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 /*
  57  * sa_entry is not refcounted: statahead thread allocates it and do async stat,
  58  * and in async stat callback ll_statahead_interpret() will add it into
  59  * sai_interim_entries, later statahead thread will call sa_handle_callback() to
  60  * instantiate entry and move it into sai_entries, and then only scanner process
  61  * can access and free it.
  62  */
  63 struct sa_entry {
  64         /* link into sai_interim_entries or sai_entries */
  65         struct list_head        se_list;
  66         /* link into sai hash table locally */
  67         struct list_head        se_hash;
  68         /* entry index in the sai */
  69         __u64                   se_index;
  70         /* low layer ldlm lock handle */
  71         __u64                   se_handle;
  72         /* entry status */
  73         se_state_t              se_state;
  74         /* entry size, contains name */
  75         int                     se_size;
  76         /* pointer to async getattr enqueue info */
  77         struct md_enqueue_info *se_minfo;
  78         /* pointer to the async getattr request */
  79         struct ptlrpc_request  *se_req;
  80         /* pointer to the target inode */
  81         struct inode           *se_inode;
  82         /* entry name */
  83         struct qstr             se_qstr;
  84         /* entry fid */
  85         struct lu_fid           se_fid;
  86 };
  87
  88 static unsigned int sai_generation;
  89 static DEFINE_SPINLOCK(sai_generation_lock);
  90
  91 static inline int sa_unhashed(struct sa_entry *entry)
  92 {
  93         return list_empty(&entry->se_hash);
  94 }
  95
  96 /* sa_entry is ready to use */
  97 static inline int sa_ready(struct sa_entry *entry)
  98 {
  99         /* Make sure sa_entry is updated and ready to use */
 100         smp_rmb();
 101         return (entry->se_state != SA_ENTRY_INIT);
 102 }
 103
 104 /* hash value to put in sai_cache */
 105 static inline int sa_hash(int val)
 106 {
 107         return val & LL_SA_CACHE_MASK;
 108 }
 109
 110 /* hash entry into sai_cache */
 111 static inline void
 112 sa_rehash(struct ll_statahead_info *sai, struct sa_entry *entry)
 113 {
 114         int i = sa_hash(entry->se_qstr.hash);
 115
 116         spin_lock(&sai->sai_cache_lock[i]);
 117         list_add_tail(&entry->se_hash, &sai->sai_cache[i]);
 118         spin_unlock(&sai->sai_cache_lock[i]);
 119 }
 120
 121 /* unhash entry from sai_cache */
 122 static inline void
 123 sa_unhash(struct ll_statahead_info *sai, struct sa_entry *entry)
 124 {
 125         int i = sa_hash(entry->se_qstr.hash);
 126
 127         spin_lock(&sai->sai_cache_lock[i]);
 128         list_del_init(&entry->se_hash);
 129         spin_unlock(&sai->sai_cache_lock[i]);
 130 }
 131
 132 static inline int agl_should_run(struct ll_statahead_info *sai,
 133                                  struct inode *inode)
 134 {
 135         return inode && S_ISREG(inode->i_mode) && sai->sai_agl_task;
 136 }
 137
 138 static inline struct ll_inode_info *
 139 agl_first_entry(struct ll_statahead_info *sai)
 140 {
 141         return list_entry(sai->sai_agls.next, struct ll_inode_info,
 142                           lli_agl_list);
 143 }
 144
 145 /* statahead window is full */
 146 static inline int sa_sent_full(struct ll_statahead_info *sai)
 147 {
 148         return atomic_read(&sai->sai_cache_count) >= sai->sai_max;
 149 }
 150
 151 /* got async stat replies */
 152 static inline int sa_has_callback(struct ll_statahead_info *sai)
 153 {
 154         return !list_empty(&sai->sai_interim_entries);
 155 }
 156
 157 static inline int agl_list_empty(struct ll_statahead_info *sai)
 158 {
 159         return list_empty(&sai->sai_agls);
 160 }
 161
 162 /**
 163  * (1) hit ratio less than 80%
 164  * or
 165  * (2) consecutive miss more than 8
 166  * then means low hit.
 167  */
 168 static inline int sa_low_hit(struct ll_statahead_info *sai)
 169 {
 170         return ((sai->sai_hit > 7 && sai->sai_hit < 4 * sai->sai_miss) ||
 171                 (sai->sai_consecutive_miss > 8));
 172 }
 173
 174 /*
 175  * if the given index is behind of statahead window more than
 176  * SA_OMITTED_ENTRY_MAX, then it is old.
 177  */
 178 static inline int is_omitted_entry(struct ll_statahead_info *sai, __u64 index)
 179 {
 180         return ((__u64)sai->sai_max + index + SA_OMITTED_ENTRY_MAX <
 181                 sai->sai_index);
 182 }
 183
 184 /* allocate sa_entry and hash it to allow scanner process to find it */
 185 static struct sa_entry *
 186 sa_alloc(struct dentry *parent, struct ll_statahead_info *sai, __u64 index,
 187          const char *name, int len, const struct lu_fid *fid)
 188 {
 189         struct ll_inode_info *lli;
 190         struct sa_entry *entry;
 191         int entry_size;
 192         char *dname;
 193
 194         ENTRY;
 195
 196         entry_size = sizeof(struct sa_entry) + (len & ~3) + 4;
 197         OBD_ALLOC(entry, entry_size);
 198         if (unlikely(!entry))
 199                 RETURN(ERR_PTR(-ENOMEM));
 200
 201         CDEBUG(D_READA, "alloc sa entry %.*s(%p) index %llu\n",
 202                len, name, entry, index);
 203
 204         entry->se_index = index;
 205
 206         entry->se_state = SA_ENTRY_INIT;
 207         entry->se_size = entry_size;
 208         dname = (char *)entry + sizeof(struct sa_entry);
 209         memcpy(dname, name, len);
 210         dname[len] = 0;
 211         entry->se_qstr.hash = ll_full_name_hash(parent, name, len);
 212         entry->se_qstr.len = len;
 213         entry->se_qstr.name = dname;
 214         entry->se_fid = *fid;
 215
 216         lli = ll_i2info(sai->sai_dentry->d_inode);
 217
 218         spin_lock(&lli->lli_sa_lock);
 219         INIT_LIST_HEAD(&entry->se_list);
 220         sa_rehash(sai, entry);
 221         spin_unlock(&lli->lli_sa_lock);
 222
 223         atomic_inc(&sai->sai_cache_count);
 224
 225         RETURN(entry);
 226 }
 227
 228 /* free sa_entry, which should have been unhashed and not in any list */
 229 static void sa_free(struct ll_statahead_info *sai, struct sa_entry *entry)
 230 {
 231         CDEBUG(D_READA, "free sa entry %.*s(%p) index %llu\n",
 232                entry->se_qstr.len, entry->se_qstr.name, entry,
 233                entry->se_index);
 234
 235         LASSERT(list_empty(&entry->se_list));
 236         LASSERT(sa_unhashed(entry));
 237
 238         OBD_FREE(entry, entry->se_size);
 239         atomic_dec(&sai->sai_cache_count);
 240 }
 241
 242 /*
 243  * find sa_entry by name, used by directory scanner, lock is not needed because
 244  * only scanner can remove the entry from cache.
