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, 2013, 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/mm.h>
  40 #include <linux/highmem.h>
  41 #include <linux/pagemap.h>
  42
  43 #define DEBUG_SUBSYSTEM S_LLITE
  44
  45 #include <obd_support.h>
  46 #include <lustre_dlm.h>
  47 #include "llite_internal.h"
  48
  49 #define SA_OMITTED_ENTRY_MAX 8ULL
  50
  51 typedef enum {
  52         /** negative values are for error cases */
  53         SA_ENTRY_INIT = 0,      /** init entry */
  54         SA_ENTRY_SUCC = 1,      /** stat succeed */
  55         SA_ENTRY_INVA = 2,      /** invalid entry */
  56 } se_state_t;
  57
  58 /* sa_entry is not refcounted: statahead thread allocates it and do async stat,
  59  * and in async stat callback ll_statahead_interpret() will add it into
  60  * sai_cb_entries, later statahead thread will call sa_handle_callback() to
  61  * instantiate entry and move it into sai_entries, and then only scanner process
  62  * can access and free it. */
  63 struct sa_entry {
  64         /* link into sai_cb_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 };
  85
  86 static unsigned int sai_generation = 0;
  87 static DEFINE_SPINLOCK(sai_generation_lock);
  88
  89 static inline int sa_unhashed(struct sa_entry *entry)
  90 {
  91         return list_empty(&entry->se_hash);
  92 }
  93
  94 /*
  95  * The entry only can be released by the caller, it is necessary to hold lock.
  96  */
  97 static inline int sa_ready(struct sa_entry *entry)
  98 {
  99         smp_rmb();
 100         return (entry->se_state != SA_ENTRY_INIT);
 101 }
 102
 103 static inline int sa_hash(int val)
 104 {
 105         return val & LL_SA_CACHE_MASK;
 106 }
 107
 108 /*
 109  * Insert entry to hash SA table.
 110  */
 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 /*
 122  * Remove entry from SA table.
 123  */
 124 static inline void
 125 sa_unhash(struct ll_statahead_info *sai, struct sa_entry *entry)
 126 {
 127         int i = sa_hash(entry->se_qstr.hash);
 128
 129         spin_lock(&sai->sai_cache_lock[i]);
 130         list_del_init(&entry->se_hash);
 131         spin_unlock(&sai->sai_cache_lock[i]);
 132 }
 133
 134 static inline int agl_should_run(struct ll_statahead_info *sai,
 135                                  struct inode *inode)
 136 {
 137         return (inode != NULL && S_ISREG(inode->i_mode) && sai->sai_agl_valid);
 138 }
 139
 140 static inline struct ll_inode_info *
 141 agl_first_entry(struct ll_statahead_info *sai)
 142 {
 143         return list_entry(sai->sai_agls.next, struct ll_inode_info,
 144                           lli_agl_list);
 145 }
 146
 147 static inline int sa_sent_full(struct ll_statahead_info *sai)
 148 {
 149         return atomic_read(&sai->sai_cache_count) >= sai->sai_max;
 150 }
 151
 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 add it into hash to let 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 {
 188         struct ll_inode_info *lli;
 189         struct sa_entry *entry;
 190         int entry_size;
 191         char *dname;
 192         ENTRY;
 193
 194         entry_size = sizeof(struct sa_entry) + (len & ~3) + 4;
 195         OBD_ALLOC(entry, entry_size);
 196         if (unlikely(entry == NULL))
 197                 RETURN(ERR_PTR(-ENOMEM));
 198
 199         CDEBUG(D_READA, "alloc sa entry %.*s(%p) index "LPU64"\n",
 200                len, name, entry, index);
 201
 202         entry->se_index = index;
 203
 204         entry->se_state = SA_ENTRY_INIT;
 205         entry->se_size = entry_size;
 206         dname = (char *)entry + sizeof(struct sa_entry);
 207         memcpy(dname, name, len);
 208         dname[len] = 0;
 209         entry->se_qstr.hash = full_name_hash(name, len);
 210         entry->se_qstr.len = len;
 211         entry->se_qstr.name = dname;
 212
 213         lli = ll_i2info(sai->sai_inode);
 214         spin_lock(&lli->lli_sa_lock);
 215         INIT_LIST_HEAD(&entry->se_list);
 216         sa_rehash(sai, entry);
 217         spin_unlock(&lli->lli_sa_lock);
 218
 219         atomic_inc(&sai->sai_cache_count);
 220
 221         RETURN(entry);
 222 }
 223
 224 /* free sa_entry which should have been unhashed and not in any list */
 225 static void sa_free(struct ll_statahead_info *sai, struct sa_entry *entry)
 226 {
 227         CDEBUG(D_READA, "free sa entry %.*s(%p) index "LPU64"\n",
 228                entry->se_qstr.len, entry->se_qstr.name, entry,
 229                entry->se_index);
 230
 231         LASSERT(list_empty(&entry->se_list));
 232         LASSERT(sa_unhashed(entry));
 233
 234         OBD_FREE(entry, entry->se_size);
 235         atomic_dec(&sai->sai_cache_count);
 236 }
 237
 238 /* find sa_entry by name, used by directory scanner, lock is not needed because
 239  * only scanner can remove the entry from hash.
