1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright 2008 Sun Microsystems, Inc. All rights reserved
30 * Use is subject to license terms.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
38 #include <linux/sched.h>
40 #include <linux/smp_lock.h>
41 #include <linux/highmem.h>
42 #include <linux/pagemap.h>
44 #define DEBUG_SUBSYSTEM S_LLITE
46 #include <obd_support.h>
47 #include <lustre_lite.h>
48 #include <lustre_dlm.h>
49 #include <linux/lustre_version.h>
50 #include "llite_internal.h"
54 unsigned int se_index;
56 struct ptlrpc_request *se_req;
57 struct md_enqueue_info *se_minfo;
61 SA_ENTRY_UNSTATED = 0,
65 static unsigned int sai_generation = 0;
66 static cfs_spinlock_t sai_generation_lock = CFS_SPIN_LOCK_UNLOCKED;
69 * Check whether first entry was stated already or not.
70 * No need to hold lli_lock, for:
71 * (1) it is me that remove entry from the list
72 * (2) the statahead thread only add new entry to the list
74 static int ll_sai_entry_stated(struct ll_statahead_info *sai)
76 struct ll_sai_entry *entry;
79 if (!cfs_list_empty(&sai->sai_entries_stated)) {
80 entry = cfs_list_entry(sai->sai_entries_stated.next,
81 struct ll_sai_entry, se_list);
82 if (entry->se_index == sai->sai_index_next)
88 static inline int sa_received_empty(struct ll_statahead_info *sai)
90 return cfs_list_empty(&sai->sai_entries_received);
93 static inline int sa_not_full(struct ll_statahead_info *sai)
95 return (sai->sai_index < sai->sai_hit + sai->sai_miss + sai->sai_max);
98 static inline int sa_is_running(struct ll_statahead_info *sai)
100 return !!(sai->sai_thread.t_flags & SVC_RUNNING);
103 static inline int sa_is_stopping(struct ll_statahead_info *sai)
105 return !!(sai->sai_thread.t_flags & SVC_STOPPING);
108 static inline int sa_is_stopped(struct ll_statahead_info *sai)
110 return !!(sai->sai_thread.t_flags & SVC_STOPPED);
114 * (1) hit ratio less than 80%
116 * (2) consecutive miss more than 8
118 static inline int sa_low_hit(struct ll_statahead_info *sai)
120 return ((sai->sai_hit > 7 && sai->sai_hit < 4 * sai->sai_miss) ||
121 (sai->sai_consecutive_miss > 8));
125 * process the deleted entry's member and free the entry.
127 * (2) free md_enqueue_info
128 * (3) drop dentry's ref count
129 * (4) release request's ref count
131 static void ll_sai_entry_cleanup(struct ll_sai_entry *entry, int free)
133 struct md_enqueue_info *minfo = entry->se_minfo;
134 struct ptlrpc_request *req = entry->se_req;
138 entry->se_minfo = NULL;
139 ll_intent_release(&minfo->mi_it);
140 dput(minfo->mi_dentry);
145 entry->se_req = NULL;
146 ptlrpc_req_finished(req);
149 LASSERT(cfs_list_empty(&entry->se_list));
156 static struct ll_statahead_info *ll_sai_alloc(void)
158 struct ll_statahead_info *sai;
164 cfs_spin_lock(&sai_generation_lock);
165 sai->sai_generation = ++sai_generation;
166 if (unlikely(sai_generation == 0))
167 sai->sai_generation = ++sai_generation;
168 cfs_spin_unlock(&sai_generation_lock);
169 cfs_atomic_set(&sai->sai_refcount, 1);
170 sai->sai_max = LL_SA_RPC_MIN;
171 cfs_waitq_init(&sai->sai_waitq);
172 cfs_waitq_init(&sai->sai_thread.t_ctl_waitq);
173 CFS_INIT_LIST_HEAD(&sai->sai_entries_sent);
174 CFS_INIT_LIST_HEAD(&sai->sai_entries_received);
175 CFS_INIT_LIST_HEAD(&sai->sai_entries_stated);
180 struct ll_statahead_info *ll_sai_get(struct ll_statahead_info *sai)
183 cfs_atomic_inc(&sai->sai_refcount);
187 static void ll_sai_put(struct ll_statahead_info *sai)
189 struct inode *inode = sai->sai_inode;
