lustre/llite/statahead.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
  30  * Use is subject to license terms.
  31  *
  32  * Copyright (c) 2011 Whamcloud, Inc.
  33  *
  34  */
  35 /*
  36  * This file is part of Lustre, http://www.lustre.org/
  37  * Lustre is a trademark of Sun Microsystems, Inc.
  38  */
  39
  40 #include <linux/fs.h>
  41 #include <linux/sched.h>
  42 #include <linux/mm.h>
  43 #include <linux/smp_lock.h>
  44 #include <linux/highmem.h>
  45 #include <linux/pagemap.h>
  46
  47 #define DEBUG_SUBSYSTEM S_LLITE
  48
  49 #include <obd_support.h>
  50 #include <lustre_lite.h>
  51 #include <lustre_dlm.h>
  52 #include <linux/lustre_version.h>
  53 #include "llite_internal.h"
  54
  55 #define SA_OMITTED_ENTRY_MAX 8ULL
  56
  57 typedef enum {
  58         /** negative values are for error cases */
  59         SA_ENTRY_INIT = 0,      /** init entry */
  60         SA_ENTRY_SUCC = 1,      /** stat succeed */
  61         SA_ENTRY_INVA = 2,      /** invalid entry */
  62         SA_ENTRY_DEST = 3,      /** entry to be destroyed */
  63 } se_stat_t;
  64
  65 struct ll_sa_entry {
  66         /* link into sai->sai_entries_{sent,received,stated} */
  67         cfs_list_t              se_list;
  68         /* link into sai hash table locally */
  69         cfs_list_t              se_hash;
  70         /* entry reference count */
  71         cfs_atomic_t            se_refcount;
  72         /* entry index in the sai */
  73         __u64                   se_index;
  74         /* low layer ldlm lock handle */
  75         __u64                   se_handle;
  76         /* entry status */
  77         se_stat_t               se_stat;
  78         /* entry size, contains name */
  79         int                     se_size;
  80         /* pointer to async getattr enqueue info */
  81         struct md_enqueue_info *se_minfo;
  82         /* pointer to the async getattr request */
  83         struct ptlrpc_request  *se_req;
  84         /* pointer to the target inode */
  85         struct inode           *se_inode;
  86         /* entry name */
  87         struct qstr             se_qstr;
  88 };
  89
  90 static unsigned int sai_generation = 0;
  91 static cfs_spinlock_t sai_generation_lock = CFS_SPIN_LOCK_UNLOCKED;
  92
  93 static inline int ll_sa_entry_unlinked(struct ll_sa_entry *entry)
  94 {
  95         return cfs_list_empty(&entry->se_list);
  96 }
  97
  98 static inline int ll_sa_entry_unhashed(struct ll_sa_entry *entry)
  99 {
 100         return cfs_list_empty(&entry->se_hash);
 101 }
 102
 103 /*
 104  * The entry only can be released by the caller, it is necessary to hold lock.
 105  */
 106 static inline int ll_sa_entry_stated(struct ll_sa_entry *entry)
 107 {
 108         smp_rmb();
 109         return (entry->se_stat != SA_ENTRY_INIT);
 110 }
 111
 112 static inline int ll_sa_entry_hash(int val)
 113 {
 114         return val & LL_SA_CACHE_MASK;
 115 }
 116
 117 /*
 118  * Insert entry to hash SA table.
 119  */
 120 static inline void
 121 ll_sa_entry_enhash(struct ll_statahead_info *sai, struct ll_sa_entry *entry)
 122 {
 123         int i = ll_sa_entry_hash(entry->se_qstr.hash);
 124
 125         cfs_spin_lock(&sai->sai_cache_lock[i]);
 126         cfs_list_add_tail(&entry->se_hash, &sai->sai_cache[i]);
 127         cfs_spin_unlock(&sai->sai_cache_lock[i]);
 128 }
 129
 130 /*
 131  * Remove entry from SA table.
 132  */
 133 static inline void
 134 ll_sa_entry_unhash(struct ll_statahead_info *sai, struct ll_sa_entry *entry)
 135 {
 136         int i = ll_sa_entry_hash(entry->se_qstr.hash);
 137
 138         cfs_spin_lock(&sai->sai_cache_lock[i]);
 139         cfs_list_del_init(&entry->se_hash);
 140         cfs_spin_unlock(&sai->sai_cache_lock[i]);
 141 }
 142
 143 static inline int agl_should_run(struct ll_statahead_info *sai,
 144                                  struct inode *inode)
 145 {
 146         if (inode != NULL && S_ISREG(inode->i_mode) &&
 147             ll_i2info(inode)->lli_smd != NULL && sai->sai_agl_valid)
 148                 return 1;
 149         return 0;
 150 }
 151
 152 static inline struct ll_sa_entry *
 153 sa_first_received_entry(struct ll_statahead_info *sai)
 154 {
 155         return cfs_list_entry(sai->sai_entries_received.next,
 156                               struct ll_sa_entry, se_list);
 157 }
 158
 159 static inline struct ll_inode_info *
 160 agl_first_entry(struct ll_statahead_info *sai)
 161 {
 162         return cfs_list_entry(sai->sai_entries_agl.next,
 163                               struct ll_inode_info, lli_agl_list);
 164 }
 165
 166 static inline int sa_sent_full(struct ll_statahead_info *sai)
 167 {
 168         return cfs_atomic_read(&sai->sai_cache_count) >= sai->sai_max;
 169 }
 170
 171 static inline int sa_received_empty(struct ll_statahead_info *sai)
 172 {
 173         return cfs_list_empty(&sai->sai_entries_received);
 174 }
 175
 176 static inline int agl_list_empty(struct ll_statahead_info *sai)
 177 {
 178         return cfs_list_empty(&sai->sai_entries_agl);
 179 }
 180
 181 /**
 182  * (1) hit ratio less than 80%
 183  * or
 184  * (2) consecutive miss more than 8
 185  * then means low hit.
 186  */
 187 static inline int sa_low_hit(struct ll_statahead_info *sai)
 188 {
 189         return ((sai->sai_hit > 7 && sai->sai_hit < 4 * sai->sai_miss) ||
 190                 (sai->sai_consecutive_miss > 8));
 191 }
 192
 193 /*
 194  * If the given index is behind of statahead window more than
 195  * SA_OMITTED_ENTRY_MAX, then it is old.
 196  */
 197 static inline int is_omitted_entry(struct ll_statahead_info *sai, __u64 index)
 198 {
 199         return ((__u64)sai->sai_max + index + SA_OMITTED_ENTRY_MAX <
 200                  sai->sai_index);
 201 }
 202
 203 /*
 204  * Insert it into sai_entries_sent tail when init.
 205  */
 206 static struct ll_sa_entry *
 207 ll_sa_entry_alloc(struct ll_statahead_info *sai, __u64 index,
 208                   const char *name, int len)
 209 {
 210         struct ll_inode_info *lli;
 211         struct ll_sa_entry   *entry;
 212         int                   entry_size;
 213         char                 *dname;
 214         ENTRY;
 215
 216         entry_size = sizeof(struct ll_sa_entry) + (len & ~3) + 4;
 217         OBD_ALLOC(entry, entry_size);
 218         if (unlikely(entry == NULL))
 219                 RETURN(ERR_PTR(-ENOMEM));
 220
 221         CDEBUG(D_READA, "alloc sai entry %.*s(%p) index "LPU64"\n",
 222                len, name, entry, index);
 223
 224         entry->se_index = index;
 225
 226         /*
 227          * Statahead entry reference rules:
 228          *
 229          * 1) When statahead entry is initialized, its reference is set as 2.
 230          *    One reference is used by the directory scanner. When the scanner
 231          *    searches the statahead cache for the given name, it can perform
 232          *    lockless hash lookup (only the scanner can remove entry from hash
 233          *    list), and once found, it needn't to call "atomic_inc()" for the
 234          *    entry reference. So the performance is improved. After using the
 235          *    statahead entry, the scanner will call "atomic_dec()" to drop the
 236          *    reference held when initialization. If it is the last reference,
 237          *    the statahead entry will be freed.
 238          *
 239          * 2) All other threads, including statahead thread and ptlrpcd thread,
 240          *    when they process the statahead entry, the reference for target
 241          *    should be held to guarantee the entry will not be released by the
 242          *    directory scanner. After processing the entry, these threads will
 243          *    drop the entry reference. If it is the last reference, the entry
 244          *    will be freed.
 245          *
 246          *    The second reference when initializes the statahead entry is used
 247          *    by the statahead thread, following the rule 2).
