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"
53 struct list_head se_list;
54 unsigned int se_index;
56 struct ptlrpc_request *se_req;
57 struct md_enqueue_info *se_minfo;
61 SA_ENTRY_UNSTATED = 0,
65 struct dentry_operations ll_sai_d_ops = {
66 .d_release = ll_release,
69 static unsigned int sai_generation = 0;
70 static spinlock_t sai_generation_lock = SPIN_LOCK_UNLOCKED;
73 * Check whether first entry was stated already or not.
74 * No need to hold lli_lock, for:
75 * (1) it is me that remove entry from the list
76 * (2) the statahead thread only add new entry to the list
78 static int ll_sai_entry_stated(struct ll_statahead_info *sai)
80 struct ll_sai_entry *entry;
83 if (!list_empty(&sai->sai_entries_stated)) {
84 entry = list_entry(sai->sai_entries_stated.next,
85 struct ll_sai_entry, se_list);
86 if (entry->se_index == sai->sai_index_next)
92 static inline int sa_received_empty(struct ll_statahead_info *sai)
94 return list_empty(&sai->sai_entries_received);
97 static inline int sa_not_full(struct ll_statahead_info *sai)
99 return sai->sai_index < sai->sai_hit + sai->sai_miss + sai->sai_max;
102 static inline int sa_is_running(struct ll_statahead_info *sai)
104 return !!(sai->sai_thread.t_flags & SVC_RUNNING);
107 static inline int sa_is_stopping(struct ll_statahead_info *sai)
109 return !!(sai->sai_thread.t_flags & SVC_STOPPING);
112 static inline int sa_is_stopped(struct ll_statahead_info *sai)
114 return !!(sai->sai_thread.t_flags & SVC_STOPPED);
118 * (1) hit ratio less than 80%
120 * (2) consecutive miss more than 8
122 static inline int sa_low_hit(struct ll_statahead_info *sai)
124 return ((sai->sai_hit < 4 * sai->sai_miss && sai->sai_hit > 7) ||
125 (sai->sai_consecutive_miss > 8));
129 * process the deleted entry's member and free the entry.
131 * (2) free md_enqueue_info
132 * (3) drop dentry's ref count
133 * (4) release request's ref count
135 static void ll_sai_entry_cleanup(struct ll_sai_entry *entry)
137 struct ptlrpc_request *req = entry->se_req;
138 struct md_enqueue_info *minfo = entry->se_minfo;
142 struct dentry *dentry = minfo->mi_dentry;
143 struct lookup_intent *it = &minfo->mi_it;
145 entry->se_minfo = NULL;
146 ll_intent_release(it);
151 entry->se_req = NULL;
152 ptlrpc_req_finished(req);
159 static struct ll_statahead_info *ll_sai_alloc(void)
161 struct ll_statahead_info *sai;
167 spin_lock(&sai_generation_lock);
168 sai->sai_generation = ++sai_generation;
169 if (unlikely(sai_generation == 0))
170 sai->sai_generation = ++sai_generation;
171 spin_unlock(&sai_generation_lock);
172 atomic_set(&sai->sai_refcount, 1);
173 sai->sai_max = LL_SA_RPC_MIN;
174 cfs_waitq_init(&sai->sai_waitq);
175 cfs_waitq_init(&sai->sai_thread.t_ctl_waitq);
176 CFS_INIT_LIST_HEAD(&sai->sai_entries_sent);
177 CFS_INIT_LIST_HEAD(&sai->sai_entries_received);
178 CFS_INIT_LIST_HEAD(&sai->sai_entries_stated);
183 struct ll_statahead_info *ll_sai_get(struct ll_statahead_info *sai)
186 atomic_inc(&sai->sai_refcount);
190 static void ll_sai_put(struct ll_statahead_info *sai)
192 struct inode *inode = sai->sai_inode;
193 struct ll_inode_info *lli = ll_i2info(inode);
196 if (atomic_dec_and_lock(&sai->sai_refcount, &lli->lli_lock)) {
197 struct ll_sai_entry *entry, *next;
200 spin_unlock(&lli->lli_lock);
202 LASSERT(sa_is_stopped(sai));
204 if (sai->sai_sent > sai->sai_replied)
205 CDEBUG(D_READA,"statahead for dir "DFID" does not "
206 "finish: [sent:%u] [replied:%u]\n",
208 sai->sai_sent, sai->sai_replied);
210 list_for_each_entry_safe(entry, next, &sai->sai_entries_sent,
212 list_del(&entry->se_list);
213 ll_sai_entry_cleanup(entry);
215 list_for_each_entry_safe(entry, next, &sai->sai_entries_received,
217 list_del(&entry->se_list);
218 ll_sai_entry_cleanup(entry);
220 list_for_each_entry_safe(entry, next, &sai->sai_entries_stated,
222 list_del(&entry->se_list);
223 ll_sai_entry_cleanup(entry);
225 dput(sai->sai_first);
233 * insert it into sai_entries_sent tail when init.
