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 static unsigned int sai_generation = 0;
66 static spinlock_t sai_generation_lock = 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 (!list_empty(&sai->sai_entries_stated)) {
80 entry = 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 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(list_empty(&entry->se_list));
156 static struct ll_statahead_info *ll_sai_alloc(void)
158 struct ll_statahead_info *sai;
164 spin_lock(&sai_generation_lock);
165 sai->sai_generation = ++sai_generation;
166 if (unlikely(sai_generation == 0))
167 sai->sai_generation = ++sai_generation;
168 spin_unlock(&sai_generation_lock);
169 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 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 (atomic_dec_and_test(&sai->sai_refcount)) {
198 struct ll_sai_entry *entry, *next;
200 spin_lock(&lli->lli_lock);
201 if (unlikely(atomic_read(&sai->sai_refcount) > 0)) {
202 /* It is race case, the interpret callback just hold
203 * a reference count */
204 spin_unlock(&lli->lli_lock);
209 LASSERT(lli->lli_opendir_key == NULL);
211 lli->lli_opendir_pid = 0;
212 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 list_for_each_entry_safe(entry, next, &sai->sai_entries_sent,
224 list_del_init(&entry->se_list);
225 ll_sai_entry_cleanup(entry, 1);
227 list_for_each_entry_safe(entry, next, &sai->sai_entries_received,
229 list_del_init(&entry->se_list);
230 ll_sai_entry_cleanup(entry, 1);
232 list_for_each_entry_safe(entry, next, &sai->sai_entries_stated,
234 list_del_init(&entry->se_list);
235 ll_sai_entry_cleanup(entry, 1);
244 * insert it into sai_entries_sent tail when init.
246 static struct ll_sai_entry *
247 ll_sai_entry_init(struct ll_statahead_info *sai, unsigned int index)
249 struct ll_inode_info *lli = ll_i2info(sai->sai_inode);
250 struct ll_sai_entry *entry;
253 OBD_ALLOC_PTR(entry);
255 RETURN(ERR_PTR(-ENOMEM));
257 CDEBUG(D_READA, "alloc sai entry %p index %u\n",
259 entry->se_index = index;
260 entry->se_stat = SA_ENTRY_UNSTATED;
262 spin_lock(&lli->lli_lock);
263 list_add_tail(&entry->se_list, &sai->sai_entries_sent);
264 spin_unlock(&lli->lli_lock);
270 * delete it from sai_entries_stated head when fini, it need not
271 * to process entry's member.
273 static int ll_sai_entry_fini(struct ll_statahead_info *sai)
275 struct ll_inode_info *lli = ll_i2info(sai->sai_inode);
276 struct ll_sai_entry *entry;
280 spin_lock(&lli->lli_lock);
281 sai->sai_index_next++;
282 if (likely(!list_empty(&sai->sai_entries_stated))) {
283 entry = list_entry(sai->sai_entries_stated.next,
284 struct ll_sai_entry, se_list);
285 if (entry->se_index < sai->sai_index_next) {
286 list_del(&entry->se_list);
291 LASSERT(sa_is_stopped(sai));
293 spin_unlock(&lli->lli_lock);
300 * \retval NULL : can not find the entry in sai_entries_sent with the index
301 * \retval entry: find the entry in sai_entries_sent with the index
303 static struct ll_sai_entry *
304 ll_sai_entry_set(struct ll_statahead_info *sai, unsigned int index, int stat,
305 struct ptlrpc_request *req, struct md_enqueue_info *minfo)
307 struct ll_sai_entry *entry;
310 if (!list_empty(&sai->sai_entries_sent)) {
311 list_for_each_entry(entry, &sai->sai_entries_sent, se_list) {
312 if (entry->se_index == index) {
313 entry->se_stat = stat;
314 entry->se_req = ptlrpc_request_addref(req);
315 entry->se_minfo = minfo;
317 } else if (entry->se_index > index) {
327 * Move entry to sai_entries_received and
328 * insert it into sai_entries_received tail.
