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 <obd_support.h>
  46 #include <lustre_lite.h>
  47 #include <lustre_dlm.h>
  48 #include "llite_internal.h"
  49
  50 #define SA_OMITTED_ENTRY_MAX 8ULL
  51
  52 typedef enum {
  53         /** negative values are for error cases */
  54         SA_ENTRY_INIT = 0,      /** init entry */
  55         SA_ENTRY_SUCC = 1,      /** stat succeed */
  56         SA_ENTRY_INVA = 2,      /** invalid entry */
  57         SA_ENTRY_DEST = 3,      /** entry to be destroyed */
  58 } se_stat_t;
  59
  60 struct ll_sa_entry {
  61         /* link into sai->sai_entries_{sent,received,stated} */
  62         cfs_list_t              se_list;
  63         /* link into sai hash table locally */
  64         cfs_list_t              se_hash;
  65         /* entry reference count */
  66         cfs_atomic_t            se_refcount;
  67         /* entry index in the sai */
  68         __u64                   se_index;
  69         /* low layer ldlm lock handle */
  70         __u64                   se_handle;
  71         /* entry status */
  72         se_stat_t               se_stat;
  73         /* entry size, contains name */
  74         int                     se_size;
  75         /* pointer to async getattr enqueue info */
  76         struct md_enqueue_info *se_minfo;
  77         /* pointer to the async getattr request */
  78         struct ptlrpc_request  *se_req;
  79         /* pointer to the target inode */
  80         struct inode           *se_inode;
  81         /* entry name */
  82         struct qstr             se_qstr;
  83 };
  84
  85 static unsigned int sai_generation = 0;
  86 static DEFINE_SPINLOCK(sai_generation_lock);
  87
  88 static inline int ll_sa_entry_unlinked(struct ll_sa_entry *entry)
  89 {
  90         return cfs_list_empty(&entry->se_list);
  91 }
  92
  93 static inline int ll_sa_entry_unhashed(struct ll_sa_entry *entry)
  94 {
  95         return cfs_list_empty(&entry->se_hash);
  96 }
  97
  98 /*
  99  * The entry only can be released by the caller, it is necessary to hold lock.
 100  */
 101 static inline int ll_sa_entry_stated(struct ll_sa_entry *entry)
 102 {
 103         smp_rmb();
 104         return (entry->se_stat != SA_ENTRY_INIT);
 105 }
 106
 107 static inline int ll_sa_entry_hash(int val)
 108 {
 109         return val & LL_SA_CACHE_MASK;
 110 }
 111
 112 /*
 113  * Insert entry to hash SA table.
 114  */
 115 static inline void
 116 ll_sa_entry_enhash(struct ll_statahead_info *sai, struct ll_sa_entry *entry)
 117 {
 118         int i = ll_sa_entry_hash(entry->se_qstr.hash);
 119
 120         spin_lock(&sai->sai_cache_lock[i]);
 121         cfs_list_add_tail(&entry->se_hash, &sai->sai_cache[i]);
 122         spin_unlock(&sai->sai_cache_lock[i]);
 123 }
 124
 125 /*
 126  * Remove entry from SA table.
 127  */
 128 static inline void
 129 ll_sa_entry_unhash(struct ll_statahead_info *sai, struct ll_sa_entry *entry)
 130 {
 131         int i = ll_sa_entry_hash(entry->se_qstr.hash);
 132
 133         spin_lock(&sai->sai_cache_lock[i]);
 134         cfs_list_del_init(&entry->se_hash);
 135         spin_unlock(&sai->sai_cache_lock[i]);
 136 }
 137
 138 static inline int agl_should_run(struct ll_statahead_info *sai,
 139                                  struct inode *inode)
 140 {
 141         return (inode != NULL && S_ISREG(inode->i_mode) && sai->sai_agl_valid);
 142 }
 143
 144 static inline struct ll_sa_entry *
 145 sa_first_received_entry(struct ll_statahead_info *sai)
 146 {
 147         return cfs_list_entry(sai->sai_entries_received.next,
 148                               struct ll_sa_entry, se_list);
 149 }
 150
 151 static inline struct ll_inode_info *
 152 agl_first_entry(struct ll_statahead_info *sai)
 153 {
 154         return cfs_list_entry(sai->sai_entries_agl.next,
 155                               struct ll_inode_info, lli_agl_list);
 156 }
 157
 158 static inline int sa_sent_full(struct ll_statahead_info *sai)
 159 {
 160         return cfs_atomic_read(&sai->sai_cache_count) >= sai->sai_max;
 161 }
 162
 163 static inline int sa_received_empty(struct ll_statahead_info *sai)
 164 {
 165         return cfs_list_empty(&sai->sai_entries_received);
 166 }
 167
 168 static inline int agl_list_empty(struct ll_statahead_info *sai)
 169 {
 170         return cfs_list_empty(&sai->sai_entries_agl);
 171 }
 172
 173 /**
 174  * (1) hit ratio less than 80%
 175  * or
 176  * (2) consecutive miss more than 8
 177  * then means low hit.
 178  */
 179 static inline int sa_low_hit(struct ll_statahead_info *sai)
 180 {
 181         return ((sai->sai_hit > 7 && sai->sai_hit < 4 * sai->sai_miss) ||
 182                 (sai->sai_consecutive_miss > 8));
 183 }
 184
 185 /*
 186  * If the given index is behind of statahead window more than
 187  * SA_OMITTED_ENTRY_MAX, then it is old.
 188  */
 189 static inline int is_omitted_entry(struct ll_statahead_info *sai, __u64 index)
 190 {
 191         return ((__u64)sai->sai_max + index + SA_OMITTED_ENTRY_MAX <
 192                  sai->sai_index);
 193 }
 194
 195 /*
 196  * Insert it into sai_entries_sent tail when init.
 197  */
 198 static struct ll_sa_entry *
 199 ll_sa_entry_alloc(struct ll_statahead_info *sai, __u64 index,
 200                   const char *name, int len)
 201 {
 202         struct ll_inode_info *lli;
 203         struct ll_sa_entry   *entry;
 204         int                   entry_size;
 205         char                 *dname;
 206         ENTRY;
 207
 208         entry_size = sizeof(struct ll_sa_entry) + (len & ~3) + 4;
 209         OBD_ALLOC(entry, entry_size);
 210         if (unlikely(entry == NULL))
 211                 RETURN(ERR_PTR(-ENOMEM));
 212
 213         CDEBUG(D_READA, "alloc sai entry %.*s(%p) index "LPU64"\n",
 214                len, name, entry, index);
 215
 216         entry->se_index = index;
 217
 218         /*
 219          * Statahead entry reference rules:
 220          *
 221          * 1) When statahead entry is initialized, its reference is set as 2.
 222          *    One reference is used by the directory scanner. When the scanner
 223          *    searches the statahead cache for the given name, it can perform
 224          *    lockless hash lookup (only the scanner can remove entry from hash
 225          *    list), and once found, it needn't to call "atomic_inc()" for the
 226          *    entry reference. So the performance is improved. After using the
 227          *    statahead entry, the scanner will call "atomic_dec()" to drop the
 228          *    reference held when initialization. If it is the last reference,
 229          *    the statahead entry will be freed.
 230          *
 231          * 2) All other threads, including statahead thread and ptlrpcd thread,
 232          *    when they process the statahead entry, the reference for target
 233          *    should be held to guarantee the entry will not be released by the
 234          *    directory scanner. After processing the entry, these threads will
 235          *    drop the entry reference. If it is the last reference, the entry
 236          *    will be freed.
 237          *
 238          *    The second reference when initializes the statahead entry is used
 239          *    by the statahead thread, following the rule 2).
 240          */
 241         cfs_atomic_set(&entry->se_refcount, 2);
 242         entry->se_stat = SA_ENTRY_INIT;
 243         entry->se_size = entry_size;
 244         dname = (char *)entry + sizeof(struct ll_sa_entry);
 245         memcpy(dname, name, len);
 246         dname[len] = 0;
 247         entry->se_qstr.hash = full_name_hash(name, len);
 248         entry->se_qstr.len = len;
 249         entry->se_qstr.name = dname;
 250
 251         lli = ll_i2info(sai->sai_inode);
 252         spin_lock(&lli->lli_sa_lock);
 253         cfs_list_add_tail(&entry->se_list, &sai->sai_entries_sent);
 254         spin_unlock(&lli->lli_sa_lock);
 255
 256         cfs_atomic_inc(&sai->sai_cache_count);
 257         ll_sa_entry_enhash(sai, entry);
 258
 259         RETURN(entry);
 260 }
 261
 262 /*
 263  * Used by the directory scanner to search entry with name.
