kernel/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                 kfree(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                 kfree(entry);
 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                 kfree(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         __u64                     handle = 0;
 710         int                    wakeup;
 711
 712         if (it_disposition(it, DISP_LOOKUP_NEG))
 713                 rc = -ENOENT;
 714
 715         if (rc == 0) {
 716                 /* release ibits lock ASAP to avoid deadlock when statahead
 717                  * thread enqueues lock on parent in readdir and another
 718                  * process enqueues lock on child with parent lock held, eg.
 719                  * unlink. */
 720                 handle = it->d.lustre.it_lock_handle;
 721                 ll_intent_drop_lock(it);
 722         }
 723
 724         spin_lock(&lli->lli_sa_lock);
 725         /* stale entry */
 726         if (unlikely(lli->lli_sai == NULL ||
 727                      lli->lli_sai->sai_generation != minfo->mi_generation)) {
 728                 spin_unlock(&lli->lli_sa_lock);
 729                 rc = -ESTALE;
 730                 goto out;
 731         } else {
 732                 sai = ll_sai_get(lli->lli_sai);
 733                 if (unlikely(!thread_is_running(&sai->sai_thread))) {
 734                         sai->sai_replied++;
 735                         spin_unlock(&lli->lli_sa_lock);
 736                         rc = -EBADFD;
 737                         goto out;
 738                 }
 739
 740                 entry = ll_sa_entry_get_byindex(sai, minfo->mi_cbdata);
 741                 if (entry == NULL) {
 742                         sai->sai_replied++;
 743                         spin_unlock(&lli->lli_sa_lock);
 744                         rc = -EIDRM;
 745                         goto out;
 746                 }
 747
 748                 if (rc != 0) {
 749                         do_sa_entry_to_stated(sai, entry, SA_ENTRY_INVA);
 750                         wakeup = (entry->se_index == sai->sai_index_wait);
 751                 } else {
 752                         entry->se_minfo = minfo;
 753                         entry->se_req = ptlrpc_request_addref(req);
 754                         /* Release the async ibits lock ASAP to avoid deadlock
 755                          * when statahead thread tries to enqueue lock on parent
 756                          * for readpage and other tries to enqueue lock on child
 757                          * with parent's lock held, for example: unlink. */
 758                         entry->se_handle = handle;
 759                         wakeup = sa_received_empty(sai);
 760                         list_add_tail(&entry->se_list,
 761                                           &sai->sai_entries_received);
 762                 }
 763                 sai->sai_replied++;
 764                 spin_unlock(&lli->lli_sa_lock);
 765
 766                 ll_sa_entry_put(sai, entry);
 767                 if (wakeup)
 768                         wake_up(&sai->sai_thread.t_ctl_waitq);
 769         }
 770
 771 out:
 772         if (rc != 0) {
 773                 ll_intent_release(it);
 774                 iput(dir);
 775                 kfree(minfo);
 776         }
 777         if (sai != NULL)
 778                 ll_sai_put(sai);
 779         return rc;
 780 }
 781
 782 static void sa_args_fini(struct md_enqueue_info *minfo,
 783                          struct ldlm_enqueue_info *einfo)
 784 {
 785         LASSERT(minfo && einfo);
 786         iput(minfo->mi_dir);
 787         kfree(minfo);
 788         kfree(einfo);
 789 }
 790
 791 /**
 792  * prepare arguments for async stat RPC.
 793  */
 794 static int sa_args_init(struct inode *dir, struct inode *child,
 795                         struct ll_sa_entry *entry, struct md_enqueue_info **pmi,
 796                         struct ldlm_enqueue_info **pei)
 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                 kfree(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                 kfree(einfo);
 818                 kfree(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
 838         return 0;
 839 }
 840
 841 static int do_sa_lookup(struct inode *dir, struct ll_sa_entry *entry)
 842 {
 843         struct md_enqueue_info   *minfo;
 844         struct ldlm_enqueue_info *einfo;
 845         int                    rc;
 846
 847         rc = sa_args_init(dir, NULL, entry, &minfo, &einfo);
 848         if (rc)
 849                 return rc;
 850
 851         rc = md_intent_getattr_async(ll_i2mdexp(dir), minfo, einfo);
 852         if (rc < 0)
 853                 sa_args_fini(minfo, einfo);
 854
 855         return rc;
 856 }
 857
 858 /**
 859  * similar to ll_revalidate_it().
