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

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.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, 2012, 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 "../include/obd_support.h"
  46 #include "../include/lustre_lite.h"
  47 #include "../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         struct list_head              se_link;
  63         /* link into sai->sai_entries_{received,stated} */
  64         struct list_head              se_list;
  65         /* link into sai hash table locally */
  66         struct list_head              se_hash;
  67         /* entry reference count */
  68         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 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         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         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 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 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 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 list_empty(&sai->sai_entries_received);
 163 }
 164
 165 static inline int agl_list_empty(struct ll_statahead_info *sai)
 166 {
 167         return 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
 204         entry_size = sizeof(struct ll_sa_entry) + (len & ~3) + 4;
 205         entry = kzalloc(entry_size, GFP_NOFS);
 206         if (unlikely(!entry))
 207                 return ERR_PTR(-ENOMEM);
 208
 209         CDEBUG(D_READA, "alloc sa entry %.*s(%p) index %llu\n",
 210                len, name, entry, index);
 211
 212         entry->se_index = index;
 213
 214         /*
 215          * Statahead entry reference rules:
 216          *
 217          * 1) When statahead entry is initialized, its reference is set as 2.
 218          *    One reference is used by the directory scanner. When the scanner
 219          *    searches the statahead cache for the given name, it can perform
 220          *    lockless hash lookup (only the scanner can remove entry from hash
 221          *    list), and once found, it needn't to call "atomic_inc()" for the
 222          *    entry reference. So the performance is improved. After using the
 223          *    statahead entry, the scanner will call "atomic_dec()" to drop the
 224          *    reference held when initialization. If it is the last reference,
 225          *    the statahead entry will be freed.
 226          *
 227          * 2) All other threads, including statahead thread and ptlrpcd thread,
 228          *    when they process the statahead entry, the reference for target
 229          *    should be held to guarantee the entry will not be released by the
 230          *    directory scanner. After processing the entry, these threads will
 231          *    drop the entry reference. If it is the last reference, the entry
 232          *    will be freed.
 233          *
 234          *    The second reference when initializes the statahead entry is used
 235          *    by the statahead thread, following the rule 2).
 236          */
 237         atomic_set(&entry->se_refcount, 2);
 238         entry->se_stat = SA_ENTRY_INIT;
 239         entry->se_size = entry_size;
 240         dname = (char *)entry + sizeof(struct ll_sa_entry);
 241         memcpy(dname, name, len);
 242         dname[len] = 0;
 243         entry->se_qstr.hash = full_name_hash(name, len);
 244         entry->se_qstr.len = len;
 245         entry->se_qstr.name = dname;
 246
 247         lli = ll_i2info(sai->sai_inode);
 248         spin_lock(&lli->lli_sa_lock);
 249         list_add_tail(&entry->se_link, &sai->sai_entries);
 250         INIT_LIST_HEAD(&entry->se_list);
 251         ll_sa_entry_enhash(sai, entry);
 252         spin_unlock(&lli->lli_sa_lock);
 253
 254         atomic_inc(&sai->sai_cache_count);
 255
 256         return entry;
 257 }
 258
 259 /*
 260  * Used by the directory scanner to search entry with name.
 261  *
 262  * Only the caller can remove the entry from hash, so it is unnecessary to hold
 263  * hash lock. It is caller's duty to release the init refcount on the entry, so
 264  * it is also unnecessary to increase refcount on the entry.
 265  */
 266 static struct ll_sa_entry *
 267 ll_sa_entry_get_byname(struct ll_statahead_info *sai, const struct qstr *qstr)
 268 {
 269         struct ll_sa_entry *entry;
 270         int i = ll_sa_entry_hash(qstr->hash);
 271
 272         list_for_each_entry(entry, &sai->sai_cache[i], se_hash) {
 273                 if (entry->se_qstr.hash == qstr->hash &&
 274                     entry->se_qstr.len == qstr->len &&
 275                     memcmp(entry->se_qstr.name, qstr->name, qstr->len) == 0)
 276                         return entry;
 277         }
 278         return NULL;
 279 }
 280
 281 /*
 282  * Used by the async getattr request callback to find entry with index.
 283  *
 284  * Inside lli_sa_lock to prevent others to change the list during the search.
 285  * It needs to increase entry refcount before returning to guarantee that the
 286  * entry cannot be freed by others.
 287  */
 288 static struct ll_sa_entry *
 289 ll_sa_entry_get_byindex(struct ll_statahead_info *sai, __u64 index)
 290 {
 291         struct ll_sa_entry *entry;
 292
 293         list_for_each_entry(entry, &sai->sai_entries, se_link) {
 294                 if (entry->se_index == index) {
 295                         LASSERT(atomic_read(&entry->se_refcount) > 0);
 296                         atomic_inc(&entry->se_refcount);
 297                         return entry;
 298                 }
 299                 if (entry->se_index > index)
 300                         break;
 301         }
 302         return NULL;
 303 }
 304
 305 static void ll_sa_entry_cleanup(struct ll_statahead_info *sai,
 306                                  struct ll_sa_entry *entry)
 307 {
 308         struct md_enqueue_info *minfo = entry->se_minfo;
 309         struct ptlrpc_request  *req   = entry->se_req;
 310
 311         if (minfo) {
 312                 entry->se_minfo = NULL;
 313                 ll_intent_release(&minfo->mi_it);
 314                 iput(minfo->mi_dir);
 315                 OBD_FREE_PTR(minfo);
 316         }
 317
 318         if (req) {
 319                 entry->se_req = NULL;
 320                 ptlrpc_req_finished(req);
 321         }
 322 }
 323
 324 static void ll_sa_entry_put(struct ll_statahead_info *sai,
 325                              struct ll_sa_entry *entry)
 326 {
 327         if (atomic_dec_and_test(&entry->se_refcount)) {
 328                 CDEBUG(D_READA, "free sa entry %.*s(%p) index %llu\n",
 329                        entry->se_qstr.len, entry->se_qstr.name, entry,
 330                        entry->se_index);
 331
 332                 LASSERT(list_empty(&entry->se_link));
 333                 LASSERT(list_empty(&entry->se_list));
 334                 LASSERT(ll_sa_entry_unhashed(entry));
 335
 336                 ll_sa_entry_cleanup(sai, entry);
 337                 if (entry->se_inode)
 338                         iput(entry->se_inode);
 339
 340                 OBD_FREE(entry, entry->se_size);
 341                 atomic_dec(&sai->sai_cache_count);
 342         }
 343 }
 344
 345 static inline void
 346 do_sa_entry_fini(struct ll_statahead_info *sai, struct ll_sa_entry *entry)
 347 {
 348         struct ll_inode_info *lli = ll_i2info(sai->sai_inode);
 349
 350         LASSERT(!ll_sa_entry_unhashed(entry));
 351         LASSERT(!list_empty(&entry->se_link));
 352
 353         ll_sa_entry_unhash(sai, entry);
 354
 355         spin_lock(&lli->lli_sa_lock);
 356         entry->se_stat = SA_ENTRY_DEST;
 357         list_del_init(&entry->se_link);
 358         if (likely(!list_empty(&entry->se_list)))
 359                 list_del_init(&entry->se_list);
 360         spin_unlock(&lli->lli_sa_lock);
 361
 362         ll_sa_entry_put(sai, entry);
 363 }
 364
 365 /*
 366  * Delete it from sai_entries_stated list when fini.
