lustre/llite/statahead.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2011, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  */
  32
  33 #include <linux/fs.h>
  34 #include <linux/sched.h>
  35 #include <linux/kthread.h>
  36 #include <linux/mm.h>
  37 #include <linux/highmem.h>
  38 #include <linux/pagemap.h>
  39 #include <linux/delay.h>
  40
  41 #define DEBUG_SUBSYSTEM S_LLITE
  42
  43 #include <obd_support.h>
  44 #include <lustre_dlm.h>
  45 #include "llite_internal.h"
  46
  47 #define SA_OMITTED_ENTRY_MAX 8ULL
  48
  49 typedef enum {
  50         /** negative values are for error cases */
  51         SA_ENTRY_INIT = 0,      /** init entry */
  52         SA_ENTRY_SUCC = 1,      /** stat succeed */
  53         SA_ENTRY_INVA = 2,      /** invalid entry */
  54 } se_state_t;
  55
  56 /*
  57  * sa_entry is not refcounted: statahead thread allocates it and do async stat,
  58  * and in async stat callback ll_statahead_interpret() will add it into
  59  * sai_interim_entries, later statahead thread will call sa_handle_callback() to
  60  * instantiate entry and move it into sai_entries, and then only scanner process
  61  * can access and free it.
  62  */
  63 struct sa_entry {
  64         /* link into sai_interim_entries or sai_entries */
  65         struct list_head        se_list;
  66         /* link into sai hash table locally */
  67         struct list_head        se_hash;
  68         /* entry index in the sai */
  69         __u64                   se_index;
  70         /* low layer ldlm lock handle */
  71         __u64                   se_handle;
  72         /* entry status */
  73         se_state_t              se_state;
  74         /* entry size, contains name */
  75         int                     se_size;
  76         /* pointer to async getattr enqueue info */
  77         struct md_enqueue_info *se_minfo;
  78         /* pointer to the async getattr request */
  79         struct ptlrpc_request  *se_req;
  80         /* pointer to the target inode */
  81         struct inode           *se_inode;
  82         /* entry name */
  83         struct qstr             se_qstr;
  84         /* entry fid */
  85         struct lu_fid           se_fid;
  86 };
  87
  88 static unsigned int sai_generation;
  89 static DEFINE_SPINLOCK(sai_generation_lock);
  90
  91 static inline int sa_unhashed(struct sa_entry *entry)
  92 {
  93         return list_empty(&entry->se_hash);
  94 }
  95
  96 /* sa_entry is ready to use */
  97 static inline int sa_ready(struct sa_entry *entry)
  98 {
  99         /* Make sure sa_entry is updated and ready to use */
 100         smp_rmb();
 101         return (entry->se_state != SA_ENTRY_INIT);
 102 }
 103
 104 /* hash value to put in sai_cache */
 105 static inline int sa_hash(int val)
 106 {
 107         return val & LL_SA_CACHE_MASK;
 108 }
 109
 110 /* hash entry into sai_cache */
 111 static inline void
 112 sa_rehash(struct ll_statahead_info *sai, struct sa_entry *entry)
 113 {
 114         int i = sa_hash(entry->se_qstr.hash);
 115
 116         spin_lock(&sai->sai_cache_lock[i]);
 117         list_add_tail(&entry->se_hash, &sai->sai_cache[i]);
 118         spin_unlock(&sai->sai_cache_lock[i]);
 119 }
 120
 121 /* unhash entry from sai_cache */
 122 static inline void
 123 sa_unhash(struct ll_statahead_info *sai, struct sa_entry *entry)
 124 {
 125         int i = sa_hash(entry->se_qstr.hash);
 126
 127         spin_lock(&sai->sai_cache_lock[i]);
 128         list_del_init(&entry->se_hash);
 129         spin_unlock(&sai->sai_cache_lock[i]);
 130 }
 131
 132 static inline int agl_should_run(struct ll_statahead_info *sai,
 133                                  struct inode *inode)
 134 {
 135         return inode && S_ISREG(inode->i_mode) && sai->sai_agl_task;
 136 }
 137
 138 static inline struct ll_inode_info *
 139 agl_first_entry(struct ll_statahead_info *sai)
 140 {
 141         return list_entry(sai->sai_agls.next, struct ll_inode_info,
 142                           lli_agl_list);
 143 }
 144
 145 /* statahead window is full */
 146 static inline int sa_sent_full(struct ll_statahead_info *sai)
 147 {
 148         return atomic_read(&sai->sai_cache_count) >= sai->sai_max;
 149 }
 150
 151 /* got async stat replies */
 152 static inline int sa_has_callback(struct ll_statahead_info *sai)
 153 {
 154         return !list_empty(&sai->sai_interim_entries);
 155 }
 156
 157 static inline int agl_list_empty(struct ll_statahead_info *sai)
 158 {
 159         return list_empty(&sai->sai_agls);
 160 }
 161
 162 /**
 163  * (1) hit ratio less than 80%
 164  * or
 165  * (2) consecutive miss more than 8
 166  * then means low hit.
 167  */
 168 static inline int sa_low_hit(struct ll_statahead_info *sai)
 169 {
 170         return ((sai->sai_hit > 7 && sai->sai_hit < 4 * sai->sai_miss) ||
 171                 (sai->sai_consecutive_miss > 8));
 172 }
 173
 174 /*
 175  * if the given index is behind of statahead window more than
 176  * SA_OMITTED_ENTRY_MAX, then it is old.
 177  */
 178 static inline int is_omitted_entry(struct ll_statahead_info *sai, __u64 index)
 179 {
 180         return ((__u64)sai->sai_max + index + SA_OMITTED_ENTRY_MAX <
 181                 sai->sai_index);
 182 }
 183
 184 /* allocate sa_entry and hash it to allow scanner process to find it */
 185 static struct sa_entry *
 186 sa_alloc(struct dentry *parent, struct ll_statahead_info *sai, __u64 index,
 187          const char *name, int len, const struct lu_fid *fid)
 188 {
 189         struct ll_inode_info *lli;
 190         struct sa_entry *entry;
 191         int entry_size;
 192         char *dname;
 193
 194         ENTRY;
 195
 196         entry_size = sizeof(struct sa_entry) + (len & ~3) + 4;
 197         OBD_ALLOC(entry, entry_size);
 198         if (unlikely(!entry))
 199                 RETURN(ERR_PTR(-ENOMEM));
 200
 201         CDEBUG(D_READA, "alloc sa entry %.*s(%p) index %llu\n",
 202                len, name, entry, index);
 203
 204         entry->se_index = index;
 205
 206         entry->se_state = SA_ENTRY_INIT;
 207         entry->se_size = entry_size;
 208         dname = (char *)entry + sizeof(struct sa_entry);
 209         memcpy(dname, name, len);
 210         dname[len] = 0;
 211         entry->se_qstr.hash = ll_full_name_hash(parent, name, len);
 212         entry->se_qstr.len = len;
 213         entry->se_qstr.name = dname;
 214         entry->se_fid = *fid;
 215
 216         lli = ll_i2info(sai->sai_dentry->d_inode);
 217
 218         spin_lock(&lli->lli_sa_lock);
 219         INIT_LIST_HEAD(&entry->se_list);
 220         sa_rehash(sai, entry);
 221         spin_unlock(&lli->lli_sa_lock);
 222
 223         atomic_inc(&sai->sai_cache_count);
 224
 225         RETURN(entry);
 226 }
 227
 228 /* free sa_entry, which should have been unhashed and not in any list */
 229 static void sa_free(struct ll_statahead_info *sai, struct sa_entry *entry)
 230 {
 231         CDEBUG(D_READA, "free sa entry %.*s(%p) index %llu\n",
 232                entry->se_qstr.len, entry->se_qstr.name, entry,
 233                entry->se_index);
 234
 235         LASSERT(list_empty(&entry->se_list));
 236         LASSERT(sa_unhashed(entry));
 237
 238         OBD_FREE(entry, entry->se_size);
 239         atomic_dec(&sai->sai_cache_count);
 240 }
 241
 242 /*
 243  * find sa_entry by name, used by directory scanner, lock is not needed because
 244  * only scanner can remove the entry from cache.
