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                 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         __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                 OBD_FREE_PTR(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         capa_put(minfo->mi_data.op_capa1);
 788         capa_put(minfo->mi_data.op_capa2);
 789         OBD_FREE_PTR(minfo);
 790         OBD_FREE_PTR(einfo);
 791 }
 792
 793 /**
 794  * There is race condition between "capa_put" and "ll_statahead_interpret" for
 795  * accessing "op_data.op_capa[1,2]" as following:
 796  * "capa_put" releases "op_data.op_capa[1,2]"'s reference count after calling
 797  * "md_intent_getattr_async". But "ll_statahead_interpret" maybe run first, and
 798  * fill "op_data.op_capa[1,2]" as POISON, then cause "capa_put" access invalid
 799  * "ocapa". So here reserve "op_data.op_capa[1,2]" in "pcapa" before calling
 800  * "md_intent_getattr_async".
 801  */
 802 static int sa_args_init(struct inode *dir, struct inode *child,
 803                         struct ll_sa_entry *entry, struct md_enqueue_info **pmi,
 804                         struct ldlm_enqueue_info **pei,
 805                         struct obd_capa **pcapa)
 806 {
 807         struct qstr           *qstr = &entry->se_qstr;
 808         struct ll_inode_info     *lli  = ll_i2info(dir);
 809         struct md_enqueue_info   *minfo;
 810         struct ldlm_enqueue_info *einfo;
 811         struct md_op_data       *op_data;
 812
 813         einfo = kzalloc(sizeof(*einfo), GFP_NOFS);
 814         if (!einfo)
 815                 return -ENOMEM;
 816
 817         minfo = kzalloc(sizeof(*minfo), GFP_NOFS);
 818         if (!minfo) {
 819                 OBD_FREE_PTR(einfo);
 820                 return -ENOMEM;
 821         }
 822
 823         op_data = ll_prep_md_op_data(&minfo->mi_data, dir, child, qstr->name,
 824                                      qstr->len, 0, LUSTRE_OPC_ANY, NULL);
 825         if (IS_ERR(op_data)) {
 826                 OBD_FREE_PTR(einfo);
 827                 OBD_FREE_PTR(minfo);
 828                 return PTR_ERR(op_data);
 829         }
 830
 831         minfo->mi_it.it_op = IT_GETATTR;
 832         minfo->mi_dir = igrab(dir);
 833         minfo->mi_cb = ll_statahead_interpret;
 834         minfo->mi_generation = lli->lli_sai->sai_generation;
 835         minfo->mi_cbdata = entry->se_index;
 836
 837         einfo->ei_type   = LDLM_IBITS;
 838         einfo->ei_mode   = it_to_lock_mode(&minfo->mi_it);
 839         einfo->ei_cb_bl  = ll_md_blocking_ast;
 840         einfo->ei_cb_cp  = ldlm_completion_ast;
 841         einfo->ei_cb_gl  = NULL;
 842         einfo->ei_cbdata = NULL;
 843
 844         *pmi = minfo;
 845         *pei = einfo;
 846         pcapa[0] = op_data->op_capa1;
 847         pcapa[1] = op_data->op_capa2;
 848
 849         return 0;
 850 }
 851
 852 static int do_sa_lookup(struct inode *dir, struct ll_sa_entry *entry)
 853 {
 854         struct md_enqueue_info   *minfo;
 855         struct ldlm_enqueue_info *einfo;
 856         struct obd_capa   *capas[2];
 857         int                    rc;
 858
 859         rc = sa_args_init(dir, NULL, entry, &minfo, &einfo, capas);
 860         if (rc)
 861                 return rc;
 862
 863         rc = md_intent_getattr_async(ll_i2mdexp(dir), minfo, einfo);
 864         if (!rc) {
 865                 capa_put(capas[0]);
 866                 capa_put(capas[1]);
 867         } else {
 868                 sa_args_fini(minfo, einfo);
 869         }
 870
 871         return rc;
 872 }
 873
 874 /**
 875  * similar to ll_revalidate_it().
