spinlock_t ll_lookup_lock = SPIN_LOCK_UNLOCKED;
/* should NOT be called with the dcache lock, see fs/dcache.c */
-void ll_release(struct dentry *de)
+static void ll_release(struct dentry *de)
{
struct ll_dentry_data *lld;
ENTRY;
OBD_ALLOC_PTR(lld);
if (likely(lld != NULL)) {
- cfs_waitq_init(&lld->lld_waitq);
lock_dentry(de);
if (likely(de->d_fsdata == NULL))
de->d_fsdata = lld;
ll_finish_md_op_data(op_data);
if (it->it_op == IT_GETATTR && !first)
ll_statahead_exit(de, rc);
+ else if (first == -EEXIST)
+ ll_statahead_mark(de);
/* If req is NULL, then md_intent_lock only tried to do a lock match;
* if all was well, it will return 1 if it found locks, 0 otherwise. */
first = ll_statahead_enter(de->d_parent->d_inode, &de, 0);
if (!first)
ll_statahead_exit(de, rc);
+ else if (first == -EEXIST)
+ ll_statahead_mark(de);
}
return rc;
.d_unpin = ll_unpin,
#endif
};
-
-static int ll_fini_revalidate_nd(struct dentry *dentry, struct nameidata *nd)
-{
- ENTRY;
- /* need lookup */
- RETURN(0);
-}
-
-struct dentry_operations ll_fini_d_ops = {
- .d_revalidate = ll_fini_revalidate_nd,
- .d_release = ll_release,
-};
-
-/*
- * It is for the following race condition:
- * When someone (maybe statahead thread) adds the dentry to the dentry hash
- * table, the dentry's "d_op" maybe NULL, at the same time, another (maybe
- * "ls -l") process finds such dentry by "do_lookup()" without "do_revalidate()"
- * called. It causes statahead window lost, and maybe other issues. --Fan Yong
- */
-static int ll_init_revalidate_nd(struct dentry *dentry, struct nameidata *nd)
-{
- struct l_wait_info lwi = { 0 };
- struct ll_dentry_data *lld;
- ENTRY;
-
- ll_set_dd(dentry);
- lld = ll_d2d(dentry);
- if (unlikely(lld == NULL))
- RETURN(-ENOMEM);
-
- l_wait_event(lld->lld_waitq, dentry->d_op != &ll_init_d_ops, &lwi);
- if (likely(dentry->d_op == &ll_d_ops))
- RETURN(ll_revalidate_nd(dentry, nd));
- else
- RETURN(dentry->d_op == &ll_fini_d_ops ? 0 : -EINVAL);
-}
-
-struct dentry_operations ll_init_d_ops = {
- .d_revalidate = ll_init_revalidate_nd,
- .d_release = ll_release,
-};
struct ll_inode_info *lli = ll_i2info(inode);
struct lov_stripe_md *lsm = lli->lli_smd;
int rc;
-
ENTRY;
+
CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p)\n", inode->i_ino,
inode->i_generation, inode);
/* The last ref on @file, maybe not the the owner pid of statahead.
* Different processes can open the same dir, "ll_opendir_key" means:
* it is me that should stop the statahead thread. */
- if (lli->lli_opendir_key == fd)
+ if (lli->lli_opendir_key == fd && lli->lli_opendir_pid != 0)
ll_stop_statahead(inode, fd);
if (inode->i_sb->s_root == file->f_dentry) {
RETURN(-ENOMEM);
if (S_ISDIR(inode->i_mode)) {
+again:
spin_lock(&lli->lli_lock);
- /* "lli->lli_opendir_pid != 0" means someone has set it.
