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, 2012, Whamcloud, Inc.
  33  */
  34 /*
  35  * This file is part of Lustre, http://www.lustre.org/
  36  * Lustre is a trademark of Sun Microsystems, Inc.
  37  */
  38
  39 #include <linux/fs.h>
  40 #include <linux/sched.h>
  41 #include <linux/mm.h>
  42 #include <linux/smp_lock.h>
  43 #include <linux/highmem.h>
  44 #include <linux/pagemap.h>
  45
  46 #define DEBUG_SUBSYSTEM S_LLITE
  47
  48 #include <obd_support.h>
  49 #include <lustre_lite.h>
  50 #include <lustre_dlm.h>
  51 #include <linux/lustre_version.h>
  52 #include "llite_internal.h"
  53
  54 #define SA_OMITTED_ENTRY_MAX 8ULL
  55
  56 typedef enum {
  57         /** negative values are for error cases */
  58         SA_ENTRY_INIT = 0,      /** init entry */
  59         SA_ENTRY_SUCC = 1,      /** stat succeed */
  60         SA_ENTRY_INVA = 2,      /** invalid entry */
  61         SA_ENTRY_DEST = 3,      /** entry to be destroyed */
  62 } se_stat_t;
  63
  64 struct ll_sa_entry {
  65         /* link into sai->sai_entries_{sent,received,stated} */
  66         cfs_list_t              se_list;
  67         /* link into sai hash table locally */
  68         cfs_list_t              se_hash;
  69         /* entry reference count */
  70         cfs_atomic_t            se_refcount;
  71         /* entry index in the sai */
  72         __u64                   se_index;
  73         /* low layer ldlm lock handle */
  74         __u64                   se_handle;
  75         /* entry status */
  76         se_stat_t               se_stat;
  77         /* entry size, contains name */
  78         int                     se_size;
  79         /* pointer to async getattr enqueue info */
  80         struct md_enqueue_info *se_minfo;
  81         /* pointer to the async getattr request */
  82         struct ptlrpc_request  *se_req;
  83         /* pointer to the target inode */
  84         struct inode           *se_inode;
  85         /* entry name */
  86         struct qstr             se_qstr;
  87 };
  88
  89 static unsigned int sai_generation = 0;
  90 static cfs_spinlock_t sai_generation_lock = CFS_SPIN_LOCK_UNLOCKED;
  91
  92 static inline int ll_sa_entry_unlinked(struct ll_sa_entry *entry)
  93 {
  94         return cfs_list_empty(&entry->se_list);
  95 }
  96
  97 static inline int ll_sa_entry_unhashed(struct ll_sa_entry *entry)
  98 {
  99         return cfs_list_empty(&entry->se_hash);
 100 }
 101
 102 /*
 103  * The entry only can be released by the caller, it is necessary to hold lock.
 104  */
 105 static inline int ll_sa_entry_stated(struct ll_sa_entry *entry)
 106 {
 107         smp_rmb();
 108         return (entry->se_stat != SA_ENTRY_INIT);
 109 }
 110
 111 static inline int ll_sa_entry_hash(int val)
 112 {
 113         return val & LL_SA_CACHE_MASK;
 114 }
 115
 116 /*
 117  * Insert entry to hash SA table.
 118  */
 119 static inline void
 120 ll_sa_entry_enhash(struct ll_statahead_info *sai, struct ll_sa_entry *entry)
 121 {
 122         int i = ll_sa_entry_hash(entry->se_qstr.hash);
 123
 124         cfs_spin_lock(&sai->sai_cache_lock[i]);
 125         cfs_list_add_tail(&entry->se_hash, &sai->sai_cache[i]);
 126         cfs_spin_unlock(&sai->sai_cache_lock[i]);
 127 }
 128
 129 /*
 130  * Remove entry from SA table.
 131  */
 132 static inline void
 133 ll_sa_entry_unhash(struct ll_statahead_info *sai, struct ll_sa_entry *entry)
 134 {
 135         int i = ll_sa_entry_hash(entry->se_qstr.hash);
 136
 137         cfs_spin_lock(&sai->sai_cache_lock[i]);
 138         cfs_list_del_init(&entry->se_hash);
 139         cfs_spin_unlock(&sai->sai_cache_lock[i]);
 140 }
 141
 142 static inline int agl_should_run(struct ll_statahead_info *sai,
 143                                  struct inode *inode)
 144 {
 145         if (inode != NULL && S_ISREG(inode->i_mode) &&
 146             ll_i2info(inode)->lli_smd != NULL && sai->sai_agl_valid)
 147                 return 1;
 148         return 0;
 149 }
 150
 151 static inline struct ll_sa_entry *
 152 sa_first_received_entry(struct ll_statahead_info *sai)
 153 {
 154         return cfs_list_entry(sai->sai_entries_received.next,
 155                               struct ll_sa_entry, se_list);
 156 }
 157
 158 static inline struct ll_inode_info *
 159 agl_first_entry(struct ll_statahead_info *sai)
 160 {
 161         return cfs_list_entry(sai->sai_entries_agl.next,
 162                               struct ll_inode_info, lli_agl_list);
 163 }
 164
 165 static inline int sa_sent_full(struct ll_statahead_info *sai)
 166 {
 167         return cfs_atomic_read(&sai->sai_cache_count) >= sai->sai_max;
 168 }
 169
 170 static inline int sa_received_empty(struct ll_statahead_info *sai)
 171 {
 172         return cfs_list_empty(&sai->sai_entries_received);
 173 }
 174
 175 static inline int agl_list_empty(struct ll_statahead_info *sai)
 176 {
 177         return cfs_list_empty(&sai->sai_entries_agl);
 178 }
 179
 180 /**
 181  * (1) hit ratio less than 80%
 182  * or
 183  * (2) consecutive miss more than 8
 184  * then means low hit.
 185  */
 186 static inline int sa_low_hit(struct ll_statahead_info *sai)
 187 {
 188         return ((sai->sai_hit > 7 && sai->sai_hit < 4 * sai->sai_miss) ||
 189                 (sai->sai_consecutive_miss > 8));
 190 }
 191
 192 /*
 193  * If the given index is behind of statahead window more than
 194  * SA_OMITTED_ENTRY_MAX, then it is old.
 195  */
 196 static inline int is_omitted_entry(struct ll_statahead_info *sai, __u64 index)
 197 {
 198         return ((__u64)sai->sai_max + index + SA_OMITTED_ENTRY_MAX <
 199                  sai->sai_index);
 200 }
 201
 202 /*
 203  * Insert it into sai_entries_sent tail when init.
 204  */
 205 static struct ll_sa_entry *
 206 ll_sa_entry_alloc(struct ll_statahead_info *sai, __u64 index,
 207                   const char *name, int len)
 208 {
 209         struct ll_inode_info *lli;
 210         struct ll_sa_entry   *entry;
 211         int                   entry_size;
 212         char                 *dname;
 213         ENTRY;
 214
 215         entry_size = sizeof(struct ll_sa_entry) + (len & ~3) + 4;
 216         OBD_ALLOC(entry, entry_size);
 217         if (unlikely(entry == NULL))
 218                 RETURN(ERR_PTR(-ENOMEM));
 219
 220         CDEBUG(D_READA, "alloc sai entry %.*s(%p) index "LPU64"\n",
 221                len, name, entry, index);
 222
 223         entry->se_index = index;
 224
 225         /*
 226          * Statahead entry reference rules:
 227          *
 228          * 1) When statahead entry is initialized, its reference is set as 2.
 229          *    One reference is used by the directory scanner. When the scanner
 230          *    searches the statahead cache for the given name, it can perform
 231          *    lockless hash lookup (only the scanner can remove entry from hash
 232          *    list), and once found, it needn't to call "atomic_inc()" for the
 233          *    entry reference. So the performance is improved. After using the
 234          *    statahead entry, the scanner will call "atomic_dec()" to drop the
 235          *    reference held when initialization. If it is the last reference,
 236          *    the statahead entry will be freed.
 237          *
 238          * 2) All other threads, including statahead thread and ptlrpcd thread,
 239          *    when they process the statahead entry, the reference for target
 240          *    should be held to guarantee the entry will not be released by the
 241          *    directory scanner. After processing the entry, these threads will
 242          *    drop the entry reference. If it is the last reference, the entry
 243          *    will be freed.
 244          *
 245          *    The second reference when initializes the statahead entry is used
 246          *    by the statahead thread, following the rule 2).
