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_lite.h>
  47 #include <lustre_dlm.h>
  48 #include "llite_internal.h"
  49
  50 #define SA_OMITTED_ENTRY_MAX 8ULL
  51
  52 typedef enum {
  53         /** negative values are for error cases */
  54         SA_ENTRY_INIT = 0,      /** init entry */
  55         SA_ENTRY_SUCC = 1,      /** stat succeed */
  56         SA_ENTRY_INVA = 2,      /** invalid entry */
  57         SA_ENTRY_DEST = 3,      /** entry to be destroyed */
  58 } se_stat_t;
  59
  60 struct ll_sa_entry {
  61         /* link into sai->sai_entries */
  62         cfs_list_t              se_link;
  63         /* link into sai->sai_entries_{received,stated} */
  64         cfs_list_t              se_list;
  65         /* link into sai hash table locally */
  66         cfs_list_t              se_hash;
  67         /* entry reference count */
  68         cfs_atomic_t            se_refcount;
  69         /* entry index in the sai */
  70         __u64                   se_index;
  71         /* low layer ldlm lock handle */
  72         __u64                   se_handle;
  73         /* entry status */
  74         se_stat_t               se_stat;
  75         /* entry size, contains name */
  76         int                     se_size;
  77         /* pointer to async getattr enqueue info */
  78         struct md_enqueue_info *se_minfo;
  79         /* pointer to the async getattr request */
  80         struct ptlrpc_request  *se_req;
  81         /* pointer to the target inode */
  82         struct inode           *se_inode;
  83         /* entry name */
  84         struct qstr             se_qstr;
  85 };
  86
  87 static unsigned int sai_generation = 0;
  88 static DEFINE_SPINLOCK(sai_generation_lock);
  89
  90 static inline int ll_sa_entry_unhashed(struct ll_sa_entry *entry)
  91 {
  92         return cfs_list_empty(&entry->se_hash);
  93 }
  94
  95 /*
  96  * The entry only can be released by the caller, it is necessary to hold lock.
  97  */
  98 static inline int ll_sa_entry_stated(struct ll_sa_entry *entry)
  99 {
 100         smp_rmb();
 101         return (entry->se_stat != SA_ENTRY_INIT);
 102 }
 103
 104 static inline int ll_sa_entry_hash(int val)
 105 {
 106         return val & LL_SA_CACHE_MASK;
 107 }
 108
 109 /*
 110  * Insert entry to hash SA table.
 111  */
 112 static inline void
 113 ll_sa_entry_enhash(struct ll_statahead_info *sai, struct ll_sa_entry *entry)
 114 {
 115         int i = ll_sa_entry_hash(entry->se_qstr.hash);
 116
 117         spin_lock(&sai->sai_cache_lock[i]);
 118         cfs_list_add_tail(&entry->se_hash, &sai->sai_cache[i]);
 119         spin_unlock(&sai->sai_cache_lock[i]);
 120 }
 121
 122 /*
 123  * Remove entry from SA table.
 124  */
 125 static inline void
 126 ll_sa_entry_unhash(struct ll_statahead_info *sai, struct ll_sa_entry *entry)
 127 {
 128         int i = ll_sa_entry_hash(entry->se_qstr.hash);
 129
 130         spin_lock(&sai->sai_cache_lock[i]);
 131         cfs_list_del_init(&entry->se_hash);
 132         spin_unlock(&sai->sai_cache_lock[i]);
 133 }
 134
 135 static inline int agl_should_run(struct ll_statahead_info *sai,
 136                                  struct inode *inode)
 137 {
 138         return (inode != NULL && S_ISREG(inode->i_mode) && sai->sai_agl_valid);
 139 }
 140
 141 static inline struct ll_sa_entry *
 142 sa_first_received_entry(struct ll_statahead_info *sai)
 143 {
 144         return cfs_list_entry(sai->sai_entries_received.next,
 145                               struct ll_sa_entry, se_list);
 146 }
 147
 148 static inline struct ll_inode_info *
 149 agl_first_entry(struct ll_statahead_info *sai)
 150 {
 151         return cfs_list_entry(sai->sai_entries_agl.next,
 152                               struct ll_inode_info, lli_agl_list);
 153 }
 154
 155 static inline int sa_sent_full(struct ll_statahead_info *sai)
 156 {
 157         return cfs_atomic_read(&sai->sai_cache_count) >= sai->sai_max;
 158 }
 159
 160 static inline int sa_received_empty(struct ll_statahead_info *sai)
 161 {
 162         return cfs_list_empty(&sai->sai_entries_received);
 163 }
 164
 165 static inline int agl_list_empty(struct ll_statahead_info *sai)
 166 {
 167         return cfs_list_empty(&sai->sai_entries_agl);
 168 }
 169
 170 /**
 171  * (1) hit ratio less than 80%
 172  * or
 173  * (2) consecutive miss more than 8
 174  * then means low hit.
 175  */
 176 static inline int sa_low_hit(struct ll_statahead_info *sai)
 177 {
 178         return ((sai->sai_hit > 7 && sai->sai_hit < 4 * sai->sai_miss) ||
 179                 (sai->sai_consecutive_miss > 8));
 180 }
 181
 182 /*
 183  * If the given index is behind of statahead window more than
 184  * SA_OMITTED_ENTRY_MAX, then it is old.
 185  */
 186 static inline int is_omitted_entry(struct ll_statahead_info *sai, __u64 index)
 187 {
 188         return ((__u64)sai->sai_max + index + SA_OMITTED_ENTRY_MAX <
 189                  sai->sai_index);
 190 }
 191
 192 /*
 193  * Insert it into sai_entries tail when init.
 194  */
 195 static struct ll_sa_entry *
 196 ll_sa_entry_alloc(struct ll_statahead_info *sai, __u64 index,
 197                   const char *name, int len)
 198 {
 199         struct ll_inode_info *lli;
 200         struct ll_sa_entry   *entry;
 201         int                   entry_size;
 202         char                 *dname;
 203         ENTRY;
 204
 205         entry_size = sizeof(struct ll_sa_entry) + (len & ~3) + 4;
 206         OBD_ALLOC(entry, entry_size);
 207         if (unlikely(entry == NULL))
 208                 RETURN(ERR_PTR(-ENOMEM));
 209
 210         CDEBUG(D_READA, "alloc sa entry %.*s(%p) index "LPU64"\n",
 211                len, name, entry, index);
 212
 213         entry->se_index = index;
 214
 215         /*
 216          * Statahead entry reference rules:
 217          *
 218          * 1) When statahead entry is initialized, its reference is set as 2.
 219          *    One reference is used by the directory scanner. When the scanner
 220          *    searches the statahead cache for the given name, it can perform
 221          *    lockless hash lookup (only the scanner can remove entry from hash
 222          *    list), and once found, it needn't to call "atomic_inc()" for the
 223          *    entry reference. So the performance is improved. After using the
 224          *    statahead entry, the scanner will call "atomic_dec()" to drop the
 225          *    reference held when initialization. If it is the last reference,
 226          *    the statahead entry will be freed.
 227          *
 228          * 2) All other threads, including statahead thread and ptlrpcd thread,
 229          *    when they process the statahead entry, the reference for target
 230          *    should be held to guarantee the entry will not be released by the
 231          *    directory scanner. After processing the entry, these threads will
 232          *    drop the entry reference. If it is the last reference, the entry
 233          *    will be freed.
 234          *
 235          *    The second reference when initializes the statahead entry is used
 236          *    by the statahead thread, following the rule 2).
 237          */
 238         cfs_atomic_set(&entry->se_refcount, 2);
 239         entry->se_stat = SA_ENTRY_INIT;
 240         entry->se_size = entry_size;
 241         dname = (char *)entry + sizeof(struct ll_sa_entry);
 242         memcpy(dname, name, len);
 243         dname[len] = 0;
 244         entry->se_qstr.hash = full_name_hash(name, len);
 245         entry->se_qstr.len = len;
 246         entry->se_qstr.name = dname;
 247
 248         lli = ll_i2info(sai->sai_inode);
 249         spin_lock(&lli->lli_sa_lock);
 250         cfs_list_add_tail(&entry->se_link, &sai->sai_entries);
 251         CFS_INIT_LIST_HEAD(&entry->se_list);
 252         ll_sa_entry_enhash(sai, entry);
 253         spin_unlock(&lli->lli_sa_lock);
 254
 255         cfs_atomic_inc(&sai->sai_cache_count);
 256
 257         RETURN(entry);
 258 }
 259
 260 /*
 261  * Used by the directory scanner to search entry with name.