 245  */
 246 static struct sa_entry *
 247 sa_get(struct ll_statahead_info *sai, const struct qstr *qstr)
 248 {
 249         struct sa_entry *entry;
 250         int i = sa_hash(qstr->hash);
 251
 252         list_for_each_entry(entry, &sai->sai_cache[i], se_hash) {
 253                 if (entry->se_qstr.hash == qstr->hash &&
 254                     entry->se_qstr.len == qstr->len &&
 255                     memcmp(entry->se_qstr.name, qstr->name, qstr->len) == 0)
 256                         return entry;
 257         }
 258         return NULL;
 259 }
 260
 261 /* unhash and unlink sa_entry, and then free it */
 262 static inline void
 263 sa_kill(struct ll_statahead_info *sai, struct sa_entry *entry)
 264 {
 265         struct ll_inode_info *lli = ll_i2info(sai->sai_dentry->d_inode);
 266
 267         LASSERT(!sa_unhashed(entry));
 268         LASSERT(!list_empty(&entry->se_list));
 269         LASSERT(sa_ready(entry));
 270
 271         sa_unhash(sai, entry);
 272
 273         spin_lock(&lli->lli_sa_lock);
 274         list_del_init(&entry->se_list);
 275         spin_unlock(&lli->lli_sa_lock);
 276
 277         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 && 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)
 300                 sa_kill(sai, entry);
 301
 302         /*
 303          * kill old completed entries, only scanner process does this, no need
 304          * to lock
 305          */
 306         list_for_each_entry_safe(tmp, next, &sai->sai_entries, se_list) {
 307                 if (!is_omitted_entry(sai, tmp->se_index))
 308                         break;
 309                 sa_kill(sai, tmp);
 310         }
 311 }
 312
 313 /*
 314  * update state and sort add entry to sai_entries by index, return true if
 315  * scanner is waiting on this entry.
 316  */
 317 static bool
 318 __sa_make_ready(struct ll_statahead_info *sai, struct sa_entry *entry, int ret)
 319 {
 320         struct sa_entry *se;
 321         struct list_head *pos = &sai->sai_entries;
 322         __u64 index = entry->se_index;
 323
 324         LASSERT(!sa_ready(entry));
 325         LASSERT(list_empty(&entry->se_list));
 326
 327         list_for_each_entry_reverse(se, &sai->sai_entries, se_list) {
 328                 if (se->se_index < entry->se_index) {
 329                         pos = &se->se_list;
 330                         break;
 331                 }
 332         }
 333         list_add(&entry->se_list, pos);
 334         /*
 335          * LU-9210: ll_statahead_interpet must be able to see this before
 336          * we wake it up
 337          */
 338         smp_store_release(&entry->se_state,
 339                           ret < 0 ? SA_ENTRY_INVA : SA_ENTRY_SUCC);
 340
 341         return (index == sai->sai_index_wait);
 342 }
 343
 344 /* finish async stat RPC arguments */
 345 static void sa_fini_data(struct md_enqueue_info *minfo)
 346 {
 347         ll_unlock_md_op_lsm(&minfo->mi_data);
 348         iput(minfo->mi_dir);
 349         OBD_FREE_PTR(minfo);
 350 }
 351
 352 static int ll_statahead_interpret(struct ptlrpc_request *req,
 353                                   struct md_enqueue_info *minfo, int rc);
 354
 355 /*
 356  * prepare arguments for async stat RPC.
 357  */
 358 static struct md_enqueue_info *
 359 sa_prep_data(struct inode *dir, struct inode *child, struct sa_entry *entry)
 360 {
 361         struct md_enqueue_info   *minfo;
 362         struct ldlm_enqueue_info *einfo;
 363         struct md_op_data        *op_data;
 364
 365         OBD_ALLOC_PTR(minfo);
 366         if (!minfo)
 367                 return ERR_PTR(-ENOMEM);
 368
 369         op_data = ll_prep_md_op_data(&minfo->mi_data, dir, child,
 370                                      entry->se_qstr.name, entry->se_qstr.len, 0,
 371                                      LUSTRE_OPC_ANY, NULL);
 372         if (IS_ERR(op_data)) {
 373                 OBD_FREE_PTR(minfo);
 374                 return (struct md_enqueue_info *)op_data;
 375         }
 376
 377         if (!child)
 378                 op_data->op_fid2 = entry->se_fid;
 379
 380         minfo->mi_it.it_op = IT_GETATTR;
 381         minfo->mi_dir = igrab(dir);
 382         minfo->mi_cb = ll_statahead_interpret;
 383         minfo->mi_cbdata = entry;
 384
 385         einfo = &minfo->mi_einfo;
 386         einfo->ei_type   = LDLM_IBITS;
 387         einfo->ei_mode   = it_to_lock_mode(&minfo->mi_it);
 388         einfo->ei_cb_bl  = ll_md_blocking_ast;
 389         einfo->ei_cb_cp  = ldlm_completion_ast;
 390         einfo->ei_cb_gl  = NULL;
 391         einfo->ei_cbdata = NULL;
 392
 393         return minfo;
 394 }
 395
 396 /*
 397  * release resources used in async stat RPC, update entry state and wakeup if
 398  * scanner process it waiting on this entry.
 399  */
 400 static void
 401 sa_make_ready(struct ll_statahead_info *sai, struct sa_entry *entry, int ret)
 402 {
 403         struct ll_inode_info *lli = ll_i2info(sai->sai_dentry->d_inode);
 404         struct md_enqueue_info *minfo = entry->se_minfo;
 405         struct ptlrpc_request *req = entry->se_req;
 406         bool wakeup;
 407
 408         /* release resources used in RPC */
 409         if (minfo) {
 410                 entry->se_minfo = NULL;
 411                 ll_intent_release(&minfo->mi_it);
 412                 sa_fini_data(minfo);
 413         }
 414
 415         if (req) {
 416                 entry->se_req = NULL;
 417                 ptlrpc_req_finished(req);
 418         }
 419
 420         spin_lock(&lli->lli_sa_lock);
 421         wakeup = __sa_make_ready(sai, entry, ret);
 422         spin_unlock(&lli->lli_sa_lock);
 423
 424         if (wakeup)
 425                 wake_up(&sai->sai_waitq);
 426 }
 427
 428 /* insert inode into the list of sai_agls */
 429 static void ll_agl_add(struct ll_statahead_info *sai,
 430                        struct inode *inode, int index)
 431 {
 432         struct ll_inode_info *child  = ll_i2info(inode);
 433         struct ll_inode_info *parent = ll_i2info(sai->sai_dentry->d_inode);
 434
 435         spin_lock(&child->lli_agl_lock);
 436         if (child->lli_agl_index == 0) {
 437                 child->lli_agl_index = index;
 438                 spin_unlock(&child->lli_agl_lock);
 439
 440                 LASSERT(list_empty(&child->lli_agl_list));
 441
 442                 spin_lock(&parent->lli_agl_lock);
 443                 /* Re-check under the lock */
 444                 if (agl_should_run(sai, inode)) {
 445                         if (agl_list_empty(sai))
 446                                 wake_up_process(sai->sai_agl_task);
 447                         igrab(inode);
 448                         list_add_tail(&child->lli_agl_list, &sai->sai_agls);
 449                 } else
 450                         child->lli_agl_index = 0;
 451                 spin_unlock(&parent->lli_agl_lock);
 452         } else {
 453                 spin_unlock(&child->lli_agl_lock);
 454         }
 455 }
 456
 457 /* allocate sai */
 458 static struct ll_statahead_info *ll_sai_alloc(struct dentry *dentry)
 459 {
 460         struct ll_statahead_info *sai;
 461         struct ll_inode_info *lli = ll_i2info(dentry->d_inode);
 462         int i;
 463
 464         ENTRY;
 465
 466         OBD_ALLOC_PTR(sai);
 467         if (!sai)
 468                 RETURN(NULL);
 469
 470         sai->sai_dentry = dget(dentry);
 471         atomic_set(&sai->sai_refcount, 1);
 472         sai->sai_max = LL_SA_RPC_MIN;
 473         sai->sai_index = 1;
 474         init_waitqueue_head(&sai->sai_waitq);
 475
 476         INIT_LIST_HEAD(&sai->sai_interim_entries);
 477         INIT_LIST_HEAD(&sai->sai_entries);
 478         INIT_LIST_HEAD(&sai->sai_agls);
 479
 480         for (i = 0; i < LL_SA_CACHE_SIZE; i++) {
 481                 INIT_LIST_HEAD(&sai->sai_cache[i]);
 482                 spin_lock_init(&sai->sai_cache_lock[i]);
 483         }
 484         atomic_set(&sai->sai_cache_count, 0);
 485
 486         spin_lock(&sai_generation_lock);
 487         lli->lli_sa_generation = ++sai_generation;
 488         if (unlikely(sai_generation == 0))
 489                 lli->lli_sa_generation = ++sai_generation;
 490         spin_unlock(&sai_generation_lock);
 491
 492         RETURN(sai);
 493 }
 494
 495 /* free sai */
 496 static inline void ll_sai_free(struct ll_statahead_info *sai)
 497 {
 498         LASSERT(sai->sai_dentry != NULL);
 499         dput(sai->sai_dentry);
 500         OBD_FREE_PTR(sai);
 501 }
 502
 503 /*
 504  * take refcount of sai if sai for @dir exists, which means statahead is on for
 505  * this directory.