331 ll_sai_entry_to_received(struct ll_statahead_info *sai, struct ll_sai_entry *entry)
333 if (!list_empty(&entry->se_list))
334 list_del_init(&entry->se_list);
335 list_add_tail(&entry->se_list, &sai->sai_entries_received);
339 * Move entry to sai_entries_stated and
340 * sort with the index.
343 ll_sai_entry_to_stated(struct ll_statahead_info *sai, struct ll_sai_entry *entry)
345 struct ll_inode_info *lli = ll_i2info(sai->sai_inode);
346 struct ll_sai_entry *se;
349 ll_sai_entry_cleanup(entry, 0);
351 spin_lock(&lli->lli_lock);
352 if (!list_empty(&entry->se_list))
353 list_del_init(&entry->se_list);
355 if (unlikely(entry->se_index < sai->sai_index_next)) {
356 spin_unlock(&lli->lli_lock);
361 list_for_each_entry_reverse(se, &sai->sai_entries_stated, se_list) {
362 if (se->se_index < entry->se_index) {
363 list_add(&entry->se_list, &se->se_list);
364 spin_unlock(&lli->lli_lock);
370 * I am the first entry.
372 list_add(&entry->se_list, &sai->sai_entries_stated);
373 spin_unlock(&lli->lli_lock);
378 * finish lookup/revalidate.
380 static int do_statahead_interpret(struct ll_statahead_info *sai)
382 struct ll_inode_info *lli = ll_i2info(sai->sai_inode);
383 struct ll_sai_entry *entry;
384 struct ptlrpc_request *req;
385 struct md_enqueue_info *minfo;
386 struct lookup_intent *it;
387 struct dentry *dentry;
389 struct mdt_body *body;
392 spin_lock(&lli->lli_lock);
393 LASSERT(!sa_received_empty(sai));
394 entry = list_entry(sai->sai_entries_received.next, struct ll_sai_entry,
396 list_del_init(&entry->se_list);
397 spin_unlock(&lli->lli_lock);
399 if (unlikely(entry->se_index < sai->sai_index_next)) {
400 CWARN("Found stale entry: [index %u] [next %u]\n",
401 entry->se_index, sai->sai_index_next);
402 ll_sai_entry_cleanup(entry, 1);
406 if (entry->se_stat != SA_ENTRY_STATED)
407 GOTO(out, rc = entry->se_stat);
410 minfo = entry->se_minfo;
412 dentry = minfo->mi_dentry;
414 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
416 GOTO(out, rc = -EFAULT);
418 if (dentry->d_inode == NULL) {
422 struct dentry *save = dentry;
423 struct it_cb_data icbd = {
424 .icbd_parent = minfo->mi_dir,
425 .icbd_childp = &dentry
428 LASSERT(fid_is_zero(&minfo->mi_data.op_fid2));
431 * XXX: No fid in reply, this is probaly cross-ref case.
432 * SA can't handle it yet.
434 if (body->valid & OBD_MD_MDS)
435 GOTO(out, rc = -EAGAIN);
437 /* BUG 15962: if statahead insert dentry into dcache (for
438 * lookup),it should hold parent dir's i_mutex to synchronize
439 * with other operations from VFS layer.
440 * E.g.: create/delete/rename/lookup, and so on. */
441 mutex_lock(&minfo->mi_dir->i_mutex);
442 rc = ll_lookup_it_finish(req, it, &icbd);
443 mutex_unlock(&minfo->mi_dir->i_mutex);
446 * Here dentry->d_inode might be NULL,
447 * because the entry may have been removed before
448 * we start doing stat ahead.
