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, Whamcloud, Inc.
  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/smp_lock.h>
  41 #include <linux/highmem.h>
  42 #include <linux/pagemap.h>
  43
  44 #define DEBUG_SUBSYSTEM S_LLITE
  45
  46 #include <obd_support.h>
  47 #include <lustre_lite.h>
  48 #include <lustre_dlm.h>
  49 #include "llite_internal.h"
  50
  51 #define SA_OMITTED_ENTRY_MAX 8ULL
  52
  53 typedef enum {
  54         /** negative values are for error cases */
  55         SA_ENTRY_INIT = 0,      /** init entry */
  56         SA_ENTRY_SUCC = 1,      /** stat succeed */
  57         SA_ENTRY_INVA = 2,      /** invalid entry */
  58         SA_ENTRY_DEST = 3,      /** entry to be destroyed */
  59 } se_stat_t;
  60
  61 struct ll_sa_entry {
  62         /* link into sai->sai_entries_{sent,received,stated} */
  63         cfs_list_t              se_list;
  64         /* link into sai hash table locally */
  65         cfs_list_t              se_hash;
  66         /* entry reference count */
  67         cfs_atomic_t            se_refcount;
  68         /* entry index in the sai */
  69         __u64                   se_index;
  70         /* low layer ldlm lock handle */
  71         __u64                   se_handle;
  72         /* entry status */
  73         se_stat_t               se_stat;
  74         /* entry size, contains name */
  75         int                     se_size;
  76         /* pointer to async getattr enqueue info */
  77         struct md_enqueue_info *se_minfo;
  78         /* pointer to the async getattr request */
  79         struct ptlrpc_request  *se_req;
  80         /* pointer to the target inode */
  81         struct inode           *se_inode;
  82         /* entry name */
  83         struct qstr             se_qstr;
  84 };
  85
  86 static unsigned int sai_generation = 0;
  87 static cfs_spinlock_t sai_generation_lock = CFS_SPIN_LOCK_UNLOCKED;
  88
  89 static inline int ll_sa_entry_unlinked(struct ll_sa_entry *entry)
  90 {
  91         return cfs_list_empty(&entry->se_list);
  92 }
  93
  94 static inline int ll_sa_entry_unhashed(struct ll_sa_entry *entry)
  95 {
  96         return cfs_list_empty(&entry->se_hash);
  97 }
  98
  99 /*
 100  * The entry only can be released by the caller, it is necessary to hold lock.
 101  */
 102 static inline int ll_sa_entry_stated(struct ll_sa_entry *entry)
 103 {
 104         smp_rmb();
 105         return (entry->se_stat != SA_ENTRY_INIT);
 106 }
 107
 108 static inline int ll_sa_entry_hash(int val)
 109 {
 110         return val & LL_SA_CACHE_MASK;
 111 }
 112
 113 /*
 114  * Insert entry to hash SA table.
 115  */
 116 static inline void
 117 ll_sa_entry_enhash(struct ll_statahead_info *sai, struct ll_sa_entry *entry)
 118 {
 119         int i = ll_sa_entry_hash(entry->se_qstr.hash);
 120
 121         cfs_spin_lock(&sai->sai_cache_lock[i]);
 122         cfs_list_add_tail(&entry->se_hash, &sai->sai_cache[i]);
 123         cfs_spin_unlock(&sai->sai_cache_lock[i]);
 124 }
 125
 126 /*
 127  * Remove entry from SA table.
 128  */
 129 static inline void
 130 ll_sa_entry_unhash(struct ll_statahead_info *sai, struct ll_sa_entry *entry)
 131 {
 132         int i = ll_sa_entry_hash(entry->se_qstr.hash);
 133
 134         cfs_spin_lock(&sai->sai_cache_lock[i]);
 135         cfs_list_del_init(&entry->se_hash);
 136         cfs_spin_unlock(&sai->sai_cache_lock[i]);
 137 }
 138
 139 static inline int agl_should_run(struct ll_statahead_info *sai,
 140                                  struct inode *inode)
 141 {
 142         if (inode != NULL && S_ISREG(inode->i_mode) &&
 143             ll_i2info(inode)->lli_smd != NULL && sai->sai_agl_valid)
 144                 return 1;
 145         return 0;
 146 }
 147
 148 static inline struct ll_sa_entry *
 149 sa_first_received_entry(struct ll_statahead_info *sai)
 150 {
 151         return cfs_list_entry(sai->sai_entries_received.next,
 152                               struct ll_sa_entry, se_list);
 153 }
 154
 155 static inline struct ll_inode_info *
 156 agl_first_entry(struct ll_statahead_info *sai)
 157 {
 158         return cfs_list_entry(sai->sai_entries_agl.next,
 159                               struct ll_inode_info, lli_agl_list);
 160 }
 161
 162 static inline int sa_sent_full(struct ll_statahead_info *sai)
 163 {
 164         return cfs_atomic_read(&sai->sai_cache_count) >= sai->sai_max;
 165 }
 166
 167 static inline int sa_received_empty(struct ll_statahead_info *sai)
 168 {
 169         return cfs_list_empty(&sai->sai_entries_received);
 170 }
 171
 172 static inline int agl_list_empty(struct ll_statahead_info *sai)
 173 {
 174         return cfs_list_empty(&sai->sai_entries_agl);
 175 }
 176
 177 /**
 178  * (1) hit ratio less than 80%
 179  * or
 180  * (2) consecutive miss more than 8
 181  * then means low hit.
 182  */
 183 static inline int sa_low_hit(struct ll_statahead_info *sai)
 184 {
 185         return ((sai->sai_hit > 7 && sai->sai_hit < 4 * sai->sai_miss) ||
 186                 (sai->sai_consecutive_miss > 8));
 187 }
 188
 189 /*
 190  * If the given index is behind of statahead window more than
 191  * SA_OMITTED_ENTRY_MAX, then it is old.
 192  */
 193 static inline int is_omitted_entry(struct ll_statahead_info *sai, __u64 index)
 194 {
 195         return ((__u64)sai->sai_max + index + SA_OMITTED_ENTRY_MAX <
 196                  sai->sai_index);
 197 }
 198
 199 /*
 200  * Insert it into sai_entries_sent tail when init.
 201  */
 202 static struct ll_sa_entry *
 203 ll_sa_entry_alloc(struct ll_statahead_info *sai, __u64 index,
 204                   const char *name, int len)
 205 {
 206         struct ll_inode_info *lli;
 207         struct ll_sa_entry   *entry;
 208         int                   entry_size;
 209         char                 *dname;
 210         ENTRY;
 211
 212         entry_size = sizeof(struct ll_sa_entry) + (len & ~3) + 4;
 213         OBD_ALLOC(entry, entry_size);
 214         if (unlikely(entry == NULL))
 215                 RETURN(ERR_PTR(-ENOMEM));
 216
 217         CDEBUG(D_READA, "alloc sai entry %.*s(%p) index "LPU64"\n",
 218                len, name, entry, index);
 219
 220         entry->se_index = index;
 221
 222         /*
 223          * Statahead entry reference rules:
 224          *
 225          * 1) When statahead entry is initialized, its reference is set as 2.
 226          *    One reference is used by the directory scanner. When the scanner
 227          *    searches the statahead cache for the given name, it can perform
 228          *    lockless hash lookup (only the scanner can remove entry from hash
 229          *    list), and once found, it needn't to call "atomic_inc()" for the
 230          *    entry reference. So the performance is improved. After using the
 231          *    statahead entry, the scanner will call "atomic_dec()" to drop the
 232          *    reference held when initialization. If it is the last reference,
 233          *    the statahead entry will be freed.
 234          *
 235          * 2) All other threads, including statahead thread and ptlrpcd thread,
 236          *    when they process the statahead entry, the reference for target
 237          *    should be held to guarantee the entry will not be released by the
 238          *    directory scanner. After processing the entry, these threads will
 239          *    drop the entry reference. If it is the last reference, the entry
 240          *    will be freed.
 241          *
 242          *    The second reference when initializes the statahead entry is used
 243          *    by the statahead thread, following the rule 2).
 244          */
 245         cfs_atomic_set(&entry->se_refcount, 2);
 246         entry->se_stat = SA_ENTRY_INIT;
 247         entry->se_size = entry_size;
 248         dname = (char *)entry + sizeof(struct ll_sa_entry);
 249         memcpy(dname, name, len);
 250         dname[len] = 0;
 251         entry->se_qstr.hash = full_name_hash(name, len);
 252         entry->se_qstr.len = len;
 253         entry->se_qstr.name = dname;
 254
 255         lli = ll_i2info(sai->sai_inode);
 256         cfs_spin_lock(&lli->lli_sa_lock);
 257         cfs_list_add_tail(&entry->se_list, &sai->sai_entries_sent);
 258         cfs_spin_unlock(&lli->lli_sa_lock);
 259
 260         cfs_atomic_inc(&sai->sai_cache_count);
 261         ll_sa_entry_enhash(sai, entry);
 262
 263         RETURN(entry);
 264 }
 265
 266 /*
 267  * Used by the directory scanner to search entry with name.