235 static struct ll_sai_entry *
236 ll_sai_entry_init(struct ll_statahead_info *sai, unsigned int index)
238 struct ll_inode_info *lli = ll_i2info(sai->sai_inode);
239 struct ll_sai_entry *entry;
242 OBD_ALLOC_PTR(entry);
244 RETURN(ERR_PTR(-ENOMEM));
246 CDEBUG(D_READA, "alloc sai entry %p index %u\n",
248 entry->se_index = index;
249 entry->se_stat = SA_ENTRY_UNSTATED;
251 spin_lock(&lli->lli_lock);
252 list_add_tail(&entry->se_list, &sai->sai_entries_sent);
253 spin_unlock(&lli->lli_lock);
259 * delete it from sai_entries_stated head when fini, it need not
260 * to process entry's member.
262 static void ll_sai_entry_fini(struct ll_statahead_info *sai)
264 struct ll_inode_info *lli = ll_i2info(sai->sai_inode);
265 struct ll_sai_entry *entry;
268 spin_lock(&lli->lli_lock);
269 sai->sai_index_next++;
270 if (likely(!list_empty(&sai->sai_entries_stated))) {
271 entry = list_entry(sai->sai_entries_stated.next,
272 struct ll_sai_entry, se_list);
273 if (entry->se_index < sai->sai_index_next) {
274 list_del(&entry->se_list);
278 LASSERT(sa_is_stopped(sai));
279 spin_unlock(&lli->lli_lock);
286 * \retval NULL : can not find the entry in sai_entries_sent with the index
287 * \retval entry: find the entry in sai_entries_sent with the index
289 static struct ll_sai_entry *
290 ll_sai_entry_set(struct ll_statahead_info *sai, unsigned int index, int stat,
291 struct ptlrpc_request *req, struct md_enqueue_info *minfo)
293 struct ll_sai_entry *entry;
296 if (!list_empty(&sai->sai_entries_sent)) {
297 list_for_each_entry(entry, &sai->sai_entries_sent,
299 if (entry->se_index == index) {
300 entry->se_stat = stat;
301 entry->se_req = ptlrpc_request_addref(req);
302 entry->se_minfo = minfo;
304 } else if (entry->se_index > index)
313 * Move entry to sai_entries_received and
314 * insert it into sai_entries_received tail.
317 ll_sai_entry_to_received(struct ll_statahead_info *sai, struct ll_sai_entry *entry)
319 if (!list_empty(&entry->se_list))
320 list_del_init(&entry->se_list);
321 list_add_tail(&entry->se_list, &sai->sai_entries_received);
325 * Move entry to sai_entries_stated and
326 * sort with the index.
