drivers/staging/lustre/lustre/llite/statahead.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2012, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  */
  36
  37 #include <linux/fs.h>
  38 #include <linux/sched.h>
  39 #include <linux/mm.h>
  40 #include <linux/highmem.h>
  41 #include <linux/pagemap.h>
  42
  43 #define DEBUG_SUBSYSTEM S_LLITE
  44
  45 #include "../include/obd_support.h"
  46 #include "../include/lustre_lite.h"
  47 #include "../include/lustre_dlm.h"
  48 #include "llite_internal.h"
  49
  50 #define SA_OMITTED_ENTRY_MAX 8ULL
  51
  52 typedef enum {
  53         /** negative values are for error cases */
  54         SA_ENTRY_INIT = 0,      /** init entry */
  55         SA_ENTRY_SUCC = 1,      /** stat succeed */
  56         SA_ENTRY_INVA = 2,      /** invalid entry */
  57         SA_ENTRY_DEST = 3,      /** entry to be destroyed */
  58 } se_stat_t;
  59
  60 struct ll_sa_entry {
  61         /* link into sai->sai_entries */
  62         struct list_head              se_link;
  63         /* link into sai->sai_entries_{received,stated} */
  64         struct list_head              se_list;
  65         /* link into sai hash table locally */
  66         struct list_head              se_hash;
  67         /* entry reference count */
  68         atomic_t            se_refcount;
  69         /* entry index in the sai */
  70         __u64              se_index;
  71         /* low layer ldlm lock handle */
  72         __u64              se_handle;
  73         /* entry status */
  74         se_stat_t              se_stat;
  75         /* entry size, contains name */
  76         int                  se_size;
  77         /* pointer to async getattr enqueue info */
  78         struct md_enqueue_info *se_minfo;
  79         /* pointer to the async getattr request */
  80         struct ptlrpc_request  *se_req;
  81         /* pointer to the target inode */
  82         struct inode       *se_inode;
  83         /* entry name */
  84         struct qstr          se_qstr;
  85 };
  86
  87 static unsigned int sai_generation;
  88 static DEFINE_SPINLOCK(sai_generation_lock);
  89
  90 static inline int ll_sa_entry_unhashed(struct ll_sa_entry *entry)
  91 {
  92         return list_empty(&entry->se_hash);
  93 }
  94
  95 /*
  96  * The entry only can be released by the caller, it is necessary to hold lock.
  97  */
  98 static inline int ll_sa_entry_stated(struct ll_sa_entry *entry)
  99 {
 100         smp_rmb();
 101         return (entry->se_stat != SA_ENTRY_INIT);
 102 }
 103
 104 static inline int ll_sa_entry_hash(int val)
 105 {
 106         return val & LL_SA_CACHE_MASK;
 107 }
 108
 109 /*
 110  * Insert entry to hash SA table.
 111  */
 112 static inline void
 113 ll_sa_entry_enhash(struct ll_statahead_info *sai, struct ll_sa_entry *entry)
 114 {
 115         int i = ll_sa_entry_hash(entry->se_qstr.hash);
 116
 117         spin_lock(&sai->sai_cache_lock[i]);
 118         list_add_tail(&entry->se_hash, &sai->sai_cache[i]);
 119         spin_unlock(&sai->sai_cache_lock[i]);
 120 }
 121
 122 /*
 123  * Remove entry from SA table.
 124  */
 125 static inline void
 126 ll_sa_entry_unhash(struct ll_statahead_info *sai, struct ll_sa_entry *entry)
 127 {
 128         int i = ll_sa_entry_hash(entry->se_qstr.hash);
 129
 130         spin_lock(&sai->sai_cache_lock[i]);
 131         list_del_init(&entry->se_hash);
 132         spin_unlock(&sai->sai_cache_lock[i]);
 133 }
 134
 135 static inline int agl_should_run(struct ll_statahead_info *sai,
 136                                  struct inode *inode)
 137 {
 138         return (inode != NULL && S_ISREG(inode->i_mode) && sai->sai_agl_valid);
 139 }
 140
 141 static inline struct ll_sa_entry *
 142 sa_first_received_entry(struct ll_statahead_info *sai)
 143 {
 144         return list_entry(sai->sai_entries_received.next,
 145                               struct ll_sa_entry, se_list);
 146 }
 147
 148 static inline struct ll_inode_info *
 149 agl_first_entry(struct ll_statahead_info *sai)
 150 {
 151         return list_entry(sai->sai_entries_agl.next,
 152                               struct ll_inode_info, lli_agl_list);
 153 }
 154
 155 static inline int sa_sent_full(struct ll_statahead_info *sai)
 156 {
 157         return atomic_read(&sai->sai_cache_count) >= sai->sai_max;
 158 }
 159
 160 static inline int sa_received_empty(struct ll_statahead_info *sai)
 161 {
 162         return list_empty(&sai->sai_entries_received);
 163 }
 164
 165 static inline int agl_list_empty(struct ll_statahead_info *sai)
 166 {
 167         return list_empty(&sai->sai_entries_agl);
 168 }
 169
 170 /**
 171  * (1) hit ratio less than 80%
 172  * or
 173  * (2) consecutive miss more than 8
 174  * then means low hit.
 175  */
 176 static inline int sa_low_hit(struct ll_statahead_info *sai)
 177 {
 178         return ((sai->sai_hit > 7 && sai->sai_hit < 4 * sai->sai_miss) ||
 179                 (sai->sai_consecutive_miss > 8));
 180 }
 181
 182 /*
 183  * If the given index is behind of statahead window more than
 184  * SA_OMITTED_ENTRY_MAX, then it is old.
 185  */
 186 static inline int is_omitted_entry(struct ll_statahead_info *sai, __u64 index)
 187 {
 188         return ((__u64)sai->sai_max + index + SA_OMITTED_ENTRY_MAX <
 189                  sai->sai_index);
 190 }
 191
 192 /*
 193  * Insert it into sai_entries tail when init.
 194  */
 195 static struct ll_sa_entry *
 196 ll_sa_entry_alloc(struct ll_statahead_info *sai, __u64 index,
 197                   const char *name, int len)
 198 {
 199         struct ll_inode_info *lli;
 200         struct ll_sa_entry   *entry;
 201         int                entry_size;
 202         char             *dname;
 203
 204         entry_size = sizeof(struct ll_sa_entry) + (len & ~3) + 4;
 205         entry = kzalloc(entry_size, GFP_NOFS);
 206         if (unlikely(!entry))
 207                 return ERR_PTR(-ENOMEM);
 208
 209         CDEBUG(D_READA, "alloc sa entry %.*s(%p) index %llu\n",
 210                len, name, entry, index);
 211
 212         entry->se_index = index;
 213
 214         /*
 215          * Statahead entry reference rules:
 216          *
 217          * 1) When statahead entry is initialized, its reference is set as 2.
 218          *    One reference is used by the directory scanner. When the scanner
 219          *    searches the statahead cache for the given name, it can perform
 220          *    lockless hash lookup (only the scanner can remove entry from hash
 221          *    list), and once found, it needn't to call "atomic_inc()" for the
 222          *    entry reference. So the performance is improved. After using the
 223          *    statahead entry, the scanner will call "atomic_dec()" to drop the
 224          *    reference held when initialization. If it is the last reference,
 225          *    the statahead entry will be freed.
 226          *
 227          * 2) All other threads, including statahead thread and ptlrpcd thread,
 228          *    when they process the statahead entry, the reference for target
 229          *    should be held to guarantee the entry will not be released by the
 230          *    directory scanner. After processing the entry, these threads will
 231          *    drop the entry reference. If it is the last reference, the entry
 232          *    will be freed.
 233          *
 234          *    The second reference when initializes the statahead entry is used
 235          *    by the statahead thread, following the rule 2).
 236          */
 237         atomic_set(&entry->se_refcount, 2);
 238         entry->se_stat = SA_ENTRY_INIT;
 239         entry->se_size = entry_size;
 240         dname = (char *)entry + sizeof(struct ll_sa_entry);
 241         memcpy(dname, name, len);
 242         dname[len] = 0;
 243         entry->se_qstr.hash = full_name_hash(name, len);
 244         entry->se_qstr.len = len;
 245         entry->se_qstr.name = dname;
 246
 247         lli = ll_i2info(sai->sai_inode);
 248         spin_lock(&lli->lli_sa_lock);
 249         list_add_tail(&entry->se_link, &sai->sai_entries);
 250         INIT_LIST_HEAD(&entry->se_list);
 251         ll_sa_entry_enhash(sai, entry);
 252         spin_unlock(&lli->lli_sa_lock);
 253
 254         atomic_inc(&sai->sai_cache_count);
 255
 256         return entry;
 257 }
 258
 259 /*
 260  * Used by the directory scanner to search entry with name.