 268  *
 269  * Only the caller can remove the entry from hash, so it is unnecessary to hold
 270  * hash lock. It is caller's duty to release the init refcount on the entry, so
 271  * it is also unnecessary to increase refcount on the entry.
 272  */
 273 static struct ll_sa_entry *
 274 ll_sa_entry_get_byname(struct ll_statahead_info *sai, const struct qstr *qstr)
 275 {
 276         struct ll_sa_entry *entry;
 277         int i = ll_sa_entry_hash(qstr->hash);
 278
 279         cfs_list_for_each_entry(entry, &sai->sai_cache[i], se_hash) {
 280                 if (entry->se_qstr.hash == qstr->hash &&
 281                     entry->se_qstr.len == qstr->len &&
 282                     memcmp(entry->se_qstr.name, qstr->name, qstr->len) == 0)
 283                         return entry;
 284         }
 285         return NULL;
 286 }
 287
 288 /*
 289  * Used by the async getattr request callback to find entry with index.
 290  *
 291  * Inside lli_sa_lock to prevent others to change the list during the search.
 292  * It needs to increase entry refcount before returning to guarantee that the
 293  * entry cannot be freed by others.
 294  */
 295 static struct ll_sa_entry *
 296 ll_sa_entry_get_byindex(struct ll_statahead_info *sai, __u64 index)
 297 {
 298         struct ll_sa_entry *entry;
 299
 300         cfs_list_for_each_entry(entry, &sai->sai_entries_sent, se_list) {
 301                 if (entry->se_index == index) {
 302                         cfs_atomic_inc(&entry->se_refcount);
 303                         return entry;
 304                 }
 305                 if (entry->se_index > index)
 306                         break;
 307         }
 308         return NULL;
 309 }
 310
 311 static void ll_sa_entry_cleanup(struct ll_statahead_info *sai,
 312                                  struct ll_sa_entry *entry)
 313 {
 314         struct md_enqueue_info *minfo = entry->se_minfo;
 315         struct ptlrpc_request  *req   = entry->se_req;
 316
 317         if (minfo) {
 318                 entry->se_minfo = NULL;
 319                 ll_intent_release(&minfo->mi_it);
 320                 iput(minfo->mi_dir);
 321                 OBD_FREE_PTR(minfo);
 322         }
 323
 324         if (req) {
 325                 entry->se_req = NULL;
 326                 ptlrpc_req_finished(req);
 327         }
 328 }
 329
 330 static void ll_sa_entry_put(struct ll_statahead_info *sai,
 331                              struct ll_sa_entry *entry)
 332 {
 333         if (cfs_atomic_dec_and_test(&entry->se_refcount)) {
 334                 CDEBUG(D_READA, "free sai entry %.*s(%p) index "LPU64"\n",
 335                        entry->se_qstr.len, entry->se_qstr.name, entry,
 336                        entry->se_index);
 337
 338                 LASSERT(ll_sa_entry_unhashed(entry));
 339                 LASSERT(ll_sa_entry_unlinked(entry));
 340
 341                 ll_sa_entry_cleanup(sai, entry);
 342                 if (entry->se_inode)
 343                         iput(entry->se_inode);
 344
 345                 OBD_FREE(entry, entry->se_size);
 346                 cfs_atomic_dec(&sai->sai_cache_count);
 347         }
 348 }
 349
 350 static inline void
 351 do_sai_entry_fini(struct ll_statahead_info *sai, struct ll_sa_entry *entry)
 352 {
 353         struct ll_inode_info *lli = ll_i2info(sai->sai_inode);
 354
 355         ll_sa_entry_unhash(sai, entry);
 356
 357         cfs_spin_lock(&lli->lli_sa_lock);
 358         entry->se_stat = SA_ENTRY_DEST;
 359         if (likely(!ll_sa_entry_unlinked(entry)))
 360                 cfs_list_del_init(&entry->se_list);
 361         cfs_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_sai_entry_fini(sai, entry);
 376
 377         /* drop old entry from sent list */
 378         cfs_list_for_each_entry_safe(pos, next, &sai->sai_entries_sent,
 379                                      se_list) {
 380                 if (is_omitted_entry(sai, pos->se_index))
 381                         do_sai_entry_fini(sai, pos);
 382                 else
 383                         break;
 384         }
 385
 386         /* drop old entry from stated list */
 387         cfs_list_for_each_entry_safe(pos, next, &sai->sai_entries_stated,
 388                                      se_list) {
 389                 if (is_omitted_entry(sai, pos->se_index))
 390                         do_sai_entry_fini(sai, pos);
 391                 else
 392                         break;
 393         }
 394 }
 395
 396 /*
 397  * Inside lli_sa_lock.
 398  */
 399 static void
 400 do_sai_entry_to_stated(struct ll_statahead_info *sai,
 401                        struct ll_sa_entry *entry, int rc)
 402 {
 403         struct ll_sa_entry *se;
 404         cfs_list_t         *pos = &sai->sai_entries_stated;
 405
 406         if (!ll_sa_entry_unlinked(entry))
 407                 cfs_list_del_init(&entry->se_list);
 408
 409         cfs_list_for_each_entry_reverse(se, &sai->sai_entries_stated, se_list) {
 410                 if (se->se_index < entry->se_index) {
 411                         pos = &se->se_list;
 412                         break;
 413                 }
 414         }
 415
 416         cfs_list_add(&entry->se_list, pos);
 417         entry->se_stat = rc;
 418 }
 419
 420 /*
 421  * Move entry to sai_entries_stated and sort with the index.
 422  * \retval 1    -- entry to be destroyed.
 423  * \retval 0    -- entry is inserted into stated list.
 424  */
 425 static int
 426 ll_sa_entry_to_stated(struct ll_statahead_info *sai,
 427                        struct ll_sa_entry *entry, int rc)
 428 {
 429         struct ll_inode_info *lli = ll_i2info(sai->sai_inode);
 430         int                   ret = 1;
 431
 432         ll_sa_entry_cleanup(sai, entry);
 433
 434         cfs_spin_lock(&lli->lli_sa_lock);
 435         if (likely(entry->se_stat != SA_ENTRY_DEST)) {
 436                 do_sai_entry_to_stated(sai, entry, rc);
 437                 ret = 0;
 438         }
 439         cfs_spin_unlock(&lli->lli_sa_lock);
 440
 441         return ret;
 442 }
 443
 444 /*
 445  * Insert inode into the list of sai_entries_agl.
