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 /* Someone is in glimpse (sync or async), do nothing. */
510 rc = down_write_trylock(&lli->lli_glimpse_sem);
512 lli->lli_agl_index = 0;
518 * Someone triggered glimpse within 1 sec before.
519 * 1) The former glimpse succeeded with glimpse lock granted by OST, and
520 * if the lock is still cached on client, AGL needs to do nothing. If
521 * it is cancelled by other client, AGL maybe cannot obtaion new lock
522 * for no glimpse callback triggered by AGL.
523 * 2) The former glimpse succeeded, but OST did not grant glimpse lock.
524 * Under such case, it is quite possible that the OST will not grant
525 * glimpse lock for AGL also.
526 * 3) The former glimpse failed, compared with other two cases, it is
527 * relative rare. AGL can ignore such case, and it will not muchly
528 * affect the performance.
530 if (lli->lli_glimpse_time != 0 &&
531 cfs_time_before(cfs_time_shift(-1), lli->lli_glimpse_time)) {
532 up_write(&lli->lli_glimpse_sem);
533 lli->lli_agl_index = 0;
538 CDEBUG(D_READA, "Handling (init) async glimpse: inode = "
539 DFID", idx = %llu\n", PFID(&lli->lli_fid), index);
542 lli->lli_agl_index = 0;
543 lli->lli_glimpse_time = cfs_time_current();
544 up_write(&lli->lli_glimpse_sem);
546 CDEBUG(D_READA, "Handled (init) async glimpse: inode= "
547 DFID", idx = %llu, rc = %d\n",
548 PFID(&lli->lli_fid), index, rc);
556 * prepare inode for sa entry, add it into agl list, now sa_entry is ready
557 * to be used by scanner process.
559 static void sa_instantiate(struct ll_statahead_info *sai,
560 struct sa_entry *entry)
562 struct inode *dir = sai->sai_dentry->d_inode;
564 struct md_enqueue_info *minfo;
565 struct lookup_intent *it;
566 struct ptlrpc_request *req;
567 struct mdt_body *body;
571 LASSERT(entry->se_handle != 0);
573 minfo = entry->se_minfo;
576 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
578 GOTO(out, rc = -EFAULT);
580 child = entry->se_inode;
582 /* revalidate; unlinked and re-created with the same name */
583 if (unlikely(!lu_fid_eq(&minfo->mi_data.op_fid2,
585 entry->se_inode = NULL;
591 it->it_lock_handle = entry->se_handle;
592 rc = md_revalidate_lock(ll_i2mdexp(dir), it, ll_inode2fid(dir), NULL);
594 GOTO(out, rc = -EAGAIN);
596 rc = ll_prep_inode(&child, req, dir->i_sb, it);
600 CDEBUG(D_READA, "%s: setting %.*s"DFID" l_data to inode %p\n",
601 ll_get_fsname(child->i_sb, NULL, 0),
602 entry->se_qstr.len, entry->se_qstr.name,
603 PFID(ll_inode2fid(child)), child);
604 ll_set_lock_data(ll_i2sbi(dir)->ll_md_exp, child, it, NULL);
606 entry->se_inode = child;
608 if (agl_should_run(sai, child))
609 ll_agl_add(sai, child, entry->se_index);
614 /* sa_make_ready() will drop ldlm ibits lock refcount by calling
615 * ll_intent_drop_lock() in spite of failures. Do not worry about
616 * calling ll_intent_drop_lock() more than once. */
617 sa_make_ready(sai, entry, rc);
620 /* once there are async stat replies, instantiate sa_entry from replies */
621 static void sa_handle_callback(struct ll_statahead_info *sai)
623 struct ll_inode_info *lli;
625 lli = ll_i2info(sai->sai_dentry->d_inode);
627 while (sa_has_callback(sai)) {
628 struct sa_entry *entry;
630 spin_lock(&lli->lli_sa_lock);
631 if (unlikely(!sa_has_callback(sai))) {
632 spin_unlock(&lli->lli_sa_lock);
635 entry = list_entry(sai->sai_interim_entries.next,
636 struct sa_entry, se_list);
637 list_del_init(&entry->se_list);
638 spin_unlock(&lli->lli_sa_lock);
640 sa_instantiate(sai, entry);
645 * callback for async stat RPC, because this is called in ptlrpcd context, we
646 * only put sa_entry in sai_interim_entries, and wake up statahead thread to
647 * really prepare inode and instantiate sa_entry later.
