Pileus Git - ~andy/linux/blob - drivers/staging/lustre/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, 2012, 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
 204         entry_size = sizeof(struct ll_sa_entry) + (len & ~3) + 4;
 205         OBD_ALLOC(entry, entry_size);
 206         if (unlikely(entry == NULL))
 207                 return ERR_PTR(-ENOMEM);
 208
 209         CDEBUG(D_READA, "alloc sa entry %.*s(%p) index "LPU64"\n",
 210                len, name, entry, index);
 211
 212         entry->se_index = index;
 213
 214         /*
 215          * Statahead entry reference rules:
 216          *
 217          * 1) When statahead entry is initialized, its reference is set as 2.
 218          *    One reference is used by the directory scanner. When the scanner
 219          *    searches the statahead cache for the given name, it can perform
 220          *    lockless hash lookup (only the scanner can remove entry from hash
 221          *    list), and once found, it needn't to call "atomic_inc()" for the
 222          *    entry reference. So the performance is improved. After using the
 223          *    statahead entry, the scanner will call "atomic_dec()" to drop the
 224          *    reference held when initialization. If it is the last reference,
 225          *    the statahead entry will be freed.
 226          *
 227          * 2) All other threads, including statahead thread and ptlrpcd thread,
 228          *    when they process the statahead entry, the reference for target
 229          *    should be held to guarantee the entry will not be released by the
 230          *    directory scanner. After processing the entry, these threads will
 231          *    drop the entry reference. If it is the last reference, the entry
 232          *    will be freed.
 233          *
 234          *    The second reference when initializes the statahead entry is used
 235          *    by the statahead thread, following the rule 2).
 236          */
 237         atomic_set(&entry->se_refcount, 2);
 238         entry->se_stat = SA_ENTRY_INIT;
 239         entry->se_size = entry_size;
 240         dname = (char *)entry + sizeof(struct ll_sa_entry);
 241         memcpy(dname, name, len);
 242         dname[len] = 0;
 243         entry->se_qstr.hash = full_name_hash(name, len);
 244         entry->se_qstr.len = len;
 245         entry->se_qstr.name = dname;
 246
 247         lli = ll_i2info(sai->sai_inode);
 248         spin_lock(&lli->lli_sa_lock);
 249         list_add_tail(&entry->se_link, &sai->sai_entries);
 250         INIT_LIST_HEAD(&entry->se_list);
 251         ll_sa_entry_enhash(sai, entry);
 252         spin_unlock(&lli->lli_sa_lock);
 253
 254         atomic_inc(&sai->sai_cache_count);
 255
 256         return entry;
 257 }
 258
 259 /*
 260  * Used by the directory scanner to search entry with name.
 261  *
 262  * Only the caller can remove the entry from hash, so it is unnecessary to hold
 263  * hash lock. It is caller's duty to release the init refcount on the entry, so
 264  * it is also unnecessary to increase refcount on the entry.
 265  */
 266 static struct ll_sa_entry *
 267 ll_sa_entry_get_byname(struct ll_statahead_info *sai, const struct qstr *qstr)
 268 {
 269         struct ll_sa_entry *entry;
 270         int i = ll_sa_entry_hash(qstr->hash);
 271
 272         list_for_each_entry(entry, &sai->sai_cache[i], se_hash) {
 273                 if (entry->se_qstr.hash == qstr->hash &&
 274                     entry->se_qstr.len == qstr->len &&
 275                     memcmp(entry->se_qstr.name, qstr->name, qstr->len) == 0)
 276                         return entry;
 277         }
 278         return NULL;
 279 }
 280
 281 /*
 282  * Used by the async getattr request callback to find entry with index.
 283  *
 284  * Inside lli_sa_lock to prevent others to change the list during the search.
 285  * It needs to increase entry refcount before returning to guarantee that the
 286  * entry cannot be freed by others.
 287  */
 288 static struct ll_sa_entry *
 289 ll_sa_entry_get_byindex(struct ll_statahead_info *sai, __u64 index)
 290 {
 291         struct ll_sa_entry *entry;
 292
 293         list_for_each_entry(entry, &sai->sai_entries, se_link) {
 294                 if (entry->se_index == index) {
 295                         LASSERT(atomic_read(&entry->se_refcount) > 0);
 296                         atomic_inc(&entry->se_refcount);
 297                         return entry;
 298                 }
 299                 if (entry->se_index > index)
 300                         break;
 301         }
 302         return NULL;
 303 }
 304
 305 static void ll_sa_entry_cleanup(struct ll_statahead_info *sai,
 306                                  struct ll_sa_entry *entry)
 307 {
 308         struct md_enqueue_info *minfo = entry->se_minfo;
 309         struct ptlrpc_request  *req   = entry->se_req;
 310
 311         if (minfo) {
 312                 entry->se_minfo = NULL;
 313                 ll_intent_release(&minfo->mi_it);
 314                 iput(minfo->mi_dir);
 315                 OBD_FREE_PTR(minfo);
 316         }
 317
 318         if (req) {
 319                 entry->se_req = NULL;
 320                 ptlrpc_req_finished(req);
 321         }
 322 }
 323
 324 static void ll_sa_entry_put(struct ll_statahead_info *sai,
 325                              struct ll_sa_entry *entry)
 326 {
 327         if (atomic_dec_and_test(&entry->se_refcount)) {
 328                 CDEBUG(D_READA, "free sa entry %.*s(%p) index "LPU64"\n",
 329                        entry->se_qstr.len, entry->se_qstr.name, entry,
 330                        entry->se_index);
 331
 332                 LASSERT(list_empty(&entry->se_link));
 333                 LASSERT(list_empty(&entry->se_list));
 334                 LASSERT(ll_sa_entry_unhashed(entry));
 335
 336                 ll_sa_entry_cleanup(sai, entry);
 337                 if (entry->se_inode)
 338                         iput(entry->se_inode);
 339
 340                 OBD_FREE(entry, entry->se_size);
 341                 atomic_dec(&sai->sai_cache_count);
 342         }
 343 }
 344
 345 static inline void
 346 do_sa_entry_fini(struct ll_statahead_info *sai, struct ll_sa_entry *entry)
 347 {
 348         struct ll_inode_info *lli = ll_i2info(sai->sai_inode);
 349
 350         LASSERT(!ll_sa_entry_unhashed(entry));
 351         LASSERT(!list_empty(&entry->se_link));
 352
 353         ll_sa_entry_unhash(sai, entry);
 354
 355         spin_lock(&lli->lli_sa_lock);
 356         entry->se_stat = SA_ENTRY_DEST;
 357         list_del_init(&entry->se_link);
 358         if (likely(!list_empty(&entry->se_list)))
 359                 list_del_init(&entry->se_list);
 360         spin_unlock(&lli->lli_sa_lock);
 361
 362         ll_sa_entry_put(sai, entry);
 363 }
 364
 365 /*
 366  * Delete it from sai_entries_stated list when fini.
 367  */
 368 static void
 369 ll_sa_entry_fini(struct ll_statahead_info *sai, struct ll_sa_entry *entry)
 370 {
 371         struct ll_sa_entry *pos, *next;
 372
 373         if (entry)
 374                 do_sa_entry_fini(sai, entry);
 375
 376         /* drop old entry, only 'scanner' process does this, no need to lock */
 377         list_for_each_entry_safe(pos, next, &sai->sai_entries, se_link) {
 378                 if (!is_omitted_entry(sai, pos->se_index))
 379                         break;
 380                 do_sa_entry_fini(sai, pos);
 381         }
 382 }
 383
 384 /*
 385  * Inside lli_sa_lock.