 240  */
 241 static struct sa_entry *
 242 sa_get(struct ll_statahead_info *sai, const struct qstr *qstr)
 243 {
 244         struct sa_entry *entry;
 245         int i = sa_hash(qstr->hash);
 246
 247         list_for_each_entry(entry, &sai->sai_cache[i], se_hash) {
 248                 if (entry->se_qstr.hash == qstr->hash &&
 249                     entry->se_qstr.len == qstr->len &&
 250                     memcmp(entry->se_qstr.name, qstr->name, qstr->len) == 0)
 251                         return entry;
 252         }
 253         return NULL;
 254 }
 255
 256 static inline void
 257 sa_kill(struct ll_statahead_info *sai, struct sa_entry *entry)
 258 {
 259         struct ll_inode_info *lli = ll_i2info(sai->sai_inode);
 260
 261         LASSERT(!sa_unhashed(entry));
 262         LASSERT(!list_empty(&entry->se_list));
 263         LASSERT(sa_ready(entry));
 264
 265         sa_unhash(sai, entry);
 266
 267         spin_lock(&lli->lli_sa_lock);
 268         list_del_init(&entry->se_list);
 269         spin_unlock(&lli->lli_sa_lock);
 270
 271         if (entry->se_inode != NULL)
 272                 iput(entry->se_inode);
 273
 274         sa_free(sai, entry);
 275 }
 276
 277 /* called by scanner after use, sa_entry will be killed */
 278 static void
 279 sa_put(struct ll_statahead_info *sai, struct sa_entry *entry)
 280 {
 281         struct sa_entry *tmp, *next;
 282
 283         if (entry != NULL && entry->se_state == SA_ENTRY_SUCC) {
 284                 struct ll_sb_info *sbi = ll_i2sbi(sai->sai_inode);
 285
 286                 sai->sai_hit++;
 287                 sai->sai_consecutive_miss = 0;
 288                 sai->sai_max = min(2 * sai->sai_max, sbi->ll_sa_max);
 289         } else {
 290                 sai->sai_miss++;
 291                 sai->sai_consecutive_miss++;
 292         }
 293
 294         if (entry != NULL)
 295                 sa_kill(sai, entry);
 296
 297         /* kill old completed entries, only scanner process does this, no need
 298          * to lock */
 299         list_for_each_entry_safe(tmp, next, &sai->sai_entries, se_list) {
 300                 if (!is_omitted_entry(sai, tmp->se_index))
 301                         break;
 302                 sa_kill(sai, tmp);
 303         }
 304
 305         wake_up(&sai->sai_thread.t_ctl_waitq);
 306 }
 307
 308 /* update state and sort add entry to sai_entries by index, return true if
 309  * scanner is waiting on this entry. */
 310 static bool
 311 __sa_make_ready(struct ll_statahead_info *sai, struct sa_entry *entry, int ret)
 312 {
 313         struct sa_entry *se;
 314         struct list_head *pos = &sai->sai_entries;
 315
 316         LASSERT(!sa_ready(entry));
 317         LASSERT(list_empty(&entry->se_list));
 318
 319         entry->se_state = ret < 0 ? SA_ENTRY_INVA : SA_ENTRY_SUCC;
 320
 321         list_for_each_entry_reverse(se, &sai->sai_entries, se_list) {
 322                 if (se->se_index < entry->se_index) {
 323                         pos = &se->se_list;
 324                         break;
 325                 }
 326         }
 327         list_add(&entry->se_list, pos);
 328
 329         return (entry->se_index == sai->sai_index_wait);
 330 }
 331
 332 /* release resources used in async stat RPC, complete entry information and
 333  * wakeup if necessary */
 334 static void
 335 sa_make_ready(struct ll_statahead_info *sai, struct sa_entry *entry, int ret)
 336 {
 337         struct ll_inode_info *lli = ll_i2info(sai->sai_inode);
 338         struct md_enqueue_info *minfo = entry->se_minfo;
 339         struct ptlrpc_request *req = entry->se_req;
 340         bool wakeup;
 341
 342         /* release resources used in RPC */
 343         if (minfo) {
 344                 entry->se_minfo = NULL;
 345                 ll_intent_release(&minfo->mi_it);
 346                 iput(minfo->mi_dir);
 347                 OBD_FREE_PTR(minfo);
 348         }
 349
 350         if (req) {
 351                 entry->se_req = NULL;
 352                 ptlrpc_req_finished(req);
 353         }
 354
 355         spin_lock(&lli->lli_sa_lock);
 356         wakeup = __sa_make_ready(sai, entry, ret);
 357         spin_unlock(&lli->lli_sa_lock);
 358
 359         if (wakeup)
 360                 wake_up(&sai->sai_waitq);
 361 }
 362
 363 /*
 364  * Insert inode into the list of sai_agls.
 365  */
 366 static void ll_agl_add(struct ll_statahead_info *sai,
 367                        struct inode *inode, int index)
 368 {
 369         struct ll_inode_info *child  = ll_i2info(inode);
 370         struct ll_inode_info *parent = ll_i2info(sai->sai_inode);
 371         int                   added  = 0;
 372
 373         spin_lock(&child->lli_agl_lock);
 374         if (child->lli_agl_index == 0) {
 375                 child->lli_agl_index = index;
 376                 spin_unlock(&child->lli_agl_lock);
 377
 378                 LASSERT(list_empty(&child->lli_agl_list));
 379
 380                 igrab(inode);
 381                 spin_lock(&parent->lli_agl_lock);
 382                 if (agl_list_empty(sai))
 383                         added = 1;
 384                 list_add_tail(&child->lli_agl_list, &sai->sai_agls);
 385                 spin_unlock(&parent->lli_agl_lock);
 386         } else {
 387                 spin_unlock(&child->lli_agl_lock);
 388         }
 389
 390         if (added > 0)
 391                 wake_up(&sai->sai_agl_thread.t_ctl_waitq);
 392 }
 393
 394 static struct ll_statahead_info *ll_sai_alloc(void)
 395 {
 396         struct ll_statahead_info *sai;
 397         int                       i;
 398         ENTRY;
 399
 400         OBD_ALLOC_PTR(sai);
 401         if (!sai)
 402                 RETURN(NULL);
 403
 404         atomic_set(&sai->sai_refcount, 1);
 405
 406         spin_lock(&sai_generation_lock);
 407         sai->sai_generation = ++sai_generation;
 408         if (unlikely(sai_generation == 0))
 409                 sai->sai_generation = ++sai_generation;
 410         spin_unlock(&sai_generation_lock);
 411
 412         sai->sai_max = LL_SA_RPC_MIN;
 413         sai->sai_index = 1;
 414         init_waitqueue_head(&sai->sai_waitq);
 415         init_waitqueue_head(&sai->sai_thread.t_ctl_waitq);
 416         init_waitqueue_head(&sai->sai_agl_thread.t_ctl_waitq);
 417
 418         INIT_LIST_HEAD(&sai->sai_interim_entries);
 419         INIT_LIST_HEAD(&sai->sai_entries);
 420         INIT_LIST_HEAD(&sai->sai_agls);
 421
 422         for (i = 0; i < LL_SA_CACHE_SIZE; i++) {
 423                 INIT_LIST_HEAD(&sai->sai_cache[i]);
 424                 spin_lock_init(&sai->sai_cache_lock[i]);
 425         }
 426         atomic_set(&sai->sai_cache_count, 0);
 427
 428         RETURN(sai);
 429 }
 430
 431 static inline struct ll_statahead_info *ll_sai_get(struct inode *dir)
 432 {
 433         struct ll_inode_info *lli = ll_i2info(dir);
 434         struct ll_statahead_info *sai = NULL;
 435
 436         spin_lock(&lli->lli_sa_lock);
 437         sai = lli->lli_sai;
 438         if (sai != NULL)
 439                 atomic_inc(&sai->sai_refcount);
 440         spin_unlock(&lli->lli_sa_lock);
 441
 442         return sai;
 443 }
 444
 445 static void ll_sai_put(struct ll_statahead_info *sai)
 446 {
 447         struct ll_inode_info *lli = ll_i2info(sai->sai_inode);
 448
 449         if (atomic_dec_and_lock(&sai->sai_refcount, &lli->lli_sa_lock)) {
 450                 struct sa_entry *entry, *next;
 451                 struct ll_sb_info *sbi = ll_i2sbi(sai->sai_inode);
 452
 453                 lli->lli_sai = NULL;
 454                 spin_unlock(&lli->lli_sa_lock);
 455
 456                 LASSERT(thread_is_stopped(&sai->sai_thread));
 457                 LASSERT(thread_is_stopped(&sai->sai_agl_thread));
 458                 LASSERT(sai->sai_sent == sai->sai_replied);
 459                 LASSERT(!sa_has_callback(sai));
 460
 461                 list_for_each_entry_safe(entry, next, &sai->sai_entries,
 462                                          se_list)
 463                         sa_kill(sai, entry);
 464
 465                 LASSERT(atomic_read(&sai->sai_cache_count) == 0);
 466                 LASSERT(agl_list_empty(sai));
 467
 468                 iput(sai->sai_inode);
 469                 OBD_FREE_PTR(sai);
 470                 atomic_dec(&sbi->ll_sa_running);
 471         }
 472 }
 473
 474 /* Do NOT forget to drop inode refcount when into sai_agls. */
 475 static void ll_agl_trigger(struct inode *inode, struct ll_statahead_info *sai)
 476 {
 477         struct ll_inode_info *lli = ll_i2info(inode);
 478         __u64 index = lli->lli_agl_index;
 479         int rc;
 480         ENTRY;
 481
 482         LASSERT(list_empty(&lli->lli_agl_list));
 483
 484         /* AGL maybe fall behind statahead with one entry */
 485         if (is_omitted_entry(sai, index + 1)) {
 486                 lli->lli_agl_index = 0;
 487                 iput(inode);
 488                 RETURN_EXIT;
 489         }
 490
 491         /* Someone is in glimpse (sync or async), do nothing. */
 492         rc = down_write_trylock(&lli->lli_glimpse_sem);
 493         if (rc == 0) {
 494                 lli->lli_agl_index = 0;
 495                 iput(inode);
 496                 RETURN_EXIT;
 497         }
 498
 499         /*
 500          * Someone triggered glimpse within 1 sec before.