450 ll_finish_locks(it, dentry);
452 if (dentry != save) {
453 minfo->mi_dentry = dentry;
460 if (!lu_fid_eq(&minfo->mi_data.op_fid2, &body->fid1)) {
461 ll_unhash_aliases(dentry->d_inode);
462 GOTO(out, rc = -EAGAIN);
465 rc = ll_revalidate_it_finish(req, it, dentry);
467 ll_unhash_aliases(dentry->d_inode);
471 spin_lock(&dcache_lock);
474 dentry->d_flags &= ~DCACHE_LUSTRE_INVALID;
475 unlock_dentry(dentry);
476 d_rehash_cond(dentry, 0);
477 spin_unlock(&dcache_lock);
479 ll_finish_locks(it, dentry);
484 if (likely(ll_sai_entry_to_stated(sai, entry)))
485 cfs_waitq_signal(&sai->sai_waitq);
489 static int ll_statahead_interpret(struct ptlrpc_request *req,
490 struct md_enqueue_info *minfo,
493 struct lookup_intent *it = &minfo->mi_it;
494 struct dentry *dentry = minfo->mi_dentry;
495 struct inode *dir = minfo->mi_dir;
496 struct ll_inode_info *lli = ll_i2info(dir);
497 struct ll_statahead_info *sai;
498 struct ll_sai_entry *entry;
501 CDEBUG(D_READA, "interpret statahead %.*s rc %d\n",
502 dentry->d_name.len, dentry->d_name.name, rc);
504 spin_lock(&lli->lli_lock);
505 if (unlikely(lli->lli_sai == NULL ||
506 lli->lli_sai->sai_generation != minfo->mi_generation)) {
507 spin_unlock(&lli->lli_lock);
508 ll_intent_release(it);
514 sai = ll_sai_get(lli->lli_sai);
515 entry = ll_sai_entry_set(sai,
516 (unsigned int)(long)minfo->mi_cbdata,
517 rc < 0 ? rc : SA_ENTRY_STATED, req,
519 LASSERT(entry != NULL);
520 if (likely(sa_is_running(sai))) {
521 ll_sai_entry_to_received(sai, entry);
523 spin_unlock(&lli->lli_lock);
524 cfs_waitq_signal(&sai->sai_thread.t_ctl_waitq);
526 if (!list_empty(&entry->se_list))
527 list_del_init(&entry->se_list);
529 spin_unlock(&lli->lli_lock);
530 ll_sai_entry_cleanup(entry, 1);
537 static void sa_args_fini(struct md_enqueue_info *minfo,
538 struct ldlm_enqueue_info *einfo)
540 LASSERT(minfo && einfo);
542 capa_put(minfo->mi_data.op_capa1);
543 capa_put(minfo->mi_data.op_capa2);
549 * There is race condition between "capa_put" and "ll_statahead_interpret" for
550 * accessing "op_data.op_capa[1,2]" as following:
551 * "capa_put" releases "op_data.op_capa[1,2]"'s reference count after calling
552 * "md_intent_getattr_async". But "ll_statahead_interpret" maybe run first, and
553 * fill "op_data.op_capa[1,2]" as POISON, then cause "capa_put" access invalid
554 * "ocapa". So here reserve "op_data.op_capa[1,2]" in "pcapa" before calling
555 * "md_intent_getattr_async".
557 static int sa_args_init(struct inode *dir, struct dentry *dentry,
558 struct md_enqueue_info **pmi,
559 struct ldlm_enqueue_info **pei,
560 struct obd_capa **pcapa)
562 struct ll_inode_info *lli = ll_i2info(dir);
563 struct md_enqueue_info *minfo;
564 struct ldlm_enqueue_info *einfo;
565 struct md_op_data *op_data;
567 OBD_ALLOC_PTR(einfo);
571 OBD_ALLOC_PTR(minfo);
577 op_data = ll_prep_md_op_data(&minfo->mi_data, dir, dentry->d_inode,
578 dentry->d_name.name, dentry->d_name.len,
579 0, LUSTRE_OPC_ANY, NULL);
580 if (IS_ERR(op_data)) {
583 return PTR_ERR(op_data);
586 minfo->mi_it.it_op = IT_GETATTR;
587 minfo->mi_dentry = dentry;
588 minfo->mi_dir = igrab(dir);
589 minfo->mi_cb = ll_statahead_interpret;
590 minfo->mi_generation = lli->lli_sai->sai_generation;
591 minfo->mi_cbdata = (void *)(long)lli->lli_sai->sai_index;
593 einfo->ei_type = LDLM_IBITS;
594 einfo->ei_mode = it_to_lock_mode(&minfo->mi_it);
595 einfo->ei_cb_bl = ll_md_blocking_ast;
596 einfo->ei_cb_cp = ldlm_completion_ast;
597 einfo->ei_cb_gl = NULL;
598 einfo->ei_cbdata = NULL;
602 pcapa[0] = op_data->op_capa1;
603 pcapa[1] = op_data->op_capa2;
609 * similar to ll_lookup_it().
