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,
 197                   const char *name, 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_cleanup(struct ll_statahead_info *sai,
 307                                  struct ll_sa_entry *entry)
 308 {
 309         struct md_enqueue_info *minfo = entry->se_minfo;
 310         struct ptlrpc_request  *req   = entry->se_req;
 311
 312         if (minfo) {
 313                 entry->se_minfo = NULL;
 314                 ll_intent_release(&minfo->mi_it);
 315                 iput(minfo->mi_dir);
 316                 OBD_FREE_PTR(minfo);
 317         }
 318
 319         if (req) {
 320                 entry->se_req = NULL;
 321                 ptlrpc_req_finished(req);
 322         }
 323 }
 324
 325 static void ll_sa_entry_put(struct ll_statahead_info *sai,
 326                              struct ll_sa_entry *entry)
 327 {
 328         if (atomic_dec_and_test(&entry->se_refcount)) {
 329                 CDEBUG(D_READA, "free sa entry %.*s(%p) index "LPU64"\n",
 330                        entry->se_qstr.len, entry->se_qstr.name, entry,
 331                        entry->se_index);
 332
 333                 LASSERT(list_empty(&entry->se_link));
 334                 LASSERT(list_empty(&entry->se_list));
 335                 LASSERT(ll_sa_entry_unhashed(entry));
 336
 337                 ll_sa_entry_cleanup(sai, entry);
 338                 if (entry->se_inode)
 339                         iput(entry->se_inode);
 340
 341                 OBD_FREE(entry, entry->se_size);
 342                 atomic_dec(&sai->sai_cache_count);
 343         }
 344 }
 345
 346 static inline void
 347 do_sa_entry_fini(struct ll_statahead_info *sai, struct ll_sa_entry *entry)
 348 {
 349         struct ll_inode_info *lli = ll_i2info(sai->sai_inode);
 350
 351         LASSERT(!ll_sa_entry_unhashed(entry));
 352         LASSERT(!list_empty(&entry->se_link));
 353
 354         ll_sa_entry_unhash(sai, entry);
 355
 356         spin_lock(&lli->lli_sa_lock);
 357         entry->se_stat = SA_ENTRY_DEST;
 358         list_del_init(&entry->se_link);
 359         if (likely(!list_empty(&entry->se_list)))
 360                 list_del_init(&entry->se_list);
 361         spin_unlock(&lli->lli_sa_lock);
 362
 363         ll_sa_entry_put(sai, entry);
 364 }
 365
 366 /*
 367  * Delete it from sai_entries_stated list when fini.
 368  */
 369 static void
 370 ll_sa_entry_fini(struct ll_statahead_info *sai, struct ll_sa_entry *entry)
 371 {
 372         struct ll_sa_entry *pos, *next;
 373
 374         if (entry)
 375                 do_sa_entry_fini(sai, entry);
 376
 377         /* drop old entry, only 'scanner' process does this, no need to lock */
 378         list_for_each_entry_safe(pos, next, &sai->sai_entries, se_link) {
 379                 if (!is_omitted_entry(sai, pos->se_index))
 380                         break;
 381                 do_sa_entry_fini(sai, pos);
 382         }
 383 }
 384
 385 /*
 386  * Inside lli_sa_lock.
 387  */
 388 static void
 389 do_sa_entry_to_stated(struct ll_statahead_info *sai,
 390                       struct ll_sa_entry *entry, se_stat_t stat)
 391 {
 392         struct ll_sa_entry *se;
 393         struct list_head   *pos = &sai->sai_entries_stated;
 394
 395         if (!list_empty(&entry->se_list))
 396                 list_del_init(&entry->se_list);
 397
 398         list_for_each_entry_reverse(se, &sai->sai_entries_stated, se_list) {
 399                 if (se->se_index < entry->se_index) {
 400                         pos = &se->se_list;
 401                         break;
 402                 }
 403         }
 404
 405         list_add(&entry->se_list, pos);
 406         entry->se_stat = stat;
 407 }
 408
 409 /*
 410  * Move entry to sai_entries_stated and sort with the index.
 411  * \retval 1    -- entry to be destroyed.
 412  * \retval 0    -- entry is inserted into stated list.
 413  */
 414 static int
 415 ll_sa_entry_to_stated(struct ll_statahead_info *sai,
 416                       struct ll_sa_entry *entry, se_stat_t stat)
 417 {
 418         struct ll_inode_info *lli = ll_i2info(sai->sai_inode);
 419         int                   ret = 1;
 420
 421         ll_sa_entry_cleanup(sai, entry);
 422
 423         spin_lock(&lli->lli_sa_lock);
 424         if (likely(entry->se_stat != SA_ENTRY_DEST)) {
 425                 do_sa_entry_to_stated(sai, entry, stat);
 426                 ret = 0;
 427         }
 428         spin_unlock(&lli->lli_sa_lock);
 429
 430         return ret;
 431 }
 432
 433 /*
 434  * Insert inode into the list of sai_entries_agl.
 435  */
 436 static void ll_agl_add(struct ll_statahead_info *sai,
 437                        struct inode *inode, int index)
 438 {
 439         struct ll_inode_info *child  = ll_i2info(inode);
 440         struct ll_inode_info *parent = ll_i2info(sai->sai_inode);
 441         int                   added  = 0;
 442
 443         spin_lock(&child->lli_agl_lock);
 444         if (child->lli_agl_index == 0) {
 445                 child->lli_agl_index = index;
 446                 spin_unlock(&child->lli_agl_lock);
 447
 448                 LASSERT(list_empty(&child->lli_agl_list));
 449
 450                 igrab(inode);
 451                 spin_lock(&parent->lli_agl_lock);
 452                 if (agl_list_empty(sai))
 453                         added = 1;
 454                 list_add_tail(&child->lli_agl_list, &sai->sai_entries_agl);
 455                 spin_unlock(&parent->lli_agl_lock);
 456         } else {
 457                 spin_unlock(&child->lli_agl_lock);
 458         }
 459
 460         if (added > 0)
 461                 wake_up(&sai->sai_agl_thread.t_ctl_waitq);
 462 }
 463
 464 static struct ll_statahead_info *ll_sai_alloc(void)
 465 {
 466         struct ll_statahead_info *sai;
 467         int                       i;
 468         ENTRY;
 469
 470         OBD_ALLOC_PTR(sai);
 471         if (!sai)
 472                 RETURN(NULL);
 473
 474         atomic_set(&sai->sai_refcount, 1);
 475
 476         spin_lock(&sai_generation_lock);
 477         sai->sai_generation = ++sai_generation;
 478         if (unlikely(sai_generation == 0))
 479                 sai->sai_generation = ++sai_generation;
 480         spin_unlock(&sai_generation_lock);
 481
 482         sai->sai_max = LL_SA_RPC_MIN;
 483         sai->sai_index = 1;
 484         init_waitqueue_head(&sai->sai_waitq);
 485         init_waitqueue_head(&sai->sai_thread.t_ctl_waitq);
 486         init_waitqueue_head(&sai->sai_agl_thread.t_ctl_waitq);
 487
 488         INIT_LIST_HEAD(&sai->sai_entries);
 489         INIT_LIST_HEAD(&sai->sai_entries_received);
 490         INIT_LIST_HEAD(&sai->sai_entries_stated);
