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

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.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, 2015, 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/mm.h>
  36 #include <linux/highmem.h>
  37 #include <linux/pagemap.h>
  38
  39 #define DEBUG_SUBSYSTEM S_LLITE
  40
  41 #include <obd_support.h>
  42 #include <lustre_dlm.h>
  43 #include "llite_internal.h"
  44
  45 #define SA_OMITTED_ENTRY_MAX 8ULL
  46
  47 enum se_stat {
  48         /** negative values are for error cases */
  49         SA_ENTRY_INIT = 0,      /** init entry */
  50         SA_ENTRY_SUCC = 1,      /** stat succeed */
  51         SA_ENTRY_INVA = 2,      /** invalid entry */
  52 };
  53
  54 /*
  55  * sa_entry is not refcounted: statahead thread allocates it and do async stat,
  56  * and in async stat callback ll_statahead_interpret() will add it into
  57  * sai_interim_entries, later statahead thread will call sa_handle_callback() to
  58  * instantiate entry and move it into sai_entries, and then only scanner process
  59  * can access and free it.
  60  */
  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         enum se_stat            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;
  87 static DEFINE_SPINLOCK(sai_generation_lock);
  88
  89 /* sa_entry is ready to use */
  90 static inline int sa_ready(struct sa_entry *entry)
  91 {
  92         smp_rmb();
  93         return (entry->se_state != SA_ENTRY_INIT);
  94 }
  95
  96 /* hash value to put in sai_cache */
  97 static inline int sa_hash(int val)
  98 {
  99         return val & LL_SA_CACHE_MASK;
 100 }
 101
 102 /* hash entry into sai_cache */
 103 static inline void
 104 sa_rehash(struct ll_statahead_info *sai, struct sa_entry *entry)
 105 {
 106         int i = sa_hash(entry->se_qstr.hash);
 107
 108         spin_lock(&sai->sai_cache_lock[i]);
 109         list_add_tail(&entry->se_hash, &sai->sai_cache[i]);
 110         spin_unlock(&sai->sai_cache_lock[i]);
 111 }
 112
 113 /*
 114  * Remove entry from SA table.
 115  */
 116 static inline void
 117 sa_unhash(struct ll_statahead_info *sai, struct sa_entry *entry)
 118 {
 119         int i = sa_hash(entry->se_qstr.hash);
 120
 121         spin_lock(&sai->sai_cache_lock[i]);
 122         list_del_init(&entry->se_hash);
 123         spin_unlock(&sai->sai_cache_lock[i]);
 124 }
 125
 126 static inline int agl_should_run(struct ll_statahead_info *sai,
 127                                  struct inode *inode)
 128 {
 129         return (inode && S_ISREG(inode->i_mode) && sai->sai_agl_valid);
 130 }
 131
 132 /* statahead window is full */
 133 static inline int sa_sent_full(struct ll_statahead_info *sai)
 134 {
 135         return atomic_read(&sai->sai_cache_count) >= sai->sai_max;
 136 }
 137
 138 /* got async stat replies */
 139 static inline int sa_has_callback(struct ll_statahead_info *sai)
 140 {
 141         return !list_empty(&sai->sai_interim_entries);
 142 }
 143
 144 static inline int agl_list_empty(struct ll_statahead_info *sai)
 145 {
 146         return list_empty(&sai->sai_agls);
 147 }
 148
 149 /**
 150  * (1) hit ratio less than 80%
 151  * or
 152  * (2) consecutive miss more than 8
 153  * then means low hit.
 154  */
 155 static inline int sa_low_hit(struct ll_statahead_info *sai)
 156 {
 157         return ((sai->sai_hit > 7 && sai->sai_hit < 4 * sai->sai_miss) ||
 158                 (sai->sai_consecutive_miss > 8));
 159 }
 160
 161 /*
 162  * if the given index is behind of statahead window more than
 163  * SA_OMITTED_ENTRY_MAX, then it is old.
 164  */
 165 static inline int is_omitted_entry(struct ll_statahead_info *sai, __u64 index)
 166 {
 167         return ((__u64)sai->sai_max + index + SA_OMITTED_ENTRY_MAX <
 168                  sai->sai_index);
 169 }
 170
 171 /* allocate sa_entry and hash it to allow scanner process to find it */
 172 static struct sa_entry *
 173 sa_alloc(struct dentry *parent, struct ll_statahead_info *sai, __u64 index,
 174          const char *name, int len, const struct lu_fid *fid)
 175 {
 176         struct ll_inode_info *lli;
 177         struct sa_entry   *entry;
 178         int                entry_size;
 179         char             *dname;
 180
 181         entry_size = sizeof(struct sa_entry) + (len & ~3) + 4;
 182         entry = kzalloc(entry_size, GFP_NOFS);
 183         if (unlikely(!entry))
 184                 return ERR_PTR(-ENOMEM);
 185
 186         CDEBUG(D_READA, "alloc sa entry %.*s(%p) index %llu\n",
 187                len, name, entry, index);
 188
 189         entry->se_index = index;
 190         entry->se_state = SA_ENTRY_INIT;
 191         entry->se_size = entry_size;
 192         dname = (char *)entry + sizeof(struct sa_entry);
 193         memcpy(dname, name, len);
 194         dname[len] = 0;
 195
 196         entry->se_qstr.hash = full_name_hash(parent, name, len);
 197         entry->se_qstr.len = len;
 198         entry->se_qstr.name = dname;
 199         entry->se_fid = *fid;
 200
 201         lli = ll_i2info(sai->sai_dentry->d_inode);
 202         spin_lock(&lli->lli_sa_lock);
 203         INIT_LIST_HEAD(&entry->se_list);
 204         sa_rehash(sai, entry);
 205         spin_unlock(&lli->lli_sa_lock);
 206
 207         atomic_inc(&sai->sai_cache_count);
 208
 209         return entry;
 210 }
 211
 212 /* free sa_entry, which should have been unhashed and not in any list */
 213 static void sa_free(struct ll_statahead_info *sai, struct sa_entry *entry)
 214 {
 215         CDEBUG(D_READA, "free sa entry %.*s(%p) index %llu\n",
 216                entry->se_qstr.len, entry->se_qstr.name, entry,
 217                entry->se_index);
 218
 219         LASSERT(list_empty(&entry->se_list));
 220         LASSERT(list_empty(&entry->se_hash));
 221
 222         kfree(entry);
 223         atomic_dec(&sai->sai_cache_count);
 224 }
 225
 226 /*
 227  * find sa_entry by name, used by directory scanner, lock is not needed because
 228  * only scanner can remove the entry from cache.