 247          */
 248         cfs_atomic_set(&entry->se_refcount, 2);
 249         entry->se_stat = SA_ENTRY_INIT;
 250         entry->se_size = entry_size;
 251         dname = (char *)entry + sizeof(struct ll_sa_entry);
 252         memcpy(dname, name, len);
 253         dname[len] = 0;
 254         entry->se_qstr.hash = full_name_hash(name, len);
 255         entry->se_qstr.len = len;
 256         entry->se_qstr.name = dname;
 257
 258         lli = ll_i2info(sai->sai_inode);
 259         cfs_spin_lock(&lli->lli_sa_lock);
 260         cfs_list_add_tail(&entry->se_list, &sai->sai_entries_sent);
 261         cfs_spin_unlock(&lli->lli_sa_lock);
 262
 263         cfs_atomic_inc(&sai->sai_cache_count);
 264         ll_sa_entry_enhash(sai, entry);
 265
 266         RETURN(entry);
 267 }
 268
 269 /*
 270  * Used by the directory scanner to search entry with name.
 271  *
 272  * Only the caller can remove the entry from hash, so it is unnecessary to hold
 273  * hash lock. It is caller's duty to release the init refcount on the entry, so
 274  * it is also unnecessary to increase refcount on the entry.
 275  */
 276 static struct ll_sa_entry *
 277 ll_sa_entry_get_byname(struct ll_statahead_info *sai, const struct qstr *qstr)
 278 {
 279         struct ll_sa_entry *entry;
 280         int i = ll_sa_entry_hash(qstr->hash);
 281
 282         cfs_list_for_each_entry(entry, &sai->sai_cache[i], se_hash) {
 283                 if (entry->se_qstr.hash == qstr->hash &&
 284                     entry->se_qstr.len == qstr->len &&
 285                     memcmp(entry->se_qstr.name, qstr->name, qstr->len) == 0)
 286                         return entry;
 287         }
 288         return NULL;
 289 }
 290
 291 /*
 292  * Used by the async getattr request callback to find entry with index.
 293  *
 294  * Inside lli_sa_lock to prevent others to change the list during the search.
 295  * It needs to increase entry refcount before returning to guarantee that the
 296  * entry cannot be freed by others.
 297  */
 298 static struct ll_sa_entry *
 299 ll_sa_entry_get_byindex(struct ll_statahead_info *sai, __u64 index)
 300 {
 301         struct ll_sa_entry *entry;
 302
 303         cfs_list_for_each_entry(entry, &sai->sai_entries_sent, se_list) {
 304                 if (entry->se_index == index) {
 305                         cfs_atomic_inc(&entry->se_refcount);
 306                         return entry;
 307                 }
 308                 if (entry->se_index > index)
 309                         break;
 310         }
 311         return NULL;
 312 }
 313
 314 static void ll_sa_entry_cleanup(struct ll_statahead_info *sai,
 315                                  struct ll_sa_entry *entry)
 316 {
 317         struct md_enqueue_info *minfo = entry->se_minfo;
 318         struct ptlrpc_request  *req   = entry->se_req;
 319
 320         if (minfo) {
 321                 entry->se_minfo = NULL;
 322                 ll_intent_release(&minfo->mi_it);
 323                 iput(minfo->mi_dir);
 324                 OBD_FREE_PTR(minfo);
 325         }
 326
 327         if (req) {
 328                 entry->se_req = NULL;
 329                 ptlrpc_req_finished(req);
 330         }
 331 }
 332
 333 static void ll_sa_entry_put(struct ll_statahead_info *sai,
 334                              struct ll_sa_entry *entry)
 335 {
 336         if (cfs_atomic_dec_and_test(&entry->se_refcount)) {
 337                 CDEBUG(D_READA, "free sai entry %.*s(%p) index "LPU64"\n",
 338                        entry->se_qstr.len, entry->se_qstr.name, entry,
 339                        entry->se_index);
 340
 341                 LASSERT(ll_sa_entry_unhashed(entry));
 342                 LASSERT(ll_sa_entry_unlinked(entry));
 343
 344                 ll_sa_entry_cleanup(sai, entry);
 345                 if (entry->se_inode)
 346                         iput(entry->se_inode);
 347
 348                 OBD_FREE(entry, entry->se_size);
 349                 cfs_atomic_dec(&sai->sai_cache_count);
 350         }
 351 }
 352
 353 static inline void
 354 do_sai_entry_fini(struct ll_statahead_info *sai, struct ll_sa_entry *entry)
 355 {
 356         struct ll_inode_info *lli = ll_i2info(sai->sai_inode);
 357
 358         ll_sa_entry_unhash(sai, entry);
 359
 360         cfs_spin_lock(&lli->lli_sa_lock);
 361         entry->se_stat = SA_ENTRY_DEST;
 362         if (likely(!ll_sa_entry_unlinked(entry)))
 363                 cfs_list_del_init(&entry->se_list);
 364         cfs_spin_unlock(&lli->lli_sa_lock);
 365
 366         ll_sa_entry_put(sai, entry);
 367 }
 368
 369 /*
 370  * Delete it from sai_entries_stated list when fini.
 371  */
 372 static void
 373 ll_sa_entry_fini(struct ll_statahead_info *sai, struct ll_sa_entry *entry)
 374 {
 375         struct ll_sa_entry *pos, *next;
 376
 377         if (entry)
 378                 do_sai_entry_fini(sai, entry);
 379
 380         /* drop old entry from sent list */
 381         cfs_list_for_each_entry_safe(pos, next, &sai->sai_entries_sent,
 382                                      se_list) {
 383                 if (is_omitted_entry(sai, pos->se_index))
 384                         do_sai_entry_fini(sai, pos);
 385                 else
 386                         break;
 387         }
 388
 389         /* drop old entry from stated list */
 390         cfs_list_for_each_entry_safe(pos, next, &sai->sai_entries_stated,
 391                                      se_list) {
 392                 if (is_omitted_entry(sai, pos->se_index))
 393                         do_sai_entry_fini(sai, pos);
 394                 else
 395                         break;
 396         }
 397 }
 398
 399 /*
 400  * Inside lli_sa_lock.
 401  */
 402 static void
 403 do_sai_entry_to_stated(struct ll_statahead_info *sai,
 404                        struct ll_sa_entry *entry, int rc)
 405 {
 406         struct ll_sa_entry *se;
 407         cfs_list_t         *pos = &sai->sai_entries_stated;
 408
 409         if (!ll_sa_entry_unlinked(entry))
 410                 cfs_list_del_init(&entry->se_list);
 411
 412         cfs_list_for_each_entry_reverse(se, &sai->sai_entries_stated, se_list) {
 413                 if (se->se_index < entry->se_index) {
 414                         pos = &se->se_list;
 415                         break;
 416                 }
 417         }
 418
 419         cfs_list_add(&entry->se_list, pos);
 420         entry->se_stat = rc;
 421 }
 422
 423 /*
 424  * Move entry to sai_entries_stated and sort with the index.
 425  * \retval 1    -- entry to be destroyed.
 426  * \retval 0    -- entry is inserted into stated list.
 427  */
 428 static int
 429 ll_sa_entry_to_stated(struct ll_statahead_info *sai,
 430                        struct ll_sa_entry *entry, int rc)
 431 {
 432         struct ll_inode_info *lli = ll_i2info(sai->sai_inode);
 433         int                   ret = 1;
 434
 435         ll_sa_entry_cleanup(sai, entry);
 436
 437         cfs_spin_lock(&lli->lli_sa_lock);
 438         if (likely(entry->se_stat != SA_ENTRY_DEST)) {
 439                 do_sai_entry_to_stated(sai, entry, rc);
 440                 ret = 0;
 441         }
 442         cfs_spin_unlock(&lli->lli_sa_lock);
 443
 444         return ret;
 445 }
 446
 447 /*
 448  * Insert inode into the list of sai_entries_agl.
