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.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2011, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  */
  32
  33 #include <linux/fs.h>
  34 #include <linux/sched.h>
  35 #include <linux/kthread.h>
  36 #include <linux/mm.h>
  37 #include <linux/highmem.h>
  38 #include <linux/pagemap.h>
  39 #include <linux/delay.h>
  40
  41 #define DEBUG_SUBSYSTEM S_LLITE
  42
  43 #include <obd_support.h>
  44 #include <lustre_dlm.h>
  45 #include "llite_internal.h"
  46
  47 #define SA_OMITTED_ENTRY_MAX 8ULL
  48
  49 typedef enum {
  50         /** negative values are for error cases */
  51         SA_ENTRY_INIT = 0,      /** init entry */
  52         SA_ENTRY_SUCC = 1,      /** stat succeed */
  53         SA_ENTRY_INVA = 2,      /** invalid entry */
  54 } se_state_t;
  55
  56 /* sa_entry is not refcounted: statahead thread allocates it and do async stat,
  57  * and in async stat callback ll_statahead_interpret() will add it into
  58  * sai_interim_entries, later statahead thread will call sa_handle_callback() to
  59  * instantiate entry and move it into sai_entries, and then only scanner process
  60  * can access and free it. */
  61 struct sa_entry {
  62         /* link into sai_interim_entries or sai_entries */
  63         struct list_head        se_list;
  64         /* link into sai hash table locally */
  65         struct list_head        se_hash;
  66         /* entry index in the sai */
  67         __u64                   se_index;
  68         /* low layer ldlm lock handle */
  69         __u64                   se_handle;
  70         /* entry status */
  71         se_state_t              se_state;
  72         /* entry size, contains name */
  73         int                     se_size;
  74         /* pointer to async getattr enqueue info */
  75         struct md_enqueue_info *se_minfo;
  76         /* pointer to the async getattr request */
  77         struct ptlrpc_request  *se_req;
  78         /* pointer to the target inode */
  79         struct inode           *se_inode;
  80         /* entry name */
  81         struct qstr             se_qstr;
  82         /* entry fid */
  83         struct lu_fid           se_fid;
  84 };
  85
  86 static unsigned int sai_generation = 0;
  87 static DEFINE_SPINLOCK(sai_generation_lock);
  88
  89 static inline int sa_unhashed(struct sa_entry *entry)
  90 {
  91         return list_empty(&entry->se_hash);
  92 }
  93
  94 /* sa_entry is ready to use */
  95 static inline int sa_ready(struct sa_entry *entry)
  96 {
  97         smp_rmb();
  98         return (entry->se_state != SA_ENTRY_INIT);
  99 }
 100
 101 /* hash value to put in sai_cache */
 102 static inline int sa_hash(int val)
 103 {
 104         return val & LL_SA_CACHE_MASK;
 105 }
 106
 107 /* hash entry into sai_cache */
 108 static inline void
 109 sa_rehash(struct ll_statahead_info *sai, struct sa_entry *entry)
 110 {
 111         int i = sa_hash(entry->se_qstr.hash);
 112
 113         spin_lock(&sai->sai_cache_lock[i]);
 114         list_add_tail(&entry->se_hash, &sai->sai_cache[i]);
 115         spin_unlock(&sai->sai_cache_lock[i]);
 116 }
 117
 118 /* unhash entry from sai_cache */
 119 static inline void
 120 sa_unhash(struct ll_statahead_info *sai, struct sa_entry *entry)
 121 {
 122         int i = sa_hash(entry->se_qstr.hash);
 123
 124         spin_lock(&sai->sai_cache_lock[i]);
 125         list_del_init(&entry->se_hash);
 126         spin_unlock(&sai->sai_cache_lock[i]);
 127 }
 128
 129 static inline int agl_should_run(struct ll_statahead_info *sai,
 130                                  struct inode *inode)
 131 {
 132         return (inode != NULL && S_ISREG(inode->i_mode) && sai->sai_agl_valid);
 133 }
 134
 135 static inline struct ll_inode_info *
 136 agl_first_entry(struct ll_statahead_info *sai)
 137 {
 138         return list_entry(sai->sai_agls.next, struct ll_inode_info,
 139                           lli_agl_list);
 140 }
 141
 142 /* statahead window is full */
 143 static inline int sa_sent_full(struct ll_statahead_info *sai)
 144 {
 145         return atomic_read(&sai->sai_cache_count) >= sai->sai_max;
 146 }
 147
 148 /* got async stat replies */
 149 static inline int sa_has_callback(struct ll_statahead_info *sai)
 150 {
 151         return !list_empty(&sai->sai_interim_entries);
 152 }
 153
 154 static inline int agl_list_empty(struct ll_statahead_info *sai)
 155 {
 156         return list_empty(&sai->sai_agls);
 157 }
 158
 159 /**
 160  * (1) hit ratio less than 80%
 161  * or
 162  * (2) consecutive miss more than 8
 163  * then means low hit.
 164  */
 165 static inline int sa_low_hit(struct ll_statahead_info *sai)
 166 {
 167         return ((sai->sai_hit > 7 && sai->sai_hit < 4 * sai->sai_miss) ||
 168                 (sai->sai_consecutive_miss > 8));
 169 }
 170
 171 /*
 172  * if the given index is behind of statahead window more than
 173  * SA_OMITTED_ENTRY_MAX, then it is old.
 174  */
 175 static inline int is_omitted_entry(struct ll_statahead_info *sai, __u64 index)
 176 {
 177         return ((__u64)sai->sai_max + index + SA_OMITTED_ENTRY_MAX <
 178                  sai->sai_index);
 179 }
 180
 181 /* allocate sa_entry and hash it to allow scanner process to find it */
 182 static struct sa_entry *
 183 sa_alloc(struct dentry *parent, struct ll_statahead_info *sai, __u64 index,
 184          const char *name, int len, const struct lu_fid *fid)
 185 {
 186         struct ll_inode_info *lli;
 187         struct sa_entry *entry;
 188         int entry_size;
 189         char *dname;
 190         ENTRY;
 191
 192         entry_size = sizeof(struct sa_entry) + (len & ~3) + 4;
 193         OBD_ALLOC(entry, entry_size);
 194         if (unlikely(entry == NULL))
 195                 RETURN(ERR_PTR(-ENOMEM));
 196
 197         CDEBUG(D_READA, "alloc sa entry %.*s(%p) index %llu\n",
 198                len, name, entry, index);
 199
 200         entry->se_index = index;
 201
 202         entry->se_state = SA_ENTRY_INIT;
 203         entry->se_size = entry_size;
 204         dname = (char *)entry + sizeof(struct sa_entry);
 205         memcpy(dname, name, len);
 206         dname[len] = 0;
 207         entry->se_qstr.hash = ll_full_name_hash(parent, name, len);
 208         entry->se_qstr.len = len;
 209         entry->se_qstr.name = dname;
 210         entry->se_fid = *fid;
 211
 212         lli = ll_i2info(sai->sai_dentry->d_inode);
 213
 214         spin_lock(&lli->lli_sa_lock);
 215         INIT_LIST_HEAD(&entry->se_list);
 216         sa_rehash(sai, entry);
 217         spin_unlock(&lli->lli_sa_lock);
 218
 219         atomic_inc(&sai->sai_cache_count);
 220
 221         RETURN(entry);
 222 }
 223
 224 /* free sa_entry, which should have been unhashed and not in any list */
 225 static void sa_free(struct ll_statahead_info *sai, struct sa_entry *entry)
 226 {
 227         CDEBUG(D_READA, "free sa entry %.*s(%p) index %llu\n",
 228                entry->se_qstr.len, entry->se_qstr.name, entry,
 229                entry->se_index);
 230
 231         LASSERT(list_empty(&entry->se_list));
 232         LASSERT(sa_unhashed(entry));
 233
 234         OBD_FREE(entry, entry->se_size);
 235         atomic_dec(&sai->sai_cache_count);
 236 }
 237
 238 /*
 239  * find sa_entry by name, used by directory scanner, lock is not needed because
 240  * only scanner can remove the entry from cache.
