4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2011, 2016, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
34 #include <linux/sched.h>
35 #include <linux/kthread.h>
37 #include <linux/highmem.h>
38 #include <linux/pagemap.h>
40 #define DEBUG_SUBSYSTEM S_LLITE
42 #include <obd_support.h>
43 #include <lustre_dlm.h>
44 #include "llite_internal.h"
46 #define SA_OMITTED_ENTRY_MAX 8ULL
49 /** negative values are for error cases */
50 SA_ENTRY_INIT = 0, /** init entry */
51 SA_ENTRY_SUCC = 1, /** stat succeed */
52 SA_ENTRY_INVA = 2, /** invalid entry */
55 /* sa_entry is not refcounted: statahead thread allocates it and do async stat,
56 * and in async stat callback ll_statahead_interpret() will add it into
57 * sai_interim_entries, later statahead thread will call sa_handle_callback() to
58 * instantiate entry and move it into sai_entries, and then only scanner process
59 * can access and free it. */
61 /* link into sai_interim_entries or sai_entries */
62 struct list_head se_list;
63 /* link into sai hash table locally */
64 struct list_head se_hash;
65 /* entry index in the sai */
67 /* low layer ldlm lock handle */
71 /* entry size, contains name */
73 /* pointer to async getattr enqueue info */
74 struct md_enqueue_info *se_minfo;
75 /* pointer to the async getattr request */
76 struct ptlrpc_request *se_req;
77 /* pointer to the target inode */
78 struct inode *se_inode;
85 static unsigned int sai_generation = 0;
86 static DEFINE_SPINLOCK(sai_generation_lock);
88 static inline int sa_unhashed(struct sa_entry *entry)
90 return list_empty(&entry->se_hash);
93 /* sa_entry is ready to use */
94 static inline int sa_ready(struct sa_entry *entry)
97 return (entry->se_state != SA_ENTRY_INIT);
100 /* hash value to put in sai_cache */
101 static inline int sa_hash(int val)
103 return val & LL_SA_CACHE_MASK;
106 /* hash entry into sai_cache */
108 sa_rehash(struct ll_statahead_info *sai, struct sa_entry *entry)
110 int i = sa_hash(entry->se_qstr.hash);
112 spin_lock(&sai->sai_cache_lock[i]);
113 list_add_tail(&entry->se_hash, &sai->sai_cache[i]);
114 spin_unlock(&sai->sai_cache_lock[i]);
117 /* unhash entry from sai_cache */
119 sa_unhash(struct ll_statahead_info *sai, struct sa_entry *entry)
121 int i = sa_hash(entry->se_qstr.hash);
123 spin_lock(&sai->sai_cache_lock[i]);
124 list_del_init(&entry->se_hash);
125 spin_unlock(&sai->sai_cache_lock[i]);
128 static inline int agl_should_run(struct ll_statahead_info *sai,
131 return (inode != NULL && S_ISREG(inode->i_mode) && sai->sai_agl_valid);
134 static inline struct ll_inode_info *
135 agl_first_entry(struct ll_statahead_info *sai)
137 return list_entry(sai->sai_agls.next, struct ll_inode_info,
141 /* statahead window is full */
142 static inline int sa_sent_full(struct ll_statahead_info *sai)
144 return atomic_read(&sai->sai_cache_count) >= sai->sai_max;
147 /* got async stat replies */
148 static inline int sa_has_callback(struct ll_statahead_info *sai)
150 return !list_empty(&sai->sai_interim_entries);
153 static inline int agl_list_empty(struct ll_statahead_info *sai)
155 return list_empty(&sai->sai_agls);
159 * (1) hit ratio less than 80%
161 * (2) consecutive miss more than 8
162 * then means low hit.
164 static inline int sa_low_hit(struct ll_statahead_info *sai)
166 return ((sai->sai_hit > 7 && sai->sai_hit < 4 * sai->sai_miss) ||
167 (sai->sai_consecutive_miss > 8));
171 * if the given index is behind of statahead window more than
172 * SA_OMITTED_ENTRY_MAX, then it is old.
174 static inline int is_omitted_entry(struct ll_statahead_info *sai, __u64 index)
176 return ((__u64)sai->sai_max + index + SA_OMITTED_ENTRY_MAX <
180 /* allocate sa_entry and hash it to allow scanner process to find it */
181 static struct sa_entry *
182 sa_alloc(struct ll_statahead_info *sai, __u64 index, const char *name, int len,
183 const struct lu_fid *fid)
185 struct ll_inode_info *lli;
186 struct sa_entry *entry;
191 entry_size = sizeof(struct sa_entry) + (len & ~3) + 4;
192 OBD_ALLOC(entry, entry_size);
193 if (unlikely(entry == NULL))
194 RETURN(ERR_PTR(-ENOMEM));
196 CDEBUG(D_READA, "alloc sa entry %.*s(%p) index %llu\n",
197 len, name, entry, index);
199 entry->se_index = index;
201 entry->se_state = SA_ENTRY_INIT;
202 entry->se_size = entry_size;
203 dname = (char *)entry + sizeof(struct sa_entry);
204 memcpy(dname, name, len);
206 entry->se_qstr.hash = full_name_hash(name, len);
207 entry->se_qstr.len = len;
208 entry->se_qstr.name = dname;
209 entry->se_fid = *fid;
211 lli = ll_i2info(sai->sai_dentry->d_inode);
213 spin_lock(&lli->lli_sa_lock);
214 INIT_LIST_HEAD(&entry->se_list);
215 sa_rehash(sai, entry);
216 spin_unlock(&lli->lli_sa_lock);
218 atomic_inc(&sai->sai_cache_count);
223 /* free sa_entry, which should have been unhashed and not in any list */
224 static void sa_free(struct ll_statahead_info *sai, struct sa_entry *entry)
226 CDEBUG(D_READA, "free sa entry %.*s(%p) index %llu\n",
227 entry->se_qstr.len, entry->se_qstr.name, entry,
230 LASSERT(list_empty(&entry->se_list));
231 LASSERT(sa_unhashed(entry));
233 OBD_FREE(entry, entry->se_size);
234 atomic_dec(&sai->sai_cache_count);
238 * find sa_entry by name, used by directory scanner, lock is not needed because
239 * only scanner can remove the entry from cache.