190 struct ll_inode_info *lli;
193 LASSERT(inode != NULL);
194 lli = ll_i2info(inode);
195 LASSERT(lli->lli_sai == sai);
197 if (cfs_atomic_dec_and_test(&sai->sai_refcount)) {
198 struct ll_sai_entry *entry, *next;
200 cfs_spin_lock(&lli->lli_lock);
201 if (unlikely(cfs_atomic_read(&sai->sai_refcount) > 0)) {
202 /* It is race case, the interpret callback just hold
203 * a reference count */
204 cfs_spin_unlock(&lli->lli_lock);
209 LASSERT(lli->lli_opendir_key == NULL);
211 lli->lli_opendir_pid = 0;
212 cfs_spin_unlock(&lli->lli_lock);
214 LASSERT(sa_is_stopped(sai));
216 if (sai->sai_sent > sai->sai_replied)
217 CDEBUG(D_READA,"statahead for dir "DFID" does not "
218 "finish: [sent:%u] [replied:%u]\n",
220 sai->sai_sent, sai->sai_replied);
222 cfs_list_for_each_entry_safe(entry, next,
223 &sai->sai_entries_sent, se_list) {
224 cfs_list_del_init(&entry->se_list);
225 ll_sai_entry_cleanup(entry, 1);
227 cfs_list_for_each_entry_safe(entry, next,
228 &sai->sai_entries_received,
230 cfs_list_del_init(&entry->se_list);
231 ll_sai_entry_cleanup(entry, 1);
233 cfs_list_for_each_entry_safe(entry, next,
234 &sai->sai_entries_stated,
236 cfs_list_del_init(&entry->se_list);
237 ll_sai_entry_cleanup(entry, 1);
246 * insert it into sai_entries_sent tail when init.
248 static struct ll_sai_entry *
249 ll_sai_entry_init(struct ll_statahead_info *sai, unsigned int index)
251 struct ll_inode_info *lli = ll_i2info(sai->sai_inode);
252 struct ll_sai_entry *entry;
255 OBD_ALLOC_PTR(entry);
257 RETURN(ERR_PTR(-ENOMEM));
259 CDEBUG(D_READA, "alloc sai entry %p index %u\n",
261 entry->se_index = index;
262 entry->se_stat = SA_ENTRY_UNSTATED;
264 cfs_spin_lock(&lli->lli_lock);
265 cfs_list_add_tail(&entry->se_list, &sai->sai_entries_sent);
266 cfs_spin_unlock(&lli->lli_lock);
272 * delete it from sai_entries_stated head when fini, it need not
273 * to process entry's member.
275 static int ll_sai_entry_fini(struct ll_statahead_info *sai)
277 struct ll_inode_info *lli = ll_i2info(sai->sai_inode);
278 struct ll_sai_entry *entry;
282 cfs_spin_lock(&lli->lli_lock);
283 sai->sai_index_next++;
284 if (likely(!cfs_list_empty(&sai->sai_entries_stated))) {
285 entry = cfs_list_entry(sai->sai_entries_stated.next,
286 struct ll_sai_entry, se_list);
287 if (entry->se_index < sai->sai_index_next) {
288 cfs_list_del(&entry->se_list);
293 LASSERT(sa_is_stopped(sai));
295 cfs_spin_unlock(&lli->lli_lock);
302 * \retval NULL : can not find the entry in sai_entries_sent with the index
303 * \retval entry: find the entry in sai_entries_sent with the index
305 static struct ll_sai_entry *
306 ll_sai_entry_set(struct ll_statahead_info *sai, unsigned int index, int stat,
307 struct ptlrpc_request *req, struct md_enqueue_info *minfo)
309 struct ll_sai_entry *entry;
312 if (!cfs_list_empty(&sai->sai_entries_sent)) {
313 cfs_list_for_each_entry(entry, &sai->sai_entries_sent,
315 if (entry->se_index == index) {
316 entry->se_stat = stat;
317 entry->se_req = ptlrpc_request_addref(req);
318 entry->se_minfo = minfo;
320 } else if (entry->se_index > index) {
330 * Move entry to sai_entries_received and
331 * insert it into sai_entries_received tail.
334 ll_sai_entry_to_received(struct ll_statahead_info *sai, struct ll_sai_entry *entry)
336 if (!cfs_list_empty(&entry->se_list))
337 cfs_list_del_init(&entry->se_list);
338 cfs_list_add_tail(&entry->se_list, &sai->sai_entries_received);
342 * Move entry to sai_entries_stated and
343 * sort with the index.