 241  */
 242 static struct sa_entry *
 243 sa_get(struct ll_statahead_info *sai, const struct qstr *qstr)
 244 {
 245         struct sa_entry *entry;
 246         int i = sa_hash(qstr->hash);
 247
 248         list_for_each_entry(entry, &sai->sai_cache[i], se_hash) {
 249                 if (entry->se_qstr.hash == qstr->hash &&
 250                     entry->se_qstr.len == qstr->len &&
 251                     memcmp(entry->se_qstr.name, qstr->name, qstr->len) == 0)
 252                         return entry;
 253         }
 254         return NULL;
 255 }
 256
 257 /* unhash and unlink sa_entry, and then free it */
 258 static inline void
 259 sa_kill(struct ll_statahead_info *sai, struct sa_entry *entry)
 260 {
 261         struct ll_inode_info *lli = ll_i2info(sai->sai_dentry->d_inode);
 262
 263         LASSERT(!sa_unhashed(entry));
 264         LASSERT(!list_empty(&entry->se_list));
 265         LASSERT(sa_ready(entry));
 266
 267         sa_unhash(sai, entry);
 268
 269         spin_lock(&lli->lli_sa_lock);
 270         list_del_init(&entry->se_list);
 271         spin_unlock(&lli->lli_sa_lock);
 272
 273         if (entry->se_inode != NULL)
 274                 iput(entry->se_inode);
 275
 276         sa_free(sai, entry);
 277 }
 278
 279 /* called by scanner after use, sa_entry will be killed */
 280 static void
 281 sa_put(struct ll_statahead_info *sai, struct sa_entry *entry)
 282 {
 283         struct sa_entry *tmp, *next;
 284
 285         if (entry != NULL && entry->se_state == SA_ENTRY_SUCC) {
 286                 struct ll_sb_info *sbi = ll_i2sbi(sai->sai_dentry->d_inode);
 287
 288                 sai->sai_hit++;
 289                 sai->sai_consecutive_miss = 0;
 290                 sai->sai_max = min(2 * sai->sai_max, sbi->ll_sa_max);
 291         } else {
 292                 sai->sai_miss++;
 293                 sai->sai_consecutive_miss++;
 294         }
 295
 296         if (entry != NULL)
 297                 sa_kill(sai, entry);
 298
 299         /* kill old completed entries, only scanner process does this, no need
 300          * to lock */
 301         list_for_each_entry_safe(tmp, next, &sai->sai_entries, se_list) {
 302                 if (!is_omitted_entry(sai, tmp->se_index))
 303                         break;
 304                 sa_kill(sai, tmp);
 305         }
 306 }
 307
 308 /* update state and sort add entry to sai_entries by index, return true if
 309  * scanner is waiting on this entry. */
 310 static bool
 311 __sa_make_ready(struct ll_statahead_info *sai, struct sa_entry *entry, int ret)
 312 {
 313         struct sa_entry *se;
 314         struct list_head *pos = &sai->sai_entries;
 315         __u64 index = entry->se_index;
 316
 317         LASSERT(!sa_ready(entry));
 318         LASSERT(list_empty(&entry->se_list));
 319
 320         list_for_each_entry_reverse(se, &sai->sai_entries, se_list) {
 321                 if (se->se_index < entry->se_index) {
 322                         pos = &se->se_list;
 323                         break;
 324                 }
 325         }
 326         list_add(&entry->se_list, pos);
 327         /*
 328          * LU-9210: ll_statahead_interpet must be able to see this before
 329          * we wake it up
 330          */
 331         smp_store_release(&entry->se_state, ret < 0 ? SA_ENTRY_INVA : SA_ENTRY_SUCC);
 332
 333         return (index == sai->sai_index_wait);
 334 }
 335
 336 /* finish async stat RPC arguments */
 337 static void sa_fini_data(struct md_enqueue_info *minfo)
 338 {
 339         ll_unlock_md_op_lsm(&minfo->mi_data);
 340         iput(minfo->mi_dir);
 341         OBD_FREE_PTR(minfo);
 342 }
 343
 344 static int ll_statahead_interpret(struct ptlrpc_request *req,
 345                                   struct md_enqueue_info *minfo, int rc);
 346
 347 /*
 348  * prepare arguments for async stat RPC.
 349  */
 350 static struct md_enqueue_info *
 351 sa_prep_data(struct inode *dir, struct inode *child, struct sa_entry *entry)
 352 {
 353         struct md_enqueue_info   *minfo;
 354         struct ldlm_enqueue_info *einfo;
 355         struct md_op_data        *op_data;
 356
 357         OBD_ALLOC_PTR(minfo);
 358         if (minfo == NULL)
 359                 return ERR_PTR(-ENOMEM);
 360
 361         op_data = ll_prep_md_op_data(&minfo->mi_data, dir, child,
 362                                      entry->se_qstr.name, entry->se_qstr.len, 0,
 363                                      LUSTRE_OPC_ANY, NULL);
 364         if (IS_ERR(op_data)) {
 365                 OBD_FREE_PTR(minfo);
 366                 return (struct md_enqueue_info *)op_data;
 367         }
 368
 369         if (child == NULL)
 370                 op_data->op_fid2 = entry->se_fid;
 371
 372         minfo->mi_it.it_op = IT_GETATTR;
 373         minfo->mi_dir = igrab(dir);
 374         minfo->mi_cb = ll_statahead_interpret;
 375         minfo->mi_cbdata = entry;
 376
 377         einfo = &minfo->mi_einfo;
 378         einfo->ei_type   = LDLM_IBITS;
 379         einfo->ei_mode   = it_to_lock_mode(&minfo->mi_it);
 380         einfo->ei_cb_bl  = ll_md_blocking_ast;
 381         einfo->ei_cb_cp  = ldlm_completion_ast;
 382         einfo->ei_cb_gl  = NULL;
 383         einfo->ei_cbdata = NULL;
 384
 385         return minfo;
 386 }
 387
 388 /*
 389  * release resources used in async stat RPC, update entry state and wakeup if
 390  * scanner process it waiting on this entry.