 264  *
 265  * Only the caller can remove the entry from hash, so it is unnecessary to hold
 266  * hash lock. It is caller's duty to release the init refcount on the entry, so
 267  * it is also unnecessary to increase refcount on the entry.
 268  */
 269 static struct ll_sa_entry *
 270 ll_sa_entry_get_byname(struct ll_statahead_info *sai, const struct qstr *qstr)
 271 {
 272         struct ll_sa_entry *entry;
 273         int i = ll_sa_entry_hash(qstr->hash);
 274
 275         cfs_list_for_each_entry(entry, &sai->sai_cache[i], se_hash) {
 276                 if (entry->se_qstr.hash == qstr->hash &&
 277                     entry->se_qstr.len == qstr->len &&
 278                     memcmp(entry->se_qstr.name, qstr->name, qstr->len) == 0)
 279                         return entry;
 280         }
 281         return NULL;
 282 }
 283
 284 /*
 285  * Used by the async getattr request callback to find entry with index.
 286  *
 287  * Inside lli_sa_lock to prevent others to change the list during the search.
 288  * It needs to increase entry refcount before returning to guarantee that the
 289  * entry cannot be freed by others.
 290  */
 291 static struct ll_sa_entry *
 292 ll_sa_entry_get_byindex(struct ll_statahead_info *sai, __u64 index)
 293 {
 294         struct ll_sa_entry *entry;
 295
 296         cfs_list_for_each_entry(entry, &sai->sai_entries_sent, se_list) {
 297                 if (entry->se_index == index) {
 298                         cfs_atomic_inc(&entry->se_refcount);
 299                         return entry;
 300                 }
 301                 if (entry->se_index > index)
 302                         break;
 303         }
 304         return NULL;
 305 }
 306
 307 static void ll_sa_entry_cleanup(struct ll_statahead_info *sai,
 308                                  struct ll_sa_entry *entry)
 309 {
 310         struct md_enqueue_info *minfo = entry->se_minfo;
 311         struct ptlrpc_request  *req   = entry->se_req;
 312
 313         if (minfo) {
 314                 entry->se_minfo = NULL;
 315                 ll_intent_release(&minfo->mi_it);
 316                 iput(minfo->mi_dir);
 317                 OBD_FREE_PTR(minfo);
 318         }
 319
 320         if (req) {
 321                 entry->se_req = NULL;
 322                 ptlrpc_req_finished(req);
 323         }
 324 }
 325
 326 static void ll_sa_entry_put(struct ll_statahead_info *sai,
 327                              struct ll_sa_entry *entry)
 328 {
 329         if (cfs_atomic_dec_and_test(&entry->se_refcount)) {
 330                 CDEBUG(D_READA, "free sai entry %.*s(%p) index "LPU64"\n",
 331                        entry->se_qstr.len, entry->se_qstr.name, entry,
 332                        entry->se_index);
 333
 334                 LASSERT(ll_sa_entry_unhashed(entry));
 335                 LASSERT(ll_sa_entry_unlinked(entry));
 336
 337                 ll_sa_entry_cleanup(sai, entry);
 338                 if (entry->se_inode)
 339                         iput(entry->se_inode);
 340
 341                 OBD_FREE(entry, entry->se_size);
 342                 cfs_atomic_dec(&sai->sai_cache_count);
 343         }
 344 }
 345
 346 static inline void
 347 do_sai_entry_fini(struct ll_statahead_info *sai, struct ll_sa_entry *entry)
 348 {
 349         struct ll_inode_info *lli = ll_i2info(sai->sai_inode);
 350
 351         ll_sa_entry_unhash(sai, entry);
 352
 353         spin_lock(&lli->lli_sa_lock);
 354         entry->se_stat = SA_ENTRY_DEST;
 355         if (likely(!ll_sa_entry_unlinked(entry)))
 356                 cfs_list_del_init(&entry->se_list);
 357         spin_unlock(&lli->lli_sa_lock);
 358
 359         ll_sa_entry_put(sai, entry);
 360 }
 361
 362 /*
 363  * Delete it from sai_entries_stated list when fini.
 364  */
 365 static void
 366 ll_sa_entry_fini(struct ll_statahead_info *sai, struct ll_sa_entry *entry)
 367 {
 368         struct ll_sa_entry *pos, *next;
 369
 370         if (entry)
 371                 do_sai_entry_fini(sai, entry);
 372
 373         /* drop old entry from sent list */
 374         cfs_list_for_each_entry_safe(pos, next, &sai->sai_entries_sent,
 375                                      se_list) {
 376                 if (is_omitted_entry(sai, pos->se_index))
 377                         do_sai_entry_fini(sai, pos);
 378                 else
 379                         break;
 380         }
 381
 382         /* drop old entry from stated list */
 383         cfs_list_for_each_entry_safe(pos, next, &sai->sai_entries_stated,
 384                                      se_list) {
 385                 if (is_omitted_entry(sai, pos->se_index))
 386                         do_sai_entry_fini(sai, pos);
 387                 else
 388                         break;
 389         }
 390 }
 391
 392 /*
 393  * Inside lli_sa_lock.
 394  */
 395 static void
 396 do_sai_entry_to_stated(struct ll_statahead_info *sai,
 397                        struct ll_sa_entry *entry, int rc)
 398 {
 399         struct ll_sa_entry *se;
 400         cfs_list_t         *pos = &sai->sai_entries_stated;
 401
 402         if (!ll_sa_entry_unlinked(entry))
 403                 cfs_list_del_init(&entry->se_list);
 404
 405         cfs_list_for_each_entry_reverse(se, &sai->sai_entries_stated, se_list) {
 406                 if (se->se_index < entry->se_index) {
 407                         pos = &se->se_list;
 408                         break;
 409                 }
 410         }
 411
 412         cfs_list_add(&entry->se_list, pos);
 413         entry->se_stat = rc;
 414 }
 415
 416 /*
 417  * Move entry to sai_entries_stated and sort with the index.
 418  * \retval 1    -- entry to be destroyed.
 419  * \retval 0    -- entry is inserted into stated list.
 420  */
 421 static int
 422 ll_sa_entry_to_stated(struct ll_statahead_info *sai,
 423                       struct ll_sa_entry *entry, int rc)
 424 {
 425         struct ll_inode_info *lli = ll_i2info(sai->sai_inode);
 426         int                   ret = 1;
 427
 428         ll_sa_entry_cleanup(sai, entry);
 429
 430         spin_lock(&lli->lli_sa_lock);
 431         if (likely(entry->se_stat != SA_ENTRY_DEST)) {
 432                 do_sai_entry_to_stated(sai, entry, rc);
 433                 ret = 0;
 434         }
 435         spin_unlock(&lli->lli_sa_lock);
 436
 437         return ret;
 438 }
 439
 440 /*
 441  * Insert inode into the list of sai_entries_agl.
