-/* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
- * vim:expandtab:shiftwidth=8:tabstop=8:
- *
+/*
* GPL HEADER START
*
* DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
*
* 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.
+ * http://www.gnu.org/licenses/gpl-2.0.html
*
* GPL HEADER END
*/
* Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
* Use is subject to license terms.
*
- * Copyright (c) 2011 Whamcloud, Inc.
- *
+ * Copyright (c) 2011, 2017, Intel Corporation.
*/
/*
* This file is part of Lustre, http://www.lustre.org/
#include <linux/fs.h>
#include <linux/sched.h>
+#include <linux/kthread.h>
#include <linux/mm.h>
-#include <linux/smp_lock.h>
#include <linux/highmem.h>
#include <linux/pagemap.h>
+#include <linux/delay.h>
#define DEBUG_SUBSYSTEM S_LLITE
#include <obd_support.h>
-#include <lustre_lite.h>
#include <lustre_dlm.h>
-#include <linux/lustre_version.h>
#include "llite_internal.h"
-struct ll_sai_entry {
- cfs_list_t se_list;
- unsigned int se_index;
- int se_stat;
- struct ptlrpc_request *se_req;
- struct md_enqueue_info *se_minfo;
-};
+#define SA_OMITTED_ENTRY_MAX 8ULL
-enum {
- SA_ENTRY_UNSTATED = 0,
- SA_ENTRY_STATED
+typedef enum {
+ /** negative values are for error cases */
+ SA_ENTRY_INIT = 0, /** init entry */
+ SA_ENTRY_SUCC = 1, /** stat succeed */
+ SA_ENTRY_INVA = 2, /** invalid entry */
+} se_state_t;
+
+/*
+ * sa_entry is not refcounted: statahead thread allocates it and do async stat,
+ * and in async stat callback ll_statahead_interpret() will add it into
+ * sai_interim_entries, later statahead thread will call sa_handle_callback() to
+ * instantiate entry and move it into sai_entries, and then only scanner process
+ * can access and free it.
+ */
+struct sa_entry {
+ /* link into sai_interim_entries or sai_entries */
+ struct list_head se_list;
+ /* link into sai hash table locally */
+ struct list_head se_hash;
+ /* entry index in the sai */
+ __u64 se_index;
+ /* low layer ldlm lock handle */
+ __u64 se_handle;
+ /* entry status */
+ se_state_t se_state;
+ /* entry size, contains name */
+ int se_size;
+ /* pointer to async getattr enqueue info */
+ struct md_enqueue_info *se_minfo;
+ /* pointer to the async getattr request */
+ struct ptlrpc_request *se_req;
+ /* pointer to the target inode */
+ struct inode *se_inode;
+ /* entry name */
+ struct qstr se_qstr;
+ /* entry fid */
+ struct lu_fid se_fid;
};
-static unsigned int sai_generation = 0;
-static cfs_spinlock_t sai_generation_lock = CFS_SPIN_LOCK_UNLOCKED;
+static unsigned int sai_generation;
+static DEFINE_SPINLOCK(sai_generation_lock);
-/**
- * Check whether first entry was stated already or not.
- * 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 int ll_sai_entry_stated(struct ll_statahead_info *sai)
+static inline int sa_unhashed(struct sa_entry *entry)
{
- 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;
+ return list_empty(&entry->se_hash);
}
-static inline int sa_received_empty(struct ll_statahead_info *sai)
+/* sa_entry is ready to use */
+static inline int sa_ready(struct sa_entry *entry)
{
- return cfs_list_empty(&sai->sai_entries_received);
+ /* Make sure sa_entry is updated and ready to use */
+ smp_rmb();
+ return (entry->se_state != SA_ENTRY_INIT);
}
-static inline int sa_not_full(struct ll_statahead_info *sai)
+/* hash value to put in sai_cache */
+static inline int sa_hash(int val)
{
- return !!(sai->sai_index < sai->sai_index_next + sai->sai_max);
+ return val & LL_SA_CACHE_MASK;
}
-static inline int sa_is_running(struct ll_statahead_info *sai)
+/* hash entry into sai_cache */
+static inline void
+sa_rehash(struct ll_statahead_info *sai, struct sa_entry *entry)
{
- return !!(sai->sai_thread.t_flags & SVC_RUNNING);
+ int i = sa_hash(entry->se_qstr.hash);
+
+ spin_lock(&sai->sai_cache_lock[i]);
+ list_add_tail(&entry->se_hash, &sai->sai_cache[i]);
+ spin_unlock(&sai->sai_cache_lock[i]);
}
-static inline int sa_is_stopping(struct ll_statahead_info *sai)
+/* unhash entry from sai_cache */
+static inline void
+sa_unhash(struct ll_statahead_info *sai, struct sa_entry *entry)
{
- return !!(sai->sai_thread.t_flags & SVC_STOPPING);
+ int i = sa_hash(entry->se_qstr.hash);
+
+ spin_lock(&sai->sai_cache_lock[i]);
+ list_del_init(&entry->se_hash);
+ spin_unlock(&sai->sai_cache_lock[i]);
}
-static inline int sa_is_stopped(struct ll_statahead_info *sai)
+static inline int agl_should_run(struct ll_statahead_info *sai,
+ struct inode *inode)
{
- return !!(sai->sai_thread.t_flags & SVC_STOPPED);
+ return inode && S_ISREG(inode->i_mode) && sai->sai_agl_task;
+}
+
+static inline struct ll_inode_info *
+agl_first_entry(struct ll_statahead_info *sai)
+{
+ return list_entry(sai->sai_agls.next, struct ll_inode_info,
+ lli_agl_list);
+}
+
+/* statahead window is full */
+static inline int sa_sent_full(struct ll_statahead_info *sai)
+{
+ return atomic_read(&sai->sai_cache_count) >= sai->sai_max;
+}
+
+/* got async stat replies */
+static inline int sa_has_callback(struct ll_statahead_info *sai)
+{
+ return !list_empty(&sai->sai_interim_entries);
+}
+
+static inline int agl_list_empty(struct ll_statahead_info *sai)
+{
+ return list_empty(&sai->sai_agls);
}
/**
* (1) hit ratio less than 80%
* or
* (2) consecutive miss more than 8
+ * then means low hit.
*/
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));
+ return ((sai->sai_hit > 7 && sai->sai_hit < 4 * sai->sai_miss) ||
+ (sai->sai_consecutive_miss > 8));
}
-/**
- * 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
+/*
+ * if the given index is behind of statahead window more than
+ * SA_OMITTED_ENTRY_MAX, then it is old.
*/
-static void ll_sai_entry_cleanup(struct ll_sai_entry *entry, int free)
+static inline int is_omitted_entry(struct ll_statahead_info *sai, __u64 index)
{
- 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) {
- LASSERT(cfs_list_empty(&entry->se_list));
- OBD_FREE_PTR(entry);
- }
-
- EXIT;
+ return ((__u64)sai->sai_max + index + SA_OMITTED_ENTRY_MAX <
+ sai->sai_index);
}
-static struct ll_statahead_info *ll_sai_alloc(void)
+/* allocate sa_entry and hash it to allow scanner process to find it */
+static struct sa_entry *
+sa_alloc(struct dentry *parent, struct ll_statahead_info *sai, __u64 index,
+ const char *name, int len, const struct lu_fid *fid)
{
- struct ll_statahead_info *sai;
-
- OBD_ALLOC_PTR(sai);
- if (!sai)
- return NULL;
-
- cfs_spin_lock(&sai_generation_lock);
- sai->sai_generation = ++sai_generation;
- if (unlikely(sai_generation == 0))
- sai->sai_generation = ++sai_generation;
- 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_sent);
- CFS_INIT_LIST_HEAD(&sai->sai_entries_received);
- CFS_INIT_LIST_HEAD(&sai->sai_entries_stated);
- return sai;
+ struct ll_inode_info *lli;
+ struct sa_entry *entry;
+ int entry_size;
+ char *dname;
+
+ ENTRY;
+
+ entry_size = sizeof(struct sa_entry) + (len & ~3) + 4;
+ OBD_ALLOC(entry, entry_size);
+ if (unlikely(!entry))
+ RETURN(ERR_PTR(-ENOMEM));
+
+ CDEBUG(D_READA, "alloc sa entry %.*s(%p) index %llu\n",
+ len, name, entry, index);
+
+ entry->se_index = index;
+
+ entry->se_state = SA_ENTRY_INIT;
+ entry->se_size = entry_size;
+ dname = (char *)entry + sizeof(struct sa_entry);
+ memcpy(dname, name, len);
+ dname[len] = 0;
+ entry->se_qstr.hash = ll_full_name_hash(parent, name, len);
+ entry->se_qstr.len = len;
+ entry->se_qstr.name = dname;
+ entry->se_fid = *fid;
+
+ lli = ll_i2info(sai->sai_dentry->d_inode);
+
+ spin_lock(&lli->lli_sa_lock);
+ INIT_LIST_HEAD(&entry->se_list);
+ sa_rehash(sai, entry);
+ spin_unlock(&lli->lli_sa_lock);
+
+ atomic_inc(&sai->sai_cache_count);
+
+ RETURN(entry);
}
-static inline
-struct ll_statahead_info *ll_sai_get(struct ll_statahead_info *sai)
+/* free sa_entry, which should have been unhashed and not in any list */
+static void sa_free(struct ll_statahead_info *sai, struct sa_entry *entry)
{
- LASSERT(sai);
- cfs_atomic_inc(&sai->sai_refcount);
- return sai;
+ CDEBUG(D_READA, "free sa entry %.*s(%p) index %llu\n",
+ entry->se_qstr.len, entry->se_qstr.name, entry,
+ entry->se_index);
+
+ LASSERT(list_empty(&entry->se_list));
+ LASSERT(sa_unhashed(entry));
+
+ OBD_FREE(entry, entry->se_size);
+ atomic_dec(&sai->sai_cache_count);
}
-static void ll_sai_put(struct ll_statahead_info *sai)
+/*
+ * find sa_entry by name, used by directory scanner, lock is not needed because
+ * only scanner can remove the entry from cache.
