lustre/llite/statahead.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2014, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  */
  36
  37 #include <linux/fs.h>
  38 #include <linux/sched.h>
  39 #include <linux/kthread.h>
  40 #include <linux/mm.h>
  41 #include <linux/highmem.h>
  42 #include <linux/pagemap.h>
  43
  44 #define DEBUG_SUBSYSTEM S_LLITE
  45
  46 #include <obd_support.h>
  47 #include <lustre_dlm.h>
  48 #include "llite_internal.h"
  49
  50 #define SA_OMITTED_ENTRY_MAX 8ULL
  51
  52 typedef enum {
  53         /** negative values are for error cases */
  54         SA_ENTRY_INIT = 0,      /** init entry */
  55         SA_ENTRY_SUCC = 1,      /** stat succeed */
  56         SA_ENTRY_INVA = 2,      /** invalid entry */
  57 } se_state_t;
  58
  59 /* sa_entry is not refcounted: statahead thread allocates it and do async stat,
  60  * and in async stat callback ll_statahead_interpret() will add it into
  61  * sai_interim_entries, later statahead thread will call sa_handle_callback() to
  62  * instantiate entry and move it into sai_entries, and then only scanner process
  63  * can access and free it. */
  64 struct sa_entry {
  65         /* link into sai_interim_entries or sai_entries */
  66         struct list_head        se_list;
  67         /* link into sai hash table locally */
  68         struct list_head        se_hash;
  69         /* entry index in the sai */
  70         __u64                   se_index;
  71         /* low layer ldlm lock handle */
  72         __u64                   se_handle;
  73         /* entry status */
  74         se_state_t              se_state;
  75         /* entry size, contains name */
  76         int                     se_size;
  77         /* pointer to async getattr enqueue info */
  78         struct md_enqueue_info *se_minfo;
  79         /* pointer to the async getattr request */
  80         struct ptlrpc_request  *se_req;
  81         /* pointer to the target inode */
  82         struct inode           *se_inode;
  83         /* entry name */
  84         struct qstr             se_qstr;
  85         /* entry fid */
  86         struct lu_fid           se_fid;
  87 };
  88
  89 static unsigned int sai_generation = 0;
  90 static DEFINE_SPINLOCK(sai_generation_lock);
  91
  92 static inline int sa_unhashed(struct sa_entry *entry)
  93 {
  94         return list_empty(&entry->se_hash);
  95 }
  96
  97 /* sa_entry is ready to use */
  98 static inline int sa_ready(struct sa_entry *entry)
  99 {
 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 != NULL && S_ISREG(inode->i_mode) && sai->sai_agl_valid);
 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 ll_statahead_info *sai, __u64 index, const char *name, int len,
 187          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         ENTRY;
 194
 195         entry_size = sizeof(struct sa_entry) + (len & ~3) + 4;
 196         OBD_ALLOC(entry, entry_size);
 197         if (unlikely(entry == NULL))
 198                 RETURN(ERR_PTR(-ENOMEM));
 199
 200         CDEBUG(D_READA, "alloc sa entry %.*s(%p) index "LPU64"\n",
 201                len, name, entry, index);
 202
 203         entry->se_index = index;
 204
 205         entry->se_state = SA_ENTRY_INIT;
 206         entry->se_size = entry_size;
 207         dname = (char *)entry + sizeof(struct sa_entry);
 208         memcpy(dname, name, len);
 209         dname[len] = 0;
 210         entry->se_qstr.hash = full_name_hash(name, len);
 211         entry->se_qstr.len = len;
 212         entry->se_qstr.name = dname;
 213         entry->se_fid = *fid;
 214
 215         lli = ll_i2info(sai->sai_dentry->d_inode);
 216
 217         spin_lock(&lli->lli_sa_lock);
 218         INIT_LIST_HEAD(&entry->se_list);
 219         sa_rehash(sai, entry);
 220         spin_unlock(&lli->lli_sa_lock);
 221
 222         atomic_inc(&sai->sai_cache_count);
 223
 224         RETURN(entry);
 225 }
 226
 227 /* free sa_entry, which should have been unhashed and not in any list */
 228 static void sa_free(struct ll_statahead_info *sai, struct sa_entry *entry)
 229 {
 230         CDEBUG(D_READA, "free sa entry %.*s(%p) index "LPU64"\n",
 231                entry->se_qstr.len, entry->se_qstr.name, entry,
 232                entry->se_index);
 233
 234         LASSERT(list_empty(&entry->se_list));
 235         LASSERT(sa_unhashed(entry));
 236
 237         OBD_FREE(entry, entry->se_size);
 238         atomic_dec(&sai->sai_cache_count);
 239 }
 240
 241 /*
 242  * find sa_entry by name, used by directory scanner, lock is not needed because
 243  * only scanner can remove the entry from cache.
 244  */
 245 static struct sa_entry *
 246 sa_get(struct ll_statahead_info *sai, const struct qstr *qstr)
 247 {
 248         struct sa_entry *entry;
 249         int i = sa_hash(qstr->hash);
 250
 251         list_for_each_entry(entry, &sai->sai_cache[i], se_hash) {
 252                 if (entry->se_qstr.hash == qstr->hash &&
 253                     entry->se_qstr.len == qstr->len &&
 254                     memcmp(entry->se_qstr.name, qstr->name, qstr->len) == 0)
 255                         return entry;
 256         }
 257         return NULL;
 258 }
 259
 260 /* unhash and unlink sa_entry, and then free it */
 261 static inline void
 262 sa_kill(struct ll_statahead_info *sai, struct sa_entry *entry)
 263 {
 264         struct ll_inode_info *lli = ll_i2info(sai->sai_dentry->d_inode);
 265
 266         LASSERT(!sa_unhashed(entry));
 267         LASSERT(!list_empty(&entry->se_list));
 268         LASSERT(sa_ready(entry));
 269
 270         sa_unhash(sai, entry);
 271
 272         spin_lock(&lli->lli_sa_lock);
 273         list_del_init(&entry->se_list);
 274         spin_unlock(&lli->lli_sa_lock);
 275
 276         if (entry->se_inode != NULL)
 277                 iput(entry->se_inode);
 278
 279         sa_free(sai, entry);
 280 }
 281
 282 /* called by scanner after use, sa_entry will be killed */
 283 static void
 284 sa_put(struct ll_statahead_info *sai, struct sa_entry *entry)
 285 {
 286         struct sa_entry *tmp, *next;
 287
 288         if (entry != NULL && entry->se_state == SA_ENTRY_SUCC) {
 289                 struct ll_sb_info *sbi = ll_i2sbi(sai->sai_dentry->d_inode);
 290
 291                 sai->sai_hit++;
 292                 sai->sai_consecutive_miss = 0;
 293                 sai->sai_max = min(2 * sai->sai_max, sbi->ll_sa_max);
 294         } else {
 295                 sai->sai_miss++;
 296                 sai->sai_consecutive_miss++;
 297         }
 298
 299         if (entry != NULL)
 300                 sa_kill(sai, entry);
 301
 302         /* kill old completed entries, only scanner process does this, no need
 303          * to lock */
 304         list_for_each_entry_safe(tmp, next, &sai->sai_entries, se_list) {
 305                 if (!is_omitted_entry(sai, tmp->se_index))
 306                         break;
 307                 sa_kill(sai, tmp);
 308         }
 309
 310         wake_up(&sai->sai_thread.t_ctl_waitq);
 311 }
 312
 313 /* update state and sort add entry to sai_entries by index, return true if
 314  * scanner is waiting on this entry. */
 315 static bool
 316 __sa_make_ready(struct ll_statahead_info *sai, struct sa_entry *entry, int ret)
 317 {
 318         struct sa_entry *se;
 319         struct list_head *pos = &sai->sai_entries;
 320         __u64 index = entry->se_index;
 321
 322         LASSERT(!sa_ready(entry));
 323         LASSERT(list_empty(&entry->se_list));
 324
 325         list_for_each_entry_reverse(se, &sai->sai_entries, se_list) {
 326                 if (se->se_index < entry->se_index) {
 327                         pos = &se->se_list;
 328                         break;
 329                 }
 330         }
 331         list_add(&entry->se_list, pos);
 332         entry->se_state = ret < 0 ? SA_ENTRY_INVA : SA_ENTRY_SUCC;
 333
 334         return (index == sai->sai_index_wait);
 335 }
 336
 337 /*
 338  * release resources used in async stat RPC, update entry state and wakeup if
 339  * scanner process it waiting on this entry.