 386  */
 387 static void
 388 do_sa_entry_to_stated(struct ll_statahead_info *sai,
 389                       struct ll_sa_entry *entry, se_stat_t stat)
 390 {
 391         struct ll_sa_entry *se;
 392         struct list_head         *pos = &sai->sai_entries_stated;
 393
 394         if (!list_empty(&entry->se_list))
 395                 list_del_init(&entry->se_list);
 396
 397         list_for_each_entry_reverse(se, &sai->sai_entries_stated, se_list) {
 398                 if (se->se_index < entry->se_index) {
 399                         pos = &se->se_list;
 400                         break;
 401                 }
 402         }
 403
 404         list_add(&entry->se_list, pos);
 405         entry->se_stat = stat;
 406 }
 407
 408 /*
 409  * Move entry to sai_entries_stated and sort with the index.
 410  * \retval 1    -- entry to be destroyed.
 411  * \retval 0    -- entry is inserted into stated list.
 412  */
 413 static int
 414 ll_sa_entry_to_stated(struct ll_statahead_info *sai,
 415                       struct ll_sa_entry *entry, se_stat_t stat)
 416 {
 417         struct ll_inode_info *lli = ll_i2info(sai->sai_inode);
 418         int                ret = 1;
 419
 420         ll_sa_entry_cleanup(sai, entry);
 421
 422         spin_lock(&lli->lli_sa_lock);
 423         if (likely(entry->se_stat != SA_ENTRY_DEST)) {
 424                 do_sa_entry_to_stated(sai, entry, stat);
 425                 ret = 0;
 426         }
 427         spin_unlock(&lli->lli_sa_lock);
 428
 429         return ret;
 430 }
 431
 432 /*
 433  * Insert inode into the list of sai_entries_agl.
 434  */
 435 static void ll_agl_add(struct ll_statahead_info *sai,
 436                        struct inode *inode, int index)
 437 {
 438         struct ll_inode_info *child  = ll_i2info(inode);
 439         struct ll_inode_info *parent = ll_i2info(sai->sai_inode);
 440         int                added  = 0;
 441
 442         spin_lock(&child->lli_agl_lock);
 443         if (child->lli_agl_index == 0) {
 444                 child->lli_agl_index = index;
 445                 spin_unlock(&child->lli_agl_lock);
 446
 447                 LASSERT(list_empty(&child->lli_agl_list));
 448
 449                 igrab(inode);
 450                 spin_lock(&parent->lli_agl_lock);
 451                 if (agl_list_empty(sai))
 452                         added = 1;
 453                 list_add_tail(&child->lli_agl_list, &sai->sai_entries_agl);
 454                 spin_unlock(&parent->lli_agl_lock);
 455         } else {
 456                 spin_unlock(&child->lli_agl_lock);
 457         }
 458
 459         if (added > 0)
 460                 wake_up(&sai->sai_agl_thread.t_ctl_waitq);
 461 }
 462
 463 static struct ll_statahead_info *ll_sai_alloc(void)
 464 {
 465         struct ll_statahead_info *sai;
 466         int                    i;
 467
 468         OBD_ALLOC_PTR(sai);
 469         if (!sai)
 470                 return NULL;
 471
 472         atomic_set(&sai->sai_refcount, 1);
 473
 474         spin_lock(&sai_generation_lock);
 475         sai->sai_generation = ++sai_generation;
 476         if (unlikely(sai_generation == 0))
 477                 sai->sai_generation = ++sai_generation;
 478         spin_unlock(&sai_generation_lock);
 479
 480         sai->sai_max = LL_SA_RPC_MIN;
 481         sai->sai_index = 1;
 482         init_waitqueue_head(&sai->sai_waitq);
 483         init_waitqueue_head(&sai->sai_thread.t_ctl_waitq);
 484         init_waitqueue_head(&sai->sai_agl_thread.t_ctl_waitq);
 485
 486         INIT_LIST_HEAD(&sai->sai_entries);
 487         INIT_LIST_HEAD(&sai->sai_entries_received);
 488         INIT_LIST_HEAD(&sai->sai_entries_stated);
 489         INIT_LIST_HEAD(&sai->sai_entries_agl);
 490
 491         for (i = 0; i < LL_SA_CACHE_SIZE; i++) {
 492                 INIT_LIST_HEAD(&sai->sai_cache[i]);
 493                 spin_lock_init(&sai->sai_cache_lock[i]);
 494         }
 495         atomic_set(&sai->sai_cache_count, 0);
 496
 497         return sai;
 498 }
 499
 500 static inline struct ll_statahead_info *
 501 ll_sai_get(struct ll_statahead_info *sai)
 502 {
 503         atomic_inc(&sai->sai_refcount);
 504         return sai;
 505 }
 506
 507 static void ll_sai_put(struct ll_statahead_info *sai)
 508 {
 509         struct inode     *inode = sai->sai_inode;
 510         struct ll_inode_info *lli   = ll_i2info(inode);
 511
 512         if (atomic_dec_and_lock(&sai->sai_refcount, &lli->lli_sa_lock)) {
 513                 struct ll_sa_entry *entry, *next;
 514
 515                 if (unlikely(atomic_read(&sai->sai_refcount) > 0)) {
 516                         /* It is race case, the interpret callback just hold
 517                          * a reference count */
 518                         spin_unlock(&lli->lli_sa_lock);
 519                         return;
 520                 }
 521
 522                 LASSERT(lli->lli_opendir_key == NULL);
 523                 LASSERT(thread_is_stopped(&sai->sai_thread));
 524                 LASSERT(thread_is_stopped(&sai->sai_agl_thread));
 525
 526                 lli->lli_sai = NULL;
 527                 lli->lli_opendir_pid = 0;
 528                 spin_unlock(&lli->lli_sa_lock);
 529
 530                 if (sai->sai_sent > sai->sai_replied)
 531                         CDEBUG(D_READA,"statahead for dir "DFID" does not "
 532                               "finish: [sent:"LPU64"] [replied:"LPU64"]\n",
 533                               PFID(&lli->lli_fid),
 534                               sai->sai_sent, sai->sai_replied);
 535
 536                 list_for_each_entry_safe(entry, next,
 537                                              &sai->sai_entries, se_link)
 538                         do_sa_entry_fini(sai, entry);
 539
 540                 LASSERT(list_empty(&sai->sai_entries));
 541                 LASSERT(sa_received_empty(sai));
 542                 LASSERT(list_empty(&sai->sai_entries_stated));
 543
 544                 LASSERT(atomic_read(&sai->sai_cache_count) == 0);
 545                 LASSERT(agl_list_empty(sai));
 546
 547                 iput(inode);
 548                 OBD_FREE_PTR(sai);
 549         }
 550 }
 551
 552 /* Do NOT forget to drop inode refcount when into sai_entries_agl. */
 553 static void ll_agl_trigger(struct inode *inode, struct ll_statahead_info *sai)
 554 {
 555         struct ll_inode_info *lli   = ll_i2info(inode);
 556         __u64            index = lli->lli_agl_index;
 557         int                rc;
 558
 559         LASSERT(list_empty(&lli->lli_agl_list));
 560
 561         /* AGL maybe fall behind statahead with one entry */
 562         if (is_omitted_entry(sai, index + 1)) {
 563                 lli->lli_agl_index = 0;
 564                 iput(inode);
 565                 return;
 566         }
 567
 568         /* Someone is in glimpse (sync or async), do nothing. */
 569         rc = down_write_trylock(&lli->lli_glimpse_sem);
 570         if (rc == 0) {
 571                 lli->lli_agl_index = 0;
 572                 iput(inode);
 573                 return;
 574         }
 575
 576         /*
 577          * Someone triggered glimpse within 1 sec before.