+ */
+static struct sa_entry *
+sa_get(struct ll_statahead_info *sai, const struct qstr *qstr)
{
- struct inode *inode = sai->sai_inode;
- struct ll_inode_info *lli;
- ENTRY;
-
- LASSERT(inode != NULL);
- lli = ll_i2info(inode);
- LASSERT(lli->lli_sai == sai);
-
- if (cfs_atomic_dec_and_lock(&sai->sai_refcount, &lli->lli_sa_lock)) {
- struct ll_sai_entry *entry, *next;
-
- 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_sa_lock);
- RETURN_EXIT;
- }
-
- LASSERT(lli->lli_opendir_key == NULL);
- lli->lli_sai = NULL;
- lli->lli_opendir_pid = 0;
- cfs_spin_unlock(&lli->lli_sa_lock);
-
- LASSERT(sa_is_stopped(sai));
-
- if (sai->sai_sent > sai->sai_replied)
- CDEBUG(D_READA,"statahead for dir "DFID" does not "
- "finish: [sent:%u] [replied:%u]\n",
- PFID(&lli->lli_fid),
- sai->sai_sent, sai->sai_replied);
-
- 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);
- }
- iput(inode);
- OBD_FREE_PTR(sai);
- }
- EXIT;
+ struct sa_entry *entry;
+ int i = sa_hash(qstr->hash);
+
+ list_for_each_entry(entry, &sai->sai_cache[i], se_hash) {
+ if (entry->se_qstr.hash == qstr->hash &&
+ entry->se_qstr.len == qstr->len &&
+ memcmp(entry->se_qstr.name, qstr->name, qstr->len) == 0)
+ return entry;
+ }
+ return NULL;
}
-/**
- * insert it into sai_entries_sent tail when init.
- */
-static struct ll_sai_entry *
-ll_sai_entry_init(struct ll_statahead_info *sai, unsigned int index)
+/* unhash and unlink sa_entry, and then free it */
+static inline void
+sa_kill(struct ll_statahead_info *sai, struct sa_entry *entry)
{
- struct ll_inode_info *lli = ll_i2info(sai->sai_inode);
- struct ll_sai_entry *entry;
- ENTRY;
+ struct ll_inode_info *lli = ll_i2info(sai->sai_dentry->d_inode);
- OBD_ALLOC_PTR(entry);
- if (entry == NULL)
- RETURN(ERR_PTR(-ENOMEM));
+ LASSERT(!sa_unhashed(entry));
+ LASSERT(!list_empty(&entry->se_list));
+ LASSERT(sa_ready(entry));
- CDEBUG(D_READA, "alloc sai entry %p index %u\n",
- entry, index);
- entry->se_index = index;
- entry->se_stat = SA_ENTRY_UNSTATED;
+ sa_unhash(sai, entry);
- cfs_spin_lock(&lli->lli_sa_lock);
- cfs_list_add_tail(&entry->se_list, &sai->sai_entries_sent);
- cfs_spin_unlock(&lli->lli_sa_lock);
+ spin_lock(&lli->lli_sa_lock);
+ list_del_init(&entry->se_list);
+ spin_unlock(&lli->lli_sa_lock);
- RETURN(entry);
+ if (entry->se_inode)
+ iput(entry->se_inode);
+
+ sa_free(sai, entry);
}
-/**
- * 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)
+/* called by scanner after use, sa_entry will be killed */
+static void
+sa_put(struct ll_statahead_info *sai, struct sa_entry *entry)
{
- struct ll_inode_info *lli = ll_i2info(sai->sai_inode);
- struct ll_sai_entry *entry;
- int rc = 0;
- ENTRY;
-
- 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,
- 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;
- OBD_FREE_PTR(entry);
- }
- } else {
- LASSERT(sa_is_stopped(sai));
- }
- cfs_spin_unlock(&lli->lli_sa_lock);
-
- RETURN(rc);
+ struct sa_entry *tmp, *next;
+
+ if (entry && entry->se_state == SA_ENTRY_SUCC) {
+ struct ll_sb_info *sbi = ll_i2sbi(sai->sai_dentry->d_inode);
+
+ 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 (entry)
+ sa_kill(sai, entry);
+
+ /*
+ * kill old completed entries, only scanner process does this, no need
+ * to lock
+ */
+ list_for_each_entry_safe(tmp, next, &sai->sai_entries, se_list) {
+ if (!is_omitted_entry(sai, tmp->se_index))
+ break;
+ sa_kill(sai, tmp);
+ }
}
-/**
- * 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
+/*
+ * update state and sort add entry to sai_entries by index, return true if
+ * scanner is waiting on this entry.
*/
-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)
+static bool
+__sa_make_ready(struct ll_statahead_info *sai, struct sa_entry *entry, int ret)
{
- struct ll_sai_entry *entry;
- ENTRY;
-
- 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);
+ struct sa_entry *se;
+ struct list_head *pos = &sai->sai_entries;
+ __u64 index = entry->se_index;
+
+ LASSERT(!sa_ready(entry));
+ LASSERT(list_empty(&entry->se_list));
+
+ list_for_each_entry_reverse(se, &sai->sai_entries, se_list) {
+ if (se->se_index < entry->se_index) {
+ pos = &se->se_list;
+ break;
+ }
+ }
+ list_add(&entry->se_list, pos);
+ /*
+ * LU-9210: ll_statahead_interpet must be able to see this before
+ * we wake it up
+ */
+ smp_store_release(&entry->se_state,
+ ret < 0 ? SA_ENTRY_INVA : SA_ENTRY_SUCC);
+
+ return (index == sai->sai_index_wait);
}
-/**
- * inside lli_sa_lock.
- * Move entry to sai_entries_received and
- * insert it into sai_entries_received tail.
+/* finish async stat RPC arguments */
+static void sa_fini_data(struct md_enqueue_info *minfo)
+{
+ ll_unlock_md_op_lsm(&minfo->mi_data);
+ iput(minfo->mi_dir);
+ OBD_FREE_PTR(minfo);
+}
+
+static int ll_statahead_interpret(struct ptlrpc_request *req,
+ struct md_enqueue_info *minfo, int rc);
+
+/*
+ * prepare arguments for async stat RPC.
*/
-static inline void
-ll_sai_entry_to_received(struct ll_statahead_info *sai, struct ll_sai_entry *entry)
+static struct md_enqueue_info *
+sa_prep_data(struct inode *dir, struct inode *child, struct sa_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);
+ struct md_enqueue_info *minfo;
+ struct ldlm_enqueue_info *einfo;
+ struct md_op_data *op_data;
+
+ OBD_ALLOC_PTR(minfo);
+ if (!minfo)
+ return ERR_PTR(-ENOMEM);
+
+ op_data = ll_prep_md_op_data(&minfo->mi_data, dir, child,
+ entry->se_qstr.name, entry->se_qstr.len, 0,
+ LUSTRE_OPC_ANY, NULL);
+ if (IS_ERR(op_data)) {
+ OBD_FREE_PTR(minfo);
+ return (struct md_enqueue_info *)op_data;
+ }
+
+ if (!child)
+ op_data->op_fid2 = entry->se_fid;
+
+ minfo->mi_it.it_op = IT_GETATTR;
+ minfo->mi_dir = igrab(dir);
+ minfo->mi_cb = ll_statahead_interpret;
+ minfo->mi_cbdata = entry;
+
+ einfo = &minfo->mi_einfo;
+ einfo->ei_type = LDLM_IBITS;
+ einfo->ei_mode = it_to_lock_mode(&minfo->mi_it);
+ einfo->ei_cb_bl = ll_md_blocking_ast;
+ einfo->ei_cb_cp = ldlm_completion_ast;
+ einfo->ei_cb_gl = NULL;
+ einfo->ei_cbdata = NULL;
+
+ return minfo;
}
-/**
- * Move entry to sai_entries_stated and
- * sort with the index.
+/*
+ * release resources used in async stat RPC, update entry state and wakeup if
+ * scanner process it waiting on this entry.
*/
-static int
-ll_sai_entry_to_stated(struct ll_statahead_info *sai, struct ll_sai_entry *entry)
+static void
+sa_make_ready(struct ll_statahead_info *sai, struct sa_entry *entry, int ret)
{
- struct ll_inode_info *lli = ll_i2info(sai->sai_inode);
- struct ll_sai_entry *se;
- ENTRY;
-
- ll_sai_entry_cleanup(entry, 0);
-
- 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_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_sa_lock);
- RETURN(1);
- }
- }
-
- /*
- * I am the first entry.
- */
- cfs_list_add(&entry->se_list, &sai->sai_entries_stated);
- cfs_spin_unlock(&lli->lli_sa_lock);
- RETURN(1);
+ struct ll_inode_info *lli = ll_i2info(sai->sai_dentry->d_inode);
+ struct md_enqueue_info *minfo = entry->se_minfo;
+ struct ptlrpc_request *req = entry->se_req;
+ bool wakeup;
+
+ /* release resources used in RPC */
+ if (minfo) {
+ entry->se_minfo = NULL;
+ ll_intent_release(&minfo->mi_it);
+ sa_fini_data(minfo);
+ }
+
+ if (req) {
+ entry->se_req = NULL;
+ ptlrpc_req_finished(req);
+ }
+
+ spin_lock(&lli->lli_sa_lock);
+ wakeup = __sa_make_ready(sai, entry, ret);
+ spin_unlock(&lli->lli_sa_lock);
+
+ if (wakeup)
+ wake_up(&sai->sai_waitq);
}
-/**
- * finish lookup/revalidate.