 245  */
 246 static struct sa_entry *
 247 sa_get(struct ll_statahead_info *sai, const struct qstr *qstr)
 248 {
 249         struct sa_entry *entry;
 250         int i = sa_hash(qstr->hash);
 251
 252         list_for_each_entry(entry, &sai->sai_cache[i], se_hash) {
 253                 if (entry->se_qstr.hash == qstr->hash &&
 254                     entry->se_qstr.len == qstr->len &&
 255                     memcmp(entry->se_qstr.name, qstr->name, qstr->len) == 0)
 256                         return entry;
 257         }
 258         return NULL;
 259 }
 260
 261 /* unhash and unlink sa_entry, and then free it */
 262 static inline void
 263 sa_kill(struct ll_statahead_info *sai, struct sa_entry *entry)
 264 {
 265         struct ll_inode_info *lli = ll_i2info(sai->sai_dentry->d_inode);
 266
 267         LASSERT(!sa_unhashed(entry));
 268         LASSERT(!list_empty(&entry->se_list));
 269         LASSERT(sa_ready(entry));
 270
 271         sa_unhash(sai, entry);
 272
 273         spin_lock(&lli->lli_sa_lock);
 274         list_del_init(&entry->se_list);
 275         spin_unlock(&lli->lli_sa_lock);
 276
 277         if (entry->se_inode)
 278                 iput(entry->se_inode);
 279
 280         sa_free(sai, entry);
 281 }
 282
 283 /* called by scanner after use, sa_entry will be killed */
 284 static void
 285 sa_put(struct ll_statahead_info *sai, struct sa_entry *entry)
 286 {
 287         struct sa_entry *tmp, *next;
 288
 289         if (entry && entry->se_state == SA_ENTRY_SUCC) {
 290                 struct ll_sb_info *sbi = ll_i2sbi(sai->sai_dentry->d_inode);
 291
 292                 sai->sai_hit++;
 293                 sai->sai_consecutive_miss = 0;
 294                 sai->sai_max = min(2 * sai->sai_max, sbi->ll_sa_max);
 295         } else {
 296                 sai->sai_miss++;
 297                 sai->sai_consecutive_miss++;
 298         }
 299
 300         if (entry)
 301                 sa_kill(sai, entry);
 302
 303         /*
 304          * kill old completed entries, only scanner process does this, no need
 305          * to lock
 306          */
 307         list_for_each_entry_safe(tmp, next, &sai->sai_entries, se_list) {
 308                 if (!is_omitted_entry(sai, tmp->se_index))
 309                         break;
 310                 sa_kill(sai, tmp);
 311         }
 312 }
 313
 314 /*
 315  * update state and sort add entry to sai_entries by index, return true if
 316  * scanner is waiting on this entry.
 317  */
 318 static bool
 319 __sa_make_ready(struct ll_statahead_info *sai, struct sa_entry *entry, int ret)
 320 {
 321         struct sa_entry *se;
 322         struct list_head *pos = &sai->sai_entries;
 323         __u64 index = entry->se_index;
 324
 325         LASSERT(!sa_ready(entry));
 326         LASSERT(list_empty(&entry->se_list));
 327
 328         list_for_each_entry_reverse(se, &sai->sai_entries, se_list) {
 329                 if (se->se_index < entry->se_index) {
 330                         pos = &se->se_list;
 331                         break;
 332                 }
 333         }
 334         list_add(&entry->se_list, pos);
 335         /*
 336          * LU-9210: ll_statahead_interpet must be able to see this before
 337          * we wake it up
 338          */
 339         smp_store_release(&entry->se_state,
 340                           ret < 0 ? SA_ENTRY_INVA : SA_ENTRY_SUCC);
 341
 342         return (index == sai->sai_index_wait);
 343 }
 344
 345 /* finish async stat RPC arguments */
 346 static void sa_fini_data(struct md_enqueue_info *minfo)
 347 {
 348         ll_unlock_md_op_lsm(&minfo->mi_data);
 349         iput(minfo->mi_dir);
 350         OBD_FREE_PTR(minfo);
 351 }
 352
 353 static int ll_statahead_interpret(struct ptlrpc_request *req,
 354                                   struct md_enqueue_info *minfo, int rc);
 355
 356 /*
 357  * prepare arguments for async stat RPC.
 358  */
 359 static struct md_enqueue_info *
 360 sa_prep_data(struct inode *dir, struct inode *child, struct sa_entry *entry)
 361 {
 362         struct md_enqueue_info   *minfo;
 363         struct ldlm_enqueue_info *einfo;
 364         struct md_op_data        *op_data;
 365
 366         OBD_ALLOC_PTR(minfo);
 367         if (!minfo)
 368                 return ERR_PTR(-ENOMEM);
 369
 370         op_data = ll_prep_md_op_data(&minfo->mi_data, dir, child,
 371                                      entry->se_qstr.name, entry->se_qstr.len, 0,
 372                                      LUSTRE_OPC_ANY, NULL);
 373         if (IS_ERR(op_data)) {
 374                 OBD_FREE_PTR(minfo);
 375                 return (struct md_enqueue_info *)op_data;
 376         }
 377
 378         if (!child)
 379                 op_data->op_fid2 = entry->se_fid;
 380
 381         minfo->mi_it.it_op = IT_GETATTR;
 382         minfo->mi_dir = igrab(dir);
 383         minfo->mi_cb = ll_statahead_interpret;
 384         minfo->mi_cbdata = entry;
 385
 386         einfo = &minfo->mi_einfo;
 387         einfo->ei_type   = LDLM_IBITS;
 388         einfo->ei_mode   = it_to_lock_mode(&minfo->mi_it);
 389         einfo->ei_cb_bl  = ll_md_blocking_ast;
 390         einfo->ei_cb_cp  = ldlm_completion_ast;
 391         einfo->ei_cb_gl  = NULL;
 392         einfo->ei_cbdata = NULL;
 393
 394         return minfo;
 395 }
 396
 397 /*
 398  * release resources used in async stat RPC, update entry state and wakeup if
 399  * scanner process it waiting on this entry.
 400  */
 401 static void
 402 sa_make_ready(struct ll_statahead_info *sai, struct sa_entry *entry, int ret)
 403 {
 404         struct ll_inode_info *lli = ll_i2info(sai->sai_dentry->d_inode);
 405         struct md_enqueue_info *minfo = entry->se_minfo;
 406         struct ptlrpc_request *req = entry->se_req;
 407         bool wakeup;
 408
 409         /* release resources used in RPC */
 410         if (minfo) {
 411                 entry->se_minfo = NULL;
 412                 ll_intent_release(&minfo->mi_it);
 413                 sa_fini_data(minfo);
 414         }
 415
 416         if (req) {
 417                 entry->se_req = NULL;
 418                 ptlrpc_req_finished(req);
 419         }
 420
 421         spin_lock(&lli->lli_sa_lock);
 422         wakeup = __sa_make_ready(sai, entry, ret);
 423         spin_unlock(&lli->lli_sa_lock);
 424
 425         if (wakeup)
 426                 wake_up(&sai->sai_waitq);
 427 }
 428
 429 /* insert inode into the list of sai_agls */
 430 static void ll_agl_add(struct ll_statahead_info *sai,
 431                        struct inode *inode, int index)
 432 {
 433         struct ll_inode_info *child  = ll_i2info(inode);
 434         struct ll_inode_info *parent = ll_i2info(sai->sai_dentry->d_inode);
 435
 436         spin_lock(&child->lli_agl_lock);
 437         if (child->lli_agl_index == 0) {
 438                 child->lli_agl_index = index;
 439                 spin_unlock(&child->lli_agl_lock);
 440
 441                 LASSERT(list_empty(&child->lli_agl_list));
 442
 443                 spin_lock(&parent->lli_agl_lock);
 444                 /* Re-check under the lock */
 445                 if (agl_should_run(sai, inode)) {
 446                         if (agl_list_empty(sai))
 447                                 wake_up_process(sai->sai_agl_task);
 448                         igrab(inode);
 449                         list_add_tail(&child->lli_agl_list, &sai->sai_agls);
 450                 } else
 451                         child->lli_agl_index = 0;
 452                 spin_unlock(&parent->lli_agl_lock);
 453         } else {
 454                 spin_unlock(&child->lli_agl_lock);
 455         }
 456 }
 457
 458 /* allocate sai */
 459 static struct ll_statahead_info *ll_sai_alloc(struct dentry *dentry)
 460 {
 461         struct ll_statahead_info *sai;
 462         struct ll_inode_info *lli = ll_i2info(dentry->d_inode);
 463         int i;
 464
 465         ENTRY;
 466
 467         OBD_ALLOC_PTR(sai);
 468         if (!sai)
 469                 RETURN(NULL);
 470
 471         sai->sai_dentry = dget(dentry);
 472         atomic_set(&sai->sai_refcount, 1);
 473         sai->sai_max = LL_SA_RPC_MIN;
 474         sai->sai_index = 1;
 475         init_waitqueue_head(&sai->sai_waitq);
 476
 477         INIT_LIST_HEAD(&sai->sai_interim_entries);
 478         INIT_LIST_HEAD(&sai->sai_entries);
 479         INIT_LIST_HEAD(&sai->sai_agls);
 480
 481         for (i = 0; i < LL_SA_CACHE_SIZE; i++) {
 482                 INIT_LIST_HEAD(&sai->sai_cache[i]);
 483                 spin_lock_init(&sai->sai_cache_lock[i]);
 484         }
 485         atomic_set(&sai->sai_cache_count, 0);
 486
 487         spin_lock(&sai_generation_lock);
 488         lli->lli_sa_generation = ++sai_generation;
 489         if (unlikely(sai_generation == 0))
 490                 lli->lli_sa_generation = ++sai_generation;
 491         spin_unlock(&sai_generation_lock);
 492
 493         RETURN(sai);
 494 }
 495
 496 /* free sai */
 497 static inline void ll_sai_free(struct ll_statahead_info *sai)
 498 {
 499         LASSERT(sai->sai_dentry != NULL);
 500         dput(sai->sai_dentry);
 501         OBD_FREE_PTR(sai);
 502 }
 503
 504 /*
 505  * take refcount of sai if sai for @dir exists, which means statahead is on for
 506  * this directory.