649 static int ll_statahead_interpret(struct ptlrpc_request *req,
650 struct md_enqueue_info *minfo, int rc)
652 struct lookup_intent *it = &minfo->mi_it;
653 struct inode *dir = minfo->mi_dir;
654 struct ll_inode_info *lli = ll_i2info(dir);
655 struct ll_statahead_info *sai = lli->lli_sai;
656 struct sa_entry *entry = (struct sa_entry *)minfo->mi_cbdata;
658 wait_queue_head_t *waitq = NULL;
661 if (it_disposition(it, DISP_LOOKUP_NEG))
664 /* because statahead thread will wait for all inflight RPC to finish,
665 * sai should be always valid, no need to refcount */
666 LASSERT(sai != NULL);
667 LASSERT(!thread_is_stopped(&sai->sai_thread));
668 LASSERT(entry != NULL);
670 CDEBUG(D_READA, "sa_entry %.*s rc %d\n",
671 entry->se_qstr.len, entry->se_qstr.name, rc);
674 ll_intent_release(it);
678 /* release ibits lock ASAP to avoid deadlock when statahead
679 * thread enqueues lock on parent in readdir and another
680 * process enqueues lock on child with parent lock held, eg.
682 handle = it->it_lock_handle;
683 ll_intent_drop_lock(it);
686 spin_lock(&lli->lli_sa_lock);
688 if (__sa_make_ready(sai, entry, rc))
689 waitq = &sai->sai_waitq;
691 entry->se_minfo = minfo;
692 entry->se_req = ptlrpc_request_addref(req);
693 /* Release the async ibits lock ASAP to avoid deadlock
694 * when statahead thread tries to enqueue lock on parent
695 * for readpage and other tries to enqueue lock on child
696 * with parent's lock held, for example: unlink. */
697 entry->se_handle = handle;
698 if (!sa_has_callback(sai))
699 waitq = &sai->sai_thread.t_ctl_waitq;
701 list_add_tail(&entry->se_list, &sai->sai_interim_entries);
706 spin_unlock(&lli->lli_sa_lock);
711 /* finish async stat RPC arguments */
712 static void sa_fini_data(struct md_enqueue_info *minfo)
719 * prepare arguments for async stat RPC.
721 static struct md_enqueue_info *
722 sa_prep_data(struct inode *dir, struct inode *child, struct sa_entry *entry)
724 struct md_enqueue_info *minfo;
725 struct ldlm_enqueue_info *einfo;
726 struct md_op_data *op_data;
728 OBD_ALLOC_PTR(minfo);
730 return ERR_PTR(-ENOMEM);
732 op_data = ll_prep_md_op_data(&minfo->mi_data, dir, child, NULL, 0, 0,
733 LUSTRE_OPC_ANY, NULL);
734 if (IS_ERR(op_data)) {
736 return (struct md_enqueue_info *)op_data;
740 op_data->op_fid2 = entry->se_fid;
742 minfo->mi_it.it_op = IT_GETATTR;
743 minfo->mi_dir = igrab(dir);
744 minfo->mi_cb = ll_statahead_interpret;
745 minfo->mi_cbdata = entry;
747 einfo = &minfo->mi_einfo;
748 einfo->ei_type = LDLM_IBITS;
749 einfo->ei_mode = it_to_lock_mode(&minfo->mi_it);
750 einfo->ei_cb_bl = ll_md_blocking_ast;
751 einfo->ei_cb_cp = ldlm_completion_ast;
752 einfo->ei_cb_gl = NULL;
753 einfo->ei_cbdata = NULL;
758 /* async stat for file not found in dcache */
759 static int sa_lookup(struct inode *dir, struct sa_entry *entry)
761 struct md_enqueue_info *minfo;
765 minfo = sa_prep_data(dir, NULL, entry);
767 RETURN(PTR_ERR(minfo));
769 rc = md_intent_getattr_async(ll_i2mdexp(dir), minfo);
777 * async stat for file found in dcache, similar to .revalidate
779 * \retval 1 dentry valid, no RPC sent
780 * \retval 0 dentry invalid, will send async stat RPC
781 * \retval negative number upon error
783 static int sa_revalidate(struct inode *dir, struct sa_entry *entry,