 501          * 1) The former glimpse succeeded with glimpse lock granted by OST, and
 502          *    if the lock is still cached on client, AGL needs to do nothing. If
 503          *    it is cancelled by other client, AGL maybe cannot obtaion new lock
 504          *    for no glimpse callback triggered by AGL.
 505          * 2) The former glimpse succeeded, but OST did not grant glimpse lock.
 506          *    Under such case, it is quite possible that the OST will not grant
 507          *    glimpse lock for AGL also.
 508          * 3) The former glimpse failed, compared with other two cases, it is
 509          *    relative rare. AGL can ignore such case, and it will not muchly
 510          *    affect the performance.
 511          */
 512         if (lli->lli_glimpse_time != 0 &&
 513             cfs_time_before(cfs_time_shift(-1), lli->lli_glimpse_time)) {
 514                 up_write(&lli->lli_glimpse_sem);
 515                 lli->lli_agl_index = 0;
 516                 iput(inode);
 517                 RETURN_EXIT;
 518         }
 519
 520         CDEBUG(D_READA, "Handling (init) async glimpse: inode = "
 521                DFID", idx = "LPU64"\n", PFID(&lli->lli_fid), index);
 522
 523         cl_agl(inode);
 524         lli->lli_agl_index = 0;
 525         lli->lli_glimpse_time = cfs_time_current();
 526         up_write(&lli->lli_glimpse_sem);
 527
 528         CDEBUG(D_READA, "Handled (init) async glimpse: inode= "
 529                DFID", idx = "LPU64", rc = %d\n",
 530                PFID(&lli->lli_fid), index, rc);
 531
 532         iput(inode);
 533
 534         EXIT;
 535 }
 536
 537 /* prepare inode for sa entry, add it into agl list, now sa_entry is ready
 538  * to be used by scanner process. */
 539 static void sa_instantiate(struct ll_statahead_info *sai,
 540                                  struct sa_entry *entry)
 541 {
 542         struct inode *dir = sai->sai_inode;
 543         struct inode *child;
 544         struct md_enqueue_info *minfo;
 545         struct lookup_intent *it;
 546         struct ptlrpc_request *req;
 547         struct mdt_body *body;
 548         int rc = 0;
 549         ENTRY;
 550
 551         LASSERT(entry->se_handle != 0);
 552
 553         minfo = entry->se_minfo;
 554         it = &minfo->mi_it;
 555         req = entry->se_req;
 556         body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
 557         if (body == NULL)
 558                 GOTO(out, rc = -EFAULT);
 559
 560         child = entry->se_inode;
 561         if (child == NULL) {
 562                 /*
 563                  * lookup.
 564                  */
 565                 LASSERT(fid_is_zero(&minfo->mi_data.op_fid2));
 566
 567                 /* XXX: No fid in reply, this is probaly cross-ref case.
 568                  * SA can't handle it yet. */
 569                 if (body->mbo_valid & OBD_MD_MDS)
 570                         GOTO(out, rc = -EAGAIN);
 571         } else {
 572                 /*
 573                  * revalidate.
 574                  */
 575                 /* unlinked and re-created with the same name */
 576                 if (unlikely(!lu_fid_eq(&minfo->mi_data.op_fid2,
 577                                         &body->mbo_fid1))) {
 578                         entry->se_inode = NULL;
 579                         iput(child);
 580                         child = NULL;
 581                 }
 582         }
 583
 584         it->d.lustre.it_lock_handle = entry->se_handle;
 585         rc = md_revalidate_lock(ll_i2mdexp(dir), it, ll_inode2fid(dir), NULL);
 586         if (rc != 1)
 587                 GOTO(out, rc = -EAGAIN);
 588
 589         rc = ll_prep_inode(&child, req, dir->i_sb, it);
 590         if (rc)
 591                 GOTO(out, rc);
 592
 593         CDEBUG(D_READA, "%s: setting %.*s"DFID" l_data to inode %p\n",
 594                ll_get_fsname(child->i_sb, NULL, 0),
 595                entry->se_qstr.len, entry->se_qstr.name,
 596                PFID(ll_inode2fid(child)), child);
 597         ll_set_lock_data(ll_i2sbi(dir)->ll_md_exp, child, it, NULL);
 598
 599         entry->se_inode = child;
 600
 601         if (agl_should_run(sai, child))
 602                 ll_agl_add(sai, child, entry->se_index);
 603
 604         EXIT;
 605
 606 out:
 607         /* sa_make_ready() will drop ldlm ibits lock refcount by calling
 608          * ll_intent_drop_lock() in spite of failures. Do not worry about
 609          * calling ll_intent_drop_lock() more than once. */
 610         sa_make_ready(sai, entry, rc);
 611 }
 612
 613 /* once there are async stat replies, instantiate sa_entry */
 614 static void sa_handle_callback(struct ll_statahead_info *sai)
 615 {
 616         struct ll_inode_info *lli;
 617
 618         lli = ll_i2info(sai->sai_inode);
 619
 620         while (sa_has_callback(sai)) {
 621                 struct sa_entry *entry;
 622
 623                 spin_lock(&lli->lli_sa_lock);
 624                 if (unlikely(!sa_has_callback(sai))) {
 625                         spin_unlock(&lli->lli_sa_lock);
 626                         break;
 627                 }
 628                 entry = list_entry(sai->sai_interim_entries.next,
 629                                    struct sa_entry, se_list);
 630                 list_del_init(&entry->se_list);
 631                 spin_unlock(&lli->lli_sa_lock);
 632
 633                 sa_instantiate(sai, entry);
 634         }
 635
 636         spin_lock(&lli->lli_agl_lock);
 637         while (!agl_list_empty(sai)) {
 638                 struct ll_inode_info *clli;
 639
 640                 clli = agl_first_entry(sai);
