lustre/llite/statahead.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2013, Intel Corporation.
  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/highmem.h>
  41 #include <linux/pagemap.h>
  42
  43 #define DEBUG_SUBSYSTEM S_LLITE
  44
  45 #include <obd_support.h>
  46 #include <lustre_lite.h>
  47 #include <lustre_dlm.h>
  48 #include "llite_internal.h"
  49
  50 #define SA_OMITTED_ENTRY_MAX 8ULL
  51
  52 typedef enum {
  53         /** negative values are for error cases */
  54         SA_ENTRY_INIT = 0,      /** init entry */
  55         SA_ENTRY_SUCC = 1,      /** stat succeed */
  56         SA_ENTRY_INVA = 2,      /** invalid entry */
  57         SA_ENTRY_DEST = 3,      /** entry to be destroyed */
  58 } se_stat_t;
  59
  60 struct ll_sa_entry {
  61         /* link into sai->sai_entries */
  62         struct list_head        se_link;
  63         /* link into sai->sai_entries_{received,stated} */
  64         struct list_head        se_list;
  65         /* link into sai hash table locally */
  66         struct list_head        se_hash;
  67         /* entry reference count */
  68         atomic_t                se_refcount;
  69         /* entry index in the sai */
  70         __u64                   se_index;
  71         /* low layer ldlm lock handle */
  72         __u64                   se_handle;
  73         /* entry status */
  74         se_stat_t               se_stat;
  75         /* entry size, contains name */
  76         int                     se_size;
  77         /* pointer to async getattr enqueue info */
  78         struct md_enqueue_info *se_minfo;
  79         /* pointer to the async getattr request */
  80         struct ptlrpc_request  *se_req;
  81         /* pointer to the target inode */
  82         struct inode           *se_inode;
  83         /* entry name */
  84         struct qstr             se_qstr;
  85 };
  86
  87 static unsigned int sai_generation = 0;
  88 static DEFINE_SPINLOCK(sai_generation_lock);
  89
  90 static inline int ll_sa_entry_unhashed(struct ll_sa_entry *entry)
  91 {
  92         return list_empty(&entry->se_hash);
  93 }
  94
  95 /*
  96  * The entry only can be released by the caller, it is necessary to hold lock.
  97  */
  98 static inline int ll_sa_entry_stated(struct ll_sa_entry *entry)
  99 {
 100         smp_rmb();
 101         return (entry->se_stat != SA_ENTRY_INIT);
 102 }
 103
 104 static inline int ll_sa_entry_hash(int val)
 105 {
 106         return val & LL_SA_CACHE_MASK;
 107 }
 108
 109 /*
 110  * Insert entry to hash SA table.
 111  */
 112 static inline void
 113 ll_sa_entry_enhash(struct ll_statahead_info *sai, struct ll_sa_entry *entry)
 114 {
 115         int i = ll_sa_entry_hash(entry->se_qstr.hash);
 116
 117         spin_lock(&sai->sai_cache_lock[i]);
 118         list_add_tail(&entry->se_hash, &sai->sai_cache[i]);
 119         spin_unlock(&sai->sai_cache_lock[i]);
 120 }
 121
 122 /*
 123  * Remove entry from SA table.
 124  */
 125 static inline void
 126 ll_sa_entry_unhash(struct ll_statahead_info *sai, struct ll_sa_entry *entry)
 127 {
 128         int i = ll_sa_entry_hash(entry->se_qstr.hash);
 129
 130         spin_lock(&sai->sai_cache_lock[i]);
 131         list_del_init(&entry->se_hash);
 132         spin_unlock(&sai->sai_cache_lock[i]);
 133 }
 134
 135 static inline int agl_should_run(struct ll_statahead_info *sai,
 136                                  struct inode *inode)
 137 {
 138         return (inode != NULL && S_ISREG(inode->i_mode) && sai->sai_agl_valid);
 139 }
 140
 141 static inline struct ll_sa_entry *
 142 sa_first_received_entry(struct ll_statahead_info *sai)
 143 {
 144         return list_entry(sai->sai_entries_received.next,
 145                           struct ll_sa_entry, se_list);
 146 }
 147
 148 static inline struct ll_inode_info *
 149 agl_first_entry(struct ll_statahead_info *sai)
 150 {
 151         return list_entry(sai->sai_entries_agl.next,
 152                           struct ll_inode_info, lli_agl_list);
 153 }
 154
 155 static inline int sa_sent_full(struct ll_statahead_info *sai)
 156 {
 157         return atomic_read(&sai->sai_cache_count) >= sai->sai_max;
 158 }
 159
 160 static inline int sa_received_empty(struct ll_statahead_info *sai)
 161 {
 162         return list_empty(&sai->sai_entries_received);
 163 }
 164
 165 static inline int agl_list_empty(struct ll_statahead_info *sai)
 166 {
 167         return list_empty(&sai->sai_entries_agl);
 168 }
 169
 170 /**
 171  * (1) hit ratio less than 80%
 172  * or
 173  * (2) consecutive miss more than 8
 174  * then means low hit.
 175  */
 176 static inline int sa_low_hit(struct ll_statahead_info *sai)
 177 {
 178         return ((sai->sai_hit > 7 && sai->sai_hit < 4 * sai->sai_miss) ||
 179                 (sai->sai_consecutive_miss > 8));
 180 }
 181
 182 /*
 183  * If the given index is behind of statahead window more than
 184  * SA_OMITTED_ENTRY_MAX, then it is old.
 185  */
 186 static inline int is_omitted_entry(struct ll_statahead_info *sai, __u64 index)
 187 {
 188         return ((__u64)sai->sai_max + index + SA_OMITTED_ENTRY_MAX <
 189                  sai->sai_index);
 190 }
 191
 192 /*
 193  * Insert it into sai_entries tail when init.
 194  */
 195 static struct ll_sa_entry *
 196 ll_sa_entry_alloc(struct ll_statahead_info *sai, __u64 index, const char *name,
 197                  int len)
 198 {
 199         struct ll_inode_info *lli;
 200         struct ll_sa_entry   *entry;
 201         int                   entry_size;
 202         char                 *dname;
 203         ENTRY;
 204
 205         entry_size = sizeof(struct ll_sa_entry) + (len & ~3) + 4;
 206         OBD_ALLOC(entry, entry_size);
 207         if (unlikely(entry == NULL))
 208                 RETURN(ERR_PTR(-ENOMEM));
 209
 210         CDEBUG(D_READA, "alloc sa entry %.*s(%p) index "LPU64"\n",
 211                len, name, entry, index);
 212
 213         entry->se_index = index;
 214
 215         /*
 216          * Statahead entry reference rules:
 217          *
 218          * 1) When statahead entry is initialized, its reference is set as 2.
 219          *    One reference is used by the directory scanner. When the scanner
 220          *    searches the statahead cache for the given name, it can perform
 221          *    lockless hash lookup (only the scanner can remove entry from hash
 222          *    list), and once found, it needn't to call "atomic_inc()" for the
 223          *    entry reference. So the performance is improved. After using the
 224          *    statahead entry, the scanner will call "atomic_dec()" to drop the
 225          *    reference held when initialization. If it is the last reference,
 226          *    the statahead entry will be freed.
 227          *
 228          * 2) All other threads, including statahead thread and ptlrpcd thread,
 229          *    when they process the statahead entry, the reference for target
 230          *    should be held to guarantee the entry will not be released by the
 231          *    directory scanner. After processing the entry, these threads will
 232          *    drop the entry reference. If it is the last reference, the entry
 233          *    will be freed.
 234          *
 235          *    The second reference when initializes the statahead entry is used
 236          *    by the statahead thread, following the rule 2).
 237          */
 238         atomic_set(&entry->se_refcount, 2);
 239         entry->se_stat = SA_ENTRY_INIT;
 240         entry->se_size = entry_size;
 241         dname = (char *)entry + sizeof(struct ll_sa_entry);
 242         memcpy(dname, name, len);
 243         dname[len] = 0;
 244         entry->se_qstr.hash = full_name_hash(name, len);
 245         entry->se_qstr.len = len;
 246         entry->se_qstr.name = dname;
 247
 248         lli = ll_i2info(sai->sai_inode);
 249         spin_lock(&lli->lli_sa_lock);
 250         list_add_tail(&entry->se_link, &sai->sai_entries);
 251         INIT_LIST_HEAD(&entry->se_list);
 252         ll_sa_entry_enhash(sai, entry);
 253         spin_unlock(&lli->lli_sa_lock);
 254
 255         atomic_inc(&sai->sai_cache_count);
 256
 257         RETURN(entry);
 258 }
 259
 260 /*
 261  * Used by the directory scanner to search entry with name.
