lustre/llite/statahead.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2012, Whamcloud, Inc.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  */
  36
  37 #include <linux/fs.h>
  38 #include <linux/sched.h>
  39 #include <linux/mm.h>
  40 #include <linux/smp_lock.h>
  41 #include <linux/highmem.h>
  42 #include <linux/pagemap.h>
  43
  44 #define DEBUG_SUBSYSTEM S_LLITE
  45
  46 #include <obd_support.h>
  47 #include <lustre_lite.h>
  48 #include <lustre_dlm.h>
  49 #include <linux/lustre_version.h>
  50 #include "llite_internal.h"
  51
  52 #define SA_OMITTED_ENTRY_MAX 8ULL
  53
  54 typedef enum {
  55         /** negative values are for error cases */
  56         SA_ENTRY_INIT = 0,      /** init entry */
  57         SA_ENTRY_SUCC = 1,      /** stat succeed */
  58         SA_ENTRY_INVA = 2,      /** invalid entry */
  59         SA_ENTRY_DEST = 3,      /** entry to be destroyed */
  60 } se_stat_t;
  61
  62 struct ll_sa_entry {
  63         /* link into sai->sai_entries_{sent,received,stated} */
  64         cfs_list_t              se_list;
  65         /* link into sai hash table locally */
  66         cfs_list_t              se_hash;
  67         /* entry reference count */
  68         cfs_atomic_t            se_refcount;
  69         /* entry index in the sai */
  70         __u64                   se_index;
  71         /* low layer ldlm lock handle */
  72         __u64                   se_handle;
  73         /* entry status */
  74         se_stat_t               se_stat;
  75         /* entry size, contains name */
  76         int                     se_size;
  77         /* pointer to async getattr enqueue info */
  78         struct md_enqueue_info *se_minfo;
  79         /* pointer to the async getattr request */
  80         struct ptlrpc_request  *se_req;
  81         /* pointer to the target inode */
  82         struct inode           *se_inode;
  83         /* entry name */
  84         struct qstr             se_qstr;
  85 };
  86
  87 static unsigned int sai_generation = 0;
  88 static cfs_spinlock_t sai_generation_lock = CFS_SPIN_LOCK_UNLOCKED;
  89
  90 static inline int ll_sa_entry_unlinked(struct ll_sa_entry *entry)
  91 {
  92         return cfs_list_empty(&entry->se_list);
  93 }
  94
  95 static inline int ll_sa_entry_unhashed(struct ll_sa_entry *entry)
  96 {
  97         return cfs_list_empty(&entry->se_hash);
  98 }
  99
 100 /*
 101  * The entry only can be released by the caller, it is necessary to hold lock.
 102  */
 103 static inline int ll_sa_entry_stated(struct ll_sa_entry *entry)
 104 {
 105         smp_rmb();
 106         return (entry->se_stat != SA_ENTRY_INIT);
 107 }
 108
 109 static inline int ll_sa_entry_hash(int val)
 110 {
 111         return val & LL_SA_CACHE_MASK;
 112 }
 113
 114 /*
 115  * Insert entry to hash SA table.
 116  */
 117 static inline void
 118 ll_sa_entry_enhash(struct ll_statahead_info *sai, struct ll_sa_entry *entry)
 119 {
 120         int i = ll_sa_entry_hash(entry->se_qstr.hash);
 121
 122         cfs_spin_lock(&sai->sai_cache_lock[i]);
 123         cfs_list_add_tail(&entry->se_hash, &sai->sai_cache[i]);
 124         cfs_spin_unlock(&sai->sai_cache_lock[i]);
 125 }
 126
 127 /*
 128  * Remove entry from SA table.
 129  */
 130 static inline void
 131 ll_sa_entry_unhash(struct ll_statahead_info *sai, struct ll_sa_entry *entry)
 132 {
 133         int i = ll_sa_entry_hash(entry->se_qstr.hash);
 134
 135         cfs_spin_lock(&sai->sai_cache_lock[i]);
 136         cfs_list_del_init(&entry->se_hash);
 137         cfs_spin_unlock(&sai->sai_cache_lock[i]);
 138 }
 139
 140 static inline int agl_should_run(struct ll_statahead_info *sai,
 141                                  struct inode *inode)
 142 {
 143         if (inode != NULL && S_ISREG(inode->i_mode) &&
 144             ll_i2info(inode)->lli_smd != NULL && sai->sai_agl_valid)
 145                 return 1;
 146         return 0;
 147 }
 148
 149 static inline struct ll_sa_entry *
 150 sa_first_received_entry(struct ll_statahead_info *sai)
 151 {
 152         return cfs_list_entry(sai->sai_entries_received.next,
 153                               struct ll_sa_entry, se_list);
 154 }
 155
 156 static inline struct ll_inode_info *
 157 agl_first_entry(struct ll_statahead_info *sai)
 158 {
 159         return cfs_list_entry(sai->sai_entries_agl.next,
 160                               struct ll_inode_info, lli_agl_list);
 161 }
 162
 163 static inline int sa_sent_full(struct ll_statahead_info *sai)
 164 {
 165         return cfs_atomic_read(&sai->sai_cache_count) >= sai->sai_max;
 166 }
 167
 168 static inline int sa_received_empty(struct ll_statahead_info *sai)
 169 {
 170         return cfs_list_empty(&sai->sai_entries_received);
 171 }
 172
 173 static inline int agl_list_empty(struct ll_statahead_info *sai)
 174 {
 175         return cfs_list_empty(&sai->sai_entries_agl);
 176 }
 177
 178 /**
 179  * (1) hit ratio less than 80%
 180  * or
 181  * (2) consecutive miss more than 8
 182  * then means low hit.
 183  */
 184 static inline int sa_low_hit(struct ll_statahead_info *sai)
 185 {
 186         return ((sai->sai_hit > 7 && sai->sai_hit < 4 * sai->sai_miss) ||
 187                 (sai->sai_consecutive_miss > 8));
 188 }
 189
 190 /*
 191  * If the given index is behind of statahead window more than
 192  * SA_OMITTED_ENTRY_MAX, then it is old.