 262  *
 263  * Only the caller can remove the entry from hash, so it is unnecessary to hold
 264  * hash lock. It is caller's duty to release the init refcount on the entry, so
 265  * it is also unnecessary to increase refcount on the entry.
 266  */
 267 static struct ll_sa_entry *
 268 ll_sa_entry_get_byname(struct ll_statahead_info *sai, const struct qstr *qstr)
 269 {
 270         struct ll_sa_entry *entry;
 271         int i = ll_sa_entry_hash(qstr->hash);
 272
 273         cfs_list_for_each_entry(entry, &sai->sai_cache[i], se_hash) {
 274                 if (entry->se_qstr.hash == qstr->hash &&
 275                     entry->se_qstr.len == qstr->len &&
 276                     memcmp(entry->se_qstr.name, qstr->name, qstr->len) == 0)
 277                         return entry;
 278         }
 279         return NULL;
 280 }
 281
 282 /*
 283  * Used by the async getattr request callback to find entry with index.
 284  *
 285  * Inside lli_sa_lock to prevent others to change the list during the search.
 286  * It needs to increase entry refcount before returning to guarantee that the
 287  * entry cannot be freed by others.
 288  */
 289 static struct ll_sa_entry *
 290 ll_sa_entry_get_byindex(struct ll_statahead_info *sai, __u64 index)
 291 {
 292         struct ll_sa_entry *entry;
 293
 294         cfs_list_for_each_entry(entry, &sai->sai_entries, se_link) {
 295                 if (entry->se_index == index) {
 296                         LASSERT(atomic_read(&entry->se_refcount) > 0);
 297                         cfs_atomic_inc(&entry->se_refcount);
 298                         return entry;
 299                 }
 300                 if (entry->se_index > index)
 301                         break;
 302         }
 303         return NULL;
 304 }
 305
 306 static void ll_sa_entry_cleanup(struct ll_statahead_info *sai,
 307                                  struct ll_sa_entry *entry)
 308 {
 309         struct md_enqueue_info *minfo = entry->se_minfo;
 310         struct ptlrpc_request  *req   = entry->se_req;
 311
 312         if (minfo) {
 313                 entry->se_minfo = NULL;
 314                 ll_intent_release(&minfo->mi_it);
 315                 iput(minfo->mi_dir);
 316                 OBD_FREE_PTR(minfo);
 317         }
 318
 319         if (req) {
 320                 entry->se_req = NULL;
 321                 ptlrpc_req_finished(req);
 322         }
 323 }
 324
 325 static void ll_sa_entry_put(struct ll_statahead_info *sai,
 326                              struct ll_sa_entry *entry)
 327 {
 328         if (cfs_atomic_dec_and_test(&entry->se_refcount)) {
 329                 CDEBUG(D_READA, "free sa entry %.*s(%p) index "LPU64"\n",
 330                        entry->se_qstr.len, entry->se_qstr.name, entry,
 331                        entry->se_index);
 332
 333                 LASSERT(cfs_list_empty(&entry->se_link));
 334                 LASSERT(cfs_list_empty(&entry->se_list));
 335                 LASSERT(ll_sa_entry_unhashed(entry));
 336
 337                 ll_sa_entry_cleanup(sai, entry);
 338                 if (entry->se_inode)
 339                         iput(entry->se_inode);
 340
 341                 OBD_FREE(entry, entry->se_size);
 342                 cfs_atomic_dec(&sai->sai_cache_count);
 343         }
 344 }
 345
 346 static inline void
 347 do_sa_entry_fini(struct ll_statahead_info *sai, struct ll_sa_entry *entry)
 348 {
 349         struct ll_inode_info *lli = ll_i2info(sai->sai_inode);
 350
 351         LASSERT(!ll_sa_entry_unhashed(entry));
 352         LASSERT(!cfs_list_empty(&entry->se_link));
 353
 354         ll_sa_entry_unhash(sai, entry);
 355
 356         spin_lock(&lli->lli_sa_lock);
 357         entry->se_stat = SA_ENTRY_DEST;
 358         cfs_list_del_init(&entry->se_link);
 359         if (likely(!cfs_list_empty(&entry->se_list)))
 360                 cfs_list_del_init(&entry->se_list);
 361         spin_unlock(&lli->lli_sa_lock);
 362
 363         ll_sa_entry_put(sai, entry);
 364 }
 365
 366 /*
 367  * Delete it from sai_entries_stated list when fini.
 368  */
 369 static void
 370 ll_sa_entry_fini(struct ll_statahead_info *sai, struct ll_sa_entry *entry)
 371 {
 372         struct ll_sa_entry *pos, *next;
 373
 374         if (entry)
 375                 do_sa_entry_fini(sai, entry);
 376
 377         /* drop old entry, only 'scanner' process does this, no need to lock */
 378         cfs_list_for_each_entry_safe(pos, next, &sai->sai_entries, se_link) {
 379                 if (!is_omitted_entry(sai, pos->se_index))
 380                         break;
 381                 do_sa_entry_fini(sai, pos);
 382         }
 383 }
 384
 385 /*
 386  * Inside lli_sa_lock.
 387  */
 388 static void
 389 do_sa_entry_to_stated(struct ll_statahead_info *sai,
 390                       struct ll_sa_entry *entry, se_stat_t stat)
 391 {
 392         struct ll_sa_entry *se;
 393         cfs_list_t         *pos = &sai->sai_entries_stated;
 394
 395         if (!cfs_list_empty(&entry->se_list))
 396                 cfs_list_del_init(&entry->se_list);
 397
 398         cfs_list_for_each_entry_reverse(se, &sai->sai_entries_stated, se_list) {
 399                 if (se->se_index < entry->se_index) {
 400                         pos = &se->se_list;
 401                         break;
 402                 }
 403         }
 404
 405         cfs_list_add(&entry->se_list, pos);
 406         entry->se_stat = stat;
 407 }
 408
 409 /*
 410  * Move entry to sai_entries_stated and sort with the index.
 411  * \retval 1    -- entry to be destroyed.
 412  * \retval 0    -- entry is inserted into stated list.
 413  */
 414 static int
 415 ll_sa_entry_to_stated(struct ll_statahead_info *sai,
 416                       struct ll_sa_entry *entry, se_stat_t stat)
 417 {
 418         struct ll_inode_info *lli = ll_i2info(sai->sai_inode);
 419         int                   ret = 1;
 420
 421         ll_sa_entry_cleanup(sai, entry);
 422
 423         spin_lock(&lli->lli_sa_lock);
 424         if (likely(entry->se_stat != SA_ENTRY_DEST)) {
 425                 do_sa_entry_to_stated(sai, entry, stat);
 426                 ret = 0;
 427         }
 428         spin_unlock(&lli->lli_sa_lock);
 429
 430         return ret;
 431 }
 432
 433 /*
 434  * Insert inode into the list of sai_entries_agl.