 391  */
 392 static void
 393 sa_make_ready(struct ll_statahead_info *sai, struct sa_entry *entry, int ret)
 394 {
 395         struct ll_inode_info *lli = ll_i2info(sai->sai_dentry->d_inode);
 396         struct md_enqueue_info *minfo = entry->se_minfo;
 397         struct ptlrpc_request *req = entry->se_req;
 398         bool wakeup;
 399
 400         /* release resources used in RPC */
 401         if (minfo) {
 402                 entry->se_minfo = NULL;
 403                 ll_intent_release(&minfo->mi_it);
 404                 sa_fini_data(minfo);
 405         }
 406
 407         if (req) {
 408                 entry->se_req = NULL;
 409                 ptlrpc_req_finished(req);
 410         }
 411
 412         spin_lock(&lli->lli_sa_lock);
 413         wakeup = __sa_make_ready(sai, entry, ret);
 414         spin_unlock(&lli->lli_sa_lock);
 415
 416         if (wakeup)
 417                 wake_up(&sai->sai_waitq);
 418 }
 419
 420 /* insert inode into the list of sai_agls */
 421 static void ll_agl_add(struct ll_statahead_info *sai,
 422                        struct inode *inode, int index)
 423 {
 424         struct ll_inode_info *child  = ll_i2info(inode);
 425         struct ll_inode_info *parent = ll_i2info(sai->sai_dentry->d_inode);
 426         int                   added  = 0;
 427
 428         spin_lock(&child->lli_agl_lock);
 429         if (child->lli_agl_index == 0) {
 430                 child->lli_agl_index = index;
 431                 spin_unlock(&child->lli_agl_lock);
 432
 433                 LASSERT(list_empty(&child->lli_agl_list));
 434
 435                 igrab(inode);
 436                 spin_lock(&parent->lli_agl_lock);
 437                 if (agl_list_empty(sai))
 438                         added = 1;
 439                 list_add_tail(&child->lli_agl_list, &sai->sai_agls);
 440                 if (added && sai->sai_agl_task)
 441                         wake_up_process(sai->sai_agl_task);
 442                 spin_unlock(&parent->lli_agl_lock);
 443         } else {
 444                 spin_unlock(&child->lli_agl_lock);
 445         }
 446 }
 447
 448 /* allocate sai */
 449 static struct ll_statahead_info *ll_sai_alloc(struct dentry *dentry)
 450 {
 451         struct ll_statahead_info *sai;
 452         struct ll_inode_info *lli = ll_i2info(dentry->d_inode);
 453         int i;
 454         ENTRY;
 455
 456         OBD_ALLOC_PTR(sai);
 457         if (!sai)
 458                 RETURN(NULL);
 459
 460         sai->sai_dentry = dget(dentry);
 461         atomic_set(&sai->sai_refcount, 1);
 462         sai->sai_max = LL_SA_RPC_MIN;
 463         sai->sai_index = 1;
 464         init_waitqueue_head(&sai->sai_waitq);
 465
 466         INIT_LIST_HEAD(&sai->sai_interim_entries);
 467         INIT_LIST_HEAD(&sai->sai_entries);
 468         INIT_LIST_HEAD(&sai->sai_agls);
 469
 470         for (i = 0; i < LL_SA_CACHE_SIZE; i++) {
 471                 INIT_LIST_HEAD(&sai->sai_cache[i]);
 472                 spin_lock_init(&sai->sai_cache_lock[i]);
 473         }
 474         atomic_set(&sai->sai_cache_count, 0);
 475
 476         spin_lock(&sai_generation_lock);
 477         lli->lli_sa_generation = ++sai_generation;
 478         if (unlikely(sai_generation == 0))
 479                 lli->lli_sa_generation = ++sai_generation;
 480         spin_unlock(&sai_generation_lock);
 481
 482         RETURN(sai);
 483 }
 484
 485 /* free sai */
 486 static inline void ll_sai_free(struct ll_statahead_info *sai)
 487 {
 488         LASSERT(sai->sai_dentry != NULL);
 489         dput(sai->sai_dentry);
 490         OBD_FREE_PTR(sai);
 491 }
 492
 493 /*
 494  * take refcount of sai if sai for @dir exists, which means statahead is on for
 495  * this directory.
 496  */
 497 static inline struct ll_statahead_info *ll_sai_get(struct inode *dir)
 498 {
 499         struct ll_inode_info *lli = ll_i2info(dir);
 500         struct ll_statahead_info *sai = NULL;
 501
 502         spin_lock(&lli->lli_sa_lock);
 503         sai = lli->lli_sai;
 504         if (sai != NULL)
 505                 atomic_inc(&sai->sai_refcount);
 506         spin_unlock(&lli->lli_sa_lock);
 507
 508         return sai;
 509 }
 510
 511 /*
 512  * put sai refcount after use, if refcount reaches zero, free sai and sa_entries
 513  * attached to it.
 514  */
 515 static void ll_sai_put(struct ll_statahead_info *sai)
 516 {
 517         struct ll_inode_info *lli = ll_i2info(sai->sai_dentry->d_inode);
 518
 519         if (atomic_dec_and_lock(&sai->sai_refcount, &lli->lli_sa_lock)) {
 520                 struct sa_entry *entry, *next;
 521                 struct ll_sb_info *sbi = ll_i2sbi(sai->sai_dentry->d_inode);
 522
 523                 lli->lli_sai = NULL;
 524                 spin_unlock(&lli->lli_sa_lock);
 525
 526                 LASSERT(!sai->sai_task);
 527                 LASSERT(!sai->sai_agl_task);
 528                 LASSERT(sai->sai_sent == sai->sai_replied);
 529                 LASSERT(!sa_has_callback(sai));
 530
 531                 list_for_each_entry_safe(entry, next, &sai->sai_entries,
 532                                          se_list)
 533                         sa_kill(sai, entry);
 534
 535                 LASSERT(atomic_read(&sai->sai_cache_count) == 0);
 536                 LASSERT(agl_list_empty(sai));
 537
 538                 ll_sai_free(sai);
 539                 atomic_dec(&sbi->ll_sa_running);
 540         }
 541 }
 542
 543 /* Do NOT forget to drop inode refcount when into sai_agls. */
 544 static void ll_agl_trigger(struct inode *inode, struct ll_statahead_info *sai)
 545 {
 546         struct ll_inode_info *lli = ll_i2info(inode);
 547         u64 index = lli->lli_agl_index;
 548         ktime_t expire;
 549         int rc;
 550
 551         ENTRY;
 552         LASSERT(list_empty(&lli->lli_agl_list));
 553
 554         /* AGL maybe fall behind statahead with one entry */
 555         if (is_omitted_entry(sai, index + 1)) {
 556                 lli->lli_agl_index = 0;
 557                 iput(inode);
 558                 RETURN_EXIT;
 559         }
 560
 561         /* In case of restore, the MDT has the right size and has already
 562          * sent it back without granting the layout lock, inode is up-to-date.
 563          * Then AGL (async glimpse lock) is useless.
 564          * Also to glimpse we need the layout, in case of a runninh restore
 565          * the MDT holds the layout lock so the glimpse will block up to the
 566          * end of restore (statahead/agl will block) */
 567         if (ll_file_test_flag(lli, LLIF_FILE_RESTORING)) {
 568                 lli->lli_agl_index = 0;
 569                 iput(inode);
 570                 RETURN_EXIT;
 571         }
 572
 573         /* Someone is in glimpse (sync or async), do nothing. */
 574         rc = down_write_trylock(&lli->lli_glimpse_sem);
 575         if (rc == 0) {
 576                 lli->lli_agl_index = 0;
 577                 iput(inode);
 578                 RETURN_EXIT;
 579         }
 580
 581         /*
 582          * Someone triggered glimpse within 1 sec before.
 583          * 1) The former glimpse succeeded with glimpse lock granted by OST, and
 584          *    if the lock is still cached on client, AGL needs to do nothing. If
 585          *    it is cancelled by other client, AGL maybe cannot obtaion new lock
 586          *    for no glimpse callback triggered by AGL.
 587          * 2) The former glimpse succeeded, but OST did not grant glimpse lock.
 588          *    Under such case, it is quite possible that the OST will not grant
 589          *    glimpse lock for AGL also.
 590          * 3) The former glimpse failed, compared with other two cases, it is
 591          *    relative rare. AGL can ignore such case, and it will not muchly
 592          *    affect the performance.
 593          */
 594         expire = ktime_sub_ns(ktime_get(), NSEC_PER_SEC);
 595         if (ktime_to_ns(lli->lli_glimpse_time) &&
 596             ktime_before(expire, lli->lli_glimpse_time)) {
 597                 up_write(&lli->lli_glimpse_sem);
 598                 lli->lli_agl_index = 0;
 599                 iput(inode);
 600                 RETURN_EXIT;
 601         }
 602
 603         CDEBUG(D_READA, "Handling (init) async glimpse: inode = "
 604                DFID", idx = %llu\n", PFID(&lli->lli_fid), index);
 605
 606         cl_agl(inode);
 607         lli->lli_agl_index = 0;
 608         lli->lli_glimpse_time = ktime_get();
 609         up_write(&lli->lli_glimpse_sem);
 610
 611         CDEBUG(D_READA, "Handled (init) async glimpse: inode= "
 612                DFID", idx = %llu, rc = %d\n",
 613                PFID(&lli->lli_fid), index, rc);
 614
 615         iput(inode);
 616
 617         EXIT;
 618 }
 619
 620 /*
 621  * prepare inode for sa entry, add it into agl list, now sa_entry is ready
 622  * to be used by scanner process.
