/* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
* vim:expandtab:shiftwidth=8:tabstop=8:
*
- * Copyright (c) 2007 Cluster File Systems, Inc.
+ * GPL HEADER START
*
- * This file is part of Lustre, http://www.lustre.org.
+ * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
*
- * Lustre is free software; you can redistribute it and/or
- * modify it under the terms of version 2 of the GNU General Public
- * License as published by the Free Software Foundation.
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License version 2 only,
+ * as published by the Free Software Foundation.
*
- * Lustre is distributed in the hope that it will be useful,
- * but WITHOUT ANY WARRANTY; without even the implied warranty of
- * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
- * GNU General Public License for more details.
+ * This program is distributed in the hope that it will be useful, but
+ * WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+ * General Public License version 2 for more details (a copy is included
+ * in the LICENSE file that accompanied this code).
*
- * You should have received a copy of the GNU General Public License
- * along with Lustre; if not, write to the Free Software
- * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
+ * You should have received a copy of the GNU General Public License
+ * version 2 along with this program; If not, see
+ * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
+ *
+ * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
+ * CA 95054 USA or visit www.sun.com if you need additional information or
+ * have any questions.
+ *
+ * GPL HEADER END
+ */
+/*
+ * Copyright 2008 Sun Microsystems, Inc. All rights reserved
+ * Use is subject to license terms.
+ */
+/*
+ * This file is part of Lustre, http://www.lustre.org/
+ * Lustre is a trademark of Sun Microsystems, Inc.
*/
#include <linux/fs.h>
#include "llite_internal.h"
struct ll_sai_entry {
- struct list_head se_list;
+ cfs_list_t se_list;
unsigned int se_index;
int se_stat;
+ struct ptlrpc_request *se_req;
+ struct md_enqueue_info *se_minfo;
+ struct dentry *se_dentry;
+ struct inode *se_inode;
};
enum {
};
static unsigned int sai_generation = 0;
-static spinlock_t sai_generation_lock = SPIN_LOCK_UNLOCKED;
+static cfs_spinlock_t sai_generation_lock = CFS_SPIN_LOCK_UNLOCKED;
+
+/**
+ * Check whether first entry was stated already or not.
+ * No need to hold lli_lock, for:
+ * (1) it is me that remove entry from the list
+ * (2) the statahead thread only add new entry to the list
+ */
+static int ll_sai_entry_stated(struct ll_statahead_info *sai)
+{
+ struct ll_sai_entry *entry;
+ int rc = 0;
+
+ if (!cfs_list_empty(&sai->sai_entries_stated)) {
+ entry = cfs_list_entry(sai->sai_entries_stated.next,
+ struct ll_sai_entry, se_list);
+ if (entry->se_index == sai->sai_index_next)
+ rc = 1;
+ }
+ return rc;
+}
+
+static inline int sa_received_empty(struct ll_statahead_info *sai)
+{
+ return cfs_list_empty(&sai->sai_entries_received);
+}
+
+static inline int sa_not_full(struct ll_statahead_info *sai)
+{
+ return (sai->sai_index < sai->sai_hit + sai->sai_miss + sai->sai_max);
+}
+
+static inline int sa_is_running(struct ll_statahead_info *sai)
+{
+ return !!(sai->sai_thread.t_flags & SVC_RUNNING);
+}
+
+static inline int sa_is_stopping(struct ll_statahead_info *sai)
+{
+ return !!(sai->sai_thread.t_flags & SVC_STOPPING);
+}
+
+static inline int sa_is_stopped(struct ll_statahead_info *sai)
+{
+ return !!(sai->sai_thread.t_flags & SVC_STOPPED);
+}
+
+/**
+ * (1) hit ratio less than 80%
+ * or
+ * (2) consecutive miss more than 8
+ */
+static inline int sa_low_hit(struct ll_statahead_info *sai)
+{
+ return ((sai->sai_hit > 7 && sai->sai_hit < 4 * sai->sai_miss) ||
+ (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
+ * (2) free md_enqueue_info
+ * (3) drop dentry's ref count
+ * (4) release request's ref count
+ */
+static void ll_sai_entry_cleanup(struct ll_sai_entry *entry, int free)
+{
+ struct md_enqueue_info *minfo = entry->se_minfo;
+ struct ptlrpc_request *req = entry->se_req;
+ ENTRY;
+
+ if (minfo) {
+ entry->se_minfo = NULL;
+ ll_intent_release(&minfo->mi_it);
+ dput(minfo->mi_dentry);
+ iput(minfo->mi_dir);
+ OBD_FREE_PTR(minfo);
+ }
+ if (req) {
+ entry->se_req = NULL;
+ ptlrpc_req_finished(req);
+ }
+ if (free)
+ ll_sai_entry_free(entry);
+
+ EXIT;
+}
static struct ll_statahead_info *ll_sai_alloc(void)
{
if (!sai)
return NULL;
- spin_lock(&sai_generation_lock);
+ cfs_spin_lock(&sai_generation_lock);
sai->sai_generation = ++sai_generation;
if (unlikely(sai_generation == 0))
sai->sai_generation = ++sai_generation;
- spin_unlock(&sai_generation_lock);
- atomic_set(&sai->sai_refcount, 1);
+ cfs_spin_unlock(&sai_generation_lock);
+ cfs_atomic_set(&sai->sai_refcount, 1);
sai->sai_max = LL_SA_RPC_MIN;
cfs_waitq_init(&sai->sai_waitq);
cfs_waitq_init(&sai->sai_thread.t_ctl_waitq);
- CFS_INIT_LIST_HEAD(&sai->sai_entries);
+ CFS_INIT_LIST_HEAD(&sai->sai_entries_sent);
+ CFS_INIT_LIST_HEAD(&sai->sai_entries_received);
+ CFS_INIT_LIST_HEAD(&sai->sai_entries_stated);