329 ll_sai_entry_to_stated(struct ll_statahead_info *sai, struct ll_sai_entry *entry)
331 struct ll_inode_info *lli = ll_i2info(sai->sai_inode);
332 struct ll_sai_entry *se;
335 spin_lock(&lli->lli_lock);
336 if (!list_empty(&entry->se_list))
337 list_del_init(&entry->se_list);
339 if (unlikely(entry->se_index < sai->sai_index_next)) {
340 spin_unlock(&lli->lli_lock);
341 ll_sai_entry_cleanup(entry);
345 list_for_each_entry_reverse(se, &sai->sai_entries_stated, se_list) {
346 if (se->se_index < entry->se_index) {
347 list_add(&entry->se_list, &se->se_list);
348 spin_unlock(&lli->lli_lock);
354 * I am the first entry.
356 list_add(&entry->se_list, &sai->sai_entries_stated);
357 spin_unlock(&lli->lli_lock);
362 * finish lookup/revalidate.
364 static int do_statahead_interpret(struct ll_statahead_info *sai)
366 struct ll_inode_info *lli = ll_i2info(sai->sai_inode);
367 struct ll_sai_entry *entry;
368 struct ptlrpc_request *req;
369 struct md_enqueue_info *minfo;
370 struct dentry *dentry;
371 struct lookup_intent *it;
373 struct mdt_body *body;
376 spin_lock(&lli->lli_lock);
377 LASSERT(!sa_received_empty(sai));
378 entry = list_entry(sai->sai_entries_received.next, struct ll_sai_entry,
380 list_del_init(&entry->se_list);
381 spin_unlock(&lli->lli_lock);
383 if (unlikely(entry->se_index < sai->sai_index_next)) {
384 ll_sai_entry_cleanup(entry);
389 minfo = entry->se_minfo;
390 dentry = minfo->mi_dentry;
393 if (entry->se_stat != SA_ENTRY_STATED)
394 GOTO(out, rc = entry->se_stat);
396 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
398 GOTO(out, rc = -EFAULT);
400 if (dentry->d_inode == NULL) {
404 struct dentry *save = dentry;
405 struct it_cb_data icbd = {
406 .icbd_parent = dentry->d_parent->d_inode,
407 .icbd_childp = &dentry
410 LASSERT(fid_is_zero(&minfo->mi_data.op_fid2));
413 * XXX: No fid in reply, this is probaly cross-ref case.
414 * SA can't handle it yet.
416 if (body->valid & OBD_MD_MDS)
417 GOTO(out, rc = -EAGAIN);
419 rc = ll_lookup_it_finish(req, it, &icbd);
422 * Here dentry->d_inode might be NULL,
423 * because the entry may have been removed before
424 * we start doing stat ahead.
426 ll_lookup_finish_locks(it, dentry);
428 if (dentry != save) {
429 minfo->mi_dentry = dentry;
436 if (!lu_fid_eq(&minfo->mi_data.op_fid2, &body->fid1)) {
437 ll_unhash_aliases(dentry->d_inode);
438 GOTO(out, rc = -EAGAIN);
441 rc = ll_revalidate_it_finish(req, it, dentry);
443 ll_unhash_aliases(dentry->d_inode);
447 spin_lock(&dcache_lock);
450 #ifdef DCACHE_LUSTRE_INVALID
451 dentry->d_flags &= ~DCACHE_LUSTRE_INVALID;
453 unlock_dentry(dentry);
454 d_rehash_cond(dentry, 0);
455 spin_unlock(&dcache_lock);
457 ll_lookup_finish_locks(it, dentry);
462 if (likely(ll_sai_entry_to_stated(sai, entry))) {
463 entry->se_minfo = NULL;
464 entry->se_req = NULL;
465 cfs_waitq_signal(&sai->sai_waitq);
466 ll_intent_release(it);
469 ptlrpc_req_finished(req);
474 static int ll_statahead_interpret(struct ptlrpc_request *req,