 578          * 1) The former glimpse succeeded with glimpse lock granted by OST, and
 579          *    if the lock is still cached on client, AGL needs to do nothing. If
 580          *    it is cancelled by other client, AGL maybe cannot obtaion new lock
 581          *    for no glimpse callback triggered by AGL.
 582          * 2) The former glimpse succeeded, but OST did not grant glimpse lock.
 583          *    Under such case, it is quite possible that the OST will not grant
 584          *    glimpse lock for AGL also.
 585          * 3) The former glimpse failed, compared with other two cases, it is
 586          *    relative rare. AGL can ignore such case, and it will not muchly
 587          *    affect the performance.
 588          */
 589         if (lli->lli_glimpse_time != 0 &&
 590             cfs_time_before(cfs_time_shift(-1), lli->lli_glimpse_time)) {
 591                 up_write(&lli->lli_glimpse_sem);
 592                 lli->lli_agl_index = 0;
 593                 iput(inode);
 594                 return;
 595         }
 596
 597         CDEBUG(D_READA, "Handling (init) async glimpse: inode = "
 598                DFID", idx = "LPU64"\n", PFID(&lli->lli_fid), index);
 599
 600         cl_agl(inode);
 601         lli->lli_agl_index = 0;
 602         lli->lli_glimpse_time = cfs_time_current();
 603         up_write(&lli->lli_glimpse_sem);
 604
 605         CDEBUG(D_READA, "Handled (init) async glimpse: inode= "
 606                DFID", idx = "LPU64", rc = %d\n",
 607                PFID(&lli->lli_fid), index, rc);
 608
 609         iput(inode);
 610 }
 611
 612 static void ll_post_statahead(struct ll_statahead_info *sai)
 613 {
 614         struct inode       *dir   = sai->sai_inode;
 615         struct inode       *child;
 616         struct ll_inode_info   *lli   = ll_i2info(dir);
 617         struct ll_sa_entry     *entry;
 618         struct md_enqueue_info *minfo;
 619         struct lookup_intent   *it;
 620         struct ptlrpc_request  *req;
 621         struct mdt_body *body;
 622         int                  rc    = 0;
 623
 624         spin_lock(&lli->lli_sa_lock);
 625         if (unlikely(sa_received_empty(sai))) {
 626                 spin_unlock(&lli->lli_sa_lock);
 627                 return;
 628         }
 629         entry = sa_first_received_entry(sai);
 630         atomic_inc(&entry->se_refcount);
 631         list_del_init(&entry->se_list);
 632         spin_unlock(&lli->lli_sa_lock);
 633
 634         LASSERT(entry->se_handle != 0);
 635
 636         minfo = entry->se_minfo;
 637         it = &minfo->mi_it;
 638         req = entry->se_req;
 639         body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
 640         if (body == NULL)
 641                 GOTO(out, rc = -EFAULT);
 642
 643         child = entry->se_inode;
 644         if (child == NULL) {
 645                 /*
 646                  * lookup.
 647                  */
 648                 LASSERT(fid_is_zero(&minfo->mi_data.op_fid2));
 649
 650                 /* XXX: No fid in reply, this is probably cross-ref case.
 651                  * SA can't handle it yet. */
 652                 if (body->valid & OBD_MD_MDS)
 653                         GOTO(out, rc = -EAGAIN);
 654         } else {
 655                 /*
 656                  * revalidate.
 657                  */
 658                 /* unlinked and re-created with the same name */
 659                 if (unlikely(!lu_fid_eq(&minfo->mi_data.op_fid2, &body->fid1))){
 660                         entry->se_inode = NULL;
 661                         iput(child);
 662                         child = NULL;
 663                 }
 664         }
 665
 666         it->d.lustre.it_lock_handle = entry->se_handle;
 667         rc = md_revalidate_lock(ll_i2mdexp(dir), it, ll_inode2fid(dir), NULL);
 668         if (rc != 1)
 669                 GOTO(out, rc = -EAGAIN);
 670
 671         rc = ll_prep_inode(&child, req, dir->i_sb, it);
 672         if (rc)
 673                 GOTO(out, rc);
 674
 675         CDEBUG(D_DLMTRACE, "setting l_data to inode %p (%lu/%u)\n",
 676                child, child->i_ino, child->i_generation);
 677         ll_set_lock_data(ll_i2sbi(dir)->ll_md_exp, child, it, NULL);
 678
 679         entry->se_inode = child;
 680
 681         if (agl_should_run(sai, child))
 682                 ll_agl_add(sai, child, entry->se_index);
 683
 684 out:
 685         /* The "ll_sa_entry_to_stated()" will drop related ldlm ibits lock
 686          * reference count by calling "ll_intent_drop_lock()" in spite of the
 687          * above operations failed or not. Do not worry about calling
 688          * "ll_intent_drop_lock()" more than once. */
 689         rc = ll_sa_entry_to_stated(sai, entry,
 690                                    rc < 0 ? SA_ENTRY_INVA : SA_ENTRY_SUCC);
 691         if (rc == 0 && entry->se_index == sai->sai_index_wait)
 692                 wake_up(&sai->sai_waitq);
 693         ll_sa_entry_put(sai, entry);
 694 }
 695
 696 static int ll_statahead_interpret(struct ptlrpc_request *req,
 697                                   struct md_enqueue_info *minfo, int rc)
 698 {
 699         struct lookup_intent     *it  = &minfo->mi_it;
 700         struct inode         *dir = minfo->mi_dir;
 701         struct ll_inode_info     *lli = ll_i2info(dir);
 702         struct ll_statahead_info *sai = NULL;
 703         struct ll_sa_entry       *entry;
 704         int                    wakeup;
 705
 706         if (it_disposition(it, DISP_LOOKUP_NEG))
 707                 rc = -ENOENT;
 708
 709         spin_lock(&lli->lli_sa_lock);
 710         /* stale entry */
 711         if (unlikely(lli->lli_sai == NULL ||
 712                      lli->lli_sai->sai_generation != minfo->mi_generation)) {
 713                 spin_unlock(&lli->lli_sa_lock);
 714                 GOTO(out, rc = -ESTALE);
 715         } else {
 716                 sai = ll_sai_get(lli->lli_sai);
 717                 if (unlikely(!thread_is_running(&sai->sai_thread))) {
 718                         sai->sai_replied++;
 719                         spin_unlock(&lli->lli_sa_lock);
 720                         GOTO(out, rc = -EBADFD);
 721                 }
 722
 723                 entry = ll_sa_entry_get_byindex(sai, minfo->mi_cbdata);
 724                 if (entry == NULL) {
 725                         sai->sai_replied++;
 726                         spin_unlock(&lli->lli_sa_lock);
 727                         GOTO(out, rc = -EIDRM);
 728                 }
 729
 730                 if (rc != 0) {
 731                         do_sa_entry_to_stated(sai, entry, SA_ENTRY_INVA);
 732                         wakeup = (entry->se_index == sai->sai_index_wait);
 733                 } else {
 734                         entry->se_minfo = minfo;
 735                         entry->se_req = ptlrpc_request_addref(req);
 736                         /* Release the async ibits lock ASAP to avoid deadlock
 737                          * when statahead thread tries to enqueue lock on parent
 738                          * for readpage and other tries to enqueue lock on child
 739                          * with parent's lock held, for example: unlink. */
 740                         entry->se_handle = it->d.lustre.it_lock_handle;
 741                         ll_intent_drop_lock(it);
 742                         wakeup = sa_received_empty(sai);
 743                         list_add_tail(&entry->se_list,
 744                                           &sai->sai_entries_received);
 745                 }
 746                 sai->sai_replied++;
 747                 spin_unlock(&lli->lli_sa_lock);
 748
 749                 ll_sa_entry_put(sai, entry);
 750                 if (wakeup)
 751                         wake_up(&sai->sai_thread.t_ctl_waitq);
 752         }
 753
 754 out:
 755         if (rc != 0) {
 756                 ll_intent_release(it);
 757                 iput(dir);
 758                 OBD_FREE_PTR(minfo);
 759         }
 760         if (sai != NULL)
 761                 ll_sai_put(sai);
 762         return rc;
 763 }
 764
 765 static void sa_args_fini(struct md_enqueue_info *minfo,
 766                          struct ldlm_enqueue_info *einfo)
 767 {
 768         LASSERT(minfo && einfo);
 769         iput(minfo->mi_dir);
 770         capa_put(minfo->mi_data.op_capa1);
 771         capa_put(minfo->mi_data.op_capa2);
 772         OBD_FREE_PTR(minfo);
 773         OBD_FREE_PTR(einfo);
 774 }
 775
 776 /**
 777  * There is race condition between "capa_put" and "ll_statahead_interpret" for
 778  * accessing "op_data.op_capa[1,2]" as following:
 779  * "capa_put" releases "op_data.op_capa[1,2]"'s reference count after calling
 780  * "md_intent_getattr_async". But "ll_statahead_interpret" maybe run first, and
 781  * fill "op_data.op_capa[1,2]" as POISON, then cause "capa_put" access invalid
 782  * "ocapa". So here reserve "op_data.op_capa[1,2]" in "pcapa" before calling
 783  * "md_intent_getattr_async".