 623  */
 624 static void sa_instantiate(struct ll_statahead_info *sai,
 625                                  struct sa_entry *entry)
 626 {
 627         struct inode *dir = sai->sai_dentry->d_inode;
 628         struct inode *child;
 629         struct md_enqueue_info *minfo;
 630         struct lookup_intent *it;
 631         struct ptlrpc_request *req;
 632         struct mdt_body *body;
 633         int rc = 0;
 634         ENTRY;
 635
 636         LASSERT(entry->se_handle != 0);
 637
 638         minfo = entry->se_minfo;
 639         it = &minfo->mi_it;
 640         req = entry->se_req;
 641         body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
 642         if (body == NULL)
 643                 GOTO(out, rc = -EFAULT);
 644
 645         child = entry->se_inode;
 646         if (child != NULL) {
 647                 /* revalidate; unlinked and re-created with the same name */
 648                 if (unlikely(!lu_fid_eq(&minfo->mi_data.op_fid2,
 649                                         &body->mbo_fid1))) {
 650                         entry->se_inode = NULL;
 651                         iput(child);
 652                         child = NULL;
 653                 }
 654         }
 655
 656         it->it_lock_handle = entry->se_handle;
 657         rc = md_revalidate_lock(ll_i2mdexp(dir), it, ll_inode2fid(dir), NULL);
 658         if (rc != 1)
 659                 GOTO(out, rc = -EAGAIN);
 660
 661         rc = ll_prep_inode(&child, req, dir->i_sb, it);
 662         if (rc)
 663                 GOTO(out, rc);
 664
 665         CDEBUG(D_READA, "%s: setting %.*s"DFID" l_data to inode %p\n",
 666                ll_i2sbi(dir)->ll_fsname, entry->se_qstr.len,
 667                entry->se_qstr.name, PFID(ll_inode2fid(child)), child);
 668         ll_set_lock_data(ll_i2sbi(dir)->ll_md_exp, child, it, NULL);
 669
 670         entry->se_inode = child;
 671
 672         if (agl_should_run(sai, child))
 673                 ll_agl_add(sai, child, entry->se_index);
 674
 675         EXIT;
 676
 677 out:
 678         /* sa_make_ready() will drop ldlm ibits lock refcount by calling
 679          * ll_intent_drop_lock() in spite of failures. Do not worry about
 680          * calling ll_intent_drop_lock() more than once. */
 681         sa_make_ready(sai, entry, rc);
 682 }
 683
 684 /* once there are async stat replies, instantiate sa_entry from replies */
 685 static void sa_handle_callback(struct ll_statahead_info *sai)
 686 {
 687         struct ll_inode_info *lli;
 688
 689         lli = ll_i2info(sai->sai_dentry->d_inode);
 690
 691         spin_lock(&lli->lli_sa_lock);
 692         while (sa_has_callback(sai)) {
 693                 struct sa_entry *entry;
 694
 695                 entry = list_entry(sai->sai_interim_entries.next,
 696                                    struct sa_entry, se_list);
 697                 list_del_init(&entry->se_list);
 698                 spin_unlock(&lli->lli_sa_lock);
 699
 700                 sa_instantiate(sai, entry);
 701                 spin_lock(&lli->lli_sa_lock);
 702         }
 703         spin_unlock(&lli->lli_sa_lock);
 704 }
 705
 706 /*
 707  * callback for async stat RPC, because this is called in ptlrpcd context, we
 708  * only put sa_entry in sai_interim_entries, and wake up statahead thread to
 709  * really prepare inode and instantiate sa_entry later.
 710  */
 711 static int ll_statahead_interpret(struct ptlrpc_request *req,
 712                                   struct md_enqueue_info *minfo, int rc)
 713 {
 714         struct lookup_intent *it = &minfo->mi_it;
 715         struct inode *dir = minfo->mi_dir;
 716         struct ll_inode_info *lli = ll_i2info(dir);
 717         struct ll_statahead_info *sai = lli->lli_sai;
 718         struct sa_entry *entry = (struct sa_entry *)minfo->mi_cbdata;
 719         __u64 handle = 0;
 720         ENTRY;
 721
 722         if (it_disposition(it, DISP_LOOKUP_NEG))
 723                 rc = -ENOENT;
 724
 725         /* because statahead thread will wait for all inflight RPC to finish,
 726          * sai should be always valid, no need to refcount */
 727         LASSERT(sai != NULL);
 728         LASSERT(entry != NULL);
 729
 730         CDEBUG(D_READA, "sa_entry %.*s rc %d\n",
 731                entry->se_qstr.len, entry->se_qstr.name, rc);
 732
 733         if (rc != 0) {
 734                 ll_intent_release(it);
 735                 sa_fini_data(minfo);
 736         } else {
 737                 /* release ibits lock ASAP to avoid deadlock when statahead
 738                  * thread enqueues lock on parent in readdir and another
 739                  * process enqueues lock on child with parent lock held, eg.
 740                  * unlink. */
 741                 handle = it->it_lock_handle;
 742                 ll_intent_drop_lock(it);
 743                 ll_unlock_md_op_lsm(&minfo->mi_data);
 744         }
 745
 746         spin_lock(&lli->lli_sa_lock);
 747         if (rc != 0) {
 748                 if (__sa_make_ready(sai, entry, rc))
 749                         wake_up(&sai->sai_waitq);
 750         } else {
 751                 int first = 0;
 752
 753                 entry->se_minfo = minfo;
 754                 entry->se_req = ptlrpc_request_addref(req);
 755                 /* Release the async ibits lock ASAP to avoid deadlock
 756                  * when statahead thread tries to enqueue lock on parent
 757                  * for readpage and other tries to enqueue lock on child
 758                  * with parent's lock held, for example: unlink. */
 759                 entry->se_handle = handle;
 760                 if (!sa_has_callback(sai))
 761                         first = 1;
 762
 763                 list_add_tail(&entry->se_list, &sai->sai_interim_entries);
 764                 if (first && sai->sai_task)
 765                         wake_up_process(sai->sai_task);
 766         }
 767         sai->sai_replied++;
 768
 769         spin_unlock(&lli->lli_sa_lock);
 770
 771         RETURN(rc);
 772 }
 773
 774 /* async stat for file not found in dcache */
 775 static int sa_lookup(struct inode *dir, struct sa_entry *entry)
 776 {
 777         struct md_enqueue_info   *minfo;
 778         int                       rc;
 779         ENTRY;
 780
 781         minfo = sa_prep_data(dir, NULL, entry);
 782         if (IS_ERR(minfo))
 783                 RETURN(PTR_ERR(minfo));
 784
 785         rc = md_intent_getattr_async(ll_i2mdexp(dir), minfo);
 786         if (rc < 0)
 787                 sa_fini_data(minfo);
 788
 789         RETURN(rc);
 790 }
 791
 792 /**
 793  * async stat for file found in dcache, similar to .revalidate
 794  *
 795  * \retval      1 dentry valid, no RPC sent
 796  * \retval      0 dentry invalid, will send async stat RPC
 797  * \retval      negative number upon error
 798  */
 799 static int sa_revalidate(struct inode *dir, struct sa_entry *entry,
 800                          struct dentry *dentry)
 801 {
 802         struct inode *inode = dentry->d_inode;
 803         struct lookup_intent it = { .it_op = IT_GETATTR,
 804                                     .it_lock_handle = 0 };
 805         struct md_enqueue_info *minfo;
 806         int rc;
 807         ENTRY;
 808
 809         if (unlikely(inode == NULL))
 810                 RETURN(1);
 811
 812         if (d_mountpoint(dentry))
 813                 RETURN(1);
 814
 815         minfo = sa_prep_data(dir, inode, entry);
 816         if (IS_ERR(minfo))
 817                 RETURN(PTR_ERR(minfo));
 818
 819         entry->se_inode = igrab(inode);
 820         rc = md_revalidate_lock(ll_i2mdexp(dir), &it, ll_inode2fid(inode),
 821                                 NULL);
 822         if (rc == 1) {
 823                 entry->se_handle = it.it_lock_handle;
 824                 ll_intent_release(&it);
 825                 sa_fini_data(minfo);
 826                 RETURN(1);
 827         }
 828
 829         rc = md_intent_getattr_async(ll_i2mdexp(dir), minfo);
 830         if (rc < 0) {
 831                 entry->se_inode = NULL;
 832                 iput(inode);
 833                 sa_fini_data(minfo);
 834         }
 835
 836         RETURN(rc);
 837 }
 838
 839 /* async stat for file with @name */
 840 static void sa_statahead(struct dentry *parent, const char *name, int len,
 841                          const struct lu_fid *fid)
 842 {
 843         struct inode *dir = parent->d_inode;
 844         struct ll_inode_info *lli = ll_i2info(dir);
 845         struct ll_statahead_info *sai = lli->lli_sai;
 846         struct dentry *dentry = NULL;
 847         struct sa_entry *entry;
 848         int rc;
 849         ENTRY;
 850
 851         entry = sa_alloc(parent, sai, sai->sai_index, name, len, fid);
 852         if (IS_ERR(entry))
 853                 RETURN_EXIT;
 854
 855         dentry = d_lookup(parent, &entry->se_qstr);
 856         if (!dentry) {
 857                 rc = sa_lookup(dir, entry);
 858         } else {
 859                 rc = sa_revalidate(dir, entry, dentry);
 860                 if (rc == 1 && agl_should_run(sai, dentry->d_inode))
 861                         ll_agl_add(sai, dentry->d_inode, entry->se_index);
 862         }
 863
 864         if (dentry != NULL)
 865                 dput(dentry);
 866
 867         if (rc != 0)
 868                 sa_make_ready(sai, entry, rc);
 869         else
 870                 sai->sai_sent++;
 871
 872         sai->sai_index++;
 873
 874         EXIT;
 875 }
 876
 877 #ifndef TASK_IDLE
 878 #define TASK_IDLE TASK_INTERRUPTIBLE
 879 #endif
 880
 881 /* async glimpse (agl) thread main function */
 882 static int ll_agl_thread(void *arg)
 883 {
 884         struct dentry *parent = (struct dentry *)arg;
 885         struct inode *dir = parent->d_inode;
 886         struct ll_inode_info *plli = ll_i2info(dir);
 887         struct ll_inode_info *clli;
 888         /* We already own this reference, so it is safe to take it
 889          * without a lock.