 435  */
 436 static void ll_agl_add(struct ll_statahead_info *sai,
 437                        struct inode *inode, int index)
 438 {
 439         struct ll_inode_info *child  = ll_i2info(inode);
 440         struct ll_inode_info *parent = ll_i2info(sai->sai_inode);
 441         int                   added  = 0;
 442
 443         spin_lock(&child->lli_agl_lock);
 444         if (child->lli_agl_index == 0) {
 445                 child->lli_agl_index = index;
 446                 spin_unlock(&child->lli_agl_lock);
 447
 448                 LASSERT(cfs_list_empty(&child->lli_agl_list));
 449
 450                 igrab(inode);
 451                 spin_lock(&parent->lli_agl_lock);
 452                 if (agl_list_empty(sai))
 453                         added = 1;
 454                 cfs_list_add_tail(&child->lli_agl_list, &sai->sai_entries_agl);
 455                 spin_unlock(&parent->lli_agl_lock);
 456         } else {
 457                 spin_unlock(&child->lli_agl_lock);
 458         }
 459
 460         if (added > 0)
 461                 cfs_waitq_signal(&sai->sai_agl_thread.t_ctl_waitq);
 462 }
 463
 464 static struct ll_statahead_info *ll_sai_alloc(void)
 465 {
 466         struct ll_statahead_info *sai;
 467         int                       i;
 468         ENTRY;
 469
 470         OBD_ALLOC_PTR(sai);
 471         if (!sai)
 472                 RETURN(NULL);
 473
 474         cfs_atomic_set(&sai->sai_refcount, 1);
 475
 476         spin_lock(&sai_generation_lock);
 477         sai->sai_generation = ++sai_generation;
 478         if (unlikely(sai_generation == 0))
 479                 sai->sai_generation = ++sai_generation;
 480         spin_unlock(&sai_generation_lock);
 481
 482         sai->sai_max = LL_SA_RPC_MIN;
 483         sai->sai_index = 1;
 484         cfs_waitq_init(&sai->sai_waitq);
 485         cfs_waitq_init(&sai->sai_thread.t_ctl_waitq);
 486         cfs_waitq_init(&sai->sai_agl_thread.t_ctl_waitq);
 487
 488         CFS_INIT_LIST_HEAD(&sai->sai_entries);
 489         CFS_INIT_LIST_HEAD(&sai->sai_entries_received);
 490         CFS_INIT_LIST_HEAD(&sai->sai_entries_stated);
 491         CFS_INIT_LIST_HEAD(&sai->sai_entries_agl);
 492
 493         for (i = 0; i < LL_SA_CACHE_SIZE; i++) {
 494                 CFS_INIT_LIST_HEAD(&sai->sai_cache[i]);
 495                 spin_lock_init(&sai->sai_cache_lock[i]);
 496         }
 497         cfs_atomic_set(&sai->sai_cache_count, 0);
 498
 499         RETURN(sai);
 500 }
 501
 502 static inline struct ll_statahead_info *
 503 ll_sai_get(struct ll_statahead_info *sai)
 504 {
 505         cfs_atomic_inc(&sai->sai_refcount);
 506         return sai;
 507 }
 508
 509 static void ll_sai_put(struct ll_statahead_info *sai)
 510 {
 511         struct inode         *inode = sai->sai_inode;
 512         struct ll_inode_info *lli   = ll_i2info(inode);
 513         ENTRY;
 514
 515         if (cfs_atomic_dec_and_lock(&sai->sai_refcount, &lli->lli_sa_lock)) {
 516                 struct ll_sa_entry *entry, *next;
 517
 518                 if (unlikely(cfs_atomic_read(&sai->sai_refcount) > 0)) {
 519                         /* It is race case, the interpret callback just hold
 520                          * a reference count */
 521                         spin_unlock(&lli->lli_sa_lock);
 522                         RETURN_EXIT;
 523                 }
 524
 525                 LASSERT(lli->lli_opendir_key == NULL);
 526                 LASSERT(thread_is_stopped(&sai->sai_thread));
 527                 LASSERT(thread_is_stopped(&sai->sai_agl_thread));
 528
 529                 lli->lli_sai = NULL;
 530                 lli->lli_opendir_pid = 0;
 531                 spin_unlock(&lli->lli_sa_lock);
 532
 533                 if (sai->sai_sent > sai->sai_replied)
 534                         CDEBUG(D_READA,"statahead for dir "DFID" does not "
 535                               "finish: [sent:"LPU64"] [replied:"LPU64"]\n",
 536                               PFID(&lli->lli_fid),
 537                               sai->sai_sent, sai->sai_replied);
 538
 539                 cfs_list_for_each_entry_safe(entry, next,
 540                                              &sai->sai_entries, se_link)
 541                         do_sa_entry_fini(sai, entry);
 542
 543                 LASSERT(list_empty(&sai->sai_entries));
 544                 LASSERT(sa_received_empty(sai));
 545                 LASSERT(list_empty(&sai->sai_entries_stated));
 546
 547                 LASSERT(cfs_atomic_read(&sai->sai_cache_count) == 0);
 548                 LASSERT(agl_list_empty(sai));
 549
 550                 iput(inode);
 551                 OBD_FREE_PTR(sai);
 552         }
 553
 554         EXIT;
 555 }
 556
 557 /* Do NOT forget to drop inode refcount when into sai_entries_agl. */
 558 static void ll_agl_trigger(struct inode *inode, struct ll_statahead_info *sai)
 559 {
 560         struct ll_inode_info *lli   = ll_i2info(inode);
 561         __u64                 index = lli->lli_agl_index;
 562         int                   rc;
 563         ENTRY;
 564
 565         LASSERT(cfs_list_empty(&lli->lli_agl_list));
 566
 567         /* AGL maybe fall behind statahead with one entry */
 568         if (is_omitted_entry(sai, index + 1)) {
 569                 lli->lli_agl_index = 0;
 570                 iput(inode);
 571                 RETURN_EXIT;
 572         }
 573
 574         /* Someone is in glimpse (sync or async), do nothing. */
 575         rc = down_write_trylock(&lli->lli_glimpse_sem);
 576         if (rc == 0) {
 577                 lli->lli_agl_index = 0;
 578                 iput(inode);
 579                 RETURN_EXIT;
 580         }
 581
 582         /*
 583          * Someone triggered glimpse within 1 sec before.
 584          * 1) The former glimpse succeeded with glimpse lock granted by OST, and
 585          *    if the lock is still cached on client, AGL needs to do nothing. If
 586          *    it is cancelled by other client, AGL maybe cannot obtaion new lock
 587          *    for no glimpse callback triggered by AGL.
 588          * 2) The former glimpse succeeded, but OST did not grant glimpse lock.
 589          *    Under such case, it is quite possible that the OST will not grant
 590          *    glimpse lock for AGL also.
 591          * 3) The former glimpse failed, compared with other two cases, it is
 592          *    relative rare. AGL can ignore such case, and it will not muchly
 593          *    affect the performance.
 594          */
 595         if (lli->lli_glimpse_time != 0 &&
 596             cfs_time_before(cfs_time_shift(-1), lli->lli_glimpse_time)) {
 597                 up_write(&lli->lli_glimpse_sem);
 598                 lli->lli_agl_index = 0;
 599                 iput(inode);
 600                 RETURN_EXIT;
 601         }
 602
 603         CDEBUG(D_READA, "Handling (init) async glimpse: inode = "
 604                DFID", idx = "LPU64"\n", PFID(&lli->lli_fid), index);
 605
 606         cl_agl(inode);
 607         lli->lli_agl_index = 0;
 608         lli->lli_glimpse_time = cfs_time_current();
 609         up_write(&lli->lli_glimpse_sem);
 610
 611         CDEBUG(D_READA, "Handled (init) async glimpse: inode= "
 612                DFID", idx = "LPU64", rc = %d\n",
 613                PFID(&lli->lli_fid), index, rc);
 614
 615         iput(inode);
 616
 617         EXIT;
 618 }
 619
 620 static void ll_post_statahead(struct ll_statahead_info *sai)
 621 {
 622         struct inode           *dir   = sai->sai_inode;
 623         struct inode           *child;
 624         struct ll_inode_info   *lli   = ll_i2info(dir);
 625         struct ll_sa_entry     *entry;
 626         struct md_enqueue_info *minfo;
 627         struct lookup_intent   *it;
 628         struct ptlrpc_request  *req;
 629         struct mdt_body        *body;
 630         int                     rc    = 0;
 631         ENTRY;
 632
 633         spin_lock(&lli->lli_sa_lock);
 634         if (unlikely(sa_received_empty(sai))) {
 635                 spin_unlock(&lli->lli_sa_lock);
 636                 RETURN_EXIT;
 637         }
 638         entry = sa_first_received_entry(sai);
 639         cfs_atomic_inc(&entry->se_refcount);
 640         cfs_list_del_init(&entry->se_list);
 641         spin_unlock(&lli->lli_sa_lock);
 642
 643         LASSERT(entry->se_handle != 0);
 644
 645         minfo = entry->se_minfo;
 646         it = &minfo->mi_it;
 647         req = entry->se_req;
 648         body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
 649         if (body == NULL)
 650                 GOTO(out, rc = -EFAULT);
 651
 652         child = entry->se_inode;
 653         if (child == NULL) {
 654                 /*
 655                  * lookup.
 656                  */
 657                 LASSERT(fid_is_zero(&minfo->mi_data.op_fid2));
 658
 659                 /* XXX: No fid in reply, this is probaly cross-ref case.
 660                  * SA can't handle it yet. */
 661                 if (body->valid & OBD_MD_MDS)
 662                         GOTO(out, rc = -EAGAIN);
 663         } else {
 664                 /*
 665                  * revalidate.