784 struct dentry *dentry)
786 struct inode *inode = dentry->d_inode;
787 struct lookup_intent it = { .it_op = IT_GETATTR,
788 .it_lock_handle = 0 };
789 struct md_enqueue_info *minfo;
793 if (unlikely(inode == NULL))
796 if (d_mountpoint(dentry))
799 entry->se_inode = igrab(inode);
800 rc = md_revalidate_lock(ll_i2mdexp(dir), &it, ll_inode2fid(inode),
803 entry->se_handle = it.it_lock_handle;
804 ll_intent_release(&it);
808 minfo = sa_prep_data(dir, inode, entry);
810 entry->se_inode = NULL;
812 RETURN(PTR_ERR(minfo));
815 rc = md_intent_getattr_async(ll_i2mdexp(dir), minfo);
817 entry->se_inode = NULL;
825 /* async stat for file with @name */
826 static void sa_statahead(struct dentry *parent, const char *name, int len,
827 const struct lu_fid *fid)
829 struct inode *dir = parent->d_inode;
830 struct ll_inode_info *lli = ll_i2info(dir);
831 struct ll_statahead_info *sai = lli->lli_sai;
832 struct dentry *dentry = NULL;
833 struct sa_entry *entry;
837 entry = sa_alloc(sai, sai->sai_index, name, len, fid);
841 dentry = d_lookup(parent, &entry->se_qstr);
843 rc = sa_lookup(dir, entry);
845 rc = sa_revalidate(dir, entry, dentry);
846 if (rc == 1 && agl_should_run(sai, dentry->d_inode))
847 ll_agl_add(sai, dentry->d_inode, entry->se_index);
854 sa_make_ready(sai, entry, rc);
863 /* async glimpse (agl) thread main function */
864 static int ll_agl_thread(void *arg)
866 struct dentry *parent = (struct dentry *)arg;
867 struct inode *dir = parent->d_inode;
868 struct ll_inode_info *plli = ll_i2info(dir);
869 struct ll_inode_info *clli;
870 struct ll_sb_info *sbi = ll_i2sbi(dir);
871 struct ll_statahead_info *sai;
872 struct ptlrpc_thread *thread;
873 struct l_wait_info lwi = { 0 };
877 sai = ll_sai_get(dir);
878 thread = &sai->sai_agl_thread;
879 thread->t_pid = current_pid();
880 CDEBUG(D_READA, "agl thread started: sai %p, parent %.*s\n",
881 sai, parent->d_name.len, parent->d_name.name);
883 atomic_inc(&sbi->ll_agl_total);
884 spin_lock(&plli->lli_agl_lock);
885 sai->sai_agl_valid = 1;
886 if (thread_is_init(thread))
887 /* If someone else has changed the thread state
888 * (e.g. already changed to SVC_STOPPING), we can't just
889 * blindly overwrite that setting. */
890 thread_set_flags(thread, SVC_RUNNING);
891 spin_unlock(&plli->lli_agl_lock);
892 wake_up(&thread->t_ctl_waitq);
895 l_wait_event(thread->t_ctl_waitq,
896 !agl_list_empty(sai) ||
897 !thread_is_running(thread),
900 if (!thread_is_running(thread))
903 spin_lock(&plli->lli_agl_lock);
904 /* The statahead thread maybe help to process AGL entries,
905 * so check whether list empty again. */
906 if (!agl_list_empty(sai)) {
907 clli = agl_first_entry(sai);
908 list_del_init(&clli->lli_agl_list);
909 spin_unlock(&plli->lli_agl_lock);
910 ll_agl_trigger(&clli->lli_vfs_inode, sai);
912 spin_unlock(&plli->lli_agl_lock);
916 spin_lock(&plli->lli_agl_lock);
917 sai->sai_agl_valid = 0;
918 while (!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 clli->lli_agl_index = 0;
923 iput(&clli->lli_vfs_inode);
924 spin_lock(&plli->lli_agl_lock);
926 thread_set_flags(thread, SVC_STOPPED);
927 spin_unlock(&plli->lli_agl_lock);
928 wake_up(&thread->t_ctl_waitq);
930 CDEBUG(D_READA, "agl thread stopped: sai %p, parent %.*s\n",