 491         INIT_LIST_HEAD(&sai->sai_entries_agl);
 492
 493         for (i = 0; i < LL_SA_CACHE_SIZE; i++) {
 494                 INIT_LIST_HEAD(&sai->sai_cache[i]);
 495                 spin_lock_init(&sai->sai_cache_lock[i]);
 496         }
 497         atomic_set(&sai->sai_cache_count, 0);
 498
 499         RETURN(sai);
 500 }
 501
 502 static inline struct ll_statahead_info *
 503 ll_sai_get(struct ll_statahead_info *sai)
 504 {
 505         atomic_inc(&sai->sai_refcount);
 506         return sai;
 507 }
 508
 509 static void ll_sai_put(struct ll_statahead_info *sai)
 510 {
 511         struct inode         *inode = sai->sai_inode;
 512         struct ll_inode_info *lli   = ll_i2info(inode);
 513         ENTRY;
 514
 515         if (atomic_dec_and_lock(&sai->sai_refcount, &lli->lli_sa_lock)) {
 516                 struct ll_sa_entry *entry, *next;
 517
 518                 if (unlikely(atomic_read(&sai->sai_refcount) > 0)) {
 519                         /* It is race case, the interpret callback just hold
 520                          * a reference count */
 521                         spin_unlock(&lli->lli_sa_lock);
 522                         RETURN_EXIT;
 523                 }
 524
 525                 LASSERT(lli->lli_opendir_key == NULL);
 526                 LASSERT(thread_is_stopped(&sai->sai_thread));
 527                 LASSERT(thread_is_stopped(&sai->sai_agl_thread));
 528
 529                 lli->lli_sai = NULL;
 530                 lli->lli_opendir_pid = 0;
 531                 spin_unlock(&lli->lli_sa_lock);
 532
 533                 if (sai->sai_sent > sai->sai_replied)
 534                         CDEBUG(D_READA,"statahead for dir "DFID" does not "
 535                               "finish: [sent:"LPU64"] [replied:"LPU64"]\n",
 536                               PFID(&lli->lli_fid),
 537                               sai->sai_sent, sai->sai_replied);
 538
 539                 list_for_each_entry_safe(entry, next,
 540                                          &sai->sai_entries, se_link)
 541                         do_sa_entry_fini(sai, entry);
 542
 543                 LASSERT(list_empty(&sai->sai_entries));
 544                 LASSERT(sa_received_empty(sai));
 545                 LASSERT(list_empty(&sai->sai_entries_stated));
 546
 547                 LASSERT(atomic_read(&sai->sai_cache_count) == 0);
 548                 LASSERT(agl_list_empty(sai));
 549
 550                 iput(inode);
 551                 OBD_FREE_PTR(sai);
 552         }
 553
 554         EXIT;
 555 }
 556
 557 /* Do NOT forget to drop inode refcount when into sai_entries_agl. */
 558 static void ll_agl_trigger(struct inode *inode, struct ll_statahead_info *sai)
 559 {
 560         struct ll_inode_info *lli   = ll_i2info(inode);
 561         __u64                 index = lli->lli_agl_index;
 562         int                   rc;
 563         ENTRY;
 564
 565         LASSERT(list_empty(&lli->lli_agl_list));
 566
 567         /* AGL maybe fall behind statahead with one entry */
 568         if (is_omitted_entry(sai, index + 1)) {
 569                 lli->lli_agl_index = 0;
 570                 iput(inode);
 571                 RETURN_EXIT;
 572         }
 573
 574         /* Someone is in glimpse (sync or async), do nothing. */
 575         rc = down_write_trylock(&lli->lli_glimpse_sem);
 576         if (rc == 0) {
 577                 lli->lli_agl_index = 0;
 578                 iput(inode);
 579                 RETURN_EXIT;
 580         }
 581
 582         /*
 583          * Someone triggered glimpse within 1 sec before.
 584          * 1) The former glimpse succeeded with glimpse lock granted by OST, and
 585          *    if the lock is still cached on client, AGL needs to do nothing. If
 586          *    it is cancelled by other client, AGL maybe cannot obtaion new lock
 587          *    for no glimpse callback triggered by AGL.
 588          * 2) The former glimpse succeeded, but OST did not grant glimpse lock.
 589          *    Under such case, it is quite possible that the OST will not grant
 590          *    glimpse lock for AGL also.
 591          * 3) The former glimpse failed, compared with other two cases, it is
 592          *    relative rare. AGL can ignore such case, and it will not muchly
 593          *    affect the performance.
 594          */
 595         if (lli->lli_glimpse_time != 0 &&
 596             cfs_time_before(cfs_time_shift(-1), lli->lli_glimpse_time)) {
 597                 up_write(&lli->lli_glimpse_sem);
 598                 lli->lli_agl_index = 0;
 599                 iput(inode);
 600                 RETURN_EXIT;
 601         }
 602
 603         CDEBUG(D_READA, "Handling (init) async glimpse: inode = "
 604                DFID", idx = "LPU64"\n", PFID(&lli->lli_fid), index);
 605
 606         cl_agl(inode);
 607         lli->lli_agl_index = 0;
 608         lli->lli_glimpse_time = cfs_time_current();
 609         up_write(&lli->lli_glimpse_sem);
 610
 611         CDEBUG(D_READA, "Handled (init) async glimpse: inode= "
 612                DFID", idx = "LPU64", rc = %d\n",
 613                PFID(&lli->lli_fid), index, rc);
 614
 615         iput(inode);
 616
 617         EXIT;
 618 }
 619
 620 static void ll_post_statahead(struct ll_statahead_info *sai)
 621 {
 622         struct inode           *dir   = sai->sai_inode;
 623         struct inode           *child;
 624         struct ll_inode_info   *lli   = ll_i2info(dir);
 625         struct ll_sa_entry     *entry;
 626         struct md_enqueue_info *minfo;
 627         struct lookup_intent   *it;
 628         struct ptlrpc_request  *req;
 629         struct mdt_body        *body;
 630         int                     rc    = 0;
 631         ENTRY;
 632
 633         spin_lock(&lli->lli_sa_lock);
 634         if (unlikely(sa_received_empty(sai))) {
 635                 spin_unlock(&lli->lli_sa_lock);
 636                 RETURN_EXIT;
 637         }
 638         entry = sa_first_received_entry(sai);
 639         atomic_inc(&entry->se_refcount);
 640         list_del_init(&entry->se_list);
 641         spin_unlock(&lli->lli_sa_lock);
 642
 643         LASSERT(entry->se_handle != 0);
 644
 645         minfo = entry->se_minfo;
 646         it = &minfo->mi_it;
 647         req = entry->se_req;
 648         body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
 649         if (body == NULL)
 650                 GOTO(out, rc = -EFAULT);
 651
 652         child = entry->se_inode;
 653         if (child == NULL) {
 654                 /*
 655                  * lookup.
 656                  */
 657                 LASSERT(fid_is_zero(&minfo->mi_data.op_fid2));
 658
 659                 /* XXX: No fid in reply, this is probaly cross-ref case.