 442  */
 443 static void ll_agl_add(struct ll_statahead_info *sai,
 444                        struct inode *inode, int index)
 445 {
 446         struct ll_inode_info *child  = ll_i2info(inode);
 447         struct ll_inode_info *parent = ll_i2info(sai->sai_inode);
 448         int                   added  = 0;
 449
 450         spin_lock(&child->lli_agl_lock);
 451         if (child->lli_agl_index == 0) {
 452                 child->lli_agl_index = index;
 453                 spin_unlock(&child->lli_agl_lock);
 454
 455                 LASSERT(cfs_list_empty(&child->lli_agl_list));
 456
 457                 igrab(inode);
 458                 spin_lock(&parent->lli_agl_lock);
 459                 if (agl_list_empty(sai))
 460                         added = 1;
 461                 cfs_list_add_tail(&child->lli_agl_list, &sai->sai_entries_agl);
 462                 spin_unlock(&parent->lli_agl_lock);
 463         } else {
 464                 spin_unlock(&child->lli_agl_lock);
 465         }
 466
 467         if (added > 0)
 468                 cfs_waitq_signal(&sai->sai_agl_thread.t_ctl_waitq);
 469 }
 470
 471 static struct ll_statahead_info *ll_sai_alloc(void)
 472 {
 473         struct ll_statahead_info *sai;
 474         int                       i;
 475         ENTRY;
 476
 477         OBD_ALLOC_PTR(sai);
 478         if (!sai)
 479                 RETURN(NULL);
 480
 481         cfs_atomic_set(&sai->sai_refcount, 1);
 482
 483         spin_lock(&sai_generation_lock);
 484         sai->sai_generation = ++sai_generation;
 485         if (unlikely(sai_generation == 0))
 486                 sai->sai_generation = ++sai_generation;
 487         spin_unlock(&sai_generation_lock);
 488
 489         sai->sai_max = LL_SA_RPC_MIN;
 490         sai->sai_index = 1;
 491         cfs_waitq_init(&sai->sai_waitq);
 492         cfs_waitq_init(&sai->sai_thread.t_ctl_waitq);
 493         cfs_waitq_init(&sai->sai_agl_thread.t_ctl_waitq);
 494
 495         CFS_INIT_LIST_HEAD(&sai->sai_entries_sent);
 496         CFS_INIT_LIST_HEAD(&sai->sai_entries_received);
 497         CFS_INIT_LIST_HEAD(&sai->sai_entries_stated);
 498         CFS_INIT_LIST_HEAD(&sai->sai_entries_agl);
 499
 500         for (i = 0; i < LL_SA_CACHE_SIZE; i++) {
 501                 CFS_INIT_LIST_HEAD(&sai->sai_cache[i]);
 502                 spin_lock_init(&sai->sai_cache_lock[i]);
 503         }
 504         cfs_atomic_set(&sai->sai_cache_count, 0);
 505
 506         RETURN(sai);
 507 }
 508
 509 static inline struct ll_statahead_info *
 510 ll_sai_get(struct ll_statahead_info *sai)
 511 {
 512         cfs_atomic_inc(&sai->sai_refcount);
 513         return sai;
 514 }
 515
 516 static void ll_sai_put(struct ll_statahead_info *sai)
 517 {
 518         struct inode         *inode = sai->sai_inode;
 519         struct ll_inode_info *lli   = ll_i2info(inode);
 520         ENTRY;
 521
 522         if (cfs_atomic_dec_and_lock(&sai->sai_refcount, &lli->lli_sa_lock)) {
 523                 struct ll_sa_entry *entry, *next;
 524
 525                 if (unlikely(cfs_atomic_read(&sai->sai_refcount) > 0)) {
 526                         /* It is race case, the interpret callback just hold
 527                          * a reference count */
 528                         spin_unlock(&lli->lli_sa_lock);
 529                         RETURN_EXIT;
 530                 }
 531
 532                 LASSERT(lli->lli_opendir_key == NULL);
 533                 LASSERT(thread_is_stopped(&sai->sai_thread));
 534                 LASSERT(thread_is_stopped(&sai->sai_agl_thread));
 535
 536                 lli->lli_sai = NULL;
 537                 lli->lli_opendir_pid = 0;
 538                 spin_unlock(&lli->lli_sa_lock);
 539
 540                 if (sai->sai_sent > sai->sai_replied)
 541                         CDEBUG(D_READA,"statahead for dir "DFID" does not "
 542                               "finish: [sent:"LPU64"] [replied:"LPU64"]\n",
 543                               PFID(&lli->lli_fid),
 544                               sai->sai_sent, sai->sai_replied);
 545
 546                 cfs_list_for_each_entry_safe(entry, next,
 547                                              &sai->sai_entries_sent, se_list)
 548                         do_sai_entry_fini(sai, entry);
 549
 550                 LASSERT(sa_received_empty(sai));
 551
 552                 cfs_list_for_each_entry_safe(entry, next,
 553                                              &sai->sai_entries_stated, se_list)
 554                         do_sai_entry_fini(sai, entry);
 555
 556                 LASSERT(cfs_atomic_read(&sai->sai_cache_count) == 0);
 557                 LASSERT(agl_list_empty(sai));
 558
 559                 iput(inode);
 560                 OBD_FREE_PTR(sai);
 561         }
 562
 563         EXIT;
 564 }
 565
 566 /* Do NOT forget to drop inode refcount when into sai_entries_agl. */
 567 static void ll_agl_trigger(struct inode *inode, struct ll_statahead_info *sai)
 568 {
 569         struct ll_inode_info *lli   = ll_i2info(inode);
 570         __u64                 index = lli->lli_agl_index;
 571         int                   rc;
 572         ENTRY;
 573
 574         LASSERT(cfs_list_empty(&lli->lli_agl_list));
 575
 576         /* AGL maybe fall behind statahead with one entry */
 577         if (is_omitted_entry(sai, index + 1)) {
 578                 lli->lli_agl_index = 0;
 579                 iput(inode);
 580                 RETURN_EXIT;
 581         }
 582
 583         /* Someone is in glimpse (sync or async), do nothing. */
 584         rc = down_write_trylock(&lli->lli_glimpse_sem);
 585         if (rc == 0) {
 586                 lli->lli_agl_index = 0;
 587                 iput(inode);
 588                 RETURN_EXIT;
 589         }
 590
 591         /*
 592          * Someone triggered glimpse within 1 sec before.
 593          * 1) The former glimpse succeeded with glimpse lock granted by OST, and
 594          *    if the lock is still cached on client, AGL needs to do nothing. If
 595          *    it is cancelled by other client, AGL maybe cannot obtaion new lock
 596          *    for no glimpse callback triggered by AGL.
 597          * 2) The former glimpse succeeded, but OST did not grant glimpse lock.
 598          *    Under such case, it is quite possible that the OST will not grant
 599          *    glimpse lock for AGL also.
 600          * 3) The former glimpse failed, compared with other two cases, it is
 601          *    relative rare. AGL can ignore such case, and it will not muchly
 602          *    affect the performance.
 603          */
 604         if (lli->lli_glimpse_time != 0 &&
 605             cfs_time_before(cfs_time_shift(-1), lli->lli_glimpse_time)) {
 606                 up_write(&lli->lli_glimpse_sem);
 607                 lli->lli_agl_index = 0;
 608                 iput(inode);
 609                 RETURN_EXIT;
 610         }
 611
 612         CDEBUG(D_READA, "Handling (init) async glimpse: inode = "
 613                DFID", idx = "LPU64"\n", PFID(&lli->lli_fid), index);
 614
 615         cl_agl(inode);
 616         lli->lli_agl_index = 0;
 617         lli->lli_glimpse_time = cfs_time_current();
 618         up_write(&lli->lli_glimpse_sem);
 619
 620         CDEBUG(D_READA, "Handled (init) async glimpse: inode= "
 621                DFID", idx = "LPU64", rc = %d\n",
 622                PFID(&lli->lli_fid), index, rc);
 623
 624         iput(inode);
 625
 626         EXIT;
 627 }
 628
 629 static void do_statahead_interpret(struct ll_statahead_info *sai,
 630                                    struct ll_sa_entry *target)
 631 {
 632         struct inode           *dir   = sai->sai_inode;
 633         struct inode           *child;
 634         struct ll_inode_info   *lli   = ll_i2info(dir);
 635         struct ll_sa_entry     *entry;
 636         struct md_enqueue_info *minfo;
 637         struct lookup_intent   *it;
 638         struct ptlrpc_request  *req;
 639         struct mdt_body        *body;
 640         int                     rc    = 0;
 641         ENTRY;
 642
 643         spin_lock(&lli->lli_sa_lock);
 644         if (target != NULL && target->se_req != NULL &&
 645             !cfs_list_empty(&target->se_list)) {
 646                 entry = target;
 647         } else if (unlikely(sa_received_empty(sai))) {
 648                 spin_unlock(&lli->lli_sa_lock);
 649                 RETURN_EXIT;
 650         } else {
 651                 entry = sa_first_received_entry(sai);
 652         }
 653
 654         cfs_atomic_inc(&entry->se_refcount);
 655         cfs_list_del_init(&entry->se_list);
 656         spin_unlock(&lli->lli_sa_lock);
 657
 658         LASSERT(entry->se_handle != 0);
 659
 660         minfo = entry->se_minfo;
 661         it = &minfo->mi_it;
 662         req = entry->se_req;
 663         body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
 664         if (body == NULL)
 665                 GOTO(out, rc = -EFAULT);
 666
 667         child = entry->se_inode;
 668         if (child == NULL) {
 669                 /*
 670                  * lookup.
 671                  */
 672                 LASSERT(fid_is_zero(&minfo->mi_data.op_fid2));
 673
 674                 /* XXX: No fid in reply, this is probaly cross-ref case.
 675                  * SA can't handle it yet. */
 676                 if (body->valid & OBD_MD_MDS)
 677                         GOTO(out, rc = -EAGAIN);
 678         } else {
 679                 /*
 680                  * revalidate.
