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.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2011, 2016, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  */
  32
  33 #include <linux/fs.h>
  34 #include <linux/sched.h>
  35 #include <linux/kthread.h>
  36 #include <linux/mm.h>
  37 #include <linux/highmem.h>
  38 #include <linux/pagemap.h>
  39
  40 #define DEBUG_SUBSYSTEM S_LLITE
  41
  42 #include <obd_support.h>
  43 #include <lustre_dlm.h>
  44 #include "llite_internal.h"
  45
  46 #define SA_OMITTED_ENTRY_MAX 8ULL
  47
  48 typedef enum {
  49         /** negative values are for error cases */
  50         SA_ENTRY_INIT = 0,      /** init entry */
  51         SA_ENTRY_SUCC = 1,      /** stat succeed */
  52         SA_ENTRY_INVA = 2,      /** invalid entry */
  53 } se_state_t;
  54
  55 /* sa_entry is not refcounted: statahead thread allocates it and do async stat,
  56  * and in async stat callback ll_statahead_interpret() will add it into
  57  * sai_interim_entries, later statahead thread will call sa_handle_callback() to
  58  * instantiate entry and move it into sai_entries, and then only scanner process
  59  * can access and free it. */
  60 struct sa_entry {
  61         /* link into sai_interim_entries or sai_entries */
  62         struct list_head        se_list;
  63         /* link into sai hash table locally */
  64         struct list_head        se_hash;
  65         /* entry index in the sai */
  66         __u64                   se_index;
  67         /* low layer ldlm lock handle */
  68         __u64                   se_handle;
  69         /* entry status */
  70         se_state_t              se_state;
  71         /* entry size, contains name */
  72         int                     se_size;
  73         /* pointer to async getattr enqueue info */
  74         struct md_enqueue_info *se_minfo;
  75         /* pointer to the async getattr request */
  76         struct ptlrpc_request  *se_req;
  77         /* pointer to the target inode */
  78         struct inode           *se_inode;
  79         /* entry name */
  80         struct qstr             se_qstr;
  81         /* entry fid */
  82         struct lu_fid           se_fid;
  83 };
  84
  85 static unsigned int sai_generation = 0;
  86 static DEFINE_SPINLOCK(sai_generation_lock);
  87
  88 static inline int sa_unhashed(struct sa_entry *entry)
  89 {
  90         return list_empty(&entry->se_hash);
  91 }
  92
  93 /* sa_entry is ready to use */
  94 static inline int sa_ready(struct sa_entry *entry)
  95 {
  96         smp_rmb();
  97         return (entry->se_state != SA_ENTRY_INIT);
  98 }
  99
 100 /* hash value to put in sai_cache */
 101 static inline int sa_hash(int val)
 102 {
 103         return val & LL_SA_CACHE_MASK;
 104 }
 105
 106 /* hash entry into sai_cache */
 107 static inline void
 108 sa_rehash(struct ll_statahead_info *sai, struct sa_entry *entry)
 109 {
 110         int i = sa_hash(entry->se_qstr.hash);
 111
 112         spin_lock(&sai->sai_cache_lock[i]);
 113         list_add_tail(&entry->se_hash, &sai->sai_cache[i]);
 114         spin_unlock(&sai->sai_cache_lock[i]);
 115 }
 116
 117 /* unhash entry from sai_cache */
 118 static inline void
 119 sa_unhash(struct ll_statahead_info *sai, struct sa_entry *entry)
 120 {
 121         int i = sa_hash(entry->se_qstr.hash);
 122
 123         spin_lock(&sai->sai_cache_lock[i]);
 124         list_del_init(&entry->se_hash);
 125         spin_unlock(&sai->sai_cache_lock[i]);
 126 }
 127
 128 static inline int agl_should_run(struct ll_statahead_info *sai,
 129                                  struct inode *inode)
 130 {
 131         return (inode != NULL && S_ISREG(inode->i_mode) && sai->sai_agl_valid);
 132 }
 133
 134 static inline struct ll_inode_info *
 135 agl_first_entry(struct ll_statahead_info *sai)
 136 {
 137         return list_entry(sai->sai_agls.next, struct ll_inode_info,
 138                           lli_agl_list);
 139 }
 140
 141 /* statahead window is full */
 142 static inline int sa_sent_full(struct ll_statahead_info *sai)
 143 {
 144         return atomic_read(&sai->sai_cache_count) >= sai->sai_max;
 145 }
 146
 147 /* got async stat replies */
 148 static inline int sa_has_callback(struct ll_statahead_info *sai)
 149 {
 150         return !list_empty(&sai->sai_interim_entries);
 151 }
 152
 153 static inline int agl_list_empty(struct ll_statahead_info *sai)
 154 {
 155         return list_empty(&sai->sai_agls);
 156 }
 157
 158 /**
 159  * (1) hit ratio less than 80%
 160  * or
 161  * (2) consecutive miss more than 8
 162  * then means low hit.