 876  * \retval      1 -- dentry valid
 877  * \retval      0 -- will send stat-ahead request
 878  * \retval others -- prepare stat-ahead request failed
 879  */
 880 static int do_sa_revalidate(struct inode *dir, struct ll_sa_entry *entry,
 881                             struct dentry *dentry)
 882 {
 883         struct inode         *inode = d_inode(dentry);
 884         struct lookup_intent      it = { .it_op = IT_GETATTR,
 885                                          .d.lustre.it_lock_handle = 0 };
 886         struct md_enqueue_info   *minfo;
 887         struct ldlm_enqueue_info *einfo;
 888         struct obd_capa   *capas[2];
 889         int rc;
 890
 891         if (unlikely(inode == NULL))
 892                 return 1;
 893
 894         if (d_mountpoint(dentry))
 895                 return 1;
 896
 897         entry->se_inode = igrab(inode);
 898         rc = md_revalidate_lock(ll_i2mdexp(dir), &it, ll_inode2fid(inode),
 899                                 NULL);
 900         if (rc == 1) {
 901                 entry->se_handle = it.d.lustre.it_lock_handle;
 902                 ll_intent_release(&it);
 903                 return 1;
 904         }
 905
 906         rc = sa_args_init(dir, inode, entry, &minfo, &einfo, capas);
 907         if (rc) {
 908                 entry->se_inode = NULL;
 909                 iput(inode);
 910                 return rc;
 911         }
 912
 913         rc = md_intent_getattr_async(ll_i2mdexp(dir), minfo, einfo);
 914         if (!rc) {
 915                 capa_put(capas[0]);
 916                 capa_put(capas[1]);
 917         } else {
 918                 entry->se_inode = NULL;
 919                 iput(inode);
 920                 sa_args_fini(minfo, einfo);
 921         }
 922
 923         return rc;
 924 }
 925
 926 static void ll_statahead_one(struct dentry *parent, const char *entry_name,
 927                              int entry_name_len)
 928 {
 929         struct inode         *dir    = d_inode(parent);
 930         struct ll_inode_info     *lli    = ll_i2info(dir);
 931         struct ll_statahead_info *sai    = lli->lli_sai;
 932         struct dentry       *dentry = NULL;
 933         struct ll_sa_entry       *entry;
 934         int                    rc;
 935         int                    rc1;
 936
 937         entry = ll_sa_entry_alloc(sai, sai->sai_index, entry_name,
 938                                   entry_name_len);
 939         if (IS_ERR(entry))
 940                 return;
 941
 942         dentry = d_lookup(parent, &entry->se_qstr);
 943         if (!dentry) {
 944                 rc = do_sa_lookup(dir, entry);
 945         } else {
 946                 rc = do_sa_revalidate(dir, entry, dentry);
 947                 if (rc == 1 && agl_should_run(sai, d_inode(dentry)))
 948                         ll_agl_add(sai, d_inode(dentry), entry->se_index);
 949         }
 950
 951         if (dentry != NULL)
 952                 dput(dentry);
 953
 954         if (rc) {
 955                 rc1 = ll_sa_entry_to_stated(sai, entry,
 956                                         rc < 0 ? SA_ENTRY_INVA : SA_ENTRY_SUCC);
 957                 if (rc1 == 0 && entry->se_index == sai->sai_index_wait)
 958                         wake_up(&sai->sai_waitq);
 959         } else {
 960                 sai->sai_sent++;
 961         }
 962
 963         sai->sai_index++;
 964         /* drop one refcount on entry by ll_sa_entry_alloc */
 965         ll_sa_entry_put(sai, entry);
 966 }
 967
 968 static int ll_agl_thread(void *arg)
 969 {
 970         struct dentry       *parent = (struct dentry *)arg;
 971         struct inode         *dir    = d_inode(parent);
 972         struct ll_inode_info     *plli   = ll_i2info(dir);
 973         struct ll_inode_info     *clli;
 974         struct ll_sb_info       *sbi    = ll_i2sbi(dir);
 975         struct ll_statahead_info *sai    = ll_sai_get(plli->lli_sai);
 976         struct ptlrpc_thread     *thread = &sai->sai_agl_thread;
 977         struct l_wait_info      lwi    = { 0 };
 978
 979         thread->t_pid = current_pid();
 980         CDEBUG(D_READA, "agl thread started: sai %p, parent %pd\n",
 981                sai, parent);
 982
 983         atomic_inc(&sbi->ll_agl_total);
 984         spin_lock(&plli->lli_agl_lock);
 985         sai->sai_agl_valid = 1;
 986         if (thread_is_init(thread))
 987                 /* If someone else has changed the thread state
 988                  * (e.g. already changed to SVC_STOPPING), we can't just
 989                  * blindly overwrite that setting. */
 990                 thread_set_flags(thread, SVC_RUNNING);
 991         spin_unlock(&plli->lli_agl_lock);
 992         wake_up(&thread->t_ctl_waitq);
 993
 994         while (1) {
 995                 l_wait_event(thread->t_ctl_waitq,
 996                              !agl_list_empty(sai) ||
 997                              !thread_is_running(thread),
 998                              &lwi);
 999
1000                 if (!thread_is_running(thread))
1001                         break;
1002
1003                 spin_lock(&plli->lli_agl_lock);
