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, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
33 #include <linux/sched.h>
34 #include <linux/kthread.h>
36 #include <linux/highmem.h>
37 #include <linux/pagemap.h>
38 #include <linux/delay.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 */
56 * sa_entry is not refcounted: statahead thread allocates it and do async stat,
57 * and in async stat callback ll_statahead_interpret() will prepare the inode
58 * and set lock data in the ptlrpcd context. Then the scanner process will be
59 * woken up if this entry is the waiting one, can access and free it.
62 /* link into sai_entries */
63 struct list_head se_list;
64 /* link into sai hash table locally */
65 struct list_head se_hash;
66 /* entry index in the sai */
68 /* low layer ldlm lock handle */
72 /* entry size, contains name */
74 /* pointer to the target inode */
75 struct inode *se_inode;
82 static unsigned int sai_generation;
83 static DEFINE_SPINLOCK(sai_generation_lock);
85 static inline int sa_unhashed(struct sa_entry *entry)
87 return list_empty(&entry->se_hash);
90 /* sa_entry is ready to use */
91 static inline int sa_ready(struct sa_entry *entry)
93 /* Make sure sa_entry is updated and ready to use */
95 return (entry->se_state != SA_ENTRY_INIT);
98 /* hash value to put in sai_cache */
99 static inline int sa_hash(int val)
101 return val & LL_SA_CACHE_MASK;
104 /* hash entry into sai_cache */
106 sa_rehash(struct ll_statahead_info *sai, struct sa_entry *entry)
108 int i = sa_hash(entry->se_qstr.hash);
110 spin_lock(&sai->sai_cache_lock[i]);
111 list_add_tail(&entry->se_hash, &sai->sai_cache[i]);
112 spin_unlock(&sai->sai_cache_lock[i]);
115 /* unhash entry from sai_cache */
117 sa_unhash(struct ll_statahead_info *sai, struct sa_entry *entry)
119 int i = sa_hash(entry->se_qstr.hash);
121 spin_lock(&sai->sai_cache_lock[i]);
122 list_del_init(&entry->se_hash);
123 spin_unlock(&sai->sai_cache_lock[i]);
126 static inline int agl_should_run(struct ll_statahead_info *sai,
129 return inode && S_ISREG(inode->i_mode) && sai->sai_agl_task;
132 static inline struct ll_inode_info *
133 agl_first_entry(struct ll_statahead_info *sai)
135 return list_first_entry(&sai->sai_agls, struct ll_inode_info,
139 /* statahead window is full */
140 static inline int sa_sent_full(struct ll_statahead_info *sai)
142 return atomic_read(&sai->sai_cache_count) >= sai->sai_max;
145 static inline int agl_list_empty(struct ll_statahead_info *sai)
147 return list_empty(&sai->sai_agls);
151 * (1) hit ratio less than 80%
153 * (2) consecutive miss more than 8
154 * then means low hit.
156 static inline int sa_low_hit(struct ll_statahead_info *sai)
158 return ((sai->sai_hit > 7 && sai->sai_hit < 4 * sai->sai_miss) ||
159 (sai->sai_consecutive_miss > 8));
163 * if the given index is behind of statahead window more than
164 * SA_OMITTED_ENTRY_MAX, then it is old.
166 static inline int is_omitted_entry(struct ll_statahead_info *sai, __u64 index)
168 return ((__u64)sai->sai_max + index + SA_OMITTED_ENTRY_MAX <
172 /* allocate sa_entry and hash it to allow scanner process to find it */
173 static struct sa_entry *
174 sa_alloc(struct dentry *parent, struct ll_statahead_info *sai, __u64 index,
175 const char *name, int len, const struct lu_fid *fid)
177 struct ll_inode_info *lli;
178 struct sa_entry *entry;
184 entry_size = sizeof(struct sa_entry) + (len & ~3) + 4;
185 OBD_ALLOC(entry, entry_size);
186 if (unlikely(!entry))
187 RETURN(ERR_PTR(-ENOMEM));
189 CDEBUG(D_READA, "alloc sa entry %.*s(%p) index %llu\n",
190 len, name, entry, index);
192 entry->se_index = index;
194 entry->se_state = SA_ENTRY_INIT;
195 entry->se_size = entry_size;
196 dname = (char *)entry + sizeof(struct sa_entry);
197 memcpy(dname, name, len);
199 entry->se_qstr.hash = ll_full_name_hash(parent, name, len);
200 entry->se_qstr.len = len;
201 entry->se_qstr.name = dname;
202 entry->se_fid = *fid;
204 lli = ll_i2info(sai->sai_dentry->d_inode);
206 spin_lock(&lli->lli_sa_lock);
207 INIT_LIST_HEAD(&entry->se_list);
208 sa_rehash(sai, entry);
209 spin_unlock(&lli->lli_sa_lock);
211 atomic_inc(&sai->sai_cache_count);
216 /* free sa_entry, which should have been unhashed and not in any list */
217 static void sa_free(struct ll_statahead_info *sai, struct sa_entry *entry)
219 CDEBUG(D_READA, "free sa entry %.*s(%p) index %llu\n",
220 entry->se_qstr.len, entry->se_qstr.name, entry,
223 LASSERT(list_empty(&entry->se_list));
224 LASSERT(sa_unhashed(entry));
226 OBD_FREE(entry, entry->se_size);
227 atomic_dec(&sai->sai_cache_count);
231 * find sa_entry by name, used by directory scanner, lock is not needed because
232 * only scanner can remove the entry from cache.