346 ll_sai_entry_to_stated(struct ll_statahead_info *sai, struct ll_sai_entry *entry)
348 struct ll_inode_info *lli = ll_i2info(sai->sai_inode);
349 struct ll_sai_entry *se;
352 ll_sai_entry_cleanup(entry, 0);
354 cfs_spin_lock(&lli->lli_lock);
355 if (!cfs_list_empty(&entry->se_list))
356 cfs_list_del_init(&entry->se_list);
358 if (unlikely(entry->se_index < sai->sai_index_next)) {
359 cfs_spin_unlock(&lli->lli_lock);
364 cfs_list_for_each_entry_reverse(se, &sai->sai_entries_stated, se_list) {
365 if (se->se_index < entry->se_index) {
366 cfs_list_add(&entry->se_list, &se->se_list);
367 cfs_spin_unlock(&lli->lli_lock);
373 * I am the first entry.
375 cfs_list_add(&entry->se_list, &sai->sai_entries_stated);
376 cfs_spin_unlock(&lli->lli_lock);
381 * finish lookup/revalidate.
383 static int do_statahead_interpret(struct ll_statahead_info *sai)
385 struct ll_inode_info *lli = ll_i2info(sai->sai_inode);
386 struct ll_sai_entry *entry;
387 struct ptlrpc_request *req;
388 struct md_enqueue_info *minfo;
389 struct lookup_intent *it;
390 struct dentry *dentry;
392 struct mdt_body *body;
395 cfs_spin_lock(&lli->lli_lock);
396 LASSERT(!sa_received_empty(sai));
397 entry = cfs_list_entry(sai->sai_entries_received.next,
398 struct ll_sai_entry, se_list);
399 cfs_list_del_init(&entry->se_list);
400 cfs_spin_unlock(&lli->lli_lock);
402 if (unlikely(entry->se_index < sai->sai_index_next)) {
403 CWARN("Found stale entry: [index %u] [next %u]\n",
404 entry->se_index, sai->sai_index_next);
405 ll_sai_entry_cleanup(entry, 1);
409 if (entry->se_stat != SA_ENTRY_STATED)
410 GOTO(out, rc = entry->se_stat);
413 minfo = entry->se_minfo;
415 dentry = minfo->mi_dentry;
417 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
419 GOTO(out, rc = -EFAULT);
421 if (dentry->d_inode == NULL) {
425 struct dentry *save = dentry;
426 struct it_cb_data icbd = {
427 .icbd_parent = minfo->mi_dir,
428 .icbd_childp = &dentry
431 LASSERT(fid_is_zero(&minfo->mi_data.op_fid2));
434 * XXX: No fid in reply, this is probaly cross-ref case.
435 * SA can't handle it yet.
437 if (body->valid & OBD_MD_MDS)
438 GOTO(out, rc = -EAGAIN);
440 /* BUG 15962: if statahead insert dentry into dcache (for
441 * lookup),it should hold parent dir's i_mutex to synchronize
442 * with other operations from VFS layer.
443 * E.g.: create/delete/rename/lookup, and so on. */
444 mutex_lock(&minfo->mi_dir->i_mutex);
445 rc = ll_lookup_it_finish(req, it, &icbd);
446 mutex_unlock(&minfo->mi_dir->i_mutex);
449 * Here dentry->d_inode might be NULL,
450 * because the entry may have been removed before
451 * we start doing stat ahead.