 261  *
 262  * Only the caller can remove the entry from hash, so it is unnecessary to hold
 263  * hash lock. It is caller's duty to release the init refcount on the entry, so
 264  * it is also unnecessary to increase refcount on the entry.
 265  */
 266 static struct ll_sa_entry *
 267 ll_sa_entry_get_byname(struct ll_statahead_info *sai, const struct qstr *qstr)
 268 {
 269         struct ll_sa_entry *entry;
 270         int i = ll_sa_entry_hash(qstr->hash);
 271
 272         list_for_each_entry(entry, &sai->sai_cache[i], se_hash) {
 273                 if (entry->se_qstr.hash == qstr->hash &&
 274                     entry->se_qstr.len == qstr->len &&
 275                     memcmp(entry->se_qstr.name, qstr->name, qstr->len) == 0)
 276                         return entry;
 277         }
 278         return NULL;
 279 }
 280
 281 /*
 282  * Used by the async getattr request callback to find entry with index.
 283  *
 284  * Inside lli_sa_lock to prevent others to change the list during the search.
 285  * It needs to increase entry refcount before returning to guarantee that the
 286  * entry cannot be freed by others.
 287  */
 288 static struct ll_sa_entry *
 289 ll_sa_entry_get_byindex(struct ll_statahead_info *sai, __u64 index)
 290 {
 291         struct ll_sa_entry *entry;
 292
 293         list_for_each_entry(entry, &sai->sai_entries, se_link) {
 294                 if (entry->se_index == index) {
 295                         LASSERT(atomic_read(&entry->se_refcount) > 0);
 296                         atomic_inc(&entry->se_refcount);
 297                         return entry;
 298                 }
 299                 if (entry->se_index > index)
 300                         break;
 301         }
 302         return NULL;
 303 }
 304
 305 static void ll_sa_entry_cleanup(struct ll_statahead_info *sai,
 306                                  struct ll_sa_entry *entry)
 307 {
 308         struct md_enqueue_info *minfo = entry->se_minfo;
 309         struct ptlrpc_request  *req   = entry->se_req;
 310
 311         if (minfo) {
 312                 entry->se_minfo = NULL;
 313                 ll_intent_release(&minfo->mi_it);
 314                 iput(minfo->mi_dir);
 315                 OBD_FREE_PTR(minfo);
 316         }
 317
 318         if (req) {
 319                 entry->se_req = NULL;
 320                 ptlrpc_req_finished(req);
 321         }
 322 }
 323
 324 static void ll_sa_entry_put(struct ll_statahead_info *sai,
 325                              struct ll_sa_entry *entry)
 326 {
 327         if (atomic_dec_and_test(&entry->se_refcount)) {
 328                 CDEBUG(D_READA, "free sa entry %.*s(%p) index %llu\n",
 329                        entry->se_qstr.len, entry->se_qstr.name, entry,
 330                        entry->se_index);
 331
 332                 LASSERT(list_empty(&entry->se_link));
 333                 LASSERT(list_empty(&entry->se_list));
 334                 LASSERT(ll_sa_entry_unhashed(entry));
 335
 336                 ll_sa_entry_cleanup(sai, entry);
 337                 iput(entry->se_inode);
 338
 339                 OBD_FREE(entry, entry->se_size);
 340                 atomic_dec(&sai->sai_cache_count);
 341         }
 342 }
 343
 344 static inline void
 345 do_sa_entry_fini(struct ll_statahead_info *sai, struct ll_sa_entry *entry)
 346 {
 347         struct ll_inode_info *lli = ll_i2info(sai->sai_inode);
 348
 349         LASSERT(!ll_sa_entry_unhashed(entry));
 350         LASSERT(!list_empty(&entry->se_link));
 351
 352         ll_sa_entry_unhash(sai, entry);
 353
 354         spin_lock(&lli->lli_sa_lock);
 355         entry->se_stat = SA_ENTRY_DEST;
 356         list_del_init(&entry->se_link);
 357         if (likely(!list_empty(&entry->se_list)))
 358                 list_del_init(&entry->se_list);
 359         spin_unlock(&lli->lli_sa_lock);
 360
 361         ll_sa_entry_put(sai, entry);
 362 }
 363
 364 /*
 365  * Delete it from sai_entries_stated list when fini.
 366  */
 367 static void
 368 ll_sa_entry_fini(struct ll_statahead_info *sai, struct ll_sa_entry *entry)
 369 {
 370         struct ll_sa_entry *pos, *next;
 371
 372         if (entry)
 373                 do_sa_entry_fini(sai, entry);
 374
 375         /* drop old entry, only 'scanner' process does this, no need to lock */
 376         list_for_each_entry_safe(pos, next, &sai->sai_entries, se_link) {
 377                 if (!is_omitted_entry(sai, pos->se_index))
 378                         break;
 379                 do_sa_entry_fini(sai, pos);
 380         }
 381 }
 382
 383 /*
 384  * Inside lli_sa_lock.
 385  */
 386 static void
 387 do_sa_entry_to_stated(struct ll_statahead_info *sai,
 388                       struct ll_sa_entry *entry, se_stat_t stat)
 389 {
 390         struct ll_sa_entry *se;
 391         struct list_head         *pos = &sai->sai_entries_stated;
 392
 393         if (!list_empty(&entry->se_list))
 394                 list_del_init(&entry->se_list);
 395
 396         list_for_each_entry_reverse(se, &sai->sai_entries_stated, se_list) {
 397                 if (se->se_index < entry->se_index) {
 398                         pos = &se->se_list;
 399                         break;
 400                 }
 401         }
 402
 403         list_add(&entry->se_list, pos);
 404         entry->se_stat = stat;
 405 }
 406
 407 /*
 408  * Move entry to sai_entries_stated and sort with the index.
 409  * \retval 1    -- entry to be destroyed.
 410  * \retval 0    -- entry is inserted into stated list.
 411  */
 412 static int
 413 ll_sa_entry_to_stated(struct ll_statahead_info *sai,
 414                       struct ll_sa_entry *entry, se_stat_t stat)
 415 {
 416         struct ll_inode_info *lli = ll_i2info(sai->sai_inode);
 417         int                ret = 1;
 418
 419         ll_sa_entry_cleanup(sai, entry);
 420
 421         spin_lock(&lli->lli_sa_lock);
 422         if (likely(entry->se_stat != SA_ENTRY_DEST)) {
 423                 do_sa_entry_to_stated(sai, entry, stat);
 424                 ret = 0;
 425         }
 426         spin_unlock(&lli->lli_sa_lock);
 427
 428         return ret;
 429 }
 430
 431 /*
 432  * Insert inode into the list of sai_entries_agl.