 367  */
 368 static void
 369 ll_sa_entry_fini(struct ll_statahead_info *sai, struct ll_sa_entry *entry)
 370 {
 371         struct ll_sa_entry *pos, *next;
 372
 373         if (entry)
 374                 do_sa_entry_fini(sai, entry);
 375
 376         /* drop old entry, only 'scanner' process does this, no need to lock */
 377         list_for_each_entry_safe(pos, next, &sai->sai_entries, se_link) {
 378                 if (!is_omitted_entry(sai, pos->se_index))
 379                         break;
 380                 do_sa_entry_fini(sai, pos);
 381         }
 382 }
 383
 384 /*
 385  * Inside lli_sa_lock.
 386  */
 387 static void
 388 do_sa_entry_to_stated(struct ll_statahead_info *sai,
 389                       struct ll_sa_entry *entry, se_stat_t stat)
 390 {
 391         struct ll_sa_entry *se;
 392         struct list_head         *pos = &sai->sai_entries_stated;
 393
 394         if (!list_empty(&entry->se_list))
 395                 list_del_init(&entry->se_list);
 396
 397         list_for_each_entry_reverse(se, &sai->sai_entries_stated, se_list) {
 398                 if (se->se_index < entry->se_index) {
 399                         pos = &se->se_list;
 400                         break;
 401                 }
 402         }
 403
 404         list_add(&entry->se_list, pos);
 405         entry->se_stat = stat;
 406 }
 407
 408 /*
 409  * Move entry to sai_entries_stated and sort with the index.
 410  * \retval 1    -- entry to be destroyed.
 411  * \retval 0    -- entry is inserted into stated list.
 412  */
 413 static int
 414 ll_sa_entry_to_stated(struct ll_statahead_info *sai,
 415                       struct ll_sa_entry *entry, se_stat_t stat)
 416 {
 417         struct ll_inode_info *lli = ll_i2info(sai->sai_inode);
 418         int                ret = 1;
 419
 420         ll_sa_entry_cleanup(sai, entry);
 421
 422         spin_lock(&lli->lli_sa_lock);
 423         if (likely(entry->se_stat != SA_ENTRY_DEST)) {
 424                 do_sa_entry_to_stated(sai, entry, stat);
 425                 ret = 0;
 426         }
 427         spin_unlock(&lli->lli_sa_lock);
 428
 429         return ret;
 430 }
 431
 432 /*
 433  * Insert inode into the list of sai_entries_agl.
 434  */
 435 static void ll_agl_add(struct ll_statahead_info *sai,
 436                        struct inode *inode, int index)
 437 {
 438         struct ll_inode_info *child  = ll_i2info(inode);
 439         struct ll_inode_info *parent = ll_i2info(sai->sai_inode);
 440         int                added  = 0;
 441
 442         spin_lock(&child->lli_agl_lock);
 443         if (child->lli_agl_index == 0) {
 444                 child->lli_agl_index = index;
 445                 spin_unlock(&child->lli_agl_lock);
 446
 447                 LASSERT(list_empty(&child->lli_agl_list));
 448
 449                 igrab(inode);
 450                 spin_lock(&parent->lli_agl_lock);
 451                 if (agl_list_empty(sai))
 452                         added = 1;
 453                 list_add_tail(&child->lli_agl_list, &sai->sai_entries_agl);
 454                 spin_unlock(&parent->lli_agl_lock);
 455         } else {
 456                 spin_unlock(&child->lli_agl_lock);
 457         }
 458
 459         if (added > 0)
 460                 wake_up(&sai->sai_agl_thread.t_ctl_waitq);
 461 }
 462
 463 static struct ll_statahead_info *ll_sai_alloc(void)
 464 {
 465         struct ll_statahead_info *sai;
 466         int                    i;
 467
 468         sai = kzalloc(sizeof(*sai), GFP_NOFS);
 469         if (!sai)
 470                 return NULL;
 471
 472         atomic_set(&sai->sai_refcount, 1);
 473
 474         spin_lock(&sai_generation_lock);
 475         sai->sai_generation = ++sai_generation;
 476         if (unlikely(sai_generation == 0))
 477                 sai->sai_generation = ++sai_generation;
 478         spin_unlock(&sai_generation_lock);
 479
 480         sai->sai_max = LL_SA_RPC_MIN;
 481         sai->sai_index = 1;
 482         init_waitqueue_head(&sai->sai_waitq);
 483         init_waitqueue_head(&sai->sai_thread.t_ctl_waitq);
 484         init_waitqueue_head(&sai->sai_agl_thread.t_ctl_waitq);
 485
 486         INIT_LIST_HEAD(&sai->sai_entries);
 487         INIT_LIST_HEAD(&sai->sai_entries_received);
 488         INIT_LIST_HEAD(&sai->sai_entries_stated);
 489         INIT_LIST_HEAD(&sai->sai_entries_agl);
 490
 491         for (i = 0; i < LL_SA_CACHE_SIZE; i++) {
 492                 INIT_LIST_HEAD(&sai->sai_cache[i]);
 493                 spin_lock_init(&sai->sai_cache_lock[i]);
 494         }
 495         atomic_set(&sai->sai_cache_count, 0);
 496
 497         return sai;
 498 }
 499
 500 static inline struct ll_statahead_info *
 501 ll_sai_get(struct ll_statahead_info *sai)
 502 {
 503         atomic_inc(&sai->sai_refcount);
 504         return sai;
 505 }
 506
 507 static void ll_sai_put(struct ll_statahead_info *sai)
 508 {
 509         struct inode     *inode = sai->sai_inode;
 510         struct ll_inode_info *lli   = ll_i2info(inode);
 511
 512         if (atomic_dec_and_lock(&sai->sai_refcount, &lli->lli_sa_lock)) {
 513                 struct ll_sa_entry *entry, *next;
 514
 515                 if (unlikely(atomic_read(&sai->sai_refcount) > 0)) {
 516                         /* It is race case, the interpret callback just hold
 517                          * a reference count */
 518                         spin_unlock(&lli->lli_sa_lock);
 519                         return;
 520                 }
 521
 522                 LASSERT(lli->lli_opendir_key == NULL);
 523                 LASSERT(thread_is_stopped(&sai->sai_thread));
 524                 LASSERT(thread_is_stopped(&sai->sai_agl_thread));
 525
 526                 lli->lli_sai = NULL;
 527                 lli->lli_opendir_pid = 0;
 528                 spin_unlock(&lli->lli_sa_lock);
 529
 530                 if (sai->sai_sent > sai->sai_replied)
 531                         CDEBUG(D_READA, "statahead for dir "DFID
 532                               " does not finish: [sent:%llu] [replied:%llu]\n",
 533                               PFID(&lli->lli_fid),
 534                               sai->sai_sent, sai->sai_replied);
 535
 536                 list_for_each_entry_safe(entry, next,
 537                                              &sai->sai_entries, se_link)
 538                         do_sa_entry_fini(sai, entry);
 539
 540                 LASSERT(list_empty(&sai->sai_entries));
 541                 LASSERT(sa_received_empty(sai));
 542                 LASSERT(list_empty(&sai->sai_entries_stated));
 543
 544                 LASSERT(atomic_read(&sai->sai_cache_count) == 0);
 545                 LASSERT(agl_list_empty(sai));
 546
 547                 iput(inode);
 548                 OBD_FREE_PTR(sai);
 549         }
 550 }
 551
 552 /* Do NOT forget to drop inode refcount when into sai_entries_agl. */
 553 static void ll_agl_trigger(struct inode *inode, struct ll_statahead_info *sai)
 554 {
 555         struct ll_inode_info *lli   = ll_i2info(inode);
 556         __u64            index = lli->lli_agl_index;
 557         int                rc;
 558
 559         LASSERT(list_empty(&lli->lli_agl_list));
 560
 561         /* AGL maybe fall behind statahead with one entry */
 562         if (is_omitted_entry(sai, index + 1)) {
 563                 lli->lli_agl_index = 0;
 564                 iput(inode);
 565                 return;
 566         }
 567
 568         /* Someone is in glimpse (sync or async), do nothing. */
 569         rc = down_write_trylock(&lli->lli_glimpse_sem);
 570         if (rc == 0) {
 571                 lli->lli_agl_index = 0;
 572                 iput(inode);
 573                 return;
 574         }
 575
 576         /*
 577          * Someone triggered glimpse within 1 sec before.