 660                  * SA can't handle it yet. */
 661                 if (body->valid & OBD_MD_MDS)
 662                         GOTO(out, rc = -EAGAIN);
 663         } else {
 664                 /*
 665                  * revalidate.
 666                  */
 667                 /* unlinked and re-created with the same name */
 668                 if (unlikely(!lu_fid_eq(&minfo->mi_data.op_fid2, &body->fid1))){
 669                         entry->se_inode = NULL;
 670                         iput(child);
 671                         child = NULL;
 672                 }
 673         }
 674
 675         it->d.lustre.it_lock_handle = entry->se_handle;
 676         rc = md_revalidate_lock(ll_i2mdexp(dir), it, ll_inode2fid(dir), NULL);
 677         if (rc != 1)
 678                 GOTO(out, rc = -EAGAIN);
 679
 680         rc = ll_prep_inode(&child, req, dir->i_sb, it);
 681         if (rc)
 682                 GOTO(out, rc);
 683
 684         CDEBUG(D_DLMTRACE, "%s: setting l_data to inode "DFID"(%p)\n",
 685                ll_get_fsname(child->i_sb, NULL, 0),
 686                PFID(ll_inode2fid(child)), child);
 687         ll_set_lock_data(ll_i2sbi(dir)->ll_md_exp, child, it, NULL);
 688
 689         entry->se_inode = child;
 690
 691         if (agl_should_run(sai, child))
 692                 ll_agl_add(sai, child, entry->se_index);
 693
 694         EXIT;
 695
 696 out:
 697         /* The "ll_sa_entry_to_stated()" will drop related ldlm ibits lock
 698          * reference count by calling "ll_intent_drop_lock()" in spite of the
 699          * above operations failed or not. Do not worry about calling
 700          * "ll_intent_drop_lock()" more than once. */
 701         rc = ll_sa_entry_to_stated(sai, entry,
 702                                    rc < 0 ? SA_ENTRY_INVA : SA_ENTRY_SUCC);
 703         if (rc == 0 && entry->se_index == sai->sai_index_wait)
 704                 wake_up(&sai->sai_waitq);
 705         ll_sa_entry_put(sai, entry);
 706 }
 707
 708 static int ll_statahead_interpret(struct ptlrpc_request *req,
 709                                   struct md_enqueue_info *minfo, int rc)
 710 {
 711         struct lookup_intent     *it  = &minfo->mi_it;
 712         struct inode             *dir = minfo->mi_dir;
 713         struct ll_inode_info     *lli = ll_i2info(dir);
 714         struct ll_statahead_info *sai = NULL;
 715         struct ll_sa_entry       *entry;
 716         int                       wakeup;
 717         ENTRY;
 718
 719         if (it_disposition(it, DISP_LOOKUP_NEG))
 720                 rc = -ENOENT;
 721
 722         spin_lock(&lli->lli_sa_lock);
 723         /* stale entry */
 724         if (unlikely(lli->lli_sai == NULL ||
 725                      lli->lli_sai->sai_generation != minfo->mi_generation)) {
 726                 spin_unlock(&lli->lli_sa_lock);
 727                 GOTO(out, rc = -ESTALE);
 728         } else {
 729                 sai = ll_sai_get(lli->lli_sai);
 730                 if (unlikely(!thread_is_running(&sai->sai_thread))) {
 731                         sai->sai_replied++;
 732                         spin_unlock(&lli->lli_sa_lock);
 733                         GOTO(out, rc = -EBADFD);
 734                 }
 735
 736                 entry = ll_sa_entry_get_byindex(sai, minfo->mi_cbdata);
 737                 if (entry == NULL) {
 738                         sai->sai_replied++;
 739                         spin_unlock(&lli->lli_sa_lock);
 740                         GOTO(out, rc = -EIDRM);
 741                 }
 742
 743                 if (rc != 0) {
 744                         do_sa_entry_to_stated(sai, entry, SA_ENTRY_INVA);
 745                         wakeup = (entry->se_index == sai->sai_index_wait);
 746                 } else {
 747                         entry->se_minfo = minfo;
 748                         entry->se_req = ptlrpc_request_addref(req);
 749                         /* Release the async ibits lock ASAP to avoid deadlock
 750                          * when statahead thread tries to enqueue lock on parent
 751                          * for readpage and other tries to enqueue lock on child
 752                          * with parent's lock held, for example: unlink. */
 753                         entry->se_handle = it->d.lustre.it_lock_handle;
 754                         ll_intent_drop_lock(it);
 755                         wakeup = sa_received_empty(sai);
 756                         list_add_tail(&entry->se_list,
 757                                           &sai->sai_entries_received);
 758                 }
 759                 sai->sai_replied++;
 760                 spin_unlock(&lli->lli_sa_lock);
 761
 762                 ll_sa_entry_put(sai, entry);
 763                 if (wakeup)
 764                         wake_up(&sai->sai_thread.t_ctl_waitq);
 765         }
 766
 767         EXIT;
 768
 769 out:
 770         if (rc != 0) {
 771                 ll_intent_release(it);
 772                 iput(dir);
 773                 OBD_FREE_PTR(minfo);
 774         }
 775         if (sai != NULL)
 776                 ll_sai_put(sai);
 777         return rc;
 778 }
 779
 780 static void sa_args_fini(struct md_enqueue_info *minfo,
 781                          struct ldlm_enqueue_info *einfo)
 782 {
 783         LASSERT(minfo && einfo);
 784         iput(minfo->mi_dir);
 785         capa_put(minfo->mi_data.op_capa1);
 786         capa_put(minfo->mi_data.op_capa2);
 787         OBD_FREE_PTR(minfo);
 788         OBD_FREE_PTR(einfo);
 789 }
 790
 791 /**
 792  * There is race condition between "capa_put" and "ll_statahead_interpret" for
 793  * accessing "op_data.op_capa[1,2]" as following:
 794  * "capa_put" releases "op_data.op_capa[1,2]"'s reference count after calling
 795  * "md_intent_getattr_async". But "ll_statahead_interpret" maybe run first, and
 796  * fill "op_data.op_capa[1,2]" as POISON, then cause "capa_put" access invalid
 797  * "ocapa". So here reserve "op_data.op_capa[1,2]" in "pcapa" before calling
 798  * "md_intent_getattr_async".