 860  * \retval      1 -- dentry valid
 861  * \retval      0 -- will send stat-ahead request
 862  * \retval others -- prepare stat-ahead request failed
 863  */
 864 static int do_sa_revalidate(struct inode *dir, struct ll_sa_entry *entry,
 865                             struct dentry *dentry)
 866 {
 867         struct inode         *inode = d_inode(dentry);
 868         struct lookup_intent      it = { .it_op = IT_GETATTR,
 869                                          .d.lustre.it_lock_handle = 0 };
 870         struct md_enqueue_info   *minfo;
 871         struct ldlm_enqueue_info *einfo;
 872         int rc;
 873
 874         if (unlikely(inode == NULL))
 875                 return 1;
 876
 877         if (d_mountpoint(dentry))
 878                 return 1;
 879
 880         entry->se_inode = igrab(inode);
 881         rc = md_revalidate_lock(ll_i2mdexp(dir), &it, ll_inode2fid(inode),
 882                                 NULL);
 883         if (rc == 1) {
 884                 entry->se_handle = it.d.lustre.it_lock_handle;
 885                 ll_intent_release(&it);
 886                 return 1;
 887         }
 888
 889         rc = sa_args_init(dir, inode, entry, &minfo, &einfo);
 890         if (rc) {
 891                 entry->se_inode = NULL;
 892                 iput(inode);
 893                 return rc;
 894         }
 895
 896         rc = md_intent_getattr_async(ll_i2mdexp(dir), minfo, einfo);
 897         if (rc < 0) {
 898                 entry->se_inode = NULL;
 899                 iput(inode);
 900                 sa_args_fini(minfo, einfo);
 901         }
 902
 903         return rc;
 904 }
 905
 906 static void ll_statahead_one(struct dentry *parent, const char *entry_name,
 907                              int entry_name_len)
 908 {
 909         struct inode         *dir    = d_inode(parent);
 910         struct ll_inode_info     *lli    = ll_i2info(dir);
 911         struct ll_statahead_info *sai    = lli->lli_sai;
 912         struct dentry       *dentry = NULL;
 913         struct ll_sa_entry       *entry;
 914         int                    rc;
 915         int                    rc1;
 916
 917         entry = ll_sa_entry_alloc(sai, sai->sai_index, entry_name,
 918                                   entry_name_len);
 919         if (IS_ERR(entry))
 920                 return;
 921
 922         dentry = d_lookup(parent, &entry->se_qstr);
 923         if (!dentry) {
 924                 rc = do_sa_lookup(dir, entry);
 925         } else {
 926                 rc = do_sa_revalidate(dir, entry, dentry);
 927                 if (rc == 1 && agl_should_run(sai, d_inode(dentry)))
 928                         ll_agl_add(sai, d_inode(dentry), entry->se_index);
 929         }
 930
 931         if (dentry != NULL)
 932                 dput(dentry);
 933
 934         if (rc) {
 935                 rc1 = ll_sa_entry_to_stated(sai, entry,
 936                                         rc < 0 ? SA_ENTRY_INVA : SA_ENTRY_SUCC);
 937                 if (rc1 == 0 && entry->se_index == sai->sai_index_wait)
 938                         wake_up(&sai->sai_waitq);
 939         } else {
 940                 sai->sai_sent++;
 941         }
 942
 943         sai->sai_index++;
 944         /* drop one refcount on entry by ll_sa_entry_alloc */
 945         ll_sa_entry_put(sai, entry);
 946 }
 947
 948 static int ll_agl_thread(void *arg)
 949 {
 950         struct dentry       *parent = arg;
 951         struct inode         *dir    = d_inode(parent);
 952         struct ll_inode_info     *plli   = ll_i2info(dir);
 953         struct ll_inode_info     *clli;
 954         struct ll_sb_info       *sbi    = ll_i2sbi(dir);
 955         struct ll_statahead_info *sai    = ll_sai_get(plli->lli_sai);
 956         struct ptlrpc_thread     *thread = &sai->sai_agl_thread;
 957         struct l_wait_info      lwi    = { 0 };
 958
 959         thread->t_pid = current_pid();
 960         CDEBUG(D_READA, "agl thread started: sai %p, parent %pd\n",
 961                sai, parent);
 962
 963         atomic_inc(&sbi->ll_agl_total);
 964         spin_lock(&plli->lli_agl_lock);
 965         sai->sai_agl_valid = 1;
 966         if (thread_is_init(thread))
 967                 /* If someone else has changed the thread state
 968                  * (e.g. already changed to SVC_STOPPING), we can't just
 969                  * blindly overwrite that setting. */
 970                 thread_set_flags(thread, SVC_RUNNING);
 971         spin_unlock(&plli->lli_agl_lock);
 972         wake_up(&thread->t_ctl_waitq);
 973
 974         while (1) {
 975                 l_wait_event(thread->t_ctl_waitq,
 976                              !agl_list_empty(sai) ||
 977                              !thread_is_running(thread),
 978                              &lwi);
 979
 980                 if (!thread_is_running(thread))
 981                         break;
 982
 983                 spin_lock(&plli->lli_agl_lock);
 984                 /* The statahead thread maybe help to process AGL entries,