 666                  */
 667                 /* unlinked and re-created with the same name */
 668                 if (unlikely(!lu_fid_eq(&minfo->mi_data.op_fid2, &body->fid1))){
 669                         entry->se_inode = NULL;
 670                         iput(child);
 671                         child = NULL;
 672                 }
 673         }
 674
 675         it->d.lustre.it_lock_handle = entry->se_handle;
 676         rc = md_revalidate_lock(ll_i2mdexp(dir), it, ll_inode2fid(dir), NULL);
 677         if (rc != 1)
 678                 GOTO(out, rc = -EAGAIN);
 679
 680         rc = ll_prep_inode(&child, req, dir->i_sb, it);
 681         if (rc)
 682                 GOTO(out, rc);
 683
 684         CDEBUG(D_DLMTRACE, "setting l_data to inode %p (%lu/%u)\n",
 685                child, child->i_ino, child->i_generation);
 686         ll_set_lock_data(ll_i2sbi(dir)->ll_md_exp, child, it, NULL);
 687
 688         entry->se_inode = child;
 689
 690         if (agl_should_run(sai, child))
 691                 ll_agl_add(sai, child, entry->se_index);
 692
 693         EXIT;
 694
 695 out:
 696         /* The "ll_sa_entry_to_stated()" will drop related ldlm ibits lock
 697          * reference count by calling "ll_intent_drop_lock()" in spite of the
 698          * above operations failed or not. Do not worry about calling
 699          * "ll_intent_drop_lock()" more than once. */
 700         rc = ll_sa_entry_to_stated(sai, entry,
 701                                    rc < 0 ? SA_ENTRY_INVA : SA_ENTRY_SUCC);
 702         if (rc == 0 && entry->se_index == sai->sai_index_wait)
 703                 cfs_waitq_signal(&sai->sai_waitq);
 704         ll_sa_entry_put(sai, entry);
 705 }
 706
 707 static int ll_statahead_interpret(struct ptlrpc_request *req,
 708                                   struct md_enqueue_info *minfo, int rc)
 709 {
 710         struct lookup_intent     *it  = &minfo->mi_it;
 711         struct inode             *dir = minfo->mi_dir;
 712         struct ll_inode_info     *lli = ll_i2info(dir);
 713         struct ll_statahead_info *sai = NULL;
 714         struct ll_sa_entry       *entry;
 715         int                       wakeup;
 716         ENTRY;
 717
 718         if (it_disposition(it, DISP_LOOKUP_NEG))
 719                 rc = -ENOENT;
 720
 721         spin_lock(&lli->lli_sa_lock);
 722         /* stale entry */
 723         if (unlikely(lli->lli_sai == NULL ||
 724                      lli->lli_sai->sai_generation != minfo->mi_generation)) {
 725                 spin_unlock(&lli->lli_sa_lock);
 726                 GOTO(out, rc = -ESTALE);
 727         } else {
 728                 sai = ll_sai_get(lli->lli_sai);
 729                 if (unlikely(!thread_is_running(&sai->sai_thread))) {
 730                         sai->sai_replied++;
 731                         spin_unlock(&lli->lli_sa_lock);
 732                         GOTO(out, rc = -EBADFD);
 733                 }
 734
 735                 entry = ll_sa_entry_get_byindex(sai, minfo->mi_cbdata);
 736                 if (entry == NULL) {
 737                         sai->sai_replied++;
 738                         spin_unlock(&lli->lli_sa_lock);
 739                         GOTO(out, rc = -EIDRM);
 740                 }
 741
 742                 if (rc != 0) {
 743                         do_sa_entry_to_stated(sai, entry, SA_ENTRY_INVA);
 744                         wakeup = (entry->se_index == sai->sai_index_wait);
 745                 } else {
 746                         entry->se_minfo = minfo;
 747                         entry->se_req = ptlrpc_request_addref(req);
 748                         /* Release the async ibits lock ASAP to avoid deadlock
 749                          * when statahead thread tries to enqueue lock on parent
 750                          * for readpage and other tries to enqueue lock on child
 751                          * with parent's lock held, for example: unlink. */
 752                         entry->se_handle = it->d.lustre.it_lock_handle;
 753                         ll_intent_drop_lock(it);
 754                         wakeup = sa_received_empty(sai);
 755                         cfs_list_add_tail(&entry->se_list,
 756                                           &sai->sai_entries_received);
 757                 }
 758                 sai->sai_replied++;
 759                 spin_unlock(&lli->lli_sa_lock);
 760
 761                 ll_sa_entry_put(sai, entry);
 762                 if (wakeup)
 763                         cfs_waitq_signal(&sai->sai_thread.t_ctl_waitq);
 764         }
 765
 766         EXIT;
 767
 768 out:
 769         if (rc != 0) {
 770                 ll_intent_release(it);
 771                 iput(dir);
 772                 OBD_FREE_PTR(minfo);
 773         }
 774         if (sai != NULL)
 775                 ll_sai_put(sai);
 776         return rc;
 777 }
 778
 779 static void sa_args_fini(struct md_enqueue_info *minfo,
 780                          struct ldlm_enqueue_info *einfo)
 781 {
 782         LASSERT(minfo && einfo);
 783         iput(minfo->mi_dir);
 784         capa_put(minfo->mi_data.op_capa1);
 785         capa_put(minfo->mi_data.op_capa2);
 786         OBD_FREE_PTR(minfo);
 787         OBD_FREE_PTR(einfo);
 788 }
 789
 790 /**
 791  * There is race condition between "capa_put" and "ll_statahead_interpret" for
 792  * accessing "op_data.op_capa[1,2]" as following:
 793  * "capa_put" releases "op_data.op_capa[1,2]"'s reference count after calling
 794  * "md_intent_getattr_async". But "ll_statahead_interpret" maybe run first, and
 795  * fill "op_data.op_capa[1,2]" as POISON, then cause "capa_put" access invalid
 796  * "ocapa". So here reserve "op_data.op_capa[1,2]" in "pcapa" before calling
 797  * "md_intent_getattr_async".
 798  */
 799 static int sa_args_init(struct inode *dir, struct inode *child,
 800                         struct ll_sa_entry *entry, struct md_enqueue_info **pmi,
 801                         struct ldlm_enqueue_info **pei,
 802                         struct obd_capa **pcapa)
 803 {
 804         struct qstr              *qstr = &entry->se_qstr;
 805         struct ll_inode_info     *lli  = ll_i2info(dir);
 806         struct md_enqueue_info   *minfo;
 807         struct ldlm_enqueue_info *einfo;
 808         struct md_op_data        *op_data;
 809
 810         OBD_ALLOC_PTR(einfo);
 811         if (einfo == NULL)
 812                 return -ENOMEM;
 813
 814         OBD_ALLOC_PTR(minfo);
 815         if (minfo == NULL) {
 816                 OBD_FREE_PTR(einfo);
 817                 return -ENOMEM;
 818         }
 819
 820         op_data = ll_prep_md_op_data(&minfo->mi_data, dir, child, qstr->name,
 821                                      qstr->len, 0, LUSTRE_OPC_ANY, NULL);
 822         if (IS_ERR(op_data)) {
 823                 OBD_FREE_PTR(einfo);
 824                 OBD_FREE_PTR(minfo);
 825                 return PTR_ERR(op_data);
 826         }
 827
 828         minfo->mi_it.it_op = IT_GETATTR;
 829         minfo->mi_dir = igrab(dir);
 830         minfo->mi_cb = ll_statahead_interpret;
 831         minfo->mi_generation = lli->lli_sai->sai_generation;
 832         minfo->mi_cbdata = entry->se_index;
 833
 834         einfo->ei_type   = LDLM_IBITS;
 835         einfo->ei_mode   = it_to_lock_mode(&minfo->mi_it);
 836         einfo->ei_cb_bl  = ll_md_blocking_ast;
 837         einfo->ei_cb_cp  = ldlm_completion_ast;
 838         einfo->ei_cb_gl  = NULL;
 839         einfo->ei_cbdata = NULL;
 840
 841         *pmi = minfo;
 842         *pei = einfo;
 843         pcapa[0] = op_data->op_capa1;
 844         pcapa[1] = op_data->op_capa2;
 845
 846         return 0;
 847 }
 848
 849 static int do_sa_lookup(struct inode *dir, struct ll_sa_entry *entry)
 850 {
 851         struct md_enqueue_info   *minfo;
 852         struct ldlm_enqueue_info *einfo;
 853         struct obd_capa          *capas[2];
 854         int                       rc;
 855         ENTRY;
 856
 857         rc = sa_args_init(dir, NULL, entry, &minfo, &einfo, capas);
 858         if (rc)
 859                 RETURN(rc);
 860
 861         rc = md_intent_getattr_async(ll_i2mdexp(dir), minfo, einfo);
 862         if (!rc) {
 863                 capa_put(capas[0]);
 864                 capa_put(capas[1]);
 865         } else {
 866                 sa_args_fini(minfo, einfo);
 867         }
 868
 869         RETURN(rc);
 870 }
 871
 872 /**
 873  * similar to ll_revalidate_it().