 681                  */
 682                 /* unlinked and re-created with the same name */
 683                 if (unlikely(!lu_fid_eq(&minfo->mi_data.op_fid2, &body->fid1))){
 684                         entry->se_inode = NULL;
 685                         iput(child);
 686                         child = NULL;
 687                 }
 688         }
 689
 690         it->d.lustre.it_lock_handle = entry->se_handle;
 691         rc = md_revalidate_lock(ll_i2mdexp(dir), it, ll_inode2fid(dir), NULL);
 692         if (rc != 1)
 693                 GOTO(out, rc = -EAGAIN);
 694
 695         rc = ll_prep_inode(&child, req, dir->i_sb, it);
 696         if (rc)
 697                 GOTO(out, rc);
 698
 699         CDEBUG(D_DLMTRACE, "setting l_data to inode %p (%lu/%u)\n",
 700                child, child->i_ino, child->i_generation);
 701         ll_set_lock_data(ll_i2sbi(dir)->ll_md_exp, child, it, NULL);
 702
 703         entry->se_inode = child;
 704
 705         if (agl_should_run(sai, child))
 706                 ll_agl_add(sai, child, entry->se_index);
 707
 708         EXIT;
 709
 710 out:
 711         /* The "ll_sa_entry_to_stated()" will drop related ldlm ibits lock
 712          * reference count by calling "ll_intent_drop_lock()" in spite of the
 713          * above operations failed or not. Do not worry about calling
 714          * "ll_intent_drop_lock()" more than once. */
 715         rc = ll_sa_entry_to_stated(sai, entry, rc < 0 ? rc : SA_ENTRY_SUCC);
 716         if (rc == 0 && entry->se_index == sai->sai_index_wait && target == NULL)
 717                 cfs_waitq_signal(&sai->sai_waitq);
 718         ll_sa_entry_put(sai, entry);
 719 }
 720
 721 static int ll_statahead_interpret(struct ptlrpc_request *req,
 722                                   struct md_enqueue_info *minfo, int rc)
 723 {
 724         struct lookup_intent     *it  = &minfo->mi_it;
 725         struct inode             *dir = minfo->mi_dir;
 726         struct ll_inode_info     *lli = ll_i2info(dir);
 727         struct ll_statahead_info *sai = NULL;
 728         struct ll_sa_entry       *entry;
 729         int                       wakeup;
 730         ENTRY;
 731
 732         if (it_disposition(it, DISP_LOOKUP_NEG))
 733                 rc = -ENOENT;
 734
 735         spin_lock(&lli->lli_sa_lock);
 736         /* stale entry */
 737         if (unlikely(lli->lli_sai == NULL ||
 738                      lli->lli_sai->sai_generation != minfo->mi_generation)) {
 739                 spin_unlock(&lli->lli_sa_lock);
 740                 GOTO(out, rc = -ESTALE);
 741         } else {
 742                 sai = ll_sai_get(lli->lli_sai);
 743                 if (unlikely(!thread_is_running(&sai->sai_thread))) {
 744                         sai->sai_replied++;
 745                         spin_unlock(&lli->lli_sa_lock);
 746                         GOTO(out, rc = -EBADFD);
 747                 }
 748
 749                 entry = ll_sa_entry_get_byindex(sai, minfo->mi_cbdata);
 750                 if (entry == NULL) {
 751                         sai->sai_replied++;
 752                         spin_unlock(&lli->lli_sa_lock);
 753                         GOTO(out, rc = -EIDRM);
 754                 }
 755
 756                 cfs_list_del_init(&entry->se_list);
 757                 if (rc != 0) {
 758                         sai->sai_replied++;
 759                         do_sai_entry_to_stated(sai, entry, rc);
 760                         spin_unlock(&lli->lli_sa_lock);
 761                         if (entry->se_index == sai->sai_index_wait)
 762                                 cfs_waitq_signal(&sai->sai_waitq);
 763                 } else {
 764                         entry->se_minfo = minfo;
 765                         entry->se_req = ptlrpc_request_addref(req);
 766                         /* Release the async ibits lock ASAP to avoid deadlock
 767                          * when statahead thread tries to enqueue lock on parent
 768                          * for readpage and other tries to enqueue lock on child
 769                          * with parent's lock held, for example: unlink. */
 770                         entry->se_handle = it->d.lustre.it_lock_handle;
 771                         ll_intent_drop_lock(it);
 772                         wakeup = sa_received_empty(sai);
 773                         cfs_list_add_tail(&entry->se_list,
 774                                           &sai->sai_entries_received);
 775                         sai->sai_replied++;
 776                         spin_unlock(&lli->lli_sa_lock);
 777                         if (wakeup)
 778                                 cfs_waitq_signal(&sai->sai_thread.t_ctl_waitq);
 779                 }
 780                 ll_sa_entry_put(sai, entry);
 781         }
 782
 783         EXIT;
 784
 785 out:
 786         if (rc != 0) {
 787                 ll_intent_release(it);
 788                 iput(dir);
 789                 OBD_FREE_PTR(minfo);
 790         }
 791         if (sai != NULL)
 792                 ll_sai_put(sai);
 793         return rc;
 794 }
 795
 796 static void sa_args_fini(struct md_enqueue_info *minfo,
 797                          struct ldlm_enqueue_info *einfo)
 798 {
 799         LASSERT(minfo && einfo);
 800         iput(minfo->mi_dir);
 801         capa_put(minfo->mi_data.op_capa1);
 802         capa_put(minfo->mi_data.op_capa2);
 803         OBD_FREE_PTR(minfo);
 804         OBD_FREE_PTR(einfo);
 805 }
 806
 807 /**
 808  * There is race condition between "capa_put" and "ll_statahead_interpret" for
 809  * accessing "op_data.op_capa[1,2]" as following:
 810  * "capa_put" releases "op_data.op_capa[1,2]"'s reference count after calling
 811  * "md_intent_getattr_async". But "ll_statahead_interpret" maybe run first, and
 812  * fill "op_data.op_capa[1,2]" as POISON, then cause "capa_put" access invalid
 813  * "ocapa". So here reserve "op_data.op_capa[1,2]" in "pcapa" before calling
 814  * "md_intent_getattr_async".
 815  */
 816 static int sa_args_init(struct inode *dir, struct inode *child,
 817                         struct ll_sa_entry *entry, struct md_enqueue_info **pmi,
 818                         struct ldlm_enqueue_info **pei,
 819                         struct obd_capa **pcapa)
 820 {
 821         struct qstr              *qstr = &entry->se_qstr;
 822         struct ll_inode_info     *lli  = ll_i2info(dir);
 823         struct md_enqueue_info   *minfo;
 824         struct ldlm_enqueue_info *einfo;
 825         struct md_op_data        *op_data;
 826
 827         OBD_ALLOC_PTR(einfo);
 828         if (einfo == NULL)
 829                 return -ENOMEM;
 830
 831         OBD_ALLOC_PTR(minfo);
 832         if (minfo == NULL) {
 833                 OBD_FREE_PTR(einfo);
 834                 return -ENOMEM;
 835         }
 836
 837         op_data = ll_prep_md_op_data(&minfo->mi_data, dir, child, qstr->name,
 838                                      qstr->len, 0, LUSTRE_OPC_ANY, NULL);
 839         if (IS_ERR(op_data)) {
 840                 OBD_FREE_PTR(einfo);
 841                 OBD_FREE_PTR(minfo);
 842                 return PTR_ERR(op_data);
 843         }
 844
 845         minfo->mi_it.it_op = IT_GETATTR;
 846         minfo->mi_dir = igrab(dir);
 847         minfo->mi_cb = ll_statahead_interpret;
 848         minfo->mi_generation = lli->lli_sai->sai_generation;
 849         minfo->mi_cbdata = entry->se_index;
 850
 851         einfo->ei_type   = LDLM_IBITS;
 852         einfo->ei_mode   = it_to_lock_mode(&minfo->mi_it);
 853         einfo->ei_cb_bl  = ll_md_blocking_ast;
 854         einfo->ei_cb_cp  = ldlm_completion_ast;
 855         einfo->ei_cb_gl  = NULL;
 856         einfo->ei_cbdata = NULL;
 857
 858         *pmi = minfo;
 859         *pei = einfo;
 860         pcapa[0] = op_data->op_capa1;
 861         pcapa[1] = op_data->op_capa2;
 862
 863         return 0;
 864 }
 865
 866 static int do_sa_lookup(struct inode *dir, struct ll_sa_entry *entry)
 867 {
 868         struct md_enqueue_info   *minfo;
 869         struct ldlm_enqueue_info *einfo;
 870         struct obd_capa          *capas[2];
 871         int                       rc;
 872         ENTRY;
 873
 874         rc = sa_args_init(dir, NULL, entry, &minfo, &einfo, capas);
 875         if (rc)
 876                 RETURN(rc);
 877
 878         rc = md_intent_getattr_async(ll_i2mdexp(dir), minfo, einfo);
 879         if (!rc) {
 880                 capa_put(capas[0]);
 881                 capa_put(capas[1]);
 882         } else {
 883                 sa_args_fini(minfo, einfo);
 884         }
 885
 886         RETURN(rc);
 887 }
 888
 889 /**
 890  * similar to ll_revalidate_it().
