* GPL HEADER END
*/
/*
- * Copyright 2008 Sun Microsystems, Inc. All rights reserved
+ * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
* Use is subject to license terms.
+ *
+ * Copyright (c) 2011 Whamcloud, Inc.
+ *
*/
/*
* This file is part of Lustre, http://www.lustre.org/
int se_stat;
struct ptlrpc_request *se_req;
struct md_enqueue_info *se_minfo;
- struct dentry *se_dentry;
- struct inode *se_inode;
};
enum {
/**
* Check whether first entry was stated already or not.
- * No need to hold lli_lock, for:
+ * No need to hold lli_sa_lock, for:
* (1) it is me that remove entry from the list
* (2) the statahead thread only add new entry to the list
*/
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_index_next + sai->sai_max);
}
static inline int sa_is_running(struct ll_statahead_info *sai)
(sai->sai_consecutive_miss > 8));
}
-static void ll_sai_entry_free(struct ll_sai_entry *entry)
-{
- struct dentry *dentry = entry->se_dentry;
- struct inode *inode = entry->se_inode;
-
- if (dentry) {
- entry->se_dentry = NULL;
- dput(dentry);
- }
- if (inode) {
- entry->se_inode = NULL;
- iput(inode);
- }
- LASSERT(cfs_list_empty(&entry->se_list));
- OBD_FREE_PTR(entry);
-}
-
/**
* process the deleted entry's member and free the entry.
* (1) release intent
entry->se_req = NULL;
ptlrpc_req_finished(req);
}
- if (free)
- ll_sai_entry_free(entry);
+ if (free) {
+ LASSERT(cfs_list_empty(&entry->se_list));
+ OBD_FREE_PTR(entry);
+ }
EXIT;
}
lli = ll_i2info(inode);
LASSERT(lli->lli_sai == sai);
- if (cfs_atomic_dec_and_test(&sai->sai_refcount)) {
+ if (cfs_atomic_dec_and_lock(&sai->sai_refcount, &lli->lli_sa_lock)) {
struct ll_sai_entry *entry, *next;
- cfs_spin_lock(&lli->lli_lock);
if (unlikely(cfs_atomic_read(&sai->sai_refcount) > 0)) {
/* It is race case, the interpret callback just hold
* a reference count */
- cfs_spin_unlock(&lli->lli_lock);
- EXIT;
- return;
+ cfs_spin_unlock(&lli->lli_sa_lock);
+ RETURN_EXIT;
}
LASSERT(lli->lli_opendir_key == NULL);
lli->lli_sai = NULL;
lli->lli_opendir_pid = 0;
- cfs_spin_unlock(&lli->lli_lock);
+ cfs_spin_unlock(&lli->lli_sa_lock);
LASSERT(sa_is_stopped(sai));
entry->se_index = index;
entry->se_stat = SA_ENTRY_UNSTATED;
- cfs_spin_lock(&lli->lli_lock);
+ cfs_spin_lock(&lli->lli_sa_lock);
cfs_list_add_tail(&entry->se_list, &sai->sai_entries_sent);
- cfs_spin_unlock(&lli->lli_lock);
+ cfs_spin_unlock(&lli->lli_sa_lock);
RETURN(entry);
}
int rc = 0;
ENTRY;
- cfs_spin_lock(&lli->lli_lock);
+ cfs_spin_lock(&lli->lli_sa_lock);
sai->sai_index_next++;
if (likely(!cfs_list_empty(&sai->sai_entries_stated))) {
entry = cfs_list_entry(sai->sai_entries_stated.next,
if (entry->se_index < sai->sai_index_next) {
cfs_list_del_init(&entry->se_list);
rc = entry->se_stat;
- ll_sai_entry_free(entry);
+ OBD_FREE_PTR(entry);
}
} else {
LASSERT(sa_is_stopped(sai));
}
- cfs_spin_unlock(&lli->lli_lock);
+ cfs_spin_unlock(&lli->lli_sa_lock);
RETURN(rc);
}
/**
- * inside lli_lock.
+ * inside lli_sa_lock.