 506  */
 507 static inline struct ll_statahead_info *ll_sai_get(struct inode *dir)
 508 {
 509         struct ll_inode_info *lli = ll_i2info(dir);
 510         struct ll_statahead_info *sai = NULL;
 511
 512         spin_lock(&lli->lli_sa_lock);
 513         sai = lli->lli_sai;
 514         if (sai)
 515                 atomic_inc(&sai->sai_refcount);
 516         spin_unlock(&lli->lli_sa_lock);
 517
 518         return sai;
 519 }
 520
 521 /*
 522  * put sai refcount after use, if refcount reaches zero, free sai and sa_entries
 523  * attached to it.
 524  */
 525 static void ll_sai_put(struct ll_statahead_info *sai)
 526 {
 527         struct ll_inode_info *lli = ll_i2info(sai->sai_dentry->d_inode);
 528
 529         if (atomic_dec_and_lock(&sai->sai_refcount, &lli->lli_sa_lock)) {
 530                 struct sa_entry *entry, *next;
 531                 struct ll_sb_info *sbi = ll_i2sbi(sai->sai_dentry->d_inode);
 532
 533                 lli->lli_sai = NULL;
 534                 spin_unlock(&lli->lli_sa_lock);
 535
 536                 LASSERT(!sai->sai_task);
 537                 LASSERT(!sai->sai_agl_task);
 538                 LASSERT(sai->sai_sent == sai->sai_replied);
 539                 LASSERT(!sa_has_callback(sai));
 540
 541                 list_for_each_entry_safe(entry, next, &sai->sai_entries,
 542                                          se_list)
 543                         sa_kill(sai, entry);
 544
 545                 LASSERT(atomic_read(&sai->sai_cache_count) == 0);
 546                 LASSERT(agl_list_empty(sai));
 547
 548                 ll_sai_free(sai);
 549                 atomic_dec(&sbi->ll_sa_running);
 550         }
 551 }
 552
 553 /* Do NOT forget to drop inode refcount when into sai_agls. */
 554 static void ll_agl_trigger(struct inode *inode, struct ll_statahead_info *sai)
 555 {
 556         struct ll_inode_info *lli = ll_i2info(inode);
 557         u64 index = lli->lli_agl_index;
 558         ktime_t expire;
 559         int rc;
 560
 561         ENTRY;
 562
 563         LASSERT(list_empty(&lli->lli_agl_list));
 564
 565         /* AGL maybe fall behind statahead with one entry */
 566         if (is_omitted_entry(sai, index + 1)) {
 567                 lli->lli_agl_index = 0;
 568                 iput(inode);
 569                 RETURN_EXIT;
 570         }
 571
 572         /*
 573          * In case of restore, the MDT has the right size and has already
 574          * sent it back without granting the layout lock, inode is up-to-date.
 575          * Then AGL (async glimpse lock) is useless.
 576          * Also to glimpse we need the layout, in case of a runninh restore
 577          * the MDT holds the layout lock so the glimpse will block up to the
 578          * end of restore (statahead/agl will block)
 579          */
 580         if (test_bit(LLIF_FILE_RESTORING, &lli->lli_flags)) {
 581                 lli->lli_agl_index = 0;
 582                 iput(inode);
 583                 RETURN_EXIT;
 584         }
 585
 586         /* Someone is in glimpse (sync or async), do nothing. */
 587         rc = down_write_trylock(&lli->lli_glimpse_sem);
 588         if (rc == 0) {
 589                 lli->lli_agl_index = 0;
 590                 iput(inode);
 591                 RETURN_EXIT;
 592         }
 593
 594         /*
 595          * Someone triggered glimpse within 1 sec before.
 596          * 1) The former glimpse succeeded with glimpse lock granted by OST, and
 597          *    if the lock is still cached on client, AGL needs to do nothing. If
 598          *    it is cancelled by other client, AGL maybe cannot obtaion new lock
 599          *    for no glimpse callback triggered by AGL.
 600          * 2) The former glimpse succeeded, but OST did not grant glimpse lock.
 601          *    Under such case, it is quite possible that the OST will not grant
 602          *    glimpse lock for AGL also.
 603          * 3) The former glimpse failed, compared with other two cases, it is
 604          *    relative rare. AGL can ignore such case, and it will not muchly
 605          *    affect the performance.
 606          */
 607         expire = ktime_sub_ns(ktime_get(), NSEC_PER_SEC);
 608         if (ktime_to_ns(lli->lli_glimpse_time) &&
 609             ktime_before(expire, lli->lli_glimpse_time)) {
 610                 up_write(&lli->lli_glimpse_sem);
 611                 lli->lli_agl_index = 0;
 612                 iput(inode);
 613                 RETURN_EXIT;
 614         }
 615
 616         CDEBUG(D_READA,
 617                "Handling (init) async glimpse: inode = " DFID", idx = %llu\n",
 618                PFID(&lli->lli_fid), index);
 619
 620         cl_agl(inode);
 621         lli->lli_agl_index = 0;
 622         lli->lli_glimpse_time = ktime_get();
 623         up_write(&lli->lli_glimpse_sem);
 624
 625         CDEBUG(D_READA,
 626                "Handled (init) async glimpse: inode= " DFID", idx = %llu, rc = %d\n",
 627                PFID(&lli->lli_fid), index, rc);
 628
 629         iput(inode);
 630
 631         EXIT;
 632 }
 633
 634 /*
 635  * prepare inode for sa entry, add it into agl list, now sa_entry is ready
 636  * to be used by scanner process.