 248          */
 249         cfs_atomic_set(&entry->se_refcount, 2);
 250         entry->se_stat = SA_ENTRY_INIT;
 251         entry->se_size = entry_size;
 252         dname = (char *)entry + sizeof(struct ll_sa_entry);
 253         memcpy(dname, name, len);
 254         dname[len] = 0;
 255         entry->se_qstr.hash = full_name_hash(name, len);
 256         entry->se_qstr.len = len;
 257         entry->se_qstr.name = dname;
 258
 259         lli = ll_i2info(sai->sai_inode);
 260         cfs_spin_lock(&lli->lli_sa_lock);
 261         cfs_list_add_tail(&entry->se_list, &sai->sai_entries_sent);
 262         cfs_spin_unlock(&lli->lli_sa_lock);
 263
 264         cfs_atomic_inc(&sai->sai_cache_count);
 265         ll_sa_entry_enhash(sai, entry);
 266
 267         RETURN(entry);
 268 }
 269
 270 /*
 271  * Used by the directory scanner to search entry with name.
 272  *
 273  * Only the caller can remove the entry from hash, so it is unnecessary to hold
 274  * hash lock. It is caller's duty to release the init refcount on the entry, so
 275  * it is also unnecessary to increase refcount on the entry.
 276  */
 277 static struct ll_sa_entry *
 278 ll_sa_entry_get_byname(struct ll_statahead_info *sai, const struct qstr *qstr)
 279 {
 280         struct ll_sa_entry *entry;
 281         int i = ll_sa_entry_hash(qstr->hash);
 282
 283         cfs_list_for_each_entry(entry, &sai->sai_cache[i], se_hash) {
 284                 if (entry->se_qstr.hash == qstr->hash &&
 285                     entry->se_qstr.len == qstr->len &&
 286                     memcmp(entry->se_qstr.name, qstr->name, qstr->len) == 0)
 287                         return entry;
 288         }
 289         return NULL;
 290 }
 291
 292 /*
 293  * Used by the async getattr request callback to find entry with index.
 294  *
 295  * Inside lli_sa_lock to prevent others to change the list during the search.
 296  * It needs to increase entry refcount before returning to guarantee that the
 297  * entry cannot be freed by others.
 298  */
 299 static struct ll_sa_entry *
 300 ll_sa_entry_get_byindex(struct ll_statahead_info *sai, __u64 index)
 301 {
 302         struct ll_sa_entry *entry;
 303
 304         cfs_list_for_each_entry(entry, &sai->sai_entries_sent, se_list) {
 305                 if (entry->se_index == index) {
 306                         cfs_atomic_inc(&entry->se_refcount);
 307                         return entry;
 308                 }
 309                 if (entry->se_index > index)
 310                         break;
 311         }
 312         return NULL;
 313 }
 314
 315 static void ll_sa_entry_cleanup(struct ll_statahead_info *sai,
 316                                  struct ll_sa_entry *entry)
 317 {
 318         struct md_enqueue_info *minfo = entry->se_minfo;
 319         struct ptlrpc_request  *req   = entry->se_req;
 320
 321         if (minfo) {
 322                 entry->se_minfo = NULL;
 323                 ll_intent_release(&minfo->mi_it);
 324                 iput(minfo->mi_dir);
 325                 OBD_FREE_PTR(minfo);
 326         }
 327
 328         if (req) {
 329                 entry->se_req = NULL;
 330                 ptlrpc_req_finished(req);
 331         }
 332 }
 333
 334 static void ll_sa_entry_put(struct ll_statahead_info *sai,
 335                              struct ll_sa_entry *entry)
 336 {
 337         if (cfs_atomic_dec_and_test(&entry->se_refcount)) {
 338                 CDEBUG(D_READA, "free sai entry %.*s(%p) index "LPU64"\n",
 339                        entry->se_qstr.len, entry->se_qstr.name, entry,
 340                        entry->se_index);
 341
 342                 LASSERT(ll_sa_entry_unhashed(entry));
 343                 LASSERT(ll_sa_entry_unlinked(entry));
 344
 345                 ll_sa_entry_cleanup(sai, entry);
 346                 if (entry->se_inode)
 347                         iput(entry->se_inode);
 348
 349                 OBD_FREE(entry, entry->se_size);
 350                 cfs_atomic_dec(&sai->sai_cache_count);
 351         }
 352 }
 353
 354 static inline void
 355 do_sai_entry_fini(struct ll_statahead_info *sai, struct ll_sa_entry *entry)
 356 {
 357         struct ll_inode_info *lli = ll_i2info(sai->sai_inode);
 358
 359         ll_sa_entry_unhash(sai, entry);
 360
 361         cfs_spin_lock(&lli->lli_sa_lock);
 362         entry->se_stat = SA_ENTRY_DEST;
 363         if (likely(!ll_sa_entry_unlinked(entry)))
 364                 cfs_list_del_init(&entry->se_list);
 365         cfs_spin_unlock(&lli->lli_sa_lock);
 366
 367         ll_sa_entry_put(sai, entry);
 368 }
 369
 370 /*
 371  * Delete it from sai_entries_stated list when fini.
 372  */
 373 static void
 374 ll_sa_entry_fini(struct ll_statahead_info *sai, struct ll_sa_entry *entry)
 375 {
 376         struct ll_sa_entry *pos, *next;
 377
 378         if (entry)
 379                 do_sai_entry_fini(sai, entry);
 380
 381         /* drop old entry from sent list */
 382         cfs_list_for_each_entry_safe(pos, next, &sai->sai_entries_sent,
 383                                      se_list) {
 384                 if (is_omitted_entry(sai, pos->se_index))
 385                         do_sai_entry_fini(sai, pos);
 386                 else
 387                         break;
 388         }
 389
 390         /* drop old entry from stated list */
 391         cfs_list_for_each_entry_safe(pos, next, &sai->sai_entries_stated,
 392                                      se_list) {
 393                 if (is_omitted_entry(sai, pos->se_index))
 394                         do_sai_entry_fini(sai, pos);
 395                 else
 396                         break;
 397         }
 398 }
 399
 400 /*
 401  * Inside lli_sa_lock.
 402  */
 403 static void
 404 do_sai_entry_to_stated(struct ll_statahead_info *sai,
 405                        struct ll_sa_entry *entry, int rc)
 406 {
 407         struct ll_sa_entry *se;
 408         cfs_list_t         *pos = &sai->sai_entries_stated;
 409
 410         if (!ll_sa_entry_unlinked(entry))
 411                 cfs_list_del_init(&entry->se_list);
 412
 413         cfs_list_for_each_entry_reverse(se, &sai->sai_entries_stated, se_list) {
 414                 if (se->se_index < entry->se_index) {
 415                         pos = &se->se_list;
 416                         break;
 417                 }
 418         }
 419
 420         cfs_list_add(&entry->se_list, pos);
 421         entry->se_stat = rc;
 422 }
 423
 424 /*
 425  * Move entry to sai_entries_stated and sort with the index.
 426  * \retval 1    -- entry to be destroyed.
 427  * \retval 0    -- entry is inserted into stated list.
 428  */
 429 static int
 430 ll_sa_entry_to_stated(struct ll_statahead_info *sai,
 431                        struct ll_sa_entry *entry, int rc)
 432 {
 433         struct ll_inode_info *lli = ll_i2info(sai->sai_inode);
 434         int                   ret = 1;
 435
 436         ll_sa_entry_cleanup(sai, entry);
 437
 438         cfs_spin_lock(&lli->lli_sa_lock);
 439         if (likely(entry->se_stat != SA_ENTRY_DEST)) {
 440                 do_sai_entry_to_stated(sai, entry, rc);
 441                 ret = 0;
 442         }
 443         cfs_spin_unlock(&lli->lli_sa_lock);
 444
 445         return ret;
 446 }
 447
 448 /*
 449  * Insert inode into the list of sai_entries_agl.