 799  */
 800 static int sa_args_init(struct inode *dir, struct inode *child,
 801                         struct ll_sa_entry *entry, struct md_enqueue_info **pmi,
 802                         struct ldlm_enqueue_info **pei,
 803                         struct obd_capa **pcapa)
 804 {
 805         struct qstr              *qstr = &entry->se_qstr;
 806         struct ll_inode_info     *lli  = ll_i2info(dir);
 807         struct md_enqueue_info   *minfo;
 808         struct ldlm_enqueue_info *einfo;
 809         struct md_op_data        *op_data;
 810
 811         OBD_ALLOC_PTR(einfo);
 812         if (einfo == NULL)
 813                 return -ENOMEM;
 814
 815         OBD_ALLOC_PTR(minfo);
 816         if (minfo == NULL) {
 817                 OBD_FREE_PTR(einfo);
 818                 return -ENOMEM;
 819         }
 820
 821         op_data = ll_prep_md_op_data(&minfo->mi_data, dir, child, qstr->name,
 822                                      qstr->len, 0, LUSTRE_OPC_ANY, NULL);
 823         if (IS_ERR(op_data)) {
 824                 OBD_FREE_PTR(einfo);
 825                 OBD_FREE_PTR(minfo);
 826                 return PTR_ERR(op_data);
 827         }
 828
 829         minfo->mi_it.it_op = IT_GETATTR;
 830         minfo->mi_dir = igrab(dir);
 831         minfo->mi_cb = ll_statahead_interpret;
 832         minfo->mi_generation = lli->lli_sai->sai_generation;
 833         minfo->mi_cbdata = entry->se_index;
 834
 835         einfo->ei_type   = LDLM_IBITS;
 836         einfo->ei_mode   = it_to_lock_mode(&minfo->mi_it);
 837         einfo->ei_cb_bl  = ll_md_blocking_ast;
 838         einfo->ei_cb_cp  = ldlm_completion_ast;
 839         einfo->ei_cb_gl  = NULL;
 840         einfo->ei_cbdata = NULL;
 841
 842         *pmi = minfo;
 843         *pei = einfo;
 844         pcapa[0] = op_data->op_capa1;
 845         pcapa[1] = op_data->op_capa2;
 846
 847         return 0;
 848 }
 849
 850 static int do_sa_lookup(struct inode *dir, struct ll_sa_entry *entry)
 851 {
 852         struct md_enqueue_info   *minfo;
 853         struct ldlm_enqueue_info *einfo;
 854         struct obd_capa          *capas[2];
 855         int                       rc;
 856         ENTRY;
 857
 858         rc = sa_args_init(dir, NULL, entry, &minfo, &einfo, capas);
 859         if (rc)
 860                 RETURN(rc);
 861
 862         rc = md_intent_getattr_async(ll_i2mdexp(dir), minfo, einfo);
 863         if (!rc) {
 864                 capa_put(capas[0]);
 865                 capa_put(capas[1]);
 866         } else {
 867                 sa_args_fini(minfo, einfo);
 868         }
 869
 870         RETURN(rc);
 871 }
 872
 873 /**
 874  * similar to ll_revalidate_it().
 875  * \retval      1 -- dentry valid
 876  * \retval      0 -- will send stat-ahead request
 877  * \retval others -- prepare stat-ahead request failed
 878  */
 879 static int do_sa_revalidate(struct inode *dir, struct ll_sa_entry *entry,
 880                             struct dentry *dentry)
 881 {
 882         struct inode             *inode = dentry->d_inode;
 883         struct lookup_intent      it = { .it_op = IT_GETATTR,
 884                                          .d.lustre.it_lock_handle = 0 };
 885         struct md_enqueue_info   *minfo;
 886         struct ldlm_enqueue_info *einfo;
 887         struct obd_capa          *capas[2];
 888         int rc;
 889         ENTRY;
 890
 891         if (unlikely(inode == NULL))
 892                 RETURN(1);
 893
 894         if (d_mountpoint(dentry))
 895                 RETURN(1);
 896
 897         entry->se_inode = igrab(inode);
 898         rc = md_revalidate_lock(ll_i2mdexp(dir), &it, ll_inode2fid(inode),NULL);
 899         if (rc == 1) {
 900                 entry->se_handle = it.d.lustre.it_lock_handle;
 901                 ll_intent_release(&it);
 902                 RETURN(1);
 903         }
 904
 905         rc = sa_args_init(dir, inode, entry, &minfo, &einfo, capas);
 906         if (rc) {
 907                 entry->se_inode = NULL;
 908                 iput(inode);
 909                 RETURN(rc);
 910         }
 911
 912         rc = md_intent_getattr_async(ll_i2mdexp(dir), minfo, einfo);
 913         if (!rc) {
 914                 capa_put(capas[0]);
 915                 capa_put(capas[1]);
 916         } else {
 917                 entry->se_inode = NULL;
 918                 iput(inode);
 919                 sa_args_fini(minfo, einfo);
 920         }
 921
 922         RETURN(rc);
 923 }
 924
 925 static void ll_statahead_one(struct dentry *parent, const char* entry_name,
 926                              int entry_name_len)
 927 {
 928         struct inode             *dir    = parent->d_inode;
 929         struct ll_inode_info     *lli    = ll_i2info(dir);
 930         struct ll_statahead_info *sai    = lli->lli_sai;
 931         struct dentry            *dentry = NULL;
 932         struct ll_sa_entry       *entry;
 933         int                       rc;
 934         int                       rc1;
 935         ENTRY;
 936
 937         entry = ll_sa_entry_alloc(sai, sai->sai_index, entry_name,
 938                                   entry_name_len);
 939         if (IS_ERR(entry))
 940                 RETURN_EXIT;
 941
 942         dentry = d_lookup(parent, &entry->se_qstr);
 943         if (!dentry) {
 944                 rc = do_sa_lookup(dir, entry);
 945         } else {
 946                 rc = do_sa_revalidate(dir, entry, dentry);
 947                 if (rc == 1 && agl_should_run(sai, dentry->d_inode))
 948                         ll_agl_add(sai, dentry->d_inode, entry->se_index);
 949         }
 950
 951         if (dentry != NULL)
 952                 dput(dentry);
 953
 954         if (rc) {
 955                 rc1 = ll_sa_entry_to_stated(sai, entry,
 956                                         rc < 0 ? SA_ENTRY_INVA : SA_ENTRY_SUCC);
 957                 if (rc1 == 0 && entry->se_index == sai->sai_index_wait)
 958                         wake_up(&sai->sai_waitq);
 959         } else {
 960                 sai->sai_sent++;
 961         }
 962
 963         sai->sai_index++;
 964         /* drop one refcount on entry by ll_sa_entry_alloc */
 965         ll_sa_entry_put(sai, entry);
 966
 967         EXIT;
 968 }
 969
 970 static int ll_agl_thread(void *arg)
 971 {
 972         struct dentry            *parent = (struct dentry *)arg;
 973         struct inode             *dir    = parent->d_inode;
 974         struct ll_inode_info     *plli   = ll_i2info(dir);
 975         struct ll_inode_info     *clli;
 976         struct ll_sb_info        *sbi    = ll_i2sbi(dir);
 977         struct ll_statahead_info *sai    = ll_sai_get(plli->lli_sai);
 978         struct ptlrpc_thread     *thread = &sai->sai_agl_thread;
 979         struct l_wait_info        lwi    = { 0 };
 980         ENTRY;
 981
 982         thread->t_pid = current_pid();
 983         CDEBUG(D_READA, "agl thread started: sai %p, parent %.*s\n",
 984                sai, parent->d_name.len, parent->d_name.name);
 985
 986         atomic_inc(&sbi->ll_agl_total);
 987         spin_lock(&plli->lli_agl_lock);
 988         sai->sai_agl_valid = 1;
 989         if (thread_is_init(thread))