 193  */
 194 static inline int is_omitted_entry(struct ll_statahead_info *sai, __u64 index)
 195 {
 196         return ((__u64)sai->sai_max + index + SA_OMITTED_ENTRY_MAX <
 197                  sai->sai_index);
 198 }
 199
 200 /*
 201  * Insert it into sai_entries_sent tail when init.
 202  */
 203 static struct ll_sa_entry *
 204 ll_sa_entry_alloc(struct ll_statahead_info *sai, __u64 index,
 205                   const char *name, int len)
 206 {
 207         struct ll_inode_info *lli;
 208         struct ll_sa_entry   *entry;
 209         int                   entry_size;
 210         char                 *dname;
 211         ENTRY;
 212
 213         entry_size = sizeof(struct ll_sa_entry) + (len & ~3) + 4;
 214         OBD_ALLOC(entry, entry_size);
 215         if (unlikely(entry == NULL))
 216                 RETURN(ERR_PTR(-ENOMEM));
 217
 218         CDEBUG(D_READA, "alloc sai entry %.*s(%p) index "LPU64"\n",
 219                len, name, entry, index);
 220
 221         entry->se_index = index;
 222
 223         /*
 224          * Statahead entry reference rules:
 225          *
 226          * 1) When statahead entry is initialized, its reference is set as 2.
 227          *    One reference is used by the directory scanner. When the scanner
 228          *    searches the statahead cache for the given name, it can perform
 229          *    lockless hash lookup (only the scanner can remove entry from hash
 230          *    list), and once found, it needn't to call "atomic_inc()" for the
 231          *    entry reference. So the performance is improved. After using the
 232          *    statahead entry, the scanner will call "atomic_dec()" to drop the
 233          *    reference held when initialization. If it is the last reference,
 234          *    the statahead entry will be freed.
 235          *
 236          * 2) All other threads, including statahead thread and ptlrpcd thread,
 237          *    when they process the statahead entry, the reference for target
 238          *    should be held to guarantee the entry will not be released by the
 239          *    directory scanner. After processing the entry, these threads will
 240          *    drop the entry reference. If it is the last reference, the entry
 241          *    will be freed.
 242          *
 243          *    The second reference when initializes the statahead entry is used
 244          *    by the statahead thread, following the rule 2).
 245          */
 246         cfs_atomic_set(&entry->se_refcount, 2);
 247         entry->se_stat = SA_ENTRY_INIT;
 248         entry->se_size = entry_size;
 249         dname = (char *)entry + sizeof(struct ll_sa_entry);
 250         memcpy(dname, name, len);
 251         dname[len] = 0;
 252         entry->se_qstr.hash = full_name_hash(name, len);
 253         entry->se_qstr.len = len;
 254         entry->se_qstr.name = dname;
 255
 256         lli = ll_i2info(sai->sai_inode);
 257         cfs_spin_lock(&lli->lli_sa_lock);
 258         cfs_list_add_tail(&entry->se_list, &sai->sai_entries_sent);
 259         cfs_spin_unlock(&lli->lli_sa_lock);
 260
 261         cfs_atomic_inc(&sai->sai_cache_count);
 262         ll_sa_entry_enhash(sai, entry);
 263
 264         RETURN(entry);
 265 }
 266
 267 /*
 268  * Used by the directory scanner to search entry with name.
 269  *
 270  * Only the caller can remove the entry from hash, so it is unnecessary to hold
 271  * hash lock. It is caller's duty to release the init refcount on the entry, so
 272  * it is also unnecessary to increase refcount on the entry.
 273  */
 274 static struct ll_sa_entry *
 275 ll_sa_entry_get_byname(struct ll_statahead_info *sai, const struct qstr *qstr)
 276 {
 277         struct ll_sa_entry *entry;
 278         int i = ll_sa_entry_hash(qstr->hash);
 279
 280         cfs_list_for_each_entry(entry, &sai->sai_cache[i], se_hash) {
 281                 if (entry->se_qstr.hash == qstr->hash &&
 282                     entry->se_qstr.len == qstr->len &&
 283                     memcmp(entry->se_qstr.name, qstr->name, qstr->len) == 0)
 284                         return entry;
 285         }
 286         return NULL;
 287 }
 288
 289 /*
 290  * Used by the async getattr request callback to find entry with index.
 291  *
 292  * Inside lli_sa_lock to prevent others to change the list during the search.
 293  * It needs to increase entry refcount before returning to guarantee that the
 294  * entry cannot be freed by others.
 295  */
 296 static struct ll_sa_entry *
 297 ll_sa_entry_get_byindex(struct ll_statahead_info *sai, __u64 index)
 298 {
 299         struct ll_sa_entry *entry;
 300
 301         cfs_list_for_each_entry(entry, &sai->sai_entries_sent, se_list) {
 302                 if (entry->se_index == index) {
 303                         cfs_atomic_inc(&entry->se_refcount);
 304                         return entry;
 305                 }
 306                 if (entry->se_index > index)
 307                         break;
 308         }
 309         return NULL;
 310 }
 311
 312 static void ll_sa_entry_cleanup(struct ll_statahead_info *sai,
 313                                  struct ll_sa_entry *entry)
 314 {
 315         struct md_enqueue_info *minfo = entry->se_minfo;
 316         struct ptlrpc_request  *req   = entry->se_req;
 317
 318         if (minfo) {
 319                 entry->se_minfo = NULL;
 320                 ll_intent_release(&minfo->mi_it);
 321                 iput(minfo->mi_dir);
 322                 OBD_FREE_PTR(minfo);
 323         }
 324
 325         if (req) {
 326                 entry->se_req = NULL;
 327                 ptlrpc_req_finished(req);
 328         }
 329 }
 330
 331 static void ll_sa_entry_put(struct ll_statahead_info *sai,
 332                              struct ll_sa_entry *entry)
 333 {
 334         if (cfs_atomic_dec_and_test(&entry->se_refcount)) {
 335                 CDEBUG(D_READA, "free sai entry %.*s(%p) index "LPU64"\n",
 336                        entry->se_qstr.len, entry->se_qstr.name, entry,
 337                        entry->se_index);
 338
 339                 LASSERT(ll_sa_entry_unhashed(entry));
 340                 LASSERT(ll_sa_entry_unlinked(entry));
 341
 342                 ll_sa_entry_cleanup(sai, entry);
 343                 if (entry->se_inode)
 344                         iput(entry->se_inode);
 345
 346                 OBD_FREE(entry, entry->se_size);
 347                 cfs_atomic_dec(&sai->sai_cache_count);
 348         }
 349 }
 350
 351 static inline void
 352 do_sai_entry_fini(struct ll_statahead_info *sai, struct ll_sa_entry *entry)
 353 {
 354         struct ll_inode_info *lli = ll_i2info(sai->sai_inode);
 355
 356         ll_sa_entry_unhash(sai, entry);
 357
 358         cfs_spin_lock(&lli->lli_sa_lock);
 359         entry->se_stat = SA_ENTRY_DEST;
 360         if (likely(!ll_sa_entry_unlinked(entry)))
 361                 cfs_list_del_init(&entry->se_list);
 362         cfs_spin_unlock(&lli->lli_sa_lock);
 363
 364         ll_sa_entry_put(sai, entry);
 365 }
 366
 367 /*
 368  * Delete it from sai_entries_stated list when fini.