234 static struct sa_entry *
235 sa_get(struct ll_statahead_info *sai, const struct qstr *qstr)
237 struct sa_entry *entry;
238 int i = sa_hash(qstr->hash);
240 list_for_each_entry(entry, &sai->sai_cache[i], se_hash) {
241 if (entry->se_qstr.hash == qstr->hash &&
242 entry->se_qstr.len == qstr->len &&
243 memcmp(entry->se_qstr.name, qstr->name, qstr->len) == 0)
249 /* unhash and unlink sa_entry, and then free it */
251 sa_kill(struct ll_statahead_info *sai, struct sa_entry *entry)
253 struct ll_inode_info *lli = ll_i2info(sai->sai_dentry->d_inode);
255 LASSERT(!sa_unhashed(entry));
256 LASSERT(!list_empty(&entry->se_list));
257 LASSERT(sa_ready(entry));
259 sa_unhash(sai, entry);
261 spin_lock(&lli->lli_sa_lock);
262 list_del_init(&entry->se_list);
263 spin_unlock(&lli->lli_sa_lock);
265 iput(entry->se_inode);
270 /* called by scanner after use, sa_entry will be killed */
272 sa_put(struct ll_statahead_info *sai, struct sa_entry *entry)
274 struct sa_entry *tmp, *next;
276 if (entry && entry->se_state == SA_ENTRY_SUCC) {
277 struct ll_sb_info *sbi = ll_i2sbi(sai->sai_dentry->d_inode);
280 sai->sai_consecutive_miss = 0;
281 sai->sai_max = min(2 * sai->sai_max, sbi->ll_sa_max);
284 sai->sai_consecutive_miss++;
291 * kill old completed entries, only scanner process does this, no need
294 list_for_each_entry_safe(tmp, next, &sai->sai_entries, se_list) {
295 if (!is_omitted_entry(sai, tmp->se_index))
302 * update state and sort add entry to sai_entries by index, return true if
303 * scanner is waiting on this entry.
306 __sa_make_ready(struct ll_statahead_info *sai, struct sa_entry *entry, int ret)
309 struct list_head *pos = &sai->sai_entries;
310 __u64 index = entry->se_index;
312 LASSERT(!sa_ready(entry));
313 LASSERT(list_empty(&entry->se_list));
315 list_for_each_entry_reverse(se, &sai->sai_entries, se_list) {
316 if (se->se_index < entry->se_index) {
321 list_add(&entry->se_list, pos);
323 * LU-9210: ll_statahead_interpet must be able to see this before
326 smp_store_release(&entry->se_state,
327 ret < 0 ? SA_ENTRY_INVA : SA_ENTRY_SUCC);
329 return (index == sai->sai_index_wait);
332 /* finish async stat RPC arguments */
333 static void sa_fini_data(struct md_op_item *item)
335 struct md_op_data *op_data = &item->mop_data;
337 if (op_data->op_flags & MF_OPNAME_KMALLOCED)
338 /* allocated via ll_setup_filename called from sa_prep_data */
339 kfree(op_data->op_name);
340 ll_unlock_md_op_lsm(&item->mop_data);
345 static int ll_statahead_interpret(struct md_op_item *item, int rc);
348 * prepare arguments for async stat RPC.
350 static struct md_op_item *
351 sa_prep_data(struct inode *dir, struct inode *child, struct sa_entry *entry)
353 struct md_op_item *item;
354 struct ldlm_enqueue_info *einfo;
355 struct md_op_data *op_data;
359 return ERR_PTR(-ENOMEM);
361 op_data = ll_prep_md_op_data(&item->mop_data, dir, child,
362 entry->se_qstr.name, entry->se_qstr.len, 0,
363 LUSTRE_OPC_ANY, NULL);
364 if (IS_ERR(op_data)) {
366 return (struct md_op_item *)op_data;
370 op_data->op_fid2 = entry->se_fid;
372 item->mop_it.it_op = IT_GETATTR;
373 item->mop_dir = igrab(dir);
374 item->mop_cb = ll_statahead_interpret;
375 item->mop_cbdata = entry;
377 einfo = &item->mop_einfo;
378 einfo->ei_type = LDLM_IBITS;
379 einfo->ei_mode = it_to_lock_mode(&item->mop_it);
380 einfo->ei_cb_bl = ll_md_blocking_ast;
381 einfo->ei_cb_cp = ldlm_completion_ast;
382 einfo->ei_cb_gl = NULL;
383 einfo->ei_cbdata = NULL;
384 einfo->ei_req_slot = 1;
390 * release resources used in async stat RPC, update entry state and wakeup if
391 * scanner process it waiting on this entry.
394 sa_make_ready(struct ll_statahead_info *sai, struct sa_entry *entry, int ret)
396 struct ll_inode_info *lli = ll_i2info(sai->sai_dentry->d_inode);
399 spin_lock(&lli->lli_sa_lock);
400 wakeup = __sa_make_ready(sai, entry, ret);
401 spin_unlock(&lli->lli_sa_lock);
404 wake_up(&sai->sai_waitq);
407 /* insert inode into the list of sai_agls */
408 static void ll_agl_add(struct ll_statahead_info *sai,
409 struct inode *inode, int index)
411 struct ll_inode_info *child = ll_i2info(inode);
412 struct ll_inode_info *parent = ll_i2info(sai->sai_dentry->d_inode);
414 spin_lock(&child->lli_agl_lock);
415 if (child->lli_agl_index == 0) {
416 child->lli_agl_index = index;
417 spin_unlock(&child->lli_agl_lock);
419 LASSERT(list_empty(&child->lli_agl_list));
421 spin_lock(&parent->lli_agl_lock);
422 /* Re-check under the lock */
423 if (agl_should_run(sai, inode)) {
424 if (agl_list_empty(sai))
425 wake_up_process(sai->sai_agl_task);
427 list_add_tail(&child->lli_agl_list, &sai->sai_agls);
429 child->lli_agl_index = 0;
430 spin_unlock(&parent->lli_agl_lock);
432 spin_unlock(&child->lli_agl_lock);
437 static struct ll_statahead_info *ll_sai_alloc(struct dentry *dentry)
439 struct ll_statahead_info *sai;
440 struct ll_inode_info *lli = ll_i2info(dentry->d_inode);
449 sai->sai_dentry = dget(dentry);
450 atomic_set(&sai->sai_refcount, 1);
451 sai->sai_max = LL_SA_RPC_MIN;
453 init_waitqueue_head(&sai->sai_waitq);
455 INIT_LIST_HEAD(&sai->sai_entries);
456 INIT_LIST_HEAD(&sai->sai_agls);
458 for (i = 0; i < LL_SA_CACHE_SIZE; i++) {
459 INIT_LIST_HEAD(&sai->sai_cache[i]);
460 spin_lock_init(&sai->sai_cache_lock[i]);
462 atomic_set(&sai->sai_cache_count, 0);
464 spin_lock(&sai_generation_lock);