 340  */
 341 static void
 342 sa_make_ready(struct ll_statahead_info *sai, struct sa_entry *entry, int ret)
 343 {
 344         struct ll_inode_info *lli = ll_i2info(sai->sai_dentry->d_inode);
 345         struct md_enqueue_info *minfo = entry->se_minfo;
 346         struct ptlrpc_request *req = entry->se_req;
 347         bool wakeup;
 348
 349         /* release resources used in RPC */
 350         if (minfo) {
 351                 entry->se_minfo = NULL;
 352                 ll_intent_release(&minfo->mi_it);
 353                 iput(minfo->mi_dir);
 354                 OBD_FREE_PTR(minfo);
 355         }
 356
 357         if (req) {
 358                 entry->se_req = NULL;
 359                 ptlrpc_req_finished(req);
 360         }
 361
 362         spin_lock(&lli->lli_sa_lock);
 363         wakeup = __sa_make_ready(sai, entry, ret);
 364         spin_unlock(&lli->lli_sa_lock);
 365
 366         if (wakeup)
 367                 wake_up(&sai->sai_waitq);
 368 }
 369
 370 /* insert inode into the list of sai_agls */
 371 static void ll_agl_add(struct ll_statahead_info *sai,
 372                        struct inode *inode, int index)
 373 {
 374         struct ll_inode_info *child  = ll_i2info(inode);
 375         struct ll_inode_info *parent = ll_i2info(sai->sai_dentry->d_inode);
 376         int                   added  = 0;
 377
 378         spin_lock(&child->lli_agl_lock);
 379         if (child->lli_agl_index == 0) {
 380                 child->lli_agl_index = index;
 381                 spin_unlock(&child->lli_agl_lock);
 382
 383                 LASSERT(list_empty(&child->lli_agl_list));
 384
 385                 igrab(inode);
 386                 spin_lock(&parent->lli_agl_lock);
 387                 if (agl_list_empty(sai))
 388                         added = 1;
 389                 list_add_tail(&child->lli_agl_list, &sai->sai_agls);
 390                 spin_unlock(&parent->lli_agl_lock);
 391         } else {
 392                 spin_unlock(&child->lli_agl_lock);
 393         }
 394
 395         if (added > 0)
 396                 wake_up(&sai->sai_agl_thread.t_ctl_waitq);
 397 }
 398
 399 /* allocate sai */
 400 static struct ll_statahead_info *ll_sai_alloc(struct dentry *dentry)
 401 {
 402         struct ll_statahead_info *sai;
 403         struct ll_inode_info *lli = ll_i2info(dentry->d_inode);
 404         int i;
 405         ENTRY;
 406
 407         OBD_ALLOC_PTR(sai);
 408         if (!sai)
 409                 RETURN(NULL);
 410
 411         sai->sai_dentry = dget(dentry);
 412         atomic_set(&sai->sai_refcount, 1);
 413         sai->sai_max = LL_SA_RPC_MIN;
 414         sai->sai_index = 1;
 415         init_waitqueue_head(&sai->sai_waitq);
 416         init_waitqueue_head(&sai->sai_thread.t_ctl_waitq);
 417         init_waitqueue_head(&sai->sai_agl_thread.t_ctl_waitq);
 418
 419         INIT_LIST_HEAD(&sai->sai_interim_entries);
 420         INIT_LIST_HEAD(&sai->sai_entries);
 421         INIT_LIST_HEAD(&sai->sai_agls);
 422
 423         for (i = 0; i < LL_SA_CACHE_SIZE; i++) {
 424                 INIT_LIST_HEAD(&sai->sai_cache[i]);
 425                 spin_lock_init(&sai->sai_cache_lock[i]);
 426         }
 427         atomic_set(&sai->sai_cache_count, 0);
 428
 429         spin_lock(&sai_generation_lock);
 430         lli->lli_sa_generation = ++sai_generation;
 431         if (unlikely(sai_generation == 0))
 432                 lli->lli_sa_generation = ++sai_generation;
 433         spin_unlock(&sai_generation_lock);
 434
 435         RETURN(sai);
 436 }
 437
 438 /* free sai */
 439 static inline void ll_sai_free(struct ll_statahead_info *sai)
 440 {
 441         LASSERT(sai->sai_dentry != NULL);
 442         dput(sai->sai_dentry);
 443         OBD_FREE_PTR(sai);
 444 }
 445
 446 /*
 447  * take refcount of sai if sai for @dir exists, which means statahead is on for
 448  * this directory.
 449  */
 450 static inline struct ll_statahead_info *ll_sai_get(struct inode *dir)
 451 {
 452         struct ll_inode_info *lli = ll_i2info(dir);
 453         struct ll_statahead_info *sai = NULL;
 454
 455         spin_lock(&lli->lli_sa_lock);
 456         sai = lli->lli_sai;
 457         if (sai != NULL)
 458                 atomic_inc(&sai->sai_refcount);
 459         spin_unlock(&lli->lli_sa_lock);
 460
 461         return sai;
 462 }
 463
 464 /*
 465  * put sai refcount after use, if refcount reaches zero, free sai and sa_entries
 466  * attached to it.
 467  */
 468 static void ll_sai_put(struct ll_statahead_info *sai)
 469 {
 470         struct ll_inode_info *lli = ll_i2info(sai->sai_dentry->d_inode);
 471
 472         if (atomic_dec_and_lock(&sai->sai_refcount, &lli->lli_sa_lock)) {
 473                 struct sa_entry *entry, *next;
 474                 struct ll_sb_info *sbi = ll_i2sbi(sai->sai_dentry->d_inode);
 475
 476                 lli->lli_sai = NULL;
 477                 spin_unlock(&lli->lli_sa_lock);
 478
 479                 LASSERT(thread_is_stopped(&sai->sai_thread));
 480                 LASSERT(thread_is_stopped(&sai->sai_agl_thread));
 481                 LASSERT(sai->sai_sent == sai->sai_replied);
 482                 LASSERT(!sa_has_callback(sai));
 483
 484                 list_for_each_entry_safe(entry, next, &sai->sai_entries,
 485                                          se_list)
 486                         sa_kill(sai, entry);
 487
 488                 LASSERT(atomic_read(&sai->sai_cache_count) == 0);
 489                 LASSERT(agl_list_empty(sai));
 490
 491                 ll_sai_free(sai);
 492                 atomic_dec(&sbi->ll_sa_running);
 493         }
 494 }
 495
 496 /* Do NOT forget to drop inode refcount when into sai_agls. */
 497 static void ll_agl_trigger(struct inode *inode, struct ll_statahead_info *sai)
 498 {
 499         struct ll_inode_info *lli = ll_i2info(inode);
 500         __u64 index = lli->lli_agl_index;
 501         int rc;
 502         ENTRY;
 503
 504         LASSERT(list_empty(&lli->lli_agl_list));
 505
 506         /* AGL maybe fall behind statahead with one entry */
 507         if (is_omitted_entry(sai, index + 1)) {
 508                 lli->lli_agl_index = 0;
 509                 iput(inode);
 510                 RETURN_EXIT;
 511         }
 512
 513         /* Someone is in glimpse (sync or async), do nothing. */
 514         rc = down_write_trylock(&lli->lli_glimpse_sem);
 515         if (rc == 0) {
 516                 lli->lli_agl_index = 0;
 517                 iput(inode);
 518                 RETURN_EXIT;
 519         }
 520
 521         /*
 522          * Someone triggered glimpse within 1 sec before.