 369  */
 370 static void
 371 ll_sa_entry_fini(struct ll_statahead_info *sai, struct ll_sa_entry *entry)
 372 {
 373         struct ll_sa_entry *pos, *next;
 374
 375         if (entry)
 376                 do_sai_entry_fini(sai, entry);
 377
 378         /* drop old entry from sent list */
 379         cfs_list_for_each_entry_safe(pos, next, &sai->sai_entries_sent,
 380                                      se_list) {
 381                 if (is_omitted_entry(sai, pos->se_index))
 382                         do_sai_entry_fini(sai, pos);
 383                 else
 384                         break;
 385         }
 386
 387         /* drop old entry from stated list */
 388         cfs_list_for_each_entry_safe(pos, next, &sai->sai_entries_stated,
 389                                      se_list) {
 390                 if (is_omitted_entry(sai, pos->se_index))
 391                         do_sai_entry_fini(sai, pos);
 392                 else
 393                         break;
 394         }
 395 }
 396
 397 /*
 398  * Inside lli_sa_lock.
 399  */
 400 static void
 401 do_sai_entry_to_stated(struct ll_statahead_info *sai,
 402                        struct ll_sa_entry *entry, int rc)
 403 {
 404         struct ll_sa_entry *se;
 405         cfs_list_t         *pos = &sai->sai_entries_stated;
 406
 407         if (!ll_sa_entry_unlinked(entry))
 408                 cfs_list_del_init(&entry->se_list);
 409
 410         cfs_list_for_each_entry_reverse(se, &sai->sai_entries_stated, se_list) {
 411                 if (se->se_index < entry->se_index) {
 412                         pos = &se->se_list;
 413                         break;
 414                 }
 415         }
 416
 417         cfs_list_add(&entry->se_list, pos);
 418         entry->se_stat = rc;
 419 }
 420
 421 /*
 422  * Move entry to sai_entries_stated and sort with the index.
 423  * \retval 1    -- entry to be destroyed.
 424  * \retval 0    -- entry is inserted into stated list.
 425  */
 426 static int
 427 ll_sa_entry_to_stated(struct ll_statahead_info *sai,
 428                        struct ll_sa_entry *entry, int rc)
 429 {
 430         struct ll_inode_info *lli = ll_i2info(sai->sai_inode);
 431         int                   ret = 1;
 432
 433         ll_sa_entry_cleanup(sai, entry);
 434
 435         cfs_spin_lock(&lli->lli_sa_lock);
 436         if (likely(entry->se_stat != SA_ENTRY_DEST)) {
 437                 do_sai_entry_to_stated(sai, entry, rc);
 438                 ret = 0;
 439         }
 440         cfs_spin_unlock(&lli->lli_sa_lock);
 441
 442         return ret;
 443 }
 444
 445 /*
 446  * Insert inode into the list of sai_entries_agl.
 447  */
 448 static void ll_agl_add(struct ll_statahead_info *sai,
 449                        struct inode *inode, int index)
 450 {
 451         struct ll_inode_info *child  = ll_i2info(inode);
 452         struct ll_inode_info *parent = ll_i2info(sai->sai_inode);
 453         int                   added  = 0;
 454
 455         cfs_spin_lock(&child->lli_agl_lock);
 456         if (child->lli_agl_index == 0) {
 457                 child->lli_agl_index = index;
 458                 cfs_spin_unlock(&child->lli_agl_lock);
 459
 460                 LASSERT(cfs_list_empty(&child->lli_agl_list));
 461
 462                 igrab(inode);
 463                 cfs_spin_lock(&parent->lli_agl_lock);
 464                 if (agl_list_empty(sai))
 465                         added = 1;
 466                 cfs_list_add_tail(&child->lli_agl_list, &sai->sai_entries_agl);
 467                 cfs_spin_unlock(&parent->lli_agl_lock);
 468         } else {
 469                 cfs_spin_unlock(&child->lli_agl_lock);
 470         }
 471
 472         if (added > 0)
 473                 cfs_waitq_signal(&sai->sai_agl_thread.t_ctl_waitq);
 474 }
 475
 476 static struct ll_statahead_info *ll_sai_alloc(void)
 477 {
 478         struct ll_statahead_info *sai;
 479         int                       i;
 480         ENTRY;
 481
 482         OBD_ALLOC_PTR(sai);
 483         if (!sai)
 484                 RETURN(NULL);
 485
 486         cfs_atomic_set(&sai->sai_refcount, 1);
 487
 488         cfs_spin_lock(&sai_generation_lock);
 489         sai->sai_generation = ++sai_generation;
 490         if (unlikely(sai_generation == 0))
 491                 sai->sai_generation = ++sai_generation;
 492         cfs_spin_unlock(&sai_generation_lock);
 493
 494         sai->sai_max = LL_SA_RPC_MIN;
 495         sai->sai_index = 1;
 496         cfs_waitq_init(&sai->sai_waitq);
 497         cfs_waitq_init(&sai->sai_thread.t_ctl_waitq);
 498         cfs_waitq_init(&sai->sai_agl_thread.t_ctl_waitq);
 499
 500         CFS_INIT_LIST_HEAD(&sai->sai_entries_sent);
 501         CFS_INIT_LIST_HEAD(&sai->sai_entries_received);
 502         CFS_INIT_LIST_HEAD(&sai->sai_entries_stated);
 503         CFS_INIT_LIST_HEAD(&sai->sai_entries_agl);
 504
 505         for (i = 0; i < LL_SA_CACHE_SIZE; i++) {
 506                 CFS_INIT_LIST_HEAD(&sai->sai_cache[i]);
 507                 cfs_spin_lock_init(&sai->sai_cache_lock[i]);
 508         }
 509         cfs_atomic_set(&sai->sai_cache_count, 0);
 510
 511         RETURN(sai);
 512 }
 513
 514 static inline struct ll_statahead_info *
 515 ll_sai_get(struct ll_statahead_info *sai)
 516 {
 517         cfs_atomic_inc(&sai->sai_refcount);
 518         return sai;
 519 }
 520
 521 static void ll_sai_put(struct ll_statahead_info *sai)
 522 {
 523         struct inode         *inode = sai->sai_inode;
 524         struct ll_inode_info *lli   = ll_i2info(inode);
 525         ENTRY;
 526
 527         if (cfs_atomic_dec_and_lock(&sai->sai_refcount, &lli->lli_sa_lock)) {
 528                 struct ll_sa_entry *entry, *next;
 529
 530                 if (unlikely(cfs_atomic_read(&sai->sai_refcount) > 0)) {
 531                         /* It is race case, the interpret callback just hold
 532                          * a reference count */