 874  * \retval      1 -- dentry valid
 875  * \retval      0 -- will send stat-ahead request
 876  * \retval others -- prepare stat-ahead request failed
 877  */
 878 static int do_sa_revalidate(struct inode *dir, struct ll_sa_entry *entry,
 879                             struct dentry *dentry)
 880 {
 881         struct inode             *inode = dentry->d_inode;
 882         struct lookup_intent      it = { .it_op = IT_GETATTR,
 883                                          .d.lustre.it_lock_handle = 0 };
 884         struct md_enqueue_info   *minfo;
 885         struct ldlm_enqueue_info *einfo;
 886         struct obd_capa          *capas[2];
 887         int rc;
 888         ENTRY;
 889
 890         if (unlikely(inode == NULL))
 891                 RETURN(1);
 892
 893         if (d_mountpoint(dentry))
 894                 RETURN(1);
 895
 896         entry->se_inode = igrab(inode);
 897         rc = md_revalidate_lock(ll_i2mdexp(dir), &it, ll_inode2fid(inode),NULL);
 898         if (rc == 1) {
 899                 entry->se_handle = it.d.lustre.it_lock_handle;
 900                 ll_intent_release(&it);
 901                 RETURN(1);
 902         }
 903
 904         rc = sa_args_init(dir, inode, entry, &minfo, &einfo, capas);
 905         if (rc) {
 906                 entry->se_inode = NULL;
 907                 iput(inode);
 908                 RETURN(rc);
 909         }
 910
 911         rc = md_intent_getattr_async(ll_i2mdexp(dir), minfo, einfo);
 912         if (!rc) {
 913                 capa_put(capas[0]);
 914                 capa_put(capas[1]);
 915         } else {
 916                 entry->se_inode = NULL;
 917                 iput(inode);
 918                 sa_args_fini(minfo, einfo);
 919         }
 920
 921         RETURN(rc);
 922 }
 923
 924 static void ll_statahead_one(struct dentry *parent, const char* entry_name,
 925                              int entry_name_len)
 926 {
 927         struct inode             *dir    = parent->d_inode;
 928         struct ll_inode_info     *lli    = ll_i2info(dir);
 929         struct ll_statahead_info *sai    = lli->lli_sai;
 930         struct dentry            *dentry = NULL;
 931         struct ll_sa_entry       *entry;
 932         int                       rc;
 933         int                       rc1;
 934         ENTRY;
 935
 936         entry = ll_sa_entry_alloc(sai, sai->sai_index, entry_name,
 937                                   entry_name_len);
 938         if (IS_ERR(entry))
 939                 RETURN_EXIT;
 940
 941         dentry = d_lookup(parent, &entry->se_qstr);
 942         if (!dentry) {
 943                 rc = do_sa_lookup(dir, entry);
 944         } else {
 945                 rc = do_sa_revalidate(dir, entry, dentry);
 946                 if (rc == 1 && agl_should_run(sai, dentry->d_inode))
 947                         ll_agl_add(sai, dentry->d_inode, entry->se_index);
 948         }
 949
 950         if (dentry != NULL)
 951                 dput(dentry);
 952
 953         if (rc) {
 954                 rc1 = ll_sa_entry_to_stated(sai, entry,
 955                                         rc < 0 ? SA_ENTRY_INVA : SA_ENTRY_SUCC);
 956                 if (rc1 == 0 && entry->se_index == sai->sai_index_wait)
 957                         cfs_waitq_signal(&sai->sai_waitq);
 958         } else {
 959                 sai->sai_sent++;
 960         }
 961
 962         sai->sai_index++;
 963         /* drop one refcount on entry by ll_sa_entry_alloc */
 964         ll_sa_entry_put(sai, entry);
 965
 966         EXIT;
 967 }
 968
 969 static int ll_agl_thread(void *arg)
 970 {
 971         struct dentry            *parent = (struct dentry *)arg;
 972         struct inode             *dir    = parent->d_inode;
 973         struct ll_inode_info     *plli   = ll_i2info(dir);
 974         struct ll_inode_info     *clli;
 975         struct ll_sb_info        *sbi    = ll_i2sbi(dir);
 976         struct ll_statahead_info *sai    = ll_sai_get(plli->lli_sai);
 977         struct ptlrpc_thread     *thread = &sai->sai_agl_thread;
 978         struct l_wait_info        lwi    = { 0 };
 979         ENTRY;
 980
 981         CDEBUG(D_READA, "agl thread started: [pid %d] [parent %.*s]\n",
 982                current_pid(), parent->d_name.len, parent->d_name.name);
 983
 984         atomic_inc(&sbi->ll_agl_total);
 985         spin_lock(&plli->lli_agl_lock);
 986         sai->sai_agl_valid = 1;
 987         thread_set_flags(thread, SVC_RUNNING);
 988         spin_unlock(&plli->lli_agl_lock);
 989         cfs_waitq_signal(&thread->t_ctl_waitq);
 990
 991         while (1) {
 992                 l_wait_event(thread->t_ctl_waitq,
 993                              !agl_list_empty(sai) ||
 994                              !thread_is_running(thread),
 995                              &lwi);
 996
 997                 if (!thread_is_running(thread))
 998                         break;
 999
1000                 spin_lock(&plli->lli_agl_lock);
1001                 /* The statahead thread maybe help to process AGL entries,
1002                  * so check whether list empty again. */
1003                 if (!agl_list_empty(sai)) {
1004                         clli = agl_first_entry(sai);
1005                         cfs_list_del_init(&clli->lli_agl_list);
1006                         spin_unlock(&plli->lli_agl_lock);
1007                         ll_agl_trigger(&clli->lli_vfs_inode, sai);
1008                 } else {
1009                         spin_unlock(&plli->lli_agl_lock);
1010                 }
1011         }
1012
1013         spin_lock(&plli->lli_agl_lock);
1014         sai->sai_agl_valid = 0;
1015         while (!agl_list_empty(sai)) {
1016                 clli = agl_first_entry(sai);
1017                 cfs_list_del_init(&clli->lli_agl_list);
1018                 spin_unlock(&plli->lli_agl_lock);
1019                 clli->lli_agl_index = 0;
1020                 iput(&clli->lli_vfs_inode);
1021                 spin_lock(&plli->lli_agl_lock);
1022         }
1023         thread_set_flags(thread, SVC_STOPPED);
1024         spin_unlock(&plli->lli_agl_lock);
1025         cfs_waitq_signal(&thread->t_ctl_waitq);
1026         ll_sai_put(sai);
1027         CDEBUG(D_READA, "agl thread stopped: [pid %d] [parent %.*s]\n",
1028                current_pid(), parent->d_name.len, parent->d_name.name);
1029         RETURN(0);
1030 }
1031
1032 static void ll_start_agl(struct dentry *parent, struct ll_statahead_info *sai)
1033 {
1034         struct ptlrpc_thread *thread = &sai->sai_agl_thread;
1035         struct l_wait_info    lwi    = { 0 };
1036         struct ll_inode_info  *plli;
1037         cfs_task_t            *task;
1038         ENTRY;
1039
1040         CDEBUG(D_READA, "start agl thread: [pid %d] [parent %.*s]\n",
1041                current_pid(), parent->d_name.len, parent->d_name.name);
1042
1043         plli = ll_i2info(parent->d_inode);
1044         task = kthread_run(ll_agl_thread, parent,
1045                                "ll_agl_%u", plli->lli_opendir_pid);
1046         if (IS_ERR(task)) {
1047                 CERROR("can't start ll_agl thread, rc: %ld\n", PTR_ERR(task));
1048                 thread_set_flags(thread, SVC_STOPPED);
1049                 RETURN_EXIT;
1050         }
1051
1052         l_wait_event(thread->t_ctl_waitq,
1053                      thread_is_running(thread) || thread_is_stopped(thread),
1054                      &lwi);
1055         EXIT;
1056 }
1057
1058 static int ll_statahead_thread(void *arg)
1059 {
1060         struct dentry            *parent = (struct dentry *)arg;
1061         struct inode             *dir    = parent->d_inode;
1062         struct ll_inode_info     *plli   = ll_i2info(dir);
1063         struct ll_inode_info     *clli;
1064         struct ll_sb_info        *sbi    = ll_i2sbi(dir);
1065         struct ll_statahead_info *sai    = ll_sai_get(plli->lli_sai);
1066         struct ptlrpc_thread     *thread = &sai->sai_thread;
1067         struct ptlrpc_thread *agl_thread = &sai->sai_agl_thread;
1068         struct page              *page;
1069         __u64                     pos    = 0;
1070         int                       first  = 0;
1071         int                       rc     = 0;
1072         struct ll_dir_chain       chain;
1073         struct l_wait_info        lwi    = { 0 };
1074         ENTRY;
1075
1076         CDEBUG(D_READA, "statahead thread started: [pid %d] [parent %.*s]\n",
1077                current_pid(), parent->d_name.len, parent->d_name.name);
1078
1079         if (sbi->ll_flags & LL_SBI_AGL_ENABLED)
1080                 ll_start_agl(parent, sai);
1081
1082         atomic_inc(&sbi->ll_sa_total);
1083         spin_lock(&plli->lli_sa_lock);
1084         thread_set_flags(thread, SVC_RUNNING);
1085         spin_unlock(&plli->lli_sa_lock);
1086         cfs_waitq_signal(&thread->t_ctl_waitq);
1087
1088         ll_dir_chain_init(&chain);
1089         page = ll_get_dir_page(dir, pos, &chain);
1090
1091         while (1) {
1092                 struct lu_dirpage *dp;
1093                 struct lu_dirent  *ent;
1094
1095                 if (IS_ERR(page)) {
1096                         rc = PTR_ERR(page);
1097                         CDEBUG(D_READA, "error reading dir "DFID" at "LPU64
1098                                "/"LPU64": [rc %d] [parent %u]\n",
1099                                PFID(ll_inode2fid(dir)), pos, sai->sai_index,
1100                                rc, plli->lli_opendir_pid);
1101                         GOTO(out, rc);
1102                 }
1103
1104                 dp = page_address(page);
1105                 for (ent = lu_dirent_start(dp); ent != NULL;
1106                      ent = lu_dirent_next(ent)) {
1107                         __u64 hash;
1108                         int namelen;
1109                         char *name;
1110
1111                         hash = le64_to_cpu(ent->lde_hash);
1112                         if (unlikely(hash < pos))
1113                                 /*
1114                                  * Skip until we find target hash value.