 449  */
 450 static void ll_agl_add(struct ll_statahead_info *sai,
 451                        struct inode *inode, int index)
 452 {
 453         struct ll_inode_info *child  = ll_i2info(inode);
 454         struct ll_inode_info *parent = ll_i2info(sai->sai_inode);
 455         int                   added  = 0;
 456
 457         cfs_spin_lock(&child->lli_agl_lock);
 458         if (child->lli_agl_index == 0) {
 459                 child->lli_agl_index = index;
 460                 cfs_spin_unlock(&child->lli_agl_lock);
 461
 462                 LASSERT(cfs_list_empty(&child->lli_agl_list));
 463
 464                 igrab(inode);
 465                 cfs_spin_lock(&parent->lli_agl_lock);
 466                 if (agl_list_empty(sai))
 467                         added = 1;
 468                 cfs_list_add_tail(&child->lli_agl_list, &sai->sai_entries_agl);
 469                 cfs_spin_unlock(&parent->lli_agl_lock);
 470         } else {
 471                 cfs_spin_unlock(&child->lli_agl_lock);
 472         }
 473
 474         if (added > 0)
 475                 cfs_waitq_signal(&sai->sai_agl_thread.t_ctl_waitq);
 476 }
 477
 478 static struct ll_statahead_info *ll_sai_alloc(void)
 479 {
 480         struct ll_statahead_info *sai;
 481         int                       i;
 482         ENTRY;
 483
 484         OBD_ALLOC_PTR(sai);
 485         if (!sai)
 486                 RETURN(NULL);
 487
 488         cfs_atomic_set(&sai->sai_refcount, 1);
 489
 490         cfs_spin_lock(&sai_generation_lock);
 491         sai->sai_generation = ++sai_generation;
 492         if (unlikely(sai_generation == 0))
 493                 sai->sai_generation = ++sai_generation;
 494         cfs_spin_unlock(&sai_generation_lock);
 495
 496         sai->sai_max = LL_SA_RPC_MIN;
 497         sai->sai_index = 1;
 498         cfs_waitq_init(&sai->sai_waitq);
 499         cfs_waitq_init(&sai->sai_thread.t_ctl_waitq);
 500         cfs_waitq_init(&sai->sai_agl_thread.t_ctl_waitq);
 501
 502         CFS_INIT_LIST_HEAD(&sai->sai_entries_sent);
 503         CFS_INIT_LIST_HEAD(&sai->sai_entries_received);
 504         CFS_INIT_LIST_HEAD(&sai->sai_entries_stated);
 505         CFS_INIT_LIST_HEAD(&sai->sai_entries_agl);
 506
 507         for (i = 0; i < LL_SA_CACHE_SIZE; i++) {
 508                 CFS_INIT_LIST_HEAD(&sai->sai_cache[i]);
 509                 cfs_spin_lock_init(&sai->sai_cache_lock[i]);
 510         }
 511         cfs_atomic_set(&sai->sai_cache_count, 0);
 512
 513         RETURN(sai);
 514 }
 515
 516 static inline struct ll_statahead_info *
 517 ll_sai_get(struct ll_statahead_info *sai)
 518 {
 519         cfs_atomic_inc(&sai->sai_refcount);
 520         return sai;
 521 }
 522
 523 static void ll_sai_put(struct ll_statahead_info *sai)
 524 {
 525         struct inode         *inode = sai->sai_inode;
 526         struct ll_inode_info *lli   = ll_i2info(inode);
 527         ENTRY;
 528
 529         if (cfs_atomic_dec_and_lock(&sai->sai_refcount, &lli->lli_sa_lock)) {
 530                 struct ll_sa_entry *entry, *next;
 531
 532                 if (unlikely(cfs_atomic_read(&sai->sai_refcount) > 0)) {
 533                         /* It is race case, the interpret callback just hold
 534                          * a reference count */
 535                         cfs_spin_unlock(&lli->lli_sa_lock);
 536                         RETURN_EXIT;
 537                 }
 538
 539                 LASSERT(lli->lli_opendir_key == NULL);
 540                 LASSERT(thread_is_stopped(&sai->sai_thread));
 541                 LASSERT(thread_is_stopped(&sai->sai_agl_thread));
 542
 543                 lli->lli_sai = NULL;
 544                 lli->lli_opendir_pid = 0;
 545                 cfs_spin_unlock(&lli->lli_sa_lock);
 546
 547                 if (sai->sai_sent > sai->sai_replied)
 548                         CDEBUG(D_READA,"statahead for dir "DFID" does not "
 549                               "finish: [sent:"LPU64"] [replied:"LPU64"]\n",
 550                               PFID(&lli->lli_fid),
 551                               sai->sai_sent, sai->sai_replied);
 552
 553                 cfs_list_for_each_entry_safe(entry, next,
 554                                              &sai->sai_entries_sent, se_list)
 555                         do_sai_entry_fini(sai, entry);
 556
 557                 LASSERT(sa_received_empty(sai));
 558
 559                 cfs_list_for_each_entry_safe(entry, next,
 560                                              &sai->sai_entries_stated, se_list)
 561                         do_sai_entry_fini(sai, entry);
 562
 563                 LASSERT(cfs_atomic_read(&sai->sai_cache_count) == 0);
 564                 LASSERT(agl_list_empty(sai));
 565
 566                 iput(inode);
 567                 OBD_FREE_PTR(sai);
 568         }
 569
 570         EXIT;
 571 }
 572
 573 /* Do NOT forget to drop inode refcount when into sai_entries_agl. */
 574 static void ll_agl_trigger(struct inode *inode, struct ll_statahead_info *sai)
 575 {
 576         struct ll_inode_info *lli   = ll_i2info(inode);
 577         __u64                 index = lli->lli_agl_index;
 578         int                   rc;
 579         ENTRY;
 580
 581         LASSERT(cfs_list_empty(&lli->lli_agl_list));
 582
 583         /* AGL maybe fall behind statahead with one entry */
 584         if (is_omitted_entry(sai, index + 1)) {
 585                 lli->lli_agl_index = 0;
 586                 iput(inode);
 587                 RETURN_EXIT;
 588         }
 589
 590         /* Someone is in glimpse (sync or async), do nothing. */
 591         rc = cfs_down_write_trylock(&lli->lli_glimpse_sem);
 592         if (rc == 0) {
 593                 lli->lli_agl_index = 0;
 594                 iput(inode);
 595                 RETURN_EXIT;
 596         }
 597
 598         /*
 599          * Someone triggered glimpse within 1 sec before.
 600          * 1) The former glimpse succeeded with glimpse lock granted by OST, and
 601          *    if the lock is still cached on client, AGL needs to do nothing. If
 602          *    it is cancelled by other client, AGL maybe cannot obtaion new lock
 603          *    for no glimpse callback triggered by AGL.
 604          * 2) The former glimpse succeeded, but OST did not grant glimpse lock.
 605          *    Under such case, it is quite possible that the OST will not grant
 606          *    glimpse lock for AGL also.
 607          * 3) The former glimpse failed, compared with other two cases, it is
 608          *    relative rare. AGL can ignore such case, and it will not muchly
 609          *    affect the performance.
 610          */
 611         if (lli->lli_glimpse_time != 0 &&
 612             cfs_time_before(cfs_time_shift(-1), lli->lli_glimpse_time)) {
 613                 cfs_up_write(&lli->lli_glimpse_sem);
 614                 lli->lli_agl_index = 0;
 615                 iput(inode);
 616                 RETURN_EXIT;
 617         }
 618
 619         CDEBUG(D_READA, "Handling (init) async glimpse: inode = "
 620                DFID", idx = "LPU64"\n", PFID(&lli->lli_fid), index);
 621
 622         cl_agl(inode);
 623         lli->lli_agl_index = 0;
 624         lli->lli_glimpse_time = cfs_time_current();
 625         cfs_up_write(&lli->lli_glimpse_sem);
 626
 627         CDEBUG(D_READA, "Handled (init) async glimpse: inode= "
 628                DFID", idx = "LPU64", rc = %d\n",
 629                PFID(&lli->lli_fid), index, rc);
 630
 631         iput(inode);
 632
 633         EXIT;
 634 }
 635
 636 static void do_statahead_interpret(struct ll_statahead_info *sai,
 637                                    struct ll_sa_entry *target)
 638 {
 639         struct inode           *dir   = sai->sai_inode;
 640         struct inode           *child;
 641         struct ll_inode_info   *lli   = ll_i2info(dir);
 642         struct ll_sa_entry     *entry;
 643         struct md_enqueue_info *minfo;
 644         struct lookup_intent   *it;
 645         struct ptlrpc_request  *req;
 646         struct mdt_body        *body;
 647         int                     rc    = 0;
 648         ENTRY;
 649
 650         cfs_spin_lock(&lli->lli_sa_lock);
 651         if (target != NULL && target->se_req != NULL &&
 652             !cfs_list_empty(&target->se_list)) {
 653                 entry = target;
 654         } else if (unlikely(sa_received_empty(sai))) {
 655                 cfs_spin_unlock(&lli->lli_sa_lock);
 656                 RETURN_EXIT;
 657         } else {
 658                 entry = sa_first_received_entry(sai);
 659         }
 660
 661         cfs_atomic_inc(&entry->se_refcount);
 662         cfs_list_del_init(&entry->se_list);
 663         cfs_spin_unlock(&lli->lli_sa_lock);
 664
 665         LASSERT(entry->se_handle != 0);
 666
 667         minfo = entry->se_minfo;
 668         it = &minfo->mi_it;
 669         req = entry->se_req;
 670         body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
 671         if (body == NULL)
 672                 GOTO(out, rc = -EFAULT);
 673
 674         child = entry->se_inode;
 675         if (child == NULL) {
 676                 /*
 677                  * lookup.
 678                  */
 679                 LASSERT(fid_is_zero(&minfo->mi_data.op_fid2));
 680
 681                 /* XXX: No fid in reply, this is probaly cross-ref case.