 890          */
 891         struct ll_statahead_info *sai = plli->lli_sai;
 892         ENTRY;
 893
 894         CDEBUG(D_READA, "agl thread started: sai %p, parent %.*s\n",
 895                sai, parent->d_name.len, parent->d_name.name);
 896
 897         while (({set_current_state(TASK_IDLE);
 898                  !kthread_should_stop(); })) {
 899                 spin_lock(&plli->lli_agl_lock);
 900                 if (!agl_list_empty(sai)) {
 901                         __set_current_state(TASK_RUNNING);
 902                         clli = agl_first_entry(sai);
 903                         list_del_init(&clli->lli_agl_list);
 904                         spin_unlock(&plli->lli_agl_lock);
 905                         ll_agl_trigger(&clli->lli_vfs_inode, sai);
 906                         cond_resched();
 907                 } else {
 908                         spin_unlock(&plli->lli_agl_lock);
 909                         schedule();
 910                 }
 911         }
 912         __set_current_state(TASK_RUNNING);
 913         RETURN(0);
 914 }
 915
 916 static void ll_stop_agl(struct ll_statahead_info *sai)
 917 {
 918         struct dentry *parent = sai->sai_dentry;
 919         struct ll_inode_info *plli = ll_i2info(parent->d_inode);
 920         struct ll_inode_info *clli;
 921         struct task_struct *agl_task;
 922
 923         spin_lock(&plli->lli_agl_lock);
 924         agl_task = sai->sai_agl_task;
 925         sai->sai_agl_task = NULL;
 926         spin_unlock(&plli->lli_agl_lock);
 927         if (!agl_task)
 928                 return;
 929
 930         CDEBUG(D_READA, "stop agl thread: sai %p pid %u\n",
 931                sai, (unsigned int)agl_task->pid);
 932         kthread_stop(agl_task);
 933
 934         spin_lock(&plli->lli_agl_lock);
 935         sai->sai_agl_valid = 0;
 936         while (!agl_list_empty(sai)) {
 937                 clli = agl_first_entry(sai);
 938                 list_del_init(&clli->lli_agl_list);
 939                 spin_unlock(&plli->lli_agl_lock);
 940                 clli->lli_agl_index = 0;
 941                 iput(&clli->lli_vfs_inode);
 942                 spin_lock(&plli->lli_agl_lock);
 943         }
 944         spin_unlock(&plli->lli_agl_lock);
 945         CDEBUG(D_READA, "agl thread stopped: sai %p, parent %.*s\n",
 946                sai, parent->d_name.len, parent->d_name.name);
 947         ll_sai_put(sai);
 948 }
 949
 950 /* start agl thread */
 951 static void ll_start_agl(struct dentry *parent, struct ll_statahead_info *sai)
 952 {
 953         struct ll_inode_info  *plli;
 954         struct task_struct    *task;
 955         ENTRY;
 956
 957         CDEBUG(D_READA, "start agl thread: sai %p, parent %.*s\n",
 958                sai, parent->d_name.len, parent->d_name.name);
 959
 960         plli = ll_i2info(parent->d_inode);
 961         task = kthread_create(ll_agl_thread, parent,
 962                               "ll_agl_%u", plli->lli_opendir_pid);
 963         if (IS_ERR(task)) {
 964                 CERROR("can't start ll_agl thread, rc: %ld\n", PTR_ERR(task));
 965                 RETURN_EXIT;
 966         }
 967         sai->sai_agl_task = task;
 968         sai->sai_agl_valid = 1;
 969         atomic_inc(&ll_i2sbi(d_inode(parent))->ll_agl_total);
 970         /* Get an extra reference that the thread holds */
 971         ll_sai_get(d_inode(parent));
 972
 973         wake_up_process(task);
 974
 975         EXIT;
 976 }
 977
 978 /* statahead thread main function */
 979 static int ll_statahead_thread(void *arg)
 980 {
 981         struct dentry *parent = (struct dentry *)arg;
 982         struct inode *dir = parent->d_inode;
 983         struct ll_inode_info *lli = ll_i2info(dir);
 984         struct ll_sb_info *sbi = ll_i2sbi(dir);
 985         struct ll_statahead_info *sai = lli->lli_sai;
 986         int first = 0;
 987         struct md_op_data *op_data;
 988         struct ll_dir_chain chain;
 989         struct page *page = NULL;
 990         __u64 pos = 0;
 991         int rc = 0;
 992         ENTRY;
 993
 994         CDEBUG(D_READA, "statahead thread starting: sai %p, parent %.*s\n",
 995                sai, parent->d_name.len, parent->d_name.name);
 996
 997         OBD_ALLOC_PTR(op_data);
 998         if (!op_data)
 999                 GOTO(out, rc = -ENOMEM);
1000
1001         ll_dir_chain_init(&chain);
1002         while (pos != MDS_DIR_END_OFF && sai->sai_task) {
1003                 struct lu_dirpage *dp;
1004                 struct lu_dirent  *ent;
1005
1006                 op_data = ll_prep_md_op_data(op_data, dir, dir, NULL, 0, 0,
1007                                      LUSTRE_OPC_ANY, dir);
1008                 if (IS_ERR(op_data)) {
1009                         rc = PTR_ERR(op_data);
1010                         break;
1011                 }
1012
1013                 sai->sai_in_readpage = 1;
1014                 page = ll_get_dir_page(dir, op_data, pos, &chain);
1015                 ll_unlock_md_op_lsm(op_data);
1016                 sai->sai_in_readpage = 0;
1017                 if (IS_ERR(page)) {
1018                         rc = PTR_ERR(page);
1019                         CDEBUG(D_READA, "error reading dir "DFID" at %llu"
1020                                "/%llu opendir_pid = %u: rc = %d\n",
1021                                PFID(ll_inode2fid(dir)), pos, sai->sai_index,
1022                                lli->lli_opendir_pid, rc);
1023                         break;
1024                 }
1025
1026                 dp = page_address(page);
1027                 for (ent = lu_dirent_start(dp);
1028                      ent != NULL && sai->sai_task &&
1029                      !sa_low_hit(sai);
1030                      ent = lu_dirent_next(ent)) {
1031                         __u64 hash;
1032                         int namelen;
1033                         char *name;
1034                         struct lu_fid fid;
1035
1036                         hash = le64_to_cpu(ent->lde_hash);
1037                         if (unlikely(hash < pos))
1038                                 /*
1039                                  * Skip until we find target hash value.