241 static struct sa_entry *
242 sa_get(struct ll_statahead_info *sai, const struct qstr *qstr)
244 struct sa_entry *entry;
245 int i = sa_hash(qstr->hash);
247 list_for_each_entry(entry, &sai->sai_cache[i], se_hash) {
248 if (entry->se_qstr.hash == qstr->hash &&
249 entry->se_qstr.len == qstr->len &&
250 memcmp(entry->se_qstr.name, qstr->name, qstr->len) == 0)
256 /* unhash and unlink sa_entry, and then free it */
258 sa_kill(struct ll_statahead_info *sai, struct sa_entry *entry)
260 struct ll_inode_info *lli = ll_i2info(sai->sai_dentry->d_inode);
262 LASSERT(!sa_unhashed(entry));
263 LASSERT(!list_empty(&entry->se_list));
264 LASSERT(sa_ready(entry));
266 sa_unhash(sai, entry);
268 spin_lock(&lli->lli_sa_lock);
269 list_del_init(&entry->se_list);
270 spin_unlock(&lli->lli_sa_lock);
272 if (entry->se_inode != NULL)
273 iput(entry->se_inode);
278 /* called by scanner after use, sa_entry will be killed */
280 sa_put(struct ll_statahead_info *sai, struct sa_entry *entry)
282 struct sa_entry *tmp, *next;
284 if (entry != NULL && entry->se_state == SA_ENTRY_SUCC) {
285 struct ll_sb_info *sbi = ll_i2sbi(sai->sai_dentry->d_inode);
288 sai->sai_consecutive_miss = 0;
289 sai->sai_max = min(2 * sai->sai_max, sbi->ll_sa_max);
292 sai->sai_consecutive_miss++;
298 /* kill old completed entries, only scanner process does this, no need
300 list_for_each_entry_safe(tmp, next, &sai->sai_entries, se_list) {
301 if (!is_omitted_entry(sai, tmp->se_index))
306 wake_up(&sai->sai_thread.t_ctl_waitq);
309 /* update state and sort add entry to sai_entries by index, return true if
310 * scanner is waiting on this entry. */
312 __sa_make_ready(struct ll_statahead_info *sai, struct sa_entry *entry, int ret)
315 struct list_head *pos = &sai->sai_entries;
316 __u64 index = entry->se_index;
318 LASSERT(!sa_ready(entry));
319 LASSERT(list_empty(&entry->se_list));
321 list_for_each_entry_reverse(se, &sai->sai_entries, se_list) {
322 if (se->se_index < entry->se_index) {
327 list_add(&entry->se_list, pos);
328 entry->se_state = ret < 0 ? SA_ENTRY_INVA : SA_ENTRY_SUCC;
330 return (index == sai->sai_index_wait);
334 * release resources used in async stat RPC, update entry state and wakeup if
335 * scanner process it waiting on this entry.
338 sa_make_ready(struct ll_statahead_info *sai, struct sa_entry *entry, int ret)
340 struct ll_inode_info *lli = ll_i2info(sai->sai_dentry->d_inode);
341 struct md_enqueue_info *minfo = entry->se_minfo;
342 struct ptlrpc_request *req = entry->se_req;
345 /* release resources used in RPC */
347 entry->se_minfo = NULL;
348 ll_intent_release(&minfo->mi_it);
354 entry->se_req = NULL;
355 ptlrpc_req_finished(req);
358 spin_lock(&lli->lli_sa_lock);
359 wakeup = __sa_make_ready(sai, entry, ret);
360 spin_unlock(&lli->lli_sa_lock);
363 wake_up(&sai->sai_waitq);
366 /* insert inode into the list of sai_agls */
367 static void ll_agl_add(struct ll_statahead_info *sai,
368 struct inode *inode, int index)
370 struct ll_inode_info *child = ll_i2info(inode);
371 struct ll_inode_info *parent = ll_i2info(sai->sai_dentry->d_inode);
374 spin_lock(&child->lli_agl_lock);
375 if (child->lli_agl_index == 0) {
376 child->lli_agl_index = index;
377 spin_unlock(&child->lli_agl_lock);
379 LASSERT(list_empty(&child->lli_agl_list));
382 spin_lock(&parent->lli_agl_lock);
383 if (agl_list_empty(sai))
385 list_add_tail(&child->lli_agl_list, &sai->sai_agls);
386 spin_unlock(&parent->lli_agl_lock);
388 spin_unlock(&child->lli_agl_lock);
392 wake_up(&sai->sai_agl_thread.t_ctl_waitq);
396 static struct ll_statahead_info *ll_sai_alloc(struct dentry *dentry)
398 struct ll_statahead_info *sai;
399 struct ll_inode_info *lli = ll_i2info(dentry->d_inode);
407 sai->sai_dentry = dget(dentry);
408 atomic_set(&sai->sai_refcount, 1);
409 sai->sai_max = LL_SA_RPC_MIN;
411 init_waitqueue_head(&sai->sai_waitq);
412 init_waitqueue_head(&sai->sai_thread.t_ctl_waitq);
413 init_waitqueue_head(&sai->sai_agl_thread.t_ctl_waitq);
415 INIT_LIST_HEAD(&sai->sai_interim_entries);
416 INIT_LIST_HEAD(&sai->sai_entries);
417 INIT_LIST_HEAD(&sai->sai_agls);
419 for (i = 0; i < LL_SA_CACHE_SIZE; i++) {
420 INIT_LIST_HEAD(&sai->sai_cache[i]);
421 spin_lock_init(&sai->sai_cache_lock[i]);
423 atomic_set(&sai->sai_cache_count, 0);
425 spin_lock(&sai_generation_lock);
426 lli->lli_sa_generation = ++sai_generation;
427 if (unlikely(sai_generation == 0))
428 lli->lli_sa_generation = ++sai_generation;
429 spin_unlock(&sai_generation_lock);
435 static inline void ll_sai_free(struct ll_statahead_info *sai)
437 LASSERT(sai->sai_dentry != NULL);
438 dput(sai->sai_dentry);
443 * take refcount of sai if sai for @dir exists, which means statahead is on for
446 static inline struct ll_statahead_info *ll_sai_get(struct inode *dir)
448 struct ll_inode_info *lli = ll_i2info(dir);
449 struct ll_statahead_info *sai = NULL;
451 spin_lock(&lli->lli_sa_lock);
454 atomic_inc(&sai->sai_refcount);
455 spin_unlock(&lli->lli_sa_lock);
461 * put sai refcount after use, if refcount reaches zero, free sai and sa_entries
464 static void ll_sai_put(struct ll_statahead_info *sai)
466 struct ll_inode_info *lli = ll_i2info(sai->sai_dentry->d_inode);