 578          * 1) The former glimpse succeeded with glimpse lock granted by OST, and
 579          *    if the lock is still cached on client, AGL needs to do nothing. If
 580          *    it is cancelled by other client, AGL maybe cannot obtain new lock
 581          *    for no glimpse callback triggered by AGL.
 582          * 2) The former glimpse succeeded, but OST did not grant glimpse lock.
 583          *    Under such case, it is quite possible that the OST will not grant
 584          *    glimpse lock for AGL also.
 585          * 3) The former glimpse failed, compared with other two cases, it is
 586          *    relative rare. AGL can ignore such case, and it will not muchly
 587          *    affect the performance.
 588          */
 589         if (lli->lli_glimpse_time != 0 &&
 590             time_before(cfs_time_shift(-1), lli->lli_glimpse_time)) {
 591                 up_write(&lli->lli_glimpse_sem);
 592                 lli->lli_agl_index = 0;
 593                 iput(inode);
 594                 return;
 595         }
 596
 597         CDEBUG(D_READA, "Handling (init) async glimpse: inode = "
 598                DFID", idx = %llu\n", PFID(&lli->lli_fid), index);
 599
 600         cl_agl(inode);
 601         lli->lli_agl_index = 0;
 602         lli->lli_glimpse_time = cfs_time_current();
 603         up_write(&lli->lli_glimpse_sem);
 604
 605         CDEBUG(D_READA, "Handled (init) async glimpse: inode= "
 606                DFID", idx = %llu, rc = %d\n",
 607                PFID(&lli->lli_fid), index, rc);
 608
 609         iput(inode);
 610 }
 611
 612 static void ll_post_statahead(struct ll_statahead_info *sai)
 613 {
 614         struct inode       *dir   = sai->sai_inode;
 615         struct inode       *child;
 616         struct ll_inode_info   *lli   = ll_i2info(dir);
 617         struct ll_sa_entry     *entry;
 618         struct md_enqueue_info *minfo;
 619         struct lookup_intent   *it;
 620         struct ptlrpc_request  *req;
 621         struct mdt_body *body;
 622         int                  rc    = 0;
 623
 624         spin_lock(&lli->lli_sa_lock);
 625         if (unlikely(sa_received_empty(sai))) {
 626                 spin_unlock(&lli->lli_sa_lock);
 627                 return;
 628         }
 629         entry = sa_first_received_entry(sai);
 630         atomic_inc(&entry->se_refcount);
 631         list_del_init(&entry->se_list);
 632         spin_unlock(&lli->lli_sa_lock);
 633
 634         LASSERT(entry->se_handle != 0);
 635
 636         minfo = entry->se_minfo;
 637         it = &minfo->mi_it;
 638         req = entry->se_req;
 639         body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
 640         if (body == NULL) {
 641                 rc = -EFAULT;
 642                 goto out;
 643         }
 644
 645         child = entry->se_inode;
 646         if (child == NULL) {
 647                 /*
 648                  * lookup.
 649                  */
 650                 LASSERT(fid_is_zero(&minfo->mi_data.op_fid2));
 651
 652                 /* XXX: No fid in reply, this is probably cross-ref case.
 653                  * SA can't handle it yet. */
 654                 if (body->valid & OBD_MD_MDS) {
 655                         rc = -EAGAIN;
 656                         goto out;
 657                 }
 658         } else {
 659                 /*
 660                  * revalidate.
 661                  */
 662                 /* unlinked and re-created with the same name */
 663                 if (unlikely(!lu_fid_eq(&minfo->mi_data.op_fid2, &body->fid1))){
 664                         entry->se_inode = NULL;
 665                         iput(child);
 666                         child = NULL;
 667                 }
 668         }
 669
 670         it->d.lustre.it_lock_handle = entry->se_handle;
 671         rc = md_revalidate_lock(ll_i2mdexp(dir), it, ll_inode2fid(dir), NULL);
 672         if (rc != 1) {
 673                 rc = -EAGAIN;
 674                 goto out;
 675         }
 676
 677         rc = ll_prep_inode(&child, req, dir->i_sb, it);
 678         if (rc)
 679                 goto out;
 680
 681         CDEBUG(D_DLMTRACE, "setting l_data to inode %p (%lu/%u)\n",
 682                child, child->i_ino, child->i_generation);
 683         ll_set_lock_data(ll_i2sbi(dir)->ll_md_exp, child, it, NULL);
 684
 685         entry->se_inode = child;
 686
 687         if (agl_should_run(sai, child))
 688                 ll_agl_add(sai, child, entry->se_index);
 689
 690 out:
 691         /* The "ll_sa_entry_to_stated()" will drop related ldlm ibits lock
 692          * reference count by calling "ll_intent_drop_lock()" in spite of the
 693          * above operations failed or not. Do not worry about calling
 694          * "ll_intent_drop_lock()" more than once. */
 695         rc = ll_sa_entry_to_stated(sai, entry,
 696                                    rc < 0 ? SA_ENTRY_INVA : SA_ENTRY_SUCC);
 697         if (rc == 0 && entry->se_index == sai->sai_index_wait)
 698                 wake_up(&sai->sai_waitq);
 699         ll_sa_entry_put(sai, entry);
 700 }
 701
 702 static int ll_statahead_interpret(struct ptlrpc_request *req,
 703                                   struct md_enqueue_info *minfo, int rc)
 704 {
 705         struct lookup_intent     *it  = &minfo->mi_it;
 706         struct inode         *dir = minfo->mi_dir;
 707         struct ll_inode_info     *lli = ll_i2info(dir);
 708         struct ll_statahead_info *sai = NULL;
 709         struct ll_sa_entry       *entry;
 710         int                    wakeup;
 711
 712         if (it_disposition(it, DISP_LOOKUP_NEG))
 713                 rc = -ENOENT;
 714
 715         spin_lock(&lli->lli_sa_lock);
 716         /* stale entry */
 717         if (unlikely(lli->lli_sai == NULL ||
 718                      lli->lli_sai->sai_generation != minfo->mi_generation)) {
 719                 spin_unlock(&lli->lli_sa_lock);
 720                 rc = -ESTALE;
 721                 goto out;
 722         } else {
 723                 sai = ll_sai_get(lli->lli_sai);
 724                 if (unlikely(!thread_is_running(&sai->sai_thread))) {
 725                         sai->sai_replied++;
 726                         spin_unlock(&lli->lli_sa_lock);
 727                         rc = -EBADFD;
 728                         goto out;
 729                 }
 730
 731                 entry = ll_sa_entry_get_byindex(sai, minfo->mi_cbdata);
 732                 if (entry == NULL) {
 733                         sai->sai_replied++;
 734                         spin_unlock(&lli->lli_sa_lock);
 735                         rc = -EIDRM;
 736                         goto out;
 737                 }
 738
 739                 if (rc != 0) {
 740                         do_sa_entry_to_stated(sai, entry, SA_ENTRY_INVA);
 741                         wakeup = (entry->se_index == sai->sai_index_wait);
 742                 } else {
 743                         entry->se_minfo = minfo;
 744                         entry->se_req = ptlrpc_request_addref(req);
 745                         /* Release the async ibits lock ASAP to avoid deadlock
 746                          * when statahead thread tries to enqueue lock on parent
 747                          * for readpage and other tries to enqueue lock on child
 748                          * with parent's lock held, for example: unlink. */
 749                         entry->se_handle = it->d.lustre.it_lock_handle;
 750                         ll_intent_drop_lock(it);
 751                         wakeup = sa_received_empty(sai);
 752                         list_add_tail(&entry->se_list,
 753                                           &sai->sai_entries_received);
 754                 }
 755                 sai->sai_replied++;
 756                 spin_unlock(&lli->lli_sa_lock);
 757
 758                 ll_sa_entry_put(sai, entry);
 759                 if (wakeup)
 760                         wake_up(&sai->sai_thread.t_ctl_waitq);
 761         }
 762
 763 out:
 764         if (rc != 0) {
 765                 ll_intent_release(it);
 766                 iput(dir);
 767                 OBD_FREE_PTR(minfo);
 768         }
 769         if (sai != NULL)
 770                 ll_sai_put(sai);
 771         return rc;
 772 }
 773
 774 static void sa_args_fini(struct md_enqueue_info *minfo,
 775                          struct ldlm_enqueue_info *einfo)
 776 {
 777         LASSERT(minfo && einfo);
 778         iput(minfo->mi_dir);
 779         capa_put(minfo->mi_data.op_capa1);
 780         capa_put(minfo->mi_data.op_capa2);
 781         OBD_FREE_PTR(minfo);
 782         OBD_FREE_PTR(einfo);
 783 }
 784
 785 /**
 786  * There is race condition between "capa_put" and "ll_statahead_interpret" for
 787  * accessing "op_data.op_capa[1,2]" as following:
 788  * "capa_put" releases "op_data.op_capa[1,2]"'s reference count after calling
 789  * "md_intent_getattr_async". But "ll_statahead_interpret" maybe run first, and
 790  * fill "op_data.op_capa[1,2]" as POISON, then cause "capa_put" access invalid
 791  * "ocapa". So here reserve "op_data.op_capa[1,2]" in "pcapa" before calling
 792  * "md_intent_getattr_async".
