/*
* 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);
}
}
/*
* 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;
}
/*
* 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);
}
}
/*
* 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;
}
/*
* 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);
}
/*
* 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);
}
/*
* 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.
*/
int
hammer2_inode_create(hammer2_inode_t *dip,
struct vattr *vap, struct ucred *cred,
const uint8_t *name, size_t name_len,
hammer2_inode_t **nipp)
{
hammer2_mount_t *hmp = dip->hmp;
hammer2_chain_t *chain;
hammer2_chain_t *parent;
hammer2_inode_t *nip;
hammer2_key_t lhc;
int error;
uid_t xuid;
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.
*/
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;
if ((lhc & HAMMER2_DIRHASH_VISIBLE) == 0)
error = ENOSPC;
if ((lhc & HAMMER2_DIRHASH_LOMASK) == HAMMER2_DIRHASH_LOMASK)
error = ENOSPC;
hammer2_chain_unlock(hmp, chain);
chain = NULL;
++lhc;
}
if (error == 0) {
chain = hammer2_chain_create(hmp, parent, NULL, lhc, 0,
HAMMER2_BREF_TYPE_INODE,
HAMMER2_INODE_BYTES);
if (chain == NULL)
error = EIO;
}
hammer2_chain_unlock(hmp, parent);
/*
* Handle the error case
*/
if (error) {
KKASSERT(chain == NULL);
*nipp = NULL;
return (error);
}
/*
* Set up the new inode
*/
nip = chain->u.ip;
*nipp = nip;
hammer2_voldata_lock(hmp);
if (vap) {
nip->ip_data.type = hammer2_get_obj_type(vap->va_type);
nip->ip_data.inum = hmp->voldata.alloc_tid++;
/* XXX modify/lock */
} else {
nip->ip_data.type = HAMMER2_OBJTYPE_DIRECTORY;
nip->ip_data.inum = 1;
}
hammer2_voldata_unlock(hmp);
nip->ip_data.version = HAMMER2_INODE_VERSION_ONE;
hammer2_update_time(&nip->ip_data.ctime);
nip->ip_data.mtime = nip->ip_data.ctime;
if (vap)
nip->ip_data.mode = vap->va_mode;
nip->ip_data.nlinks = 1;
if (vap) {
if (dip) {
xuid = hammer2_to_unix_xid(&dip->ip_data.uid);
xuid = vop_helper_create_uid(dip->pmp->mp,
dip->ip_data.mode,
xuid,
cred,
&vap->va_mode);
} else {
xuid = 0;
}
/* XXX
if (vap->va_vaflags & VA_UID_UUID_VALID)
nip->ip_data.uid = vap->va_uid_uuid;
else if (vap->va_uid != (uid_t)VNOVAL)
hammer2_guid_to_uuid(&nip->ip_data.uid, vap->va_uid);
else
*/
hammer2_guid_to_uuid(&nip->ip_data.uid, xuid);
/* XXX if (vap->va_vaflags & VA_GID_UUID_VALID)
nip->ip_data.gid = vap->va_gid_uuid;
else if (vap->va_gid != (gid_t)VNOVAL)
hammer2_guid_to_uuid(&nip->ip_data.gid, vap->va_gid);
else */ if (dip)
nip->ip_data.gid = dip->ip_data.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 (nip->ip_data.type == HAMMER2_OBJTYPE_REGFILE ||
nip->ip_data.type == HAMMER2_OBJTYPE_SOFTLINK) {
nip->ip_data.op_flags |= HAMMER2_OPFLAG_DIRECTDATA;
}
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;
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;
}
/*
* 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));
}
}
}
/*
* 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;
}
/*
* Locate first match or overlap under parent, return a new cluster
*/
hammer2_cluster_t *
hammer2_cluster_lookup(hammer2_cluster_t *cparent, hammer2_key_t *key_nextp,