468 if (atomic_dec_and_lock(&sai->sai_refcount, &lli->lli_sa_lock)) {
469 struct sa_entry *entry, *next;
470 struct ll_sb_info *sbi = ll_i2sbi(sai->sai_dentry->d_inode);
473 spin_unlock(&lli->lli_sa_lock);
475 LASSERT(thread_is_stopped(&sai->sai_thread));
476 LASSERT(thread_is_stopped(&sai->sai_agl_thread));
477 LASSERT(sai->sai_sent == sai->sai_replied);
478 LASSERT(!sa_has_callback(sai));
480 list_for_each_entry_safe(entry, next, &sai->sai_entries,
484 LASSERT(atomic_read(&sai->sai_cache_count) == 0);
485 LASSERT(agl_list_empty(sai));
488 atomic_dec(&sbi->ll_sa_running);
492 /* Do NOT forget to drop inode refcount when into sai_agls. */
493 static void ll_agl_trigger(struct inode *inode, struct ll_statahead_info *sai)
495 struct ll_inode_info *lli = ll_i2info(inode);
496 __u64 index = lli->lli_agl_index;
500 LASSERT(list_empty(&lli->lli_agl_list));
502 /* AGL maybe fall behind statahead with one entry */
503 if (is_omitted_entry(sai, index + 1)) {
504 lli->lli_agl_index = 0;
509 /* In case of restore, the MDT has the right size and has already
510 * sent it back without granting the layout lock, inode is up-to-date.
511 * Then AGL (async glimpse lock) is useless.
512 * Also to glimpse we need the layout, in case of a runninh restore
513 * the MDT holds the layout lock so the glimpse will block up to the
514 * end of restore (statahead/agl will block) */
515 if (ll_file_test_flag(lli, LLIF_FILE_RESTORING)) {
516 lli->lli_agl_index = 0;
521 /* Someone is in glimpse (sync or async), do nothing. */
522 rc = down_write_trylock(&lli->lli_glimpse_sem);
524 lli->lli_agl_index = 0;
530 * Someone triggered glimpse within 1 sec before.
531 * 1) The former glimpse succeeded with glimpse lock granted by OST, and
532 * if the lock is still cached on client, AGL needs to do nothing. If
533 * it is cancelled by other client, AGL maybe cannot obtaion new lock
534 * for no glimpse callback triggered by AGL.
535 * 2) The former glimpse succeeded, but OST did not grant glimpse lock.
536 * Under such case, it is quite possible that the OST will not grant
537 * glimpse lock for AGL also.
538 * 3) The former glimpse failed, compared with other two cases, it is
539 * relative rare. AGL can ignore such case, and it will not muchly
540 * affect the performance.
542 if (lli->lli_glimpse_time != 0 &&
543 cfs_time_before(cfs_time_shift(-1), lli->lli_glimpse_time)) {
544 up_write(&lli->lli_glimpse_sem);
545 lli->lli_agl_index = 0;
550 CDEBUG(D_READA, "Handling (init) async glimpse: inode = "
551 DFID", idx = %llu\n", PFID(&lli->lli_fid), index);
554 lli->lli_agl_index = 0;
555 lli->lli_glimpse_time = cfs_time_current();
556 up_write(&lli->lli_glimpse_sem);
558 CDEBUG(D_READA, "Handled (init) async glimpse: inode= "
559 DFID", idx = %llu, rc = %d\n",
560 PFID(&lli->lli_fid), index, rc);
568 * prepare inode for sa entry, add it into agl list, now sa_entry is ready
569 * to be used by scanner process.
571 static void sa_instantiate(struct ll_statahead_info *sai,
572 struct sa_entry *entry)
574 struct inode *dir = sai->sai_dentry->d_inode;
576 struct md_enqueue_info *minfo;
577 struct lookup_intent *it;
578 struct ptlrpc_request *req;
579 struct mdt_body *body;
583 LASSERT(entry->se_handle != 0);
585 minfo = entry->se_minfo;
588 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
590 GOTO(out, rc = -EFAULT);
592 child = entry->se_inode;
594 /* revalidate; unlinked and re-created with the same name */
595 if (unlikely(!lu_fid_eq(&minfo->mi_data.op_fid2,
597 entry->se_inode = NULL;
603 it->it_lock_handle = entry->se_handle;
604 rc = md_revalidate_lock(ll_i2mdexp(dir), it, ll_inode2fid(dir), NULL);
606 GOTO(out, rc = -EAGAIN);
608 rc = ll_prep_inode(&child, req, dir->i_sb, it);
612 CDEBUG(D_READA, "%s: setting %.*s"DFID" l_data to inode %p\n",
613 ll_get_fsname(child->i_sb, NULL, 0),
614 entry->se_qstr.len, entry->se_qstr.name,
615 PFID(ll_inode2fid(child)), child);
616 ll_set_lock_data(ll_i2sbi(dir)->ll_md_exp, child, it, NULL);
618 entry->se_inode = child;
620 if (agl_should_run(sai, child))
621 ll_agl_add(sai, child, entry->se_index);
626 /* sa_make_ready() will drop ldlm ibits lock refcount by calling
627 * ll_intent_drop_lock() in spite of failures. Do not worry about
628 * calling ll_intent_drop_lock() more than once. */
629 sa_make_ready(sai, entry, rc);
632 /* once there are async stat replies, instantiate sa_entry from replies */
633 static void sa_handle_callback(struct ll_statahead_info *sai)
635 struct ll_inode_info *lli;
637 lli = ll_i2info(sai->sai_dentry->d_inode);
639 while (sa_has_callback(sai)) {
640 struct sa_entry *entry;
642 spin_lock(&lli->lli_sa_lock);
643 if (unlikely(!sa_has_callback(sai))) {
644 spin_unlock(&lli->lli_sa_lock);
647 entry = list_entry(sai->sai_interim_entries.next,
648 struct sa_entry, se_list);
649 list_del_init(&entry->se_list);
650 spin_unlock(&lli->lli_sa_lock);
652 sa_instantiate(sai, entry);
657 * callback for async stat RPC, because this is called in ptlrpcd context, we
658 * only put sa_entry in sai_interim_entries, and wake up statahead thread to
659 * really prepare inode and instantiate sa_entry later.