 262  *
 263  * Only the caller can remove the entry from hash, so it is unnecessary to hold
 264  * hash lock. It is caller's duty to release the init refcount on the entry, so
 265  * it is also unnecessary to increase refcount on the entry.
 266  */
 267 static struct ll_sa_entry *
 268 ll_sa_entry_get_byname(struct ll_statahead_info *sai, const struct qstr *qstr)
 269 {
 270         struct ll_sa_entry *entry;
 271         int i = ll_sa_entry_hash(qstr->hash);
 272
 273         list_for_each_entry(entry, &sai->sai_cache[i], se_hash) {
 274                 if (entry->se_qstr.hash == qstr->hash &&
 275                     entry->se_qstr.len == qstr->len &&
 276                     memcmp(entry->se_qstr.name, qstr->name, qstr->len) == 0)
 277                         return entry;
 278         }
 279         return NULL;
 280 }
 281
 282 /*
 283  * Used by the async getattr request callback to find entry with index.
 284  *
 285  * Inside lli_sa_lock to prevent others to change the list during the search.
 286  * It needs to increase entry refcount before returning to guarantee that the
 287  * entry cannot be freed by others.
 288  */
 289 static struct ll_sa_entry *
 290 ll_sa_entry_get_byindex(struct ll_statahead_info *sai, __u64 index)
 291 {
 292         struct ll_sa_entry *entry;
 293
 294         list_for_each_entry(entry, &sai->sai_entries, se_link) {
 295                 if (entry->se_index == index) {
 296                         LASSERT(atomic_read(&entry->se_refcount) > 0);
 297                         atomic_inc(&entry->se_refcount);
 298                         return entry;
 299                 }
 300                 if (entry->se_index > index)
 301                         break;
 302         }
 303         return NULL;
 304 }
 305
 306 static void ll_sa_entry_put(struct ll_statahead_info *sai,
 307                              struct ll_sa_entry *entry)
 308 {
 309         if (atomic_dec_and_test(&entry->se_refcount)) {
 310                 CDEBUG(D_READA, "free sa entry %.*s(%p) index "LPU64"\n",
 311                        entry->se_qstr.len, entry->se_qstr.name, entry,
 312                        entry->se_index);
 313
 314                 LASSERT(list_empty(&entry->se_link));
 315                 LASSERT(list_empty(&entry->se_list));
 316                 LASSERT(ll_sa_entry_unhashed(entry));
 317
 318                 if (entry->se_inode)
 319                         iput(entry->se_inode);
 320
 321                 OBD_FREE(entry, entry->se_size);
 322                 atomic_dec(&sai->sai_cache_count);
 323         }
 324 }
 325
 326 static inline void
 327 do_sa_entry_fini(struct ll_statahead_info *sai, struct ll_sa_entry *entry)
 328 {
 329         struct ll_inode_info *lli = ll_i2info(sai->sai_inode);
 330
 331         LASSERT(!ll_sa_entry_unhashed(entry));
 332         LASSERT(!list_empty(&entry->se_link));
 333
 334         ll_sa_entry_unhash(sai, entry);
 335
 336         spin_lock(&lli->lli_sa_lock);
 337         entry->se_stat = SA_ENTRY_DEST;
 338         list_del_init(&entry->se_link);
 339         if (likely(!list_empty(&entry->se_list)))
 340                 list_del_init(&entry->se_list);
 341         spin_unlock(&lli->lli_sa_lock);
 342
 343         ll_sa_entry_put(sai, entry);
 344 }
 345
 346 /*
 347  * Delete it from sai_entries_stated list when fini.
 348  */
 349 static void
 350 ll_sa_entry_fini(struct ll_statahead_info *sai, struct ll_sa_entry *entry)
 351 {
 352         struct ll_sa_entry *pos, *next;
 353
 354         if (entry)
 355                 do_sa_entry_fini(sai, entry);
 356
 357         /* drop old entry, only 'scanner' process does this, no need to lock */
 358         list_for_each_entry_safe(pos, next, &sai->sai_entries, se_link) {
 359                 if (!is_omitted_entry(sai, pos->se_index))
 360                         break;
 361                 /* keep those whose statahead RPC not finished */
 362                 if (pos->se_stat == SA_ENTRY_SUCC ||
 363                     pos->se_stat == SA_ENTRY_INVA)
 364                         do_sa_entry_fini(sai, pos);
 365         }
 366 }
 367
 368 /*
 369  * Inside lli_sa_lock.
 370  */
 371 static void
 372 __sa_entry_post_stat(struct ll_statahead_info *sai, struct ll_sa_entry *entry,
 373                      se_stat_t stat)
 374 {
 375         struct ll_sa_entry *se;
 376         struct list_head *pos = &sai->sai_entries_stated;
 377
 378         LASSERT(entry->se_stat == SA_ENTRY_INIT);
 379
 380         if (!list_empty(&entry->se_list))
 381                 list_del_init(&entry->se_list);
 382
 383         list_for_each_entry_reverse(se, &sai->sai_entries_stated, se_list) {
 384                 if (se->se_index < entry->se_index) {
 385                         pos = &se->se_list;
 386                         break;
 387                 }
 388         }
 389
 390         list_add(&entry->se_list, pos);
 391         entry->se_stat = stat;
 392 }
 393
 394 /*
 395  * Move entry to sai_entries_stated and sort with the index.
 396  * \retval 1    -- entry to be destroyed.
 397  * \retval 0    -- entry is inserted into stated list.
 398  */
 399 static void
 400 sa_entry_post_stat(struct ll_statahead_info *sai, struct ll_sa_entry *entry,
 401                    se_stat_t stat)
 402 {
 403         struct ll_inode_info *lli = ll_i2info(sai->sai_inode);
 404         struct md_enqueue_info *minfo = entry->se_minfo;
 405         struct ptlrpc_request *req = entry->se_req;
 406
 407         /* release resources used in RPC */
 408         if (minfo) {
 409                 entry->se_minfo = NULL;
 410                 ll_intent_release(&minfo->mi_it);
 411                 iput(minfo->mi_dir);
 412                 OBD_FREE_PTR(minfo);
 413         }
 414
 415         if (req) {
 416                 entry->se_req = NULL;
 417                 ptlrpc_req_finished(req);
 418         }
 419
 420         spin_lock(&lli->lli_sa_lock);
 421         __sa_entry_post_stat(sai, entry, stat);
 422         spin_unlock(&lli->lli_sa_lock);
 423 }
 424
 425 /*
 426  * Insert inode into the list of sai_entries_agl.