hammer2_key_t key_beg, hammer2_key_t key_end,
int flags, int *ddflagp)
{
hammer2_pfsmount_t *pmp;
hammer2_cluster_t *cluster;
hammer2_chain_t *chain;
hammer2_key_t key_accum;
hammer2_key_t key_next;
hammer2_key_t bref_key;
int bref_keybits;
int null_count;
int ddflag;
int i;
uint8_t bref_type;
u_int bytes;
pmp = cparent->pmp; /* can be NULL */
key_accum = *key_nextp;
null_count = 0;
bref_type = 0;
bref_key = 0;
bref_keybits = 0;
bytes = 0;
cluster = kmalloc(sizeof(*cluster), M_HAMMER2, M_WAITOK | M_ZERO);
cluster->pmp = pmp; /* can be NULL */
cluster->refs = 1;
/* cluster->focus = NULL; already null */
cparent->focus = NULL;
*ddflagp = 0;
for (i = 0; i < cparent->nchains; ++i) {
key_next = *key_nextp;
if (cparent->array[i] == NULL) {
++null_count;
continue;
}
chain = hammer2_chain_lookup(&cparent->array[i], &key_next,
key_beg, key_end,
&cparent->cache_index[i],
flags, &ddflag);
if (cparent->focus == NULL)
cparent->focus = cparent->array[i];
cluster->array[i] = chain;
if (chain == NULL) {
++null_count;
} else {
if (cluster->focus == NULL) {
bref_type = chain->bref.type;
bref_key = chain->bref.key;
bref_keybits = chain->bref.keybits;
bytes = chain->bytes;
*ddflagp = ddflag;
cluster->focus = chain;
}
KKASSERT(bref_type == chain->bref.type);
KKASSERT(bref_key == chain->bref.key);
KKASSERT(bref_keybits == chain->bref.keybits);
KKASSERT(bytes == chain->bytes);
KKASSERT(*ddflagp == ddflag);
}
if (key_accum > key_next)
key_accum = key_next;
}
*key_nextp = key_accum;
cluster->nchains = i;
if (null_count == i) {
hammer2_cluster_drop(cluster);
cluster = NULL;
}
return (cluster);
}
/*
* 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.
*/
int
hammer2_unlink_file(hammer2_inode_t *dip,
const uint8_t *name, size_t name_len,
int isdir, hammer2_inode_t *retain_ip)
{
hammer2_mount_t *hmp;
hammer2_chain_t *parent;
hammer2_chain_t *chain;
hammer2_chain_t *dparent;
hammer2_chain_t *dchain;
hammer2_key_t lhc;
hammer2_inode_t *ip;
hammer2_inode_t *oip;
int error;
uint8_t type;
error = 0;
oip = NULL;
hmp = dip->hmp;
lhc = hammer2_dirhash(name, name_len);
/*
* Search for the filename in the directory
*/
parent = &dip->chain;
hammer2_chain_lock(hmp, parent, HAMMER2_RESOLVE_ALWAYS);
chain = hammer2_chain_lookup(hmp, &parent,
lhc, lhc + HAMMER2_DIRHASH_LOMASK,
0);
while (chain) {
if (chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
chain->u.ip &&
name_len == chain->data->ipdata.name_len &&
bcmp(name, chain->data->ipdata.filename, name_len) == 0) {
break;
}
chain = hammer2_chain_next(hmp, &parent, chain,
lhc, lhc + HAMMER2_DIRHASH_LOMASK,
0);
}
/*
* Not found or wrong type (isdir < 0 disables the type check).
*/
if (chain == NULL) {
hammer2_chain_unlock(hmp, parent);
return ENOENT;
}
if ((type = chain->data->ipdata.type) == HAMMER2_OBJTYPE_HARDLINK)
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.
*/
if (chain->data->ipdata.type == HAMMER2_OBJTYPE_HARDLINK) {
hammer2_chain_unlock(hmp, parent);
parent = NULL;
error = hammer2_hardlink_find(dip, &chain, &oip);
}
/*
* 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.