1004                 /* The statahead thread maybe help to process AGL entries,
1005                  * so check whether list empty again. */
1006                 if (!agl_list_empty(sai)) {
1007                         clli = agl_first_entry(sai);
1008                         list_del_init(&clli->lli_agl_list);
1009                         spin_unlock(&plli->lli_agl_lock);
1010                         ll_agl_trigger(&clli->lli_vfs_inode, sai);
1011                 } else {
1012                         spin_unlock(&plli->lli_agl_lock);
1013                 }
1014         }
1015
1016         spin_lock(&plli->lli_agl_lock);
1017         sai->sai_agl_valid = 0;
1018         while (!agl_list_empty(sai)) {
1019                 clli = agl_first_entry(sai);
1020                 list_del_init(&clli->lli_agl_list);
1021                 spin_unlock(&plli->lli_agl_lock);
1022                 clli->lli_agl_index = 0;
1023                 iput(&clli->lli_vfs_inode);
1024                 spin_lock(&plli->lli_agl_lock);
1025         }
1026         thread_set_flags(thread, SVC_STOPPED);
1027         spin_unlock(&plli->lli_agl_lock);
1028         wake_up(&thread->t_ctl_waitq);
1029         ll_sai_put(sai);
1030         CDEBUG(D_READA, "agl thread stopped: sai %p, parent %pd\n",
1031                sai, parent);
1032         return 0;
1033 }
1034
1035 static void ll_start_agl(struct dentry *parent, struct ll_statahead_info *sai)
1036 {
1037         struct ptlrpc_thread *thread = &sai->sai_agl_thread;
1038         struct l_wait_info    lwi    = { 0 };
1039         struct ll_inode_info  *plli;
1040         struct task_struct *task;
1041
1042         CDEBUG(D_READA, "start agl thread: sai %p, parent %pd\n",
1043                sai, parent);
1044
1045         plli = ll_i2info(d_inode(parent));
1046         task = kthread_run(ll_agl_thread, parent,
1047                                "ll_agl_%u", plli->lli_opendir_pid);
1048         if (IS_ERR(task)) {
1049                 CERROR("can't start ll_agl thread, rc: %ld\n", PTR_ERR(task));
1050                 thread_set_flags(thread, SVC_STOPPED);
1051                 return;
1052         }
1053
1054         l_wait_event(thread->t_ctl_waitq,
1055                      thread_is_running(thread) || thread_is_stopped(thread),
1056                      &lwi);
1057 }
1058
1059 static int ll_statahead_thread(void *arg)
1060 {
1061         struct dentry       *parent = (struct dentry *)arg;
1062         struct inode         *dir    = d_inode(parent);
1063         struct ll_inode_info     *plli   = ll_i2info(dir);
1064         struct ll_inode_info     *clli;
1065         struct ll_sb_info       *sbi    = ll_i2sbi(dir);
1066         struct ll_statahead_info *sai    = ll_sai_get(plli->lli_sai);
1067         struct ptlrpc_thread     *thread = &sai->sai_thread;
1068         struct ptlrpc_thread *agl_thread = &sai->sai_agl_thread;
1069         struct page           *page;
1070         __u64                pos    = 0;
1071         int                    first  = 0;
1072         int                    rc     = 0;
1073         struct ll_dir_chain       chain;
1074         struct l_wait_info      lwi    = { 0 };
1075
1076         thread->t_pid = current_pid();
1077         CDEBUG(D_READA, "statahead thread starting: sai %p, parent %pd\n",
1078                sai, parent);
1079
1080         if (sbi->ll_flags & LL_SBI_AGL_ENABLED)
1081                 ll_start_agl(parent, sai);
1082
1083         atomic_inc(&sbi->ll_sa_total);
1084         spin_lock(&plli->lli_sa_lock);
1085         if (thread_is_init(thread))
1086                 /* If someone else has changed the thread state
1087                  * (e.g. already changed to SVC_STOPPING), we can't just
1088                  * blindly overwrite that setting. */
1089                 thread_set_flags(thread, SVC_RUNNING);
1090         spin_unlock(&plli->lli_sa_lock);
1091         wake_up(&thread->t_ctl_waitq);
1092
1093         ll_dir_chain_init(&chain);
1094         page = ll_get_dir_page(dir, pos, &chain);
1095
1096         while (1) {
1097                 struct lu_dirpage *dp;
1098                 struct lu_dirent  *ent;
1099
1100                 if (IS_ERR(page)) {
1101                         rc = PTR_ERR(page);
1102                         CDEBUG(D_READA, "error reading dir "DFID" at %llu/%llu: [rc %d] [parent %u]\n",
1103                                PFID(ll_inode2fid(dir)), pos, sai->sai_index,
1104                                rc, plli->lli_opendir_pid);
1105                         goto out;
1106                 }
1107
1108                 dp = page_address(page);
1109                 for (ent = lu_dirent_start(dp); ent != NULL;
1110                      ent = lu_dirent_next(ent)) {
1111                         __u64 hash;
1112                         int namelen;
1113                         char *name;
1114
1115                         hash = le64_to_cpu(ent->lde_hash);
1116                         if (unlikely(hash < pos))
1117                                 /*
1118                                  * Skip until we find target hash value.