return sai;
}
-static inline
+static inline
struct ll_statahead_info *ll_sai_get(struct ll_statahead_info *sai)
{
LASSERT(sai);
- atomic_inc(&sai->sai_refcount);
+ cfs_atomic_inc(&sai->sai_refcount);
return sai;
}
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;
- if (atomic_dec_and_lock(&sai->sai_refcount, &lli->lli_lock)) {
+ LASSERT(inode != NULL);
+ lli = ll_i2info(inode);
+ LASSERT(lli->lli_sai == sai);
+
+ if (cfs_atomic_dec_and_test(&sai->sai_refcount)) {
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;
+ }
+
+ LASSERT(lli->lli_opendir_key == NULL);
lli->lli_sai = NULL;
- spin_unlock(&lli->lli_lock);
+ lli->lli_opendir_pid = 0;
+ cfs_spin_unlock(&lli->lli_lock);
- LASSERT(sai->sai_thread.t_flags & SVC_STOPPED);
+ LASSERT(sa_is_stopped(sai));
if (sai->sai_sent > sai->sai_replied)
CDEBUG(D_READA,"statahead for dir "DFID" does not "
PFID(&lli->lli_fid),
sai->sai_sent, sai->sai_replied);
- list_for_each_entry_safe(entry, next, &sai->sai_entries,
- se_list) {
- list_del(&entry->se_list);
- OBD_FREE_PTR(entry);
+ cfs_list_for_each_entry_safe(entry, next,
+ &sai->sai_entries_sent, se_list) {
+ cfs_list_del_init(&entry->se_list);
+ ll_sai_entry_cleanup(entry, 1);
+ }
+ cfs_list_for_each_entry_safe(entry, next,
+ &sai->sai_entries_received,
+ se_list) {
+ cfs_list_del_init(&entry->se_list);
+ ll_sai_entry_cleanup(entry, 1);
+ }
+ cfs_list_for_each_entry_safe(entry, next,
+ &sai->sai_entries_stated,
+ se_list) {
+ cfs_list_del_init(&entry->se_list);
+ ll_sai_entry_cleanup(entry, 1);
}
- OBD_FREE_PTR(sai);
iput(inode);
+ OBD_FREE_PTR(sai);
}
EXIT;
}
+/**
+ * insert it into sai_entries_sent tail when init.
+ */
static struct ll_sai_entry *
-ll_sai_entry_get(struct ll_statahead_info *sai, unsigned int index, int stat)
+ll_sai_entry_init(struct ll_statahead_info *sai, unsigned int index)
{
struct ll_inode_info *lli = ll_i2info(sai->sai_inode);
struct ll_sai_entry *entry;
if (entry == NULL)
RETURN(ERR_PTR(-ENOMEM));
- CDEBUG(D_READA, "alloc sai entry %p index %u, stat %d\n",
- entry, index, stat);
+ CDEBUG(D_READA, "alloc sai entry %p index %u\n",
+ entry, index);
entry->se_index = index;
- entry->se_stat = stat;
+ entry->se_stat = SA_ENTRY_UNSTATED;
- spin_lock(&lli->lli_lock);
- list_add_tail(&entry->se_list, &sai->sai_entries);
- spin_unlock(&lli->lli_lock);
+ cfs_spin_lock(&lli->lli_lock);
+ cfs_list_add_tail(&entry->se_list, &sai->sai_entries_sent);
+ cfs_spin_unlock(&lli->lli_lock);
RETURN(entry);
}
-/* inside lli_lock
- * return value:
- * 0: can not find the entry with the index
- * 1: it is the first entry
- * 2: it is not the first entry */
-static int
-ll_sai_entry_set(struct ll_statahead_info *sai, unsigned int index, int stat)
+/**
+ * delete it from sai_entries_stated head when fini, it need not
+ * to process entry's member.
+ */
+static int ll_sai_entry_fini(struct ll_statahead_info *sai)
{
- struct ll_sai_entry *entry;
- int rc = 0;
+ struct ll_inode_info *lli = ll_i2info(sai->sai_inode);
+ struct ll_sai_entry *entry;
+ int rc = 0;
ENTRY;
- if (list_empty(&sai->sai_entries))
- RETURN(0);
-
- entry = list_entry(sai->sai_entries.next, struct ll_sai_entry, se_list);
- if (entry->se_index == index)
- GOTO(out, rc = 1);
-
- while (entry->se_list.next != &sai->sai_entries &&
- entry->se_index < index) {
- entry = list_entry(entry->se_list.next, struct ll_sai_entry,
- se_list);
- if (entry->se_index == index)
- GOTO(out, rc = 2);
- }
-
- EXIT;
-
-out:
- if (rc) {
- LASSERT(entry->se_stat == SA_ENTRY_UNSTATED);
- entry->se_stat = stat;
+ cfs_spin_lock(&lli->lli_lock);
+ sai->sai_index_next++;
+ if (likely(!cfs_list_empty(&sai->sai_entries_stated))) {
+ entry = cfs_list_entry(sai->sai_entries_stated.next,
+ struct ll_sai_entry, se_list);
+ if (entry->se_index < sai->sai_index_next) {
+ cfs_list_del_init(&entry->se_list);
+ rc = entry->se_stat;
+ ll_sai_entry_free(entry);
+ }
+ } else {
+ LASSERT(sa_is_stopped(sai));
}
+ cfs_spin_unlock(&lli->lli_lock);
- return rc;
+ RETURN(rc);
}
-/* Check whether first entry was stated already or not.
- * No need to hold lli_lock, for:
- * (1) it is me that remove entry from the list (ll_sai_entry_put)
- * (2) the statahead thread only add new entry to the list tail */
-static int ll_sai_entry_stated(struct ll_statahead_info *sai)
+/**
+ * inside lli_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
+ */
+static struct ll_sai_entry *
+ll_sai_entry_set(struct ll_statahead_info *sai, unsigned int index, int stat,
+ struct ptlrpc_request *req, struct md_enqueue_info *minfo)
{
- struct ll_sai_entry *entry;
- int rc = 0;
+ struct ll_sai_entry *entry;
ENTRY;
- if (!list_empty(&sai->sai_entries)) {
- entry = list_entry(sai->sai_entries.next, struct ll_sai_entry,
- se_list);
- rc = (entry->se_stat != SA_ENTRY_UNSTATED);
+ if (!cfs_list_empty(&sai->sai_entries_sent)) {
+ cfs_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);
+ entry->se_minfo = minfo;
+ RETURN(entry);
+ } else if (entry->se_index > index) {
+ RETURN(NULL);
+ }
+ }
}
+ RETURN(NULL);
+}
- RETURN(rc);
+/**
+ * inside lli_lock.
