lustre/llite/statahead.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   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.
  11  *
  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).
  17  *
  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
  21  *
  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
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
  30  * Use is subject to license terms.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  */
  36
  37 #include <linux/fs.h>
  38 #include <linux/sched.h>
  39 #include <linux/mm.h>
  40 #include <linux/smp_lock.h>
  41 #include <linux/highmem.h>
  42 #include <linux/pagemap.h>
  43
  44 #define DEBUG_SUBSYSTEM S_LLITE
  45
  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"
  51
  52 struct ll_sai_entry {
  53         cfs_list_t              se_list;
  54         unsigned int            se_index;
  55         int                     se_stat;
  56         struct ptlrpc_request  *se_req;
  57         struct md_enqueue_info *se_minfo;
  58 };
  59
  60 enum {
  61         SA_ENTRY_UNSTATED = 0,
  62         SA_ENTRY_STATED
  63 };
  64
  65 static unsigned int sai_generation = 0;
  66 static cfs_spinlock_t sai_generation_lock = CFS_SPIN_LOCK_UNLOCKED;
  67
  68 /**
  69  * Check whether first entry was stated already or not.
  70  * No need to hold lli_sa_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
  73  */
  74 static int ll_sai_entry_stated(struct ll_statahead_info *sai)
  75 {
  76         struct ll_sai_entry  *entry;
  77         int                   rc = 0;
  78
  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)
  83                         rc = 1;
  84         }
  85         return rc;
  86 }
  87
  88 static inline int sa_received_empty(struct ll_statahead_info *sai)
  89 {
  90         return cfs_list_empty(&sai->sai_entries_received);
  91 }
  92
  93 static inline int sa_not_full(struct ll_statahead_info *sai)
  94 {
  95         return (sai->sai_index < sai->sai_hit + sai->sai_miss + sai->sai_max);
  96 }
  97
  98 static inline int sa_is_running(struct ll_statahead_info *sai)
  99 {
 100         return !!(sai->sai_thread.t_flags & SVC_RUNNING);
 101 }
 102
 103 static inline int sa_is_stopping(struct ll_statahead_info *sai)
 104 {
 105         return !!(sai->sai_thread.t_flags & SVC_STOPPING);
 106 }
 107
 108 static inline int sa_is_stopped(struct ll_statahead_info *sai)
 109 {
 110         return !!(sai->sai_thread.t_flags & SVC_STOPPED);
 111 }
 112
 113 /**
 114  * (1) hit ratio less than 80%
 115  * or
 116  * (2) consecutive miss more than 8
 117  */
 118 static inline int sa_low_hit(struct ll_statahead_info *sai)
 119 {
 120         return ((sai->sai_hit > 7 && sai->sai_hit < 4 * sai->sai_miss) ||
 121                 (sai->sai_consecutive_miss > 8));
 122 }
 123
 124 /**
 125  * process the deleted entry's member and free the entry.
 126  * (1) release intent
 127  * (2) free md_enqueue_info
 128  * (3) drop dentry's ref count
 129  * (4) release request's ref count
 130  */
 131 static void ll_sai_entry_cleanup(struct ll_sai_entry *entry, int free)
 132 {
 133         struct md_enqueue_info *minfo = entry->se_minfo;
 134         struct ptlrpc_request  *req   = entry->se_req;
 135         ENTRY;
 136
 137         if (minfo) {
 138                 entry->se_minfo = NULL;
 139                 ll_intent_release(&minfo->mi_it);
 140                 dput(minfo->mi_dentry);
 141                 iput(minfo->mi_dir);
 142                 OBD_FREE_PTR(minfo);
 143         }
 144         if (req) {
 145                 entry->se_req = NULL;
 146                 ptlrpc_req_finished(req);
 147         }
 148         if (free) {
 149                 LASSERT(cfs_list_empty(&entry->se_list));
 150                 OBD_FREE_PTR(entry);
 151         }
 152
 153         EXIT;
 154 }
 155
 156 static struct ll_statahead_info *ll_sai_alloc(void)
 157 {
 158         struct ll_statahead_info *sai;
 159
 160         OBD_ALLOC_PTR(sai);
 161         if (!sai)
 162                 return NULL;
 163
 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);
 176         return sai;
 177 }
 178
 179 static inline
 180 struct ll_statahead_info *ll_sai_get(struct ll_statahead_info *sai)
 181 {
 182         LASSERT(sai);
 183         cfs_atomic_inc(&sai->sai_refcount);
 184         return sai;
 185 }
 186
 187 static void ll_sai_put(struct ll_statahead_info *sai)
 188 {
 189         struct inode         *inode = sai->sai_inode;
 190         struct ll_inode_info *lli;
 191         ENTRY;
 192
 193         LASSERT(inode != NULL);
 194         lli = ll_i2info(inode);
 195         LASSERT(lli->lli_sai == sai);
 196
 197         if (cfs_atomic_dec_and_test(&sai->sai_refcount)) {
 198                 struct ll_sai_entry *entry, *next;
 199
 200                 cfs_spin_lock(&lli->lli_sa_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_sa_lock);
 205                         EXIT;
 206                         return;
 207                 }
 208
 209                 LASSERT(lli->lli_opendir_key == NULL);
 210                 lli->lli_sai = NULL;
 211                 lli->lli_opendir_pid = 0;
 212                 cfs_spin_unlock(&lli->lli_sa_lock);
 213
 214                 LASSERT(sa_is_stopped(sai));
 215
 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",
 219                               PFID(&lli->lli_fid),
 220                               sai->sai_sent, sai->sai_replied);
 221
 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);
 226                 }
 227                 cfs_list_for_each_entry_safe(entry, next,
 228                                              &sai->sai_entries_received,
 229                                              se_list) {
 230                         cfs_list_del_init(&entry->se_list);
 231                         ll_sai_entry_cleanup(entry, 1);
 232                 }
 233                 cfs_list_for_each_entry_safe(entry, next,
 234                                              &sai->sai_entries_stated,
 235                                              se_list) {
 236                         cfs_list_del_init(&entry->se_list);
 237                         ll_sai_entry_cleanup(entry, 1);
 238                 }
 239                 iput(inode);
 240                 OBD_FREE_PTR(sai);
 241         }
 242         EXIT;
 243 }
 244
 245 /**
 246  * insert it into sai_entries_sent tail when init.