 229  */
 230 static struct sa_entry *
 231 sa_get(struct ll_statahead_info *sai, const struct qstr *qstr)
 232 {
 233         struct sa_entry *entry;
 234         int i = sa_hash(qstr->hash);
 235
 236         list_for_each_entry(entry, &sai->sai_cache[i], se_hash) {
 237                 if (entry->se_qstr.hash == qstr->hash &&
 238                     entry->se_qstr.len == qstr->len &&
 239                     memcmp(entry->se_qstr.name, qstr->name, qstr->len) == 0)
 240                         return entry;
 241         }
 242         return NULL;
 243 }
 244
 245 /* unhash and unlink sa_entry, and then free it */
 246 static inline void
 247 sa_kill(struct ll_statahead_info *sai, struct sa_entry *entry)
 248 {
 249         struct ll_inode_info *lli = ll_i2info(sai->sai_dentry->d_inode);
 250
 251         LASSERT(!list_empty(&entry->se_hash));
 252         LASSERT(!list_empty(&entry->se_list));
 253         LASSERT(sa_ready(entry));
 254
 255         sa_unhash(sai, entry);
 256
 257         spin_lock(&lli->lli_sa_lock);
 258         list_del_init(&entry->se_list);
 259         spin_unlock(&lli->lli_sa_lock);
 260
 261         if (entry->se_inode)
 262                 iput(entry->se_inode);
 263
 264         sa_free(sai, entry);
 265 }
 266
 267 /* called by scanner after use, sa_entry will be killed */
 268 static void
 269 sa_put(struct ll_statahead_info *sai, struct sa_entry *entry)
 270 {
 271         struct sa_entry *tmp, *next;
 272
 273         if (entry && entry->se_state == SA_ENTRY_SUCC) {
 274                 struct ll_sb_info *sbi = ll_i2sbi(sai->sai_dentry->d_inode);
 275
 276                 sai->sai_hit++;
 277                 sai->sai_consecutive_miss = 0;
 278                 sai->sai_max = min(2 * sai->sai_max, sbi->ll_sa_max);
 279         } else {
 280                 sai->sai_miss++;
 281                 sai->sai_consecutive_miss++;
 282         }
 283
 284         if (entry)
 285                 sa_kill(sai, entry);
 286
 287         /*
 288          * kill old completed entries, only scanner process does this, no need
 289          * to lock
 290          */
 291         list_for_each_entry_safe(tmp, next, &sai->sai_entries, se_list) {
 292                 if (!is_omitted_entry(sai, tmp->se_index))
 293                         break;
 294                 sa_kill(sai, tmp);
 295         }
 296
 297         wake_up(&sai->sai_thread.t_ctl_waitq);
 298 }
 299
 300 /*
 301  * update state and sort add entry to sai_entries by index, return true if
 302  * scanner is waiting on this entry.
 303  */
 304 static bool
 305 __sa_make_ready(struct ll_statahead_info *sai, struct sa_entry *entry, int ret)
 306 {
 307         struct list_head *pos = &sai->sai_entries;
 308         __u64 index = entry->se_index;
 309         struct sa_entry *se;
 310
 311         LASSERT(!sa_ready(entry));
 312         LASSERT(list_empty(&entry->se_list));
 313
 314         list_for_each_entry_reverse(se, &sai->sai_entries, se_list) {
 315                 if (se->se_index < entry->se_index) {
 316                         pos = &se->se_list;
 317                         break;
 318                 }
 319         }
 320         list_add(&entry->se_list, pos);
 321         entry->se_state = ret < 0 ? SA_ENTRY_INVA : SA_ENTRY_SUCC;
 322
 323         return (index == sai->sai_index_wait);
 324 }
 325
 326 /*
 327  * release resources used in async stat RPC, update entry state and wakeup if
 328  * scanner process it waiting on this entry.
 329  */
 330 static void
 331 sa_make_ready(struct ll_statahead_info *sai, struct sa_entry *entry, int ret)
 332 {
 333         struct ll_inode_info *lli = ll_i2info(sai->sai_dentry->d_inode);
 334         struct md_enqueue_info *minfo = entry->se_minfo;
 335         struct ptlrpc_request *req = entry->se_req;
 336         bool wakeup;
 337
 338         /* release resources used in RPC */
 339         if (minfo) {
 340                 entry->se_minfo = NULL;
 341                 ll_intent_release(&minfo->mi_it);
 342                 iput(minfo->mi_dir);
 343                 kfree(minfo);
 344         }
 345
 346         if (req) {
 347                 entry->se_req = NULL;
 348                 ptlrpc_req_finished(req);
 349         }
 350
 351         spin_lock(&lli->lli_sa_lock);
 352         wakeup = __sa_make_ready(sai, entry, ret);
 353         spin_unlock(&lli->lli_sa_lock);
 354
 355         if (wakeup)
 356                 wake_up(&sai->sai_waitq);
 357 }
 358
 359 /* Insert inode into the list of sai_agls. */
 360 static void ll_agl_add(struct ll_statahead_info *sai,
 361                        struct inode *inode, int index)
 362 {
 363         struct ll_inode_info *child  = ll_i2info(inode);
 364         struct ll_inode_info *parent = ll_i2info(sai->sai_dentry->d_inode);
 365         int                added  = 0;
 366
 367         spin_lock(&child->lli_agl_lock);
 368         if (child->lli_agl_index == 0) {
 369                 child->lli_agl_index = index;
 370                 spin_unlock(&child->lli_agl_lock);
 371
 372                 LASSERT(list_empty(&child->lli_agl_list));
 373
 374                 igrab(inode);
 375                 spin_lock(&parent->lli_agl_lock);
 376                 if (list_empty(&sai->sai_agls))
 377                         added = 1;
 378                 list_add_tail(&child->lli_agl_list, &sai->sai_agls);
 379                 spin_unlock(&parent->lli_agl_lock);
 380         } else {
 381                 spin_unlock(&child->lli_agl_lock);
 382         }
 383
 384         if (added > 0)
 385                 wake_up(&sai->sai_agl_thread.t_ctl_waitq);
 386 }
 387
 388 /* allocate sai */
 389 static struct ll_statahead_info *ll_sai_alloc(struct dentry *dentry)
 390 {
 391         struct ll_inode_info *lli = ll_i2info(dentry->d_inode);
 392         struct ll_statahead_info *sai;
 393         int                    i;
 394
 395         sai = kzalloc(sizeof(*sai), GFP_NOFS);
 396         if (!sai)
 397                 return NULL;
 398
 399         sai->sai_dentry = dget(dentry);
 400         atomic_set(&sai->sai_refcount, 1);
 401
 402         sai->sai_max = LL_SA_RPC_MIN;
 403         sai->sai_index = 1;
 404         init_waitqueue_head(&sai->sai_waitq);
 405         init_waitqueue_head(&sai->sai_thread.t_ctl_waitq);
 406         init_waitqueue_head(&sai->sai_agl_thread.t_ctl_waitq);
 407
 408         INIT_LIST_HEAD(&sai->sai_interim_entries);
 409         INIT_LIST_HEAD(&sai->sai_entries);
 410         INIT_LIST_HEAD(&sai->sai_agls);
 411
 412         for (i = 0; i < LL_SA_CACHE_SIZE; i++) {
 413                 INIT_LIST_HEAD(&sai->sai_cache[i]);
 414                 spin_lock_init(&sai->sai_cache_lock[i]);
 415         }
 416         atomic_set(&sai->sai_cache_count, 0);
 417
 418         spin_lock(&sai_generation_lock);
 419         lli->lli_sa_generation = ++sai_generation;
 420         if (unlikely(!sai_generation))
 421                 lli->lli_sa_generation = ++sai_generation;
 422         spin_unlock(&sai_generation_lock);
 423
 424         return sai;
 425 }
 426
 427 /* free sai */
 428 static inline void ll_sai_free(struct ll_statahead_info *sai)
 429 {
 430         LASSERT(sai->sai_dentry);
 431         dput(sai->sai_dentry);
 432         kfree(sai);
 433 }
 434
 435 /*
 436  * take refcount of sai if sai for @dir exists, which means statahead is on for
 437  * this directory.
 438  */
 439 static inline struct ll_statahead_info *ll_sai_get(struct inode *dir)
 440 {
 441         struct ll_inode_info *lli = ll_i2info(dir);
 442         struct ll_statahead_info *sai = NULL;
 443
 444         spin_lock(&lli->lli_sa_lock);
 445         sai = lli->lli_sai;
 446         if (sai)
 447                 atomic_inc(&sai->sai_refcount);
 448         spin_unlock(&lli->lli_sa_lock);
 449
 450         return sai;
 451 }
 452
 453 /*
 454  * put sai refcount after use, if refcount reaches zero, free sai and sa_entries
 455  * attached to it.