+ * Move entry to sai_entries_received and
+ * insert it into sai_entries_received tail.
+ */
+static inline void
+ll_sai_entry_to_received(struct ll_statahead_info *sai, struct ll_sai_entry *entry)
+{
+ if (!cfs_list_empty(&entry->se_list))
+ cfs_list_del_init(&entry->se_list);
+ cfs_list_add_tail(&entry->se_list, &sai->sai_entries_received);
}
-static void ll_sai_entry_put(struct ll_statahead_info *sai)
+/**
+ * Move entry to sai_entries_stated and
+ * sort with the index.
+ */
+static int
+ll_sai_entry_to_stated(struct ll_statahead_info *sai, struct ll_sai_entry *entry)
{
struct ll_inode_info *lli = ll_i2info(sai->sai_inode);
- struct ll_sai_entry *entry;
+ struct ll_sai_entry *se;
ENTRY;
- spin_lock(&lli->lli_lock);
- if (!list_empty(&sai->sai_entries)) {
- entry = list_entry(sai->sai_entries.next,
- struct ll_sai_entry, se_list);
- list_del(&entry->se_list);
- OBD_FREE_PTR(entry);
+ ll_sai_entry_cleanup(entry, 0);
+
+ cfs_spin_lock(&lli->lli_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);
+ RETURN(0);
}
- spin_unlock(&lli->lli_lock);
- EXIT;
+ 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);
+ RETURN(1);
+ }
+ }
+
+ /*
+ * I am the first entry.
+ */
+ cfs_list_add(&entry->se_list, &sai->sai_entries_stated);
+ cfs_spin_unlock(&lli->lli_lock);
+ RETURN(1);
}
-/* finish lookup/revalidate */
-static int ll_statahead_interpret(struct ptlrpc_request *req,
- struct md_enqueue_info *minfo,
- int rc)
+/**
+ * finish lookup/revalidate.
+ */
+static int do_statahead_interpret(struct ll_statahead_info *sai)
{
- struct lookup_intent *it = &minfo->mi_it;
- struct dentry *dentry = minfo->mi_dentry;
- struct inode *dir = dentry->d_parent->d_inode;
- struct ll_inode_info *lli = ll_i2info(dir);
- struct ll_statahead_info *sai = NULL;
+ struct ll_inode_info *lli = ll_i2info(sai->sai_inode);
+ struct ll_sai_entry *entry;
+ struct ptlrpc_request *req;
+ struct md_enqueue_info *minfo;
+ struct lookup_intent *it;
+ struct dentry *dentry;
+ int rc = 0;
+ struct mdt_body *body;
ENTRY;
- CDEBUG(D_READA, "interpret statahead %.*s rc %d\n",
- dentry->d_name.len, dentry->d_name.name, rc);
-
- spin_lock(&lli->lli_lock);
- if (unlikely(lli->lli_sai == NULL ||
- lli->lli_sai->sai_generation != minfo->mi_generation)) {
- spin_unlock(&lli->lli_lock);
- GOTO(out_free, rc = -ESTALE);
- } else {
- sai = ll_sai_get(lli->lli_sai);
- spin_unlock(&lli->lli_lock);
+ cfs_spin_lock(&lli->lli_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);
+
+ if (unlikely(entry->se_index < sai->sai_index_next)) {
+ CWARN("Found stale entry: [index %u] [next %u]\n",
+ entry->se_index, sai->sai_index_next);
+ ll_sai_entry_cleanup(entry, 1);
+ RETURN(0);
}
- if (rc || dir == NULL)
- GOTO(out, rc);
+ if (entry->se_stat != SA_ENTRY_STATED)
+ GOTO(out, rc = entry->se_stat);
+
+ req = entry->se_req;
+ minfo = entry->se_minfo;
+ it = &minfo->mi_it;
+ dentry = minfo->mi_dentry;
+
+ body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
+ if (body == NULL)
+ GOTO(out, rc = -EFAULT);
if (dentry->d_inode == NULL) {
- /* lookup */
+ /*
+ * lookup.
+ */
struct dentry *save = dentry;
struct it_cb_data icbd = {
- .icbd_parent = dir,
+ .icbd_parent = minfo->mi_dir,
.icbd_childp = &dentry
};
LASSERT(fid_is_zero(&minfo->mi_data.op_fid2));
- rc = ll_lookup_it_finish(req, it, &icbd);
- if (!rc)
- /* Here dentry->d_inode might be NULL,
- * because the entry may have been removed before
- * we start doing stat ahead. */
- ll_lookup_finish_locks(it, dentry);
+ /* XXX: No fid in reply, this is probaly cross-ref case.
+ * SA can't handle it yet. */
+ if (body->valid & OBD_MD_MDS)
+ GOTO(out, rc = -EAGAIN);
- if (dentry != save)
- dput(save);
+ /* Here dentry->d_inode might be NULL, because the entry may
+ * have been removed before we start doing stat ahead. */
+
+ /* 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);
} else {
- /* revalidate */
- struct mdt_body *body;
-
- body = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF,
- sizeof(*body));
+ /*
+ * revalidate.