+/* insert inode into the list of sai_agls */
+static void ll_agl_add(struct ll_statahead_info *sai,
+ struct inode *inode, int index)
+{
+ struct ll_inode_info *child = ll_i2info(inode);
+ struct ll_inode_info *parent = ll_i2info(sai->sai_dentry->d_inode);
+
+ spin_lock(&child->lli_agl_lock);
+ if (child->lli_agl_index == 0) {
+ child->lli_agl_index = index;
+ spin_unlock(&child->lli_agl_lock);
+
+ LASSERT(list_empty(&child->lli_agl_list));
+
+ spin_lock(&parent->lli_agl_lock);
+ /* Re-check under the lock */
+ if (agl_should_run(sai, inode)) {
+ if (agl_list_empty(sai))
+ wake_up_process(sai->sai_agl_task);
+ igrab(inode);
+ list_add_tail(&child->lli_agl_list, &sai->sai_agls);
+ } else
+ child->lli_agl_index = 0;
+ spin_unlock(&parent->lli_agl_lock);
+ } else {
+ spin_unlock(&child->lli_agl_lock);
+ }
+}
+
+/* allocate sai */
+static struct ll_statahead_info *ll_sai_alloc(struct dentry *dentry)
+{
+ struct ll_statahead_info *sai;
+ struct ll_inode_info *lli = ll_i2info(dentry->d_inode);
+ int i;
+
+ ENTRY;
+
+ OBD_ALLOC_PTR(sai);
+ if (!sai)
+ RETURN(NULL);
+
+ sai->sai_dentry = dget(dentry);
+ atomic_set(&sai->sai_refcount, 1);
+ sai->sai_max = LL_SA_RPC_MIN;
+ sai->sai_index = 1;
+ init_waitqueue_head(&sai->sai_waitq);
+
+ INIT_LIST_HEAD(&sai->sai_interim_entries);
+ INIT_LIST_HEAD(&sai->sai_entries);
+ INIT_LIST_HEAD(&sai->sai_agls);
+
+ for (i = 0; i < LL_SA_CACHE_SIZE; i++) {
+ INIT_LIST_HEAD(&sai->sai_cache[i]);
+ spin_lock_init(&sai->sai_cache_lock[i]);
+ }
+ atomic_set(&sai->sai_cache_count, 0);
+
+ spin_lock(&sai_generation_lock);
+ lli->lli_sa_generation = ++sai_generation;
+ if (unlikely(sai_generation == 0))
+ lli->lli_sa_generation = ++sai_generation;
+ spin_unlock(&sai_generation_lock);
+
+ RETURN(sai);
+}
+
+/* free sai */
+static inline void ll_sai_free(struct ll_statahead_info *sai)
+{
+ LASSERT(sai->sai_dentry != NULL);
+ dput(sai->sai_dentry);
+ OBD_FREE_PTR(sai);
+}
+
+/*
+ * take refcount of sai if sai for @dir exists, which means statahead is on for
+ * this directory.
*/
-static int do_statahead_interpret(struct ll_statahead_info *sai)
+static inline struct ll_statahead_info *ll_sai_get(struct inode *dir)
{
- 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;
-
- 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_sa_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 (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.
- */
- struct dentry *save = dentry;
- struct it_cb_data icbd = {
- .icbd_parent = minfo->mi_dir,
- .icbd_childp = &dentry
- };
-
- LASSERT(fid_is_zero(&minfo->mi_data.op_fid2));
-
- /* 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);
-
- /* 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);
-
- if (dentry != save) {
- minfo->mi_dentry = dentry;
- dput(save);
- }
- } else {
- /*
- * revalidate.
- */
- if (!lu_fid_eq(&minfo->mi_data.op_fid2, &body->fid1)) {
- ll_unhash_aliases(dentry->d_inode);
- GOTO(out, rc = -EAGAIN);
- }
-
- rc = ll_revalidate_it_finish(req, it, dentry);
- if (rc) {
- ll_unhash_aliases(dentry->d_inode);
- GOTO(out, rc);
- }
-
- cfs_spin_lock(&ll_lookup_lock);
- spin_lock(&dcache_lock);
- lock_dentry(dentry);
- __d_drop(dentry);
- dentry->d_flags &= ~DCACHE_LUSTRE_INVALID;
- 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;
+ struct ll_inode_info *lli = ll_i2info(dir);
+ struct ll_statahead_info *sai = NULL;
-out:
- /* 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;
+ spin_lock(&lli->lli_sa_lock);
+ sai = lli->lli_sai;
+ if (sai)
+ atomic_inc(&sai->sai_refcount);
+ spin_unlock(&lli->lli_sa_lock);
+
+ return sai;
}
-static int ll_statahead_interpret(struct ptlrpc_request *req,
- struct md_enqueue_info *minfo,
- int rc)
+/*
+ * put sai refcount after use, if refcount reaches zero, free sai and sa_entries
+ * attached to it.
+ */
+static void ll_sai_put(struct ll_statahead_info *sai)
{
- 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_sa_lock);
- /* stale entry */
- if (unlikely(lli->lli_sai == NULL ||
- lli->lli_sai->sai_generation != minfo->mi_generation)) {
- cfs_spin_unlock(&lli->lli_sa_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_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_sa_lock);
- ll_sai_entry_cleanup(entry, 1);
- }
- ll_sai_put(sai);
- RETURN(rc);
- }
+ struct ll_inode_info *lli = ll_i2info(sai->sai_dentry->d_inode);
+
+ if (atomic_dec_and_lock(&sai->sai_refcount, &lli->lli_sa_lock)) {
+ struct sa_entry *entry, *next;
+ struct ll_sb_info *sbi = ll_i2sbi(sai->sai_dentry->d_inode);
+
+ lli->lli_sai = NULL;
+ spin_unlock(&lli->lli_sa_lock);
+
+ LASSERT(!sai->sai_task);
+ LASSERT(!sai->sai_agl_task);
+ LASSERT(sai->sai_sent == sai->sai_replied);
+ LASSERT(!sa_has_callback(sai));
+
+ list_for_each_entry_safe(entry, next, &sai->sai_entries,
+ se_list)
+ sa_kill(sai, entry);
+
+ LASSERT(atomic_read(&sai->sai_cache_count) == 0);
+ LASSERT(agl_list_empty(sai));
+
+ ll_sai_free(sai);
+ atomic_dec(&sbi->ll_sa_running);
+ }
}
-static void sa_args_fini(struct md_enqueue_info *minfo,
- struct ldlm_enqueue_info *einfo)
+/* Do NOT forget to drop inode refcount when into sai_agls. */
+static void ll_agl_trigger(struct inode *inode, struct ll_statahead_info *sai)
{
- 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);
+ struct ll_inode_info *lli = ll_i2info(inode);
+ u64 index = lli->lli_agl_index;
+ ktime_t expire;
+ int rc;
+
+ ENTRY;
+
+ LASSERT(list_empty(&lli->lli_agl_list));
+
+ /* AGL maybe fall behind statahead with one entry */
+ if (is_omitted_entry(sai, index + 1)) {
+ lli->lli_agl_index = 0;
+ iput(inode);
+ RETURN_EXIT;
+ }
+
+ /*
+ * In case of restore, the MDT has the right size and has already
+ * sent it back without granting the layout lock, inode is up-to-date.
+ * Then AGL (async glimpse lock) is useless.
+ * Also to glimpse we need the layout, in case of a runninh restore
+ * the MDT holds the layout lock so the glimpse will block up to the
+ * end of restore (statahead/agl will block)
+ */
+ if (ll_file_test_flag(lli, LLIF_FILE_RESTORING)) {
+ lli->lli_agl_index = 0;
+ iput(inode);
+ RETURN_EXIT;
+ }
+
+ /* Someone is in glimpse (sync or async), do nothing. */
+ rc = down_write_trylock(&lli->lli_glimpse_sem);
+ if (rc == 0) {
+ lli->lli_agl_index = 0;
+ iput(inode);
+ RETURN_EXIT;
+ }
+
+ /*
+ * Someone triggered glimpse within 1 sec before.
+ * 1) The former glimpse succeeded with glimpse lock granted by OST, and
+ * if the lock is still cached on client, AGL needs to do nothing. If
+ * it is cancelled by other client, AGL maybe cannot obtaion new lock
+ * for no glimpse callback triggered by AGL.
+ * 2) The former glimpse succeeded, but OST did not grant glimpse lock.
+ * Under such case, it is quite possible that the OST will not grant
+ * glimpse lock for AGL also.
+ * 3) The former glimpse failed, compared with other two cases, it is
+ * relative rare. AGL can ignore such case, and it will not muchly
+ * affect the performance.
+ */
+ expire = ktime_sub_ns(ktime_get(), NSEC_PER_SEC);
+ if (ktime_to_ns(lli->lli_glimpse_time) &&
+ ktime_before(expire, lli->lli_glimpse_time)) {
+ up_write(&lli->lli_glimpse_sem);
+ lli->lli_agl_index = 0;
+ iput(inode);
+ RETURN_EXIT;
+ }
+
+ CDEBUG(D_READA,
+ "Handling (init) async glimpse: inode = " DFID", idx = %llu\n",
+ PFID(&lli->lli_fid), index);
+
+ cl_agl(inode);
+ lli->lli_agl_index = 0;
+ lli->lli_glimpse_time = ktime_get();
+ up_write(&lli->lli_glimpse_sem);
+
+ CDEBUG(D_READA,
+ "Handled (init) async glimpse: inode= " DFID", idx = %llu, rc = %d\n",
+ PFID(&lli->lli_fid), index, rc);
+
+ iput(inode);
+
+ EXIT;
}
-/**
- * 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".