 456  */
 457 static void ll_sai_put(struct ll_statahead_info *sai)
 458 {
 459         struct ll_inode_info *lli = ll_i2info(sai->sai_dentry->d_inode);
 460
 461         if (atomic_dec_and_lock(&sai->sai_refcount, &lli->lli_sa_lock)) {
 462                 struct ll_sb_info *sbi = ll_i2sbi(sai->sai_dentry->d_inode);
 463                 struct sa_entry *entry, *next;
 464
 465                 lli->lli_sai = NULL;
 466                 spin_unlock(&lli->lli_sa_lock);
 467
 468                 LASSERT(thread_is_stopped(&sai->sai_thread));
 469                 LASSERT(thread_is_stopped(&sai->sai_agl_thread));
 470                 LASSERT(sai->sai_sent == sai->sai_replied);
 471                 LASSERT(!sa_has_callback(sai));
 472
 473                 list_for_each_entry_safe(entry, next, &sai->sai_entries,
 474                                          se_list)
 475                         sa_kill(sai, entry);
 476
 477                 LASSERT(atomic_read(&sai->sai_cache_count) == 0);
 478                 LASSERT(list_empty(&sai->sai_agls));
 479
 480                 ll_sai_free(sai);
 481                 atomic_dec(&sbi->ll_sa_running);
 482         }
 483 }
 484
 485 /* Do NOT forget to drop inode refcount when into sai_agls. */
 486 static void ll_agl_trigger(struct inode *inode, struct ll_statahead_info *sai)
 487 {
 488         struct ll_inode_info *lli   = ll_i2info(inode);
 489         __u64            index = lli->lli_agl_index;
 490         int                rc;
 491
 492         LASSERT(list_empty(&lli->lli_agl_list));
 493
 494         /* AGL maybe fall behind statahead with one entry */
 495         if (is_omitted_entry(sai, index + 1)) {
 496                 lli->lli_agl_index = 0;
 497                 iput(inode);
 498                 return;
 499         }
 500
 501         /* Someone is in glimpse (sync or async), do nothing. */
 502         rc = down_write_trylock(&lli->lli_glimpse_sem);
 503         if (rc == 0) {
 504                 lli->lli_agl_index = 0;
 505                 iput(inode);
 506                 return;
 507         }
 508
 509         /*
 510          * Someone triggered glimpse within 1 sec before.
 511          * 1) The former glimpse succeeded with glimpse lock granted by OST, and
 512          *    if the lock is still cached on client, AGL needs to do nothing. If
 513          *    it is cancelled by other client, AGL maybe cannot obtain new lock
 514          *    for no glimpse callback triggered by AGL.
 515          * 2) The former glimpse succeeded, but OST did not grant glimpse lock.
 516          *    Under such case, it is quite possible that the OST will not grant
 517          *    glimpse lock for AGL also.
 518          * 3) The former glimpse failed, compared with other two cases, it is
 519          *    relative rare. AGL can ignore such case, and it will not muchly
 520          *    affect the performance.
 521          */
 522         if (lli->lli_glimpse_time != 0 &&
 523             time_before(cfs_time_shift(-1), lli->lli_glimpse_time)) {
 524                 up_write(&lli->lli_glimpse_sem);
 525                 lli->lli_agl_index = 0;
 526                 iput(inode);
 527                 return;
 528         }
 529
 530         CDEBUG(D_READA, "Handling (init) async glimpse: inode = "
 531                DFID ", idx = %llu\n", PFID(&lli->lli_fid), index);
 532
 533         cl_agl(inode);
 534         lli->lli_agl_index = 0;
 535         lli->lli_glimpse_time = cfs_time_current();
 536         up_write(&lli->lli_glimpse_sem);
 537
 538         CDEBUG(D_READA, "Handled (init) async glimpse: inode= "
 539                DFID ", idx = %llu, rc = %d\n",
 540                PFID(&lli->lli_fid), index, rc);
 541
 542         iput(inode);
 543 }
 544
 545 /*
 546  * prepare inode for sa entry, add it into agl list, now sa_entry is ready
 547  * to be used by scanner process.
 548  */
 549 static void sa_instantiate(struct ll_statahead_info *sai,
 550                            struct sa_entry *entry)
 551 {
 552         struct inode *dir = sai->sai_dentry->d_inode;
 553         struct inode       *child;
 554         struct md_enqueue_info *minfo;
 555         struct lookup_intent   *it;
 556         struct ptlrpc_request  *req;
 557         struct mdt_body *body;
 558         int                  rc    = 0;
 559
 560         LASSERT(entry->se_handle != 0);
 561
 562         minfo = entry->se_minfo;
 563         it = &minfo->mi_it;
 564         req = entry->se_req;
 565         body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
 566         if (!body) {
 567                 rc = -EFAULT;
 568                 goto out;
 569         }
 570
 571         child = entry->se_inode;
 572         if (child) {
 573                 /* revalidate; unlinked and re-created with the same name */
 574                 if (unlikely(!lu_fid_eq(&minfo->mi_data.op_fid2, &body->mbo_fid1))) {
 575                         entry->se_inode = NULL;
 576                         iput(child);
 577                         child = NULL;
 578                 }
 579         }
 580
 581         it->it_lock_handle = entry->se_handle;
 582         rc = md_revalidate_lock(ll_i2mdexp(dir), it, ll_inode2fid(dir), NULL);
 583         if (rc != 1) {
 584                 rc = -EAGAIN;
 585                 goto out;
 586         }
 587
 588         rc = ll_prep_inode(&child, req, dir->i_sb, it);
 589         if (rc)
 590                 goto out;
 591
 592         CDEBUG(D_READA, "%s: setting %.*s" DFID " l_data to inode %p\n",
 593                ll_get_fsname(child->i_sb, NULL, 0),
 594                entry->se_qstr.len, entry->se_qstr.name,
 595                PFID(ll_inode2fid(child)), child);
 596         ll_set_lock_data(ll_i2sbi(dir)->ll_md_exp, child, it, NULL);
 597
 598         entry->se_inode = child;
 599
 600         if (agl_should_run(sai, child))
 601                 ll_agl_add(sai, child, entry->se_index);
 602
 603 out:
 604         /*
 605          * sa_make_ready() will drop ldlm ibits lock refcount by calling
 606          * ll_intent_drop_lock() in spite of failures. Do not worry about
 607          * calling ll_intent_drop_lock() more than once.
 608          */
 609         sa_make_ready(sai, entry, rc);
 610 }
 611
 612 /* once there are async stat replies, instantiate sa_entry from replies */
 613 static void sa_handle_callback(struct ll_statahead_info *sai)
 614 {
 615         struct ll_inode_info *lli;
 616
 617         lli = ll_i2info(sai->sai_dentry->d_inode);
 618
 619         while (sa_has_callback(sai)) {
 620                 struct sa_entry *entry;
 621
 622                 spin_lock(&lli->lli_sa_lock);
 623                 if (unlikely(!sa_has_callback(sai))) {
 624                         spin_unlock(&lli->lli_sa_lock);
 625                         break;
 626                 }
 627                 entry = list_entry(sai->sai_interim_entries.next,
 628                                    struct sa_entry, se_list);
 629                 list_del_init(&entry->se_list);
 630                 spin_unlock(&lli->lli_sa_lock);
 631
 632                 sa_instantiate(sai, entry);
 633         }
 634 }
 635
 636 /*
 637  * callback for async stat, because this is called in ptlrpcd context, we only
 638  * put sa_entry in sai_cb_entries list, and let sa_handle_callback() to really
 639  * prepare inode and instantiate sa_entry later.