 793  */
 794 static int sa_args_init(struct inode *dir, struct inode *child,
 795                         struct ll_sa_entry *entry, struct md_enqueue_info **pmi,
 796                         struct ldlm_enqueue_info **pei,
 797                         struct obd_capa **pcapa)
 798 {
 799         struct qstr           *qstr = &entry->se_qstr;
 800         struct ll_inode_info     *lli  = ll_i2info(dir);
 801         struct md_enqueue_info   *minfo;
 802         struct ldlm_enqueue_info *einfo;
 803         struct md_op_data       *op_data;
 804
 805         einfo = kzalloc(sizeof(*einfo), GFP_NOFS);
 806         if (!einfo)
 807                 return -ENOMEM;
 808
 809         minfo = kzalloc(sizeof(*minfo), GFP_NOFS);
 810         if (!minfo) {
 811                 OBD_FREE_PTR(einfo);
 812                 return -ENOMEM;
 813         }
 814
 815         op_data = ll_prep_md_op_data(&minfo->mi_data, dir, child, qstr->name,
 816                                      qstr->len, 0, LUSTRE_OPC_ANY, NULL);
 817         if (IS_ERR(op_data)) {
 818                 OBD_FREE_PTR(einfo);
 819                 OBD_FREE_PTR(minfo);
 820                 return PTR_ERR(op_data);
 821         }
 822
 823         minfo->mi_it.it_op = IT_GETATTR;
 824         minfo->mi_dir = igrab(dir);
 825         minfo->mi_cb = ll_statahead_interpret;
 826         minfo->mi_generation = lli->lli_sai->sai_generation;
 827         minfo->mi_cbdata = entry->se_index;
 828
 829         einfo->ei_type   = LDLM_IBITS;
 830         einfo->ei_mode   = it_to_lock_mode(&minfo->mi_it);
 831         einfo->ei_cb_bl  = ll_md_blocking_ast;
 832         einfo->ei_cb_cp  = ldlm_completion_ast;
 833         einfo->ei_cb_gl  = NULL;
 834         einfo->ei_cbdata = NULL;
 835
 836         *pmi = minfo;
 837         *pei = einfo;
 838         pcapa[0] = op_data->op_capa1;
 839         pcapa[1] = op_data->op_capa2;
 840
 841         return 0;
 842 }
 843
 844 static int do_sa_lookup(struct inode *dir, struct ll_sa_entry *entry)
 845 {
 846         struct md_enqueue_info   *minfo;
 847         struct ldlm_enqueue_info *einfo;
 848         struct obd_capa   *capas[2];
 849         int                    rc;
 850
 851         rc = sa_args_init(dir, NULL, entry, &minfo, &einfo, capas);
 852         if (rc)
 853                 return rc;
 854
 855         rc = md_intent_getattr_async(ll_i2mdexp(dir), minfo, einfo);
 856         if (!rc) {
 857                 capa_put(capas[0]);
 858                 capa_put(capas[1]);
 859         } else {
 860                 sa_args_fini(minfo, einfo);
 861         }
 862
 863         return rc;
 864 }
 865
 866 /**
 867  * similar to ll_revalidate_it().
 868  * \retval      1 -- dentry valid
 869  * \retval      0 -- will send stat-ahead request
 870  * \retval others -- prepare stat-ahead request failed
 871  */
 872 static int do_sa_revalidate(struct inode *dir, struct ll_sa_entry *entry,
 873                             struct dentry *dentry)
 874 {
 875         struct inode         *inode = dentry->d_inode;
 876         struct lookup_intent      it = { .it_op = IT_GETATTR,
 877                                          .d.lustre.it_lock_handle = 0 };
 878         struct md_enqueue_info   *minfo;
 879         struct ldlm_enqueue_info *einfo;
 880         struct obd_capa   *capas[2];
 881         int rc;
 882
 883         if (unlikely(inode == NULL))
 884                 return 1;
 885
 886         if (d_mountpoint(dentry))
 887                 return 1;
 888
 889         entry->se_inode = igrab(inode);
 890         rc = md_revalidate_lock(ll_i2mdexp(dir), &it, ll_inode2fid(inode),
 891                                 NULL);
 892         if (rc == 1) {
 893                 entry->se_handle = it.d.lustre.it_lock_handle;
 894                 ll_intent_release(&it);
 895                 return 1;
 896         }
 897
 898         rc = sa_args_init(dir, inode, entry, &minfo, &einfo, capas);
 899         if (rc) {
 900                 entry->se_inode = NULL;
 901                 iput(inode);
 902                 return rc;
 903         }
 904
 905         rc = md_intent_getattr_async(ll_i2mdexp(dir), minfo, einfo);
 906         if (!rc) {
 907                 capa_put(capas[0]);
 908                 capa_put(capas[1]);
 909         } else {
 910                 entry->se_inode = NULL;
 911                 iput(inode);
 912                 sa_args_fini(minfo, einfo);
 913         }
 914
 915         return rc;
 916 }
 917
 918 static void ll_statahead_one(struct dentry *parent, const char* entry_name,
 919                              int entry_name_len)
 920 {
 921         struct inode         *dir    = parent->d_inode;
 922         struct ll_inode_info     *lli    = ll_i2info(dir);
 923         struct ll_statahead_info *sai    = lli->lli_sai;
 924         struct dentry       *dentry = NULL;
 925         struct ll_sa_entry       *entry;
 926         int                    rc;
 927         int                    rc1;
 928
 929         entry = ll_sa_entry_alloc(sai, sai->sai_index, entry_name,
 930                                   entry_name_len);
 931         if (IS_ERR(entry))
 932                 return;
 933
 934         dentry = d_lookup(parent, &entry->se_qstr);
 935         if (!dentry) {
 936                 rc = do_sa_lookup(dir, entry);
 937         } else {
 938                 rc = do_sa_revalidate(dir, entry, dentry);
 939                 if (rc == 1 && agl_should_run(sai, dentry->d_inode))
 940                         ll_agl_add(sai, dentry->d_inode, entry->se_index);
 941         }
 942
 943         if (dentry != NULL)
 944                 dput(dentry);
 945
 946         if (rc) {
 947                 rc1 = ll_sa_entry_to_stated(sai, entry,
 948                                         rc < 0 ? SA_ENTRY_INVA : SA_ENTRY_SUCC);
 949                 if (rc1 == 0 && entry->se_index == sai->sai_index_wait)
 950                         wake_up(&sai->sai_waitq);
 951         } else {
 952                 sai->sai_sent++;
 953         }
 954
 955         sai->sai_index++;
 956         /* drop one refcount on entry by ll_sa_entry_alloc */
 957         ll_sa_entry_put(sai, entry);
 958 }
 959
 960 static int ll_agl_thread(void *arg)
 961 {
 962         struct dentry       *parent = (struct dentry *)arg;
 963         struct inode         *dir    = parent->d_inode;
 964         struct ll_inode_info     *plli   = ll_i2info(dir);
 965         struct ll_inode_info     *clli;