- * "lli->lli_sai != NULL" means the previous statahead has not
- * been cleanup. */
- if (lli->lli_opendir_pid == 0 && lli->lli_sai == NULL) {
- opendir_set = 1;
- lli->lli_opendir_pid = cfs_curproc_pid();
+ if (lli->lli_opendir_key == NULL && lli->lli_opendir_pid == 0) {
+ LASSERT(lli->lli_sai == NULL);
lli->lli_opendir_key = fd;
- } else if (unlikely(lli->lli_opendir_pid == cfs_curproc_pid())) {
+ lli->lli_opendir_pid = cfs_curproc_pid();
+ opendir_set = 1;
+ } else if (unlikely(lli->lli_opendir_pid == cfs_curproc_pid() &&
+ lli->lli_opendir_key != NULL)) {
/* Two cases for this:
* (1) The same process open such directory many times.
* (2) The old process opened the directory, and exited
* before its children processes. Then new process
* with the same pid opens such directory before the
* old process's children processes exit.
- * Change the owner to the latest one. */
- opendir_set = 2;
- lli->lli_opendir_key = fd;
+ * reset stat ahead for such cases. */
+ spin_unlock(&lli->lli_lock);
+ CDEBUG(D_INFO, "Conflict statahead for %.*s "DFID
+ " reset it.\n", file->f_dentry->d_name.len,
+ file->f_dentry->d_name.name,
+ PFID(&lli->lli_fid));
+ ll_stop_statahead(inode, lli->lli_opendir_key);
+ goto again;
}
spin_unlock(&lli->lli_lock);
}
}
up(&lli->lli_och_sem);
out_openerr:
- if (opendir_set == 1) {
- lli->lli_opendir_key = NULL;
- lli->lli_opendir_pid = 0;
- } else if (unlikely(opendir_set == 2)) {
+ if (opendir_set != 0)
ll_stop_statahead(inode, fd);
- }
}
return rc;
struct lookup_intent *lld_it;
#endif
unsigned int lld_sa_generation;
- cfs_waitq_t lld_waitq;
};
#define ll_d2d(de) ((struct ll_dentry_data*)((de)->d_fsdata))
* protect race ll_find_aliases vs ll_revalidate_it vs ll_unhash_aliases
*/
extern spinlock_t ll_lookup_lock;
-extern struct dentry_operations ll_init_d_ops;
extern struct dentry_operations ll_d_ops;
-extern struct dentry_operations ll_fini_d_ops;
-void ll_release(struct dentry *de);
void ll_intent_drop_lock(struct lookup_intent *);
void ll_intent_release(struct lookup_intent *);
int ll_drop_dentry(struct dentry *dentry);
/* per inode struct, for dir only */
struct ll_statahead_info {
struct inode *sai_inode;
- struct dentry *sai_first; /* first dentry item */
unsigned int sai_generation; /* generation for statahead */
atomic_t sai_refcount; /* when access this struct, hold
* refcount */
void ll_stop_statahead(struct inode *inode, void *key);
static inline
-void ll_d_wakeup(struct dentry *dentry)
+int ll_statahead_mark(struct dentry *dentry)
{
- struct ll_dentry_data *lld = ll_d2d(dentry);
+ struct ll_inode_info *lli = ll_i2info(dentry->d_parent->d_inode);
+ struct ll_statahead_info *sai = lli->lli_sai;
+ struct ll_dentry_data *ldd = ll_d2d(dentry);
+ int rc = 0;