+ */
if (!lu_fid_eq(&minfo->mi_data.op_fid2, &body->fid1)) {
ll_unhash_aliases(dentry->d_inode);
GOTO(out, rc = -EAGAIN);
GOTO(out, rc);
}
+ cfs_spin_lock(&ll_lookup_lock);
spin_lock(&dcache_lock);
lock_dentry(dentry);
__d_drop(dentry);
-#ifdef DCACHE_LUSTRE_INVALID
dentry->d_flags &= ~DCACHE_LUSTRE_INVALID;
-#endif
unlock_dentry(dentry);
d_rehash_cond(dentry, 0);
spin_unlock(&dcache_lock);
+ cfs_spin_unlock(&ll_lookup_lock);
ll_lookup_finish_locks(it, dentry);
}
EXIT;
out:
- if (sai != NULL) {
- int first;
+ /* The "ll_sai_entry_to_stated()" will drop related ldlm ibits lock
+ * reference count with ll_intent_drop_lock() called in spite of the
+ * above operations failed or not. Do not worry about calling
+ * "ll_intent_drop_lock()" more than once. */
+ if (likely(ll_sai_entry_to_stated(sai, entry)))
+ cfs_waitq_signal(&sai->sai_waitq);
+ return rc;
+}
- sai->sai_replied++;
- spin_lock(&lli->lli_lock);
- first = ll_sai_entry_set(sai,
- (unsigned int)(long)minfo->mi_cbdata,
- SA_ENTRY_STATED);
- spin_unlock(&lli->lli_lock);
- if (first == 1)
- /* wake up the "ls -l" process only when the first entry
- * returned. */
- cfs_waitq_signal(&sai->sai_waitq);
- else if (first == 0)
- CDEBUG(D_READA, "can't find sai entry for dir "
- DFID" generation %u index %u\n",
- PFID(&lli->lli_fid),
- minfo->mi_generation,
- (unsigned int)(long)minfo->mi_cbdata);
+static int ll_statahead_interpret(struct ptlrpc_request *req,
+ struct md_enqueue_info *minfo,
+ int rc)
+{
+ struct lookup_intent *it = &minfo->mi_it;
+ struct dentry *dentry = minfo->mi_dentry;
+ struct inode *dir = minfo->mi_dir;
+ struct ll_inode_info *lli = ll_i2info(dir);
+ struct ll_statahead_info *sai;
+ struct ll_sai_entry *entry;
+ ENTRY;
+
+ CDEBUG(D_READA, "interpret statahead %.*s rc %d\n",
+ dentry->d_name.len, dentry->d_name.name, rc);
+ cfs_spin_lock(&lli->lli_lock);
+ /* stale entry */
+ if (unlikely(lli->lli_sai == NULL ||
+ lli->lli_sai->sai_generation != minfo->mi_generation)) {
+ cfs_spin_unlock(&lli->lli_lock);
+ ll_intent_release(it);
+ dput(dentry);
+ iput(dir);
+ OBD_FREE_PTR(minfo);
+ RETURN(-ESTALE);
+ } else {
+ sai = ll_sai_get(lli->lli_sai);
+ entry = ll_sai_entry_set(sai,
+ (unsigned int)(long)minfo->mi_cbdata,
+ rc < 0 ? rc : SA_ENTRY_STATED, req,
+ minfo);
+ LASSERT(entry != NULL);
+ if (likely(sa_is_running(sai))) {
+ ll_sai_entry_to_received(sai, entry);
+ sai->sai_replied++;
+ cfs_spin_unlock(&lli->lli_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);
+ ll_sai_entry_cleanup(entry, 1);
+ }
ll_sai_put(sai);
+ RETURN(rc);
}
-out_free:
- ll_intent_release(it);
- OBD_FREE_PTR(minfo);
-
- dput(dentry);
- return rc;
}
static void sa_args_fini(struct md_enqueue_info *minfo,
struct ldlm_enqueue_info *einfo)
{
LASSERT(minfo && einfo);
+ iput(minfo->mi_dir);
capa_put(minfo->mi_data.op_capa1);
capa_put(minfo->mi_data.op_capa2);
OBD_FREE_PTR(minfo);
OBD_FREE_PTR(einfo);
}
-/* There is race condition between "capa_put" and "ll_statahead_interpret" for
+/**
+ * There is race condition between "capa_put" and "ll_statahead_interpret" for
* accessing "op_data.op_capa[1,2]" as following:
* "capa_put" releases "op_data.op_capa[1,2]"'s reference count after calling
* "md_intent_getattr_async". But "ll_statahead_interpret" maybe run first, and
* fill "op_data.op_capa[1,2]" as POISON, then cause "capa_put" access invalid
* "ocapa". So here reserve "op_data.op_capa[1,2]" in "pcapa" before calling
- * "md_intent_getattr_async". */
+ * "md_intent_getattr_async".
+ */
static int sa_args_init(struct inode *dir, struct dentry *dentry,
struct md_enqueue_info **pmi,
struct ldlm_enqueue_info **pei,
minfo->mi_it.it_op = IT_GETATTR;
minfo->mi_dentry = dentry;
+ minfo->mi_dir = igrab(dir);
minfo->mi_cb = ll_statahead_interpret;
minfo->mi_generation = lli->lli_sai->sai_generation;
minfo->mi_cbdata = (void *)(long)lli->lli_sai->sai_index;
return 0;
}
-/* similar to ll_lookup_it(). */
+/**
+ * similar to ll_lookup_it().
+ */
static int do_sa_lookup(struct inode *dir, struct dentry *dentry)
{
struct md_enqueue_info *minfo;
RETURN(rc);
}
-/* similar to ll_revalidate_it().
- * return value:
- * 1 -- dentry valid
- * 0 -- will send stat-ahead request
- * others -- prepare stat-ahead request failed */
-static int do_sa_revalidate(struct dentry *dentry)
+/**
+ * similar to ll_revalidate_it().