 784  */
 785 static int sa_args_init(struct inode *dir, struct inode *child,
 786                         struct ll_sa_entry *entry, struct md_enqueue_info **pmi,
 787                         struct ldlm_enqueue_info **pei,
 788                         struct obd_capa **pcapa)
 789 {
 790         struct qstr           *qstr = &entry->se_qstr;
 791         struct ll_inode_info     *lli  = ll_i2info(dir);
 792         struct md_enqueue_info   *minfo;
 793         struct ldlm_enqueue_info *einfo;
 794         struct md_op_data       *op_data;
 795
 796         OBD_ALLOC_PTR(einfo);
 797         if (einfo == NULL)
 798                 return -ENOMEM;
 799
 800         OBD_ALLOC_PTR(minfo);
 801         if (minfo == NULL) {
 802                 OBD_FREE_PTR(einfo);
 803                 return -ENOMEM;
 804         }
 805
 806         op_data = ll_prep_md_op_data(&minfo->mi_data, dir, child, qstr->name,
 807                                      qstr->len, 0, LUSTRE_OPC_ANY, NULL);
 808         if (IS_ERR(op_data)) {
 809                 OBD_FREE_PTR(einfo);
 810                 OBD_FREE_PTR(minfo);
 811                 return PTR_ERR(op_data);
 812         }
 813
 814         minfo->mi_it.it_op = IT_GETATTR;
 815         minfo->mi_dir = igrab(dir);
 816         minfo->mi_cb = ll_statahead_interpret;
 817         minfo->mi_generation = lli->lli_sai->sai_generation;
 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
 842         rc = sa_args_init(dir, NULL, entry, &minfo, &einfo, capas);
 843         if (rc)
 844                 return rc;
 845
 846         rc = md_intent_getattr_async(ll_i2mdexp(dir), minfo, einfo);
 847         if (!rc) {
 848                 capa_put(capas[0]);
 849                 capa_put(capas[1]);
 850         } else {
 851                 sa_args_fini(minfo, einfo);
 852         }
 853
 854         return rc;
 855 }
 856
 857 /**
 858  * similar to ll_revalidate_it().
 859  * \retval      1 -- dentry valid
 860  * \retval      0 -- will send stat-ahead request
 861  * \retval others -- prepare stat-ahead request failed
 862  */
 863 static int do_sa_revalidate(struct inode *dir, struct ll_sa_entry *entry,
 864                             struct dentry *dentry)
 865 {
 866         struct inode         *inode = dentry->d_inode;
 867         struct lookup_intent      it = { .it_op = IT_GETATTR,
 868                                          .d.lustre.it_lock_handle = 0 };
 869         struct md_enqueue_info   *minfo;
 870         struct ldlm_enqueue_info *einfo;
 871         struct obd_capa   *capas[2];
 872         int rc;
 873
 874         if (unlikely(inode == NULL))
 875                 return 1;
 876
 877         if (d_mountpoint(dentry))
 878                 return 1;
 879
 880         if (unlikely(dentry == dentry->d_sb->s_root))
 881                 return 1;
 882
 883         entry->se_inode = igrab(inode);
 884         rc = md_revalidate_lock(ll_i2mdexp(dir), &it, ll_inode2fid(inode),NULL);
 885         if (rc == 1) {
 886                 entry->se_handle = it.d.lustre.it_lock_handle;
 887                 ll_intent_release(&it);
 888                 return 1;
 889         }
 890
 891         rc = sa_args_init(dir, inode, entry, &minfo, &einfo, capas);
 892         if (rc) {
 893                 entry->se_inode = NULL;
 894                 iput(inode);
 895                 return rc;
 896         }
 897
 898         rc = md_intent_getattr_async(ll_i2mdexp(dir), minfo, einfo);
 899         if (!rc) {
 900                 capa_put(capas[0]);
 901                 capa_put(capas[1]);
 902         } else {
 903                 entry->se_inode = NULL;
 904                 iput(inode);
 905                 sa_args_fini(minfo, einfo);
 906         }
 907
 908         return rc;
 909 }
 910
 911 static void ll_statahead_one(struct dentry *parent, const char* entry_name,
 912                              int entry_name_len)
 913 {
 914         struct inode         *dir    = parent->d_inode;
 915         struct ll_inode_info     *lli    = ll_i2info(dir);
 916         struct ll_statahead_info *sai    = lli->lli_sai;
 917         struct dentry       *dentry = NULL;
 918         struct ll_sa_entry       *entry;
 919         int                    rc;
 920         int                    rc1;
 921
 922         entry = ll_sa_entry_alloc(sai, sai->sai_index, entry_name,
 923                                   entry_name_len);
 924         if (IS_ERR(entry))
 925                 return;
 926
 927         dentry = d_lookup(parent, &entry->se_qstr);
 928         if (!dentry) {
 929                 rc = do_sa_lookup(dir, entry);
 930         } else {
 931                 rc = do_sa_revalidate(dir, entry, dentry);
 932                 if (rc == 1 && agl_should_run(sai, dentry->d_inode))
 933                         ll_agl_add(sai, dentry->d_inode, entry->se_index);
 934         }
 935
 936         if (dentry != NULL)
 937                 dput(dentry);
 938
 939         if (rc) {
 940                 rc1 = ll_sa_entry_to_stated(sai, entry,
 941                                         rc < 0 ? SA_ENTRY_INVA : SA_ENTRY_SUCC);
 942                 if (rc1 == 0 && entry->se_index == sai->sai_index_wait)
 943                         wake_up(&sai->sai_waitq);
 944         } else {
 945                 sai->sai_sent++;
 946         }
 947
 948         sai->sai_index++;
 949         /* drop one refcount on entry by ll_sa_entry_alloc */
 950         ll_sa_entry_put(sai, entry);
 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    = ll_sai_get(plli->lli_sai);
 961         struct ptlrpc_thread     *thread = &sai->sai_agl_thread;
 962         struct l_wait_info      lwi    = { 0 };
 963
 964         CDEBUG(D_READA, "agl thread started: [pid %d] [parent %.*s]\n",
 965                current_pid(), parent->d_name.len, parent->d_name.name);
 966
 967         atomic_inc(&sbi->ll_agl_total);
 968         spin_lock(&plli->lli_agl_lock);
 969         sai->sai_agl_valid = 1;
 970         thread_set_flags(thread, SVC_RUNNING);
 971         spin_unlock(&plli->lli_agl_lock);
 972         wake_up(&thread->t_ctl_waitq);
 973
 974         while (1) {
 975                 l_wait_event(thread->t_ctl_waitq,
 976                              !agl_list_empty(sai) ||
 977                              !thread_is_running(thread),
 978                              &lwi);
 979
 980                 if (!thread_is_running(thread))
 981                         break;
 982
 983                 spin_lock(&plli->lli_agl_lock);