 507  */
 508 static inline struct ll_statahead_info *ll_sai_get(struct inode *dir)
 509 {
 510         struct ll_inode_info *lli = ll_i2info(dir);
 511         struct ll_statahead_info *sai = NULL;
 512
 513         spin_lock(&lli->lli_sa_lock);
 514         sai = lli->lli_sai;
 515         if (sai)
 516                 atomic_inc(&sai->sai_refcount);
 517         spin_unlock(&lli->lli_sa_lock);
 518
 519         return sai;
 520 }
 521
 522 /*
 523  * put sai refcount after use, if refcount reaches zero, free sai and sa_entries
 524  * attached to it.
 525  */
 526 static void ll_sai_put(struct ll_statahead_info *sai)
 527 {
 528         struct ll_inode_info *lli = ll_i2info(sai->sai_dentry->d_inode);
 529
 530         if (atomic_dec_and_lock(&sai->sai_refcount, &lli->lli_sa_lock)) {
 531                 struct sa_entry *entry, *next;
 532                 struct ll_sb_info *sbi = ll_i2sbi(sai->sai_dentry->d_inode);
 533
 534                 lli->lli_sai = NULL;
 535                 spin_unlock(&lli->lli_sa_lock);
 536
 537                 LASSERT(!sai->sai_task);
 538                 LASSERT(!sai->sai_agl_task);
 539                 LASSERT(sai->sai_sent == sai->sai_replied);
 540                 LASSERT(!sa_has_callback(sai));
 541
 542                 list_for_each_entry_safe(entry, next, &sai->sai_entries,
 543                                          se_list)
 544                         sa_kill(sai, entry);
 545
 546                 LASSERT(atomic_read(&sai->sai_cache_count) == 0);
 547                 LASSERT(agl_list_empty(sai));
 548
 549                 ll_sai_free(sai);
 550                 atomic_dec(&sbi->ll_sa_running);
 551         }
 552 }
 553
 554 /* Do NOT forget to drop inode refcount when into sai_agls. */
 555 static void ll_agl_trigger(struct inode *inode, struct ll_statahead_info *sai)
 556 {
 557         struct ll_inode_info *lli = ll_i2info(inode);
 558         u64 index = lli->lli_agl_index;
 559         ktime_t expire;
 560         int rc;
 561
 562         ENTRY;
 563
 564         LASSERT(list_empty(&lli->lli_agl_list));
 565
 566         /* AGL maybe fall behind statahead with one entry */
 567         if (is_omitted_entry(sai, index + 1)) {
 568                 lli->lli_agl_index = 0;
 569                 iput(inode);
 570                 RETURN_EXIT;
 571         }
 572
 573         /*
 574          * In case of restore, the MDT has the right size and has already
 575          * sent it back without granting the layout lock, inode is up-to-date.
 576          * Then AGL (async glimpse lock) is useless.
 577          * Also to glimpse we need the layout, in case of a runninh restore
 578          * the MDT holds the layout lock so the glimpse will block up to the
 579          * end of restore (statahead/agl will block)
 580          */
 581         if (ll_file_test_flag(lli, LLIF_FILE_RESTORING)) {
 582                 lli->lli_agl_index = 0;
 583                 iput(inode);
 584                 RETURN_EXIT;
 585         }
 586
 587         /* Someone is in glimpse (sync or async), do nothing. */
 588         rc = down_write_trylock(&lli->lli_glimpse_sem);
 589         if (rc == 0) {
 590                 lli->lli_agl_index = 0;
 591                 iput(inode);
 592                 RETURN_EXIT;
 593         }
 594
 595         /*
 596          * Someone triggered glimpse within 1 sec before.
 597          * 1) The former glimpse succeeded with glimpse lock granted by OST, and
 598          *    if the lock is still cached on client, AGL needs to do nothing. If
 599          *    it is cancelled by other client, AGL maybe cannot obtaion new lock
 600          *    for no glimpse callback triggered by AGL.
 601          * 2) The former glimpse succeeded, but OST did not grant glimpse lock.
 602          *    Under such case, it is quite possible that the OST will not grant
 603          *    glimpse lock for AGL also.
 604          * 3) The former glimpse failed, compared with other two cases, it is
 605          *    relative rare. AGL can ignore such case, and it will not muchly
 606          *    affect the performance.
 607          */
 608         expire = ktime_sub_ns(ktime_get(), NSEC_PER_SEC);
 609         if (ktime_to_ns(lli->lli_glimpse_time) &&
 610             ktime_before(expire, lli->lli_glimpse_time)) {
 611                 up_write(&lli->lli_glimpse_sem);
 612                 lli->lli_agl_index = 0;
 613                 iput(inode);
 614                 RETURN_EXIT;
 615         }
 616
 617         CDEBUG(D_READA,
 618                "Handling (init) async glimpse: inode = " DFID", idx = %llu\n",
 619                PFID(&lli->lli_fid), index);
 620
 621         cl_agl(inode);
 622         lli->lli_agl_index = 0;
 623         lli->lli_glimpse_time = ktime_get();
 624         up_write(&lli->lli_glimpse_sem);
 625
 626         CDEBUG(D_READA,
 627                "Handled (init) async glimpse: inode= " DFID", idx = %llu, rc = %d\n",
 628                PFID(&lli->lli_fid), index, rc);
 629
 630         iput(inode);
 631
 632         EXIT;
 633 }
 634
 635 /*
 636  * prepare inode for sa entry, add it into agl list, now sa_entry is ready
 637  * to be used by scanner process.
 638  */
 639 static void sa_instantiate(struct ll_statahead_info *sai,
 640                            struct sa_entry *entry)
 641 {
 642         struct inode *dir = sai->sai_dentry->d_inode;
 643         struct inode *child;
 644         struct md_enqueue_info *minfo;
 645         struct lookup_intent *it;
 646         struct ptlrpc_request *req;
 647         struct mdt_body *body;
 648         int rc = 0;
 649
 650         ENTRY;
 651
 652         LASSERT(entry->se_handle != 0);
 653
 654         minfo = entry->se_minfo;
 655         it = &minfo->mi_it;
 656         req = entry->se_req;
 657         body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
 658         if (!body)
 659                 GOTO(out, rc = -EFAULT);
 660
 661         child = entry->se_inode;
 662         /* revalidate; unlinked and re-created with the same name */
 663         if (unlikely(!lu_fid_eq(&minfo->mi_data.op_fid2, &body->mbo_fid1))) {
 664                 if (child) {
 665                         entry->se_inode = NULL;
 666                         iput(child);
 667                 }
 668                 /* The mdt_body is invalid. Skip this entry */
 669                 GOTO(out, rc = -EAGAIN);
 670         }
 671
 672         it->it_lock_handle = entry->se_handle;
 673         rc = md_revalidate_lock(ll_i2mdexp(dir), it, ll_inode2fid(dir), NULL);
 674         if (rc != 1)
 675                 GOTO(out, rc = -EAGAIN);
 676
 677         rc = ll_prep_inode(&child, req, dir->i_sb, it);
 678         if (rc)
 679                 GOTO(out, rc);
 680
 681         CDEBUG(D_READA, "%s: setting %.*s"DFID" l_data to inode %p\n",
 682                ll_i2sbi(dir)->ll_fsname, entry->se_qstr.len,
 683                entry->se_qstr.name, PFID(ll_inode2fid(child)), child);
 684         ll_set_lock_data(ll_i2sbi(dir)->ll_md_exp, child, it, NULL);
 685
 686         entry->se_inode = child;
 687
 688         if (agl_should_run(sai, child))
 689                 ll_agl_add(sai, child, entry->se_index);
 690
 691         EXIT;
 692
 693 out:
 694         /*
 695          * sa_make_ready() will drop ldlm ibits lock refcount by calling
 696          * ll_intent_drop_lock() in spite of failures. Do not worry about
 697          * calling ll_intent_drop_lock() more than once.