 523          * 1) The former glimpse succeeded with glimpse lock granted by OST, and
 524          *    if the lock is still cached on client, AGL needs to do nothing. If
 525          *    it is cancelled by other client, AGL maybe cannot obtaion new lock
 526          *    for no glimpse callback triggered by AGL.
 527          * 2) The former glimpse succeeded, but OST did not grant glimpse lock.
 528          *    Under such case, it is quite possible that the OST will not grant
 529          *    glimpse lock for AGL also.
 530          * 3) The former glimpse failed, compared with other two cases, it is
 531          *    relative rare. AGL can ignore such case, and it will not muchly
 532          *    affect the performance.
 533          */
 534         if (lli->lli_glimpse_time != 0 &&
 535             cfs_time_before(cfs_time_shift(-1), lli->lli_glimpse_time)) {
 536                 up_write(&lli->lli_glimpse_sem);
 537                 lli->lli_agl_index = 0;
 538                 iput(inode);
 539                 RETURN_EXIT;
 540         }
 541
 542         CDEBUG(D_READA, "Handling (init) async glimpse: inode = "
 543                DFID", idx = "LPU64"\n", PFID(&lli->lli_fid), index);
 544
 545         cl_agl(inode);
 546         lli->lli_agl_index = 0;
 547         lli->lli_glimpse_time = cfs_time_current();
 548         up_write(&lli->lli_glimpse_sem);
 549
 550         CDEBUG(D_READA, "Handled (init) async glimpse: inode= "
 551                DFID", idx = "LPU64", rc = %d\n",
 552                PFID(&lli->lli_fid), index, rc);
 553
 554         iput(inode);
 555
 556         EXIT;
 557 }
 558
 559 /*
 560  * prepare inode for sa entry, add it into agl list, now sa_entry is ready
 561  * to be used by scanner process.
 562  */
 563 static void sa_instantiate(struct ll_statahead_info *sai,
 564                                  struct sa_entry *entry)
 565 {
 566         struct inode *dir = sai->sai_dentry->d_inode;
 567         struct inode *child;
 568         struct md_enqueue_info *minfo;
 569         struct lookup_intent *it;
 570         struct ptlrpc_request *req;
 571         struct mdt_body *body;
 572         int rc = 0;
 573         ENTRY;
 574
 575         LASSERT(entry->se_handle != 0);
 576
 577         minfo = entry->se_minfo;
 578         it = &minfo->mi_it;
 579         req = entry->se_req;
 580         body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
 581         if (body == NULL)
 582                 GOTO(out, rc = -EFAULT);
 583
 584         child = entry->se_inode;
 585         if (child != NULL) {
 586                 /* revalidate; unlinked and re-created with the same name */
 587                 if (unlikely(!lu_fid_eq(&minfo->mi_data.op_fid2,
 588                                         &body->mbo_fid1))) {
 589                         entry->se_inode = NULL;
 590                         iput(child);
 591                         child = NULL;
 592                 }
 593         }
 594
 595         it->d.lustre.it_lock_handle = entry->se_handle;
 596         rc = md_revalidate_lock(ll_i2mdexp(dir), it, ll_inode2fid(dir), NULL);
 597         if (rc != 1)
 598                 GOTO(out, rc = -EAGAIN);
 599
 600         rc = ll_prep_inode(&child, req, dir->i_sb, it);
 601         if (rc)
 602                 GOTO(out, rc);
 603
 604         CDEBUG(D_READA, "%s: setting %.*s"DFID" l_data to inode %p\n",
 605                ll_get_fsname(child->i_sb, NULL, 0),
 606                entry->se_qstr.len, entry->se_qstr.name,
 607                PFID(ll_inode2fid(child)), child);
 608         ll_set_lock_data(ll_i2sbi(dir)->ll_md_exp, child, it, NULL);
 609
 610         entry->se_inode = child;
 611
 612         if (agl_should_run(sai, child))
 613                 ll_agl_add(sai, child, entry->se_index);
 614
 615         EXIT;
 616
 617 out:
 618         /* sa_make_ready() will drop ldlm ibits lock refcount by calling
 619          * ll_intent_drop_lock() in spite of failures. Do not worry about
 620          * calling ll_intent_drop_lock() more than once. */
 621         sa_make_ready(sai, entry, rc);
 622 }
 623
 624 /* once there are async stat replies, instantiate sa_entry from replies */
 625 static void sa_handle_callback(struct ll_statahead_info *sai)
 626 {
 627         struct ll_inode_info *lli;
 628
 629         lli = ll_i2info(sai->sai_dentry->d_inode);
 630
 631         while (sa_has_callback(sai)) {
 632                 struct sa_entry *entry;
 633
 634                 spin_lock(&lli->lli_sa_lock);
 635                 if (unlikely(!sa_has_callback(sai))) {
 636                         spin_unlock(&lli->lli_sa_lock);
 637                         break;
 638                 }
 639                 entry = list_entry(sai->sai_interim_entries.next,
 640                                    struct sa_entry, se_list);
 641                 list_del_init(&entry->se_list);
 642                 spin_unlock(&lli->lli_sa_lock);
 643
 644                 sa_instantiate(sai, entry);
 645         }
 646 }
 647
 648 /*
 649  * callback for async stat RPC, because this is called in ptlrpcd context, we
 650  * only put sa_entry in sai_interim_entries, and wake up statahead thread to
 651  * really prepare inode and instantiate sa_entry later.
 652  */
 653 static int ll_statahead_interpret(struct ptlrpc_request *req,
 654                                   struct md_enqueue_info *minfo, int rc)
 655 {
 656         struct lookup_intent *it = &minfo->mi_it;
 657         struct inode *dir = minfo->mi_dir;
 658         struct ll_inode_info *lli = ll_i2info(dir);
 659         struct ll_statahead_info *sai = lli->lli_sai;
 660         struct sa_entry *entry = (struct sa_entry *)minfo->mi_cbdata;
 661         __u64 handle = 0;
 662         bool wakeup;
 663         ENTRY;
 664
 665         if (it_disposition(it, DISP_LOOKUP_NEG))
 666                 rc = -ENOENT;
 667
 668         /* because statahead thread will wait for all inflight RPC to finish,
 669          * sai should be always valid, no need to refcount */
 670         LASSERT(sai != NULL);
 671         LASSERT(!thread_is_stopped(&sai->sai_thread));
 672         LASSERT(entry != NULL);
 673
 674         CDEBUG(D_READA, "sa_entry %.*s rc %d\n",
 675                entry->se_qstr.len, entry->se_qstr.name, rc);
 676
 677         if (rc != 0) {
 678                 ll_intent_release(it);
 679                 iput(dir);
 680                 OBD_FREE_PTR(minfo);
 681         } else {
 682                 /* release ibits lock ASAP to avoid deadlock when statahead
 683                  * thread enqueues lock on parent in readdir and another
 684                  * process enqueues lock on child with parent lock held, eg.