1119                                  */
1120                                 continue;
1121
1122                         namelen = le16_to_cpu(ent->lde_namelen);
1123                         if (unlikely(namelen == 0))
1124                                 /*
1125                                  * Skip dummy record.
1126                                  */
1127                                 continue;
1128
1129                         name = ent->lde_name;
1130                         if (name[0] == '.') {
1131                                 if (namelen == 1) {
1132                                         /*
1133                                          * skip "."
1134                                          */
1135                                         continue;
1136                                 } else if (name[1] == '.' && namelen == 2) {
1137                                         /*
1138                                          * skip ".."
1139                                          */
1140                                         continue;
1141                                 } else if (!sai->sai_ls_all) {
1142                                         /*
1143                                          * skip hidden files.
1144                                          */
1145                                         sai->sai_skip_hidden++;
1146                                         continue;
1147                                 }
1148                         }
1149
1150                         /*
1151                          * don't stat-ahead first entry.
1152                          */
1153                         if (unlikely(++first == 1))
1154                                 continue;
1155
1156 keep_it:
1157                         l_wait_event(thread->t_ctl_waitq,
1158                                      !sa_sent_full(sai) ||
1159                                      !sa_received_empty(sai) ||
1160                                      !agl_list_empty(sai) ||
1161                                      !thread_is_running(thread),
1162                                      &lwi);
1163
1164 interpret_it:
1165                         while (!sa_received_empty(sai))
1166                                 ll_post_statahead(sai);
1167
1168                         if (unlikely(!thread_is_running(thread))) {
1169                                 ll_release_page(page, 0);
1170                                 rc = 0;
1171                                 goto out;
1172                         }
1173
1174                         /* If no window for metadata statahead, but there are
1175                          * some AGL entries to be triggered, then try to help
1176                          * to process the AGL entries. */
1177                         if (sa_sent_full(sai)) {
1178                                 spin_lock(&plli->lli_agl_lock);
1179                                 while (!agl_list_empty(sai)) {
1180                                         clli = agl_first_entry(sai);
1181                                         list_del_init(&clli->lli_agl_list);
1182                                         spin_unlock(&plli->lli_agl_lock);
1183                                         ll_agl_trigger(&clli->lli_vfs_inode,
1184                                                        sai);
1185
1186                                         if (!sa_received_empty(sai))
1187                                                 goto interpret_it;
1188
1189                                         if (unlikely(
1190                                                 !thread_is_running(thread))) {
1191                                                 ll_release_page(page, 0);
1192                                                 rc = 0;
1193                                                 goto out;
1194                                         }
1195
1196                                         if (!sa_sent_full(sai))
1197                                                 goto do_it;
1198
1199                                         spin_lock(&plli->lli_agl_lock);
1200                                 }
1201                                 spin_unlock(&plli->lli_agl_lock);
1202
1203                                 goto keep_it;
1204                         }
1205
1206 do_it:
1207                         ll_statahead_one(parent, name, namelen);
1208                 }
1209                 pos = le64_to_cpu(dp->ldp_hash_end);
1210                 if (pos == MDS_DIR_END_OFF) {
1211                         /*
1212                          * End of directory reached.
1213                          */
1214                         ll_release_page(page, 0);
1215                         while (1) {
1216                                 l_wait_event(thread->t_ctl_waitq,
1217                                              !sa_received_empty(sai) ||
1218                                              sai->sai_sent == sai->sai_replied||
1219                                              !thread_is_running(thread),
1220                                              &lwi);
1221
1222                                 while (!sa_received_empty(sai))
1223                                         ll_post_statahead(sai);
1224
1225                                 if (unlikely(!thread_is_running(thread))) {
1226                                         rc = 0;
1227                                         goto out;
1228                                 }
1229
1230                                 if (sai->sai_sent == sai->sai_replied &&
1231                                     sa_received_empty(sai))
1232                                         break;
1233                         }
1234
1235                         spin_lock(&plli->lli_agl_lock);
1236                         while (!agl_list_empty(sai) &&
1237                                thread_is_running(thread)) {
1238                                 clli = agl_first_entry(sai);
1239                                 list_del_init(&clli->lli_agl_list);
1240                                 spin_unlock(&plli->lli_agl_lock);
1241                                 ll_agl_trigger(&clli->lli_vfs_inode, sai);
1242                                 spin_lock(&plli->lli_agl_lock);
1243                         }
1244                         spin_unlock(&plli->lli_agl_lock);
1245
1246                         rc = 0;
1247                         goto out;
1248                 } else if (1) {
1249                         /*
1250                          * chain is exhausted.
1251                          * Normal case: continue to the next page.