1040                                  */
1041                                 continue;
1042
1043                         namelen = le16_to_cpu(ent->lde_namelen);
1044                         if (unlikely(namelen == 0))
1045                                 /*
1046                                  * Skip dummy record.
1047                                  */
1048                                 continue;
1049
1050                         name = ent->lde_name;
1051                         if (name[0] == '.') {
1052                                 if (namelen == 1) {
1053                                         /*
1054                                          * skip "."
1055                                          */
1056                                         continue;
1057                                 } else if (name[1] == '.' && namelen == 2) {
1058                                         /*
1059                                          * skip ".."
1060                                          */
1061                                         continue;
1062                                 } else if (!sai->sai_ls_all) {
1063                                         /*
1064                                          * skip hidden files.
1065                                          */
1066                                         sai->sai_skip_hidden++;
1067                                         continue;
1068                                 }
1069                         }
1070
1071                         /*
1072                          * don't stat-ahead first entry.
1073                          */
1074                         if (unlikely(++first == 1))
1075                                 continue;
1076
1077                         fid_le_to_cpu(&fid, &ent->lde_fid);
1078
1079                         while (({set_current_state(TASK_IDLE);
1080                                  sai->sai_task; })) {
1081                                 if (sa_has_callback(sai)) {
1082                                         __set_current_state(TASK_RUNNING);
1083                                         sa_handle_callback(sai);
1084                                 }
1085
1086                                 spin_lock(&lli->lli_agl_lock);
1087                                 while (sa_sent_full(sai) &&
1088                                        !agl_list_empty(sai)) {
1089                                         struct ll_inode_info *clli;
1090
1091                                         __set_current_state(TASK_RUNNING);
1092                                         clli = agl_first_entry(sai);
1093                                         list_del_init(&clli->lli_agl_list);
1094                                         spin_unlock(&lli->lli_agl_lock);
1095
1096                                         ll_agl_trigger(&clli->lli_vfs_inode,
1097                                                         sai);
1098                                         cond_resched();
1099                                         spin_lock(&lli->lli_agl_lock);
1100                                 }
1101                                 spin_unlock(&lli->lli_agl_lock);
1102
1103                                 if (!sa_sent_full(sai))
1104                                         break;
1105                                 schedule();
1106                         }
1107                         __set_current_state(TASK_RUNNING);
1108
1109                         sa_statahead(parent, name, namelen, &fid);
1110                 }
1111
1112                 pos = le64_to_cpu(dp->ldp_hash_end);
1113                 ll_release_page(dir, page,
1114                                 le32_to_cpu(dp->ldp_flags) & LDF_COLLIDE);
1115
1116                 if (sa_low_hit(sai)) {
1117                         rc = -EFAULT;
1118                         atomic_inc(&sbi->ll_sa_wrong);
1119                         CDEBUG(D_READA, "Statahead for dir "DFID" hit "
1120                                "ratio too low: hit/miss %llu/%llu"
1121                                ", sent/replied %llu/%llu, stopping "
1122                                "statahead thread: pid %d\n",
1123                                PFID(&lli->lli_fid), sai->sai_hit,
1124                                sai->sai_miss, sai->sai_sent,
1125                                sai->sai_replied, current->pid);
1126                         break;
1127                 }
1128         }
1129         ll_dir_chain_fini(&chain);
1130         ll_finish_md_op_data(op_data);
1131
1132         if (rc < 0) {
1133                 spin_lock(&lli->lli_sa_lock);
1134                 sai->sai_task = NULL;
1135                 lli->lli_sa_enabled = 0;
1136                 spin_unlock(&lli->lli_sa_lock);
1137         }
1138
1139         /* statahead is finished, but statahead entries need to be cached, wait
1140          * for file release to stop me.
1141          */
1142         while (({set_current_state(TASK_IDLE);
1143                  sai->sai_task; })) {
1144                 if (sa_has_callback(sai)) {
1145                         __set_current_state(TASK_RUNNING);
1146                         sa_handle_callback(sai);
1147                 } else {
1148                         schedule();
1149                 }
1150         }
1151         __set_current_state(TASK_RUNNING);
1152
1153         EXIT;
1154 out:
1155         ll_stop_agl(sai);
1156
1157         /* wait for inflight statahead RPCs to finish, and then we can free sai
1158          * safely because statahead RPC will access sai data */
1159         while (sai->sai_sent != sai->sai_replied)
1160                 /* in case we're not woken up, timeout wait */
1161                 msleep(125);
1162
1163         /* release resources held by statahead RPCs */
1164         sa_handle_callback(sai);
1165
1166         CDEBUG(D_READA, "statahead thread stopped: sai %p, parent %.*s\n",
1167                sai, parent->d_name.len, parent->d_name.name);
1168
1169         spin_lock(&lli->lli_sa_lock);
1170         sai->sai_task = NULL;
1171         spin_unlock(&lli->lli_sa_lock);
1172         wake_up(&sai->sai_waitq);
1173
1174         ll_sai_put(sai);
1175
1176         return rc;
1177 }
1178
1179 /* authorize opened dir handle @key to statahead */
1180 void ll_authorize_statahead(struct inode *dir, void *key)
1181 {
1182         struct ll_inode_info *lli = ll_i2info(dir);
1183
1184         spin_lock(&lli->lli_sa_lock);
1185         if (lli->lli_opendir_key == NULL && lli->lli_sai == NULL) {
1186                 /*
1187                  * if lli_sai is not NULL, it means previous statahead is not
1188                  * finished yet, we'd better not start a new statahead for now.
1189                  */
1190                 LASSERT(lli->lli_opendir_pid == 0);
1191                 lli->lli_opendir_key = key;
1192                 lli->lli_opendir_pid = current->pid;
1193                 lli->lli_sa_enabled = 1;
1194         }
1195         spin_unlock(&lli->lli_sa_lock);
1196 }
1197
1198 /*
1199  * deauthorize opened dir handle @key to statahead, and notify statahead thread
1200  * to quit if it's running.
1201  */
1202 void ll_deauthorize_statahead(struct inode *dir, void *key)
1203 {
1204         struct ll_inode_info *lli = ll_i2info(dir);
1205         struct ll_statahead_info *sai;
1206
1207         LASSERT(lli->lli_opendir_key == key);
1208         LASSERT(lli->lli_opendir_pid != 0);
1209
1210         CDEBUG(D_READA, "deauthorize statahead for "DFID"\n",
1211                 PFID(&lli->lli_fid));
1212
1213         spin_lock(&lli->lli_sa_lock);
1214         lli->lli_opendir_key = NULL;
1215         lli->lli_opendir_pid = 0;
1216         lli->lli_sa_enabled = 0;
1217         sai = lli->lli_sai;
1218         if (sai && sai->sai_task) {
1219                 /*
1220                  * statahead thread may not have quit yet because it needs to
1221                  * cache entries, now it's time to tell it to quit.
1222                  *
1223                  * wake_up_process() provides the necessary barriers
1224                  * to pair with set_current_state().
1225                  */
1226                 struct task_struct *task = sai->sai_task;
1227
1228                 sai->sai_task = NULL;
1229                 wake_up_process(task);
1230         }
1231         spin_unlock(&lli->lli_sa_lock);
1232 }
1233
1234 enum {
1235         /**
1236          * not first dirent, or is "."
1237          */
1238         LS_NOT_FIRST_DE = 0,
1239         /**
1240          * the first non-hidden dirent
1241          */
1242         LS_FIRST_DE,
1243         /**
1244          * the first hidden dirent, that is "."