 641                 list_del_init(&clli->lli_agl_list);
 642                 spin_unlock(&lli->lli_agl_lock);
 643
 644                 ll_agl_trigger(&clli->lli_vfs_inode, sai);
 645
 646                 spin_lock(&lli->lli_agl_lock);
 647         }
 648         spin_unlock(&lli->lli_agl_lock);
 649 }
 650
 651 /* callback for async stat, because this is called in ptlrpcd context, we only
 652  * put sa_entry in sai_cb_entries list, and let sa_handle_callback() to really
 653  * prepare inode and instantiate sa_entry later. */
 654 static int ll_statahead_interpret(struct ptlrpc_request *req,
 655                                   struct md_enqueue_info *minfo, int rc)
 656 {
 657         struct lookup_intent *it = &minfo->mi_it;
 658         struct inode *dir = minfo->mi_dir;
 659         struct ll_inode_info *lli = ll_i2info(dir);
 660         struct ll_statahead_info *sai = lli->lli_sai;
 661         struct sa_entry *entry = (struct sa_entry *)minfo->mi_cbdata;
 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         spin_lock(&lli->lli_sa_lock);
 678         if (rc != 0) {
 679                 wakeup = __sa_make_ready(sai, entry, rc);
 680         } else {
 681                 entry->se_minfo = minfo;
 682                 entry->se_req = ptlrpc_request_addref(req);
 683                 /* Release the async ibits lock ASAP to avoid deadlock
 684                  * when statahead thread tries to enqueue lock on parent
 685                  * for readpage and other tries to enqueue lock on child
 686                  * with parent's lock held, for example: unlink. */
 687                 entry->se_handle = it->d.lustre.it_lock_handle;
 688                 ll_intent_drop_lock(it);
 689                 wakeup = !sa_has_callback(sai);
 690                 list_add_tail(&entry->se_list, &sai->sai_interim_entries);
 691         }
 692         sai->sai_replied++;
 693         if (wakeup)
 694                 wake_up(&sai->sai_thread.t_ctl_waitq);
 695         spin_unlock(&lli->lli_sa_lock);
 696
 697         if (rc != 0) {
 698                 ll_intent_release(it);
 699                 iput(dir);
 700                 OBD_FREE_PTR(minfo);
 701         }
 702         RETURN(rc);
 703 }
 704
 705 static void sa_fini_data(struct md_enqueue_info *minfo,
 706                          struct ldlm_enqueue_info *einfo)
 707 {
 708         LASSERT(minfo && einfo);
 709         iput(minfo->mi_dir);
 710         capa_put(minfo->mi_data.op_capa1);
 711         capa_put(minfo->mi_data.op_capa2);
 712         OBD_FREE_PTR(minfo);
 713         OBD_FREE_PTR(einfo);
 714 }
 715
 716 /**
 717  * There is race condition between "capa_put" and "ll_statahead_interpret" for
 718  * accessing "op_data.op_capa[1,2]" as following:
 719  * "capa_put" releases "op_data.op_capa[1,2]"'s reference count after calling
 720  * "md_intent_getattr_async". But "ll_statahead_interpret" maybe run first, and
 721  * fill "op_data.op_capa[1,2]" as POISON, then cause "capa_put" access invalid
 722  * "ocapa". So here reserve "op_data.op_capa[1,2]" in "pcapa" before calling
 723  * "md_intent_getattr_async".
 724  */
 725 static int sa_prep_data(struct inode *dir, struct inode *child,
 726                         struct sa_entry *entry, struct md_enqueue_info **pmi,
 727                         struct ldlm_enqueue_info **pei,
 728                         struct obd_capa **pcapa)
 729 {
 730         struct qstr              *qstr = &entry->se_qstr;
 731         struct md_enqueue_info   *minfo;
 732         struct ldlm_enqueue_info *einfo;
 733         struct md_op_data        *op_data;
 734
 735         OBD_ALLOC_PTR(einfo);
 736         if (einfo == NULL)
 737                 return -ENOMEM;
 738
 739         OBD_ALLOC_PTR(minfo);
 740         if (minfo == NULL) {
 741                 OBD_FREE_PTR(einfo);
 742                 return -ENOMEM;
 743         }
 744
 745         op_data = ll_prep_md_op_data(&minfo->mi_data, dir, child, qstr->name,
 746                                      qstr->len, 0, LUSTRE_OPC_ANY, NULL);
 747         if (IS_ERR(op_data)) {
 748                 OBD_FREE_PTR(einfo);
 749                 OBD_FREE_PTR(minfo);
 750                 return PTR_ERR(op_data);
 751         }
 752
 753         minfo->mi_it.it_op = IT_GETATTR;
 754         minfo->mi_dir = igrab(dir);
 755         minfo->mi_cb = ll_statahead_interpret;
 756         minfo->mi_cbdata = entry;
 757
 758         einfo->ei_type   = LDLM_IBITS;
 759         einfo->ei_mode   = it_to_lock_mode(&minfo->mi_it);
 760         einfo->ei_cb_bl  = ll_md_blocking_ast;
 761         einfo->ei_cb_cp  = ldlm_completion_ast;
 762         einfo->ei_cb_gl  = NULL;
 763         einfo->ei_cbdata = NULL;
 764
 765         *pmi = minfo;
 766         *pei = einfo;
 767         pcapa[0] = op_data->op_capa1;
 768         pcapa[1] = op_data->op_capa2;
 769
 770         return 0;
 771 }
 772
 773 static int sa_lookup(struct inode *dir, struct sa_entry *entry)
 774 {
 775         struct md_enqueue_info   *minfo;
 776         struct ldlm_enqueue_info *einfo;
 777         struct obd_capa          *capas[2];
 778         int                       rc;
 779         ENTRY;
 780
 781         rc = sa_prep_data(dir, NULL, entry, &minfo, &einfo, capas);
 782         if (rc)
 783                 RETURN(rc);
 784
 785         rc = md_intent_getattr_async(ll_i2mdexp(dir), minfo, einfo);
 786         if (!rc) {
 787                 capa_put(capas[0]);
 788                 capa_put(capas[1]);
 789         } else {
 790                 sa_fini_data(minfo, einfo);
 791         }
 792
 793         RETURN(rc);
 794 }
 795
 796 /**
 797  * similar to ll_revalidate_it().