 533                         cfs_spin_unlock(&lli->lli_sa_lock);
 534                         RETURN_EXIT;
 535                 }
 536
 537                 LASSERT(lli->lli_opendir_key == NULL);
 538                 LASSERT(thread_is_stopped(&sai->sai_thread));
 539                 LASSERT(thread_is_stopped(&sai->sai_agl_thread));
 540
 541                 lli->lli_sai = NULL;
 542                 lli->lli_opendir_pid = 0;
 543                 cfs_spin_unlock(&lli->lli_sa_lock);
 544
 545                 if (sai->sai_sent > sai->sai_replied)
 546                         CDEBUG(D_READA,"statahead for dir "DFID" does not "
 547                               "finish: [sent:"LPU64"] [replied:"LPU64"]\n",
 548                               PFID(&lli->lli_fid),
 549                               sai->sai_sent, sai->sai_replied);
 550
 551                 cfs_list_for_each_entry_safe(entry, next,
 552                                              &sai->sai_entries_sent, se_list)
 553                         do_sai_entry_fini(sai, entry);
 554
 555                 LASSERT(sa_received_empty(sai));
 556
 557                 cfs_list_for_each_entry_safe(entry, next,
 558                                              &sai->sai_entries_stated, se_list)
 559                         do_sai_entry_fini(sai, entry);
 560
 561                 LASSERT(cfs_atomic_read(&sai->sai_cache_count) == 0);
 562                 LASSERT(agl_list_empty(sai));
 563
 564                 iput(inode);
 565                 OBD_FREE_PTR(sai);
 566         }
 567
 568         EXIT;
 569 }
 570
 571 /* Do NOT forget to drop inode refcount when into sai_entries_agl. */
 572 static void ll_agl_trigger(struct inode *inode, struct ll_statahead_info *sai)
 573 {
 574         struct ll_inode_info *lli   = ll_i2info(inode);
 575         __u64                 index = lli->lli_agl_index;
 576         int                   rc;
 577         ENTRY;
 578
 579         LASSERT(cfs_list_empty(&lli->lli_agl_list));
 580
 581         /* AGL maybe fall behind statahead with one entry */
 582         if (is_omitted_entry(sai, index + 1)) {
 583                 lli->lli_agl_index = 0;
 584                 iput(inode);
 585                 RETURN_EXIT;
 586         }
 587
 588         /* Someone is in glimpse (sync or async), do nothing. */
 589         rc = cfs_down_write_trylock(&lli->lli_glimpse_sem);
 590         if (rc == 0) {
 591                 lli->lli_agl_index = 0;
 592                 iput(inode);
 593                 RETURN_EXIT;
 594         }
 595
 596         /*
 597          * Someone triggered glimpse within 1 sec before.
 598          * 1) The former glimpse succeeded with glimpse lock granted by OST, and
 599          *    if the lock is still cached on client, AGL needs to do nothing. If
 600          *    it is cancelled by other client, AGL maybe cannot obtaion new lock
 601          *    for no glimpse callback triggered by AGL.
 602          * 2) The former glimpse succeeded, but OST did not grant glimpse lock.
 603          *    Under such case, it is quite possible that the OST will not grant
 604          *    glimpse lock for AGL also.
 605          * 3) The former glimpse failed, compared with other two cases, it is
 606          *    relative rare. AGL can ignore such case, and it will not muchly
 607          *    affect the performance.
 608          */
 609         if (lli->lli_glimpse_time != 0 &&
 610             cfs_time_before(cfs_time_shift(-1), lli->lli_glimpse_time)) {
 611                 cfs_up_write(&lli->lli_glimpse_sem);
 612                 lli->lli_agl_index = 0;
 613                 iput(inode);
 614                 RETURN_EXIT;
 615         }
 616
 617         CDEBUG(D_READA, "Handling (init) async glimpse: inode = "
 618                DFID", idx = "LPU64"\n", PFID(&lli->lli_fid), index);
 619
 620         cl_agl(inode);
 621         lli->lli_agl_index = 0;
 622         lli->lli_glimpse_time = cfs_time_current();
 623         cfs_up_write(&lli->lli_glimpse_sem);
 624
 625         CDEBUG(D_READA, "Handled (init) async glimpse: inode= "
 626                DFID", idx = "LPU64", rc = %d\n",
 627                PFID(&lli->lli_fid), index, rc);
 628
 629         iput(inode);
 630
 631         EXIT;
 632 }
 633
 634 static void do_statahead_interpret(struct ll_statahead_info *sai,
 635                                    struct ll_sa_entry *target)
 636 {
 637         struct inode           *dir   = sai->sai_inode;
 638         struct inode           *child;
 639         struct ll_inode_info   *lli   = ll_i2info(dir);
 640         struct ll_sa_entry     *entry;
 641         struct md_enqueue_info *minfo;
 642         struct lookup_intent   *it;
 643         struct ptlrpc_request  *req;
 644         struct mdt_body        *body;
 645         int                     rc    = 0;
 646         ENTRY;
 647
 648         cfs_spin_lock(&lli->lli_sa_lock);
 649         if (target != NULL && target->se_req != NULL &&
 650             !cfs_list_empty(&target->se_list)) {
 651                 entry = target;
 652         } else if (unlikely(sa_received_empty(sai))) {
 653                 cfs_spin_unlock(&lli->lli_sa_lock);
 654                 RETURN_EXIT;
 655         } else {
 656                 entry = sa_first_received_entry(sai);
 657         }
 658
 659         cfs_atomic_inc(&entry->se_refcount);
 660         cfs_list_del_init(&entry->se_list);
 661         cfs_spin_unlock(&lli->lli_sa_lock);
 662
 663         LASSERT(entry->se_handle != 0);
 664
 665         minfo = entry->se_minfo;
 666         it = &minfo->mi_it;
 667         req = entry->se_req;
 668         body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
 669         if (body == NULL)
 670                 GOTO(out, rc = -EFAULT);
 671
 672         child = entry->se_inode;
 673         if (child == NULL) {
 674                 /*
 675                  * lookup.