931 sai, parent->d_name.len, parent->d_name.name);
935 /* start agl thread */
936 static void ll_start_agl(struct dentry *parent, struct ll_statahead_info *sai)
938 struct ptlrpc_thread *thread = &sai->sai_agl_thread;
939 struct l_wait_info lwi = { 0 };
940 struct ll_inode_info *plli;
941 struct task_struct *task;
944 CDEBUG(D_READA, "start agl thread: sai %p, parent %.*s\n",
945 sai, parent->d_name.len, parent->d_name.name);
947 plli = ll_i2info(parent->d_inode);
948 task = kthread_run(ll_agl_thread, parent,
949 "ll_agl_%u", plli->lli_opendir_pid);
951 CERROR("can't start ll_agl thread, rc: %ld\n", PTR_ERR(task));
952 thread_set_flags(thread, SVC_STOPPED);
956 l_wait_event(thread->t_ctl_waitq,
957 thread_is_running(thread) || thread_is_stopped(thread),
962 /* statahead thread main function */
963 static int ll_statahead_thread(void *arg)
965 struct dentry *parent = (struct dentry *)arg;
966 struct inode *dir = parent->d_inode;
967 struct ll_inode_info *lli = ll_i2info(dir);
968 struct ll_sb_info *sbi = ll_i2sbi(dir);
969 struct ll_statahead_info *sai;
970 struct ptlrpc_thread *sa_thread;
971 struct ptlrpc_thread *agl_thread;
973 struct md_op_data *op_data;
974 struct ll_dir_chain chain;
975 struct l_wait_info lwi = { 0 };
976 struct page *page = NULL;
981 sai = ll_sai_get(dir);
982 sa_thread = &sai->sai_thread;
983 agl_thread = &sai->sai_agl_thread;
984 sa_thread->t_pid = current_pid();
985 CDEBUG(D_READA, "statahead thread starting: sai %p, parent %.*s\n",
986 sai, parent->d_name.len, parent->d_name.name);
988 op_data = ll_prep_md_op_data(NULL, dir, dir, NULL, 0, 0,
989 LUSTRE_OPC_ANY, dir);
991 GOTO(out, rc = PTR_ERR(op_data));
993 op_data->op_max_pages = ll_i2sbi(dir)->ll_md_brw_pages;
995 if (sbi->ll_flags & LL_SBI_AGL_ENABLED)
996 ll_start_agl(parent, sai);
998 atomic_inc(&sbi->ll_sa_total);
999 spin_lock(&lli->lli_sa_lock);
1000 if (thread_is_init(sa_thread))
1001 /* If someone else has changed the thread state
1002 * (e.g. already changed to SVC_STOPPING), we can't just
1003 * blindly overwrite that setting. */
1004 thread_set_flags(sa_thread, SVC_RUNNING);
1005 spin_unlock(&lli->lli_sa_lock);
1006 wake_up(&sa_thread->t_ctl_waitq);
1008 ll_dir_chain_init(&chain);
1009 while (pos != MDS_DIR_END_OFF && thread_is_running(sa_thread)) {
1010 struct lu_dirpage *dp;
1011 struct lu_dirent *ent;
1013 sai->sai_in_readpage = 1;
1014 page = ll_get_dir_page(dir, op_data, pos, &chain);
1015 sai->sai_in_readpage = 0;
1018 CDEBUG(D_READA, "error reading dir "DFID" at %llu"
1019 "/%llu opendir_pid = %u: rc = %d\n",
1020 PFID(ll_inode2fid(dir)), pos, sai->sai_index,
1021 lli->lli_opendir_pid, rc);
1025 dp = page_address(page);
1026 for (ent = lu_dirent_start(dp);
1027 ent != NULL && thread_is_running(sa_thread) &&
1029 ent = lu_dirent_next(ent)) {
1035 hash = le64_to_cpu(ent->lde_hash);
1036 if (unlikely(hash < pos))
1038 * Skip until we find target hash value.
1042 namelen = le16_to_cpu(ent->lde_namelen);
1043 if (unlikely(namelen == 0))
1045 * Skip dummy record.
1049 name = ent->lde_name;
1050 if (name[0] == '.') {
1056 } else if (name[1] == '.' && namelen == 2) {
1061 } else if (!sai->sai_ls_all) {
1063 * skip hidden files.
1065 sai->sai_skip_hidden++;
1071 * don't stat-ahead first entry.