 798  * \retval      1 -- dentry valid
 799  * \retval      0 -- will send stat-ahead request
 800  * \retval others -- prepare stat-ahead request failed
 801  */
 802 static int sa_revalidate(struct inode *dir, struct sa_entry *entry,
 803                          struct dentry *dentry)
 804 {
 805         struct inode *inode = dentry->d_inode;
 806         struct lookup_intent it = { .it_op = IT_GETATTR,
 807                                     .d.lustre.it_lock_handle = 0 };
 808         struct md_enqueue_info *minfo;
 809         struct ldlm_enqueue_info *einfo;
 810         struct obd_capa *capas[2];
 811         int rc;
 812         ENTRY;
 813
 814         if (unlikely(inode == NULL))
 815                 RETURN(1);
 816
 817         if (d_mountpoint(dentry))
 818                 RETURN(1);
 819
 820         entry->se_inode = igrab(inode);
 821         rc = md_revalidate_lock(ll_i2mdexp(dir), &it, ll_inode2fid(inode),
 822                                 NULL);
 823         if (rc == 1) {
 824                 entry->se_handle = it.d.lustre.it_lock_handle;
 825                 ll_intent_release(&it);
 826                 RETURN(1);
 827         }
 828
 829         rc = sa_prep_data(dir, inode, entry, &minfo, &einfo, capas);
 830         if (rc) {
 831                 entry->se_inode = NULL;
 832                 iput(inode);
 833                 RETURN(rc);
 834         }
 835
 836         rc = md_intent_getattr_async(ll_i2mdexp(dir), minfo, einfo);
 837         if (!rc) {
 838                 capa_put(capas[0]);
 839                 capa_put(capas[1]);
 840         } else {
 841                 entry->se_inode = NULL;
 842                 iput(inode);
 843                 sa_fini_data(minfo, einfo);
 844         }
 845
 846         RETURN(rc);
 847 }
 848
 849 static void sa_statahead(struct dentry *parent, const char *name, int len)
 850 {
 851         struct inode *dir = parent->d_inode;
 852         struct ll_inode_info *lli = ll_i2info(dir);
 853         struct ll_statahead_info *sai = lli->lli_sai;
 854         struct dentry *dentry = NULL;
 855         struct sa_entry *entry;
 856         int rc;
 857         ENTRY;
 858
 859         entry = sa_alloc(sai, sai->sai_index, name, len);
 860         if (IS_ERR(entry))
 861                 RETURN_EXIT;
 862
 863         dentry = d_lookup(parent, &entry->se_qstr);
 864         if (!dentry) {
 865                 rc = sa_lookup(dir, entry);
 866         } else {
 867                 rc = sa_revalidate(dir, entry, dentry);
 868                 if (rc == 1 && agl_should_run(sai, dentry->d_inode))
 869                         ll_agl_add(sai, dentry->d_inode, entry->se_index);
 870         }
 871
 872         if (dentry != NULL)
 873                 dput(dentry);
 874
 875         if (rc != 0)
 876                 sa_make_ready(sai, entry, rc);
 877         else
 878                 sai->sai_sent++;
 879
 880         sai->sai_index++;
 881
 882         EXIT;
 883 }
 884
 885 static int ll_agl_thread(void *arg)
 886 {
 887         struct dentry *parent = (struct dentry *)arg;
 888         struct inode *dir = parent->d_inode;
 889         struct ll_inode_info *plli = ll_i2info(dir);
 890         struct ll_inode_info *clli;
 891         struct ll_sb_info *sbi = ll_i2sbi(dir);
 892         struct ll_statahead_info *sai;
 893         struct ptlrpc_thread *thread;
 894         struct l_wait_info lwi = { 0 };
 895         ENTRY;
 896
 897
 898         sai = ll_sai_get(dir);
 899         thread = &sai->sai_agl_thread;
 900         thread->t_pid = current_pid();
 901         CDEBUG(D_READA, "agl thread started: sai %p, parent %.*s\n",
 902                sai, parent->d_name.len, parent->d_name.name);
 903
 904         atomic_inc(&sbi->ll_agl_total);
 905         spin_lock(&plli->lli_agl_lock);
 906         sai->sai_agl_valid = 1;
 907         if (thread_is_init(thread))
 908                 /* If someone else has changed the thread state
 909                  * (e.g. already changed to SVC_STOPPING), we can't just
 910                  * blindly overwrite that setting. */
 911                 thread_set_flags(thread, SVC_RUNNING);
 912         spin_unlock(&plli->lli_agl_lock);
 913         wake_up(&thread->t_ctl_waitq);
 914
 915         while (1) {
 916                 l_wait_event(thread->t_ctl_waitq,
 917                              !agl_list_empty(sai) ||
 918                              !thread_is_running(thread),
 919                              &lwi);
 920
 921                 if (!thread_is_running(thread))
 922                         break;
 923
 924                 spin_lock(&plli->lli_agl_lock);
 925                 /* The statahead thread maybe help to process AGL entries,
 926                  * so check whether list empty again. */
 927                 if (!agl_list_empty(sai)) {
 928                         clli = agl_first_entry(sai);
 929                         list_del_init(&clli->lli_agl_list);
 930                         spin_unlock(&plli->lli_agl_lock);
 931                         ll_agl_trigger(&clli->lli_vfs_inode, sai);
 932                 } else {
 933                         spin_unlock(&plli->lli_agl_lock);
 934                 }
 935         }
 936
 937         spin_lock(&plli->lli_agl_lock);
 938         sai->sai_agl_valid = 0;
 939         while (!agl_list_empty(sai)) {
 940                 clli = agl_first_entry(sai);
 941                 list_del_init(&clli->lli_agl_list);
 942                 spin_unlock(&plli->lli_agl_lock);
 943                 clli->lli_agl_index = 0;
 944                 iput(&clli->lli_vfs_inode);
 945                 spin_lock(&plli->lli_agl_lock);
 946         }
 947         thread_set_flags(thread, SVC_STOPPED);
 948         spin_unlock(&plli->lli_agl_lock);
 949         wake_up(&thread->t_ctl_waitq);
 950         ll_sai_put(sai);
 951         CDEBUG(D_READA, "agl thread stopped: sai %p, parent %.*s\n",
 952                sai, parent->d_name.len, parent->d_name.name);
 953         RETURN(0);
 954 }
 955
 956 static void ll_start_agl(struct dentry *parent, struct ll_statahead_info *sai)
 957 {
 958         struct ptlrpc_thread *thread = &sai->sai_agl_thread;
 959         struct l_wait_info    lwi    = { 0 };
 960         struct ll_inode_info  *plli;
 961         struct task_struct            *task;
 962         ENTRY;
 963
 964         CDEBUG(D_READA, "start agl thread: sai %p, parent %.*s\n",
 965                sai, parent->d_name.len, parent->d_name.name);
 966
 967         plli = ll_i2info(parent->d_inode);
 968         task = kthread_run(ll_agl_thread, parent,
 969                                "ll_agl_%u", plli->lli_opendir_pid);
 970         if (IS_ERR(task)) {
 971                 CERROR("can't start ll_agl thread, rc: %ld\n", PTR_ERR(task));
 972                 thread_set_flags(thread, SVC_STOPPED);
 973                 RETURN_EXIT;
 974         }
 975
 976         l_wait_event(thread->t_ctl_waitq,
 977                      thread_is_running(thread) || thread_is_stopped(thread),
 978                      &lwi);
 979         EXIT;
 980 }
 981
 982 static int ll_statahead_thread(void *arg)
 983 {
 984         struct dentry *parent = (struct dentry *)arg;
 985         struct inode *dir = parent->d_inode;
 986         struct ll_inode_info *lli = ll_i2info(dir);
 987         struct ll_sb_info *sbi = ll_i2sbi(dir);
 988         struct ll_statahead_info *sai;
 989         struct ptlrpc_thread *thread;
 990         struct ptlrpc_thread *agl_thread;
 991         int first = 0;
 992         struct md_op_data *op_data;
 993         struct ll_dir_chain chain;
 994         struct l_wait_info lwi = { 0 };
 995         struct page *page = NULL;
 996         __u64 pos = 0;
 997         int rc = 0;
 998         ENTRY;
 999
1000         sai = ll_sai_get(dir);
1001         thread = &sai->sai_thread;
1002         agl_thread = &sai->sai_agl_thread;
1003         thread->t_pid = current_pid();
1004         CDEBUG(D_READA, "statahead thread starting: sai %p, parent %.*s\n",
1005                sai, parent->d_name.len, parent->d_name.name);
1006
1007         op_data = ll_prep_md_op_data(NULL, dir, dir, NULL, 0, 0,
1008                                      LUSTRE_OPC_ANY, dir);
1009         if (IS_ERR(op_data))
1010                 GOTO(out, rc = PTR_ERR(op_data));
1011
1012         op_data->op_max_pages = ll_i2sbi(dir)->ll_md_brw_pages;
1013
1014         if (sbi->ll_flags & LL_SBI_AGL_ENABLED)
1015                 ll_start_agl(parent, sai);
1016
1017         atomic_inc(&sbi->ll_sa_total);
1018         spin_lock(&lli->lli_sa_lock);
1019         if (thread_is_init(thread))
1020                 /* If someone else has changed the thread state
1021                  * (e.g. already changed to SVC_STOPPING), we can't just
1022                  * blindly overwrite that setting. */
1023                 thread_set_flags(thread, SVC_RUNNING);
1024         spin_unlock(&lli->lli_sa_lock);
1025         wake_up(&thread->t_ctl_waitq);
1026
1027         ll_dir_chain_init(&chain);
1028         while (pos != MDS_DIR_END_OFF && thread_is_running(thread)) {
1029                 struct lu_dirpage *dp;
1030                 struct lu_dirent  *ent;
1031
1032                 sai->sai_in_readpage = 1;
1033                 page = ll_get_dir_page(dir, op_data, pos, &chain);
1034                 sai->sai_in_readpage = 0;
1035                 if (IS_ERR(page)) {
1036                         rc = PTR_ERR(page);
1037                         CDEBUG(D_READA, "error reading dir "DFID" at "LPU64
1038                                "/"LPU64" opendir_pid = %u: rc = %d\n",
1039                                PFID(ll_inode2fid(dir)), pos, sai->sai_index,
1040                                lli->lli_opendir_pid, rc);
1041                         break;
1042                 }
1043
1044                 dp = page_address(page);
1045                 for (ent = lu_dirent_start(dp);
1046                      ent != NULL && thread_is_running(thread) &&
1047                      !sa_low_hit(sai);
1048                      ent = lu_dirent_next(ent)) {
1049                         __u64 hash;
1050                         int namelen;
1051                         char *name;
1052
1053                         hash = le64_to_cpu(ent->lde_hash);
1054                         if (unlikely(hash < pos))
1055                                 /*
1056                                  * Skip until we find target hash value.