 966         struct ll_sb_info       *sbi    = ll_i2sbi(dir);
 967         struct ll_statahead_info *sai    = ll_sai_get(plli->lli_sai);
 968         struct ptlrpc_thread     *thread = &sai->sai_agl_thread;
 969         struct l_wait_info      lwi    = { 0 };
 970
 971         thread->t_pid = current_pid();
 972         CDEBUG(D_READA, "agl thread started: sai %p, parent %pd\n",
 973                sai, parent);
 974
 975         atomic_inc(&sbi->ll_agl_total);
 976         spin_lock(&plli->lli_agl_lock);
 977         sai->sai_agl_valid = 1;
 978         if (thread_is_init(thread))
 979                 /* If someone else has changed the thread state
 980                  * (e.g. already changed to SVC_STOPPING), we can't just
 981                  * blindly overwrite that setting. */
 982                 thread_set_flags(thread, SVC_RUNNING);
 983         spin_unlock(&plli->lli_agl_lock);
 984         wake_up(&thread->t_ctl_waitq);
 985
 986         while (1) {
 987                 l_wait_event(thread->t_ctl_waitq,
 988                              !agl_list_empty(sai) ||
 989                              !thread_is_running(thread),
 990                              &lwi);
 991
 992                 if (!thread_is_running(thread))
 993                         break;
 994
 995                 spin_lock(&plli->lli_agl_lock);
 996                 /* The statahead thread maybe help to process AGL entries,
 997                  * so check whether list empty again. */
 998                 if (!agl_list_empty(sai)) {
 999                         clli = agl_first_entry(sai);
1000                         list_del_init(&clli->lli_agl_list);
1001                         spin_unlock(&plli->lli_agl_lock);
1002                         ll_agl_trigger(&clli->lli_vfs_inode, sai);
1003                 } else {
1004                         spin_unlock(&plli->lli_agl_lock);
1005                 }
1006         }
1007
1008         spin_lock(&plli->lli_agl_lock);
1009         sai->sai_agl_valid = 0;
1010         while (!agl_list_empty(sai)) {
1011                 clli = agl_first_entry(sai);
1012                 list_del_init(&clli->lli_agl_list);
1013                 spin_unlock(&plli->lli_agl_lock);
1014                 clli->lli_agl_index = 0;
1015                 iput(&clli->lli_vfs_inode);
1016                 spin_lock(&plli->lli_agl_lock);
1017         }
1018         thread_set_flags(thread, SVC_STOPPED);
1019         spin_unlock(&plli->lli_agl_lock);
1020         wake_up(&thread->t_ctl_waitq);
1021         ll_sai_put(sai);
1022         CDEBUG(D_READA, "agl thread stopped: sai %p, parent %pd\n",
1023                sai, parent);
1024         return 0;
1025 }
1026
1027 static void ll_start_agl(struct dentry *parent, struct ll_statahead_info *sai)
1028 {
1029         struct ptlrpc_thread *thread = &sai->sai_agl_thread;
1030         struct l_wait_info    lwi    = { 0 };
1031         struct ll_inode_info  *plli;
1032         struct task_struct *task;
1033
1034         CDEBUG(D_READA, "start agl thread: sai %p, parent %pd\n",
1035                sai, parent);
1036
1037         plli = ll_i2info(parent->d_inode);
1038         task = kthread_run(ll_agl_thread, parent,
1039                                "ll_agl_%u", plli->lli_opendir_pid);
1040         if (IS_ERR(task)) {
1041                 CERROR("can't start ll_agl thread, rc: %ld\n", PTR_ERR(task));
1042                 thread_set_flags(thread, SVC_STOPPED);
1043                 return;
1044         }
1045
1046         l_wait_event(thread->t_ctl_waitq,
1047                      thread_is_running(thread) || thread_is_stopped(thread),
1048                      &lwi);
1049 }
1050
1051 static int ll_statahead_thread(void *arg)
1052 {
1053         struct dentry       *parent = (struct dentry *)arg;
1054         struct inode         *dir    = parent->d_inode;
1055         struct ll_inode_info     *plli   = ll_i2info(dir);
1056         struct ll_inode_info     *clli;
1057         struct ll_sb_info       *sbi    = ll_i2sbi(dir);
1058         struct ll_statahead_info *sai    = ll_sai_get(plli->lli_sai);
1059         struct ptlrpc_thread     *thread = &sai->sai_thread;
1060         struct ptlrpc_thread *agl_thread = &sai->sai_agl_thread;
1061         struct page           *page;
1062         __u64                pos    = 0;
1063         int                    first  = 0;
1064         int                    rc     = 0;
1065         struct ll_dir_chain       chain;
1066         struct l_wait_info      lwi    = { 0 };
1067
1068         thread->t_pid = current_pid();
1069         CDEBUG(D_READA, "statahead thread starting: sai %p, parent %pd\n",
1070                sai, parent);
1071
1072         if (sbi->ll_flags & LL_SBI_AGL_ENABLED)
1073                 ll_start_agl(parent, sai);
1074
1075         atomic_inc(&sbi->ll_sa_total);
1076         spin_lock(&plli->lli_sa_lock);
1077         if (thread_is_init(thread))
1078                 /* If someone else has changed the thread state
1079                  * (e.g. already changed to SVC_STOPPING), we can't just
1080                  * blindly overwrite that setting. */
1081                 thread_set_flags(thread, SVC_RUNNING);
1082         spin_unlock(&plli->lli_sa_lock);
1083         wake_up(&thread->t_ctl_waitq);
1084
1085         ll_dir_chain_init(&chain);
1086         page = ll_get_dir_page(dir, pos, &chain);
1087
1088         while (1) {
1089                 struct lu_dirpage *dp;
1090                 struct lu_dirent  *ent;
1091
1092                 if (IS_ERR(page)) {
1093                         rc = PTR_ERR(page);
1094                         CDEBUG(D_READA, "error reading dir "DFID" at %llu/%llu: [rc %d] [parent %u]\n",
1095                                PFID(ll_inode2fid(dir)), pos, sai->sai_index,
1096                                rc, plli->lli_opendir_pid);
1097                         goto out;
1098                 }
1099
1100                 dp = page_address(page);
1101                 for (ent = lu_dirent_start(dp); ent != NULL;
1102                      ent = lu_dirent_next(ent)) {
1103                         __u64 hash;
1104                         int namelen;
1105                         char *name;
1106
1107                         hash = le64_to_cpu(ent->lde_hash);
1108                         if (unlikely(hash < pos))
1109                                 /*
1110                                  * Skip until we find target hash value.