 247  */
 248 static struct ll_sai_entry *
 249 ll_sai_entry_init(struct ll_statahead_info *sai, unsigned int index)
 250 {
 251         struct ll_inode_info *lli = ll_i2info(sai->sai_inode);
 252         struct ll_sai_entry  *entry;
 253         ENTRY;
 254
 255         OBD_ALLOC_PTR(entry);
 256         if (entry == NULL)
 257                 RETURN(ERR_PTR(-ENOMEM));
 258
 259         CDEBUG(D_READA, "alloc sai entry %p index %u\n",
 260                entry, index);
 261         entry->se_index = index;
 262         entry->se_stat = SA_ENTRY_UNSTATED;
 263
 264         cfs_spin_lock(&lli->lli_sa_lock);
 265         cfs_list_add_tail(&entry->se_list, &sai->sai_entries_sent);
 266         cfs_spin_unlock(&lli->lli_sa_lock);
 267
 268         RETURN(entry);
 269 }
 270
 271 /**
 272  * delete it from sai_entries_stated head when fini, it need not
 273  * to process entry's member.
 274  */
 275 static int ll_sai_entry_fini(struct ll_statahead_info *sai)
 276 {
 277         struct ll_inode_info *lli = ll_i2info(sai->sai_inode);
 278         struct ll_sai_entry  *entry;
 279         int rc = 0;
 280         ENTRY;
 281
 282         cfs_spin_lock(&lli->lli_sa_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_init(&entry->se_list);
 289                         rc = entry->se_stat;
 290                         OBD_FREE_PTR(entry);
 291                 }
 292         } else {
 293                 LASSERT(sa_is_stopped(sai));
 294         }
 295         cfs_spin_unlock(&lli->lli_sa_lock);
 296
 297         RETURN(rc);
 298 }
 299
 300 /**
 301  * inside lli_sa_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
 304  */
 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)
 308 {
 309         struct ll_sai_entry *entry;
 310         ENTRY;
 311
 312         if (!cfs_list_empty(&sai->sai_entries_sent)) {
 313                 cfs_list_for_each_entry(entry, &sai->sai_entries_sent,
 314                                         se_list) {
 315                         if (entry->se_index == index) {
 316                                 entry->se_stat = stat;
 317                                 entry->se_req = ptlrpc_request_addref(req);
 318                                 entry->se_minfo = minfo;
 319                                 RETURN(entry);
 320                         } else if (entry->se_index > index) {
 321                                 RETURN(NULL);
 322                         }
 323                 }
 324         }
 325         RETURN(NULL);
 326 }
 327
 328 /**
 329  * inside lli_sa_lock.
 330  * Move entry to sai_entries_received and
 331  * insert it into sai_entries_received tail.
 332  */
 333 static inline void
 334 ll_sai_entry_to_received(struct ll_statahead_info *sai, struct ll_sai_entry *entry)
 335 {
 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);
 339 }
 340
 341 /**
 342  * Move entry to sai_entries_stated and
 343  * sort with the index.
 344  */
 345 static int
 346 ll_sai_entry_to_stated(struct ll_statahead_info *sai, struct ll_sai_entry *entry)
 347 {
 348         struct ll_inode_info *lli = ll_i2info(sai->sai_inode);
 349         struct ll_sai_entry  *se;
 350         ENTRY;
 351
 352         ll_sai_entry_cleanup(entry, 0);
 353
 354         cfs_spin_lock(&lli->lli_sa_lock);
 355         if (!cfs_list_empty(&entry->se_list))
 356                 cfs_list_del_init(&entry->se_list);
 357
 358         /* stale entry */
 359         if (unlikely(entry->se_index < sai->sai_index_next)) {
 360                 cfs_spin_unlock(&lli->lli_sa_lock);
 361                 OBD_FREE_PTR(entry);
 362                 RETURN(0);
 363         }
 364
 365         cfs_list_for_each_entry_reverse(se, &sai->sai_entries_stated, se_list) {
 366                 if (se->se_index < entry->se_index) {
 367                         cfs_list_add(&entry->se_list, &se->se_list);
 368                         cfs_spin_unlock(&lli->lli_sa_lock);
 369                         RETURN(1);
 370                 }
 371         }
 372
 373         /*
 374          * I am the first entry.