 990                 /* If someone else has changed the thread state
 991                  * (e.g. already changed to SVC_STOPPING), we can't just
 992                  * blindly overwrite that setting. */
 993                 thread_set_flags(thread, SVC_RUNNING);
 994         spin_unlock(&plli->lli_agl_lock);
 995         wake_up(&thread->t_ctl_waitq);
 996
 997         while (1) {
 998                 l_wait_event(thread->t_ctl_waitq,
 999                              !agl_list_empty(sai) ||
1000                              !thread_is_running(thread),
1001                              &lwi);
1002
1003                 if (!thread_is_running(thread))
1004                         break;
1005
1006                 spin_lock(&plli->lli_agl_lock);
1007                 /* The statahead thread maybe help to process AGL entries,
1008                  * so check whether list empty again. */
1009                 if (!agl_list_empty(sai)) {
1010                         clli = agl_first_entry(sai);
1011                         list_del_init(&clli->lli_agl_list);
1012                         spin_unlock(&plli->lli_agl_lock);
1013                         ll_agl_trigger(&clli->lli_vfs_inode, sai);
1014                 } else {
1015                         spin_unlock(&plli->lli_agl_lock);
1016                 }
1017         }
1018
1019         spin_lock(&plli->lli_agl_lock);
1020         sai->sai_agl_valid = 0;
1021         while (!agl_list_empty(sai)) {
1022                 clli = agl_first_entry(sai);
1023                 list_del_init(&clli->lli_agl_list);
1024                 spin_unlock(&plli->lli_agl_lock);
1025                 clli->lli_agl_index = 0;
1026                 iput(&clli->lli_vfs_inode);
1027                 spin_lock(&plli->lli_agl_lock);
1028         }
1029         thread_set_flags(thread, SVC_STOPPED);
1030         spin_unlock(&plli->lli_agl_lock);
1031         wake_up(&thread->t_ctl_waitq);
1032         ll_sai_put(sai);
1033         CDEBUG(D_READA, "agl thread stopped: sai %p, parent %.*s\n",
1034                sai, parent->d_name.len, parent->d_name.name);
1035         RETURN(0);
1036 }
1037
1038 static void ll_start_agl(struct dentry *parent, struct ll_statahead_info *sai)
1039 {
1040         struct ptlrpc_thread *thread = &sai->sai_agl_thread;
1041         struct l_wait_info    lwi    = { 0 };
1042         struct ll_inode_info  *plli;
1043         struct task_struct            *task;
1044         ENTRY;
1045
1046         CDEBUG(D_READA, "start agl thread: sai %p, parent %.*s\n",
1047                sai, parent->d_name.len, parent->d_name.name);
1048
1049         plli = ll_i2info(parent->d_inode);
1050         task = kthread_run(ll_agl_thread, parent,
1051                                "ll_agl_%u", plli->lli_opendir_pid);
1052         if (IS_ERR(task)) {
1053                 CERROR("can't start ll_agl thread, rc: %ld\n", PTR_ERR(task));
1054                 thread_set_flags(thread, SVC_STOPPED);
1055                 RETURN_EXIT;
1056         }
1057
1058         l_wait_event(thread->t_ctl_waitq,
1059                      thread_is_running(thread) || thread_is_stopped(thread),
1060                      &lwi);
1061         EXIT;
1062 }
1063
1064 static int ll_statahead_thread(void *arg)
1065 {
1066         struct dentry            *parent = (struct dentry *)arg;
1067         struct inode             *dir    = parent->d_inode;
1068         struct ll_inode_info     *plli   = ll_i2info(dir);
1069         struct ll_inode_info     *clli;
1070         struct ll_sb_info        *sbi    = ll_i2sbi(dir);
1071         struct ll_statahead_info *sai    = ll_sai_get(plli->lli_sai);
1072         struct ptlrpc_thread     *thread = &sai->sai_thread;
1073         struct ptlrpc_thread *agl_thread = &sai->sai_agl_thread;
1074         int                       first  = 0;
1075         int                       rc     = 0;
1076         struct md_op_data        *op_data;
1077         struct ll_dir_chain       chain;
1078         struct l_wait_info        lwi    = { 0 };
1079         struct lu_dirent         *ent;
1080         struct page             *page = NULL;
1081         ENTRY;
1082
1083         thread->t_pid = current_pid();
1084         CDEBUG(D_READA, "statahead thread starting: sai %p, parent %.*s\n",
1085                sai, parent->d_name.len, parent->d_name.name);
1086
1087         op_data = ll_prep_md_op_data(NULL, dir, dir, NULL, 0, 0,
1088                                      LUSTRE_OPC_ANY, dir);
1089         if (IS_ERR(op_data))
1090                 RETURN(PTR_ERR(op_data));
1091
1092         op_data->op_hash_offset = 0;
1093         op_data->op_max_pages = ll_i2sbi(dir)->ll_md_brw_pages;
1094
1095         if (sbi->ll_flags & LL_SBI_AGL_ENABLED)
1096                 ll_start_agl(parent, sai);
1097
1098         atomic_inc(&sbi->ll_sa_total);
1099         spin_lock(&plli->lli_sa_lock);
1100         if (thread_is_init(thread))
1101                 /* If someone else has changed the thread state
1102                  * (e.g. already changed to SVC_STOPPING), we can't just
1103                  * blindly overwrite that setting. */
1104                 thread_set_flags(thread, SVC_RUNNING);
1105         spin_unlock(&plli->lli_sa_lock);
1106         wake_up(&thread->t_ctl_waitq);
1107
1108         ll_dir_chain_init(&chain);
1109         for (ent = ll_dir_entry_start(dir, op_data, &page);
1110              ent != NULL && !IS_ERR(ent);
1111              ent = ll_dir_entry_next(dir, op_data, ent, &page)) {
1112                 __u64 hash;
1113                 int namelen;
1114                 char *name;
1115
1116                 hash = le64_to_cpu(ent->lde_hash);
1117                 if (unlikely(hash < op_data->op_hash_offset))
1118                         /*
1119                          * Skip until we find target hash value.
1120                          */
1121                         continue;
1122
1123                 namelen = le16_to_cpu(ent->lde_namelen);
1124                 if (unlikely(namelen == 0))
1125                         /*
1126                          * Skip dummy record.
1127                          */
1128                         continue;
1129
1130                 name = ent->lde_name;
1131                 if (name[0] == '.') {
1132                         if (namelen == 1) {
1133                                 /*
1134                                  * skip "."
1135                                  */
1136                                 continue;
1137                         } else if (name[1] == '.' && namelen == 2) {
1138                                 /*
1139                                  * skip ".."
1140                                  */
1141                                 continue;
1142                         } else if (!sai->sai_ls_all) {
1143                                 /*
1144                                  * skip hidden files.
1145                                  */
1146                                 sai->sai_skip_hidden++;
1147                                 continue;
1148                         }
1149                 }
1150
1151                 /*
1152                  * don't stat-ahead first entry.