 433  */
 434 static void ll_agl_add(struct ll_statahead_info *sai,
 435                        struct inode *inode, int index)
 436 {
 437         struct ll_inode_info *child  = ll_i2info(inode);
 438         struct ll_inode_info *parent = ll_i2info(sai->sai_inode);
 439         int                added  = 0;
 440
 441         spin_lock(&child->lli_agl_lock);
 442         if (child->lli_agl_index == 0) {
 443                 child->lli_agl_index = index;
 444                 spin_unlock(&child->lli_agl_lock);
 445
 446                 LASSERT(list_empty(&child->lli_agl_list));
 447
 448                 igrab(inode);
 449                 spin_lock(&parent->lli_agl_lock);
 450                 if (agl_list_empty(sai))
 451                         added = 1;
 452                 list_add_tail(&child->lli_agl_list, &sai->sai_entries_agl);
 453                 spin_unlock(&parent->lli_agl_lock);
 454         } else {
 455                 spin_unlock(&child->lli_agl_lock);
 456         }
 457
 458         if (added > 0)
 459                 wake_up(&sai->sai_agl_thread.t_ctl_waitq);
 460 }
 461
 462 static struct ll_statahead_info *ll_sai_alloc(void)
 463 {
 464         struct ll_statahead_info *sai;
 465         int                    i;
 466
 467         sai = kzalloc(sizeof(*sai), GFP_NOFS);
 468         if (!sai)
 469                 return NULL;
 470
 471         atomic_set(&sai->sai_refcount, 1);
 472
 473         spin_lock(&sai_generation_lock);
 474         sai->sai_generation = ++sai_generation;
 475         if (unlikely(sai_generation == 0))
 476                 sai->sai_generation = ++sai_generation;
 477         spin_unlock(&sai_generation_lock);
 478
 479         sai->sai_max = LL_SA_RPC_MIN;
 480         sai->sai_index = 1;
 481         init_waitqueue_head(&sai->sai_waitq);
 482         init_waitqueue_head(&sai->sai_thread.t_ctl_waitq);
 483         init_waitqueue_head(&sai->sai_agl_thread.t_ctl_waitq);
 484
 485         INIT_LIST_HEAD(&sai->sai_entries);
 486         INIT_LIST_HEAD(&sai->sai_entries_received);
 487         INIT_LIST_HEAD(&sai->sai_entries_stated);
 488         INIT_LIST_HEAD(&sai->sai_entries_agl);
 489
 490         for (i = 0; i < LL_SA_CACHE_SIZE; i++) {
 491                 INIT_LIST_HEAD(&sai->sai_cache[i]);
 492                 spin_lock_init(&sai->sai_cache_lock[i]);
 493         }
 494         atomic_set(&sai->sai_cache_count, 0);
 495
 496         return sai;
 497 }
 498
 499 static inline struct ll_statahead_info *
 500 ll_sai_get(struct ll_statahead_info *sai)
 501 {
 502         atomic_inc(&sai->sai_refcount);
 503         return sai;
 504 }
 505
 506 static void ll_sai_put(struct ll_statahead_info *sai)
 507 {
 508         struct inode     *inode = sai->sai_inode;
 509         struct ll_inode_info *lli   = ll_i2info(inode);
 510
 511         if (atomic_dec_and_lock(&sai->sai_refcount, &lli->lli_sa_lock)) {
 512                 struct ll_sa_entry *entry, *next;
 513
 514                 if (unlikely(atomic_read(&sai->sai_refcount) > 0)) {
 515                         /* It is race case, the interpret callback just hold
 516                          * a reference count */
 517                         spin_unlock(&lli->lli_sa_lock);
 518                         return;
 519                 }
 520
 521                 LASSERT(lli->lli_opendir_key == NULL);
 522                 LASSERT(thread_is_stopped(&sai->sai_thread));
 523                 LASSERT(thread_is_stopped(&sai->sai_agl_thread));
 524
 525                 lli->lli_sai = NULL;
 526                 lli->lli_opendir_pid = 0;
 527                 spin_unlock(&lli->lli_sa_lock);
 528
 529                 if (sai->sai_sent > sai->sai_replied)
 530                         CDEBUG(D_READA, "statahead for dir "DFID
 531                               " does not finish: [sent:%llu] [replied:%llu]\n",
 532                               PFID(&lli->lli_fid),
 533                               sai->sai_sent, sai->sai_replied);
 534
 535                 list_for_each_entry_safe(entry, next,
 536                                              &sai->sai_entries, se_link)
 537                         do_sa_entry_fini(sai, entry);
 538
 539                 LASSERT(list_empty(&sai->sai_entries));
 540                 LASSERT(sa_received_empty(sai));
 541                 LASSERT(list_empty(&sai->sai_entries_stated));
 542
 543                 LASSERT(atomic_read(&sai->sai_cache_count) == 0);
 544                 LASSERT(agl_list_empty(sai));
 545
 546                 iput(inode);
 547                 OBD_FREE_PTR(sai);
 548         }
 549 }
 550
 551 /* Do NOT forget to drop inode refcount when into sai_entries_agl. */
 552 static void ll_agl_trigger(struct inode *inode, struct ll_statahead_info *sai)
 553 {
 554         struct ll_inode_info *lli   = ll_i2info(inode);
 555         __u64            index = lli->lli_agl_index;
 556         int                rc;
 557
 558         LASSERT(list_empty(&lli->lli_agl_list));
 559
 560         /* AGL maybe fall behind statahead with one entry */
 561         if (is_omitted_entry(sai, index + 1)) {
 562                 lli->lli_agl_index = 0;
 563                 iput(inode);
 564                 return;
 565         }
 566
 567         /* Someone is in glimpse (sync or async), do nothing. */
 568         rc = down_write_trylock(&lli->lli_glimpse_sem);
 569         if (rc == 0) {
 570                 lli->lli_agl_index = 0;
 571                 iput(inode);
 572                 return;
 573         }
 574
 575         /*
 576          * Someone triggered glimpse within 1 sec before.
 577          * 1) The former glimpse succeeded with glimpse lock granted by OST, and
 578          *    if the lock is still cached on client, AGL needs to do nothing. If
 579          *    it is cancelled by other client, AGL maybe cannot obtain new lock
 580          *    for no glimpse callback triggered by AGL.
 581          * 2) The former glimpse succeeded, but OST did not grant glimpse lock.
 582          *    Under such case, it is quite possible that the OST will not grant
 583          *    glimpse lock for AGL also.
 584          * 3) The former glimpse failed, compared with other two cases, it is
 585          *    relative rare. AGL can ignore such case, and it will not muchly
 586          *    affect the performance.
 587          */
 588         if (lli->lli_glimpse_time != 0 &&
 589             time_before(cfs_time_shift(-1), lli->lli_glimpse_time)) {
 590                 up_write(&lli->lli_glimpse_sem);
 591                 lli->lli_agl_index = 0;
 592                 iput(inode);
 593                 return;
 594         }
 595
 596         CDEBUG(D_READA, "Handling (init) async glimpse: inode = "
 597                DFID", idx = %llu\n", PFID(&lli->lli_fid), index);
 598
 599         cl_agl(inode);
 600         lli->lli_agl_index = 0;
 601         lli->lli_glimpse_time = cfs_time_current();
 602         up_write(&lli->lli_glimpse_sem);
 603
 604         CDEBUG(D_READA, "Handled (init) async glimpse: inode= "
 605                DFID", idx = %llu, rc = %d\n",
 606                PFID(&lli->lli_fid), index, rc);
 607
 608         iput(inode);
 609 }
 610
 611 static void ll_post_statahead(struct ll_statahead_info *sai)
 612 {
 613         struct inode       *dir   = sai->sai_inode;
 614         struct inode       *child;
 615         struct ll_inode_info   *lli   = ll_i2info(dir);
 616         struct ll_sa_entry     *entry;
 617         struct md_enqueue_info *minfo;
 618         struct lookup_intent   *it;
 619         struct ptlrpc_request  *req;
 620         struct mdt_body *body;
 621         int                  rc    = 0;
 622
 623         spin_lock(&lli->lli_sa_lock);
 624         if (unlikely(sa_received_empty(sai))) {
 625                 spin_unlock(&lli->lli_sa_lock);
 626                 return;
 627         }
 628         entry = sa_first_received_entry(sai);
 629         atomic_inc(&entry->se_refcount);
 630         list_del_init(&entry->se_list);
 631         spin_unlock(&lli->lli_sa_lock);
 632
 633         LASSERT(entry->se_handle != 0);
 634
 635         minfo = entry->se_minfo;
 636         it = &minfo->mi_it;
 637         req = entry->se_req;
 638         body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
 639         if (body == NULL) {
 640                 rc = -EFAULT;
 641                 goto out;
 642         }
 643
 644         child = entry->se_inode;
 645         if (child == NULL) {
 646                 /*
 647                  * lookup.
 648                  */
 649                 LASSERT(fid_is_zero(&minfo->mi_data.op_fid2));
 650
 651                 /* XXX: No fid in reply, this is probably cross-ref case.
 652                  * SA can't handle it yet. */
 653                 if (body->valid & OBD_MD_MDS) {
 654                         rc = -EAGAIN;
 655                         goto out;
 656                 }
 657         } else {
 658                 /*
 659                  * revalidate.