 985                  * so check whether list empty again. */
 986                 if (!agl_list_empty(sai)) {
 987                         clli = agl_first_entry(sai);
 988                         list_del_init(&clli->lli_agl_list);
 989                         spin_unlock(&plli->lli_agl_lock);
 990                         ll_agl_trigger(&clli->lli_vfs_inode, sai);
 991                 } else {
 992                         spin_unlock(&plli->lli_agl_lock);
 993                 }
 994         }
 995
 996         spin_lock(&plli->lli_agl_lock);
 997         sai->sai_agl_valid = 0;
 998         while (!agl_list_empty(sai)) {
 999                 clli = agl_first_entry(sai);
1000                 list_del_init(&clli->lli_agl_list);
1001                 spin_unlock(&plli->lli_agl_lock);
1002                 clli->lli_agl_index = 0;
1003                 iput(&clli->lli_vfs_inode);
1004                 spin_lock(&plli->lli_agl_lock);
1005         }
1006         thread_set_flags(thread, SVC_STOPPED);
1007         spin_unlock(&plli->lli_agl_lock);
1008         wake_up(&thread->t_ctl_waitq);
1009         ll_sai_put(sai);
1010         CDEBUG(D_READA, "agl thread stopped: sai %p, parent %pd\n",
1011                sai, parent);
1012         return 0;
1013 }
1014
1015 static void ll_start_agl(struct dentry *parent, struct ll_statahead_info *sai)
1016 {
1017         struct ptlrpc_thread *thread = &sai->sai_agl_thread;
1018         struct l_wait_info    lwi    = { 0 };
1019         struct ll_inode_info  *plli;
1020         struct task_struct *task;
1021
1022         CDEBUG(D_READA, "start agl thread: sai %p, parent %pd\n",
1023                sai, parent);
1024
1025         plli = ll_i2info(d_inode(parent));
1026         task = kthread_run(ll_agl_thread, parent,
1027                                "ll_agl_%u", plli->lli_opendir_pid);
1028         if (IS_ERR(task)) {
1029                 CERROR("can't start ll_agl thread, rc: %ld\n", PTR_ERR(task));
1030                 thread_set_flags(thread, SVC_STOPPED);
1031                 return;
1032         }
1033
1034         l_wait_event(thread->t_ctl_waitq,
1035                      thread_is_running(thread) || thread_is_stopped(thread),
1036                      &lwi);
1037 }
1038
1039 static int ll_statahead_thread(void *arg)
1040 {
1041         struct dentry       *parent = arg;
1042         struct inode         *dir    = d_inode(parent);
1043         struct ll_inode_info     *plli   = ll_i2info(dir);
1044         struct ll_inode_info     *clli;
1045         struct ll_sb_info       *sbi    = ll_i2sbi(dir);
1046         struct ll_statahead_info *sai    = ll_sai_get(plli->lli_sai);
1047         struct ptlrpc_thread     *thread = &sai->sai_thread;
1048         struct ptlrpc_thread *agl_thread = &sai->sai_agl_thread;
1049         struct page           *page;
1050         __u64                pos    = 0;
1051         int                    first  = 0;
1052         int                    rc     = 0;
1053         struct ll_dir_chain       chain;
1054         struct l_wait_info      lwi    = { 0 };
1055
1056         thread->t_pid = current_pid();
1057         CDEBUG(D_READA, "statahead thread starting: sai %p, parent %pd\n",
1058                sai, parent);
1059
1060         if (sbi->ll_flags & LL_SBI_AGL_ENABLED)
1061                 ll_start_agl(parent, sai);
1062
1063         atomic_inc(&sbi->ll_sa_total);
1064         spin_lock(&plli->lli_sa_lock);
1065         if (thread_is_init(thread))
1066                 /* If someone else has changed the thread state
1067                  * (e.g. already changed to SVC_STOPPING), we can't just
1068                  * blindly overwrite that setting. */
1069                 thread_set_flags(thread, SVC_RUNNING);
1070         spin_unlock(&plli->lli_sa_lock);
1071         wake_up(&thread->t_ctl_waitq);
1072
1073         ll_dir_chain_init(&chain);
1074         page = ll_get_dir_page(dir, pos, &chain);
1075
1076         while (1) {
1077                 struct lu_dirpage *dp;
1078                 struct lu_dirent  *ent;
1079
1080                 if (IS_ERR(page)) {
1081                         rc = PTR_ERR(page);
1082                         CDEBUG(D_READA, "error reading dir "DFID" at %llu/%llu: [rc %d] [parent %u]\n",
1083                                PFID(ll_inode2fid(dir)), pos, sai->sai_index,
1084                                rc, plli->lli_opendir_pid);
1085                         goto out;
1086                 }
1087
1088                 dp = page_address(page);
1089                 for (ent = lu_dirent_start(dp); ent != NULL;
1090                      ent = lu_dirent_next(ent)) {
1091                         __u64 hash;
1092                         int namelen;
1093                         char *name;
1094
1095                         hash = le64_to_cpu(ent->lde_hash);
1096                         if (unlikely(hash < pos))
1097                                 /*
1098                                  * Skip until we find target hash value.