 450  */
 451 static void ll_agl_add(struct ll_statahead_info *sai,
 452                        struct inode *inode, int index)
 453 {
 454         struct ll_inode_info *child  = ll_i2info(inode);
 455         struct ll_inode_info *parent = ll_i2info(sai->sai_inode);
 456         int                   added  = 0;
 457
 458         cfs_spin_lock(&child->lli_agl_lock);
 459         if (child->lli_agl_index == 0) {
 460                 child->lli_agl_index = index;
 461                 cfs_spin_unlock(&child->lli_agl_lock);
 462
 463                 LASSERT(cfs_list_empty(&child->lli_agl_list));
 464
 465                 igrab(inode);
 466                 cfs_spin_lock(&parent->lli_agl_lock);
 467                 if (agl_list_empty(sai))
 468                         added = 1;
 469                 cfs_list_add_tail(&child->lli_agl_list, &sai->sai_entries_agl);
 470                 cfs_spin_unlock(&parent->lli_agl_lock);
 471         } else {
 472                 cfs_spin_unlock(&child->lli_agl_lock);
 473         }
 474
 475         if (added > 0)
 476                 cfs_waitq_signal(&sai->sai_agl_thread.t_ctl_waitq);
 477 }
 478
 479 static struct ll_statahead_info *ll_sai_alloc(void)
 480 {
 481         struct ll_statahead_info *sai;
 482         int                       i;
 483         ENTRY;
 484
 485         OBD_ALLOC_PTR(sai);
 486         if (!sai)
 487                 RETURN(NULL);
 488
 489         cfs_atomic_set(&sai->sai_refcount, 1);
 490
 491         cfs_spin_lock(&sai_generation_lock);
 492         sai->sai_generation = ++sai_generation;
 493         if (unlikely(sai_generation == 0))
 494                 sai->sai_generation = ++sai_generation;
 495         cfs_spin_unlock(&sai_generation_lock);
 496
 497         sai->sai_max = LL_SA_RPC_MIN;
 498         sai->sai_index = 1;
 499         cfs_waitq_init(&sai->sai_waitq);
 500         cfs_waitq_init(&sai->sai_thread.t_ctl_waitq);
 501         cfs_waitq_init(&sai->sai_agl_thread.t_ctl_waitq);
 502
 503         CFS_INIT_LIST_HEAD(&sai->sai_entries_sent);
 504         CFS_INIT_LIST_HEAD(&sai->sai_entries_received);
 505         CFS_INIT_LIST_HEAD(&sai->sai_entries_stated);
 506         CFS_INIT_LIST_HEAD(&sai->sai_entries_agl);
 507
 508         for (i = 0; i < LL_SA_CACHE_SIZE; i++) {
 509                 CFS_INIT_LIST_HEAD(&sai->sai_cache[i]);
 510                 cfs_spin_lock_init(&sai->sai_cache_lock[i]);
 511         }
 512         cfs_atomic_set(&sai->sai_cache_count, 0);
 513
 514         RETURN(sai);
 515 }
 516
 517 static inline struct ll_statahead_info *
 518 ll_sai_get(struct ll_statahead_info *sai)
 519 {
 520         cfs_atomic_inc(&sai->sai_refcount);
 521         return sai;
 522 }
 523
 524 static void ll_sai_put(struct ll_statahead_info *sai)
 525 {
 526         struct inode         *inode = sai->sai_inode;
 527         struct ll_inode_info *lli   = ll_i2info(inode);
 528         ENTRY;
 529
 530         if (cfs_atomic_dec_and_lock(&sai->sai_refcount, &lli->lli_sa_lock)) {
 531                 struct ll_sa_entry *entry, *next;
 532
 533                 if (unlikely(cfs_atomic_read(&sai->sai_refcount) > 0)) {
 534                         /* It is race case, the interpret callback just hold
 535                          * a reference count */
 536                         cfs_spin_unlock(&lli->lli_sa_lock);
 537                         RETURN_EXIT;
 538                 }
 539
 540                 LASSERT(lli->lli_opendir_key == NULL);
 541                 LASSERT(thread_is_stopped(&sai->sai_thread));
 542                 LASSERT(thread_is_stopped(&sai->sai_agl_thread));
 543
 544                 lli->lli_sai = NULL;
 545                 lli->lli_opendir_pid = 0;
 546                 cfs_spin_unlock(&lli->lli_sa_lock);
 547
 548                 if (sai->sai_sent > sai->sai_replied)
 549                         CDEBUG(D_READA,"statahead for dir "DFID" does not "
 550                               "finish: [sent:"LPU64"] [replied:"LPU64"]\n",
 551                               PFID(&lli->lli_fid),
 552                               sai->sai_sent, sai->sai_replied);
 553
 554                 cfs_list_for_each_entry_safe(entry, next,
 555                                              &sai->sai_entries_sent, se_list)
 556                         do_sai_entry_fini(sai, entry);
 557
 558                 LASSERT(sa_received_empty(sai));
 559
 560                 cfs_list_for_each_entry_safe(entry, next,
 561                                              &sai->sai_entries_stated, se_list)
 562                         do_sai_entry_fini(sai, entry);
 563
 564                 LASSERT(cfs_atomic_read(&sai->sai_cache_count) == 0);
 565                 LASSERT(agl_list_empty(sai));
 566
 567                 iput(inode);
 568                 OBD_FREE_PTR(sai);
 569         }
 570
 571         EXIT;
 572 }
 573
 574 #ifndef HAVE_AGL_SUPPORT
 575 # define cl_agl(inode)  do {} while (0)
 576 #endif
 577
 578 /* Do NOT forget to drop inode refcount when into sai_entries_agl. */
 579 static void ll_agl_trigger(struct inode *inode, struct ll_statahead_info *sai)
 580 {
 581         struct ll_inode_info *lli   = ll_i2info(inode);
 582         __u64                 index = lli->lli_agl_index;
 583         int                   rc;
 584         ENTRY;
 585
 586         LASSERT(cfs_list_empty(&lli->lli_agl_list));
 587
 588         /* AGL maybe fall behind statahead with one entry */
 589         if (is_omitted_entry(sai, index + 1)) {
 590                 lli->lli_agl_index = 0;
 591                 iput(inode);
 592                 RETURN_EXIT;
 593         }
 594
 595         /* Someone is in glimpse (sync or async), do nothing. */
 596         rc = cfs_down_write_trylock(&lli->lli_glimpse_sem);
 597         if (rc == 0) {
 598                 lli->lli_agl_index = 0;
 599                 iput(inode);
 600                 RETURN_EXIT;
 601         }
 602
 603         /*
 604          * Someone triggered glimpse within 1 sec before.
 605          * 1) The former glimpse succeeded with glimpse lock granted by OST, and
 606          *    if the lock is still cached on client, AGL needs to do nothing. If
 607          *    it is cancelled by other client, AGL maybe cannot obtaion new lock
 608          *    for no glimpse callback triggered by AGL.
 609          * 2) The former glimpse succeeded, but OST did not grant glimpse lock.
 610          *    Under such case, it is quite possible that the OST will not grant
 611          *    glimpse lock for AGL also.
 612          * 3) The former glimpse failed, compared with other two cases, it is
 613          *    relative rare. AGL can ignore such case, and it will not muchly
 614          *    affect the performance.
 615          */
 616         if (lli->lli_glimpse_time != 0 &&
 617             cfs_time_before(cfs_time_shift(-1), lli->lli_glimpse_time)) {
 618                 cfs_up_write(&lli->lli_glimpse_sem);
 619                 lli->lli_agl_index = 0;
 620                 iput(inode);
 621                 RETURN_EXIT;
 622         }
 623
 624         CDEBUG(D_READA, "Handling (init) async glimpse: inode = "
 625                DFID", idx = "LPU64"\n", PFID(&lli->lli_fid), index);
 626
 627         cl_agl(inode);
 628         lli->lli_agl_index = 0;
 629         lli->lli_glimpse_time = cfs_time_current();
 630         cfs_up_write(&lli->lli_glimpse_sem);
 631
 632         CDEBUG(D_READA, "Handled (init) async glimpse: inode= "
 633                DFID", idx = "LPU64", rc = %d\n",
 634                PFID(&lli->lli_fid), index, rc);
 635
 636         iput(inode);
 637
 638         EXIT;
 639 }
 640
 641 static void do_statahead_interpret(struct ll_statahead_info *sai,
 642                                    struct ll_sa_entry *target)
 643 {
 644         struct inode           *dir   = sai->sai_inode;
 645         struct inode           *child;
 646         struct ll_inode_info   *lli   = ll_i2info(dir);
 647         struct ll_sa_entry     *entry;
 648         struct md_enqueue_info *minfo;
 649         struct lookup_intent   *it;
 650         struct ptlrpc_request  *req;
 651         struct mdt_body        *body;
 652         int                     rc    = 0;
 653         ENTRY;
 654
 655         cfs_spin_lock(&lli->lli_sa_lock);
 656         if (target != NULL && target->se_req != NULL &&
 657             !cfs_list_empty(&target->se_list)) {
 658                 entry = target;
 659         } else if (unlikely(sa_received_empty(sai))) {
 660                 cfs_spin_unlock(&lli->lli_sa_lock);
 661                 RETURN_EXIT;
 662         } else {
 663                 entry = sa_first_received_entry(sai);
 664         }
 665
 666         cfs_atomic_inc(&entry->se_refcount);
 667         cfs_list_del_init(&entry->se_list);
 668         cfs_spin_unlock(&lli->lli_sa_lock);
 669
 670         LASSERT(entry->se_handle != 0);
 671
 672         minfo = entry->se_minfo;
 673         it = &minfo->mi_it;
 674         req = entry->se_req;
 675         body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
 676         if (body == NULL)
 677                 GOTO(out, rc = -EFAULT);
 678
 679         child = entry->se_inode;
 680         if (child == NULL) {
 681                 /*
 682                  * lookup.
 683                  */
 684                 LASSERT(fid_is_zero(&minfo->mi_data.op_fid2));
 685
 686                 /* XXX: No fid in reply, this is probaly cross-ref case.
