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;
375 spin_lock(&lli->lli_lock);
376 LASSERT(!sa_received_empty(sai));
377 entry = list_entry(sai->sai_entries_received.next, struct ll_sai_entry,
379 list_del_init(&entry->se_list);
380 spin_unlock(&lli->lli_lock);
382 if (unlikely(entry->se_index < sai->sai_index_next)) {
383 ll_sai_entry_cleanup(entry);
388 minfo = entry->se_minfo;
389 dentry = minfo->mi_dentry;
392 if (entry->se_stat != SA_ENTRY_STATED)
393 GOTO(out, rc = entry->se_stat);
395 if (dentry->d_inode == NULL) {
399 struct dentry *save = dentry;
400 struct it_cb_data icbd = {
401 .icbd_parent = dentry->d_parent->d_inode,
402 .icbd_childp = &dentry
405 LASSERT(fid_is_zero(&minfo->mi_data.op_fid2));
407 rc = ll_lookup_it_finish(req, it, &icbd);
410 * Here dentry->d_inode might be NULL,
411 * because the entry may have been removed before
412 * we start doing stat ahead.
414 ll_lookup_finish_locks(it, dentry);
416 if (dentry != save) {
417 minfo->mi_dentry = dentry;
424 struct mdt_body *body;
426 body = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF,
428 if (!lu_fid_eq(&minfo->mi_data.op_fid2, &body->fid1)) {
429 ll_unhash_aliases(dentry->d_inode);
430 GOTO(out, rc = -EAGAIN);
433 rc = ll_revalidate_it_finish(req, it, dentry);
435 ll_unhash_aliases(dentry->d_inode);
439 spin_lock(&dcache_lock);
442 #ifdef DCACHE_LUSTRE_INVALID
443 dentry->d_flags &= ~DCACHE_LUSTRE_INVALID;
445 unlock_dentry(dentry);
446 d_rehash_cond(dentry, 0);
447 spin_unlock(&dcache_lock);
449 ll_lookup_finish_locks(it, dentry);
454 if (likely(ll_sai_entry_to_stated(sai, entry))) {
455 entry->se_minfo = NULL;
456 entry->se_req = NULL;
457 cfs_waitq_signal(&sai->sai_waitq);
458 ll_intent_release(it);
461 ptlrpc_req_finished(req);
466 static int ll_statahead_interpret(struct ptlrpc_request *req,
467 struct md_enqueue_info *minfo,
470 struct dentry *dentry = minfo->mi_dentry;
471 struct lookup_intent *it = &minfo->mi_it;
472 struct inode *dir = dentry->d_parent->d_inode;
473 struct ll_inode_info *lli = ll_i2info(dir);
474 struct ll_statahead_info *sai;
475 struct ll_sai_entry *entry;
478 CDEBUG(D_READA, "interpret statahead %.*s rc %d\n",
479 dentry->d_name.len, dentry->d_name.name, rc);
481 spin_lock(&lli->lli_lock);
482 if (unlikely(lli->lli_sai == NULL ||
483 lli->lli_sai->sai_generation != minfo->mi_generation)) {
484 spin_unlock(&lli->lli_lock);
485 ll_intent_release(it);
491 if (rc || dir == NULL)
494 entry = ll_sai_entry_set(sai,
495 (unsigned int)(long)minfo->mi_cbdata,
496 rc ? SA_ENTRY_UNSTATED :
497 SA_ENTRY_STATED, req, minfo);
498 LASSERT(entry != NULL);
499 if (likely(sa_is_running(sai))) {
500 ll_sai_entry_to_received(sai, entry);
502 spin_unlock(&lli->lli_lock);
503 cfs_waitq_signal(&sai->sai_thread.t_ctl_waitq);
505 if (!list_empty(&entry->se_list))
506 list_del_init(&entry->se_list);
508 spin_unlock(&lli->lli_lock);
509 ll_sai_entry_cleanup(entry);
515 static void sa_args_fini(struct md_enqueue_info *minfo,
516 struct ldlm_enqueue_info *einfo)
518 LASSERT(minfo && einfo);
519 capa_put(minfo->mi_data.op_capa1);
520 capa_put(minfo->mi_data.op_capa2);
526 * There is race condition between "capa_put" and "ll_statahead_interpret" for
527 * accessing "op_data.op_capa[1,2]" as following:
528 * "capa_put" releases "op_data.op_capa[1,2]"'s reference count after calling
529 * "md_intent_getattr_async". But "ll_statahead_interpret" maybe run first, and
530 * fill "op_data.op_capa[1,2]" as POISON, then cause "capa_put" access invalid
531 * "ocapa". So here reserve "op_data.op_capa[1,2]" in "pcapa" before calling
532 * "md_intent_getattr_async".