661 static int ll_statahead_interpret(struct ptlrpc_request *req,
662 struct md_enqueue_info *minfo, int rc)
664 struct lookup_intent *it = &minfo->mi_it;
665 struct inode *dir = minfo->mi_dir;
666 struct ll_inode_info *lli = ll_i2info(dir);
667 struct ll_statahead_info *sai = lli->lli_sai;
668 struct sa_entry *entry = (struct sa_entry *)minfo->mi_cbdata;
670 wait_queue_head_t *waitq = NULL;
673 if (it_disposition(it, DISP_LOOKUP_NEG))
676 /* because statahead thread will wait for all inflight RPC to finish,
677 * sai should be always valid, no need to refcount */
678 LASSERT(sai != NULL);
679 LASSERT(!thread_is_stopped(&sai->sai_thread));
680 LASSERT(entry != NULL);
682 CDEBUG(D_READA, "sa_entry %.*s rc %d\n",
683 entry->se_qstr.len, entry->se_qstr.name, rc);
686 ll_intent_release(it);
690 /* release ibits lock ASAP to avoid deadlock when statahead
691 * thread enqueues lock on parent in readdir and another
692 * process enqueues lock on child with parent lock held, eg.
694 handle = it->it_lock_handle;
695 ll_intent_drop_lock(it);
698 spin_lock(&lli->lli_sa_lock);
700 if (__sa_make_ready(sai, entry, rc))
701 waitq = &sai->sai_waitq;
703 entry->se_minfo = minfo;
704 entry->se_req = ptlrpc_request_addref(req);
705 /* Release the async ibits lock ASAP to avoid deadlock
706 * when statahead thread tries to enqueue lock on parent
707 * for readpage and other tries to enqueue lock on child
708 * with parent's lock held, for example: unlink. */
709 entry->se_handle = handle;
710 if (!sa_has_callback(sai))
711 waitq = &sai->sai_thread.t_ctl_waitq;
713 list_add_tail(&entry->se_list, &sai->sai_interim_entries);
718 spin_unlock(&lli->lli_sa_lock);
723 /* finish async stat RPC arguments */
724 static void sa_fini_data(struct md_enqueue_info *minfo)
731 * prepare arguments for async stat RPC.
733 static struct md_enqueue_info *
734 sa_prep_data(struct inode *dir, struct inode *child, struct sa_entry *entry)
736 struct md_enqueue_info *minfo;
737 struct ldlm_enqueue_info *einfo;
738 struct md_op_data *op_data;
740 OBD_ALLOC_PTR(minfo);
742 return ERR_PTR(-ENOMEM);
744 op_data = ll_prep_md_op_data(&minfo->mi_data, dir, child, NULL, 0, 0,
745 LUSTRE_OPC_ANY, NULL);
746 if (IS_ERR(op_data)) {
748 return (struct md_enqueue_info *)op_data;
752 op_data->op_fid2 = entry->se_fid;
754 minfo->mi_it.it_op = IT_GETATTR;
755 minfo->mi_dir = igrab(dir);
756 minfo->mi_cb = ll_statahead_interpret;
757 minfo->mi_cbdata = entry;
759 einfo = &minfo->mi_einfo;
760 einfo->ei_type = LDLM_IBITS;
761 einfo->ei_mode = it_to_lock_mode(&minfo->mi_it);
762 einfo->ei_cb_bl = ll_md_blocking_ast;
763 einfo->ei_cb_cp = ldlm_completion_ast;
764 einfo->ei_cb_gl = NULL;
765 einfo->ei_cbdata = NULL;
770 /* async stat for file not found in dcache */
771 static int sa_lookup(struct inode *dir, struct sa_entry *entry)
773 struct md_enqueue_info *minfo;
777 minfo = sa_prep_data(dir, NULL, entry);
779 RETURN(PTR_ERR(minfo));
781 rc = md_intent_getattr_async(ll_i2mdexp(dir), minfo);
789 * async stat for file found in dcache, similar to .revalidate
791 * \retval 1 dentry valid, no RPC sent
792 * \retval 0 dentry invalid, will send async stat RPC
793 * \retval negative number upon error
795 static int sa_revalidate(struct inode *dir, struct sa_entry *entry,
796 struct dentry *dentry)
798 struct inode *inode = dentry->d_inode;
799 struct lookup_intent it = { .it_op = IT_GETATTR,
800 .it_lock_handle = 0 };
801 struct md_enqueue_info *minfo;
805 if (unlikely(inode == NULL))
808 if (d_mountpoint(dentry))
811 entry->se_inode = igrab(inode);
812 rc = md_revalidate_lock(ll_i2mdexp(dir), &it, ll_inode2fid(inode),
815 entry->se_handle = it.it_lock_handle;
816 ll_intent_release(&it);
820 minfo = sa_prep_data(dir, inode, entry);
822 entry->se_inode = NULL;
824 RETURN(PTR_ERR(minfo));
827 rc = md_intent_getattr_async(ll_i2mdexp(dir), minfo);
829 entry->se_inode = NULL;
837 /* async stat for file with @name */
838 static void sa_statahead(struct dentry *parent, const char *name, int len,
839 const struct lu_fid *fid)
841 struct inode *dir = parent->d_inode;
842 struct ll_inode_info *lli = ll_i2info(dir);
843 struct ll_statahead_info *sai = lli->lli_sai;
844 struct dentry *dentry = NULL;
845 struct sa_entry *entry;
849 entry = sa_alloc(sai, sai->sai_index, name, len, fid);
853 dentry = d_lookup(parent, &entry->se_qstr);
855 rc = sa_lookup(dir, entry);
857 rc = sa_revalidate(dir, entry, dentry);
858 if (rc == 1 && agl_should_run(sai, dentry->d_inode))
859 ll_agl_add(sai, dentry->d_inode, entry->se_index);
866 sa_make_ready(sai, entry, rc);
875 /* async glimpse (agl) thread main function */
876 static int ll_agl_thread(void *arg)
878 struct dentry *parent = (struct dentry *)arg;
879 struct inode *dir = parent->d_inode;