465 lli->lli_sa_generation = ++sai_generation;
466 if (unlikely(sai_generation == 0))
467 lli->lli_sa_generation = ++sai_generation;
468 spin_unlock(&sai_generation_lock);
474 static inline void ll_sai_free(struct ll_statahead_info *sai)
476 LASSERT(sai->sai_dentry != NULL);
477 dput(sai->sai_dentry);
482 * take refcount of sai if sai for @dir exists, which means statahead is on for
485 static inline struct ll_statahead_info *ll_sai_get(struct inode *dir)
487 struct ll_inode_info *lli = ll_i2info(dir);
488 struct ll_statahead_info *sai = NULL;
490 spin_lock(&lli->lli_sa_lock);
493 atomic_inc(&sai->sai_refcount);
494 spin_unlock(&lli->lli_sa_lock);
500 * put sai refcount after use, if refcount reaches zero, free sai and sa_entries
503 static void ll_sai_put(struct ll_statahead_info *sai)
505 struct ll_inode_info *lli = ll_i2info(sai->sai_dentry->d_inode);
507 if (atomic_dec_and_lock(&sai->sai_refcount, &lli->lli_sa_lock)) {
508 struct sa_entry *entry, *next;
509 struct ll_sb_info *sbi = ll_i2sbi(sai->sai_dentry->d_inode);
512 spin_unlock(&lli->lli_sa_lock);
514 LASSERT(!sai->sai_task);
515 LASSERT(!sai->sai_agl_task);
516 LASSERT(sai->sai_sent == sai->sai_replied);
518 list_for_each_entry_safe(entry, next, &sai->sai_entries,
522 LASSERT(atomic_read(&sai->sai_cache_count) == 0);
523 LASSERT(agl_list_empty(sai));
526 atomic_dec(&sbi->ll_sa_running);
530 /* Do NOT forget to drop inode refcount when into sai_agls. */
531 static void ll_agl_trigger(struct inode *inode, struct ll_statahead_info *sai)
533 struct ll_inode_info *lli = ll_i2info(inode);
534 u64 index = lli->lli_agl_index;
540 LASSERT(list_empty(&lli->lli_agl_list));
542 /* AGL maybe fall behind statahead with one entry */
543 if (is_omitted_entry(sai, index + 1)) {
544 lli->lli_agl_index = 0;
550 * In case of restore, the MDT has the right size and has already
551 * sent it back without granting the layout lock, inode is up-to-date.
552 * Then AGL (async glimpse lock) is useless.
553 * Also to glimpse we need the layout, in case of a runninh restore
554 * the MDT holds the layout lock so the glimpse will block up to the
555 * end of restore (statahead/agl will block)
557 if (test_bit(LLIF_FILE_RESTORING, &lli->lli_flags)) {
558 lli->lli_agl_index = 0;
563 /* Someone is in glimpse (sync or async), do nothing. */
564 rc = down_write_trylock(&lli->lli_glimpse_sem);
566 lli->lli_agl_index = 0;
572 * Someone triggered glimpse within 1 sec before.
573 * 1) The former glimpse succeeded with glimpse lock granted by OST, and
574 * if the lock is still cached on client, AGL needs to do nothing. If
575 * it is cancelled by other client, AGL maybe cannot obtaion new lock
576 * for no glimpse callback triggered by AGL.
577 * 2) The former glimpse succeeded, but OST did not grant glimpse lock.
578 * Under such case, it is quite possible that the OST will not grant
579 * glimpse lock for AGL also.
580 * 3) The former glimpse failed, compared with other two cases, it is
581 * relative rare. AGL can ignore such case, and it will not muchly
582 * affect the performance.
584 expire = ktime_sub_ns(ktime_get(), NSEC_PER_SEC);
585 if (ktime_to_ns(lli->lli_glimpse_time) &&
586 ktime_before(expire, lli->lli_glimpse_time)) {
587 up_write(&lli->lli_glimpse_sem);
588 lli->lli_agl_index = 0;
594 "Handling (init) async glimpse: inode = " DFID", idx = %llu\n",
595 PFID(&lli->lli_fid), index);
598 lli->lli_agl_index = 0;
599 lli->lli_glimpse_time = ktime_get();
600 up_write(&lli->lli_glimpse_sem);
603 "Handled (init) async glimpse: inode= " DFID", idx = %llu, rc = %d\n",
604 PFID(&lli->lli_fid), index, rc);
611 static void ll_statahead_interpret_fini(struct ll_inode_info *lli,
612 struct ll_statahead_info *sai,
613 struct md_op_item *item,
614 struct sa_entry *entry,
615 struct ptlrpc_request *req,
619 * First it will drop ldlm ibits lock refcount by calling
620 * ll_intent_drop_lock() in spite of failures. Do not worry about
621 * calling ll_intent_drop_lock() more than once.
623 ll_intent_release(&item->mop_it);
626 ptlrpc_req_finished(req);
627 sa_make_ready(sai, entry, rc);
629 spin_lock(&lli->lli_sa_lock);
631 spin_unlock(&lli->lli_sa_lock);
634 static void ll_statahead_interpret_work(struct work_struct *work)
636 struct md_op_item *item = container_of(work, struct md_op_item,
638 struct req_capsule *pill = item->mop_pill;
639 struct inode *dir = item->mop_dir;
640 struct ll_inode_info *lli = ll_i2info(dir);
641 struct ll_statahead_info *sai = lli->lli_sai;
642 struct lookup_intent *it;
643 struct sa_entry *entry;
644 struct mdt_body *body;
650 entry = (struct sa_entry *)item->mop_cbdata;
651 LASSERT(entry->se_handle != 0);
654 body = req_capsule_server_get(pill, &RMF_MDT_BODY);
656 GOTO(out, rc = -EFAULT);
658 child = entry->se_inode;
659 /* revalidate; unlinked and re-created with the same name */
660 if (unlikely(!lu_fid_eq(&item->mop_data.op_fid2, &body->mbo_fid1))) {
662 entry->se_inode = NULL;
665 /* The mdt_body is invalid. Skip this entry */
666 GOTO(out, rc = -EAGAIN);
669 it->it_lock_handle = entry->se_handle;
670 rc = md_revalidate_lock(ll_i2mdexp(dir), it, ll_inode2fid(dir), NULL);
672 GOTO(out, rc = -EAGAIN);
674 rc = ll_prep_inode(&child, pill, dir->i_sb, it);
678 /* If encryption context was returned by MDT, put it in
679 * inode now to save an extra getxattr.