1111                                  */
1112                                 continue;
1113
1114                         namelen = le16_to_cpu(ent->lde_namelen);
1115                         if (unlikely(namelen == 0))
1116                                 /*
1117                                  * Skip dummy record.
1118                                  */
1119                                 continue;
1120
1121                         name = ent->lde_name;
1122                         if (name[0] == '.') {
1123                                 if (namelen == 1) {
1124                                         /*
1125                                          * skip "."
1126                                          */
1127                                         continue;
1128                                 } else if (name[1] == '.' && namelen == 2) {
1129                                         /*
1130                                          * skip ".."
1131                                          */
1132                                         continue;
1133                                 } else if (!sai->sai_ls_all) {
1134                                         /*
1135                                          * skip hidden files.
1136                                          */
1137                                         sai->sai_skip_hidden++;
1138                                         continue;
1139                                 }
1140                         }
1141
1142                         /*
1143                          * don't stat-ahead first entry.
1144                          */
1145                         if (unlikely(++first == 1))
1146                                 continue;
1147
1148 keep_it:
1149                         l_wait_event(thread->t_ctl_waitq,
1150                                      !sa_sent_full(sai) ||
1151                                      !sa_received_empty(sai) ||
1152                                      !agl_list_empty(sai) ||
1153                                      !thread_is_running(thread),
1154                                      &lwi);
1155
1156 interpret_it:
1157                         while (!sa_received_empty(sai))
1158                                 ll_post_statahead(sai);
1159
1160                         if (unlikely(!thread_is_running(thread))) {
1161                                 ll_release_page(page, 0);
1162                                 rc = 0;
1163                                 goto out;
1164                         }
1165
1166                         /* If no window for metadata statahead, but there are
1167                          * some AGL entries to be triggered, then try to help
1168                          * to process the AGL entries. */
1169                         if (sa_sent_full(sai)) {
1170                                 spin_lock(&plli->lli_agl_lock);
1171                                 while (!agl_list_empty(sai)) {
1172                                         clli = agl_first_entry(sai);
1173                                         list_del_init(&clli->lli_agl_list);
1174                                         spin_unlock(&plli->lli_agl_lock);
1175                                         ll_agl_trigger(&clli->lli_vfs_inode,
1176                                                        sai);
1177
1178                                         if (!sa_received_empty(sai))
1179                                                 goto interpret_it;
1180
1181                                         if (unlikely(
1182                                                 !thread_is_running(thread))) {
1183                                                 ll_release_page(page, 0);
1184                                                 rc = 0;
1185                                                 goto out;
1186                                         }
1187
1188                                         if (!sa_sent_full(sai))
1189                                                 goto do_it;
1190
1191                                         spin_lock(&plli->lli_agl_lock);
1192                                 }
1193                                 spin_unlock(&plli->lli_agl_lock);
1194
1195                                 goto keep_it;
1196                         }
1197
1198 do_it:
1199                         ll_statahead_one(parent, name, namelen);
1200                 }
1201                 pos = le64_to_cpu(dp->ldp_hash_end);
1202                 if (pos == MDS_DIR_END_OFF) {
1203                         /*
1204                          * End of directory reached.
1205                          */
1206                         ll_release_page(page, 0);
1207                         while (1) {
1208                                 l_wait_event(thread->t_ctl_waitq,
1209                                              !sa_received_empty(sai) ||
1210                                              sai->sai_sent == sai->sai_replied||
1211                                              !thread_is_running(thread),
1212                                              &lwi);
1213
1214                                 while (!sa_received_empty(sai))
1215                                         ll_post_statahead(sai);
1216
1217                                 if (unlikely(!thread_is_running(thread))) {
1218                                         rc = 0;
1219                                         goto out;
1220                                 }
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                                 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                         rc = 0;
1239                         goto out;
1240                 } else if (1) {
1241                         /*
1242                          * chain is exhausted.
1243                          * Normal case: continue to the next page.
1244                          */
1245                         ll_release_page(page, le32_to_cpu(dp->ldp_flags) &
1246                                               LDF_COLLIDE);
1247                         page = ll_get_dir_page(dir, pos, &chain);
1248                 } else {
1249                         LASSERT(le32_to_cpu(dp->ldp_flags) & LDF_COLLIDE);
1250                         ll_release_page(page, 1);
1251                         /*
1252                          * go into overflow page.
1253                          */
1254                 }
1255         }
1256
1257 out:
1258         if (sai->sai_agl_valid) {
1259                 spin_lock(&plli->lli_agl_lock);
1260                 thread_set_flags(agl_thread, SVC_STOPPING);
1261                 spin_unlock(&plli->lli_agl_lock);
1262                 wake_up(&agl_thread->t_ctl_waitq);
1263
1264                 CDEBUG(D_READA, "stop agl thread: sai %p pid %u\n",
1265                        sai, (unsigned int)agl_thread->t_pid);
1266                 l_wait_event(agl_thread->t_ctl_waitq,
1267                              thread_is_stopped(agl_thread),
1268                              &lwi);
1269         } else {
1270                 /* Set agl_thread flags anyway. */
1271                 thread_set_flags(&sai->sai_agl_thread, SVC_STOPPED);
1272         }
1273         ll_dir_chain_fini(&chain);
1274         spin_lock(&plli->lli_sa_lock);
1275         if (!sa_received_empty(sai)) {
1276                 thread_set_flags(thread, SVC_STOPPING);
1277                 spin_unlock(&plli->lli_sa_lock);
1278
1279                 /* To release the resources held by received entries. */
1280                 while (!sa_received_empty(sai))
1281                         ll_post_statahead(sai);
1282
1283                 spin_lock(&plli->lli_sa_lock);
1284         }
1285         thread_set_flags(thread, SVC_STOPPED);
1286         spin_unlock(&plli->lli_sa_lock);
1287         wake_up(&sai->sai_waitq);
1288         wake_up(&thread->t_ctl_waitq);
1289         ll_sai_put(sai);
1290         dput(parent);
1291         CDEBUG(D_READA, "statahead thread stopped: sai %p, parent %pd\n",
1292                sai, parent);
1293         return rc;
1294 }
1295
1296 /**
1297  * called in ll_file_release().
1298  */
1299 void ll_stop_statahead(struct inode *dir, void *key)
1300 {
1301         struct ll_inode_info *lli = ll_i2info(dir);
1302
1303         if (unlikely(key == NULL))
1304                 return;
1305
1306         spin_lock(&lli->lli_sa_lock);
1307         if (lli->lli_opendir_key != key || lli->lli_opendir_pid == 0) {
1308                 spin_unlock(&lli->lli_sa_lock);
1309                 return;
1310         }
1311
1312         lli->lli_opendir_key = NULL;
1313
1314         if (lli->lli_sai) {
1315                 struct l_wait_info lwi = { 0 };
1316                 struct ptlrpc_thread *thread = &lli->lli_sai->sai_thread;
1317
1318                 if (!thread_is_stopped(thread)) {
1319                         thread_set_flags(thread, SVC_STOPPING);
1320                         spin_unlock(&lli->lli_sa_lock);
1321                         wake_up(&thread->t_ctl_waitq);
1322
1323                         CDEBUG(D_READA, "stop statahead thread: sai %p pid %u\n",
1324                                lli->lli_sai, (unsigned int)thread->t_pid);
1325                         l_wait_event(thread->t_ctl_waitq,
1326                                      thread_is_stopped(thread),
1327                                      &lwi);
1328                 } else {
1329                         spin_unlock(&lli->lli_sa_lock);
1330                 }
1331
1332                 /*
1333                  * Put the ref which was held when first statahead_enter.