 637  */
 638 static void sa_instantiate(struct ll_statahead_info *sai,
 639                            struct sa_entry *entry)
 640 {
 641         struct inode *dir = sai->sai_dentry->d_inode;
 642         struct inode *child;
 643         struct md_enqueue_info *minfo;
 644         struct lookup_intent *it;
 645         struct ptlrpc_request *req;
 646         struct mdt_body *body;
 647         int rc = 0;
 648
 649         ENTRY;
 650
 651         LASSERT(entry->se_handle != 0);
 652
 653         minfo = entry->se_minfo;
 654         it = &minfo->mi_it;
 655         req = entry->se_req;
 656         body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
 657         if (!body)
 658                 GOTO(out, rc = -EFAULT);
 659
 660         child = entry->se_inode;
 661         /* revalidate; unlinked and re-created with the same name */
 662         if (unlikely(!lu_fid_eq(&minfo->mi_data.op_fid2, &body->mbo_fid1))) {
 663                 if (child) {
 664                         entry->se_inode = NULL;
 665                         iput(child);
 666                 }
 667                 /* The mdt_body is invalid. Skip this entry */
 668                 GOTO(out, rc = -EAGAIN);
 669         }
 670
 671         it->it_lock_handle = entry->se_handle;
 672         rc = md_revalidate_lock(ll_i2mdexp(dir), it, ll_inode2fid(dir), NULL);
 673         if (rc != 1)
 674                 GOTO(out, rc = -EAGAIN);
 675
 676         rc = ll_prep_inode(&child, req, dir->i_sb, it);
 677         if (rc)
 678                 GOTO(out, rc);
 679
 680         CDEBUG(D_READA, "%s: setting %.*s"DFID" l_data to inode %p\n",
 681                ll_i2sbi(dir)->ll_fsname, entry->se_qstr.len,
 682                entry->se_qstr.name, PFID(ll_inode2fid(child)), child);
 683         ll_set_lock_data(ll_i2sbi(dir)->ll_md_exp, child, it, NULL);
 684
 685         entry->se_inode = child;
 686
 687         if (agl_should_run(sai, child))
 688                 ll_agl_add(sai, child, entry->se_index);
 689
 690         EXIT;
 691
 692 out:
 693         /*
 694          * sa_make_ready() will drop ldlm ibits lock refcount by calling
 695          * ll_intent_drop_lock() in spite of failures. Do not worry about
 696          * calling ll_intent_drop_lock() more than once.
 697          */
 698         sa_make_ready(sai, entry, rc);
 699 }
 700
 701 /* once there are async stat replies, instantiate sa_entry from replies */
 702 static void sa_handle_callback(struct ll_statahead_info *sai)
 703 {
 704         struct ll_inode_info *lli;
 705
 706         lli = ll_i2info(sai->sai_dentry->d_inode);
 707
 708         spin_lock(&lli->lli_sa_lock);
 709         while (sa_has_callback(sai)) {
 710                 struct sa_entry *entry;
 711
 712                 entry = list_entry(sai->sai_interim_entries.next,
 713                                    struct sa_entry, se_list);
 714                 list_del_init(&entry->se_list);
 715                 spin_unlock(&lli->lli_sa_lock);
 716
 717                 sa_instantiate(sai, entry);
 718                 spin_lock(&lli->lli_sa_lock);
 719         }
 720         spin_unlock(&lli->lli_sa_lock);
 721 }
 722
 723 /*
 724  * callback for async stat RPC, because this is called in ptlrpcd context, we
 725  * only put sa_entry in sai_interim_entries, and wake up statahead thread to
 726  * really prepare inode and instantiate sa_entry later.
 727  */
 728 static int ll_statahead_interpret(struct ptlrpc_request *req,
 729                                   struct md_enqueue_info *minfo, int rc)
 730 {
 731         struct lookup_intent *it = &minfo->mi_it;
 732         struct inode *dir = minfo->mi_dir;
 733         struct ll_inode_info *lli = ll_i2info(dir);
 734         struct ll_statahead_info *sai = lli->lli_sai;
 735         struct sa_entry *entry = (struct sa_entry *)minfo->mi_cbdata;
 736         __u64 handle = 0;
 737
 738         ENTRY;
 739
 740         if (it_disposition(it, DISP_LOOKUP_NEG))
 741                 rc = -ENOENT;
 742
 743         /*
 744          * because statahead thread will wait for all inflight RPC to finish,
 745          * sai should be always valid, no need to refcount
 746          */
 747         LASSERT(sai != NULL);
 748         LASSERT(entry != NULL);
 749
 750         CDEBUG(D_READA, "sa_entry %.*s rc %d\n",
 751                entry->se_qstr.len, entry->se_qstr.name, rc);
 752
 753         if (rc != 0) {
 754                 ll_intent_release(it);
 755                 sa_fini_data(minfo);
 756         } else {
 757                 /*
 758                  * release ibits lock ASAP to avoid deadlock when statahead
 759                  * thread enqueues lock on parent in readdir and another
 760                  * process enqueues lock on child with parent lock held, eg.
 761                  * unlink.
 762                  */
 763                 handle = it->it_lock_handle;
 764                 ll_intent_drop_lock(it);
 765                 ll_unlock_md_op_lsm(&minfo->mi_data);
 766         }
 767
 768         spin_lock(&lli->lli_sa_lock);
 769         if (rc != 0) {
 770                 if (__sa_make_ready(sai, entry, rc))
 771                         wake_up(&sai->sai_waitq);
 772         } else {
 773                 int first = 0;
 774
 775                 entry->se_minfo = minfo;
 776                 entry->se_req = ptlrpc_request_addref(req);
 777                 /*
 778                  * Release the async ibits lock ASAP to avoid deadlock
 779                  * when statahead thread tries to enqueue lock on parent
 780                  * for readpage and other tries to enqueue lock on child
 781                  * with parent's lock held, for example: unlink.
 782                  */
 783                 entry->se_handle = handle;
 784                 if (!sa_has_callback(sai))
 785                         first = 1;
 786
 787                 list_add_tail(&entry->se_list, &sai->sai_interim_entries);
 788                 if (first && sai->sai_task)
 789                         wake_up_process(sai->sai_task);
 790         }
 791         sai->sai_replied++;
 792
 793         spin_unlock(&lli->lli_sa_lock);
 794
 795         RETURN(rc);
 796 }
 797
 798 /* async stat for file not found in dcache */
 799 static int sa_lookup(struct inode *dir, struct sa_entry *entry)
 800 {
 801         struct md_enqueue_info   *minfo;
 802         int                       rc;
 803
 804         ENTRY;
 805
 806         minfo = sa_prep_data(dir, NULL, entry);
 807         if (IS_ERR(minfo))
 808                 RETURN(PTR_ERR(minfo));
 809
 810         rc = md_intent_getattr_async(ll_i2mdexp(dir), minfo);
 811         if (rc < 0)
 812                 sa_fini_data(minfo);
 813
 814         RETURN(rc);
 815 }
 816
 817 /**
 818  * async stat for file found in dcache, similar to .revalidate
 819  *
 820  * \retval      1 dentry valid, no RPC sent
 821  * \retval      0 dentry invalid, will send async stat RPC
 822  * \retval      negative number upon error
 823  */
 824 static int sa_revalidate(struct inode *dir, struct sa_entry *entry,
 825                          struct dentry *dentry)
 826 {
 827         struct inode *inode = dentry->d_inode;
 828         struct lookup_intent it = { .it_op = IT_GETATTR,
 829                                     .it_lock_handle = 0 };
 830         struct md_enqueue_info *minfo;
 831         int rc;
 832
 833         ENTRY;
 834
 835         if (unlikely(!inode))
 836                 RETURN(1);
 837
 838         if (d_mountpoint(dentry))
 839                 RETURN(1);
 840
 841         minfo = sa_prep_data(dir, inode, entry);
 842         if (IS_ERR(minfo))
 843                 RETURN(PTR_ERR(minfo));
 844
 845         entry->se_inode = igrab(inode);
 846         rc = md_revalidate_lock(ll_i2mdexp(dir), &it, ll_inode2fid(inode),
 847                                 NULL);
 848         if (rc == 1) {
 849                 entry->se_handle = it.it_lock_handle;
 850                 ll_intent_release(&it);
 851                 sa_fini_data(minfo);
 852                 RETURN(1);
 853         }
 854
 855         rc = md_intent_getattr_async(ll_i2mdexp(dir), minfo);
 856         if (rc < 0) {
 857                 entry->se_inode = NULL;
 858                 iput(inode);
 859                 sa_fini_data(minfo);
 860         }
 861
 862         RETURN(rc);
 863 }
 864
 865 /* async stat for file with @name */
 866 static void sa_statahead(struct dentry *parent, const char *name, int len,
 867                          const struct lu_fid *fid)
 868 {
 869         struct inode *dir = parent->d_inode;
 870         struct ll_inode_info *lli = ll_i2info(dir);
 871         struct ll_statahead_info *sai = lli->lli_sai;
 872         struct dentry *dentry = NULL;
 873         struct sa_entry *entry;
 874         int rc;
 875
 876         ENTRY;
 877
 878         entry = sa_alloc(parent, sai, sai->sai_index, name, len, fid);
 879         if (IS_ERR(entry))
 880                 RETURN_EXIT;
 881
 882         dentry = d_lookup(parent, &entry->se_qstr);
 883         if (!dentry) {
 884                 rc = sa_lookup(dir, entry);
 885         } else {
 886                 rc = sa_revalidate(dir, entry, dentry);
 887                 if (rc == 1 && agl_should_run(sai, dentry->d_inode))
 888                         ll_agl_add(sai, dentry->d_inode, entry->se_index);
 889         }
 890
 891         if (dentry)
 892                 dput(dentry);
 893
 894         if (rc != 0)
 895                 sa_make_ready(sai, entry, rc);
 896         else
 897                 sai->sai_sent++;
 898
 899         sai->sai_index++;
 900
 901         EXIT;
 902 }
 903
 904 /* async glimpse (agl) thread main function */
 905 static int ll_agl_thread(void *arg)
 906 {
 907         struct dentry *parent = (struct dentry *)arg;
 908         struct inode *dir = parent->d_inode;
 909         struct ll_inode_info *plli = ll_i2info(dir);
 910         struct ll_inode_info *clli;
 911         /*
 912          * We already own this reference, so it is safe to take it
 913          * without a lock.