611 static int do_sa_lookup(struct inode *dir, struct dentry *dentry)
613 struct md_enqueue_info *minfo;
614 struct ldlm_enqueue_info *einfo;
615 struct obd_capa *capas[2];
619 rc = sa_args_init(dir, dentry, &minfo, &einfo, capas);
623 rc = md_intent_getattr_async(ll_i2mdexp(dir), minfo, einfo);
628 sa_args_fini(minfo, einfo);
635 * similar to ll_revalidate_it().
636 * \retval 1 -- dentry valid
637 * \retval 0 -- will send stat-ahead request
638 * \retval others -- prepare stat-ahead request failed
640 static int do_sa_revalidate(struct inode *dir, struct dentry *dentry)
642 struct inode *inode = dentry->d_inode;
643 struct lookup_intent it = { .it_op = IT_GETATTR };
644 struct md_enqueue_info *minfo;
645 struct ldlm_enqueue_info *einfo;
646 struct obd_capa *capas[2];
650 if (unlikely(inode == NULL))
653 if (d_mountpoint(dentry))
656 if (unlikely(dentry == dentry->d_sb->s_root))
659 rc = md_revalidate_lock(ll_i2mdexp(dir), &it, ll_inode2fid(inode));
661 ll_intent_release(&it);
665 rc = sa_args_init(dir, dentry, &minfo, &einfo, capas);
669 rc = md_intent_getattr_async(ll_i2mdexp(dir), minfo, einfo);
674 sa_args_fini(minfo, einfo);
680 static inline void ll_name2qstr(struct qstr *q, const char *name, int namelen)
684 q->hash = full_name_hash(name, namelen);
687 static int ll_statahead_one(struct dentry *parent, const char* entry_name,
690 struct inode *dir = parent->d_inode;
691 struct ll_inode_info *lli = ll_i2info(dir);
692 struct ll_statahead_info *sai = lli->lli_sai;
694 struct dentry *dentry;
695 struct ll_sai_entry *se;
699 if (parent->d_flags & DCACHE_LUSTRE_INVALID) {
700 CDEBUG(D_READA, "parent dentry@%p %.*s is "
701 "invalid, skip statahead\n",
702 parent, parent->d_name.len, parent->d_name.name);
706 se = ll_sai_entry_init(sai, sai->sai_index);
710 ll_name2qstr(&name, entry_name, entry_name_len);
711 dentry = d_lookup(parent, &name);
713 dentry = d_alloc(parent, &name);
715 rc = do_sa_lookup(dir, dentry);
719 GOTO(out, rc = -ENOMEM);
722 rc = do_sa_revalidate(dir, dentry);
731 CDEBUG(D_READA, "set sai entry %p index %u stat %d rc %d\n",
732 se, se->se_index, se->se_stat, rc);
733 se->se_stat = rc < 0 ? rc : SA_ENTRY_STATED;
734 if (ll_sai_entry_to_stated(sai, se))
735 cfs_waitq_signal(&sai->sai_waitq);
744 static int ll_statahead_thread(void *arg)
746 struct dentry *parent = (struct dentry *)arg;
747 struct inode *dir = parent->d_inode;
748 struct ll_inode_info *lli = ll_i2info(dir);
749 struct ll_sb_info *sbi = ll_i2sbi(dir);
750 struct ll_statahead_info *sai = ll_sai_get(lli->lli_sai);
751 struct ptlrpc_thread *thread = &sai->sai_thread;
756 struct ll_dir_chain chain;
761 snprintf(pname, 15, "ll_sa_%u", lli->lli_opendir_pid);
762 cfs_daemonize(pname);
765 atomic_inc(&sbi->ll_sa_total);
766 spin_lock(&lli->lli_lock);
767 thread->t_flags = SVC_RUNNING;
768 spin_unlock(&lli->lli_lock);
769 cfs_waitq_signal(&thread->t_ctl_waitq);
770 CDEBUG(D_READA, "start doing statahead for %s\n", parent->d_name.name);
772 ll_dir_chain_init(&chain);
773 page = ll_get_dir_page(dir, pos, 0, &chain);
776 struct l_wait_info lwi = { 0 };
777 struct lu_dirpage *dp;
778 struct lu_dirent *ent;
782 CDEBUG(D_READA, "error reading dir "DFID" at "LPU64
783 "/%u: [rc %d] [parent %u]\n",
784 PFID(ll_inode2fid(dir)), pos, sai->sai_index,
785 rc, lli->lli_opendir_pid);
789 dp = page_address(page);
790 for (ent = lu_dirent_start(dp); ent != NULL;
791 ent = lu_dirent_next(ent)) {
792 char *name = ent->lde_name;
793 int namelen = le16_to_cpu(ent->lde_namelen);
801 if (name[0] == '.') {
807 } else if (name[1] == '.' && namelen == 2) {
812 } else if (!sai->sai_ls_all) {
816 sai->sai_skip_hidden++;
822 * don't stat-ahead first entry.