 427  */
 428 static void ll_agl_add(struct ll_statahead_info *sai,
 429                        struct inode *inode, int index)
 430 {
 431         struct ll_inode_info *child  = ll_i2info(inode);
 432         struct ll_inode_info *parent = ll_i2info(sai->sai_inode);
 433         int                   added  = 0;
 434
 435         spin_lock(&child->lli_agl_lock);
 436         if (child->lli_agl_index == 0) {
 437                 child->lli_agl_index = index;
 438                 spin_unlock(&child->lli_agl_lock);
 439
 440                 LASSERT(list_empty(&child->lli_agl_list));
 441
 442                 igrab(inode);
 443                 spin_lock(&parent->lli_agl_lock);
 444                 if (agl_list_empty(sai))
 445                         added = 1;
 446                 list_add_tail(&child->lli_agl_list, &sai->sai_entries_agl);
 447                 spin_unlock(&parent->lli_agl_lock);
 448         } else {
 449                 spin_unlock(&child->lli_agl_lock);
 450         }
 451
 452         if (added > 0)
 453                 wake_up(&sai->sai_agl_thread.t_ctl_waitq);
 454 }
 455
 456 static struct ll_statahead_info *ll_sai_alloc(void)
 457 {
 458         struct ll_statahead_info *sai;
 459         int                       i;
 460         ENTRY;
 461
 462         OBD_ALLOC_PTR(sai);
 463         if (!sai)
 464                 RETURN(NULL);
 465
 466         atomic_set(&sai->sai_refcount, 1);
 467
 468         spin_lock(&sai_generation_lock);
 469         sai->sai_generation = ++sai_generation;
 470         if (unlikely(sai_generation == 0))
 471                 sai->sai_generation = ++sai_generation;
 472         spin_unlock(&sai_generation_lock);
 473
 474         sai->sai_max = LL_SA_RPC_MIN;
 475         sai->sai_index = 1;
 476         init_waitqueue_head(&sai->sai_waitq);
 477         init_waitqueue_head(&sai->sai_thread.t_ctl_waitq);
 478         init_waitqueue_head(&sai->sai_agl_thread.t_ctl_waitq);
 479
 480         INIT_LIST_HEAD(&sai->sai_entries);
 481         INIT_LIST_HEAD(&sai->sai_entries_received);
 482         INIT_LIST_HEAD(&sai->sai_entries_stated);
 483         INIT_LIST_HEAD(&sai->sai_entries_agl);
 484
 485         for (i = 0; i < LL_SA_CACHE_SIZE; i++) {
 486                 INIT_LIST_HEAD(&sai->sai_cache[i]);
 487                 spin_lock_init(&sai->sai_cache_lock[i]);
 488         }
 489         atomic_set(&sai->sai_cache_count, 0);
 490
 491         RETURN(sai);
 492 }
 493
 494 static inline struct ll_statahead_info *ll_sai_get(struct inode *dir)
 495 {
 496         struct ll_inode_info *lli = ll_i2info(dir);
 497         struct ll_statahead_info *sai = NULL;
 498
 499         spin_lock(&lli->lli_sa_lock);
 500         sai = lli->lli_sai;
 501         if (sai != NULL)
 502                 atomic_inc(&sai->sai_refcount);
 503         spin_unlock(&lli->lli_sa_lock);
 504
 505         return sai;
 506 }
 507
 508 static void ll_sai_put(struct ll_statahead_info *sai)
 509 {
 510         struct ll_inode_info *lli = ll_i2info(sai->sai_inode);
 511
 512         if (atomic_dec_and_lock(&sai->sai_refcount, &lli->lli_sa_lock)) {
 513                 struct ll_sa_entry *entry, *next;
 514                 struct ll_sb_info *sbi = ll_i2sbi(sai->sai_inode);
 515
 516                 lli->lli_sai = NULL;
 517                 spin_unlock(&lli->lli_sa_lock);
 518
 519                 LASSERT(thread_is_stopped(&sai->sai_thread));
 520                 LASSERT(thread_is_stopped(&sai->sai_agl_thread));
 521                 LASSERT(sai->sai_sent == sai->sai_replied);
 522
 523                 list_for_each_entry_safe(entry, next, &sai->sai_entries,
 524                                          se_link)
 525                         do_sa_entry_fini(sai, entry);
 526
 527                 LASSERT(atomic_read(&sai->sai_cache_count) == 0);
 528                 LASSERT(agl_list_empty(sai));
 529                 LASSERT(atomic_read(&sai->sai_refcount) == 0);
 530
 531                 iput(sai->sai_inode);
 532                 OBD_FREE_PTR(sai);
 533                 atomic_dec(&sbi->ll_sa_running);
 534         }
 535 }
 536
 537 /* Do NOT forget to drop inode refcount when into sai_entries_agl. */
 538 static void ll_agl_trigger(struct inode *inode, struct ll_statahead_info *sai)
 539 {
 540         struct ll_inode_info *lli   = ll_i2info(inode);
 541         __u64                 index = lli->lli_agl_index;
 542         int                   rc;
 543         ENTRY;
 544
 545         LASSERT(list_empty(&lli->lli_agl_list));
 546
 547         /* AGL maybe fall behind statahead with one entry */
 548         if (is_omitted_entry(sai, index + 1)) {
 549                 lli->lli_agl_index = 0;
 550                 iput(inode);
 551                 RETURN_EXIT;
 552         }
 553
 554         /* Someone is in glimpse (sync or async), do nothing. */
 555         rc = down_write_trylock(&lli->lli_glimpse_sem);
 556         if (rc == 0) {
 557                 lli->lli_agl_index = 0;
 558                 iput(inode);
 559                 RETURN_EXIT;
 560         }
 561
 562         /*
 563          * Someone triggered glimpse within 1 sec before.
 564          * 1) The former glimpse succeeded with glimpse lock granted by OST, and
 565          *    if the lock is still cached on client, AGL needs to do nothing. If
 566          *    it is cancelled by other client, AGL maybe cannot obtaion new lock
 567          *    for no glimpse callback triggered by AGL.
 568          * 2) The former glimpse succeeded, but OST did not grant glimpse lock.
 569          *    Under such case, it is quite possible that the OST will not grant
 570          *    glimpse lock for AGL also.
 571          * 3) The former glimpse failed, compared with other two cases, it is
 572          *    relative rare. AGL can ignore such case, and it will not muchly
 573          *    affect the performance.
 574          */
 575         if (lli->lli_glimpse_time != 0 &&
 576             cfs_time_before(cfs_time_shift(-1), lli->lli_glimpse_time)) {
 577                 up_write(&lli->lli_glimpse_sem);
 578                 lli->lli_agl_index = 0;
 579                 iput(inode);
 580                 RETURN_EXIT;
 581         }
 582
 583         CDEBUG(D_READA, "Handling (init) async glimpse: inode = "
 584                DFID", idx = "LPU64"\n", PFID(&lli->lli_fid), index);
 585
 586         cl_agl(inode);
 587         lli->lli_agl_index = 0;
 588         lli->lli_glimpse_time = cfs_time_current();
 589         up_write(&lli->lli_glimpse_sem);
 590
 591         CDEBUG(D_READA, "Handled (init) async glimpse: inode= "
 592                DFID", idx = "LPU64", rc = %d\n",
 593                PFID(&lli->lli_fid), index, rc);
 594
 595         iput(inode);
 596
 597         EXIT;
 598 }
 599
 600 /* prepare inode for received statahead entry, and add it into agl list */
 601 static void sa_post_one(struct ll_statahead_info *sai,
 602                         struct ll_sa_entry *entry)
 603 {
 604         struct inode           *dir   = sai->sai_inode;
 605         struct inode           *child;
 606         struct md_enqueue_info *minfo;
 607         struct lookup_intent   *it;
 608         struct ptlrpc_request  *req;
 609         struct mdt_body        *body;
 610         int                     rc    = 0;
 611         ENTRY;
 612
 613         LASSERT(entry->se_handle != 0);
 614
 615         minfo = entry->se_minfo;
 616         it = &minfo->mi_it;
 617         req = entry->se_req;
 618         body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
 619         if (body == NULL)
 620                 GOTO(out, rc = -EFAULT);
 621
 622         child = entry->se_inode;
 623         if (child == NULL) {
 624                 /*
 625                  * lookup.
 626                  */
 627                 LASSERT(fid_is_zero(&minfo->mi_data.op_fid2));
 628
 629                 /* XXX: No fid in reply, this is probaly cross-ref case.
 630                  * SA can't handle it yet. */
 631                 if (body->mbo_valid & OBD_MD_MDS)
 632                         GOTO(out, rc = -EAGAIN);
 633         } else {
 634                 /*
 635                  * revalidate.