 660                  */
 661                 /* unlinked and re-created with the same name */
 662                 if (unlikely(!lu_fid_eq(&minfo->mi_data.op_fid2, &body->fid1))){
 663                         entry->se_inode = NULL;
 664                         iput(child);
 665                         child = NULL;
 666                 }
 667         }
 668
 669         it->d.lustre.it_lock_handle = entry->se_handle;
 670         rc = md_revalidate_lock(ll_i2mdexp(dir), it, ll_inode2fid(dir), NULL);
 671         if (rc != 1) {
 672                 rc = -EAGAIN;
 673                 goto out;
 674         }
 675
 676         rc = ll_prep_inode(&child, req, dir->i_sb, it);
 677         if (rc)
 678                 goto out;
 679
 680         CDEBUG(D_DLMTRACE, "setting l_data to inode %p (%lu/%u)\n",
 681                child, child->i_ino, child->i_generation);
 682         ll_set_lock_data(ll_i2sbi(dir)->ll_md_exp, child, it, NULL);
 683
 684         entry->se_inode = child;
 685
 686         if (agl_should_run(sai, child))
 687                 ll_agl_add(sai, child, entry->se_index);
 688
 689 out:
 690         /* The "ll_sa_entry_to_stated()" will drop related ldlm ibits lock
 691          * reference count by calling "ll_intent_drop_lock()" in spite of the
 692          * above operations failed or not. Do not worry about calling
 693          * "ll_intent_drop_lock()" more than once. */
 694         rc = ll_sa_entry_to_stated(sai, entry,
 695                                    rc < 0 ? SA_ENTRY_INVA : SA_ENTRY_SUCC);
 696         if (rc == 0 && entry->se_index == sai->sai_index_wait)
 697                 wake_up(&sai->sai_waitq);
 698         ll_sa_entry_put(sai, entry);
 699 }
 700
 701 static int ll_statahead_interpret(struct ptlrpc_request *req,
 702                                   struct md_enqueue_info *minfo, int rc)
 703 {
 704         struct lookup_intent     *it  = &minfo->mi_it;
 705         struct inode         *dir = minfo->mi_dir;
 706         struct ll_inode_info     *lli = ll_i2info(dir);
 707         struct ll_statahead_info *sai = NULL;
 708         struct ll_sa_entry       *entry;
 709         int                    wakeup;
 710
 711         if (it_disposition(it, DISP_LOOKUP_NEG))
 712                 rc = -ENOENT;
 713
 714         spin_lock(&lli->lli_sa_lock);
 715         /* stale entry */
 716         if (unlikely(lli->lli_sai == NULL ||
 717                      lli->lli_sai->sai_generation != minfo->mi_generation)) {
 718                 spin_unlock(&lli->lli_sa_lock);
 719                 rc = -ESTALE;
 720                 goto out;
 721         } else {
 722                 sai = ll_sai_get(lli->lli_sai);
 723                 if (unlikely(!thread_is_running(&sai->sai_thread))) {
 724                         sai->sai_replied++;
 725                         spin_unlock(&lli->lli_sa_lock);
 726                         rc = -EBADFD;
 727                         goto out;
 728                 }
 729
 730                 entry = ll_sa_entry_get_byindex(sai, minfo->mi_cbdata);
 731                 if (entry == NULL) {
 732                         sai->sai_replied++;
 733                         spin_unlock(&lli->lli_sa_lock);
 734                         rc = -EIDRM;
 735                         goto out;
 736                 }
 737
 738                 if (rc != 0) {
 739                         do_sa_entry_to_stated(sai, entry, SA_ENTRY_INVA);
 740                         wakeup = (entry->se_index == sai->sai_index_wait);
 741                 } else {
 742                         entry->se_minfo = minfo;
 743                         entry->se_req = ptlrpc_request_addref(req);
 744                         /* Release the async ibits lock ASAP to avoid deadlock
 745                          * when statahead thread tries to enqueue lock on parent
 746                          * for readpage and other tries to enqueue lock on child
 747                          * with parent's lock held, for example: unlink. */
 748                         entry->se_handle = it->d.lustre.it_lock_handle;
 749                         ll_intent_drop_lock(it);
 750                         wakeup = sa_received_empty(sai);
 751                         list_add_tail(&entry->se_list,
 752                                           &sai->sai_entries_received);
 753                 }
 754                 sai->sai_replied++;
 755                 spin_unlock(&lli->lli_sa_lock);
 756
 757                 ll_sa_entry_put(sai, entry);
 758                 if (wakeup)
 759                         wake_up(&sai->sai_thread.t_ctl_waitq);
 760         }
 761
 762 out:
 763         if (rc != 0) {
 764                 ll_intent_release(it);
 765                 iput(dir);
 766                 OBD_FREE_PTR(minfo);
 767         }
 768         if (sai != NULL)
 769                 ll_sai_put(sai);
 770         return rc;
 771 }
 772
 773 static void sa_args_fini(struct md_enqueue_info *minfo,
 774                          struct ldlm_enqueue_info *einfo)
 775 {
 776         LASSERT(minfo && einfo);
 777         iput(minfo->mi_dir);
 778         capa_put(minfo->mi_data.op_capa1);
 779         capa_put(minfo->mi_data.op_capa2);
 780         OBD_FREE_PTR(minfo);
 781         OBD_FREE_PTR(einfo);
 782 }
 783
 784 /**
 785  * There is race condition between "capa_put" and "ll_statahead_interpret" for
 786  * accessing "op_data.op_capa[1,2]" as following:
 787  * "capa_put" releases "op_data.op_capa[1,2]"'s reference count after calling
 788  * "md_intent_getattr_async". But "ll_statahead_interpret" maybe run first, and
 789  * fill "op_data.op_capa[1,2]" as POISON, then cause "capa_put" access invalid
 790  * "ocapa". So here reserve "op_data.op_capa[1,2]" in "pcapa" before calling
 791  * "md_intent_getattr_async".
 792  */
 793 static int sa_args_init(struct inode *dir, struct inode *child,
 794                         struct ll_sa_entry *entry, struct md_enqueue_info **pmi,
 795                         struct ldlm_enqueue_info **pei,
 796                         struct obd_capa **pcapa)
 797 {
 798         struct qstr           *qstr = &entry->se_qstr;
 799         struct ll_inode_info     *lli  = ll_i2info(dir);
 800         struct md_enqueue_info   *minfo;
 801         struct ldlm_enqueue_info *einfo;
 802         struct md_op_data       *op_data;
 803
 804         einfo = kzalloc(sizeof(*einfo), GFP_NOFS);
 805         if (!einfo)
 806                 return -ENOMEM;
 807
 808         minfo = kzalloc(sizeof(*minfo), GFP_NOFS);
 809         if (!minfo) {
 810                 OBD_FREE_PTR(einfo);
 811                 return -ENOMEM;
 812         }
 813
 814         op_data = ll_prep_md_op_data(&minfo->mi_data, dir, child, qstr->name,
 815                                      qstr->len, 0, LUSTRE_OPC_ANY, NULL);
 816         if (IS_ERR(op_data)) {
 817                 OBD_FREE_PTR(einfo);
 818                 OBD_FREE_PTR(minfo);
 819                 return PTR_ERR(op_data);
 820         }
 821
 822         minfo->mi_it.it_op = IT_GETATTR;
 823         minfo->mi_dir = igrab(dir);
 824         minfo->mi_cb = ll_statahead_interpret;
 825         minfo->mi_generation = lli->lli_sai->sai_generation;
 826         minfo->mi_cbdata = entry->se_index;
 827
 828         einfo->ei_type   = LDLM_IBITS;
 829         einfo->ei_mode   = it_to_lock_mode(&minfo->mi_it);
 830         einfo->ei_cb_bl  = ll_md_blocking_ast;
 831         einfo->ei_cb_cp  = ldlm_completion_ast;
 832         einfo->ei_cb_gl  = NULL;
 833         einfo->ei_cbdata = NULL;
 834
 835         *pmi = minfo;
 836         *pei = einfo;
 837         pcapa[0] = op_data->op_capa1;
 838         pcapa[1] = op_data->op_capa2;
 839
 840         return 0;
 841 }
 842
 843 static int do_sa_lookup(struct inode *dir, struct ll_sa_entry *entry)
 844 {
 845         struct md_enqueue_info   *minfo;
 846         struct ldlm_enqueue_info *einfo;
 847         struct obd_capa   *capas[2];
 848         int                    rc;
 849
 850         rc = sa_args_init(dir, NULL, entry, &minfo, &einfo, capas);
 851         if (rc)
 852                 return rc;
 853
 854         rc = md_intent_getattr_async(ll_i2mdexp(dir), minfo, einfo);
 855         if (!rc) {
 856                 capa_put(capas[0]);
 857                 capa_put(capas[1]);
 858         } else {
 859                 sa_args_fini(minfo, einfo);
 860         }
 861
 862         return rc;
 863 }
 864
 865 /**
 866  * similar to ll_revalidate_it().