475 struct md_enqueue_info *minfo,
478 struct dentry *dentry = minfo->mi_dentry;
479 struct lookup_intent *it = &minfo->mi_it;
480 struct inode *dir = dentry->d_parent->d_inode;
481 struct ll_inode_info *lli = ll_i2info(dir);
482 struct ll_statahead_info *sai;
483 struct ll_sai_entry *entry;
486 CDEBUG(D_READA, "interpret statahead %.*s rc %d\n",
487 dentry->d_name.len, dentry->d_name.name, rc);
489 spin_lock(&lli->lli_lock);
490 if (unlikely(lli->lli_sai == NULL ||
491 lli->lli_sai->sai_generation != minfo->mi_generation)) {
492 spin_unlock(&lli->lli_lock);
493 ll_intent_release(it);
499 if (rc || dir == NULL)
502 entry = ll_sai_entry_set(sai,
503 (unsigned int)(long)minfo->mi_cbdata,
504 rc ? SA_ENTRY_UNSTATED :
505 SA_ENTRY_STATED, req, minfo);
506 LASSERT(entry != NULL);
507 if (likely(sa_is_running(sai))) {
508 ll_sai_entry_to_received(sai, entry);
510 spin_unlock(&lli->lli_lock);
511 cfs_waitq_signal(&sai->sai_thread.t_ctl_waitq);
513 if (!list_empty(&entry->se_list))
514 list_del_init(&entry->se_list);
516 spin_unlock(&lli->lli_lock);
517 ll_sai_entry_cleanup(entry);
523 static void sa_args_fini(struct md_enqueue_info *minfo,
524 struct ldlm_enqueue_info *einfo)
526 LASSERT(minfo && einfo);
527 capa_put(minfo->mi_data.op_capa1);
528 capa_put(minfo->mi_data.op_capa2);
534 * There is race condition between "capa_put" and "ll_statahead_interpret" for
535 * accessing "op_data.op_capa[1,2]" as following:
536 * "capa_put" releases "op_data.op_capa[1,2]"'s reference count after calling
537 * "md_intent_getattr_async". But "ll_statahead_interpret" maybe run first, and
538 * fill "op_data.op_capa[1,2]" as POISON, then cause "capa_put" access invalid
539 * "ocapa". So here reserve "op_data.op_capa[1,2]" in "pcapa" before calling
540 * "md_intent_getattr_async".
542 static int sa_args_init(struct inode *dir, struct dentry *dentry,
543 struct md_enqueue_info **pmi,
544 struct ldlm_enqueue_info **pei,
545 struct obd_capa **pcapa)
547 struct ll_inode_info *lli = ll_i2info(dir);
548 struct md_enqueue_info *minfo;
549 struct ldlm_enqueue_info *einfo;
550 struct md_op_data *op_data;
552 OBD_ALLOC_PTR(einfo);
556 OBD_ALLOC_PTR(minfo);
562 op_data = ll_prep_md_op_data(&minfo->mi_data, dir, dentry->d_inode,
563 dentry->d_name.name, dentry->d_name.len,
564 0, LUSTRE_OPC_ANY, NULL);
565 if (IS_ERR(op_data)) {
568 return PTR_ERR(op_data);
571 minfo->mi_it.it_op = IT_GETATTR;
572 minfo->mi_dentry = dentry;
573 minfo->mi_cb = ll_statahead_interpret;
574 minfo->mi_generation = lli->lli_sai->sai_generation;
575 minfo->mi_cbdata = (void *)(long)lli->lli_sai->sai_index;
577 einfo->ei_type = LDLM_IBITS;
578 einfo->ei_mode = it_to_lock_mode(&minfo->mi_it);
579 einfo->ei_cb_bl = ll_md_blocking_ast;
580 einfo->ei_cb_cp = ldlm_completion_ast;
581 einfo->ei_cb_gl = NULL;
582 einfo->ei_cbdata = NULL;
586 pcapa[0] = op_data->op_capa1;
587 pcapa[1] = op_data->op_capa2;
593 * similar to ll_lookup_it().