 636                  */
 637                 /* unlinked and re-created with the same name */
 638                 if (unlikely(!lu_fid_eq(&minfo->mi_data.op_fid2,
 639                                         &body->mbo_fid1))) {
 640                         entry->se_inode = NULL;
 641                         iput(child);
 642                         child = NULL;
 643                 }
 644         }
 645
 646         it->d.lustre.it_lock_handle = entry->se_handle;
 647         rc = md_revalidate_lock(ll_i2mdexp(dir), it, ll_inode2fid(dir), NULL);
 648         if (rc != 1)
 649                 GOTO(out, rc = -EAGAIN);
 650
 651         rc = ll_prep_inode(&child, req, dir->i_sb, it);
 652         if (rc)
 653                 GOTO(out, rc);
 654
 655         CDEBUG(D_DLMTRACE, "%s: setting l_data to inode "DFID"(%p)\n",
 656                ll_get_fsname(child->i_sb, NULL, 0),
 657                PFID(ll_inode2fid(child)), child);
 658         ll_set_lock_data(ll_i2sbi(dir)->ll_md_exp, child, it, NULL);
 659
 660         entry->se_inode = child;
 661
 662         if (agl_should_run(sai, child))
 663                 ll_agl_add(sai, child, entry->se_index);
 664
 665         EXIT;
 666
 667 out:
 668         /* The "sa_entry_post_stat()" will drop related ldlm ibits lock
 669          * reference count by calling "ll_intent_drop_lock()" in spite of the
 670          * above operations failed or not. Do not worry about calling
 671          * "ll_intent_drop_lock()" more than once. */
 672         sa_entry_post_stat(sai, entry, rc < 0 ? SA_ENTRY_INVA : SA_ENTRY_SUCC);
 673         if (entry->se_index == sai->sai_index_wait)
 674                 wake_up(&sai->sai_waitq);
 675         ll_sa_entry_put(sai, entry);
 676 }
 677
 678 static void ll_post_statahead(struct ll_statahead_info *sai)
 679 {
 680         struct ll_inode_info *lli;
 681
 682         lli = ll_i2info(sai->sai_inode);
 683
 684         while (!sa_received_empty(sai)) {
 685                 struct ll_sa_entry *entry;
 686
 687                 spin_lock(&lli->lli_sa_lock);
 688                 if (unlikely(sa_received_empty(sai))) {
 689                         spin_unlock(&lli->lli_sa_lock);
 690                         break;
 691                 }
 692                 entry = sa_first_received_entry(sai);
 693                 atomic_inc(&entry->se_refcount);
 694                 list_del_init(&entry->se_list);
 695                 spin_unlock(&lli->lli_sa_lock);
 696
 697                 sa_post_one(sai, entry);
 698         }
 699
 700         spin_lock(&lli->lli_agl_lock);
 701         while (!agl_list_empty(sai)) {
 702                 struct ll_inode_info *clli;
 703
 704                 clli = agl_first_entry(sai);
 705                 list_del_init(&clli->lli_agl_list);
 706                 spin_unlock(&lli->lli_agl_lock);
 707
 708                 ll_agl_trigger(&clli->lli_vfs_inode, sai);
 709
 710                 spin_lock(&lli->lli_agl_lock);
 711         }
 712         spin_unlock(&lli->lli_agl_lock);
 713 }
 714
 715 static int ll_statahead_interpret(struct ptlrpc_request *req,
 716                                   struct md_enqueue_info *minfo, int rc)
 717 {
 718         struct lookup_intent *it = &minfo->mi_it;
 719         struct inode *dir = minfo->mi_dir;
 720         struct ll_inode_info *lli = ll_i2info(dir);
 721         struct ll_statahead_info *sai;
 722         struct ll_sa_entry *entry;
 723         int wakeup;
 724         ENTRY;
 725
 726         if (it_disposition(it, DISP_LOOKUP_NEG))
 727                 rc = -ENOENT;
 728
 729         sai = ll_sai_get(dir);
 730         LASSERT(sai != NULL);
 731         LASSERT(!thread_is_stopped(&sai->sai_thread));
 732
 733         spin_lock(&lli->lli_sa_lock);
 734         entry = ll_sa_entry_get_byindex(sai, minfo->mi_cbdata);
 735         LASSERT(entry != NULL);
 736         if (rc != 0) {
 737                 __sa_entry_post_stat(sai, entry, SA_ENTRY_INVA);
 738                 wakeup = (entry->se_index == sai->sai_index_wait);
 739         } else {
 740                 entry->se_minfo = minfo;
 741                 entry->se_req = ptlrpc_request_addref(req);
 742                 /* Release the async ibits lock ASAP to avoid deadlock
 743                  * when statahead thread tries to enqueue lock on parent
 744                  * for readpage and other tries to enqueue lock on child
 745                  * with parent's lock held, for example: unlink. */
 746                 entry->se_handle = it->d.lustre.it_lock_handle;
 747                 ll_intent_drop_lock(it);
 748                 wakeup = sa_received_empty(sai);
 749                 list_add_tail(&entry->se_list, &sai->sai_entries_received);
 750         }
 751         sai->sai_replied++;
 752         spin_unlock(&lli->lli_sa_lock);
 753
 754         ll_sa_entry_put(sai, entry);
 755         if (wakeup)
 756                 wake_up(&sai->sai_thread.t_ctl_waitq);
 757
 758         if (rc != 0) {
 759                 ll_intent_release(it);
 760                 iput(dir);
 761                 OBD_FREE_PTR(minfo);
 762         }
 763         ll_sai_put(sai);
 764         RETURN(rc);
 765 }
 766
 767 static void sa_args_fini(struct md_enqueue_info *minfo,
 768                          struct ldlm_enqueue_info *einfo)
 769 {
 770         LASSERT(minfo && einfo);
 771         iput(minfo->mi_dir);
 772         capa_put(minfo->mi_data.op_capa1);
 773         capa_put(minfo->mi_data.op_capa2);
 774         OBD_FREE_PTR(minfo);
 775         OBD_FREE_PTR(einfo);
 776 }
 777
 778 /**
 779  * There is race condition between "capa_put" and "ll_statahead_interpret" for
 780  * accessing "op_data.op_capa[1,2]" as following:
 781  * "capa_put" releases "op_data.op_capa[1,2]"'s reference count after calling
 782  * "md_intent_getattr_async". But "ll_statahead_interpret" maybe run first, and
 783  * fill "op_data.op_capa[1,2]" as POISON, then cause "capa_put" access invalid
 784  * "ocapa". So here reserve "op_data.op_capa[1,2]" in "pcapa" before calling
 785  * "md_intent_getattr_async".
 786  */
 787 static int sa_args_init(struct inode *dir, struct inode *child,
 788                         struct ll_sa_entry *entry, struct md_enqueue_info **pmi,
 789                         struct ldlm_enqueue_info **pei,
 790                         struct obd_capa **pcapa)
 791 {
 792         struct qstr              *qstr = &entry->se_qstr;
 793         struct md_enqueue_info   *minfo;
 794         struct ldlm_enqueue_info *einfo;
 795         struct md_op_data        *op_data;
 796
 797         OBD_ALLOC_PTR(einfo);
 798         if (einfo == NULL)
 799                 return -ENOMEM;
 800
 801         OBD_ALLOC_PTR(minfo);
 802         if (minfo == NULL) {
 803                 OBD_FREE_PTR(einfo);
 804                 return -ENOMEM;
 805         }
 806
 807         op_data = ll_prep_md_op_data(&minfo->mi_data, dir, child, qstr->name,
 808                                      qstr->len, 0, LUSTRE_OPC_ANY, NULL);
 809         if (IS_ERR(op_data)) {
 810                 OBD_FREE_PTR(einfo);
 811                 OBD_FREE_PTR(minfo);
 812                 return PTR_ERR(op_data);
 813         }
 814
 815         minfo->mi_it.it_op = IT_GETATTR;
 816         minfo->mi_dir = igrab(dir);
 817         minfo->mi_cb = ll_statahead_interpret;
 818         minfo->mi_cbdata = entry->se_index;
 819
 820         einfo->ei_type   = LDLM_IBITS;
 821         einfo->ei_mode   = it_to_lock_mode(&minfo->mi_it);
 822         einfo->ei_cb_bl  = ll_md_blocking_ast;
 823         einfo->ei_cb_cp  = ldlm_completion_ast;
 824         einfo->ei_cb_gl  = NULL;
 825         einfo->ei_cbdata = NULL;
 826
 827         *pmi = minfo;
 828         *pei = einfo;
 829         pcapa[0] = op_data->op_capa1;
 830         pcapa[1] = op_data->op_capa2;
 831
 832         return 0;
 833 }
 834
 835 static int do_sa_lookup(struct inode *dir, struct ll_sa_entry *entry)
 836 {
 837         struct md_enqueue_info   *minfo;
 838         struct ldlm_enqueue_info *einfo;
 839         struct obd_capa          *capas[2];
 840         int                       rc;
 841         ENTRY;
 842
 843         rc = sa_args_init(dir, NULL, entry, &minfo, &einfo, capas);
 844         if (rc)
 845                 RETURN(rc);
 846
 847         rc = md_intent_getattr_async(ll_i2mdexp(dir), minfo, einfo);
 848         if (!rc) {
 849                 capa_put(capas[0]);
 850                 capa_put(capas[1]);
 851         } else {
 852                 sa_args_fini(minfo, einfo);
 853         }
 854
 855         RETURN(rc);
 856 }
 857
 858 /**
 859  * similar to ll_revalidate_it().