 676                  */
 677                 LASSERT(fid_is_zero(&minfo->mi_data.op_fid2));
 678
 679                 /* XXX: No fid in reply, this is probaly cross-ref case.
 680                  * SA can't handle it yet. */
 681                 if (body->valid & OBD_MD_MDS)
 682                         GOTO(out, rc = -EAGAIN);
 683         } else {
 684                 /*
 685                  * revalidate.
 686                  */
 687                 /* unlinked and re-created with the same name */
 688                 if (unlikely(!lu_fid_eq(&minfo->mi_data.op_fid2, &body->fid1))){
 689                         entry->se_inode = NULL;
 690                         iput(child);
 691                         child = NULL;
 692                 }
 693         }
 694
 695         it->d.lustre.it_lock_handle = entry->se_handle;
 696         rc = md_revalidate_lock(ll_i2mdexp(dir), it, NULL, NULL);
 697         if (rc != 1)
 698                 GOTO(out, rc = -EAGAIN);
 699
 700         rc = ll_prep_inode(&child, req, dir->i_sb);
 701         if (rc)
 702                 GOTO(out, rc);
 703
 704         CDEBUG(D_DLMTRACE, "setting l_data to inode %p (%lu/%u)\n",
 705                child, child->i_ino, child->i_generation);
 706         ll_set_lock_data(ll_i2sbi(dir)->ll_md_exp, child, it, NULL);
 707
 708         entry->se_inode = child;
 709
 710         if (agl_should_run(sai, child))
 711                 ll_agl_add(sai, child, entry->se_index);
 712
 713         EXIT;
 714
 715 out:
 716         /* The "ll_sa_entry_to_stated()" will drop related ldlm ibits lock
 717          * reference count by calling "ll_intent_drop_lock()" in spite of the
 718          * above operations failed or not. Do not worry about calling
 719          * "ll_intent_drop_lock()" more than once. */
 720         rc = ll_sa_entry_to_stated(sai, entry, rc < 0 ? rc : SA_ENTRY_SUCC);
 721         if (rc == 0 && entry->se_index == sai->sai_index_wait && target == NULL)
 722                 cfs_waitq_signal(&sai->sai_waitq);
 723         ll_sa_entry_put(sai, entry);
 724 }
 725
 726 static int ll_statahead_interpret(struct ptlrpc_request *req,
 727                                   struct md_enqueue_info *minfo, int rc)
 728 {
 729         struct lookup_intent     *it  = &minfo->mi_it;
 730         struct inode             *dir = minfo->mi_dir;
 731         struct ll_inode_info     *lli = ll_i2info(dir);
 732         struct ll_statahead_info *sai = NULL;
 733         struct ll_sa_entry       *entry;
 734         int                       wakeup;
 735         ENTRY;
 736
 737         if (it_disposition(it, DISP_LOOKUP_NEG))
 738                 rc = -ENOENT;
 739
 740         cfs_spin_lock(&lli->lli_sa_lock);
 741         /* stale entry */
 742         if (unlikely(lli->lli_sai == NULL ||
 743                      lli->lli_sai->sai_generation != minfo->mi_generation)) {
 744                 cfs_spin_unlock(&lli->lli_sa_lock);
 745                 GOTO(out, rc = -ESTALE);
 746         } else {
 747                 sai = ll_sai_get(lli->lli_sai);
 748                 if (unlikely(!thread_is_running(&sai->sai_thread))) {
 749                         sai->sai_replied++;
 750                         cfs_spin_unlock(&lli->lli_sa_lock);
 751                         GOTO(out, rc = -EBADFD);
 752                 }
 753
 754                 entry = ll_sa_entry_get_byindex(sai, minfo->mi_cbdata);
 755                 if (entry == NULL) {
 756                         sai->sai_replied++;
 757                         cfs_spin_unlock(&lli->lli_sa_lock);
 758                         GOTO(out, rc = -EIDRM);
 759                 }
 760
 761                 cfs_list_del_init(&entry->se_list);
 762                 if (rc != 0) {
 763                         sai->sai_replied++;
 764                         do_sai_entry_to_stated(sai, entry, rc);
 765                         cfs_spin_unlock(&lli->lli_sa_lock);
 766                         if (entry->se_index == sai->sai_index_wait)
 767                                 cfs_waitq_signal(&sai->sai_waitq);
 768                 } else {
 769                         entry->se_minfo = minfo;
 770                         entry->se_req = ptlrpc_request_addref(req);
 771                         /* Release the async ibits lock ASAP to avoid deadlock
 772                          * when statahead thread tries to enqueue lock on parent
 773                          * for readpage and other tries to enqueue lock on child
 774                          * with parent's lock held, for example: unlink. */
 775                         entry->se_handle = it->d.lustre.it_lock_handle;
 776                         ll_intent_drop_lock(it);
 777                         wakeup = sa_received_empty(sai);
 778                         cfs_list_add_tail(&entry->se_list,
 779                                           &sai->sai_entries_received);
 780                         sai->sai_replied++;
 781                         cfs_spin_unlock(&lli->lli_sa_lock);
 782                         if (wakeup)
 783                                 cfs_waitq_signal(&sai->sai_thread.t_ctl_waitq);
 784                 }
 785                 ll_sa_entry_put(sai, entry);
 786         }
 787
 788         EXIT;
 789
 790 out:
 791         if (rc != 0) {
 792                 ll_intent_release(it);
 793                 iput(dir);
 794                 OBD_FREE_PTR(minfo);
 795         }
 796         if (sai != NULL)
 797                 ll_sai_put(sai);
 798         return rc;
 799 }
 800
 801 static void sa_args_fini(struct md_enqueue_info *minfo,
 802                          struct ldlm_enqueue_info *einfo)
 803 {
 804         LASSERT(minfo && einfo);
 805         iput(minfo->mi_dir);
 806         capa_put(minfo->mi_data.op_capa1);
 807         capa_put(minfo->mi_data.op_capa2);
 808         OBD_FREE_PTR(minfo);
 809         OBD_FREE_PTR(einfo);
 810 }
 811
 812 /**
 813  * There is race condition between "capa_put" and "ll_statahead_interpret" for
 814  * accessing "op_data.op_capa[1,2]" as following:
 815  * "capa_put" releases "op_data.op_capa[1,2]"'s reference count after calling
 816  * "md_intent_getattr_async". But "ll_statahead_interpret" maybe run first, and
 817  * fill "op_data.op_capa[1,2]" as POISON, then cause "capa_put" access invalid
 818  * "ocapa". So here reserve "op_data.op_capa[1,2]" in "pcapa" before calling
 819  * "md_intent_getattr_async".