*
* 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 >= 0) {
dparent = chain;
hammer2_chain_lock(hmp, dparent, HAMMER2_RESOLVE_ALWAYS);
dchain = hammer2_chain_lookup(hmp, &dparent,
0, (hammer2_key_t)-1,
HAMMER2_LOOKUP_NODATA);
if (dchain) {
hammer2_chain_unlock(hmp, dchain);
hammer2_chain_unlock(hmp, dparent);
error = ENOTEMPTY;
goto done;
}
hammer2_chain_unlock(hmp, 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.
*/
if (oip) {
/*
* If this was a hardlink we first delete the hardlink
* pointer entry.
*/
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,
(retain_ip == oip));
hammer2_chain_unlock(hmp, &oip->chain);
hammer2_chain_unlock(hmp, parent);
parent = NULL;
/*
* Then decrement nlinks on hardlink target.
*/
ip = chain->u.ip;
if (ip->ip_data.nlinks == 1) {
dparent = chain->parent;
hammer2_chain_ref(hmp, chain);
hammer2_chain_unlock(hmp, chain);
hammer2_chain_lock(hmp, dparent,
HAMMER2_RESOLVE_ALWAYS);
hammer2_chain_lock(hmp, chain, HAMMER2_RESOLVE_ALWAYS);
hammer2_chain_drop(hmp, chain);
hammer2_chain_modify(hmp, chain, 0);
--ip->ip_data.nlinks;
hammer2_chain_delete(hmp, dparent, chain, 0);
hammer2_chain_unlock(hmp, dparent);
} else {
hammer2_chain_modify(hmp, chain, 0);
--ip->ip_data.nlinks;
}
} else {
/*
* Otherwise this was not a hardlink and we can just
* remove the entry and decrement nlinks.
*/
ip = chain->u.ip;
hammer2_chain_modify(hmp, chain, 0);
--ip->ip_data.nlinks;
hammer2_chain_delete(hmp, parent, chain,
(retain_ip == ip));
}
error = 0;
done:
if (chain)
hammer2_chain_unlock(hmp, chain);
if (parent)
hammer2_chain_unlock(hmp, parent);
if (oip)
hammer2_chain_drop(oip->hmp, &oip->chain);
return error;
}
/*
* Connect inode (oip) to the specified directory using the specified name.
* (oip) must be locked.
*
* If (oip) is not currently connected we simply connect it up.
*
* If (oip) is already connected we create a OBJTYPE_HARDLINK entry which
* points to (oip)'s inode number. (oip) is expected to be the terminus of
* the hardlink sitting as a hidden file in a common parent directory
* in this situation (thus the lock order is correct).
*/
int
hammer2_inode_connect(hammer2_inode_t *dip, hammer2_inode_t *oip,
const uint8_t *name, size_t name_len)
{
hammer2_mount_t *hmp = dip->hmp;
hammer2_chain_t *chain;
hammer2_chain_t *parent;
hammer2_inode_t *nip;
hammer2_key_t lhc;
int error;
int hlink;
lhc = hammer2_dirhash(name, name_len);
hlink = (oip->chain.parent != NULL);
/*
* In fake mode flush oip so we can just snapshot it downbelow.
*/
if (hlink && hammer2_hardlink_enable < 0)
hammer2_chain_flush(hmp, &oip->chain, 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.
*/
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;
if ((lhc & HAMMER2_DIRHASH_LOMASK) == HAMMER2_DIRHASH_LOMASK)
error = ENOSPC;
hammer2_chain_unlock(hmp, chain);
chain = NULL;
++lhc;
}
/*
* Passing a non-NULL chain to hammer2_chain_create() reconnects the
* existing chain instead of creating a new one. The chain's bref
* will be properly updated.