+/*
+ * prepare inode for sa entry, add it into agl list, now sa_entry is ready
+ * to be used by scanner process.
*/
-static int sa_args_init(struct inode *dir, struct dentry *dentry,
- struct md_enqueue_info **pmi,
- struct ldlm_enqueue_info **pei,
- struct obd_capa **pcapa)
+static void sa_instantiate(struct ll_statahead_info *sai,
+ struct sa_entry *entry)
{
- struct ll_inode_info *lli = ll_i2info(dir);
- struct md_enqueue_info *minfo;
- struct ldlm_enqueue_info *einfo;
- struct md_op_data *op_data;
-
- OBD_ALLOC_PTR(einfo);
- if (einfo == NULL)
- return -ENOMEM;
-
- OBD_ALLOC_PTR(minfo);
- if (minfo == NULL) {
- OBD_FREE_PTR(einfo);
- return -ENOMEM;
- }
-
- op_data = ll_prep_md_op_data(&minfo->mi_data, dir, dentry->d_inode,
- dentry->d_name.name, dentry->d_name.len,
- 0, LUSTRE_OPC_ANY, NULL);
- if (IS_ERR(op_data)) {
- OBD_FREE_PTR(einfo);
- OBD_FREE_PTR(minfo);
- return PTR_ERR(op_data);
- }
-
- 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;
-
- einfo->ei_type = LDLM_IBITS;
- einfo->ei_mode = it_to_lock_mode(&minfo->mi_it);
- einfo->ei_cb_bl = ll_md_blocking_ast;
- einfo->ei_cb_cp = ldlm_completion_ast;
- einfo->ei_cb_gl = NULL;
- einfo->ei_cbdata = NULL;
-
- *pmi = minfo;
- *pei = einfo;
- pcapa[0] = op_data->op_capa1;
- pcapa[1] = op_data->op_capa2;
-
- return 0;
+ struct inode *dir = sai->sai_dentry->d_inode;
+ struct inode *child;
+ struct md_enqueue_info *minfo;
+ struct lookup_intent *it;
+ struct ptlrpc_request *req;
+ struct mdt_body *body;
+ int rc = 0;
+
+ ENTRY;
+
+ LASSERT(entry->se_handle != 0);
+
+ minfo = entry->se_minfo;
+ it = &minfo->mi_it;
+ req = entry->se_req;
+ body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
+ if (!body)
+ GOTO(out, rc = -EFAULT);
+
+ child = entry->se_inode;
+ /* revalidate; unlinked and re-created with the same name */
+ if (unlikely(!lu_fid_eq(&minfo->mi_data.op_fid2, &body->mbo_fid1))) {
+ if (child) {
+ entry->se_inode = NULL;
+ iput(child);
+ }
+ /* The mdt_body is invalid. Skip this entry */
+ GOTO(out, rc = -EAGAIN);
+ }
+
+ it->it_lock_handle = entry->se_handle;
+ rc = md_revalidate_lock(ll_i2mdexp(dir), it, ll_inode2fid(dir), NULL);
+ if (rc != 1)
+ GOTO(out, rc = -EAGAIN);
+
+ rc = ll_prep_inode(&child, req, dir->i_sb, it);
+ if (rc)
+ GOTO(out, rc);
+
+ CDEBUG(D_READA, "%s: setting %.*s"DFID" l_data to inode %p\n",
+ ll_i2sbi(dir)->ll_fsname, entry->se_qstr.len,
+ entry->se_qstr.name, PFID(ll_inode2fid(child)), child);
+ ll_set_lock_data(ll_i2sbi(dir)->ll_md_exp, child, it, NULL);
+
+ entry->se_inode = child;
+
+ if (agl_should_run(sai, child))
+ ll_agl_add(sai, child, entry->se_index);
+
+ EXIT;
+
+out:
+ /*
+ * sa_make_ready() will drop ldlm ibits lock refcount by calling
+ * ll_intent_drop_lock() in spite of failures. Do not worry about
+ * calling ll_intent_drop_lock() more than once.
+ */
+ sa_make_ready(sai, entry, rc);
}
-/**
- * similar to ll_lookup_it().
+/* once there are async stat replies, instantiate sa_entry from replies */
+static void sa_handle_callback(struct ll_statahead_info *sai)
+{
+ struct ll_inode_info *lli;
+
+ lli = ll_i2info(sai->sai_dentry->d_inode);
+
+ spin_lock(&lli->lli_sa_lock);
+ while (sa_has_callback(sai)) {
+ struct sa_entry *entry;
+
+ entry = list_entry(sai->sai_interim_entries.next,
+ struct sa_entry, se_list);
+ list_del_init(&entry->se_list);
+ spin_unlock(&lli->lli_sa_lock);
+
+ sa_instantiate(sai, entry);
+ spin_lock(&lli->lli_sa_lock);
+ }
+ spin_unlock(&lli->lli_sa_lock);
+}
+
+/*
+ * callback for async stat RPC, because this is called in ptlrpcd context, we
+ * only put sa_entry in sai_interim_entries, and wake up statahead thread to
+ * really prepare inode and instantiate sa_entry later.
*/
-static int do_sa_lookup(struct inode *dir, struct dentry *dentry)
+static int ll_statahead_interpret(struct ptlrpc_request *req,
+ struct md_enqueue_info *minfo, int rc)
+{
+ struct lookup_intent *it = &minfo->mi_it;
+ struct inode *dir = minfo->mi_dir;
+ struct ll_inode_info *lli = ll_i2info(dir);
+ struct ll_statahead_info *sai = lli->lli_sai;
+ struct sa_entry *entry = (struct sa_entry *)minfo->mi_cbdata;
+ __u64 handle = 0;
+
+ ENTRY;
+
+ if (it_disposition(it, DISP_LOOKUP_NEG))
+ rc = -ENOENT;
+
+ /*
+ * because statahead thread will wait for all inflight RPC to finish,
+ * sai should be always valid, no need to refcount
+ */
+ LASSERT(sai != NULL);
+ LASSERT(entry != NULL);
+
+ CDEBUG(D_READA, "sa_entry %.*s rc %d\n",
+ entry->se_qstr.len, entry->se_qstr.name, rc);
+
+ if (rc != 0) {
+ ll_intent_release(it);
+ sa_fini_data(minfo);
+ } else {
+ /*
+ * release ibits lock ASAP to avoid deadlock when statahead
+ * thread enqueues lock on parent in readdir and another
+ * process enqueues lock on child with parent lock held, eg.
+ * unlink.
+ */
+ handle = it->it_lock_handle;
+ ll_intent_drop_lock(it);
+ ll_unlock_md_op_lsm(&minfo->mi_data);
+ }
+
+ spin_lock(&lli->lli_sa_lock);
+ if (rc != 0) {
+ if (__sa_make_ready(sai, entry, rc))
+ wake_up(&sai->sai_waitq);
+ } else {
+ int first = 0;
+
+ entry->se_minfo = minfo;
+ entry->se_req = ptlrpc_request_addref(req);
+ /*
+ * Release the async ibits lock ASAP to avoid deadlock
+ * when statahead thread tries to enqueue lock on parent
+ * for readpage and other tries to enqueue lock on child
+ * with parent's lock held, for example: unlink.
+ */
+ entry->se_handle = handle;
+ if (!sa_has_callback(sai))
+ first = 1;
+
+ list_add_tail(&entry->se_list, &sai->sai_interim_entries);
+ if (first && sai->sai_task)
+ wake_up_process(sai->sai_task);
+ }
+ sai->sai_replied++;
+
+ spin_unlock(&lli->lli_sa_lock);
+
+ RETURN(rc);
+}
+
+/* async stat for file not found in dcache */
+static int sa_lookup(struct inode *dir, struct sa_entry *entry)
{
- struct md_enqueue_info *minfo;
- struct ldlm_enqueue_info *einfo;
- struct obd_capa *capas[2];
- int rc;
- ENTRY;
-
- rc = sa_args_init(dir, dentry, &minfo, &einfo, capas);
- if (rc)
- RETURN(rc);
-
- rc = md_intent_getattr_async(ll_i2mdexp(dir), minfo, einfo);
- if (!rc) {
- capa_put(capas[0]);
- capa_put(capas[1]);
- } else {
- sa_args_fini(minfo, einfo);
- }
-
- RETURN(rc);
+ struct md_enqueue_info *minfo;
+ int rc;
+
+ ENTRY;
+
+ minfo = sa_prep_data(dir, NULL, entry);
+ if (IS_ERR(minfo))
+ RETURN(PTR_ERR(minfo));
+
+ rc = md_intent_getattr_async(ll_i2mdexp(dir), minfo);
+ if (rc < 0)
+ sa_fini_data(minfo);
+
+ RETURN(rc);
}
/**
- * similar to ll_revalidate_it().