681 if (body->mbo_valid & OBD_MD_ENCCTX) {
682 void *encctx = req_capsule_server_get(pill, &RMF_FILE_ENCCTX);
683 __u32 encctxlen = req_capsule_get_size(pill, &RMF_FILE_ENCCTX,
688 "server returned encryption ctx for "DFID"\n",
689 PFID(ll_inode2fid(child)));
690 rc = ll_xattr_cache_insert(child,
691 xattr_for_enc(child),
694 CWARN("%s: cannot set enc ctx for "DFID": rc = %d\n",
695 ll_i2sbi(child)->ll_fsname,
696 PFID(ll_inode2fid(child)), rc);
700 CDEBUG(D_READA, "%s: setting %.*s"DFID" l_data to inode %p\n",
701 ll_i2sbi(dir)->ll_fsname, entry->se_qstr.len,
702 entry->se_qstr.name, PFID(ll_inode2fid(child)), child);
703 ll_set_lock_data(ll_i2sbi(dir)->ll_md_exp, child, it, NULL);
705 entry->se_inode = child;
707 if (agl_should_run(sai, child))
708 ll_agl_add(sai, child, entry->se_index);
710 ll_statahead_interpret_fini(lli, sai, item, entry, pill->rc_req, rc);
714 * Callback for async stat RPC, this is called in ptlrpcd context. It prepares
715 * the inode and set lock data directly in the ptlrpcd context. It will wake up
716 * the directory listing process if the dentry is the waiting one.
718 static int ll_statahead_interpret(struct md_op_item *item, int rc)
720 struct req_capsule *pill = item->mop_pill;
721 struct lookup_intent *it = &item->mop_it;
722 struct inode *dir = item->mop_dir;
723 struct ll_inode_info *lli = ll_i2info(dir);
724 struct ll_statahead_info *sai = lli->lli_sai;
725 struct sa_entry *entry = (struct sa_entry *)item->mop_cbdata;
726 struct work_struct *work = &item->mop_work;
727 struct mdt_body *body;
733 if (it_disposition(it, DISP_LOOKUP_NEG))
737 * because statahead thread will wait for all inflight RPC to finish,
738 * sai should be always valid, no need to refcount
740 LASSERT(sai != NULL);
741 LASSERT(entry != NULL);
743 CDEBUG(D_READA, "sa_entry %.*s rc %d\n",
744 entry->se_qstr.len, entry->se_qstr.name, rc);
749 body = req_capsule_server_get(pill, &RMF_MDT_BODY);
751 GOTO(out, rc = -EFAULT);
753 child = entry->se_inode;
754 /* revalidate; unlinked and re-created with the same name */
755 if (unlikely(!lu_fid_eq(&item->mop_data.op_fid2, &body->mbo_fid1))) {
757 entry->se_inode = NULL;
760 /* The mdt_body is invalid. Skip this entry */
761 GOTO(out, rc = -EAGAIN);
764 entry->se_handle = it->it_lock_handle;
766 * In ptlrpcd context, it is not allowed to generate new RPCs
767 * especially for striped directories or regular files with layout
771 * release ibits lock ASAP to avoid deadlock when statahead
772 * thread enqueues lock on parent in readdir and another
773 * process enqueues lock on child with parent lock held, eg.
776 handle = it->it_lock_handle;
777 ll_intent_drop_lock(it);
778 ll_unlock_md_op_lsm(&item->mop_data);
781 * If the statahead entry is a striped directory or regular file with
782 * layout change, it will generate a new RPC and long wait in the
784 * However, it is dangerous of blocking in ptlrpcd thread.
785 * Here we use work queue or the separate statahead thread to handle
786 * the extra RPC and long wait:
787 * (@ll_prep_inode->@lmv_revalidate_slaves);
788 * (@ll_prep_inode->@lov_layout_change->osc_cache_wait_range);
790 INIT_WORK(work, ll_statahead_interpret_work);
791 ptlrpc_request_addref(pill->rc_req);
795 ll_statahead_interpret_fini(lli, sai, item, entry, NULL, rc);
799 /* async stat for file not found in dcache */
800 static int sa_lookup(struct inode *dir, struct sa_entry *entry)
802 struct md_op_item *item;
807 item = sa_prep_data(dir, NULL, entry);
809 RETURN(PTR_ERR(item));
811 rc = md_intent_getattr_async(ll_i2mdexp(dir), item);
819 * async stat for file found in dcache, similar to .revalidate
821 * \retval 1 dentry valid, no RPC sent
822 * \retval 0 dentry invalid, will send async stat RPC
823 * \retval negative number upon error
825 static int sa_revalidate(struct inode *dir, struct sa_entry *entry,
826 struct dentry *dentry)
828 struct inode *inode = dentry->d_inode;
829 struct lookup_intent it = { .it_op = IT_GETATTR,
830 .it_lock_handle = 0 };
831 struct md_op_item *item;
836 if (unlikely(!inode))
839 if (d_mountpoint(dentry))
842 item = sa_prep_data(dir, inode, entry);
844 RETURN(PTR_ERR(item));
846 entry->se_inode = igrab(inode);
847 rc = md_revalidate_lock(ll_i2mdexp(dir), &it, ll_inode2fid(inode),
850 entry->se_handle = it.it_lock_handle;
851 ll_intent_release(&it);
856 rc = md_intent_getattr_async(ll_i2mdexp(dir), item);
858 entry->se_inode = NULL;
866 /* async stat for file with @name */
867 static void sa_statahead(struct dentry *parent, const char *name, int len,
868 const struct lu_fid *fid)
870 struct inode *dir = parent->d_inode;
871 struct ll_inode_info *lli = ll_i2info(dir);