 867  * \retval      1 -- dentry valid
 868  * \retval      0 -- will send stat-ahead request
 869  * \retval others -- prepare stat-ahead request failed
 870  */
 871 static int do_sa_revalidate(struct inode *dir, struct ll_sa_entry *entry,
 872                             struct dentry *dentry)
 873 {
 874         struct inode         *inode = dentry->d_inode;
 875         struct lookup_intent      it = { .it_op = IT_GETATTR,
 876                                          .d.lustre.it_lock_handle = 0 };
 877         struct md_enqueue_info   *minfo;
 878         struct ldlm_enqueue_info *einfo;
 879         struct obd_capa   *capas[2];
 880         int rc;
 881
 882         if (unlikely(inode == NULL))
 883                 return 1;
 884
 885         if (d_mountpoint(dentry))
 886                 return 1;
 887
 888         entry->se_inode = igrab(inode);
 889         rc = md_revalidate_lock(ll_i2mdexp(dir), &it, ll_inode2fid(inode),
 890                                 NULL);
 891         if (rc == 1) {
 892                 entry->se_handle = it.d.lustre.it_lock_handle;
 893                 ll_intent_release(&it);
 894                 return 1;
 895         }
 896
 897         rc = sa_args_init(dir, inode, entry, &minfo, &einfo, capas);
 898         if (rc) {
 899                 entry->se_inode = NULL;
 900                 iput(inode);
 901                 return rc;
 902         }
 903
 904         rc = md_intent_getattr_async(ll_i2mdexp(dir), minfo, einfo);
 905         if (!rc) {
 906                 capa_put(capas[0]);
 907                 capa_put(capas[1]);
 908         } else {
 909                 entry->se_inode = NULL;
 910                 iput(inode);
 911                 sa_args_fini(minfo, einfo);
 912         }
 913
 914         return rc;
 915 }
 916
 917 static void ll_statahead_one(struct dentry *parent, const char *entry_name,
 918                              int entry_name_len)
 919 {
 920         struct inode         *dir    = parent->d_inode;
 921         struct ll_inode_info     *lli    = ll_i2info(dir);
 922         struct ll_statahead_info *sai    = lli->lli_sai;
 923         struct dentry       *dentry = NULL;
 924         struct ll_sa_entry       *entry;
 925         int                    rc;
 926         int                    rc1;
 927
 928         entry = ll_sa_entry_alloc(sai, sai->sai_index, entry_name,
 929                                   entry_name_len);
 930         if (IS_ERR(entry))
 931                 return;
 932
 933         dentry = d_lookup(parent, &entry->se_qstr);
 934         if (!dentry) {
 935                 rc = do_sa_lookup(dir, entry);
 936         } else {
 937                 rc = do_sa_revalidate(dir, entry, dentry);
 938                 if (rc == 1 && agl_should_run(sai, dentry->d_inode))
 939                         ll_agl_add(sai, dentry->d_inode, entry->se_index);
 940         }
 941
 942         if (dentry != NULL)
 943                 dput(dentry);
 944
 945         if (rc) {
 946                 rc1 = ll_sa_entry_to_stated(sai, entry,
 947                                         rc < 0 ? SA_ENTRY_INVA : SA_ENTRY_SUCC);
 948                 if (rc1 == 0 && entry->se_index == sai->sai_index_wait)
 949                         wake_up(&sai->sai_waitq);
 950         } else {
 951                 sai->sai_sent++;
 952         }
 953
 954         sai->sai_index++;
 955         /* drop one refcount on entry by ll_sa_entry_alloc */
 956         ll_sa_entry_put(sai, entry);
 957 }
 958
 959 static int ll_agl_thread(void *arg)
 960 {
 961         struct dentry       *parent = (struct dentry *)arg;
 962         struct inode         *dir    = parent->d_inode;
 963         struct ll_inode_info     *plli   = ll_i2info(dir);
 964         struct ll_inode_info     *clli;
 965         struct ll_sb_info       *sbi    = ll_i2sbi(dir);
 966         struct ll_statahead_info *sai    = ll_sai_get(plli->lli_sai);
 967         struct ptlrpc_thread     *thread = &sai->sai_agl_thread;
 968         struct l_wait_info      lwi    = { 0 };
 969
 970         thread->t_pid = current_pid();
 971         CDEBUG(D_READA, "agl thread started: sai %p, parent %pd\n",
 972                sai, parent);
 973
 974         atomic_inc(&sbi->ll_agl_total);
 975         spin_lock(&plli->lli_agl_lock);
 976         sai->sai_agl_valid = 1;
 977         if (thread_is_init(thread))
 978                 /* If someone else has changed the thread state
 979                  * (e.g. already changed to SVC_STOPPING), we can't just
 980                  * blindly overwrite that setting. */
 981                 thread_set_flags(thread, SVC_RUNNING);
 982         spin_unlock(&plli->lli_agl_lock);
 983         wake_up(&thread->t_ctl_waitq);
 984
 985         while (1) {
 986                 l_wait_event(thread->t_ctl_waitq,
 987                              !agl_list_empty(sai) ||
 988                              !thread_is_running(thread),
 989                              &lwi);
 990
 991                 if (!thread_is_running(thread))
 992                         break;
 993
 994                 spin_lock(&plli->lli_agl_lock);
 995                 /* The statahead thread maybe help to process AGL entries,
 996                  * so check whether list empty again. */
 997                 if (!agl_list_empty(sai)) {
 998                         clli = agl_first_entry(sai);
 999                         list_del_init(&clli->lli_agl_list);
1000                         spin_unlock(&plli->lli_agl_lock);
1001                         ll_agl_trigger(&clli->lli_vfs_inode, sai);
1002                 } else {
1003                         spin_unlock(&plli->lli_agl_lock);
1004                 }
1005         }
1006
1007         spin_lock(&plli->lli_agl_lock);
1008         sai->sai_agl_valid = 0;
1009         while (!agl_list_empty(sai)) {
1010                 clli = agl_first_entry(sai);
1011                 list_del_init(&clli->lli_agl_list);
1012                 spin_unlock(&plli->lli_agl_lock);
1013                 clli->lli_agl_index = 0;
1014                 iput(&clli->lli_vfs_inode);
1015                 spin_lock(&plli->lli_agl_lock);
1016         }
1017         thread_set_flags(thread, SVC_STOPPED);
1018         spin_unlock(&plli->lli_agl_lock);
1019         wake_up(&thread->t_ctl_waitq);
1020         ll_sai_put(sai);
1021         CDEBUG(D_READA, "agl thread stopped: sai %p, parent %pd\n",
1022                sai, parent);
1023         return 0;
1024 }
1025
1026 static void ll_start_agl(struct dentry *parent, struct ll_statahead_info *sai)
1027 {
1028         struct ptlrpc_thread *thread = &sai->sai_agl_thread;
1029         struct l_wait_info    lwi    = { 0 };
1030         struct ll_inode_info  *plli;
1031         struct task_struct *task;
1032
1033         CDEBUG(D_READA, "start agl thread: sai %p, parent %pd\n",
1034                sai, parent);
1035
1036         plli = ll_i2info(parent->d_inode);
1037         task = kthread_run(ll_agl_thread, parent,
1038                                "ll_agl_%u", plli->lli_opendir_pid);
1039         if (IS_ERR(task)) {
1040                 CERROR("can't start ll_agl thread, rc: %ld\n", PTR_ERR(task));
1041                 thread_set_flags(thread, SVC_STOPPED);
1042                 return;
1043         }
1044
1045         l_wait_event(thread->t_ctl_waitq,
1046                      thread_is_running(thread) || thread_is_stopped(thread),
1047                      &lwi);
1048 }
1049
1050 static int ll_statahead_thread(void *arg)
1051 {
1052         struct dentry       *parent = (struct dentry *)arg;
1053         struct inode         *dir    = parent->d_inode;
1054         struct ll_inode_info     *plli   = ll_i2info(dir);
1055         struct ll_inode_info     *clli;
1056         struct ll_sb_info       *sbi    = ll_i2sbi(dir);
1057         struct ll_statahead_info *sai    = ll_sai_get(plli->lli_sai);
1058         struct ptlrpc_thread     *thread = &sai->sai_thread;
1059         struct ptlrpc_thread *agl_thread = &sai->sai_agl_thread;
1060         struct page           *page;
1061         __u64                pos    = 0;
1062         int                    first  = 0;
1063         int                    rc     = 0;
1064         struct ll_dir_chain       chain;
1065         struct l_wait_info      lwi    = { 0 };
1066
1067         thread->t_pid = current_pid();
1068         CDEBUG(D_READA, "statahead thread starting: sai %p, parent %pd\n",
1069                sai, parent);
1070
1071         if (sbi->ll_flags & LL_SBI_AGL_ENABLED)
1072                 ll_start_agl(parent, sai);
1073
1074         atomic_inc(&sbi->ll_sa_total);
1075         spin_lock(&plli->lli_sa_lock);
1076         if (thread_is_init(thread))
1077                 /* If someone else has changed the thread state
1078                  * (e.g. already changed to SVC_STOPPING), we can't just
1079                  * blindly overwrite that setting. */
1080                 thread_set_flags(thread, SVC_RUNNING);
1081         spin_unlock(&plli->lli_sa_lock);
1082         wake_up(&thread->t_ctl_waitq);
1083
1084         ll_dir_chain_init(&chain);
1085         page = ll_get_dir_page(dir, pos, &chain);
1086
1087         while (1) {
1088                 struct lu_dirpage *dp;
1089                 struct lu_dirent  *ent;
1090
1091                 if (IS_ERR(page)) {
1092                         rc = PTR_ERR(page);
1093                         CDEBUG(D_READA, "error reading dir "DFID" at %llu/%llu: [rc %d] [parent %u]\n",
1094                                PFID(ll_inode2fid(dir)), pos, sai->sai_index,
1095                                rc, plli->lli_opendir_pid);
1096                         goto out;
1097                 }
1098
1099                 dp = page_address(page);
1100                 for (ent = lu_dirent_start(dp); ent != NULL;
1101                      ent = lu_dirent_next(ent)) {
1102                         __u64 hash;
1103                         int namelen;
1104                         char *name;
1105
1106                         hash = le64_to_cpu(ent->lde_hash);
1107                         if (unlikely(hash < pos))
1108                                 /*
1109                                  * Skip until we find target hash value.