- * \retval 1 -- dentry valid
- * \retval 0 -- will send stat-ahead request
- * \retval others -- prepare stat-ahead request failed
+ * async stat for file found in dcache, similar to .revalidate
+ *
+ * \retval 1 dentry valid, no RPC sent
+ * \retval 0 dentry invalid, will send async stat RPC
+ * \retval negative number upon error
*/
-static int do_sa_revalidate(struct inode *dir, struct dentry *dentry)
+static int sa_revalidate(struct inode *dir, struct sa_entry *entry,
+ struct dentry *dentry)
{
- struct inode *inode = dentry->d_inode;
- struct lookup_intent it = { .it_op = IT_GETATTR };
- struct md_enqueue_info *minfo;
- struct ldlm_enqueue_info *einfo;
- struct obd_capa *capas[2];
- int rc;
- ENTRY;
-
- if (unlikely(inode == NULL))
- RETURN(1);
-
- if (d_mountpoint(dentry))
- RETURN(1);
-
- if (unlikely(dentry == dentry->d_sb->s_root))
- RETURN(1);
-
- rc = md_revalidate_lock(ll_i2mdexp(dir), &it, ll_inode2fid(inode));
- if (rc == 1) {
- ll_intent_release(&it);
- RETURN(1);
- }
-
- rc = sa_args_init(dir, dentry, &minfo, &einfo, capas);
- if (rc)
- RETURN(rc);
-
- rc = md_intent_getattr_async(ll_i2mdexp(dir), minfo, einfo);
- if (!rc) {
- capa_put(capas[0]);
- capa_put(capas[1]);
- } else {
- sa_args_fini(minfo, einfo);
- }
-
- RETURN(rc);
+ struct inode *inode = dentry->d_inode;
+ struct lookup_intent it = { .it_op = IT_GETATTR,
+ .it_lock_handle = 0 };
+ struct md_enqueue_info *minfo;
+ int rc;
+
+ ENTRY;
+
+ if (unlikely(!inode))
+ RETURN(1);
+
+ if (d_mountpoint(dentry))
+ RETURN(1);
+
+ minfo = sa_prep_data(dir, inode, entry);
+ if (IS_ERR(minfo))
+ RETURN(PTR_ERR(minfo));
+
+ entry->se_inode = igrab(inode);
+ rc = md_revalidate_lock(ll_i2mdexp(dir), &it, ll_inode2fid(inode),
+ NULL);
+ if (rc == 1) {
+ entry->se_handle = it.it_lock_handle;
+ ll_intent_release(&it);
+ sa_fini_data(minfo);
+ RETURN(1);
+ }
+
+ rc = md_intent_getattr_async(ll_i2mdexp(dir), minfo);
+ if (rc < 0) {
+ entry->se_inode = NULL;
+ iput(inode);
+ sa_fini_data(minfo);
+ }
+
+ RETURN(rc);
}
-static inline void ll_name2qstr(struct qstr *q, const char *name, int namelen)
+/* async stat for file with @name */
+static void sa_statahead(struct dentry *parent, const char *name, int len,
+ const struct lu_fid *fid)
{
- q->name = name;
- q->len = namelen;
- q->hash = full_name_hash(name, namelen);
+ struct inode *dir = parent->d_inode;
+ struct ll_inode_info *lli = ll_i2info(dir);
+ struct ll_statahead_info *sai = lli->lli_sai;
+ struct dentry *dentry = NULL;
+ struct sa_entry *entry;
+ int rc;
+
+ ENTRY;
+
+ entry = sa_alloc(parent, sai, sai->sai_index, name, len, fid);
+ if (IS_ERR(entry))
+ RETURN_EXIT;
+
+ dentry = d_lookup(parent, &entry->se_qstr);
+ if (!dentry) {
+ rc = sa_lookup(dir, entry);
+ } else {
+ rc = sa_revalidate(dir, entry, dentry);
+ if (rc == 1 && agl_should_run(sai, dentry->d_inode))
+ ll_agl_add(sai, dentry->d_inode, entry->se_index);
+ }
+
+ if (dentry)
+ dput(dentry);
+
+ if (rc != 0)
+ sa_make_ready(sai, entry, rc);
+ else
+ sai->sai_sent++;
+
+ sai->sai_index++;
+
+ EXIT;
}
-static int ll_statahead_one(struct dentry *parent, const char* entry_name,
- int entry_name_len)
+/* async glimpse (agl) thread main function */
+static int ll_agl_thread(void *arg)
{
- struct inode *dir = parent->d_inode;
- struct ll_inode_info *lli = ll_i2info(dir);
- struct ll_statahead_info *sai = lli->lli_sai;
- struct qstr name;
- struct dentry *dentry = NULL;
- struct ll_sai_entry *se;
- int rc;
- ENTRY;
-
- if (parent->d_flags & DCACHE_LUSTRE_INVALID) {
- 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_init(sai, sai->sai_index);
- if (IS_ERR(se))
- RETURN(PTR_ERR(se));
-
- ll_name2qstr(&name, entry_name, entry_name_len);
- dentry = d_lookup(parent, &name);
- if (!dentry) {
- dentry = d_alloc(parent, &name);
- if (dentry)
- rc = do_sa_lookup(dir, dentry);
- else
- GOTO(out, rc = -ENOMEM);
- } else {
- rc = do_sa_revalidate(dir, dentry);
- }
-
- EXIT;
+ struct dentry *parent = (struct dentry *)arg;
+ struct inode *dir = parent->d_inode;
+ struct ll_inode_info *plli = ll_i2info(dir);
+ struct ll_inode_info *clli;
+ /*
+ * We already own this reference, so it is safe to take it
+ * without a lock.
+ */
+ struct ll_statahead_info *sai = plli->lli_sai;
+
+ ENTRY;
+
+ CDEBUG(D_READA, "agl thread started: sai %p, parent %pd\n",
+ sai, parent);
+
+ while (({set_current_state(TASK_IDLE);
+ !kthread_should_stop(); })) {
+ spin_lock(&plli->lli_agl_lock);
+ if (!agl_list_empty(sai)) {
+ __set_current_state(TASK_RUNNING);
+ clli = agl_first_entry(sai);
+ list_del_init(&clli->lli_agl_list);
+ spin_unlock(&plli->lli_agl_lock);
+ ll_agl_trigger(&clli->lli_vfs_inode, sai);
+ cond_resched();
+ } else {
+ spin_unlock(&plli->lli_agl_lock);
+ schedule();
+ }
+ }
+ __set_current_state(TASK_RUNNING);
+ RETURN(0);
+}
-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);
- if (ll_sai_entry_to_stated(sai, se))
- cfs_waitq_signal(&sai->sai_waitq);
- } else {
- sai->sai_sent++;
- }
-
- sai->sai_index++;
- return rc;
+static void ll_stop_agl(struct ll_statahead_info *sai)
+{
+ struct dentry *parent = sai->sai_dentry;
+ struct ll_inode_info *plli = ll_i2info(parent->d_inode);
+ struct ll_inode_info *clli;
+ struct task_struct *agl_task;
+
+ spin_lock(&plli->lli_agl_lock);
+ agl_task = sai->sai_agl_task;
+ sai->sai_agl_task = NULL;
+ spin_unlock(&plli->lli_agl_lock);
+ if (!agl_task)
+ return;
+
+ CDEBUG(D_READA, "stop agl thread: sai %p pid %u\n",
+ sai, (unsigned int)agl_task->pid);
+ kthread_stop(agl_task);
+
+ spin_lock(&plli->lli_agl_lock);
+ while (!agl_list_empty(sai)) {
+ clli = agl_first_entry(sai);
+ list_del_init(&clli->lli_agl_list);
+ spin_unlock(&plli->lli_agl_lock);
+ clli->lli_agl_index = 0;
+ iput(&clli->lli_vfs_inode);
+ spin_lock(&plli->lli_agl_lock);
+ }
+ spin_unlock(&plli->lli_agl_lock);
+ CDEBUG(D_READA, "agl thread stopped: sai %p, parent %pd\n",
+ sai, parent);
+ ll_sai_put(sai);
}
-static int ll_statahead_thread(void *arg)
+/* start agl thread */
+static void ll_start_agl(struct dentry *parent, struct ll_statahead_info *sai)
{
- 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);
- struct ll_statahead_info *sai = ll_sai_get(lli->lli_sai);
- struct ptlrpc_thread *thread = &sai->sai_thread;
- struct page *page;
- __u64 pos = 0;
- int first = 0;
- int rc = 0;
- struct ll_dir_chain chain;
- ENTRY;
-
- {
- char pname[16];
- snprintf(pname, 15, "ll_sa_%u", lli->lli_opendir_pid);
- cfs_daemonize(pname);
- }
-
- atomic_inc(&sbi->ll_sa_total);
- cfs_spin_lock(&lli->lli_sa_lock);
- thread->t_flags = SVC_RUNNING;
- 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(NULL, 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);
- 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)) {
- __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) {
- /*
- * skip "."
- */
- continue;
- } else if (name[1] == '.' && namelen == 2) {
- /*
- * skip ".."
- */
- continue;
- } else if (!sai->sai_ls_all) {
- /*
- * skip hidden files.
- */
- sai->sai_skip_hidden++;
- continue;
- }
- }
-
- /*
- * don't stat-ahead first entry.
- */
- if (unlikely(!first)) {
- first++;
- continue;
- }
-
-keep_de:
- l_wait_event(thread->t_ctl_waitq,
- !sa_is_running(sai) || sa_not_full(sai) ||
- !sa_received_empty(sai),
- &lwi);
-
- 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);
- GOTO(out, rc);
- }
- }
- pos = le64_to_cpu(dp->ldp_hash_end);
- ll_put_page(page);
- if (pos == MDS_DIR_END_OFF) {
- /*
- * 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.
- */
- lli->lli_sa_pos = pos;
- page = ll_get_dir_page(NULL, dir, pos, 1, &chain);
- } else {
- /*
- * go into overflow page.