 984                 /* The statahead thread maybe help to process AGL entries,
 985                  * so check whether list empty again. */
 986                 if (!agl_list_empty(sai)) {
 987                         clli = agl_first_entry(sai);
 988                         list_del_init(&clli->lli_agl_list);
 989                         spin_unlock(&plli->lli_agl_lock);
 990                         ll_agl_trigger(&clli->lli_vfs_inode, sai);
 991                 } else {
 992                         spin_unlock(&plli->lli_agl_lock);
 993                 }
 994         }
 995
 996         spin_lock(&plli->lli_agl_lock);
 997         sai->sai_agl_valid = 0;
 998         while (!agl_list_empty(sai)) {
 999                 clli = agl_first_entry(sai);
1000                 list_del_init(&clli->lli_agl_list);
1001                 spin_unlock(&plli->lli_agl_lock);
1002                 clli->lli_agl_index = 0;
1003                 iput(&clli->lli_vfs_inode);
1004                 spin_lock(&plli->lli_agl_lock);
1005         }
1006         thread_set_flags(thread, SVC_STOPPED);
1007         spin_unlock(&plli->lli_agl_lock);
1008         wake_up(&thread->t_ctl_waitq);
1009         ll_sai_put(sai);
1010         CDEBUG(D_READA, "agl thread stopped: [pid %d] [parent %.*s]\n",
1011                current_pid(), parent->d_name.len, parent->d_name.name);
1012         return 0;
1013 }
1014
1015 static void ll_start_agl(struct dentry *parent, struct ll_statahead_info *sai)
1016 {
1017         struct ptlrpc_thread *thread = &sai->sai_agl_thread;
1018         struct l_wait_info    lwi    = { 0 };
1019         struct ll_inode_info  *plli;
1020         struct task_struct *task;
1021
1022         CDEBUG(D_READA, "start agl thread: [pid %d] [parent %.*s]\n",
1023                current_pid(), parent->d_name.len, parent->d_name.name);
1024
1025         plli = ll_i2info(parent->d_inode);
1026         task = kthread_run(ll_agl_thread, parent,
1027                                "ll_agl_%u", plli->lli_opendir_pid);
1028         if (IS_ERR(task)) {
1029                 CERROR("can't start ll_agl thread, rc: %ld\n", PTR_ERR(task));
1030                 thread_set_flags(thread, SVC_STOPPED);
1031                 return;
1032         }
1033
1034         l_wait_event(thread->t_ctl_waitq,
1035                      thread_is_running(thread) || thread_is_stopped(thread),
1036                      &lwi);
1037 }
1038
1039 static int ll_statahead_thread(void *arg)
1040 {
1041         struct dentry       *parent = (struct dentry *)arg;
1042         struct inode         *dir    = parent->d_inode;
1043         struct ll_inode_info     *plli   = ll_i2info(dir);
1044         struct ll_inode_info     *clli;
1045         struct ll_sb_info       *sbi    = ll_i2sbi(dir);
1046         struct ll_statahead_info *sai    = ll_sai_get(plli->lli_sai);
1047         struct ptlrpc_thread     *thread = &sai->sai_thread;
1048         struct ptlrpc_thread *agl_thread = &sai->sai_agl_thread;
1049         struct page           *page;
1050         __u64                pos    = 0;
1051         int                    first  = 0;
1052         int                    rc     = 0;
1053         struct ll_dir_chain       chain;
1054         struct l_wait_info      lwi    = { 0 };
1055
1056         CDEBUG(D_READA, "statahead thread started: [pid %d] [parent %.*s]\n",
1057                current_pid(), parent->d_name.len, parent->d_name.name);
1058
1059         if (sbi->ll_flags & LL_SBI_AGL_ENABLED)
1060                 ll_start_agl(parent, sai);
1061
1062         atomic_inc(&sbi->ll_sa_total);
1063         spin_lock(&plli->lli_sa_lock);
1064         thread_set_flags(thread, SVC_RUNNING);
1065         spin_unlock(&plli->lli_sa_lock);
1066         wake_up(&thread->t_ctl_waitq);
1067
1068         ll_dir_chain_init(&chain);
1069         page = ll_get_dir_page(dir, pos, &chain);
1070
1071         while (1) {
1072                 struct lu_dirpage *dp;
1073                 struct lu_dirent  *ent;
1074
1075                 if (IS_ERR(page)) {
1076                         rc = PTR_ERR(page);
1077                         CDEBUG(D_READA, "error reading dir "DFID" at "LPU64
1078                                "/"LPU64": [rc %d] [parent %u]\n",
1079                                PFID(ll_inode2fid(dir)), pos, sai->sai_index,
1080                                rc, plli->lli_opendir_pid);
1081                         GOTO(out, rc);
1082                 }
1083
1084                 dp = page_address(page);
1085                 for (ent = lu_dirent_start(dp); ent != NULL;
1086                      ent = lu_dirent_next(ent)) {
1087                         __u64 hash;
1088                         int namelen;
1089                         char *name;
1090
1091                         hash = le64_to_cpu(ent->lde_hash);
1092                         if (unlikely(hash < pos))
1093                                 /*
1094                                  * Skip until we find target hash value.
1095                                  */
1096                                 continue;
1097
1098                         namelen = le16_to_cpu(ent->lde_namelen);
1099                         if (unlikely(namelen == 0))
1100                                 /*
1101                                  * Skip dummy record.
1102                                  */
1103                                 continue;
1104
1105                         name = ent->lde_name;
1106                         if (name[0] == '.') {
1107                                 if (namelen == 1) {
1108                                         /*
1109                                          * skip "."
1110                                          */
1111                                         continue;
1112                                 } else if (name[1] == '.' && namelen == 2) {
1113                                         /*
1114                                          * skip ".."
1115                                          */
1116                                         continue;
1117                                 } else if (!sai->sai_ls_all) {
1118                                         /*
1119                                          * skip hidden files.
1120                                          */
1121                                         sai->sai_skip_hidden++;
1122                                         continue;
1123                                 }
1124                         }
1125
1126                         /*
1127                          * don't stat-ahead first entry.