453 ll_finish_locks(it, dentry);
455 if (dentry != save) {
456 minfo->mi_dentry = dentry;
463 if (!lu_fid_eq(&minfo->mi_data.op_fid2, &body->fid1)) {
464 ll_unhash_aliases(dentry->d_inode);
465 GOTO(out, rc = -EAGAIN);
468 rc = ll_revalidate_it_finish(req, it, dentry);
470 ll_unhash_aliases(dentry->d_inode);
474 spin_lock(&dcache_lock);
477 dentry->d_flags &= ~DCACHE_LUSTRE_INVALID;
478 unlock_dentry(dentry);
479 d_rehash_cond(dentry, 0);
480 spin_unlock(&dcache_lock);
482 ll_finish_locks(it, dentry);
487 if (likely(ll_sai_entry_to_stated(sai, entry)))
488 cfs_waitq_signal(&sai->sai_waitq);
492 static int ll_statahead_interpret(struct ptlrpc_request *req,
493 struct md_enqueue_info *minfo,
496 struct lookup_intent *it = &minfo->mi_it;
497 struct dentry *dentry = minfo->mi_dentry;
498 struct inode *dir = minfo->mi_dir;
499 struct ll_inode_info *lli = ll_i2info(dir);
500 struct ll_statahead_info *sai;
501 struct ll_sai_entry *entry;
504 CDEBUG(D_READA, "interpret statahead %.*s rc %d\n",
505 dentry->d_name.len, dentry->d_name.name, rc);
507 cfs_spin_lock(&lli->lli_lock);
508 if (unlikely(lli->lli_sai == NULL ||
509 lli->lli_sai->sai_generation != minfo->mi_generation)) {
510 cfs_spin_unlock(&lli->lli_lock);
511 ll_intent_release(it);
517 sai = ll_sai_get(lli->lli_sai);
518 entry = ll_sai_entry_set(sai,
519 (unsigned int)(long)minfo->mi_cbdata,
520 rc < 0 ? rc : SA_ENTRY_STATED, req,
522 LASSERT(entry != NULL);
523 if (likely(sa_is_running(sai))) {
524 ll_sai_entry_to_received(sai, entry);
526 cfs_spin_unlock(&lli->lli_lock);
527 cfs_waitq_signal(&sai->sai_thread.t_ctl_waitq);
529 if (!cfs_list_empty(&entry->se_list))
530 cfs_list_del_init(&entry->se_list);
532 cfs_spin_unlock(&lli->lli_lock);
533 ll_sai_entry_cleanup(entry, 1);
540 static void sa_args_fini(struct md_enqueue_info *minfo,
541 struct ldlm_enqueue_info *einfo)
543 LASSERT(minfo && einfo);
545 capa_put(minfo->mi_data.op_capa1);
546 capa_put(minfo->mi_data.op_capa2);
552 * There is race condition between "capa_put" and "ll_statahead_interpret" for
553 * accessing "op_data.op_capa[1,2]" as following:
554 * "capa_put" releases "op_data.op_capa[1,2]"'s reference count after calling
555 * "md_intent_getattr_async". But "ll_statahead_interpret" maybe run first, and
556 * fill "op_data.op_capa[1,2]" as POISON, then cause "capa_put" access invalid
557 * "ocapa". So here reserve "op_data.op_capa[1,2]" in "pcapa" before calling
558 * "md_intent_getattr_async".
560 static int sa_args_init(struct inode *dir, struct dentry *dentry,
561 struct md_enqueue_info **pmi,
562 struct ldlm_enqueue_info **pei,
563 struct obd_capa **pcapa)
565 struct ll_inode_info *lli = ll_i2info(dir);
566 struct md_enqueue_info *minfo;
567 struct ldlm_enqueue_info *einfo;
568 struct md_op_data *op_data;
570 OBD_ALLOC_PTR(einfo);
574 OBD_ALLOC_PTR(minfo);
580 op_data = ll_prep_md_op_data(&minfo->mi_data, dir, dentry->d_inode,
581 dentry->d_name.name, dentry->d_name.len,
582 0, LUSTRE_OPC_ANY, NULL);
583 if (IS_ERR(op_data)) {
586 return PTR_ERR(op_data);
589 minfo->mi_it.it_op = IT_GETATTR;
590 minfo->mi_dentry = dentry;
591 minfo->mi_dir = igrab(dir);
592 minfo->mi_cb = ll_statahead_interpret;
593 minfo->mi_generation = lli->lli_sai->sai_generation;
594 minfo->mi_cbdata = (void *)(long)lli->lli_sai->sai_index;
596 einfo->ei_type = LDLM_IBITS;
597 einfo->ei_mode = it_to_lock_mode(&minfo->mi_it);
598 einfo->ei_cb_bl = ll_md_blocking_ast;
599 einfo->ei_cb_cp = ldlm_completion_ast;
600 einfo->ei_cb_gl = NULL;
601 einfo->ei_cbdata = NULL;
605 pcapa[0] = op_data->op_capa1;
606 pcapa[1] = op_data->op_capa2;
612 * similar to ll_lookup_it().