*/
if (error == 0) {
if (hlink) {
chain = hammer2_chain_create(hmp, parent,
NULL, lhc, 0,
HAMMER2_BREF_TYPE_INODE,
HAMMER2_INODE_BYTES);
} else {
chain = hammer2_chain_create(hmp, parent,
&oip->chain, lhc, 0,
HAMMER2_BREF_TYPE_INODE,
HAMMER2_INODE_BYTES);
if (chain)
KKASSERT(chain == &oip->chain);
}
if (chain == NULL)
error = EIO;
}
hammer2_chain_unlock(hmp, parent);
/*
* Handle the error case
*/
if (error) {
KKASSERT(chain == 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.
*/
if (hlink && hammer2_hardlink_enable >= 0) {
/*
* Create the HARDLINK pointer. oip represents the hardlink
* target in this situation.
*/
nip = chain->u.ip;
hammer2_chain_modify(hmp, chain, 0);
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;
nip->ip_data.target_type = oip->ip_data.type;
nip->ip_data.type = HAMMER2_OBJTYPE_HARDLINK;
nip->ip_data.inum = oip->ip_data.inum;
nip->ip_data.nlinks = 1;
kprintf("created hardlink %*.*s\n",
(int)name_len, (int)name_len, name);
hammer2_chain_unlock(hmp, chain);
} else if (hlink && hammer2_hardlink_enable < 0) {
/*
* Create a snapshot (hardlink fake mode for debugging).
*/
nip = chain->u.ip;
nip->ip_data = oip->ip_data;
hammer2_chain_modify(hmp, chain, 0);
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;
kprintf("created fake hardlink %*.*s\n",
(int)name_len, (int)name_len, name);
hammer2_chain_unlock(hmp, chain);
} else {
/*
* Normally disconnected inode (e.g. during a rename) that
* was reconnected. We must fixup the name stored in
* oip.
*
* We are using oip as chain, already locked by caller,
* do not unlock it.
*/
hammer2_chain_modify(hmp, chain, 0);
if (oip->ip_data.name_len != name_len ||
bcmp(oip->ip_data.filename, name, name_len) != 0) {
KKASSERT(name_len < HAMMER2_INODE_MAXNAME);
bcopy(name, oip->ip_data.filename, name_len);
oip->ip_data.name_key = lhc;
oip->ip_data.name_len = name_len;
}
oip->ip_data.nlinks = 1;
}
return (0);
}
/*
* 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);
}
/*
* 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);
}
/*
* ochain represents the target file inode. We need to move it to the
* specified common parent directory (dip) and rename it to a special
* invisible "0xINODENUMBER" filename.
*
* We use chain_duplicate and duplicate ochain at the new location,
* renaming it appropriately. We create a temporary chain and
* then delete it to placemark where the duplicate will go. Both of
* these use the inode number for (lhc) (the key), generating the
* invisible filename.
*/
static
hammer2_chain_t *
hammer2_hardlink_shiftup(hammer2_trans_t *trans, hammer2_chain_t **ochainp,
hammer2_inode_t *dip, int *errorp)
{
hammer2_inode_data_t *nipdata;
hammer2_chain_t *parent;
hammer2_chain_t *ochain;
hammer2_chain_t *nchain;
hammer2_chain_t *tmp;
hammer2_key_t lhc;
hammer2_blockref_t bref;
ochain = *ochainp;
*errorp = 0;
lhc = ochain->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.
*/
retry:
parent = hammer2_chain_lookup_init(dip->chain, 0);
nchain = hammer2_chain_lookup(&parent, lhc, lhc, 0);
if (nchain) {
kprintf("X3 chain %p parent %p dip %p dip->chain %p\n",
nchain, parent, dip, dip->chain);
hammer2_chain_unlock(nchain);
nchain = NULL;
*errorp = ENOSPC;
#if 1
Debugger("X3");
#endif
}
/*
* Create entry in common parent directory using the seek position
* calculated above.