1128                          */
1129                         if (unlikely(++first == 1))
1130                                 continue;
1131
1132 keep_it:
1133                         l_wait_event(thread->t_ctl_waitq,
1134                                      !sa_sent_full(sai) ||
1135                                      !sa_received_empty(sai) ||
1136                                      !agl_list_empty(sai) ||
1137                                      !thread_is_running(thread),
1138                                      &lwi);
1139
1140 interpret_it:
1141                         while (!sa_received_empty(sai))
1142                                 ll_post_statahead(sai);
1143
1144                         if (unlikely(!thread_is_running(thread))) {
1145                                 ll_release_page(page, 0);
1146                                 GOTO(out, rc = 0);
1147                         }
1148
1149                         /* If no window for metadata statahead, but there are
1150                          * some AGL entries to be triggered, then try to help
1151                          * to process the AGL entries. */
1152                         if (sa_sent_full(sai)) {
1153                                 spin_lock(&plli->lli_agl_lock);
1154                                 while (!agl_list_empty(sai)) {
1155                                         clli = agl_first_entry(sai);
1156                                         list_del_init(&clli->lli_agl_list);
1157                                         spin_unlock(&plli->lli_agl_lock);
1158                                         ll_agl_trigger(&clli->lli_vfs_inode,
1159                                                        sai);
1160
1161                                         if (!sa_received_empty(sai))
1162                                                 goto interpret_it;
1163
1164                                         if (unlikely(
1165                                                 !thread_is_running(thread))) {
1166                                                 ll_release_page(page, 0);
1167                                                 GOTO(out, rc = 0);
1168                                         }
1169
1170                                         if (!sa_sent_full(sai))
1171                                                 goto do_it;
1172
1173                                         spin_lock(&plli->lli_agl_lock);
1174                                 }
1175                                 spin_unlock(&plli->lli_agl_lock);
1176
1177                                 goto keep_it;
1178                         }
1179
1180 do_it:
1181                         ll_statahead_one(parent, name, namelen);
1182                 }
1183                 pos = le64_to_cpu(dp->ldp_hash_end);
1184                 if (pos == MDS_DIR_END_OFF) {
1185                         /*
1186                          * End of directory reached.
1187                          */
1188                         ll_release_page(page, 0);
1189                         while (1) {
1190                                 l_wait_event(thread->t_ctl_waitq,
1191                                              !sa_received_empty(sai) ||
1192                                              sai->sai_sent == sai->sai_replied||
1193                                              !thread_is_running(thread),
1194                                              &lwi);
1195
1196                                 while (!sa_received_empty(sai))
1197                                         ll_post_statahead(sai);
1198
1199                                 if (unlikely(!thread_is_running(thread)))
1200                                         GOTO(out, rc = 0);
1201
1202                                 if (sai->sai_sent == sai->sai_replied &&
1203                                     sa_received_empty(sai))
1204                                         break;
1205                         }
1206
1207                         spin_lock(&plli->lli_agl_lock);
1208                         while (!agl_list_empty(sai) &&
1209                                thread_is_running(thread)) {
1210                                 clli = agl_first_entry(sai);
1211                                 list_del_init(&clli->lli_agl_list);
1212                                 spin_unlock(&plli->lli_agl_lock);
1213                                 ll_agl_trigger(&clli->lli_vfs_inode, sai);
1214                                 spin_lock(&plli->lli_agl_lock);
1215                         }
1216                         spin_unlock(&plli->lli_agl_lock);
1217
1218                         GOTO(out, rc = 0);
1219                 } else if (1) {
1220                         /*
1221                          * chain is exhausted.
1222                          * Normal case: continue to the next page.
1223                          */
1224                         ll_release_page(page, le32_to_cpu(dp->ldp_flags) &
1225                                               LDF_COLLIDE);
1226                         sai->sai_in_readpage = 1;
1227                         page = ll_get_dir_page(dir, pos, &chain);
1228                         sai->sai_in_readpage = 0;
1229                 } else {
1230                         LASSERT(le32_to_cpu(dp->ldp_flags) & LDF_COLLIDE);
1231                         ll_release_page(page, 1);
1232                         /*
1233                          * go into overflow page.
1234                          */
1235                 }
1236         }
1237
1238 out:
1239         if (sai->sai_agl_valid) {
1240                 spin_lock(&plli->lli_agl_lock);
1241                 thread_set_flags(agl_thread, SVC_STOPPING);
1242                 spin_unlock(&plli->lli_agl_lock);
1243                 wake_up(&agl_thread->t_ctl_waitq);
1244
1245                 CDEBUG(D_READA, "stop agl thread: [pid %d]\n",
1246                        current_pid());
1247                 l_wait_event(agl_thread->t_ctl_waitq,
1248                              thread_is_stopped(agl_thread),
1249                              &lwi);
1250         } else {
1251                 /* Set agl_thread flags anyway. */
1252                 thread_set_flags(&sai->sai_agl_thread, SVC_STOPPED);
1253         }
1254         ll_dir_chain_fini(&chain);
1255         spin_lock(&plli->lli_sa_lock);
1256         if (!sa_received_empty(sai)) {
1257                 thread_set_flags(thread, SVC_STOPPING);
1258                 spin_unlock(&plli->lli_sa_lock);
1259
1260                 /* To release the resources held by received entries. */
1261                 while (!sa_received_empty(sai))
1262                         ll_post_statahead(sai);
1263
1264                 spin_lock(&plli->lli_sa_lock);
1265         }
1266         thread_set_flags(thread, SVC_STOPPED);
1267         spin_unlock(&plli->lli_sa_lock);
1268         wake_up(&sai->sai_waitq);
1269         wake_up(&thread->t_ctl_waitq);
1270         ll_sai_put(sai);
1271         dput(parent);
1272         CDEBUG(D_READA, "statahead thread stopped: [pid %d] [parent %.*s]\n",
1273                current_pid(), parent->d_name.len, parent->d_name.name);
1274         return rc;
1275 }
1276
1277 /**
1278  * called in ll_file_release().
1279  */
1280 void ll_stop_statahead(struct inode *dir, void *key)
1281 {
1282         struct ll_inode_info *lli = ll_i2info(dir);
1283
1284         if (unlikely(key == NULL))
1285                 return;
1286
1287         spin_lock(&lli->lli_sa_lock);
1288         if (lli->lli_opendir_key != key || lli->lli_opendir_pid == 0) {
1289                 spin_unlock(&lli->lli_sa_lock);
1290                 return;
1291         }
1292
1293         lli->lli_opendir_key = NULL;
1294
1295         if (lli->lli_sai) {
1296                 struct l_wait_info lwi = { 0 };
1297                 struct ptlrpc_thread *thread = &lli->lli_sai->sai_thread;
1298
1299                 if (!thread_is_stopped(thread)) {
1300                         thread_set_flags(thread, SVC_STOPPING);
1301                         spin_unlock(&lli->lli_sa_lock);
1302                         wake_up(&thread->t_ctl_waitq);
1303
1304                         CDEBUG(D_READA, "stop statahead thread: [pid %d]\n",
1305                                current_pid());
1306                         l_wait_event(thread->t_ctl_waitq,
1307                                      thread_is_stopped(thread),
1308                                      &lwi);
1309                 } else {
1310                         spin_unlock(&lli->lli_sa_lock);
1311                 }
1312
1313                 /*
1314                  * Put the ref which was held when first statahead_enter.
1315                  * It maybe not the last ref for some statahead requests
1316                  * maybe inflight.