534 static int sa_args_init(struct inode *dir, struct dentry *dentry,
535 struct md_enqueue_info **pmi,
536 struct ldlm_enqueue_info **pei,
537 struct obd_capa **pcapa)
539 struct ll_inode_info *lli = ll_i2info(dir);
540 struct md_enqueue_info *minfo;
541 struct ldlm_enqueue_info *einfo;
542 struct md_op_data *op_data;
544 OBD_ALLOC_PTR(einfo);
548 OBD_ALLOC_PTR(minfo);
554 op_data = ll_prep_md_op_data(&minfo->mi_data, dir, dentry->d_inode,
555 dentry->d_name.name, dentry->d_name.len,
556 0, LUSTRE_OPC_ANY, NULL);
557 if (IS_ERR(op_data)) {
560 return PTR_ERR(op_data);
563 minfo->mi_it.it_op = IT_GETATTR;
564 minfo->mi_dentry = dentry;
565 minfo->mi_cb = ll_statahead_interpret;
566 minfo->mi_generation = lli->lli_sai->sai_generation;
567 minfo->mi_cbdata = (void *)(long)lli->lli_sai->sai_index;
569 einfo->ei_type = LDLM_IBITS;
570 einfo->ei_mode = it_to_lock_mode(&minfo->mi_it);
571 einfo->ei_cb_bl = ll_md_blocking_ast;
572 einfo->ei_cb_cp = ldlm_completion_ast;
573 einfo->ei_cb_gl = NULL;
574 einfo->ei_cbdata = NULL;
578 pcapa[0] = op_data->op_capa1;
579 pcapa[1] = op_data->op_capa2;
585 * similar to ll_lookup_it().
587 static int do_sa_lookup(struct inode *dir, struct dentry *dentry)
589 struct md_enqueue_info *minfo;
590 struct ldlm_enqueue_info *einfo;
591 struct obd_capa *capas[2];
595 rc = sa_args_init(dir, dentry, &minfo, &einfo, capas);
599 rc = md_intent_getattr_async(ll_i2mdexp(dir), minfo, einfo);
604 sa_args_fini(minfo, einfo);
611 * similar to ll_revalidate_it().
612 * \retval 1 -- dentry valid
613 * \retval 0 -- will send stat-ahead request
614 * \retval others -- prepare stat-ahead request failed
616 static int do_sa_revalidate(struct dentry *dentry)
618 struct inode *inode = dentry->d_inode;
619 struct inode *dir = dentry->d_parent->d_inode;
620 struct lookup_intent it = { .it_op = IT_GETATTR };
621 struct md_enqueue_info *minfo;
622 struct ldlm_enqueue_info *einfo;
623 struct obd_capa *capas[2];
630 if (d_mountpoint(dentry))
633 if (dentry == dentry->d_sb->s_root)
636 rc = md_revalidate_lock(ll_i2mdexp(dir), &it, ll_inode2fid(inode));
638 ll_intent_release(&it);
642 rc = sa_args_init(dir, dentry, &minfo, &einfo, capas);
646 rc = md_intent_getattr_async(ll_i2mdexp(dir), minfo, einfo);
651 sa_args_fini(minfo, einfo);
657 static inline void ll_name2qstr(struct qstr *this, const char *name, int namelen)
659 unsigned long hash = init_name_hash();
664 for (; namelen > 0; namelen--, name++) {
665 c = *(const unsigned char *)name;
666 hash = partial_name_hash(c, hash);
668 this->hash = end_name_hash(hash);
671 static int ll_statahead_one(struct dentry *parent, const char* entry_name,
674 struct inode *dir = parent->d_inode;
675 struct ll_inode_info *lli = ll_i2info(dir);
676 struct ll_statahead_info *sai = lli->lli_sai;
678 struct dentry *dentry;
679 struct ll_sai_entry *se;
683 #ifdef DCACHE_LUSTRE_INVALID
684 if (parent->d_flags & DCACHE_LUSTRE_INVALID) {
686 if (d_unhashed(parent)) {
688 CDEBUG(D_READA, "parent dentry@%p %.*s is "
689 "invalid, skip statahead\n",