 640  */
 641 static int ll_statahead_interpret(struct ptlrpc_request *req,
 642                                   struct md_enqueue_info *minfo, int rc)
 643 {
 644         struct lookup_intent     *it  = &minfo->mi_it;
 645         struct inode         *dir = minfo->mi_dir;
 646         struct ll_inode_info     *lli = ll_i2info(dir);
 647         struct ll_statahead_info *sai = lli->lli_sai;
 648         struct sa_entry *entry = (struct sa_entry *)minfo->mi_cbdata;
 649         wait_queue_head_t *waitq = NULL;
 650         __u64 handle = 0;
 651
 652         if (it_disposition(it, DISP_LOOKUP_NEG))
 653                 rc = -ENOENT;
 654
 655         /*
 656          * because statahead thread will wait for all inflight RPC to finish,
 657          * sai should be always valid, no need to refcount
 658          */
 659         LASSERT(sai);
 660         LASSERT(!thread_is_stopped(&sai->sai_thread));
 661         LASSERT(entry);
 662
 663         CDEBUG(D_READA, "sa_entry %.*s rc %d\n",
 664                entry->se_qstr.len, entry->se_qstr.name, rc);
 665
 666         if (rc) {
 667                 ll_intent_release(it);
 668                 iput(dir);
 669                 kfree(minfo);
 670         } else {
 671                 /*
 672                  * release ibits lock ASAP to avoid deadlock when statahead
 673                  * thread enqueues lock on parent in readdir and another
 674                  * process enqueues lock on child with parent lock held, eg.
 675                  * unlink.
 676                  */
 677                 handle = it->it_lock_handle;
 678                 ll_intent_drop_lock(it);
 679         }
 680
 681         spin_lock(&lli->lli_sa_lock);
 682         if (rc) {
 683                 if (__sa_make_ready(sai, entry, rc))
 684                         waitq = &sai->sai_waitq;
 685         } else {
 686                 entry->se_minfo = minfo;
 687                 entry->se_req = ptlrpc_request_addref(req);
 688                 /*
 689                  * Release the async ibits lock ASAP to avoid deadlock
 690                  * when statahead thread tries to enqueue lock on parent
 691                  * for readpage and other tries to enqueue lock on child
 692                  * with parent's lock held, for example: unlink.
 693                  */
 694                 entry->se_handle = handle;
 695                 if (!sa_has_callback(sai))
 696                         waitq = &sai->sai_thread.t_ctl_waitq;
 697
 698                 list_add_tail(&entry->se_list, &sai->sai_interim_entries);
 699         }
 700         sai->sai_replied++;
 701
 702         if (waitq)
 703                 wake_up(waitq);
 704         spin_unlock(&lli->lli_sa_lock);
 705
 706         return rc;
 707 }
 708
 709 /* finish async stat RPC arguments */
 710 static void sa_fini_data(struct md_enqueue_info *minfo)
 711 {
 712         iput(minfo->mi_dir);
 713         kfree(minfo);
 714 }
 715
 716 /**
 717  * prepare arguments for async stat RPC.
 718  */
 719 static struct md_enqueue_info *
 720 sa_prep_data(struct inode *dir, struct inode *child, struct sa_entry *entry)
 721 {
 722         struct md_enqueue_info   *minfo;
 723         struct ldlm_enqueue_info *einfo;
 724         struct md_op_data       *op_data;
 725
 726         minfo = kzalloc(sizeof(*minfo), GFP_NOFS);
 727         if (!minfo)
 728                 return ERR_PTR(-ENOMEM);
 729
 730         op_data = ll_prep_md_op_data(&minfo->mi_data, dir, child, NULL, 0, 0,
 731                                      LUSTRE_OPC_ANY, NULL);
 732         if (IS_ERR(op_data)) {
 733                 kfree(minfo);
 734                 return (struct md_enqueue_info *)op_data;
 735         }
 736
 737         if (!child)
 738                 op_data->op_fid2 = entry->se_fid;
 739
 740         minfo->mi_it.it_op = IT_GETATTR;
 741         minfo->mi_dir = igrab(dir);
 742         minfo->mi_cb = ll_statahead_interpret;
 743         minfo->mi_cbdata = entry;
 744
 745         einfo = &minfo->mi_einfo;
 746         einfo->ei_type   = LDLM_IBITS;
 747         einfo->ei_mode   = it_to_lock_mode(&minfo->mi_it);
 748         einfo->ei_cb_bl  = ll_md_blocking_ast;
 749         einfo->ei_cb_cp  = ldlm_completion_ast;
 750         einfo->ei_cb_gl  = NULL;
 751         einfo->ei_cbdata = NULL;
 752
 753         return minfo;
 754 }
 755
 756 /* async stat for file not found in dcache */
 757 static int sa_lookup(struct inode *dir, struct sa_entry *entry)
 758 {
 759         struct md_enqueue_info   *minfo;
 760         int                    rc;
 761
 762         minfo = sa_prep_data(dir, NULL, entry);
 763         if (IS_ERR(minfo))
 764                 return PTR_ERR(minfo);
 765
 766         rc = md_intent_getattr_async(ll_i2mdexp(dir), minfo);
 767         if (rc)
 768                 sa_fini_data(minfo);
 769
 770         return rc;
 771 }
 772
 773 /**
 774  * async stat for file found in dcache, similar to .revalidate
 775  *
 776  * \retval      1 dentry valid, no RPC sent
 777  * \retval      0 dentry invalid, will send async stat RPC
 778  * \retval      negative number upon error
 779  */
 780 static int sa_revalidate(struct inode *dir, struct sa_entry *entry,
 781                          struct dentry *dentry)
 782 {
 783         struct inode         *inode = d_inode(dentry);
 784         struct lookup_intent      it = { .it_op = IT_GETATTR,
 785                                          .it_lock_handle = 0 };
 786         struct md_enqueue_info   *minfo;
 787         int rc;
 788
 789         if (unlikely(!inode))
 790                 return 1;
 791
 792         if (d_mountpoint(dentry))
 793                 return 1;
 794
 795         entry->se_inode = igrab(inode);
 796         rc = md_revalidate_lock(ll_i2mdexp(dir), &it, ll_inode2fid(inode),
 797                                 NULL);
 798         if (rc == 1) {
 799                 entry->se_handle = it.it_lock_handle;
 800                 ll_intent_release(&it);
 801                 return 1;
 802         }
 803
 804         minfo = sa_prep_data(dir, inode, entry);
 805         if (IS_ERR(minfo)) {
 806                 entry->se_inode = NULL;
 807                 iput(inode);
 808                 return PTR_ERR(minfo);
 809         }
 810
 811         rc = md_intent_getattr_async(ll_i2mdexp(dir), minfo);
 812         if (rc) {
 813                 entry->se_inode = NULL;
 814                 iput(inode);
 815                 sa_fini_data(minfo);
 816         }
 817
 818         return rc;
 819 }
 820
 821 /* async stat for file with @name */
 822 static void sa_statahead(struct dentry *parent, const char *name, int len,
 823                          const struct lu_fid *fid)
 824 {
 825         struct inode         *dir    = d_inode(parent);
 826         struct ll_inode_info     *lli    = ll_i2info(dir);
 827         struct ll_statahead_info *sai    = lli->lli_sai;
 828         struct dentry       *dentry = NULL;
 829         struct sa_entry *entry;
 830         int                    rc;
 831
 832         entry = sa_alloc(parent, sai, sai->sai_index, name, len, fid);
 833         if (IS_ERR(entry))
 834                 return;
 835
 836         dentry = d_lookup(parent, &entry->se_qstr);
 837         if (!dentry) {
 838                 rc = sa_lookup(dir, entry);
 839         } else {
 840                 rc = sa_revalidate(dir, entry, dentry);
 841                 if (rc == 1 && agl_should_run(sai, d_inode(dentry)))
 842                         ll_agl_add(sai, d_inode(dentry), entry->se_index);
 843         }
 844
 845         if (dentry)
 846                 dput(dentry);
 847
 848         if (rc)
 849                 sa_make_ready(sai, entry, rc);
 850         else
 851                 sai->sai_sent++;
 852
 853         sai->sai_index++;
 854 }
 855
 856 /* async glimpse (agl) thread main function */
 857 static int ll_agl_thread(void *arg)
 858 {
 859         struct dentry       *parent = arg;
 860         struct inode         *dir    = d_inode(parent);
 861         struct ll_inode_info     *plli   = ll_i2info(dir);
 862         struct ll_inode_info     *clli;
 863         struct ll_sb_info       *sbi    = ll_i2sbi(dir);
 864         struct ll_statahead_info *sai;
 865         struct ptlrpc_thread *thread;
 866         struct l_wait_info      lwi    = { 0 };
 867
 868         sai = ll_sai_get(dir);
 869         thread = &sai->sai_agl_thread;
 870         thread->t_pid = current_pid();
 871         CDEBUG(D_READA, "agl thread started: sai %p, parent %pd\n",
 872                sai, parent);
 873
 874         atomic_inc(&sbi->ll_agl_total);
 875         spin_lock(&plli->lli_agl_lock);
 876         sai->sai_agl_valid = 1;
 877         if (thread_is_init(thread))
 878                 /* If someone else has changed the thread state
 879                  * (e.g. already changed to SVC_STOPPING), we can't just
 880                  * blindly overwrite that setting.