 914          */
 915         struct ll_statahead_info *sai = plli->lli_sai;
 916
 917         ENTRY;
 918
 919         CDEBUG(D_READA, "agl thread started: sai %p, parent %pd\n",
 920                sai, parent);
 921
 922         while (({set_current_state(TASK_IDLE);
 923                  !kthread_should_stop(); })) {
 924                 spin_lock(&plli->lli_agl_lock);
 925                 if (!agl_list_empty(sai)) {
 926                         __set_current_state(TASK_RUNNING);
 927                         clli = agl_first_entry(sai);
 928                         list_del_init(&clli->lli_agl_list);
 929                         spin_unlock(&plli->lli_agl_lock);
 930                         ll_agl_trigger(&clli->lli_vfs_inode, sai);
 931                         cond_resched();
 932                 } else {
 933                         spin_unlock(&plli->lli_agl_lock);
 934                         schedule();
 935                 }
 936         }
 937         __set_current_state(TASK_RUNNING);
 938         RETURN(0);
 939 }
 940
 941 static void ll_stop_agl(struct ll_statahead_info *sai)
 942 {
 943         struct dentry *parent = sai->sai_dentry;
 944         struct ll_inode_info *plli = ll_i2info(parent->d_inode);
 945         struct ll_inode_info *clli;
 946         struct task_struct *agl_task;
 947
 948         spin_lock(&plli->lli_agl_lock);
 949         agl_task = sai->sai_agl_task;
 950         sai->sai_agl_task = NULL;
 951         spin_unlock(&plli->lli_agl_lock);
 952         if (!agl_task)
 953                 return;
 954
 955         CDEBUG(D_READA, "stop agl thread: sai %p pid %u\n",
 956                sai, (unsigned int)agl_task->pid);
 957         kthread_stop(agl_task);
 958
 959         spin_lock(&plli->lli_agl_lock);
 960         while (!agl_list_empty(sai)) {
 961                 clli = agl_first_entry(sai);
 962                 list_del_init(&clli->lli_agl_list);
 963                 spin_unlock(&plli->lli_agl_lock);
 964                 clli->lli_agl_index = 0;
 965                 iput(&clli->lli_vfs_inode);
 966                 spin_lock(&plli->lli_agl_lock);
 967         }
 968         spin_unlock(&plli->lli_agl_lock);
 969         CDEBUG(D_READA, "agl thread stopped: sai %p, parent %pd\n",
 970                sai, parent);
 971         ll_sai_put(sai);
 972 }
 973
 974 /* start agl thread */
 975 static void ll_start_agl(struct dentry *parent, struct ll_statahead_info *sai)
 976 {
 977         int node = cfs_cpt_spread_node(cfs_cpt_tab, CFS_CPT_ANY);
 978         struct ll_inode_info *plli;
 979         struct task_struct *task;
 980
 981         ENTRY;
 982
 983         CDEBUG(D_READA, "start agl thread: sai %p, parent %pd\n",
 984                sai, parent);
 985
 986         plli = ll_i2info(parent->d_inode);
 987         task = kthread_create_on_node(ll_agl_thread, parent, node, "ll_agl_%d",
 988                                       plli->lli_opendir_pid);
 989         if (IS_ERR(task)) {
 990                 CERROR("can't start ll_agl thread, rc: %ld\n", PTR_ERR(task));
 991                 RETURN_EXIT;
 992         }
 993         sai->sai_agl_task = task;
 994         atomic_inc(&ll_i2sbi(d_inode(parent))->ll_agl_total);
 995         /* Get an extra reference that the thread holds */
 996         ll_sai_get(d_inode(parent));
 997
 998         wake_up_process(task);
 999
1000         EXIT;
1001 }
1002
1003 /* statahead thread main function */
1004 static int ll_statahead_thread(void *arg)
1005 {
1006         struct dentry *parent = (struct dentry *)arg;
1007         struct inode *dir = parent->d_inode;
1008         struct ll_inode_info *lli = ll_i2info(dir);
1009         struct ll_sb_info *sbi = ll_i2sbi(dir);
1010         struct ll_statahead_info *sai = lli->lli_sai;
1011         int first = 0;
1012         struct md_op_data *op_data;
1013         struct page *page = NULL;
1014         __u64 pos = 0;
1015         int rc = 0;
1016
1017         ENTRY;
1018
1019         CDEBUG(D_READA, "statahead thread starting: sai %p, parent %pd\n",
1020                sai, parent);
1021
1022         OBD_ALLOC_PTR(op_data);
1023         if (!op_data)
1024                 GOTO(out, rc = -ENOMEM);
1025
1026         while (pos != MDS_DIR_END_OFF && sai->sai_task) {
1027                 struct lu_dirpage *dp;
1028                 struct lu_dirent  *ent;
1029
1030                 op_data = ll_prep_md_op_data(op_data, dir, dir, NULL, 0, 0,
1031                                              LUSTRE_OPC_ANY, dir);
1032                 if (IS_ERR(op_data)) {
1033                         rc = PTR_ERR(op_data);
1034                         break;
1035                 }
1036
1037                 sai->sai_in_readpage = 1;
1038                 page = ll_get_dir_page(dir, op_data, pos);
1039                 ll_unlock_md_op_lsm(op_data);
1040                 sai->sai_in_readpage = 0;
1041                 if (IS_ERR(page)) {
1042                         rc = PTR_ERR(page);
1043                         CDEBUG(D_READA,
1044                                "error reading dir "DFID" at %llu /%llu opendir_pid = %u: rc = %d\n",
1045                                PFID(ll_inode2fid(dir)), pos, sai->sai_index,
1046                                lli->lli_opendir_pid, rc);
1047                         break;
1048                 }
1049
1050                 dp = page_address(page);
1051                 for (ent = lu_dirent_start(dp);
1052                      ent != NULL && sai->sai_task &&
1053                      !sa_low_hit(sai);
1054                      ent = lu_dirent_next(ent)) {
1055                         __u64 hash;
1056                         int namelen;
1057                         char *name;
1058                         struct lu_fid fid;
1059
1060                         hash = le64_to_cpu(ent->lde_hash);
1061                         if (unlikely(hash < pos))
1062                                 /*
1063                                  * Skip until we find target hash value.
1064                                  */
1065                                 continue;
1066
1067                         namelen = le16_to_cpu(ent->lde_namelen);
1068                         if (unlikely(namelen == 0))
1069                                 /*
1070                                  * Skip dummy record.
1071                                  */
1072                                 continue;
1073
1074                         name = ent->lde_name;
1075                         if (name[0] == '.') {
1076                                 if (namelen == 1) {
1077                                         /*
1078                                          * skip "."
1079                                          */
1080                                         continue;
1081                                 } else if (name[1] == '.' && namelen == 2) {
1082                                         /*
1083                                          * skip ".."