824 if (unlikely(!first)) {
830 l_wait_event(thread->t_ctl_waitq,
831 !sa_is_running(sai) || sa_not_full(sai) ||
832 !sa_received_empty(sai),
835 while (!sa_received_empty(sai) && sa_is_running(sai))
836 do_statahead_interpret(sai);
838 if (unlikely(!sa_is_running(sai))) {
843 if (!sa_not_full(sai))
845 * do not skip the current de.
849 rc = ll_statahead_one(parent, name, namelen);
855 pos = le64_to_cpu(dp->ldp_hash_end);
857 if (pos == DIR_END_OFF) {
859 * End of directory reached.
862 l_wait_event(thread->t_ctl_waitq,
863 !sa_is_running(sai) ||
864 !sa_received_empty(sai) ||
865 sai->sai_sent == sai->sai_replied,
867 if (!sa_received_empty(sai) &&
869 do_statahead_interpret(sai);
875 * chain is exhausted.
876 * Normal case: continue to the next page.
878 page = ll_get_dir_page(dir, pos, 1, &chain);
881 * go into overflow page.
888 ll_dir_chain_fini(&chain);
889 spin_lock(&lli->lli_lock);
890 thread->t_flags = SVC_STOPPED;
891 spin_unlock(&lli->lli_lock);
892 cfs_waitq_signal(&sai->sai_waitq);
893 cfs_waitq_signal(&thread->t_ctl_waitq);
896 CDEBUG(D_READA, "statahead thread stopped, pid %d\n",
902 * called in ll_file_release().
904 void ll_stop_statahead(struct inode *inode, void *key)
906 struct ll_inode_info *lli = ll_i2info(inode);
908 if (unlikely(key == NULL))
911 spin_lock(&lli->lli_lock);
912 if (lli->lli_opendir_key != key || lli->lli_opendir_pid == 0) {
913 spin_unlock(&lli->lli_lock);
917 lli->lli_opendir_key = NULL;
920 struct l_wait_info lwi = { 0 };
921 struct ptlrpc_thread *thread = &lli->lli_sai->sai_thread;
923 if (!sa_is_stopped(lli->lli_sai)) {
924 thread->t_flags = SVC_STOPPING;
925 spin_unlock(&lli->lli_lock);
926 cfs_waitq_signal(&thread->t_ctl_waitq);
928 CDEBUG(D_READA, "stopping statahead thread, pid %d\n",
930 l_wait_event(thread->t_ctl_waitq,
931 sa_is_stopped(lli->lli_sai),
934 spin_unlock(&lli->lli_lock);
938 * Put the ref which was held when first statahead_enter.
939 * It maybe not the last ref for some statahead requests
942 ll_sai_put(lli->lli_sai);
944 lli->lli_opendir_pid = 0;
945 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 "."