 820  */
 821 static int sa_args_init(struct inode *dir, struct inode *child,
 822                         struct ll_sa_entry *entry, struct md_enqueue_info **pmi,
 823                         struct ldlm_enqueue_info **pei,
 824                         struct obd_capa **pcapa)
 825 {
 826         struct qstr              *qstr = &entry->se_qstr;
 827         struct ll_inode_info     *lli  = ll_i2info(dir);
 828         struct md_enqueue_info   *minfo;
 829         struct ldlm_enqueue_info *einfo;
 830         struct md_op_data        *op_data;
 831
 832         OBD_ALLOC_PTR(einfo);
 833         if (einfo == NULL)
 834                 return -ENOMEM;
 835
 836         OBD_ALLOC_PTR(minfo);
 837         if (minfo == NULL) {
 838                 OBD_FREE_PTR(einfo);
 839                 return -ENOMEM;
 840         }
 841
 842         op_data = ll_prep_md_op_data(&minfo->mi_data, dir, child, qstr->name,
 843                                      qstr->len, 0, LUSTRE_OPC_ANY, NULL);
 844         if (IS_ERR(op_data)) {
 845                 OBD_FREE_PTR(einfo);
 846                 OBD_FREE_PTR(minfo);
 847                 return PTR_ERR(op_data);
 848         }
 849
 850         minfo->mi_it.it_op = IT_GETATTR;
 851         minfo->mi_dir = igrab(dir);
 852         minfo->mi_cb = ll_statahead_interpret;
 853         minfo->mi_generation = lli->lli_sai->sai_generation;
 854         minfo->mi_cbdata = entry->se_index;
 855
 856         einfo->ei_type   = LDLM_IBITS;
 857         einfo->ei_mode   = it_to_lock_mode(&minfo->mi_it);
 858         einfo->ei_cb_bl  = ll_md_blocking_ast;
 859         einfo->ei_cb_cp  = ldlm_completion_ast;
 860         einfo->ei_cb_gl  = NULL;
 861         einfo->ei_cbdata = NULL;
 862
 863         *pmi = minfo;
 864         *pei = einfo;
 865         pcapa[0] = op_data->op_capa1;
 866         pcapa[1] = op_data->op_capa2;
 867
 868         return 0;
 869 }
 870
 871 static int do_sa_lookup(struct inode *dir, struct ll_sa_entry *entry)
 872 {
 873         struct md_enqueue_info   *minfo;
 874         struct ldlm_enqueue_info *einfo;
 875         struct obd_capa          *capas[2];
 876         int                       rc;
 877         ENTRY;
 878
 879         rc = sa_args_init(dir, NULL, entry, &minfo, &einfo, capas);
 880         if (rc)
 881                 RETURN(rc);
 882
 883         rc = md_intent_getattr_async(ll_i2mdexp(dir), minfo, einfo);
 884         if (!rc) {
 885                 capa_put(capas[0]);
 886                 capa_put(capas[1]);
 887         } else {
 888                 sa_args_fini(minfo, einfo);
 889         }
 890
 891         RETURN(rc);
 892 }
 893
 894 /**
 895  * similar to ll_revalidate_it().