1110                                  */
1111                                 continue;
1112
1113                         namelen = le16_to_cpu(ent->lde_namelen);
1114                         if (unlikely(namelen == 0))
1115                                 /*
1116                                  * Skip dummy record.
1117                                  */
1118                                 continue;
1119
1120                         name = ent->lde_name;
1121                         if (name[0] == '.') {
1122                                 if (namelen == 1) {
1123                                         /*
1124                                          * skip "."
1125                                          */
1126                                         continue;
1127                                 } else if (name[1] == '.' && namelen == 2) {
1128                                         /*
1129                                          * skip ".."
1130                                          */
1131                                         continue;
1132                                 } else if (!sai->sai_ls_all) {
1133                                         /*
1134                                          * skip hidden files.
1135                                          */
1136                                         sai->sai_skip_hidden++;
1137                                         continue;
1138                                 }
1139                         }
1140
1141                         /*
1142                          * don't stat-ahead first entry.
1143                          */
1144                         if (unlikely(++first == 1))
1145                                 continue;
1146
1147 keep_it:
1148                         l_wait_event(thread->t_ctl_waitq,
1149                                      !sa_sent_full(sai) ||
1150                                      !sa_received_empty(sai) ||
1151                                      !agl_list_empty(sai) ||
1152                                      !thread_is_running(thread),
1153                                      &lwi);
1154
1155 interpret_it:
1156                         while (!sa_received_empty(sai))
1157                                 ll_post_statahead(sai);
1158
1159                         if (unlikely(!thread_is_running(thread))) {
1160                                 ll_release_page(page, 0);
1161                                 rc = 0;
1162                                 goto out;
1163                         }
1164
1165                         /* If no window for metadata statahead, but there are
1166                          * some AGL entries to be triggered, then try to help
1167                          * to process the AGL entries. */
1168                         if (sa_sent_full(sai)) {
1169                                 spin_lock(&plli->lli_agl_lock);
1170                                 while (!agl_list_empty(sai)) {
1171                                         clli = agl_first_entry(sai);
1172                                         list_del_init(&clli->lli_agl_list);
1173                                         spin_unlock(&plli->lli_agl_lock);
1174                                         ll_agl_trigger(&clli->lli_vfs_inode,
1175                                                        sai);
1176
1177                                         if (!sa_received_empty(sai))
1178                                                 goto interpret_it;
1179
1180                                         if (unlikely(
1181                                                 !thread_is_running(thread))) {
1182                                                 ll_release_page(page, 0);
1183                                                 rc = 0;
1184                                                 goto out;
1185                                         }
1186
1187                                         if (!sa_sent_full(sai))
1188                                                 goto do_it;
1189
1190                                         spin_lock(&plli->lli_agl_lock);
1191                                 }
1192                                 spin_unlock(&plli->lli_agl_lock);
1193
1194                                 goto keep_it;
1195                         }
1196
1197 do_it:
1198                         ll_statahead_one(parent, name, namelen);
1199                 }
1200                 pos = le64_to_cpu(dp->ldp_hash_end);
1201                 if (pos == MDS_DIR_END_OFF) {
1202                         /*
1203                          * End of directory reached.
1204                          */
1205                         ll_release_page(page, 0);
1206                         while (1) {
1207                                 l_wait_event(thread->t_ctl_waitq,
1208                                              !sa_received_empty(sai) ||
1209                                              sai->sai_sent == sai->sai_replied||
1210                                              !thread_is_running(thread),
1211                                              &lwi);
1212
1213                                 while (!sa_received_empty(sai))
1214                                         ll_post_statahead(sai);
1215
1216                                 if (unlikely(!thread_is_running(thread))) {
1217                                         rc = 0;
1218                                         goto out;
1219                                 }
1220
1221                                 if (sai->sai_sent == sai->sai_replied &&
1222                                     sa_received_empty(sai))
1223                                         break;
1224                         }
1225
1226                         spin_lock(&plli->lli_agl_lock);
1227                         while (!agl_list_empty(sai) &&
1228                                thread_is_running(thread)) {
1229                                 clli = agl_first_entry(sai);
1230                                 list_del_init(&clli->lli_agl_list);
1231                                 spin_unlock(&plli->lli_agl_lock);
1232                                 ll_agl_trigger(&clli->lli_vfs_inode, sai);
1233                                 spin_lock(&plli->lli_agl_lock);
1234                         }
1235                         spin_unlock(&plli->lli_agl_lock);
1236
1237                         rc = 0;
1238                         goto out;
1239                 } else if (1) {
1240                         /*
1241                          * chain is exhausted.
1242                          * Normal case: continue to the next page.
1243                          */
1244                         ll_release_page(page, le32_to_cpu(dp->ldp_flags) &
1245                                               LDF_COLLIDE);
1246                         page = ll_get_dir_page(dir, pos, &chain);
1247                 } else {
1248                         LASSERT(le32_to_cpu(dp->ldp_flags) & LDF_COLLIDE);
1249                         ll_release_page(page, 1);
1250                         /*
1251                          * go into overflow page.
1252                          */
1253                 }
1254         }
1255
1256 out:
1257         if (sai->sai_agl_valid) {
1258                 spin_lock(&plli->lli_agl_lock);
1259                 thread_set_flags(agl_thread, SVC_STOPPING);
1260                 spin_unlock(&plli->lli_agl_lock);
1261                 wake_up(&agl_thread->t_ctl_waitq);
1262
1263                 CDEBUG(D_READA, "stop agl thread: sai %p pid %u\n",
1264                        sai, (unsigned int)agl_thread->t_pid);
1265                 l_wait_event(agl_thread->t_ctl_waitq,
1266                              thread_is_stopped(agl_thread),
1267                              &lwi);
1268         } else {
1269                 /* Set agl_thread flags anyway. */
1270                 thread_set_flags(&sai->sai_agl_thread, SVC_STOPPED);
1271         }
1272         ll_dir_chain_fini(&chain);
1273         spin_lock(&plli->lli_sa_lock);
1274         if (!sa_received_empty(sai)) {
1275                 thread_set_flags(thread, SVC_STOPPING);
1276                 spin_unlock(&plli->lli_sa_lock);
1277
1278                 /* To release the resources held by received entries. */
1279                 while (!sa_received_empty(sai))
1280                         ll_post_statahead(sai);
1281
1282                 spin_lock(&plli->lli_sa_lock);
1283         }
1284         thread_set_flags(thread, SVC_STOPPED);
1285         spin_unlock(&plli->lli_sa_lock);
1286         wake_up(&sai->sai_waitq);
1287         wake_up(&thread->t_ctl_waitq);
1288         ll_sai_put(sai);
1289         dput(parent);
1290         CDEBUG(D_READA, "statahead thread stopped: sai %p, parent %pd\n",
1291                sai, parent);
1292         return rc;
1293 }
1294
1295 /**
1296  * called in ll_file_release().