 698          */
 699         sa_make_ready(sai, entry, rc);
 700 }
 701
 702 /* once there are async stat replies, instantiate sa_entry from replies */
 703 static void sa_handle_callback(struct ll_statahead_info *sai)
 704 {
 705         struct ll_inode_info *lli;
 706
 707         lli = ll_i2info(sai->sai_dentry->d_inode);
 708
 709         spin_lock(&lli->lli_sa_lock);
 710         while (sa_has_callback(sai)) {
 711                 struct sa_entry *entry;
 712
 713                 entry = list_entry(sai->sai_interim_entries.next,
 714                                    struct sa_entry, se_list);
 715                 list_del_init(&entry->se_list);
 716                 spin_unlock(&lli->lli_sa_lock);
 717
 718                 sa_instantiate(sai, entry);
 719                 spin_lock(&lli->lli_sa_lock);
 720         }
 721         spin_unlock(&lli->lli_sa_lock);
 722 }
 723
 724 /*
 725  * callback for async stat RPC, because this is called in ptlrpcd context, we
 726  * only put sa_entry in sai_interim_entries, and wake up statahead thread to
 727  * really prepare inode and instantiate sa_entry later.
 728  */
 729 static int ll_statahead_interpret(struct ptlrpc_request *req,
 730                                   struct md_enqueue_info *minfo, int rc)
 731 {
 732         struct lookup_intent *it = &minfo->mi_it;
 733         struct inode *dir = minfo->mi_dir;
 734         struct ll_inode_info *lli = ll_i2info(dir);
 735         struct ll_statahead_info *sai = lli->lli_sai;
 736         struct sa_entry *entry = (struct sa_entry *)minfo->mi_cbdata;
 737         __u64 handle = 0;
 738
 739         ENTRY;
 740
 741         if (it_disposition(it, DISP_LOOKUP_NEG))
 742                 rc = -ENOENT;
 743
 744         /*
 745          * because statahead thread will wait for all inflight RPC to finish,
 746          * sai should be always valid, no need to refcount
 747          */
 748         LASSERT(sai != NULL);
 749         LASSERT(entry != NULL);
 750
 751         CDEBUG(D_READA, "sa_entry %.*s rc %d\n",
 752                entry->se_qstr.len, entry->se_qstr.name, rc);
 753
 754         if (rc != 0) {
 755                 ll_intent_release(it);
 756                 sa_fini_data(minfo);
 757         } else {
 758                 /*
 759                  * release ibits lock ASAP to avoid deadlock when statahead
 760                  * thread enqueues lock on parent in readdir and another
 761                  * process enqueues lock on child with parent lock held, eg.
 762                  * unlink.
 763                  */
 764                 handle = it->it_lock_handle;
 765                 ll_intent_drop_lock(it);
 766                 ll_unlock_md_op_lsm(&minfo->mi_data);
 767         }
 768
 769         spin_lock(&lli->lli_sa_lock);
 770         if (rc != 0) {
 771                 if (__sa_make_ready(sai, entry, rc))
 772                         wake_up(&sai->sai_waitq);
 773         } else {
 774                 int first = 0;
 775
 776                 entry->se_minfo = minfo;
 777                 entry->se_req = ptlrpc_request_addref(req);
 778                 /*
 779                  * Release the async ibits lock ASAP to avoid deadlock
 780                  * when statahead thread tries to enqueue lock on parent
 781                  * for readpage and other tries to enqueue lock on child
 782                  * with parent's lock held, for example: unlink.
 783                  */
 784                 entry->se_handle = handle;
 785                 if (!sa_has_callback(sai))
 786                         first = 1;
 787
 788                 list_add_tail(&entry->se_list, &sai->sai_interim_entries);
 789                 if (first && sai->sai_task)
 790                         wake_up_process(sai->sai_task);
 791         }
 792         sai->sai_replied++;
 793
 794         spin_unlock(&lli->lli_sa_lock);
 795
 796         RETURN(rc);
 797 }
 798
 799 /* async stat for file not found in dcache */
 800 static int sa_lookup(struct inode *dir, struct sa_entry *entry)
 801 {
 802         struct md_enqueue_info   *minfo;
 803         int                       rc;
 804
 805         ENTRY;
 806
 807         minfo = sa_prep_data(dir, NULL, entry);
 808         if (IS_ERR(minfo))
 809                 RETURN(PTR_ERR(minfo));
 810
 811         rc = md_intent_getattr_async(ll_i2mdexp(dir), minfo);
 812         if (rc < 0)
 813                 sa_fini_data(minfo);
 814
 815         RETURN(rc);
 816 }
 817
 818 /**
 819  * async stat for file found in dcache, similar to .revalidate
 820  *
 821  * \retval      1 dentry valid, no RPC sent
 822  * \retval      0 dentry invalid, will send async stat RPC
 823  * \retval      negative number upon error
 824  */
 825 static int sa_revalidate(struct inode *dir, struct sa_entry *entry,
 826                          struct dentry *dentry)
 827 {
 828         struct inode *inode = dentry->d_inode;
 829         struct lookup_intent it = { .it_op = IT_GETATTR,
 830                                     .it_lock_handle = 0 };
 831         struct md_enqueue_info *minfo;
 832         int rc;
 833
 834         ENTRY;
 835
 836         if (unlikely(!inode))
 837                 RETURN(1);
 838
 839         if (d_mountpoint(dentry))
 840                 RETURN(1);
 841
 842         minfo = sa_prep_data(dir, inode, entry);
 843         if (IS_ERR(minfo))
 844                 RETURN(PTR_ERR(minfo));
 845
 846         entry->se_inode = igrab(inode);
 847         rc = md_revalidate_lock(ll_i2mdexp(dir), &it, ll_inode2fid(inode),
 848                                 NULL);
 849         if (rc == 1) {
 850                 entry->se_handle = it.it_lock_handle;
 851                 ll_intent_release(&it);
 852                 sa_fini_data(minfo);
 853                 RETURN(1);
 854         }
 855
 856         rc = md_intent_getattr_async(ll_i2mdexp(dir), minfo);
 857         if (rc < 0) {
 858                 entry->se_inode = NULL;
 859                 iput(inode);
 860                 sa_fini_data(minfo);
 861         }
 862
 863         RETURN(rc);
 864 }
 865
 866 /* async stat for file with @name */
 867 static void sa_statahead(struct dentry *parent, const char *name, int len,
 868                          const struct lu_fid *fid)
 869 {
 870         struct inode *dir = parent->d_inode;
 871         struct ll_inode_info *lli = ll_i2info(dir);
 872         struct ll_statahead_info *sai = lli->lli_sai;
 873         struct dentry *dentry = NULL;
 874         struct sa_entry *entry;
 875         int rc;
 876
 877         ENTRY;
 878
 879         entry = sa_alloc(parent, sai, sai->sai_index, name, len, fid);
 880         if (IS_ERR(entry))
 881                 RETURN_EXIT;
 882
 883         dentry = d_lookup(parent, &entry->se_qstr);
 884         if (!dentry) {
 885                 rc = sa_lookup(dir, entry);
 886         } else {
 887                 rc = sa_revalidate(dir, entry, dentry);
 888                 if (rc == 1 && agl_should_run(sai, dentry->d_inode))
 889                         ll_agl_add(sai, dentry->d_inode, entry->se_index);
 890         }
 891
 892         if (dentry)
 893                 dput(dentry);
 894
 895         if (rc != 0)
 896                 sa_make_ready(sai, entry, rc);
 897         else
 898                 sai->sai_sent++;
 899
 900         sai->sai_index++;
 901
 902         EXIT;
 903 }
 904
 905 /* async glimpse (agl) thread main function */
 906 static int ll_agl_thread(void *arg)
 907 {
 908         struct dentry *parent = (struct dentry *)arg;
 909         struct inode *dir = parent->d_inode;
 910         struct ll_inode_info *plli = ll_i2info(dir);
 911         struct ll_inode_info *clli;
 912         /*
 913          * We already own this reference, so it is safe to take it
 914          * without a lock.