880 struct ll_inode_info *plli = ll_i2info(dir);
881 struct ll_inode_info *clli;
882 struct ll_sb_info *sbi = ll_i2sbi(dir);
883 struct ll_statahead_info *sai;
884 struct ptlrpc_thread *thread;
885 struct l_wait_info lwi = { 0 };
889 sai = ll_sai_get(dir);
890 thread = &sai->sai_agl_thread;
891 thread->t_pid = current_pid();
892 CDEBUG(D_READA, "agl thread started: sai %p, parent %.*s\n",
893 sai, parent->d_name.len, parent->d_name.name);
895 atomic_inc(&sbi->ll_agl_total);
896 spin_lock(&plli->lli_agl_lock);
897 sai->sai_agl_valid = 1;
898 if (thread_is_init(thread))
899 /* If someone else has changed the thread state
900 * (e.g. already changed to SVC_STOPPING), we can't just
901 * blindly overwrite that setting. */
902 thread_set_flags(thread, SVC_RUNNING);
903 spin_unlock(&plli->lli_agl_lock);
904 wake_up(&thread->t_ctl_waitq);
907 l_wait_event(thread->t_ctl_waitq,
908 !agl_list_empty(sai) ||
909 !thread_is_running(thread),
912 if (!thread_is_running(thread))
915 spin_lock(&plli->lli_agl_lock);
916 /* The statahead thread maybe help to process AGL entries,
917 * so check whether list empty again. */
918 if (!agl_list_empty(sai)) {
919 clli = agl_first_entry(sai);
920 list_del_init(&clli->lli_agl_list);
921 spin_unlock(&plli->lli_agl_lock);
922 ll_agl_trigger(&clli->lli_vfs_inode, sai);
924 spin_unlock(&plli->lli_agl_lock);
928 spin_lock(&plli->lli_agl_lock);
929 sai->sai_agl_valid = 0;
930 while (!agl_list_empty(sai)) {
931 clli = agl_first_entry(sai);
932 list_del_init(&clli->lli_agl_list);
933 spin_unlock(&plli->lli_agl_lock);
934 clli->lli_agl_index = 0;
935 iput(&clli->lli_vfs_inode);
936 spin_lock(&plli->lli_agl_lock);
938 thread_set_flags(thread, SVC_STOPPED);
939 spin_unlock(&plli->lli_agl_lock);
940 wake_up(&thread->t_ctl_waitq);
942 CDEBUG(D_READA, "agl thread stopped: sai %p, parent %.*s\n",
943 sai, parent->d_name.len, parent->d_name.name);
947 /* start agl thread */
948 static void ll_start_agl(struct dentry *parent, struct ll_statahead_info *sai)
950 struct ptlrpc_thread *thread = &sai->sai_agl_thread;
951 struct l_wait_info lwi = { 0 };
952 struct ll_inode_info *plli;
953 struct task_struct *task;
956 CDEBUG(D_READA, "start agl thread: sai %p, parent %.*s\n",
957 sai, parent->d_name.len, parent->d_name.name);
959 plli = ll_i2info(parent->d_inode);
960 task = kthread_run(ll_agl_thread, parent,
961 "ll_agl_%u", plli->lli_opendir_pid);
963 CERROR("can't start ll_agl thread, rc: %ld\n", PTR_ERR(task));
964 thread_set_flags(thread, SVC_STOPPED);
968 l_wait_event(thread->t_ctl_waitq,
969 thread_is_running(thread) || thread_is_stopped(thread),
974 /* statahead thread main function */
975 static int ll_statahead_thread(void *arg)
977 struct dentry *parent = (struct dentry *)arg;
978 struct inode *dir = parent->d_inode;
979 struct ll_inode_info *lli = ll_i2info(dir);
980 struct ll_sb_info *sbi = ll_i2sbi(dir);
981 struct ll_statahead_info *sai;
982 struct ptlrpc_thread *sa_thread;
983 struct ptlrpc_thread *agl_thread;
985 struct md_op_data *op_data;
986 struct ll_dir_chain chain;
987 struct l_wait_info lwi = { 0 };
988 struct page *page = NULL;
993 sai = ll_sai_get(dir);
994 sa_thread = &sai->sai_thread;
995 agl_thread = &sai->sai_agl_thread;
996 sa_thread->t_pid = current_pid();
997 CDEBUG(D_READA, "statahead thread starting: sai %p, parent %.*s\n",
998 sai, parent->d_name.len, parent->d_name.name);
1000 op_data = ll_prep_md_op_data(NULL, dir, dir, NULL, 0, 0,
1001 LUSTRE_OPC_ANY, dir);
1002 if (IS_ERR(op_data))
1003 GOTO(out, rc = PTR_ERR(op_data));
1005 op_data->op_max_pages = ll_i2sbi(dir)->ll_md_brw_pages;
1007 if (sbi->ll_flags & LL_SBI_AGL_ENABLED)
1008 ll_start_agl(parent, sai);
1010 atomic_inc(&sbi->ll_sa_total);
1011 spin_lock(&lli->lli_sa_lock);
1012 if (thread_is_init(sa_thread))
1013 /* If someone else has changed the thread state
1014 * (e.g. already changed to SVC_STOPPING), we can't just
1015 * blindly overwrite that setting. */
1016 thread_set_flags(sa_thread, SVC_RUNNING);
1017 spin_unlock(&lli->lli_sa_lock);
1018 wake_up(&sa_thread->t_ctl_waitq);
1020 ll_dir_chain_init(&chain);
1021 while (pos != MDS_DIR_END_OFF && thread_is_running(sa_thread)) {
1022 struct lu_dirpage *dp;
1023 struct lu_dirent *ent;
1025 sai->sai_in_readpage = 1;
1026 page = ll_get_dir_page(dir, op_data, pos, &chain);
1027 sai->sai_in_readpage = 0;
1030 CDEBUG(D_READA, "error reading dir "DFID" at %llu"
1031 "/%llu opendir_pid = %u: rc = %d\n",
1032 PFID(ll_inode2fid(dir)), pos, sai->sai_index,
1033 lli->lli_opendir_pid, rc);
1037 dp = page_address(page);
1038 for (ent = lu_dirent_start(dp);
1039 ent != NULL && thread_is_running(sa_thread) &&
1041 ent = lu_dirent_next(ent)) {
1047 hash = le64_to_cpu(ent->lde_hash);
1048 if (unlikely(hash < pos))
1050 * Skip until we find target hash value.