1153                  */
1154                 if (unlikely(++first == 1))
1155                         continue;
1156
1157 keep_it:
1158                 l_wait_event(thread->t_ctl_waitq,
1159                              !sa_sent_full(sai) ||
1160                              !sa_received_empty(sai) ||
1161                              !agl_list_empty(sai) ||
1162                              !thread_is_running(thread),
1163                              &lwi);
1164
1165 interpret_it:
1166                 while (!sa_received_empty(sai))
1167                         ll_post_statahead(sai);
1168
1169                 if (unlikely(!thread_is_running(thread)))
1170                         GOTO(out, rc = 0);
1171
1172                 /* If no window for metadata statahead, but there are
1173                  * some AGL entries to be triggered, then try to help
1174                  * to process the AGL entries. */
1175                 if (sa_sent_full(sai)) {
1176                         spin_lock(&plli->lli_agl_lock);
1177                         while (!agl_list_empty(sai)) {
1178                                 clli = agl_first_entry(sai);
1179                                 list_del_init(&clli->lli_agl_list);
1180                                 spin_unlock(&plli->lli_agl_lock);
1181                                 ll_agl_trigger(&clli->lli_vfs_inode,
1182                                                sai);
1183
1184                                 if (!sa_received_empty(sai))
1185                                         goto interpret_it;
1186
1187                                 if (unlikely(
1188                                         !thread_is_running(thread)))
1189                                         GOTO(out, rc = 0);
1190
1191                                 if (!sa_sent_full(sai))
1192                                         goto do_it;
1193
1194                                 spin_lock(&plli->lli_agl_lock);
1195                         }
1196                         spin_unlock(&plli->lli_agl_lock);
1197
1198                         goto keep_it;
1199                 }
1200
1201 do_it:
1202                 ll_statahead_one(parent, name, namelen);
1203         }
1204
1205         if (page != NULL) {
1206                 kunmap(page);
1207                 page_cache_release(page);
1208         }
1209
1210          /*
1211          * End of directory reached.
1212          */
1213         while (1) {
1214                 l_wait_event(thread->t_ctl_waitq,
1215                              !sa_received_empty(sai) ||
1216                              sai->sai_sent == sai->sai_replied ||
1217                              !thread_is_running(thread),
1218                              &lwi);
1219
1220                 while (!sa_received_empty(sai))
1221                         ll_post_statahead(sai);
1222
1223                 if (unlikely(!thread_is_running(thread)))
1224                         GOTO(out, rc = 0);
1225
1226                 if (sai->sai_sent == sai->sai_replied &&
1227                     sa_received_empty(sai))
1228                         break;
1229         }
1230
1231         spin_lock(&plli->lli_agl_lock);
1232         while (!agl_list_empty(sai) &&
1233                thread_is_running(thread)) {
1234                 clli = agl_first_entry(sai);
1235                 list_del_init(&clli->lli_agl_list);
1236                 spin_unlock(&plli->lli_agl_lock);
1237                 ll_agl_trigger(&clli->lli_vfs_inode, sai);
1238                 spin_lock(&plli->lli_agl_lock);
1239         }
1240         spin_unlock(&plli->lli_agl_lock);
1241 out:
1242         EXIT;
1243         ll_finish_md_op_data(op_data);
1244         if (sai->sai_agl_valid) {
1245                 spin_lock(&plli->lli_agl_lock);
1246                 thread_set_flags(agl_thread, SVC_STOPPING);
1247                 spin_unlock(&plli->lli_agl_lock);
1248                 wake_up(&agl_thread->t_ctl_waitq);
1249
1250                 CDEBUG(D_READA, "stop agl thread: sai %p pid %u\n",
1251                        sai, (unsigned int)agl_thread->t_pid);
1252                 l_wait_event(agl_thread->t_ctl_waitq,
1253                              thread_is_stopped(agl_thread),
1254                              &lwi);
1255         } else {
1256                 /* Set agl_thread flags anyway. */
1257                 thread_set_flags(&sai->sai_agl_thread, SVC_STOPPED);
1258         }
1259         ll_dir_chain_fini(&chain);
1260         spin_lock(&plli->lli_sa_lock);
1261         if (!sa_received_empty(sai)) {
1262                 thread_set_flags(thread, SVC_STOPPING);
1263                 spin_unlock(&plli->lli_sa_lock);
1264
1265                 /* To release the resources held by received entries. */
1266                 while (!sa_received_empty(sai))
1267                         ll_post_statahead(sai);
1268
1269                 spin_lock(&plli->lli_sa_lock);
1270         }
1271         thread_set_flags(thread, SVC_STOPPED);
1272         spin_unlock(&plli->lli_sa_lock);
1273         wake_up(&sai->sai_waitq);
1274         wake_up(&thread->t_ctl_waitq);
1275         ll_sai_put(sai);
1276         dput(parent);
1277         CDEBUG(D_READA, "statahead thread stopped: sai %p, parent %.*s\n",
1278                sai, parent->d_name.len, parent->d_name.name);
1279         return rc;
1280 }
1281
1282 /**
1283  * called in ll_file_release().
1284  */
1285 void ll_stop_statahead(struct inode *dir, void *key)
1286 {
1287         struct ll_inode_info *lli = ll_i2info(dir);
1288
1289         if (unlikely(key == NULL))
1290                 return;
1291
1292         spin_lock(&lli->lli_sa_lock);
1293         if (lli->lli_opendir_key != key || lli->lli_opendir_pid == 0) {
1294                 spin_unlock(&lli->lli_sa_lock);
1295                 return;
1296         }
1297
1298         lli->lli_opendir_key = NULL;
1299
1300         if (lli->lli_sai) {
1301                 struct l_wait_info lwi = { 0 };
1302                 struct ptlrpc_thread *thread = &lli->lli_sai->sai_thread;
1303
1304                 if (!thread_is_stopped(thread)) {
1305                         thread_set_flags(thread, SVC_STOPPING);
1306                         spin_unlock(&lli->lli_sa_lock);
1307                         wake_up(&thread->t_ctl_waitq);
1308
1309                         CDEBUG(D_READA, "stop statahead thread: sai %p pid %u\n",
1310                                lli->lli_sai, (unsigned int)thread->t_pid);
1311                         l_wait_event(thread->t_ctl_waitq,
1312                                      thread_is_stopped(thread),
1313                                      &lwi);
1314                 } else {
1315                         spin_unlock(&lli->lli_sa_lock);
1316                 }
1317
1318                 /*
1319                  * Put the ref which was held when first statahead_enter.
1320                  * It maybe not the last ref for some statahead requests
1321                  * maybe inflight.
1322                  */
1323                 ll_sai_put(lli->lli_sai);
1324         } else {
1325                 lli->lli_opendir_pid = 0;
1326                 spin_unlock(&lli->lli_sa_lock);
1327         }
1328 }
1329
1330 enum {
1331         /**
1332          * not first dirent, or is "."
1333          */
1334         LS_NONE_FIRST_DE = 0,
1335         /**
1336          * the first non-hidden dirent
1337          */
1338         LS_FIRST_DE,
1339         /**
1340          * the first hidden dirent, that is "."