 915          */
 916         struct ll_statahead_info *sai = plli->lli_sai;
 917
 918         ENTRY;
 919
 920         CDEBUG(D_READA, "agl thread started: sai %p, parent %pd\n",
 921                sai, parent);
 922
 923         while (({set_current_state(TASK_IDLE);
 924                  !kthread_should_stop(); })) {
 925                 spin_lock(&plli->lli_agl_lock);
 926                 if (!agl_list_empty(sai)) {
 927                         __set_current_state(TASK_RUNNING);
 928                         clli = agl_first_entry(sai);
 929                         list_del_init(&clli->lli_agl_list);
 930                         spin_unlock(&plli->lli_agl_lock);
 931                         ll_agl_trigger(&clli->lli_vfs_inode, sai);
 932                         cond_resched();
 933                 } else {
 934                         spin_unlock(&plli->lli_agl_lock);
 935                         schedule();
 936                 }
 937         }
 938         __set_current_state(TASK_RUNNING);
 939         RETURN(0);
 940 }
 941
 942 static void ll_stop_agl(struct ll_statahead_info *sai)
 943 {
 944         struct dentry *parent = sai->sai_dentry;
 945         struct ll_inode_info *plli = ll_i2info(parent->d_inode);
 946         struct ll_inode_info *clli;
 947         struct task_struct *agl_task;
 948
 949         spin_lock(&plli->lli_agl_lock);
 950         agl_task = sai->sai_agl_task;
 951         sai->sai_agl_task = NULL;
 952         spin_unlock(&plli->lli_agl_lock);
 953         if (!agl_task)
 954                 return;
 955
 956         CDEBUG(D_READA, "stop agl thread: sai %p pid %u\n",
 957                sai, (unsigned int)agl_task->pid);
 958         kthread_stop(agl_task);
 959
 960         spin_lock(&plli->lli_agl_lock);
 961         while (!agl_list_empty(sai)) {
 962                 clli = agl_first_entry(sai);
 963                 list_del_init(&clli->lli_agl_list);
 964                 spin_unlock(&plli->lli_agl_lock);
 965                 clli->lli_agl_index = 0;
 966                 iput(&clli->lli_vfs_inode);
 967                 spin_lock(&plli->lli_agl_lock);
 968         }
 969         spin_unlock(&plli->lli_agl_lock);
 970         CDEBUG(D_READA, "agl thread stopped: sai %p, parent %pd\n",
 971                sai, parent);
 972         ll_sai_put(sai);
 973 }
 974
 975 /* start agl thread */
 976 static void ll_start_agl(struct dentry *parent, struct ll_statahead_info *sai)
 977 {
 978         int node = cfs_cpt_spread_node(cfs_cpt_tab, CFS_CPT_ANY);
 979         struct ll_inode_info *plli;
 980         struct task_struct *task;
 981
 982         ENTRY;
 983
 984         CDEBUG(D_READA, "start agl thread: sai %p, parent %pd\n",
 985                sai, parent);
 986
 987         plli = ll_i2info(parent->d_inode);
 988         task = kthread_create_on_node(ll_agl_thread, parent, node, "ll_agl_%d",
 989                                       plli->lli_opendir_pid);
 990         if (IS_ERR(task)) {
 991                 CERROR("can't start ll_agl thread, rc: %ld\n", PTR_ERR(task));
 992                 RETURN_EXIT;
 993         }
 994         sai->sai_agl_task = task;
 995         atomic_inc(&ll_i2sbi(d_inode(parent))->ll_agl_total);
 996         /* Get an extra reference that the thread holds */
 997         ll_sai_get(d_inode(parent));
 998
 999         wake_up_process(task);
1000
1001         EXIT;
1002 }
1003
1004 /* statahead thread main function */
1005 static int ll_statahead_thread(void *arg)
1006 {
1007         struct dentry *parent = (struct dentry *)arg;
1008         struct inode *dir = parent->d_inode;
1009         struct ll_inode_info *lli = ll_i2info(dir);
1010         struct ll_sb_info *sbi = ll_i2sbi(dir);
1011         struct ll_statahead_info *sai = lli->lli_sai;
1012         int first = 0;
1013         struct md_op_data *op_data;
1014         struct ll_dir_chain chain;
1015         struct page *page = NULL;
1016         __u64 pos = 0;
1017         int rc = 0;
1018
1019         ENTRY;
1020
1021         CDEBUG(D_READA, "statahead thread starting: sai %p, parent %pd\n",
1022                sai, parent);
1023
1024         OBD_ALLOC_PTR(op_data);
1025         if (!op_data)
1026                 GOTO(out, rc = -ENOMEM);
1027
1028         ll_dir_chain_init(&chain);
1029         while (pos != MDS_DIR_END_OFF && sai->sai_task) {
1030                 struct lu_dirpage *dp;
1031                 struct lu_dirent  *ent;
1032
1033                 op_data = ll_prep_md_op_data(op_data, dir, dir, NULL, 0, 0,
1034                                              LUSTRE_OPC_ANY, dir);
1035                 if (IS_ERR(op_data)) {
1036                         rc = PTR_ERR(op_data);
1037                         break;
1038                 }
1039
1040                 sai->sai_in_readpage = 1;
1041                 page = ll_get_dir_page(dir, op_data, pos, &chain);
1042                 ll_unlock_md_op_lsm(op_data);
1043                 sai->sai_in_readpage = 0;
1044                 if (IS_ERR(page)) {
1045                         rc = PTR_ERR(page);
1046                         CDEBUG(D_READA,
1047                                "error reading dir "DFID" at %llu /%llu opendir_pid = %u: rc = %d\n",
1048                                PFID(ll_inode2fid(dir)), pos, sai->sai_index,
1049                                lli->lli_opendir_pid, rc);
1050                         break;
1051                 }
1052
1053                 dp = page_address(page);
1054                 for (ent = lu_dirent_start(dp);
1055                      ent != NULL && sai->sai_task &&
1056                      !sa_low_hit(sai);
1057                      ent = lu_dirent_next(ent)) {
1058                         __u64 hash;
1059                         int namelen;
1060                         char *name;
1061                         struct lu_fid fid;
1062
1063                         hash = le64_to_cpu(ent->lde_hash);
1064                         if (unlikely(hash < pos))
1065                                 /*
1066                                  * Skip until we find target hash value.
1067                                  */
1068                                 continue;
1069
1070                         namelen = le16_to_cpu(ent->lde_namelen);
1071                         if (unlikely(namelen == 0))
1072                                 /*
1073                                  * Skip dummy record.
1074                                  */
1075                                 continue;
1076
1077                         name = ent->lde_name;
1078                         if (name[0] == '.') {
1079                                 if (namelen == 1) {
1080                                         /*
1081                                          * skip "."
1082                                          */
1083                                         continue;
1084                                 } else if (name[1] == '.' && namelen == 2) {
1085                                         /*
1086                                          * skip ".."
1087                                          */
1088                                         continue;
1089                                 } else if (!sai->sai_ls_all) {
1090                                         /*
1091                                          * skip hidden files.