 860  * \retval      1 -- dentry valid
 861  * \retval      0 -- will send stat-ahead request
 862  * \retval others -- prepare stat-ahead request failed
 863  */
 864 static int do_sa_revalidate(struct inode *dir, struct ll_sa_entry *entry,
 865                             struct dentry *dentry)
 866 {
 867         struct inode             *inode = dentry->d_inode;
 868         struct lookup_intent      it = { .it_op = IT_GETATTR,
 869                                          .d.lustre.it_lock_handle = 0 };
 870         struct md_enqueue_info   *minfo;
 871         struct ldlm_enqueue_info *einfo;
 872         struct obd_capa          *capas[2];
 873         int rc;
 874         ENTRY;
 875
 876         if (unlikely(inode == NULL))
 877                 RETURN(1);
 878
 879         if (d_mountpoint(dentry))
 880                 RETURN(1);
 881
 882         entry->se_inode = igrab(inode);
 883         rc = md_revalidate_lock(ll_i2mdexp(dir), &it, ll_inode2fid(inode),NULL);
 884         if (rc == 1) {
 885                 entry->se_handle = it.d.lustre.it_lock_handle;
 886                 ll_intent_release(&it);
 887                 RETURN(1);
 888         }
 889
 890         rc = sa_args_init(dir, inode, entry, &minfo, &einfo, capas);
 891         if (rc) {
 892                 entry->se_inode = NULL;
 893                 iput(inode);
 894                 RETURN(rc);
 895         }
 896
 897         rc = md_intent_getattr_async(ll_i2mdexp(dir), minfo, einfo);
 898         if (!rc) {
 899                 capa_put(capas[0]);
 900                 capa_put(capas[1]);
 901         } else {
 902                 entry->se_inode = NULL;
 903                 iput(inode);
 904                 sa_args_fini(minfo, einfo);
 905         }
 906
 907         RETURN(rc);
 908 }
 909
 910 static void ll_statahead_one(struct dentry *parent, const char *name,
 911                              const int namelen)
 912 {
 913         struct inode             *dir    = parent->d_inode;
 914         struct ll_inode_info     *lli    = ll_i2info(dir);
 915         struct ll_statahead_info *sai    = lli->lli_sai;
 916         struct dentry            *dentry = NULL;
 917         struct ll_sa_entry       *entry;
 918         int                       rc;
 919         ENTRY;
 920
 921         entry = ll_sa_entry_alloc(sai, sai->sai_index, name,namelen);
 922         if (IS_ERR(entry))
 923                 RETURN_EXIT;
 924
 925         dentry = d_lookup(parent, &entry->se_qstr);
 926         if (!dentry) {
 927                 rc = do_sa_lookup(dir, entry);
 928         } else {
 929                 rc = do_sa_revalidate(dir, entry, dentry);
 930                 if (rc == 1 && agl_should_run(sai, dentry->d_inode))
 931                         ll_agl_add(sai, dentry->d_inode, entry->se_index);
 932         }
 933
 934         if (dentry != NULL)
 935                 dput(dentry);
 936
 937         if (rc) {
 938                 sa_entry_post_stat(sai, entry,
 939                                    rc < 0 ? SA_ENTRY_INVA : SA_ENTRY_SUCC);
 940                 if (entry->se_index == sai->sai_index_wait)
 941                         wake_up(&sai->sai_waitq);
 942         } else {
 943                 sai->sai_sent++;
 944         }
 945
 946         sai->sai_index++;
 947         /* drop one refcount on entry by ll_sa_entry_alloc */
 948         ll_sa_entry_put(sai, entry);
 949
 950         EXIT;
 951 }
 952
 953 static int ll_agl_thread(void *arg)
 954 {
 955         struct dentry *parent = (struct dentry *)arg;
 956         struct inode *dir = parent->d_inode;
 957         struct ll_inode_info *plli = ll_i2info(dir);
 958         struct ll_inode_info *clli;
 959         struct ll_sb_info *sbi = ll_i2sbi(dir);
 960         struct ll_statahead_info *sai;
 961         struct ptlrpc_thread *thread;
 962         struct l_wait_info lwi = { 0 };
 963         ENTRY;
 964
 965
 966         sai = ll_sai_get(dir);
 967         thread = &sai->sai_agl_thread;
 968         thread->t_pid = current_pid();
 969         CDEBUG(D_READA, "agl thread started: sai %p, parent %.*s\n",
 970                sai, parent->d_name.len, parent->d_name.name);
 971
 972         atomic_inc(&sbi->ll_agl_total);
 973         spin_lock(&plli->lli_agl_lock);
 974         sai->sai_agl_valid = 1;
 975         if (thread_is_init(thread))
 976                 /* If someone else has changed the thread state
 977                  * (e.g. already changed to SVC_STOPPING), we can't just
 978                  * blindly overwrite that setting. */
 979                 thread_set_flags(thread, SVC_RUNNING);
 980         spin_unlock(&plli->lli_agl_lock);
 981         wake_up(&thread->t_ctl_waitq);
 982
 983         while (1) {
 984                 l_wait_event(thread->t_ctl_waitq,
 985                              !agl_list_empty(sai) ||
 986                              !thread_is_running(thread),
 987                              &lwi);
 988
 989                 if (!thread_is_running(thread))
 990                         break;
 991
 992                 spin_lock(&plli->lli_agl_lock);
 993                 /* The statahead thread maybe help to process AGL entries,
 994                  * so check whether list empty again. */
 995                 if (!agl_list_empty(sai)) {
 996                         clli = agl_first_entry(sai);
 997                         list_del_init(&clli->lli_agl_list);
 998                         spin_unlock(&plli->lli_agl_lock);
 999                         ll_agl_trigger(&clli->lli_vfs_inode, sai);
1000                 } else {
1001                         spin_unlock(&plli->lli_agl_lock);
1002                 }
1003         }
1004
1005         spin_lock(&plli->lli_agl_lock);
1006         sai->sai_agl_valid = 0;
1007         while (!agl_list_empty(sai)) {
1008                 clli = agl_first_entry(sai);
1009                 list_del_init(&clli->lli_agl_list);
1010                 spin_unlock(&plli->lli_agl_lock);
1011                 clli->lli_agl_index = 0;
1012                 iput(&clli->lli_vfs_inode);
1013                 spin_lock(&plli->lli_agl_lock);
1014         }
1015         thread_set_flags(thread, SVC_STOPPED);
1016         spin_unlock(&plli->lli_agl_lock);
1017         wake_up(&thread->t_ctl_waitq);
1018         ll_sai_put(sai);
1019         CDEBUG(D_READA, "agl thread stopped: sai %p, parent %.*s\n",
1020                sai, parent->d_name.len, parent->d_name.name);
1021         RETURN(0);
1022 }
1023
1024 static void ll_start_agl(struct dentry *parent, struct ll_statahead_info *sai)
1025 {
1026         struct ptlrpc_thread *thread = &sai->sai_agl_thread;
1027         struct l_wait_info    lwi    = { 0 };
1028         struct ll_inode_info  *plli;
1029         struct task_struct            *task;
1030         ENTRY;
1031
1032         CDEBUG(D_READA, "start agl thread: sai %p, parent %.*s\n",
1033                sai, parent->d_name.len, parent->d_name.name);
1034
1035         plli = ll_i2info(parent->d_inode);
1036         task = kthread_run(ll_agl_thread, parent,
1037                                "ll_agl_%u", plli->lli_opendir_pid);
1038         if (IS_ERR(task)) {
1039                 CERROR("can't start ll_agl thread, rc: %ld\n", PTR_ERR(task));
1040                 thread_set_flags(thread, SVC_STOPPED);
1041                 RETURN_EXIT;
1042         }
1043
1044         l_wait_event(thread->t_ctl_waitq,
1045                      thread_is_running(thread) || thread_is_stopped(thread),
1046                      &lwi);
1047         EXIT;
1048 }
1049
1050 static int ll_statahead_thread(void *arg)
1051 {
1052         struct dentry *parent = (struct dentry *)arg;
1053         struct inode *dir = parent->d_inode;
1054         struct ll_inode_info *lli = ll_i2info(dir);
1055         struct ll_sb_info *sbi = ll_i2sbi(dir);
1056         struct ll_statahead_info *sai;
1057         struct ptlrpc_thread *thread;
1058         struct ptlrpc_thread *agl_thread;
1059         int first = 0;
1060         struct md_op_data *op_data;
1061         struct ll_dir_chain chain;
1062         struct l_wait_info lwi = { 0 };
1063         struct page *page = NULL;
1064         __u64 pos = 0;
1065         int rc = 0;
1066         ENTRY;
1067
1068         sai = ll_sai_get(dir);
1069         thread = &sai->sai_thread;
1070         agl_thread = &sai->sai_agl_thread;
1071         thread->t_pid = current_pid();
1072         CDEBUG(D_READA, "statahead thread starting: sai %p, parent %.*s\n",
1073                sai, parent->d_name.len, parent->d_name.name);
1074
1075         op_data = ll_prep_md_op_data(NULL, dir, dir, NULL, 0, 0,
1076                                      LUSTRE_OPC_ANY, dir);
1077         if (IS_ERR(op_data))
1078                 GOTO(out, rc = PTR_ERR(op_data));
1079