1252                          */
1253                         ll_release_page(page, le32_to_cpu(dp->ldp_flags) &
1254                                               LDF_COLLIDE);
1255                         page = ll_get_dir_page(dir, pos, &chain);
1256                 } else {
1257                         LASSERT(le32_to_cpu(dp->ldp_flags) & LDF_COLLIDE);
1258                         ll_release_page(page, 1);
1259                         /*
1260                          * go into overflow page.
1261                          */
1262                 }
1263         }
1264
1265 out:
1266         if (sai->sai_agl_valid) {
1267                 spin_lock(&plli->lli_agl_lock);
1268                 thread_set_flags(agl_thread, SVC_STOPPING);
1269                 spin_unlock(&plli->lli_agl_lock);
1270                 wake_up(&agl_thread->t_ctl_waitq);
1271
1272                 CDEBUG(D_READA, "stop agl thread: sai %p pid %u\n",
1273                        sai, (unsigned int)agl_thread->t_pid);
1274                 l_wait_event(agl_thread->t_ctl_waitq,
1275                              thread_is_stopped(agl_thread),
1276                              &lwi);
1277         } else {
1278                 /* Set agl_thread flags anyway. */
1279                 thread_set_flags(&sai->sai_agl_thread, SVC_STOPPED);
1280         }
1281         ll_dir_chain_fini(&chain);
1282         spin_lock(&plli->lli_sa_lock);
1283         if (!sa_received_empty(sai)) {
1284                 thread_set_flags(thread, SVC_STOPPING);
1285                 spin_unlock(&plli->lli_sa_lock);
1286
1287                 /* To release the resources held by received entries. */
1288                 while (!sa_received_empty(sai))
1289                         ll_post_statahead(sai);
1290
1291                 spin_lock(&plli->lli_sa_lock);
1292         }
1293         thread_set_flags(thread, SVC_STOPPED);
1294         spin_unlock(&plli->lli_sa_lock);
1295         wake_up(&sai->sai_waitq);
1296         wake_up(&thread->t_ctl_waitq);
1297         ll_sai_put(sai);
1298         dput(parent);
1299         CDEBUG(D_READA, "statahead thread stopped: sai %p, parent %pd\n",
1300                sai, parent);
1301         return rc;
1302 }
1303
1304 /**
1305  * called in ll_file_release().
1306  */
1307 void ll_stop_statahead(struct inode *dir, void *key)
1308 {
1309         struct ll_inode_info *lli = ll_i2info(dir);
1310
1311         if (unlikely(key == NULL))
1312                 return;
1313
1314         spin_lock(&lli->lli_sa_lock);
1315         if (lli->lli_opendir_key != key || lli->lli_opendir_pid == 0) {
1316                 spin_unlock(&lli->lli_sa_lock);
1317                 return;
1318         }
1319
1320         lli->lli_opendir_key = NULL;
1321
1322         if (lli->lli_sai) {
1323                 struct l_wait_info lwi = { 0 };
1324                 struct ptlrpc_thread *thread = &lli->lli_sai->sai_thread;
1325
1326                 if (!thread_is_stopped(thread)) {
1327                         thread_set_flags(thread, SVC_STOPPING);
1328                         spin_unlock(&lli->lli_sa_lock);
1329                         wake_up(&thread->t_ctl_waitq);
1330
1331                         CDEBUG(D_READA, "stop statahead thread: sai %p pid %u\n",
1332                                lli->lli_sai, (unsigned int)thread->t_pid);
1333                         l_wait_event(thread->t_ctl_waitq,
1334                                      thread_is_stopped(thread),
1335                                      &lwi);
1336                 } else {
1337                         spin_unlock(&lli->lli_sa_lock);
1338                 }
1339
1340                 /*
1341                  * Put the ref which was held when first statahead_enter.
1342                  * It maybe not the last ref for some statahead requests
1343                  * maybe inflight.
1344                  */
1345                 ll_sai_put(lli->lli_sai);
1346         } else {
1347                 lli->lli_opendir_pid = 0;
1348                 spin_unlock(&lli->lli_sa_lock);
1349         }
1350 }
1351
1352 enum {
1353         /**
1354          * not first dirent, or is "."
1355          */
1356         LS_NONE_FIRST_DE = 0,
1357         /**
1358          * the first non-hidden dirent
1359          */
1360         LS_FIRST_DE,
1361         /**
1362          * the first hidden dirent, that is "."