1115                                  */
1116                                 continue;
1117
1118                         namelen = le16_to_cpu(ent->lde_namelen);
1119                         if (unlikely(namelen == 0))
1120                                 /*
1121                                  * Skip dummy record.
1122                                  */
1123                                 continue;
1124
1125                         name = ent->lde_name;
1126                         if (name[0] == '.') {
1127                                 if (namelen == 1) {
1128                                         /*
1129                                          * skip "."
1130                                          */
1131                                         continue;
1132                                 } else if (name[1] == '.' && namelen == 2) {
1133                                         /*
1134                                          * skip ".."
1135                                          */
1136                                         continue;
1137                                 } else if (!sai->sai_ls_all) {
1138                                         /*
1139                                          * skip hidden files.
1140                                          */
1141                                         sai->sai_skip_hidden++;
1142                                         continue;
1143                                 }
1144                         }
1145
1146                         /*
1147                          * don't stat-ahead first entry.
1148                          */
1149                         if (unlikely(++first == 1))
1150                                 continue;
1151
1152 keep_it:
1153                         l_wait_event(thread->t_ctl_waitq,
1154                                      !sa_sent_full(sai) ||
1155                                      !sa_received_empty(sai) ||
1156                                      !agl_list_empty(sai) ||
1157                                      !thread_is_running(thread),
1158                                      &lwi);
1159
1160 interpret_it:
1161                         while (!sa_received_empty(sai))
1162                                 ll_post_statahead(sai);
1163
1164                         if (unlikely(!thread_is_running(thread))) {
1165                                 ll_release_page(page, 0);
1166                                 GOTO(out, rc = 0);
1167                         }
1168
1169                         /* If no window for metadata statahead, but there are
1170                          * some AGL entries to be triggered, then try to help
1171                          * to process the AGL entries. */
1172                         if (sa_sent_full(sai)) {
1173                                 spin_lock(&plli->lli_agl_lock);
1174                                 while (!agl_list_empty(sai)) {
1175                                         clli = agl_first_entry(sai);
1176                                         cfs_list_del_init(&clli->lli_agl_list);
1177                                         spin_unlock(&plli->lli_agl_lock);
1178                                         ll_agl_trigger(&clli->lli_vfs_inode,
1179                                                        sai);
1180
1181                                         if (!sa_received_empty(sai))
1182                                                 goto interpret_it;
1183
1184                                         if (unlikely(
1185                                                 !thread_is_running(thread))) {
1186                                                 ll_release_page(page, 0);
1187                                                 GOTO(out, rc = 0);
1188                                         }
1189
1190                                         if (!sa_sent_full(sai))
1191                                                 goto do_it;
1192
1193                                         spin_lock(&plli->lli_agl_lock);
1194                                 }
1195                                 spin_unlock(&plli->lli_agl_lock);
1196
1197                                 goto keep_it;
1198                         }
1199
1200 do_it:
1201                         ll_statahead_one(parent, name, namelen);
1202                 }
1203                 pos = le64_to_cpu(dp->ldp_hash_end);
1204                 if (pos == MDS_DIR_END_OFF) {
1205                         /*
1206                          * End of directory reached.
1207                          */
1208                         ll_release_page(page, 0);
1209                         while (1) {
1210                                 l_wait_event(thread->t_ctl_waitq,
1211                                              !sa_received_empty(sai) ||
1212                                              sai->sai_sent == sai->sai_replied||
1213                                              !thread_is_running(thread),
1214                                              &lwi);
1215
1216                                 while (!sa_received_empty(sai))
1217                                         ll_post_statahead(sai);
1218
1219                                 if (unlikely(!thread_is_running(thread)))
1220                                         GOTO(out, rc = 0);
1221
1222                                 if (sai->sai_sent == sai->sai_replied &&
1223                                     sa_received_empty(sai))
1224                                         break;
1225                         }
1226
1227                         spin_lock(&plli->lli_agl_lock);
1228                         while (!agl_list_empty(sai) &&
1229                                thread_is_running(thread)) {
1230                                 clli = agl_first_entry(sai);
1231                                 cfs_list_del_init(&clli->lli_agl_list);
1232                                 spin_unlock(&plli->lli_agl_lock);
1233                                 ll_agl_trigger(&clli->lli_vfs_inode, sai);
1234                                 spin_lock(&plli->lli_agl_lock);
1235                         }
1236                         spin_unlock(&plli->lli_agl_lock);
1237
1238                         GOTO(out, rc = 0);
1239                 } else if (1) {
1240                         /*
1241                          * chain is exhausted.
1242                          * Normal case: continue to the next page.
1243                          */
1244                         ll_release_page(page, le32_to_cpu(dp->ldp_flags) &
1245                                               LDF_COLLIDE);
1246                         sai->sai_in_readpage = 1;
1247                         page = ll_get_dir_page(dir, pos, &chain);
1248                         sai->sai_in_readpage = 0;
1249                 } else {
1250                         LASSERT(le32_to_cpu(dp->ldp_flags) & LDF_COLLIDE);
1251                         ll_release_page(page, 1);
1252                         /*
1253                          * go into overflow page.
1254                          */
1255                 }
1256         }
1257         EXIT;
1258
1259 out:
1260         if (sai->sai_agl_valid) {
1261                 spin_lock(&plli->lli_agl_lock);
1262                 thread_set_flags(agl_thread, SVC_STOPPING);
1263                 spin_unlock(&plli->lli_agl_lock);
1264                 cfs_waitq_signal(&agl_thread->t_ctl_waitq);
1265
1266                 CDEBUG(D_READA, "stop agl thread: [pid %d]\n",
1267                        current_pid());
1268                 l_wait_event(agl_thread->t_ctl_waitq,
1269                              thread_is_stopped(agl_thread),
1270                              &lwi);
1271         } else {
1272                 /* Set agl_thread flags anyway. */
1273                 thread_set_flags(&sai->sai_agl_thread, SVC_STOPPED);
1274         }
1275         ll_dir_chain_fini(&chain);
1276         spin_lock(&plli->lli_sa_lock);
1277         if (!sa_received_empty(sai)) {
1278                 thread_set_flags(thread, SVC_STOPPING);
1279                 spin_unlock(&plli->lli_sa_lock);
1280
1281                 /* To release the resources held by received entries. */
1282                 while (!sa_received_empty(sai))
1283                         ll_post_statahead(sai);
1284
1285                 spin_lock(&plli->lli_sa_lock);
1286         }
1287         thread_set_flags(thread, SVC_STOPPED);
1288         spin_unlock(&plli->lli_sa_lock);
1289         cfs_waitq_signal(&sai->sai_waitq);
1290         cfs_waitq_signal(&thread->t_ctl_waitq);
1291         ll_sai_put(sai);
1292         dput(parent);
1293         CDEBUG(D_READA, "statahead thread stopped: [pid %d] [parent %.*s]\n",
1294                current_pid(), parent->d_name.len, parent->d_name.name);
1295         return rc;
1296 }
1297
1298 /**
1299  * called in ll_file_release().