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;
982 struct ll_inode_info *lli = ll_i2info(dir);
985 CERROR("error reading dir "DFID" at "LPU64": "
986 "[rc %d] [parent %u]\n",
987 PFID(ll_inode2fid(dir)), pos,
988 rc, lli->lli_opendir_pid);
992 dp = page_address(page);
993 for (ent = lu_dirent_start(dp); ent != NULL;
994 ent = lu_dirent_next(ent)) {
995 char *name = ent->lde_name;
996 int namelen = le16_to_cpu(ent->lde_namelen);
1000 * skip dummy record.
1004 if (name[0] == '.') {
1010 else if (name[1] == '.' && namelen == 2)
1021 if (dot_de && target->name[0] != '.') {
1022 CDEBUG(D_READA, "%.*s skip hidden file %.*s\n",
1023 target->len, target->name,
1028 if (target->len != namelen ||
1029 memcmp(target->name, name, namelen) != 0)
1030 rc = LS_NONE_FIRST_DE;
1034 rc = LS_FIRST_DOT_DE;
1039 pos = le64_to_cpu(dp->ldp_hash_end);
1041 if (pos == DIR_END_OFF) {
1043 * End of directory reached.
1048 * chain is exhausted
1049 * Normal case: continue to the next page.
1051 page = ll_get_dir_page(dir, pos, 1, &chain);
1054 * go into overflow page.
1061 ll_dir_chain_fini(&chain);
1066 * Start statahead thread if this is the first dir entry.
1067 * Otherwise if a thread is started already, wait it until it is ahead of me.
1068 * \retval 0 -- stat ahead thread process such dentry, for lookup, it miss
1069 * \retval 1 -- stat ahead thread process such dentry, for lookup, it hit
1070 * \retval -EEXIST -- stat ahead thread started, and this is the first dentry
1071 * \retval -EBADFD -- statahead thread exit and not dentry available
1072 * \retval -EAGAIN -- try to stat by caller
1073 * \retval others -- error
1075 int do_statahead_enter(struct inode *dir, struct dentry **dentryp, int lookup)
1077 struct ll_inode_info *lli;
1078 struct ll_statahead_info *sai;
1079 struct dentry *parent;
1080 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
1084 LASSERT(dir != NULL);
1085 lli = ll_i2info(dir);
1086 LASSERT(lli->lli_opendir_pid == cfs_curproc_pid());
1090 if (unlikely(sa_is_stopped(sai) &&
1091 list_empty(&sai->sai_entries_stated)))
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)) {
1124 * If statahead insert dentry into dcache (for lookup),
1125 * it should hold parent dir's i_mutex to synchronize
1126 * with other operations from VFS layer.
1127 * E.g.: create/delete/rename/lookup, and so on.
1129 * To prevent the dead lock between statahead and its
1130 * parent process, the parent process should release
1131 * such i_mutex before waiting for statahead to fetch
1132 * related dentry attribute from MDS.
1134 * It is no matter for parent process to release such
1135 * i_mutex temporary, if someone else create dentry for
1136 * the same item in such interval, we can find it after
1137 * woke up by statahead. */
1139 LASSERT(mutex_is_locked(&dir->i_mutex));
1140 mutex_unlock(&dir->i_mutex);
1142 rc = l_wait_event(sai->sai_waitq,
1143 ll_sai_entry_stated(sai) ||
1147 mutex_lock(&dir->i_mutex);
1151 struct dentry *result;
1153 result = d_lookup((*dentryp)->d_parent,
1154 &(*dentryp)->d_name);
1156 LASSERT(result != *dentryp);
1157 /* BUG 16303: do not drop reference count for
1158 * "*dentryp", VFS will do that by itself. */
1164 * do nothing for revalidate.
