/*
* 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);
}
}
/*
* 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);
}
/*
* 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);
}
/*
* 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);
}
}
/*
* If an open file is unlinked H2 needs to retain the file in the topology
* to ensure that its backing store is not recovered by the bulk free scan.
* This also allows us to avoid having to special-case the CHAIN_DELETED flag.
*
* To do this the file is moved to a hidden directory in the PFS root and
* renamed. The hidden directory must be created if it does not exist.
*/
static
void
hammer2_inode_move_to_hidden(hammer2_trans_t *trans, hammer2_chain_t **chainp,
hammer2_tid_t inum)
{
hammer2_chain_t *chain;
hammer2_chain_t *dchain;
hammer2_pfsmount_t *pmp;
int error;
chain = *chainp;
pmp = chain->pmp;
KKASSERT(pmp != NULL);
KKASSERT(pmp->ihidden != NULL);
hammer2_chain_delete(trans, chain, 0);
dchain = hammer2_inode_lock_ex(pmp->ihidden);
error = hammer2_inode_connect(trans, chainp, 0,
pmp->ihidden, &dchain,
NULL, 0, inum);
hammer2_inode_unlock_ex(pmp->ihidden, dchain);
KKASSERT(error == 0);
}
/*
* 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);
}
/*
* This is called from the mount code to initialize pmp->ihidden
*/
void
hammer2_inode_install_hidden(hammer2_pfsmount_t *pmp)
{
hammer2_trans_t trans;
hammer2_chain_t *parent;
hammer2_chain_t *chain;
hammer2_chain_t *scan;
hammer2_inode_data_t *ipdata;
hammer2_key_t key_dummy;
hammer2_key_t key_next;
int cache_index;
int error;
int count;
if (pmp->ihidden)
return;
/*
* Find the hidden directory
*/
bzero(&key_dummy, sizeof(key_dummy));
hammer2_trans_init(&trans, pmp, NULL, 0);
parent = hammer2_inode_lock_ex(pmp->iroot);
chain = hammer2_chain_lookup(&parent, &key_dummy,
HAMMER2_INODE_HIDDENDIR,
HAMMER2_INODE_HIDDENDIR,
&cache_index, 0);
if (chain) {
pmp->ihidden = hammer2_inode_get(pmp, pmp->iroot, chain);
hammer2_inode_ref(pmp->ihidden);
/*
* Remove any unlinked files which were left open as-of
* any system crash.
*/
count = 0;
scan = hammer2_chain_lookup(&chain, &key_next,
0, HAMMER2_MAX_TID,
&cache_index,
HAMMER2_LOOKUP_NODATA);
while (scan) {
if (scan->bref.type == HAMMER2_BREF_TYPE_INODE) {
hammer2_chain_delete(&trans, scan, 0);
++count;
}
scan = hammer2_chain_next(&chain, scan, &key_next,
0, HAMMER2_MAX_TID,
&cache_index,
HAMMER2_LOOKUP_NODATA);
}
hammer2_inode_unlock_ex(pmp->ihidden, chain);
hammer2_inode_unlock_ex(pmp->iroot, parent);
hammer2_trans_done(&trans);
kprintf("hammer2: PFS loaded hidden dir, "
"removed %d dead entries\n", count);
return;
}
/*
* Create the hidden directory
*/
error = hammer2_chain_create(&trans, &parent, &chain,
HAMMER2_INODE_HIDDENDIR, 0,
HAMMER2_BREF_TYPE_INODE,
HAMMER2_INODE_BYTES);
hammer2_inode_unlock_ex(pmp->iroot, parent);
hammer2_chain_modify(&trans, &chain, 0);
ipdata = &chain->data->ipdata;
ipdata->type = HAMMER2_OBJTYPE_DIRECTORY;
ipdata->inum = HAMMER2_INODE_HIDDENDIR;
ipdata->nlinks = 1;
kprintf("hammer2: PFS root missing hidden directory, creating\n");
pmp->ihidden = hammer2_inode_get(pmp, pmp->iroot, chain);
hammer2_inode_ref(pmp->ihidden);
hammer2_inode_unlock_ex(pmp->ihidden, chain);
hammer2_trans_done(&trans);
}
/*
* 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;
}
/*
* 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.