1334                  * It maybe not the last ref for some statahead requests
1335                  * maybe inflight.
1336                  */
1337                 ll_sai_put(lli->lli_sai);
1338         } else {
1339                 lli->lli_opendir_pid = 0;
1340                 spin_unlock(&lli->lli_sa_lock);
1341         }
1342 }
1343
1344 enum {
1345         /**
1346          * not first dirent, or is "."
1347          */
1348         LS_NONE_FIRST_DE = 0,
1349         /**
1350          * the first non-hidden dirent
1351          */
1352         LS_FIRST_DE,
1353         /**
1354          * the first hidden dirent, that is "."
1355          */
1356         LS_FIRST_DOT_DE
1357 };
1358
1359 static int is_first_dirent(struct inode *dir, struct dentry *dentry)
1360 {
1361         struct ll_dir_chain   chain;
1362         struct qstr       *target = &dentry->d_name;
1363         struct page       *page;
1364         __u64            pos    = 0;
1365         int                dot_de;
1366         int                rc     = LS_NONE_FIRST_DE;
1367
1368         ll_dir_chain_init(&chain);
1369         page = ll_get_dir_page(dir, pos, &chain);
1370
1371         while (1) {
1372                 struct lu_dirpage *dp;
1373                 struct lu_dirent  *ent;
1374
1375                 if (IS_ERR(page)) {
1376                         struct ll_inode_info *lli = ll_i2info(dir);
1377
1378                         rc = PTR_ERR(page);
1379                         CERROR("error reading dir "DFID" at %llu: [rc %d] [parent %u]\n",
1380                                PFID(ll_inode2fid(dir)), pos,
1381                                rc, lli->lli_opendir_pid);
1382                         break;
1383                 }
1384
1385                 dp = page_address(page);
1386                 for (ent = lu_dirent_start(dp); ent != NULL;
1387                      ent = lu_dirent_next(ent)) {
1388                         __u64 hash;
1389                         int namelen;
1390                         char *name;
1391
1392                         hash = le64_to_cpu(ent->lde_hash);
1393                         /* The ll_get_dir_page() can return any page containing
1394                          * the given hash which may be not the start hash. */
1395                         if (unlikely(hash < pos))
1396                                 continue;
1397
1398                         namelen = le16_to_cpu(ent->lde_namelen);
1399                         if (unlikely(namelen == 0))
1400                                 /*
1401                                  * skip dummy record.
1402                                  */
1403                                 continue;
1404
1405                         name = ent->lde_name;
1406                         if (name[0] == '.') {
1407                                 if (namelen == 1)
1408                                         /*
1409                                          * skip "."
1410                                          */
1411                                         continue;
1412                                 else if (name[1] == '.' && namelen == 2)
1413                                         /*
1414                                          * skip ".."
1415                                          */
1416                                         continue;
1417                                 else
1418                                         dot_de = 1;
1419                         } else {
1420                                 dot_de = 0;
1421                         }
1422
1423                         if (dot_de && target->name[0] != '.') {
1424                                 CDEBUG(D_READA, "%.*s skip hidden file %.*s\n",
1425                                        target->len, target->name,
1426                                        namelen, name);
1427                                 continue;
1428                         }
1429
1430                         if (target->len != namelen ||
1431                             memcmp(target->name, name, namelen) != 0)
1432                                 rc = LS_NONE_FIRST_DE;
1433                         else if (!dot_de)
1434                                 rc = LS_FIRST_DE;
1435                         else
1436                                 rc = LS_FIRST_DOT_DE;
1437
1438                         ll_release_page(page, 0);
1439                         goto out;
1440                 }
1441                 pos = le64_to_cpu(dp->ldp_hash_end);
1442                 if (pos == MDS_DIR_END_OFF) {
1443                         /*
1444                          * End of directory reached.
1445                          */
1446                         ll_release_page(page, 0);
1447                         break;
1448                 } else if (1) {
1449                         /*
1450                          * chain is exhausted
1451                          * Normal case: continue to the next page.
1452                          */
1453                         ll_release_page(page, le32_to_cpu(dp->ldp_flags) &
1454                                               LDF_COLLIDE);
1455                         page = ll_get_dir_page(dir, pos, &chain);
1456                 } else {
1457                         /*
1458                          * go into overflow page.
1459                          */
1460                         LASSERT(le32_to_cpu(dp->ldp_flags) & LDF_COLLIDE);
1461                         ll_release_page(page, 1);
1462                 }
1463         }
1464
1465 out:
1466         ll_dir_chain_fini(&chain);
1467         return rc;
1468 }
1469
1470 static void
1471 ll_sai_unplug(struct ll_statahead_info *sai, struct ll_sa_entry *entry)
1472 {
1473         struct ptlrpc_thread *thread = &sai->sai_thread;
1474         struct ll_sb_info    *sbi    = ll_i2sbi(sai->sai_inode);
1475         int                hit;
1476
1477         if (entry != NULL && entry->se_stat == SA_ENTRY_SUCC)
1478                 hit = 1;
1479         else
1480                 hit = 0;
1481
1482         ll_sa_entry_fini(sai, entry);
1483         if (hit) {
1484                 sai->sai_hit++;
1485                 sai->sai_consecutive_miss = 0;
1486                 sai->sai_max = min(2 * sai->sai_max, sbi->ll_sa_max);
1487         } else {
1488                 struct ll_inode_info *lli = ll_i2info(sai->sai_inode);
1489
1490                 sai->sai_miss++;
1491                 sai->sai_consecutive_miss++;
1492                 if (sa_low_hit(sai) && thread_is_running(thread)) {
1493                         atomic_inc(&sbi->ll_sa_wrong);
1494                         CDEBUG(D_READA, "Statahead for dir "DFID" hit "
1495                                "ratio too low: hit/miss %llu/%llu"
1496                                ", sent/replied %llu/%llu, stopping "
1497                                "statahead thread\n",
1498                                PFID(&lli->lli_fid), sai->sai_hit,
1499                                sai->sai_miss, sai->sai_sent,
1500                                sai->sai_replied);
1501                         spin_lock(&lli->lli_sa_lock);
1502                         if (!thread_is_stopped(thread))
1503                                 thread_set_flags(thread, SVC_STOPPING);
1504                         spin_unlock(&lli->lli_sa_lock);
1505                 }
1506         }
1507
1508         if (!thread_is_stopped(thread))
1509                 wake_up(&thread->t_ctl_waitq);
1510 }
1511
1512 /**
1513  * Start statahead thread if this is the first dir entry.
1514  * Otherwise if a thread is started already, wait it until it is ahead of me.
1515  * \retval 1       -- find entry with lock in cache, the caller needs to do
1516  *                  nothing.
1517  * \retval 0       -- find entry in cache, but without lock, the caller needs
1518  *                  refresh from MDS.
1519  * \retval others  -- the caller need to process as non-statahead.