 896  * \retval      1 -- dentry valid
 897  * \retval      0 -- will send stat-ahead request
 898  * \retval others -- prepare stat-ahead request failed
 899  */
 900 static int do_sa_revalidate(struct inode *dir, struct ll_sa_entry *entry,
 901                             struct dentry *dentry)
 902 {
 903         struct inode             *inode = dentry->d_inode;
 904         struct lookup_intent      it = { .it_op = IT_GETATTR,
 905                                          .d.lustre.it_lock_handle = 0 };
 906         struct md_enqueue_info   *minfo;
 907         struct ldlm_enqueue_info *einfo;
 908         struct obd_capa          *capas[2];
 909         int rc;
 910         ENTRY;
 911
 912         if (unlikely(inode == NULL))
 913                 RETURN(1);
 914
 915         if (d_mountpoint(dentry))
 916                 RETURN(1);
 917
 918         if (unlikely(dentry == dentry->d_sb->s_root))
 919                 RETURN(1);
 920
 921         entry->se_inode = igrab(inode);
 922         rc = md_revalidate_lock(ll_i2mdexp(dir), &it, ll_inode2fid(inode),NULL);
 923         if (rc == 1) {
 924                 entry->se_handle = it.d.lustre.it_lock_handle;
 925                 ll_intent_release(&it);
 926                 RETURN(1);
 927         }
 928
 929         rc = sa_args_init(dir, inode, entry, &minfo, &einfo, capas);
 930         if (rc) {
 931                 entry->se_inode = NULL;
 932                 iput(inode);
 933                 RETURN(rc);
 934         }
 935
 936         rc = md_intent_getattr_async(ll_i2mdexp(dir), minfo, einfo);
 937         if (!rc) {
 938                 capa_put(capas[0]);
 939                 capa_put(capas[1]);
 940         } else {
 941                 entry->se_inode = NULL;
 942                 iput(inode);
 943                 sa_args_fini(minfo, einfo);
 944         }
 945
 946         RETURN(rc);
 947 }
 948
 949 static void ll_statahead_one(struct dentry *parent, const char* entry_name,
 950                              int entry_name_len)
 951 {
 952         struct inode             *dir    = parent->d_inode;
 953         struct ll_inode_info     *lli    = ll_i2info(dir);
 954         struct ll_statahead_info *sai    = lli->lli_sai;
 955         struct dentry            *dentry = NULL;
 956         struct ll_sa_entry       *entry;
 957         int                       rc;
 958         int                       rc1;
 959         ENTRY;
 960
 961         entry = ll_sa_entry_alloc(sai, sai->sai_index, entry_name,
 962                                   entry_name_len);
 963         if (IS_ERR(entry))
 964                 RETURN_EXIT;
 965
 966         dentry = d_lookup(parent, &entry->se_qstr);
 967         if (!dentry) {
 968                 rc = do_sa_lookup(dir, entry);
 969         } else {
 970                 rc = do_sa_revalidate(dir, entry, dentry);
 971                 if (rc == 1 && agl_should_run(sai, dentry->d_inode))
 972                         ll_agl_add(sai, dentry->d_inode, entry->se_index);
 973         }
 974
 975         if (dentry != NULL)
 976                 dput(dentry);
 977
 978         if (rc) {
 979                 rc1 = ll_sa_entry_to_stated(sai, entry,
 980                                         rc < 0 ? SA_ENTRY_INVA : SA_ENTRY_SUCC);
 981                 if (rc1 == 0 && entry->se_index == sai->sai_index_wait)
 982                         cfs_waitq_signal(&sai->sai_waitq);
 983         } else {
 984                 sai->sai_sent++;
 985         }
 986
 987         sai->sai_index++;
 988         /* drop one refcount on entry by ll_sa_entry_alloc */
 989         ll_sa_entry_put(sai, entry);
 990
 991         EXIT;
 992 }
 993
 994 static int ll_agl_thread(void *arg)
 995 {
 996         struct dentry            *parent = (struct dentry *)arg;
 997         struct inode             *dir    = parent->d_inode;
 998         struct ll_inode_info     *plli   = ll_i2info(dir);
 999         struct ll_inode_info     *clli;
1000         struct ll_sb_info        *sbi    = ll_i2sbi(dir);
1001         struct ll_statahead_info *sai    = ll_sai_get(plli->lli_sai);
1002         struct ptlrpc_thread     *thread = &sai->sai_agl_thread;
1003         struct l_wait_info        lwi    = { 0 };
1004         ENTRY;
1005
1006         {
1007                 char pname[16];
1008                 snprintf(pname, 15, "ll_agl_%u", plli->lli_opendir_pid);
1009                 cfs_daemonize(pname);
1010         }
1011
1012         CDEBUG(D_READA, "agl thread started: [pid %d] [parent %.*s]\n",
1013                cfs_curproc_pid(), parent->d_name.len, parent->d_name.name);
1014
1015         atomic_inc(&sbi->ll_agl_total);
1016         cfs_spin_lock(&plli->lli_agl_lock);
1017         sai->sai_agl_valid = 1;
1018         thread_set_flags(thread, SVC_RUNNING);
1019         cfs_spin_unlock(&plli->lli_agl_lock);
1020         cfs_waitq_signal(&thread->t_ctl_waitq);
1021
1022         while (1) {
1023                 l_wait_event(thread->t_ctl_waitq,
1024                              !agl_list_empty(sai) ||
1025                              !thread_is_running(thread),
1026                              &lwi);
1027
1028                 if (!thread_is_running(thread))
1029                         break;
1030
1031                 cfs_spin_lock(&plli->lli_agl_lock);
1032                 /* The statahead thread maybe help to process AGL entries,
1033                  * so check whether list empty again. */
1034                 if (!agl_list_empty(sai)) {
1035                         clli = agl_first_entry(sai);
1036                         cfs_list_del_init(&clli->lli_agl_list);
1037                         cfs_spin_unlock(&plli->lli_agl_lock);
1038                         ll_agl_trigger(&clli->lli_vfs_inode, sai);
1039                 } else {
1040                         cfs_spin_unlock(&plli->lli_agl_lock);
1041                 }
1042         }
1043
1044         cfs_spin_lock(&plli->lli_agl_lock);
1045         sai->sai_agl_valid = 0;
1046         while (!agl_list_empty(sai)) {
1047                 clli = agl_first_entry(sai);
1048                 cfs_list_del_init(&clli->lli_agl_list);
1049                 cfs_spin_unlock(&plli->lli_agl_lock);
1050                 clli->lli_agl_index = 0;
1051                 iput(&clli->lli_vfs_inode);
1052                 cfs_spin_lock(&plli->lli_agl_lock);
1053         }
1054         thread_set_flags(thread, SVC_STOPPED);
1055         cfs_spin_unlock(&plli->lli_agl_lock);
1056         cfs_waitq_signal(&thread->t_ctl_waitq);
1057         ll_sai_put(sai);
1058         CDEBUG(D_READA, "agl thread stopped: [pid %d] [parent %.*s]\n",
1059                cfs_curproc_pid(), parent->d_name.len, parent->d_name.name);