*
* NOTE! This function does not prevent the underlying file from still
* being used if it has other refs (such as from an inode, or if it's
* chain is manually held). However, 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)
{
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 lhc;
int error;
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,
lhc, lhc + HAMMER2_DIRHASH_LOMASK,
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,
lhc, lhc + HAMMER2_DIRHASH_LOMASK,
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 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,
0, (hammer2_key_t)-1,
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.
*
* 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);
hammer2_chain_unlock(ochain);
/*
* Then decrement nlinks on hardlink target, deleting
* the target when nlinks drops to 0.
*/
hammer2_chain_modify(trans, &chain, 0);
--chain->data->ipdata.nlinks;
if (chain->data->ipdata.nlinks == 0)
hammer2_chain_delete(trans, chain);
} else {
/*
* Otherwise this was not a hardlink and we can just
* remove the entry and decrement nlinks.
*
* NOTE: *_get() integrates chain's lock into the inode lock.
*/
hammer2_chain_modify(trans, &chain, 0);
ipdata = &chain->data->ipdata;
--ipdata->nlinks;
hammer2_chain_delete(trans, chain);
}
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 snapshot of the specified {parent, ochain} with the specified
* label. The originating hammer2_inode must be exclusively locked for
* safety.
*
* The ioctl code has already synced the filesystem.
*/
int
hammer2_cluster_snapshot(hammer2_trans_t *trans, hammer2_cluster_t *ocluster,
hammer2_ioc_pfs_t *pfs)
{
hammer2_mount_t *hmp;
hammer2_cluster_t *ncluster;
const hammer2_inode_data_t *ipdata;
hammer2_inode_data_t *wipdata;
hammer2_inode_t *nip;
size_t name_len;
hammer2_key_t lhc;
struct vattr vat;
uuid_t opfs_clid;
int error;
kprintf("snapshot %s\n", pfs->name);
name_len = strlen(pfs->name);
lhc = hammer2_dirhash(pfs->name, name_len);
ipdata = &hammer2_cluster_data(ocluster)->ipdata;
opfs_clid = ipdata->pfs_clid;
hmp = ocluster->focus->hmp;
/*
* Create the snapshot directory under the super-root
*
* Set PFS type, generate a unique filesystem id, and generate
* a cluster id. Use the same clid when snapshotting a PFS root,
* which theoretically allows the snapshot to be used as part of
* the same cluster (perhaps as a cache).
*
* Copy the (flushed) blockref array. Theoretically we could use
* chain_duplicate() but it becomes difficult to disentangle
* the shared core so for now just brute-force it.
*/
VATTR_NULL(&vat);
vat.va_type = VDIR;
vat.va_mode = 0755;
ncluster = NULL;
nip = hammer2_inode_create(trans, hmp->spmp->iroot, &vat,
proc0.p_ucred, pfs->name, name_len,
&ncluster, &error);
if (nip) {
wipdata = hammer2_cluster_modify_ip(trans, nip, ncluster, 0);
wipdata->pfs_type = HAMMER2_PFSTYPE_SNAPSHOT;
kern_uuidgen(&wipdata->pfs_fsid, 1);
if (ocluster->focus->flags & HAMMER2_CHAIN_PFSROOT)
wipdata->pfs_clid = opfs_clid;
else
kern_uuidgen(&wipdata->pfs_clid, 1);
hammer2_cluster_set_chainflags(ncluster, HAMMER2_CHAIN_PFSROOT);
/* XXX hack blockset copy */
/* XXX doesn't work with real cluster */
KKASSERT(ocluster->nchains == 1);
wipdata->u.blockset = ocluster->focus->data->ipdata.u.blockset;
hammer2_cluster_modsync(ncluster);
hammer2_inode_unlock_ex(nip, ncluster);
}
return (error);
}
/*
* Consolidate for hard link creation. This moves the specified terminal
* hardlink inode to a directory common to its current directory and tdip
* if necessary, replacing *ipp with the new inode chain element and
* modifying the original inode chain element to OBJTYPE_HARDLINK.
*
* If the original inode chain element was a prior incarnation of a hidden
* inode it can simply be deleted instead of converted.
*
* (*ipp)'s nlinks field is locked on entry and the new (*ipp)'s nlinks
* field will be locked on return (with the original's unlocked).
*
* The link count is bumped if requested.