 163  */
 164 static inline int sa_low_hit(struct ll_statahead_info *sai)
 165 {
 166         return ((sai->sai_hit > 7 && sai->sai_hit < 4 * sai->sai_miss) ||
 167                 (sai->sai_consecutive_miss > 8));
 168 }
 169
 170 /*
 171  * if the given index is behind of statahead window more than
 172  * SA_OMITTED_ENTRY_MAX, then it is old.
 173  */
 174 static inline int is_omitted_entry(struct ll_statahead_info *sai, __u64 index)
 175 {
 176         return ((__u64)sai->sai_max + index + SA_OMITTED_ENTRY_MAX <
 177                  sai->sai_index);
 178 }
 179
 180 /* allocate sa_entry and hash it to allow scanner process to find it */
 181 static struct sa_entry *
 182 sa_alloc(struct ll_statahead_info *sai, __u64 index, const char *name, int len,
 183          const struct lu_fid *fid)
 184 {
 185         struct ll_inode_info *lli;
 186         struct sa_entry *entry;
 187         int entry_size;
 188         char *dname;
 189         ENTRY;
 190
 191         entry_size = sizeof(struct sa_entry) + (len & ~3) + 4;
 192         OBD_ALLOC(entry, entry_size);
 193         if (unlikely(entry == NULL))
 194                 RETURN(ERR_PTR(-ENOMEM));
 195
 196         CDEBUG(D_READA, "alloc sa entry %.*s(%p) index %llu\n",
 197                len, name, entry, index);
 198
 199         entry->se_index = index;
 200
 201         entry->se_state = SA_ENTRY_INIT;
 202         entry->se_size = entry_size;
 203         dname = (char *)entry + sizeof(struct sa_entry);
 204         memcpy(dname, name, len);
 205         dname[len] = 0;
 206         entry->se_qstr.hash = full_name_hash(name, len);
 207         entry->se_qstr.len = len;
 208         entry->se_qstr.name = dname;
 209         entry->se_fid = *fid;
 210
 211         lli = ll_i2info(sai->sai_dentry->d_inode);
 212
 213         spin_lock(&lli->lli_sa_lock);
 214         INIT_LIST_HEAD(&entry->se_list);
 215         sa_rehash(sai, entry);
 216         spin_unlock(&lli->lli_sa_lock);
 217
 218         atomic_inc(&sai->sai_cache_count);
 219
 220         RETURN(entry);
 221 }
 222
 223 /* free sa_entry, which should have been unhashed and not in any list */
 224 static void sa_free(struct ll_statahead_info *sai, struct sa_entry *entry)
 225 {
 226         CDEBUG(D_READA, "free sa entry %.*s(%p) index %llu\n",
 227                entry->se_qstr.len, entry->se_qstr.name, entry,
 228                entry->se_index);
 229
 230         LASSERT(list_empty(&entry->se_list));
 231         LASSERT(sa_unhashed(entry));
 232
 233         OBD_FREE(entry, entry->se_size);
 234         atomic_dec(&sai->sai_cache_count);
 235 }
 236
 237 /*
 238  * find sa_entry by name, used by directory scanner, lock is not needed because
 239  * only scanner can remove the entry from cache.