1363          */
1364         LS_FIRST_DOT_DE
1365 };
1366
1367 static int is_first_dirent(struct inode *dir, struct dentry *dentry)
1368 {
1369         struct ll_dir_chain   chain;
1370         struct qstr       *target = &dentry->d_name;
1371         struct page       *page;
1372         __u64            pos    = 0;
1373         int                dot_de;
1374         int                rc     = LS_NONE_FIRST_DE;
1375
1376         ll_dir_chain_init(&chain);
1377         page = ll_get_dir_page(dir, pos, &chain);
1378
1379         while (1) {
1380                 struct lu_dirpage *dp;
1381                 struct lu_dirent  *ent;
1382
1383                 if (IS_ERR(page)) {
1384                         struct ll_inode_info *lli = ll_i2info(dir);
1385
1386                         rc = PTR_ERR(page);
1387                         CERROR("error reading dir "DFID" at %llu: [rc %d] [parent %u]\n",
1388                                PFID(ll_inode2fid(dir)), pos,
1389                                rc, lli->lli_opendir_pid);
1390                         break;
1391                 }
1392
1393                 dp = page_address(page);
1394                 for (ent = lu_dirent_start(dp); ent != NULL;
1395                      ent = lu_dirent_next(ent)) {
1396                         __u64 hash;
1397                         int namelen;
1398                         char *name;
1399
1400                         hash = le64_to_cpu(ent->lde_hash);
1401                         /* The ll_get_dir_page() can return any page containing
1402                          * the given hash which may be not the start hash. */
1403                         if (unlikely(hash < pos))
1404                                 continue;
1405
1406                         namelen = le16_to_cpu(ent->lde_namelen);
1407                         if (unlikely(namelen == 0))
1408                                 /*
1409                                  * skip dummy record.
1410                                  */
1411                                 continue;
1412
1413                         name = ent->lde_name;
1414                         if (name[0] == '.') {
1415                                 if (namelen == 1)
1416                                         /*
1417                                          * skip "."
1418                                          */
1419                                         continue;
1420                                 else if (name[1] == '.' && namelen == 2)
1421                                         /*
1422                                          * skip ".."
1423                                          */
1424                                         continue;
1425                                 else
1426                                         dot_de = 1;
1427                         } else {
1428                                 dot_de = 0;
1429                         }
1430
1431                         if (dot_de && target->name[0] != '.') {
1432                                 CDEBUG(D_READA, "%.*s skip hidden file %.*s\n",
1433                                        target->len, target->name,
1434                                        namelen, name);
1435                                 continue;
1436                         }
1437
1438                         if (target->len != namelen ||
1439                             memcmp(target->name, name, namelen) != 0)
1440                                 rc = LS_NONE_FIRST_DE;
1441                         else if (!dot_de)
1442                                 rc = LS_FIRST_DE;
1443                         else
1444                                 rc = LS_FIRST_DOT_DE;
1445
1446                         ll_release_page(page, 0);
1447                         goto out;
1448                 }
1449                 pos = le64_to_cpu(dp->ldp_hash_end);
1450                 if (pos == MDS_DIR_END_OFF) {
1451                         /*
1452                          * End of directory reached.
1453                          */
1454                         ll_release_page(page, 0);
1455                         break;
1456                 } else if (1) {
1457                         /*
1458                          * chain is exhausted
1459                          * Normal case: continue to the next page.
1460                          */
1461                         ll_release_page(page, le32_to_cpu(dp->ldp_flags) &
1462                                               LDF_COLLIDE);
1463                         page = ll_get_dir_page(dir, pos, &chain);
1464                 } else {
1465                         /*
1466                          * go into overflow page.
1467                          */
1468                         LASSERT(le32_to_cpu(dp->ldp_flags) & LDF_COLLIDE);
1469                         ll_release_page(page, 1);
1470                 }
1471         }
1472
1473 out:
1474         ll_dir_chain_fini(&chain);
1475         return rc;
1476 }
1477
1478 static void
1479 ll_sai_unplug(struct ll_statahead_info *sai, struct ll_sa_entry *entry)
1480 {
1481         struct ptlrpc_thread *thread = &sai->sai_thread;
1482         struct ll_sb_info    *sbi    = ll_i2sbi(sai->sai_inode);
1483         int                hit;
1484
1485         if (entry != NULL && entry->se_stat == SA_ENTRY_SUCC)
1486                 hit = 1;
1487         else
1488                 hit = 0;
1489
1490         ll_sa_entry_fini(sai, entry);
1491         if (hit) {
1492                 sai->sai_hit++;
1493                 sai->sai_consecutive_miss = 0;
1494                 sai->sai_max = min(2 * sai->sai_max, sbi->ll_sa_max);
1495         } else {
1496                 struct ll_inode_info *lli = ll_i2info(sai->sai_inode);
1497
1498                 sai->sai_miss++;
1499                 sai->sai_consecutive_miss++;
1500                 if (sa_low_hit(sai) && thread_is_running(thread)) {
1501                         atomic_inc(&sbi->ll_sa_wrong);
1502                         CDEBUG(D_READA, "Statahead for dir " DFID " hit ratio too low: hit/miss %llu/%llu, sent/replied %llu/%llu, stopping statahead thread\n",
1503                                PFID(&lli->lli_fid), sai->sai_hit,
1504                                sai->sai_miss, sai->sai_sent,
1505                                sai->sai_replied);
1506                         spin_lock(&lli->lli_sa_lock);
1507                         if (!thread_is_stopped(thread))
1508                                 thread_set_flags(thread, SVC_STOPPING);
1509                         spin_unlock(&lli->lli_sa_lock);
1510                 }
1511         }
1512
1513         if (!thread_is_stopped(thread))
1514                 wake_up(&thread->t_ctl_waitq);
1515 }
1516
1517 /**
1518  * Start statahead thread if this is the first dir entry.