 375          */
 376         cfs_list_add(&entry->se_list, &sai->sai_entries_stated);
 377         cfs_spin_unlock(&lli->lli_sa_lock);
 378         RETURN(1);
 379 }
 380
 381 /**
 382  * finish lookup/revalidate.
 383  */
 384 static int do_statahead_interpret(struct ll_statahead_info *sai)
 385 {
 386         struct ll_inode_info   *lli = ll_i2info(sai->sai_inode);
 387         struct ll_sai_entry    *entry;
 388         struct ptlrpc_request  *req;
 389         struct md_enqueue_info *minfo;
 390         struct lookup_intent   *it;
 391         struct dentry          *dentry;
 392         int                     rc = 0;
 393         struct mdt_body        *body;
 394         ENTRY;
 395
 396         cfs_spin_lock(&lli->lli_sa_lock);
 397         LASSERT(!sa_received_empty(sai));
 398         entry = cfs_list_entry(sai->sai_entries_received.next,
 399                                struct ll_sai_entry, se_list);
 400         cfs_list_del_init(&entry->se_list);
 401         cfs_spin_unlock(&lli->lli_sa_lock);
 402
 403         if (unlikely(entry->se_index < sai->sai_index_next)) {
 404                 CWARN("Found stale entry: [index %u] [next %u]\n",
 405                       entry->se_index, sai->sai_index_next);
 406                 ll_sai_entry_cleanup(entry, 1);
 407                 RETURN(0);
 408         }
 409
 410         if (entry->se_stat != SA_ENTRY_STATED)
 411                 GOTO(out, rc = entry->se_stat);
 412
 413         req = entry->se_req;
 414         minfo = entry->se_minfo;
 415         it = &minfo->mi_it;
 416         dentry = minfo->mi_dentry;
 417
 418         body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
 419         if (body == NULL)
 420                 GOTO(out, rc = -EFAULT);
 421
 422         if (dentry->d_inode == NULL) {
 423                 /*
 424                  * lookup.
 425                  */
 426                 struct dentry    *save = dentry;
 427                 struct it_cb_data icbd = {
 428                         .icbd_parent   = minfo->mi_dir,
 429                         .icbd_childp   = &dentry
 430                 };
 431
 432                 LASSERT(fid_is_zero(&minfo->mi_data.op_fid2));
 433
 434                 /* XXX: No fid in reply, this is probaly cross-ref case.
 435                  * SA can't handle it yet. */
 436                 if (body->valid & OBD_MD_MDS)
 437                         GOTO(out, rc = -EAGAIN);
 438
 439                 /* Here dentry->d_inode might be NULL, because the entry may
 440                  * have been removed before we start doing stat ahead. */
 441                 rc = ll_lookup_it_finish(req, it, &icbd);
 442                 if (!rc)
 443                         ll_lookup_finish_locks(it, dentry);
 444
 445                 if (dentry != save) {
 446                         minfo->mi_dentry = dentry;
 447                         dput(save);
 448                 }
 449         } else {
 450                 /*
 451                  * revalidate.