 685                  * unlink. */
 686                 handle = it->d.lustre.it_lock_handle;
 687                 ll_intent_drop_lock(it);
 688         }
 689
 690         spin_lock(&lli->lli_sa_lock);
 691         if (rc != 0) {
 692                 wakeup = __sa_make_ready(sai, entry, rc);
 693         } else {
 694                 entry->se_minfo = minfo;
 695                 entry->se_req = ptlrpc_request_addref(req);
 696                 /* Release the async ibits lock ASAP to avoid deadlock
 697                  * when statahead thread tries to enqueue lock on parent
 698                  * for readpage and other tries to enqueue lock on child
 699                  * with parent's lock held, for example: unlink. */
 700                 entry->se_handle = handle;
 701                 wakeup = !sa_has_callback(sai);
 702                 list_add_tail(&entry->se_list, &sai->sai_interim_entries);
 703         }
 704         sai->sai_replied++;
 705         if (wakeup)
 706                 wake_up(&sai->sai_thread.t_ctl_waitq);
 707         spin_unlock(&lli->lli_sa_lock);
 708
 709         RETURN(rc);
 710 }
 711
 712 /* finish async stat RPC arguments */
 713 static void sa_fini_data(struct md_enqueue_info *minfo)
 714 {
 715         iput(minfo->mi_dir);
 716         OBD_FREE_PTR(minfo);
 717 }
 718
 719 /*
 720  * prepare arguments for async stat RPC.
 721  */
 722 static struct md_enqueue_info *
 723 sa_prep_data(struct inode *dir, struct inode *child, struct sa_entry *entry)
 724 {
 725         struct md_enqueue_info   *minfo;
 726         struct ldlm_enqueue_info *einfo;
 727         struct md_op_data        *op_data;
 728
 729         OBD_ALLOC_PTR(minfo);
 730         if (minfo == NULL)
 731                 return ERR_PTR(-ENOMEM);
 732
 733         op_data = ll_prep_md_op_data(&minfo->mi_data, dir, child, NULL, 0, 0,
 734                                      LUSTRE_OPC_ANY, NULL);
 735         if (IS_ERR(op_data)) {
 736                 OBD_FREE_PTR(minfo);
 737                 return (struct md_enqueue_info *)op_data;
 738         }
 739
 740         if (child == NULL)
 741                 op_data->op_fid2 = entry->se_fid;
 742
 743         minfo->mi_it.it_op = IT_GETATTR;
 744         minfo->mi_dir = igrab(dir);
 745         minfo->mi_cb = ll_statahead_interpret;
 746         minfo->mi_cbdata = entry;
 747
 748         einfo = &minfo->mi_einfo;
 749         einfo->ei_type   = LDLM_IBITS;
 750         einfo->ei_mode   = it_to_lock_mode(&minfo->mi_it);
 751         einfo->ei_cb_bl  = ll_md_blocking_ast;
 752         einfo->ei_cb_cp  = ldlm_completion_ast;
 753         einfo->ei_cb_gl  = NULL;
 754         einfo->ei_cbdata = NULL;
 755
 756         return minfo;
 757 }
 758
 759 /* async stat for file not found in dcache */
 760 static int sa_lookup(struct inode *dir, struct sa_entry *entry)
 761 {
 762         struct md_enqueue_info   *minfo;
 763         int                       rc;
 764         ENTRY;
 765
 766         minfo = sa_prep_data(dir, NULL, entry);
 767         if (IS_ERR(minfo))
 768                 RETURN(PTR_ERR(minfo));
 769
 770         rc = md_intent_getattr_async(ll_i2mdexp(dir), minfo);
 771         if (rc < 0)
 772                 sa_fini_data(minfo);
 773
 774         RETURN(rc);
 775 }
 776
 777 /**
 778  * async stat for file found in dcache, similar to .revalidate
 779  *
 780  * \retval      1 dentry valid, no RPC sent
 781  * \retval      0 dentry invalid, will send async stat RPC
 782  * \retval      negative number upon error
 783  */
 784 static int sa_revalidate(struct inode *dir, struct sa_entry *entry,
 785                          struct dentry *dentry)
 786 {
 787         struct inode *inode = dentry->d_inode;
 788         struct lookup_intent it = { .it_op = IT_GETATTR,
 789                                     .d.lustre.it_lock_handle = 0 };
 790         struct md_enqueue_info *minfo;
 791         int rc;
 792         ENTRY;
 793
 794         if (unlikely(inode == NULL))
 795                 RETURN(1);
 796
 797         if (d_mountpoint(dentry))
 798                 RETURN(1);
 799
 800         entry->se_inode = igrab(inode);
 801         rc = md_revalidate_lock(ll_i2mdexp(dir), &it, ll_inode2fid(inode),
 802                                 NULL);
 803         if (rc == 1) {
 804                 entry->se_handle = it.d.lustre.it_lock_handle;
 805                 ll_intent_release(&it);
 806                 RETURN(1);
 807         }
 808
 809         minfo = sa_prep_data(dir, inode, entry);
 810         if (IS_ERR(minfo)) {
 811                 entry->se_inode = NULL;
 812                 iput(inode);
 813                 RETURN(PTR_ERR(minfo));
 814         }
 815
 816         rc = md_intent_getattr_async(ll_i2mdexp(dir), minfo);
 817         if (rc < 0) {
 818                 entry->se_inode = NULL;
 819                 iput(inode);
 820                 sa_fini_data(minfo);
 821         }
 822
 823         RETURN(rc);
 824 }
 825
 826 /* async stat for file with @name */
 827 static void sa_statahead(struct dentry *parent, const char *name, int len,
 828                          const struct lu_fid *fid)
 829 {
 830         struct inode *dir = parent->d_inode;
 831         struct ll_inode_info *lli = ll_i2info(dir);
 832         struct ll_statahead_info *sai = lli->lli_sai;
 833         struct dentry *dentry = NULL;
 834         struct sa_entry *entry;
 835         int rc;
 836         ENTRY;
 837
 838         entry = sa_alloc(sai, sai->sai_index, name, len, fid);
 839         if (IS_ERR(entry))
 840                 RETURN_EXIT;
 841
 842         dentry = d_lookup(parent, &entry->se_qstr);
 843         if (!dentry) {
 844                 rc = sa_lookup(dir, entry);
 845         } else {
 846                 rc = sa_revalidate(dir, entry, dentry);
 847                 if (rc == 1 && agl_should_run(sai, dentry->d_inode))
 848                         ll_agl_add(sai, dentry->d_inode, entry->se_index);
 849         }
 850
 851         if (dentry != NULL)
 852                 dput(dentry);
 853
 854         if (rc != 0)
 855                 sa_make_ready(sai, entry, rc);
 856         else
 857                 sai->sai_sent++;
 858
 859         sai->sai_index++;
 860
 861         EXIT;
 862 }
 863
 864 /* async glimpse (agl) thread main function */
 865 static int ll_agl_thread(void *arg)
 866 {
 867         struct dentry *parent = (struct dentry *)arg;
 868         struct inode *dir = parent->d_inode;
 869         struct ll_inode_info *plli = ll_i2info(dir);
 870         struct ll_inode_info *clli;
 871         struct ll_sb_info *sbi = ll_i2sbi(dir);
 872         struct ll_statahead_info *sai;
 873         struct ptlrpc_thread *thread;
 874         struct l_wait_info lwi = { 0 };
 875         ENTRY;
 876
 877
 878         sai = ll_sai_get(dir);
 879         thread = &sai->sai_agl_thread;
 880         thread->t_pid = current_pid();
 881         CDEBUG(D_READA, "agl thread started: sai %p, parent %.*s\n",
 882                sai, parent->d_name.len, parent->d_name.name);
 883
 884         atomic_inc(&sbi->ll_agl_total);
 885         spin_lock(&plli->lli_agl_lock);
 886         sai->sai_agl_valid = 1;
 887         if (thread_is_init(thread))
 888                 /* If someone else has changed the thread state
 889                  * (e.g. already changed to SVC_STOPPING), we can't just
 890                  * blindly overwrite that setting. */
 891                 thread_set_flags(thread, SVC_RUNNING);
 892         spin_unlock(&plli->lli_agl_lock);
 893         wake_up(&thread->t_ctl_waitq);
 894
 895         while (1) {