614 static int do_sa_lookup(struct inode *dir, struct dentry *dentry)
616 struct md_enqueue_info *minfo;
617 struct ldlm_enqueue_info *einfo;
618 struct obd_capa *capas[2];
622 rc = sa_args_init(dir, dentry, &minfo, &einfo, capas);
626 rc = md_intent_getattr_async(ll_i2mdexp(dir), minfo, einfo);
631 sa_args_fini(minfo, einfo);
638 * similar to ll_revalidate_it().
639 * \retval 1 -- dentry valid
640 * \retval 0 -- will send stat-ahead request
641 * \retval others -- prepare stat-ahead request failed
643 static int do_sa_revalidate(struct inode *dir, struct dentry *dentry)
645 struct inode *inode = dentry->d_inode;
646 struct lookup_intent it = { .it_op = IT_GETATTR };
647 struct md_enqueue_info *minfo;
648 struct ldlm_enqueue_info *einfo;
649 struct obd_capa *capas[2];
653 if (unlikely(inode == NULL))
656 if (d_mountpoint(dentry))
659 if (unlikely(dentry == dentry->d_sb->s_root))
662 rc = md_revalidate_lock(ll_i2mdexp(dir), &it, ll_inode2fid(inode));
664 ll_intent_release(&it);
668 rc = sa_args_init(dir, dentry, &minfo, &einfo, capas);
672 rc = md_intent_getattr_async(ll_i2mdexp(dir), minfo, einfo);
677 sa_args_fini(minfo, einfo);
683 static inline void ll_name2qstr(struct qstr *q, const char *name, int namelen)
687 q->hash = full_name_hash(name, namelen);
690 static int ll_statahead_one(struct dentry *parent, const char* entry_name,
693 struct inode *dir = parent->d_inode;
694 struct ll_inode_info *lli = ll_i2info(dir);
695 struct ll_statahead_info *sai = lli->lli_sai;
697 struct dentry *dentry;
698 struct ll_sai_entry *se;
702 if (parent->d_flags & DCACHE_LUSTRE_INVALID) {
703 CDEBUG(D_READA, "parent dentry@%p %.*s is "
704 "invalid, skip statahead\n",
705 parent, parent->d_name.len, parent->d_name.name);
709 se = ll_sai_entry_init(sai, sai->sai_index);
713 ll_name2qstr(&name, entry_name, entry_name_len);
714 dentry = d_lookup(parent, &name);
716 dentry = d_alloc(parent, &name);
718 rc = do_sa_lookup(dir, dentry);
722 GOTO(out, rc = -ENOMEM);
725 rc = do_sa_revalidate(dir, dentry);
734 CDEBUG(D_READA, "set sai entry %p index %u stat %d rc %d\n",
735 se, se->se_index, se->se_stat, rc);
736 se->se_stat = rc < 0 ? rc : SA_ENTRY_STATED;
737 if (ll_sai_entry_to_stated(sai, se))
738 cfs_waitq_signal(&sai->sai_waitq);
747 static int ll_statahead_thread(void *arg)
749 struct dentry *parent = (struct dentry *)arg;
750 struct inode *dir = parent->d_inode;
751 struct ll_inode_info *lli = ll_i2info(dir);
752 struct ll_sb_info *sbi = ll_i2sbi(dir);
753 struct ll_statahead_info *sai = ll_sai_get(lli->lli_sai);
754 struct ptlrpc_thread *thread = &sai->sai_thread;
759 struct ll_dir_chain chain;
764 snprintf(pname, 15, "ll_sa_%u", lli->lli_opendir_pid);
765 cfs_daemonize(pname);
768 atomic_inc(&sbi->ll_sa_total);
769 cfs_spin_lock(&lli->lli_lock);
770 thread->t_flags = SVC_RUNNING;
771 cfs_spin_unlock(&lli->lli_lock);
772 cfs_waitq_signal(&thread->t_ctl_waitq);
773 CDEBUG(D_READA, "start doing statahead for %s\n", parent->d_name.name);
775 ll_dir_chain_init(&chain);
776 page = ll_get_dir_page(dir, pos, 0, &chain);
779 struct l_wait_info lwi = { 0 };
780 struct lu_dirpage *dp;
781 struct lu_dirent *ent;
785 CDEBUG(D_READA, "error reading dir "DFID" at "LPU64
786 "/%u: [rc %d] [parent %u]\n",
787 PFID(ll_inode2fid(dir)), pos, sai->sai_index,
788 rc, lli->lli_opendir_pid);
792 dp = page_address(page);
793 for (ent = lu_dirent_start(dp); ent != NULL;
794 ent = lu_dirent_next(ent)) {
795 char *name = ent->lde_name;
796 int namelen = le16_to_cpu(ent->lde_namelen);
804 if (name[0] == '.') {
810 } else if (name[1] == '.' && namelen == 2) {
815 } else if (!sai->sai_ls_all) {
819 sai->sai_skip_hidden++;
825 * don't stat-ahead first entry.