 891  * \retval      1 -- dentry valid
 892  * \retval      0 -- will send stat-ahead request
 893  * \retval others -- prepare stat-ahead request failed
 894  */
 895 static int do_sa_revalidate(struct inode *dir, struct ll_sa_entry *entry,
 896                             struct dentry *dentry)
 897 {
 898         struct inode             *inode = dentry->d_inode;
 899         struct lookup_intent      it = { .it_op = IT_GETATTR,
 900                                          .d.lustre.it_lock_handle = 0 };
 901         struct md_enqueue_info   *minfo;
 902         struct ldlm_enqueue_info *einfo;
 903         struct obd_capa          *capas[2];
 904         int rc;
 905         ENTRY;
 906
 907         if (unlikely(inode == NULL))
 908                 RETURN(1);
 909
 910         if (d_mountpoint(dentry))
 911                 RETURN(1);
 912
 913         if (unlikely(dentry == dentry->d_sb->s_root))
 914                 RETURN(1);
 915
 916         entry->se_inode = igrab(inode);
 917         rc = md_revalidate_lock(ll_i2mdexp(dir), &it, ll_inode2fid(inode),NULL);
 918         if (rc == 1) {
 919                 entry->se_handle = it.d.lustre.it_lock_handle;
 920                 ll_intent_release(&it);
 921                 RETURN(1);
 922         }
 923
 924         rc = sa_args_init(dir, inode, entry, &minfo, &einfo, capas);
 925         if (rc) {
 926                 entry->se_inode = NULL;
 927                 iput(inode);
 928                 RETURN(rc);
 929         }
 930
 931         rc = md_intent_getattr_async(ll_i2mdexp(dir), minfo, einfo);
 932         if (!rc) {
 933                 capa_put(capas[0]);
 934                 capa_put(capas[1]);
 935         } else {
 936                 entry->se_inode = NULL;
 937                 iput(inode);
 938                 sa_args_fini(minfo, einfo);
 939         }
 940
 941         RETURN(rc);
 942 }
 943
 944 static void ll_statahead_one(struct dentry *parent, const char* entry_name,
 945                              int entry_name_len)
 946 {
 947         struct inode             *dir    = parent->d_inode;
 948         struct ll_inode_info     *lli    = ll_i2info(dir);
 949         struct ll_statahead_info *sai    = lli->lli_sai;
 950         struct dentry            *dentry = NULL;
 951         struct ll_sa_entry       *entry;
 952         int                       rc;
 953         int                       rc1;
 954         ENTRY;
 955
 956         entry = ll_sa_entry_alloc(sai, sai->sai_index, entry_name,
 957                                   entry_name_len);
 958         if (IS_ERR(entry))
 959                 RETURN_EXIT;
 960
 961         dentry = d_lookup(parent, &entry->se_qstr);
 962         if (!dentry) {
 963                 rc = do_sa_lookup(dir, entry);
 964         } else {
 965                 rc = do_sa_revalidate(dir, entry, dentry);
 966                 if (rc == 1 && agl_should_run(sai, dentry->d_inode))
 967                         ll_agl_add(sai, dentry->d_inode, entry->se_index);
 968         }
 969
 970         if (dentry != NULL)
 971                 dput(dentry);
 972
 973         if (rc) {
 974                 rc1 = ll_sa_entry_to_stated(sai, entry,
 975                                         rc < 0 ? SA_ENTRY_INVA : SA_ENTRY_SUCC);
 976                 if (rc1 == 0 && entry->se_index == sai->sai_index_wait)
 977                         cfs_waitq_signal(&sai->sai_waitq);
 978         } else {
 979                 sai->sai_sent++;
 980         }
 981
 982         sai->sai_index++;
 983         /* drop one refcount on entry by ll_sa_entry_alloc */
 984         ll_sa_entry_put(sai, entry);
 985
 986         EXIT;
 987 }
 988
 989 static int ll_agl_thread(void *arg)
 990 {
 991         struct dentry            *parent = (struct dentry *)arg;
 992         struct inode             *dir    = parent->d_inode;
 993         struct ll_inode_info     *plli   = ll_i2info(dir);
 994         struct ll_inode_info     *clli;
 995         struct ll_sb_info        *sbi    = ll_i2sbi(dir);
 996         struct ll_statahead_info *sai    = ll_sai_get(plli->lli_sai);
 997         struct ptlrpc_thread     *thread = &sai->sai_agl_thread;
 998         struct l_wait_info        lwi    = { 0 };
 999         ENTRY;
1000
1001         {
1002                 char pname[16];
1003                 snprintf(pname, 15, "ll_agl_%u", plli->lli_opendir_pid);
1004                 cfs_daemonize(pname);
1005         }
1006
1007         CDEBUG(D_READA, "agl thread started: [pid %d] [parent %.*s]\n",
1008                cfs_curproc_pid(), parent->d_name.len, parent->d_name.name);
1009
1010         atomic_inc(&sbi->ll_agl_total);
1011         spin_lock(&plli->lli_agl_lock);
1012         sai->sai_agl_valid = 1;
1013         thread_set_flags(thread, SVC_RUNNING);
1014         spin_unlock(&plli->lli_agl_lock);
1015         cfs_waitq_signal(&thread->t_ctl_waitq);
1016
1017         while (1) {
1018                 l_wait_event(thread->t_ctl_waitq,
1019                              !agl_list_empty(sai) ||
1020                              !thread_is_running(thread),
1021                              &lwi);
1022
1023                 if (!thread_is_running(thread))
1024                         break;
1025
1026                 spin_lock(&plli->lli_agl_lock);
1027                 /* The statahead thread maybe help to process AGL entries,
1028                  * so check whether list empty again. */
1029                 if (!agl_list_empty(sai)) {
1030                         clli = agl_first_entry(sai);
1031                         cfs_list_del_init(&clli->lli_agl_list);
1032                         spin_unlock(&plli->lli_agl_lock);
1033                         ll_agl_trigger(&clli->lli_vfs_inode, sai);
1034                 } else {
1035                         spin_unlock(&plli->lli_agl_lock);
1036                 }
1037         }
1038
1039         spin_lock(&plli->lli_agl_lock);
1040         sai->sai_agl_valid = 0;
1041         while (!agl_list_empty(sai)) {
1042                 clli = agl_first_entry(sai);
1043                 cfs_list_del_init(&clli->lli_agl_list);
1044                 spin_unlock(&plli->lli_agl_lock);
1045                 clli->lli_agl_index = 0;
1046                 iput(&clli->lli_vfs_inode);
1047                 spin_lock(&plli->lli_agl_lock);
1048         }
1049         thread_set_flags(thread, SVC_STOPPED);
1050         spin_unlock(&plli->lli_agl_lock);
1051         cfs_waitq_signal(&thread->t_ctl_waitq);
1052         ll_sai_put(sai);
1053         CDEBUG(D_READA, "agl thread stopped: [pid %d] [parent %.*s]\n",
1054                cfs_curproc_pid(), parent->d_name.len, parent->d_name.name);
1055         RETURN(0);
1056 }
1057
1058 static void ll_start_agl(struct dentry *parent, struct ll_statahead_info *sai)
1059 {
1060         struct ptlrpc_thread *thread = &sai->sai_agl_thread;
1061         struct l_wait_info    lwi    = { 0 };
1062         int                   rc;
1063         ENTRY;
1064
1065         CDEBUG(D_READA, "start agl thread: [pid %d] [parent %.*s]\n",
1066                cfs_curproc_pid(), parent->d_name.len, parent->d_name.name);
1067
1068         rc = cfs_create_thread(ll_agl_thread, parent, 0);
1069         if (rc < 0) {
1070                 CERROR("can't start ll_agl thread, rc: %d\n", rc);
1071                 thread_set_flags(thread, SVC_STOPPED);
1072                 RETURN_EXIT;
1073         }
1074
1075         l_wait_event(thread->t_ctl_waitq,
1076                      thread_is_running(thread) || thread_is_stopped(thread),
1077                      &lwi);
1078         EXIT;
1079 }
1080
1081 static int ll_statahead_thread(void *arg)
1082 {
1083         struct dentry            *parent = (struct dentry *)arg;
1084         struct inode             *dir    = parent->d_inode;
1085         struct ll_inode_info     *plli   = ll_i2info(dir);
1086         struct ll_inode_info     *clli;
1087         struct ll_sb_info        *sbi    = ll_i2sbi(dir);
1088         struct ll_statahead_info *sai    = ll_sai_get(plli->lli_sai);
1089         struct ptlrpc_thread     *thread = &sai->sai_thread;
1090         struct ptlrpc_thread *agl_thread = &sai->sai_agl_thread;
1091         struct page              *page;
1092         __u64                     pos    = 0;
1093         int                       first  = 0;
1094         int                       rc     = 0;
1095         struct ll_dir_chain       chain;
1096         struct l_wait_info        lwi    = { 0 };
1097         ENTRY;
1098
1099         {
1100                 char pname[16];
1101                 snprintf(pname, 15, "ll_sa_%u", plli->lli_opendir_pid);
1102                 cfs_daemonize(pname);
1103         }
1104
1105         CDEBUG(D_READA, "statahead thread started: [pid %d] [parent %.*s]\n",
1106                cfs_curproc_pid(), parent->d_name.len, parent->d_name.name);
1107
1108         if (sbi->ll_flags & LL_SBI_AGL_ENABLED)
1109                 ll_start_agl(parent, sai);
1110
1111         atomic_inc(&sbi->ll_sa_total);
1112         spin_lock(&plli->lli_sa_lock);
1113         thread_set_flags(thread, SVC_RUNNING);
1114         spin_unlock(&plli->lli_sa_lock);
1115         cfs_waitq_signal(&thread->t_ctl_waitq);
1116
1117         ll_dir_chain_init(&chain);
1118         page = ll_get_dir_page(dir, pos, &chain);
1119
1120         while (1) {
1121                 struct lu_dirpage *dp;
1122                 struct lu_dirent  *ent;
1123
1124                 if (IS_ERR(page)) {
1125                         rc = PTR_ERR(page);
1126                         CDEBUG(D_READA, "error reading dir "DFID" at "LPU64
1127                                "/"LPU64": [rc %d] [parent %u]\n",
1128                                PFID(ll_inode2fid(dir)), pos, sai->sai_index,
1129                                rc, plli->lli_opendir_pid);
1130                         GOTO(out, rc);
1131                 }
1132
1133                 dp = page_address(page);
1134                 for (ent = lu_dirent_start(dp); ent != NULL;
1135                      ent = lu_dirent_next(ent)) {
1136                         __u64 hash;
1137                         int namelen;
1138                         char *name;
1139
1140                         hash = le64_to_cpu(ent->lde_hash);
1141                         if (unlikely(hash < pos))
1142                                 /*
1143                                  * Skip until we find target hash value.