1057                                  */
1058                                 continue;
1059
1060                         namelen = le16_to_cpu(ent->lde_namelen);
1061                         if (unlikely(namelen == 0))
1062                                 /*
1063                                  * Skip dummy record.
1064                                  */
1065                                 continue;
1066
1067                         name = ent->lde_name;
1068                         if (name[0] == '.') {
1069                                 if (namelen == 1) {
1070                                         /*
1071                                          * skip "."
1072                                          */
1073                                         continue;
1074                                 } else if (name[1] == '.' && namelen == 2) {
1075                                         /*
1076                                          * skip ".."
1077                                          */
1078                                         continue;
1079                                 } else if (!sai->sai_ls_all) {
1080                                         /*
1081                                          * skip hidden files.
1082                                          */
1083                                         sai->sai_skip_hidden++;
1084                                         continue;
1085                                 }
1086                         }
1087
1088                         /*
1089                          * don't stat-ahead first entry.
1090                          */
1091                         if (unlikely(++first == 1))
1092                                 continue;
1093
1094                         /* wait for spare statahead window */
1095                         do {
1096                                 l_wait_event(thread->t_ctl_waitq,
1097                                              !sa_sent_full(sai) ||
1098                                              sa_has_callback(sai) ||
1099                                              !agl_list_empty(sai) ||
1100                                              !thread_is_running(thread),
1101                                              &lwi);
1102
1103                                 sa_handle_callback(sai);
1104                         } while (sa_sent_full(sai) &&
1105                                  thread_is_running(thread));
1106
1107                         sa_statahead(parent, name, namelen);
1108                 }
1109
1110                 pos = le64_to_cpu(dp->ldp_hash_end);
1111                 ll_release_page(dir, page,
1112                                 le32_to_cpu(dp->ldp_flags) & LDF_COLLIDE);
1113
1114                 if (sa_low_hit(sai)) {
1115                         rc = -EFAULT;
1116                         atomic_inc(&sbi->ll_sa_wrong);
1117                         CDEBUG(D_READA, "Statahead for dir "DFID" hit "
1118                                "ratio too low: hit/miss "LPU64"/"LPU64
1119                                ", sent/replied "LPU64"/"LPU64", stopping "
1120                                "statahead thread: pid %d\n",
1121                                PFID(&lli->lli_fid), sai->sai_hit,
1122                                sai->sai_miss, sai->sai_sent,
1123                                sai->sai_replied, current_pid());
1124                         break;
1125                 }
1126         }
1127         ll_dir_chain_fini(&chain);
1128         ll_finish_md_op_data(op_data);
1129
1130         if (rc < 0) {
1131                 spin_lock(&lli->lli_sa_lock);
1132                 thread_set_flags(thread, SVC_STOPPING);
1133                 lli->lli_sa_enabled = 0;
1134                 spin_unlock(&lli->lli_sa_lock);
1135         }
1136
1137         /* statahead is finished, but statahead entries need to be cached, wait
1138          * for file release to stop me. */
1139         while (thread_is_running(thread)) {
1140                 l_wait_event(thread->t_ctl_waitq,
1141                              sa_has_callback(sai) ||
1142                              !agl_list_empty(sai) ||
1143                              !thread_is_running(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(HZ >> 3, NULL, NULL);
1172                 l_wait_event(thread->t_ctl_waitq,
1173                         sai->sai_sent == sai->sai_replied, &lwi);
1174         }
1175
1176         /* release resources held by statahead RPCs */
1177         sa_handle_callback(sai);
1178
1179         spin_lock(&lli->lli_sa_lock);
1180         thread_set_flags(thread, SVC_STOPPED);
1181         spin_unlock(&lli->lli_sa_lock);
1182
1183         wake_up(&sai->sai_waitq);
1184         wake_up(&thread->t_ctl_waitq);
1185         ll_sai_put(sai);
1186         CDEBUG(D_READA, "statahead thread stopped: sai %p, parent %.*s\n",
1187                sai, parent->d_name.len, parent->d_name.name);
1188         dput(parent);
1189         return rc;
1190 }
1191
1192 /* authorize opened dir handle @key to statahead later */
1193 void ll_authorize_statahead(struct inode *dir, void *key)
1194 {
1195         struct ll_inode_info *lli = ll_i2info(dir);
1196
1197         spin_lock(&lli->lli_sa_lock);
1198         if (lli->lli_opendir_key == NULL && lli->lli_sai == NULL) {
1199                 /*
1200                  * if lli_sai is not NULL, it means previous statahead is not
1201                  * finished yet, we'd better not start a new statahead for now.
1202                  */
1203                 LASSERT(lli->lli_opendir_pid == 0);
1204                 lli->lli_opendir_key = key;
1205                 lli->lli_opendir_pid = current_pid();
1206                 lli->lli_sa_enabled = 1;
1207         }
1208         spin_unlock(&lli->lli_sa_lock);
1209 }
1210
1211 /* deauthorize opened dir handle @key to statahead, but statahead thread may
1212  * still be running, notify it to quit. */
1213 void ll_deauthorize_statahead(struct inode *dir, void *key)
1214 {
1215         struct ll_inode_info *lli = ll_i2info(dir);
1216         struct ll_statahead_info *sai;
1217
1218         LASSERT(lli->lli_opendir_key == key);
1219         LASSERT(lli->lli_opendir_pid != 0);
1220
1221         CDEBUG(D_READA, "deauthorize statahead for "DFID"\n",
1222                 PFID(&lli->lli_fid));
1223
1224         spin_lock(&lli->lli_sa_lock);
1225         lli->lli_opendir_key = NULL;
1226         lli->lli_opendir_pid = 0;
1227         lli->lli_sa_enabled = 0;
1228         sai = lli->lli_sai;
1229         if (sai != NULL && thread_is_running(&sai->sai_thread)) {
1230                 /*
1231                  * statahead thread may not quit yet because it needs to cache
1232                  * entries, now it's time to tell it to quit.