1073 if (unlikely(++first == 1))
1076 fid_le_to_cpu(&fid, &ent->lde_fid);
1078 /* wait for spare statahead window */
1080 l_wait_event(sa_thread->t_ctl_waitq,
1081 !sa_sent_full(sai) ||
1082 sa_has_callback(sai) ||
1083 !agl_list_empty(sai) ||
1084 !thread_is_running(sa_thread),
1087 sa_handle_callback(sai);
1089 spin_lock(&lli->lli_agl_lock);
1090 while (sa_sent_full(sai) &&
1091 !agl_list_empty(sai)) {
1092 struct ll_inode_info *clli;
1094 clli = agl_first_entry(sai);
1095 list_del_init(&clli->lli_agl_list);
1096 spin_unlock(&lli->lli_agl_lock);
1098 ll_agl_trigger(&clli->lli_vfs_inode,
1101 spin_lock(&lli->lli_agl_lock);
1103 spin_unlock(&lli->lli_agl_lock);
1104 } while (sa_sent_full(sai) &&
1105 thread_is_running(sa_thread));
1107 sa_statahead(parent, name, namelen, &fid);
1110 pos = le64_to_cpu(dp->ldp_hash_end);
1111 ll_release_page(dir, page,
1112 le32_to_cpu(dp->ldp_flags) & LDF_COLLIDE);
1114 if (sa_low_hit(sai)) {
1116 atomic_inc(&sbi->ll_sa_wrong);
1117 CDEBUG(D_READA, "Statahead for dir "DFID" hit "
1118 "ratio too low: hit/miss %llu/%llu"
1119 ", sent/replied %llu/%llu, stopping "
1120 "statahead thread: pid %d\n",
1121 PFID(&lli->lli_fid), sai->sai_hit,
1122 sai->sai_miss, sai->sai_sent,
1123 sai->sai_replied, current_pid());
1127 ll_dir_chain_fini(&chain);
1128 ll_finish_md_op_data(op_data);
1131 spin_lock(&lli->lli_sa_lock);
1132 thread_set_flags(sa_thread, SVC_STOPPING);
1133 lli->lli_sa_enabled = 0;
1134 spin_unlock(&lli->lli_sa_lock);
1137 /* statahead is finished, but statahead entries need to be cached, wait
1138 * for file release to stop me. */
1139 while (thread_is_running(sa_thread)) {
1140 l_wait_event(sa_thread->t_ctl_waitq,
1141 sa_has_callback(sai) ||
1142 !thread_is_running(sa_thread),
1145 sa_handle_callback(sai);
1150 if (sai->sai_agl_valid) {
1151 spin_lock(&lli->lli_agl_lock);
1152 thread_set_flags(agl_thread, SVC_STOPPING);
1153 spin_unlock(&lli->lli_agl_lock);
1154 wake_up(&agl_thread->t_ctl_waitq);
1156 CDEBUG(D_READA, "stop agl thread: sai %p pid %u\n",
1157 sai, (unsigned int)agl_thread->t_pid);
1158 l_wait_event(agl_thread->t_ctl_waitq,
1159 thread_is_stopped(agl_thread),
1162 /* Set agl_thread flags anyway. */
1163 thread_set_flags(agl_thread, SVC_STOPPED);
1166 /* wait for inflight statahead RPCs to finish, and then we can free sai
1167 * safely because statahead RPC will access sai data */
1168 while (sai->sai_sent != sai->sai_replied) {
1169 /* in case we're not woken up, timeout wait */
1170 lwi = LWI_TIMEOUT(msecs_to_jiffies(MSEC_PER_SEC >> 3),
1172 l_wait_event(sa_thread->t_ctl_waitq,
1173 sai->sai_sent == sai->sai_replied, &lwi);
1176 /* release resources held by statahead RPCs */
1177 sa_handle_callback(sai);
1179 spin_lock(&lli->lli_sa_lock);
1180 thread_set_flags(sa_thread, SVC_STOPPED);
1181 spin_unlock(&lli->lli_sa_lock);
1183 CDEBUG(D_READA, "statahead thread stopped: sai %p, parent %.*s\n",
1184 sai, parent->d_name.len, parent->d_name.name);
1186 wake_up(&sai->sai_waitq);
1187 wake_up(&sa_thread->t_ctl_waitq);
1193 /* authorize opened dir handle @key to statahead */
1194 void ll_authorize_statahead(struct inode *dir, void *key)
1196 struct ll_inode_info *lli = ll_i2info(dir);
1198 spin_lock(&lli->lli_sa_lock);
1199 if (lli->lli_opendir_key == NULL && lli->lli_sai == NULL) {
1201 * if lli_sai is not NULL, it means previous statahead is not
1202 * finished yet, we'd better not start a new statahead for now.