1144                                  */
1145                                 continue;
1146
1147                         namelen = le16_to_cpu(ent->lde_namelen);
1148                         if (unlikely(namelen == 0))
1149                                 /*
1150                                  * Skip dummy record.
1151                                  */
1152                                 continue;
1153
1154                         name = ent->lde_name;
1155                         if (name[0] == '.') {
1156                                 if (namelen == 1) {
1157                                         /*
1158                                          * skip "."
1159                                          */
1160                                         continue;
1161                                 } else if (name[1] == '.' && namelen == 2) {
1162                                         /*
1163                                          * skip ".."
1164                                          */
1165                                         continue;
1166                                 } else if (!sai->sai_ls_all) {
1167                                         /*
1168                                          * skip hidden files.
1169                                          */
1170                                         sai->sai_skip_hidden++;
1171                                         continue;
1172                                 }
1173                         }
1174
1175                         /*
1176                          * don't stat-ahead first entry.
1177                          */
1178                         if (unlikely(++first == 1))
1179                                 continue;
1180
1181 keep_it:
1182                         l_wait_event(thread->t_ctl_waitq,
1183                                      !sa_sent_full(sai) ||
1184                                      !sa_received_empty(sai) ||
1185                                      !agl_list_empty(sai) ||
1186                                      !thread_is_running(thread),
1187                                      &lwi);
1188
1189 interpret_it:
1190                         while (!sa_received_empty(sai))
1191                                 do_statahead_interpret(sai, NULL);
1192
1193                         if (unlikely(!thread_is_running(thread))) {
1194                                 ll_release_page(page, 0);
1195                                 GOTO(out, rc = 0);
1196                         }
1197
1198                         /* If no window for metadata statahead, but there are
1199                          * some AGL entries to be triggered, then try to help
1200                          * to process the AGL entries. */
1201                         if (sa_sent_full(sai)) {
1202                                 spin_lock(&plli->lli_agl_lock);
1203                                 while (!agl_list_empty(sai)) {
1204                                         clli = agl_first_entry(sai);
1205                                         cfs_list_del_init(&clli->lli_agl_list);
1206                                         spin_unlock(&plli->lli_agl_lock);
1207                                         ll_agl_trigger(&clli->lli_vfs_inode,
1208                                                        sai);
1209
1210                                         if (!sa_received_empty(sai))
1211                                                 goto interpret_it;
1212
1213                                         if (unlikely(
1214                                                 !thread_is_running(thread))) {
1215                                                 ll_release_page(page, 0);
1216                                                 GOTO(out, rc = 0);
1217                                         }
1218
1219                                         if (!sa_sent_full(sai))
1220                                                 goto do_it;
1221
1222                                         spin_lock(&plli->lli_agl_lock);
1223                                 }
1224                                 spin_unlock(&plli->lli_agl_lock);
1225
1226                                 goto keep_it;
1227                         }
1228
1229 do_it:
1230                         ll_statahead_one(parent, name, namelen);
1231                 }
1232                 pos = le64_to_cpu(dp->ldp_hash_end);
1233                 if (pos == MDS_DIR_END_OFF) {
1234                         /*
1235                          * End of directory reached.
1236                          */
1237                         ll_release_page(page, 0);
1238                         while (1) {
1239                                 l_wait_event(thread->t_ctl_waitq,
1240                                              !sa_received_empty(sai) ||
1241                                              sai->sai_sent == sai->sai_replied||
1242                                              !thread_is_running(thread),
1243                                              &lwi);
1244
1245                                 while (!sa_received_empty(sai))
1246                                         do_statahead_interpret(sai, NULL);
1247
1248                                 if (unlikely(!thread_is_running(thread)))
1249                                         GOTO(out, rc = 0);
1250
1251                                 if (sai->sai_sent == sai->sai_replied &&
1252                                     sa_received_empty(sai))
1253                                         break;
1254                         }
1255
1256                         spin_lock(&plli->lli_agl_lock);
1257                         while (!agl_list_empty(sai) &&
1258                                thread_is_running(thread)) {
1259                                 clli = agl_first_entry(sai);
1260                                 cfs_list_del_init(&clli->lli_agl_list);
1261                                 spin_unlock(&plli->lli_agl_lock);
1262                                 ll_agl_trigger(&clli->lli_vfs_inode, sai);
1263                                 spin_lock(&plli->lli_agl_lock);
1264                         }
1265                         spin_unlock(&plli->lli_agl_lock);
1266
1267                         GOTO(out, rc = 0);
1268                 } else if (1) {
1269                         /*
1270                          * chain is exhausted.
1271                          * Normal case: continue to the next page.
1272                          */
1273                         ll_release_page(page, le32_to_cpu(dp->ldp_flags) &
1274                                               LDF_COLLIDE);
1275                         sai->sai_in_readpage = 1;
1276                         page = ll_get_dir_page(dir, pos, &chain);
1277                         sai->sai_in_readpage = 0;
1278                 } else {
1279                         LASSERT(le32_to_cpu(dp->ldp_flags) & LDF_COLLIDE);
1280                         ll_release_page(page, 1);
1281                         /*
1282                          * go into overflow page.
1283                          */
1284                 }
1285         }
1286         EXIT;
1287
1288 out:
1289         if (sai->sai_agl_valid) {
1290                 spin_lock(&plli->lli_agl_lock);
1291                 thread_set_flags(agl_thread, SVC_STOPPING);
1292                 spin_unlock(&plli->lli_agl_lock);
1293                 cfs_waitq_signal(&agl_thread->t_ctl_waitq);
1294
1295                 CDEBUG(D_READA, "stop agl thread: [pid %d]\n",
1296                        cfs_curproc_pid());
1297                 l_wait_event(agl_thread->t_ctl_waitq,
1298                              thread_is_stopped(agl_thread),
1299                              &lwi);
1300         } else {
1301                 /* Set agl_thread flags anyway. */
1302                 thread_set_flags(&sai->sai_agl_thread, SVC_STOPPED);
1303         }
1304         ll_dir_chain_fini(&chain);
1305         spin_lock(&plli->lli_sa_lock);
1306         if (!sa_received_empty(sai)) {
1307                 thread_set_flags(thread, SVC_STOPPING);
1308                 spin_unlock(&plli->lli_sa_lock);
1309
1310                 /* To release the resources held by received entries. */
1311                 while (!sa_received_empty(sai))
1312                         do_statahead_interpret(sai, NULL);
1313
1314                 spin_lock(&plli->lli_sa_lock);
1315         }
1316         thread_set_flags(thread, SVC_STOPPED);
1317         spin_unlock(&plli->lli_sa_lock);
1318         cfs_waitq_signal(&sai->sai_waitq);
1319         cfs_waitq_signal(&thread->t_ctl_waitq);
1320         ll_sai_put(sai);
1321         dput(parent);
1322         CDEBUG(D_READA, "statahead thread stopped: [pid %d] [parent %.*s]\n",
1323                cfs_curproc_pid(), parent->d_name.len, parent->d_name.name);
1324         return rc;
1325 }
1326
1327 /**
1328  * called in ll_file_release().