595 static int do_sa_lookup(struct inode *dir, struct dentry *dentry)
597 struct md_enqueue_info *minfo;
598 struct ldlm_enqueue_info *einfo;
599 struct obd_capa *capas[2];
603 rc = sa_args_init(dir, dentry, &minfo, &einfo, capas);
607 rc = md_intent_getattr_async(ll_i2mdexp(dir), minfo, einfo);
612 sa_args_fini(minfo, einfo);
619 * similar to ll_revalidate_it().
620 * \retval 1 -- dentry valid
621 * \retval 0 -- will send stat-ahead request
622 * \retval others -- prepare stat-ahead request failed
624 static int do_sa_revalidate(struct dentry *dentry)
626 struct inode *inode = dentry->d_inode;
627 struct inode *dir = dentry->d_parent->d_inode;
628 struct lookup_intent it = { .it_op = IT_GETATTR };
629 struct md_enqueue_info *minfo;
630 struct ldlm_enqueue_info *einfo;
631 struct obd_capa *capas[2];
638 if (d_mountpoint(dentry))
641 if (dentry == dentry->d_sb->s_root)
644 rc = md_revalidate_lock(ll_i2mdexp(dir), &it, ll_inode2fid(inode));
646 ll_intent_release(&it);
650 rc = sa_args_init(dir, dentry, &minfo, &einfo, capas);
654 rc = md_intent_getattr_async(ll_i2mdexp(dir), minfo, einfo);
659 sa_args_fini(minfo, einfo);
665 static inline void ll_name2qstr(struct qstr *this, const char *name, int namelen)
667 unsigned long hash = init_name_hash();
672 for (; namelen > 0; namelen--, name++) {
673 c = *(const unsigned char *)name;
674 hash = partial_name_hash(c, hash);
676 this->hash = end_name_hash(hash);
679 static int ll_statahead_one(struct dentry *parent, const char* entry_name,
682 struct inode *dir = parent->d_inode;
683 struct ll_inode_info *lli = ll_i2info(dir);
684 struct ll_statahead_info *sai = lli->lli_sai;
686 struct dentry *dentry;
687 struct ll_sai_entry *se;
691 #ifdef DCACHE_LUSTRE_INVALID
692 if (parent->d_flags & DCACHE_LUSTRE_INVALID) {
694 if (d_unhashed(parent)) {
696 CDEBUG(D_READA, "parent dentry@%p %.*s is "
697 "invalid, skip statahead\n",
698 parent, parent->d_name.len, parent->d_name.name);
702 se = ll_sai_entry_init(sai, sai->sai_index);
706 ll_name2qstr(&name, entry_name, entry_name_len);
707 dentry = d_lookup(parent, &name);
709 dentry = d_alloc(parent, &name);
711 rc = do_sa_lookup(dir, dentry);
715 GOTO(out, rc = -ENOMEM);
718 rc = do_sa_revalidate(dentry);
727 CDEBUG(D_READA, "set sai entry %p index %u stat %d rc %d\n",
728 se, se->se_index, se->se_stat, rc);
730 if (ll_sai_entry_to_stated(sai, se))
731 cfs_waitq_signal(&sai->sai_waitq);
740 struct ll_sa_thread_args {
741 struct dentry *sta_parent;
745 static int ll_statahead_thread(void *arg)
747 struct ll_sa_thread_args *sta = arg;
748 struct dentry *parent = dget(sta->sta_parent);
749 struct inode *dir = parent->d_inode;
750 struct ll_inode_info *lli = ll_i2info(dir);
751 struct ll_sb_info *sbi = ll_i2sbi(dir);
752 struct ll_statahead_info *sai = ll_sai_get(lli->lli_sai);
753 struct ptlrpc_thread *thread = &sai->sai_thread;
758 struct ll_dir_chain chain;
763 snprintf(pname, 15, "ll_sa_%u", sta->sta_pid);
764 cfs_daemonize(pname);
768 spin_lock(&lli->lli_lock);
769 thread->t_flags = SVC_RUNNING;
770 spin_unlock(&lli->lli_lock);
771 cfs_waitq_signal(&thread->t_ctl_waitq);
772 CDEBUG(D_READA, "start doing statahead for %s\n", parent->d_name.name);
774 ll_dir_chain_init(&chain);
775 page = ll_get_dir_page(dir, pos, 0, &chain);
778 struct l_wait_info lwi = { 0 };
779 struct lu_dirpage *dp;
780 struct lu_dirent *ent;
784 CERROR("error reading dir "DFID" at "LPU64"/%u: rc %d\n",
785 PFID(ll_inode2fid(dir)), pos,
790 dp = page_address(page);
791 for (ent = lu_dirent_start(dp); ent != NULL;
792 ent = lu_dirent_next(ent)) {
793 char *name = ent->lde_name;
794 int namelen = le16_to_cpu(ent->lde_namelen);
802 if (name[0] == '.') {
808 } else if (name[1] == '.' && namelen == 2) {
813 } else if (!sai->sai_ls_all) {
817 sai->sai_skip_hidden++;
823 * don't stat-ahead first entry.