1297  */
1298 void ll_stop_statahead(struct inode *dir, void *key)
1299 {
1300         struct ll_inode_info *lli = ll_i2info(dir);
1301
1302         if (unlikely(key == NULL))
1303                 return;
1304
1305         spin_lock(&lli->lli_sa_lock);
1306         if (lli->lli_opendir_key != key || lli->lli_opendir_pid == 0) {
1307                 spin_unlock(&lli->lli_sa_lock);
1308                 return;
1309         }
1310
1311         lli->lli_opendir_key = NULL;
1312
1313         if (lli->lli_sai) {
1314                 struct l_wait_info lwi = { 0 };
1315                 struct ptlrpc_thread *thread = &lli->lli_sai->sai_thread;
1316
1317                 if (!thread_is_stopped(thread)) {
1318                         thread_set_flags(thread, SVC_STOPPING);
1319                         spin_unlock(&lli->lli_sa_lock);
1320                         wake_up(&thread->t_ctl_waitq);
1321
1322                         CDEBUG(D_READA, "stop statahead thread: sai %p pid %u\n",
1323                                lli->lli_sai, (unsigned int)thread->t_pid);
1324                         l_wait_event(thread->t_ctl_waitq,
1325                                      thread_is_stopped(thread),
1326                                      &lwi);
1327                 } else {
1328                         spin_unlock(&lli->lli_sa_lock);
1329                 }
1330
1331                 /*
1332                  * Put the ref which was held when first statahead_enter.
1333                  * It maybe not the last ref for some statahead requests
1334                  * maybe inflight.
1335                  */
1336                 ll_sai_put(lli->lli_sai);
1337         } else {
1338                 lli->lli_opendir_pid = 0;
1339                 spin_unlock(&lli->lli_sa_lock);
1340         }
1341 }
1342
1343 enum {
1344         /**
1345          * not first dirent, or is "."
1346          */
1347         LS_NONE_FIRST_DE = 0,
1348         /**
1349          * the first non-hidden dirent
1350          */
1351         LS_FIRST_DE,
1352         /**
1353          * the first hidden dirent, that is "."
1354          */
1355         LS_FIRST_DOT_DE
1356 };
1357
1358 static int is_first_dirent(struct inode *dir, struct dentry *dentry)
1359 {
1360         struct ll_dir_chain   chain;
1361         struct qstr       *target = &dentry->d_name;
1362         struct page       *page;
1363         __u64            pos    = 0;
1364         int                dot_de;
1365         int                rc     = LS_NONE_FIRST_DE;
1366
1367         ll_dir_chain_init(&chain);
1368         page = ll_get_dir_page(dir, pos, &chain);
1369
1370         while (1) {
1371                 struct lu_dirpage *dp;
1372                 struct lu_dirent  *ent;
1373
1374                 if (IS_ERR(page)) {
1375                         struct ll_inode_info *lli = ll_i2info(dir);
1376
1377                         rc = PTR_ERR(page);
1378                         CERROR("error reading dir "DFID" at %llu: [rc %d] [parent %u]\n",
1379                                PFID(ll_inode2fid(dir)), pos,
1380                                rc, lli->lli_opendir_pid);
1381                         break;
1382                 }
1383
1384                 dp = page_address(page);
1385                 for (ent = lu_dirent_start(dp); ent != NULL;
1386                      ent = lu_dirent_next(ent)) {
1387                         __u64 hash;
1388                         int namelen;
1389                         char *name;
1390
1391                         hash = le64_to_cpu(ent->lde_hash);
1392                         /* The ll_get_dir_page() can return any page containing
1393                          * the given hash which may be not the start hash. */
1394                         if (unlikely(hash < pos))
1395                                 continue;
1396
1397                         namelen = le16_to_cpu(ent->lde_namelen);
1398                         if (unlikely(namelen == 0))
1399                                 /*
1400                                  * skip dummy record.
1401                                  */
1402                                 continue;
1403
1404                         name = ent->lde_name;
1405                         if (name[0] == '.') {
1406                                 if (namelen == 1)
1407                                         /*
1408                                          * skip "."
1409                                          */
1410                                         continue;
1411                                 else if (name[1] == '.' && namelen == 2)
1412                                         /*
1413                                          * skip ".."
1414                                          */
1415                                         continue;
1416                                 else
1417                                         dot_de = 1;
1418                         } else {
1419                                 dot_de = 0;
1420                         }
1421
1422                         if (dot_de && target->name[0] != '.') {
1423                                 CDEBUG(D_READA, "%.*s skip hidden file %.*s\n",
1424                                        target->len, target->name,
1425                                        namelen, name);
1426                                 continue;
1427                         }
1428
1429                         if (target->len != namelen ||
1430                             memcmp(target->name, name, namelen) != 0)
1431                                 rc = LS_NONE_FIRST_DE;
1432                         else if (!dot_de)
1433                                 rc = LS_FIRST_DE;
1434                         else
1435                                 rc = LS_FIRST_DOT_DE;
1436
1437                         ll_release_page(page, 0);
1438                         goto out;
1439                 }
1440                 pos = le64_to_cpu(dp->ldp_hash_end);
1441                 if (pos == MDS_DIR_END_OFF) {
1442                         /*
1443                          * End of directory reached.
1444                          */
1445                         ll_release_page(page, 0);
1446                         break;
1447                 } else if (1) {
1448                         /*
1449                          * chain is exhausted
1450                          * Normal case: continue to the next page.
1451                          */
1452                         ll_release_page(page, le32_to_cpu(dp->ldp_flags) &
1453                                               LDF_COLLIDE);
1454                         page = ll_get_dir_page(dir, pos, &chain);
1455                 } else {
1456                         /*
1457                          * go into overflow page.
1458                          */
1459                         LASSERT(le32_to_cpu(dp->ldp_flags) & LDF_COLLIDE);
1460                         ll_release_page(page, 1);
1461                 }
1462         }
1463
1464 out:
1465         ll_dir_chain_fini(&chain);
1466         return rc;
1467 }
1468
1469 static void
1470 ll_sai_unplug(struct ll_statahead_info *sai, struct ll_sa_entry *entry)
1471 {
1472         struct ptlrpc_thread *thread = &sai->sai_thread;
1473         struct ll_sb_info    *sbi    = ll_i2sbi(sai->sai_inode);
1474         int                hit;
1475
1476         if (entry != NULL && entry->se_stat == SA_ENTRY_SUCC)
1477                 hit = 1;
1478         else
1479                 hit = 0;
1480
1481         ll_sa_entry_fini(sai, entry);
1482         if (hit) {
1483                 sai->sai_hit++;
1484                 sai->sai_consecutive_miss = 0;
1485                 sai->sai_max = min(2 * sai->sai_max, sbi->ll_sa_max);
1486         } else {
1487                 struct ll_inode_info *lli = ll_i2info(sai->sai_inode);
1488
1489                 sai->sai_miss++;
1490                 sai->sai_consecutive_miss++;
1491                 if (sa_low_hit(sai) && thread_is_running(thread)) {
1492                         atomic_inc(&sbi->ll_sa_wrong);
1493                         CDEBUG(D_READA, "Statahead for dir " DFID " hit ratio too low: hit/miss %llu/%llu, sent/replied %llu/%llu, stopping statahead thread\n",
1494                                PFID(&lli->lli_fid), sai->sai_hit,
1495                                sai->sai_miss, sai->sai_sent,
1496                                sai->sai_replied);
1497                         spin_lock(&lli->lli_sa_lock);
1498                         if (!thread_is_stopped(thread))
1499                                 thread_set_flags(thread, SVC_STOPPING);
1500                         spin_unlock(&lli->lli_sa_lock);
1501                 }
1502         }
1503
1504         if (!thread_is_stopped(thread))
1505                 wake_up(&thread->t_ctl_waitq);
1506 }
1507
1508 /**
1509  * Start statahead thread if this is the first dir entry.
1510  * Otherwise if a thread is started already, wait it until it is ahead of me.
1511  * \retval 1       -- find entry with lock in cache, the caller needs to do
1512  *                  nothing.
1513  * \retval 0       -- find entry in cache, but without lock, the caller needs
1514  *                  refresh from MDS.
1515  * \retval others  -- the caller need to process as non-statahead.