 881                  */
 882                 thread_set_flags(thread, SVC_RUNNING);
 883         spin_unlock(&plli->lli_agl_lock);
 884         wake_up(&thread->t_ctl_waitq);
 885
 886         while (1) {
 887                 l_wait_event(thread->t_ctl_waitq,
 888                              !list_empty(&sai->sai_agls) ||
 889                              !thread_is_running(thread),
 890                              &lwi);
 891
 892                 if (!thread_is_running(thread))
 893                         break;
 894
 895                 spin_lock(&plli->lli_agl_lock);
 896                 /* The statahead thread maybe help to process AGL entries,
 897                  * so check whether list empty again.
 898                  */
 899                 if (!list_empty(&sai->sai_agls)) {
 900                         clli = list_entry(sai->sai_agls.next,
 901                                           struct ll_inode_info, lli_agl_list);
 902                         list_del_init(&clli->lli_agl_list);
 903                         spin_unlock(&plli->lli_agl_lock);
 904                         ll_agl_trigger(&clli->lli_vfs_inode, sai);
 905                 } else {
 906                         spin_unlock(&plli->lli_agl_lock);
 907                 }
 908         }
 909
 910         spin_lock(&plli->lli_agl_lock);
 911         sai->sai_agl_valid = 0;
 912         while (!list_empty(&sai->sai_agls)) {
 913                 clli = list_entry(sai->sai_agls.next,
 914                                   struct ll_inode_info, lli_agl_list);
 915                 list_del_init(&clli->lli_agl_list);
 916                 spin_unlock(&plli->lli_agl_lock);
 917                 clli->lli_agl_index = 0;
 918                 iput(&clli->lli_vfs_inode);
 919                 spin_lock(&plli->lli_agl_lock);
 920         }
 921         thread_set_flags(thread, SVC_STOPPED);
 922         spin_unlock(&plli->lli_agl_lock);
 923         wake_up(&thread->t_ctl_waitq);
 924         ll_sai_put(sai);
 925         CDEBUG(D_READA, "agl thread stopped: sai %p, parent %pd\n",
 926                sai, parent);
 927         return 0;
 928 }
 929
 930 /* start agl thread */
 931 static void ll_start_agl(struct dentry *parent, struct ll_statahead_info *sai)
 932 {
 933         struct ptlrpc_thread *thread = &sai->sai_agl_thread;
 934         struct l_wait_info    lwi    = { 0 };
 935         struct ll_inode_info  *plli;
 936         struct task_struct *task;
 937
 938         CDEBUG(D_READA, "start agl thread: sai %p, parent %pd\n",
 939                sai, parent);
 940
 941         plli = ll_i2info(d_inode(parent));
 942         task = kthread_run(ll_agl_thread, parent, "ll_agl_%u",
 943                            plli->lli_opendir_pid);
 944         if (IS_ERR(task)) {
 945                 CERROR("can't start ll_agl thread, rc: %ld\n", PTR_ERR(task));
 946                 thread_set_flags(thread, SVC_STOPPED);
 947                 return;
 948         }
 949
 950         l_wait_event(thread->t_ctl_waitq,
 951                      thread_is_running(thread) || thread_is_stopped(thread),
 952                      &lwi);
 953 }
 954
 955 /* statahead thread main function */
 956 static int ll_statahead_thread(void *arg)
 957 {
 958         struct dentry       *parent = arg;
 959         struct inode         *dir    = d_inode(parent);
 960         struct ll_inode_info     *lli   = ll_i2info(dir);
 961         struct ll_sb_info       *sbi    = ll_i2sbi(dir);
 962         struct ll_statahead_info *sai;
 963         struct ptlrpc_thread *sa_thread;
 964         struct ptlrpc_thread *agl_thread;
 965         struct page           *page = NULL;
 966         __u64                pos    = 0;
 967         int                    first  = 0;
 968         int                    rc     = 0;
 969         struct md_op_data *op_data;
 970         struct l_wait_info      lwi    = { 0 };
 971
 972         sai = ll_sai_get(dir);
 973         sa_thread = &sai->sai_thread;
 974         agl_thread = &sai->sai_agl_thread;
 975         sa_thread->t_pid = current_pid();
 976         CDEBUG(D_READA, "statahead thread starting: sai %p, parent %pd\n",
 977                sai, parent);
 978
 979         op_data = ll_prep_md_op_data(NULL, dir, dir, NULL, 0, 0,
 980                                      LUSTRE_OPC_ANY, dir);
 981         if (IS_ERR(op_data)) {
 982                 rc = PTR_ERR(op_data);
 983                 goto out;
 984         }
 985
 986         op_data->op_max_pages = ll_i2sbi(dir)->ll_md_brw_pages;
 987
 988         if (sbi->ll_flags & LL_SBI_AGL_ENABLED)
 989                 ll_start_agl(parent, sai);
 990
 991         atomic_inc(&sbi->ll_sa_total);
 992         spin_lock(&lli->lli_sa_lock);
 993         if (thread_is_init(sa_thread))
 994                 /* If someone else has changed the thread state
 995                  * (e.g. already changed to SVC_STOPPING), we can't just
 996                  * blindly overwrite that setting.
 997                  */
 998                 thread_set_flags(sa_thread, SVC_RUNNING);
 999         spin_unlock(&lli->lli_sa_lock);
1000         wake_up(&sa_thread->t_ctl_waitq);
1001
1002         while (pos != MDS_DIR_END_OFF && thread_is_running(sa_thread)) {
1003                 struct lu_dirpage *dp;
1004                 struct lu_dirent  *ent;
1005
1006                 sai->sai_in_readpage = 1;
1007                 page = ll_get_dir_page(dir, op_data, pos);
1008                 sai->sai_in_readpage = 0;
1009                 if (IS_ERR(page)) {
1010                         rc = PTR_ERR(page);
1011                         CDEBUG(D_READA, "error reading dir " DFID " at %llu/%llu: opendir_pid = %u: rc = %d\n",
1012                                PFID(ll_inode2fid(dir)), pos, sai->sai_index,
1013                                lli->lli_opendir_pid, rc);
1014                         break;
1015                 }
1016
1017                 dp = page_address(page);
1018                 for (ent = lu_dirent_start(dp);
1019                      ent && thread_is_running(sa_thread) && !sa_low_hit(sai);
1020                      ent = lu_dirent_next(ent)) {
1021                         struct lu_fid fid;
1022                         __u64 hash;
1023                         int namelen;
1024                         char *name;
1025
1026                         hash = le64_to_cpu(ent->lde_hash);
1027                         if (unlikely(hash < pos))
1028                                 /*
1029                                  * Skip until we find target hash value.