1099                                  */
1100                                 continue;
1101
1102                         namelen = le16_to_cpu(ent->lde_namelen);
1103                         if (unlikely(namelen == 0))
1104                                 /*
1105                                  * Skip dummy record.
1106                                  */
1107                                 continue;
1108
1109                         name = ent->lde_name;
1110                         if (name[0] == '.') {
1111                                 if (namelen == 1) {
1112                                         /*
1113                                          * skip "."
1114                                          */
1115                                         continue;
1116                                 } else if (name[1] == '.' && namelen == 2) {
1117                                         /*
1118                                          * skip ".."
1119                                          */
1120                                         continue;
1121                                 } else if (!sai->sai_ls_all) {
1122                                         /*
1123                                          * skip hidden files.
1124                                          */
1125                                         sai->sai_skip_hidden++;
1126                                         continue;
1127                                 }
1128                         }
1129
1130                         /*
1131                          * don't stat-ahead first entry.
1132                          */
1133                         if (unlikely(++first == 1))
1134                                 continue;
1135
1136 keep_it:
1137                         l_wait_event(thread->t_ctl_waitq,
1138                                      !sa_sent_full(sai) ||
1139                                      !sa_received_empty(sai) ||
1140                                      !agl_list_empty(sai) ||
1141                                      !thread_is_running(thread),
1142                                      &lwi);
1143
1144 interpret_it:
1145                         while (!sa_received_empty(sai))
1146                                 ll_post_statahead(sai);
1147
1148                         if (unlikely(!thread_is_running(thread))) {
1149                                 ll_release_page(page, 0);
1150                                 rc = 0;
1151                                 goto out;
1152                         }
1153
1154                         /* If no window for metadata statahead, but there are
1155                          * some AGL entries to be triggered, then try to help
1156                          * to process the AGL entries. */
1157                         if (sa_sent_full(sai)) {
1158                                 spin_lock(&plli->lli_agl_lock);
1159                                 while (!agl_list_empty(sai)) {
1160                                         clli = agl_first_entry(sai);
1161                                         list_del_init(&clli->lli_agl_list);
1162                                         spin_unlock(&plli->lli_agl_lock);
1163                                         ll_agl_trigger(&clli->lli_vfs_inode,
1164                                                        sai);
1165
1166                                         if (!sa_received_empty(sai))
1167                                                 goto interpret_it;
1168
1169                                         if (unlikely(
1170                                                 !thread_is_running(thread))) {
1171                                                 ll_release_page(page, 0);
1172                                                 rc = 0;
1173                                                 goto out;
1174                                         }
1175
1176                                         if (!sa_sent_full(sai))
1177                                                 goto do_it;
1178
1179                                         spin_lock(&plli->lli_agl_lock);
1180                                 }
1181                                 spin_unlock(&plli->lli_agl_lock);
1182
1183                                 goto keep_it;
1184                         }
1185
1186 do_it:
1187                         ll_statahead_one(parent, name, namelen);
1188                 }
1189                 pos = le64_to_cpu(dp->ldp_hash_end);
1190                 if (pos == MDS_DIR_END_OFF) {
1191                         /*
1192                          * End of directory reached.
1193                          */
1194                         ll_release_page(page, 0);
1195                         while (1) {
1196                                 l_wait_event(thread->t_ctl_waitq,
1197                                              !sa_received_empty(sai) ||
1198                                              sai->sai_sent == sai->sai_replied ||
1199                                              !thread_is_running(thread),
1200                                              &lwi);
1201
1202                                 while (!sa_received_empty(sai))
1203                                         ll_post_statahead(sai);
1204
1205                                 if (unlikely(!thread_is_running(thread))) {
1206                                         rc = 0;
1207                                         goto out;
1208                                 }
1209
1210                                 if (sai->sai_sent == sai->sai_replied &&
1211                                     sa_received_empty(sai))
1212                                         break;
1213                         }
1214
1215                         spin_lock(&plli->lli_agl_lock);
1216                         while (!agl_list_empty(sai) &&
1217                                thread_is_running(thread)) {
1218                                 clli = agl_first_entry(sai);
1219                                 list_del_init(&clli->lli_agl_list);
1220                                 spin_unlock(&plli->lli_agl_lock);
1221                                 ll_agl_trigger(&clli->lli_vfs_inode, sai);
1222                                 spin_lock(&plli->lli_agl_lock);
1223                         }
1224                         spin_unlock(&plli->lli_agl_lock);
1225
1226                         rc = 0;
1227                         goto out;
1228                 } else if (1) {
1229                         /*
1230                          * chain is exhausted.