1092                                          */
1093                                         sai->sai_skip_hidden++;
1094                                         continue;
1095                                 }
1096                         }
1097
1098                         /*
1099                          * don't stat-ahead first entry.
1100                          */
1101                         if (unlikely(++first == 1))
1102                                 continue;
1103
1104                         fid_le_to_cpu(&fid, &ent->lde_fid);
1105
1106                         while (({set_current_state(TASK_IDLE);
1107                                  sai->sai_task; })) {
1108                                 if (sa_has_callback(sai)) {
1109                                         __set_current_state(TASK_RUNNING);
1110                                         sa_handle_callback(sai);
1111                                 }
1112
1113                                 spin_lock(&lli->lli_agl_lock);
1114                                 while (sa_sent_full(sai) &&
1115                                        !agl_list_empty(sai)) {
1116                                         struct ll_inode_info *clli;
1117
1118                                         __set_current_state(TASK_RUNNING);
1119                                         clli = agl_first_entry(sai);
1120                                         list_del_init(&clli->lli_agl_list);
1121                                         spin_unlock(&lli->lli_agl_lock);
1122
1123                                         ll_agl_trigger(&clli->lli_vfs_inode,
1124                                                        sai);
1125                                         cond_resched();
1126                                         spin_lock(&lli->lli_agl_lock);
1127                                 }
1128                                 spin_unlock(&lli->lli_agl_lock);
1129
1130                                 if (!sa_sent_full(sai))
1131                                         break;
1132                                 schedule();
1133                         }
1134                         __set_current_state(TASK_RUNNING);
1135
1136                         sa_statahead(parent, name, namelen, &fid);
1137                 }
1138
1139                 pos = le64_to_cpu(dp->ldp_hash_end);
1140                 ll_release_page(dir, page,
1141                                 le32_to_cpu(dp->ldp_flags) & LDF_COLLIDE);
1142
1143                 if (sa_low_hit(sai)) {
1144                         rc = -EFAULT;
1145                         atomic_inc(&sbi->ll_sa_wrong);
1146                         CDEBUG(D_READA,
1147                                "Statahead for dir "DFID" hit ratio too low: hit/miss %llu/%llu, sent/replied %llu/%llu, stoppingstatahead thread: pid %d\n",
1148                                PFID(&lli->lli_fid), sai->sai_hit,
1149                                sai->sai_miss, sai->sai_sent,
1150                                sai->sai_replied, current->pid);
1151                         break;
1152                 }
1153         }
1154         ll_dir_chain_fini(&chain);
1155         ll_finish_md_op_data(op_data);
1156
1157         if (rc < 0) {
1158                 spin_lock(&lli->lli_sa_lock);
1159                 sai->sai_task = NULL;
1160                 lli->lli_sa_enabled = 0;
1161                 spin_unlock(&lli->lli_sa_lock);
1162         }
1163
1164         /*
1165          * statahead is finished, but statahead entries need to be cached, wait
1166          * for file release to stop me.
1167          */
1168         while (({set_current_state(TASK_IDLE);
1169                  sai->sai_task; })) {
1170                 if (sa_has_callback(sai)) {
1171                         __set_current_state(TASK_RUNNING);
1172                         sa_handle_callback(sai);
1173                 } else {
1174                         schedule();
1175                 }
1176         }
1177         __set_current_state(TASK_RUNNING);
1178
1179         EXIT;
1180 out:
1181         ll_stop_agl(sai);
1182
1183         /*
1184          * wait for inflight statahead RPCs to finish, and then we can free sai
1185          * safely because statahead RPC will access sai data
1186          */
1187         while (sai->sai_sent != sai->sai_replied)
1188                 /* in case we're not woken up, timeout wait */
1189                 msleep(125);
1190
1191         /* release resources held by statahead RPCs */
1192         sa_handle_callback(sai);
1193
1194         CDEBUG(D_READA, "%s: statahead thread stopped: sai %p, parent %pd\n",
1195                sbi->ll_fsname, sai, parent);
1196
1197         spin_lock(&lli->lli_sa_lock);
1198         sai->sai_task = NULL;
1199         spin_unlock(&lli->lli_sa_lock);
1200         wake_up(&sai->sai_waitq);
1201
1202         ll_sai_put(sai);
1203
1204         return rc;
1205 }
1206
1207 /* authorize opened dir handle @key to statahead */
1208 void ll_authorize_statahead(struct inode *dir, void *key)
1209 {
1210         struct ll_inode_info *lli = ll_i2info(dir);
1211
1212         spin_lock(&lli->lli_sa_lock);
1213         if (!lli->lli_opendir_key && !lli->lli_sai) {
1214                 /*
1215                  * if lli_sai is not NULL, it means previous statahead is not
1216                  * finished yet, we'd better not start a new statahead for now.
1217                  */
1218                 LASSERT(lli->lli_opendir_pid == 0);
1219                 lli->lli_opendir_key = key;
1220                 lli->lli_opendir_pid = current->pid;
1221                 lli->lli_sa_enabled = 1;
1222         }
1223         spin_unlock(&lli->lli_sa_lock);
1224 }
1225
1226 /*
1227  * deauthorize opened dir handle @key to statahead, and notify statahead thread
1228  * to quit if it's running.
1229  */
1230 void ll_deauthorize_statahead(struct inode *dir, void *key)
1231 {
1232         struct ll_inode_info *lli = ll_i2info(dir);
1233         struct ll_statahead_info *sai;
1234
1235         LASSERT(lli->lli_opendir_key == key);
1236         LASSERT(lli->lli_opendir_pid != 0);
1237
1238         CDEBUG(D_READA, "deauthorize statahead for "DFID"\n",
1239                PFID(&lli->lli_fid));
1240
1241         spin_lock(&lli->lli_sa_lock);
1242         lli->lli_opendir_key = NULL;
1243         lli->lli_opendir_pid = 0;
1244         lli->lli_sa_enabled = 0;
1245         sai = lli->lli_sai;
1246         if (sai && sai->sai_task) {
1247                 /*
1248                  * statahead thread may not have quit yet because it needs to
1249                  * cache entries, now it's time to tell it to quit.
1250                  *
1251                  * wake_up_process() provides the necessary barriers
1252                  * to pair with set_current_state().
1253                  */
1254                 struct task_struct *task = sai->sai_task;
1255
1256                 sai->sai_task = NULL;
1257                 wake_up_process(task);
1258         }
1259         spin_unlock(&lli->lli_sa_lock);
1260 }
1261
1262 enum {
1263         /**
1264          * not first dirent, or is "."
1265          */
1266         LS_NOT_FIRST_DE = 0,
1267         /**
1268          * the first non-hidden dirent
1269          */
1270         LS_FIRST_DE,
1271         /**
1272          * the first hidden dirent, that is "."
1273          */
1274         LS_FIRST_DOT_DE
1275 };
1276
1277 /* file is first dirent under @dir */
1278 static int is_first_dirent(struct inode *dir, struct dentry *dentry)
1279 {
1280         struct ll_dir_chain   chain;
1281         struct qstr          *target = &dentry->d_name;
1282         struct md_op_data    *op_data;
1283         int                   dot_de;
1284         struct page          *page = NULL;
1285         int                   rc = LS_NOT_FIRST_DE;
1286         __u64                 pos = 0;
1287
1288         ENTRY;
1289
1290         op_data = ll_prep_md_op_data(NULL, dir, dir, NULL, 0, 0,
1291                                      LUSTRE_OPC_ANY, dir);
1292         if (IS_ERR(op_data))
1293                 RETURN(PTR_ERR(op_data));
1294         /**
1295          *FIXME choose the start offset of the readdir
1296          */
1297
1298         ll_dir_chain_init(&chain);
1299         page = ll_get_dir_page(dir, op_data, 0, &chain);
1300
1301         while (1) {
1302                 struct lu_dirpage *dp;
1303                 struct lu_dirent  *ent;
1304
1305                 if (IS_ERR(page)) {
1306                         struct ll_inode_info *lli = ll_i2info(dir);
1307
1308                         rc = PTR_ERR(page);
1309                         CERROR("%s: reading dir "DFID" at %llu opendir_pid = %u : rc = %d\n",
1310                                ll_i2sbi(dir)->ll_fsname,
1311                                PFID(ll_inode2fid(dir)), pos,
1312                                lli->lli_opendir_pid, rc);
1313                         break;
1314                 }
1315
1316                 dp = page_address(page);
1317                 for (ent = lu_dirent_start(dp); ent != NULL;
1318                      ent = lu_dirent_next(ent)) {
1319                         __u64 hash;
1320                         int namelen;
1321                         char *name;
1322
1323                         hash = le64_to_cpu(ent->lde_hash);
1324                         /*
1325                          * The ll_get_dir_page() can return any page containing
1326                          * the given hash which may be not the start hash.