 446  */
 447 static void ll_agl_add(struct ll_statahead_info *sai,
 448                        struct inode *inode, int index)
 449 {
 450         struct ll_inode_info *child  = ll_i2info(inode);
 451         struct ll_inode_info *parent = ll_i2info(sai->sai_inode);
 452         int                   added  = 0;
 453
 454         cfs_spin_lock(&child->lli_agl_lock);
 455         if (child->lli_agl_index == 0) {
 456                 child->lli_agl_index = index;
 457                 cfs_spin_unlock(&child->lli_agl_lock);
 458
 459                 LASSERT(cfs_list_empty(&child->lli_agl_list));
 460
 461                 igrab(inode);
 462                 cfs_spin_lock(&parent->lli_agl_lock);
 463                 if (agl_list_empty(sai))
 464                         added = 1;
 465                 cfs_list_add_tail(&child->lli_agl_list, &sai->sai_entries_agl);
 466                 cfs_spin_unlock(&parent->lli_agl_lock);
 467         } else {
 468                 cfs_spin_unlock(&child->lli_agl_lock);
 469         }
 470
 471         if (added > 0)
 472                 cfs_waitq_signal(&sai->sai_agl_thread.t_ctl_waitq);
 473 }
 474
 475 static struct ll_statahead_info *ll_sai_alloc(void)
 476 {
 477         struct ll_statahead_info *sai;
 478         int                       i;
 479         ENTRY;
 480
 481         OBD_ALLOC_PTR(sai);
 482         if (!sai)
 483                 RETURN(NULL);
 484
 485         cfs_atomic_set(&sai->sai_refcount, 1);
 486
 487         cfs_spin_lock(&sai_generation_lock);
 488         sai->sai_generation = ++sai_generation;
 489         if (unlikely(sai_generation == 0))
 490                 sai->sai_generation = ++sai_generation;
 491         cfs_spin_unlock(&sai_generation_lock);
 492
 493         sai->sai_max = LL_SA_RPC_MIN;
 494         sai->sai_index = 1;
 495         cfs_waitq_init(&sai->sai_waitq);
 496         cfs_waitq_init(&sai->sai_thread.t_ctl_waitq);
 497         cfs_waitq_init(&sai->sai_agl_thread.t_ctl_waitq);
 498
 499         CFS_INIT_LIST_HEAD(&sai->sai_entries_sent);
 500         CFS_INIT_LIST_HEAD(&sai->sai_entries_received);
 501         CFS_INIT_LIST_HEAD(&sai->sai_entries_stated);
 502         CFS_INIT_LIST_HEAD(&sai->sai_entries_agl);
 503
 504         for (i = 0; i < LL_SA_CACHE_SIZE; i++) {
 505                 CFS_INIT_LIST_HEAD(&sai->sai_cache[i]);
 506                 cfs_spin_lock_init(&sai->sai_cache_lock[i]);
 507         }
 508         cfs_atomic_set(&sai->sai_cache_count, 0);
 509
 510         RETURN(sai);
 511 }
 512
 513 static inline struct ll_statahead_info *
 514 ll_sai_get(struct ll_statahead_info *sai)
 515 {
 516         cfs_atomic_inc(&sai->sai_refcount);
 517         return sai;
 518 }
 519
 520 static void ll_sai_put(struct ll_statahead_info *sai)
 521 {
 522         struct inode         *inode = sai->sai_inode;
 523         struct ll_inode_info *lli   = ll_i2info(inode);
 524         ENTRY;
 525
 526         if (cfs_atomic_dec_and_lock(&sai->sai_refcount, &lli->lli_sa_lock)) {
 527                 struct ll_sa_entry *entry, *next;
 528
 529                 if (unlikely(cfs_atomic_read(&sai->sai_refcount) > 0)) {
 530                         /* It is race case, the interpret callback just hold
 531                          * a reference count */
 532                         cfs_spin_unlock(&lli->lli_sa_lock);
 533                         RETURN_EXIT;
 534                 }
 535
 536                 LASSERT(lli->lli_opendir_key == NULL);
 537                 LASSERT(thread_is_stopped(&sai->sai_thread));
 538                 LASSERT(thread_is_stopped(&sai->sai_agl_thread));
 539
 540                 lli->lli_sai = NULL;
 541                 lli->lli_opendir_pid = 0;
 542                 cfs_spin_unlock(&lli->lli_sa_lock);
 543
 544                 if (sai->sai_sent > sai->sai_replied)
 545                         CDEBUG(D_READA,"statahead for dir "DFID" does not "
 546                               "finish: [sent:"LPU64"] [replied:"LPU64"]\n",
 547                               PFID(&lli->lli_fid),
 548                               sai->sai_sent, sai->sai_replied);
 549
 550                 cfs_list_for_each_entry_safe(entry, next,
 551                                              &sai->sai_entries_sent, se_list)
 552                         do_sai_entry_fini(sai, entry);
 553
 554                 LASSERT(sa_received_empty(sai));
 555
 556                 cfs_list_for_each_entry_safe(entry, next,
 557                                              &sai->sai_entries_stated, se_list)
 558                         do_sai_entry_fini(sai, entry);
 559
 560                 LASSERT(cfs_atomic_read(&sai->sai_cache_count) == 0);
 561                 LASSERT(agl_list_empty(sai));
 562
 563                 iput(inode);
 564                 OBD_FREE_PTR(sai);
 565         }
 566
 567         EXIT;
 568 }
 569
 570 /* Do NOT forget to drop inode refcount when into sai_entries_agl. */
 571 static void ll_agl_trigger(struct inode *inode, struct ll_statahead_info *sai)
 572 {
 573         struct ll_inode_info *lli   = ll_i2info(inode);
 574         __u64                 index = lli->lli_agl_index;
 575         int                   rc;
 576         ENTRY;
 577
 578         LASSERT(cfs_list_empty(&lli->lli_agl_list));
 579
 580         /* AGL maybe fall behind statahead with one entry */
 581         if (is_omitted_entry(sai, index + 1)) {
 582                 lli->lli_agl_index = 0;
 583                 iput(inode);
 584                 RETURN_EXIT;
 585         }
 586
 587         /* Someone is in glimpse (sync or async), do nothing. */
 588         rc = cfs_down_write_trylock(&lli->lli_glimpse_sem);
 589         if (rc == 0) {
 590                 lli->lli_agl_index = 0;
 591                 iput(inode);
 592                 RETURN_EXIT;
 593         }
 594
 595         /*
 596          * Someone triggered glimpse within 1 sec before.
 597          * 1) The former glimpse succeeded with glimpse lock granted by OST, and
 598          *    if the lock is still cached on client, AGL needs to do nothing. If
 599          *    it is cancelled by other client, AGL maybe cannot obtaion new lock
 600          *    for no glimpse callback triggered by AGL.
 601          * 2) The former glimpse succeeded, but OST did not grant glimpse lock.
 602          *    Under such case, it is quite possible that the OST will not grant
 603          *    glimpse lock for AGL also.
 604          * 3) The former glimpse failed, compared with other two cases, it is
 605          *    relative rare. AGL can ignore such case, and it will not muchly
 606          *    affect the performance.
 607          */
 608         if (lli->lli_glimpse_time != 0 &&
 609             cfs_time_before(cfs_time_shift(-1), lli->lli_glimpse_time)) {
 610                 cfs_up_write(&lli->lli_glimpse_sem);
 611                 lli->lli_agl_index = 0;
 612                 iput(inode);
 613                 RETURN_EXIT;
 614         }
 615
 616         CDEBUG(D_READA, "Handling (init) async glimpse: inode = "
 617                DFID", idx = "LPU64"\n", PFID(&lli->lli_fid), index);
 618
 619         cl_agl(inode);
 620         lli->lli_agl_index = 0;
 621         lli->lli_glimpse_time = cfs_time_current();
 622         cfs_up_write(&lli->lli_glimpse_sem);
 623
 624         CDEBUG(D_READA, "Handled (init) async glimpse: inode= "
 625                DFID", idx = "LPU64", rc = %d\n",
 626                PFID(&lli->lli_fid), index, rc);
 627
 628         iput(inode);
 629
 630         EXIT;
 631 }
 632
 633 static void do_statahead_interpret(struct ll_statahead_info *sai,
 634                                    struct ll_sa_entry *target)
 635 {
 636         struct inode           *dir   = sai->sai_inode;
 637         struct inode           *child;
 638         struct ll_inode_info   *lli   = ll_i2info(dir);
 639         struct ll_sa_entry     *entry;
 640         struct md_enqueue_info *minfo;
 641         struct lookup_intent   *it;
 642         struct ptlrpc_request  *req;
 643         struct mdt_body        *body;
 644         int                     rc    = 0;
 645         ENTRY;
 646
 647         cfs_spin_lock(&lli->lli_sa_lock);
 648         if (target != NULL && target->se_req != NULL &&
 649             !cfs_list_empty(&target->se_list)) {
 650                 entry = target;
 651         } else if (unlikely(sa_received_empty(sai))) {
 652                 cfs_spin_unlock(&lli->lli_sa_lock);
 653                 RETURN_EXIT;
 654         } else {
 655                 entry = sa_first_received_entry(sai);
 656         }
 657
 658         cfs_atomic_inc(&entry->se_refcount);
 659         cfs_list_del_init(&entry->se_list);
 660         cfs_spin_unlock(&lli->lli_sa_lock);
 661
 662         LASSERT(entry->se_handle != 0);
 663
 664         minfo = entry->se_minfo;
 665         it = &minfo->mi_it;
 666         req = entry->se_req;
 667         body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
 668         if (body == NULL)
 669                 GOTO(out, rc = -EFAULT);
 670
 671         child = entry->se_inode;
 672         if (child == NULL) {
 673                 /*
 674                  * lookup.
 675                  */
 676                 LASSERT(fid_is_zero(&minfo->mi_data.op_fid2));
 677
 678                 /* XXX: No fid in reply, this is probaly cross-ref case.
 679                  * SA can't handle it yet. */
 680                 if (body->valid & OBD_MD_MDS)
 681                         GOTO(out, rc = -EAGAIN);
 682         } else {
 683                 /*
 684                  * revalidate.