1204 LASSERT(lli->lli_opendir_pid == 0);
1205 lli->lli_opendir_key = key;
1206 lli->lli_opendir_pid = current_pid();
1207 lli->lli_sa_enabled = 1;
1209 spin_unlock(&lli->lli_sa_lock);
1213 * deauthorize opened dir handle @key to statahead, and notify statahead thread
1214 * to quit if it's running.
1216 void ll_deauthorize_statahead(struct inode *dir, void *key)
1218 struct ll_inode_info *lli = ll_i2info(dir);
1219 struct ll_statahead_info *sai;
1221 LASSERT(lli->lli_opendir_key == key);
1222 LASSERT(lli->lli_opendir_pid != 0);
1224 CDEBUG(D_READA, "deauthorize statahead for "DFID"\n",
1225 PFID(&lli->lli_fid));
1227 spin_lock(&lli->lli_sa_lock);
1228 lli->lli_opendir_key = NULL;
1229 lli->lli_opendir_pid = 0;
1230 lli->lli_sa_enabled = 0;
1232 if (sai != NULL && thread_is_running(&sai->sai_thread)) {
1234 * statahead thread may not quit yet because it needs to cache
1235 * entries, now it's time to tell it to quit.
1237 * In case sai is released, wake_up() is called inside spinlock,
1238 * so we have to call smp_mb() explicitely to serialize ops.
1240 thread_set_flags(&sai->sai_thread, SVC_STOPPING);
1242 wake_up(&sai->sai_thread.t_ctl_waitq);
1244 spin_unlock(&lli->lli_sa_lock);
1249 * not first dirent, or is "."
1251 LS_NOT_FIRST_DE = 0,
1253 * the first non-hidden dirent
1257 * the first hidden dirent, that is "."
1262 /* file is first dirent under @dir */
1263 static int is_first_dirent(struct inode *dir, struct dentry *dentry)
1265 struct ll_dir_chain chain;
1266 struct qstr *target = &dentry->d_name;
1267 struct md_op_data *op_data;
1269 struct page *page = NULL;
1270 int rc = LS_NOT_FIRST_DE;
1274 op_data = ll_prep_md_op_data(NULL, dir, dir, NULL, 0, 0,
1275 LUSTRE_OPC_ANY, dir);
1276 if (IS_ERR(op_data))
1277 RETURN(PTR_ERR(op_data));
1279 *FIXME choose the start offset of the readdir
1281 op_data->op_max_pages = ll_i2sbi(dir)->ll_md_brw_pages;
1283 ll_dir_chain_init(&chain);
1284 page = ll_get_dir_page(dir, op_data, 0, &chain);
1287 struct lu_dirpage *dp;
1288 struct lu_dirent *ent;
1291 struct ll_inode_info *lli = ll_i2info(dir);
1294 CERROR("%s: reading dir "DFID" at %llu"
1295 "opendir_pid = %u : rc = %d\n",
1296 ll_get_fsname(dir->i_sb, NULL, 0),
1297 PFID(ll_inode2fid(dir)), pos,
1298 lli->lli_opendir_pid, rc);
1302 dp = page_address(page);
1303 for (ent = lu_dirent_start(dp); ent != NULL;
1304 ent = lu_dirent_next(ent)) {
1309 hash = le64_to_cpu(ent->lde_hash);
1310 /* The ll_get_dir_page() can return any page containing
1311 * the given hash which may be not the start hash. */
1312 if (unlikely(hash < pos))
1315 namelen = le16_to_cpu(ent->lde_namelen);
1316 if (unlikely(namelen == 0))
1318 * skip dummy record.
1322 name = ent->lde_name;
1323 if (name[0] == '.') {
1329 else if (name[1] == '.' && namelen == 2)
1340 if (dot_de && target->name[0] != '.') {
1341 CDEBUG(D_READA, "%.*s skip hidden file %.*s\n",
1342 target->len, target->name,
1347 if (target->len != namelen ||
1348 memcmp(target->name, name, namelen) != 0)
1349 rc = LS_NOT_FIRST_DE;
1353 rc = LS_FIRST_DOT_DE;
1355 ll_release_page(dir, page, false);
1358 pos = le64_to_cpu(dp->ldp_hash_end);
1359 if (pos == MDS_DIR_END_OFF) {
1361 * End of directory reached.
1363 ll_release_page(dir, page, false);
1367 * chain is exhausted
1368 * Normal case: continue to the next page.