 452                  */
 453                 if (!lu_fid_eq(&minfo->mi_data.op_fid2, &body->fid1)) {
 454                         ll_unhash_aliases(dentry->d_inode);
 455                         GOTO(out, rc = -EAGAIN);
 456                 }
 457
 458                 rc = ll_revalidate_it_finish(req, it, dentry);
 459                 if (rc) {
 460                         ll_unhash_aliases(dentry->d_inode);
 461                         GOTO(out, rc);
 462                 }
 463
 464                 cfs_spin_lock(&ll_lookup_lock);
 465                 spin_lock(&dcache_lock);
 466                 lock_dentry(dentry);
 467                 __d_drop(dentry);
 468                 dentry->d_flags &= ~DCACHE_LUSTRE_INVALID;
 469                 unlock_dentry(dentry);
 470                 d_rehash_cond(dentry, 0);
 471                 spin_unlock(&dcache_lock);
 472                 cfs_spin_unlock(&ll_lookup_lock);
 473
 474                 ll_lookup_finish_locks(it, dentry);
 475         }
 476         EXIT;
 477
 478 out:
 479         /* The "ll_sai_entry_to_stated()" will drop related ldlm ibits lock
 480          * reference count with ll_intent_drop_lock() called in spite of the
 481          * above operations failed or not. Do not worry about calling
 482          * "ll_intent_drop_lock()" more than once. */
 483         if (likely(ll_sai_entry_to_stated(sai, entry)))
 484                 cfs_waitq_signal(&sai->sai_waitq);
 485         return rc;
 486 }
 487
 488 static int ll_statahead_interpret(struct ptlrpc_request *req,
 489                                   struct md_enqueue_info *minfo,
 490                                   int rc)
 491 {
 492         struct lookup_intent     *it = &minfo->mi_it;
 493         struct dentry            *dentry = minfo->mi_dentry;
 494         struct inode             *dir = minfo->mi_dir;
 495         struct ll_inode_info     *lli = ll_i2info(dir);
 496         struct ll_statahead_info *sai;
 497         struct ll_sai_entry      *entry;
 498         ENTRY;
 499
 500         CDEBUG(D_READA, "interpret statahead %.*s rc %d\n",
 501                dentry->d_name.len, dentry->d_name.name, rc);
 502
 503         cfs_spin_lock(&lli->lli_sa_lock);
 504         /* stale entry */
 505         if (unlikely(lli->lli_sai == NULL ||
 506             lli->lli_sai->sai_generation != minfo->mi_generation)) {
 507                 cfs_spin_unlock(&lli->lli_sa_lock);
 508                 ll_intent_release(it);
 509                 dput(dentry);
 510                 iput(dir);
 511                 OBD_FREE_PTR(minfo);
 512                 RETURN(-ESTALE);
 513         } else {
 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,
 518                                          minfo);
 519                 LASSERT(entry != NULL);
 520                 if (likely(sa_is_running(sai))) {
 521                         ll_sai_entry_to_received(sai, entry);
 522                         sai->sai_replied++;
 523                         cfs_spin_unlock(&lli->lli_sa_lock);
 524                         cfs_waitq_signal(&sai->sai_thread.t_ctl_waitq);
 525                 } else {
 526                         if (!cfs_list_empty(&entry->se_list))
 527                                 cfs_list_del_init(&entry->se_list);
 528                         sai->sai_replied++;
 529                         cfs_spin_unlock(&lli->lli_sa_lock);
 530                         ll_sai_entry_cleanup(entry, 1);
 531                 }
 532                 ll_sai_put(sai);
 533                 RETURN(rc);
 534         }
 535 }
 536
 537 static void sa_args_fini(struct md_enqueue_info *minfo,
 538                          struct ldlm_enqueue_info *einfo)
 539 {
 540         LASSERT(minfo && einfo);
 541         iput(minfo->mi_dir);
 542         capa_put(minfo->mi_data.op_capa1);
 543         capa_put(minfo->mi_data.op_capa2);
 544         OBD_FREE_PTR(minfo);
 545         OBD_FREE_PTR(einfo);
 546 }
 547
 548 /**
 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".
 556  */
 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)
 561 {
 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;
 566
 567         OBD_ALLOC_PTR(einfo);
 568         if (einfo == NULL)
 569                 return -ENOMEM;
 570
 571         OBD_ALLOC_PTR(minfo);
 572         if (minfo == NULL) {
 573                 OBD_FREE_PTR(einfo);
 574                 return -ENOMEM;
 575         }
 576
 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)) {
 581                 OBD_FREE_PTR(einfo);
 582                 OBD_FREE_PTR(minfo);
 583                 return PTR_ERR(op_data);
 584         }
 585
 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;
 592
 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;
 599
 600         *pmi = minfo;
 601         *pei = einfo;
 602         pcapa[0] = op_data->op_capa1;
 603         pcapa[1] = op_data->op_capa2;
 604
 605         return 0;
 606 }
 607
 608 /**
 609  * similar to ll_lookup_it().
 610  */
 611 static int do_sa_lookup(struct inode *dir, struct dentry *dentry)
 612 {
 613         struct md_enqueue_info   *minfo;
 614         struct ldlm_enqueue_info *einfo;
 615         struct obd_capa          *capas[2];
 616         int                       rc;
 617         ENTRY;
 618
 619         rc = sa_args_init(dir, dentry, &minfo, &einfo, capas);
 620         if (rc)
 621                 RETURN(rc);
 622
 623         rc = md_intent_getattr_async(ll_i2mdexp(dir), minfo, einfo);
 624         if (!rc) {
 625                 capa_put(capas[0]);
 626                 capa_put(capas[1]);
 627         } else {
 628                 sa_args_fini(minfo, einfo);
 629         }
 630
 631         RETURN(rc);
 632 }
 633
 634 /**
 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
 639  */
 640 static int do_sa_revalidate(struct inode *dir, struct dentry *dentry)
 641 {
 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];
 647         int rc;
 648         ENTRY;
 649
 650         if (unlikely(inode == NULL))
 651                 RETURN(1);
 652
 653         if (d_mountpoint(dentry))
 654                 RETURN(1);
 655
 656         if (unlikely(dentry == dentry->d_sb->s_root))
 657                 RETURN(1);
 658