1327                          */
1328                         if (unlikely(hash < pos))
1329                                 continue;
1330
1331                         namelen = le16_to_cpu(ent->lde_namelen);
1332                         if (unlikely(namelen == 0))
1333                                 /*
1334                                  * skip dummy record.
1335                                  */
1336                                 continue;
1337
1338                         name = ent->lde_name;
1339                         if (name[0] == '.') {
1340                                 if (namelen == 1)
1341                                         /*
1342                                          * skip "."
1343                                          */
1344                                         continue;
1345                                 else if (name[1] == '.' && namelen == 2)
1346                                         /*
1347                                          * skip ".."
1348                                          */
1349                                         continue;
1350                                 else
1351                                         dot_de = 1;
1352                         } else {
1353                                 dot_de = 0;
1354                         }
1355
1356                         if (dot_de && target->name[0] != '.') {
1357                                 CDEBUG(D_READA, "%.*s skip hidden file %.*s\n",
1358                                        target->len, target->name,
1359                                        namelen, name);
1360                                 continue;
1361                         }
1362
1363                         if (target->len != namelen ||
1364                             memcmp(target->name, name, namelen) != 0)
1365                                 rc = LS_NOT_FIRST_DE;
1366                         else if (!dot_de)
1367                                 rc = LS_FIRST_DE;
1368                         else
1369                                 rc = LS_FIRST_DOT_DE;
1370
1371                         ll_release_page(dir, page, false);
1372                         GOTO(out, rc);
1373                 }
1374                 pos = le64_to_cpu(dp->ldp_hash_end);
1375                 if (pos == MDS_DIR_END_OFF) {
1376                         /*
1377                          * End of directory reached.
1378                          */
1379                         ll_release_page(dir, page, false);
1380                         GOTO(out, rc);
1381                 } else {
1382                         /*
1383                          * chain is exhausted
1384                          * Normal case: continue to the next page.
1385                          */
1386                         ll_release_page(dir, page, le32_to_cpu(dp->ldp_flags) &
1387                                               LDF_COLLIDE);
1388                         page = ll_get_dir_page(dir, op_data, pos, &chain);
1389                 }
1390         }
1391         EXIT;
1392 out:
1393         ll_dir_chain_fini(&chain);
1394         ll_finish_md_op_data(op_data);
1395
1396         return rc;
1397 }
1398
1399 /**
1400  * revalidate @dentryp from statahead cache
1401  *
1402  * \param[in] dir       parent directory
1403  * \param[in] sai       sai structure
1404  * \param[out] dentryp  pointer to dentry which will be revalidated
1405  * \param[in] unplug    unplug statahead window only (normally for negative
1406  *                      dentry)
1407  * \retval              1 on success, dentry is saved in @dentryp
1408  * \retval              0 if revalidation failed (no proper lock on client)
1409  * \retval              negative number upon error
1410  */
1411 static int revalidate_statahead_dentry(struct inode *dir,
1412                                        struct ll_statahead_info *sai,
1413                                        struct dentry **dentryp,
1414                                        bool unplug)
1415 {
1416         struct sa_entry *entry = NULL;
1417         struct ll_dentry_data *ldd;
1418         struct ll_inode_info *lli = ll_i2info(dir);
1419         int rc = 0;
1420
1421         ENTRY;
1422
1423         if ((*dentryp)->d_name.name[0] == '.') {
1424                 if (sai->sai_ls_all ||
1425                     sai->sai_miss_hidden >= sai->sai_skip_hidden) {
1426                         /*
1427                          * Hidden dentry is the first one, or statahead
1428                          * thread does not skip so many hidden dentries
1429                          * before "sai_ls_all" enabled as below.
1430                          */
1431                 } else {
1432                         if (!sai->sai_ls_all)
1433                                 /*
1434                                  * It maybe because hidden dentry is not
1435                                  * the first one, "sai_ls_all" was not
1436                                  * set, then "ls -al" missed. Enable
1437                                  * "sai_ls_all" for such case.
1438                                  */
1439                                 sai->sai_ls_all = 1;
1440
1441                         /*
1442                          * Such "getattr" has been skipped before
1443                          * "sai_ls_all" enabled as above.
1444                          */
1445                         sai->sai_miss_hidden++;
1446                         RETURN(-EAGAIN);
1447                 }
1448         }
1449
1450         if (unplug)
1451                 GOTO(out, rc = 1);
1452
1453         entry = sa_get(sai, &(*dentryp)->d_name);
1454         if (!entry)
1455                 GOTO(out, rc = -EAGAIN);
1456
1457         /* if statahead is busy in readdir, help it do post-work */
1458         if (!sa_ready(entry) && sai->sai_in_readpage)
1459                 sa_handle_callback(sai);
1460
1461         if (!sa_ready(entry)) {
1462                 spin_lock(&lli->lli_sa_lock);
1463                 sai->sai_index_wait = entry->se_index;
1464                 spin_unlock(&lli->lli_sa_lock);
1465                 rc = wait_event_idle_timeout(sai->sai_waitq, sa_ready(entry),
1466                                              cfs_time_seconds(30));
1467                 if (rc == 0) {
1468                         /*
1469                          * entry may not be ready, so it may be used by inflight
1470                          * statahead RPC, don't free it.
1471                          */
1472                         entry = NULL;
1473                         GOTO(out, rc = -EAGAIN);
1474                 }
1475         }
1476
1477         /*
1478          * We need to see the value that was set immediately before we
1479          * were woken up.
1480          */
1481         if (smp_load_acquire(&entry->se_state) == SA_ENTRY_SUCC &&
1482             entry->se_inode) {
1483                 struct inode *inode = entry->se_inode;
1484                 struct lookup_intent it = { .it_op = IT_GETATTR,
1485                                             .it_lock_handle =
1486                                                 entry->se_handle };
1487                 __u64 bits;
1488
1489                 rc = md_revalidate_lock(ll_i2mdexp(dir), &it,
1490                                         ll_inode2fid(inode), &bits);
1491                 if (rc == 1) {
1492                         if (!(*dentryp)->d_inode) {
1493                                 struct dentry *alias;
1494
1495                                 alias = ll_splice_alias(inode, *dentryp);
1496                                 if (IS_ERR(alias)) {
1497                                         ll_intent_release(&it);
1498                                         GOTO(out, rc = PTR_ERR(alias));
1499                                 }
1500                                 *dentryp = alias;
1501                                 /*
1502                                  * statahead prepared this inode, transfer inode
1503                                  * refcount from sa_entry to dentry
1504                                  */
1505                                 entry->se_inode = NULL;
1506                         } else if ((*dentryp)->d_inode != inode) {
1507                                 /* revalidate, but inode is recreated */
1508                                 CDEBUG(D_READA,
1509                                        "%s: stale dentry %pd inode " DFID", statahead inode "DFID "\n",
1510                                        ll_i2sbi(inode)->ll_fsname, *dentryp,
1511                                        PFID(ll_inode2fid((*dentryp)->d_inode)),
1512                                        PFID(ll_inode2fid(inode)));
1513                                 ll_intent_release(&it);
1514                                 GOTO(out, rc = -ESTALE);
1515                         }
1516
1517                         if ((bits & MDS_INODELOCK_LOOKUP) &&
1518                             d_lustre_invalid(*dentryp))
1519                                 d_lustre_revalidate(*dentryp);
1520                         ll_intent_release(&it);
1521                 }
1522         }
1523 out:
1524         /*
1525          * statahead cached sa_entry can be used only once, and will be killed
1526          * right after use, so if lookup/revalidate accessed statahead cache,
1527          * set dentry ldd_sa_generation to parent lli_sa_generation, later if we
1528          * stat this file again, we know we've done statahead before, see
1529          * dentry_may_statahead().