690 parent, parent->d_name.len, parent->d_name.name);
694 se = ll_sai_entry_init(sai, sai->sai_index);
698 ll_name2qstr(&name, entry_name, entry_name_len);
699 dentry = d_lookup(parent, &name);
701 dentry = d_alloc(parent, &name);
703 rc = do_sa_lookup(dir, dentry);
707 GOTO(out, rc = -ENOMEM);
710 rc = do_sa_revalidate(dentry);
719 CDEBUG(D_READA, "set sai entry %p index %u stat %d rc %d\n",
720 se, se->se_index, se->se_stat, rc);
722 if (ll_sai_entry_to_stated(sai, se))
723 cfs_waitq_signal(&sai->sai_waitq);
732 struct ll_sa_thread_args {
733 struct dentry *sta_parent;
737 static int ll_statahead_thread(void *arg)
739 struct ll_sa_thread_args *sta = arg;
740 struct dentry *parent = dget(sta->sta_parent);
741 struct inode *dir = parent->d_inode;
742 struct ll_inode_info *lli = ll_i2info(dir);
743 struct ll_sb_info *sbi = ll_i2sbi(dir);
744 struct ll_statahead_info *sai = ll_sai_get(lli->lli_sai);
745 struct ptlrpc_thread *thread = &sai->sai_thread;
750 struct ll_dir_chain chain;
755 snprintf(pname, 15, "ll_sa_%u", sta->sta_pid);
756 cfs_daemonize(pname);
760 spin_lock(&lli->lli_lock);
761 thread->t_flags = SVC_RUNNING;
762 spin_unlock(&lli->lli_lock);
763 cfs_waitq_signal(&thread->t_ctl_waitq);
764 CDEBUG(D_READA, "start doing statahead for %s\n", parent->d_name.name);
766 ll_dir_chain_init(&chain);
767 page = ll_get_dir_page(dir, pos, 0, &chain);
770 struct l_wait_info lwi = { 0 };
771 struct lu_dirpage *dp;
772 struct lu_dirent *ent;
776 CERROR("error reading dir "DFID" at "LPU64"/%u: rc %d\n",
777 PFID(ll_inode2fid(dir)), pos,
782 dp = page_address(page);
783 for (ent = lu_dirent_start(dp); ent != NULL;
784 ent = lu_dirent_next(ent)) {
785 char *name = ent->lde_name;
786 int namelen = le16_to_cpu(ent->lde_namelen);
794 if (name[0] == '.') {
800 } else if (name[1] == '.' && namelen == 2) {
805 } else if (!sai->sai_ls_all) {
809 sai->sai_skip_hidden++;
815 * don't stat-ahead first entry.
817 if (unlikely(!first)) {
823 l_wait_event(thread->t_ctl_waitq,
824 !sa_is_running(sai) || sa_not_full(sai) ||
825 !sa_received_empty(sai),
828 while (!sa_received_empty(sai) && sa_is_running(sai))
829 do_statahead_interpret(sai);
831 if (unlikely(!sa_is_running(sai))) {
836 if (!sa_not_full(sai))
838 * do not skip the current de.
842 rc = ll_statahead_one(parent, name, namelen);
848 pos = le64_to_cpu(dp->ldp_hash_end);
850 if (pos == DIR_END_OFF) {
852 * End of directory reached.
855 l_wait_event(thread->t_ctl_waitq,
856 !sa_is_running(sai) ||
857 !sa_received_empty(sai) ||
858 sai->sai_sent == sai->sai_replied,
860 if (!sa_received_empty(sai) &&
862 do_statahead_interpret(sai);
868 * chain is exhausted.
869 * Normal case: continue to the next page.
871 page = ll_get_dir_page(dir, pos, 1, &chain);
874 * go into overflow page.
881 ll_dir_chain_fini(&chain);
882 spin_lock(&lli->lli_lock);
883 thread->t_flags = SVC_STOPPED;
884 spin_unlock(&lli->lli_lock);
885 cfs_waitq_signal(&sai->sai_waitq);
886 cfs_waitq_signal(&thread->t_ctl_waitq);
889 CDEBUG(D_READA, "statahead thread stopped, pid %d\n",
895 * called in ll_file_release().