 685                  */
 686                 /* unlinked and re-created with the same name */
 687                 if (unlikely(!lu_fid_eq(&minfo->mi_data.op_fid2, &body->fid1))){
 688                         entry->se_inode = NULL;
 689                         iput(child);
 690                         child = NULL;
 691                 }
 692         }
 693
 694         it->d.lustre.it_lock_handle = entry->se_handle;
 695         rc = md_revalidate_lock(ll_i2mdexp(dir), it, NULL, NULL);
 696         if (rc != 1)
 697                 GOTO(out, rc = -EAGAIN);
 698
 699         rc = ll_prep_inode(&child, req, dir->i_sb);
 700         if (rc)
 701                 GOTO(out, rc);
 702
 703         CDEBUG(D_DLMTRACE, "setting l_data to inode %p (%lu/%u)\n",
 704                child, child->i_ino, child->i_generation);
 705         ll_set_lock_data(ll_i2sbi(dir)->ll_md_exp, child, it, NULL);
 706
 707         entry->se_inode = child;
 708
 709         if (agl_should_run(sai, child))
 710                 ll_agl_add(sai, child, entry->se_index);
 711
 712         EXIT;
 713
 714 out:
 715         /* The "ll_sa_entry_to_stated()" will drop related ldlm ibits lock
 716          * reference count by calling "ll_intent_drop_lock()" in spite of the
 717          * above operations failed or not. Do not worry about calling
 718          * "ll_intent_drop_lock()" more than once. */
 719         rc = ll_sa_entry_to_stated(sai, entry, rc < 0 ? rc : SA_ENTRY_SUCC);
 720         if (rc == 0 && entry->se_index == sai->sai_index_wait && target == NULL)
 721                 cfs_waitq_signal(&sai->sai_waitq);
 722         ll_sa_entry_put(sai, entry);
 723 }
 724
 725 static int ll_statahead_interpret(struct ptlrpc_request *req,
 726                                   struct md_enqueue_info *minfo, int rc)
 727 {
 728         struct lookup_intent     *it  = &minfo->mi_it;
 729         struct inode             *dir = minfo->mi_dir;
 730         struct ll_inode_info     *lli = ll_i2info(dir);
 731         struct ll_statahead_info *sai = NULL;
 732         struct ll_sa_entry       *entry;
 733         int                       wakeup;
 734         ENTRY;
 735
 736         if (it_disposition(it, DISP_LOOKUP_NEG))
 737                 rc = -ENOENT;
 738
 739         cfs_spin_lock(&lli->lli_sa_lock);
 740         /* stale entry */
 741         if (unlikely(lli->lli_sai == NULL ||
 742                      lli->lli_sai->sai_generation != minfo->mi_generation)) {
 743                 cfs_spin_unlock(&lli->lli_sa_lock);
 744                 GOTO(out, rc = -ESTALE);
 745         } else {
 746                 sai = ll_sai_get(lli->lli_sai);
 747                 if (unlikely(!thread_is_running(&sai->sai_thread))) {
 748                         sai->sai_replied++;
 749                         cfs_spin_unlock(&lli->lli_sa_lock);
 750                         GOTO(out, rc = -EBADFD);
 751                 }
 752
 753                 entry = ll_sa_entry_get_byindex(sai, minfo->mi_cbdata);
 754                 if (entry == NULL) {
 755                         sai->sai_replied++;
 756                         cfs_spin_unlock(&lli->lli_sa_lock);
 757                         GOTO(out, rc = -EIDRM);
 758                 }
 759
 760                 cfs_list_del_init(&entry->se_list);
 761                 if (rc != 0) {
 762                         sai->sai_replied++;
 763                         do_sai_entry_to_stated(sai, entry, rc);
 764                         cfs_spin_unlock(&lli->lli_sa_lock);
 765                         if (entry->se_index == sai->sai_index_wait)
 766                                 cfs_waitq_signal(&sai->sai_waitq);
 767                 } else {
 768                         entry->se_minfo = minfo;
 769                         entry->se_req = ptlrpc_request_addref(req);
 770                         /* Release the async ibits lock ASAP to avoid deadlock
 771                          * when statahead thread tries to enqueue lock on parent
 772                          * for readpage and other tries to enqueue lock on child
 773                          * with parent's lock held, for example: unlink. */
 774                         entry->se_handle = it->d.lustre.it_lock_handle;
 775                         ll_intent_drop_lock(it);
 776                         wakeup = sa_received_empty(sai);
 777                         cfs_list_add_tail(&entry->se_list,
 778                                           &sai->sai_entries_received);
 779                         sai->sai_replied++;
 780                         cfs_spin_unlock(&lli->lli_sa_lock);
 781                         if (wakeup)
 782                                 cfs_waitq_signal(&sai->sai_thread.t_ctl_waitq);
 783                 }
 784                 ll_sa_entry_put(sai, entry);
 785         }
 786
 787         EXIT;
 788
 789 out:
 790         if (rc != 0) {
 791                 ll_intent_release(it);
 792                 iput(dir);
 793                 OBD_FREE_PTR(minfo);
 794         }
 795         if (sai != NULL)
 796                 ll_sai_put(sai);
 797         return rc;
 798 }
 799
 800 static void sa_args_fini(struct md_enqueue_info *minfo,
 801                          struct ldlm_enqueue_info *einfo)
 802 {
 803         LASSERT(minfo && einfo);
 804         iput(minfo->mi_dir);
 805         capa_put(minfo->mi_data.op_capa1);
 806         capa_put(minfo->mi_data.op_capa2);
 807         OBD_FREE_PTR(minfo);
 808         OBD_FREE_PTR(einfo);
 809 }
 810
 811 /**
 812  * There is race condition between "capa_put" and "ll_statahead_interpret" for
 813  * accessing "op_data.op_capa[1,2]" as following:
 814  * "capa_put" releases "op_data.op_capa[1,2]"'s reference count after calling
 815  * "md_intent_getattr_async". But "ll_statahead_interpret" maybe run first, and
 816  * fill "op_data.op_capa[1,2]" as POISON, then cause "capa_put" access invalid
 817  * "ocapa". So here reserve "op_data.op_capa[1,2]" in "pcapa" before calling
 818  * "md_intent_getattr_async".
 819  */
 820 static int sa_args_init(struct inode *dir, struct inode *child,
 821                         struct ll_sa_entry *entry, struct md_enqueue_info **pmi,
 822                         struct ldlm_enqueue_info **pei,
 823                         struct obd_capa **pcapa)
 824 {
 825         struct qstr              *qstr = &entry->se_qstr;
 826         struct ll_inode_info     *lli  = ll_i2info(dir);
 827         struct md_enqueue_info   *minfo;
 828         struct ldlm_enqueue_info *einfo;
 829         struct md_op_data        *op_data;
 830
 831         OBD_ALLOC_PTR(einfo);
 832         if (einfo == NULL)
 833                 return -ENOMEM;
 834
 835         OBD_ALLOC_PTR(minfo);
 836         if (minfo == NULL) {
 837                 OBD_FREE_PTR(einfo);
 838                 return -ENOMEM;
 839         }
 840
 841         op_data = ll_prep_md_op_data(&minfo->mi_data, dir, child, qstr->name,
 842                                      qstr->len, 0, LUSTRE_OPC_ANY, NULL);
 843         if (IS_ERR(op_data)) {
 844                 OBD_FREE_PTR(einfo);
 845                 OBD_FREE_PTR(minfo);
 846                 return PTR_ERR(op_data);
 847         }
 848
 849         minfo->mi_it.it_op = IT_GETATTR;
 850         minfo->mi_dir = igrab(dir);
 851         minfo->mi_cb = ll_statahead_interpret;
 852         minfo->mi_generation = lli->lli_sai->sai_generation;
 853         minfo->mi_cbdata = entry->se_index;
 854
 855         einfo->ei_type   = LDLM_IBITS;
 856         einfo->ei_mode   = it_to_lock_mode(&minfo->mi_it);
 857         einfo->ei_cb_bl  = ll_md_blocking_ast;
 858         einfo->ei_cb_cp  = ldlm_completion_ast;
 859         einfo->ei_cb_gl  = NULL;
 860         einfo->ei_cbdata = NULL;
 861
 862         *pmi = minfo;
 863         *pei = einfo;
 864         pcapa[0] = op_data->op_capa1;
 865         pcapa[1] = op_data->op_capa2;
 866
 867         return 0;
 868 }
 869
 870 static int do_sa_lookup(struct inode *dir, struct ll_sa_entry *entry)
 871 {
 872         struct md_enqueue_info   *minfo;
 873         struct ldlm_enqueue_info *einfo;
 874         struct obd_capa          *capas[2];
 875         int                       rc;
 876         ENTRY;
 877
 878         rc = sa_args_init(dir, NULL, entry, &minfo, &einfo, capas);
 879         if (rc)
 880                 RETURN(rc);
 881
 882         rc = md_intent_getattr_async(ll_i2mdexp(dir), minfo, einfo);
 883         if (!rc) {
 884                 capa_put(capas[0]);
 885                 capa_put(capas[1]);
 886         } else {
 887                 sa_args_fini(minfo, einfo);
 888         }
 889
 890         RETURN(rc);
 891 }
 892
 893 /**
 894  * similar to ll_revalidate_it().