1054 namelen = le16_to_cpu(ent->lde_namelen);
1055 if (unlikely(namelen == 0))
1057 * Skip dummy record.
1061 name = ent->lde_name;
1062 if (name[0] == '.') {
1068 } else if (name[1] == '.' && namelen == 2) {
1073 } else if (!sai->sai_ls_all) {
1075 * skip hidden files.
1077 sai->sai_skip_hidden++;
1083 * don't stat-ahead first entry.
1085 if (unlikely(++first == 1))
1088 fid_le_to_cpu(&fid, &ent->lde_fid);
1090 /* wait for spare statahead window */
1092 l_wait_event(sa_thread->t_ctl_waitq,
1093 !sa_sent_full(sai) ||
1094 sa_has_callback(sai) ||
1095 !agl_list_empty(sai) ||
1096 !thread_is_running(sa_thread),
1099 sa_handle_callback(sai);
1101 spin_lock(&lli->lli_agl_lock);
1102 while (sa_sent_full(sai) &&
1103 !agl_list_empty(sai)) {
1104 struct ll_inode_info *clli;
1106 clli = agl_first_entry(sai);
1107 list_del_init(&clli->lli_agl_list);
1108 spin_unlock(&lli->lli_agl_lock);
1110 ll_agl_trigger(&clli->lli_vfs_inode,
1113 spin_lock(&lli->lli_agl_lock);
1115 spin_unlock(&lli->lli_agl_lock);
1116 } while (sa_sent_full(sai) &&
1117 thread_is_running(sa_thread));
1119 sa_statahead(parent, name, namelen, &fid);
1122 pos = le64_to_cpu(dp->ldp_hash_end);
1123 ll_release_page(dir, page,
1124 le32_to_cpu(dp->ldp_flags) & LDF_COLLIDE);
1126 if (sa_low_hit(sai)) {
1128 atomic_inc(&sbi->ll_sa_wrong);
1129 CDEBUG(D_READA, "Statahead for dir "DFID" hit "
1130 "ratio too low: hit/miss %llu/%llu"
1131 ", sent/replied %llu/%llu, stopping "
1132 "statahead thread: pid %d\n",
1133 PFID(&lli->lli_fid), sai->sai_hit,
1134 sai->sai_miss, sai->sai_sent,
1135 sai->sai_replied, current_pid());
1139 ll_dir_chain_fini(&chain);
1140 ll_finish_md_op_data(op_data);
1143 spin_lock(&lli->lli_sa_lock);
1144 thread_set_flags(sa_thread, SVC_STOPPING);
1145 lli->lli_sa_enabled = 0;
1146 spin_unlock(&lli->lli_sa_lock);
1149 /* statahead is finished, but statahead entries need to be cached, wait
1150 * for file release to stop me. */
1151 while (thread_is_running(sa_thread)) {
1152 l_wait_event(sa_thread->t_ctl_waitq,
1153 sa_has_callback(sai) ||
1154 !thread_is_running(sa_thread),
1157 sa_handle_callback(sai);
1162 if (sai->sai_agl_valid) {
1163 spin_lock(&lli->lli_agl_lock);
1164 thread_set_flags(agl_thread, SVC_STOPPING);
1165 spin_unlock(&lli->lli_agl_lock);
1166 wake_up(&agl_thread->t_ctl_waitq);
1168 CDEBUG(D_READA, "stop agl thread: sai %p pid %u\n",
1169 sai, (unsigned int)agl_thread->t_pid);
1170 l_wait_event(agl_thread->t_ctl_waitq,
1171 thread_is_stopped(agl_thread),
1174 /* Set agl_thread flags anyway. */
1175 thread_set_flags(agl_thread, SVC_STOPPED);
1178 /* wait for inflight statahead RPCs to finish, and then we can free sai
1179 * safely because statahead RPC will access sai data */
1180 while (sai->sai_sent != sai->sai_replied) {
1181 /* in case we're not woken up, timeout wait */
1182 lwi = LWI_TIMEOUT(msecs_to_jiffies(MSEC_PER_SEC >> 3),
1184 l_wait_event(sa_thread->t_ctl_waitq,
1185 sai->sai_sent == sai->sai_replied, &lwi);
1188 /* release resources held by statahead RPCs */
1189 sa_handle_callback(sai);
1191 spin_lock(&lli->lli_sa_lock);
1192 thread_set_flags(sa_thread, SVC_STOPPED);
1193 spin_unlock(&lli->lli_sa_lock);
1195 CDEBUG(D_READA, "statahead thread stopped: sai %p, parent %.*s\n",
1196 sai, parent->d_name.len, parent->d_name.name);
1198 wake_up(&sai->sai_waitq);
1199 wake_up(&sa_thread->t_ctl_waitq);
1205 /* authorize opened dir handle @key to statahead */
1206 void ll_authorize_statahead(struct inode *dir, void *key)
1208 struct ll_inode_info *lli = ll_i2info(dir);
1210 spin_lock(&lli->lli_sa_lock);
1211 if (lli->lli_opendir_key == NULL && lli->lli_sai == NULL) {
1213 * if lli_sai is not NULL, it means previous statahead is not
1214 * finished yet, we'd better not start a new statahead for now.
1216 LASSERT(lli->lli_opendir_pid == 0);
1217 lli->lli_opendir_key = key;
1218 lli->lli_opendir_pid = current_pid();
1219 lli->lli_sa_enabled = 1;
1221 spin_unlock(&lli->lli_sa_lock);
1225 * deauthorize opened dir handle @key to statahead, and notify statahead thread
1226 * to quit if it's running.