1169 /* I am the "lli_opendir_pid" owner, only me can set "lli_sai". */
1170 rc = is_first_dirent(dir, *dentryp);
1171 if (rc == LS_NONE_FIRST_DE)
1172 /* It is not "ls -{a}l" operation, no need statahead for it. */
1173 GOTO(out, rc = -EAGAIN);
1175 sai = ll_sai_alloc();
1177 GOTO(out, rc = -ENOMEM);
1179 sai->sai_ls_all = (rc == LS_FIRST_DOT_DE);
1180 sai->sai_inode = igrab(dir);
1181 if (unlikely(sai->sai_inode == NULL)) {
1182 CWARN("Do not start stat ahead on dying inode "DFID" .\n",
1183 PFID(&lli->lli_fid));
1185 GOTO(out, rc = -ESTALE);
1188 /* get parent reference count here, and put it in ll_statahead_thread */
1189 parent = dget((*dentryp)->d_parent);
1190 if (unlikely(sai->sai_inode != parent->d_inode)) {
1191 struct ll_inode_info *nlli = ll_i2info(parent->d_inode);
1193 CWARN("Race condition, someone changed %.*s just now: "
1194 "old parent "DFID", new parent "DFID" .\n",
1195 (*dentryp)->d_name.len, (*dentryp)->d_name.name,
1196 PFID(&lli->lli_fid), PFID(&nlli->lli_fid));
1198 iput(sai->sai_inode);
1204 rc = cfs_kernel_thread(ll_statahead_thread, parent, 0);
1206 CERROR("can't start ll_sa thread, rc: %d\n", rc);
1208 lli->lli_opendir_key = NULL;
1209 sai->sai_thread.t_flags = SVC_STOPPED;
1211 LASSERT(lli->lli_sai == NULL);
1215 l_wait_event(sai->sai_thread.t_ctl_waitq,
1216 sa_is_running(sai) || sa_is_stopped(sai),
1220 * We don't stat-ahead for the first dirent since we are already in
1221 * lookup, and -EEXIST also indicates that this is the first dirent.
1226 spin_lock(&lli->lli_lock);
1227 lli->lli_opendir_key = NULL;
1228 lli->lli_opendir_pid = 0;
1229 spin_unlock(&lli->lli_lock);
1234 * update hit/miss count.
1236 void ll_statahead_exit(struct inode *dir, struct dentry *dentry, int result)
1238 struct ll_inode_info *lli;
1239 struct ll_statahead_info *sai;
1240 struct ll_sb_info *sbi;
1241 struct ll_dentry_data *ldd = ll_d2d(dentry);
1245 LASSERT(dir != NULL);
1246 lli = ll_i2info(dir);
1247 LASSERT(lli->lli_opendir_pid == cfs_curproc_pid());
1249 LASSERT(sai != NULL);
1250 sbi = ll_i2sbi(dir);
1252 rc = ll_sai_entry_fini(sai);
1253 /* rc == -ENOENT means such dentry was removed just between statahead
1254 * readdir and pre-fetched, count it as hit.
1256 * result == -ENOENT has two meanings:
1257 * 1. such dentry was removed just between statahead pre-fetched and
1258 * main process stat such dentry.
1259 * 2. main process stat non-exist dentry.
1260 * Since we can distinguish such two cases, just count it as miss. */
1261 if (result >= 1 || unlikely(rc == -ENOENT)) {
1263 sai->sai_consecutive_miss = 0;
1264 sai->sai_max = min(2 * sai->sai_max, sbi->ll_sa_max);
1267 sai->sai_consecutive_miss++;
1268 if (sa_low_hit(sai) && sa_is_running(sai)) {
1269 atomic_inc(&sbi->ll_sa_wrong);
1270 CDEBUG(D_READA, "Statahead for dir "DFID" hit ratio "
1271 "too low: hit/miss %u/%u, sent/replied %u/%u, "
1272 "stopping statahead thread: pid %d\n",
1273 PFID(&lli->lli_fid), sai->sai_hit,
1274 sai->sai_miss, sai->sai_sent,
1275 sai->sai_replied, cfs_curproc_pid());
1276 spin_lock(&lli->lli_lock);
1277 if (!sa_is_stopped(sai))
1278 sai->sai_thread.t_flags = SVC_STOPPING;
1279 spin_unlock(&lli->lli_lock);
1283 if (!sa_is_stopped(sai))
1284 cfs_waitq_signal(&sai->sai_thread.t_ctl_waitq);
1285 if (likely(ldd != NULL))
1286 ldd->lld_sa_generation = sai->sai_generation;