- LASSERT(dentry->d_op != &ll_init_d_ops);
- if (lld != NULL)
- cfs_waitq_broadcast(&lld->lld_waitq);
+ if (likely(ldd != NULL))
+ ldd->lld_sa_generation = sai->sai_generation;
+ else
+ rc = -ENOMEM;
+
+ return rc;
}
static inline
return do_statahead_enter(dir, dentryp, lookup);
}
+static void inline ll_dops_init(struct dentry *de, int block)
+{
+ struct ll_dentry_data *lld = ll_d2d(de);
+
+ if (lld == NULL && block != 0) {
+ ll_set_dd(de);
+ lld = ll_d2d(de);
+ }
+
+ if (lld != NULL)
+ lld->lld_sa_generation = 0;
+
+ de->d_op = &ll_d_ops;
+}
+
/* llite ioctl register support rountine */
#ifdef __KERNEL__
enum llioc_iter {
RETURN(ERR_PTR(-ENOMEM));
}
- ll_set_dd(result);
-
- lock_dentry(result);
- if (unlikely(result->d_op == &ll_init_d_ops)) {
- result->d_op = &ll_d_ops;
- unlock_dentry(result);
- smp_wmb();
- ll_d_wakeup(result);
- } else {
- result->d_op = &ll_d_ops;
- unlock_dentry(result);
- }
+ ll_dops_init(result, 1);
RETURN(result);
}
struct list_head *tmp;
struct dentry *dentry;
struct dentry *last_discon = NULL;
-
+
spin_lock(&ll_lookup_lock);
spin_lock(&dcache_lock);
list_for_each(tmp, &inode->i_dentry) {
dentry->d_flags &= ~DCACHE_LUSTRE_INVALID;
#endif
unlock_dentry(dentry);
+ ll_dops_init(dentry, 0);
d_rehash_cond(dentry, 0); /* avoid taking dcache_lock inside */
spin_unlock(&dcache_lock);
spin_unlock(&ll_lookup_lock);
dget_locked(last_discon);
spin_unlock(&dcache_lock);
spin_unlock(&ll_lookup_lock);
+ ll_dops_init(last_discon, 1);
d_rehash(de);
d_move(last_discon, de);
iput(inode);
return de;
}
-static inline void ll_dop_init(struct dentry *de, int *set)
-{
- lock_dentry(de);
- if (likely(de->d_op != &ll_d_ops)) {
- de->d_op = &ll_init_d_ops;
- *set = 1;
- }
- unlock_dentry(de);
-}
-
-static inline void ll_dop_fini(struct dentry *de, int succ)
-{
- lock_dentry(de);
- if (likely(de->d_op == &ll_init_d_ops)) {
- if (succ)
- de->d_op = &ll_d_ops;
- else
- de->d_op = &ll_fini_d_ops;
- unlock_dentry(de);
- smp_wmb();
- ll_d_wakeup(de);
- } else {
- if (succ)
- de->d_op = &ll_d_ops;
- unlock_dentry(de);
- }
-}
-
int ll_lookup_it_finish(struct ptlrpc_request *request,
struct lookup_intent *it, void *data)
{
struct inode *parent = icbd->icbd_parent;
struct ll_sb_info *sbi = ll_i2sbi(parent);
struct inode *inode = NULL;
- int set = 0, rc;
+ int rc;
ENTRY;
- ll_dop_init(*de, &set);
-
/* NB 1 request reference will be taken away by ll_intent_lock()
* when I return */
if (!it_disposition(it, DISP_LOOKUP_NEG)) {
struct dentry *save = *de;
rc = ll_prep_inode(&inode, request, (*de)->d_sb);
- if (rc) {
- if (set)
- ll_dop_fini(*de, 0);
+ if (rc)
RETURN(rc);
- }
CDEBUG(D_DLMTRACE, "setting l_data to inode %p (%lu/%u)\n",
inode, inode->i_ino, inode->i_generation);
ll_glimpse_size or some equivalent themselves anyway.