1245          */
1246         LS_FIRST_DOT_DE
1247 };
1248
1249 /* file is first dirent under @dir */
1250 static int is_first_dirent(struct inode *dir, struct dentry *dentry)
1251 {
1252         struct ll_dir_chain   chain;
1253         struct qstr          *target = &dentry->d_name;
1254         struct md_op_data    *op_data;
1255         int                   dot_de;
1256         struct page          *page = NULL;
1257         int                   rc = LS_NOT_FIRST_DE;
1258         __u64                 pos = 0;
1259         ENTRY;
1260
1261         op_data = ll_prep_md_op_data(NULL, dir, dir, NULL, 0, 0,
1262                                      LUSTRE_OPC_ANY, dir);
1263         if (IS_ERR(op_data))
1264                 RETURN(PTR_ERR(op_data));
1265         /**
1266          *FIXME choose the start offset of the readdir
1267          */
1268
1269         ll_dir_chain_init(&chain);
1270         page = ll_get_dir_page(dir, op_data, 0, &chain);
1271
1272         while (1) {
1273                 struct lu_dirpage *dp;
1274                 struct lu_dirent  *ent;
1275
1276                 if (IS_ERR(page)) {
1277                         struct ll_inode_info *lli = ll_i2info(dir);
1278
1279                         rc = PTR_ERR(page);
1280                         CERROR("%s: reading dir "DFID" at %llu opendir_pid = %u : rc = %d\n",
1281                                ll_i2sbi(dir)->ll_fsname,
1282                                PFID(ll_inode2fid(dir)), pos,
1283                                lli->lli_opendir_pid, rc);
1284                         break;
1285                 }
1286
1287                 dp = page_address(page);
1288                 for (ent = lu_dirent_start(dp); ent != NULL;
1289                      ent = lu_dirent_next(ent)) {
1290                         __u64 hash;
1291                         int namelen;
1292                         char *name;
1293
1294                         hash = le64_to_cpu(ent->lde_hash);
1295                         /* The ll_get_dir_page() can return any page containing
1296                          * the given hash which may be not the start hash. */
1297                         if (unlikely(hash < pos))
1298                                 continue;
1299
1300                         namelen = le16_to_cpu(ent->lde_namelen);
1301                         if (unlikely(namelen == 0))
1302                                 /*
1303                                  * skip dummy record.
1304                                  */
1305                                 continue;
1306
1307                         name = ent->lde_name;
1308                         if (name[0] == '.') {
1309                                 if (namelen == 1)
1310                                         /*
1311                                          * skip "."
1312                                          */
1313                                         continue;
1314                                 else if (name[1] == '.' && namelen == 2)
1315                                         /*
1316                                          * skip ".."
1317                                          */
1318                                         continue;
1319                                 else
1320                                         dot_de = 1;
1321                         } else {
1322                                 dot_de = 0;
1323                         }
1324
1325                         if (dot_de && target->name[0] != '.') {
1326                                 CDEBUG(D_READA, "%.*s skip hidden file %.*s\n",
1327                                        target->len, target->name,
1328                                        namelen, name);
1329                                 continue;
1330                         }
1331
1332                         if (target->len != namelen ||
1333                             memcmp(target->name, name, namelen) != 0)
1334                                 rc = LS_NOT_FIRST_DE;
1335                         else if (!dot_de)
1336                                 rc = LS_FIRST_DE;
1337                         else
1338                                 rc = LS_FIRST_DOT_DE;
1339
1340                         ll_release_page(dir, page, false);
1341                         GOTO(out, rc);
1342                 }
1343                 pos = le64_to_cpu(dp->ldp_hash_end);
1344                 if (pos == MDS_DIR_END_OFF) {
1345                         /*
1346                          * End of directory reached.
1347                          */
1348                         ll_release_page(dir, page, false);
1349                         GOTO(out, rc);
1350                 } else {
1351                         /*
1352                          * chain is exhausted
1353                          * Normal case: continue to the next page.
1354                          */
1355                         ll_release_page(dir, page, le32_to_cpu(dp->ldp_flags) &
1356                                               LDF_COLLIDE);
1357                         page = ll_get_dir_page(dir, op_data, pos, &chain);
1358                 }
1359         }
1360         EXIT;
1361 out:
1362         ll_dir_chain_fini(&chain);
1363         ll_finish_md_op_data(op_data);
1364         return rc;
1365 }
1366
1367 /**
1368  * revalidate @dentryp from statahead cache
1369  *
1370  * \param[in] dir       parent directory
1371  * \param[in] sai       sai structure
1372  * \param[out] dentryp  pointer to dentry which will be revalidated
1373  * \param[in] unplug    unplug statahead window only (normally for negative
1374  *                      dentry)
1375  * \retval              1 on success, dentry is saved in @dentryp
1376  * \retval              0 if revalidation failed (no proper lock on client)
1377  * \retval              negative number upon error
1378  */
1379 static int revalidate_statahead_dentry(struct inode *dir,
1380                                         struct ll_statahead_info *sai,
1381                                         struct dentry **dentryp,
1382                                         bool unplug)
1383 {
1384         struct sa_entry *entry = NULL;
1385         struct ll_dentry_data *ldd;
1386         struct ll_inode_info *lli = ll_i2info(dir);
1387         int rc = 0;
1388         ENTRY;
1389
1390         if ((*dentryp)->d_name.name[0] == '.') {
1391                 if (sai->sai_ls_all ||
1392                     sai->sai_miss_hidden >= sai->sai_skip_hidden) {
1393                         /*
1394                          * Hidden dentry is the first one, or statahead
1395                          * thread does not skip so many hidden dentries
1396                          * before "sai_ls_all" enabled as below.
1397                          */
1398                 } else {
1399                         if (!sai->sai_ls_all)
1400                                 /*
1401                                  * It maybe because hidden dentry is not
1402                                  * the first one, "sai_ls_all" was not
1403                                  * set, then "ls -al" missed. Enable
1404                                  * "sai_ls_all" for such case.
1405                                  */
1406                                 sai->sai_ls_all = 1;
1407
1408                         /*
1409                          * Such "getattr" has been skipped before
1410                          * "sai_ls_all" enabled as above.
1411                          */
1412                         sai->sai_miss_hidden++;
1413                         RETURN(-EAGAIN);
1414                 }
1415         }
1416
1417         if (unplug)
1418                 GOTO(out, rc = 1);
1419
1420         entry = sa_get(sai, &(*dentryp)->d_name);
1421         if (entry == NULL)
1422                 GOTO(out, rc = -EAGAIN);
1423
1424         /* if statahead is busy in readdir, help it do post-work */
1425         if (!sa_ready(entry) && sai->sai_in_readpage)
1426                 sa_handle_callback(sai);
1427
1428         if (!sa_ready(entry)) {
1429                 spin_lock(&lli->lli_sa_lock);
1430                 sai->sai_index_wait = entry->se_index;
1431                 spin_unlock(&lli->lli_sa_lock);
1432                 rc = wait_event_idle_timeout(sai->sai_waitq, sa_ready(entry),
1433                                              cfs_time_seconds(30));
1434                 if (rc == 0) {
1435                         /*
1436                          * entry may not be ready, so it may be used by inflight
1437                          * statahead RPC, don't free it.
1438                          */
1439                         entry = NULL;
1440                         GOTO(out, rc = -EAGAIN);
1441                 }
1442         }
1443
1444         /*
1445          * We need to see the value that was set immediately before we
1446          * were woken up.