1317                  */
1318                 ll_sai_put(lli->lli_sai);
1319         } else {
1320                 lli->lli_opendir_pid = 0;
1321                 spin_unlock(&lli->lli_sa_lock);
1322         }
1323 }
1324
1325 enum {
1326         /**
1327          * not first dirent, or is "."
1328          */
1329         LS_NONE_FIRST_DE = 0,
1330         /**
1331          * the first non-hidden dirent
1332          */
1333         LS_FIRST_DE,
1334         /**
1335          * the first hidden dirent, that is "."
1336          */
1337         LS_FIRST_DOT_DE
1338 };
1339
1340 static int is_first_dirent(struct inode *dir, struct dentry *dentry)
1341 {
1342         struct ll_dir_chain   chain;
1343         struct qstr       *target = &dentry->d_name;
1344         struct page       *page;
1345         __u64            pos    = 0;
1346         int                dot_de;
1347         int                rc     = LS_NONE_FIRST_DE;
1348
1349         ll_dir_chain_init(&chain);
1350         page = ll_get_dir_page(dir, pos, &chain);
1351
1352         while (1) {
1353                 struct lu_dirpage *dp;
1354                 struct lu_dirent  *ent;
1355
1356                 if (IS_ERR(page)) {
1357                         struct ll_inode_info *lli = ll_i2info(dir);
1358
1359                         rc = PTR_ERR(page);
1360                         CERROR("error reading dir "DFID" at "LPU64": "
1361                                "[rc %d] [parent %u]\n",
1362                                PFID(ll_inode2fid(dir)), pos,
1363                                rc, lli->lli_opendir_pid);
1364                         break;
1365                 }
1366
1367                 dp = page_address(page);
1368                 for (ent = lu_dirent_start(dp); ent != NULL;
1369                      ent = lu_dirent_next(ent)) {
1370                         __u64 hash;
1371                         int namelen;
1372                         char *name;
1373
1374                         hash = le64_to_cpu(ent->lde_hash);
1375                         /* The ll_get_dir_page() can return any page containing
1376                          * the given hash which may be not the start hash. */
1377                         if (unlikely(hash < pos))
1378                                 continue;
1379
1380                         namelen = le16_to_cpu(ent->lde_namelen);
1381                         if (unlikely(namelen == 0))
1382                                 /*
1383                                  * skip dummy record.
1384                                  */
1385                                 continue;
1386
1387                         name = ent->lde_name;
1388                         if (name[0] == '.') {
1389                                 if (namelen == 1)
1390                                         /*
1391                                          * skip "."
1392                                          */
1393                                         continue;
1394                                 else if (name[1] == '.' && namelen == 2)
1395                                         /*
1396                                          * skip ".."
1397                                          */
1398                                         continue;
1399                                 else
1400                                         dot_de = 1;
1401                         } else {
1402                                 dot_de = 0;
1403                         }
1404
1405                         if (dot_de && target->name[0] != '.') {
1406                                 CDEBUG(D_READA, "%.*s skip hidden file %.*s\n",
1407                                        target->len, target->name,
1408                                        namelen, name);
1409                                 continue;
1410                         }
1411
1412                         if (target->len != namelen ||
1413                             memcmp(target->name, name, namelen) != 0)
1414                                 rc = LS_NONE_FIRST_DE;
1415                         else if (!dot_de)
1416                                 rc = LS_FIRST_DE;
1417                         else
1418                                 rc = LS_FIRST_DOT_DE;
1419
1420                         ll_release_page(page, 0);
1421                         GOTO(out, rc);
1422                 }
1423                 pos = le64_to_cpu(dp->ldp_hash_end);
1424                 if (pos == MDS_DIR_END_OFF) {
1425                         /*
1426                          * End of directory reached.
1427                          */
1428                         ll_release_page(page, 0);
1429                         break;
1430                 } else if (1) {
1431                         /*
1432                          * chain is exhausted
1433                          * Normal case: continue to the next page.
1434                          */
1435                         ll_release_page(page, le32_to_cpu(dp->ldp_flags) &
1436                                               LDF_COLLIDE);
1437                         page = ll_get_dir_page(dir, pos, &chain);
1438                 } else {
1439                         /*
1440                          * go into overflow page.
1441                          */
1442                         LASSERT(le32_to_cpu(dp->ldp_flags) & LDF_COLLIDE);
1443                         ll_release_page(page, 1);
1444                 }
1445         }
1446
1447 out:
1448         ll_dir_chain_fini(&chain);
1449         return rc;
1450 }
1451
1452 static void
1453 ll_sai_unplug(struct ll_statahead_info *sai, struct ll_sa_entry *entry)
1454 {
1455         struct ptlrpc_thread *thread = &sai->sai_thread;
1456         struct ll_sb_info    *sbi    = ll_i2sbi(sai->sai_inode);
1457         int                hit;
1458
1459         if (entry != NULL && entry->se_stat == SA_ENTRY_SUCC)
1460                 hit = 1;
1461         else
1462                 hit = 0;
1463
1464         ll_sa_entry_fini(sai, entry);
1465         if (hit) {
1466                 sai->sai_hit++;
1467                 sai->sai_consecutive_miss = 0;
1468                 sai->sai_max = min(2 * sai->sai_max, sbi->ll_sa_max);
1469         } else {
1470                 struct ll_inode_info *lli = ll_i2info(sai->sai_inode);
1471
1472                 sai->sai_miss++;
1473                 sai->sai_consecutive_miss++;
1474                 if (sa_low_hit(sai) && thread_is_running(thread)) {
1475                         atomic_inc(&sbi->ll_sa_wrong);
1476                         CDEBUG(D_READA, "Statahead for dir "DFID" hit "
1477                                "ratio too low: hit/miss "LPU64"/"LPU64
1478                                ", sent/replied "LPU64"/"LPU64", stopping "
1479                                "statahead thread: pid %d\n",
1480                                PFID(&lli->lli_fid), sai->sai_hit,
1481                                sai->sai_miss, sai->sai_sent,
1482                                sai->sai_replied, current_pid());
1483                         spin_lock(&lli->lli_sa_lock);
1484                         if (!thread_is_stopped(thread))
1485                                 thread_set_flags(thread, SVC_STOPPING);
1486                         spin_unlock(&lli->lli_sa_lock);
1487                 }
1488         }
1489
1490         if (!thread_is_stopped(thread))
1491                 wake_up(&thread->t_ctl_waitq);
1492 }
1493
1494 /**
1495  * Start statahead thread if this is the first dir entry.
1496  * Otherwise if a thread is started already, wait it until it is ahead of me.
1497  * \retval 1       -- find entry with lock in cache, the caller needs to do
1498  *                  nothing.
1499  * \retval 0       -- find entry in cache, but without lock, the caller needs
1500  *                  refresh from MDS.
1501  * \retval others  -- the caller need to process as non-statahead.
1502  */
1503 int do_statahead_enter(struct inode *dir, struct dentry **dentryp,
1504                        int only_unplug)
1505 {
1506         struct ll_inode_info     *lli   = ll_i2info(dir);
1507         struct ll_statahead_info *sai   = lli->lli_sai;
1508         struct dentry       *parent;
1509         struct ll_sa_entry       *entry;
1510         struct ptlrpc_thread     *thread;
1511         struct l_wait_info      lwi   = { 0 };
1512         int                    rc    = 0;
1513         struct ll_inode_info     *plli;
1514
1515         LASSERT(lli->lli_opendir_pid == current_pid());
1516
1517         if (sai) {
1518                 thread = &sai->sai_thread;
1519                 if (unlikely(thread_is_stopped(thread) &&
1520                              list_empty(&sai->sai_entries_stated))) {
1521                         /* to release resource */
1522                         ll_stop_statahead(dir, lli->lli_opendir_key);
1523                         return -EAGAIN;
1524                 }
1525
1526                 if ((*dentryp)->d_name.name[0] == '.') {
1527                         if (sai->sai_ls_all ||
1528                             sai->sai_miss_hidden >= sai->sai_skip_hidden) {
1529                                 /*
1530                                  * Hidden dentry is the first one, or statahead
1531                                  * thread does not skip so many hidden dentries
1532                                  * before "sai_ls_all" enabled as below.