 659         rc = md_revalidate_lock(ll_i2mdexp(dir), &it, ll_inode2fid(inode));
 660         if (rc == 1) {
 661                 ll_intent_release(&it);
 662                 RETURN(1);
 663         }
 664
 665         rc = sa_args_init(dir, dentry, &minfo, &einfo, capas);
 666         if (rc)
 667                 RETURN(rc);
 668
 669         rc = md_intent_getattr_async(ll_i2mdexp(dir), minfo, einfo);
 670         if (!rc) {
 671                 capa_put(capas[0]);
 672                 capa_put(capas[1]);
 673         } else {
 674                 sa_args_fini(minfo, einfo);
 675         }
 676
 677         RETURN(rc);
 678 }
 679
 680 static inline void ll_name2qstr(struct qstr *q, const char *name, int namelen)
 681 {
 682         q->name = name;
 683         q->len  = namelen;
 684         q->hash = full_name_hash(name, namelen);
 685 }
 686
 687 static int ll_statahead_one(struct dentry *parent, const char* entry_name,
 688                             int entry_name_len)
 689 {
 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;
 693         struct qstr               name;
 694         struct dentry            *dentry;
 695         struct ll_sai_entry      *se;
 696         int                       rc;
 697         ENTRY;
 698
 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);
 703                 RETURN(-EINVAL);
 704         }
 705
 706         se = ll_sai_entry_init(sai, sai->sai_index);
 707         if (IS_ERR(se))
 708                 RETURN(PTR_ERR(se));
 709
 710         ll_name2qstr(&name, entry_name, entry_name_len);
 711         dentry = d_lookup(parent, &name);
 712         if (!dentry) {
 713                 dentry = d_alloc(parent, &name);
 714                 if (dentry) {
 715                         rc = do_sa_lookup(dir, dentry);
 716                         if (rc)
 717                                 dput(dentry);
 718                 } else {
 719                         GOTO(out, rc = -ENOMEM);
 720                 }
 721         } else {
 722                 rc = do_sa_revalidate(dir, dentry);
 723                 if (rc)
 724                         dput(dentry);
 725         }
 726
 727         EXIT;
 728
 729 out:
 730         if (rc) {
 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);
 736         } else {
 737                 sai->sai_sent++;
 738         }
 739
 740         sai->sai_index++;
 741         return rc;
 742 }
 743
 744 static int ll_statahead_thread(void *arg)
 745 {
 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;
 752         struct page              *page;
 753         __u64                     pos = 0;
 754         int                       first = 0;
 755         int                       rc = 0;
 756         struct ll_dir_chain       chain;
 757         ENTRY;
 758
 759         {
 760                 char pname[16];
 761                 snprintf(pname, 15, "ll_sa_%u", lli->lli_opendir_pid);
 762                 cfs_daemonize(pname);
 763         }
 764
 765         atomic_inc(&sbi->ll_sa_total);
 766         cfs_spin_lock(&lli->lli_sa_lock);
 767         thread->t_flags = SVC_RUNNING;
 768         cfs_spin_unlock(&lli->lli_sa_lock);
 769         cfs_waitq_signal(&thread->t_ctl_waitq);
 770         CDEBUG(D_READA, "start doing statahead for %s\n", parent->d_name.name);
 771
 772         ll_dir_chain_init(&chain);
 773         page = ll_get_dir_page(dir, pos, 0, &chain);
 774
 775         while (1) {
 776                 struct l_wait_info lwi = { 0 };
 777                 struct lu_dirpage *dp;
 778                 struct lu_dirent  *ent;
 779
 780                 if (IS_ERR(page)) {
 781                         rc = PTR_ERR(page);
 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);
 786                         break;
 787                 }
 788
 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);
 794
 795                         if (unlikely(namelen == 0))
 796                                 /*
 797                                  * Skip dummy record.
 798                                  */
 799                                 continue;
 800
 801                         if (name[0] == '.') {
 802                                 if (namelen == 1) {
 803                                         /*
 804                                          * skip "."
 805                                          */
 806                                         continue;
 807                                 } else if (name[1] == '.' && namelen == 2) {
 808                                         /*
 809                                          * skip ".."
 810                                          */
 811                                         continue;
 812                                 } else if (!sai->sai_ls_all) {
 813                                         /*
 814                                          * skip hidden files.
 815                                          */
 816                                         sai->sai_skip_hidden++;
 817                                         continue;
 818                                 }
 819                         }
 820
 821                         /*
 822                          * don't stat-ahead first entry.
 823                          */
 824                         if (unlikely(!first)) {
 825                                 first++;
 826                                 continue;
 827                         }
 828
 829 keep_de:
 830                         l_wait_event(thread->t_ctl_waitq,
 831                                      !sa_is_running(sai) || sa_not_full(sai) ||
 832                                      !sa_received_empty(sai),
 833                                      &lwi);
 834
 835                         while (!sa_received_empty(sai) && sa_is_running(sai))
 836                                 do_statahead_interpret(sai);
 837
 838                         if (unlikely(!sa_is_running(sai))) {
 839                                 ll_put_page(page);
 840                                 GOTO(out, rc);
 841                         }
 842
 843                         if (!sa_not_full(sai))
 844                                 /*
 845                                  * do not skip the current de.