 896                 l_wait_event(thread->t_ctl_waitq,
 897                              !agl_list_empty(sai) ||
 898                              !thread_is_running(thread),
 899                              &lwi);
 900
 901                 if (!thread_is_running(thread))
 902                         break;
 903
 904                 spin_lock(&plli->lli_agl_lock);
 905                 /* The statahead thread maybe help to process AGL entries,
 906                  * so check whether list empty again. */
 907                 if (!agl_list_empty(sai)) {
 908                         clli = agl_first_entry(sai);
 909                         list_del_init(&clli->lli_agl_list);
 910                         spin_unlock(&plli->lli_agl_lock);
 911                         ll_agl_trigger(&clli->lli_vfs_inode, sai);
 912                 } else {
 913                         spin_unlock(&plli->lli_agl_lock);
 914                 }
 915         }
 916
 917         spin_lock(&plli->lli_agl_lock);
 918         sai->sai_agl_valid = 0;
 919         while (!agl_list_empty(sai)) {
 920                 clli = agl_first_entry(sai);
 921                 list_del_init(&clli->lli_agl_list);
 922                 spin_unlock(&plli->lli_agl_lock);
 923                 clli->lli_agl_index = 0;
 924                 iput(&clli->lli_vfs_inode);
 925                 spin_lock(&plli->lli_agl_lock);
 926         }
 927         thread_set_flags(thread, SVC_STOPPED);
 928         spin_unlock(&plli->lli_agl_lock);
 929         wake_up(&thread->t_ctl_waitq);
 930         ll_sai_put(sai);
 931         CDEBUG(D_READA, "agl thread stopped: sai %p, parent %.*s\n",
 932                sai, parent->d_name.len, parent->d_name.name);
 933         RETURN(0);
 934 }
 935
 936 /* start agl thread */
 937 static void ll_start_agl(struct dentry *parent, struct ll_statahead_info *sai)
 938 {
 939         struct ptlrpc_thread *thread = &sai->sai_agl_thread;
 940         struct l_wait_info    lwi    = { 0 };
 941         struct ll_inode_info  *plli;
 942         struct task_struct            *task;
 943         ENTRY;
 944
 945         CDEBUG(D_READA, "start agl thread: sai %p, parent %.*s\n",
 946                sai, parent->d_name.len, parent->d_name.name);
 947
 948         plli = ll_i2info(parent->d_inode);
 949         task = kthread_run(ll_agl_thread, parent,
 950                                "ll_agl_%u", plli->lli_opendir_pid);
 951         if (IS_ERR(task)) {
 952                 CERROR("can't start ll_agl thread, rc: %ld\n", PTR_ERR(task));
 953                 thread_set_flags(thread, SVC_STOPPED);
 954                 RETURN_EXIT;
 955         }
 956
 957         l_wait_event(thread->t_ctl_waitq,
 958                      thread_is_running(thread) || thread_is_stopped(thread),
 959                      &lwi);
 960         EXIT;
 961 }
 962
 963 /* statahead thread main function */
 964 static int ll_statahead_thread(void *arg)
 965 {
 966         struct dentry *parent = (struct dentry *)arg;
 967         struct inode *dir = parent->d_inode;
 968         struct ll_inode_info *lli = ll_i2info(dir);
 969         struct ll_sb_info *sbi = ll_i2sbi(dir);
 970         struct ll_statahead_info *sai;
 971         struct ptlrpc_thread *sa_thread;
 972         struct ptlrpc_thread *agl_thread;
 973         int first = 0;
 974         struct md_op_data *op_data;
 975         struct ll_dir_chain chain;
 976         struct l_wait_info lwi = { 0 };
 977         struct page *page = NULL;
 978         __u64 pos = 0;
 979         int rc = 0;
 980         ENTRY;
 981
 982         sai = ll_sai_get(dir);
 983         sa_thread = &sai->sai_thread;
 984         agl_thread = &sai->sai_agl_thread;
 985         sa_thread->t_pid = current_pid();
 986         CDEBUG(D_READA, "statahead thread starting: sai %p, parent %.*s\n",
 987                sai, parent->d_name.len, parent->d_name.name);
 988
 989         op_data = ll_prep_md_op_data(NULL, dir, dir, NULL, 0, 0,
 990                                      LUSTRE_OPC_ANY, dir);
 991         if (IS_ERR(op_data))
 992                 GOTO(out, rc = PTR_ERR(op_data));
 993
 994         op_data->op_max_pages = ll_i2sbi(dir)->ll_md_brw_pages;
 995
 996         if (sbi->ll_flags & LL_SBI_AGL_ENABLED)
 997                 ll_start_agl(parent, sai);
 998
 999         atomic_inc(&sbi->ll_sa_total);
1000         spin_lock(&lli->lli_sa_lock);
1001         if (thread_is_init(sa_thread))
1002                 /* If someone else has changed the thread state
1003                  * (e.g. already changed to SVC_STOPPING), we can't just
1004                  * blindly overwrite that setting. */
1005                 thread_set_flags(sa_thread, SVC_RUNNING);
1006         spin_unlock(&lli->lli_sa_lock);
1007         wake_up(&sa_thread->t_ctl_waitq);
1008
1009         ll_dir_chain_init(&chain);
1010         while (pos != MDS_DIR_END_OFF && thread_is_running(sa_thread)) {
1011                 struct lu_dirpage *dp;
1012                 struct lu_dirent  *ent;
1013
1014                 sai->sai_in_readpage = 1;
1015                 page = ll_get_dir_page(dir, op_data, pos, &chain);
1016                 sai->sai_in_readpage = 0;
1017                 if (IS_ERR(page)) {
1018                         rc = PTR_ERR(page);
1019                         CDEBUG(D_READA, "error reading dir "DFID" at "LPU64
1020                                "/"LPU64" opendir_pid = %u: rc = %d\n",
1021                                PFID(ll_inode2fid(dir)), pos, sai->sai_index,
1022                                lli->lli_opendir_pid, rc);
1023                         break;
1024                 }
1025
1026                 dp = page_address(page);
1027                 for (ent = lu_dirent_start(dp);
1028                      ent != NULL && thread_is_running(sa_thread) &&
1029                      !sa_low_hit(sai);
1030                      ent = lu_dirent_next(ent)) {
1031                         __u64 hash;
1032                         int namelen;
1033                         char *name;
1034                         struct lu_fid fid;
1035
1036                         hash = le64_to_cpu(ent->lde_hash);
1037                         if (unlikely(hash < pos))
1038                                 /*
1039                                  * Skip until we find target hash value.
1040                                  */
1041                                 continue;
1042
1043                         namelen = le16_to_cpu(ent->lde_namelen);
1044                         if (unlikely(namelen == 0))
1045                                 /*
1046                                  * Skip dummy record.
1047                                  */
1048                                 continue;
1049
1050                         name = ent->lde_name;
1051                         if (name[0] == '.') {
1052                                 if (namelen == 1) {
1053                                         /*
1054                                          * skip "."
1055                                          */
1056                                         continue;
1057                                 } else if (name[1] == '.' && namelen == 2) {
1058                                         /*
1059                                          * skip ".."
1060                                          */
1061                                         continue;
1062                                 } else if (!sai->sai_ls_all) {
1063                                         /*
1064                                          * skip hidden files.