897 void ll_stop_statahead(struct inode *inode, void *key)
899 struct ll_inode_info *lli = ll_i2info(inode);
900 struct ptlrpc_thread *thread;
902 spin_lock(&lli->lli_lock);
903 if (lli->lli_opendir_pid == 0 ||
904 unlikely(lli->lli_opendir_key != key)) {
905 spin_unlock(&lli->lli_lock);
909 lli->lli_opendir_key = NULL;
910 lli->lli_opendir_pid = 0;
913 struct l_wait_info lwi = { 0 };
915 thread = &lli->lli_sai->sai_thread;
916 if (!sa_is_stopped(lli->lli_sai)) {
917 thread->t_flags = SVC_STOPPING;
918 spin_unlock(&lli->lli_lock);
919 cfs_waitq_signal(&thread->t_ctl_waitq);
921 CDEBUG(D_READA, "stopping statahead thread, pid %d\n",
923 l_wait_event(thread->t_ctl_waitq,
924 sa_is_stopped(lli->lli_sai),
927 spin_unlock(&lli->lli_lock);
931 * Put the ref which was held when first statahead_enter.
932 * It maybe not the last ref for some statahead requests
935 ll_sai_put(lli->lli_sai);
938 spin_unlock(&lli->lli_lock);
943 * not first dirent, or is "."
945 LS_NONE_FIRST_DE = 0,
947 * the first non-hidden dirent
951 * the first hidden dirent, that is ".xxx
956 static int is_first_dirent(struct inode *dir, struct dentry *dentry)
958 struct ll_dir_chain chain;
959 struct qstr *target = &dentry->d_name;
963 int rc = LS_NONE_FIRST_DE;
966 ll_dir_chain_init(&chain);
967 page = ll_get_dir_page(dir, pos, 0, &chain);
970 struct lu_dirpage *dp;
971 struct lu_dirent *ent;
975 CERROR("error reading dir "DFID" at "LPU64": rc %d\n",
976 PFID(ll_inode2fid(dir)), pos, rc);
980 dp = page_address(page);
981 for (ent = lu_dirent_start(dp); ent != NULL;
982 ent = lu_dirent_next(ent)) {
983 char *name = ent->lde_name;
984 int namelen = le16_to_cpu(ent->lde_namelen);
992 if (name[0] == '.') {
998 else if (name[1] == '.' && namelen == 2)
1009 if (dot_de && target->name[0] != '.') {
1010 CDEBUG(D_READA, "%.*s skip hidden file %.*s\n",
1011 target->len, target->name,
1016 if (target->len == namelen &&
1017 !strncmp(target->name, name, target->len))
1018 rc = LS_FIRST_DE + dot_de;
1020 rc = LS_NONE_FIRST_DE;
1024 pos = le64_to_cpu(dp->ldp_hash_end);
1026 if (pos == DIR_END_OFF) {
1028 * End of directory reached.
1033 * chain is exhausted
1034 * Normal case: continue to the next page.
1036 page = ll_get_dir_page(dir, pos, 1, &chain);
1039 * go into overflow page.
1046 ll_dir_chain_fini(&chain);
1051 * Start statahead thread if this is the first dir entry.
1052 * Otherwise if a thread is started already, wait it until it is ahead of me.
1053 * \retval 0 -- stat ahead thread process such dentry, for lookup, it miss
1054 * \retval 1 -- stat ahead thread process such dentry, for lookup, it hit
1055 * \retval -EEXIST -- stat ahead thread started, and this is the first dentry
1056 * \retval -EBADFD -- statahead thread exit and not dentry available
1057 * \retval others -- error
1059 int do_statahead_enter(struct inode *dir, struct dentry **dentryp, int lookup)
1061 struct ll_sb_info *sbi = ll_i2sbi(dir);
1062 struct ll_inode_info *lli = ll_i2info(dir);
1063 struct ll_statahead_info *sai = lli->lli_sai;
1064 struct ll_sa_thread_args sta;
1065 struct l_wait_info lwi = { 0 };
1069 LASSERT(lli->lli_opendir_pid == cfs_curproc_pid());
1072 if (unlikely(sa_is_stopped(sai) &&
1073 list_empty(&sai->sai_entries_stated)))
1077 * skip the first dentry.
1079 if (unlikely((*dentryp)->d_name.len ==
1080 sai->sai_first->d_name.len &&
1081 !strncmp((*dentryp)->d_name.name,
1082 sai->sai_first->d_name.name,
1083 sai->sai_first->d_name.len)))
1086 if ((*dentryp)->d_name.name[0] == '.') {
1087 if (likely(sai->sai_ls_all ||
1088 sai->sai_miss_hidden >= sai->sai_skip_hidden)) {
1090 * Hidden dentry is the first one, or statahead
1091 * thread does not skip so many hidden dentries
1092 * before "sai_ls_all" enabled as below.