- */
- }
- }
- EXIT;
+ int node = cfs_cpt_spread_node(cfs_cpt_tab, CFS_CPT_ANY);
+ struct ll_inode_info *plli;
+ struct task_struct *task;
+
+ ENTRY;
+
+ CDEBUG(D_READA, "start agl thread: sai %p, parent %pd\n",
+ sai, parent);
+
+ plli = ll_i2info(parent->d_inode);
+ task = kthread_create_on_node(ll_agl_thread, parent, node, "ll_agl_%d",
+ plli->lli_opendir_pid);
+ if (IS_ERR(task)) {
+ CERROR("can't start ll_agl thread, rc: %ld\n", PTR_ERR(task));
+ RETURN_EXIT;
+ }
+ sai->sai_agl_task = task;
+ atomic_inc(&ll_i2sbi(d_inode(parent))->ll_agl_total);
+ /* Get an extra reference that the thread holds */
+ ll_sai_get(d_inode(parent));
+
+ wake_up_process(task);
+
+ EXIT;
+}
+/* statahead thread main function */
+static int ll_statahead_thread(void *arg)
+{
+ 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);
+ struct ll_statahead_info *sai = lli->lli_sai;
+ int first = 0;
+ struct md_op_data *op_data;
+ struct page *page = NULL;
+ __u64 pos = 0;
+ int rc = 0;
+
+ ENTRY;
+
+ CDEBUG(D_READA, "statahead thread starting: sai %p, parent %pd\n",
+ sai, parent);
+
+ OBD_ALLOC_PTR(op_data);
+ if (!op_data)
+ GOTO(out, rc = -ENOMEM);
+
+ while (pos != MDS_DIR_END_OFF && sai->sai_task) {
+ struct lu_dirpage *dp;
+ struct lu_dirent *ent;
+
+ op_data = ll_prep_md_op_data(op_data, dir, dir, NULL, 0, 0,
+ LUSTRE_OPC_ANY, dir);
+ if (IS_ERR(op_data)) {
+ rc = PTR_ERR(op_data);
+ break;
+ }
+
+ sai->sai_in_readpage = 1;
+ page = ll_get_dir_page(dir, op_data, pos);
+ ll_unlock_md_op_lsm(op_data);
+ sai->sai_in_readpage = 0;
+ if (IS_ERR(page)) {
+ rc = PTR_ERR(page);
+ CDEBUG(D_READA,
+ "error reading dir "DFID" at %llu /%llu opendir_pid = %u: rc = %d\n",
+ PFID(ll_inode2fid(dir)), pos, sai->sai_index,
+ lli->lli_opendir_pid, rc);
+ break;
+ }
+
+ dp = page_address(page);
+ for (ent = lu_dirent_start(dp);
+ ent != NULL && sai->sai_task &&
+ !sa_low_hit(sai);
+ ent = lu_dirent_next(ent)) {
+ __u64 hash;
+ int namelen;
+ char *name;
+ struct lu_fid fid;
+
+ 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) {
+ /*
+ * skip "."
+ */
+ continue;
+ } else if (name[1] == '.' && namelen == 2) {
+ /*
+ * skip ".."
+ */
+ continue;
+ } else if (!sai->sai_ls_all) {
+ /*
+ * skip hidden files.
+ */
+ sai->sai_skip_hidden++;
+ continue;
+ }
+ }
+
+ /*
+ * don't stat-ahead first entry.
+ */
+ if (unlikely(++first == 1))
+ continue;
+
+ fid_le_to_cpu(&fid, &ent->lde_fid);
+
+ while (({set_current_state(TASK_IDLE);
+ sai->sai_task; })) {
+ if (sa_has_callback(sai)) {
+ __set_current_state(TASK_RUNNING);
+ sa_handle_callback(sai);
+ }
+
+ spin_lock(&lli->lli_agl_lock);
+ while (sa_sent_full(sai) &&
+ !agl_list_empty(sai)) {
+ struct ll_inode_info *clli;
+
+ __set_current_state(TASK_RUNNING);
+ clli = agl_first_entry(sai);
+ list_del_init(&clli->lli_agl_list);
+ spin_unlock(&lli->lli_agl_lock);
+
+ ll_agl_trigger(&clli->lli_vfs_inode,
+ sai);
+ cond_resched();
+ spin_lock(&lli->lli_agl_lock);
+ }
+ spin_unlock(&lli->lli_agl_lock);
+
+ if (!sa_sent_full(sai))
+ break;
+ schedule();
+ }
+ __set_current_state(TASK_RUNNING);
+
+ sa_statahead(parent, name, namelen, &fid);
+ }
+
+ pos = le64_to_cpu(dp->ldp_hash_end);
+ ll_release_page(dir, page,
+ le32_to_cpu(dp->ldp_flags) & LDF_COLLIDE);
+
+ if (sa_low_hit(sai)) {
+ rc = -EFAULT;
+ atomic_inc(&sbi->ll_sa_wrong);
+ CDEBUG(D_READA,
+ "Statahead for dir "DFID" hit ratio too low: hit/miss %llu/%llu, sent/replied %llu/%llu, stoppingstatahead thread: pid %d\n",
+ PFID(&lli->lli_fid), sai->sai_hit,
+ sai->sai_miss, sai->sai_sent,
+ sai->sai_replied, current->pid);
+ break;
+ }
+ }
+ ll_finish_md_op_data(op_data);
+
+ if (rc < 0) {
+ spin_lock(&lli->lli_sa_lock);
+ sai->sai_task = NULL;
+ lli->lli_sa_enabled = 0;
+ spin_unlock(&lli->lli_sa_lock);
+ }
+
+ /*
+ * statahead is finished, but statahead entries need to be cached, wait
+ * for file release to stop me.
+ */
+ while (({set_current_state(TASK_IDLE);
+ sai->sai_task; })) {
+ if (sa_has_callback(sai)) {
+ __set_current_state(TASK_RUNNING);
+ sa_handle_callback(sai);
+ } else {
+ schedule();
+ }
+ }
+ __set_current_state(TASK_RUNNING);
+
+ EXIT;
out:
- ll_dir_chain_fini(&chain);
- cfs_spin_lock(&lli->lli_sa_lock);
- thread->t_flags = SVC_STOPPED;
- cfs_spin_unlock(&lli->lli_sa_lock);
- cfs_waitq_signal(&sai->sai_waitq);
- cfs_waitq_signal(&thread->t_ctl_waitq);
- ll_sai_put(sai);
- dput(parent);
- CDEBUG(D_READA, "statahead thread stopped, pid %d\n",
- cfs_curproc_pid());
- return rc;
+ ll_stop_agl(sai);
+
+ /*
+ * wait for inflight statahead RPCs to finish, and then we can free sai
+ * safely because statahead RPC will access sai data
+ */
+ while (sai->sai_sent != sai->sai_replied)
+ /* in case we're not woken up, timeout wait */
+ msleep(125);
+
+ /* release resources held by statahead RPCs */
+ sa_handle_callback(sai);
+
+ CDEBUG(D_READA, "%s: statahead thread stopped: sai %p, parent %pd\n",
+ sbi->ll_fsname, sai, parent);
+
+ spin_lock(&lli->lli_sa_lock);
+ sai->sai_task = NULL;
+ spin_unlock(&lli->lli_sa_lock);
+ wake_up(&sai->sai_waitq);
+
+ ll_sai_put(sai);
+
+ return rc;
}
-/**
- * called in ll_file_release().
+/* authorize opened dir handle @key to statahead */
+void ll_authorize_statahead(struct inode *dir, void *key)
+{
+ struct ll_inode_info *lli = ll_i2info(dir);
+
+ spin_lock(&lli->lli_sa_lock);
+ if (!lli->lli_opendir_key && !lli->lli_sai) {
+ /*
+ * if lli_sai is not NULL, it means previous statahead is not
+ * finished yet, we'd better not start a new statahead for now.
+ */
+ LASSERT(lli->lli_opendir_pid == 0);
+ lli->lli_opendir_key = key;
+ lli->lli_opendir_pid = current->pid;
+ lli->lli_sa_enabled = 1;
+ }
+ spin_unlock(&lli->lli_sa_lock);
+}
+
+/*
+ * deauthorize opened dir handle @key to statahead, and notify statahead thread
+ * to quit if it's running.
*/
-void ll_stop_statahead(struct inode *dir, void *key)
+void ll_deauthorize_statahead(struct inode *dir, void *key)
{
- struct ll_inode_info *lli = ll_i2info(dir);
-
- if (unlikely(key == NULL))
- return;
-
- cfs_spin_lock(&lli->lli_sa_lock);
- if (lli->lli_opendir_key != key || lli->lli_opendir_pid == 0) {
- cfs_spin_unlock(&lli->lli_sa_lock);
- return;
- }
-
- lli->lli_opendir_key = NULL;
-
- if (lli->lli_sai) {
- struct l_wait_info lwi = { 0 };
- struct ptlrpc_thread *thread = &lli->lli_sai->sai_thread;
-
- if (!sa_is_stopped(lli->lli_sai)) {
- thread->t_flags = SVC_STOPPING;
- cfs_spin_unlock(&lli->lli_sa_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,
- sa_is_stopped(lli->lli_sai),
- &lwi);
- } else {
- cfs_spin_unlock(&lli->lli_sa_lock);
- }
-
- /*
- * Put the ref which was held when first statahead_enter.
- * It maybe not the last ref for some statahead requests
- * maybe inflight.
- */
- ll_sai_put(lli->lli_sai);
- } else {
- lli->lli_opendir_pid = 0;
- cfs_spin_unlock(&lli->lli_sa_lock);
- }
+ struct ll_inode_info *lli = ll_i2info(dir);
+ struct ll_statahead_info *sai;
+
+ LASSERT(lli->lli_opendir_key == key);
+ LASSERT(lli->lli_opendir_pid != 0);
+
+ CDEBUG(D_READA, "deauthorize statahead for "DFID"\n",
+ PFID(&lli->lli_fid));
+
+ spin_lock(&lli->lli_sa_lock);
+ lli->lli_opendir_key = NULL;
+ lli->lli_opendir_pid = 0;
+ lli->lli_sa_enabled = 0;
+ sai = lli->lli_sai;
+ if (sai && sai->sai_task) {
+ /*
+ * statahead thread may not have quit yet because it needs to
+ * cache entries, now it's time to tell it to quit.