1231                          * Normal case: continue to the next page.
1232                          */
1233                         ll_release_page(page, le32_to_cpu(dp->ldp_flags) &
1234                                               LDF_COLLIDE);
1235                         page = ll_get_dir_page(dir, pos, &chain);
1236                 } else {
1237                         LASSERT(le32_to_cpu(dp->ldp_flags) & LDF_COLLIDE);
1238                         ll_release_page(page, 1);
1239                         /*
1240                          * go into overflow page.
1241                          */
1242                 }
1243         }
1244
1245 out:
1246         if (sai->sai_agl_valid) {
1247                 spin_lock(&plli->lli_agl_lock);
1248                 thread_set_flags(agl_thread, SVC_STOPPING);
1249                 spin_unlock(&plli->lli_agl_lock);
1250                 wake_up(&agl_thread->t_ctl_waitq);
1251
1252                 CDEBUG(D_READA, "stop agl thread: sai %p pid %u\n",
1253                        sai, (unsigned int)agl_thread->t_pid);
1254                 l_wait_event(agl_thread->t_ctl_waitq,
1255                              thread_is_stopped(agl_thread),
1256                              &lwi);
1257         } else {
1258                 /* Set agl_thread flags anyway. */
1259                 thread_set_flags(&sai->sai_agl_thread, SVC_STOPPED);
1260         }
1261         ll_dir_chain_fini(&chain);
1262         spin_lock(&plli->lli_sa_lock);
1263         if (!sa_received_empty(sai)) {
1264                 thread_set_flags(thread, SVC_STOPPING);
1265                 spin_unlock(&plli->lli_sa_lock);
1266
1267                 /* To release the resources held by received entries. */
1268                 while (!sa_received_empty(sai))
1269                         ll_post_statahead(sai);
1270
1271                 spin_lock(&plli->lli_sa_lock);
1272         }
1273         thread_set_flags(thread, SVC_STOPPED);
1274         spin_unlock(&plli->lli_sa_lock);
1275         wake_up(&sai->sai_waitq);
1276         wake_up(&thread->t_ctl_waitq);
1277         ll_sai_put(sai);
1278         dput(parent);
1279         CDEBUG(D_READA, "statahead thread stopped: sai %p, parent %pd\n",
1280                sai, parent);
1281         return rc;
1282 }
1283
1284 /**
1285  * called in ll_file_release().
1286  */
1287 void ll_stop_statahead(struct inode *dir, void *key)
1288 {
1289         struct ll_inode_info *lli = ll_i2info(dir);
1290
1291         if (unlikely(key == NULL))
1292                 return;
1293
1294         spin_lock(&lli->lli_sa_lock);
1295         if (lli->lli_opendir_key != key || lli->lli_opendir_pid == 0) {
1296                 spin_unlock(&lli->lli_sa_lock);
1297                 return;
1298         }
1299
1300         lli->lli_opendir_key = NULL;
1301
1302         if (lli->lli_sai) {
1303                 struct l_wait_info lwi = { 0 };
1304                 struct ptlrpc_thread *thread = &lli->lli_sai->sai_thread;
1305
1306                 if (!thread_is_stopped(thread)) {
1307                         thread_set_flags(thread, SVC_STOPPING);
1308                         spin_unlock(&lli->lli_sa_lock);
1309                         wake_up(&thread->t_ctl_waitq);
1310
1311                         CDEBUG(D_READA, "stop statahead thread: sai %p pid %u\n",
1312                                lli->lli_sai, (unsigned int)thread->t_pid);
1313                         l_wait_event(thread->t_ctl_waitq,
1314                                      thread_is_stopped(thread),
1315                                      &lwi);
1316                 } else {
1317                         spin_unlock(&lli->lli_sa_lock);
1318                 }
1319
1320                 /*
1321                  * Put the ref which was held when first statahead_enter.
1322                  * It maybe not the last ref for some statahead requests
1323                  * maybe inflight.
1324                  */
1325                 ll_sai_put(lli->lli_sai);
1326         } else {
1327                 lli->lli_opendir_pid = 0;
1328                 spin_unlock(&lli->lli_sa_lock);
1329         }
1330 }
1331
1332 enum {
1333         /**
1334          * not first dirent, or is "."
1335          */
1336         LS_NONE_FIRST_DE = 0,
1337         /**
1338          * the first non-hidden dirent
1339          */
1340         LS_FIRST_DE,
1341         /**
1342          * the first hidden dirent, that is "."