872 struct ll_statahead_info *sai = lli->lli_sai;
873 struct dentry *dentry = NULL;
874 struct sa_entry *entry;
879 entry = sa_alloc(parent, sai, sai->sai_index, name, len, fid);
883 dentry = d_lookup(parent, &entry->se_qstr);
885 rc = sa_lookup(dir, entry);
887 rc = sa_revalidate(dir, entry, dentry);
888 if (rc == 1 && agl_should_run(sai, dentry->d_inode))
889 ll_agl_add(sai, dentry->d_inode, entry->se_index);
896 sa_make_ready(sai, entry, rc);
905 /* async glimpse (agl) thread main function */
906 static int ll_agl_thread(void *arg)
908 struct dentry *parent = (struct dentry *)arg;
909 struct inode *dir = parent->d_inode;
910 struct ll_inode_info *plli = ll_i2info(dir);
911 struct ll_inode_info *clli;
913 * We already own this reference, so it is safe to take it
916 struct ll_statahead_info *sai = plli->lli_sai;
920 CDEBUG(D_READA, "agl thread started: sai %p, parent %pd\n",
923 while (({set_current_state(TASK_IDLE);
924 !kthread_should_stop(); })) {
925 spin_lock(&plli->lli_agl_lock);
926 clli = list_first_entry_or_null(&sai->sai_agls,
927 struct ll_inode_info,
930 __set_current_state(TASK_RUNNING);
931 list_del_init(&clli->lli_agl_list);
932 spin_unlock(&plli->lli_agl_lock);
933 ll_agl_trigger(&clli->lli_vfs_inode, sai);
936 spin_unlock(&plli->lli_agl_lock);
940 __set_current_state(TASK_RUNNING);
944 static void ll_stop_agl(struct ll_statahead_info *sai)
946 struct dentry *parent = sai->sai_dentry;
947 struct ll_inode_info *plli = ll_i2info(parent->d_inode);
948 struct ll_inode_info *clli;
949 struct task_struct *agl_task;
951 spin_lock(&plli->lli_agl_lock);
952 agl_task = sai->sai_agl_task;
953 sai->sai_agl_task = NULL;
954 spin_unlock(&plli->lli_agl_lock);
958 CDEBUG(D_READA, "stop agl thread: sai %p pid %u\n",
959 sai, (unsigned int)agl_task->pid);
960 kthread_stop(agl_task);
962 spin_lock(&plli->lli_agl_lock);
963 while ((clli = list_first_entry_or_null(&sai->sai_agls,
964 struct ll_inode_info,
965 lli_agl_list)) != NULL) {
966 list_del_init(&clli->lli_agl_list);
967 spin_unlock(&plli->lli_agl_lock);
968 clli->lli_agl_index = 0;
969 iput(&clli->lli_vfs_inode);
970 spin_lock(&plli->lli_agl_lock);
972 spin_unlock(&plli->lli_agl_lock);
973 CDEBUG(D_READA, "agl thread stopped: sai %p, parent %pd\n",
978 /* start agl thread */
979 static void ll_start_agl(struct dentry *parent, struct ll_statahead_info *sai)
981 int node = cfs_cpt_spread_node(cfs_cpt_tab, CFS_CPT_ANY);
982 struct ll_inode_info *plli;
983 struct task_struct *task;
987 CDEBUG(D_READA, "start agl thread: sai %p, parent %pd\n",
990 plli = ll_i2info(parent->d_inode);
991 task = kthread_create_on_node(ll_agl_thread, parent, node, "ll_agl_%d",
992 plli->lli_opendir_pid);
994 CERROR("can't start ll_agl thread, rc: %ld\n", PTR_ERR(task));
997 sai->sai_agl_task = task;
998 atomic_inc(&ll_i2sbi(d_inode(parent))->ll_agl_total);
999 /* Get an extra reference that the thread holds */
1000 ll_sai_get(d_inode(parent));
1002 wake_up_process(task);
1007 /* statahead thread main function */
1008 static int ll_statahead_thread(void *arg)
1010 struct dentry *parent = (struct dentry *)arg;
1011 struct inode *dir = parent->d_inode;
1012 struct ll_inode_info *lli = ll_i2info(dir);
1013 struct ll_sb_info *sbi = ll_i2sbi(dir);
1014 struct ll_statahead_info *sai = lli->lli_sai;
1016 struct md_op_data *op_data;
1017 struct page *page = NULL;
1023 CDEBUG(D_READA, "statahead thread starting: sai %p, parent %pd\n",
1026 OBD_ALLOC_PTR(op_data);
1028 GOTO(out, rc = -ENOMEM);
1030 while (pos != MDS_DIR_END_OFF && sai->sai_task) {
1031 struct lu_dirpage *dp;
1032 struct lu_dirent *ent;
1034 op_data = ll_prep_md_op_data(op_data, dir, dir, NULL, 0, 0,
1035 LUSTRE_OPC_ANY, dir);
1036 if (IS_ERR(op_data)) {
1037 rc = PTR_ERR(op_data);
1041 page = ll_get_dir_page(dir, op_data, pos, NULL);
1042 ll_unlock_md_op_lsm(op_data);
1046 "error reading dir "DFID" at %llu /%llu opendir_pid = %u: rc = %d\n",
1047 PFID(ll_inode2fid(dir)), pos, sai->sai_index,
1048 lli->lli_opendir_pid, rc);
1052 dp = page_address(page);
1053 for (ent = lu_dirent_start(dp);
1054 ent != NULL && sai->sai_task &&
1056 ent = lu_dirent_next(ent)) {
1061 struct llcrypt_str lltr = LLTR_INIT(NULL, 0);
1063 hash = le64_to_cpu(ent->lde_hash);
1064 if (unlikely(hash < pos))
1066 * Skip until we find target hash value.
1070 namelen = le16_to_cpu(ent->lde_namelen);
1071 if (unlikely(namelen == 0))
1073 * Skip dummy record.
1077 name = ent->lde_name;
1078 if (name[0] == '.') {
1084 } else if (name[1] == '.' && namelen == 2) {
1089 } else if (!sai->sai_ls_all) {
1091 * skip hidden files.
1093 sai->sai_skip_hidden++;
1099 * don't stat-ahead first entry.