Also see bug 7198. */
+ ll_dops_init(*de, 1);
*de = ll_find_alias(inode, *de);
- if (set && *de != save)
- ll_dop_fini(save, 0);
+ if (*de != save) {
+ struct ll_dentry_data *lld = ll_d2d(*de);
+
+ /* just make sure the ll_dentry_data is ready */
+ if (unlikely(lld == NULL)) {
+ ll_set_dd(*de);
+ lld = ll_d2d(*de);
+ if (likely(lld != NULL))
+ lld->lld_sa_generation = 0;
+ }
+ }
} else {
+ ll_dops_init(*de, 1);
/* Check that parent has UPDATE lock. If there is none, we
cannot afford to hash this dentry (done by ll_d_add) as it
might get picked up later when UPDATE lock will appear */
}
}
- ll_set_dd(*de);
-
- ll_dop_fini(*de, 1);
-
RETURN(0);
}
struct md_op_data *op_data;
struct it_cb_data icbd;
__u32 opc;
- int rc;
+ int rc, first = 0;
ENTRY;
if (dentry->d_name.len > ll_i2sbi(parent)->ll_namelen)
}
if (it->it_op == IT_GETATTR) {
- rc = ll_statahead_enter(parent, &dentry, 1);
- if (rc >= 0) {
- ll_statahead_exit(dentry, rc);
- if (rc == 1)
+ first = ll_statahead_enter(parent, &dentry, 1);
+ if (first >= 0) {
+ ll_statahead_exit(dentry, first);
+ if (first == 1)
RETURN(retval = dentry);
}
}
GOTO(out, retval = ERR_PTR(rc));
}
+ if (first == -EEXIST)
+ ll_statahead_mark(dentry);
+
if ((it->it_op & IT_OPEN) && dentry->d_inode &&
!S_ISREG(dentry->d_inode->i_mode) &&
!S_ISDIR(dentry->d_inode->i_mode)) {
SA_ENTRY_STATED
};
-struct dentry_operations ll_sai_d_ops = {
- .d_release = ll_release,
-};
-
static unsigned int sai_generation = 0;
static spinlock_t sai_generation_lock = SPIN_LOCK_UNLOCKED;
static inline int sa_not_full(struct ll_statahead_info *sai)
{
- return sai->sai_index < sai->sai_hit + sai->sai_miss + sai->sai_max;
+ return (sai->sai_index < sai->sai_hit + sai->sai_miss + sai->sai_max);
}
static inline int sa_is_running(struct ll_statahead_info *sai)
* (3) drop dentry's ref count
* (4) release request's ref count
*/
-static void ll_sai_entry_cleanup(struct ll_sai_entry *entry)
+static void ll_sai_entry_cleanup(struct ll_sai_entry *entry, int free)
{
struct ptlrpc_request *req = entry->se_req;
struct md_enqueue_info *minfo = entry->se_minfo;
entry->se_req = NULL;
ptlrpc_req_finished(req);
}
- OBD_FREE_PTR(entry);
+ if (free) {
+ LASSERT(list_empty(&entry->se_list));
+ OBD_FREE_PTR(entry);
+ }
EXIT;
}
static void ll_sai_put(struct ll_statahead_info *sai)
{
struct inode *inode = sai->sai_inode;
- struct ll_inode_info *lli = ll_i2info(inode);
+ struct ll_inode_info *lli;
ENTRY;
+ LASSERT(inode != NULL);
+ lli = ll_i2info(inode);
if (atomic_dec_and_lock(&sai->sai_refcount, &lli->lli_lock)) {
struct ll_sai_entry *entry, *next;
+ LASSERT(lli->lli_opendir_key == NULL);
lli->lli_sai = NULL;
+ lli->lli_opendir_pid = 0;
spin_unlock(&lli->lli_lock);
LASSERT(sa_is_stopped(sai));
list_for_each_entry_safe(entry, next, &sai->sai_entries_sent,
se_list) {
- list_del(&entry->se_list);
- ll_sai_entry_cleanup(entry);
+ list_del_init(&entry->se_list);
+ ll_sai_entry_cleanup(entry, 1);
}
list_for_each_entry_safe(entry, next, &sai->sai_entries_received,
se_list) {
- list_del(&entry->se_list);
- ll_sai_entry_cleanup(entry);
+ list_del_init(&entry->se_list);
+ ll_sai_entry_cleanup(entry, 1);
}
list_for_each_entry_safe(entry, next, &sai->sai_entries_stated,
se_list) {
- list_del(&entry->se_list);
- ll_sai_entry_cleanup(entry);
+ list_del_init(&entry->se_list);
+ ll_sai_entry_cleanup(entry, 1);
}