827 if (unlikely(!first)) {
833 l_wait_event(thread->t_ctl_waitq,
834 !sa_is_running(sai) || sa_not_full(sai) ||
835 !sa_received_empty(sai),
838 while (!sa_received_empty(sai) && sa_is_running(sai))
839 do_statahead_interpret(sai);
841 if (unlikely(!sa_is_running(sai))) {
846 if (!sa_not_full(sai))
848 * do not skip the current de.
852 rc = ll_statahead_one(parent, name, namelen);
858 pos = le64_to_cpu(dp->ldp_hash_end);
860 if (pos == DIR_END_OFF) {
862 * End of directory reached.
865 l_wait_event(thread->t_ctl_waitq,
866 !sa_is_running(sai) ||
867 !sa_received_empty(sai) ||
868 sai->sai_sent == sai->sai_replied,
870 if (!sa_received_empty(sai) &&
872 do_statahead_interpret(sai);
878 * chain is exhausted.
879 * Normal case: continue to the next page.
881 page = ll_get_dir_page(dir, pos, 1, &chain);
884 * go into overflow page.
891 ll_dir_chain_fini(&chain);
892 cfs_spin_lock(&lli->lli_lock);
893 thread->t_flags = SVC_STOPPED;
894 cfs_spin_unlock(&lli->lli_lock);
895 cfs_waitq_signal(&sai->sai_waitq);
896 cfs_waitq_signal(&thread->t_ctl_waitq);
899 CDEBUG(D_READA, "statahead thread stopped, pid %d\n",
905 * called in ll_file_release().
907 void ll_stop_statahead(struct inode *inode, void *key)
909 struct ll_inode_info *lli = ll_i2info(inode);
911 if (unlikely(key == NULL))
914 cfs_spin_lock(&lli->lli_lock);
915 if (lli->lli_opendir_key != key || lli->lli_opendir_pid == 0) {
916 cfs_spin_unlock(&lli->lli_lock);
920 lli->lli_opendir_key = NULL;
923 struct l_wait_info lwi = { 0 };
924 struct ptlrpc_thread *thread = &lli->lli_sai->sai_thread;
926 if (!sa_is_stopped(lli->lli_sai)) {
927 thread->t_flags = SVC_STOPPING;
928 cfs_spin_unlock(&lli->lli_lock);
929 cfs_waitq_signal(&thread->t_ctl_waitq);
931 CDEBUG(D_READA, "stopping statahead thread, pid %d\n",
933 l_wait_event(thread->t_ctl_waitq,
934 sa_is_stopped(lli->lli_sai),
937 cfs_spin_unlock(&lli->lli_lock);
941 * Put the ref which was held when first statahead_enter.
942 * It maybe not the last ref for some statahead requests
945 ll_sai_put(lli->lli_sai);
947 lli->lli_opendir_pid = 0;
948 cfs_spin_unlock(&lli->lli_lock);
954 * not first dirent, or is "."
956 LS_NONE_FIRST_DE = 0,
958 * the first non-hidden dirent
962 * the first hidden dirent, that is "."
967 static int is_first_dirent(struct inode *dir, struct dentry *dentry)
969 struct ll_dir_chain chain;
970 struct qstr *target = &dentry->d_name;
974 int rc = LS_NONE_FIRST_DE;
977 ll_dir_chain_init(&chain);
978 page = ll_get_dir_page(dir, pos, 0, &chain);
981 struct lu_dirpage *dp;
982 struct lu_dirent *ent;
985 struct ll_inode_info *lli = ll_i2info(dir);
988 CERROR("error reading dir "DFID" at "LPU64": "
989 "[rc %d] [parent %u]\n",
990 PFID(ll_inode2fid(dir)), pos,
991 rc, lli->lli_opendir_pid);
995 dp = page_address(page);
996 for (ent = lu_dirent_start(dp); ent != NULL;
997 ent = lu_dirent_next(ent)) {
998 char *name = ent->lde_name;
999 int namelen = le16_to_cpu(ent->lde_namelen);
1003 * skip dummy record.