*/
int
hammer2_hardlink_consolidate(hammer2_inode_t **ipp, hammer2_inode_t *tdip)
{
hammer2_mount_t *hmp;
hammer2_inode_t *oip = *ipp;
hammer2_inode_t *nip = NULL;
hammer2_inode_t *fdip;
hammer2_chain_t *parent;
int error;
hmp = tdip->hmp;
if (hammer2_hardlink_enable < 0)
return (0);
if (hammer2_hardlink_enable == 0)
/*XXX return (ENOTSUP);
*/
/*
* Find the common parent directory
*/
fdip = oip->pip;
while (fdip->depth > tdip->depth) {
fdip = fdip->pip;
KKASSERT(fdip != NULL);
}
while (tdip->depth > fdip->depth) {
tdip = tdip->pip;
KKASSERT(tdip != NULL);
}
while (fdip != tdip) {
fdip = fdip->pip;
tdip = tdip->pip;
KKASSERT(fdip != NULL);
KKASSERT(tdip != NULL);
}
/*
* Nothing to do (except bump the link count) if the hardlink has
* already been consolidated in the correct place.
*/
if (oip->pip == fdip &&
(oip->ip_data.name_key & HAMMER2_DIRHASH_VISIBLE) == 0) {
kprintf("hardlink already consolidated correctly\n");
nip = oip;
hammer2_inode_lock_ex(nip);
hammer2_chain_modify(hmp, &nip->chain, 0);
++nip->ip_data.nlinks;
hammer2_inode_unlock_ex(nip);
return (0);
}
/*
* Create a hidden inode directory entry in the parent, copying
* (*oip)'s state. Then replace oip with OBJTYPE_HARDLINK.
*
* The duplication function will either flush or move any chains
* under oip to the new hardlink target inode, retiring all chains
* related to oip before returning. XXX vp->ip races.
*/
error = hammer2_inode_duplicate(fdip, oip, &nip, NULL, 0);
if (error == 0) {
/*
* Bump nlinks on duplicated hidden inode.
*/
kprintf("hardlink consolidation success in parent dir %s\n",
fdip->ip_data.filename);
hammer2_inode_lock_nlinks(nip);
hammer2_inode_unlock_nlinks(oip);
hammer2_chain_modify(hmp, &nip->chain, 0);
++nip->ip_data.nlinks;
hammer2_inode_unlock_ex(nip);
if (oip->ip_data.name_key & HAMMER2_DIRHASH_VISIBLE) {
/*
* Replace the old inode with an OBJTYPE_HARDLINK
* pointer.
*/
hammer2_inode_lock_ex(oip);
hammer2_chain_modify(hmp, &oip->chain, 0);
oip->ip_data.target_type = oip->ip_data.type;
oip->ip_data.type = HAMMER2_OBJTYPE_HARDLINK;
oip->ip_data.uflags = 0;
oip->ip_data.rmajor = 0;
oip->ip_data.rminor = 0;
oip->ip_data.ctime = 0;
oip->ip_data.mtime = 0;
oip->ip_data.atime = 0;
oip->ip_data.btime = 0;
bzero(&oip->ip_data.uid, sizeof(oip->ip_data.uid));
bzero(&oip->ip_data.gid, sizeof(oip->ip_data.gid));
oip->ip_data.op_flags = HAMMER2_OPFLAG_DIRECTDATA;
oip->ip_data.cap_flags = 0;
oip->ip_data.mode = 0;
oip->ip_data.size = 0;
oip->ip_data.nlinks = 1;
oip->ip_data.iparent = 0; /* XXX */
oip->ip_data.pfs_type = 0;
oip->ip_data.pfs_inum = 0;
bzero(&oip->ip_data.pfs_clid,
sizeof(oip->ip_data.pfs_clid));
bzero(&oip->ip_data.pfs_fsid,
sizeof(oip->ip_data.pfs_fsid));
oip->ip_data.data_quota = 0;
oip->ip_data.data_count = 0;
oip->ip_data.inode_quota = 0;
oip->ip_data.inode_count = 0;
oip->ip_data.attr_tid = 0;
oip->ip_data.dirent_tid = 0;
bzero(&oip->ip_data.u, sizeof(oip->ip_data.u));
/* XXX transaction ids */
hammer2_inode_unlock_ex(oip);
} else {
/*
* The old inode was a hardlink target, which we
* have now moved. We must delete it so the new
* hardlink target at a higher directory level
* becomes the only hardlink target for this inode.