825 if (unlikely(!first)) {
831 l_wait_event(thread->t_ctl_waitq,
832 !sa_is_running(sai) || sa_not_full(sai) ||
833 !sa_received_empty(sai),
836 while (!sa_received_empty(sai) && sa_is_running(sai))
837 do_statahead_interpret(sai);
839 if (unlikely(!sa_is_running(sai))) {
844 if (!sa_not_full(sai))
846 * do not skip the current de.
850 rc = ll_statahead_one(parent, name, namelen);
856 pos = le64_to_cpu(dp->ldp_hash_end);
858 if (pos == DIR_END_OFF) {
860 * End of directory reached.
863 l_wait_event(thread->t_ctl_waitq,
864 !sa_is_running(sai) ||
865 !sa_received_empty(sai) ||
866 sai->sai_sent == sai->sai_replied,
868 if (!sa_received_empty(sai) &&
870 do_statahead_interpret(sai);
876 * chain is exhausted.
877 * Normal case: continue to the next page.
879 page = ll_get_dir_page(dir, pos, 1, &chain);
882 * go into overflow page.
889 ll_dir_chain_fini(&chain);
890 spin_lock(&lli->lli_lock);
891 thread->t_flags = SVC_STOPPED;
892 spin_unlock(&lli->lli_lock);
893 cfs_waitq_signal(&sai->sai_waitq);
894 cfs_waitq_signal(&thread->t_ctl_waitq);
897 CDEBUG(D_READA, "statahead thread stopped, pid %d\n",
903 * called in ll_file_release().
905 void ll_stop_statahead(struct inode *inode, void *key)
907 struct ll_inode_info *lli = ll_i2info(inode);
908 struct ptlrpc_thread *thread;
910 spin_lock(&lli->lli_lock);
911 if (lli->lli_opendir_pid == 0 ||
912 unlikely(lli->lli_opendir_key != key)) {
913 spin_unlock(&lli->lli_lock);
917 lli->lli_opendir_key = NULL;
918 lli->lli_opendir_pid = 0;
921 struct l_wait_info lwi = { 0 };
923 thread = &lli->lli_sai->sai_thread;
924 if (!sa_is_stopped(lli->lli_sai)) {
925 thread->t_flags = SVC_STOPPING;
926 spin_unlock(&lli->lli_lock);
927 cfs_waitq_signal(&thread->t_ctl_waitq);
929 CDEBUG(D_READA, "stopping statahead thread, pid %d\n",
931 l_wait_event(thread->t_ctl_waitq,
932 sa_is_stopped(lli->lli_sai),
935 spin_unlock(&lli->lli_lock);
939 * Put the ref which was held when first statahead_enter.
940 * It maybe not the last ref for some statahead requests
943 ll_sai_put(lli->lli_sai);
946 spin_unlock(&lli->lli_lock);
951 * not first dirent, or is "."