1065                                          */
1066                                         sai->sai_skip_hidden++;
1067                                         continue;
1068                                 }
1069                         }
1070
1071                         /*
1072                          * don't stat-ahead first entry.
1073                          */
1074                         if (unlikely(++first == 1))
1075                                 continue;
1076
1077                         fid_le_to_cpu(&fid, &ent->lde_fid);
1078
1079                         /* wait for spare statahead window */
1080                         do {
1081                                 l_wait_event(sa_thread->t_ctl_waitq,
1082                                              !sa_sent_full(sai) ||
1083                                              sa_has_callback(sai) ||
1084                                              !agl_list_empty(sai) ||
1085                                              !thread_is_running(sa_thread),
1086                                              &lwi);
1087
1088                                 sa_handle_callback(sai);
1089
1090                                 spin_lock(&lli->lli_agl_lock);
1091                                 while (sa_sent_full(sai) &&
1092                                        !agl_list_empty(sai)) {
1093                                         struct ll_inode_info *clli;
1094
1095                                         clli = agl_first_entry(sai);
1096                                         list_del_init(&clli->lli_agl_list);
1097                                         spin_unlock(&lli->lli_agl_lock);
1098
1099                                         ll_agl_trigger(&clli->lli_vfs_inode,
1100                                                         sai);
1101
1102                                         spin_lock(&lli->lli_agl_lock);
1103                                 }
1104                                 spin_unlock(&lli->lli_agl_lock);
1105                         } while (sa_sent_full(sai) &&
1106                                  thread_is_running(sa_thread));
1107
1108                         sa_statahead(parent, name, namelen, &fid);
1109                 }
1110
1111                 pos = le64_to_cpu(dp->ldp_hash_end);
1112                 ll_release_page(dir, page,
1113                                 le32_to_cpu(dp->ldp_flags) & LDF_COLLIDE);
1114
1115                 if (sa_low_hit(sai)) {
1116                         rc = -EFAULT;
1117                         atomic_inc(&sbi->ll_sa_wrong);
1118                         CDEBUG(D_READA, "Statahead for dir "DFID" hit "
1119                                "ratio too low: hit/miss "LPU64"/"LPU64
1120                                ", sent/replied "LPU64"/"LPU64", stopping "
1121                                "statahead thread: pid %d\n",
1122                                PFID(&lli->lli_fid), sai->sai_hit,
1123                                sai->sai_miss, sai->sai_sent,
1124                                sai->sai_replied, current_pid());
1125                         break;
1126                 }
1127         }
1128         ll_dir_chain_fini(&chain);
1129         ll_finish_md_op_data(op_data);
1130
1131         if (rc < 0) {
1132                 spin_lock(&lli->lli_sa_lock);
1133                 thread_set_flags(sa_thread, SVC_STOPPING);
1134                 lli->lli_sa_enabled = 0;
1135                 spin_unlock(&lli->lli_sa_lock);
1136         }
1137
1138         /* statahead is finished, but statahead entries need to be cached, wait
1139          * for file release to stop me. */
1140         while (thread_is_running(sa_thread)) {
1141                 l_wait_event(sa_thread->t_ctl_waitq,
1142                              sa_has_callback(sai) ||
1143                              !thread_is_running(sa_thread),
1144                              &lwi);
1145
1146                 sa_handle_callback(sai);
1147         }
1148
1149         EXIT;
1150 out:
1151         if (sai->sai_agl_valid) {
1152                 spin_lock(&lli->lli_agl_lock);
1153                 thread_set_flags(agl_thread, SVC_STOPPING);
1154                 spin_unlock(&lli->lli_agl_lock);
1155                 wake_up(&agl_thread->t_ctl_waitq);
1156
1157                 CDEBUG(D_READA, "stop agl thread: sai %p pid %u\n",
1158                        sai, (unsigned int)agl_thread->t_pid);
1159                 l_wait_event(agl_thread->t_ctl_waitq,
1160                              thread_is_stopped(agl_thread),
1161                              &lwi);
1162         } else {
1163                 /* Set agl_thread flags anyway. */
1164                 thread_set_flags(agl_thread, SVC_STOPPED);
1165         }
1166
1167         /* wait for inflight statahead RPCs to finish, and then we can free sai
1168          * safely because statahead RPC will access sai data */
1169         while (sai->sai_sent != sai->sai_replied) {
1170                 /* in case we're not woken up, timeout wait */
1171                 lwi = LWI_TIMEOUT(msecs_to_jiffies(MSEC_PER_SEC >> 3),
1172                                   NULL, NULL);
1173                 l_wait_event(sa_thread->t_ctl_waitq,
1174                         sai->sai_sent == sai->sai_replied, &lwi);
1175         }
1176
1177         /* release resources held by statahead RPCs */
1178         sa_handle_callback(sai);
1179
1180         spin_lock(&lli->lli_sa_lock);
1181         thread_set_flags(sa_thread, SVC_STOPPED);
1182         spin_unlock(&lli->lli_sa_lock);
1183
1184         CDEBUG(D_READA, "statahead thread stopped: sai %p, parent %.*s\n",
1185                sai, parent->d_name.len, parent->d_name.name);
1186
1187         wake_up(&sai->sai_waitq);
1188         wake_up(&sa_thread->t_ctl_waitq);
1189         ll_sai_put(sai);
1190
1191         return rc;
1192 }
1193
1194 /* authorize opened dir handle @key to statahead */
1195 void ll_authorize_statahead(struct inode *dir, void *key)
1196 {
1197         struct ll_inode_info *lli = ll_i2info(dir);
1198
1199         spin_lock(&lli->lli_sa_lock);
1200         if (lli->lli_opendir_key == NULL && lli->lli_sai == NULL) {
1201                 /*
1202                  * if lli_sai is not NULL, it means previous statahead is not
1203                  * finished yet, we'd better not start a new statahead for now.
1204                  */
1205                 LASSERT(lli->lli_opendir_pid == 0);
1206                 lli->lli_opendir_key = key;
1207                 lli->lli_opendir_pid = current_pid();
1208                 lli->lli_sa_enabled = 1;
1209         }
1210         spin_unlock(&lli->lli_sa_lock);
1211 }
1212
1213 /*
1214  * deauthorize opened dir handle @key to statahead, and notify statahead thread
1215  * to quit if it's running.
1216  */
1217 void ll_deauthorize_statahead(struct inode *dir, void *key)
1218 {
1219         struct ll_inode_info *lli = ll_i2info(dir);
1220         struct ll_statahead_info *sai;
1221
1222         LASSERT(lli->lli_opendir_key == key);
1223         LASSERT(lli->lli_opendir_pid != 0);
1224
1225         CDEBUG(D_READA, "deauthorize statahead for "DFID"\n",
1226                 PFID(&lli->lli_fid));
1227
1228         spin_lock(&lli->lli_sa_lock);
1229         lli->lli_opendir_key = NULL;
1230         lli->lli_opendir_pid = 0;
1231         lli->lli_sa_enabled = 0;
1232         sai = lli->lli_sai;
1233         if (sai != NULL && thread_is_running(&sai->sai_thread)) {
1234                 /*
1235                  * statahead thread may not quit yet because it needs to cache
1236                  * entries, now it's time to tell it to quit.
1237                  */
1238                 thread_set_flags(&sai->sai_thread, SVC_STOPPING);
1239                 wake_up(&sai->sai_thread.t_ctl_waitq);
1240         }
1241         spin_unlock(&lli->lli_sa_lock);
1242 }
1243
1244 enum {
1245         /**
1246          * not first dirent, or is "."