1030                                  */
1031                                 continue;
1032
1033                         namelen = le16_to_cpu(ent->lde_namelen);
1034                         if (unlikely(namelen == 0))
1035                                 /*
1036                                  * Skip dummy record.
1037                                  */
1038                                 continue;
1039
1040                         name = ent->lde_name;
1041                         if (name[0] == '.') {
1042                                 if (namelen == 1) {
1043                                         /*
1044                                          * skip "."
1045                                          */
1046                                         continue;
1047                                 } else if (name[1] == '.' && namelen == 2) {
1048                                         /*
1049                                          * skip ".."
1050                                          */
1051                                         continue;
1052                                 } else if (!sai->sai_ls_all) {
1053                                         /*
1054                                          * skip hidden files.
1055                                          */
1056                                         sai->sai_skip_hidden++;
1057                                         continue;
1058                                 }
1059                         }
1060
1061                         /*
1062                          * don't stat-ahead first entry.
1063                          */
1064                         if (unlikely(++first == 1))
1065                                 continue;
1066
1067                         fid_le_to_cpu(&fid, &ent->lde_fid);
1068
1069                         /* wait for spare statahead window */
1070                         do {
1071                                 l_wait_event(sa_thread->t_ctl_waitq,
1072                                              !sa_sent_full(sai) ||
1073                                              sa_has_callback(sai) ||
1074                                              !list_empty(&sai->sai_agls) ||
1075                                              !thread_is_running(sa_thread),
1076                                              &lwi);
1077                                 sa_handle_callback(sai);
1078
1079                                 spin_lock(&lli->lli_agl_lock);
1080                                 while (sa_sent_full(sai) &&
1081                                        !agl_list_empty(sai)) {
1082                                         struct ll_inode_info *clli;
1083
1084                                         clli = list_entry(sai->sai_agls.next,
1085                                                           struct ll_inode_info, lli_agl_list);
1086                                         list_del_init(&clli->lli_agl_list);
1087                                         spin_unlock(&lli->lli_agl_lock);
1088
1089                                         ll_agl_trigger(&clli->lli_vfs_inode,
1090                                                        sai);
1091
1092                                         spin_lock(&lli->lli_agl_lock);
1093                                 }
1094                                 spin_unlock(&lli->lli_agl_lock);
1095                         } while (sa_sent_full(sai) &&
1096                                  thread_is_running(sa_thread));
1097
1098                         sa_statahead(parent, name, namelen, &fid);
1099                 }
1100
1101                 pos = le64_to_cpu(dp->ldp_hash_end);
1102                 ll_release_page(dir, page,
1103                                 le32_to_cpu(dp->ldp_flags) & LDF_COLLIDE);
1104
1105                 if (sa_low_hit(sai)) {
1106                         rc = -EFAULT;
1107                         atomic_inc(&sbi->ll_sa_wrong);
1108                         CDEBUG(D_READA, "Statahead for dir " DFID " hit ratio too low: hit/miss %llu/%llu, sent/replied %llu/%llu, stopping statahead thread: pid %d\n",
1109                                PFID(&lli->lli_fid), sai->sai_hit,
1110                                sai->sai_miss, sai->sai_sent,
1111                                sai->sai_replied, current_pid());
1112                         break;
1113                 }
1114         }
1115         ll_finish_md_op_data(op_data);
1116
1117         if (rc < 0) {
1118                 spin_lock(&lli->lli_sa_lock);
1119                 thread_set_flags(sa_thread, SVC_STOPPING);
1120                 lli->lli_sa_enabled = 0;
1121                 spin_unlock(&lli->lli_sa_lock);
1122         }
1123
1124         /*
1125          * statahead is finished, but statahead entries need to be cached, wait
1126          * for file release to stop me.
1127          */
1128         while (thread_is_running(sa_thread)) {
1129                 l_wait_event(sa_thread->t_ctl_waitq,
1130                              sa_has_callback(sai) ||
1131                              !agl_list_empty(sai) ||
1132                              !thread_is_running(sa_thread),
1133                              &lwi);
1134
1135                 sa_handle_callback(sai);
1136         }
1137 out:
1138         if (sai->sai_agl_valid) {
1139                 spin_lock(&lli->lli_agl_lock);
1140                 thread_set_flags(agl_thread, SVC_STOPPING);
1141                 spin_unlock(&lli->lli_agl_lock);
1142                 wake_up(&agl_thread->t_ctl_waitq);
1143
1144                 CDEBUG(D_READA, "stop agl thread: sai %p pid %u\n",
1145                        sai, (unsigned int)agl_thread->t_pid);
1146                 l_wait_event(agl_thread->t_ctl_waitq,
1147                              thread_is_stopped(agl_thread),
1148                              &lwi);
1149         } else {
1150                 /* Set agl_thread flags anyway. */
1151                 thread_set_flags(agl_thread, SVC_STOPPED);
1152         }
1153
1154         /*
1155          * wait for inflight statahead RPCs to finish, and then we can free sai
1156          * safely because statahead RPC will access sai data
1157          */
1158         while (sai->sai_sent != sai->sai_replied) {
1159                 /* in case we're not woken up, timeout wait */
1160                 lwi = LWI_TIMEOUT(msecs_to_jiffies(MSEC_PER_SEC >> 3),
1161                                   NULL, NULL);
1162                 l_wait_event(sa_thread->t_ctl_waitq,
1163                              sai->sai_sent == sai->sai_replied, &lwi);
1164         }
1165
1166         /* release resources held by statahead RPCs */
1167         sa_handle_callback(sai);
1168
1169         spin_lock(&lli->lli_sa_lock);
1170         thread_set_flags(sa_thread, SVC_STOPPED);
1171         spin_unlock(&lli->lli_sa_lock);
1172
1173         CDEBUG(D_READA, "statahead thread stopped: sai %p, parent %pd\n",
1174                sai, parent);
1175
1176         wake_up(&sai->sai_waitq);
1177         wake_up(&sa_thread->t_ctl_waitq);
1178         ll_sai_put(sai);
1179
1180         return rc;
1181 }
1182
1183 /* authorize opened dir handle @key to statahead */
1184 void ll_authorize_statahead(struct inode *dir, void *key)
1185 {
1186         struct ll_inode_info *lli = ll_i2info(dir);
1187
1188         spin_lock(&lli->lli_sa_lock);
1189         if (!lli->lli_opendir_key && !lli->lli_sai) {
1190                 /*
1191                  * if lli_sai is not NULL, it means previous statahead is not
1192                  * finished yet, we'd better not start a new statahead for now.
1193                  */
1194                 LASSERT(!lli->lli_opendir_pid);
1195                 lli->lli_opendir_key = key;
1196                 lli->lli_opendir_pid = current_pid();
1197                 lli->lli_sa_enabled = 1;
1198         }
1199         spin_unlock(&lli->lli_sa_lock);
1200 }
1201
1202 /*
1203  * deauthorize opened dir handle @key to statahead, but statahead thread may
1204  * still be running, notify it to quit.
1205  */
1206 void ll_deauthorize_statahead(struct inode *dir, void *key)
1207 {
1208         struct ll_inode_info *lli = ll_i2info(dir);
1209         struct ll_statahead_info *sai;
1210
1211         LASSERT(lli->lli_opendir_key == key);
1212         LASSERT(lli->lli_opendir_pid);
1213
1214         CDEBUG(D_READA, "deauthorize statahead for " DFID "\n",
1215                PFID(&lli->lli_fid));
1216
1217         spin_lock(&lli->lli_sa_lock);
1218         lli->lli_opendir_key = NULL;
1219         lli->lli_opendir_pid = 0;
1220         lli->lli_sa_enabled = 0;
1221         sai = lli->lli_sai;
1222         if (sai && thread_is_running(&sai->sai_thread)) {
1223                 /*
1224                  * statahead thread may not quit yet because it needs to cache
1225                  * entries, now it's time to tell it to quit.