1341          */
1342         LS_FIRST_DOT_DE
1343 };
1344
1345 static int is_first_dirent(struct inode *dir, struct dentry *dentry)
1346 {
1347         struct ll_dir_chain   chain;
1348         struct qstr          *target = &dentry->d_name;
1349         struct md_op_data    *op_data;
1350         int                   dot_de;
1351         struct lu_dirent     *ent;
1352         struct page          *page = NULL;
1353         int                   rc     = LS_NONE_FIRST_DE;
1354         ENTRY;
1355
1356         ll_dir_chain_init(&chain);
1357
1358         op_data = ll_prep_md_op_data(NULL, dir, dir, NULL, 0, 0,
1359                                      LUSTRE_OPC_ANY, dir);
1360         if (IS_ERR(op_data))
1361                 GOTO(out, rc = PTR_ERR(op_data));
1362         /**
1363          *FIXME choose the start offset of the readdir
1364          */
1365         op_data->op_stripe_offset = 0;
1366         op_data->op_hash_offset = 0;
1367         op_data->op_max_pages = ll_i2sbi(dir)->ll_md_brw_pages;
1368
1369         for (ent = ll_dir_entry_start(dir, op_data, &page);
1370              ent != NULL && !IS_ERR(ent);
1371              ent = ll_dir_entry_next(dir, op_data, ent, &page)) {
1372                 __u64 hash;
1373                 int namelen;
1374                 char *name;
1375
1376                 hash = le64_to_cpu(ent->lde_hash);
1377                 /* The ll_get_dir_page() can return any page containing
1378                  * the given hash which may be not the start hash. */
1379                 if (unlikely(hash < op_data->op_hash_offset))
1380                         continue;
1381
1382                 namelen = le16_to_cpu(ent->lde_namelen);
1383                 if (unlikely(namelen == 0))
1384                         /*
1385                          * skip dummy record.
1386                          */
1387                         continue;
1388
1389                 name = ent->lde_name;
1390                 if (name[0] == '.') {
1391                         if (namelen == 1)
1392                                 /*
1393                                  * skip "."
1394                                  */
1395                                 continue;
1396                         else if (name[1] == '.' && namelen == 2)
1397                                 /*
1398                                  * skip ".."
1399                                  */
1400                                 continue;
1401                         else
1402                                 dot_de = 1;
1403                 } else {
1404                         dot_de = 0;
1405                 }
1406
1407                 if (dot_de && target->name[0] != '.') {
1408                         CDEBUG(D_READA, "%.*s skip hidden file %.*s\n",
1409                                target->len, target->name,
1410                                namelen, name);
1411                         continue;
1412                 }
1413
1414                 if (target->len != namelen ||
1415                     memcmp(target->name, name, namelen) != 0)
1416                         rc = LS_NONE_FIRST_DE;
1417                 else if (!dot_de)
1418                         rc = LS_FIRST_DE;
1419                 else
1420                         rc = LS_FIRST_DOT_DE;
1421
1422                 break;
1423         }
1424         EXIT;
1425
1426         if (page != NULL) {
1427                 kunmap(page);
1428                 page_cache_release(page);
1429         }
1430         ll_finish_md_op_data(op_data);
1431 out:
1432         ll_dir_chain_fini(&chain);
1433         return rc;
1434 }
1435
1436 static void
1437 ll_sai_unplug(struct ll_statahead_info *sai, struct ll_sa_entry *entry)
1438 {
1439         struct ptlrpc_thread *thread = &sai->sai_thread;
1440         struct ll_sb_info    *sbi    = ll_i2sbi(sai->sai_inode);
1441         int                   hit;
1442         ENTRY;
1443
1444         if (entry != NULL && entry->se_stat == SA_ENTRY_SUCC)
1445                 hit = 1;
1446         else
1447                 hit = 0;
1448
1449         ll_sa_entry_fini(sai, entry);
1450         if (hit) {
1451                 sai->sai_hit++;
1452                 sai->sai_consecutive_miss = 0;
1453                 sai->sai_max = min(2 * sai->sai_max, sbi->ll_sa_max);
1454         } else {
1455                 struct ll_inode_info *lli = ll_i2info(sai->sai_inode);
1456
1457                 sai->sai_miss++;
1458                 sai->sai_consecutive_miss++;
1459                 if (sa_low_hit(sai) && thread_is_running(thread)) {
1460                         atomic_inc(&sbi->ll_sa_wrong);
1461                         CDEBUG(D_READA, "Statahead for dir "DFID" hit "
1462                                "ratio too low: hit/miss "LPU64"/"LPU64
1463                                ", sent/replied "LPU64"/"LPU64", stopping "
1464                                "statahead thread\n",
1465                                PFID(&lli->lli_fid), sai->sai_hit,
1466                                sai->sai_miss, sai->sai_sent,
1467                                sai->sai_replied);
1468                         spin_lock(&lli->lli_sa_lock);
1469                         if (!thread_is_stopped(thread))
1470                                 thread_set_flags(thread, SVC_STOPPING);
1471                         spin_unlock(&lli->lli_sa_lock);
1472                 }
1473         }
1474
1475         if (!thread_is_stopped(thread))
1476                 wake_up(&thread->t_ctl_waitq);
1477
1478         EXIT;
1479 }
1480
1481 /**
1482  * Start statahead thread if this is the first dir entry.
1483  * Otherwise if a thread is started already, wait it until it is ahead of me.
1484  * \retval 1       -- find entry with lock in cache, the caller needs to do
1485  *                    nothing.
1486  * \retval 0       -- find entry in cache, but without lock, the caller needs
1487  *                    refresh from MDS.
1488  * \retval others  -- the caller need to process as non-statahead.
1489  */
1490 int do_statahead_enter(struct inode *dir, struct dentry **dentryp,
1491                        int only_unplug)
1492 {
1493         struct ll_inode_info            *lli = ll_i2info(dir);
1494         struct ll_statahead_info        *sai = lli->lli_sai;
1495         struct dentry                   *parent;
1496         struct ll_sa_entry              *entry;
1497         struct ptlrpc_thread            *thread;
1498         struct l_wait_info               lwi = { 0 };
1499         struct task_struct              *task;
1500         int                              rc = 0;
1501         struct ll_inode_info            *plli;
1502         ENTRY;
1503
1504         LASSERT(lli->lli_opendir_pid == current_pid());
1505
1506         if (sai) {
1507                 thread = &sai->sai_thread;
1508                 if (unlikely(thread_is_stopped(thread) &&
1509                              list_empty(&sai->sai_entries_stated))) {
1510                         /* to release resource */
1511                         ll_stop_statahead(dir, lli->lli_opendir_key);
1512                         RETURN(-EAGAIN);
1513                 }
1514
1515                 if ((*dentryp)->d_name.name[0] == '.') {
1516                         if (sai->sai_ls_all ||
1517                             sai->sai_miss_hidden >= sai->sai_skip_hidden) {
1518                                 /*
1519                                  * Hidden dentry is the first one, or statahead
1520                                  * thread does not skip so many hidden dentries
1521                                  * before "sai_ls_all" enabled as below.
1522                                  */
1523                         } else {
1524                                 if (!sai->sai_ls_all)
1525                                         /*
1526                                          * It maybe because hidden dentry is not
1527                                          * the first one, "sai_ls_all" was not
1528                                          * set, then "ls -al" missed. Enable
1529                                          * "sai_ls_all" for such case.