1516  */
1517 int do_statahead_enter(struct inode *dir, struct dentry **dentryp,
1518                        int only_unplug)
1519 {
1520         struct ll_inode_info     *lli   = ll_i2info(dir);
1521         struct ll_statahead_info *sai   = lli->lli_sai;
1522         struct dentry       *parent;
1523         struct ll_sa_entry       *entry;
1524         struct ptlrpc_thread     *thread;
1525         struct l_wait_info      lwi   = { 0 };
1526         int                    rc    = 0;
1527         struct ll_inode_info     *plli;
1528
1529         LASSERT(lli->lli_opendir_pid == current_pid());
1530
1531         if (sai) {
1532                 thread = &sai->sai_thread;
1533                 if (unlikely(thread_is_stopped(thread) &&
1534                              list_empty(&sai->sai_entries_stated))) {
1535                         /* to release resource */
1536                         ll_stop_statahead(dir, lli->lli_opendir_key);
1537                         return -EAGAIN;
1538                 }
1539
1540                 if ((*dentryp)->d_name.name[0] == '.') {
1541                         if (sai->sai_ls_all ||
1542                             sai->sai_miss_hidden >= sai->sai_skip_hidden) {
1543                                 /*
1544                                  * Hidden dentry is the first one, or statahead
1545                                  * thread does not skip so many hidden dentries
1546                                  * before "sai_ls_all" enabled as below.
1547                                  */
1548                         } else {
1549                                 if (!sai->sai_ls_all)
1550                                         /*
1551                                          * It maybe because hidden dentry is not
1552                                          * the first one, "sai_ls_all" was not
1553                                          * set, then "ls -al" missed. Enable
1554                                          * "sai_ls_all" for such case.
1555                                          */
1556                                         sai->sai_ls_all = 1;
1557
1558                                 /*
1559                                  * Such "getattr" has been skipped before
1560                                  * "sai_ls_all" enabled as above.
1561                                  */
1562                                 sai->sai_miss_hidden++;
1563                                 return -EAGAIN;
1564                         }
1565                 }
1566
1567                 entry = ll_sa_entry_get_byname(sai, &(*dentryp)->d_name);
1568                 if (entry == NULL || only_unplug) {
1569                         ll_sai_unplug(sai, entry);
1570                         return entry ? 1 : -EAGAIN;
1571                 }
1572
1573                 if (!ll_sa_entry_stated(entry)) {
1574                         sai->sai_index_wait = entry->se_index;
1575                         lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(30), NULL,
1576                                                LWI_ON_SIGNAL_NOOP, NULL);
1577                         rc = l_wait_event(sai->sai_waitq,
1578                                           ll_sa_entry_stated(entry) ||
1579                                           thread_is_stopped(thread),
1580                                           &lwi);
1581                         if (rc < 0) {
1582                                 ll_sai_unplug(sai, entry);
1583                                 return -EAGAIN;
1584                         }
1585                 }
1586
1587                 if (entry->se_stat == SA_ENTRY_SUCC &&
1588                     entry->se_inode != NULL) {
1589                         struct inode *inode = entry->se_inode;
1590                         struct lookup_intent it = { .it_op = IT_GETATTR,
1591                                                     .d.lustre.it_lock_handle =
1592                                                      entry->se_handle };
1593                         __u64 bits;
1594
1595                         rc = md_revalidate_lock(ll_i2mdexp(dir), &it,
1596                                                 ll_inode2fid(inode), &bits);
1597                         if (rc == 1) {
1598                                 if ((*dentryp)->d_inode == NULL) {
1599                                         struct dentry *alias;
1600
1601                                         alias = ll_splice_alias(inode,
1602                                                                    *dentryp);
1603                                         if (IS_ERR(alias)) {
1604                                                 ll_sai_unplug(sai, entry);
1605                                                 return PTR_ERR(alias);
1606                                         }
1607                                         *dentryp = alias;
1608                                 } else if ((*dentryp)->d_inode != inode) {
1609                                         /* revalidate, but inode is recreated */
1610                                         CDEBUG(D_READA,
1611                                               "stale dentry %pd inode %lu/%u, statahead inode %lu/%u\n",
1612                                               *dentryp,
1613                                               (*dentryp)->d_inode->i_ino,
1614                                               (*dentryp)->d_inode->i_generation,
1615                                               inode->i_ino,
1616                                               inode->i_generation);
1617                                         ll_sai_unplug(sai, entry);
1618                                         return -ESTALE;
1619                                 } else {
1620                                         iput(inode);
1621                                 }
1622                                 entry->se_inode = NULL;
1623
1624                                 if ((bits & MDS_INODELOCK_LOOKUP) &&
1625                                     d_lustre_invalid(*dentryp))
1626                                         d_lustre_revalidate(*dentryp);
1627                                 ll_intent_release(&it);
1628                         }
1629                 }
1630
1631                 ll_sai_unplug(sai, entry);
1632                 return rc;
1633         }
1634
1635         /* I am the "lli_opendir_pid" owner, only me can set "lli_sai". */
1636         rc = is_first_dirent(dir, *dentryp);
1637         if (rc == LS_NONE_FIRST_DE) {
1638                 /* It is not "ls -{a}l" operation, no need statahead for it. */
1639                 rc = -EAGAIN;
1640                 goto out;
1641         }
1642
1643         sai = ll_sai_alloc();
1644         if (sai == NULL) {
1645                 rc = -ENOMEM;
1646                 goto out;
1647         }
1648
1649         sai->sai_ls_all = (rc == LS_FIRST_DOT_DE);
1650         sai->sai_inode = igrab(dir);
1651         if (unlikely(sai->sai_inode == NULL)) {
1652                 CWARN("Do not start stat ahead on dying inode "DFID"\n",
1653                       PFID(&lli->lli_fid));
1654                 rc = -ESTALE;
1655                 goto out;
1656         }
1657
1658         /* get parent reference count here, and put it in ll_statahead_thread */
1659         parent = dget((*dentryp)->d_parent);
1660         if (unlikely(sai->sai_inode != parent->d_inode)) {
1661                 struct ll_inode_info *nlli = ll_i2info(parent->d_inode);
1662
1663                 CWARN("Race condition, someone changed %pd just now: old parent "DFID", new parent "DFID"\n",
1664                       *dentryp,
1665                       PFID(&lli->lli_fid), PFID(&nlli->lli_fid));
1666                 dput(parent);
1667                 iput(sai->sai_inode);
1668                 rc = -EAGAIN;
1669                 goto out;
1670         }
1671
1672         CDEBUG(D_READA, "start statahead thread: sai %p, parent %pd\n",
1673                sai, parent);
1674
1675         /* The sai buffer already has one reference taken at allocation time,
1676          * but as soon as we expose the sai by attaching it to the lli that
1677          * default reference can be dropped by another thread calling
1678          * ll_stop_statahead. We need to take a local reference to protect
1679          * the sai buffer while we intend to access it. */
1680         ll_sai_get(sai);
1681         lli->lli_sai = sai;
1682
1683         plli = ll_i2info(parent->d_inode);
1684         rc = PTR_ERR(kthread_run(ll_statahead_thread, parent,
1685                                  "ll_sa_%u", plli->lli_opendir_pid));
1686         thread = &sai->sai_thread;
1687         if (IS_ERR_VALUE(rc)) {
1688                 CERROR("can't start ll_sa thread, rc: %d\n", rc);
1689                 dput(parent);
1690                 lli->lli_opendir_key = NULL;
1691                 thread_set_flags(thread, SVC_STOPPED);
1692                 thread_set_flags(&sai->sai_agl_thread, SVC_STOPPED);
1693                 /* Drop both our own local reference and the default
1694                  * reference from allocation time. */
1695                 ll_sai_put(sai);
1696                 ll_sai_put(sai);
1697                 LASSERT(lli->lli_sai == NULL);
1698                 return -EAGAIN;
1699         }
1700
1701         l_wait_event(thread->t_ctl_waitq,
1702                      thread_is_running(thread) || thread_is_stopped(thread),
1703                      &lwi);
1704         ll_sai_put(sai);
1705
1706         /*
1707          * We don't stat-ahead for the first dirent since we are already in
1708          * lookup.
1709          */
1710         return -EAGAIN;
1711
1712 out:
1713         if (sai != NULL)
1714                 OBD_FREE_PTR(sai);
1715         spin_lock(&lli->lli_sa_lock);
1716         lli->lli_opendir_key = NULL;
1717         lli->lli_opendir_pid = 0;
1718         spin_unlock(&lli->lli_sa_lock);
1719         return rc;
1720 }