* \retval NULL : can not find the entry in sai_entries_sent with the index
* \retval entry: find the entry in sai_entries_sent with the index
*/
}
/**
- * inside lli_lock.
+ * inside lli_sa_lock.
* Move entry to sai_entries_received and
* insert it into sai_entries_received tail.
*/
ll_sai_entry_cleanup(entry, 0);
- cfs_spin_lock(&lli->lli_lock);
+ cfs_spin_lock(&lli->lli_sa_lock);
if (!cfs_list_empty(&entry->se_list))
cfs_list_del_init(&entry->se_list);
/* stale entry */
if (unlikely(entry->se_index < sai->sai_index_next)) {
- cfs_spin_unlock(&lli->lli_lock);
- ll_sai_entry_free(entry);
+ cfs_spin_unlock(&lli->lli_sa_lock);
+ OBD_FREE_PTR(entry);
RETURN(0);
}
cfs_list_for_each_entry_reverse(se, &sai->sai_entries_stated, se_list) {
if (se->se_index < entry->se_index) {
cfs_list_add(&entry->se_list, &se->se_list);
- cfs_spin_unlock(&lli->lli_lock);
+ cfs_spin_unlock(&lli->lli_sa_lock);
RETURN(1);
}
}
* I am the first entry.
*/
cfs_list_add(&entry->se_list, &sai->sai_entries_stated);
- cfs_spin_unlock(&lli->lli_lock);
+ cfs_spin_unlock(&lli->lli_sa_lock);
RETURN(1);
}
struct mdt_body *body;
ENTRY;
- cfs_spin_lock(&lli->lli_lock);
+ cfs_spin_lock(&lli->lli_sa_lock);
LASSERT(!sa_received_empty(sai));
entry = cfs_list_entry(sai->sai_entries_received.next,
struct ll_sai_entry, se_list);
cfs_list_del_init(&entry->se_list);
- cfs_spin_unlock(&lli->lli_lock);
+ cfs_spin_unlock(&lli->lli_sa_lock);
if (unlikely(entry->se_index < sai->sai_index_next)) {
CWARN("Found stale entry: [index %u] [next %u]\n",
/* Here dentry->d_inode might be NULL, because the entry may
* have been removed before we start doing stat ahead. */
+ rc = ll_lookup_it_finish(req, it, &icbd);
+ if (!rc)
+ ll_lookup_finish_locks(it, dentry);
- /* BUG 15962, 21739: since statahead thread does not hold
- * parent's i_mutex, it can not alias the dentry to inode.
- * Here we just create/update inode in memory, and let the
- * main "ls -l" thread to alias such dentry to the inode with
- * parent's i_mutex held.
- * On the other hand, we hold ldlm ibits lock for the inode
- * yet, to allow other operations to cancel such lock in time,
- * we should drop the ldlm lock reference count, then the main
- * "ls -l" thread should check/get such ldlm ibits lock before
- * aliasing such dentry to the inode later. If we don't do such
- * drop here, it maybe cause deadlock with i_muext held by
- * others, just like bug 21739. */
- rc = ll_lookup_it_finish(req, it, &icbd, &entry->se_inode);
- if (entry->se_inode != NULL)
- entry->se_dentry = dget(dentry);
- LASSERT(dentry == save);
- ll_intent_drop_lock(it);
+ if (dentry != save) {
+ minfo->mi_dentry = dentry;
+ dput(save);
+ }
} else {
/*
* revalidate.