 895  * \retval      1 -- dentry valid
 896  * \retval      0 -- will send stat-ahead request
 897  * \retval others -- prepare stat-ahead request failed
 898  */
 899 static int do_sa_revalidate(struct inode *dir, struct ll_sa_entry *entry,
 900                             struct dentry *dentry)
 901 {
 902         struct inode             *inode = dentry->d_inode;
 903         struct lookup_intent      it = { .it_op = IT_GETATTR,
 904                                          .d.lustre.it_lock_handle = 0 };
 905         struct md_enqueue_info   *minfo;
 906         struct ldlm_enqueue_info *einfo;
 907         struct obd_capa          *capas[2];
 908         int rc;
 909         ENTRY;
 910
 911         if (unlikely(inode == NULL))
 912                 RETURN(1);
 913
 914         if (d_mountpoint(dentry))
 915                 RETURN(1);
 916
 917         if (unlikely(dentry == dentry->d_sb->s_root))
 918                 RETURN(1);
 919
 920         entry->se_inode = igrab(inode);
 921         rc = md_revalidate_lock(ll_i2mdexp(dir), &it, ll_inode2fid(inode),NULL);
 922         if (rc == 1) {
 923                 entry->se_handle = it.d.lustre.it_lock_handle;
 924                 ll_intent_release(&it);
 925                 RETURN(1);
 926         }
 927
 928         rc = sa_args_init(dir, inode, entry, &minfo, &einfo, capas);
 929         if (rc) {
 930                 entry->se_inode = NULL;
 931                 iput(inode);
 932                 RETURN(rc);
 933         }
 934
 935         rc = md_intent_getattr_async(ll_i2mdexp(dir), minfo, einfo);
 936         if (!rc) {
 937                 capa_put(capas[0]);
 938                 capa_put(capas[1]);
 939         } else {
 940                 entry->se_inode = NULL;
 941                 iput(inode);
 942                 sa_args_fini(minfo, einfo);
 943         }
 944
 945         RETURN(rc);
 946 }
 947
 948 static void ll_statahead_one(struct dentry *parent, const char* entry_name,
 949                              int entry_name_len)
 950 {
 951         struct inode             *dir    = parent->d_inode;
 952         struct ll_inode_info     *lli    = ll_i2info(dir);
 953         struct ll_statahead_info *sai    = lli->lli_sai;
 954         struct dentry            *dentry = NULL;
 955         struct ll_sa_entry       *entry;
 956         int                       rc;
 957         int                       rc1;
 958         ENTRY;
 959
 960         entry = ll_sa_entry_alloc(sai, sai->sai_index, entry_name,
 961                                   entry_name_len);
 962         if (IS_ERR(entry))
 963                 RETURN_EXIT;
 964
 965         dentry = d_lookup(parent, &entry->se_qstr);
 966         if (!dentry) {
 967                 rc = do_sa_lookup(dir, entry);
 968         } else {
 969                 rc = do_sa_revalidate(dir, entry, dentry);
 970                 if (rc == 1 && agl_should_run(sai, dentry->d_inode))
 971                         ll_agl_add(sai, dentry->d_inode, entry->se_index);
 972         }
 973
 974         if (dentry != NULL)
 975                 dput(dentry);
 976
 977         if (rc) {
 978                 rc1 = ll_sa_entry_to_stated(sai, entry,
 979                                         rc < 0 ? SA_ENTRY_INVA : SA_ENTRY_SUCC);
 980                 if (rc1 == 0 && entry->se_index == sai->sai_index_wait)
 981                         cfs_waitq_signal(&sai->sai_waitq);
 982         } else {
 983                 sai->sai_sent++;
 984         }
 985
 986         sai->sai_index++;
 987         /* drop one refcount on entry by ll_sa_entry_alloc */
 988         ll_sa_entry_put(sai, entry);
 989
 990         EXIT;
 991 }
 992
 993 static int ll_agl_thread(void *arg)
 994 {
 995         struct dentry            *parent = (struct dentry *)arg;
 996         struct inode             *dir    = parent->d_inode;
 997         struct ll_inode_info     *plli   = ll_i2info(dir);
 998         struct ll_inode_info     *clli;
 999         struct ll_sb_info        *sbi    = ll_i2sbi(dir);
1000         struct ll_statahead_info *sai    = ll_sai_get(plli->lli_sai);
1001         struct ptlrpc_thread     *thread = &sai->sai_agl_thread;
1002         struct l_wait_info        lwi    = { 0 };
1003         ENTRY;
1004
1005         {
1006                 char pname[16];
1007                 snprintf(pname, 15, "ll_agl_%u", plli->lli_opendir_pid);
1008                 cfs_daemonize(pname);
1009         }
1010
1011         CDEBUG(D_READA, "agl thread started: [pid %d] [parent %.*s]\n",
1012                cfs_curproc_pid(), parent->d_name.len, parent->d_name.name);
1013
1014         atomic_inc(&sbi->ll_agl_total);
1015         cfs_spin_lock(&plli->lli_agl_lock);
1016         sai->sai_agl_valid = 1;
1017         thread_set_flags(thread, SVC_RUNNING);
1018         cfs_spin_unlock(&plli->lli_agl_lock);
1019         cfs_waitq_signal(&thread->t_ctl_waitq);
1020
1021         while (1) {
1022                 l_wait_event(thread->t_ctl_waitq,
1023                              !agl_list_empty(sai) ||
1024                              !thread_is_running(thread),
1025                              &lwi);
1026
1027                 if (!thread_is_running(thread))
1028                         break;
1029
1030                 cfs_spin_lock(&plli->lli_agl_lock);
1031                 /* The statahead thread maybe help to process AGL entries,
1032                  * so check whether list empty again. */
1033                 if (!agl_list_empty(sai)) {
1034                         clli = agl_first_entry(sai);
1035                         cfs_list_del_init(&clli->lli_agl_list);
1036                         cfs_spin_unlock(&plli->lli_agl_lock);
1037                         ll_agl_trigger(&clli->lli_vfs_inode, sai);
1038                 } else {
1039                         cfs_spin_unlock(&plli->lli_agl_lock);
1040                 }
1041         }
1042
1043         cfs_spin_lock(&plli->lli_agl_lock);
1044         sai->sai_agl_valid = 0;
1045         while (!agl_list_empty(sai)) {
1046                 clli = agl_first_entry(sai);
1047                 cfs_list_del_init(&clli->lli_agl_list);
1048                 cfs_spin_unlock(&plli->lli_agl_lock);
1049                 clli->lli_agl_index = 0;
1050                 iput(&clli->lli_vfs_inode);
1051                 cfs_spin_lock(&plli->lli_agl_lock);
1052         }
1053         thread_set_flags(thread, SVC_STOPPED);
1054         cfs_spin_unlock(&plli->lli_agl_lock);
1055         cfs_waitq_signal(&thread->t_ctl_waitq);
1056         ll_sai_put(sai);
1057         CDEBUG(D_READA, "agl thread stopped: [pid %d] [parent %.*s]\n",
1058                cfs_curproc_pid(), parent->d_name.len, parent->d_name.name);
1059         RETURN(0);
1060 }
1061
1062 static void ll_start_agl(struct dentry *parent, struct ll_statahead_info *sai)
1063 {
1064         struct ptlrpc_thread *thread = &sai->sai_agl_thread;
1065         struct l_wait_info    lwi    = { 0 };
1066         int                   rc;
1067         ENTRY;
1068
1069         CDEBUG(D_READA, "start agl thread: [pid %d] [parent %.*s]\n",
1070                cfs_curproc_pid(), parent->d_name.len, parent->d_name.name);
1071
1072         rc = cfs_create_thread(ll_agl_thread, parent, 0);
1073         if (rc < 0) {
1074                 CERROR("can't start ll_agl thread, rc: %d\n", rc);
1075                 thread_set_flags(thread, SVC_STOPPED);
1076                 RETURN_EXIT;
1077         }
1078
1079         l_wait_event(thread->t_ctl_waitq,
1080                      thread_is_running(thread) || thread_is_stopped(thread),
1081                      &lwi);
1082         EXIT;
1083 }
1084
1085 static int ll_statahead_thread(void *arg)
1086 {
1087         struct dentry            *parent = (struct dentry *)arg;
1088         struct inode             *dir    = parent->d_inode;
1089         struct ll_inode_info     *plli   = ll_i2info(dir);
1090         struct ll_inode_info     *clli;
1091         struct ll_sb_info        *sbi    = ll_i2sbi(dir);
1092         struct ll_statahead_info *sai    = ll_sai_get(plli->lli_sai);
1093         struct ptlrpc_thread     *thread = &sai->sai_thread;
1094         struct ptlrpc_thread *agl_thread = &sai->sai_agl_thread;
1095         struct page              *page;
1096         __u64                     pos    = 0;
1097         int                       first  = 0;
1098         int                       rc     = 0;
1099         struct ll_dir_chain       chain;
1100         struct l_wait_info        lwi    = { 0 };
1101         ENTRY;
1102
1103         {
1104                 char pname[16];
1105                 snprintf(pname, 15, "ll_sa_%u", plli->lli_opendir_pid);
1106                 cfs_daemonize(pname);
1107         }
1108
1109         CDEBUG(D_READA, "statahead thread started: [pid %d] [parent %.*s]\n",
1110                cfs_curproc_pid(), parent->d_name.len, parent->d_name.name);
1111
1112         if (sbi->ll_flags & LL_SBI_AGL_ENABLED)
1113                 ll_start_agl(parent, sai);
1114
1115         atomic_inc(&sbi->ll_sa_total);
1116         cfs_spin_lock(&plli->lli_sa_lock);
1117         thread_set_flags(thread, SVC_RUNNING);
1118         cfs_spin_unlock(&plli->lli_sa_lock);
1119         cfs_waitq_signal(&thread->t_ctl_waitq);
1120
1121         plli->lli_sa_pos = 0;
1122         ll_dir_chain_init(&chain);
1123         page = ll_get_dir_page(NULL, dir, pos, &chain);
1124
1125         while (1) {
1126                 struct lu_dirpage *dp;
1127                 struct lu_dirent  *ent;
1128
1129                 if (IS_ERR(page)) {
1130                         rc = PTR_ERR(page);
1131                         CDEBUG(D_READA, "error reading dir "DFID" at "LPU64
1132                                "/"LPU64": [rc %d] [parent %u]\n",
1133                                PFID(ll_inode2fid(dir)), pos, sai->sai_index,
1134                                rc, plli->lli_opendir_pid);
1135                         GOTO(out, rc);
1136                 }
1137
1138                 dp = page_address(page);
1139                 for (ent = lu_dirent_start(dp); ent != NULL;
1140                      ent = lu_dirent_next(ent)) {
1141                         __u64 hash;
1142                         int namelen;
1143                         char *name;
1144
1145                         hash = le64_to_cpu(ent->lde_hash);
1146                         if (unlikely(hash < pos))
1147                                 /*
1148                                  * Skip until we find target hash value.