1084                                          */
1085                                         continue;
1086                                 } else if (!sai->sai_ls_all) {
1087                                         /*
1088                                          * skip hidden files.
1089                                          */
1090                                         sai->sai_skip_hidden++;
1091                                         continue;
1092                                 }
1093                         }
1094
1095                         /*
1096                          * don't stat-ahead first entry.
1097                          */
1098                         if (unlikely(++first == 1))
1099                                 continue;
1100
1101                         fid_le_to_cpu(&fid, &ent->lde_fid);
1102
1103                         while (({set_current_state(TASK_IDLE);
1104                                  sai->sai_task; })) {
1105                                 if (sa_has_callback(sai)) {
1106                                         __set_current_state(TASK_RUNNING);
1107                                         sa_handle_callback(sai);
1108                                 }
1109
1110                                 spin_lock(&lli->lli_agl_lock);
1111                                 while (sa_sent_full(sai) &&
1112                                        !agl_list_empty(sai)) {
1113                                         struct ll_inode_info *clli;
1114
1115                                         __set_current_state(TASK_RUNNING);
1116                                         clli = agl_first_entry(sai);
1117                                         list_del_init(&clli->lli_agl_list);
1118                                         spin_unlock(&lli->lli_agl_lock);
1119
1120                                         ll_agl_trigger(&clli->lli_vfs_inode,
1121                                                        sai);
1122                                         cond_resched();
1123                                         spin_lock(&lli->lli_agl_lock);
1124                                 }
1125                                 spin_unlock(&lli->lli_agl_lock);
1126
1127                                 if (!sa_sent_full(sai))
1128                                         break;
1129                                 schedule();
1130                         }
1131                         __set_current_state(TASK_RUNNING);
1132
1133                         sa_statahead(parent, name, namelen, &fid);
1134                 }
1135
1136                 pos = le64_to_cpu(dp->ldp_hash_end);
1137                 ll_release_page(dir, page,
1138                                 le32_to_cpu(dp->ldp_flags) & LDF_COLLIDE);
1139
1140                 if (sa_low_hit(sai)) {
1141                         rc = -EFAULT;
1142                         atomic_inc(&sbi->ll_sa_wrong);
1143                         CDEBUG(D_READA,
1144                                "Statahead for dir "DFID" hit ratio too low: hit/miss %llu/%llu, sent/replied %llu/%llu, stoppingstatahead thread: pid %d\n",
1145                                PFID(&lli->lli_fid), sai->sai_hit,
1146                                sai->sai_miss, sai->sai_sent,
1147                                sai->sai_replied, current->pid);
1148                         break;
1149                 }
1150         }
1151         ll_finish_md_op_data(op_data);
1152
1153         if (rc < 0) {
1154                 spin_lock(&lli->lli_sa_lock);
1155                 sai->sai_task = NULL;
1156                 lli->lli_sa_enabled = 0;
1157                 spin_unlock(&lli->lli_sa_lock);
1158         }
1159
1160         /*
1161          * statahead is finished, but statahead entries need to be cached, wait
1162          * for file release to stop me.
1163          */
1164         while (({set_current_state(TASK_IDLE);
1165                  sai->sai_task; })) {
1166                 if (sa_has_callback(sai)) {
1167                         __set_current_state(TASK_RUNNING);
1168                         sa_handle_callback(sai);
1169                 } else {
1170                         schedule();
1171                 }
1172         }
1173         __set_current_state(TASK_RUNNING);
1174
1175         EXIT;
1176 out:
1177         ll_stop_agl(sai);
1178
1179         /*
1180          * wait for inflight statahead RPCs to finish, and then we can free sai
1181          * safely because statahead RPC will access sai data
1182          */
1183         while (sai->sai_sent != sai->sai_replied)
1184                 /* in case we're not woken up, timeout wait */
1185                 msleep(125);
1186
1187         /* release resources held by statahead RPCs */
1188         sa_handle_callback(sai);
1189
1190         CDEBUG(D_READA, "%s: statahead thread stopped: sai %p, parent %pd\n",
1191                sbi->ll_fsname, sai, parent);
1192
1193         spin_lock(&lli->lli_sa_lock);
1194         sai->sai_task = NULL;
1195         spin_unlock(&lli->lli_sa_lock);
1196         wake_up(&sai->sai_waitq);
1197
1198         ll_sai_put(sai);
1199
1200         return rc;
1201 }
1202
1203 /* authorize opened dir handle @key to statahead */
1204 void ll_authorize_statahead(struct inode *dir, void *key)
1205 {
1206         struct ll_inode_info *lli = ll_i2info(dir);
1207
1208         spin_lock(&lli->lli_sa_lock);
1209         if (!lli->lli_opendir_key && !lli->lli_sai) {
1210                 /*
1211                  * if lli_sai is not NULL, it means previous statahead is not
1212                  * finished yet, we'd better not start a new statahead for now.
1213                  */
1214                 LASSERT(lli->lli_opendir_pid == 0);
1215                 lli->lli_opendir_key = key;
1216                 lli->lli_opendir_pid = current->pid;
1217                 lli->lli_sa_enabled = 1;
1218         }
1219         spin_unlock(&lli->lli_sa_lock);
1220 }
1221
1222 /*
1223  * deauthorize opened dir handle @key to statahead, and notify statahead thread
1224  * to quit if it's running.
1225  */
1226 void ll_deauthorize_statahead(struct inode *dir, void *key)
1227 {
1228         struct ll_inode_info *lli = ll_i2info(dir);
1229         struct ll_statahead_info *sai;
1230
1231         LASSERT(lli->lli_opendir_key == key);
1232         LASSERT(lli->lli_opendir_pid != 0);
1233
1234         CDEBUG(D_READA, "deauthorize statahead for "DFID"\n",
1235                PFID(&lli->lli_fid));
1236
1237         spin_lock(&lli->lli_sa_lock);
1238         lli->lli_opendir_key = NULL;
1239         lli->lli_opendir_pid = 0;
1240         lli->lli_sa_enabled = 0;
1241         sai = lli->lli_sai;
1242         if (sai && sai->sai_task) {
1243                 /*
1244                  * statahead thread may not have quit yet because it needs to
1245                  * cache entries, now it's time to tell it to quit.
1246                  *
1247                  * wake_up_process() provides the necessary barriers
1248                  * to pair with set_current_state().
1249                  */
1250                 struct task_struct *task = sai->sai_task;
1251
1252                 sai->sai_task = NULL;
1253                 wake_up_process(task);
1254         }
1255         spin_unlock(&lli->lli_sa_lock);
1256 }
1257
1258 enum {
1259         /**
1260          * not first dirent, or is "."
1261          */
1262         LS_NOT_FIRST_DE = 0,
1263         /**
1264          * the first non-hidden dirent
1265          */
1266         LS_FIRST_DE,
1267         /**
1268          * the first hidden dirent, that is "."
1269          */
1270         LS_FIRST_DOT_DE
1271 };
1272
1273 /* file is first dirent under @dir */
1274 static int is_first_dirent(struct inode *dir, struct dentry *dentry)
1275 {
1276         struct qstr          *target = &dentry->d_name;
1277         struct md_op_data    *op_data;
1278         int                   dot_de;
1279         struct page          *page = NULL;
1280         int                   rc = LS_NOT_FIRST_DE;
1281         __u64                 pos = 0;
1282
1283         ENTRY;
1284
1285         op_data = ll_prep_md_op_data(NULL, dir, dir, NULL, 0, 0,
1286                                      LUSTRE_OPC_ANY, dir);
1287         if (IS_ERR(op_data))
1288                 RETURN(PTR_ERR(op_data));
1289         /**
1290          *FIXME choose the start offset of the readdir
1291          */
1292
1293         page = ll_get_dir_page(dir, op_data, 0);
1294
1295         while (1) {
1296                 struct lu_dirpage *dp;
1297                 struct lu_dirent  *ent;
1298
1299                 if (IS_ERR(page)) {
1300                         struct ll_inode_info *lli = ll_i2info(dir);
1301
1302                         rc = PTR_ERR(page);
1303                         CERROR("%s: reading dir "DFID" at %llu opendir_pid = %u : rc = %d\n",
1304                                ll_i2sbi(dir)->ll_fsname,
1305                                PFID(ll_inode2fid(dir)), pos,
1306                                lli->lli_opendir_pid, rc);
1307                         break;
1308                 }
1309
1310                 dp = page_address(page);
1311                 for (ent = lu_dirent_start(dp); ent != NULL;
1312                      ent = lu_dirent_next(ent)) {
1313                         __u64 hash;
1314                         int namelen;
1315                         char *name;
1316
1317                         hash = le64_to_cpu(ent->lde_hash);
1318                         /*
1319                          * The ll_get_dir_page() can return any page containing
1320                          * the given hash which may be not the start hash.