+ *
+ * wake_up_process() provides the necessary barriers
+ * to pair with set_current_state().
+ */
+ struct task_struct *task = sai->sai_task;
+
+ sai->sai_task = NULL;
+ wake_up_process(task);
+ }
+ spin_unlock(&lli->lli_sa_lock);
}
enum {
- /**
- * 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
+ /**
+ * not first dirent, or is "."
+ */
+ LS_NOT_FIRST_DE = 0,
+ /**
+ * the first non-hidden dirent
+ */
+ LS_FIRST_DE,
+ /**
+ * the first hidden dirent, that is "."
+ */
+ LS_FIRST_DOT_DE
};
+/* file is first dirent under @dir */
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;
- __u64 pos = 0;
- int dot_de;
- int rc = LS_NONE_FIRST_DE;
- ENTRY;
-
- lli->lli_sa_pos = 0;
- ll_dir_chain_init(&chain);
- page = ll_get_dir_page(NULL, dir, pos, 0, &chain);
-
- while (1) {
- struct lu_dirpage *dp;
- 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 "LPU64": "
- "[rc %d] [parent %u]\n",
- PFID(ll_inode2fid(dir)), pos,
- rc, lli->lli_opendir_pid);
- break;
- }
-
- dp = page_address(page);
- for (ent = lu_dirent_start(dp); ent != NULL;
- ent = lu_dirent_next(ent)) {
- int namelen;
- char *name;
-
- 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)
- /*
- * skip "."
- */
- continue;
- else if (name[1] == '.' && namelen == 2)
- /*
- * skip ".."
- */
- continue;
- else
- dot_de = 1;
- } else {
- dot_de = 0;
- }
-
- if (dot_de && target->name[0] != '.') {
- CDEBUG(D_READA, "%.*s skip hidden file %.*s\n",
- target->len, target->name,
- namelen, name);
- continue;
- }
-
- 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 == MDS_DIR_END_OFF) {
- /*
- * End of directory reached.
- */
- break;
- } else if (1) {
- /*
- * chain is exhausted
- * Normal case: continue to the next page.
- */
- lli->lli_sa_pos = pos;
- page = ll_get_dir_page(NULL, dir, pos, 1, &chain);
- } else {
- /*
- * go into overflow page.
- */
- }
- }
- EXIT;
+ struct qstr *target = &dentry->d_name;
+ struct md_op_data *op_data;
+ int dot_de;
+ struct page *page = NULL;
+ int rc = LS_NOT_FIRST_DE;
+ __u64 pos = 0;
+
+ ENTRY;
+
+ op_data = ll_prep_md_op_data(NULL, dir, dir, NULL, 0, 0,
+ LUSTRE_OPC_ANY, dir);
+ if (IS_ERR(op_data))
+ RETURN(PTR_ERR(op_data));
+ /**
+ *FIXME choose the start offset of the readdir
+ */
+
+ page = ll_get_dir_page(dir, op_data, 0);
+
+ while (1) {
+ struct lu_dirpage *dp;
+ struct lu_dirent *ent;
+
+ if (IS_ERR(page)) {
+ struct ll_inode_info *lli = ll_i2info(dir);
+
+ rc = PTR_ERR(page);
+ CERROR("%s: reading dir "DFID" at %llu opendir_pid = %u : rc = %d\n",
+ ll_i2sbi(dir)->ll_fsname,
+ PFID(ll_inode2fid(dir)), pos,
+ lli->lli_opendir_pid, rc);
+ break;
+ }
+
+ dp = page_address(page);
+ for (ent = lu_dirent_start(dp); ent != NULL;
+ ent = lu_dirent_next(ent)) {
+ __u64 hash;
+ int namelen;
+ char *name;
+
+ hash = le64_to_cpu(ent->lde_hash);
+ /*
+ * The ll_get_dir_page() can return any page containing
+ * the given hash which may be not the start hash.
+ */
+ if (unlikely(hash < pos))
+ 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)
+ /*
+ * skip "."
+ */
+ continue;
+ else if (name[1] == '.' && namelen == 2)
+ /*
+ * skip ".."
+ */
+ continue;
+ else
+ dot_de = 1;
+ } else {
+ dot_de = 0;
+ }
+
+ if (dot_de && target->name[0] != '.') {
+ CDEBUG(D_READA, "%.*s skip hidden file %.*s\n",
+ target->len, target->name,
+ namelen, name);
+ continue;
+ }
+
+ if (target->len != namelen ||
+ memcmp(target->name, name, namelen) != 0)
+ rc = LS_NOT_FIRST_DE;
+ else if (!dot_de)
+ rc = LS_FIRST_DE;
+ else
+ rc = LS_FIRST_DOT_DE;
+
+ ll_release_page(dir, page, false);
+ GOTO(out, rc);
+ }
+ pos = le64_to_cpu(dp->ldp_hash_end);
+ if (pos == MDS_DIR_END_OFF) {
+ /*
+ * End of directory reached.
+ */
+ ll_release_page(dir, page, false);
+ GOTO(out, rc);
+ } else {
+ /*
+ * chain is exhausted
+ * Normal case: continue to the next page.
+ */
+ ll_release_page(dir, page, le32_to_cpu(dp->ldp_flags) &
+ LDF_COLLIDE);
+ page = ll_get_dir_page(dir, op_data, pos);
+ }
+ }
+ EXIT;
+out:
+ ll_finish_md_op_data(op_data);
+
+ return rc;
+}
+/**
+ * revalidate @dentryp from statahead cache
+ *
+ * \param[in] dir parent directory
+ * \param[in] sai sai structure
+ * \param[out] dentryp pointer to dentry which will be revalidated
+ * \param[in] unplug unplug statahead window only (normally for negative
+ * dentry)
+ * \retval 1 on success, dentry is saved in @dentryp
+ * \retval 0 if revalidation failed (no proper lock on client)
+ * \retval negative number upon error
+ */
+static int revalidate_statahead_dentry(struct inode *dir,
+ struct ll_statahead_info *sai,
+ struct dentry **dentryp,
+ bool unplug)
+{
+ struct sa_entry *entry = NULL;
+ struct ll_dentry_data *ldd;
+ struct ll_inode_info *lli = ll_i2info(dir);
+ int rc = 0;
+
+ ENTRY;
+
+ if ((*dentryp)->d_name.name[0] == '.') {
+ if (sai->sai_ls_all ||
+ sai->sai_miss_hidden >= sai->sai_skip_hidden) {
+ /*
+ * Hidden dentry is the first one, or statahead
+ * thread does not skip so many hidden dentries
+ * before "sai_ls_all" enabled as below.
+ */
+ } else {
+ if (!sai->sai_ls_all)
+ /*
+ * 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->sai_ls_all = 1;
+
+ /*
+ * Such "getattr" has been skipped before
+ * "sai_ls_all" enabled as above.
+ */
+ sai->sai_miss_hidden++;
+ RETURN(-EAGAIN);
+ }
+ }
+
+ if (unplug)
+ GOTO(out, rc = 1);
+
+ entry = sa_get(sai, &(*dentryp)->d_name);
+ if (!entry)
+ GOTO(out, rc = -EAGAIN);
+
+ /* if statahead is busy in readdir, help it do post-work */
+ if (!sa_ready(entry) && sai->sai_in_readpage)
+ sa_handle_callback(sai);
+
+ if (!sa_ready(entry)) {
+ spin_lock(&lli->lli_sa_lock);
+ sai->sai_index_wait = entry->se_index;
+ spin_unlock(&lli->lli_sa_lock);
+ rc = wait_event_idle_timeout(sai->sai_waitq, sa_ready(entry),
+ cfs_time_seconds(30));
+ if (rc == 0) {
+ /*
+ * entry may not be ready, so it may be used by inflight
+ * statahead RPC, don't free it.
+ */
+ entry = NULL;
+ GOTO(out, rc = -EAGAIN);
+ }
+ }
+
+ /*
+ * We need to see the value that was set immediately before we
+ * were woken up.
+ */
+ if (smp_load_acquire(&entry->se_state) == SA_ENTRY_SUCC &&
+ entry->se_inode) {
+ struct inode *inode = entry->se_inode;
+ struct lookup_intent it = { .it_op = IT_GETATTR,
+ .it_lock_handle =
+ entry->se_handle };
+ __u64 bits;
+
+ rc = md_revalidate_lock(ll_i2mdexp(dir), &it,
+ ll_inode2fid(inode), &bits);
+ if (rc == 1) {
+ if (!(*dentryp)->d_inode) {
+ struct dentry *alias;
+
+ alias = ll_splice_alias(inode, *dentryp);
+ if (IS_ERR(alias)) {
+ ll_intent_release(&it);
+ GOTO(out, rc = PTR_ERR(alias));
+ }
+ *dentryp = alias;
+ /*
+ * statahead prepared this inode, transfer inode
+ * refcount from sa_entry to dentry
+ */
+ entry->se_inode = NULL;
+ } else if ((*dentryp)->d_inode != inode) {
+ /* revalidate, but inode is recreated */
+ CDEBUG(D_READA,
+ "%s: stale dentry %pd inode " DFID", statahead inode "DFID "\n",
+ ll_i2sbi(inode)->ll_fsname, *dentryp,
+ PFID(ll_inode2fid((*dentryp)->d_inode)),
+ PFID(ll_inode2fid(inode)));
+ ll_intent_release(&it);
+ GOTO(out, rc = -ESTALE);
+ }
+
+ if ((bits & MDS_INODELOCK_LOOKUP) &&
+ d_lustre_invalid(*dentryp))
+ d_lustre_revalidate(*dentryp);
+ ll_intent_release(&it);
+ }
+ }
out:
- ll_dir_chain_fini(&chain);
- return rc;
+ /*
+ * statahead cached sa_entry can be used only once, and will be killed
+ * right after use, so if lookup/revalidate accessed statahead cache,
+ * set dentry ldd_sa_generation to parent lli_sa_generation, later if we
+ * stat this file again, we know we've done statahead before, see
+ * dentry_may_statahead().