 682                  * SA can't handle it yet. */
 683                 if (body->valid & OBD_MD_MDS)
 684                         GOTO(out, rc = -EAGAIN);
 685         } else {
 686                 /*
 687                  * revalidate.
 688                  */
 689                 /* unlinked and re-created with the same name */
 690                 if (unlikely(!lu_fid_eq(&minfo->mi_data.op_fid2, &body->fid1))){
 691                         entry->se_inode = NULL;
 692                         iput(child);
 693                         child = NULL;
 694                 }
 695         }
 696
 697         it->d.lustre.it_lock_handle = entry->se_handle;
 698         rc = md_revalidate_lock(ll_i2mdexp(dir), it, NULL, NULL);
 699         if (rc != 1)
 700                 GOTO(out, rc = -EAGAIN);
 701
 702         rc = ll_prep_inode(&child, req, dir->i_sb);
 703         if (rc)
 704                 GOTO(out, rc);
 705
 706         CDEBUG(D_DLMTRACE, "setting l_data to inode %p (%lu/%u)\n",
 707                child, child->i_ino, child->i_generation);
 708         ll_set_lock_data(ll_i2sbi(dir)->ll_md_exp, child, it, NULL);
 709
 710         entry->se_inode = child;
 711
 712         if (agl_should_run(sai, child))
 713                 ll_agl_add(sai, child, entry->se_index);
 714
 715         EXIT;
 716
 717 out:
 718         /* The "ll_sa_entry_to_stated()" will drop related ldlm ibits lock
 719          * reference count by calling "ll_intent_drop_lock()" in spite of the
 720          * above operations failed or not. Do not worry about calling
 721          * "ll_intent_drop_lock()" more than once. */
 722         rc = ll_sa_entry_to_stated(sai, entry, rc < 0 ? rc : SA_ENTRY_SUCC);
 723         if (rc == 0 && entry->se_index == sai->sai_index_wait && target == NULL)
 724                 cfs_waitq_signal(&sai->sai_waitq);
 725         ll_sa_entry_put(sai, entry);
 726 }
 727
 728 static int ll_statahead_interpret(struct ptlrpc_request *req,
 729                                   struct md_enqueue_info *minfo, int rc)
 730 {
 731         struct lookup_intent     *it  = &minfo->mi_it;
 732         struct inode             *dir = minfo->mi_dir;
 733         struct ll_inode_info     *lli = ll_i2info(dir);
 734         struct ll_statahead_info *sai = NULL;
 735         struct ll_sa_entry       *entry;
 736         int                       wakeup;
 737         ENTRY;
 738
 739         if (it_disposition(it, DISP_LOOKUP_NEG))
 740                 rc = -ENOENT;
 741
 742         cfs_spin_lock(&lli->lli_sa_lock);
 743         /* stale entry */
 744         if (unlikely(lli->lli_sai == NULL ||
 745                      lli->lli_sai->sai_generation != minfo->mi_generation)) {
 746                 cfs_spin_unlock(&lli->lli_sa_lock);
 747                 GOTO(out, rc = -ESTALE);
 748         } else {
 749                 sai = ll_sai_get(lli->lli_sai);
 750                 if (unlikely(!thread_is_running(&sai->sai_thread))) {
 751                         sai->sai_replied++;
 752                         cfs_spin_unlock(&lli->lli_sa_lock);
 753                         GOTO(out, rc = -EBADFD);
 754                 }
 755
 756                 entry = ll_sa_entry_get_byindex(sai, minfo->mi_cbdata);
 757                 if (entry == NULL) {
 758                         sai->sai_replied++;
 759                         cfs_spin_unlock(&lli->lli_sa_lock);
 760                         GOTO(out, rc = -EIDRM);
 761                 }
 762
 763                 cfs_list_del_init(&entry->se_list);
 764                 if (rc != 0) {
 765                         sai->sai_replied++;
 766                         do_sai_entry_to_stated(sai, entry, rc);
 767                         cfs_spin_unlock(&lli->lli_sa_lock);
 768                         if (entry->se_index == sai->sai_index_wait)
 769                                 cfs_waitq_signal(&sai->sai_waitq);
 770                 } else {
 771                         entry->se_minfo = minfo;
 772                         entry->se_req = ptlrpc_request_addref(req);
 773                         /* Release the async ibits lock ASAP to avoid deadlock
 774                          * when statahead thread tries to enqueue lock on parent
 775                          * for readpage and other tries to enqueue lock on child
 776                          * with parent's lock held, for example: unlink. */
 777                         entry->se_handle = it->d.lustre.it_lock_handle;
 778                         ll_intent_drop_lock(it);
 779                         wakeup = sa_received_empty(sai);
 780                         cfs_list_add_tail(&entry->se_list,
 781                                           &sai->sai_entries_received);
 782                         sai->sai_replied++;
 783                         cfs_spin_unlock(&lli->lli_sa_lock);
 784                         if (wakeup)
 785                                 cfs_waitq_signal(&sai->sai_thread.t_ctl_waitq);
 786                 }
 787                 ll_sa_entry_put(sai, entry);
 788         }
 789
 790         EXIT;
 791
 792 out:
 793         if (rc != 0) {
 794                 ll_intent_release(it);
 795                 iput(dir);
 796                 OBD_FREE_PTR(minfo);
 797         }
 798         if (sai != NULL)
 799                 ll_sai_put(sai);
 800         return rc;
 801 }
 802
 803 static void sa_args_fini(struct md_enqueue_info *minfo,
 804                          struct ldlm_enqueue_info *einfo)
 805 {
 806         LASSERT(minfo && einfo);
 807         iput(minfo->mi_dir);
 808         capa_put(minfo->mi_data.op_capa1);
 809         capa_put(minfo->mi_data.op_capa2);
 810         OBD_FREE_PTR(minfo);
 811         OBD_FREE_PTR(einfo);
 812 }
 813
 814 /**
 815  * There is race condition between "capa_put" and "ll_statahead_interpret" for
 816  * accessing "op_data.op_capa[1,2]" as following:
 817  * "capa_put" releases "op_data.op_capa[1,2]"'s reference count after calling
 818  * "md_intent_getattr_async". But "ll_statahead_interpret" maybe run first, and
 819  * fill "op_data.op_capa[1,2]" as POISON, then cause "capa_put" access invalid
 820  * "ocapa". So here reserve "op_data.op_capa[1,2]" in "pcapa" before calling
 821  * "md_intent_getattr_async".
 822  */
 823 static int sa_args_init(struct inode *dir, struct inode *child,
 824                         struct ll_sa_entry *entry, struct md_enqueue_info **pmi,
 825                         struct ldlm_enqueue_info **pei,
 826                         struct obd_capa **pcapa)
 827 {
 828         struct qstr              *qstr = &entry->se_qstr;
 829         struct ll_inode_info     *lli  = ll_i2info(dir);
 830         struct md_enqueue_info   *minfo;
 831         struct ldlm_enqueue_info *einfo;
 832         struct md_op_data        *op_data;
 833
 834         OBD_ALLOC_PTR(einfo);
 835         if (einfo == NULL)
 836                 return -ENOMEM;
 837
 838         OBD_ALLOC_PTR(minfo);
 839         if (minfo == NULL) {
 840                 OBD_FREE_PTR(einfo);
 841                 return -ENOMEM;
 842         }
 843
 844         op_data = ll_prep_md_op_data(&minfo->mi_data, dir, child, qstr->name,
 845                                      qstr->len, 0, LUSTRE_OPC_ANY, NULL);
 846         if (IS_ERR(op_data)) {
 847                 OBD_FREE_PTR(einfo);
 848                 OBD_FREE_PTR(minfo);
 849                 return PTR_ERR(op_data);
 850         }
 851
 852         minfo->mi_it.it_op = IT_GETATTR;
 853         minfo->mi_dir = igrab(dir);
 854         minfo->mi_cb = ll_statahead_interpret;
 855         minfo->mi_generation = lli->lli_sai->sai_generation;
 856         minfo->mi_cbdata = entry->se_index;
 857
 858         einfo->ei_type   = LDLM_IBITS;
 859         einfo->ei_mode   = it_to_lock_mode(&minfo->mi_it);
 860         einfo->ei_cb_bl  = ll_md_blocking_ast;
 861         einfo->ei_cb_cp  = ldlm_completion_ast;
 862         einfo->ei_cb_gl  = NULL;
 863         einfo->ei_cbdata = NULL;
 864
 865         *pmi = minfo;
 866         *pei = einfo;
 867         pcapa[0] = op_data->op_capa1;
 868         pcapa[1] = op_data->op_capa2;
 869
 870         return 0;
 871 }
 872
 873 static int do_sa_lookup(struct inode *dir, struct ll_sa_entry *entry)
 874 {
 875         struct md_enqueue_info   *minfo;
 876         struct ldlm_enqueue_info *einfo;
 877         struct obd_capa          *capas[2];
 878         int                       rc;
 879         ENTRY;
 880
 881         rc = sa_args_init(dir, NULL, entry, &minfo, &einfo, capas);
 882         if (rc)
 883                 RETURN(rc);
 884
 885         rc = md_intent_getattr_async(ll_i2mdexp(dir), minfo, einfo);
 886         if (!rc) {
 887                 capa_put(capas[0]);
 888                 capa_put(capas[1]);
 889         } else {
 890                 sa_args_fini(minfo, einfo);
 891         }
 892
 893         RETURN(rc);
 894 }
 895
 896 /**
 897  * similar to ll_revalidate_it().