1149                                  */
1150                                 continue;
1151
1152                         namelen = le16_to_cpu(ent->lde_namelen);
1153                         if (unlikely(namelen == 0))
1154                                 /*
1155                                  * Skip dummy record.
1156                                  */
1157                                 continue;
1158
1159                         name = ent->lde_name;
1160                         if (name[0] == '.') {
1161                                 if (namelen == 1) {
1162                                         /*
1163                                          * skip "."
1164                                          */
1165                                         continue;
1166                                 } else if (name[1] == '.' && namelen == 2) {
1167                                         /*
1168                                          * skip ".."
1169                                          */
1170                                         continue;
1171                                 } else if (!sai->sai_ls_all) {
1172                                         /*
1173                                          * skip hidden files.
1174                                          */
1175                                         sai->sai_skip_hidden++;
1176                                         continue;
1177                                 }
1178                         }
1179
1180                         /*
1181                          * don't stat-ahead first entry.
1182                          */
1183                         if (unlikely(++first == 1))
1184                                 continue;
1185
1186 keep_it:
1187                         l_wait_event(thread->t_ctl_waitq,
1188                                      !sa_sent_full(sai) ||
1189                                      !sa_received_empty(sai) ||
1190                                      !agl_list_empty(sai) ||
1191                                      !thread_is_running(thread),
1192                                      &lwi);
1193
1194 interpret_it:
1195                         while (!sa_received_empty(sai))
1196                                 do_statahead_interpret(sai, NULL);
1197
1198                         if (unlikely(!thread_is_running(thread))) {
1199                                 ll_release_page(page, 0);
1200                                 GOTO(out, rc = 0);
1201                         }
1202
1203                         /* If no window for metadata statahead, but there are
1204                          * some AGL entries to be triggered, then try to help
1205                          * to process the AGL entries. */
1206                         if (sa_sent_full(sai)) {
1207                                 cfs_spin_lock(&plli->lli_agl_lock);
1208                                 while (!agl_list_empty(sai)) {
1209                                         clli = agl_first_entry(sai);
1210                                         cfs_list_del_init(&clli->lli_agl_list);
1211                                         cfs_spin_unlock(&plli->lli_agl_lock);
1212                                         ll_agl_trigger(&clli->lli_vfs_inode,
1213                                                        sai);
1214
1215                                         if (!sa_received_empty(sai))
1216                                                 goto interpret_it;
1217
1218                                         if (unlikely(
1219                                                 !thread_is_running(thread))) {
1220                                                 ll_release_page(page, 0);
1221                                                 GOTO(out, rc = 0);
1222                                         }
1223
1224                                         if (!sa_sent_full(sai))
1225                                                 goto do_it;
1226
1227                                         cfs_spin_lock(&plli->lli_agl_lock);
1228                                 }
1229                                 cfs_spin_unlock(&plli->lli_agl_lock);
1230
1231                                 goto keep_it;
1232                         }
1233
1234 do_it:
1235                         ll_statahead_one(parent, name, namelen);
1236                 }
1237                 pos = le64_to_cpu(dp->ldp_hash_end);
1238                 if (pos == MDS_DIR_END_OFF) {
1239                         /*
1240                          * End of directory reached.
1241                          */
1242                         ll_release_page(page, 0);
1243                         while (1) {
1244                                 l_wait_event(thread->t_ctl_waitq,
1245                                              !sa_received_empty(sai) ||
1246                                              sai->sai_sent == sai->sai_replied||
1247                                              !thread_is_running(thread),
1248                                              &lwi);
1249
1250                                 while (!sa_received_empty(sai))
1251                                         do_statahead_interpret(sai, NULL);
1252
1253                                 if (unlikely(!thread_is_running(thread)))
1254                                         GOTO(out, rc = 0);
1255
1256                                 if (sai->sai_sent == sai->sai_replied &&
1257                                     sa_received_empty(sai))
1258                                         break;
1259                         }
1260
1261                         cfs_spin_lock(&plli->lli_agl_lock);
1262                         while (!agl_list_empty(sai) &&
1263                                thread_is_running(thread)) {
1264                                 clli = agl_first_entry(sai);
1265                                 cfs_list_del_init(&clli->lli_agl_list);
1266                                 cfs_spin_unlock(&plli->lli_agl_lock);
1267                                 ll_agl_trigger(&clli->lli_vfs_inode, sai);
1268                                 cfs_spin_lock(&plli->lli_agl_lock);
1269                         }
1270                         cfs_spin_unlock(&plli->lli_agl_lock);
1271
1272                         GOTO(out, rc = 0);
1273                 } else if (1) {
1274                         /*
1275                          * chain is exhausted.
1276                          * Normal case: continue to the next page.
1277                          */
1278                         ll_release_page(page, le32_to_cpu(dp->ldp_flags) &
1279                                               LDF_COLLIDE);
1280                         plli->lli_sa_pos = pos;
1281                         sai->sai_in_readpage = 1;
1282                         page = ll_get_dir_page(NULL, dir, pos, &chain);
1283                         sai->sai_in_readpage = 0;
1284                 } else {
1285                         LASSERT(le32_to_cpu(dp->ldp_flags) & LDF_COLLIDE);
1286                         ll_release_page(page, 1);
1287                         /*
1288                          * go into overflow page.
1289                          */
1290                 }
1291         }
1292         EXIT;
1293
1294 out:
1295         if (sai->sai_agl_valid) {
1296                 cfs_spin_lock(&plli->lli_agl_lock);
1297                 thread_set_flags(agl_thread, SVC_STOPPING);
1298                 cfs_spin_unlock(&plli->lli_agl_lock);
1299                 cfs_waitq_signal(&agl_thread->t_ctl_waitq);
1300
1301                 CDEBUG(D_READA, "stop agl thread: [pid %d]\n",
1302                        cfs_curproc_pid());
1303                 l_wait_event(agl_thread->t_ctl_waitq,
1304                              thread_is_stopped(agl_thread),
1305                              &lwi);
1306         } else {
1307                 /* Set agl_thread flags anyway. */
1308                 thread_set_flags(&sai->sai_agl_thread, SVC_STOPPED);
1309         }
1310         ll_dir_chain_fini(&chain);
1311         cfs_spin_lock(&plli->lli_sa_lock);
1312         if (!sa_received_empty(sai)) {
1313                 thread_set_flags(thread, SVC_STOPPING);
1314                 cfs_spin_unlock(&plli->lli_sa_lock);
1315
1316                 /* To release the resources held by received entries. */
1317                 while (!sa_received_empty(sai))
1318                         do_statahead_interpret(sai, NULL);
1319
1320                 cfs_spin_lock(&plli->lli_sa_lock);
1321         }
1322         thread_set_flags(thread, SVC_STOPPED);
1323         cfs_spin_unlock(&plli->lli_sa_lock);
1324         cfs_waitq_signal(&sai->sai_waitq);
1325         cfs_waitq_signal(&thread->t_ctl_waitq);
1326         ll_sai_put(sai);
1327         dput(parent);
1328         CDEBUG(D_READA, "statahead thread stopped: [pid %d] [parent %.*s]\n",
1329                cfs_curproc_pid(), parent->d_name.len, parent->d_name.name);
1330         return rc;
1331 }
1332
1333 /**
1334  * called in ll_file_release().