1101 if (unlikely(++first == 1))
1104 fid_le_to_cpu(&fid, &ent->lde_fid);
1106 while (({set_current_state(TASK_IDLE);
1107 sai->sai_task; })) {
1108 spin_lock(&lli->lli_agl_lock);
1109 while (sa_sent_full(sai) &&
1110 !agl_list_empty(sai)) {
1111 struct ll_inode_info *clli;
1113 __set_current_state(TASK_RUNNING);
1114 clli = agl_first_entry(sai);
1115 list_del_init(&clli->lli_agl_list);
1116 spin_unlock(&lli->lli_agl_lock);
1118 ll_agl_trigger(&clli->lli_vfs_inode,
1121 spin_lock(&lli->lli_agl_lock);
1123 spin_unlock(&lli->lli_agl_lock);
1125 if (!sa_sent_full(sai))
1129 __set_current_state(TASK_RUNNING);
1131 if (IS_ENCRYPTED(dir)) {
1132 struct llcrypt_str de_name =
1133 LLTR_INIT(ent->lde_name, namelen);
1136 rc = llcrypt_fname_alloc_buffer(dir, NAME_MAX,
1141 fid_le_to_cpu(&fid, &ent->lde_fid);
1142 if (ll_fname_disk_to_usr(dir, 0, 0, &de_name,
1144 llcrypt_fname_free_buffer(&lltr);
1152 sa_statahead(parent, name, namelen, &fid);
1153 llcrypt_fname_free_buffer(&lltr);
1156 pos = le64_to_cpu(dp->ldp_hash_end);
1157 down_read(&lli->lli_lsm_sem);
1158 ll_release_page(dir, page,
1159 le32_to_cpu(dp->ldp_flags) & LDF_COLLIDE);
1160 up_read(&lli->lli_lsm_sem);
1162 if (sa_low_hit(sai)) {
1164 atomic_inc(&sbi->ll_sa_wrong);
1166 "Statahead for dir "DFID" hit ratio too low: hit/miss %llu/%llu, sent/replied %llu/%llu, stoppingstatahead thread: pid %d\n",
1167 PFID(&lli->lli_fid), sai->sai_hit,
1168 sai->sai_miss, sai->sai_sent,
1169 sai->sai_replied, current->pid);
1173 ll_finish_md_op_data(op_data);
1176 spin_lock(&lli->lli_sa_lock);
1177 sai->sai_task = NULL;
1178 lli->lli_sa_enabled = 0;
1179 spin_unlock(&lli->lli_sa_lock);
1183 * statahead is finished, but statahead entries need to be cached, wait
1184 * for file release closedir() call to stop me.
1186 while (({set_current_state(TASK_IDLE);
1187 sai->sai_task; })) {
1190 __set_current_state(TASK_RUNNING);
1197 * wait for inflight statahead RPCs to finish, and then we can free sai
1198 * safely because statahead RPC will access sai data
1200 while (sai->sai_sent != sai->sai_replied)
1201 /* in case we're not woken up, timeout wait */
1204 CDEBUG(D_READA, "%s: statahead thread stopped: sai %p, parent %pd\n",
1205 sbi->ll_fsname, sai, parent);
1207 spin_lock(&lli->lli_sa_lock);
1208 sai->sai_task = NULL;
1209 spin_unlock(&lli->lli_sa_lock);
1210 wake_up(&sai->sai_waitq);
1212 atomic_add(sai->sai_hit, &sbi->ll_sa_hit_total);
1213 atomic_add(sai->sai_miss, &sbi->ll_sa_miss_total);
1220 /* authorize opened dir handle @key to statahead */
1221 void ll_authorize_statahead(struct inode *dir, void *key)
1223 struct ll_inode_info *lli = ll_i2info(dir);
1225 spin_lock(&lli->lli_sa_lock);
1226 if (!lli->lli_opendir_key && !lli->lli_sai) {
1228 * if lli_sai is not NULL, it means previous statahead is not
1229 * finished yet, we'd better not start a new statahead for now.
1231 LASSERT(lli->lli_opendir_pid == 0);
1232 lli->lli_opendir_key = key;
1233 lli->lli_opendir_pid = current->pid;
1234 lli->lli_sa_enabled = 1;
1236 spin_unlock(&lli->lli_sa_lock);
1240 * deauthorize opened dir handle @key to statahead, and notify statahead thread
1241 * to quit if it's running.
1243 void ll_deauthorize_statahead(struct inode *dir, void *key)
1245 struct ll_inode_info *lli = ll_i2info(dir);
1246 struct ll_statahead_info *sai;
1248 LASSERT(lli->lli_opendir_key == key);
1249 LASSERT(lli->lli_opendir_pid != 0);
1251 CDEBUG(D_READA, "deauthorize statahead for "DFID"\n",
1252 PFID(&lli->lli_fid));
1254 spin_lock(&lli->lli_sa_lock);
1255 lli->lli_opendir_key = NULL;
1256 lli->lli_opendir_pid = 0;
1257 lli->lli_sa_enabled = 0;
1259 if (sai && sai->sai_task) {
1261 * statahead thread may not have quit yet because it needs to
1262 * cache entries, now it's time to tell it to quit.
1264 * wake_up_process() provides the necessary barriers
1265 * to pair with set_current_state().
1267 struct task_struct *task = sai->sai_task;
1269 sai->sai_task = NULL;
1270 wake_up_process(task);
1272 spin_unlock(&lli->lli_sa_lock);
1277 * not first dirent, or is "."
1279 LS_NOT_FIRST_DE = 0,
1281 * the first non-hidden dirent
1285 * the first hidden dirent, that is "."
1290 /* file is first dirent under @dir */
1291 static int is_first_dirent(struct inode *dir, struct dentry *dentry)
1293 struct qstr *target = &dentry->d_name;
1294 struct md_op_data *op_data;
1296 struct page *page = NULL;
1297 int rc = LS_NOT_FIRST_DE;
1299 struct llcrypt_str lltr = LLTR_INIT(NULL, 0);
1303 op_data = ll_prep_md_op_data(NULL, dir, dir, NULL, 0, 0,
1304 LUSTRE_OPC_ANY, dir);
1305 if (IS_ERR(op_data))
1306 RETURN(PTR_ERR(op_data));
1308 if (IS_ENCRYPTED(dir)) {
1309 int rc2 = llcrypt_fname_alloc_buffer(dir, NAME_MAX, &lltr);
1316 *FIXME choose the start offset of the readdir
1319 page = ll_get_dir_page(dir, op_data, 0, NULL);
1322 struct lu_dirpage *dp;
1323 struct lu_dirent *ent;
1326 struct ll_inode_info *lli = ll_i2info(dir);
1329 CERROR("%s: reading dir "DFID" at %llu opendir_pid = %u : rc = %d\n",
1330 ll_i2sbi(dir)->ll_fsname,
1331 PFID(ll_inode2fid(dir)), pos,
1332 lli->lli_opendir_pid, rc);
1336 dp = page_address(page);
1337 for (ent = lu_dirent_start(dp); ent != NULL;
1338 ent = lu_dirent_next(ent)) {
1343 hash = le64_to_cpu(ent->lde_hash);
1345 * The ll_get_dir_page() can return any page containing
1346 * the given hash which may be not the start hash.