*/
if (*errorp == 0) {
KKASSERT(nchain == NULL);
*errorp = hammer2_chain_create(trans, &parent, &nchain,
lhc, 0,
HAMMER2_BREF_TYPE_INODE,/* n/a */
HAMMER2_INODE_BYTES); /* n/a */
hammer2_chain_refactor(&ochain);
*ochainp = ochain;
}
/*
* Cleanup and handle retries.
*/
if (*errorp == EAGAIN) {
hammer2_chain_ref(parent);
hammer2_chain_lookup_done(parent);
hammer2_chain_wait(parent);
hammer2_chain_drop(parent);
goto retry;
}
/*
* Handle the error case
*/
if (*errorp) {
KKASSERT(nchain == NULL);
hammer2_chain_lookup_done(parent);
return (NULL);
}
/*
* Use chain as a placeholder for (lhc), delete it and replace
* it with our duplication.
*
* Gain a second lock on ochain for the duplication function to
* unlock, maintain the caller's original lock across the call.
*
* This is a bit messy.
*/
hammer2_chain_delete(trans, nchain);
hammer2_chain_lock(ochain, HAMMER2_RESOLVE_ALWAYS);
tmp = ochain;
bref = tmp->bref;
bref.key = lhc; /* invisible dir entry key */
bref.keybits = 0;
hammer2_chain_duplicate(trans, parent, nchain->index, &tmp, &bref);
hammer2_chain_lookup_done(parent);
hammer2_chain_unlock(nchain); /* no longer needed */
/*
* Now set chain to our duplicate and modify it appropriately.
*
* 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.
*/
nchain = tmp;
tmp = NULL; /* safety */
hammer2_chain_modify(trans, &nchain, HAMMER2_MODIFY_ASSERTNOCOPY);
nipdata = &nchain->data->ipdata;
ksnprintf(nipdata->filename, sizeof(nipdata->filename),
"0x%016jx", (intmax_t)nipdata->inum);
nipdata->name_len = strlen(nipdata->filename);
nipdata->name_key = lhc;
return (nchain);
}
/*
* Connect the target inode represented by (*chainp) to the media topology
* at (dip, name, len).
*
* 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, int hlink,
hammer2_inode_t *dip, hammer2_chain_t **chainp,
const uint8_t *name, size_t name_len)
{
hammer2_inode_data_t *ipdata;
hammer2_chain_t *nchain;
hammer2_chain_t *parent;
hammer2_chain_t *ochain;
hammer2_key_t lhc;
int error;
ochain = *chainp;
/*
* 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.
*/
parent = hammer2_chain_lookup_init(dip->chain, 0);
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(&parent, lhc, lhc, 0);
if (nchain == NULL)
break;
if ((lhc & HAMMER2_DIRHASH_LOMASK) == HAMMER2_DIRHASH_LOMASK)
error = ENOSPC;
hammer2_chain_unlock(nchain);
nchain = NULL;
++lhc;
}
if (error == 0) {
if (hlink) {
/*
* Hardlink pointer needed, create totally fresh
* directory entry.
*/
KKASSERT(nchain == NULL);
error = hammer2_chain_create(trans, &parent, &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).
*/
nchain = ochain;
ochain = NULL;
hammer2_chain_duplicate(trans, NULL, -1, &nchain, NULL);
error = hammer2_chain_create(trans, &parent, &nchain,
lhc, 0,
HAMMER2_BREF_TYPE_INODE,
HAMMER2_INODE_BYTES);
}
}
/*
* Unlock stuff.
*/
KKASSERT(error != EAGAIN);
hammer2_chain_lookup_done(parent);
parent = NULL;
/*
* 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,
HAMMER2_MODIFY_ASSERTNOCOPY);
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,
HAMMER2_MODIFY_ASSERTNOCOPY);
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;
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,
HAMMER2_MODIFY_ASSERTNOCOPY);
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);
}
/*
* 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);
}
/*
* Each out of sync node sync-thread must issue an all-nodes XOP scan of
* the inode. This creates a multiplication effect since the XOP scan itself
* issues to all nodes. However, this is the only way we can safely
* synchronize nodes which might have disparate I/O bandwidths and the only
* way we can safely deal with stalled nodes.