1530                                          */
1531                                         sai->sai_ls_all = 1;
1532
1533                                 /*
1534                                  * Such "getattr" has been skipped before
1535                                  * "sai_ls_all" enabled as above.
1536                                  */
1537                                 sai->sai_miss_hidden++;
1538                                 RETURN(-EAGAIN);
1539                         }
1540                 }
1541
1542                 entry = ll_sa_entry_get_byname(sai, &(*dentryp)->d_name);
1543                 if (entry == NULL || only_unplug) {
1544                         ll_sai_unplug(sai, entry);
1545                         RETURN(entry ? 1 : -EAGAIN);
1546                 }
1547
1548                 if (!ll_sa_entry_stated(entry)) {
1549                         sai->sai_index_wait = entry->se_index;
1550                         lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(30), NULL,
1551                                                LWI_ON_SIGNAL_NOOP, NULL);
1552                         rc = l_wait_event(sai->sai_waitq,
1553                                           ll_sa_entry_stated(entry) ||
1554                                           thread_is_stopped(thread),
1555                                           &lwi);
1556                         if (rc < 0) {
1557                                 ll_sai_unplug(sai, entry);
1558                                 RETURN(-EAGAIN);
1559                         }
1560                 }
1561
1562                 if (entry->se_stat == SA_ENTRY_SUCC &&
1563                     entry->se_inode != NULL) {
1564                         struct inode *inode = entry->se_inode;
1565                         struct lookup_intent it = { .it_op = IT_GETATTR,
1566                                                     .d.lustre.it_lock_handle =
1567                                                      entry->se_handle };
1568                         __u64 bits;
1569
1570                         rc = md_revalidate_lock(ll_i2mdexp(dir), &it,
1571                                                 ll_inode2fid(inode), &bits);
1572                         if (rc == 1) {
1573                                 if ((*dentryp)->d_inode == NULL) {
1574                                         struct dentry *alias;
1575
1576                                         alias = ll_splice_alias(inode,
1577                                                                 *dentryp);
1578                                         if (IS_ERR(alias)) {
1579                                                 ll_sai_unplug(sai, entry);
1580                                                 RETURN(PTR_ERR(alias));
1581                                         }
1582                                         *dentryp = alias;
1583                                 } else if ((*dentryp)->d_inode != inode) {
1584                                         /* revalidate, but inode is recreated */
1585                                         CDEBUG(D_READA,
1586                                                "%s: stale dentry %.*s inode "
1587                                                DFID", statahead inode "DFID
1588                                                "\n",
1589                                                ll_get_fsname((*dentryp)->d_inode->i_sb, NULL, 0),
1590                                                (*dentryp)->d_name.len,
1591                                                (*dentryp)->d_name.name,
1592                                                PFID(ll_inode2fid((*dentryp)->d_inode)),
1593                                                PFID(ll_inode2fid(inode)));
1594                                         ll_intent_release(&it);
1595                                         ll_sai_unplug(sai, entry);
1596                                         RETURN(-ESTALE);
1597                                 } else {
1598                                         iput(inode);
1599                                 }
1600                                 entry->se_inode = NULL;
1601
1602                                 if ((bits & MDS_INODELOCK_LOOKUP) &&
1603                                     d_lustre_invalid(*dentryp))
1604                                         d_lustre_revalidate(*dentryp);
1605                                 ll_intent_release(&it);
1606                         }
1607                 }
1608
1609                 ll_sai_unplug(sai, entry);
1610                 RETURN(rc);
1611         }
1612
1613         /* I am the "lli_opendir_pid" owner, only me can set "lli_sai". */
1614         rc = is_first_dirent(dir, *dentryp);
1615         if (rc == LS_NONE_FIRST_DE)
1616                 /* It is not "ls -{a}l" operation, no need statahead for it. */
1617                 GOTO(out, rc = -EAGAIN);
1618
1619         sai = ll_sai_alloc();
1620         if (sai == NULL)
1621                 GOTO(out, rc = -ENOMEM);
1622
1623         sai->sai_ls_all = (rc == LS_FIRST_DOT_DE);
1624         sai->sai_inode = igrab(dir);
1625         if (unlikely(sai->sai_inode == NULL)) {
1626                 CWARN("Do not start stat ahead on dying inode "DFID"\n",
1627                       PFID(&lli->lli_fid));
1628                 GOTO(out, rc = -ESTALE);
1629         }
1630
1631         /* get parent reference count here, and put it in ll_statahead_thread */
1632         parent = dget((*dentryp)->d_parent);
1633         if (unlikely(sai->sai_inode != parent->d_inode)) {
1634                 struct ll_inode_info *nlli = ll_i2info(parent->d_inode);
1635
1636                 CWARN("Race condition, someone changed %.*s just now: "
1637                       "old parent "DFID", new parent "DFID"\n",
1638                       (*dentryp)->d_name.len, (*dentryp)->d_name.name,
1639                       PFID(&lli->lli_fid), PFID(&nlli->lli_fid));
1640                 dput(parent);
1641                 iput(sai->sai_inode);
1642                 GOTO(out, rc = -EAGAIN);
1643         }
1644
1645         CDEBUG(D_READA, "start statahead thread: sai %p, parent %.*s\n",
1646                sai, parent->d_name.len, parent->d_name.name);
1647
1648         /* The sai buffer already has one reference taken at allocation time,
1649          * but as soon as we expose the sai by attaching it to the lli that
1650          * default reference can be dropped by another thread calling
1651          * ll_stop_statahead. We need to take a local reference to protect
1652          * the sai buffer while we intend to access it. */
1653         ll_sai_get(sai);
1654         lli->lli_sai = sai;
1655
1656         plli = ll_i2info(parent->d_inode);
1657         task = kthread_run(ll_statahead_thread, parent, "ll_sa_%u",
1658                            plli->lli_opendir_pid);
1659         thread = &sai->sai_thread;
1660         if (IS_ERR(task)) {
1661                 rc = PTR_ERR(task);
1662                 CERROR("cannot start ll_sa thread: rc = %d\n", rc);
1663                 dput(parent);
1664                 lli->lli_opendir_key = NULL;
1665                 thread_set_flags(thread, SVC_STOPPED);
1666                 thread_set_flags(&sai->sai_agl_thread, SVC_STOPPED);
1667                 /* Drop both our own local reference and the default
1668                  * reference from allocation time. */
1669                 ll_sai_put(sai);
1670                 ll_sai_put(sai);
1671                 LASSERT(lli->lli_sai == NULL);
1672                 RETURN(-EAGAIN);
1673         }
1674
1675         l_wait_event(thread->t_ctl_waitq,
1676                      thread_is_running(thread) || thread_is_stopped(thread),
1677                      &lwi);
1678         ll_sai_put(sai);
1679
1680         /*
1681          * We don't stat-ahead for the first dirent since we are already in
1682          * lookup.
1683          */
1684         RETURN(-EAGAIN);
1685
1686 out:
1687         if (sai != NULL)
1688                 OBD_FREE_PTR(sai);
1689         spin_lock(&lli->lli_sa_lock);
1690         lli->lli_opendir_key = NULL;
1691         lli->lli_opendir_pid = 0;
1692         spin_unlock(&lli->lli_sa_lock);
1693         return rc;
1694 }