1247          */
1248         LS_NOT_FIRST_DE = 0,
1249         /**
1250          * the first non-hidden dirent
1251          */
1252         LS_FIRST_DE,
1253         /**
1254          * the first hidden dirent, that is "."
1255          */
1256         LS_FIRST_DOT_DE
1257 };
1258
1259 /* file is first dirent under @dir */
1260 static int is_first_dirent(struct inode *dir, struct dentry *dentry)
1261 {
1262         struct ll_dir_chain   chain;
1263         struct qstr          *target = &dentry->d_name;
1264         struct md_op_data    *op_data;
1265         int                   dot_de;
1266         struct page          *page = NULL;
1267         int                   rc = LS_NOT_FIRST_DE;
1268         __u64                 pos = 0;
1269         ENTRY;
1270
1271         op_data = ll_prep_md_op_data(NULL, dir, dir, NULL, 0, 0,
1272                                      LUSTRE_OPC_ANY, dir);
1273         if (IS_ERR(op_data))
1274                 RETURN(PTR_ERR(op_data));
1275         /**
1276          *FIXME choose the start offset of the readdir
1277          */
1278         op_data->op_max_pages = ll_i2sbi(dir)->ll_md_brw_pages;
1279
1280         ll_dir_chain_init(&chain);
1281         page = ll_get_dir_page(dir, op_data, 0, &chain);
1282
1283         while (1) {
1284                 struct lu_dirpage *dp;
1285                 struct lu_dirent  *ent;
1286
1287                 if (IS_ERR(page)) {
1288                         struct ll_inode_info *lli = ll_i2info(dir);
1289
1290                         rc = PTR_ERR(page);
1291                         CERROR("%s: reading dir "DFID" at "LPU64
1292                                "opendir_pid = %u : rc = %d\n",
1293                                ll_get_fsname(dir->i_sb, NULL, 0),
1294                                PFID(ll_inode2fid(dir)), pos,
1295                                lli->lli_opendir_pid, rc);
1296                         break;
1297                 }
1298
1299                 dp = page_address(page);
1300                 for (ent = lu_dirent_start(dp); ent != NULL;
1301                      ent = lu_dirent_next(ent)) {
1302                         __u64 hash;
1303                         int namelen;
1304                         char *name;
1305
1306                         hash = le64_to_cpu(ent->lde_hash);
1307                         /* The ll_get_dir_page() can return any page containing
1308                          * the given hash which may be not the start hash. */
1309                         if (unlikely(hash < pos))
1310                                 continue;
1311
1312                         namelen = le16_to_cpu(ent->lde_namelen);
1313                         if (unlikely(namelen == 0))
1314                                 /*
1315                                  * skip dummy record.
1316                                  */
1317                                 continue;
1318
1319                         name = ent->lde_name;
1320                         if (name[0] == '.') {
1321                                 if (namelen == 1)
1322                                         /*
1323                                          * skip "."
1324                                          */
1325                                         continue;
1326                                 else if (name[1] == '.' && namelen == 2)
1327                                         /*
1328                                          * skip ".."
1329                                          */
1330                                         continue;
1331                                 else
1332                                         dot_de = 1;
1333                         } else {
1334                                 dot_de = 0;
1335                         }
1336
1337                         if (dot_de && target->name[0] != '.') {
1338                                 CDEBUG(D_READA, "%.*s skip hidden file %.*s\n",
1339                                        target->len, target->name,
1340                                        namelen, name);
1341                                 continue;
1342                         }
1343
1344                         if (target->len != namelen ||
1345                             memcmp(target->name, name, namelen) != 0)
1346                                 rc = LS_NOT_FIRST_DE;
1347                         else if (!dot_de)
1348                                 rc = LS_FIRST_DE;
1349                         else
1350                                 rc = LS_FIRST_DOT_DE;
1351
1352                         ll_release_page(dir, page, false);
1353                         GOTO(out, rc);
1354                 }
1355                 pos = le64_to_cpu(dp->ldp_hash_end);
1356                 if (pos == MDS_DIR_END_OFF) {
1357                         /*
1358                          * End of directory reached.
1359                          */
1360                         ll_release_page(dir, page, false);
1361                         GOTO(out, rc);
1362                 } else {
1363                         /*
1364                          * chain is exhausted
1365                          * Normal case: continue to the next page.
1366                          */
1367                         ll_release_page(dir, page, le32_to_cpu(dp->ldp_flags) &
1368                                               LDF_COLLIDE);
1369                         page = ll_get_dir_page(dir, op_data, pos, &chain);
1370                 }
1371         }
1372         EXIT;
1373 out:
1374         ll_dir_chain_fini(&chain);
1375         ll_finish_md_op_data(op_data);
1376         return rc;
1377 }
1378
1379 /**
1380  * revalidate @dentryp from statahead cache
1381  *
1382  * \param[in] dir       parent directory
1383  * \param[in] sai       sai structure
1384  * \param[out] dentryp  pointer to dentry which will be revalidated
1385  * \param[in] unplug    unplug statahead window only (normally for negative
1386  *                      dentry)
1387  * \retval              1 on success, dentry is saved in @dentryp
1388  * \retval              0 if revalidation failed (no proper lock on client)
1389  * \retval              negative number upon error
1390  */
1391 static int revalidate_statahead_dentry(struct inode *dir,
1392                                         struct ll_statahead_info *sai,
1393                                         struct dentry **dentryp,
1394                                         bool unplug)
1395 {
1396         struct sa_entry *entry = NULL;
1397         struct l_wait_info lwi = { 0 };
1398         struct ll_dentry_data *ldd;
1399         struct ll_inode_info *lli;
1400         int rc = 0;
1401         ENTRY;
1402
1403         if ((*dentryp)->d_name.name[0] == '.') {
1404                 if (sai->sai_ls_all ||
1405                     sai->sai_miss_hidden >= sai->sai_skip_hidden) {
1406                         /*
1407                          * Hidden dentry is the first one, or statahead
1408                          * thread does not skip so many hidden dentries
1409                          * before "sai_ls_all" enabled as below.
1410                          */
1411                 } else {
1412                         if (!sai->sai_ls_all)
1413                                 /*
1414                                  * It maybe because hidden dentry is not
1415                                  * the first one, "sai_ls_all" was not
1416                                  * set, then "ls -al" missed. Enable
1417                                  * "sai_ls_all" for such case.
1418                                  */
1419                                 sai->sai_ls_all = 1;
1420
1421                         /*
1422                          * Such "getattr" has been skipped before
1423                          * "sai_ls_all" enabled as above.
1424                          */
1425                         sai->sai_miss_hidden++;
1426                         RETURN(-EAGAIN);
1427                 }
1428         }
1429
1430         if (unplug)
1431                 GOTO(out, rc = 1);
1432
1433         entry = sa_get(sai, &(*dentryp)->d_name);
1434         if (entry == NULL)
1435                 GOTO(out, rc = -EAGAIN);
1436
1437         /* if statahead is busy in readdir, help it do post-work */
1438         if (!sa_ready(entry) && sai->sai_in_readpage)
1439                 sa_handle_callback(sai);
1440
1441         if (!sa_ready(entry)) {
1442                 sai->sai_index_wait = entry->se_index;
1443                 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(30), NULL,
1444                                         LWI_ON_SIGNAL_NOOP, NULL);
1445                 rc = l_wait_event(sai->sai_waitq, sa_ready(entry), &lwi);
1446                 if (rc < 0) {
1447                         /*
1448                          * entry may not be ready, so it may be used by inflight
1449                          * statahead RPC, don't free it.