1329  */
1330 void ll_stop_statahead(struct inode *dir, void *key)
1331 {
1332         struct ll_inode_info *lli = ll_i2info(dir);
1333
1334         if (unlikely(key == NULL))
1335                 return;
1336
1337         spin_lock(&lli->lli_sa_lock);
1338         if (lli->lli_opendir_key != key || lli->lli_opendir_pid == 0) {
1339                 spin_unlock(&lli->lli_sa_lock);
1340                 return;
1341         }
1342
1343         lli->lli_opendir_key = NULL;
1344
1345         if (lli->lli_sai) {
1346                 struct l_wait_info lwi = { 0 };
1347                 struct ptlrpc_thread *thread = &lli->lli_sai->sai_thread;
1348
1349                 if (!thread_is_stopped(thread)) {
1350                         thread_set_flags(thread, SVC_STOPPING);
1351                         spin_unlock(&lli->lli_sa_lock);
1352                         cfs_waitq_signal(&thread->t_ctl_waitq);
1353
1354                         CDEBUG(D_READA, "stop statahead thread: [pid %d]\n",
1355                                cfs_curproc_pid());
1356                         l_wait_event(thread->t_ctl_waitq,
1357                                      thread_is_stopped(thread),
1358                                      &lwi);
1359                 } else {
1360                         spin_unlock(&lli->lli_sa_lock);
1361                 }
1362
1363                 /*
1364                  * Put the ref which was held when first statahead_enter.
1365                  * It maybe not the last ref for some statahead requests
1366                  * maybe inflight.
1367                  */
1368                 ll_sai_put(lli->lli_sai);
1369         } else {
1370                 lli->lli_opendir_pid = 0;
1371                 spin_unlock(&lli->lli_sa_lock);
1372         }
1373 }
1374
1375 enum {
1376         /**
1377          * not first dirent, or is "."
1378          */
1379         LS_NONE_FIRST_DE = 0,
1380         /**
1381          * the first non-hidden dirent
1382          */
1383         LS_FIRST_DE,
1384         /**
1385          * the first hidden dirent, that is "."
1386          */
1387         LS_FIRST_DOT_DE
1388 };
1389
1390 static int is_first_dirent(struct inode *dir, struct dentry *dentry)
1391 {
1392         struct ll_dir_chain   chain;
1393         struct qstr          *target = &dentry->d_name;
1394         struct page          *page;
1395         __u64                 pos    = 0;
1396         int                   dot_de;
1397         int                   rc     = LS_NONE_FIRST_DE;
1398         ENTRY;
1399
1400         ll_dir_chain_init(&chain);
1401         page = ll_get_dir_page(dir, pos, &chain);
1402
1403         while (1) {
1404                 struct lu_dirpage *dp;
1405                 struct lu_dirent  *ent;
1406
1407                 if (IS_ERR(page)) {
1408                         struct ll_inode_info *lli = ll_i2info(dir);
1409
1410                         rc = PTR_ERR(page);
1411                         CERROR("error reading dir "DFID" at "LPU64": "
1412                                "[rc %d] [parent %u]\n",
1413                                PFID(ll_inode2fid(dir)), pos,
1414                                rc, lli->lli_opendir_pid);
1415                         break;
1416                 }
1417
1418                 dp = page_address(page);
1419                 for (ent = lu_dirent_start(dp); ent != NULL;
1420                      ent = lu_dirent_next(ent)) {
1421                         __u64 hash;
1422                         int namelen;
1423                         char *name;
1424
1425                         hash = le64_to_cpu(ent->lde_hash);
1426                         /* The ll_get_dir_page() can return any page containing
1427                          * the given hash which may be not the start hash. */
1428                         if (unlikely(hash < pos))
1429                                 continue;
1430
1431                         namelen = le16_to_cpu(ent->lde_namelen);
1432                         if (unlikely(namelen == 0))
1433                                 /*
1434                                  * skip dummy record.
1435                                  */
1436                                 continue;
1437
1438                         name = ent->lde_name;
1439                         if (name[0] == '.') {
1440                                 if (namelen == 1)
1441                                         /*
1442                                          * skip "."
1443                                          */
1444                                         continue;
1445                                 else if (name[1] == '.' && namelen == 2)
1446                                         /*
1447                                          * skip ".."
1448                                          */
1449                                         continue;
1450                                 else
1451                                         dot_de = 1;
1452                         } else {
1453                                 dot_de = 0;
1454                         }
1455
1456                         if (dot_de && target->name[0] != '.') {
1457                                 CDEBUG(D_READA, "%.*s skip hidden file %.*s\n",
1458                                        target->len, target->name,
1459                                        namelen, name);
1460                                 continue;
1461                         }
1462
1463                         if (target->len != namelen ||
1464                             memcmp(target->name, name, namelen) != 0)
1465                                 rc = LS_NONE_FIRST_DE;
1466                         else if (!dot_de)
1467                                 rc = LS_FIRST_DE;
1468                         else
1469                                 rc = LS_FIRST_DOT_DE;
1470
1471                         ll_release_page(page, 0);
1472                         GOTO(out, rc);
1473                 }
1474                 pos = le64_to_cpu(dp->ldp_hash_end);
1475                 if (pos == MDS_DIR_END_OFF) {
1476                         /*
1477                          * End of directory reached.
1478                          */
1479                         ll_release_page(page, 0);
1480                         break;
1481                 } else if (1) {
1482                         /*
1483                          * chain is exhausted
1484                          * Normal case: continue to the next page.
1485                          */
1486                         ll_release_page(page, le32_to_cpu(dp->ldp_flags) &
1487                                               LDF_COLLIDE);
1488                         page = ll_get_dir_page(dir, pos, &chain);
1489                 } else {
1490                         /*
1491                          * go into overflow page.
1492                          */
1493                         LASSERT(le32_to_cpu(dp->ldp_flags) & LDF_COLLIDE);
1494                         ll_release_page(page, 1);
1495                 }
1496         }
1497         EXIT;
1498
1499 out:
1500         ll_dir_chain_fini(&chain);
1501         return rc;
1502 }
1503
1504 static void
1505 ll_sai_unplug(struct ll_statahead_info *sai, struct ll_sa_entry *entry)
1506 {
1507         struct ptlrpc_thread *thread = &sai->sai_thread;
1508         struct ll_sb_info    *sbi    = ll_i2sbi(sai->sai_inode);
1509         int                   hit;
1510         ENTRY;
1511
1512         if (entry != NULL && entry->se_stat == SA_ENTRY_SUCC)
1513                 hit = 1;
1514         else
1515                 hit = 0;
1516
1517         ll_sa_entry_fini(sai, entry);
1518         if (hit) {
1519                 sai->sai_hit++;
1520                 sai->sai_consecutive_miss = 0;
1521                 sai->sai_max = min(2 * sai->sai_max, sbi->ll_sa_max);
1522         } else {
1523                 struct ll_inode_info *lli = ll_i2info(sai->sai_inode);
1524
1525                 sai->sai_miss++;
1526                 sai->sai_consecutive_miss++;
1527                 if (sa_low_hit(sai) && thread_is_running(thread)) {
1528                         atomic_inc(&sbi->ll_sa_wrong);
1529                         CDEBUG(D_READA, "Statahead for dir "DFID" hit "
1530                                "ratio too low: hit/miss "LPU64"/"LPU64
1531                                ", sent/replied "LPU64"/"LPU64", stopping "
1532                                "statahead thread: pid %d\n",
1533                                PFID(&lli->lli_fid), sai->sai_hit,
1534                                sai->sai_miss, sai->sai_sent,
1535                                sai->sai_replied, cfs_curproc_pid());
1536                         spin_lock(&lli->lli_sa_lock);
1537                         if (!thread_is_stopped(thread))
1538                                 thread_set_flags(thread, SVC_STOPPING);
1539                         spin_unlock(&lli->lli_sa_lock);
1540                 }
1541         }
1542
1543         if (!thread_is_stopped(thread))
1544                 cfs_waitq_signal(&thread->t_ctl_waitq);
1545
1546         EXIT;
1547 }
1548
1549 /**
1550  * Start statahead thread if this is the first dir entry.