1321                          */
1322                         if (unlikely(hash < pos))
1323                                 continue;
1324
1325                         namelen = le16_to_cpu(ent->lde_namelen);
1326                         if (unlikely(namelen == 0))
1327                                 /*
1328                                  * skip dummy record.
1329                                  */
1330                                 continue;
1331
1332                         name = ent->lde_name;
1333                         if (name[0] == '.') {
1334                                 if (namelen == 1)
1335                                         /*
1336                                          * skip "."
1337                                          */
1338                                         continue;
1339                                 else if (name[1] == '.' && namelen == 2)
1340                                         /*
1341                                          * skip ".."
1342                                          */
1343                                         continue;
1344                                 else
1345                                         dot_de = 1;
1346                         } else {
1347                                 dot_de = 0;
1348                         }
1349
1350                         if (dot_de && target->name[0] != '.') {
1351                                 CDEBUG(D_READA, "%.*s skip hidden file %.*s\n",
1352                                        target->len, target->name,
1353                                        namelen, name);
1354                                 continue;
1355                         }
1356
1357                         if (target->len != namelen ||
1358                             memcmp(target->name, name, namelen) != 0)
1359                                 rc = LS_NOT_FIRST_DE;
1360                         else if (!dot_de)
1361                                 rc = LS_FIRST_DE;
1362                         else
1363                                 rc = LS_FIRST_DOT_DE;
1364
1365                         ll_release_page(dir, page, false);
1366                         GOTO(out, rc);
1367                 }
1368                 pos = le64_to_cpu(dp->ldp_hash_end);
1369                 if (pos == MDS_DIR_END_OFF) {
1370                         /*
1371                          * End of directory reached.
1372                          */
1373                         ll_release_page(dir, page, false);
1374                         GOTO(out, rc);
1375                 } else {
1376                         /*
1377                          * chain is exhausted
1378                          * Normal case: continue to the next page.
1379                          */
1380                         ll_release_page(dir, page, le32_to_cpu(dp->ldp_flags) &
1381                                               LDF_COLLIDE);
1382                         page = ll_get_dir_page(dir, op_data, pos);
1383                 }
1384         }
1385         EXIT;
1386 out:
1387         ll_finish_md_op_data(op_data);
1388
1389         return rc;
1390 }
1391
1392 /**
1393  * revalidate @dentryp from statahead cache
1394  *
1395  * \param[in] dir       parent directory
1396  * \param[in] sai       sai structure
1397  * \param[out] dentryp  pointer to dentry which will be revalidated
1398  * \param[in] unplug    unplug statahead window only (normally for negative
1399  *                      dentry)
1400  * \retval              1 on success, dentry is saved in @dentryp
1401  * \retval              0 if revalidation failed (no proper lock on client)
1402  * \retval              negative number upon error
1403  */
1404 static int revalidate_statahead_dentry(struct inode *dir,
1405                                        struct ll_statahead_info *sai,
1406                                        struct dentry **dentryp,
1407                                        bool unplug)
1408 {
1409         struct sa_entry *entry = NULL;
1410         struct ll_dentry_data *ldd;
1411         struct ll_inode_info *lli = ll_i2info(dir);
1412         int rc = 0;
1413
1414         ENTRY;
1415
1416         if ((*dentryp)->d_name.name[0] == '.') {
1417                 if (sai->sai_ls_all ||
1418                     sai->sai_miss_hidden >= sai->sai_skip_hidden) {
1419                         /*
1420                          * Hidden dentry is the first one, or statahead
1421                          * thread does not skip so many hidden dentries
1422                          * before "sai_ls_all" enabled as below.
1423                          */
1424                 } else {
1425                         if (!sai->sai_ls_all)
1426                                 /*
1427                                  * It maybe because hidden dentry is not
1428                                  * the first one, "sai_ls_all" was not
1429                                  * set, then "ls -al" missed. Enable
1430                                  * "sai_ls_all" for such case.
1431                                  */
1432                                 sai->sai_ls_all = 1;
1433
1434                         /*
1435                          * Such "getattr" has been skipped before
1436                          * "sai_ls_all" enabled as above.
1437                          */
1438                         sai->sai_miss_hidden++;
1439                         RETURN(-EAGAIN);
1440                 }
1441         }
1442
1443         if (unplug)
1444                 GOTO(out, rc = 1);
1445
1446         entry = sa_get(sai, &(*dentryp)->d_name);
1447         if (!entry)
1448                 GOTO(out, rc = -EAGAIN);
1449
1450         /* if statahead is busy in readdir, help it do post-work */
1451         if (!sa_ready(entry) && sai->sai_in_readpage)
1452                 sa_handle_callback(sai);
1453
1454         if (!sa_ready(entry)) {
1455                 spin_lock(&lli->lli_sa_lock);
1456                 sai->sai_index_wait = entry->se_index;
1457                 spin_unlock(&lli->lli_sa_lock);
1458                 rc = wait_event_idle_timeout(sai->sai_waitq, sa_ready(entry),
1459                                              cfs_time_seconds(30));
1460                 if (rc == 0) {
1461                         /*
1462                          * entry may not be ready, so it may be used by inflight
1463                          * statahead RPC, don't free it.
1464                          */
1465                         entry = NULL;
1466                         GOTO(out, rc = -EAGAIN);
1467                 }
1468         }
1469
1470         /*
1471          * We need to see the value that was set immediately before we
1472          * were woken up.
1473          */
1474         if (smp_load_acquire(&entry->se_state) == SA_ENTRY_SUCC &&
1475             entry->se_inode) {
1476                 struct inode *inode = entry->se_inode;
1477                 struct lookup_intent it = { .it_op = IT_GETATTR,
1478                                             .it_lock_handle =
1479                                                 entry->se_handle };
1480                 __u64 bits;
1481
1482                 rc = md_revalidate_lock(ll_i2mdexp(dir), &it,
1483                                         ll_inode2fid(inode), &bits);
1484                 if (rc == 1) {
1485                         if (!(*dentryp)->d_inode) {
1486                                 struct dentry *alias;
1487
1488                                 alias = ll_splice_alias(inode, *dentryp);
1489                                 if (IS_ERR(alias)) {
1490                                         ll_intent_release(&it);
1491                                         GOTO(out, rc = PTR_ERR(alias));
1492                                 }
1493                                 *dentryp = alias;
1494                                 /*
1495                                  * statahead prepared this inode, transfer inode
1496                                  * refcount from sa_entry to dentry
1497                                  */
1498                                 entry->se_inode = NULL;
1499                         } else if ((*dentryp)->d_inode != inode) {
1500                                 /* revalidate, but inode is recreated */
1501                                 CDEBUG(D_READA,
1502                                        "%s: stale dentry %pd inode " DFID", statahead inode "DFID "\n",
1503                                        ll_i2sbi(inode)->ll_fsname, *dentryp,
1504                                        PFID(ll_inode2fid((*dentryp)->d_inode)),
1505                                        PFID(ll_inode2fid(inode)));
1506                                 ll_intent_release(&it);
1507                                 GOTO(out, rc = -ESTALE);
1508                         }
1509
1510                         if ((bits & MDS_INODELOCK_LOOKUP) &&
1511                             d_lustre_invalid(*dentryp))
1512                                 d_lustre_revalidate(*dentryp);
1513                         ll_intent_release(&it);
1514                 }
1515         }
1516 out:
1517         /*
1518          * statahead cached sa_entry can be used only once, and will be killed
1519          * right after use, so if lookup/revalidate accessed statahead cache,
1520          * set dentry ldd_sa_generation to parent lli_sa_generation, later if we
1521          * stat this file again, we know we've done statahead before, see
1522          * dentry_may_statahead().