*/
static
int
hammer2_sync_slaves(hammer2_thread_t *thr, hammer2_inode_t *ip,
hammer2_deferred_list_t *list)
{
hammer2_xop_scanall_t *xop;
hammer2_chain_t *parent;
hammer2_chain_t *chain;
hammer2_pfs_t *pmp;
hammer2_key_t key_next;
hammer2_tid_t sync_tid;
int cache_index = -1;
int needrescan;
int wantupdate;
int error;
int nerror;
int idx;
int n;
pmp = ip->pmp;
idx = thr->clindex; /* cluster node we are responsible for */
needrescan = 0;
wantupdate = 0;
if (ip->cluster.focus == NULL)
return (EINPROGRESS);
sync_tid = ip->cluster.focus->bref.modify_tid;
#if 0
/*
* Nothing to do if all slaves are synchronized.
* Nothing to do if cluster not authoritatively readable.
*/
if (pmp->cluster_flags & HAMMER2_CLUSTER_SSYNCED)
return(0);
if ((pmp->cluster_flags & HAMMER2_CLUSTER_RDHARD) == 0)
return(HAMMER2_ERROR_INCOMPLETE);
#endif
error = 0;
/*
* The inode is left unlocked during the scan. Issue a XOP
* that does *not* include our cluster index to iterate
* properly synchronized elements and resolve our cluster index
* against it.
*/
hammer2_inode_lock(ip, HAMMER2_RESOLVE_SHARED);
xop = hammer2_xop_alloc(ip, HAMMER2_XOP_MODIFYING);
xop->key_beg = HAMMER2_KEY_MIN;
xop->key_end = HAMMER2_KEY_MAX;
hammer2_xop_start_except(&xop->head, hammer2_xop_scanall, idx);
parent = hammer2_inode_chain(ip, idx,
HAMMER2_RESOLVE_ALWAYS |
HAMMER2_RESOLVE_SHARED);
if (parent->bref.modify_tid != sync_tid)
wantupdate = 1;
hammer2_inode_unlock(ip);
chain = hammer2_chain_lookup(&parent, &key_next,
HAMMER2_KEY_MIN, HAMMER2_KEY_MAX,
&cache_index,
HAMMER2_LOOKUP_SHARED |
HAMMER2_LOOKUP_NODIRECT |
HAMMER2_LOOKUP_NODATA);
error = hammer2_xop_collect(&xop->head, 0);
kprintf("XOP_INITIAL xop=%p clindex %d on %s\n", xop, thr->clindex,
pmp->pfs_names[thr->clindex]);
for (;;) {
/*
* We are done if our scan is done and the XOP scan is done.
* We are done if the XOP scan failed (that is, we don't
* have authoritative data to synchronize with).
*/
int advance_local = 0;
int advance_xop = 0;
int dodefer = 0;
hammer2_chain_t *focus;
kprintf("loop xop=%p chain[1]=%p lockcnt=%d\n",
xop, xop->head.cluster.array[1].chain,
(xop->head.cluster.array[1].chain ?
xop->head.cluster.array[1].chain->lockcnt : -1)
);
if (chain == NULL && error == ENOENT)
break;
if (error && error != ENOENT)
break;
/*
* Compare
*/
if (chain && error == ENOENT) {
/*
* If we have local chains but the XOP scan is done,
* the chains need to be deleted.
*/
n = -1;
focus = NULL;
} else if (chain == NULL) {
/*
* If our local scan is done but the XOP scan is not,
* we need to create the missing chain(s).
*/
n = 1;
focus = xop->head.cluster.focus;
} else {
/*
* Otherwise compare to determine the action
* needed.