1060         RETURN(0);
1061 }
1062
1063 static void ll_start_agl(struct dentry *parent, struct ll_statahead_info *sai)
1064 {
1065         struct ptlrpc_thread *thread = &sai->sai_agl_thread;
1066         struct l_wait_info    lwi    = { 0 };
1067         int                   rc;
1068         ENTRY;
1069
1070         CDEBUG(D_READA, "start agl thread: [pid %d] [parent %.*s]\n",
1071                cfs_curproc_pid(), parent->d_name.len, parent->d_name.name);
1072
1073         rc = cfs_create_thread(ll_agl_thread, parent, 0);
1074         if (rc < 0) {
1075                 CERROR("can't start ll_agl thread, rc: %d\n", rc);
1076                 thread_set_flags(thread, SVC_STOPPED);
1077                 RETURN_EXIT;
1078         }
1079
1080         l_wait_event(thread->t_ctl_waitq,
1081                      thread_is_running(thread) || thread_is_stopped(thread),
1082                      &lwi);
1083         EXIT;
1084 }
1085
1086 static int ll_statahead_thread(void *arg)
1087 {
1088         struct dentry            *parent = (struct dentry *)arg;
1089         struct inode             *dir    = parent->d_inode;
1090         struct ll_inode_info     *plli   = ll_i2info(dir);
1091         struct ll_inode_info     *clli;
1092         struct ll_sb_info        *sbi    = ll_i2sbi(dir);
1093         struct ll_statahead_info *sai    = ll_sai_get(plli->lli_sai);
1094         struct ptlrpc_thread     *thread = &sai->sai_thread;
1095         struct ptlrpc_thread *agl_thread = &sai->sai_agl_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                 cfs_spin_lock(&plli->lli_agl_lock);
1298                 thread_set_flags(agl_thread, SVC_STOPPING);
1299                 cfs_spin_unlock(&plli->lli_agl_lock);
1300                 cfs_waitq_signal(&agl_thread->t_ctl_waitq);
1301
1302                 CDEBUG(D_READA, "stop agl thread: [pid %d]\n",
1303                        cfs_curproc_pid());
1304                 l_wait_event(agl_thread->t_ctl_waitq,
1305                              thread_is_stopped(agl_thread),
1306                              &lwi);
1307         } else {
1308                 /* Set agl_thread flags anyway. */
1309                 thread_set_flags(&sai->sai_agl_thread, SVC_STOPPED);
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 if ((*dentryp)->d_inode != inode) {
1662                                         /* revalidate, but inode is recreated */
1663                                         CDEBUG(D_READA,
1664                                               "stale dentry %.*s inode %lu/%u, "
1665                                               "statahead inode %lu/%u\n",
1666                                               (*dentryp)->d_name.len,
1667                                               (*dentryp)->d_name.name,
1668                                               (*dentryp)->d_inode->i_ino,
1669                                               (*dentryp)->d_inode->i_generation,
1670                                               inode->i_ino,
1671                                               inode->i_generation);
1672                                         ll_sai_unplug(sai, entry);
1673                                         RETURN(-ESTALE);
1674                                 } else {
1675                                         ll_dentry_rehash(*dentryp, 0);
1676                                         iput(inode);
1677                                 }
1678                                 entry->se_inode = NULL;
1679
1680                                 ll_dentry_reset_flags(*dentryp, bits);
1681                                 ll_intent_release(&it);
1682                         }
1683                 }
1684
1685                 ll_sai_unplug(sai, entry);
1686                 RETURN(rc);
1687         }
1688
1689         /* I am the "lli_opendir_pid" owner, only me can set "lli_sai". */
1690         rc = is_first_dirent(dir, *dentryp);
1691         if (rc == LS_NONE_FIRST_DE)
1692                 /* It is not "ls -{a}l" operation, no need statahead for it. */
1693                 GOTO(out, rc = -EAGAIN);
1694
1695         sai = ll_sai_alloc();
1696         if (sai == NULL)
1697                 GOTO(out, rc = -ENOMEM);
1698
1699         sai->sai_ls_all = (rc == LS_FIRST_DOT_DE);
1700         sai->sai_inode = igrab(dir);
1701         if (unlikely(sai->sai_inode == NULL)) {
1702                 CWARN("Do not start stat ahead on dying inode "DFID"\n",
1703                       PFID(&lli->lli_fid));
1704                 GOTO(out, rc = -ESTALE);
1705         }
1706
1707         /* get parent reference count here, and put it in ll_statahead_thread */
1708         parent = dget((*dentryp)->d_parent);
1709         if (unlikely(sai->sai_inode != parent->d_inode)) {
1710                 struct ll_inode_info *nlli = ll_i2info(parent->d_inode);
1711
1712                 CWARN("Race condition, someone changed %.*s just now: "
1713                       "old parent "DFID", new parent "DFID"\n",
1714                       (*dentryp)->d_name.len, (*dentryp)->d_name.name,
1715                       PFID(&lli->lli_fid), PFID(&nlli->lli_fid));
1716                 dput(parent);
1717                 iput(sai->sai_inode);
1718                 GOTO(out, rc = -EAGAIN);
1719         }
1720
1721         CDEBUG(D_READA, "start statahead thread: [pid %d] [parent %.*s]\n",
1722                cfs_curproc_pid(), parent->d_name.len, parent->d_name.name);
1723
1724         lli->lli_sai = sai;
1725         rc = cfs_create_thread(ll_statahead_thread, parent, 0);
1726         thread = &sai->sai_thread;
1727         if (rc < 0) {
1728                 CERROR("can't start ll_sa thread, rc: %d\n", rc);
1729                 dput(parent);
1730                 lli->lli_opendir_key = NULL;
1731                 thread_set_flags(thread, SVC_STOPPED);
1732                 thread_set_flags(&sai->sai_agl_thread, SVC_STOPPED);
1733                 ll_sai_put(sai);
1734                 LASSERT(lli->lli_sai == NULL);
1735                 RETURN(-EAGAIN);
1736         }
1737
1738         l_wait_event(thread->t_ctl_waitq,
1739                      thread_is_running(thread) || thread_is_stopped(thread),
1740                      &lwi);
1741
1742         /*
1743          * We don't stat-ahead for the first dirent since we are already in
1744          * lookup.
1745          */
1746         RETURN(-EAGAIN);
1747
1748 out:
1749         if (sai != NULL)
1750                 OBD_FREE_PTR(sai);
1751         cfs_spin_lock(&lli->lli_sa_lock);
1752         lli->lli_opendir_key = NULL;
1753         lli->lli_opendir_pid = 0;
1754         cfs_spin_unlock(&lli->lli_sa_lock);
1755         return rc;
1756 }