1447          */
1448         if (smp_load_acquire(&entry->se_state) == SA_ENTRY_SUCC &&
1449             entry->se_inode) {
1450                 struct inode *inode = entry->se_inode;
1451                 struct lookup_intent it = { .it_op = IT_GETATTR,
1452                                             .it_lock_handle =
1453                                                 entry->se_handle };
1454                 __u64 bits;
1455
1456                 rc = md_revalidate_lock(ll_i2mdexp(dir), &it,
1457                                         ll_inode2fid(inode), &bits);
1458                 if (rc == 1) {
1459                         if ((*dentryp)->d_inode == NULL) {
1460                                 struct dentry *alias;
1461
1462                                 alias = ll_splice_alias(inode, *dentryp);
1463                                 if (IS_ERR(alias)) {
1464                                         ll_intent_release(&it);
1465                                         GOTO(out, rc = PTR_ERR(alias));
1466                                 }
1467                                 *dentryp = alias;
1468                                 /* statahead prepared this inode, transfer inode
1469                                  * refcount from sa_entry to dentry */
1470                                 entry->se_inode = NULL;
1471                         } else if ((*dentryp)->d_inode != inode) {
1472                                 /* revalidate, but inode is recreated */
1473                                 CDEBUG(D_READA,
1474                                         "%s: stale dentry %.*s inode "
1475                                         DFID", statahead inode "DFID
1476                                         "\n",
1477                                         ll_i2sbi(inode)->ll_fsname,
1478                                         (*dentryp)->d_name.len,
1479                                         (*dentryp)->d_name.name,
1480                                         PFID(ll_inode2fid((*dentryp)->d_inode)),
1481                                         PFID(ll_inode2fid(inode)));
1482                                 ll_intent_release(&it);
1483                                 GOTO(out, rc = -ESTALE);
1484                         }
1485
1486                         if ((bits & MDS_INODELOCK_LOOKUP) &&
1487                             d_lustre_invalid(*dentryp))
1488                                 d_lustre_revalidate(*dentryp);
1489                         ll_intent_release(&it);
1490                 }
1491         }
1492 out:
1493         /*
1494          * statahead cached sa_entry can be used only once, and will be killed
1495          * right after use, so if lookup/revalidate accessed statahead cache,
1496          * set dentry ldd_sa_generation to parent lli_sa_generation, later if we
1497          * stat this file again, we know we've done statahead before, see
1498          * dentry_may_statahead().
1499          */
1500         ldd = ll_d2d(*dentryp);
1501         /* ldd can be NULL if llite lookup failed. */
1502         if (ldd != NULL)
1503                 ldd->lld_sa_generation = lli->lli_sa_generation;
1504         sa_put(sai, entry);
1505         spin_lock(&lli->lli_sa_lock);
1506         if (sai->sai_task)
1507                 wake_up_process(sai->sai_task);
1508         spin_unlock(&lli->lli_sa_lock);
1509
1510         RETURN(rc);
1511 }
1512
1513 /**
1514  * start statahead thread
1515  *
1516  * \param[in] dir       parent directory
1517  * \param[in] dentry    dentry that triggers statahead, normally the first
1518  *                      dirent under @dir
1519  * \retval              -EAGAIN on success, because when this function is
1520  *                      called, it's already in lookup call, so client should
1521  *                      do it itself instead of waiting for statahead thread
1522  *                      to do it asynchronously.
1523  * \retval              negative number upon error
1524  */
1525 static int start_statahead_thread(struct inode *dir, struct dentry *dentry)
1526 {
1527         struct ll_inode_info *lli = ll_i2info(dir);
1528         struct ll_statahead_info *sai = NULL;
1529         struct dentry *parent = dentry->d_parent;
1530         struct task_struct *task;
1531         struct ll_sb_info *sbi = ll_i2sbi(parent->d_inode);
1532         int first = LS_FIRST_DE;
1533         int rc = 0;
1534         ENTRY;
1535
1536         /* I am the "lli_opendir_pid" owner, only me can set "lli_sai". */
1537         first = is_first_dirent(dir, dentry);
1538         if (first == LS_NOT_FIRST_DE)
1539                 /* It is not "ls -{a}l" operation, no need statahead for it. */
1540                 GOTO(out, rc = -EFAULT);
1541
1542         if (unlikely(atomic_inc_return(&sbi->ll_sa_running) >
1543                                        sbi->ll_sa_running_max)) {
1544                 CDEBUG(D_READA,
1545                        "Too many concurrent statahead instances, "
1546                        "avoid new statahead instance temporarily.\n");
1547                 GOTO(out, rc = -EMFILE);
1548         }
1549
1550         sai = ll_sai_alloc(parent);
1551         if (sai == NULL)
1552                 GOTO(out, rc = -ENOMEM);
1553
1554         sai->sai_ls_all = (first == LS_FIRST_DOT_DE);
1555
1556         /* if current lli_opendir_key was deauthorized, or dir re-opened by
1557          * another process, don't start statahead, otherwise the newly spawned
1558          * statahead thread won't be notified to quit. */
1559         spin_lock(&lli->lli_sa_lock);
1560         if (unlikely(lli->lli_sai != NULL ||
1561                      lli->lli_opendir_key == NULL ||
1562                      lli->lli_opendir_pid != current->pid)) {
1563                 spin_unlock(&lli->lli_sa_lock);
1564                 GOTO(out, rc = -EPERM);
1565         }
1566         lli->lli_sai = sai;
1567         spin_unlock(&lli->lli_sa_lock);
1568
1569         CDEBUG(D_READA, "start statahead thread: [pid %d] [parent %.*s]\n",
1570                current->pid, parent->d_name.len, parent->d_name.name);
1571
1572         task = kthread_create(ll_statahead_thread, parent, "ll_sa_%u",
1573                               lli->lli_opendir_pid);
1574         if (IS_ERR(task)) {
1575                 spin_lock(&lli->lli_sa_lock);
1576                 lli->lli_sai = NULL;
1577                 spin_unlock(&lli->lli_sa_lock);
1578                 rc = PTR_ERR(task);
1579                 CERROR("can't start ll_sa thread, rc: %d\n", rc);
1580                 GOTO(out, rc);
1581         }
1582
1583         if (ll_i2sbi(parent->d_inode)->ll_flags & LL_SBI_AGL_ENABLED)
1584                 ll_start_agl(parent, sai);
1585
1586         atomic_inc(&ll_i2sbi(parent->d_inode)->ll_sa_total);
1587         sai->sai_task = task;
1588
1589         wake_up_process(task);
1590         /*
1591          * We don't stat-ahead for the first dirent since we are already in
1592          * lookup.
1593          */
1594         RETURN(-EAGAIN);
1595
1596 out:
1597         /* once we start statahead thread failed, disable statahead so that
1598          * subsequent stat won't waste time to try it. */
1599         spin_lock(&lli->lli_sa_lock);
1600         if (lli->lli_opendir_pid == current->pid)
1601                 lli->lli_sa_enabled = 0;
1602         spin_unlock(&lli->lli_sa_lock);
1603
1604         if (sai != NULL)
1605                 ll_sai_free(sai);
1606         if (first != LS_NOT_FIRST_DE)
1607                 atomic_dec(&sbi->ll_sa_running);
1608
1609         RETURN(rc);
1610 }
1611
1612 /**
1613  * statahead entry function, this is called when client getattr on a file, it
1614  * will start statahead thread if this is the first dir entry, else revalidate
1615  * dentry from statahead cache.
1616  *
1617  * \param[in]  dir      parent directory
1618  * \param[out] dentryp  dentry to getattr
1619  * \param[in]  unplug   unplug statahead window only (normally for negative
1620  *                      dentry)
1621  * \retval              1 on success
1622  * \retval              0 revalidation from statahead cache failed, caller needs
1623  *                      to getattr from server directly
1624  * \retval              negative number on error, caller often ignores this and
1625  *                      then getattr from server
1626  */
1627 int ll_statahead(struct inode *dir, struct dentry **dentryp, bool unplug)
1628 {
1629         struct ll_statahead_info *sai;
1630
1631         sai = ll_sai_get(dir);
1632         if (sai != NULL) {
1633                 int rc;
1634
1635                 rc = revalidate_statahead_dentry(dir, sai, dentryp, unplug);
1636                 CDEBUG(D_READA, "revalidate statahead %.*s: %d.\n",
1637                         (*dentryp)->d_name.len, (*dentryp)->d_name.name, rc);
1638                 ll_sai_put(sai);
1639                 return rc;
1640         }
1641         return start_statahead_thread(dir, *dentryp);
1642 }