1226                  */
1227                 thread_set_flags(&sai->sai_thread, SVC_STOPPING);
1228                 wake_up(&sai->sai_thread.t_ctl_waitq);
1229         }
1230         spin_unlock(&lli->lli_sa_lock);
1231 }
1232
1233 enum {
1234         /**
1235          * not first dirent, or is "."
1236          */
1237         LS_NOT_FIRST_DE = 0,
1238         /**
1239          * the first non-hidden dirent
1240          */
1241         LS_FIRST_DE,
1242         /**
1243          * the first hidden dirent, that is "."
1244          */
1245         LS_FIRST_DOT_DE
1246 };
1247
1248 /* file is first dirent under @dir */
1249 static int is_first_dirent(struct inode *dir, struct dentry *dentry)
1250 {
1251         const struct qstr  *target = &dentry->d_name;
1252         struct md_op_data *op_data;
1253         struct page       *page;
1254         __u64            pos    = 0;
1255         int                dot_de;
1256         int rc = LS_NOT_FIRST_DE;
1257
1258         op_data = ll_prep_md_op_data(NULL, dir, dir, NULL, 0, 0,
1259                                      LUSTRE_OPC_ANY, dir);
1260         if (IS_ERR(op_data))
1261                 return PTR_ERR(op_data);
1262         /**
1263          * FIXME choose the start offset of the readdir
1264          */
1265         op_data->op_max_pages = ll_i2sbi(dir)->ll_md_brw_pages;
1266
1267         page = ll_get_dir_page(dir, op_data, pos);
1268
1269         while (1) {
1270                 struct lu_dirpage *dp;
1271                 struct lu_dirent  *ent;
1272
1273                 if (IS_ERR(page)) {
1274                         struct ll_inode_info *lli = ll_i2info(dir);
1275
1276                         rc = PTR_ERR(page);
1277                         CERROR("%s: error reading dir " DFID " at %llu: opendir_pid = %u : rc = %d\n",
1278                                ll_get_fsname(dir->i_sb, NULL, 0),
1279                                PFID(ll_inode2fid(dir)), pos,
1280                                lli->lli_opendir_pid, rc);
1281                         break;
1282                 }
1283
1284                 dp = page_address(page);
1285                 for (ent = lu_dirent_start(dp); ent;
1286                      ent = lu_dirent_next(ent)) {
1287                         __u64 hash;
1288                         int namelen;
1289                         char *name;
1290
1291                         hash = le64_to_cpu(ent->lde_hash);
1292                         /* The ll_get_dir_page() can return any page containing
1293                          * the given hash which may be not the start hash.
1294                          */
1295                         if (unlikely(hash < pos))
1296                                 continue;
1297
1298                         namelen = le16_to_cpu(ent->lde_namelen);
1299                         if (unlikely(namelen == 0))
1300                                 /*
1301                                  * skip dummy record.
1302                                  */
1303                                 continue;
1304
1305                         name = ent->lde_name;
1306                         if (name[0] == '.') {
1307                                 if (namelen == 1)
1308                                         /*
1309                                          * skip "."
1310                                          */
1311                                         continue;
1312                                 else if (name[1] == '.' && namelen == 2)
1313                                         /*
1314                                          * skip ".."
1315                                          */
1316                                         continue;
1317                                 else
1318                                         dot_de = 1;
1319                         } else {
1320                                 dot_de = 0;
1321                         }
1322
1323                         if (dot_de && target->name[0] != '.') {
1324                                 CDEBUG(D_READA, "%.*s skip hidden file %.*s\n",
1325                                        target->len, target->name,
1326                                        namelen, name);
1327                                 continue;
1328                         }
1329
1330                         if (target->len != namelen ||
1331                             memcmp(target->name, name, namelen) != 0)
1332                                 rc = LS_NOT_FIRST_DE;
1333                         else if (!dot_de)
1334                                 rc = LS_FIRST_DE;
1335                         else
1336                                 rc = LS_FIRST_DOT_DE;
1337
1338                         ll_release_page(dir, page, false);
1339                         goto out;
1340                 }
1341                 pos = le64_to_cpu(dp->ldp_hash_end);
1342                 if (pos == MDS_DIR_END_OFF) {
1343                         /*
1344                          * End of directory reached.
1345                          */
1346                         ll_release_page(dir, page, false);
1347                         goto out;
1348                 } else {
1349                         /*
1350                          * chain is exhausted
1351                          * Normal case: continue to the next page.
1352                          */
1353                         ll_release_page(dir, page,
1354                                         le32_to_cpu(dp->ldp_flags) &
1355                                         LDF_COLLIDE);
1356                         page = ll_get_dir_page(dir, op_data, pos);
1357                 }
1358         }
1359 out:
1360         ll_finish_md_op_data(op_data);
1361         return rc;
1362 }
1363
1364 /**
1365  * revalidate @dentryp from statahead cache
1366  *
1367  * \param[in]  dir      parent directory
1368  * \param[in]  sai      sai structure
1369  * \param[out] dentryp  pointer to dentry which will be revalidated
1370  * \param[in]  unplug   unplug statahead window only (normally for negative
1371  *                      dentry)
1372  * \retval              1 on success, dentry is saved in @dentryp
1373  * \retval              0 if revalidation failed (no proper lock on client)
1374  * \retval              negative number upon error
1375  */
1376 static int revalidate_statahead_dentry(struct inode *dir,
1377                                        struct ll_statahead_info *sai,
1378                                        struct dentry **dentryp,
1379                                        bool unplug)
1380 {
1381         struct ll_inode_info *lli = ll_i2info(dir);
1382         struct sa_entry *entry = NULL;
1383         struct l_wait_info lwi = { 0 };
1384         struct ll_dentry_data *ldd;
1385         int rc = 0;
1386
1387         if ((*dentryp)->d_name.name[0] == '.') {
1388                 if (sai->sai_ls_all ||
1389                     sai->sai_miss_hidden >= sai->sai_skip_hidden) {
1390                         /*
1391                          * Hidden dentry is the first one, or statahead
1392                          * thread does not skip so many hidden dentries
1393                          * before "sai_ls_all" enabled as below.
1394                          */
1395                 } else {
1396                         if (!sai->sai_ls_all)
1397                                 /*
1398                                  * It maybe because hidden dentry is not
1399                                  * the first one, "sai_ls_all" was not
1400                                  * set, then "ls -al" missed. Enable
1401                                  * "sai_ls_all" for such case.
1402                                  */
1403                                 sai->sai_ls_all = 1;
1404
1405                         /*
1406                          * Such "getattr" has been skipped before
1407                          * "sai_ls_all" enabled as above.
1408                          */
1409                         sai->sai_miss_hidden++;
1410                         return -EAGAIN;
1411                 }
1412         }
1413
1414         if (unplug) {
1415                 rc = 1;
1416                 goto out_unplug;
1417         }
1418
1419         entry = sa_get(sai, &(*dentryp)->d_name);
1420         if (!entry) {
1421                 rc = -EAGAIN;
1422                 goto out_unplug;
1423         }
1424
1425         /* if statahead is busy in readdir, help it do post-work */
1426         if (!sa_ready(entry) && sai->sai_in_readpage)
1427                 sa_handle_callback(sai);
1428
1429         if (!sa_ready(entry)) {
1430                 spin_lock(&lli->lli_sa_lock);
1431                 sai->sai_index_wait = entry->se_index;
1432                 spin_unlock(&lli->lli_sa_lock);
1433                 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(30), NULL,
1434                                        LWI_ON_SIGNAL_NOOP, NULL);
1435                 rc = l_wait_event(sai->sai_waitq, sa_ready(entry), &lwi);
1436                 if (rc < 0) {
1437                         /*
1438                          * entry may not be ready, so it may be used by inflight
1439                          * statahead RPC, don't free it.