1300  */
1301 void ll_stop_statahead(struct inode *dir, void *key)
1302 {
1303         struct ll_inode_info *lli = ll_i2info(dir);
1304
1305         if (unlikely(key == NULL))
1306                 return;
1307
1308         spin_lock(&lli->lli_sa_lock);
1309         if (lli->lli_opendir_key != key || lli->lli_opendir_pid == 0) {
1310                 spin_unlock(&lli->lli_sa_lock);
1311                 return;
1312         }
1313
1314         lli->lli_opendir_key = NULL;
1315
1316         if (lli->lli_sai) {
1317                 struct l_wait_info lwi = { 0 };
1318                 struct ptlrpc_thread *thread = &lli->lli_sai->sai_thread;
1319
1320                 if (!thread_is_stopped(thread)) {
1321                         thread_set_flags(thread, SVC_STOPPING);
1322                         spin_unlock(&lli->lli_sa_lock);
1323                         cfs_waitq_signal(&thread->t_ctl_waitq);
1324
1325                         CDEBUG(D_READA, "stop statahead thread: [pid %d]\n",
1326                                current_pid());
1327                         l_wait_event(thread->t_ctl_waitq,
1328                                      thread_is_stopped(thread),
1329                                      &lwi);
1330                 } else {
1331                         spin_unlock(&lli->lli_sa_lock);
1332                 }
1333
1334                 /*
1335                  * Put the ref which was held when first statahead_enter.
1336                  * It maybe not the last ref for some statahead requests
1337                  * maybe inflight.
1338                  */
1339                 ll_sai_put(lli->lli_sai);
1340         } else {
1341                 lli->lli_opendir_pid = 0;
1342                 spin_unlock(&lli->lli_sa_lock);
1343         }
1344 }
1345
1346 enum {
1347         /**
1348          * not first dirent, or is "."
1349          */
1350         LS_NONE_FIRST_DE = 0,
1351         /**
1352          * the first non-hidden dirent
1353          */
1354         LS_FIRST_DE,
1355         /**
1356          * the first hidden dirent, that is "."
1357          */
1358         LS_FIRST_DOT_DE
1359 };
1360
1361 static int is_first_dirent(struct inode *dir, struct dentry *dentry)
1362 {
1363         struct ll_dir_chain   chain;
1364         struct qstr          *target = &dentry->d_name;
1365         struct page          *page;
1366         __u64                 pos    = 0;
1367         int                   dot_de;
1368         int                   rc     = LS_NONE_FIRST_DE;
1369         ENTRY;
1370
1371         ll_dir_chain_init(&chain);
1372         page = ll_get_dir_page(dir, pos, &chain);
1373
1374         while (1) {
1375                 struct lu_dirpage *dp;
1376                 struct lu_dirent  *ent;
1377
1378                 if (IS_ERR(page)) {
1379                         struct ll_inode_info *lli = ll_i2info(dir);
1380
1381                         rc = PTR_ERR(page);
1382                         CERROR("error reading dir "DFID" at "LPU64": "
1383                                "[rc %d] [parent %u]\n",
1384                                PFID(ll_inode2fid(dir)), pos,
1385                                rc, lli->lli_opendir_pid);
1386                         break;
1387                 }
1388
1389                 dp = page_address(page);
1390                 for (ent = lu_dirent_start(dp); ent != NULL;
1391                      ent = lu_dirent_next(ent)) {
1392                         __u64 hash;
1393                         int namelen;
1394                         char *name;
1395
1396                         hash = le64_to_cpu(ent->lde_hash);
1397                         /* The ll_get_dir_page() can return any page containing
1398                          * the given hash which may be not the start hash. */
1399                         if (unlikely(hash < pos))
1400                                 continue;
1401
1402                         namelen = le16_to_cpu(ent->lde_namelen);
1403                         if (unlikely(namelen == 0))
1404                                 /*
1405                                  * skip dummy record.
1406                                  */
1407                                 continue;
1408
1409                         name = ent->lde_name;
1410                         if (name[0] == '.') {
1411                                 if (namelen == 1)
1412                                         /*
1413                                          * skip "."
1414                                          */
1415                                         continue;
1416                                 else if (name[1] == '.' && namelen == 2)
1417                                         /*
1418                                          * skip ".."
1419                                          */
1420                                         continue;
1421                                 else
1422                                         dot_de = 1;
1423                         } else {
1424                                 dot_de = 0;
1425                         }
1426
1427                         if (dot_de && target->name[0] != '.') {
1428                                 CDEBUG(D_READA, "%.*s skip hidden file %.*s\n",
1429                                        target->len, target->name,
1430                                        namelen, name);
1431                                 continue;
1432                         }
1433
1434                         if (target->len != namelen ||
1435                             memcmp(target->name, name, namelen) != 0)
1436                                 rc = LS_NONE_FIRST_DE;
1437                         else if (!dot_de)
1438                                 rc = LS_FIRST_DE;
1439                         else
1440                                 rc = LS_FIRST_DOT_DE;
1441
1442                         ll_release_page(page, 0);
1443                         GOTO(out, rc);
1444                 }
1445                 pos = le64_to_cpu(dp->ldp_hash_end);
1446                 if (pos == MDS_DIR_END_OFF) {
1447                         /*
1448                          * End of directory reached.
1449                          */
1450                         ll_release_page(page, 0);
1451                         break;
1452                 } else if (1) {
1453                         /*
1454                          * chain is exhausted
1455                          * Normal case: continue to the next page.
1456                          */
1457                         ll_release_page(page, le32_to_cpu(dp->ldp_flags) &
1458                                               LDF_COLLIDE);
1459                         page = ll_get_dir_page(dir, pos, &chain);
1460                 } else {
1461                         /*
1462                          * go into overflow page.
1463                          */
1464                         LASSERT(le32_to_cpu(dp->ldp_flags) & LDF_COLLIDE);
1465                         ll_release_page(page, 1);
1466                 }
1467         }
1468         EXIT;
1469
1470 out:
1471         ll_dir_chain_fini(&chain);
1472         return rc;
1473 }
1474
1475 static void
1476 ll_sai_unplug(struct ll_statahead_info *sai, struct ll_sa_entry *entry)
1477 {
1478         struct ptlrpc_thread *thread = &sai->sai_thread;
1479         struct ll_sb_info    *sbi    = ll_i2sbi(sai->sai_inode);
1480         int                   hit;
1481         ENTRY;
1482
1483         if (entry != NULL && entry->se_stat == SA_ENTRY_SUCC)
1484                 hit = 1;
1485         else
1486                 hit = 0;
1487
1488         ll_sa_entry_fini(sai, entry);
1489         if (hit) {
1490                 sai->sai_hit++;
1491                 sai->sai_consecutive_miss = 0;
1492                 sai->sai_max = min(2 * sai->sai_max, sbi->ll_sa_max);
1493         } else {
1494                 struct ll_inode_info *lli = ll_i2info(sai->sai_inode);
1495
1496                 sai->sai_miss++;
1497                 sai->sai_consecutive_miss++;
1498                 if (sa_low_hit(sai) && thread_is_running(thread)) {
1499                         atomic_inc(&sbi->ll_sa_wrong);
1500                         CDEBUG(D_READA, "Statahead for dir "DFID" hit "
1501                                "ratio too low: hit/miss "LPU64"/"LPU64
1502                                ", sent/replied "LPU64"/"LPU64", stopping "
1503                                "statahead thread: pid %d\n",
1504                                PFID(&lli->lli_fid), sai->sai_hit,
1505                                sai->sai_miss, sai->sai_sent,
1506                                sai->sai_replied, current_pid());
1507                         spin_lock(&lli->lli_sa_lock);
1508                         if (!thread_is_stopped(thread))
1509                                 thread_set_flags(thread, SVC_STOPPING);
1510                         spin_unlock(&lli->lli_sa_lock);
1511                 }
1512         }
1513
1514         if (!thread_is_stopped(thread))
1515                 cfs_waitq_signal(&thread->t_ctl_waitq);
1516
1517         EXIT;
1518 }
1519
1520 /**
1521  * Start statahead thread if this is the first dir entry.
1522  * Otherwise if a thread is started already, wait it until it is ahead of me.
1523  * \retval 1       -- find entry with lock in cache, the caller needs to do
1524  *                    nothing.
1525  * \retval 0       -- find entry in cache, but without lock, the caller needs
1526  *                    refresh from MDS.
1527  * \retval others  -- the caller need to process as non-statahead.