1343          */
1344         LS_FIRST_DOT_DE
1345 };
1346
1347 static int is_first_dirent(struct inode *dir, struct dentry *dentry)
1348 {
1349         struct ll_dir_chain   chain;
1350         struct qstr       *target = &dentry->d_name;
1351         struct page       *page;
1352         __u64            pos    = 0;
1353         int                dot_de;
1354         int                rc     = LS_NONE_FIRST_DE;
1355
1356         ll_dir_chain_init(&chain);
1357         page = ll_get_dir_page(dir, pos, &chain);
1358
1359         while (1) {
1360                 struct lu_dirpage *dp;
1361                 struct lu_dirent  *ent;
1362
1363                 if (IS_ERR(page)) {
1364                         struct ll_inode_info *lli = ll_i2info(dir);
1365
1366                         rc = PTR_ERR(page);
1367                         CERROR("error reading dir "DFID" at %llu: [rc %d] [parent %u]\n",
1368                                PFID(ll_inode2fid(dir)), pos,
1369                                rc, lli->lli_opendir_pid);
1370                         break;
1371                 }
1372
1373                 dp = page_address(page);
1374                 for (ent = lu_dirent_start(dp); ent != NULL;
1375                      ent = lu_dirent_next(ent)) {
1376                         __u64 hash;
1377                         int namelen;
1378                         char *name;
1379
1380                         hash = le64_to_cpu(ent->lde_hash);
1381                         /* The ll_get_dir_page() can return any page containing
1382                          * the given hash which may be not the start hash. */
1383                         if (unlikely(hash < pos))
1384                                 continue;
1385
1386                         namelen = le16_to_cpu(ent->lde_namelen);
1387                         if (unlikely(namelen == 0))
1388                                 /*
1389                                  * skip dummy record.
1390                                  */
1391                                 continue;
1392
1393                         name = ent->lde_name;
1394                         if (name[0] == '.') {
1395                                 if (namelen == 1)
1396                                         /*
1397                                          * skip "."
1398                                          */
1399                                         continue;
1400                                 else if (name[1] == '.' && namelen == 2)
1401                                         /*
1402                                          * skip ".."
1403                                          */
1404                                         continue;
1405                                 else
1406                                         dot_de = 1;
1407                         } else {
1408                                 dot_de = 0;
1409                         }
1410
1411                         if (dot_de && target->name[0] != '.') {
1412                                 CDEBUG(D_READA, "%.*s skip hidden file %.*s\n",
1413                                        target->len, target->name,
1414                                        namelen, name);
1415                                 continue;
1416                         }
1417
1418                         if (target->len != namelen ||
1419                             memcmp(target->name, name, namelen) != 0)
1420                                 rc = LS_NONE_FIRST_DE;
1421                         else if (!dot_de)
1422                                 rc = LS_FIRST_DE;
1423                         else
1424                                 rc = LS_FIRST_DOT_DE;
1425
1426                         ll_release_page(page, 0);
1427                         goto out;
1428                 }
1429                 pos = le64_to_cpu(dp->ldp_hash_end);
1430                 if (pos == MDS_DIR_END_OFF) {
1431                         /*
1432                          * End of directory reached.
1433                          */
1434                         ll_release_page(page, 0);
1435                         break;
1436                 } else if (1) {
1437                         /*
1438                          * chain is exhausted
1439                          * Normal case: continue to the next page.
1440                          */
1441                         ll_release_page(page, le32_to_cpu(dp->ldp_flags) &
1442                                               LDF_COLLIDE);
1443                         page = ll_get_dir_page(dir, pos, &chain);
1444                 } else {
1445                         /*
1446                          * go into overflow page.
1447                          */
1448                         LASSERT(le32_to_cpu(dp->ldp_flags) & LDF_COLLIDE);
1449                         ll_release_page(page, 1);
1450                 }
1451         }
1452
1453 out:
1454         ll_dir_chain_fini(&chain);
1455         return rc;
1456 }
1457
1458 static void
1459 ll_sai_unplug(struct ll_statahead_info *sai, struct ll_sa_entry *entry)
1460 {
1461         struct ptlrpc_thread *thread = &sai->sai_thread;
1462         struct ll_sb_info    *sbi    = ll_i2sbi(sai->sai_inode);
1463         int                hit;
1464
1465         if (entry != NULL && entry->se_stat == SA_ENTRY_SUCC)
1466                 hit = 1;
1467         else
1468                 hit = 0;
1469
1470         ll_sa_entry_fini(sai, entry);
1471         if (hit) {
1472                 sai->sai_hit++;
1473                 sai->sai_consecutive_miss = 0;
1474                 sai->sai_max = min(2 * sai->sai_max, sbi->ll_sa_max);
1475         } else {
1476                 struct ll_inode_info *lli = ll_i2info(sai->sai_inode);
1477
1478                 sai->sai_miss++;
1479                 sai->sai_consecutive_miss++;
1480                 if (sa_low_hit(sai) && thread_is_running(thread)) {
1481                         atomic_inc(&sbi->ll_sa_wrong);
1482                         CDEBUG(D_READA, "Statahead for dir " DFID " hit ratio too low: hit/miss %llu/%llu, sent/replied %llu/%llu, stopping statahead thread\n",
1483                                PFID(&lli->lli_fid), sai->sai_hit,
1484                                sai->sai_miss, sai->sai_sent,
1485                                sai->sai_replied);
1486                         spin_lock(&lli->lli_sa_lock);
1487                         if (!thread_is_stopped(thread))
1488                                 thread_set_flags(thread, SVC_STOPPING);
1489                         spin_unlock(&lli->lli_sa_lock);
1490                 }
1491         }
1492
1493         if (!thread_is_stopped(thread))
1494                 wake_up(&thread->t_ctl_waitq);
1495 }
1496
1497 /**
1498  * Start statahead thread if this is the first dir entry.