1440                          */
1441                         entry = NULL;
1442                         rc = -EAGAIN;
1443                         goto out_unplug;
1444                 }
1445         }
1446
1447         if (entry->se_state == SA_ENTRY_SUCC && entry->se_inode) {
1448                 struct inode *inode = entry->se_inode;
1449                 struct lookup_intent it = { .it_op = IT_GETATTR,
1450                                             .it_lock_handle = entry->se_handle };
1451                 __u64 bits;
1452
1453                 rc = md_revalidate_lock(ll_i2mdexp(dir), &it,
1454                                         ll_inode2fid(inode), &bits);
1455                 if (rc == 1) {
1456                         if (!(*dentryp)->d_inode) {
1457                                 struct dentry *alias;
1458
1459                                 alias = ll_splice_alias(inode, *dentryp);
1460                                 if (IS_ERR(alias)) {
1461                                         ll_intent_release(&it);
1462                                         rc = PTR_ERR(alias);
1463                                         goto out_unplug;
1464                                 }
1465                                 *dentryp = alias;
1466                                 /**
1467                                  * statahead prepared this inode, transfer inode
1468                                  * refcount from sa_entry to dentry
1469                                  */
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 %pd inode " DFID ", statahead inode " DFID "\n",
1475                                        ll_get_fsname((*dentryp)->d_inode->i_sb,
1476                                                      NULL, 0),
1477                                        *dentryp,
1478                                        PFID(ll_inode2fid((*dentryp)->d_inode)),
1479                                        PFID(ll_inode2fid(inode)));
1480                                 ll_intent_release(&it);
1481                                 rc = -ESTALE;
1482                                 goto out_unplug;
1483                         }
1484
1485                         if ((bits & MDS_INODELOCK_LOOKUP) &&
1486                             d_lustre_invalid(*dentryp))
1487                                 d_lustre_revalidate(*dentryp);
1488                         ll_intent_release(&it);
1489                 }
1490         }
1491 out_unplug:
1492         /*
1493          * statahead cached sa_entry can be used only once, and will be killed
1494          * right after use, so if lookup/revalidate accessed statahead cache,
1495          * set dentry ldd_sa_generation to parent lli_sa_generation, later if we
1496          * stat this file again, we know we've done statahead before, see
1497          * dentry_may_statahead().
1498          */
1499         ldd = ll_d2d(*dentryp);
1500         ldd->lld_sa_generation = lli->lli_sa_generation;
1501         sa_put(sai, entry);
1502         return rc;
1503 }
1504
1505 /**
1506  * start statahead thread
1507  *
1508  * \param[in] dir       parent directory
1509  * \param[in] dentry    dentry that triggers statahead, normally the first
1510  *                      dirent under @dir
1511  * \retval              -EAGAIN on success, because when this function is
1512  *                      called, it's already in lookup call, so client should
1513  *                      do it itself instead of waiting for statahead thread
1514  *                      to do it asynchronously.
1515  * \retval              negative number upon error
1516  */
1517 static int start_statahead_thread(struct inode *dir, struct dentry *dentry)
1518 {
1519         struct ll_inode_info *lli = ll_i2info(dir);
1520         struct ll_statahead_info *sai = NULL;
1521         struct l_wait_info lwi = { 0 };
1522         struct ptlrpc_thread *thread;
1523         struct task_struct *task;
1524         struct dentry *parent = dentry->d_parent;
1525         int rc;
1526
1527         /* I am the "lli_opendir_pid" owner, only me can set "lli_sai". */
1528         rc = is_first_dirent(dir, dentry);
1529         if (rc == LS_NOT_FIRST_DE) {
1530                 /* It is not "ls -{a}l" operation, no need statahead for it. */
1531                 rc = -EFAULT;
1532                 goto out;
1533         }
1534
1535         sai = ll_sai_alloc(parent);
1536         if (!sai) {
1537                 rc = -ENOMEM;
1538                 goto out;
1539         }
1540
1541         sai->sai_ls_all = (rc == LS_FIRST_DOT_DE);
1542         /*
1543          * if current lli_opendir_key was deauthorized, or dir re-opened by
1544          * another process, don't start statahead, otherwise the newly spawned
1545          * statahead thread won't be notified to quit.
1546          */
1547         spin_lock(&lli->lli_sa_lock);
1548         if (unlikely(lli->lli_sai || lli->lli_opendir_key ||
1549                      lli->lli_opendir_pid != current->pid)) {
1550                 spin_unlock(&lli->lli_sa_lock);
1551                 rc = -EPERM;
1552                 goto out;
1553         }
1554         lli->lli_sai = sai;
1555         spin_unlock(&lli->lli_sa_lock);
1556
1557         atomic_inc(&ll_i2sbi(parent->d_inode)->ll_sa_running);
1558
1559         CDEBUG(D_READA, "start statahead thread: [pid %d] [parent %pd]\n",
1560                current_pid(), parent);
1561
1562         task = kthread_run(ll_statahead_thread, parent, "ll_sa_%u",
1563                            lli->lli_opendir_pid);
1564         thread = &sai->sai_thread;
1565         if (IS_ERR(task)) {
1566                 rc = PTR_ERR(task);
1567                 CERROR("can't start ll_sa thread, rc : %d\n", rc);
1568                 goto out;
1569         }
1570
1571         l_wait_event(thread->t_ctl_waitq,
1572                      thread_is_running(thread) || thread_is_stopped(thread),
1573                      &lwi);
1574         ll_sai_put(sai);
1575
1576         /*
1577          * We don't stat-ahead for the first dirent since we are already in
1578          * lookup.
1579          */
1580         return -EAGAIN;
1581
1582 out:
1583         /*
1584          * once we start statahead thread failed, disable statahead so
1585          * that subsequent stat won't waste time to try it.
1586          */
1587         spin_lock(&lli->lli_sa_lock);
1588         lli->lli_sa_enabled = 0;
1589         lli->lli_sai = NULL;
1590         spin_unlock(&lli->lli_sa_lock);
1591         if (sai)
1592                 ll_sai_free(sai);
1593         return rc;
1594 }
1595
1596 /**
1597  * statahead entry function, this is called when client getattr on a file, it
1598  * will start statahead thread if this is the first dir entry, else revalidate
1599  * dentry from statahead cache.
1600  *
1601  * \param[in]  dir      parent directory
1602  * \param[out] dentryp  dentry to getattr
1603  * \param[in]  unplug   unplug statahead window only (normally for negative
1604  *                      dentry)
1605  * \retval              1 on success
1606  * \retval              0 revalidation from statahead cache failed, caller needs
1607  *                      to getattr from server directly
1608  * \retval              negative number on error, caller often ignores this and
1609  *                      then getattr from server
1610  */
1611 int ll_statahead(struct inode *dir, struct dentry **dentryp, bool unplug)
1612 {
1613         struct ll_statahead_info *sai;
1614
1615         sai = ll_sai_get(dir);
1616         if (sai) {
1617                 int rc;
1618
1619                 rc = revalidate_statahead_dentry(dir, sai, dentryp, unplug);
1620                 CDEBUG(D_READA, "revalidate statahead %pd: %d.\n",
1621                        *dentryp, rc);
1622                 ll_sai_put(sai);
1623                 return rc;
1624         }
1625         return start_statahead_thread(dir, *dentryp);
1626 }