1528  */
1529 int do_statahead_enter(struct inode *dir, struct dentry **dentryp,
1530                        int only_unplug)
1531 {
1532         struct ll_inode_info            *lli = ll_i2info(dir);
1533         struct ll_statahead_info        *sai = lli->lli_sai;
1534         struct dentry                   *parent;
1535         struct ll_sa_entry              *entry;
1536         struct ptlrpc_thread            *thread;
1537         struct l_wait_info               lwi = { 0 };
1538         int                              rc = 0;
1539         struct ll_inode_info            *plli;
1540         ENTRY;
1541
1542         LASSERT(lli->lli_opendir_pid == current_pid());
1543
1544         if (sai) {
1545                 thread = &sai->sai_thread;
1546                 if (unlikely(thread_is_stopped(thread) &&
1547                              cfs_list_empty(&sai->sai_entries_stated))) {
1548                         /* to release resource */
1549                         ll_stop_statahead(dir, lli->lli_opendir_key);
1550                         RETURN(-EAGAIN);
1551                 }
1552
1553                 if ((*dentryp)->d_name.name[0] == '.') {
1554                         if (sai->sai_ls_all ||
1555                             sai->sai_miss_hidden >= sai->sai_skip_hidden) {
1556                                 /*
1557                                  * Hidden dentry is the first one, or statahead
1558                                  * thread does not skip so many hidden dentries
1559                                  * before "sai_ls_all" enabled as below.
1560                                  */
1561                         } else {
1562                                 if (!sai->sai_ls_all)
1563                                         /*
1564                                          * It maybe because hidden dentry is not
1565                                          * the first one, "sai_ls_all" was not
1566                                          * set, then "ls -al" missed. Enable
1567                                          * "sai_ls_all" for such case.
1568                                          */
1569                                         sai->sai_ls_all = 1;
1570
1571                                 /*
1572                                  * Such "getattr" has been skipped before
1573                                  * "sai_ls_all" enabled as above.
1574                                  */
1575                                 sai->sai_miss_hidden++;
1576                                 RETURN(-EAGAIN);
1577                         }
1578                 }
1579
1580                 entry = ll_sa_entry_get_byname(sai, &(*dentryp)->d_name);
1581                 if (entry == NULL || only_unplug) {
1582                         ll_sai_unplug(sai, entry);
1583                         RETURN(entry ? 1 : -EAGAIN);
1584                 }
1585
1586                 /* if statahead is busy in readdir, help it do post-work */
1587                 while (!ll_sa_entry_stated(entry) &&
1588                        sai->sai_in_readpage &&
1589                        !sa_received_empty(sai))
1590                         ll_post_statahead(sai);
1591
1592                 if (!ll_sa_entry_stated(entry)) {
1593                         sai->sai_index_wait = entry->se_index;
1594                         lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(30), NULL,
1595                                                LWI_ON_SIGNAL_NOOP, NULL);
1596                         rc = l_wait_event(sai->sai_waitq,
1597                                           ll_sa_entry_stated(entry) ||
1598                                           thread_is_stopped(thread),
1599                                           &lwi);
1600                         if (rc < 0) {
1601                                 ll_sai_unplug(sai, entry);
1602                                 RETURN(-EAGAIN);
1603                         }
1604                 }
1605
1606                 if (entry->se_stat == SA_ENTRY_SUCC &&
1607                     entry->se_inode != NULL) {
1608                         struct inode *inode = entry->se_inode;
1609                         struct lookup_intent it = { .it_op = IT_GETATTR,
1610                                                     .d.lustre.it_lock_handle =
1611                                                      entry->se_handle };
1612                         __u64 bits;
1613
1614                         rc = md_revalidate_lock(ll_i2mdexp(dir), &it,
1615                                                 ll_inode2fid(inode), &bits);
1616                         if (rc == 1) {
1617                                 if ((*dentryp)->d_inode == NULL) {
1618                                         struct dentry *alias;
1619
1620                                         alias = ll_splice_alias(inode,
1621                                                                 *dentryp);
1622                                         if (IS_ERR(alias)) {
1623                                                 ll_sai_unplug(sai, entry);
1624                                                 RETURN(PTR_ERR(alias));
1625                                         }
1626                                         *dentryp = alias;
1627                                 } else if ((*dentryp)->d_inode != inode) {
1628                                         /* revalidate, but inode is recreated */
1629                                         CDEBUG(D_READA,
1630                                               "stale dentry %.*s inode %lu/%u, "
1631                                               "statahead inode %lu/%u\n",
1632                                               (*dentryp)->d_name.len,
1633                                               (*dentryp)->d_name.name,
1634                                               (*dentryp)->d_inode->i_ino,
1635                                               (*dentryp)->d_inode->i_generation,
1636                                               inode->i_ino,
1637                                               inode->i_generation);
1638                                         ll_sai_unplug(sai, entry);
1639                                         RETURN(-ESTALE);
1640                                 } else {
1641                                         iput(inode);
1642                                 }
1643                                 entry->se_inode = NULL;
1644
1645                                 if ((bits & MDS_INODELOCK_LOOKUP) &&
1646                                     d_lustre_invalid(*dentryp))
1647                                         d_lustre_revalidate(*dentryp);
1648                                 ll_intent_release(&it);
1649                         }
1650                 }
1651
1652                 ll_sai_unplug(sai, entry);
1653                 RETURN(rc);
1654         }
1655
1656         /* I am the "lli_opendir_pid" owner, only me can set "lli_sai". */
1657         rc = is_first_dirent(dir, *dentryp);
1658         if (rc == LS_NONE_FIRST_DE)
1659                 /* It is not "ls -{a}l" operation, no need statahead for it. */
1660                 GOTO(out, rc = -EAGAIN);
1661
1662         sai = ll_sai_alloc();
1663         if (sai == NULL)
1664                 GOTO(out, rc = -ENOMEM);
1665
1666         sai->sai_ls_all = (rc == LS_FIRST_DOT_DE);
1667         sai->sai_inode = igrab(dir);
1668         if (unlikely(sai->sai_inode == NULL)) {
1669                 CWARN("Do not start stat ahead on dying inode "DFID"\n",
1670                       PFID(&lli->lli_fid));
1671                 GOTO(out, rc = -ESTALE);
1672         }
1673
1674         /* get parent reference count here, and put it in ll_statahead_thread */
1675         parent = dget((*dentryp)->d_parent);
1676         if (unlikely(sai->sai_inode != parent->d_inode)) {
1677                 struct ll_inode_info *nlli = ll_i2info(parent->d_inode);
1678
1679                 CWARN("Race condition, someone changed %.*s just now: "
1680                       "old parent "DFID", new parent "DFID"\n",
1681                       (*dentryp)->d_name.len, (*dentryp)->d_name.name,
1682                       PFID(&lli->lli_fid), PFID(&nlli->lli_fid));
1683                 dput(parent);
1684                 iput(sai->sai_inode);
1685                 GOTO(out, rc = -EAGAIN);
1686         }
1687
1688         CDEBUG(D_READA, "start statahead thread: [pid %d] [parent %.*s]\n",
1689                current_pid(), parent->d_name.len, parent->d_name.name);
1690
1691         lli->lli_sai = sai;
1692
1693         plli = ll_i2info(parent->d_inode);
1694         rc = PTR_ERR(kthread_run(ll_statahead_thread, parent,
1695                                  "ll_sa_%u", plli->lli_opendir_pid));
1696         thread = &sai->sai_thread;
1697         if (IS_ERR_VALUE(rc)) {
1698                 CERROR("can't start ll_sa thread, rc: %d\n", rc);
1699                 dput(parent);
1700                 lli->lli_opendir_key = NULL;
1701                 thread_set_flags(thread, SVC_STOPPED);
1702                 thread_set_flags(&sai->sai_agl_thread, SVC_STOPPED);
1703                 ll_sai_put(sai);
1704                 LASSERT(lli->lli_sai == NULL);
1705                 RETURN(-EAGAIN);
1706         }
1707
1708         l_wait_event(thread->t_ctl_waitq,
1709                      thread_is_running(thread) || thread_is_stopped(thread),
1710                      &lwi);
1711
1712         /*
1713          * We don't stat-ahead for the first dirent since we are already in
1714          * lookup.
1715          */
1716         RETURN(-EAGAIN);
1717
1718 out:
1719         if (sai != NULL)
1720                 OBD_FREE_PTR(sai);
1721         spin_lock(&lli->lli_sa_lock);
1722         lli->lli_opendir_key = NULL;
1723         lli->lli_opendir_pid = 0;
1724         spin_unlock(&lli->lli_sa_lock);
1725         return rc;
1726 }