 846                                  */
 847                                 goto keep_de;
 848
 849                         rc = ll_statahead_one(parent, name, namelen);
 850                         if (rc < 0) {
 851                                 ll_put_page(page);
 852                                 GOTO(out, rc);
 853                         }
 854                 }
 855                 pos = le64_to_cpu(dp->ldp_hash_end);
 856                 ll_put_page(page);
 857                 if (pos == DIR_END_OFF) {
 858                         /*
 859                          * End of directory reached.
 860                          */
 861                         while (1) {
 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,
 866                                              &lwi);
 867                                 if (!sa_received_empty(sai) &&
 868                                     sa_is_running(sai))
 869                                         do_statahead_interpret(sai);
 870                                 else
 871                                         GOTO(out, rc);
 872                         }
 873                 } else if (1) {
 874                         /*
 875                          * chain is exhausted.
 876                          * Normal case: continue to the next page.
 877                          */
 878                         page = ll_get_dir_page(dir, pos, 1, &chain);
 879                 } else {
 880                         /*
 881                          * go into overflow page.
 882                          */
 883                 }
 884         }
 885         EXIT;
 886
 887 out:
 888         ll_dir_chain_fini(&chain);
 889         cfs_spin_lock(&lli->lli_sa_lock);
 890         thread->t_flags = SVC_STOPPED;
 891         cfs_spin_unlock(&lli->lli_sa_lock);
 892         cfs_waitq_signal(&sai->sai_waitq);
 893         cfs_waitq_signal(&thread->t_ctl_waitq);
 894         ll_sai_put(sai);
 895         dput(parent);
 896         CDEBUG(D_READA, "statahead thread stopped, pid %d\n",
 897                cfs_curproc_pid());
 898         return rc;
 899 }
 900
 901 /**
 902  * called in ll_file_release().
 903  */
 904 void ll_stop_statahead(struct inode *dir, void *key)
 905 {
 906         struct ll_inode_info *lli = ll_i2info(dir);
 907
 908         if (unlikely(key == NULL))
 909                 return;
 910
 911         cfs_spin_lock(&lli->lli_sa_lock);
 912         if (lli->lli_opendir_key != key || lli->lli_opendir_pid == 0) {
 913                 cfs_spin_unlock(&lli->lli_sa_lock);
 914                 return;
 915         }
 916
 917         lli->lli_opendir_key = NULL;
 918
 919         if (lli->lli_sai) {
 920                 struct l_wait_info lwi = { 0 };
 921                 struct ptlrpc_thread *thread = &lli->lli_sai->sai_thread;
 922
 923                 if (!sa_is_stopped(lli->lli_sai)) {
 924                         thread->t_flags = SVC_STOPPING;
 925                         cfs_spin_unlock(&lli->lli_sa_lock);
 926                         cfs_waitq_signal(&thread->t_ctl_waitq);
 927
 928                         CDEBUG(D_READA, "stopping statahead thread, pid %d\n",
 929                                cfs_curproc_pid());
 930                         l_wait_event(thread->t_ctl_waitq,
 931                                      sa_is_stopped(lli->lli_sai),
 932                                      &lwi);
 933                 } else {
 934                         cfs_spin_unlock(&lli->lli_sa_lock);
 935                 }
 936
 937                 /*
 938                  * Put the ref which was held when first statahead_enter.
 939                  * It maybe not the last ref for some statahead requests
 940                  * maybe inflight.
 941                  */
 942                 ll_sai_put(lli->lli_sai);
 943         } else {
 944                 lli->lli_opendir_pid = 0;
 945                 cfs_spin_unlock(&lli->lli_sa_lock);
 946         }
 947 }
 948
 949 enum {
 950         /**
 951          * not first dirent, or is "."
 952          */
 953         LS_NONE_FIRST_DE = 0,
 954         /**
 955          * the first non-hidden dirent
 956          */
 957         LS_FIRST_DE,
 958         /**
 959          * the first hidden dirent, that is "."
 960          */
 961         LS_FIRST_DOT_DE
 962 };
 963
 964 static int is_first_dirent(struct inode *dir, struct dentry *dentry)
 965 {
 966         struct ll_dir_chain chain;
 967         struct qstr        *target = &dentry->d_name;
 968         struct page        *page;
 969         __u64               pos = 0;
 970         int                 dot_de;
 971         int                 rc = LS_NONE_FIRST_DE;
 972         ENTRY;
 973
 974         ll_dir_chain_init(&chain);
 975         page = ll_get_dir_page(dir, pos, 0, &chain);
 976
 977         while (1) {
 978                 struct lu_dirpage *dp;
 979                 struct lu_dirent  *ent;
 980
 981                 if (IS_ERR(page)) {
 982                         struct ll_inode_info *lli = ll_i2info(dir);
 983
 984                         rc = PTR_ERR(page);
 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);
 989                         break;
 990                 }
 991
 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);
 997
 998                         if (namelen == 0)
 999                                 /*
1000                                  * skip dummy record.