*/
focus = xop->head.cluster.focus;
n = hammer2_chain_cmp(chain, focus);
}
/*
* Take action based on comparison results.
*/
if (n < 0) {
/*
* Delete extranious local data. This will
* automatically advance the chain.
*/
nerror = hammer2_sync_destroy(thr, &parent, &chain,
0, idx);
} else if (n == 0 && chain->bref.modify_tid !=
focus->bref.modify_tid) {
/*
* Matching key but local data or meta-data requires
* updating. If we will recurse, we still need to
* update to compatible content first but we do not
* synchronize modify_tid until the entire recursion
* has completed successfully.
*/
if (focus->bref.type == HAMMER2_BREF_TYPE_INODE) {
nerror = hammer2_sync_replace(
thr, parent, chain,
0,
idx, focus);
dodefer = 1;
} else {
nerror = hammer2_sync_replace(
thr, parent, chain,
focus->bref.modify_tid,
idx, focus);
}
} else if (n == 0) {
/*
* 100% match, advance both
*/
advance_local = 1;
advance_xop = 1;
nerror = 0;
} else if (n > 0) {
/*
* Insert missing local data.
*
* If we will recurse, we still need to update to
* compatible content first but we do not synchronize
* modify_tid until the entire recursion has
* completed successfully.
*/
if (focus->bref.type == HAMMER2_BREF_TYPE_INODE) {
nerror = hammer2_sync_insert(
thr, &parent, &chain,
0,
idx, focus);
dodefer = 2;
} else {
nerror = hammer2_sync_insert(
thr, &parent, &chain,
focus->bref.modify_tid,
idx, focus);
}
advance_local = 1;
advance_xop = 1;
}
/*
* We cannot recurse depth-first because the XOP is still
* running in node threads for this scan. Create a placemarker
* by obtaining and record the hammer2_inode.
*
* We excluded our node from the XOP so we must temporarily
* add it to xop->head.cluster so it is properly incorporated
* into the inode.
*
* The deferral is pushed onto a LIFO list for bottom-up
* synchronization.
*/
if (error == 0 && dodefer) {
hammer2_inode_t *nip;
hammer2_deferred_ip_t *defer;
KKASSERT(focus->bref.type == HAMMER2_BREF_TYPE_INODE);
defer = kmalloc(sizeof(*defer), M_HAMMER2,
M_WAITOK | M_ZERO);
KKASSERT(xop->head.cluster.array[idx].chain == NULL);
xop->head.cluster.array[idx].flags =
HAMMER2_CITEM_INVALID;
xop->head.cluster.array[idx].chain = chain;
nip = hammer2_inode_get(pmp, ip,
&xop->head.cluster, idx);
xop->head.cluster.array[idx].chain = NULL;
hammer2_inode_ref(nip);
hammer2_inode_unlock(nip);
defer->next = list->base;
defer->ip = nip;
list->base = defer;
++list->count;
needrescan = 1;
}
/*
* If at least one deferral was added and the deferral
* list has grown too large, stop adding more. This
* will trigger an EAGAIN return.
*/
if (needrescan && list->count > 1000)
break;
/*
* Advancements for iteration.
*/
if (advance_xop) {
error = hammer2_xop_collect(&xop->head, 0);
}
if (advance_local) {
chain = hammer2_chain_next(&parent, chain, &key_next,
key_next, HAMMER2_KEY_MAX,
&cache_index,
HAMMER2_LOOKUP_SHARED |
HAMMER2_LOOKUP_NODIRECT |
HAMMER2_LOOKUP_NODATA);
}
}
hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
if (chain) {
hammer2_chain_unlock(chain);
hammer2_chain_drop(chain);
}
if (parent) {
hammer2_chain_unlock(parent);
hammer2_chain_drop(parent);
}
/*
* If we added deferrals we want the caller to synchronize them
* and then call us again.
*
* NOTE: In this situation we do not yet want to synchronize our
* inode, setting the error code also has that effect.