1335  */
1336 void ll_stop_statahead(struct inode *dir, void *key)
1337 {
1338         struct ll_inode_info *lli = ll_i2info(dir);
1339
1340         if (unlikely(key == NULL))
1341                 return;
1342
1343         cfs_spin_lock(&lli->lli_sa_lock);
1344         if (lli->lli_opendir_key != key || lli->lli_opendir_pid == 0) {
1345                 cfs_spin_unlock(&lli->lli_sa_lock);
1346                 return;
1347         }
1348
1349         lli->lli_opendir_key = NULL;
1350
1351         if (lli->lli_sai) {
1352                 struct l_wait_info lwi = { 0 };
1353                 struct ptlrpc_thread *thread = &lli->lli_sai->sai_thread;
1354
1355                 if (!thread_is_stopped(thread)) {
1356                         thread_set_flags(thread, SVC_STOPPING);
1357                         cfs_spin_unlock(&lli->lli_sa_lock);
1358                         cfs_waitq_signal(&thread->t_ctl_waitq);
1359
1360                         CDEBUG(D_READA, "stop statahead thread: [pid %d]\n",
1361                                cfs_curproc_pid());
1362                         l_wait_event(thread->t_ctl_waitq,
1363                                      thread_is_stopped(thread),
1364                                      &lwi);
1365                 } else {
1366                         cfs_spin_unlock(&lli->lli_sa_lock);
1367                 }
1368
1369                 /*
1370                  * Put the ref which was held when first statahead_enter.
1371                  * It maybe not the last ref for some statahead requests
1372                  * maybe inflight.
1373                  */
1374                 ll_sai_put(lli->lli_sai);
1375         } else {
1376                 lli->lli_opendir_pid = 0;
1377                 cfs_spin_unlock(&lli->lli_sa_lock);
1378         }
1379 }
1380
1381 enum {
1382         /**
1383          * not first dirent, or is "."
1384          */
1385         LS_NONE_FIRST_DE = 0,
1386         /**
1387          * the first non-hidden dirent
1388          */
1389         LS_FIRST_DE,
1390         /**
1391          * the first hidden dirent, that is "."
1392          */
1393         LS_FIRST_DOT_DE
1394 };
1395
1396 static int is_first_dirent(struct inode *dir, struct dentry *dentry)
1397 {
1398         struct ll_inode_info *lli    = ll_i2info(dir);
1399         struct ll_dir_chain   chain;
1400         struct qstr          *target = &dentry->d_name;
1401         struct page          *page;
1402         __u64                 pos    = 0;
1403         int                   dot_de;
1404         int                   rc     = LS_NONE_FIRST_DE;
1405         ENTRY;
1406
1407         lli->lli_sa_pos = 0;
1408         ll_dir_chain_init(&chain);
1409         page = ll_get_dir_page(NULL, dir, pos, &chain);
1410
1411         while (1) {
1412                 struct lu_dirpage *dp;
1413                 struct lu_dirent  *ent;
1414
1415                 if (IS_ERR(page)) {
1416                         struct ll_inode_info *lli = ll_i2info(dir);
1417
1418                         rc = PTR_ERR(page);
1419                         CERROR("error reading dir "DFID" at "LPU64": "
1420                                "[rc %d] [parent %u]\n",
1421                                PFID(ll_inode2fid(dir)), pos,
1422                                rc, lli->lli_opendir_pid);
1423                         break;
1424                 }
1425
1426                 dp = page_address(page);
1427                 for (ent = lu_dirent_start(dp); ent != NULL;
1428                      ent = lu_dirent_next(ent)) {
1429                         __u64 hash;
1430                         int namelen;
1431                         char *name;
1432
1433                         hash = le64_to_cpu(ent->lde_hash);
1434                         /* The ll_get_dir_page() can return any page containing
1435                          * the given hash which may be not the start hash. */
1436                         if (unlikely(hash < pos))
1437                                 continue;
1438
1439                         namelen = le16_to_cpu(ent->lde_namelen);
1440                         if (unlikely(namelen == 0))
1441                                 /*
1442                                  * skip dummy record.
1443                                  */
1444                                 continue;
1445
1446                         name = ent->lde_name;
1447                         if (name[0] == '.') {
1448                                 if (namelen == 1)
1449                                         /*
1450                                          * skip "."
1451                                          */
1452                                         continue;
1453                                 else if (name[1] == '.' && namelen == 2)
1454                                         /*
1455                                          * skip ".."
1456                                          */
1457                                         continue;
1458                                 else
1459                                         dot_de = 1;
1460                         } else {
1461                                 dot_de = 0;
1462                         }
1463
1464                         if (dot_de && target->name[0] != '.') {
1465                                 CDEBUG(D_READA, "%.*s skip hidden file %.*s\n",
1466                                        target->len, target->name,
1467                                        namelen, name);
1468                                 continue;
1469                         }
1470
1471                         if (target->len != namelen ||
1472                             memcmp(target->name, name, namelen) != 0)
1473                                 rc = LS_NONE_FIRST_DE;
1474                         else if (!dot_de)
1475                                 rc = LS_FIRST_DE;
1476                         else
1477                                 rc = LS_FIRST_DOT_DE;
1478
1479                         ll_release_page(page, 0);
1480                         GOTO(out, rc);
1481                 }
1482                 pos = le64_to_cpu(dp->ldp_hash_end);
1483                 if (pos == MDS_DIR_END_OFF) {
1484                         /*
1485                          * End of directory reached.
1486                          */
1487                         ll_release_page(page, 0);
1488                         break;
1489                 } else if (1) {
1490                         /*
1491                          * chain is exhausted
1492                          * Normal case: continue to the next page.
1493                          */
1494                         ll_release_page(page, le32_to_cpu(dp->ldp_flags) &
1495                                               LDF_COLLIDE);
1496                         lli->lli_sa_pos = pos;
1497                         page = ll_get_dir_page(NULL, dir, pos, &chain);
1498                 } else {
1499                         /*
1500                          * go into overflow page.
1501                          */
1502                         LASSERT(le32_to_cpu(dp->ldp_flags) & LDF_COLLIDE);
1503                         ll_release_page(page, 1);
1504                 }
1505         }
1506         EXIT;
1507
1508 out:
1509         ll_dir_chain_fini(&chain);
1510         return rc;
1511 }
1512
1513 static void
1514 ll_sai_unplug(struct ll_statahead_info *sai, struct ll_sa_entry *entry)
1515 {
1516         struct ptlrpc_thread *thread = &sai->sai_thread;
1517         struct ll_sb_info    *sbi    = ll_i2sbi(sai->sai_inode);
1518         int                   hit;
1519         ENTRY;
1520
1521         if (entry != NULL && entry->se_stat == SA_ENTRY_SUCC)
1522                 hit = 1;
1523         else
1524                 hit = 0;
1525
1526         ll_sa_entry_fini(sai, entry);
1527         if (hit) {
1528                 sai->sai_hit++;
1529                 sai->sai_consecutive_miss = 0;
1530                 sai->sai_max = min(2 * sai->sai_max, sbi->ll_sa_max);
1531         } else {
1532                 struct ll_inode_info *lli = ll_i2info(sai->sai_inode);
1533
1534                 sai->sai_miss++;
1535                 sai->sai_consecutive_miss++;
1536                 if (sa_low_hit(sai) && thread_is_running(thread)) {
1537                         atomic_inc(&sbi->ll_sa_wrong);
1538                         CDEBUG(D_READA, "Statahead for dir "DFID" hit "
1539                                "ratio too low: hit/miss "LPU64"/"LPU64
1540                                ", sent/replied "LPU64"/"LPU64", stopping "
1541                                "statahead thread: pid %d\n",
1542                                PFID(&lli->lli_fid), sai->sai_hit,
1543                                sai->sai_miss, sai->sai_sent,
1544                                sai->sai_replied, cfs_curproc_pid());
1545                         cfs_spin_lock(&lli->lli_sa_lock);
1546                         if (!thread_is_stopped(thread))
1547                                 thread_set_flags(thread, SVC_STOPPING);
1548                         cfs_spin_unlock(&lli->lli_sa_lock);
1549                 }
1550         }
1551
1552         if (!thread_is_stopped(thread))
1553                 cfs_waitq_signal(&thread->t_ctl_waitq);
1554
1555         EXIT;
1556 }
1557
1558 /**
1559  * Start statahead thread if this is the first dir entry.