1080         op_data->op_max_pages = ll_i2sbi(dir)->ll_md_brw_pages;
1081
1082         if (sbi->ll_flags & LL_SBI_AGL_ENABLED)
1083                 ll_start_agl(parent, sai);
1084
1085         atomic_inc(&sbi->ll_sa_total);
1086         spin_lock(&lli->lli_sa_lock);
1087         if (thread_is_init(thread))
1088                 /* If someone else has changed the thread state
1089                  * (e.g. already changed to SVC_STOPPING), we can't just
1090                  * blindly overwrite that setting. */
1091                 thread_set_flags(thread, SVC_RUNNING);
1092         spin_unlock(&lli->lli_sa_lock);
1093         wake_up(&thread->t_ctl_waitq);
1094
1095         ll_dir_chain_init(&chain);
1096         while (pos != MDS_DIR_END_OFF && thread_is_running(thread)) {
1097                 struct lu_dirpage *dp;
1098                 struct lu_dirent  *ent;
1099
1100                 sai->sai_in_readpage = 1;
1101                 page = ll_get_dir_page(dir, op_data, pos, &chain);
1102                 sai->sai_in_readpage = 0;
1103                 if (IS_ERR(page)) {
1104                         rc = PTR_ERR(page);
1105                         CDEBUG(D_READA, "error reading dir "DFID" at "LPU64
1106                                "/"LPU64" opendir_pid = %u: rc = %d\n",
1107                                PFID(ll_inode2fid(dir)), pos, sai->sai_index,
1108                                lli->lli_opendir_pid, rc);
1109                         break;
1110                 }
1111
1112                 dp = page_address(page);
1113                 for (ent = lu_dirent_start(dp);
1114                      ent != NULL && thread_is_running(thread) &&
1115                      !sa_low_hit(sai);
1116                      ent = lu_dirent_next(ent)) {
1117                         __u64 hash;
1118                         int namelen;
1119                         char *name;
1120
1121                         hash = le64_to_cpu(ent->lde_hash);
1122                         if (unlikely(hash < pos))
1123                                 /*
1124                                  * Skip until we find target hash value.
1125                                  */
1126                                 continue;
1127
1128                         namelen = le16_to_cpu(ent->lde_namelen);
1129                         if (unlikely(namelen == 0))
1130                                 /*
1131                                  * Skip dummy record.
1132                                  */
1133                                 continue;
1134
1135                         name = ent->lde_name;
1136                         if (name[0] == '.') {
1137                                 if (namelen == 1) {
1138                                         /*
1139                                          * skip "."
1140                                          */
1141                                         continue;
1142                                 } else if (name[1] == '.' && namelen == 2) {
1143                                         /*
1144                                          * skip ".."
1145                                          */
1146                                         continue;
1147                                 } else if (!sai->sai_ls_all) {
1148                                         /*
1149                                          * skip hidden files.
1150                                          */
1151                                         sai->sai_skip_hidden++;
1152                                         continue;
1153                                 }
1154                         }
1155
1156                         /*
1157                          * don't stat-ahead first entry.
1158                          */
1159                         if (unlikely(++first == 1))
1160                                 continue;
1161
1162                         /* wait for spare statahead window */
1163                         do {
1164                                 l_wait_event(thread->t_ctl_waitq,
1165                                              !sa_sent_full(sai) ||
1166                                              !sa_received_empty(sai) ||
1167                                              !agl_list_empty(sai) ||
1168                                              !thread_is_running(thread),
1169                                              &lwi);
1170
1171                                 ll_post_statahead(sai);
1172                         } while (sa_sent_full(sai) &&
1173                                  thread_is_running(thread));
1174
1175                         ll_statahead_one(parent, name, namelen);
1176                 }
1177
1178                 pos = le64_to_cpu(dp->ldp_hash_end);
1179                 ll_release_page(dir, page,
1180                                 le32_to_cpu(dp->ldp_flags) & LDF_COLLIDE);
1181
1182                 if (sa_low_hit(sai)) {
1183                         rc = -EFAULT;
1184                         atomic_inc(&sbi->ll_sa_wrong);
1185                         CDEBUG(D_READA, "Statahead for dir "DFID" hit "
1186                                "ratio too low: hit/miss "LPU64"/"LPU64
1187                                ", sent/replied "LPU64"/"LPU64", stopping "
1188                                "statahead thread: pid %d\n",
1189                                PFID(&lli->lli_fid), sai->sai_hit,
1190                                sai->sai_miss, sai->sai_sent,
1191                                sai->sai_replied, current_pid());
1192                         break;
1193                 }
1194         }
1195         ll_dir_chain_fini(&chain);
1196         ll_finish_md_op_data(op_data);
1197
1198         if (rc < 0) {
1199                 spin_lock(&lli->lli_sa_lock);
1200                 thread_set_flags(thread, SVC_STOPPING);
1201                 lli->lli_sa_enabled = 0;
1202                 spin_unlock(&lli->lli_sa_lock);
1203         }
1204
1205         /* statahead is finished, but statahead entries need to be cached, wait
1206          * for file release to stop me. */
1207         while (thread_is_running(thread)) {
1208                 l_wait_event(thread->t_ctl_waitq,
1209                              !sa_received_empty(sai) ||
1210                              !agl_list_empty(sai) ||
1211                              !thread_is_running(thread),
1212                              &lwi);
1213
1214                 ll_post_statahead(sai);
1215         }
1216
1217         EXIT;
1218 out:
1219         if (sai->sai_agl_valid) {
1220                 spin_lock(&lli->lli_agl_lock);
1221                 thread_set_flags(agl_thread, SVC_STOPPING);
1222                 spin_unlock(&lli->lli_agl_lock);
1223                 wake_up(&agl_thread->t_ctl_waitq);
1224
1225                 CDEBUG(D_READA, "stop agl thread: sai %p pid %u\n",
1226                        sai, (unsigned int)agl_thread->t_pid);
1227                 l_wait_event(agl_thread->t_ctl_waitq,
1228                              thread_is_stopped(agl_thread),
1229                              &lwi);
1230         } else {
1231                 /* Set agl_thread flags anyway. */
1232                 thread_set_flags(agl_thread, SVC_STOPPED);
1233         }
1234
1235         /* wait for inflight statahead RPCs to finish, and then we can free sai
1236          * safely because statahead RPC will access sai data */
1237         while (sai->sai_sent != sai->sai_replied) {
1238                 /* in case we're not woken up, timeout wait */
1239                 lwi = LWI_TIMEOUT(HZ >> 3, NULL, NULL);
1240                 l_wait_event(thread->t_ctl_waitq,
1241                         sai->sai_sent == sai->sai_replied, &lwi);
1242         }
1243
1244         /* release resources held by received entries. */
1245         ll_post_statahead(sai);
1246
1247         spin_lock(&lli->lli_sa_lock);
1248         thread_set_flags(thread, SVC_STOPPED);
1249         spin_unlock(&lli->lli_sa_lock);
1250
1251         wake_up(&sai->sai_waitq);
1252         wake_up(&thread->t_ctl_waitq);
1253         ll_sai_put(sai);
1254         CDEBUG(D_READA, "statahead thread stopped: sai %p, parent %.*s\n",
1255                sai, parent->d_name.len, parent->d_name.name);
1256         dput(parent);
1257         return rc;
1258 }
1259
1260 /* authorize opened dir handle @key to statahead later */
1261 void ll_authorize_statahead(struct inode *dir, void *key)
1262 {
1263         struct ll_inode_info *lli = ll_i2info(dir);
1264
1265         spin_lock(&lli->lli_sa_lock);
1266         if (lli->lli_opendir_key == NULL && lli->lli_sai == NULL) {
1267                 /*
1268                  * if lli_sai is not NULL, it means previous statahead is not
1269                  * finished yet, we'd better not start a new statahead for now.