+ * \retval 1 -- dentry valid
+ * \retval 0 -- will send stat-ahead request
+ * \retval others -- prepare stat-ahead request failed
+ */
+static int do_sa_revalidate(struct inode *dir, struct dentry *dentry)
{
struct inode *inode = dentry->d_inode;
- struct inode *dir = dentry->d_parent->d_inode;
struct lookup_intent it = { .it_op = IT_GETATTR };
struct md_enqueue_info *minfo;
struct ldlm_enqueue_info *einfo;
int rc;
ENTRY;
- if (inode == NULL)
+ if (unlikely(inode == NULL))
RETURN(1);
if (d_mountpoint(dentry))
RETURN(1);
- if (dentry == dentry->d_sb->s_root)
+ if (unlikely(dentry == dentry->d_sb->s_root))
RETURN(1);
- rc = md_revalidate_lock(ll_i2mdexp(dir), &it, ll_inode2fid(inode));
+ rc = md_revalidate_lock(ll_i2mdexp(dir), &it, ll_inode2fid(inode),
+ NULL);
if (rc == 1) {
ll_intent_release(&it);
RETURN(1);
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,
int rc;
ENTRY;
-#ifdef DCACHE_LUSTRE_INVALID
if (parent->d_flags & DCACHE_LUSTRE_INVALID) {
-#else
- if (d_unhashed(parent)) {
-#endif
CDEBUG(D_READA, "parent dentry@%p %.*s is "
"invalid, skip statahead\n",
parent, parent->d_name.len, parent->d_name.name);
RETURN(-EINVAL);
}
- se = ll_sai_entry_get(sai, sai->sai_index, SA_ENTRY_UNSTATED);
+ se = ll_sai_entry_init(sai, sai->sai_index);
if (IS_ERR(se))
RETURN(PTR_ERR(se));
GOTO(out, rc = -ENOMEM);
}
} else {
- rc = do_sa_revalidate(dentry);
+ rc = do_sa_revalidate(dir, dentry);
if (rc)
dput(dentry);
}
if (rc) {
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;
- cfs_waitq_signal(&sai->sai_waitq);
+ se->se_stat = rc < 0 ? rc : SA_ENTRY_STATED;
+ if (ll_sai_entry_to_stated(sai, se))
+ cfs_waitq_signal(&sai->sai_waitq);
} else {
sai->sai_sent++;
}
return rc;
}
-static inline int sa_check_stop(struct ll_statahead_info *sai)
-{
- return !!(sai->sai_thread.t_flags & SVC_STOPPING);
-}
-
-static inline int sa_not_full(struct ll_statahead_info *sai)
-{
- return sai->sai_index < sai->sai_hit + sai->sai_miss + sai->sai_max;
-}
-
-/* (1) hit ratio less than 80%
- * or
- * (2) consecutive miss more than 8 */
-static inline int sa_low_hit(struct ll_statahead_info *sai)
-{
- return ((sai->sai_hit < 4 * sai->sai_miss && sai->sai_hit > 7) ||
- (sai->sai_consecutive_miss > 8));
-}
-
-struct ll_sa_thread_args {
- struct dentry *sta_parent;
- pid_t sta_pid;
-};
-
static int ll_statahead_thread(void *arg)
{
- struct ll_sa_thread_args *sta = arg;
- struct dentry *parent = dget(sta->sta_parent);
+ struct dentry *parent = (struct dentry *)arg;
struct inode *dir = parent->d_inode;
struct ll_inode_info *lli = ll_i2info(dir);
struct ll_sb_info *sbi = ll_i2sbi(dir);
{
char pname[16];
- snprintf(pname, 15, "ll_sa_%u", sta->sta_pid);
+ snprintf(pname, 15, "ll_sa_%u", lli->lli_opendir_pid);
cfs_daemonize(pname);
}
- sbi->ll_sa_total++;
- spin_lock(&lli->lli_lock);
+ atomic_inc(&sbi->ll_sa_total);
+ cfs_spin_lock(&lli->lli_lock);
thread->t_flags = SVC_RUNNING;
- spin_unlock(&lli->lli_lock);
+ cfs_spin_unlock(&lli->lli_lock);
cfs_waitq_signal(&thread->t_ctl_waitq);
CDEBUG(D_READA, "start doing statahead for %s\n", parent->d_name.name);
page = ll_get_dir_page(dir, pos, 0, &chain);
while (1) {
+ struct l_wait_info lwi = { 0 };
struct lu_dirpage *dp;
struct lu_dirent *ent;
if (IS_ERR(page)) {
rc = PTR_ERR(page);
- CERROR("error reading dir "DFID" at %llu/%u: rc %d\n",
- PFID(ll_inode2fid(dir)), pos,
- sai->sai_index, rc);
+ CDEBUG(D_READA, "error reading dir "DFID" at "LPU64
+ "/%u: [rc %d] [parent %u]\n",
+ PFID(ll_inode2fid(dir)), pos, sai->sai_index,
+ rc, lli->lli_opendir_pid);
break;
}
dp = page_address(page);
for (ent = lu_dirent_start(dp); ent != NULL;
ent = lu_dirent_next(ent)) {
- struct l_wait_info lwi = { 0 };
char *name = ent->lde_name;
int namelen = le16_to_cpu(ent->lde_namelen);
- if (namelen == 0)
- /* Skip dummy record. */
+ if (unlikely(namelen == 0))
+ /*
+ * Skip dummy record.
+ */
continue;
if (name[0] == '.') {
if (namelen == 1) {
- /* skip . */
+ /*
+ * skip "."
+ */
continue;
} else if (name[1] == '.' && namelen == 2) {
- /* skip .. */
+ /*
+ * skip ".."
+ */
continue;
} else if (!sai->sai_ls_all) {
- /* skip hidden files */
+ /*
+ * skip hidden files.
+ */
sai->sai_skip_hidden++;
continue;
}
}
- /* don't stat-ahead first entry */
+ /*
+ * don't stat-ahead first entry.
+ */
if (unlikely(!first)) {
first++;
continue;
}
+keep_de:
l_wait_event(thread->t_ctl_waitq,
- sa_check_stop(sai) || sa_not_full(sai),
+ !sa_is_running(sai) || sa_not_full(sai) ||
+ !sa_received_empty(sai),
&lwi);
- if (unlikely(sa_check_stop(sai))) {
+ while (!sa_received_empty(sai) && sa_is_running(sai))
+ do_statahead_interpret(sai);
+
+ if (unlikely(!sa_is_running(sai))) {
ll_put_page(page);
GOTO(out, rc);
}
+ if (!sa_not_full(sai))
+ /*
+ * do not skip the current de.