 898  * \retval      1 -- dentry valid
 899  * \retval      0 -- will send stat-ahead request
 900  * \retval others -- prepare stat-ahead request failed
 901  */
 902 static int do_sa_revalidate(struct inode *dir, struct ll_sa_entry *entry,
 903                             struct dentry *dentry)
 904 {
 905         struct inode             *inode = dentry->d_inode;
 906         struct lookup_intent      it = { .it_op = IT_GETATTR,
 907                                          .d.lustre.it_lock_handle = 0 };
 908         struct md_enqueue_info   *minfo;
 909         struct ldlm_enqueue_info *einfo;
 910         struct obd_capa          *capas[2];
 911         int rc;
 912         ENTRY;
 913
 914         if (unlikely(inode == NULL))
 915                 RETURN(1);
 916
 917         if (d_mountpoint(dentry))
 918                 RETURN(1);
 919
 920         if (unlikely(dentry == dentry->d_sb->s_root))
 921                 RETURN(1);
 922
 923         entry->se_inode = igrab(inode);
 924         rc = md_revalidate_lock(ll_i2mdexp(dir), &it, ll_inode2fid(inode),NULL);
 925         if (rc == 1) {
 926                 entry->se_handle = it.d.lustre.it_lock_handle;
 927                 ll_intent_release(&it);
 928                 RETURN(1);
 929         }
 930
 931         rc = sa_args_init(dir, inode, entry, &minfo, &einfo, capas);
 932         if (rc) {
 933                 entry->se_inode = NULL;
 934                 iput(inode);
 935                 RETURN(rc);
 936         }
 937
 938         rc = md_intent_getattr_async(ll_i2mdexp(dir), minfo, einfo);
 939         if (!rc) {
 940                 capa_put(capas[0]);
 941                 capa_put(capas[1]);
 942         } else {
 943                 entry->se_inode = NULL;
 944                 iput(inode);
 945                 sa_args_fini(minfo, einfo);
 946         }
 947
 948         RETURN(rc);
 949 }
 950
 951 static void ll_statahead_one(struct dentry *parent, const char* entry_name,
 952                              int entry_name_len)
 953 {
 954         struct inode             *dir    = parent->d_inode;
 955         struct ll_inode_info     *lli    = ll_i2info(dir);
 956         struct ll_statahead_info *sai    = lli->lli_sai;
 957         struct dentry            *dentry = NULL;
 958         struct ll_sa_entry       *entry;
 959         int                       rc;
 960         int                       rc1;
 961         ENTRY;
 962
 963         entry = ll_sa_entry_alloc(sai, sai->sai_index, entry_name,
 964                                   entry_name_len);
 965         if (IS_ERR(entry))
 966                 RETURN_EXIT;
 967
 968         dentry = d_lookup(parent, &entry->se_qstr);
 969         if (!dentry) {
 970                 rc = do_sa_lookup(dir, entry);
 971         } else {
 972                 rc = do_sa_revalidate(dir, entry, dentry);
 973                 if (rc == 1 && agl_should_run(sai, dentry->d_inode))
 974                         ll_agl_add(sai, dentry->d_inode, entry->se_index);
 975         }
 976
 977         if (dentry != NULL)
 978                 dput(dentry);
 979
 980         if (rc) {
 981                 rc1 = ll_sa_entry_to_stated(sai, entry,
 982                                         rc < 0 ? SA_ENTRY_INVA : SA_ENTRY_SUCC);
 983                 if (rc1 == 0 && entry->se_index == sai->sai_index_wait)
 984                         cfs_waitq_signal(&sai->sai_waitq);
 985         } else {
 986                 sai->sai_sent++;
 987         }
 988
 989         sai->sai_index++;
 990         /* drop one refcount on entry by ll_sa_entry_alloc */
 991         ll_sa_entry_put(sai, entry);
 992
 993         EXIT;
 994 }
 995
 996 static int ll_agl_thread(void *arg)
 997 {
 998         struct dentry            *parent = (struct dentry *)arg;
 999         struct inode             *dir    = parent->d_inode;
1000         struct ll_inode_info     *plli   = ll_i2info(dir);
1001         struct ll_inode_info     *clli;
1002         struct ll_sb_info        *sbi    = ll_i2sbi(dir);
1003         struct ll_statahead_info *sai    = ll_sai_get(plli->lli_sai);
1004         struct ptlrpc_thread     *thread = &sai->sai_agl_thread;
1005         struct l_wait_info        lwi    = { 0 };
1006         ENTRY;
1007
1008         {
1009                 char pname[16];
1010                 snprintf(pname, 15, "ll_agl_%u", plli->lli_opendir_pid);
1011                 cfs_daemonize(pname);
1012         }
1013
1014         CDEBUG(D_READA, "agl thread started: [pid %d] [parent %.*s]\n",
1015                cfs_curproc_pid(), parent->d_name.len, parent->d_name.name);
1016
1017         atomic_inc(&sbi->ll_agl_total);
1018         cfs_spin_lock(&plli->lli_agl_lock);
1019         sai->sai_agl_valid = 1;
1020         thread_set_flags(thread, SVC_RUNNING);
1021         cfs_spin_unlock(&plli->lli_agl_lock);
1022         cfs_waitq_signal(&thread->t_ctl_waitq);
1023
1024         while (1) {
1025                 l_wait_event(thread->t_ctl_waitq,
1026                              !agl_list_empty(sai) ||
1027                              !thread_is_running(thread),
1028                              &lwi);
1029
1030                 if (!thread_is_running(thread))
1031                         break;
1032
1033                 cfs_spin_lock(&plli->lli_agl_lock);
1034                 /* The statahead thread maybe help to process AGL entries,
1035                  * so check whether list empty again. */
1036                 if (!agl_list_empty(sai)) {
1037                         clli = agl_first_entry(sai);
1038                         cfs_list_del_init(&clli->lli_agl_list);
1039                         cfs_spin_unlock(&plli->lli_agl_lock);
1040                         ll_agl_trigger(&clli->lli_vfs_inode, sai);
1041                 } else {
1042                         cfs_spin_unlock(&plli->lli_agl_lock);
1043                 }
1044         }
1045
1046         cfs_spin_lock(&plli->lli_agl_lock);
1047         sai->sai_agl_valid = 0;
1048         while (!agl_list_empty(sai)) {
1049                 clli = agl_first_entry(sai);
1050                 cfs_list_del_init(&clli->lli_agl_list);
1051                 cfs_spin_unlock(&plli->lli_agl_lock);
1052                 clli->lli_agl_index = 0;
1053                 iput(&clli->lli_vfs_inode);
1054                 cfs_spin_lock(&plli->lli_agl_lock);
1055         }
1056         thread_set_flags(thread, SVC_STOPPED);
1057         cfs_spin_unlock(&plli->lli_agl_lock);
1058         cfs_waitq_signal(&thread->t_ctl_waitq);
1059         ll_sai_put(sai);
1060         CDEBUG(D_READA, "agl thread stopped: [pid %d] [parent %.*s]\n",
1061                cfs_curproc_pid(), parent->d_name.len, parent->d_name.name);
1062         RETURN(0);
1063 }
1064
1065 static void ll_start_agl(struct dentry *parent, struct ll_statahead_info *sai)
1066 {
1067         struct ptlrpc_thread *thread = &sai->sai_agl_thread;
1068         struct l_wait_info    lwi    = { 0 };
1069         int                   rc;
1070         ENTRY;
1071
1072         CDEBUG(D_READA, "start agl thread: [pid %d] [parent %.*s]\n",
1073                cfs_curproc_pid(), parent->d_name.len, parent->d_name.name);
1074
1075         rc = cfs_create_thread(ll_agl_thread, parent, 0);
1076         if (rc < 0) {
1077                 CERROR("can't start ll_agl thread, rc: %d\n", rc);
1078                 RETURN_EXIT;
1079         }
1080
1081         l_wait_event(thread->t_ctl_waitq,
1082                      thread_is_running(thread) || thread_is_stopped(thread),
1083                      &lwi);
1084         EXIT;
1085 }
1086
1087 static int ll_statahead_thread(void *arg)
1088 {
1089         struct dentry            *parent = (struct dentry *)arg;
1090         struct inode             *dir    = parent->d_inode;
1091         struct ll_inode_info     *plli   = ll_i2info(dir);
1092         struct ll_inode_info     *clli;
1093         struct ll_sb_info        *sbi    = ll_i2sbi(dir);
1094         struct ll_statahead_info *sai    = ll_sai_get(plli->lli_sai);
1095         struct ptlrpc_thread     *thread = &sai->sai_thread;
1096         struct page              *page;
1097         __u64                     pos    = 0;
1098         int                       first  = 0;
1099         int                       rc     = 0;
1100         struct ll_dir_chain       chain;
1101         struct l_wait_info        lwi    = { 0 };
1102         ENTRY;
1103
1104         {
1105                 char pname[16];
1106                 snprintf(pname, 15, "ll_sa_%u", plli->lli_opendir_pid);
1107                 cfs_daemonize(pname);
1108         }
1109
1110         CDEBUG(D_READA, "statahead thread started: [pid %d] [parent %.*s]\n",
1111                cfs_curproc_pid(), parent->d_name.len, parent->d_name.name);
1112