1450                          */
1451                         entry = NULL;
1452                         GOTO(out, rc = -EAGAIN);
1453                 }
1454         }
1455
1456         if (entry->se_state == SA_ENTRY_SUCC && entry->se_inode != NULL) {
1457                 struct inode *inode = entry->se_inode;
1458                 struct lookup_intent it = { .it_op = IT_GETATTR,
1459                                             .d.lustre.it_lock_handle =
1460                                                 entry->se_handle };
1461                 __u64 bits;
1462
1463                 rc = md_revalidate_lock(ll_i2mdexp(dir), &it,
1464                                         ll_inode2fid(inode), &bits);
1465                 if (rc == 1) {
1466                         if ((*dentryp)->d_inode == NULL) {
1467                                 struct dentry *alias;
1468
1469                                 alias = ll_splice_alias(inode, *dentryp);
1470                                 if (IS_ERR(alias)) {
1471                                         ll_intent_release(&it);
1472                                         GOTO(out, rc = PTR_ERR(alias));
1473                                 }
1474                                 *dentryp = alias;
1475                                 /* statahead prepared this inode, transfer inode
1476                                  * refcount from sa_entry to dentry */
1477                                 entry->se_inode = NULL;
1478                         } else if ((*dentryp)->d_inode != inode) {
1479                                 /* revalidate, but inode is recreated */
1480                                 CDEBUG(D_READA,
1481                                         "%s: stale dentry %.*s inode "
1482                                         DFID", statahead inode "DFID
1483                                         "\n",
1484                                         ll_get_fsname((*dentryp)->d_inode->i_sb,
1485                                                       NULL, 0),
1486                                         (*dentryp)->d_name.len,
1487                                         (*dentryp)->d_name.name,
1488                                         PFID(ll_inode2fid((*dentryp)->d_inode)),
1489                                         PFID(ll_inode2fid(inode)));
1490                                 ll_intent_release(&it);
1491                                 GOTO(out, rc = -ESTALE);
1492                         }
1493
1494                         if ((bits & MDS_INODELOCK_LOOKUP) &&
1495                             d_lustre_invalid(*dentryp))
1496                                 d_lustre_revalidate(*dentryp);
1497                         ll_intent_release(&it);
1498                 }
1499         }
1500 out:
1501         /*
1502          * statahead cached sa_entry can be used only once, and will be killed
1503          * right after use, so if lookup/revalidate accessed statahead cache,
1504          * set dentry ldd_sa_generation to parent lli_sa_generation, later if we
1505          * stat this file again, we know we've done statahead before, see
1506          * dentry_may_statahead().
1507          */
1508         ldd = ll_d2d(*dentryp);
1509         lli = ll_i2info(dir);
1510         /* ldd can be NULL if llite lookup failed. */
1511         if (ldd != NULL)
1512                 ldd->lld_sa_generation = lli->lli_sa_generation;
1513         sa_put(sai, entry);
1514
1515         RETURN(rc);
1516 }
1517
1518 /**
1519  * start statahead thread
1520  *
1521  * \param[in] dir       parent directory
1522  * \param[in] dentry    dentry that triggers statahead, normally the first
1523  *                      dirent under @dir
1524  * \retval              -EAGAIN on success, because when this function is
1525  *                      called, it's already in lookup call, so client should
1526  *                      do it itself instead of waiting for statahead thread
1527  *                      to do it asynchronously.
1528  * \retval              negative number upon error
1529  */
1530 static int start_statahead_thread(struct inode *dir, struct dentry *dentry)
1531 {
1532         struct ll_inode_info *lli = ll_i2info(dir);
1533         struct ll_statahead_info *sai = NULL;
1534         struct dentry *parent = dentry->d_parent;
1535         struct ptlrpc_thread *thread;
1536         struct l_wait_info lwi = { 0 };
1537         struct task_struct *task;
1538         int rc;
1539         ENTRY;
1540
1541         /* I am the "lli_opendir_pid" owner, only me can set "lli_sai". */
1542         rc = is_first_dirent(dir, dentry);
1543         if (rc == LS_NOT_FIRST_DE)
1544                 /* It is not "ls -{a}l" operation, no need statahead for it. */
1545                 GOTO(out, rc = -EFAULT);
1546
1547         sai = ll_sai_alloc(parent);
1548         if (sai == NULL)
1549                 GOTO(out, rc = -ENOMEM);
1550
1551         sai->sai_ls_all = (rc == LS_FIRST_DOT_DE);
1552
1553         /* if current lli_opendir_key was deauthorized, or dir re-opened by
1554          * another process, don't start statahead, otherwise the newly spawned
1555          * statahead thread won't be notified to quit. */
1556         spin_lock(&lli->lli_sa_lock);
1557         if (unlikely(lli->lli_sai != NULL ||
1558                      lli->lli_opendir_key == NULL ||
1559                      lli->lli_opendir_pid != current->pid)) {
1560                 spin_unlock(&lli->lli_sa_lock);
1561                 GOTO(out, rc = -EPERM);
1562         }
1563         lli->lli_sai = sai;
1564         spin_unlock(&lli->lli_sa_lock);
1565
1566         atomic_inc(&ll_i2sbi(parent->d_inode)->ll_sa_running);
1567
1568         CDEBUG(D_READA, "start statahead thread: [pid %d] [parent %.*s]\n",
1569                current_pid(), parent->d_name.len, parent->d_name.name);
1570
1571         task = kthread_run(ll_statahead_thread, parent, "ll_sa_%u",
1572                            lli->lli_opendir_pid);
1573         thread = &sai->sai_thread;
1574         if (IS_ERR(task)) {
1575                 rc = PTR_ERR(task);
1576                 CERROR("can't start ll_sa thread, rc: %d\n", rc);
1577                 GOTO(out, rc);
1578         }
1579
1580         l_wait_event(thread->t_ctl_waitq,
1581                      thread_is_running(thread) || thread_is_stopped(thread),
1582                      &lwi);
1583         ll_sai_put(sai);
1584
1585         /*
1586          * We don't stat-ahead for the first dirent since we are already in
1587          * lookup.
1588          */
1589         RETURN(-EAGAIN);
1590
1591 out:
1592         /* once we start statahead thread failed, disable statahead so that
1593          * subsequent stat won't waste time to try it. */
1594         spin_lock(&lli->lli_sa_lock);
1595         lli->lli_sa_enabled = 0;
1596         lli->lli_sai = NULL;
1597         spin_unlock(&lli->lli_sa_lock);
1598
1599         if (sai != NULL)
1600                 ll_sai_free(sai);
1601
1602         RETURN(rc);
1603 }
1604
1605 /**
1606  * statahead entry function, this is called when client getattr on a file, it
1607  * will start statahead thread if this is the first dir entry, else revalidate
1608  * dentry from statahead cache.
1609  *
1610  * \param[in]  dir      parent directory
1611  * \param[out] dentryp  dentry to getattr
1612  * \param[in]  unplug   unplug statahead window only (normally for negative
1613  *                      dentry)
1614  * \retval              1 on success
1615  * \retval              0 revalidation from statahead cache failed, caller needs
1616  *                      to getattr from server directly
1617  * \retval              negative number on error, caller often ignores this and
1618  *                      then getattr from server
1619  */
1620 int ll_statahead(struct inode *dir, struct dentry **dentryp, bool unplug)
1621 {
1622         struct ll_statahead_info *sai;
1623
1624         sai = ll_sai_get(dir);
1625         if (sai != NULL) {
1626                 int rc;
1627
1628                 rc = revalidate_statahead_dentry(dir, sai, dentryp, unplug);
1629                 CDEBUG(D_READA, "revalidate statahead %.*s: %d.\n",
1630                         (*dentryp)->d_name.len, (*dentryp)->d_name.name, rc);
1631                 ll_sai_put(sai);
1632                 return rc;
1633         }
1634         return start_statahead_thread(dir, *dentryp);
1635 }