+ */
+ goto keep_de;
+
rc = ll_statahead_one(parent, name, namelen);
if (rc < 0) {
ll_put_page(page);
pos = le64_to_cpu(dp->ldp_hash_end);
ll_put_page(page);
if (pos == DIR_END_OFF) {
- /* End of directory reached. */
- break;
- } else if (1 /* chain is exhausted*/) {
- /* Normal case: continue to the next page. */
+ /*
+ * End of directory reached.
+ */
+ while (1) {
+ l_wait_event(thread->t_ctl_waitq,
+ !sa_is_running(sai) ||
+ !sa_received_empty(sai) ||
+ sai->sai_sent == sai->sai_replied,
+ &lwi);
+ if (!sa_received_empty(sai) &&
+ sa_is_running(sai))
+ do_statahead_interpret(sai);
+ else
+ GOTO(out, rc);
+ }
+ } else if (1) {
+ /*
+ * chain is exhausted.
+ * Normal case: continue to the next page.
+ */
page = ll_get_dir_page(dir, pos, 1, &chain);
} else {
- /* go into overflow page. */
+ /*
+ * go into overflow page.
+ */
}
}
EXIT;
out:
ll_dir_chain_fini(&chain);
- spin_lock(&lli->lli_lock);
+ cfs_spin_lock(&lli->lli_lock);
thread->t_flags = SVC_STOPPED;
- spin_unlock(&lli->lli_lock);
+ cfs_spin_unlock(&lli->lli_lock);
cfs_waitq_signal(&sai->sai_waitq);
cfs_waitq_signal(&thread->t_ctl_waitq);
ll_sai_put(sai);
return rc;
}
-/* called in ll_file_release() */
+/**
+ * called in ll_file_release().
+ */
void ll_stop_statahead(struct inode *inode, void *key)
{
struct ll_inode_info *lli = ll_i2info(inode);
- struct ptlrpc_thread *thread;
- spin_lock(&lli->lli_lock);
- if (lli->lli_opendir_pid == 0 ||
- unlikely(lli->lli_opendir_key != key)) {
- spin_unlock(&lli->lli_lock);
+ if (unlikely(key == NULL))
+ return;
+
+ cfs_spin_lock(&lli->lli_lock);
+ if (lli->lli_opendir_key != key || lli->lli_opendir_pid == 0) {
+ cfs_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 (!(thread->t_flags & SVC_STOPPED)) {
+ if (!sa_is_stopped(lli->lli_sai)) {
thread->t_flags = SVC_STOPPING;
- spin_unlock(&lli->lli_lock);
+ cfs_spin_unlock(&lli->lli_lock);
cfs_waitq_signal(&thread->t_ctl_waitq);
CDEBUG(D_READA, "stopping statahead thread, pid %d\n",
cfs_curproc_pid());
l_wait_event(thread->t_ctl_waitq,
- thread->t_flags & SVC_STOPPED,
+ sa_is_stopped(lli->lli_sai),
&lwi);
} else {
- spin_unlock(&lli->lli_lock);
+ cfs_spin_unlock(&lli->lli_lock);
}
- /* Put the ref which was held when first statahead_enter.
+ /*
+ * Put the ref which was held when first statahead_enter.
* It maybe not the last ref for some statahead requests
- * maybe inflight. */
+ * maybe inflight.
+ */
ll_sai_put(lli->lli_sai);
- return;
+ } else {
+ lli->lli_opendir_pid = 0;
+ cfs_spin_unlock(&lli->lli_lock);
}
- spin_unlock(&lli->lli_lock);
}
enum {
- LS_NONE_FIRST_DE = 0, /* not first dirent, or is "." */
- LS_FIRST_DE, /* the first non-hidden dirent */
- LS_FIRST_DOT_DE /* the first hidden dirent, that is ".xxx" */
+ /**
+ * not first dirent, or is "."
+ */
+ LS_NONE_FIRST_DE = 0,
+ /**
+ * the first non-hidden dirent
+ */
+ LS_FIRST_DE,
+ /**
+ * the first hidden dirent, that is "."
+ */
+ LS_FIRST_DOT_DE
};
static int is_first_dirent(struct inode *dir, struct dentry *dentry)
struct lu_dirent *ent;
if (IS_ERR(page)) {
+ struct ll_inode_info *lli = ll_i2info(dir);
+
rc = PTR_ERR(page);
- CERROR("error reading dir "DFID" at %llu: rc %d\n",
- PFID(ll_inode2fid(dir)), pos, rc);
+ CERROR("error reading dir "DFID" at "LPU64": "
+ "[rc %d] [parent %u]\n",
+ PFID(ll_inode2fid(dir)), pos,
+ rc, lli->lli_opendir_pid);
break;
}
int namelen = le16_to_cpu(ent->lde_namelen);
if (namelen == 0)
- /* Skip dummy record. */
+ /*
+ * skip dummy record.
+ */
continue;
if (name[0] == '.') {
if (namelen == 1)
- /* skip . */
+ /*
+ * skip "."
+ */
continue;
else if (name[1] == '.' && namelen == 2)
- /* skip .. */
+ /*
+ * skip ".."
+ */
continue;
else
dot_de = 1;
continue;
}
- if (target->len == namelen &&
- !strncmp(target->name, name, target->len))
- rc = LS_FIRST_DE + dot_de;
- else
+ if (target->len != namelen ||
+ memcmp(target->name, name, namelen) != 0)
rc = LS_NONE_FIRST_DE;
+ else if (!dot_de)
+ rc = LS_FIRST_DE;
+ else
+ rc = LS_FIRST_DOT_DE;
+
ll_put_page(page);
GOTO(out, rc);
}
pos = le64_to_cpu(dp->ldp_hash_end);
ll_put_page(page);
if (pos == DIR_END_OFF) {
- /* End of directory reached. */
+ /*
+ * End of directory reached.