1533                                  */
1534                         } else {
1535                                 if (!sai->sai_ls_all)
1536                                         /*
1537                                          * It maybe because hidden dentry is not
1538                                          * the first one, "sai_ls_all" was not
1539                                          * set, then "ls -al" missed. Enable
1540                                          * "sai_ls_all" for such case.
1541                                          */
1542                                         sai->sai_ls_all = 1;
1543
1544                                 /*
1545                                  * Such "getattr" has been skipped before
1546                                  * "sai_ls_all" enabled as above.
1547                                  */
1548                                 sai->sai_miss_hidden++;
1549                                 return -EAGAIN;
1550                         }
1551                 }
1552
1553                 entry = ll_sa_entry_get_byname(sai, &(*dentryp)->d_name);
1554                 if (entry == NULL || only_unplug) {
1555                         ll_sai_unplug(sai, entry);
1556                         return entry ? 1 : -EAGAIN;
1557                 }
1558
1559                 /* if statahead is busy in readdir, help it do post-work */
1560                 while (!ll_sa_entry_stated(entry) &&
1561                        sai->sai_in_readpage &&
1562                        !sa_received_empty(sai))
1563                         ll_post_statahead(sai);
1564
1565                 if (!ll_sa_entry_stated(entry)) {
1566                         sai->sai_index_wait = entry->se_index;
1567                         lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(30), NULL,
1568                                                LWI_ON_SIGNAL_NOOP, NULL);
1569                         rc = l_wait_event(sai->sai_waitq,
1570                                           ll_sa_entry_stated(entry) ||
1571                                           thread_is_stopped(thread),
1572                                           &lwi);
1573                         if (rc < 0) {
1574                                 ll_sai_unplug(sai, entry);
1575                                 return -EAGAIN;
1576                         }
1577                 }
1578
1579                 if (entry->se_stat == SA_ENTRY_SUCC &&
1580                     entry->se_inode != NULL) {
1581                         struct inode *inode = entry->se_inode;
1582                         struct lookup_intent it = { .it_op = IT_GETATTR,
1583                                                     .d.lustre.it_lock_handle =
1584                                                      entry->se_handle };
1585                         __u64 bits;
1586
1587                         rc = md_revalidate_lock(ll_i2mdexp(dir), &it,
1588                                                 ll_inode2fid(inode), &bits);
1589                         if (rc == 1) {
1590                                 if ((*dentryp)->d_inode == NULL) {
1591                                         *dentryp = ll_splice_alias(inode,
1592                                                                    *dentryp);
1593                                 } else if ((*dentryp)->d_inode != inode) {
1594                                         /* revalidate, but inode is recreated */
1595                                         CDEBUG(D_READA,
1596                                               "stale dentry %.*s inode %lu/%u, "
1597                                               "statahead inode %lu/%u\n",
1598                                               (*dentryp)->d_name.len,
1599                                               (*dentryp)->d_name.name,
1600                                               (*dentryp)->d_inode->i_ino,
1601                                               (*dentryp)->d_inode->i_generation,
1602                                               inode->i_ino,
1603                                               inode->i_generation);
1604                                         ll_sai_unplug(sai, entry);
1605                                         return -ESTALE;
1606                                 } else {
1607                                         iput(inode);
1608                                 }
1609                                 entry->se_inode = NULL;
1610
1611                                 if ((bits & MDS_INODELOCK_LOOKUP) &&
1612                                     d_lustre_invalid(*dentryp))
1613                                         d_lustre_revalidate(*dentryp);
1614                                 ll_intent_release(&it);
1615                         }
1616                 }
1617
1618                 ll_sai_unplug(sai, entry);
1619                 return rc;
1620         }
1621
1622         /* I am the "lli_opendir_pid" owner, only me can set "lli_sai". */
1623         rc = is_first_dirent(dir, *dentryp);
1624         if (rc == LS_NONE_FIRST_DE)
1625                 /* It is not "ls -{a}l" operation, no need statahead for it. */
1626                 GOTO(out, rc = -EAGAIN);
1627
1628         sai = ll_sai_alloc();
1629         if (sai == NULL)
1630                 GOTO(out, rc = -ENOMEM);
1631
1632         sai->sai_ls_all = (rc == LS_FIRST_DOT_DE);
1633         sai->sai_inode = igrab(dir);
1634         if (unlikely(sai->sai_inode == NULL)) {
1635                 CWARN("Do not start stat ahead on dying inode "DFID"\n",
1636                       PFID(&lli->lli_fid));
1637                 GOTO(out, rc = -ESTALE);
1638         }
1639
1640         /* get parent reference count here, and put it in ll_statahead_thread */
1641         parent = dget((*dentryp)->d_parent);
1642         if (unlikely(sai->sai_inode != parent->d_inode)) {
1643                 struct ll_inode_info *nlli = ll_i2info(parent->d_inode);
1644
1645                 CWARN("Race condition, someone changed %.*s just now: "
1646                       "old parent "DFID", new parent "DFID"\n",
1647                       (*dentryp)->d_name.len, (*dentryp)->d_name.name,
1648                       PFID(&lli->lli_fid), PFID(&nlli->lli_fid));
1649                 dput(parent);
1650                 iput(sai->sai_inode);
1651                 GOTO(out, rc = -EAGAIN);
1652         }
1653
1654         CDEBUG(D_READA, "start statahead thread: [pid %d] [parent %.*s]\n",
1655                current_pid(), parent->d_name.len, parent->d_name.name);
1656
1657         lli->lli_sai = sai;
1658
1659         plli = ll_i2info(parent->d_inode);
1660         rc = PTR_ERR(kthread_run(ll_statahead_thread, parent,
1661                                  "ll_sa_%u", plli->lli_opendir_pid));
1662         thread = &sai->sai_thread;
1663         if (IS_ERR_VALUE(rc)) {
1664                 CERROR("can't start ll_sa thread, rc: %d\n", rc);
1665                 dput(parent);
1666                 lli->lli_opendir_key = NULL;
1667                 thread_set_flags(thread, SVC_STOPPED);
1668                 thread_set_flags(&sai->sai_agl_thread, SVC_STOPPED);
1669                 ll_sai_put(sai);
1670                 LASSERT(lli->lli_sai == NULL);
1671                 return -EAGAIN;
1672         }
1673
1674         l_wait_event(thread->t_ctl_waitq,
1675                      thread_is_running(thread) || thread_is_stopped(thread),
1676                      &lwi);
1677
1678         /*
1679          * We don't stat-ahead for the first dirent since we are already in
1680          * lookup.
1681          */
1682         return -EAGAIN;
1683
1684 out:
1685         if (sai != NULL)
1686                 OBD_FREE_PTR(sai);
1687         spin_lock(&lli->lli_sa_lock);
1688         lli->lli_opendir_key = NULL;
1689         lli->lli_opendir_pid = 0;
1690         spin_unlock(&lli->lli_sa_lock);
1691         return rc;
1692 }