1095 if (!sai->sai_ls_all)
1097 * It maybe because hidden dentry is not
1098 * the first one, "sai_ls_all" was not
1099 * set, then "ls -al" missed. Enable
1100 * "sai_ls_all" for such case.
1102 sai->sai_ls_all = 1;
1105 * Such "getattr" has been skipped before
1106 * "sai_ls_all" enabled as above.
1108 sai->sai_miss_hidden++;
1113 if (ll_sai_entry_stated(sai)) {
1114 sbi->ll_sa_cached++;
1116 sbi->ll_sa_blocked++;
1118 * thread started already, avoid double-stat.
1120 l_wait_event(sai->sai_waitq,
1121 ll_sai_entry_stated(sai) || sa_is_stopped(sai),
1126 struct dentry *result;
1128 result = d_lookup((*dentryp)->d_parent,
1129 &(*dentryp)->d_name);
1131 LASSERT(result != *dentryp);
1132 /* BUG 16303: do not drop reference count for
1133 * "*dentryp", VFS will do that by itself. */
1139 * do nothing for revalidate.
1145 * I am the "lli_opendir_pid" owner, only me can set "lli_sai".
1147 LASSERT(lli->lli_sai == NULL);
1149 rc = is_first_dirent(dir, *dentryp);
1150 if (rc == LS_NONE_FIRST_DE) {
1152 * It is not "ls -{a}l" operation, no need statahead for it.
1154 spin_lock(&lli->lli_lock);
1155 lli->lli_opendir_key = NULL;
1156 lli->lli_opendir_pid = 0;
1157 spin_unlock(&lli->lli_lock);
1161 sai = ll_sai_alloc();
1165 sai->sai_inode = igrab(dir);
1166 sai->sai_first = dget(*dentryp);
1167 sai->sai_ls_all = (rc == LS_FIRST_DOT_DE);
1169 sta.sta_parent = (*dentryp)->d_parent;
1170 sta.sta_pid = cfs_curproc_pid();
1173 rc = cfs_kernel_thread(ll_statahead_thread, &sta, 0);
1175 CERROR("can't start ll_sa thread, rc: %d\n", rc);
1176 sai->sai_thread.t_flags = SVC_STOPPED;
1178 LASSERT(lli->lli_sai == NULL);
1182 l_wait_event(sai->sai_thread.t_ctl_waitq,
1183 sa_is_running(sai) || sa_is_stopped(sai),
1187 * We don't stat-ahead for the first dirent since we are already in
1188 * lookup, and -EEXIST also indicates that this is the first dirent.
1194 * update hit/miss count.
1196 int ll_statahead_exit(struct dentry *dentry, int result)
1198 struct dentry *parent = dentry->d_parent;
1199 struct ll_inode_info *lli = ll_i2info(parent->d_inode);
1200 struct ll_sb_info *sbi = ll_i2sbi(parent->d_inode);
1201 struct ll_dentry_data *ldd = ll_d2d(dentry);
1204 if (lli->lli_opendir_pid != cfs_curproc_pid())
1208 struct ll_statahead_info *sai = lli->lli_sai;
1213 sai->sai_consecutive_miss = 0;
1214 sai->sai_max = min(2 * sai->sai_max, sbi->ll_sa_max);
1218 sai->sai_consecutive_miss++;
1219 if (sa_low_hit(sai) && sa_is_running(sai)) {
1221 CDEBUG(D_READA, "statahead for dir %.*s hit "
1222 "ratio too low: hit/miss %u/%u, "
1223 "sent/replied %u/%u. stopping statahead "
1225 parent->d_name.len, parent->d_name.name,
1226 sai->sai_hit, sai->sai_miss,
1227 sai->sai_sent, sai->sai_replied,
1229 spin_lock(&lli->lli_lock);
1230 if (!sa_is_stopped(sai))
1231 sai->sai_thread.t_flags = SVC_STOPPING;
1232 spin_unlock(&lli->lli_lock);
1236 if (!sa_is_stopped(sai))
1237 cfs_waitq_signal(&sai->sai_thread.t_ctl_waitq);
1238 ll_sai_entry_fini(sai);
1240 if (unlikely(ldd == NULL)) {
1242 ldd = ll_d2d(dentry);
1243 if (ldd != NULL && dentry->d_op == NULL) {
1244 lock_dentry(dentry);
1245 dentry->d_op = dentry->d_op ? : &ll_sai_d_ops;
1246 unlock_dentry(dentry);
1250 if (likely(ldd != NULL))
1251 ldd->lld_sa_generation = sai->sai_generation;