*/
kprintf("DELETE INVISIBLE\n");
parent = oip->chain.parent;
hammer2_chain_lock(hmp, parent,
HAMMER2_RESOLVE_ALWAYS);
hammer2_chain_lock(hmp, &oip->chain,
HAMMER2_RESOLVE_ALWAYS);
hammer2_chain_delete(hmp, parent, &oip->chain, 0);
hammer2_chain_unlock(hmp, &oip->chain);
hammer2_chain_unlock(hmp, parent);
}
*ipp = nip;
} else {
KKASSERT(nip == NULL);
}
return (error);
}
/*
* Duplicate the specified existing inode in the specified target directory.
* If name is NULL the inode is duplicated as a hidden directory entry.
*
* Returns the new inode. The old inode is left alone.
*
* XXX name needs to be NULL for now.
*/
int
hammer2_inode_duplicate(hammer2_inode_t *dip, hammer2_inode_t *oip,
hammer2_inode_t **nipp,
const uint8_t *name, size_t name_len)
{
hammer2_mount_t *hmp = dip->hmp;
hammer2_inode_t *nip;
hammer2_chain_t *parent;
hammer2_chain_t *chain;
hammer2_key_t lhc;
int error;
if (name) {
lhc = hammer2_dirhash(name, name_len);
} else {
lhc = oip->ip_data.inum;
KKASSERT((lhc & HAMMER2_DIRHASH_VISIBLE) == 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.
*/
nip = NULL;
parent = &dip->chain;
hammer2_chain_lock(hmp, parent, HAMMER2_RESOLVE_ALWAYS);
error = 0;
while (error == 0) {
chain = hammer2_chain_lookup(hmp, &parent, lhc, lhc, 0);
if (chain == NULL)
break;
/* XXX bcmp name if not NULL */
if ((lhc & HAMMER2_DIRHASH_LOMASK) == HAMMER2_DIRHASH_LOMASK)
error = ENOSPC;
if ((lhc & HAMMER2_DIRHASH_VISIBLE) == 0) /* shouldn't happen */
error = ENOSPC;
hammer2_chain_unlock(hmp, chain);
chain = NULL;
++lhc;
}
/*
* Create entry in common parent directory.
*/
if (error == 0) {
chain = hammer2_chain_create(hmp, parent, NULL, lhc, 0,
HAMMER2_BREF_TYPE_INODE /* n/a */,
HAMMER2_INODE_BYTES); /* n/a */
if (chain == NULL)
error = EIO;
}
hammer2_chain_unlock(hmp, parent);
/*
* Handle the error case
*/
if (error) {
KKASSERT(chain == NULL);
return (error);
}
/*
* XXX This is currently a horrible hack. Well, if we wanted to
* duplicate a file, i.e. as in a snapshot, we definitely
* would have to flush it first.
*
* For hardlink target generation we can theoretically move any
* active chain structures without flushing, but that gets really
* iffy for code which follows chain->parent and ip->pip links.
*
* XXX only works with files. Duplicating a directory hierarchy
* requires a flush but doesn't deal with races post-flush.
* Well, it would work I guess, but you might catch some files
* mid-operation.
*
* We cannot leave oip with any in-memory chains because (for a
* hardlink), oip will become a OBJTYPE_HARDLINK which is just a
* pointer to the real hardlink's inum and can't have any sub-chains.
* XXX might be 0-ref chains left.
*/
hammer2_inode_lock_ex(oip);
hammer2_chain_flush(hmp, &oip->chain, 0);
hammer2_inode_unlock_ex(oip);
/*KKASSERT(RB_EMPTY(&oip->chain.rbhead));*/
nip = chain->u.ip;
hammer2_chain_modify(hmp, chain, 0);
nip->ip_data = oip->ip_data; /* sync media data after flush */
if (name) {
/*
* Directory entries are inodes so if the name has changed
* we have to update the inode.
*/
KKASSERT(name_len < HAMMER2_INODE_MAXNAME);
bcopy(name, nip->ip_data.filename, name_len);
nip->ip_data.name_key = lhc;
nip->ip_data.name_len = name_len;
} else {
/*
* 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.
*/
ksnprintf(nip->ip_data.filename, sizeof(nip->ip_data.filename),
"0x%016jx", (intmax_t)nip->ip_data.inum);
nip->ip_data.name_len = strlen(nip->ip_data.filename);
nip->ip_data.name_key = lhc;
}
*nipp = nip;
return (0);
}