1560  * Otherwise if a thread is started already, wait it until it is ahead of me.
1561  * \retval 1       -- find entry with lock in cache, the caller needs to do
1562  *                    nothing.
1563  * \retval 0       -- find entry in cache, but without lock, the caller needs
1564  *                    refresh from MDS.
1565  * \retval others  -- the caller need to process as non-statahead.
1566  */
1567 int do_statahead_enter(struct inode *dir, struct dentry **dentryp,
1568                        int only_unplug)
1569 {
1570         struct ll_inode_info     *lli   = ll_i2info(dir);
1571         struct ll_statahead_info *sai   = lli->lli_sai;
1572         struct dentry            *parent;
1573         struct ll_sa_entry       *entry;
1574         struct ptlrpc_thread     *thread;
1575         struct l_wait_info        lwi   = { 0 };
1576         int                       rc    = 0;
1577         ENTRY;
1578
1579         LASSERT(lli->lli_opendir_pid == cfs_curproc_pid());
1580
1581         if (sai) {
1582                 thread = &sai->sai_thread;
1583                 if (unlikely(thread_is_stopped(thread) &&
1584                              cfs_list_empty(&sai->sai_entries_stated))) {
1585                         /* to release resource */
1586                         ll_stop_statahead(dir, lli->lli_opendir_key);
1587                         RETURN(-EAGAIN);
1588                 }
1589
1590                 if ((*dentryp)->d_name.name[0] == '.') {
1591                         if (sai->sai_ls_all ||
1592                             sai->sai_miss_hidden >= sai->sai_skip_hidden) {
1593                                 /*
1594                                  * Hidden dentry is the first one, or statahead
1595                                  * thread does not skip so many hidden dentries
1596                                  * before "sai_ls_all" enabled as below.
1597                                  */
1598                         } else {
1599                                 if (!sai->sai_ls_all)
1600                                         /*
1601                                          * It maybe because hidden dentry is not
1602                                          * the first one, "sai_ls_all" was not
1603                                          * set, then "ls -al" missed. Enable
1604                                          * "sai_ls_all" for such case.
1605                                          */
1606                                         sai->sai_ls_all = 1;
1607
1608                                 /*
1609                                  * Such "getattr" has been skipped before
1610                                  * "sai_ls_all" enabled as above.
1611                                  */
1612                                 sai->sai_miss_hidden++;
1613                                 RETURN(-EAGAIN);
1614                         }
1615                 }
1616
1617                 entry = ll_sa_entry_get_byname(sai, &(*dentryp)->d_name);
1618                 if (entry == NULL || only_unplug) {
1619                         ll_sai_unplug(sai, entry);
1620                         RETURN(entry ? 1 : -EAGAIN);
1621                 }
1622
1623                 while (!ll_sa_entry_stated(entry) &&
1624                        sai->sai_in_readpage &&
1625                        !sa_received_empty(sai))
1626                         do_statahead_interpret(sai, entry);
1627
1628                 if (!ll_sa_entry_stated(entry)) {
1629                         sai->sai_index_wait = entry->se_index;
1630                         lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(30), NULL,
1631                                                LWI_ON_SIGNAL_NOOP, NULL);
1632                         rc = l_wait_event(sai->sai_waitq,
1633                                           ll_sa_entry_stated(entry) ||
1634                                           thread_is_stopped(thread),
1635                                           &lwi);
1636                         if (rc < 0) {
1637                                 ll_sai_unplug(sai, entry);
1638                                 RETURN(-EAGAIN);
1639                         }
1640                 }
1641
1642                 if (entry->se_stat == SA_ENTRY_SUCC &&
1643                     entry->se_inode != NULL) {
1644                         struct inode *inode = entry->se_inode;
1645                         struct lookup_intent it = { .it_op = IT_GETATTR,
1646                                                     .d.lustre.it_lock_handle =
1647                                                      entry->se_handle };
1648                         struct ll_dentry_data *lld;
1649                         __u64 bits;
1650
1651                         rc = md_revalidate_lock(ll_i2mdexp(dir), &it,
1652                                                 ll_inode2fid(inode), &bits);
1653                         if (rc == 1) {
1654                                 if ((*dentryp)->d_inode == NULL) {
1655                                         *dentryp = ll_find_alias(inode,
1656                                                                  *dentryp);
1657                                         lld = ll_d2d(*dentryp);
1658                                         if (unlikely(lld == NULL))
1659                                                 ll_dops_init(*dentryp, 1, 1);
1660                                 } else if ((*dentryp)->d_inode != inode) {
1661                                         /* revalidate, but inode is recreated */
1662                                         CDEBUG(D_READA,
1663                                               "stale dentry %.*s inode %lu/%u, "
1664                                               "statahead inode %lu/%u\n",
1665                                               (*dentryp)->d_name.len,
1666                                               (*dentryp)->d_name.name,
1667                                               (*dentryp)->d_inode->i_ino,
1668                                               (*dentryp)->d_inode->i_generation,
1669                                               inode->i_ino,
1670                                               inode->i_generation);
1671                                         ll_sai_unplug(sai, entry);
1672                                         RETURN(-ESTALE);
1673                                 } else {
1674                                         ll_dentry_rehash(*dentryp, 0);
1675                                         iput(inode);
1676                                 }
1677                                 entry->se_inode = NULL;
1678
1679                                 ll_dentry_reset_flags(*dentryp, bits);
1680                                 ll_intent_release(&it);
1681                         }
1682                 }
1683
1684                 ll_sai_unplug(sai, entry);
1685                 RETURN(rc);
1686         }
1687
1688         /* I am the "lli_opendir_pid" owner, only me can set "lli_sai". */
1689         rc = is_first_dirent(dir, *dentryp);
1690         if (rc == LS_NONE_FIRST_DE)
1691                 /* It is not "ls -{a}l" operation, no need statahead for it. */
1692                 GOTO(out, rc = -EAGAIN);
1693
1694         sai = ll_sai_alloc();
1695         if (sai == NULL)
1696                 GOTO(out, rc = -ENOMEM);
1697
1698         sai->sai_ls_all = (rc == LS_FIRST_DOT_DE);
1699         sai->sai_inode = igrab(dir);
1700         if (unlikely(sai->sai_inode == NULL)) {
1701                 CWARN("Do not start stat ahead on dying inode "DFID"\n",
1702                       PFID(&lli->lli_fid));
1703                 GOTO(out, rc = -ESTALE);
1704         }
1705
1706         /* get parent reference count here, and put it in ll_statahead_thread */
1707         parent = dget((*dentryp)->d_parent);
1708         if (unlikely(sai->sai_inode != parent->d_inode)) {
1709                 struct ll_inode_info *nlli = ll_i2info(parent->d_inode);
1710
1711                 CWARN("Race condition, someone changed %.*s just now: "
1712                       "old parent "DFID", new parent "DFID"\n",
1713                       (*dentryp)->d_name.len, (*dentryp)->d_name.name,
1714                       PFID(&lli->lli_fid), PFID(&nlli->lli_fid));
1715                 dput(parent);
1716                 iput(sai->sai_inode);
1717                 GOTO(out, rc = -EAGAIN);
1718         }
1719
1720         CDEBUG(D_READA, "start statahead thread: [pid %d] [parent %.*s]\n",
1721                cfs_curproc_pid(), parent->d_name.len, parent->d_name.name);
1722
1723         lli->lli_sai = sai;
1724         rc = cfs_create_thread(ll_statahead_thread, parent, 0);
1725         thread = &sai->sai_thread;
1726         if (rc < 0) {
1727                 CERROR("can't start ll_sa thread, rc: %d\n", rc);
1728                 dput(parent);
1729                 lli->lli_opendir_key = NULL;
1730                 thread_set_flags(thread, SVC_STOPPED);
1731                 thread_set_flags(&sai->sai_agl_thread, SVC_STOPPED);
1732                 ll_sai_put(sai);
1733                 LASSERT(lli->lli_sai == NULL);
1734                 RETURN(-EAGAIN);
1735         }
1736
1737         l_wait_event(thread->t_ctl_waitq,
1738                      thread_is_running(thread) || thread_is_stopped(thread),
1739                      &lwi);
1740
1741         /*
1742          * We don't stat-ahead for the first dirent since we are already in
1743          * lookup.
1744          */
1745         RETURN(-EAGAIN);
1746
1747 out:
1748         if (sai != NULL)
1749                 OBD_FREE_PTR(sai);
1750         cfs_spin_lock(&lli->lli_sa_lock);
1751         lli->lli_opendir_key = NULL;
1752         lli->lli_opendir_pid = 0;
1753         cfs_spin_unlock(&lli->lli_sa_lock);
1754         return rc;
1755 }