CDEBUG(D_READA, "interpret statahead %.*s rc %d\n",
dentry->d_name.len, dentry->d_name.name, rc);
- cfs_spin_lock(&lli->lli_lock);
+ cfs_spin_lock(&lli->lli_sa_lock);
/* stale entry */
if (unlikely(lli->lli_sai == NULL ||
lli->lli_sai->sai_generation != minfo->mi_generation)) {
- cfs_spin_unlock(&lli->lli_lock);
+ cfs_spin_unlock(&lli->lli_sa_lock);
ll_intent_release(it);
dput(dentry);
iput(dir);
if (likely(sa_is_running(sai))) {
ll_sai_entry_to_received(sai, entry);
sai->sai_replied++;
- cfs_spin_unlock(&lli->lli_lock);
+ cfs_spin_unlock(&lli->lli_sa_lock);
cfs_waitq_signal(&sai->sai_thread.t_ctl_waitq);
} else {
if (!cfs_list_empty(&entry->se_list))
cfs_list_del_init(&entry->se_list);
sai->sai_replied++;
- cfs_spin_unlock(&lli->lli_lock);
+ cfs_spin_unlock(&lli->lli_sa_lock);
ll_sai_entry_cleanup(entry, 1);
}
ll_sai_put(sai);
if (unlikely(dentry == dentry->d_sb->s_root))
RETURN(1);
- rc = md_revalidate_lock(ll_i2mdexp(dir), &it, ll_inode2fid(inode),
- NULL);
+ rc = md_revalidate_lock(ll_i2mdexp(dir), &it, ll_inode2fid(inode));
if (rc == 1) {
ll_intent_release(&it);
RETURN(1);
struct ll_inode_info *lli = ll_i2info(dir);
struct ll_statahead_info *sai = lli->lli_sai;
struct qstr name;
- struct dentry *dentry;
+ struct dentry *dentry = NULL;
struct ll_sai_entry *se;
int rc;
ENTRY;
dentry = d_lookup(parent, &name);
if (!dentry) {
dentry = d_alloc(parent, &name);
- if (dentry) {
+ if (dentry)
rc = do_sa_lookup(dir, dentry);
- if (rc)
- dput(dentry);
- } else {
+ else
GOTO(out, rc = -ENOMEM);
- }
} else {
rc = do_sa_revalidate(dir, dentry);
- if (rc)
- dput(dentry);
}
EXIT;
out:
if (rc) {
+ if (dentry != NULL)
+ dput(dentry);
+ se->se_stat = rc < 0 ? rc : SA_ENTRY_STATED;
CDEBUG(D_READA, "set sai entry %p index %u stat %d rc %d\n",
se, se->se_index, se->se_stat, rc);
- se->se_stat = rc < 0 ? rc : SA_ENTRY_STATED;
if (ll_sai_entry_to_stated(sai, se))
cfs_waitq_signal(&sai->sai_waitq);
} else {
}
atomic_inc(&sbi->ll_sa_total);
- cfs_spin_lock(&lli->lli_lock);
+ cfs_spin_lock(&lli->lli_sa_lock);
thread->t_flags = SVC_RUNNING;
- cfs_spin_unlock(&lli->lli_lock);
+ cfs_spin_unlock(&lli->lli_sa_lock);
cfs_waitq_signal(&thread->t_ctl_waitq);
CDEBUG(D_READA, "start doing statahead for %s\n", parent->d_name.name);
+ sai->sai_pid = cfs_curproc_pid();
+ lli->lli_sa_pos = 0;
ll_dir_chain_init(&chain);
- page = ll_get_dir_page(dir, pos, 0, &chain);
+ page = ll_get_dir_page(NULL, dir, pos, 0, &chain);
while (1) {
struct l_wait_info lwi = { 0 };
dp = page_address(page);
for (ent = lu_dirent_start(dp); ent != NULL;
ent = lu_dirent_next(ent)) {
- char *name = ent->lde_name;
- int namelen = le16_to_cpu(ent->lde_namelen);
+ __u64 hash;
+ int namelen;
+ char *name;
+
+ hash = le64_to_cpu(ent->lde_hash);
+ if (unlikely(hash < pos))
+ /*
+ * Skip until we find target hash value.
+ */
+ continue;
+ namelen = le16_to_cpu(ent->lde_namelen);
if (unlikely(namelen == 0))
/*
* Skip dummy record.
*/
continue;
+ name = ent->lde_name;
if (name[0] == '.') {
if (namelen == 1) {
/*
}
pos = le64_to_cpu(dp->ldp_hash_end);
ll_put_page(page);
- if (pos == DIR_END_OFF) {
+ if (pos == MDS_DIR_END_OFF) {
/*
* End of directory reached.
*/
* chain is exhausted.