1270                  */
1271                 LASSERT(lli->lli_opendir_pid == 0);
1272                 lli->lli_opendir_key = key;
1273                 lli->lli_opendir_pid = current_pid();
1274                 lli->lli_sa_enabled = 1;
1275         }
1276         spin_unlock(&lli->lli_sa_lock);
1277 }
1278
1279 /* deauthorize opened dir handle @key to statahead, but statahead thread may
1280  * still be running, notify it to quit. */
1281 void ll_deauthorize_statahead(struct inode *dir, void *key)
1282 {
1283         struct ll_inode_info *lli = ll_i2info(dir);
1284         struct ll_statahead_info *sai;
1285
1286         LASSERT(lli->lli_opendir_key == key);
1287         LASSERT(lli->lli_opendir_pid != 0);
1288
1289         CDEBUG(D_READA, "deauthorize statahead for "DFID"\n",
1290                 PFID(&lli->lli_fid));
1291
1292         spin_lock(&lli->lli_sa_lock);
1293         lli->lli_opendir_key = NULL;
1294         lli->lli_opendir_pid = 0;
1295         lli->lli_sa_enabled = 0;
1296         sai = lli->lli_sai;
1297         if (sai != NULL && thread_is_running(&sai->sai_thread)) {
1298                 /*
1299                  * statahead thread may not quit yet because it needs to cache
1300                  * stated entries, now it's time to tell it to quit.
1301                  */
1302                 thread_set_flags(&sai->sai_thread, SVC_STOPPING);
1303                 wake_up(&sai->sai_thread.t_ctl_waitq);
1304         }
1305         spin_unlock(&lli->lli_sa_lock);
1306 }
1307
1308 enum {
1309         /**
1310          * not first dirent, or is "."
1311          */
1312         LS_NONE_FIRST_DE = 0,
1313         /**
1314          * the first non-hidden dirent
1315          */
1316         LS_FIRST_DE,
1317         /**
1318          * the first hidden dirent, that is "."
1319          */
1320         LS_FIRST_DOT_DE
1321 };
1322
1323 static int is_first_dirent(struct inode *dir, struct dentry *dentry)
1324 {
1325         struct ll_dir_chain   chain;
1326         struct qstr          *target = &dentry->d_name;
1327         struct md_op_data    *op_data;
1328         int                   dot_de;
1329         struct page          *page = NULL;
1330         int                   rc     = LS_NONE_FIRST_DE;
1331         __u64                 pos = 0;
1332         ENTRY;
1333
1334         op_data = ll_prep_md_op_data(NULL, dir, dir, NULL, 0, 0,
1335                                      LUSTRE_OPC_ANY, dir);
1336         if (IS_ERR(op_data))
1337                 RETURN(PTR_ERR(op_data));
1338         /**
1339          *FIXME choose the start offset of the readdir
1340          */
1341         op_data->op_stripe_offset = 0;
1342         op_data->op_max_pages = ll_i2sbi(dir)->ll_md_brw_pages;
1343
1344         ll_dir_chain_init(&chain);
1345         page = ll_get_dir_page(dir, op_data, 0, &chain);
1346
1347         while (1) {
1348                 struct lu_dirpage *dp;
1349                 struct lu_dirent  *ent;
1350
1351                 if (IS_ERR(page)) {
1352                         struct ll_inode_info *lli = ll_i2info(dir);
1353
1354                         rc = PTR_ERR(page);
1355                         CERROR("%s: reading dir "DFID" at "LPU64
1356                                "opendir_pid = %u : rc = %d\n",
1357                                ll_get_fsname(dir->i_sb, NULL, 0),
1358                                PFID(ll_inode2fid(dir)), pos,
1359                                lli->lli_opendir_pid, rc);
1360                         break;
1361                 }
1362
1363                 dp = page_address(page);
1364                 for (ent = lu_dirent_start(dp); ent != NULL;
1365                      ent = lu_dirent_next(ent)) {
1366                         __u64 hash;
1367                         int namelen;
1368                         char *name;
1369
1370                         hash = le64_to_cpu(ent->lde_hash);
1371                         /* The ll_get_dir_page() can return any page containing
1372                          * the given hash which may be not the start hash. */
1373                         if (unlikely(hash < pos))
1374                                 continue;
1375
1376                         namelen = le16_to_cpu(ent->lde_namelen);
1377                         if (unlikely(namelen == 0))
1378                                 /*
1379                                  * skip dummy record.
1380                                  */
1381                                 continue;
1382
1383                         name = ent->lde_name;
1384                         if (name[0] == '.') {
1385                                 if (namelen == 1)
1386                                         /*
1387                                          * skip "."
1388                                          */
1389                                         continue;
1390                                 else if (name[1] == '.' && namelen == 2)
1391                                         /*
1392                                          * skip ".."
1393                                          */
1394                                         continue;
1395                                 else
1396                                         dot_de = 1;
1397                         } else {
1398                                 dot_de = 0;
1399                         }
1400
1401                         if (dot_de && target->name[0] != '.') {
1402                                 CDEBUG(D_READA, "%.*s skip hidden file %.*s\n",
1403                                        target->len, target->name,
1404                                        namelen, name);
1405                                 continue;
1406                         }
1407
1408                         if (target->len != namelen ||
1409                             memcmp(target->name, name, namelen) != 0)
1410                                 rc = LS_NONE_FIRST_DE;
1411                         else if (!dot_de)
1412                                 rc = LS_FIRST_DE;
1413                         else
1414                                 rc = LS_FIRST_DOT_DE;
1415
1416                         ll_release_page(dir, page, false);
1417                         GOTO(out, rc);
1418                 }
1419                 pos = le64_to_cpu(dp->ldp_hash_end);
1420                 if (pos == MDS_DIR_END_OFF) {
1421                         /*
1422                          * End of directory reached.
1423                          */
1424                         ll_release_page(dir, page, false);
1425                         GOTO(out, rc);
1426                 } else {
1427                         /*
1428                          * chain is exhausted
1429                          * Normal case: continue to the next page.
1430                          */
1431                         ll_release_page(dir, page, le32_to_cpu(dp->ldp_flags) &
1432                                               LDF_COLLIDE);
1433                         page = ll_get_dir_page(dir, op_data, pos, &chain);
1434                 }
1435         }
1436         EXIT;
1437 out:
1438         ll_dir_chain_fini(&chain);
1439         ll_finish_md_op_data(op_data);
1440         return rc;
1441 }
1442
1443 static void
1444 ll_sai_unplug(struct ll_statahead_info *sai, struct ll_sa_entry *entry)
1445 {
1446         if (entry != NULL && entry->se_stat == SA_ENTRY_SUCC) {
1447                 struct ll_sb_info *sbi = ll_i2sbi(sai->sai_inode);
1448
1449                 sai->sai_hit++;
1450                 sai->sai_consecutive_miss = 0;
1451                 sai->sai_max = min(2 * sai->sai_max, sbi->ll_sa_max);
1452         } else {
1453                 sai->sai_miss++;
1454                 sai->sai_consecutive_miss++;
1455         }
1456         ll_sa_entry_fini(sai, entry);
1457         wake_up(&sai->sai_thread.t_ctl_waitq);
1458 }
1459
1460 static int revalidate_statahead_dentry(struct inode *dir,
1461                                         struct ll_statahead_info *sai,
1462                                         struct dentry **dentryp,
1463                                         int only_unplug)
1464 {
1465         struct ll_sa_entry *entry = NULL;
1466         struct l_wait_info lwi = { 0 };
1467         int rc = 0;
1468         ENTRY;
1469
1470         if ((*dentryp)->d_name.name[0] == '.') {
1471                 if (sai->sai_ls_all ||
1472                     sai->sai_miss_hidden >= sai->sai_skip_hidden) {
1473                         /*
1474                          * Hidden dentry is the first one, or statahead
1475                          * thread does not skip so many hidden dentries
1476                          * before "sai_ls_all" enabled as below.
1477                          */
1478                 } else {
1479                         if (!sai->sai_ls_all)
1480                                 /*
1481                                  * It maybe because hidden dentry is not
1482                                  * the first one, "sai_ls_all" was not
1483                                  * set, then "ls -al" missed. Enable
1484                                  * "sai_ls_all" for such case.
1485                                  */
1486                                 sai->sai_ls_all = 1;
1487
1488                         /*
1489                          * Such "getattr" has been skipped before
1490                          * "sai_ls_all" enabled as above.