+ */
+ ldd = ll_d2d(*dentryp);
+ /* ldd can be NULL if llite lookup failed. */
+ if (ldd)
+ ldd->lld_sa_generation = lli->lli_sa_generation;
+ sa_put(sai, entry);
+ spin_lock(&lli->lli_sa_lock);
+ if (sai->sai_task)
+ wake_up_process(sai->sai_task);
+ spin_unlock(&lli->lli_sa_lock);
+
+ RETURN(rc);
}
/**
- * 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, 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
- * \retval others -- error
+ * start statahead thread
+ *
+ * \param[in] dir parent directory
+ * \param[in] dentry dentry that triggers statahead, normally the first
+ * dirent under @dir
+ * \param[in] agl indicate whether AGL is needed
+ * \retval -EAGAIN on success, because when this function is
+ * called, it's already in lookup call, so client should
+ * do it itself instead of waiting for statahead thread
+ * to do it asynchronously.
+ * \retval negative number upon error
*/
-int do_statahead_enter(struct inode *dir, struct dentry **dentryp, int lookup)
+static int start_statahead_thread(struct inode *dir, struct dentry *dentry,
+ bool agl)
{
- struct ll_inode_info *lli;
- struct ll_statahead_info *sai;
- struct dentry *parent;
- struct l_wait_info lwi = { 0 };
- 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(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
- * thread does not skip so many hidden dentries
- * before "sai_ls_all" enabled as below.
- */
- } else {
- if (!sai->sai_ls_all)
- /*
- * 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->sai_ls_all = 1;
-
- /*
- * 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 (lookup) {
- struct dentry *result;
-
- result = d_lookup((*dentryp)->d_parent,
- &(*dentryp)->d_name);
- if (result) {
- LASSERT(result != *dentryp);
- /* BUG 16303: do not drop reference count for
- * "*dentryp", VFS will do that by itself. */
- *dentryp = result;
- RETURN(1);
- }
- }
- /*
- * do nothing for revalidate.
- */
- RETURN(0);
- }
-
- /* 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. */
- GOTO(out, rc = -EAGAIN);
-
- sai = ll_sai_alloc();
- if (sai == NULL)
- GOTO(out, rc = -ENOMEM);
-
- 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);
- }
-
- /* 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_create_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(-EAGAIN);
- }
-
- 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.
- */
- RETURN(-EEXIST);
+ int node = cfs_cpt_spread_node(cfs_cpt_tab, CFS_CPT_ANY);
+ struct ll_inode_info *lli = ll_i2info(dir);
+ struct ll_statahead_info *sai = NULL;
+ struct dentry *parent = dentry->d_parent;
+ struct task_struct *task;
+ struct ll_sb_info *sbi = ll_i2sbi(parent->d_inode);
+ int first = LS_FIRST_DE;
+ int rc = 0;
+
+ ENTRY;
+
+ /* I am the "lli_opendir_pid" owner, only me can set "lli_sai". */
+ first = is_first_dirent(dir, dentry);
+ if (first == LS_NOT_FIRST_DE)
+ /* It is not "ls -{a}l" operation, no need statahead for it. */
+ GOTO(out, rc = -EFAULT);
+
+ if (unlikely(atomic_inc_return(&sbi->ll_sa_running) >
+ sbi->ll_sa_running_max)) {
+ CDEBUG(D_READA,
+ "Too many concurrent statahead instances, avoid new statahead instance temporarily.\n");
+ GOTO(out, rc = -EMFILE);
+ }
+
+ sai = ll_sai_alloc(parent);
+ if (!sai)
+ GOTO(out, rc = -ENOMEM);
+
+ sai->sai_ls_all = (first == LS_FIRST_DOT_DE);
+
+ /*
+ * if current lli_opendir_key was deauthorized, or dir re-opened by
+ * another process, don't start statahead, otherwise the newly spawned
+ * statahead thread won't be notified to quit.
+ */
+ spin_lock(&lli->lli_sa_lock);
+ if (unlikely(lli->lli_sai || !lli->lli_opendir_key ||
+ lli->lli_opendir_pid != current->pid)) {
+ spin_unlock(&lli->lli_sa_lock);
+ GOTO(out, rc = -EPERM);
+ }
+ lli->lli_sai = sai;
+ spin_unlock(&lli->lli_sa_lock);
+
+ CDEBUG(D_READA, "start statahead thread: [pid %d] [parent %pd]\n",
+ current->pid, parent);
+
+ task = kthread_create_on_node(ll_statahead_thread, parent, node,
+ "ll_sa_%u", lli->lli_opendir_pid);
+ if (IS_ERR(task)) {
+ spin_lock(&lli->lli_sa_lock);
+ lli->lli_sai = NULL;
+ spin_unlock(&lli->lli_sa_lock);
+ rc = PTR_ERR(task);
+ CERROR("can't start ll_sa thread, rc: %d\n", rc);
+ GOTO(out, rc);
+ }
+
+ if (ll_i2sbi(parent->d_inode)->ll_flags & LL_SBI_AGL_ENABLED && agl)
+ ll_start_agl(parent, sai);
+
+ atomic_inc(&ll_i2sbi(parent->d_inode)->ll_sa_total);
+ sai->sai_task = task;
+
+ wake_up_process(task);
+ /*
+ * We don't stat-ahead for the first dirent since we are already in
+ * lookup.
+ */
+ RETURN(-EAGAIN);
out:
- cfs_spin_lock(&lli->lli_sa_lock);
- lli->lli_opendir_key = NULL;
- lli->lli_opendir_pid = 0;
- cfs_spin_unlock(&lli->lli_sa_lock);
- return rc;
+ /*
+ * once we start statahead thread failed, disable statahead so that
+ * subsequent stat won't waste time to try it.
+ */
+ spin_lock(&lli->lli_sa_lock);
+ if (lli->lli_opendir_pid == current->pid)
+ lli->lli_sa_enabled = 0;
+ spin_unlock(&lli->lli_sa_lock);
+
+ if (sai)
+ ll_sai_free(sai);
+ if (first != LS_NOT_FIRST_DE)
+ atomic_dec(&sbi->ll_sa_running);
+
+ RETURN(rc);
+}
+
+/*
+ * Check whether statahead for @dir was started.
+ */
+static inline bool ll_statahead_started(struct inode *dir, bool agl)
+{
+ struct ll_inode_info *lli = ll_i2info(dir);
+ struct ll_statahead_info *sai;
+
+ spin_lock(&lli->lli_sa_lock);
+ sai = lli->lli_sai;
+ if (sai && (sai->sai_agl_task != NULL) != agl)
+ CDEBUG(D_READA,
+ "%s: Statahead AGL hint changed from %d to %d\n",
+ ll_i2sbi(dir)->ll_fsname,
+ sai->sai_agl_task != NULL, agl);
+ spin_unlock(&lli->lli_sa_lock);
+
+ return !!sai;
}
/**
- * update hit/miss count.
+ * statahead entry function, this is called when client getattr on a file, it
+ * will start statahead thread if this is the first dir entry, else revalidate
+ * dentry from statahead cache.
+ *
+ * \param[in] dir parent directory
+ * \param[out] dentryp dentry to getattr
+ * \param[in] agl whether start the agl thread
+ *
+ * \retval 1 on success
+ * \retval 0 revalidation from statahead cache failed, caller needs
+ * to getattr from server directly
+ * \retval negative number on error, caller often ignores this and
+ * then getattr from server
+ */
+int ll_start_statahead(struct inode *dir, struct dentry *dentry, bool agl)
+{
+ if (!ll_statahead_started(dir, agl))
+ return start_statahead_thread(dir, dentry, agl);
+ return 0;
+}
+
+/**
+ * revalidate dentry from statahead cache.
+ *
+ * \param[in] dir parent directory
+ * \param[out] dentryp dentry to getattr
+ * \param[in] unplug unplug statahead window only (normally for negative
+ * dentry)
+ * \retval 1 on success
+ * \retval 0 revalidation from statahead cache failed, caller needs
+ * to getattr from server directly
+ * \retval negative number on error, caller often ignores this and
+ * then getattr from server
*/
-void ll_statahead_exit(struct inode *dir, struct dentry *dentry, int result)
+int ll_revalidate_statahead(struct inode *dir, struct dentry **dentryp,
+ bool unplug)
{
- 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;
-
- 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_sa_lock);
- if (!sa_is_stopped(sai))
- sai->sai_thread.t_flags = SVC_STOPPING;
- cfs_spin_unlock(&lli->lli_sa_lock);
- }
- }
-
- if (!sa_is_stopped(sai))
- cfs_waitq_signal(&sai->sai_thread.t_ctl_waitq);
- if (likely(ldd != NULL))
- ldd->lld_sa_generation = sai->sai_generation;
-
- EXIT;
+ struct ll_statahead_info *sai;
+ int rc = 0;
+
+ sai = ll_sai_get(dir);
+ if (sai) {
+ rc = revalidate_statahead_dentry(dir, sai, dentryp, unplug);
+ CDEBUG(D_READA, "revalidate statahead %pd: rc = %d.\n",
+ *dentryp, rc);
+ ll_sai_put(sai);
+ }
+ return rc;
}