+ */
break;
- } else if (1 /* chain is exhausted*/) {
- /* Normal case: continue to the next page. */
+ } else if (1) {
+ /*
+ * chain is exhausted
+ * Normal case: continue to the next page.
+ */
page = ll_get_dir_page(dir, pos, 1, &chain);
} else {
- /* go into overflow page. */
+ /*
+ * go into overflow page.
+ */
}
}
EXIT;
return rc;
}
-/* Start statahead thread if this is the first dir entry.
+static int is_same_dentry(struct dentry *d1, struct dentry *d2)
+{
+ if (unlikely(d1 == d2))
+ return 1;
+ if (d1->d_parent == d2->d_parent &&
+ d1->d_name.hash == d2->d_name.hash &&
+ 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.
- * Return value:
- * 0 -- stat ahead thread process such dentry, for lookup, it miss
- * 1 -- stat ahead thread process such dentry, for lookup, it hit
- * -EEXIST -- stat ahead thread started, and this is the first dentry
- * -EBADFD -- statahead thread exit and not dentry available
- * others -- error */
+ * \retval 0 -- stat ahead thread process such dentry, miss for lookup
+ * \retval 1 -- stat ahead thread process such dentry, hit for any case
+ * \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
+ * \retval others -- error
+ */
int do_statahead_enter(struct inode *dir, struct dentry **dentryp, int lookup)
{
- struct ll_sb_info *sbi = ll_i2sbi(dir);
- struct ll_inode_info *lli = ll_i2info(dir);
- struct ll_statahead_info *sai = lli->lli_sai;
- struct ll_sa_thread_args sta;
+ struct ll_inode_info *lli;
+ struct ll_statahead_info *sai;
+ struct dentry *parent;
struct l_wait_info lwi = { 0 };
- int rc;
+ int rc = 0;
ENTRY;
+ LASSERT(dir != NULL);
+ lli = ll_i2info(dir);
LASSERT(lli->lli_opendir_pid == cfs_curproc_pid());
+ sai = lli->lli_sai;
if (sai) {
- if (unlikely(sai->sai_thread.t_flags & SVC_STOPPED &&
- list_empty(&sai->sai_entries)))
+ if (unlikely(sa_is_stopped(sai) &&
+ cfs_list_empty(&sai->sai_entries_stated)))
RETURN(-EBADFD);
if ((*dentryp)->d_name.name[0] == '.') {
if (likely(sai->sai_ls_all ||
sai->sai_miss_hidden >= sai->sai_skip_hidden)) {
- /* Hidden dentry is the first one, or statahead
+ /*
+ * Hidden dentry is the first one, or statahead
* thread does not skip so many hidden dentries
- * before "sai_ls_all" enabled as below. */
+ * before "sai_ls_all" enabled as below.
+ */
} else {
if (!sai->sai_ls_all)
- /* It maybe because hidden dentry is not
+ /*
+ * It maybe because hidden dentry is not
* the first one, "sai_ls_all" was not
* set, then "ls -al" missed. Enable
- * "sai_ls_all" for such case. */
+ * "sai_ls_all" for such case.
+ */
sai->sai_ls_all = 1;
- /* Such "getattr" has been skipped before
- * "sai_ls_all" enabled as above. */
+ /*
+ * Such "getattr" has been skipped before
+ * "sai_ls_all" enabled as above.
+ */
sai->sai_miss_hidden++;
RETURN(-ENOENT);
}
}
+ if (!ll_sai_entry_stated(sai)) {
+ /*
+ * thread started already, avoid double-stat.
+ */
+ lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
+ rc = l_wait_event(sai->sai_waitq,
+ ll_sai_entry_stated(sai) ||
+ sa_is_stopped(sai),
+ &lwi);
+ if (unlikely(rc == -EINTR))
+ RETURN(rc);
+ }
+
if (ll_sai_entry_stated(sai)) {
- sbi->ll_sa_cached++;
- } else {
- sbi->ll_sa_blocked++;
- /* thread started already, avoid double-stat */
- l_wait_event(sai->sai_waitq,
- ll_sai_entry_stated(sai) ||
- sai->sai_thread.t_flags & SVC_STOPPED,
- &lwi);
+ 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) {
&(*dentryp)->d_name);
if (result) {
LASSERT(result != *dentryp);
- dput(*dentryp);
+ /* BUG 16303: do not drop reference count for
+ * "*dentryp", VFS will do that by itself. */
*dentryp = result;
RETURN(1);
}
}
- /* do nothing for revalidate */
+ /*
+ * do nothing for revalidate.
