,
0, 3);
error = hammer2_chain_create(trans, dchainp, &nchain,
lhc, 0,
HAMMER2_BREF_TYPE_INODE,
HAMMER2_INODE_BYTES);
}
}
/*
* Unlock stuff.
*/
KKASSERT(error != EAGAIN);
/*
* nchain should be NULL on error, leave ochain (== *chainp) alone.
*/
if (error) {
KKASSERT(nchain == NULL);
return (error);
}
/*
* Directory entries are inodes so if the name has changed we have
* to update the inode.
*
* When creating an OBJTYPE_HARDLINK entry remember to unlock the
* chain, the caller will access the hardlink via the actual hardlink
* target file and not the hardlink pointer entry, so we must still
* return ochain.
*/
if (hlink && hammer2_hardlink_enable >= 0) {
/*
* Create the HARDLINK pointer. oip represents the hardlink
* target in this situation.
*
* We will return ochain (the hardlink target).
*/
hammer2_chain_modify(trans, &nchain, 0);
KKASSERT(name_len < HAMMER2_INODE_MAXNAME);
ipdata = &nchain->data->ipdata;
bcopy(name, ipdata->filename, name_len);
ipdata->name_key = lhc;
ipdata->name_len = name_len;
ipdata->target_type = ochain->data->ipdata.type;
ipdata->type = HAMMER2_OBJTYPE_HARDLINK;
ipdata->inum = ochain->data->ipdata.inum;
ipdata->nlinks = 1;
hammer2_chain_unlock(nchain);
nchain = ochain;
ochain = NULL;
} else if (hlink && hammer2_hardlink_enable < 0) {
/*
* Create a snapshot (hardlink fake mode for debugging).
* (ochain already flushed above so we can just copy the
* bref XXX).
*
* Since this is a snapshot we return nchain in the fake
* hardlink case.
*/
hammer2_chain_modify(trans, &nchain, 0);
KKASSERT(name_len < HAMMER2_INODE_MAXNAME);
ipdata = &nchain->data->ipdata;
*ipdata = ochain->data->ipdata;
bcopy(name, ipdata->filename, name_len);
ipdata->name_key = lhc;
ipdata->name_len = name_len;
atomic_clear_int(&nchain->core->flags,
HAMMER2_CORE_COUNTEDBREFS);
kprintf("created fake hardlink %*.*s\n",
(int)name_len, (int)name_len, name);
} else {
/*
* nchain is a duplicate of ochain at the new location.
* We must fixup the name stored in oip. The bref key
* has already been set up.
*/
hammer2_chain_modify(trans, &nchain, 0);
ipdata = &nchain->data->ipdata;
KKASSERT(name_len < HAMMER2_INODE_MAXNAME);
bcopy(name, ipdata->filename, name_len);
ipdata->name_key = lhc;
ipdata->name_len = name_len;
ipdata->nlinks = 1;
}
/*
* We are replacing ochain with nchain, unlock ochain. In the
* case where ochain is left unchanged the code above sets
* nchain to ochain and ochain to NULL, resulting in a NOP here.
*/
if (ochain)
hammer2_chain_unlock(ochain);
*chainp = nchain;
return (0);
}
/*
* Find a specific PFS by name
*/
static int
hammer2_ioctl_pfs_lookup(hammer2_inode_t *ip, void *data)
{
const hammer2_inode_data_t *ripdata;
hammer2_dev_t *hmp;
hammer2_ioc_pfs_t *pfs;
hammer2_chain_t *parent;
hammer2_chain_t *chain;
hammer2_key_t key_next;
hammer2_key_t lhc;
int cache_index = -1;
int error;
size_t len;
hmp = ip->pmp->pfs_hmps[0];
if (hmp == NULL)
return (EINVAL);
pfs = data;
error = 0;
hammer2_inode_lock(hmp->spmp->iroot, HAMMER2_RESOLVE_SHARED);
parent = hammer2_inode_chain(hmp->spmp->iroot, 0,
HAMMER2_RESOLVE_ALWAYS |
HAMMER2_RESOLVE_SHARED);
pfs->name[sizeof(pfs->name) - 1] = 0;
len = strlen(pfs->name);
lhc = hammer2_dirhash(pfs->name, len);
chain = hammer2_chain_lookup(&parent, &key_next,
lhc, lhc + HAMMER2_DIRHASH_LOMASK,
&cache_index,
HAMMER2_LOOKUP_SHARED);
while (chain) {
if (chain->bref.type == HAMMER2_BREF_TYPE_INODE) {
ripdata = &chain->data->ipdata;
if (ripdata->meta.name_len == len &&
bcmp(ripdata->filename, pfs->name, len) == 0) {
break;
}
ripdata = NULL; /* safety */
}
chain = hammer2_chain_next(&parent, chain, &key_next,
key_next,
lhc + HAMMER2_DIRHASH_LOMASK,
&cache_index,
HAMMER2_LOOKUP_SHARED);
}
/*
* Load the data being returned by the ioctl.
*/
if (chain) {
ripdata = &chain->data->ipdata;
pfs->name_key = ripdata->meta.name_key;
pfs->pfs_type = ripdata->meta.pfs_type;
pfs->pfs_subtype = ripdata->meta.pfs_subtype;
pfs->pfs_clid = ripdata->meta.pfs_clid;
pfs->pfs_fsid = ripdata->meta.pfs_fsid;
ripdata = NULL;
hammer2_chain_unlock(chain);
hammer2_chain_drop(chain);
} else {
error = ENOENT;
}
if (parent) {
hammer2_chain_unlock(parent);
hammer2_chain_drop(parent);
}
hammer2_inode_unlock(hmp->spmp->iroot);
return (error);
}