1370 ll_release_page(dir, page, le32_to_cpu(dp->ldp_flags) &
1372 page = ll_get_dir_page(dir, op_data, pos, &chain);
1377 ll_dir_chain_fini(&chain);
1378 ll_finish_md_op_data(op_data);
1383 * revalidate @dentryp from statahead cache
1385 * \param[in] dir parent directory
1386 * \param[in] sai sai structure
1387 * \param[out] dentryp pointer to dentry which will be revalidated
1388 * \param[in] unplug unplug statahead window only (normally for negative
1390 * \retval 1 on success, dentry is saved in @dentryp
1391 * \retval 0 if revalidation failed (no proper lock on client)
1392 * \retval negative number upon error
1394 static int revalidate_statahead_dentry(struct inode *dir,
1395 struct ll_statahead_info *sai,
1396 struct dentry **dentryp,
1399 struct sa_entry *entry = NULL;
1400 struct l_wait_info lwi = { 0 };
1401 struct ll_dentry_data *ldd;
1402 struct ll_inode_info *lli = ll_i2info(dir);
1406 if ((*dentryp)->d_name.name[0] == '.') {
1407 if (sai->sai_ls_all ||
1408 sai->sai_miss_hidden >= sai->sai_skip_hidden) {
1410 * Hidden dentry is the first one, or statahead
1411 * thread does not skip so many hidden dentries
1412 * before "sai_ls_all" enabled as below.
1415 if (!sai->sai_ls_all)
1417 * It maybe because hidden dentry is not
1418 * the first one, "sai_ls_all" was not
1419 * set, then "ls -al" missed. Enable
1420 * "sai_ls_all" for such case.
1422 sai->sai_ls_all = 1;
1425 * Such "getattr" has been skipped before
1426 * "sai_ls_all" enabled as above.
1428 sai->sai_miss_hidden++;
1436 entry = sa_get(sai, &(*dentryp)->d_name);
1438 GOTO(out, rc = -EAGAIN);
1440 /* if statahead is busy in readdir, help it do post-work */
1441 if (!sa_ready(entry) && sai->sai_in_readpage)
1442 sa_handle_callback(sai);
1444 if (!sa_ready(entry)) {
1445 spin_lock(&lli->lli_sa_lock);
1446 sai->sai_index_wait = entry->se_index;
1447 spin_unlock(&lli->lli_sa_lock);
1448 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(30), NULL,
1449 LWI_ON_SIGNAL_NOOP, NULL);
1450 rc = l_wait_event(sai->sai_waitq, sa_ready(entry), &lwi);
1453 * entry may not be ready, so it may be used by inflight
1454 * statahead RPC, don't free it.
1457 GOTO(out, rc = -EAGAIN);
1461 if (entry->se_state == SA_ENTRY_SUCC && entry->se_inode != NULL) {
1462 struct inode *inode = entry->se_inode;
1463 struct lookup_intent it = { .it_op = IT_GETATTR,
1468 rc = md_revalidate_lock(ll_i2mdexp(dir), &it,
1469 ll_inode2fid(inode), &bits);
1471 if ((*dentryp)->d_inode == NULL) {
1472 struct dentry *alias;
1474 alias = ll_splice_alias(inode, *dentryp);
1475 if (IS_ERR(alias)) {
1476 ll_intent_release(&it);
1477 GOTO(out, rc = PTR_ERR(alias));
1480 /* statahead prepared this inode, transfer inode
1481 * refcount from sa_entry to dentry */
1482 entry->se_inode = NULL;
1483 } else if ((*dentryp)->d_inode != inode) {
1484 /* revalidate, but inode is recreated */
1486 "%s: stale dentry %.*s inode "
1487 DFID", statahead inode "DFID
1489 ll_get_fsname((*dentryp)->d_inode->i_sb,
1491 (*dentryp)->d_name.len,
1492 (*dentryp)->d_name.name,
1493 PFID(ll_inode2fid((*dentryp)->d_inode)),
1494 PFID(ll_inode2fid(inode)));
1495 ll_intent_release(&it);
1496 GOTO(out, rc = -ESTALE);
1499 if ((bits & MDS_INODELOCK_LOOKUP) &&
1500 d_lustre_invalid(*dentryp))
1501 d_lustre_revalidate(*dentryp);
1502 ll_intent_release(&it);
1507 * statahead cached sa_entry can be used only once, and will be killed
1508 * right after use, so if lookup/revalidate accessed statahead cache,
1509 * set dentry ldd_sa_generation to parent lli_sa_generation, later if we
1510 * stat this file again, we know we've done statahead before, see
1511 * dentry_may_statahead().