1113         if (sbi->ll_flags & LL_SBI_AGL_ENABLED)
1114                 ll_start_agl(parent, sai);
1115
1116         atomic_inc(&sbi->ll_sa_total);
1117         cfs_spin_lock(&plli->lli_sa_lock);
1118         thread_set_flags(thread, SVC_RUNNING);
1119         cfs_spin_unlock(&plli->lli_sa_lock);
1120         cfs_waitq_signal(&thread->t_ctl_waitq);
1121
1122         plli->lli_sa_pos = 0;
1123         ll_dir_chain_init(&chain);
1124         page = ll_get_dir_page(NULL, dir, pos, &chain);
1125
1126         while (1) {
1127                 struct lu_dirpage *dp;
1128                 struct lu_dirent  *ent;
1129
1130                 if (IS_ERR(page)) {
1131                         rc = PTR_ERR(page);
1132                         CDEBUG(D_READA, "error reading dir "DFID" at "LPU64
1133                                "/"LPU64": [rc %d] [parent %u]\n",
1134                                PFID(ll_inode2fid(dir)), pos, sai->sai_index,
1135                                rc, plli->lli_opendir_pid);
1136                         GOTO(out, rc);
1137                 }
1138
1139                 dp = page_address(page);
1140                 for (ent = lu_dirent_start(dp); ent != NULL;
1141                      ent = lu_dirent_next(ent)) {
1142                         __u64 hash;
1143                         int namelen;
1144                         char *name;
1145
1146                         hash = le64_to_cpu(ent->lde_hash);
1147                         if (unlikely(hash < pos))
1148                                 /*
1149                                  * Skip until we find target hash value.
1150                                  */
1151                                 continue;
1152
1153                         namelen = le16_to_cpu(ent->lde_namelen);
1154                         if (unlikely(namelen == 0))
1155                                 /*
1156                                  * Skip dummy record.
1157                                  */
1158                                 continue;
1159
1160                         name = ent->lde_name;
1161                         if (name[0] == '.') {
1162                                 if (namelen == 1) {
1163                                         /*
1164                                          * skip "."
1165                                          */
1166                                         continue;
1167                                 } else if (name[1] == '.' && namelen == 2) {
1168                                         /*
1169                                          * skip ".."
1170                                          */
1171                                         continue;
1172                                 } else if (!sai->sai_ls_all) {
1173                                         /*
1174                                          * skip hidden files.
1175                                          */
1176                                         sai->sai_skip_hidden++;
1177                                         continue;
1178                                 }
1179                         }
1180
1181                         /*
1182                          * don't stat-ahead first entry.
1183                          */
1184                         if (unlikely(++first == 1))
1185                                 continue;
1186
1187 keep_it:
1188                         l_wait_event(thread->t_ctl_waitq,
1189                                      !sa_sent_full(sai) ||
1190                                      !sa_received_empty(sai) ||
1191                                      !agl_list_empty(sai) ||
1192                                      !thread_is_running(thread),
1193                                      &lwi);
1194
1195 interpret_it:
1196                         while (!sa_received_empty(sai))
1197                                 do_statahead_interpret(sai, NULL);
1198
1199                         if (unlikely(!thread_is_running(thread))) {
1200                                 ll_release_page(page, 0);
1201                                 GOTO(out, rc = 0);
1202                         }
1203
1204                         /* If no window for metadata statahead, but there are
1205                          * some AGL entries to be triggered, then try to help
1206                          * to process the AGL entries. */
1207                         if (sa_sent_full(sai)) {
1208                                 cfs_spin_lock(&plli->lli_agl_lock);
1209                                 while (!agl_list_empty(sai)) {
1210                                         clli = agl_first_entry(sai);
1211                                         cfs_list_del_init(&clli->lli_agl_list);
1212                                         cfs_spin_unlock(&plli->lli_agl_lock);
1213                                         ll_agl_trigger(&clli->lli_vfs_inode,
1214                                                        sai);
1215
1216                                         if (!sa_received_empty(sai))
1217                                                 goto interpret_it;
1218
1219                                         if (unlikely(
1220                                                 !thread_is_running(thread))) {
1221                                                 ll_release_page(page, 0);
1222                                                 GOTO(out, rc = 0);
1223                                         }
1224
1225                                         if (!sa_sent_full(sai))
1226                                                 goto do_it;
1227
1228                                         cfs_spin_lock(&plli->lli_agl_lock);
1229                                 }
1230                                 cfs_spin_unlock(&plli->lli_agl_lock);
1231
1232                                 goto keep_it;
1233                         }
1234
1235 do_it:
1236                         ll_statahead_one(parent, name, namelen);
1237                 }
1238                 pos = le64_to_cpu(dp->ldp_hash_end);
1239                 if (pos == MDS_DIR_END_OFF) {
1240                         /*
1241                          * End of directory reached.
1242                          */
1243                         ll_release_page(page, 0);
1244                         while (1) {
1245                                 l_wait_event(thread->t_ctl_waitq,
1246                                              !sa_received_empty(sai) ||
1247                                              sai->sai_sent == sai->sai_replied||
1248                                              !thread_is_running(thread),
1249                                              &lwi);
1250
1251                                 while (!sa_received_empty(sai))
1252                                         do_statahead_interpret(sai, NULL);
1253
1254                                 if (unlikely(!thread_is_running(thread)))
1255                                         GOTO(out, rc = 0);
1256
1257                                 if (sai->sai_sent == sai->sai_replied &&
1258                                     sa_received_empty(sai))
1259                                         break;
1260                         }
1261
1262                         cfs_spin_lock(&plli->lli_agl_lock);
1263                         while (!agl_list_empty(sai) &&
1264                                thread_is_running(thread)) {
1265                                 clli = agl_first_entry(sai);
1266                                 cfs_list_del_init(&clli->lli_agl_list);
1267                                 cfs_spin_unlock(&plli->lli_agl_lock);
1268                                 ll_agl_trigger(&clli->lli_vfs_inode, sai);
1269                                 cfs_spin_lock(&plli->lli_agl_lock);
1270                         }
1271                         cfs_spin_unlock(&plli->lli_agl_lock);
1272
1273                         GOTO(out, rc = 0);
1274                 } else if (1) {
1275                         /*
1276                          * chain is exhausted.
1277                          * Normal case: continue to the next page.
1278                          */
1279                         ll_release_page(page, le32_to_cpu(dp->ldp_flags) &
1280                                               LDF_COLLIDE);
1281                         plli->lli_sa_pos = pos;
1282                         sai->sai_in_readpage = 1;
1283                         page = ll_get_dir_page(NULL, dir, pos, &chain);
1284                         sai->sai_in_readpage = 0;
1285                 } else {
1286                         LASSERT(le32_to_cpu(dp->ldp_flags) & LDF_COLLIDE);
1287                         ll_release_page(page, 1);
1288                         /*
1289                          * go into overflow page.