* Normal case: continue to the next page.
*/
- page = ll_get_dir_page(dir, pos, 1, &chain);
+ lli->lli_sa_pos = pos;
+ page = ll_get_dir_page(NULL, dir, pos, 1, &chain);
} else {
/*
* go into overflow page.
out:
ll_dir_chain_fini(&chain);
- cfs_spin_lock(&lli->lli_lock);
+ cfs_spin_lock(&lli->lli_sa_lock);
thread->t_flags = SVC_STOPPED;
- cfs_spin_unlock(&lli->lli_lock);
+ cfs_spin_unlock(&lli->lli_sa_lock);
cfs_waitq_signal(&sai->sai_waitq);
cfs_waitq_signal(&thread->t_ctl_waitq);
ll_sai_put(sai);
/**
* called in ll_file_release().
*/
-void ll_stop_statahead(struct inode *inode, void *key)
+void ll_stop_statahead(struct inode *dir, void *key)
{
- struct ll_inode_info *lli = ll_i2info(inode);
+ struct ll_inode_info *lli = ll_i2info(dir);
if (unlikely(key == NULL))
return;
- cfs_spin_lock(&lli->lli_lock);
+ cfs_spin_lock(&lli->lli_sa_lock);
if (lli->lli_opendir_key != key || lli->lli_opendir_pid == 0) {
- cfs_spin_unlock(&lli->lli_lock);
+ cfs_spin_unlock(&lli->lli_sa_lock);
return;
}
if (!sa_is_stopped(lli->lli_sai)) {
thread->t_flags = SVC_STOPPING;
- cfs_spin_unlock(&lli->lli_lock);
+ cfs_spin_unlock(&lli->lli_sa_lock);
cfs_waitq_signal(&thread->t_ctl_waitq);
CDEBUG(D_READA, "stopping statahead thread, pid %d\n",
sa_is_stopped(lli->lli_sai),
&lwi);
} else {
- cfs_spin_unlock(&lli->lli_lock);
+ cfs_spin_unlock(&lli->lli_sa_lock);
}
/*
ll_sai_put(lli->lli_sai);
} else {
lli->lli_opendir_pid = 0;
- cfs_spin_unlock(&lli->lli_lock);
+ cfs_spin_unlock(&lli->lli_sa_lock);
}
}
static int is_first_dirent(struct inode *dir, struct dentry *dentry)
{
+ struct ll_inode_info *lli = ll_i2info(dir);
struct ll_dir_chain chain;
struct qstr *target = &dentry->d_name;
struct page *page;
int rc = LS_NONE_FIRST_DE;
ENTRY;
+ lli->lli_sa_pos = 0;
ll_dir_chain_init(&chain);
- page = ll_get_dir_page(dir, pos, 0, &chain);
+ page = ll_get_dir_page(NULL, dir, pos, 0, &chain);
while (1) {
struct lu_dirpage *dp;
dp = page_address(page);
for (ent = lu_dirent_start(dp); ent != NULL;
ent = lu_dirent_next(ent)) {
- char *name = ent->lde_name;
- int namelen = le16_to_cpu(ent->lde_namelen);
+ int namelen;
+ char *name;
- if (namelen == 0)
+ namelen = le16_to_cpu(ent->lde_namelen);
+ if (unlikely(namelen == 0))
/*
* skip dummy record.
*/
continue;
+ name = ent->lde_name;
if (name[0] == '.') {
if (namelen == 1)
/*
}
pos = le64_to_cpu(dp->ldp_hash_end);
ll_put_page(page);
- if (pos == DIR_END_OFF) {
+ if (pos == MDS_DIR_END_OFF) {
/*
* End of directory reached.
*/
* chain is exhausted
* Normal case: continue to the next page.
*/
- page = ll_get_dir_page(dir, pos, 1, &chain);
+ lli->lli_sa_pos = pos;
+ page = ll_get_dir_page(NULL, dir, pos, 1, &chain);
} else {
/*
* go into overflow page.
return rc;
}
-static int is_same_dentry(struct dentry *d1, struct dentry *d2)
-{
- if (unlikely(d1 == d2))
- return 1;
- if (d1->d_name.len == d2->d_name.len &&
- memcmp(d1->d_name.name, d2->d_name.name, d1->d_name.len) == 0)
- return 1;
- return 0;
-}
-
/**
* Start statahead thread if this is the first dir entry.