1499  * Otherwise if a thread is started already, wait it until it is ahead of me.
1500  * \retval 1       -- find entry with lock in cache, the caller needs to do
1501  *                  nothing.
1502  * \retval 0       -- find entry in cache, but without lock, the caller needs
1503  *                  refresh from MDS.
1504  * \retval others  -- the caller need to process as non-statahead.
1505  */
1506 int do_statahead_enter(struct inode *dir, struct dentry **dentryp,
1507                        int only_unplug)
1508 {
1509         struct ll_inode_info     *lli   = ll_i2info(dir);
1510         struct ll_statahead_info *sai   = lli->lli_sai;
1511         struct dentry       *parent;
1512         struct ll_sa_entry       *entry;
1513         struct ptlrpc_thread     *thread;
1514         struct l_wait_info      lwi   = { 0 };
1515         int                    rc    = 0;
1516         struct ll_inode_info     *plli;
1517
1518         LASSERT(lli->lli_opendir_pid == current_pid());
1519
1520         if (sai) {
1521                 thread = &sai->sai_thread;
1522                 if (unlikely(thread_is_stopped(thread) &&
1523                              list_empty(&sai->sai_entries_stated))) {
1524                         /* to release resource */
1525                         ll_stop_statahead(dir, lli->lli_opendir_key);
1526                         return -EAGAIN;
1527                 }
1528
1529                 if ((*dentryp)->d_name.name[0] == '.') {
1530                         if (sai->sai_ls_all ||
1531                             sai->sai_miss_hidden >= sai->sai_skip_hidden) {
1532                                 /*
1533                                  * Hidden dentry is the first one, or statahead
1534                                  * thread does not skip so many hidden dentries
1535                                  * before "sai_ls_all" enabled as below.
1536                                  */
1537                         } else {
1538                                 if (!sai->sai_ls_all)
1539                                         /*
1540                                          * It maybe because hidden dentry is not
1541                                          * the first one, "sai_ls_all" was not
1542                                          * set, then "ls -al" missed. Enable
1543                                          * "sai_ls_all" for such case.
1544                                          */
1545                                         sai->sai_ls_all = 1;
1546
1547                                 /*
1548                                  * Such "getattr" has been skipped before
1549                                  * "sai_ls_all" enabled as above.
1550                                  */
1551                                 sai->sai_miss_hidden++;
1552                                 return -EAGAIN;
1553                         }
1554                 }
1555
1556                 entry = ll_sa_entry_get_byname(sai, &(*dentryp)->d_name);
1557                 if (entry == NULL || only_unplug) {
1558                         ll_sai_unplug(sai, entry);
1559                         return entry ? 1 : -EAGAIN;
1560                 }
1561
1562                 if (!ll_sa_entry_stated(entry)) {
1563                         sai->sai_index_wait = entry->se_index;
1564                         lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(30), NULL,
1565                                                LWI_ON_SIGNAL_NOOP, NULL);
1566                         rc = l_wait_event(sai->sai_waitq,
1567                                           ll_sa_entry_stated(entry) ||
1568                                           thread_is_stopped(thread),
1569                                           &lwi);
1570                         if (rc < 0) {
1571                                 ll_sai_unplug(sai, entry);
1572                                 return -EAGAIN;
1573                         }
1574                 }
1575
1576                 if (entry->se_stat == SA_ENTRY_SUCC &&
1577                     entry->se_inode != NULL) {
1578                         struct inode *inode = entry->se_inode;
1579                         struct lookup_intent it = { .it_op = IT_GETATTR,
1580                                                     .d.lustre.it_lock_handle =
1581                                                      entry->se_handle };
1582                         __u64 bits;
1583
1584                         rc = md_revalidate_lock(ll_i2mdexp(dir), &it,
1585                                                 ll_inode2fid(inode), &bits);
1586                         if (rc == 1) {
1587                                 if (d_inode(*dentryp) == NULL) {
1588                                         struct dentry *alias;
1589
1590                                         alias = ll_splice_alias(inode,
1591                                                                    *dentryp);
1592                                         if (IS_ERR(alias)) {
1593                                                 ll_sai_unplug(sai, entry);
1594                                                 return PTR_ERR(alias);