1520  */
1521 int do_statahead_enter(struct inode *dir, struct dentry **dentryp,
1522                        int only_unplug)
1523 {
1524         struct ll_inode_info     *lli   = ll_i2info(dir);
1525         struct ll_statahead_info *sai   = lli->lli_sai;
1526         struct dentry       *parent;
1527         struct ll_sa_entry       *entry;
1528         struct ptlrpc_thread     *thread;
1529         struct l_wait_info      lwi   = { 0 };
1530         int                    rc    = 0;
1531         struct ll_inode_info     *plli;
1532
1533         LASSERT(lli->lli_opendir_pid == current_pid());
1534
1535         if (sai) {
1536                 thread = &sai->sai_thread;
1537                 if (unlikely(thread_is_stopped(thread) &&
1538                              list_empty(&sai->sai_entries_stated))) {
1539                         /* to release resource */
1540                         ll_stop_statahead(dir, lli->lli_opendir_key);
1541                         return -EAGAIN;
1542                 }
1543
1544                 if ((*dentryp)->d_name.name[0] == '.') {
1545                         if (sai->sai_ls_all ||
1546                             sai->sai_miss_hidden >= sai->sai_skip_hidden) {
1547                                 /*
1548                                  * Hidden dentry is the first one, or statahead
1549                                  * thread does not skip so many hidden dentries
1550                                  * before "sai_ls_all" enabled as below.
1551                                  */
1552                         } else {
1553                                 if (!sai->sai_ls_all)
1554                                         /*
1555                                          * It maybe because hidden dentry is not
1556                                          * the first one, "sai_ls_all" was not
1557                                          * set, then "ls -al" missed. Enable
1558                                          * "sai_ls_all" for such case.
1559                                          */
1560                                         sai->sai_ls_all = 1;
1561
1562                                 /*
1563                                  * Such "getattr" has been skipped before
1564                                  * "sai_ls_all" enabled as above.
1565                                  */
1566                                 sai->sai_miss_hidden++;
1567                                 return -EAGAIN;
1568                         }
1569                 }
1570
1571                 entry = ll_sa_entry_get_byname(sai, &(*dentryp)->d_name);
1572                 if (entry == NULL || only_unplug) {
1573                         ll_sai_unplug(sai, entry);
1574                         return entry ? 1 : -EAGAIN;
1575                 }
1576
1577                 if (!ll_sa_entry_stated(entry)) {
1578                         sai->sai_index_wait = entry->se_index;
1579                         lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(30), NULL,
1580                                                LWI_ON_SIGNAL_NOOP, NULL);
1581                         rc = l_wait_event(sai->sai_waitq,
1582                                           ll_sa_entry_stated(entry) ||
1583                                           thread_is_stopped(thread),
1584                                           &lwi);
1585                         if (rc < 0) {
1586                                 ll_sai_unplug(sai, entry);
1587                                 return -EAGAIN;
1588                         }
1589                 }
1590
1591                 if (entry->se_stat == SA_ENTRY_SUCC &&
1592                     entry->se_inode != NULL) {
1593                         struct inode *inode = entry->se_inode;
1594                         struct lookup_intent it = { .it_op = IT_GETATTR,
1595                                                     .d.lustre.it_lock_handle =
1596                                                      entry->se_handle };
1597                         __u64 bits;
1598
1599                         rc = md_revalidate_lock(ll_i2mdexp(dir), &it,
1600                                                 ll_inode2fid(inode), &bits);
1601                         if (rc == 1) {
1602                                 if ((*dentryp)->d_inode == NULL) {
1603                                         struct dentry *alias;
1604
1605                                         alias = ll_splice_alias(inode,
1606                                                                    *dentryp);
1607                                         if (IS_ERR(alias)) {
1608                                                 ll_sai_unplug(sai, entry);
1609                                                 return PTR_ERR(alias);
1610                                         }
1611                                         *dentryp = alias;
1612                                 } else if ((*dentryp)->d_inode != inode) {
1613                                         /* revalidate, but inode is recreated */
1614                                         CDEBUG(D_READA,
1615                                               "stale dentry %pd inode %lu/%u, "
1616                                               "statahead inode %lu/%u\n",
1617                                               *dentryp,
1618                                               (*dentryp)->d_inode->i_ino,
1619                                               (*dentryp)->d_inode->i_generation,
1620                                               inode->i_ino,
1621                                               inode->i_generation);
1622                                         ll_sai_unplug(sai, entry);
1623                                         return -ESTALE;
1624                                 } else {
1625                                         iput(inode);
1626                                 }
1627                                 entry->se_inode = NULL;
1628
1629                                 if ((bits & MDS_INODELOCK_LOOKUP) &&
1630                                     d_lustre_invalid(*dentryp))
1631                                         d_lustre_revalidate(*dentryp);
1632                                 ll_intent_release(&it);
1633                         }
1634                 }
1635
1636                 ll_sai_unplug(sai, entry);
1637                 return rc;
1638         }
1639
1640         /* I am the "lli_opendir_pid" owner, only me can set "lli_sai". */
1641         rc = is_first_dirent(dir, *dentryp);
1642         if (rc == LS_NONE_FIRST_DE) {
1643                 /* It is not "ls -{a}l" operation, no need statahead for it. */
1644                 rc = -EAGAIN;
1645                 goto out;
1646         }
1647
1648         sai = ll_sai_alloc();
1649         if (sai == NULL) {
1650                 rc = -ENOMEM;
1651                 goto out;
1652         }
1653
1654         sai->sai_ls_all = (rc == LS_FIRST_DOT_DE);
1655         sai->sai_inode = igrab(dir);
1656         if (unlikely(sai->sai_inode == NULL)) {
1657                 CWARN("Do not start stat ahead on dying inode "DFID"\n",
1658                       PFID(&lli->lli_fid));
1659                 rc = -ESTALE;
1660                 goto out;
1661         }
1662
1663         /* get parent reference count here, and put it in ll_statahead_thread */
1664         parent = dget((*dentryp)->d_parent);
1665         if (unlikely(sai->sai_inode != parent->d_inode)) {
1666                 struct ll_inode_info *nlli = ll_i2info(parent->d_inode);
1667
1668                 CWARN("Race condition, someone changed %pd just now: "
1669                       "old parent "DFID", new parent "DFID"\n",
1670                       *dentryp,
1671                       PFID(&lli->lli_fid), PFID(&nlli->lli_fid));
1672                 dput(parent);
1673                 iput(sai->sai_inode);
1674                 rc = -EAGAIN;
1675                 goto out;
1676         }
1677
1678         CDEBUG(D_READA, "start statahead thread: sai %p, parent %pd\n",
1679                sai, parent);
1680
1681         /* The sai buffer already has one reference taken at allocation time,
1682          * but as soon as we expose the sai by attaching it to the lli that
1683          * default reference can be dropped by another thread calling
1684          * ll_stop_statahead. We need to take a local reference to protect
1685          * the sai buffer while we intend to access it. */
1686         ll_sai_get(sai);
1687         lli->lli_sai = sai;
1688
1689         plli = ll_i2info(parent->d_inode);
1690         rc = PTR_ERR(kthread_run(ll_statahead_thread, parent,
1691                                  "ll_sa_%u", plli->lli_opendir_pid));
1692         thread = &sai->sai_thread;
1693         if (IS_ERR_VALUE(rc)) {
1694                 CERROR("can't start ll_sa thread, rc: %d\n", rc);
1695                 dput(parent);
1696                 lli->lli_opendir_key = NULL;
1697                 thread_set_flags(thread, SVC_STOPPED);
1698                 thread_set_flags(&sai->sai_agl_thread, SVC_STOPPED);
1699                 /* Drop both our own local reference and the default
1700                  * reference from allocation time. */
1701                 ll_sai_put(sai);
1702                 ll_sai_put(sai);
1703                 LASSERT(lli->lli_sai == NULL);
1704                 return -EAGAIN;
1705         }
1706
1707         l_wait_event(thread->t_ctl_waitq,
1708                      thread_is_running(thread) || thread_is_stopped(thread),
1709                      &lwi);
1710         ll_sai_put(sai);
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 }