* Otherwise if a thread is started already, wait it until it is ahead of me.
- * \retval 0 -- stat ahead thread process such dentry, miss for lookup
- * \retval 1 -- stat ahead thread process such dentry, hit for any case
+ * \retval 0 -- stat ahead thread process such dentry, for lookup, it miss
+ * \retval 1 -- stat ahead thread process such dentry, for lookup, it hit
* \retval -EEXIST -- stat ahead thread started, and this is the first dentry
* \retval -EBADFD -- statahead thread exit and not dentry available
* \retval -EAGAIN -- try to stat by caller
RETURN(rc);
}
- if (ll_sai_entry_stated(sai)) {
- struct ll_sai_entry *entry;
-
- entry = cfs_list_entry(sai->sai_entries_stated.next,
- struct ll_sai_entry, se_list);
- /* This is for statahead lookup */
- if (entry->se_inode != NULL) {
- struct lookup_intent it = {.it_op = IT_GETATTR};
- struct dentry *dchild = entry->se_dentry;
- struct inode *ichild = entry->se_inode;
- int found = 0;
- __u32 bits;
-
- LASSERT(dchild != *dentryp);
-
- if (!lookup)
- mutex_lock(&dir->i_mutex);
-
- rc = md_revalidate_lock(ll_i2mdexp(dir), &it,
- ll_inode2fid(ichild),
- &bits);
- if (rc == 1) {
- struct dentry *save = dchild;
-
- ll_lookup_it_alias(&dchild, ichild,
- bits);
- ll_lookup_finish_locks(&it, dchild);
- if (dchild != save)
- dput(save);
- found = is_same_dentry(dchild,
- *dentryp);
- } else {
- /* Someone has canceled related ldlm
- * lock before the real "revalidate"
- * using it.
- * Drop the inode reference count held
- * by interpreter. */
- iput(ichild);
- }
-
- if (!lookup)
- mutex_unlock(&dir->i_mutex);
-
- entry->se_dentry = NULL;
- entry->se_inode = NULL;
- if (found) {
- if (lookup) {
- LASSERT(dchild != *dentryp);
- /* VFS will drop the reference
- * count for dchild and *dentryp
- * by itself. */
- *dentryp = dchild;
- } else {
- LASSERT(dchild == *dentryp);
- /* Drop the dentry reference
- * count held by statahead. */
- dput(dchild);
- }
- RETURN(1);
- } else {
- /* Drop the dentry reference count held
- * by statahead. */
- dput(dchild);
- }
- }
- }
-
if (lookup) {
struct dentry *result;
}
lli->lli_sai = sai;
- rc = cfs_kernel_thread(ll_statahead_thread, parent, 0);
+ rc = cfs_create_thread(ll_statahead_thread, parent, 0);
if (rc < 0) {
CERROR("can't start ll_sa thread, rc: %d\n", rc);
dput(parent);
RETURN(-EEXIST);
out:
- cfs_spin_lock(&lli->lli_lock);
+ cfs_spin_lock(&lli->lli_sa_lock);
lli->lli_opendir_key = NULL;
lli->lli_opendir_pid = 0;
- cfs_spin_unlock(&lli->lli_lock);
+ cfs_spin_unlock(&lli->lli_sa_lock);
return rc;
}
PFID(&lli->lli_fid), sai->sai_hit,
sai->sai_miss, sai->sai_sent,
sai->sai_replied, cfs_curproc_pid());
- cfs_spin_lock(&lli->lli_lock);
+ cfs_spin_lock(&lli->lli_sa_lock);
if (!sa_is_stopped(sai))
sai->sai_thread.t_flags = SVC_STOPPING;
- cfs_spin_unlock(&lli->lli_lock);
+ cfs_spin_unlock(&lli->lli_sa_lock);
}
}