1595                                         }
1596                                         *dentryp = alias;
1597                                 } else if (d_inode(*dentryp) != inode) {
1598                                         /* revalidate, but inode is recreated */
1599                                         CDEBUG(D_READA,
1600                                               "stale dentry %pd inode %lu/%u, statahead inode %lu/%u\n",
1601                                               *dentryp,
1602                                               d_inode(*dentryp)->i_ino,
1603                                               d_inode(*dentryp)->i_generation,
1604                                               inode->i_ino,
1605                                               inode->i_generation);
1606                                         ll_sai_unplug(sai, entry);
1607                                         return -ESTALE;
1608                                 } else {
1609                                         iput(inode);
1610                                 }
1611                                 entry->se_inode = NULL;
1612
1613                                 if ((bits & MDS_INODELOCK_LOOKUP) &&
1614                                     d_lustre_invalid(*dentryp))
1615                                         d_lustre_revalidate(*dentryp);
1616                                 ll_intent_release(&it);
1617                         }
1618                 }
1619
1620                 ll_sai_unplug(sai, entry);
1621                 return rc;
1622         }
1623
1624         /* I am the "lli_opendir_pid" owner, only me can set "lli_sai". */
1625         rc = is_first_dirent(dir, *dentryp);
1626         if (rc == LS_NONE_FIRST_DE) {
1627                 /* It is not "ls -{a}l" operation, no need statahead for it. */
1628                 rc = -EAGAIN;
1629                 goto out;
1630         }
1631
1632         sai = ll_sai_alloc();
1633         if (sai == NULL) {
1634                 rc = -ENOMEM;
1635                 goto out;
1636         }
1637
1638         sai->sai_ls_all = (rc == LS_FIRST_DOT_DE);
1639         sai->sai_inode = igrab(dir);
1640         if (unlikely(sai->sai_inode == NULL)) {
1641                 CWARN("Do not start stat ahead on dying inode "DFID"\n",
1642                       PFID(&lli->lli_fid));
1643                 rc = -ESTALE;
1644                 goto out;
1645         }
1646
1647         /* get parent reference count here, and put it in ll_statahead_thread */
1648         parent = dget((*dentryp)->d_parent);
1649         if (unlikely(sai->sai_inode != d_inode(parent))) {
1650                 struct ll_inode_info *nlli = ll_i2info(d_inode(parent));
1651
1652                 CWARN("Race condition, someone changed %pd just now: old parent "DFID", new parent "DFID"\n",
1653                       *dentryp,
1654                       PFID(&lli->lli_fid), PFID(&nlli->lli_fid));
1655                 dput(parent);
1656                 iput(sai->sai_inode);
1657                 rc = -EAGAIN;
1658                 goto out;
1659         }
1660
1661         CDEBUG(D_READA, "start statahead thread: sai %p, parent %pd\n",
1662                sai, parent);
1663
1664         /* The sai buffer already has one reference taken at allocation time,
1665          * but as soon as we expose the sai by attaching it to the lli that
1666          * default reference can be dropped by another thread calling
1667          * ll_stop_statahead. We need to take a local reference to protect
1668          * the sai buffer while we intend to access it. */
1669         ll_sai_get(sai);
1670         lli->lli_sai = sai;
1671
1672         plli = ll_i2info(d_inode(parent));
1673         rc = PTR_ERR(kthread_run(ll_statahead_thread, parent,
1674                                  "ll_sa_%u", plli->lli_opendir_pid));
1675         thread = &sai->sai_thread;
1676         if (IS_ERR_VALUE(rc)) {
1677                 CERROR("can't start ll_sa thread, rc: %d\n", rc);
1678                 dput(parent);
1679                 lli->lli_opendir_key = NULL;
1680                 thread_set_flags(thread, SVC_STOPPED);
1681                 thread_set_flags(&sai->sai_agl_thread, SVC_STOPPED);
1682                 /* Drop both our own local reference and the default
1683                  * reference from allocation time. */
1684                 ll_sai_put(sai);
1685                 ll_sai_put(sai);
1686                 LASSERT(lli->lli_sai == NULL);
1687                 return -EAGAIN;
1688         }
1689
1690         l_wait_event(thread->t_ctl_waitq,
1691                      thread_is_running(thread) || thread_is_stopped(thread),
1692                      &lwi);
1693         ll_sai_put(sai);
1694
1695         /*
1696          * We don't stat-ahead for the first dirent since we are already in
1697          * lookup.
1698          */
1699         return -EAGAIN;
1700
1701 out:
1702         kfree(sai);
1703         spin_lock(&lli->lli_sa_lock);
1704         lli->lli_opendir_key = NULL;
1705         lli->lli_opendir_pid = 0;
1706         spin_unlock(&lli->lli_sa_lock);
1707         return rc;
1708 }