1233                  */
1234                 thread_set_flags(&sai->sai_thread, SVC_STOPPING);
1235                 wake_up(&sai->sai_thread.t_ctl_waitq);
1236         }
1237         spin_unlock(&lli->lli_sa_lock);
1238 }
1239
1240 enum {
1241         /**
1242          * not first dirent, or is "."
1243          */
1244         LS_NONE_FIRST_DE = 0,
1245         /**
1246          * the first non-hidden dirent
1247          */
1248         LS_FIRST_DE,
1249         /**
1250          * the first hidden dirent, that is "."
1251          */
1252         LS_FIRST_DOT_DE
1253 };
1254
1255 static int is_first_dirent(struct inode *dir, struct dentry *dentry)
1256 {
1257         struct ll_dir_chain   chain;
1258         struct qstr          *target = &dentry->d_name;
1259         struct md_op_data    *op_data;
1260         int                   dot_de;
1261         struct page          *page = NULL;
1262         int                   rc     = LS_NONE_FIRST_DE;
1263         __u64                 pos = 0;
1264         ENTRY;
1265
1266         op_data = ll_prep_md_op_data(NULL, dir, dir, NULL, 0, 0,
1267                                      LUSTRE_OPC_ANY, dir);
1268         if (IS_ERR(op_data))
1269                 RETURN(PTR_ERR(op_data));
1270         /**
1271          *FIXME choose the start offset of the readdir
1272          */
1273         op_data->op_stripe_offset = 0;
1274         op_data->op_max_pages = ll_i2sbi(dir)->ll_md_brw_pages;
1275
1276         ll_dir_chain_init(&chain);
1277         page = ll_get_dir_page(dir, op_data, 0, &chain);
1278
1279         while (1) {
1280                 struct lu_dirpage *dp;
1281                 struct lu_dirent  *ent;
1282
1283                 if (IS_ERR(page)) {
1284                         struct ll_inode_info *lli = ll_i2info(dir);
1285
1286                         rc = PTR_ERR(page);
1287                         CERROR("%s: reading dir "DFID" at "LPU64
1288                                "opendir_pid = %u : rc = %d\n",
1289                                ll_get_fsname(dir->i_sb, NULL, 0),
1290                                PFID(ll_inode2fid(dir)), pos,
1291                                lli->lli_opendir_pid, rc);
1292                         break;
1293                 }
1294
1295                 dp = page_address(page);
1296                 for (ent = lu_dirent_start(dp); ent != NULL;
1297                      ent = lu_dirent_next(ent)) {
1298                         __u64 hash;
1299                         int namelen;
1300                         char *name;
1301
1302                         hash = le64_to_cpu(ent->lde_hash);
1303                         /* The ll_get_dir_page() can return any page containing
1304                          * the given hash which may be not the start hash. */
1305                         if (unlikely(hash < pos))
1306                                 continue;
1307
1308                         namelen = le16_to_cpu(ent->lde_namelen);
1309                         if (unlikely(namelen == 0))
1310                                 /*
1311                                  * skip dummy record.
1312                                  */
1313                                 continue;
1314
1315                         name = ent->lde_name;
1316                         if (name[0] == '.') {
1317                                 if (namelen == 1)
1318                                         /*
1319                                          * skip "."
1320                                          */
1321                                         continue;
1322                                 else if (name[1] == '.' && namelen == 2)
1323                                         /*
1324                                          * skip ".."
1325                                          */
1326                                         continue;
1327                                 else
1328                                         dot_de = 1;
1329                         } else {
1330                                 dot_de = 0;
1331                         }
1332
1333                         if (dot_de && target->name[0] != '.') {
1334                                 CDEBUG(D_READA, "%.*s skip hidden file %.*s\n",
1335                                        target->len, target->name,
1336                                        namelen, name);
1337                                 continue;
1338                         }
1339
1340                         if (target->len != namelen ||
1341                             memcmp(target->name, name, namelen) != 0)
1342                                 rc = LS_NONE_FIRST_DE;
1343                         else if (!dot_de)
1344                                 rc = LS_FIRST_DE;
1345                         else
1346                                 rc = LS_FIRST_DOT_DE;
1347
1348                         ll_release_page(dir, page, false);
1349                         GOTO(out, rc);
1350                 }
1351                 pos = le64_to_cpu(dp->ldp_hash_end);
1352                 if (pos == MDS_DIR_END_OFF) {
1353                         /*
1354                          * End of directory reached.
1355                          */
1356                         ll_release_page(dir, page, false);
1357                         GOTO(out, rc);
1358                 } else {
1359                         /*
1360                          * chain is exhausted
1361                          * Normal case: continue to the next page.
1362                          */
1363                         ll_release_page(dir, page, le32_to_cpu(dp->ldp_flags) &
1364                                               LDF_COLLIDE);
1365                         page = ll_get_dir_page(dir, op_data, pos, &chain);
1366                 }
1367         }
1368         EXIT;
1369 out:
1370         ll_dir_chain_fini(&chain);
1371         ll_finish_md_op_data(op_data);
1372         return rc;
1373 }
1374
1375 static int revalidate_statahead_dentry(struct inode *dir,
1376                                         struct ll_statahead_info *sai,
1377                                         struct dentry **dentryp,
1378                                         int only_unplug)
1379 {
1380         struct sa_entry *entry = NULL;
1381         struct l_wait_info lwi = { 0 };
1382         int rc = 0;
1383         ENTRY;
1384
1385         if ((*dentryp)->d_name.name[0] == '.') {
1386                 if (sai->sai_ls_all ||
1387                     sai->sai_miss_hidden >= sai->sai_skip_hidden) {
1388                         /*
1389                          * Hidden dentry is the first one, or statahead
1390                          * thread does not skip so many hidden dentries
1391                          * before "sai_ls_all" enabled as below.
1392                          */
1393                 } else {
1394                         if (!sai->sai_ls_all)
1395                                 /*
1396                                  * It maybe because hidden dentry is not
1397                                  * the first one, "sai_ls_all" was not
1398                                  * set, then "ls -al" missed. Enable
1399                                  * "sai_ls_all" for such case.
1400                                  */
1401                                 sai->sai_ls_all = 1;
1402
1403                         /*
1404                          * Such "getattr" has been skipped before
1405                          * "sai_ls_all" enabled as above.