1290                          */
1291                 }
1292         }
1293         EXIT;
1294
1295 out:
1296         if (sai->sai_agl_valid) {
1297                 struct ptlrpc_thread *agl_thread = &sai->sai_agl_thread;
1298
1299                 cfs_spin_lock(&plli->lli_agl_lock);
1300                 thread_set_flags(agl_thread, SVC_STOPPING);
1301                 cfs_spin_unlock(&plli->lli_agl_lock);
1302                 cfs_waitq_signal(&agl_thread->t_ctl_waitq);
1303
1304                 CDEBUG(D_READA, "stop agl thread: [pid %d]\n",
1305                        cfs_curproc_pid());
1306                 l_wait_event(agl_thread->t_ctl_waitq,
1307                              thread_is_stopped(agl_thread),
1308                              &lwi);
1309         }
1310
1311         ll_dir_chain_fini(&chain);
1312         cfs_spin_lock(&plli->lli_sa_lock);
1313         if (!sa_received_empty(sai)) {
1314                 thread_set_flags(thread, SVC_STOPPING);
1315                 cfs_spin_unlock(&plli->lli_sa_lock);
1316
1317                 /* To release the resources held by received entries. */
1318                 while (!sa_received_empty(sai))
1319                         do_statahead_interpret(sai, NULL);
1320
1321                 cfs_spin_lock(&plli->lli_sa_lock);
1322         }
1323         thread_set_flags(thread, SVC_STOPPED);
1324         cfs_spin_unlock(&plli->lli_sa_lock);
1325         cfs_waitq_signal(&sai->sai_waitq);
1326         cfs_waitq_signal(&thread->t_ctl_waitq);
1327         ll_sai_put(sai);
1328         dput(parent);
1329         CDEBUG(D_READA, "statahead thread stopped: [pid %d] [parent %.*s]\n",
1330                cfs_curproc_pid(), parent->d_name.len, parent->d_name.name);
1331         return rc;
1332 }
1333
1334 /**
1335  * called in ll_file_release().
1336  */
1337 void ll_stop_statahead(struct inode *dir, void *key)
1338 {
1339         struct ll_inode_info *lli = ll_i2info(dir);
1340
1341         if (unlikely(key == NULL))
1342                 return;
1343
1344         cfs_spin_lock(&lli->lli_sa_lock);
1345         if (lli->lli_opendir_key != key || lli->lli_opendir_pid == 0) {
1346                 cfs_spin_unlock(&lli->lli_sa_lock);
1347                 return;
1348         }
1349
1350         lli->lli_opendir_key = NULL;
1351
1352         if (lli->lli_sai) {
1353                 struct l_wait_info lwi = { 0 };
1354                 struct ptlrpc_thread *thread = &lli->lli_sai->sai_thread;
1355
1356                 if (!thread_is_stopped(thread)) {
1357                         thread_set_flags(thread, SVC_STOPPING);
1358                         cfs_spin_unlock(&lli->lli_sa_lock);
1359                         cfs_waitq_signal(&thread->t_ctl_waitq);
1360
1361                         CDEBUG(D_READA, "stop statahead thread: [pid %d]\n",
1362                                cfs_curproc_pid());
1363                         l_wait_event(thread->t_ctl_waitq,
1364                                      thread_is_stopped(thread),
1365                                      &lwi);
1366                 } else {
1367                         cfs_spin_unlock(&lli->lli_sa_lock);
1368                 }
1369
1370                 /*
1371                  * Put the ref which was held when first statahead_enter.
1372                  * It maybe not the last ref for some statahead requests
1373                  * maybe inflight.
1374                  */
1375                 ll_sai_put(lli->lli_sai);
1376         } else {
1377                 lli->lli_opendir_pid = 0;
1378                 cfs_spin_unlock(&lli->lli_sa_lock);
1379         }
1380 }
1381
1382 enum {
1383         /**
1384          * not first dirent, or is "."
1385          */
1386         LS_NONE_FIRST_DE = 0,
1387         /**
1388          * the first non-hidden dirent
1389          */
1390         LS_FIRST_DE,
1391         /**
1392          * the first hidden dirent, that is "."
1393          */
1394         LS_FIRST_DOT_DE
1395 };
1396
1397 static int is_first_dirent(struct inode *dir, struct dentry *dentry)
1398 {
1399         struct ll_inode_info *lli    = ll_i2info(dir);
1400         struct ll_dir_chain   chain;
1401         struct qstr          *target = &dentry->d_name;
1402         struct page          *page;
1403         __u64                 pos    = 0;
1404         int                   dot_de;
1405         int                   rc     = LS_NONE_FIRST_DE;
1406         ENTRY;
1407
1408         lli->lli_sa_pos = 0;
1409         ll_dir_chain_init(&chain);
1410         page = ll_get_dir_page(NULL, dir, pos, &chain);
1411
1412         while (1) {
1413                 struct lu_dirpage *dp;
1414                 struct lu_dirent  *ent;
1415
1416                 if (IS_ERR(page)) {
1417                         struct ll_inode_info *lli = ll_i2info(dir);
1418
1419                         rc = PTR_ERR(page);
1420                         CERROR("error reading dir "DFID" at "LPU64": "
1421                                "[rc %d] [parent %u]\n",
1422                                PFID(ll_inode2fid(dir)), pos,
1423                                rc, lli->lli_opendir_pid);
1424                         break;
1425                 }
1426
1427                 dp = page_address(page);
1428                 for (ent = lu_dirent_start(dp); ent != NULL;
1429                      ent = lu_dirent_next(ent)) {
1430                         __u64 hash;
1431                         int namelen;
1432                         char *name;
1433
1434                         hash = le64_to_cpu(ent->lde_hash);
1435                         /* The ll_get_dir_page() can return any page containing
1436                          * the given hash which may be not the start hash. */
1437                         if (unlikely(hash < pos))
1438                                 continue;
1439
1440                         namelen = le16_to_cpu(ent->lde_namelen);
1441                         if (unlikely(namelen == 0))
1442                                 /*
1443                                  * skip dummy record.
1444                                  */
1445                                 continue;
1446
1447                         name = ent->lde_name;
1448                         if (name[0] == '.') {
1449                                 if (namelen == 1)
1450                                         /*
1451                                          * skip "."
1452                                          */
1453                                         continue;
1454                                 else if (name[1] == '.' && namelen == 2)
1455                                         /*
1456                                          * skip ".."
1457                                          */
1458                                         continue;
1459                                 else
1460                                         dot_de = 1;
1461                         } else {
1462                                 dot_de = 0;
1463                         }
1464
1465                         if (dot_de && target->name[0] != '.') {
1466                                 CDEBUG(D_READA, "%.*s skip hidden file %.*s\n",
1467                                        target->len, target->name,
1468                                        namelen, name);
1469                                 continue;
1470                         }
1471
1472                         if (target->len != namelen ||
1473                             memcmp(target->name, name, namelen) != 0)
1474                                 rc = LS_NONE_FIRST_DE;
1475                         else if (!dot_de)
1476                                 rc = LS_FIRST_DE;
1477                         else
1478                                 rc = LS_FIRST_DOT_DE;
1479
1480                         ll_release_page(page, 0);
1481                         GOTO(out, rc);
1482                 }
1483                 pos = le64_to_cpu(dp->ldp_hash_end);
1484                 if (pos == MDS_DIR_END_OFF) {
1485                         /*
1486                          * End of directory reached.
1487                          */
1488                         ll_release_page(page, 0);
1489                         break;
1490                 } else if (1) {
1491                         /*
1492                          * chain is exhausted
1493                          * Normal case: continue to the next page.
1494                          */
1495                         ll_release_page(page, le32_to_cpu(dp->ldp_flags) &
1496                                               LDF_COLLIDE);
1497                         lli->lli_sa_pos = pos;
1498                         page = ll_get_dir_page(NULL, dir, pos, &chain);
1499                 } else {
1500                         /*
1501                          * go into overflow page.
1502                          */
1503                         LASSERT(le32_to_cpu(dp->ldp_flags) & LDF_COLLIDE);
1504                         ll_release_page(page, 1);
1505                 }
1506         }
1507         EXIT;
1508
1509 out:
1510         ll_dir_chain_fini(&chain);
1511         return rc;
1512 }
1513
1514 static void
1515 ll_sai_unplug(struct ll_statahead_info *sai, struct ll_sa_entry *entry)
1516 {
1517         struct ptlrpc_thread *thread = &sai->sai_thread;
1518         struct ll_sb_info    *sbi    = ll_i2sbi(sai->sai_inode);
1519         int                   hit;
1520         ENTRY;
1521
1522         if (entry != NULL && entry->se_stat == SA_ENTRY_SUCC)
1523                 hit = 1;
1524         else
1525                 hit = 0;
1526
1527         ll_sa_entry_fini(sai, entry);
1528         if (hit) {
1529                 sai->sai_hit++;
1530                 sai->sai_consecutive_miss = 0;
1531                 sai->sai_max = min(2 * sai->sai_max, sbi->ll_sa_max);
1532         } else {
1533                 struct ll_inode_info *lli = ll_i2info(sai->sai_inode);
1534
1535                 sai->sai_miss++;
1536                 sai->sai_consecutive_miss++;
1537                 if (sa_low_hit(sai) && thread_is_running(thread)) {
1538                         atomic_inc(&sbi->ll_sa_wrong);
1539                         CDEBUG(D_READA, "Statahead for dir "DFID" hit "
1540                                "ratio too low: hit/miss "LPU64"/"LPU64
1541                                ", sent/replied "LPU64"/"LPU64", stopping "
1542                                "statahead thread: pid %d\n",
1543                                PFID(&lli->lli_fid), sai->sai_hit,
1544                                sai->sai_miss, sai->sai_sent,
1545                                sai->sai_replied, cfs_curproc_pid());
1546                         cfs_spin_lock(&lli->lli_sa_lock);
1547                         if (!thread_is_stopped(thread))
1548                                 thread_set_flags(thread, SVC_STOPPING);
1549                         cfs_spin_unlock(&lli->lli_sa_lock);
1550                 }
1551         }
1552
1553         if (!thread_is_stopped(thread))
1554                 cfs_waitq_signal(&thread->t_ctl_waitq);
1555
1556         EXIT;
1557 }
1558
1559 /**
1560  * Start statahead thread if this is the first dir entry.