953 LS_NONE_FIRST_DE = 0,
955 * the first non-hidden dirent
959 * the first hidden dirent, that is ".xxx
964 static int is_first_dirent(struct inode *dir, struct dentry *dentry)
966 struct ll_dir_chain chain;
967 struct qstr *target = &dentry->d_name;
971 int rc = LS_NONE_FIRST_DE;
974 ll_dir_chain_init(&chain);
975 page = ll_get_dir_page(dir, pos, 0, &chain);
978 struct lu_dirpage *dp;
979 struct lu_dirent *ent;
983 CERROR("error reading dir "DFID" at "LPU64": rc %d\n",
984 PFID(ll_inode2fid(dir)), pos, rc);
988 dp = page_address(page);
989 for (ent = lu_dirent_start(dp); ent != NULL;
990 ent = lu_dirent_next(ent)) {
991 char *name = ent->lde_name;
992 int namelen = le16_to_cpu(ent->lde_namelen);
1000 if (name[0] == '.') {
1006 else if (name[1] == '.' && namelen == 2)
1017 if (dot_de && target->name[0] != '.') {
1018 CDEBUG(D_READA, "%.*s skip hidden file %.*s\n",
1019 target->len, target->name,
1024 if (target->len == namelen &&
1025 !strncmp(target->name, name, target->len))
1026 rc = LS_FIRST_DE + dot_de;
1028 rc = LS_NONE_FIRST_DE;
1032 pos = le64_to_cpu(dp->ldp_hash_end);
1034 if (pos == DIR_END_OFF) {
1036 * End of directory reached.
1041 * chain is exhausted
1042 * Normal case: continue to the next page.
1044 page = ll_get_dir_page(dir, pos, 1, &chain);
1047 * go into overflow page.
1054 ll_dir_chain_fini(&chain);
1059 * Start statahead thread if this is the first dir entry.
1060 * Otherwise if a thread is started already, wait it until it is ahead of me.
1061 * \retval 0 -- stat ahead thread process such dentry, for lookup, it miss
1062 * \retval 1 -- stat ahead thread process such dentry, for lookup, it hit
1063 * \retval -EEXIST -- stat ahead thread started, and this is the first dentry
1064 * \retval -EBADFD -- statahead thread exit and not dentry available
1065 * \retval others -- error
1067 int do_statahead_enter(struct inode *dir, struct dentry **dentryp, int lookup)
1069 struct ll_sb_info *sbi = ll_i2sbi(dir);
1070 struct ll_inode_info *lli = ll_i2info(dir);
1071 struct ll_statahead_info *sai = lli->lli_sai;
1072 struct ll_sa_thread_args sta;
1073 struct l_wait_info lwi = { 0 };
1077 LASSERT(lli->lli_opendir_pid == cfs_curproc_pid());
1080 if (unlikely(sa_is_stopped(sai) &&
1081 list_empty(&sai->sai_entries_stated)))
1085 * skip the first dentry.
1087 if (unlikely((*dentryp)->d_name.len ==
1088 sai->sai_first->d_name.len &&
1089 !strncmp((*dentryp)->d_name.name,
1090 sai->sai_first->d_name.name,
1091 sai->sai_first->d_name.len)))
1094 if ((*dentryp)->d_name.name[0] == '.') {
1095 if (likely(sai->sai_ls_all ||
1096 sai->sai_miss_hidden >= sai->sai_skip_hidden)) {
1098 * Hidden dentry is the first one, or statahead
1099 * thread does not skip so many hidden dentries
1100 * before "sai_ls_all" enabled as below.
1103 if (!sai->sai_ls_all)
1105 * It maybe because hidden dentry is not
1106 * the first one, "sai_ls_all" was not
1107 * set, then "ls -al" missed. Enable
1108 * "sai_ls_all" for such case.
1110 sai->sai_ls_all = 1;
1113 * Such "getattr" has been skipped before
1114 * "sai_ls_all" enabled as above.
1116 sai->sai_miss_hidden++;
1121 if (ll_sai_entry_stated(sai)) {
1122 sbi->ll_sa_cached++;
1124 sbi->ll_sa_blocked++;
1126 * thread started already, avoid double-stat.