 687                  * SA can't handle it yet. */
 688                 if (body->valid & OBD_MD_MDS)
 689                         GOTO(out, rc = -EAGAIN);
 690         } else {
 691                 /*
 692                  * revalidate.
 693                  */
 694                 /* unlinked and re-created with the same name */
 695                 if (unlikely(!lu_fid_eq(&minfo->mi_data.op_fid2, &body->fid1))){
 696                         entry->se_inode = NULL;
 697                         iput(child);
 698                         child = NULL;
 699                 }
 700         }
 701
 702         it->d.lustre.it_lock_handle = entry->se_handle;
 703         rc = md_revalidate_lock(ll_i2mdexp(dir), it, NULL, NULL);
 704         if (rc != 1)
 705                 GOTO(out, rc = -EAGAIN);
 706
 707         rc = ll_prep_inode(&child, req, dir->i_sb);
 708         if (rc)
 709                 GOTO(out, rc);
 710
 711         CDEBUG(D_DLMTRACE, "setting l_data to inode %p (%lu/%u)\n",
 712                child, child->i_ino, child->i_generation);
 713         ll_set_lock_data(ll_i2sbi(dir)->ll_md_exp, child, it, NULL);
 714
 715         entry->se_inode = child;
 716
 717         if (agl_should_run(sai, child))
 718                 ll_agl_add(sai, child, entry->se_index);
 719
 720         EXIT;
 721
 722 out:
 723         /* The "ll_sa_entry_to_stated()" will drop related ldlm ibits lock
 724          * reference count by calling "ll_intent_drop_lock()" in spite of the
 725          * above operations failed or not. Do not worry about calling
 726          * "ll_intent_drop_lock()" more than once. */
 727         rc = ll_sa_entry_to_stated(sai, entry, rc < 0 ? rc : SA_ENTRY_SUCC);
 728         if (rc == 0 && entry->se_index == sai->sai_index_wait && target == NULL)
 729                 cfs_waitq_signal(&sai->sai_waitq);
 730         ll_sa_entry_put(sai, entry);
 731 }
 732
 733 static int ll_statahead_interpret(struct ptlrpc_request *req,
 734                                   struct md_enqueue_info *minfo, int rc)
 735 {
 736         struct lookup_intent     *it  = &minfo->mi_it;
 737         struct inode             *dir = minfo->mi_dir;
 738         struct ll_inode_info     *lli = ll_i2info(dir);
 739         struct ll_statahead_info *sai = NULL;
 740         struct ll_sa_entry       *entry;
 741         int                       wakeup;
 742         ENTRY;
 743
 744         if (it_disposition(it, DISP_LOOKUP_NEG))
 745                 rc = -ENOENT;
 746
 747         cfs_spin_lock(&lli->lli_sa_lock);
 748         /* stale entry */
 749         if (unlikely(lli->lli_sai == NULL ||
 750                      lli->lli_sai->sai_generation != minfo->mi_generation)) {
 751                 cfs_spin_unlock(&lli->lli_sa_lock);
 752                 GOTO(out, rc = -ESTALE);
 753         } else {
 754                 sai = ll_sai_get(lli->lli_sai);
 755                 if (unlikely(!thread_is_running(&sai->sai_thread))) {
 756                         sai->sai_replied++;
 757                         cfs_spin_unlock(&lli->lli_sa_lock);
 758                         GOTO(out, rc = -EBADFD);
 759                 }
 760
 761                 entry = ll_sa_entry_get_byindex(sai, minfo->mi_cbdata);
 762                 if (entry == NULL) {
 763                         sai->sai_replied++;
 764                         cfs_spin_unlock(&lli->lli_sa_lock);
 765                         GOTO(out, rc = -EIDRM);
 766                 }
 767
 768                 cfs_list_del_init(&entry->se_list);
 769                 if (rc != 0) {
 770                         sai->sai_replied++;
 771                         do_sai_entry_to_stated(sai, entry, rc);
 772                         cfs_spin_unlock(&lli->lli_sa_lock);
 773                         if (entry->se_index == sai->sai_index_wait)
 774                                 cfs_waitq_signal(&sai->sai_waitq);
 775                 } else {
 776                         entry->se_minfo = minfo;
 777                         entry->se_req = ptlrpc_request_addref(req);
 778                         /* Release the async ibits lock ASAP to avoid deadlock
 779                          * when statahead thread tries to enqueue lock on parent
 780                          * for readpage and other tries to enqueue lock on child
 781                          * with parent's lock held, for example: unlink. */
 782                         entry->se_handle = it->d.lustre.it_lock_handle;
 783                         ll_intent_drop_lock(it);
 784                         wakeup = sa_received_empty(sai);
 785                         cfs_list_add_tail(&entry->se_list,
 786                                           &sai->sai_entries_received);
 787                         sai->sai_replied++;
 788                         cfs_spin_unlock(&lli->lli_sa_lock);
 789                         if (wakeup)
 790                                 cfs_waitq_signal(&sai->sai_thread.t_ctl_waitq);
 791                 }
 792                 ll_sa_entry_put(sai, entry);
 793         }
 794
 795         EXIT;
 796
 797 out:
 798         if (rc != 0) {
 799                 ll_intent_release(it);
 800                 iput(dir);
 801                 OBD_FREE_PTR(minfo);
 802         }
 803         if (sai != NULL)
 804                 ll_sai_put(sai);
 805         return rc;
 806 }
 807
 808 static void sa_args_fini(struct md_enqueue_info *minfo,
 809                          struct ldlm_enqueue_info *einfo)
 810 {
 811         LASSERT(minfo && einfo);
 812         iput(minfo->mi_dir);
 813         capa_put(minfo->mi_data.op_capa1);
 814         capa_put(minfo->mi_data.op_capa2);
 815         OBD_FREE_PTR(minfo);
 816         OBD_FREE_PTR(einfo);
 817 }
 818
 819 /**
 820  * There is race condition between "capa_put" and "ll_statahead_interpret" for
 821  * accessing "op_data.op_capa[1,2]" as following:
 822  * "capa_put" releases "op_data.op_capa[1,2]"'s reference count after calling
 823  * "md_intent_getattr_async". But "ll_statahead_interpret" maybe run first, and
 824  * fill "op_data.op_capa[1,2]" as POISON, then cause "capa_put" access invalid
 825  * "ocapa". So here reserve "op_data.op_capa[1,2]" in "pcapa" before calling
 826  * "md_intent_getattr_async".
 827  */
 828 static int sa_args_init(struct inode *dir, struct inode *child,
 829                         struct ll_sa_entry *entry, struct md_enqueue_info **pmi,
 830                         struct ldlm_enqueue_info **pei,
 831                         struct obd_capa **pcapa)
 832 {
 833         struct qstr              *qstr = &entry->se_qstr;
 834         struct ll_inode_info     *lli  = ll_i2info(dir);
 835         struct md_enqueue_info   *minfo;
 836         struct ldlm_enqueue_info *einfo;
 837         struct md_op_data        *op_data;
 838
 839         OBD_ALLOC_PTR(einfo);
 840         if (einfo == NULL)
 841                 return -ENOMEM;
 842
 843         OBD_ALLOC_PTR(minfo);
 844         if (minfo == NULL) {
 845                 OBD_FREE_PTR(einfo);
 846                 return -ENOMEM;
 847         }
 848
 849         op_data = ll_prep_md_op_data(&minfo->mi_data, dir, child, qstr->name,
 850                                      qstr->len, 0, LUSTRE_OPC_ANY, NULL);
 851         if (IS_ERR(op_data)) {
 852                 OBD_FREE_PTR(einfo);
 853                 OBD_FREE_PTR(minfo);
 854                 return PTR_ERR(op_data);
 855         }
 856
 857         minfo->mi_it.it_op = IT_GETATTR;
 858         minfo->mi_dir = igrab(dir);
 859         minfo->mi_cb = ll_statahead_interpret;
 860         minfo->mi_generation = lli->lli_sai->sai_generation;
 861         minfo->mi_cbdata = entry->se_index;
 862
 863         einfo->ei_type   = LDLM_IBITS;
 864         einfo->ei_mode   = it_to_lock_mode(&minfo->mi_it);
 865         einfo->ei_cb_bl  = ll_md_blocking_ast;
 866         einfo->ei_cb_cp  = ldlm_completion_ast;
 867         einfo->ei_cb_gl  = NULL;
 868         einfo->ei_cbdata = NULL;
 869
 870         *pmi = minfo;
 871         *pei = einfo;
 872         pcapa[0] = op_data->op_capa1;
 873         pcapa[1] = op_data->op_capa2;
 874
 875         return 0;
 876 }
 877
 878 static int do_sa_lookup(struct inode *dir, struct ll_sa_entry *entry)
 879 {
 880         struct md_enqueue_info   *minfo;
 881         struct ldlm_enqueue_info *einfo;
 882         struct obd_capa          *capas[2];
 883         int                       rc;
 884         ENTRY;
 885
 886         rc = sa_args_init(dir, NULL, entry, &minfo, &einfo, capas);
 887         if (rc)
 888                 RETURN(rc);
 889
 890         rc = md_intent_getattr_async(ll_i2mdexp(dir), minfo, einfo);
 891         if (!rc) {
 892                 capa_put(capas[0]);
 893                 capa_put(capas[1]);
 894         } else {
 895                 sa_args_fini(minfo, einfo);
 896         }
 897
 898         RETURN(rc);
 899 }
 900
 901 /**
 902  * similar to ll_revalidate_it().