1228 void ll_deauthorize_statahead(struct inode *dir, void *key)
1230 struct ll_inode_info *lli = ll_i2info(dir);
1231 struct ll_statahead_info *sai;
1233 LASSERT(lli->lli_opendir_key == key);
1234 LASSERT(lli->lli_opendir_pid != 0);
1236 CDEBUG(D_READA, "deauthorize statahead for "DFID"\n",
1237 PFID(&lli->lli_fid));
1239 spin_lock(&lli->lli_sa_lock);
1240 lli->lli_opendir_key = NULL;
1241 lli->lli_opendir_pid = 0;
1242 lli->lli_sa_enabled = 0;
1244 if (sai != NULL && thread_is_running(&sai->sai_thread)) {
1246 * statahead thread may not quit yet because it needs to cache
1247 * entries, now it's time to tell it to quit.
1249 * In case sai is released, wake_up() is called inside spinlock,
1250 * so we have to call smp_mb() explicitely to serialize ops.
1252 thread_set_flags(&sai->sai_thread, SVC_STOPPING);
1254 wake_up(&sai->sai_thread.t_ctl_waitq);
1256 spin_unlock(&lli->lli_sa_lock);
1261 * not first dirent, or is "."
1263 LS_NOT_FIRST_DE = 0,
1265 * the first non-hidden dirent
1269 * the first hidden dirent, that is "."
1274 /* file is first dirent under @dir */
1275 static int is_first_dirent(struct inode *dir, struct dentry *dentry)
1277 struct ll_dir_chain chain;
1278 struct qstr *target = &dentry->d_name;
1279 struct md_op_data *op_data;
1281 struct page *page = NULL;
1282 int rc = LS_NOT_FIRST_DE;
1286 op_data = ll_prep_md_op_data(NULL, dir, dir, NULL, 0, 0,
1287 LUSTRE_OPC_ANY, dir);
1288 if (IS_ERR(op_data))
1289 RETURN(PTR_ERR(op_data));
1291 *FIXME choose the start offset of the readdir
1293 op_data->op_max_pages = ll_i2sbi(dir)->ll_md_brw_pages;
1295 ll_dir_chain_init(&chain);
1296 page = ll_get_dir_page(dir, op_data, 0, &chain);
1299 struct lu_dirpage *dp;
1300 struct lu_dirent *ent;
1303 struct ll_inode_info *lli = ll_i2info(dir);
1306 CERROR("%s: reading dir "DFID" at %llu"
1307 "opendir_pid = %u : rc = %d\n",
1308 ll_get_fsname(dir->i_sb, NULL, 0),
1309 PFID(ll_inode2fid(dir)), pos,
1310 lli->lli_opendir_pid, rc);
1314 dp = page_address(page);
1315 for (ent = lu_dirent_start(dp); ent != NULL;
1316 ent = lu_dirent_next(ent)) {
1321 hash = le64_to_cpu(ent->lde_hash);
1322 /* The ll_get_dir_page() can return any page containing
1323 * the given hash which may be not the start hash. */
1324 if (unlikely(hash < pos))
1327 namelen = le16_to_cpu(ent->lde_namelen);
1328 if (unlikely(namelen == 0))
1330 * skip dummy record.
1334 name = ent->lde_name;
1335 if (name[0] == '.') {
1341 else if (name[1] == '.' && namelen == 2)
1352 if (dot_de && target->name[0] != '.') {
1353 CDEBUG(D_READA, "%.*s skip hidden file %.*s\n",
1354 target->len, target->name,
1359 if (target->len != namelen ||
1360 memcmp(target->name, name, namelen) != 0)
1361 rc = LS_NOT_FIRST_DE;
1365 rc = LS_FIRST_DOT_DE;
1367 ll_release_page(dir, page, false);
1370 pos = le64_to_cpu(dp->ldp_hash_end);
1371 if (pos == MDS_DIR_END_OFF) {
1373 * End of directory reached.
1375 ll_release_page(dir, page, false);
1379 * chain is exhausted
1380 * Normal case: continue to the next page.
1382 ll_release_page(dir, page, le32_to_cpu(dp->ldp_flags) &
1384 page = ll_get_dir_page(dir, op_data, pos, &chain);
1389 ll_dir_chain_fini(&chain);
1390 ll_finish_md_op_data(op_data);
1395 * revalidate @dentryp from statahead cache
1397 * \param[in] dir parent directory
1398 * \param[in] sai sai structure
1399 * \param[out] dentryp pointer to dentry which will be revalidated
1400 * \param[in] unplug unplug statahead window only (normally for negative
1402 * \retval 1 on success, dentry is saved in @dentryp
1403 * \retval 0 if revalidation failed (no proper lock on client)
1404 * \retval negative number upon error
1406 static int revalidate_statahead_dentry(struct inode *dir,
1407 struct ll_statahead_info *sai,
1408 struct dentry **dentryp,
1411 struct sa_entry *entry = NULL;
1412 struct l_wait_info lwi = { 0 };
1413 struct ll_dentry_data *ldd;
1414 struct ll_inode_info *lli = ll_i2info(dir);
1418 if ((*dentryp)->d_name.name[0] == '.') {
1419 if (sai->sai_ls_all ||
1420 sai->sai_miss_hidden >= sai->sai_skip_hidden) {
1422 * Hidden dentry is the first one, or statahead
1423 * thread does not skip so many hidden dentries
1424 * before "sai_ls_all" enabled as below.
1427 if (!sai->sai_ls_all)
1429 * It maybe because hidden dentry is not
1430 * the first one, "sai_ls_all" was not
1431 * set, then "ls -al" missed. Enable
1432 * "sai_ls_all" for such case.
1434 sai->sai_ls_all = 1;
1437 * Such "getattr" has been skipped before
1438 * "sai_ls_all" enabled as above.
1440 sai->sai_miss_hidden++;
1448 entry = sa_get(sai, &(*dentryp)->d_name);
1450 GOTO(out, rc = -EAGAIN);
1452 /* if statahead is busy in readdir, help it do post-work */
1453 if (!sa_ready(entry) && sai->sai_in_readpage)
1454 sa_handle_callback(sai);
1456 if (!sa_ready(entry)) {
1457 spin_lock(&lli->lli_sa_lock);
1458 sai->sai_index_wait = entry->se_index;
1459 spin_unlock(&lli->lli_sa_lock);
1460 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(30), NULL,
1461 LWI_ON_SIGNAL_NOOP, NULL);
1462 rc = l_wait_event(sai->sai_waitq, sa_ready(entry), &lwi);
1465 * entry may not be ready, so it may be used by inflight
1466 * statahead RPC, don't free it.