*/
if (error == 0 && needrescan)
error = EAGAIN;
/*
* If no error occurred and work was performed, synchronize the
* inode meta-data itself.
*
* XXX inode lock was lost
*/
if (error == 0 && wantupdate) {
hammer2_xop_ipcluster_t *xop2;
hammer2_chain_t *focus;
xop2 = hammer2_xop_alloc(ip, HAMMER2_XOP_MODIFYING);
hammer2_xop_start_except(&xop2->head, hammer2_xop_ipcluster,
idx);
error = hammer2_xop_collect(&xop2->head, 0);
if (error == 0) {
focus = xop2->head.cluster.focus;
kprintf("syncthr: update inode %p (%s)\n",
focus,
(focus ?
(char *)focus->data->ipdata.filename : "?"));
chain = hammer2_inode_chain_and_parent(ip, idx,
&parent,
HAMMER2_RESOLVE_ALWAYS |
HAMMER2_RESOLVE_SHARED);
KKASSERT(parent != NULL);
nerror = hammer2_sync_replace(
thr, parent, chain,
sync_tid,
idx, focus);
hammer2_chain_unlock(chain);
hammer2_chain_drop(chain);
hammer2_chain_unlock(parent);
hammer2_chain_drop(parent);
/* XXX */
}
hammer2_xop_retire(&xop2->head, HAMMER2_XOPMASK_VOP);
}
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);
}
/*
* Update LNK_SPAN state
*/
static void
hammer2_update_spans(hammer2_dev_t *hmp, kdmsg_state_t *state)
{
const hammer2_inode_data_t *ripdata;
hammer2_chain_t *parent;
hammer2_chain_t *chain;
hammer2_pfs_t *spmp;
hammer2_key_t key_next;
kdmsg_msg_t *rmsg;
size_t name_len;
int cache_index = -1;
/*
* Lookup mount point under the media-localized super-root.
*
* cluster->pmp will incorrectly point to spmp and must be fixed
* up later on.
*/
spmp = hmp->spmp;
hammer2_inode_lock(spmp->iroot, 0);
parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS);
chain = NULL;
if (parent == NULL)
goto done;
chain = hammer2_chain_lookup(&parent, &key_next,
HAMMER2_KEY_MIN, HAMMER2_KEY_MAX,
&cache_index,
0);
while (chain) {
if (chain->bref.type != HAMMER2_BREF_TYPE_INODE)
continue;
ripdata = &chain->data->ipdata;
kprintf("UPDATE SPANS: %s\n", ripdata->filename);
rmsg = kdmsg_msg_alloc(&hmp->iocom.state0,
DMSG_LNK_SPAN | DMSGF_CREATE,
hammer2_lnk_span_reply, NULL);
rmsg->any.lnk_span.peer_id = ripdata->meta.pfs_clid;
rmsg->any.lnk_span.pfs_id = ripdata->meta.pfs_fsid;
rmsg->any.lnk_span.pfs_type = ripdata->meta.pfs_type;
rmsg->any.lnk_span.peer_type = DMSG_PEER_HAMMER2;
rmsg->any.lnk_span.proto_version = DMSG_SPAN_PROTO_1;
name_len = ripdata->meta.name_len;
if (name_len >= sizeof(rmsg->any.lnk_span.peer_label))
name_len = sizeof(rmsg->any.lnk_span.peer_label) - 1;
bcopy(ripdata->filename,
rmsg->any.lnk_span.peer_label,
name_len);
kdmsg_msg_write(rmsg);
chain = hammer2_chain_next(&parent, chain, &key_next,
key_next, HAMMER2_KEY_MAX,
&cache_index,
0);
}
hammer2_inode_unlock(spmp->iroot);
done:
if (chain) {
hammer2_chain_unlock(chain);
hammer2_chain_drop(chain);
}
if (parent) {
hammer2_chain_unlock(parent);
hammer2_chain_drop(parent);
}
}