 903  * \retval      1 -- dentry valid
 904  * \retval      0 -- will send stat-ahead request
 905  * \retval others -- prepare stat-ahead request failed
 906  */
 907 static int do_sa_revalidate(struct inode *dir, struct ll_sa_entry *entry,
 908                             struct dentry *dentry)
 909 {
 910         struct inode             *inode = dentry->d_inode;
 911         struct lookup_intent      it = { .it_op = IT_GETATTR,
 912                                          .d.lustre.it_lock_handle = 0 };
 913         struct md_enqueue_info   *minfo;
 914         struct ldlm_enqueue_info *einfo;
 915         struct obd_capa          *capas[2];
 916         int rc;
 917         ENTRY;
 918
 919         if (unlikely(inode == NULL))
 920                 RETURN(1);
 921
 922         if (d_mountpoint(dentry))
 923                 RETURN(1);
 924
 925         if (unlikely(dentry == dentry->d_sb->s_root))
 926                 RETURN(1);
 927
 928         entry->se_inode = igrab(inode);
 929         rc = md_revalidate_lock(ll_i2mdexp(dir), &it, ll_inode2fid(inode),NULL);
 930         if (rc == 1) {
 931                 entry->se_handle = it.d.lustre.it_lock_handle;
 932                 ll_intent_release(&it);
 933                 RETURN(1);
 934         }
 935
 936         rc = sa_args_init(dir, inode, entry, &minfo, &einfo, capas);
 937         if (rc) {
 938                 entry->se_inode = NULL;
 939                 iput(inode);
 940                 RETURN(rc);
 941         }
 942
 943         rc = md_intent_getattr_async(ll_i2mdexp(dir), minfo, einfo);
 944         if (!rc) {
 945                 capa_put(capas[0]);
 946                 capa_put(capas[1]);
 947         } else {
 948                 entry->se_inode = NULL;
 949                 iput(inode);
 950                 sa_args_fini(minfo, einfo);
 951         }
 952
 953         RETURN(rc);
 954 }
 955
 956 static void ll_statahead_one(struct dentry *parent, const char* entry_name,
 957                              int entry_name_len)
 958 {
 959         struct inode             *dir    = parent->d_inode;
 960         struct ll_inode_info     *lli    = ll_i2info(dir);
 961         struct ll_statahead_info *sai    = lli->lli_sai;
 962         struct dentry            *dentry = NULL;
 963         struct ll_sa_entry       *entry;
 964         int                       rc;
 965         int                       rc1;
 966         ENTRY;
 967
 968         entry = ll_sa_entry_alloc(sai, sai->sai_index, entry_name,
 969                                   entry_name_len);
 970         if (IS_ERR(entry))
 971                 RETURN_EXIT;
 972
 973         dentry = d_lookup(parent, &entry->se_qstr);
 974         if (!dentry) {
 975                 rc = do_sa_lookup(dir, entry);
 976         } else {
 977                 rc = do_sa_revalidate(dir, entry, dentry);
 978                 if (rc == 1 && agl_should_run(sai, dentry->d_inode))
 979                         ll_agl_add(sai, dentry->d_inode, entry->se_index);
 980         }
 981
 982         if (dentry != NULL)
 983                 dput(dentry);
 984
 985         if (rc) {
 986                 rc1 = ll_sa_entry_to_stated(sai, entry,
 987                                         rc < 0 ? SA_ENTRY_INVA : SA_ENTRY_SUCC);
 988                 if (rc1 == 0 && entry->se_index == sai->sai_index_wait)
 989                         cfs_waitq_signal(&sai->sai_waitq);
 990         } else {
 991                 sai->sai_sent++;
 992         }
 993
 994         sai->sai_index++;
 995         /* drop one refcount on entry by ll_sa_entry_alloc */
 996         ll_sa_entry_put(sai, entry);
 997
 998         EXIT;
 999 }
1000
1001 static int ll_agl_thread(void *arg)
1002 {
1003         struct dentry            *parent = (struct dentry *)arg;
1004         struct inode             *dir    = parent->d_inode;
1005         struct ll_inode_info     *plli   = ll_i2info(dir);
1006         struct ll_inode_info     *clli;
1007         struct ll_sb_info        *sbi    = ll_i2sbi(dir);
1008         struct ll_statahead_info *sai    = ll_sai_get(plli->lli_sai);
1009         struct ptlrpc_thread     *thread = &sai->sai_agl_thread;
1010         struct l_wait_info        lwi    = { 0 };
1011         ENTRY;
1012
1013         {
1014                 char pname[16];
1015                 snprintf(pname, 15, "ll_agl_%u", plli->lli_opendir_pid);
1016                 cfs_daemonize(pname);
1017         }
1018
1019         CDEBUG(D_READA, "agl thread started: [pid %d] [parent %.*s]\n",
1020                cfs_curproc_pid(), parent->d_name.len, parent->d_name.name);
1021
1022         atomic_inc(&sbi->ll_agl_total);
1023         cfs_spin_lock(&plli->lli_agl_lock);
1024         sai->sai_agl_valid = 1;
1025         thread_set_flags(thread, SVC_RUNNING);
1026         cfs_spin_unlock(&plli->lli_agl_lock);
1027         cfs_waitq_signal(&thread->t_ctl_waitq);
1028
1029         while (1) {
1030                 l_wait_event(thread->t_ctl_waitq,
1031                              !agl_list_empty(sai) ||
1032                              !thread_is_running(thread),
1033                              &lwi);
1034
1035                 if (!thread_is_running(thread))
1036                         break;
1037
1038                 cfs_spin_lock(&plli->lli_agl_lock);
1039                 /* The statahead thread maybe help to process AGL entries,
1040                  * so check whether list empty again. */
1041                 if (!agl_list_empty(sai)) {
1042                         clli = agl_first_entry(sai);
1043                         cfs_list_del_init(&clli->lli_agl_list);
1044                         cfs_spin_unlock(&plli->lli_agl_lock);
1045                         ll_agl_trigger(&clli->lli_vfs_inode, sai);
1046                 } else {
1047                         cfs_spin_unlock(&plli->lli_agl_lock);
1048                 }
1049         }
1050
1051         cfs_spin_lock(&plli->lli_agl_lock);
1052         sai->sai_agl_valid = 0;
1053         while (!agl_list_empty(sai)) {
1054                 clli = agl_first_entry(sai);
1055                 cfs_list_del_init(&clli->lli_agl_list);
1056                 cfs_spin_unlock(&plli->lli_agl_lock);
1057                 clli->lli_agl_index = 0;
1058                 iput(&clli->lli_vfs_inode);
1059                 cfs_spin_lock(&plli->lli_agl_lock);
1060         }
1061         thread_set_flags(thread, SVC_STOPPED);
1062         cfs_spin_unlock(&plli->lli_agl_lock);
1063         cfs_waitq_signal(&thread->t_ctl_waitq);
1064         ll_sai_put(sai);
1065         CDEBUG(D_READA, "agl thread stopped: [pid %d] [parent %.*s]\n",
1066                cfs_curproc_pid(), parent->d_name.len, parent->d_name.name);
1067         RETURN(0);
1068 }
1069
1070 static void ll_start_agl(struct dentry *parent, struct ll_statahead_info *sai)
1071 {
1072         struct ptlrpc_thread *thread = &sai->sai_agl_thread;
1073         struct l_wait_info    lwi    = { 0 };
1074         int                   rc;
1075         ENTRY;
1076
1077         CDEBUG(D_READA, "start agl thread: [pid %d] [parent %.*s]\n",
1078                cfs_curproc_pid(), parent->d_name.len, parent->d_name.name);
1079
1080         rc = cfs_create_thread(ll_agl_thread, parent, 0);
1081         if (rc < 0) {
1082                 CERROR("can't start ll_agl thread, rc: %d\n", rc);
1083                 RETURN_EXIT;
1084         }
1085
1086         l_wait_event(thread->t_ctl_waitq,
1087                      thread_is_running(thread) || thread_is_stopped(thread),
1088                      &lwi);
1089         EXIT;
1090 }
1091
1092 static int ll_statahead_thread(void *arg)
1093 {
1094         struct dentry            *parent = (struct dentry *)arg;
1095         struct inode             *dir    = parent->d_inode;
1096         struct ll_inode_info     *plli   = ll_i2info(dir);
1097         struct ll_inode_info     *clli;
1098         struct ll_sb_info        *sbi    = ll_i2sbi(dir);
1099         struct ll_statahead_info *sai    = ll_sai_get(plli->lli_sai);
1100         struct ptlrpc_thread     *thread = &sai->sai_thread;
1101         struct page              *page;
1102         __u64                     pos    = 0;
1103         int                       first  = 0;
1104         int                       rc     = 0;
1105         struct ll_dir_chain       chain;
1106         struct l_wait_info        lwi    = { 0 };
1107         ENTRY;
1108
1109         {
1110                 char pname[16];
1111                 snprintf(pname, 15, "ll_sa_%u", plli->lli_opendir_pid);
1112                 cfs_daemonize(pname);
1113         }
1114
1115         CDEBUG(D_READA, "statahead thread started: [pid %d] [parent %.*s]\n",
1116                cfs_curproc_pid(), parent->d_name.len, parent->d_name.name);
1117