1469 GOTO(out, rc = -EAGAIN);
1473 if (entry->se_state == SA_ENTRY_SUCC && entry->se_inode != NULL) {
1474 struct inode *inode = entry->se_inode;
1475 struct lookup_intent it = { .it_op = IT_GETATTR,
1480 rc = md_revalidate_lock(ll_i2mdexp(dir), &it,
1481 ll_inode2fid(inode), &bits);
1483 if ((*dentryp)->d_inode == NULL) {
1484 struct dentry *alias;
1486 alias = ll_splice_alias(inode, *dentryp);
1487 if (IS_ERR(alias)) {
1488 ll_intent_release(&it);
1489 GOTO(out, rc = PTR_ERR(alias));
1492 /* statahead prepared this inode, transfer inode
1493 * refcount from sa_entry to dentry */
1494 entry->se_inode = NULL;
1495 } else if ((*dentryp)->d_inode != inode) {
1496 /* revalidate, but inode is recreated */
1498 "%s: stale dentry %.*s inode "
1499 DFID", statahead inode "DFID
1501 ll_get_fsname((*dentryp)->d_inode->i_sb,
1503 (*dentryp)->d_name.len,
1504 (*dentryp)->d_name.name,
1505 PFID(ll_inode2fid((*dentryp)->d_inode)),
1506 PFID(ll_inode2fid(inode)));
1507 ll_intent_release(&it);
1508 GOTO(out, rc = -ESTALE);
1511 if ((bits & MDS_INODELOCK_LOOKUP) &&
1512 d_lustre_invalid(*dentryp))
1513 d_lustre_revalidate(*dentryp);
1514 ll_intent_release(&it);
1519 * statahead cached sa_entry can be used only once, and will be killed
1520 * right after use, so if lookup/revalidate accessed statahead cache,
1521 * set dentry ldd_sa_generation to parent lli_sa_generation, later if we
1522 * stat this file again, we know we've done statahead before, see
1523 * dentry_may_statahead().
1525 ldd = ll_d2d(*dentryp);
1526 /* ldd can be NULL if llite lookup failed. */
1528 ldd->lld_sa_generation = lli->lli_sa_generation;
1535 * start statahead thread
1537 * \param[in] dir parent directory
1538 * \param[in] dentry dentry that triggers statahead, normally the first
1540 * \retval -EAGAIN on success, because when this function is
1541 * called, it's already in lookup call, so client should
1542 * do it itself instead of waiting for statahead thread
1543 * to do it asynchronously.
1544 * \retval negative number upon error
1546 static int start_statahead_thread(struct inode *dir, struct dentry *dentry)
1548 struct ll_inode_info *lli = ll_i2info(dir);
1549 struct ll_statahead_info *sai = NULL;
1550 struct dentry *parent = dentry->d_parent;
1551 struct ptlrpc_thread *thread;
1552 struct l_wait_info lwi = { 0 };
1553 struct task_struct *task;
1557 /* I am the "lli_opendir_pid" owner, only me can set "lli_sai". */
1558 rc = is_first_dirent(dir, dentry);
1559 if (rc == LS_NOT_FIRST_DE)
1560 /* It is not "ls -{a}l" operation, no need statahead for it. */
1561 GOTO(out, rc = -EFAULT);
1563 sai = ll_sai_alloc(parent);
1565 GOTO(out, rc = -ENOMEM);
1567 sai->sai_ls_all = (rc == LS_FIRST_DOT_DE);
1569 /* if current lli_opendir_key was deauthorized, or dir re-opened by
1570 * another process, don't start statahead, otherwise the newly spawned
1571 * statahead thread won't be notified to quit. */
1572 spin_lock(&lli->lli_sa_lock);
1573 if (unlikely(lli->lli_sai != NULL ||
1574 lli->lli_opendir_key == NULL ||
1575 lli->lli_opendir_pid != current->pid)) {
1576 spin_unlock(&lli->lli_sa_lock);
1577 GOTO(out, rc = -EPERM);
1580 spin_unlock(&lli->lli_sa_lock);
1582 atomic_inc(&ll_i2sbi(parent->d_inode)->ll_sa_running);
1584 CDEBUG(D_READA, "start statahead thread: [pid %d] [parent %.*s]\n",
1585 current_pid(), parent->d_name.len, parent->d_name.name);
1587 task = kthread_run(ll_statahead_thread, parent, "ll_sa_%u",
1588 lli->lli_opendir_pid);
1589 thread = &sai->sai_thread;
1592 CERROR("can't start ll_sa thread, rc: %d\n", rc);
1596 l_wait_event(thread->t_ctl_waitq,
1597 thread_is_running(thread) || thread_is_stopped(thread),
1602 * We don't stat-ahead for the first dirent since we are already in
1608 /* once we start statahead thread failed, disable statahead so that
1609 * subsequent stat won't waste time to try it. */
1610 spin_lock(&lli->lli_sa_lock);
1611 lli->lli_sa_enabled = 0;
1612 lli->lli_sai = NULL;
1613 spin_unlock(&lli->lli_sa_lock);
1622 * statahead entry function, this is called when client getattr on a file, it
1623 * will start statahead thread if this is the first dir entry, else revalidate
1624 * dentry from statahead cache.
1626 * \param[in] dir parent directory
1627 * \param[out] dentryp dentry to getattr
1628 * \param[in] unplug unplug statahead window only (normally for negative
1630 * \retval 1 on success
1631 * \retval 0 revalidation from statahead cache failed, caller needs
1632 * to getattr from server directly
1633 * \retval negative number on error, caller often ignores this and
1634 * then getattr from server
1636 int ll_statahead(struct inode *dir, struct dentry **dentryp, bool unplug)
1638 struct ll_statahead_info *sai;
1640 sai = ll_sai_get(dir);
1644 rc = revalidate_statahead_dentry(dir, sai, dentryp, unplug);
1645 CDEBUG(D_READA, "revalidate statahead %.*s: %d.\n",
1646 (*dentryp)->d_name.len, (*dentryp)->d_name.name, rc);
1650 return start_statahead_thread(dir, *dentryp);