1551  * Otherwise if a thread is started already, wait it until it is ahead of me.
1552  * \retval 1       -- find entry with lock in cache, the caller needs to do
1553  *                    nothing.
1554  * \retval 0       -- find entry in cache, but without lock, the caller needs
1555  *                    refresh from MDS.
1556  * \retval others  -- the caller need to process as non-statahead.
1557  */
1558 int do_statahead_enter(struct inode *dir, struct dentry **dentryp,
1559                        int only_unplug)
1560 {
1561         struct ll_inode_info     *lli   = ll_i2info(dir);
1562         struct ll_statahead_info *sai   = lli->lli_sai;
1563         struct dentry            *parent;
1564         struct ll_sa_entry       *entry;
1565         struct ptlrpc_thread     *thread;
1566         struct l_wait_info        lwi   = { 0 };
1567         int                       rc    = 0;
1568         ENTRY;
1569
1570         LASSERT(lli->lli_opendir_pid == cfs_curproc_pid());
1571
1572         if (sai) {
1573                 thread = &sai->sai_thread;
1574                 if (unlikely(thread_is_stopped(thread) &&
1575                              cfs_list_empty(&sai->sai_entries_stated))) {
1576                         /* to release resource */
1577                         ll_stop_statahead(dir, lli->lli_opendir_key);
1578                         RETURN(-EAGAIN);
1579                 }
1580
1581                 if ((*dentryp)->d_name.name[0] == '.') {
1582                         if (sai->sai_ls_all ||
1583                             sai->sai_miss_hidden >= sai->sai_skip_hidden) {
1584                                 /*
1585                                  * Hidden dentry is the first one, or statahead
1586                                  * thread does not skip so many hidden dentries
1587                                  * before "sai_ls_all" enabled as below.
1588                                  */
1589                         } else {
1590                                 if (!sai->sai_ls_all)
1591                                         /*
1592                                          * It maybe because hidden dentry is not
1593                                          * the first one, "sai_ls_all" was not
1594                                          * set, then "ls -al" missed. Enable
1595                                          * "sai_ls_all" for such case.
1596                                          */
1597                                         sai->sai_ls_all = 1;
1598
1599                                 /*
1600                                  * Such "getattr" has been skipped before
1601                                  * "sai_ls_all" enabled as above.
1602                                  */
1603                                 sai->sai_miss_hidden++;
1604                                 RETURN(-EAGAIN);
1605                         }
1606                 }
1607
1608                 entry = ll_sa_entry_get_byname(sai, &(*dentryp)->d_name);
1609                 if (entry == NULL || only_unplug) {
1610                         ll_sai_unplug(sai, entry);
1611                         RETURN(entry ? 1 : -EAGAIN);
1612                 }
1613
1614                 while (!ll_sa_entry_stated(entry) &&
1615                        sai->sai_in_readpage &&
1616                        !sa_received_empty(sai))
1617                         do_statahead_interpret(sai, entry);
1618
1619                 if (!ll_sa_entry_stated(entry)) {
1620                         sai->sai_index_wait = entry->se_index;
1621                         lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(30), NULL,
1622                                                LWI_ON_SIGNAL_NOOP, NULL);
1623                         rc = l_wait_event(sai->sai_waitq,
1624                                           ll_sa_entry_stated(entry) ||
1625                                           thread_is_stopped(thread),
1626                                           &lwi);
1627                         if (rc < 0) {
1628                                 ll_sai_unplug(sai, entry);
1629                                 RETURN(-EAGAIN);
1630                         }
1631                 }
1632
1633                 if (entry->se_stat == SA_ENTRY_SUCC &&
1634                     entry->se_inode != NULL) {
1635                         struct inode *inode = entry->se_inode;
1636                         struct lookup_intent it = { .it_op = IT_GETATTR,
1637                                                     .d.lustre.it_lock_handle =
1638                                                      entry->se_handle };
1639                         __u64 bits;
1640
1641                         rc = md_revalidate_lock(ll_i2mdexp(dir), &it,
1642                                                 ll_inode2fid(inode), &bits);
1643                         if (rc == 1) {
1644                                 if ((*dentryp)->d_inode == NULL) {
1645                                         *dentryp = ll_splice_alias(inode,
1646                                                                    *dentryp);
1647                                 } else if ((*dentryp)->d_inode != inode) {
1648                                         /* revalidate, but inode is recreated */
1649                                         CDEBUG(D_READA,
1650                                               "stale dentry %.*s inode %lu/%u, "
1651                                               "statahead inode %lu/%u\n",
1652                                               (*dentryp)->d_name.len,
1653                                               (*dentryp)->d_name.name,
1654                                               (*dentryp)->d_inode->i_ino,
1655                                               (*dentryp)->d_inode->i_generation,
1656                                               inode->i_ino,
1657                                               inode->i_generation);
1658                                         ll_sai_unplug(sai, entry);
1659                                         RETURN(-ESTALE);
1660                                 } else {
1661                                         iput(inode);
1662                                 }
1663                                 entry->se_inode = NULL;
1664
1665                                 if ((bits & MDS_INODELOCK_LOOKUP) &&
1666                                     d_lustre_invalid(*dentryp))
1667                                         d_lustre_revalidate(*dentryp);
1668                                 ll_intent_release(&it);
1669                         }
1670                 }
1671
1672                 ll_sai_unplug(sai, entry);
1673                 RETURN(rc);
1674         }
1675
1676         /* I am the "lli_opendir_pid" owner, only me can set "lli_sai". */
1677         rc = is_first_dirent(dir, *dentryp);
1678         if (rc == LS_NONE_FIRST_DE)
1679                 /* It is not "ls -{a}l" operation, no need statahead for it. */
1680                 GOTO(out, rc = -EAGAIN);
1681
1682         sai = ll_sai_alloc();
1683         if (sai == NULL)
1684                 GOTO(out, rc = -ENOMEM);
1685
1686         sai->sai_ls_all = (rc == LS_FIRST_DOT_DE);
1687         sai->sai_inode = igrab(dir);
1688         if (unlikely(sai->sai_inode == NULL)) {
1689                 CWARN("Do not start stat ahead on dying inode "DFID"\n",
1690                       PFID(&lli->lli_fid));
1691                 GOTO(out, rc = -ESTALE);
1692         }
1693
1694         /* get parent reference count here, and put it in ll_statahead_thread */
1695         parent = dget((*dentryp)->d_parent);
1696         if (unlikely(sai->sai_inode != parent->d_inode)) {
1697                 struct ll_inode_info *nlli = ll_i2info(parent->d_inode);
1698
1699                 CWARN("Race condition, someone changed %.*s just now: "
1700                       "old parent "DFID", new parent "DFID"\n",
1701                       (*dentryp)->d_name.len, (*dentryp)->d_name.name,
1702                       PFID(&lli->lli_fid), PFID(&nlli->lli_fid));
1703                 dput(parent);
1704                 iput(sai->sai_inode);
1705                 GOTO(out, rc = -EAGAIN);
1706         }
1707
1708         CDEBUG(D_READA, "start statahead thread: [pid %d] [parent %.*s]\n",
1709                cfs_curproc_pid(), parent->d_name.len, parent->d_name.name);
1710
1711         lli->lli_sai = sai;
1712         rc = cfs_create_thread(ll_statahead_thread, parent, 0);
1713         thread = &sai->sai_thread;
1714         if (rc < 0) {
1715                 CERROR("can't start ll_sa thread, rc: %d\n", rc);
1716                 dput(parent);
1717                 lli->lli_opendir_key = NULL;
1718                 thread_set_flags(thread, SVC_STOPPED);
1719                 thread_set_flags(&sai->sai_agl_thread, SVC_STOPPED);
1720                 ll_sai_put(sai);
1721                 LASSERT(lli->lli_sai == NULL);
1722                 RETURN(-EAGAIN);
1723         }
1724
1725         l_wait_event(thread->t_ctl_waitq,
1726                      thread_is_running(thread) || thread_is_stopped(thread),
1727                      &lwi);
1728
1729         /*
1730          * We don't stat-ahead for the first dirent since we are already in
1731          * lookup.
1732          */
1733         RETURN(-EAGAIN);
1734
1735 out:
1736         if (sai != NULL)
1737                 OBD_FREE_PTR(sai);
1738         spin_lock(&lli->lli_sa_lock);
1739         lli->lli_opendir_key = NULL;
1740         lli->lli_opendir_pid = 0;
1741         spin_unlock(&lli->lli_sa_lock);
1742         return rc;
1743 }