1118         if (sbi->ll_flags & LL_SBI_AGL_ENABLED)
1119                 ll_start_agl(parent, sai);
1120
1121         atomic_inc(&sbi->ll_sa_total);
1122         cfs_spin_lock(&plli->lli_sa_lock);
1123         thread_set_flags(thread, SVC_RUNNING);
1124         cfs_spin_unlock(&plli->lli_sa_lock);
1125         cfs_waitq_signal(&thread->t_ctl_waitq);
1126
1127         plli->lli_sa_pos = 0;
1128         ll_dir_chain_init(&chain);
1129         page = ll_get_dir_page(NULL, dir, pos, &chain);
1130
1131         while (1) {
1132                 struct lu_dirpage *dp;
1133                 struct lu_dirent  *ent;
1134
1135                 if (IS_ERR(page)) {
1136                         rc = PTR_ERR(page);
1137                         CDEBUG(D_READA, "error reading dir "DFID" at "LPU64
1138                                "/"LPU64": [rc %d] [parent %u]\n",
1139                                PFID(ll_inode2fid(dir)), pos, sai->sai_index,
1140                                rc, plli->lli_opendir_pid);
1141                         GOTO(out, rc);
1142                 }
1143
1144                 dp = page_address(page);
1145                 for (ent = lu_dirent_start(dp); ent != NULL;
1146                      ent = lu_dirent_next(ent)) {
1147                         __u64 hash;
1148                         int namelen;
1149                         char *name;
1150
1151                         hash = le64_to_cpu(ent->lde_hash);
1152                         if (unlikely(hash < pos))
1153                                 /*
1154                                  * Skip until we find target hash value.
1155                                  */
1156                                 continue;
1157
1158                         namelen = le16_to_cpu(ent->lde_namelen);
1159                         if (unlikely(namelen == 0))
1160                                 /*
1161                                  * Skip dummy record.
1162                                  */
1163                                 continue;
1164
1165                         name = ent->lde_name;
1166                         if (name[0] == '.') {
1167                                 if (namelen == 1) {
1168                                         /*
1169                                          * skip "."
1170                                          */
1171                                         continue;
1172                                 } else if (name[1] == '.' && namelen == 2) {
1173                                         /*
1174                                          * skip ".."
1175                                          */
1176                                         continue;
1177                                 } else if (!sai->sai_ls_all) {
1178                                         /*
1179                                          * skip hidden files.
1180                                          */
1181                                         sai->sai_skip_hidden++;
1182                                         continue;
1183                                 }
1184                         }
1185
1186                         /*
1187                          * don't stat-ahead first entry.
1188                          */
1189                         if (unlikely(++first == 1))
1190                                 continue;
1191
1192 keep_it:
1193                         l_wait_event(thread->t_ctl_waitq,
1194                                      !sa_sent_full(sai) ||
1195                                      !sa_received_empty(sai) ||
1196                                      !agl_list_empty(sai) ||
1197                                      !thread_is_running(thread),
1198                                      &lwi);
1199
1200 interpret_it:
1201                         while (!sa_received_empty(sai))
1202                                 do_statahead_interpret(sai, NULL);
1203
1204                         if (unlikely(!thread_is_running(thread))) {
1205                                 ll_release_page(page, 0);
1206                                 GOTO(out, rc = 0);
1207                         }
1208
1209                         /* If no window for metadata statahead, but there are
1210                          * some AGL entries to be triggered, then try to help
1211                          * to process the AGL entries. */
1212                         if (sa_sent_full(sai)) {
1213                                 cfs_spin_lock(&plli->lli_agl_lock);
1214                                 while (!agl_list_empty(sai)) {
1215                                         clli = agl_first_entry(sai);
1216                                         cfs_list_del_init(&clli->lli_agl_list);
1217                                         cfs_spin_unlock(&plli->lli_agl_lock);
1218                                         ll_agl_trigger(&clli->lli_vfs_inode,
1219                                                        sai);
1220
1221                                         if (!sa_received_empty(sai))
1222                                                 goto interpret_it;
1223
1224                                         if (unlikely(
1225                                                 !thread_is_running(thread))) {
1226                                                 ll_release_page(page, 0);
1227                                                 GOTO(out, rc = 0);
1228                                         }
1229
1230                                         if (!sa_sent_full(sai))
1231                                                 goto do_it;
1232
1233                                         cfs_spin_lock(&plli->lli_agl_lock);
1234                                 }
1235                                 cfs_spin_unlock(&plli->lli_agl_lock);
1236
1237                                 goto keep_it;
1238                         }
1239
1240 do_it:
1241                         ll_statahead_one(parent, name, namelen);
1242                 }
1243                 pos = le64_to_cpu(dp->ldp_hash_end);
1244                 if (pos == MDS_DIR_END_OFF) {
1245                         /*
1246                          * End of directory reached.
1247                          */
1248                         ll_release_page(page, 0);
1249                         while (1) {
1250                                 l_wait_event(thread->t_ctl_waitq,
1251                                              !sa_received_empty(sai) ||
1252                                              sai->sai_sent == sai->sai_replied||
1253                                              !thread_is_running(thread),
1254                                              &lwi);
1255
1256                                 while (!sa_received_empty(sai))
1257                                         do_statahead_interpret(sai, NULL);
1258
1259                                 if (unlikely(!thread_is_running(thread)))
1260                                         GOTO(out, rc = 0);
1261
1262                                 if (sai->sai_sent == sai->sai_replied &&
1263                                     sa_received_empty(sai))
1264                                         break;
1265                         }
1266
1267                         cfs_spin_lock(&plli->lli_agl_lock);
1268                         while (!agl_list_empty(sai) &&
1269                                thread_is_running(thread)) {
1270                                 clli = agl_first_entry(sai);
1271                                 cfs_list_del_init(&clli->lli_agl_list);
1272                                 cfs_spin_unlock(&plli->lli_agl_lock);
1273                                 ll_agl_trigger(&clli->lli_vfs_inode, sai);
1274                                 cfs_spin_lock(&plli->lli_agl_lock);
1275                         }
1276                         cfs_spin_unlock(&plli->lli_agl_lock);
1277
1278                         GOTO(out, rc = 0);
1279                 } else if (1) {
1280                         /*
1281                          * chain is exhausted.
1282                          * Normal case: continue to the next page.
1283                          */
1284                         ll_release_page(page, le32_to_cpu(dp->ldp_flags) &
1285                                               LDF_COLLIDE);
1286                         plli->lli_sa_pos = pos;
1287                         sai->sai_in_readpage = 1;
1288                         page = ll_get_dir_page(NULL, dir, pos, &chain);
1289                         sai->sai_in_readpage = 0;
1290                 } else {
1291                         LASSERT(le32_to_cpu(dp->ldp_flags) & LDF_COLLIDE);
1292                         ll_release_page(page, 1);
1293                         /*
1294                          * go into overflow page.