1491                          */
1492                         sai->sai_miss_hidden++;
1493                         RETURN(-EAGAIN);
1494                 }
1495         }
1496
1497         entry = ll_sa_entry_get_byname(sai, &(*dentryp)->d_name);
1498         if (entry == NULL || only_unplug) {
1499                 ll_sai_unplug(sai, entry);
1500                 RETURN(entry ? 1 : -EAGAIN);
1501         }
1502
1503         /* if statahead is busy in readdir, help it do post-work */
1504         if (!ll_sa_entry_stated(entry) && sai->sai_in_readpage)
1505                 ll_post_statahead(sai);
1506
1507         if (!ll_sa_entry_stated(entry)) {
1508                 sai->sai_index_wait = entry->se_index;
1509                 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(30), NULL,
1510                                         LWI_ON_SIGNAL_NOOP, NULL);
1511                 rc = l_wait_event(sai->sai_waitq,
1512                                 ll_sa_entry_stated(entry) ||
1513                                 thread_is_stopped(&sai->sai_thread),
1514                                 &lwi);
1515                 if (rc < 0) {
1516                         ll_sai_unplug(sai, entry);
1517                         RETURN(-EAGAIN);
1518                 }
1519         }
1520
1521         if (entry->se_stat == SA_ENTRY_SUCC && entry->se_inode != NULL) {
1522                 struct inode *inode = entry->se_inode;
1523                 struct lookup_intent it = { .it_op = IT_GETATTR,
1524                                             .d.lustre.it_lock_handle =
1525                                                 entry->se_handle };
1526                 __u64 bits;
1527
1528                 rc = md_revalidate_lock(ll_i2mdexp(dir), &it,
1529                                         ll_inode2fid(inode), &bits);
1530                 if (rc == 1) {
1531                         if ((*dentryp)->d_inode == NULL) {
1532                                 struct dentry *alias;
1533
1534                                 alias = ll_splice_alias(inode, *dentryp);
1535                                 if (IS_ERR(alias)) {
1536                                         ll_sai_unplug(sai, entry);
1537                                         RETURN(PTR_ERR(alias));
1538                                 }
1539                                 *dentryp = alias;
1540                         } else if ((*dentryp)->d_inode != inode) {
1541                                 /* revalidate, but inode is recreated */
1542                                 CDEBUG(D_READA,
1543                                         "%s: stale dentry %.*s inode "
1544                                         DFID", statahead inode "DFID
1545                                         "\n",
1546                                         ll_get_fsname((*dentryp)->d_inode->i_sb,
1547                                                       NULL, 0),
1548                                         (*dentryp)->d_name.len,
1549                                         (*dentryp)->d_name.name,
1550                                         PFID(ll_inode2fid((*dentryp)->d_inode)),
1551                                         PFID(ll_inode2fid(inode)));
1552                                 ll_sai_unplug(sai, entry);
1553                                 RETURN(-ESTALE);
1554                         } else {
1555                                 iput(inode);
1556                         }
1557                         entry->se_inode = NULL;
1558
1559                         if ((bits & MDS_INODELOCK_LOOKUP) &&
1560                             d_lustre_invalid(*dentryp))
1561                                 d_lustre_revalidate(*dentryp);
1562                         ll_intent_release(&it);
1563                 }
1564         }
1565
1566         ll_sai_unplug(sai, entry);
1567         RETURN(rc);
1568 }
1569
1570 static int start_statahead_thread(struct inode *dir, struct dentry *dentry)
1571 {
1572         struct ll_inode_info *lli = ll_i2info(dir);
1573         struct ll_statahead_info *sai = NULL;
1574         struct dentry *parent;
1575         struct ptlrpc_thread *thread;
1576         struct l_wait_info lwi = { 0 };
1577         struct task_struct *task;
1578         int rc;
1579         ENTRY;
1580
1581         /* I am the "lli_opendir_pid" owner, only me can set "lli_sai". */
1582         rc = is_first_dirent(dir, dentry);
1583         if (rc == LS_NONE_FIRST_DE)
1584                 /* It is not "ls -{a}l" operation, no need statahead for it. */
1585                 GOTO(out, rc = -EAGAIN);
1586
1587         sai = ll_sai_alloc();
1588         if (sai == NULL)
1589                 GOTO(out, rc = -ENOMEM);
1590
1591         sai->sai_ls_all = (rc == LS_FIRST_DOT_DE);
1592         sai->sai_inode = igrab(dir);
1593         if (unlikely(sai->sai_inode == NULL)) {
1594                 CWARN("Do not start stat ahead on dying inode "DFID"\n",
1595                         PFID(&lli->lli_fid));
1596                 GOTO(out, rc = -ESTALE);
1597         }
1598
1599         /* get parent reference count here, and put it in ll_statahead_thread */
1600         parent = dget(dentry->d_parent);
1601         if (unlikely(sai->sai_inode != parent->d_inode)) {
1602                 struct ll_inode_info *nlli = ll_i2info(parent->d_inode);
1603
1604                 CWARN("Race condition, someone changed %.*s just now: "
1605                         "old parent "DFID", new parent "DFID"\n",
1606                         dentry->d_name.len, dentry->d_name.name,
1607                         PFID(&lli->lli_fid), PFID(&nlli->lli_fid));
1608                 dput(parent);
1609                 iput(sai->sai_inode);
1610                 GOTO(out, rc = -EAGAIN);
1611         }
1612
1613         CDEBUG(D_READA, "start statahead thread: sai %p, parent %.*s\n",
1614                sai, parent->d_name.len, parent->d_name.name);
1615
1616         lli->lli_sai = sai;
1617
1618         task = kthread_run(ll_statahead_thread, parent, "ll_sa_%u",
1619                            lli->lli_opendir_pid);
1620         thread = &sai->sai_thread;
1621         if (IS_ERR(task)) {
1622                 rc = PTR_ERR(task);
1623                 CERROR("cannot start ll_sa thread: rc = %d\n", rc);
1624                 dput(parent);
1625                 lli->lli_opendir_key = NULL;
1626                 thread_set_flags(thread, SVC_STOPPED);
1627                 thread_set_flags(&sai->sai_agl_thread, SVC_STOPPED);
1628                 ll_sai_put(sai);
1629                 LASSERT(lli->lli_sai == NULL);
1630                 RETURN(-EAGAIN);
1631         }
1632
1633         l_wait_event(thread->t_ctl_waitq,
1634                      thread_is_running(thread) || thread_is_stopped(thread),
1635                      &lwi);
1636         atomic_inc(&ll_i2sbi(parent->d_inode)->ll_sa_running);
1637         ll_sai_put(sai);
1638
1639         /*
1640          * We don't stat-ahead for the first dirent since we are already in
1641          * lookup.
1642          */
1643         RETURN(-EAGAIN);
1644
1645 out:
1646         if (sai != NULL)
1647                 OBD_FREE_PTR(sai);
1648         spin_lock(&lli->lli_sa_lock);
1649         lli->lli_opendir_key = NULL;
1650         lli->lli_opendir_pid = 0;
1651         lli->lli_sa_enabled = 0;
1652         spin_unlock(&lli->lli_sa_lock);
1653
1654         RETURN(rc);
1655 }
1656
1657 /**
1658  * Start statahead thread if this is the first dir entry.
1659  * Otherwise if a thread is started already, wait it until it is ahead of me.
1660  * \retval 1       -- find entry with lock in cache, the caller needs to do
1661  *                    nothing.
1662  * \retval 0       -- find entry in cache, but without lock, the caller needs
1663  *                    refresh from MDS.
1664  * \retval others  -- the caller need to process as non-statahead.
1665  */
1666 int do_statahead_enter(struct inode *dir, struct dentry **dentryp,
1667                        int only_unplug)
1668 {
1669         struct ll_statahead_info *sai;
1670
1671         sai = ll_sai_get(dir);
1672         if (sai != NULL) {
1673                 int rc;
1674
1675                 rc = revalidate_statahead_dentry(dir, sai, dentryp,
1676                                                  only_unplug);
1677                 CDEBUG(D_READA, "revalidate statahead %.*s: %d.\n",
1678                         (*dentryp)->d_name.len, (*dentryp)->d_name.name, rc);
1679                 ll_sai_put(sai);
1680                 return rc;
1681         }
1682
1683         return start_statahead_thread(dir, *dentryp);
1684 }