1128 l_wait_event(sai->sai_waitq,
1129 ll_sai_entry_stated(sai) || sa_is_stopped(sai),
1134 struct dentry *result;
1136 result = d_lookup((*dentryp)->d_parent,
1137 &(*dentryp)->d_name);
1139 LASSERT(result != *dentryp);
1140 /* BUG 16303: do not drop reference count for
1141 * "*dentryp", VFS will do that by itself. */
1147 * do nothing for revalidate.
1153 * I am the "lli_opendir_pid" owner, only me can set "lli_sai".
1155 LASSERT(lli->lli_sai == NULL);
1157 rc = is_first_dirent(dir, *dentryp);
1158 if (rc == LS_NONE_FIRST_DE) {
1160 * It is not "ls -{a}l" operation, no need statahead for it.
1162 spin_lock(&lli->lli_lock);
1163 lli->lli_opendir_key = NULL;
1164 lli->lli_opendir_pid = 0;
1165 spin_unlock(&lli->lli_lock);
1169 sai = ll_sai_alloc();
1173 sai->sai_inode = igrab(dir);
1174 sai->sai_first = dget(*dentryp);
1175 sai->sai_ls_all = (rc == LS_FIRST_DOT_DE);
1177 sta.sta_parent = (*dentryp)->d_parent;
1178 sta.sta_pid = cfs_curproc_pid();
1181 rc = cfs_kernel_thread(ll_statahead_thread, &sta, 0);
1183 CERROR("can't start ll_sa thread, rc: %d\n", rc);
1184 sai->sai_thread.t_flags = SVC_STOPPED;
1186 LASSERT(lli->lli_sai == NULL);
1190 l_wait_event(sai->sai_thread.t_ctl_waitq,
1191 sa_is_running(sai) || sa_is_stopped(sai),
1195 * We don't stat-ahead for the first dirent since we are already in
1196 * lookup, and -EEXIST also indicates that this is the first dirent.
1202 * update hit/miss count.
1204 int ll_statahead_exit(struct dentry *dentry, int result)
1206 struct dentry *parent = dentry->d_parent;
1207 struct ll_inode_info *lli = ll_i2info(parent->d_inode);
1208 struct ll_sb_info *sbi = ll_i2sbi(parent->d_inode);
1209 struct ll_dentry_data *ldd = ll_d2d(dentry);
1212 if (lli->lli_opendir_pid != cfs_curproc_pid())
1216 struct ll_statahead_info *sai = lli->lli_sai;
1221 sai->sai_consecutive_miss = 0;
1222 sai->sai_max = min(2 * sai->sai_max, sbi->ll_sa_max);
1226 sai->sai_consecutive_miss++;
1227 if (sa_low_hit(sai) && sa_is_running(sai)) {
1229 CDEBUG(D_READA, "statahead for dir %.*s hit "
1230 "ratio too low: hit/miss %u/%u, "
1231 "sent/replied %u/%u. stopping statahead "
1233 parent->d_name.len, parent->d_name.name,
1234 sai->sai_hit, sai->sai_miss,
1235 sai->sai_sent, sai->sai_replied,
1237 spin_lock(&lli->lli_lock);
1238 if (!sa_is_stopped(sai))
1239 sai->sai_thread.t_flags = SVC_STOPPING;
1240 spin_unlock(&lli->lli_lock);
1244 if (!sa_is_stopped(sai))
1245 cfs_waitq_signal(&sai->sai_thread.t_ctl_waitq);
1246 ll_sai_entry_fini(sai);
1248 if (unlikely(ldd == NULL)) {
1250 ldd = ll_d2d(dentry);
1251 if (ldd != NULL && dentry->d_op == NULL) {
1252 lock_dentry(dentry);
1253 dentry->d_op = dentry->d_op ? : &ll_sai_d_ops;
1254 unlock_dentry(dentry);
1258 if (likely(ldd != NULL))
1259 ldd->lld_sa_generation = sai->sai_generation;