- dput(sai->sai_first);
OBD_FREE_PTR(sai);
iput(inode);
}
ENTRY;
if (!list_empty(&sai->sai_entries_sent)) {
- list_for_each_entry(entry, &sai->sai_entries_sent,
- se_list) {
+ list_for_each_entry(entry, &sai->sai_entries_sent, se_list) {
if (entry->se_index == index) {
entry->se_stat = stat;
entry->se_req = ptlrpc_request_addref(req);
struct ll_sai_entry *se;
ENTRY;
+ ll_sai_entry_cleanup(entry, 0);
+
spin_lock(&lli->lli_lock);
if (!list_empty(&entry->se_list))
list_del_init(&entry->se_list);
if (unlikely(entry->se_index < sai->sai_index_next)) {
spin_unlock(&lli->lli_lock);
- ll_sai_entry_cleanup(entry);
+ OBD_FREE_PTR(entry);
RETURN(0);
}
spin_unlock(&lli->lli_lock);
if (unlikely(entry->se_index < sai->sai_index_next)) {
- ll_sai_entry_cleanup(entry);
+ ll_sai_entry_cleanup(entry, 1);
RETURN(0);
}
+ if (entry->se_stat != SA_ENTRY_STATED)
+ GOTO(out, rc = entry->se_stat);
+
req = entry->se_req;
minfo = entry->se_minfo;
dentry = minfo->mi_dentry;
it = &minfo->mi_it;
- if (entry->se_stat != SA_ENTRY_STATED)
- GOTO(out, rc = entry->se_stat);
-
body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
if (body == NULL)
GOTO(out, rc = -EFAULT);
GOTO(out, rc);
}
+ spin_lock(&ll_lookup_lock);
spin_lock(&dcache_lock);
lock_dentry(dentry);
__d_drop(dentry);
unlock_dentry(dentry);
d_rehash_cond(dentry, 0);
spin_unlock(&dcache_lock);
+ spin_unlock(&ll_lookup_lock);
ll_lookup_finish_locks(it, dentry);
}
EXIT;
out:
- if (likely(ll_sai_entry_to_stated(sai, entry))) {
- entry->se_minfo = NULL;
- entry->se_req = NULL;
+ if (likely(ll_sai_entry_to_stated(sai, entry)))
cfs_waitq_signal(&sai->sai_waitq);
- ll_intent_release(it);
- OBD_FREE_PTR(minfo);
- dput(dentry);
- ptlrpc_req_finished(req);
- }
return rc;
}
OBD_FREE_PTR(minfo);
RETURN(-ESTALE);
} else {
- sai = lli->lli_sai;
+ sai = ll_sai_get(lli->lli_sai);
if (rc || dir == NULL)
rc = -ESTALE;
list_del_init(&entry->se_list);
sai->sai_replied++;
spin_unlock(&lli->lli_lock);
- ll_sai_entry_cleanup(entry);
+ ll_sai_entry_cleanup(entry, 1);
}
+ ll_sai_put(sai);
RETURN(rc);
}
}
RETURN(rc);
}
-static inline void ll_name2qstr(struct qstr *this, const char *name, int namelen)
+static inline void ll_name2qstr(struct qstr *q, const char *name, int namelen)
{
- unsigned long hash = init_name_hash();
- unsigned int c;
-
- this->name = name;
- this->len = namelen;
- for (; namelen > 0; namelen--, name++) {
- c = *(const unsigned char *)name;
- hash = partial_name_hash(c, hash);
- }
- this->hash = end_name_hash(hash);
+ q->name = name;
+ q->len = namelen;
+ q->hash = full_name_hash(name, namelen);
}
static int ll_statahead_one(struct dentry *parent, const char* entry_name,
void ll_stop_statahead(struct inode *inode, void *key)
{
struct ll_inode_info *lli = ll_i2info(inode);
- struct ptlrpc_thread *thread;
+
+ if (unlikely(key == NULL))
+ return;
spin_lock(&lli->lli_lock);
- if (lli->lli_opendir_pid == 0 ||
- unlikely(lli->lli_opendir_key != key)) {
+ if (lli->lli_opendir_key != key || lli->lli_opendir_pid == 0) {
spin_unlock(&lli->lli_lock);
return;
}
lli->lli_opendir_key = NULL;
- lli->lli_opendir_pid = 0;
if (lli->lli_sai) {
struct l_wait_info lwi = { 0 };
+ struct ptlrpc_thread *thread = &lli->lli_sai->sai_thread;
- thread = &lli->lli_sai->sai_thread;
if (!sa_is_stopped(lli->lli_sai)) {
thread->t_flags = SVC_STOPPING;
spin_unlock(&lli->lli_lock);
* maybe inflight.