1295                          */
1296                 }
1297         }
1298         EXIT;
1299
1300 out:
1301         if (sai->sai_agl_valid) {
1302                 struct ptlrpc_thread *agl_thread = &sai->sai_agl_thread;
1303
1304                 cfs_spin_lock(&plli->lli_agl_lock);
1305                 thread_set_flags(agl_thread, SVC_STOPPING);
1306                 cfs_spin_unlock(&plli->lli_agl_lock);
1307                 cfs_waitq_signal(&agl_thread->t_ctl_waitq);
1308
1309                 CDEBUG(D_READA, "stop agl thread: [pid %d]\n",
1310                        cfs_curproc_pid());
1311                 l_wait_event(agl_thread->t_ctl_waitq,
1312                              thread_is_stopped(agl_thread),
1313                              &lwi);
1314         }
1315
1316         ll_dir_chain_fini(&chain);
1317         cfs_spin_lock(&plli->lli_sa_lock);
1318         if (!sa_received_empty(sai)) {
1319                 thread_set_flags(thread, SVC_STOPPING);
1320                 cfs_spin_unlock(&plli->lli_sa_lock);
1321
1322                 /* To release the resources held by received entries. */
1323                 while (!sa_received_empty(sai))
1324                         do_statahead_interpret(sai, NULL);
1325
1326                 cfs_spin_lock(&plli->lli_sa_lock);
1327         }
1328         thread_set_flags(thread, SVC_STOPPED);
1329         cfs_spin_unlock(&plli->lli_sa_lock);
1330         cfs_waitq_signal(&sai->sai_waitq);
1331         cfs_waitq_signal(&thread->t_ctl_waitq);
1332         ll_sai_put(sai);
1333         dput(parent);
1334         CDEBUG(D_READA, "statahead thread stopped: [pid %d] [parent %.*s]\n",
1335                cfs_curproc_pid(), parent->d_name.len, parent->d_name.name);
1336         return rc;
1337 }
1338
1339 /**
1340  * called in ll_file_release().
1341  */
1342 void ll_stop_statahead(struct inode *dir, void *key)
1343 {
1344         struct ll_inode_info *lli = ll_i2info(dir);
1345
1346         if (unlikely(key == NULL))
1347                 return;
1348
1349         cfs_spin_lock(&lli->lli_sa_lock);
1350         if (lli->lli_opendir_key != key || lli->lli_opendir_pid == 0) {
1351                 cfs_spin_unlock(&lli->lli_sa_lock);
1352                 return;
1353         }
1354
1355         lli->lli_opendir_key = NULL;
1356
1357         if (lli->lli_sai) {
1358                 struct l_wait_info lwi = { 0 };
1359                 struct ptlrpc_thread *thread = &lli->lli_sai->sai_thread;
1360
1361                 if (!thread_is_stopped(thread)) {
1362                         thread_set_flags(thread, SVC_STOPPING);
1363                         cfs_spin_unlock(&lli->lli_sa_lock);
1364                         cfs_waitq_signal(&thread->t_ctl_waitq);
1365
1366                         CDEBUG(D_READA, "stop statahead thread: [pid %d]\n",
1367                                cfs_curproc_pid());
1368                         l_wait_event(thread->t_ctl_waitq,
1369                                      thread_is_stopped(thread),
1370                                      &lwi);
1371                 } else {
1372                         cfs_spin_unlock(&lli->lli_sa_lock);
1373                 }
1374
1375                 /*
1376                  * Put the ref which was held when first statahead_enter.
1377                  * It maybe not the last ref for some statahead requests
1378                  * maybe inflight.
1379                  */
1380                 ll_sai_put(lli->lli_sai);
1381         } else {
1382                 lli->lli_opendir_pid = 0;
1383                 cfs_spin_unlock(&lli->lli_sa_lock);
1384         }
1385 }
1386
1387 enum {
1388         /**
1389          * not first dirent, or is "."
1390          */
1391         LS_NONE_FIRST_DE = 0,
1392         /**
1393          * the first non-hidden dirent
1394          */
1395         LS_FIRST_DE,
1396         /**
1397          * the first hidden dirent, that is "."
1398          */
1399         LS_FIRST_DOT_DE
1400 };
1401
1402 static int is_first_dirent(struct inode *dir, struct dentry *dentry)
1403 {
1404         struct ll_inode_info *lli    = ll_i2info(dir);
1405         struct ll_dir_chain   chain;
1406         struct qstr          *target = &dentry->d_name;
1407         struct page          *page;
1408         __u64                 pos    = 0;
1409         int                   dot_de;
1410         int                   rc     = LS_NONE_FIRST_DE;
1411         ENTRY;
1412
1413         lli->lli_sa_pos = 0;
1414         ll_dir_chain_init(&chain);
1415         page = ll_get_dir_page(NULL, dir, pos, &chain);
1416
1417         while (1) {
1418                 struct lu_dirpage *dp;
1419                 struct lu_dirent  *ent;
1420
1421                 if (IS_ERR(page)) {
1422                         struct ll_inode_info *lli = ll_i2info(dir);
1423
1424                         rc = PTR_ERR(page);
1425                         CERROR("error reading dir "DFID" at "LPU64": "
1426                                "[rc %d] [parent %u]\n",
1427                                PFID(ll_inode2fid(dir)), pos,
1428                                rc, lli->lli_opendir_pid);
1429                         break;
1430                 }
1431
1432                 dp = page_address(page);
1433                 for (ent = lu_dirent_start(dp); ent != NULL;
1434                      ent = lu_dirent_next(ent)) {
1435                         __u64 hash;
1436                         int namelen;
1437                         char *name;
1438
1439                         hash = le64_to_cpu(ent->lde_hash);
1440                         /* The ll_get_dir_page() can return any page containing
1441                          * the given hash which may be not the start hash. */
1442                         if (unlikely(hash < pos))
1443                                 continue;
1444
1445                         namelen = le16_to_cpu(ent->lde_namelen);
1446                         if (unlikely(namelen == 0))
1447                                 /*
1448                                  * skip dummy record.
1449                                  */
1450                                 continue;
1451
1452                         name = ent->lde_name;
1453                         if (name[0] == '.') {
1454                                 if (namelen == 1)
1455                                         /*
1456                                          * skip "."
1457                                          */
1458                                         continue;
1459                                 else if (name[1] == '.' && namelen == 2)
1460                                         /*
1461                                          * skip ".."
1462                                          */
1463                                         continue;
1464                                 else
1465                                         dot_de = 1;
1466                         } else {
1467                                 dot_de = 0;
1468                         }
1469
1470                         if (dot_de && target->name[0] != '.') {
1471                                 CDEBUG(D_READA, "%.*s skip hidden file %.*s\n",
1472                                        target->len, target->name,
1473                                        namelen, name);
1474                                 continue;
1475                         }
1476
1477                         if (target->len != namelen ||
1478                             memcmp(target->name, name, namelen) != 0)
1479                                 rc = LS_NONE_FIRST_DE;
1480                         else if (!dot_de)
1481                                 rc = LS_FIRST_DE;
1482                         else
1483                                 rc = LS_FIRST_DOT_DE;
1484
1485                         ll_release_page(page, 0);
1486                         GOTO(out, rc);
1487                 }
1488                 pos = le64_to_cpu(dp->ldp_hash_end);
1489                 if (pos == MDS_DIR_END_OFF) {
1490                         /*
1491                          * End of directory reached.
1492                          */
1493                         ll_release_page(page, 0);
1494                         break;
1495                 } else if (1) {
1496                         /*
1497                          * chain is exhausted
1498                          * Normal case: continue to the next page.
1499                          */
1500                         ll_release_page(page, le32_to_cpu(dp->ldp_flags) &
1501                                               LDF_COLLIDE);
1502                         lli->lli_sa_pos = pos;
1503                         page = ll_get_dir_page(NULL, dir, pos, &chain);
1504                 } else {
1505                         /*
1506                          * go into overflow page.
1507                          */
1508                         LASSERT(le32_to_cpu(dp->ldp_flags) & LDF_COLLIDE);
1509                         ll_release_page(page, 1);
1510                 }
1511         }
1512         EXIT;
1513
1514 out:
1515         ll_dir_chain_fini(&chain);
1516         return rc;
1517 }
1518
1519 static void
1520 ll_sai_unplug(struct ll_statahead_info *sai, struct ll_sa_entry *entry)
1521 {
1522         struct ptlrpc_thread *thread = &sai->sai_thread;
1523         struct ll_sb_info    *sbi    = ll_i2sbi(sai->sai_inode);
1524         int                   hit;
1525         ENTRY;
1526
1527         if (entry != NULL && entry->se_stat == SA_ENTRY_SUCC)
1528                 hit = 1;
1529         else
1530                 hit = 0;
1531
1532         ll_sa_entry_fini(sai, entry);
1533         if (hit) {
1534                 sai->sai_hit++;
1535                 sai->sai_consecutive_miss = 0;
1536                 sai->sai_max = min(2 * sai->sai_max, sbi->ll_sa_max);
1537         } else {
1538                 struct ll_inode_info *lli = ll_i2info(sai->sai_inode);
1539
1540                 sai->sai_miss++;
1541                 sai->sai_consecutive_miss++;
1542                 if (sa_low_hit(sai) && thread_is_running(thread)) {
1543                         atomic_inc(&sbi->ll_sa_wrong);
1544                         CDEBUG(D_READA, "Statahead for dir "DFID" hit "
1545                                "ratio too low: hit/miss "LPU64"/"LPU64
1546                                ", sent/replied "LPU64"/"LPU64", stopping "
1547                                "statahead thread: pid %d\n",
1548                                PFID(&lli->lli_fid), sai->sai_hit,
1549                                sai->sai_miss, sai->sai_sent,
1550                                sai->sai_replied, cfs_curproc_pid());
1551                         cfs_spin_lock(&lli->lli_sa_lock);
1552                         if (!thread_is_stopped(thread))
1553                                 thread_set_flags(thread, SVC_STOPPING);
1554                         cfs_spin_unlock(&lli->lli_sa_lock);
1555                 }
1556         }
1557
1558         if (!thread_is_stopped(thread))
1559                 cfs_waitq_signal(&thread->t_ctl_waitq);
1560
1561         EXIT;
1562 }
1563
1564 /**
1565  * Start statahead thread if this is the first dir entry.