1001                                  */
1002                                 continue;
1003
1004                         if (name[0] == '.') {
1005                                 if (namelen == 1)
1006                                         /*
1007                                          * skip "."
1008                                          */
1009                                         continue;
1010                                 else if (name[1] == '.' && namelen == 2)
1011                                         /*
1012                                          * skip ".."
1013                                          */
1014                                         continue;
1015                                 else
1016                                         dot_de = 1;
1017                         } else {
1018                                 dot_de = 0;
1019                         }
1020
1021                         if (dot_de && target->name[0] != '.') {
1022                                 CDEBUG(D_READA, "%.*s skip hidden file %.*s\n",
1023                                        target->len, target->name,
1024                                        namelen, name);
1025                                 continue;
1026                         }
1027
1028                         if (target->len != namelen ||
1029                             memcmp(target->name, name, namelen) != 0)
1030                                 rc = LS_NONE_FIRST_DE;
1031                         else if (!dot_de)
1032                                 rc = LS_FIRST_DE;
1033                         else
1034                                 rc = LS_FIRST_DOT_DE;
1035
1036                         ll_put_page(page);
1037                         GOTO(out, rc);
1038                 }
1039                 pos = le64_to_cpu(dp->ldp_hash_end);
1040                 ll_put_page(page);
1041                 if (pos == DIR_END_OFF) {
1042                         /*
1043                          * End of directory reached.
1044                          */
1045                         break;
1046                 } else if (1) {
1047                         /*
1048                          * chain is exhausted
1049                          * Normal case: continue to the next page.
1050                          */
1051                         page = ll_get_dir_page(dir, pos, 1, &chain);
1052                 } else {
1053                         /*
1054                          * go into overflow page.
1055                          */
1056                 }
1057         }
1058         EXIT;
1059
1060 out:
1061         ll_dir_chain_fini(&chain);
1062         return rc;
1063 }
1064
1065 /**
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
1074  */
1075 int do_statahead_enter(struct inode *dir, struct dentry **dentryp, int lookup)
1076 {
1077         struct ll_inode_info     *lli;
1078         struct ll_statahead_info *sai;
1079         struct dentry            *parent;
1080         struct l_wait_info        lwi = { 0 };
1081         int                       rc = 0;
1082         ENTRY;
1083
1084         LASSERT(dir != NULL);
1085         lli = ll_i2info(dir);
1086         LASSERT(lli->lli_opendir_pid == cfs_curproc_pid());
1087         sai = lli->lli_sai;
1088
1089         if (sai) {
1090                 if (unlikely(sa_is_stopped(sai) &&
1091                              cfs_list_empty(&sai->sai_entries_stated)))
1092                         RETURN(-EBADFD);
1093
1094                 if ((*dentryp)->d_name.name[0] == '.') {
1095                         if (likely(sai->sai_ls_all ||
1096                             sai->sai_miss_hidden >= sai->sai_skip_hidden)) {
1097                                 /*
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.
1101                                  */
1102                         } else {
1103                                 if (!sai->sai_ls_all)
1104                                         /*
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.
1109                                          */
1110                                         sai->sai_ls_all = 1;
1111
1112                                 /*
1113                                  * Such "getattr" has been skipped before
1114                                  * "sai_ls_all" enabled as above.
1115                                  */
1116                                 sai->sai_miss_hidden++;
1117                                 RETURN(-ENOENT);
1118                         }
1119                 }
1120
1121                 if (!ll_sai_entry_stated(sai)) {
1122                         /*
1123                          * thread started already, avoid double-stat.
1124                          */
1125                         lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
1126                         rc = l_wait_event(sai->sai_waitq,
1127                                           ll_sai_entry_stated(sai) ||
1128                                           sa_is_stopped(sai),
1129                                           &lwi);
1130                         if (unlikely(rc == -EINTR))
1131                                 RETURN(rc);
1132                 }
1133
1134                 if (lookup) {
1135                         struct dentry *result;
1136
1137                         result = d_lookup((*dentryp)->d_parent,
1138                                           &(*dentryp)->d_name);
1139                         if (result) {
1140                                 LASSERT(result != *dentryp);
1141                                 /* BUG 16303: do not drop reference count for
1142                                  * "*dentryp", VFS will do that by itself. */
1143                                 *dentryp = result;
1144                                 RETURN(1);
1145                         }
1146                 }
1147                 /*
1148                  * do nothing for revalidate.