1519  * Otherwise if a thread is started already, wait it until it is ahead of me.
1520  * \retval 1       -- find entry with lock in cache, the caller needs to do
1521  *                  nothing.
1522  * \retval 0       -- find entry in cache, but without lock, the caller needs
1523  *                  refresh from MDS.
1524  * \retval others  -- the caller need to process as non-statahead.
1525  */
1526 int do_statahead_enter(struct inode *dir, struct dentry **dentryp,
1527                        int only_unplug)
1528 {
1529         struct ll_inode_info     *lli   = ll_i2info(dir);
1530         struct ll_statahead_info *sai   = lli->lli_sai;
1531         struct dentry       *parent;
1532         struct ll_sa_entry       *entry;
1533         struct ptlrpc_thread     *thread;
1534         struct l_wait_info      lwi   = { 0 };
1535         int                    rc    = 0;
1536         struct ll_inode_info     *plli;
1537
1538         LASSERT(lli->lli_opendir_pid == current_pid());
1539
1540         if (sai) {
1541                 thread = &sai->sai_thread;
1542                 if (unlikely(thread_is_stopped(thread) &&
1543                              list_empty(&sai->sai_entries_stated))) {
1544                         /* to release resource */
1545                         ll_stop_statahead(dir, lli->lli_opendir_key);
1546                         return -EAGAIN;
1547                 }
1548
1549                 if ((*dentryp)->d_name.name[0] == '.') {
1550                         if (sai->sai_ls_all ||
1551                             sai->sai_miss_hidden >= sai->sai_skip_hidden) {
1552                                 /*
1553                                  * Hidden dentry is the first one, or statahead
1554                                  * thread does not skip so many hidden dentries
1555                                  * before "sai_ls_all" enabled as below.
1556                                  */
1557                         } else {
1558                                 if (!sai->sai_ls_all)
1559                                         /*
1560                                          * It maybe because hidden dentry is not
1561                                          * the first one, "sai_ls_all" was not
1562                                          * set, then "ls -al" missed. Enable
1563                                          * "sai_ls_all" for such case.
1564                                          */
1565                                         sai->sai_ls_all = 1;
1566
1567                                 /*
1568                                  * Such "getattr" has been skipped before
1569                                  * "sai_ls_all" enabled as above.
1570                                  */
1571                                 sai->sai_miss_hidden++;
1572                                 return -EAGAIN;
1573                         }
1574                 }
1575
1576                 entry = ll_sa_entry_get_byname(sai, &(*dentryp)->d_name);
1577                 if (entry == NULL || only_unplug) {
1578                         ll_sai_unplug(sai, entry);
1579                         return entry ? 1 : -EAGAIN;
1580                 }
1581
1582                 if (!ll_sa_entry_stated(entry)) {
1583                         sai->sai_index_wait = entry->se_index;
1584                         lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(30), NULL,
1585                                                LWI_ON_SIGNAL_NOOP, NULL);
1586                         rc = l_wait_event(sai->sai_waitq,
1587                                           ll_sa_entry_stated(entry) ||
1588                                           thread_is_stopped(thread),
1589                                           &lwi);
1590                         if (rc < 0) {
1591                                 ll_sai_unplug(sai, entry);
1592                                 return -EAGAIN;
1593                         }
1594                 }
1595
1596                 if (entry->se_stat == SA_ENTRY_SUCC &&
1597                     entry->se_inode != NULL) {
1598                         struct inode *inode = entry->se_inode;
1599                         struct lookup_intent it = { .it_op = IT_GETATTR,
1600                                                     .d.lustre.it_lock_handle =
1601                                                      entry->se_handle };
1602                         __u64 bits;
1603
1604                         rc = md_revalidate_lock(ll_i2mdexp(dir), &it,
1605                                                 ll_inode2fid(inode), &bits);
1606                         if (rc == 1) {
1607                                 if (d_inode(*dentryp) == NULL) {
1608                                         struct dentry *alias;
1609
1610                                         alias = ll_splice_alias(inode,
1611                                                                    *dentryp);
1612                                         if (IS_ERR(alias)) {
1613                                                 ll_sai_unplug(sai, entry);
1614                                                 return PTR_ERR(alias);