1566  * Otherwise if a thread is started already, wait it until it is ahead of me.
1567  * \retval 1       -- find entry with lock in cache, the caller needs to do
1568  *                    nothing.
1569  * \retval 0       -- find entry in cache, but without lock, the caller needs
1570  *                    refresh from MDS.
1571  * \retval others  -- the caller need to process as non-statahead.
1572  */
1573 int do_statahead_enter(struct inode *dir, struct dentry **dentryp,
1574                        int only_unplug)
1575 {
1576         struct ll_inode_info     *lli   = ll_i2info(dir);
1577         struct ll_statahead_info *sai   = lli->lli_sai;
1578         struct dentry            *parent;
1579         struct ll_sa_entry       *entry;
1580         struct ptlrpc_thread     *thread;
1581         struct l_wait_info        lwi   = { 0 };
1582         int                       rc    = 0;
1583         ENTRY;
1584
1585         LASSERT(lli->lli_opendir_pid == cfs_curproc_pid());
1586
1587         if (sai) {
1588                 thread = &sai->sai_thread;
1589                 if (unlikely(thread_is_stopped(thread) &&
1590                              cfs_list_empty(&sai->sai_entries_stated))) {
1591                         /* to release resource */
1592                         ll_stop_statahead(dir, lli->lli_opendir_key);
1593                         RETURN(-EAGAIN);
1594                 }
1595
1596                 if ((*dentryp)->d_name.name[0] == '.') {
1597                         if (sai->sai_ls_all ||
1598                             sai->sai_miss_hidden >= sai->sai_skip_hidden) {
1599                                 /*
1600                                  * Hidden dentry is the first one, or statahead
1601                                  * thread does not skip so many hidden dentries
1602                                  * before "sai_ls_all" enabled as below.
1603                                  */
1604                         } else {
1605                                 if (!sai->sai_ls_all)
1606                                         /*
1607                                          * It maybe because hidden dentry is not
1608                                          * the first one, "sai_ls_all" was not
1609                                          * set, then "ls -al" missed. Enable
1610                                          * "sai_ls_all" for such case.
1611                                          */
1612                                         sai->sai_ls_all = 1;
1613
1614                                 /*
1615                                  * Such "getattr" has been skipped before
1616                                  * "sai_ls_all" enabled as above.
1617                                  */
1618                                 sai->sai_miss_hidden++;
1619                                 RETURN(-EAGAIN);
1620                         }
1621                 }
1622
1623                 entry = ll_sa_entry_get_byname(sai, &(*dentryp)->d_name);
1624                 if (entry == NULL || only_unplug) {
1625                         ll_sai_unplug(sai, entry);
1626                         RETURN(entry ? 1 : -EAGAIN);
1627                 }
1628
1629                 while (!ll_sa_entry_stated(entry) &&
1630                        sai->sai_in_readpage &&
1631                        !sa_received_empty(sai))
1632                         do_statahead_interpret(sai, entry);
1633
1634                 if (!ll_sa_entry_stated(entry)) {
1635                         sai->sai_index_wait = entry->se_index;
1636                         lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(30), NULL,
1637                                                LWI_ON_SIGNAL_NOOP, NULL);
1638                         rc = l_wait_event(sai->sai_waitq,
1639                                           ll_sa_entry_stated(entry) ||
1640                                           thread_is_stopped(thread),
1641                                           &lwi);
1642                         if (rc < 0) {
1643                                 ll_sai_unplug(sai, entry);
1644                                 RETURN(-EAGAIN);
1645                         }
1646                 }
1647
1648                 if (entry->se_stat == SA_ENTRY_SUCC &&
1649                     entry->se_inode != NULL) {
1650                         struct inode *inode = entry->se_inode;
1651                         struct lookup_intent it = { .it_op = IT_GETATTR,
1652                                                     .d.lustre.it_lock_handle =
1653                                                      entry->se_handle };
1654                         struct ll_dentry_data *lld;
1655                         __u64 bits;
1656
1657                         rc = md_revalidate_lock(ll_i2mdexp(dir), &it,
1658                                                 ll_inode2fid(inode), &bits);
1659                         if (rc == 1) {
1660                                 if ((*dentryp)->d_inode == NULL) {
1661                                         *dentryp = ll_find_alias(inode,
1662                                                                  *dentryp);
1663                                         lld = ll_d2d(*dentryp);
1664                                         if (unlikely(lld == NULL))
1665                                                 ll_dops_init(*dentryp, 1, 1);
1666                                 } else {
1667                                         LASSERT((*dentryp)->d_inode == inode);
1668
1669                                         ll_dentry_rehash(*dentryp, 0);
1670                                         iput(inode);
1671                                 }
1672                                 entry->se_inode = NULL;
1673
1674                                 ll_dentry_reset_flags(*dentryp, bits);
1675                                 ll_intent_release(&it);
1676                         }
1677                 }
1678
1679                 ll_sai_unplug(sai, entry);
1680                 RETURN(rc);
1681         }
1682
1683         /* I am the "lli_opendir_pid" owner, only me can set "lli_sai". */
1684         rc = is_first_dirent(dir, *dentryp);
1685         if (rc == LS_NONE_FIRST_DE)
1686                 /* It is not "ls -{a}l" operation, no need statahead for it. */
1687                 GOTO(out, rc = -EAGAIN);
1688
1689         sai = ll_sai_alloc();
1690         if (sai == NULL)
1691                 GOTO(out, rc = -ENOMEM);
1692
1693         sai->sai_ls_all = (rc == LS_FIRST_DOT_DE);
1694         sai->sai_inode = igrab(dir);
1695         if (unlikely(sai->sai_inode == NULL)) {
1696                 CWARN("Do not start stat ahead on dying inode "DFID"\n",
1697                       PFID(&lli->lli_fid));
1698                 GOTO(out, rc = -ESTALE);
1699         }
1700
1701         /* get parent reference count here, and put it in ll_statahead_thread */
1702         parent = dget((*dentryp)->d_parent);
1703         if (unlikely(sai->sai_inode != parent->d_inode)) {
1704                 struct ll_inode_info *nlli = ll_i2info(parent->d_inode);
1705
1706                 CWARN("Race condition, someone changed %.*s just now: "
1707                       "old parent "DFID", new parent "DFID"\n",
1708                       (*dentryp)->d_name.len, (*dentryp)->d_name.name,
1709                       PFID(&lli->lli_fid), PFID(&nlli->lli_fid));
1710                 dput(parent);
1711                 iput(sai->sai_inode);
1712                 GOTO(out, rc = -EAGAIN);
1713         }
1714
1715         CDEBUG(D_READA, "start statahead thread: [pid %d] [parent %.*s]\n",
1716                cfs_curproc_pid(), parent->d_name.len, parent->d_name.name);
1717
1718         lli->lli_sai = sai;
1719         rc = cfs_create_thread(ll_statahead_thread, parent, 0);
1720         thread = &sai->sai_thread;
1721         if (rc < 0) {
1722                 CERROR("can't start ll_sa thread, rc: %d\n", rc);
1723                 dput(parent);
1724                 lli->lli_opendir_key = NULL;
1725                 thread_set_flags(thread, SVC_STOPPED);
1726                 ll_sai_put(sai);
1727                 LASSERT(lli->lli_sai == NULL);
1728                 RETURN(-EAGAIN);
1729         }
1730
1731         l_wait_event(thread->t_ctl_waitq,
1732                      thread_is_running(thread) || thread_is_stopped(thread),
1733                      &lwi);
1734
1735         /*
1736          * We don't stat-ahead for the first dirent since we are already in
1737          * lookup, and -EEXIST also indicates that this is the first dirent.
1738          */
1739         RETURN(-EEXIST);
1740
1741 out:
1742         if (sai != NULL)
1743                 OBD_FREE_PTR(sai);
1744         cfs_spin_lock(&lli->lli_sa_lock);
1745         lli->lli_opendir_key = NULL;
1746         lli->lli_opendir_pid = 0;
1747         cfs_spin_unlock(&lli->lli_sa_lock);
1748         return rc;
1749 }