1513 ldd = ll_d2d(*dentryp);
1514 /* ldd can be NULL if llite lookup failed. */
1516 ldd->lld_sa_generation = lli->lli_sa_generation;
1523 * start statahead thread
1525 * \param[in] dir parent directory
1526 * \param[in] dentry dentry that triggers statahead, normally the first
1528 * \retval -EAGAIN on success, because when this function is
1529 * called, it's already in lookup call, so client should
1530 * do it itself instead of waiting for statahead thread
1531 * to do it asynchronously.
1532 * \retval negative number upon error
1534 static int start_statahead_thread(struct inode *dir, struct dentry *dentry)
1536 struct ll_inode_info *lli = ll_i2info(dir);
1537 struct ll_statahead_info *sai = NULL;
1538 struct dentry *parent = dentry->d_parent;
1539 struct ptlrpc_thread *thread;
1540 struct l_wait_info lwi = { 0 };
1541 struct task_struct *task;
1545 /* I am the "lli_opendir_pid" owner, only me can set "lli_sai". */
1546 rc = is_first_dirent(dir, dentry);
1547 if (rc == LS_NOT_FIRST_DE)
1548 /* It is not "ls -{a}l" operation, no need statahead for it. */
1549 GOTO(out, rc = -EFAULT);
1551 sai = ll_sai_alloc(parent);
1553 GOTO(out, rc = -ENOMEM);
1555 sai->sai_ls_all = (rc == LS_FIRST_DOT_DE);
1557 /* if current lli_opendir_key was deauthorized, or dir re-opened by
1558 * another process, don't start statahead, otherwise the newly spawned
1559 * statahead thread won't be notified to quit. */
1560 spin_lock(&lli->lli_sa_lock);
1561 if (unlikely(lli->lli_sai != NULL ||
1562 lli->lli_opendir_key == NULL ||
1563 lli->lli_opendir_pid != current->pid)) {
1564 spin_unlock(&lli->lli_sa_lock);
1565 GOTO(out, rc = -EPERM);
1568 spin_unlock(&lli->lli_sa_lock);
1570 atomic_inc(&ll_i2sbi(parent->d_inode)->ll_sa_running);
1572 CDEBUG(D_READA, "start statahead thread: [pid %d] [parent %.*s]\n",
1573 current_pid(), parent->d_name.len, parent->d_name.name);
1575 task = kthread_run(ll_statahead_thread, parent, "ll_sa_%u",
1576 lli->lli_opendir_pid);
1577 thread = &sai->sai_thread;
1580 CERROR("can't start ll_sa thread, rc: %d\n", rc);
1584 l_wait_event(thread->t_ctl_waitq,
1585 thread_is_running(thread) || thread_is_stopped(thread),
1590 * We don't stat-ahead for the first dirent since we are already in
1596 /* once we start statahead thread failed, disable statahead so that
1597 * subsequent stat won't waste time to try it. */
1598 spin_lock(&lli->lli_sa_lock);
1599 lli->lli_sa_enabled = 0;
1600 lli->lli_sai = NULL;
1601 spin_unlock(&lli->lli_sa_lock);
1610 * statahead entry function, this is called when client getattr on a file, it
1611 * will start statahead thread if this is the first dir entry, else revalidate
1612 * dentry from statahead cache.
1614 * \param[in] dir parent directory
1615 * \param[out] dentryp dentry to getattr
1616 * \param[in] unplug unplug statahead window only (normally for negative
1618 * \retval 1 on success
1619 * \retval 0 revalidation from statahead cache failed, caller needs
1620 * to getattr from server directly
1621 * \retval negative number on error, caller often ignores this and
1622 * then getattr from server
1624 int ll_statahead(struct inode *dir, struct dentry **dentryp, bool unplug)
1626 struct ll_statahead_info *sai;
1628 sai = ll_sai_get(dir);
1632 rc = revalidate_statahead_dentry(dir, sai, dentryp, unplug);
1633 CDEBUG(D_READA, "revalidate statahead %.*s: %d.\n",
1634 (*dentryp)->d_name.len, (*dentryp)->d_name.name, rc);
1638 return start_statahead_thread(dir, *dentryp);