1530          */
1531         ldd = ll_d2d(*dentryp);
1532         /* ldd can be NULL if llite lookup failed. */
1533         if (ldd)
1534                 ldd->lld_sa_generation = lli->lli_sa_generation;
1535         sa_put(sai, entry);
1536         spin_lock(&lli->lli_sa_lock);
1537         if (sai->sai_task)
1538                 wake_up_process(sai->sai_task);
1539         spin_unlock(&lli->lli_sa_lock);
1540
1541         RETURN(rc);
1542 }
1543
1544 /**
1545  * start statahead thread
1546  *
1547  * \param[in] dir       parent directory
1548  * \param[in] dentry    dentry that triggers statahead, normally the first
1549  *                      dirent under @dir
1550  * \param[in] agl       indicate whether AGL is needed
1551  * \retval              -EAGAIN on success, because when this function is
1552  *                      called, it's already in lookup call, so client should
1553  *                      do it itself instead of waiting for statahead thread
1554  *                      to do it asynchronously.
1555  * \retval              negative number upon error
1556  */
1557 static int start_statahead_thread(struct inode *dir, struct dentry *dentry,
1558                                   bool agl)
1559 {
1560         int node = cfs_cpt_spread_node(cfs_cpt_tab, CFS_CPT_ANY);
1561         struct ll_inode_info *lli = ll_i2info(dir);
1562         struct ll_statahead_info *sai = NULL;
1563         struct dentry *parent = dentry->d_parent;
1564         struct task_struct *task;
1565         struct ll_sb_info *sbi = ll_i2sbi(parent->d_inode);
1566         int first = LS_FIRST_DE;
1567         int rc = 0;
1568
1569         ENTRY;
1570
1571         /* I am the "lli_opendir_pid" owner, only me can set "lli_sai". */
1572         first = is_first_dirent(dir, dentry);
1573         if (first == LS_NOT_FIRST_DE)
1574                 /* It is not "ls -{a}l" operation, no need statahead for it. */
1575                 GOTO(out, rc = -EFAULT);
1576
1577         if (unlikely(atomic_inc_return(&sbi->ll_sa_running) >
1578                                        sbi->ll_sa_running_max)) {
1579                 CDEBUG(D_READA,
1580                        "Too many concurrent statahead instances, avoid new statahead instance temporarily.\n");
1581                 GOTO(out, rc = -EMFILE);
1582         }
1583
1584         sai = ll_sai_alloc(parent);
1585         if (!sai)
1586                 GOTO(out, rc = -ENOMEM);
1587
1588         sai->sai_ls_all = (first == LS_FIRST_DOT_DE);
1589
1590         /*
1591          * if current lli_opendir_key was deauthorized, or dir re-opened by
1592          * another process, don't start statahead, otherwise the newly spawned
1593          * statahead thread won't be notified to quit.
1594          */
1595         spin_lock(&lli->lli_sa_lock);
1596         if (unlikely(lli->lli_sai || !lli->lli_opendir_key ||
1597                      lli->lli_opendir_pid != current->pid)) {
1598                 spin_unlock(&lli->lli_sa_lock);
1599                 GOTO(out, rc = -EPERM);
1600         }
1601         lli->lli_sai = sai;
1602         spin_unlock(&lli->lli_sa_lock);
1603
1604         CDEBUG(D_READA, "start statahead thread: [pid %d] [parent %pd]\n",
1605                current->pid, parent);
1606
1607         task = kthread_create_on_node(ll_statahead_thread, parent, node,
1608                                       "ll_sa_%u", lli->lli_opendir_pid);
1609         if (IS_ERR(task)) {
1610                 spin_lock(&lli->lli_sa_lock);
1611                 lli->lli_sai = NULL;
1612                 spin_unlock(&lli->lli_sa_lock);
1613                 rc = PTR_ERR(task);
1614                 CERROR("can't start ll_sa thread, rc: %d\n", rc);
1615                 GOTO(out, rc);
1616         }
1617
1618         if (ll_i2sbi(parent->d_inode)->ll_flags & LL_SBI_AGL_ENABLED && agl)
1619                 ll_start_agl(parent, sai);
1620
1621         atomic_inc(&ll_i2sbi(parent->d_inode)->ll_sa_total);
1622         sai->sai_task = task;
1623
1624         wake_up_process(task);
1625         /*
1626          * We don't stat-ahead for the first dirent since we are already in
1627          * lookup.
1628          */
1629         RETURN(-EAGAIN);
1630
1631 out:
1632         /*
1633          * once we start statahead thread failed, disable statahead so that
1634          * subsequent stat won't waste time to try it.
1635          */
1636         spin_lock(&lli->lli_sa_lock);
1637         if (lli->lli_opendir_pid == current->pid)
1638                 lli->lli_sa_enabled = 0;
1639         spin_unlock(&lli->lli_sa_lock);
1640
1641         if (sai)
1642                 ll_sai_free(sai);
1643         if (first != LS_NOT_FIRST_DE)
1644                 atomic_dec(&sbi->ll_sa_running);
1645
1646         RETURN(rc);
1647 }
1648
1649 /*
1650  * Check whether statahead for @dir was started.
1651  */
1652 static inline bool ll_statahead_started(struct inode *dir, bool agl)
1653 {
1654         struct ll_inode_info *lli = ll_i2info(dir);
1655         struct ll_statahead_info *sai;
1656
1657         spin_lock(&lli->lli_sa_lock);
1658         sai = lli->lli_sai;
1659         if (sai && (sai->sai_agl_task != NULL) != agl)
1660                 CDEBUG(D_READA,
1661                        "%s: Statahead AGL hint changed from %d to %d\n",
1662                        ll_i2sbi(dir)->ll_fsname,
1663                        sai->sai_agl_task != NULL, agl);
1664         spin_unlock(&lli->lli_sa_lock);
1665
1666         return !!sai;
1667 }
1668
1669 /**
1670  * statahead entry function, this is called when client getattr on a file, it
1671  * will start statahead thread if this is the first dir entry, else revalidate
1672  * dentry from statahead cache.
1673  *
1674  * \param[in]  dir      parent directory
1675  * \param[out] dentryp  dentry to getattr
1676  * \param[in]  agl      whether start the agl thread
1677  *
1678  * \retval              1 on success
1679  * \retval              0 revalidation from statahead cache failed, caller needs
1680  *                      to getattr from server directly
1681  * \retval              negative number on error, caller often ignores this and
1682  *                      then getattr from server
1683  */
1684 int ll_start_statahead(struct inode *dir, struct dentry *dentry, bool agl)
1685 {
1686         if (!ll_statahead_started(dir, agl))
1687                 return start_statahead_thread(dir, dentry, agl);
1688         return 0;
1689 }
1690
1691 /**
1692  * revalidate dentry from statahead cache.
1693  *
1694  * \param[in]  dir      parent directory
1695  * \param[out] dentryp  dentry to getattr
1696  * \param[in]  unplug   unplug statahead window only (normally for negative
1697  *                      dentry)
1698  * \retval              1 on success
1699  * \retval              0 revalidation from statahead cache failed, caller needs
1700  *                      to getattr from server directly
1701  * \retval              negative number on error, caller often ignores this and
1702  *                      then getattr from server
1703  */
1704 int ll_revalidate_statahead(struct inode *dir, struct dentry **dentryp,
1705                             bool unplug)
1706 {
1707         struct ll_statahead_info *sai;
1708         int rc = 0;
1709
1710         sai = ll_sai_get(dir);
1711         if (sai) {
1712                 rc = revalidate_statahead_dentry(dir, sai, dentryp, unplug);
1713                 CDEBUG(D_READA, "revalidate statahead %pd: rc = %d.\n",
1714                        *dentryp, rc);
1715                 ll_sai_put(sai);
1716         }
1717         return rc;
1718 }