1007 if (name[0] == '.') {
1013 else if (name[1] == '.' && namelen == 2)
1024 if (dot_de && target->name[0] != '.') {
1025 CDEBUG(D_READA, "%.*s skip hidden file %.*s\n",
1026 target->len, target->name,
1031 if (target->len != namelen ||
1032 memcmp(target->name, name, namelen) != 0)
1033 rc = LS_NONE_FIRST_DE;
1037 rc = LS_FIRST_DOT_DE;
1042 pos = le64_to_cpu(dp->ldp_hash_end);
1044 if (pos == DIR_END_OFF) {
1046 * End of directory reached.
1051 * chain is exhausted
1052 * Normal case: continue to the next page.
1054 page = ll_get_dir_page(dir, pos, 1, &chain);
1057 * go into overflow page.
1064 ll_dir_chain_fini(&chain);
1069 * Start statahead thread if this is the first dir entry.
1070 * Otherwise if a thread is started already, wait it until it is ahead of me.
1071 * \retval 0 -- stat ahead thread process such dentry, for lookup, it miss
1072 * \retval 1 -- stat ahead thread process such dentry, for lookup, it hit
1073 * \retval -EEXIST -- stat ahead thread started, and this is the first dentry
1074 * \retval -EBADFD -- statahead thread exit and not dentry available
1075 * \retval -EAGAIN -- try to stat by caller
1076 * \retval others -- error
1078 int do_statahead_enter(struct inode *dir, struct dentry **dentryp, int lookup)
1080 struct ll_inode_info *lli;
1081 struct ll_statahead_info *sai;
1082 struct dentry *parent;
1083 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
1087 LASSERT(dir != NULL);
1088 lli = ll_i2info(dir);
1089 LASSERT(lli->lli_opendir_pid == cfs_curproc_pid());
1093 if (unlikely(sa_is_stopped(sai) &&
1094 cfs_list_empty(&sai->sai_entries_stated)))
1097 if ((*dentryp)->d_name.name[0] == '.') {
1098 if (likely(sai->sai_ls_all ||
1099 sai->sai_miss_hidden >= sai->sai_skip_hidden)) {
1101 * Hidden dentry is the first one, or statahead
1102 * thread does not skip so many hidden dentries
1103 * before "sai_ls_all" enabled as below.
1106 if (!sai->sai_ls_all)
1108 * It maybe because hidden dentry is not
1109 * the first one, "sai_ls_all" was not
1110 * set, then "ls -al" missed. Enable
1111 * "sai_ls_all" for such case.
1113 sai->sai_ls_all = 1;
1116 * Such "getattr" has been skipped before
1117 * "sai_ls_all" enabled as above.
1119 sai->sai_miss_hidden++;
1124 if (!ll_sai_entry_stated(sai)) {
1127 * If statahead insert dentry into dcache (for lookup),
1128 * it should hold parent dir's i_mutex to synchronize
1129 * with other operations from VFS layer.
1130 * E.g.: create/delete/rename/lookup, and so on.
1132 * To prevent the dead lock between statahead and its
1133 * parent process, the parent process should release
1134 * such i_mutex before waiting for statahead to fetch
1135 * related dentry attribute from MDS.
1137 * It is no matter for parent process to release such
1138 * i_mutex temporary, if someone else create dentry for
1139 * the same item in such interval, we can find it after
1140 * woke up by statahead. */
1142 LASSERT(mutex_is_locked(&dir->i_mutex));
1143 mutex_unlock(&dir->i_mutex);
1145 rc = l_wait_event(sai->sai_waitq,
1146 ll_sai_entry_stated(sai) ||
1150 mutex_lock(&dir->i_mutex);
1154 struct dentry *result;
1156 result = d_lookup((*dentryp)->d_parent,
1157 &(*dentryp)->d_name);
1159 LASSERT(result != *dentryp);
1160 /* BUG 16303: do not drop reference count for
1161 * "*dentryp", VFS will do that by itself. */
1167 * do nothing for revalidate.