1406                          */
1407                         sai->sai_miss_hidden++;
1408                         RETURN(-EAGAIN);
1409                 }
1410         }
1411
1412         entry = sa_get(sai, &(*dentryp)->d_name);
1413         if (entry == NULL || only_unplug) {
1414                 sa_put(sai, entry);
1415                 RETURN(entry ? 1 : -EAGAIN);
1416         }
1417
1418         /* if statahead is busy in readdir, help it do post-work */
1419         if (!sa_ready(entry) && sai->sai_in_readpage)
1420                 sa_handle_callback(sai);
1421
1422         if (!sa_ready(entry)) {
1423                 sai->sai_index_wait = entry->se_index;
1424                 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(30), NULL,
1425                                         LWI_ON_SIGNAL_NOOP, NULL);
1426                 rc = l_wait_event(sai->sai_waitq,
1427                                 sa_ready(entry) ||
1428                                 thread_is_stopped(&sai->sai_thread),
1429                                 &lwi);
1430                 if (rc < 0) {
1431                         sa_put(sai, entry);
1432                         RETURN(-EAGAIN);
1433                 }
1434         }
1435
1436         if (entry->se_state == SA_ENTRY_SUCC && entry->se_inode != NULL) {
1437                 struct inode *inode = entry->se_inode;
1438                 struct lookup_intent it = { .it_op = IT_GETATTR,
1439                                             .d.lustre.it_lock_handle =
1440                                                 entry->se_handle };
1441                 __u64 bits;
1442
1443                 rc = md_revalidate_lock(ll_i2mdexp(dir), &it,
1444                                         ll_inode2fid(inode), &bits);
1445                 if (rc == 1) {
1446                         if ((*dentryp)->d_inode == NULL) {
1447                                 struct dentry *alias;
1448
1449                                 alias = ll_splice_alias(inode, *dentryp);
1450                                 if (IS_ERR(alias)) {
1451                                         sa_put(sai, entry);
1452                                         RETURN(PTR_ERR(alias));
1453                                 }
1454                                 *dentryp = alias;
1455                         } else if ((*dentryp)->d_inode != inode) {
1456                                 /* revalidate, but inode is recreated */
1457                                 CDEBUG(D_READA,
1458                                         "%s: stale dentry %.*s inode "
1459                                         DFID", statahead inode "DFID
1460                                         "\n",
1461                                         ll_get_fsname((*dentryp)->d_inode->i_sb,
1462                                                       NULL, 0),
1463                                         (*dentryp)->d_name.len,
1464                                         (*dentryp)->d_name.name,
1465                                         PFID(ll_inode2fid((*dentryp)->d_inode)),
1466                                         PFID(ll_inode2fid(inode)));
1467                                 sa_put(sai, entry);
1468                                 RETURN(-ESTALE);
1469                         } else {
1470                                 iput(inode);
1471                         }
1472                         entry->se_inode = NULL;
1473
1474                         if ((bits & MDS_INODELOCK_LOOKUP) &&
1475                             d_lustre_invalid(*dentryp))
1476                                 d_lustre_revalidate(*dentryp);
1477                         ll_intent_release(&it);
1478                 }
1479         }
1480
1481         sa_put(sai, entry);
1482         RETURN(rc);
1483 }
1484
1485 static int start_statahead_thread(struct inode *dir, struct dentry *dentry)
1486 {
1487         struct ll_inode_info *lli = ll_i2info(dir);
1488         struct ll_statahead_info *sai = NULL;
1489         struct dentry *parent;
1490         struct ptlrpc_thread *thread;
1491         struct l_wait_info lwi = { 0 };
1492         struct task_struct *task;
1493         int rc;
1494         ENTRY;
1495
1496         /* I am the "lli_opendir_pid" owner, only me can set "lli_sai". */
1497         rc = is_first_dirent(dir, dentry);
1498         if (rc == LS_NONE_FIRST_DE)
1499                 /* It is not "ls -{a}l" operation, no need statahead for it. */
1500                 GOTO(out, rc = -EAGAIN);
1501
1502         sai = ll_sai_alloc();
1503         if (sai == NULL)
1504                 GOTO(out, rc = -ENOMEM);
1505
1506         sai->sai_ls_all = (rc == LS_FIRST_DOT_DE);
1507         sai->sai_inode = igrab(dir);
1508         if (unlikely(sai->sai_inode == NULL)) {
1509                 CWARN("Do not start stat ahead on dying inode "DFID"\n",
1510                         PFID(&lli->lli_fid));
1511                 GOTO(out, rc = -ESTALE);
1512         }
1513
1514         /* get parent reference count here, and put it in ll_statahead_thread */
1515         parent = dget(dentry->d_parent);
1516         if (unlikely(sai->sai_inode != parent->d_inode)) {
1517                 struct ll_inode_info *nlli = ll_i2info(parent->d_inode);
1518
1519                 CWARN("Race condition, someone changed %.*s just now: "
1520                         "old parent "DFID", new parent "DFID"\n",
1521                         dentry->d_name.len, dentry->d_name.name,
1522                         PFID(&lli->lli_fid), PFID(&nlli->lli_fid));
1523                 dput(parent);
1524                 iput(sai->sai_inode);
1525                 GOTO(out, rc = -EAGAIN);
1526         }
1527
1528         CDEBUG(D_READA, "start statahead thread: sai %p, parent %.*s\n",
1529                sai, parent->d_name.len, parent->d_name.name);
1530
1531         lli->lli_sai = sai;
1532
1533         task = kthread_run(ll_statahead_thread, parent, "ll_sa_%u",
1534                            lli->lli_opendir_pid);
1535         thread = &sai->sai_thread;
1536         if (IS_ERR(task)) {
1537                 rc = PTR_ERR(task);
1538                 CERROR("cannot start ll_sa thread: rc = %d\n", rc);
1539                 dput(parent);
1540                 lli->lli_opendir_key = NULL;
1541                 thread_set_flags(thread, SVC_STOPPED);
1542                 thread_set_flags(&sai->sai_agl_thread, SVC_STOPPED);
1543                 ll_sai_put(sai);
1544                 LASSERT(lli->lli_sai == NULL);
1545                 RETURN(-EAGAIN);
1546         }
1547
1548         l_wait_event(thread->t_ctl_waitq,
1549                      thread_is_running(thread) || thread_is_stopped(thread),
1550                      &lwi);
1551         atomic_inc(&ll_i2sbi(parent->d_inode)->ll_sa_running);
1552         ll_sai_put(sai);
1553
1554         /*
1555          * We don't stat-ahead for the first dirent since we are already in
1556          * lookup.
1557          */
1558         RETURN(-EAGAIN);
1559
1560 out:
1561         if (sai != NULL)
1562                 OBD_FREE_PTR(sai);
1563         spin_lock(&lli->lli_sa_lock);
1564         lli->lli_opendir_key = NULL;
1565         lli->lli_opendir_pid = 0;
1566         lli->lli_sa_enabled = 0;
1567         spin_unlock(&lli->lli_sa_lock);
1568
1569         RETURN(rc);
1570 }
1571
1572 /**
1573  * Start statahead thread if this is the first dir entry.
1574  * Otherwise if a thread is started already, wait it until it is ahead of me.
1575  * \retval 1       -- find entry with lock in cache, the caller needs to do
1576  *                    nothing.
1577  * \retval 0       -- find entry in cache, but without lock, the caller needs
1578  *                    refresh from MDS.
1579  * \retval others  -- the caller need to process as non-statahead.
1580  */
1581 int do_statahead_enter(struct inode *dir, struct dentry **dentryp,
1582                        int only_unplug)
1583 {
1584         struct ll_statahead_info *sai;
1585
1586         sai = ll_sai_get(dir);
1587         if (sai != NULL) {
1588                 int rc;
1589
1590                 rc = revalidate_statahead_dentry(dir, sai, dentryp,
1591                                                  only_unplug);
1592                 CDEBUG(D_READA, "revalidate statahead %.*s: %d.\n",
1593                         (*dentryp)->d_name.len, (*dentryp)->d_name.name, rc);
1594                 ll_sai_put(sai);
1595                 return rc;
1596         }
1597
1598         return start_statahead_thread(dir, *dentryp);
1599 }