1348 if (unlikely(hash < pos))
1351 namelen = le16_to_cpu(ent->lde_namelen);
1352 if (unlikely(namelen == 0))
1354 * skip dummy record.
1358 name = ent->lde_name;
1359 if (name[0] == '.') {
1365 else if (name[1] == '.' && namelen == 2)
1376 if (dot_de && target->name[0] != '.') {
1377 CDEBUG(D_READA, "%.*s skip hidden file %.*s\n",
1378 target->len, target->name,
1383 if (IS_ENCRYPTED(dir)) {
1384 struct llcrypt_str de_name =
1385 LLTR_INIT(ent->lde_name, namelen);
1388 fid_le_to_cpu(&fid, &ent->lde_fid);
1389 if (ll_fname_disk_to_usr(dir, 0, 0, &de_name,
1396 if (target->len != namelen ||
1397 memcmp(target->name, name, namelen) != 0)
1398 rc = LS_NOT_FIRST_DE;
1402 rc = LS_FIRST_DOT_DE;
1404 ll_release_page(dir, page, false);
1407 pos = le64_to_cpu(dp->ldp_hash_end);
1408 if (pos == MDS_DIR_END_OFF) {
1410 * End of directory reached.
1412 ll_release_page(dir, page, false);
1416 * chain is exhausted
1417 * Normal case: continue to the next page.
1419 ll_release_page(dir, page, le32_to_cpu(dp->ldp_flags) &
1421 page = ll_get_dir_page(dir, op_data, pos, NULL);
1426 llcrypt_fname_free_buffer(&lltr);
1427 ll_finish_md_op_data(op_data);
1433 * revalidate @dentryp from statahead cache
1435 * \param[in] dir parent directory
1436 * \param[in] sai sai structure
1437 * \param[out] dentryp pointer to dentry which will be revalidated
1438 * \param[in] unplug unplug statahead window only (normally for negative
1440 * \retval 1 on success, dentry is saved in @dentryp
1441 * \retval 0 if revalidation failed (no proper lock on client)
1442 * \retval negative number upon error
1444 static int revalidate_statahead_dentry(struct inode *dir,
1445 struct ll_statahead_info *sai,
1446 struct dentry **dentryp,
1449 struct sa_entry *entry = NULL;
1450 struct ll_inode_info *lli = ll_i2info(dir);
1455 if ((*dentryp)->d_name.name[0] == '.') {
1456 if (sai->sai_ls_all ||
1457 sai->sai_miss_hidden >= sai->sai_skip_hidden) {
1459 * Hidden dentry is the first one, or statahead
1460 * thread does not skip so many hidden dentries
1461 * before "sai_ls_all" enabled as below.
1464 if (!sai->sai_ls_all)
1466 * It maybe because hidden dentry is not
1467 * the first one, "sai_ls_all" was not
1468 * set, then "ls -al" missed. Enable
1469 * "sai_ls_all" for such case.
1471 sai->sai_ls_all = 1;
1474 * Such "getattr" has been skipped before
1475 * "sai_ls_all" enabled as above.
1477 sai->sai_miss_hidden++;
1485 entry = sa_get(sai, &(*dentryp)->d_name);
1487 GOTO(out, rc = -EAGAIN);
1489 if (!sa_ready(entry)) {
1490 spin_lock(&lli->lli_sa_lock);
1491 sai->sai_index_wait = entry->se_index;
1492 spin_unlock(&lli->lli_sa_lock);
1493 rc = wait_event_idle_timeout(sai->sai_waitq, sa_ready(entry),
1494 cfs_time_seconds(30));
1497 * entry may not be ready, so it may be used by inflight
1498 * statahead RPC, don't free it.
1501 GOTO(out, rc = -EAGAIN);
1506 * We need to see the value that was set immediately before we
1509 if (smp_load_acquire(&entry->se_state) == SA_ENTRY_SUCC &&
1511 struct inode *inode = entry->se_inode;
1512 struct lookup_intent it = { .it_op = IT_GETATTR,
1517 rc = md_revalidate_lock(ll_i2mdexp(dir), &it,
1518 ll_inode2fid(inode), &bits);
1520 if (!(*dentryp)->d_inode) {
1521 struct dentry *alias;
1523 alias = ll_splice_alias(inode, *dentryp);
1524 if (IS_ERR(alias)) {
1525 ll_intent_release(&it);
1526 GOTO(out, rc = PTR_ERR(alias));
1530 * statahead prepared this inode, transfer inode
1531 * refcount from sa_entry to dentry
1533 entry->se_inode = NULL;
1534 } else if ((*dentryp)->d_inode != inode) {
1535 /* revalidate, but inode is recreated */
1537 "%s: stale dentry %pd inode " DFID", statahead inode "DFID "\n",
1538 ll_i2sbi(inode)->ll_fsname, *dentryp,
1539 PFID(ll_inode2fid((*dentryp)->d_inode)),
1540 PFID(ll_inode2fid(inode)));
1541 ll_intent_release(&it);
1542 GOTO(out, rc = -ESTALE);
1545 if ((bits & MDS_INODELOCK_LOOKUP) &&
1546 d_lustre_invalid(*dentryp)) {
1547 d_lustre_revalidate(*dentryp);
1548 ll_update_dir_depth(dir, (*dentryp)->d_inode);
1551 ll_intent_release(&it);
1556 * statahead cached sa_entry can be used only once, and will be killed
1557 * right after use, so if lookup/revalidate accessed statahead cache,
1558 * set dentry ldd_sa_generation to parent lli_sa_generation, later if we
1559 * stat this file again, we know we've done statahead before, see
1560 * dentry_may_statahead().