*/
ll_sai_put(lli->lli_sai);
- return;
+ } else {
+ lli->lli_opendir_pid = 0;
+ spin_unlock(&lli->lli_lock);
}
- spin_unlock(&lli->lli_lock);
}
enum {
}
if (target->len == namelen &&
- !strncmp(target->name, name, target->len))
+ memcmp(target->name, name, namelen) == 0)
rc = LS_FIRST_DE + dot_de;
else
rc = LS_NONE_FIRST_DE;
list_empty(&sai->sai_entries_stated)))
RETURN(-EBADFD);
- /*
- * skip the first dentry.
- */
- if (unlikely((*dentryp)->d_name.len ==
- sai->sai_first->d_name.len &&
- !strncmp((*dentryp)->d_name.name,
- sai->sai_first->d_name.name,
- sai->sai_first->d_name.len)))
- RETURN(-EEXIST);
-
if ((*dentryp)->d_name.name[0] == '.') {
if (likely(sai->sai_ls_all ||
sai->sai_miss_hidden >= sai->sai_skip_hidden)) {
if (sai == NULL)
RETURN(-ENOMEM);
- sai->sai_inode = igrab(dir);
- sai->sai_first = dget(*dentryp);
sai->sai_ls_all = (rc == LS_FIRST_DOT_DE);
+ sai->sai_inode = igrab(dir);
+ if (unlikely(sai->sai_inode == NULL)) {
+ CWARN("Do not start stat ahead on dying inode "DFID" .\n",
+ PFID(&lli->lli_fid));
+ OBD_FREE_PTR(sai);
+ RETURN(-ESTALE);
+ }
+
+ LASSERT(sai->sai_inode == (*dentryp)->d_parent->d_inode);
sta.sta_parent = (*dentryp)->d_parent;
sta.sta_pid = cfs_curproc_pid();
rc = cfs_kernel_thread(ll_statahead_thread, &sta, 0);
if (rc < 0) {
CERROR("can't start ll_sa thread, rc: %d\n", rc);
+ lli->lli_opendir_key = NULL;
sai->sai_thread.t_flags = SVC_STOPPED;
ll_sai_put(sai);
LASSERT(lli->lli_sai == NULL);
struct dentry *parent = dentry->d_parent;
struct ll_inode_info *lli = ll_i2info(parent->d_inode);
struct ll_sb_info *sbi = ll_i2sbi(parent->d_inode);
- struct ll_dentry_data *ldd = ll_d2d(dentry);
+ int rc = 0;
ENTRY;
if (lli->lli_opendir_pid != cfs_curproc_pid())
if (!sa_is_stopped(sai))
cfs_waitq_signal(&sai->sai_thread.t_ctl_waitq);
ll_sai_entry_fini(sai);
-
- if (unlikely(ldd == NULL)) {
- ll_set_dd(dentry);
- ldd = ll_d2d(dentry);
- if (ldd != NULL && dentry->d_op == NULL) {
- lock_dentry(dentry);
- dentry->d_op = dentry->d_op ? : &ll_sai_d_ops;
- unlock_dentry(dentry);
- }
- }
-
- if (likely(ldd != NULL))
- ldd->lld_sa_generation = sai->sai_generation;
- else
- RETURN(-ENOMEM);
+ rc = ll_statahead_mark(dentry);
}
- RETURN(0);
+ RETURN(rc);
}