1149                  */
1150                 RETURN(0);
1151         }
1152
1153         /* I am the "lli_opendir_pid" owner, only me can set "lli_sai". */
1154         rc = is_first_dirent(dir, *dentryp);
1155         if (rc == LS_NONE_FIRST_DE)
1156                 /* It is not "ls -{a}l" operation, no need statahead for it. */
1157                 GOTO(out, rc = -EAGAIN);
1158
1159         sai = ll_sai_alloc();
1160         if (sai == NULL)
1161                 GOTO(out, rc = -ENOMEM);
1162
1163         sai->sai_ls_all = (rc == LS_FIRST_DOT_DE);
1164         sai->sai_inode = igrab(dir);
1165         if (unlikely(sai->sai_inode == NULL)) {
1166                 CWARN("Do not start stat ahead on dying inode "DFID"\n",
1167                       PFID(&lli->lli_fid));
1168                 OBD_FREE_PTR(sai);
1169                 GOTO(out, rc = -ESTALE);
1170         }
1171
1172         /* get parent reference count here, and put it in ll_statahead_thread */
1173         parent = dget((*dentryp)->d_parent);
1174         if (unlikely(sai->sai_inode != parent->d_inode)) {
1175                 struct ll_inode_info *nlli = ll_i2info(parent->d_inode);
1176
1177                 CWARN("Race condition, someone changed %.*s just now: "
1178                       "old parent "DFID", new parent "DFID"\n",
1179                       (*dentryp)->d_name.len, (*dentryp)->d_name.name,
1180                       PFID(&lli->lli_fid), PFID(&nlli->lli_fid));
1181                 dput(parent);
1182                 iput(sai->sai_inode);
1183                 OBD_FREE_PTR(sai);
1184                 RETURN(-EAGAIN);
1185         }
1186
1187         lli->lli_sai = sai;
1188         rc = cfs_kernel_thread(ll_statahead_thread, parent, 0);
1189         if (rc < 0) {
1190                 CERROR("can't start ll_sa thread, rc: %d\n", rc);
1191                 dput(parent);
1192                 lli->lli_opendir_key = NULL;
1193                 sai->sai_thread.t_flags = SVC_STOPPED;
1194                 ll_sai_put(sai);
1195                 LASSERT(lli->lli_sai == NULL);
1196                 RETURN(-EAGAIN);
1197         }
1198
1199         l_wait_event(sai->sai_thread.t_ctl_waitq,
1200                      sa_is_running(sai) || sa_is_stopped(sai),
1201                      &lwi);
1202
1203         /*
1204          * We don't stat-ahead for the first dirent since we are already in
1205          * lookup, and -EEXIST also indicates that this is the first dirent.
1206          */
1207         RETURN(-EEXIST);
1208
1209 out:
1210         cfs_spin_lock(&lli->lli_sa_lock);
1211         lli->lli_opendir_key = NULL;
1212         lli->lli_opendir_pid = 0;
1213         cfs_spin_unlock(&lli->lli_sa_lock);
1214         return rc;
1215 }
1216
1217 /**
1218  * update hit/miss count.
1219  */
1220 void ll_statahead_exit(struct inode *dir, struct dentry *dentry, int result)
1221 {
1222         struct ll_inode_info     *lli;
1223         struct ll_statahead_info *sai;
1224         struct ll_sb_info        *sbi;
1225         struct ll_dentry_data    *ldd = ll_d2d(dentry);
1226         int                       rc;
1227         ENTRY;
1228
1229         LASSERT(dir != NULL);
1230         lli = ll_i2info(dir);
1231         LASSERT(lli->lli_opendir_pid == cfs_curproc_pid());
1232         sai = lli->lli_sai;
1233         LASSERT(sai != NULL);
1234         sbi = ll_i2sbi(dir);
1235
1236         rc = ll_sai_entry_fini(sai);
1237         /* rc == -ENOENT means such dentry was removed just between statahead
1238          * readdir and pre-fetched, count it as hit.
1239          *
1240          * result == -ENOENT has two meanings:
1241          * 1. such dentry was removed just between statahead pre-fetched and
1242          *    main process stat such dentry.
1243          * 2. main process stat non-exist dentry.
1244          * We can not distinguish such two cases, just count them as miss. */
1245         if (result >= 1 || unlikely(rc == -ENOENT)) {
1246                 sai->sai_hit++;
1247                 sai->sai_consecutive_miss = 0;
1248                 sai->sai_max = min(2 * sai->sai_max, sbi->ll_sa_max);
1249         } else {
1250                 sai->sai_miss++;
1251                 sai->sai_consecutive_miss++;
1252                 if (sa_low_hit(sai) && sa_is_running(sai)) {
1253                         atomic_inc(&sbi->ll_sa_wrong);
1254                         CDEBUG(D_READA, "Statahead for dir "DFID" hit ratio "
1255                                "too low: hit/miss %u/%u, sent/replied %u/%u, "
1256                                "stopping statahead thread: pid %d\n",
1257                                PFID(&lli->lli_fid), sai->sai_hit,
1258                                sai->sai_miss, sai->sai_sent,
1259                                sai->sai_replied, cfs_curproc_pid());
1260                         cfs_spin_lock(&lli->lli_sa_lock);
1261                         if (!sa_is_stopped(sai))
1262                                 sai->sai_thread.t_flags = SVC_STOPPING;
1263                         cfs_spin_unlock(&lli->lli_sa_lock);
1264                 }
1265         }
1266
1267         if (!sa_is_stopped(sai))
1268                 cfs_waitq_signal(&sai->sai_thread.t_ctl_waitq);
1269         if (likely(ldd != NULL))
1270                 ldd->lld_sa_generation = sai->sai_generation;
1271
1272         EXIT;
1273 }