1561  * Otherwise if a thread is started already, wait it until it is ahead of me.
1562  * \retval 1       -- find entry with lock in cache, the caller needs to do
1563  *                    nothing.
1564  * \retval 0       -- find entry in cache, but without lock, the caller needs
1565  *                    refresh from MDS.
1566  * \retval others  -- the caller need to process as non-statahead.
1567  */
1568 int do_statahead_enter(struct inode *dir, struct dentry **dentryp,
1569                        int only_unplug)
1570 {
1571         struct ll_inode_info     *lli   = ll_i2info(dir);
1572         struct ll_statahead_info *sai   = lli->lli_sai;
1573         struct dentry            *parent;
1574         struct ll_sa_entry       *entry;
1575         struct ptlrpc_thread     *thread;
1576         struct l_wait_info        lwi   = { 0 };
1577         int                       rc    = 0;
1578         ENTRY;
1579
1580         LASSERT(lli->lli_opendir_pid == cfs_curproc_pid());
1581
1582         if (sai) {
1583                 thread = &sai->sai_thread;
1584                 if (unlikely(thread_is_stopped(thread) &&
1585                              cfs_list_empty(&sai->sai_entries_stated))) {
1586                         /* to release resource */
1587                         ll_stop_statahead(dir, lli->lli_opendir_key);
1588                         RETURN(-EAGAIN);
1589                 }
1590
1591                 if ((*dentryp)->d_name.name[0] == '.') {
1592                         if (sai->sai_ls_all ||
1593                             sai->sai_miss_hidden >= sai->sai_skip_hidden) {
1594                                 /*
1595                                  * Hidden dentry is the first one, or statahead
1596                                  * thread does not skip so many hidden dentries
1597                                  * before "sai_ls_all" enabled as below.
1598                                  */
1599                         } else {
1600                                 if (!sai->sai_ls_all)
1601                                         /*
1602                                          * It maybe because hidden dentry is not
1603                                          * the first one, "sai_ls_all" was not
1604                                          * set, then "ls -al" missed. Enable
1605                                          * "sai_ls_all" for such case.
1606                                          */
1607                                         sai->sai_ls_all = 1;
1608
1609                                 /*
1610                                  * Such "getattr" has been skipped before
1611                                  * "sai_ls_all" enabled as above.
1612                                  */
1613                                 sai->sai_miss_hidden++;
1614                                 RETURN(-EAGAIN);
1615                         }
1616                 }
1617
1618                 entry = ll_sa_entry_get_byname(sai, &(*dentryp)->d_name);
1619                 if (entry == NULL || only_unplug) {
1620                         ll_sai_unplug(sai, entry);
1621                         RETURN(entry ? 1 : -EAGAIN);
1622                 }
1623
1624                 while (!ll_sa_entry_stated(entry) &&
1625                        sai->sai_in_readpage &&
1626                        !sa_received_empty(sai))
1627                         do_statahead_interpret(sai, entry);
1628
1629                 if (!ll_sa_entry_stated(entry)) {
1630                         sai->sai_index_wait = entry->se_index;
1631                         lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(30), NULL,
1632                                                LWI_ON_SIGNAL_NOOP, NULL);
1633                         rc = l_wait_event(sai->sai_waitq,
1634                                           ll_sa_entry_stated(entry) ||
1635                                           thread_is_stopped(thread),
1636                                           &lwi);
1637                         if (rc < 0) {
1638                                 ll_sai_unplug(sai, entry);
1639                                 RETURN(-EAGAIN);
1640                         }
1641                 }
1642
1643                 if (entry->se_stat == SA_ENTRY_SUCC &&
1644                     entry->se_inode != NULL) {
1645                         struct inode *inode = entry->se_inode;
1646                         struct lookup_intent it = { .it_op = IT_GETATTR,
1647                                                     .d.lustre.it_lock_handle =
1648                                                      entry->se_handle };
1649                         struct ll_dentry_data *lld;
1650                         __u64 bits;
1651
1652                         rc = md_revalidate_lock(ll_i2mdexp(dir), &it,
1653                                                 ll_inode2fid(inode), &bits);
1654                         if (rc == 1) {
1655                                 if ((*dentryp)->d_inode == NULL) {
1656                                         *dentryp = ll_find_alias(inode,
1657                                                                  *dentryp);
1658                                         lld = ll_d2d(*dentryp);
1659                                         if (unlikely(lld == NULL))
1660                                                 ll_dops_init(*dentryp, 1, 1);
1661                                 } else {
1662                                         LASSERT((*dentryp)->d_inode == inode);
1663
1664                                         ll_dentry_rehash(*dentryp, 0);
1665                                         iput(inode);
1666                                 }
1667                                 entry->se_inode = NULL;
1668
1669                                 ll_dentry_reset_flags(*dentryp, bits);
1670                                 ll_intent_release(&it);
1671                         }
1672                 }
1673
1674                 ll_sai_unplug(sai, entry);
1675                 RETURN(rc);
1676         }
1677
1678         /* I am the "lli_opendir_pid" owner, only me can set "lli_sai". */
1679         rc = is_first_dirent(dir, *dentryp);
1680         if (rc == LS_NONE_FIRST_DE)
1681                 /* It is not "ls -{a}l" operation, no need statahead for it. */
1682                 GOTO(out, rc = -EAGAIN);
1683
1684         sai = ll_sai_alloc();
1685         if (sai == NULL)
1686                 GOTO(out, rc = -ENOMEM);
1687
1688         sai->sai_ls_all = (rc == LS_FIRST_DOT_DE);
1689         sai->sai_inode = igrab(dir);
1690         if (unlikely(sai->sai_inode == NULL)) {
1691                 CWARN("Do not start stat ahead on dying inode "DFID"\n",
1692                       PFID(&lli->lli_fid));
1693                 GOTO(out, rc = -ESTALE);
1694         }
1695
1696         /* get parent reference count here, and put it in ll_statahead_thread */
1697         parent = dget((*dentryp)->d_parent);
1698         if (unlikely(sai->sai_inode != parent->d_inode)) {
1699                 struct ll_inode_info *nlli = ll_i2info(parent->d_inode);
1700
1701                 CWARN("Race condition, someone changed %.*s just now: "
1702                       "old parent "DFID", new parent "DFID"\n",
1703                       (*dentryp)->d_name.len, (*dentryp)->d_name.name,
1704                       PFID(&lli->lli_fid), PFID(&nlli->lli_fid));
1705                 dput(parent);
1706                 iput(sai->sai_inode);
1707                 GOTO(out, rc = -EAGAIN);
1708         }
1709
1710         CDEBUG(D_READA, "start statahead thread: [pid %d] [parent %.*s]\n",
1711                cfs_curproc_pid(), parent->d_name.len, parent->d_name.name);
1712
1713         lli->lli_sai = sai;
1714         rc = cfs_create_thread(ll_statahead_thread, parent, 0);
1715         thread = &sai->sai_thread;
1716         if (rc < 0) {
1717                 CERROR("can't start ll_sa thread, rc: %d\n", rc);
1718                 dput(parent);
1719                 lli->lli_opendir_key = NULL;
1720                 thread_set_flags(thread, SVC_STOPPED);
1721                 ll_sai_put(sai);
1722                 LASSERT(lli->lli_sai == NULL);
1723                 RETURN(-EAGAIN);
1724         }
1725
1726         l_wait_event(thread->t_ctl_waitq,
1727                      thread_is_running(thread) || thread_is_stopped(thread),
1728                      &lwi);
1729
1730         /*
1731          * We don't stat-ahead for the first dirent since we are already in
1732          * lookup, and -EEXIST also indicates that this is the first dirent.
1733          */
1734         RETURN(-EEXIST);
1735
1736 out:
1737         if (sai != NULL)
1738                 OBD_FREE_PTR(sai);
1739         cfs_spin_lock(&lli->lli_sa_lock);
1740         lli->lli_opendir_key = NULL;
1741         lli->lli_opendir_pid = 0;
1742         cfs_spin_unlock(&lli->lli_sa_lock);
1743         return rc;
1744 }