+ */
RETURN(0);
}
- /* I am the "lli_opendir_pid" owner, only me can set "lli_sai". */
- LASSERT(lli->lli_sai == NULL);
-
+ /* I am the "lli_opendir_pid" owner, only me can set "lli_sai". */
rc = is_first_dirent(dir, *dentryp);
- if (rc == LS_NONE_FIRST_DE) {
- /* It is not "ls -{a}l" operation, no need statahead for it */
- spin_lock(&lli->lli_lock);
- lli->lli_opendir_key = NULL;
- lli->lli_opendir_pid = 0;
- spin_unlock(&lli->lli_lock);
- RETURN(-EBADF);
- }
+ if (rc == LS_NONE_FIRST_DE)
+ /* It is not "ls -{a}l" operation, no need statahead for it. */
+ GOTO(out, rc = -EAGAIN);
sai = ll_sai_alloc();
if (sai == NULL)
- RETURN(-ENOMEM);
+ GOTO(out, rc = -ENOMEM);
- sai->sai_inode = igrab(dir);
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);
+ GOTO(out, rc = -ESTALE);
+ }
- sta.sta_parent = (*dentryp)->d_parent;
- sta.sta_pid = cfs_curproc_pid();
+ /* get parent reference count here, and put it in ll_statahead_thread */
+ parent = dget((*dentryp)->d_parent);
+ if (unlikely(sai->sai_inode != parent->d_inode)) {
+ struct ll_inode_info *nlli = ll_i2info(parent->d_inode);
+
+ CWARN("Race condition, someone changed %.*s just now: "
+ "old parent "DFID", new parent "DFID"\n",
+ (*dentryp)->d_name.len, (*dentryp)->d_name.name,
+ PFID(&lli->lli_fid), PFID(&nlli->lli_fid));
+ dput(parent);
+ iput(sai->sai_inode);
+ OBD_FREE_PTR(sai);
+ RETURN(-EAGAIN);
+ }
lli->lli_sai = sai;
- rc = cfs_kernel_thread(ll_statahead_thread, &sta, 0);
+ rc = cfs_kernel_thread(ll_statahead_thread, parent, 0);
if (rc < 0) {
CERROR("can't start ll_sa thread, rc: %d\n", rc);
+ dput(parent);
+ lli->lli_opendir_key = NULL;
sai->sai_thread.t_flags = SVC_STOPPED;
ll_sai_put(sai);
LASSERT(lli->lli_sai == NULL);
- RETURN(rc);
+ RETURN(-EAGAIN);
}
- l_wait_event(sai->sai_thread.t_ctl_waitq,
- sai->sai_thread.t_flags & (SVC_RUNNING | SVC_STOPPED),
+ l_wait_event(sai->sai_thread.t_ctl_waitq,
+ sa_is_running(sai) || sa_is_stopped(sai),
&lwi);
- /* We don't stat-ahead for the first dirent since we are already in
- * lookup, and -EEXIST also indicates that this is the first dirent. */
+ /*
+ * We don't stat-ahead for the first dirent since we are already in
+ * lookup, and -EEXIST also indicates that this is the first dirent.
+ */
RETURN(-EEXIST);
+
+out:
+ cfs_spin_lock(&lli->lli_lock);
+ lli->lli_opendir_key = NULL;
+ lli->lli_opendir_pid = 0;
+ cfs_spin_unlock(&lli->lli_lock);
+ return rc;
}
-/* update hit/miss count */
-void ll_statahead_exit(struct dentry *dentry, int result)
+/**
+ * update hit/miss count.
+ */
+void ll_statahead_exit(struct inode *dir, struct dentry *dentry, int result)
{
- 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);
-
- if (lli->lli_opendir_pid != cfs_curproc_pid())
- return;
-
- if (lli->lli_sai) {
- struct ll_statahead_info *sai = lli->lli_sai;
+ struct ll_inode_info *lli;
+ struct ll_statahead_info *sai;
+ struct ll_sb_info *sbi;
+ struct ll_dentry_data *ldd = ll_d2d(dentry);
+ int rc;
+ ENTRY;
- if (result == 1) {
- sbi->ll_sa_hit++;
- sai->sai_hit++;
- sai->sai_consecutive_miss = 0;
- sai->sai_max = min(2 * sai->sai_max, sbi->ll_sa_max);
- } else {
- sbi->ll_sa_miss++;
- sai->sai_miss++;
- sai->sai_consecutive_miss++;
- if (sa_low_hit(sai) &&
- sai->sai_thread.t_flags & SVC_RUNNING) {
- sbi->ll_sa_wrong++;
- CDEBUG(D_READA, "statahead for dir %.*s hit "
- "ratio too low: hit/miss %u/%u, "
- "sent/replied %u/%u. stopping statahead "
- "thread: pid %d\n",
- parent->d_name.len, parent->d_name.name,
- sai->sai_hit, sai->sai_miss,
- sai->sai_sent, sai->sai_replied,
- cfs_curproc_pid());
- spin_lock(&lli->lli_lock);
- if (!(sai->sai_thread.t_flags & SVC_STOPPED))
- sai->sai_thread.t_flags = SVC_STOPPING;
- spin_unlock(&lli->lli_lock);
- }
+ LASSERT(dir != NULL);
+ lli = ll_i2info(dir);
+ LASSERT(lli->lli_opendir_pid == cfs_curproc_pid());
+ sai = lli->lli_sai;
+ LASSERT(sai != NULL);
+ sbi = ll_i2sbi(dir);
+
+ rc = ll_sai_entry_fini(sai);
+ /* rc == -ENOENT means such dentry was removed just between statahead
+ * readdir and pre-fetched, count it as hit.
+ *
+ * result == -ENOENT has two meanings:
+ * 1. such dentry was removed just between statahead pre-fetched and
+ * main process stat such dentry.
+ * 2. main process stat non-exist dentry.
+ * We can not distinguish such two cases, just count them as miss. */
+ if (result >= 1 || unlikely(rc == -ENOENT)) {
+ sai->sai_hit++;
+ sai->sai_consecutive_miss = 0;
+ sai->sai_max = min(2 * sai->sai_max, sbi->ll_sa_max);
+ } else {
+ sai->sai_miss++;
+ sai->sai_consecutive_miss++;
+ if (sa_low_hit(sai) && sa_is_running(sai)) {
+ atomic_inc(&sbi->ll_sa_wrong);
+ CDEBUG(D_READA, "Statahead for dir "DFID" hit ratio "
+ "too low: hit/miss %u/%u, sent/replied %u/%u, "
+ "stopping statahead thread: pid %d\n",
+ 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);
+ if (!sa_is_stopped(sai))
+ sai->sai_thread.t_flags = SVC_STOPPING;
+ cfs_spin_unlock(&lli->lli_lock);
}
+ }
+ if (!sa_is_stopped(sai))
cfs_waitq_signal(&sai->sai_thread.t_ctl_waitq);
- ll_sai_entry_put(sai);
+ if (likely(ldd != NULL))
+ ldd->lld_sa_generation = sai->sai_generation;
- if (likely(ldd != NULL))
- ldd->lld_sa_generation = sai->sai_generation;
- }
+ EXIT;
}