1615                                         }
1616                                         *dentryp = alias;
1617                                 } else if (d_inode(*dentryp) != inode) {
1618                                         /* revalidate, but inode is recreated */
1619                                         CDEBUG(D_READA,
1620                                               "stale dentry %pd inode %lu/%u, statahead inode %lu/%u\n",
1621                                               *dentryp,
1622                                               d_inode(*dentryp)->i_ino,
1623                                               d_inode(*dentryp)->i_generation,
1624                                               inode->i_ino,
1625                                               inode->i_generation);
1626                                         ll_sai_unplug(sai, entry);
1627                                         return -ESTALE;
1628                                 } else {
1629                                         iput(inode);
1630                                 }
1631                                 entry->se_inode = NULL;
1632
1633                                 if ((bits & MDS_INODELOCK_LOOKUP) &&
1634                                     d_lustre_invalid(*dentryp))
1635                                         d_lustre_revalidate(*dentryp);
1636                                 ll_intent_release(&it);
1637                         }
1638                 }
1639
1640                 ll_sai_unplug(sai, entry);
1641                 return rc;
1642         }
1643
1644         /* I am the "lli_opendir_pid" owner, only me can set "lli_sai". */
1645         rc = is_first_dirent(dir, *dentryp);
1646         if (rc == LS_NONE_FIRST_DE) {
1647                 /* It is not "ls -{a}l" operation, no need statahead for it. */
1648                 rc = -EAGAIN;
1649                 goto out;
1650         }
1651
1652         sai = ll_sai_alloc();
1653         if (sai == NULL) {
1654                 rc = -ENOMEM;
1655                 goto out;
1656         }
1657
1658         sai->sai_ls_all = (rc == LS_FIRST_DOT_DE);
1659         sai->sai_inode = igrab(dir);
1660         if (unlikely(sai->sai_inode == NULL)) {
1661                 CWARN("Do not start stat ahead on dying inode "DFID"\n",
1662                       PFID(&lli->lli_fid));
1663                 rc = -ESTALE;
1664                 goto out;
1665         }
1666
1667         /* get parent reference count here, and put it in ll_statahead_thread */
1668         parent = dget((*dentryp)->d_parent);
1669         if (unlikely(sai->sai_inode != d_inode(parent))) {
1670                 struct ll_inode_info *nlli = ll_i2info(d_inode(parent));
1671
1672                 CWARN("Race condition, someone changed %pd just now: old parent "DFID", new parent "DFID"\n",
1673                       *dentryp,
1674                       PFID(&lli->lli_fid), PFID(&nlli->lli_fid));
1675                 dput(parent);
1676                 iput(sai->sai_inode);
1677                 rc = -EAGAIN;
1678                 goto out;
1679         }
1680
1681         CDEBUG(D_READA, "start statahead thread: sai %p, parent %pd\n",
1682                sai, parent);
1683
1684         /* The sai buffer already has one reference taken at allocation time,
1685          * but as soon as we expose the sai by attaching it to the lli that
1686          * default reference can be dropped by another thread calling
1687          * ll_stop_statahead. We need to take a local reference to protect
1688          * the sai buffer while we intend to access it. */
1689         ll_sai_get(sai);
1690         lli->lli_sai = sai;
1691
1692         plli = ll_i2info(d_inode(parent));
1693         rc = PTR_ERR(kthread_run(ll_statahead_thread, parent,
1694                                  "ll_sa_%u", plli->lli_opendir_pid));
1695         thread = &sai->sai_thread;
1696         if (IS_ERR_VALUE(rc)) {
1697                 CERROR("can't start ll_sa thread, rc: %d\n", rc);
1698                 dput(parent);
1699                 lli->lli_opendir_key = NULL;
1700                 thread_set_flags(thread, SVC_STOPPED);
1701                 thread_set_flags(&sai->sai_agl_thread, SVC_STOPPED);
1702                 /* Drop both our own local reference and the default
1703                  * reference from allocation time. */
1704                 ll_sai_put(sai);
1705                 ll_sai_put(sai);
1706                 LASSERT(lli->lli_sai == NULL);
1707                 return -EAGAIN;
1708         }
1709
1710         l_wait_event(thread->t_ctl_waitq,
1711                      thread_is_running(thread) || thread_is_stopped(thread),
1712                      &lwi);
1713         ll_sai_put(sai);
1714
1715         /*
1716          * We don't stat-ahead for the first dirent since we are already in
1717          * lookup.
1718          */
1719         return -EAGAIN;
1720
1721 out:
1722         if (sai != NULL)
1723                 OBD_FREE_PTR(sai);
1724         spin_lock(&lli->lli_sa_lock);
1725         lli->lli_opendir_key = NULL;
1726         lli->lli_opendir_pid = 0;
1727         spin_unlock(&lli->lli_sa_lock);
1728         return rc;
1729 }