1172 /* I am the "lli_opendir_pid" owner, only me can set "lli_sai". */
1173 rc = is_first_dirent(dir, *dentryp);
1174 if (rc == LS_NONE_FIRST_DE)
1175 /* It is not "ls -{a}l" operation, no need statahead for it. */
1176 GOTO(out, rc = -EAGAIN);
1178 sai = ll_sai_alloc();
1180 GOTO(out, rc = -ENOMEM);
1182 sai->sai_ls_all = (rc == LS_FIRST_DOT_DE);
1183 sai->sai_inode = igrab(dir);
1184 if (unlikely(sai->sai_inode == NULL)) {
1185 CWARN("Do not start stat ahead on dying inode "DFID" .\n",
1186 PFID(&lli->lli_fid));
1188 GOTO(out, rc = -ESTALE);
1191 /* get parent reference count here, and put it in ll_statahead_thread */
1192 parent = dget((*dentryp)->d_parent);
1193 if (unlikely(sai->sai_inode != parent->d_inode)) {
1194 struct ll_inode_info *nlli = ll_i2info(parent->d_inode);
1196 CWARN("Race condition, someone changed %.*s just now: "
1197 "old parent "DFID", new parent "DFID" .\n",
1198 (*dentryp)->d_name.len, (*dentryp)->d_name.name,
1199 PFID(&lli->lli_fid), PFID(&nlli->lli_fid));
1201 iput(sai->sai_inode);
1207 rc = cfs_kernel_thread(ll_statahead_thread, parent, 0);
1209 CERROR("can't start ll_sa thread, rc: %d\n", rc);
1211 lli->lli_opendir_key = NULL;
1212 sai->sai_thread.t_flags = SVC_STOPPED;
1214 LASSERT(lli->lli_sai == NULL);
1218 l_wait_event(sai->sai_thread.t_ctl_waitq,
1219 sa_is_running(sai) || sa_is_stopped(sai),
1223 * We don't stat-ahead for the first dirent since we are already in
1224 * lookup, and -EEXIST also indicates that this is the first dirent.
1229 cfs_spin_lock(&lli->lli_lock);
1230 lli->lli_opendir_key = NULL;
1231 lli->lli_opendir_pid = 0;
1232 cfs_spin_unlock(&lli->lli_lock);
1237 * update hit/miss count.
1239 void ll_statahead_exit(struct inode *dir, struct dentry *dentry, int result)
1241 struct ll_inode_info *lli;
1242 struct ll_statahead_info *sai;
1243 struct ll_sb_info *sbi;
1244 struct ll_dentry_data *ldd = ll_d2d(dentry);
1248 LASSERT(dir != NULL);
1249 lli = ll_i2info(dir);
1250 LASSERT(lli->lli_opendir_pid == cfs_curproc_pid());
1252 LASSERT(sai != NULL);
1253 sbi = ll_i2sbi(dir);
1255 rc = ll_sai_entry_fini(sai);
1256 /* rc == -ENOENT means such dentry was removed just between statahead
1257 * readdir and pre-fetched, count it as hit.
1259 * result == -ENOENT has two meanings:
1260 * 1. such dentry was removed just between statahead pre-fetched and
1261 * main process stat such dentry.
1262 * 2. main process stat non-exist dentry.
1263 * Since we can distinguish such two cases, just count it as miss. */
1264 if (result >= 1 || unlikely(rc == -ENOENT)) {
1266 sai->sai_consecutive_miss = 0;
1267 sai->sai_max = min(2 * sai->sai_max, sbi->ll_sa_max);
1270 sai->sai_consecutive_miss++;
1271 if (sa_low_hit(sai) && sa_is_running(sai)) {
1272 atomic_inc(&sbi->ll_sa_wrong);
1273 CDEBUG(D_READA, "Statahead for dir "DFID" hit ratio "
1274 "too low: hit/miss %u/%u, sent/replied %u/%u, "
1275 "stopping statahead thread: pid %d\n",
1276 PFID(&lli->lli_fid), sai->sai_hit,
1277 sai->sai_miss, sai->sai_sent,
1278 sai->sai_replied, cfs_curproc_pid());
1279 cfs_spin_lock(&lli->lli_lock);
1280 if (!sa_is_stopped(sai))
1281 sai->sai_thread.t_flags = SVC_STOPPING;
1282 cfs_spin_unlock(&lli->lli_lock);
1286 if (!sa_is_stopped(sai))
1287 cfs_waitq_signal(&sai->sai_thread.t_ctl_waitq);
1288 if (likely(ldd != NULL))
1289 ldd->lld_sa_generation = sai->sai_generation;