1523          */
1524         ldd = ll_d2d(*dentryp);
1525         /* ldd can be NULL if llite lookup failed. */
1526         if (ldd)
1527                 ldd->lld_sa_generation = lli->lli_sa_generation;
1528         sa_put(sai, entry);
1529         spin_lock(&lli->lli_sa_lock);
1530         if (sai->sai_task)
1531                 wake_up_process(sai->sai_task);
1532         spin_unlock(&lli->lli_sa_lock);
1533
1534         RETURN(rc);
1535 }
1536
1537 /**
1538  * start statahead thread
1539  *
1540  * \param[in] dir       parent directory
1541  * \param[in] dentry    dentry that triggers statahead, normally the first
1542  *                      dirent under @dir
1543  * \param[in] agl       indicate whether AGL is needed
1544  * \retval              -EAGAIN on success, because when this function is
1545  *                      called, it's already in lookup call, so client should
1546  *                      do it itself instead of waiting for statahead thread
1547  *                      to do it asynchronously.
1548  * \retval              negative number upon error
1549  */
1550 static int start_statahead_thread(struct inode *dir, struct dentry *dentry,
1551                                   bool agl)
1552 {
1553         int node = cfs_cpt_spread_node(cfs_cpt_tab, CFS_CPT_ANY);
1554         struct ll_inode_info *lli = ll_i2info(dir);
1555         struct ll_statahead_info *sai = NULL;
1556         struct dentry *parent = dentry->d_parent;
1557         struct task_struct *task;
1558         struct ll_sb_info *sbi = ll_i2sbi(parent->d_inode);
1559         int first = LS_FIRST_DE;
1560         int rc = 0;
1561
1562         ENTRY;
1563
1564         /* I am the "lli_opendir_pid" owner, only me can set "lli_sai". */
1565         first = is_first_dirent(dir, dentry);
1566         if (first == LS_NOT_FIRST_DE)
1567                 /* It is not "ls -{a}l" operation, no need statahead for it. */
1568                 GOTO(out, rc = -EFAULT);
1569
1570         if (unlikely(atomic_inc_return(&sbi->ll_sa_running) >
1571                                        sbi->ll_sa_running_max)) {
1572                 CDEBUG(D_READA,
1573                        "Too many concurrent statahead instances, avoid new statahead instance temporarily.\n");
1574                 GOTO(out, rc = -EMFILE);
1575         }
1576
1577         sai = ll_sai_alloc(parent);
1578         if (!sai)
1579                 GOTO(out, rc = -ENOMEM);
1580
1581         sai->sai_ls_all = (first == LS_FIRST_DOT_DE);
1582
1583         /*
1584          * if current lli_opendir_key was deauthorized, or dir re-opened by
1585          * another process, don't start statahead, otherwise the newly spawned
1586          * statahead thread won't be notified to quit.
1587          */
1588         spin_lock(&lli->lli_sa_lock);
1589         if (unlikely(lli->lli_sai || !lli->lli_opendir_key ||
1590                      lli->lli_opendir_pid != current->pid)) {
1591                 spin_unlock(&lli->lli_sa_lock);
1592                 GOTO(out, rc = -EPERM);
1593         }
1594         lli->lli_sai = sai;
1595         spin_unlock(&lli->lli_sa_lock);
1596
1597         CDEBUG(D_READA, "start statahead thread: [pid %d] [parent %pd]\n",
1598                current->pid, parent);
1599
1600         task = kthread_create_on_node(ll_statahead_thread, parent, node,
1601                                       "ll_sa_%u", lli->lli_opendir_pid);
1602         if (IS_ERR(task)) {
1603                 spin_lock(&lli->lli_sa_lock);
1604                 lli->lli_sai = NULL;
1605                 spin_unlock(&lli->lli_sa_lock);
1606                 rc = PTR_ERR(task);
1607                 CERROR("can't start ll_sa thread, rc: %d\n", rc);
1608                 GOTO(out, rc);
1609         }
1610
1611         if (ll_i2sbi(parent->d_inode)->ll_flags & LL_SBI_AGL_ENABLED && agl)
1612                 ll_start_agl(parent, sai);
1613
1614         atomic_inc(&ll_i2sbi(parent->d_inode)->ll_sa_total);
1615         sai->sai_task = task;
1616
1617         wake_up_process(task);
1618         /*
1619          * We don't stat-ahead for the first dirent since we are already in
1620          * lookup.
1621          */
1622         RETURN(-EAGAIN);
1623
1624 out:
1625         /*
1626          * once we start statahead thread failed, disable statahead so that
1627          * subsequent stat won't waste time to try it.
1628          */
1629         spin_lock(&lli->lli_sa_lock);
1630         if (lli->lli_opendir_pid == current->pid)
1631                 lli->lli_sa_enabled = 0;
1632         spin_unlock(&lli->lli_sa_lock);
1633
1634         if (sai)
1635                 ll_sai_free(sai);
1636         if (first != LS_NOT_FIRST_DE)
1637                 atomic_dec(&sbi->ll_sa_running);
1638
1639         RETURN(rc);
1640 }
1641
1642 /*
1643  * Check whether statahead for @dir was started.
1644  */
1645 static inline bool ll_statahead_started(struct inode *dir, bool agl)
1646 {
1647         struct ll_inode_info *lli = ll_i2info(dir);
1648         struct ll_statahead_info *sai;
1649
1650         spin_lock(&lli->lli_sa_lock);
1651         sai = lli->lli_sai;
1652         if (sai && (sai->sai_agl_task != NULL) != agl)
1653                 CDEBUG(D_READA,
1654                        "%s: Statahead AGL hint changed from %d to %d\n",
1655                        ll_i2sbi(dir)->ll_fsname,
1656                        sai->sai_agl_task != NULL, agl);
1657         spin_unlock(&lli->lli_sa_lock);
1658
1659         return !!sai;
1660 }
1661
1662 /**
1663  * statahead entry function, this is called when client getattr on a file, it
1664  * will start statahead thread if this is the first dir entry, else revalidate
1665  * dentry from statahead cache.
1666  *
1667  * \param[in]  dir      parent directory
1668  * \param[out] dentryp  dentry to getattr
1669  * \param[in]  agl      whether start the agl thread
1670  *
1671  * \retval              1 on success
1672  * \retval              0 revalidation from statahead cache failed, caller needs
1673  *                      to getattr from server directly
1674  * \retval              negative number on error, caller often ignores this and
1675  *                      then getattr from server
1676  */
1677 int ll_start_statahead(struct inode *dir, struct dentry *dentry, bool agl)
1678 {
1679         if (!ll_statahead_started(dir, agl))
1680                 return start_statahead_thread(dir, dentry, agl);
1681         return 0;
1682 }
1683
1684 /**
1685  * revalidate dentry from statahead cache.
1686  *
1687  * \param[in]  dir      parent directory
1688  * \param[out] dentryp  dentry to getattr
1689  * \param[in]  unplug   unplug statahead window only (normally for negative
1690  *                      dentry)
1691  * \retval              1 on success
1692  * \retval              0 revalidation from statahead cache failed, caller needs
1693  *                      to getattr from server directly
1694  * \retval              negative number on error, caller often ignores this and
1695  *                      then getattr from server
1696  */
1697 int ll_revalidate_statahead(struct inode *dir, struct dentry **dentryp,
1698                             bool unplug)
1699 {
1700         struct ll_statahead_info *sai;
1701         int rc = 0;
1702
1703         sai = ll_sai_get(dir);
1704         if (sai) {
1705                 rc = revalidate_statahead_dentry(dir, sai, dentryp, unplug);
1706                 CDEBUG(D_READA, "revalidate statahead %pd: rc = %d.\n",
1707                        *dentryp, rc);
1708                 ll_sai_put(sai);
1709         }
1710         return rc;
1711 }