1562 if (lld_is_init(*dentryp))
1563 ll_d2d(*dentryp)->lld_sa_generation = lli->lli_sa_generation;
1565 spin_lock(&lli->lli_sa_lock);
1567 wake_up_process(sai->sai_task);
1568 spin_unlock(&lli->lli_sa_lock);
1574 * start statahead thread
1576 * \param[in] dir parent directory
1577 * \param[in] dentry dentry that triggers statahead, normally the first
1579 * \param[in] agl indicate whether AGL is needed
1580 * \retval -EAGAIN on success, because when this function is
1581 * called, it's already in lookup call, so client should
1582 * do it itself instead of waiting for statahead thread
1583 * to do it asynchronously.
1584 * \retval negative number upon error
1586 static int start_statahead_thread(struct inode *dir, struct dentry *dentry,
1589 int node = cfs_cpt_spread_node(cfs_cpt_tab, CFS_CPT_ANY);
1590 struct ll_inode_info *lli = ll_i2info(dir);
1591 struct ll_statahead_info *sai = NULL;
1592 struct dentry *parent = dentry->d_parent;
1593 struct task_struct *task;
1594 struct ll_sb_info *sbi = ll_i2sbi(parent->d_inode);
1595 int first = LS_FIRST_DE;
1600 /* I am the "lli_opendir_pid" owner, only me can set "lli_sai". */
1601 first = is_first_dirent(dir, dentry);
1602 if (first == LS_NOT_FIRST_DE)
1603 /* It is not "ls -{a}l" operation, no need statahead for it. */
1604 GOTO(out, rc = -EFAULT);
1606 if (unlikely(atomic_inc_return(&sbi->ll_sa_running) >
1607 sbi->ll_sa_running_max)) {
1609 "Too many concurrent statahead instances, avoid new statahead instance temporarily.\n");
1610 GOTO(out, rc = -EMFILE);
1613 sai = ll_sai_alloc(parent);
1615 GOTO(out, rc = -ENOMEM);
1617 sai->sai_ls_all = (first == LS_FIRST_DOT_DE);
1620 * if current lli_opendir_key was deauthorized, or dir re-opened by
1621 * another process, don't start statahead, otherwise the newly spawned
1622 * statahead thread won't be notified to quit.
1624 spin_lock(&lli->lli_sa_lock);
1625 if (unlikely(lli->lli_sai || !lli->lli_opendir_key ||
1626 lli->lli_opendir_pid != current->pid)) {
1627 spin_unlock(&lli->lli_sa_lock);
1628 GOTO(out, rc = -EPERM);
1631 spin_unlock(&lli->lli_sa_lock);
1633 CDEBUG(D_READA, "start statahead thread: [pid %d] [parent %pd]\n",
1634 current->pid, parent);
1636 task = kthread_create_on_node(ll_statahead_thread, parent, node,
1637 "ll_sa_%u", lli->lli_opendir_pid);
1639 spin_lock(&lli->lli_sa_lock);
1640 lli->lli_sai = NULL;
1641 spin_unlock(&lli->lli_sa_lock);
1643 CERROR("can't start ll_sa thread, rc: %d\n", rc);
1647 if (test_bit(LL_SBI_AGL_ENABLED, ll_i2sbi(parent->d_inode)->ll_flags) &&
1649 ll_start_agl(parent, sai);
1651 atomic_inc(&ll_i2sbi(parent->d_inode)->ll_sa_total);
1652 sai->sai_task = task;
1654 wake_up_process(task);
1656 * We don't stat-ahead for the first dirent since we are already in
1663 * once we start statahead thread failed, disable statahead so that
1664 * subsequent stat won't waste time to try it.
1666 spin_lock(&lli->lli_sa_lock);
1667 if (lli->lli_opendir_pid == current->pid)
1668 lli->lli_sa_enabled = 0;
1669 spin_unlock(&lli->lli_sa_lock);
1673 if (first != LS_NOT_FIRST_DE)
1674 atomic_dec(&sbi->ll_sa_running);
1680 * Check whether statahead for @dir was started.
1682 static inline bool ll_statahead_started(struct inode *dir, bool agl)
1684 struct ll_inode_info *lli = ll_i2info(dir);
1685 struct ll_statahead_info *sai;
1687 spin_lock(&lli->lli_sa_lock);
1689 if (sai && (sai->sai_agl_task != NULL) != agl)
1691 "%s: Statahead AGL hint changed from %d to %d\n",
1692 ll_i2sbi(dir)->ll_fsname,
1693 sai->sai_agl_task != NULL, agl);
1694 spin_unlock(&lli->lli_sa_lock);
1700 * statahead entry function, this is called when client getattr on a file, it
1701 * will start statahead thread if this is the first dir entry, else revalidate
1702 * dentry from statahead cache.
1704 * \param[in] dir parent directory
1705 * \param[out] dentryp dentry to getattr
1706 * \param[in] agl whether start the agl thread
1708 * \retval 1 on success
1709 * \retval 0 revalidation from statahead cache failed, caller needs
1710 * to getattr from server directly
1711 * \retval negative number on error, caller often ignores this and
1712 * then getattr from server
1714 int ll_start_statahead(struct inode *dir, struct dentry *dentry, bool agl)
1716 if (!ll_statahead_started(dir, agl))
1717 return start_statahead_thread(dir, dentry, agl);
1722 * revalidate dentry from statahead cache.
1724 * \param[in] dir parent directory
1725 * \param[out] dentryp dentry to getattr
1726 * \param[in] unplug unplug statahead window only (normally for negative
1728 * \retval 1 on success
1729 * \retval 0 revalidation from statahead cache failed, caller needs
1730 * to getattr from server directly
1731 * \retval negative number on error, caller often ignores this and
1732 * then getattr from server
1734 int ll_revalidate_statahead(struct inode *dir, struct dentry **dentryp,
1737 struct ll_statahead_info *sai;
1740 sai = ll_sai_get(dir);
1742 rc = revalidate_statahead_dentry(dir, sai, dentryp, unplug);
1743 CDEBUG(D_READA, "revalidate statahead %pd: rc = %d.\n",