 240  */
 241 static struct sa_entry *
 242 sa_get(struct ll_statahead_info *sai, const struct qstr *qstr)
 243 {
 244         struct sa_entry *entry;
 245         int i = sa_hash(qstr->hash);
 246
 247         list_for_each_entry(entry, &sai->sai_cache[i], se_hash) {
 248                 if (entry->se_qstr.hash == qstr->hash &&
 249                     entry->se_qstr.len == qstr->len &&
 250                     memcmp(entry->se_qstr.name, qstr->name, qstr->len) == 0)
 251                         return entry;
 252         }
 253         return NULL;
 254 }
 255
 256 /* unhash and unlink sa_entry, and then free it */
 257 static inline void
 258 sa_kill(struct ll_statahead_info *sai, struct sa_entry *entry)
 259 {
 260         struct ll_inode_info *lli = ll_i2info(sai->sai_dentry->d_inode);
 261
 262         LASSERT(!sa_unhashed(entry));
 263         LASSERT(!list_empty(&entry->se_list));
 264         LASSERT(sa_ready(entry));
 265
 266         sa_unhash(sai, entry);
 267
 268         spin_lock(&lli->lli_sa_lock);
 269         list_del_init(&entry->se_list);
 270         spin_unlock(&lli->lli_sa_lock);
 271
 272         if (entry->se_inode != NULL)
 273                 iput(entry->se_inode);
 274
 275         sa_free(sai, entry);
 276 }
 277
 278 /* called by scanner after use, sa_entry will be killed */
 279 static void
 280 sa_put(struct ll_statahead_info *sai, struct sa_entry *entry)
 281 {
 282         struct sa_entry *tmp, *next;
 283
 284         if (entry != NULL && entry->se_state == SA_ENTRY_SUCC) {
 285                 struct ll_sb_info *sbi = ll_i2sbi(sai->sai_dentry->d_inode);
 286
 287                 sai->sai_hit++;
 288                 sai->sai_consecutive_miss = 0;
 289                 sai->sai_max = min(2 * sai->sai_max, sbi->ll_sa_max);
 290         } else {
 291                 sai->sai_miss++;
 292                 sai->sai_consecutive_miss++;
 293         }
 294
 295         if (entry != NULL)
 296                 sa_kill(sai, entry);
 297
 298         /* kill old completed entries, only scanner process does this, no need
 299          * to lock */
 300         list_for_each_entry_safe(tmp, next, &sai->sai_entries, se_list) {
 301                 if (!is_omitted_entry(sai, tmp->se_index))
 302                         break;
 303                 sa_kill(sai, tmp);
 304         }
 305
 306         wake_up(&sai->sai_thread.t_ctl_waitq);
 307 }
 308
 309 /* update state and sort add entry to sai_entries by index, return true if
 310  * scanner is waiting on this entry. */
 311 static bool
 312 __sa_make_ready(struct ll_statahead_info *sai, struct sa_entry *entry, int ret)
 313 {
 314         struct sa_entry *se;
 315         struct list_head *pos = &sai->sai_entries;
 316         __u64 index = entry->se_index;
 317
 318         LASSERT(!sa_ready(entry));
 319         LASSERT(list_empty(&entry->se_list));
 320
 321         list_for_each_entry_reverse(se, &sai->sai_entries, se_list) {
 322                 if (se->se_index < entry->se_index) {
 323                         pos = &se->se_list;
 324                         break;
 325                 }
 326         }
 327         list_add(&entry->se_list, pos);
 328         entry->se_state = ret < 0 ? SA_ENTRY_INVA : SA_ENTRY_SUCC;
 329
 330         return (index == sai->sai_index_wait);
 331 }
 332
 333 /*
 334  * release resources used in async stat RPC, update entry state and wakeup if
 335  * scanner process it waiting on this entry.
 336  */
 337 static void
 338 sa_make_ready(struct ll_statahead_info *sai, struct sa_entry *entry, int ret)
 339 {
 340         struct ll_inode_info *lli = ll_i2info(sai->sai_dentry->d_inode);
 341         struct md_enqueue_info *minfo = entry->se_minfo;
 342         struct ptlrpc_request *req = entry->se_req;
 343         bool wakeup;
 344
 345         /* release resources used in RPC */
 346         if (minfo) {
 347                 entry->se_minfo = NULL;
 348                 ll_intent_release(&minfo->mi_it);
 349                 iput(minfo->mi_dir);
 350                 OBD_FREE_PTR(minfo);
 351         }
 352
 353         if (req) {
 354                 entry->se_req = NULL;
 355                 ptlrpc_req_finished(req);
 356         }
 357
 358         spin_lock(&lli->lli_sa_lock);
 359         wakeup = __sa_make_ready(sai, entry, ret);
 360         spin_unlock(&lli->lli_sa_lock);
 361
 362         if (wakeup)
 363                 wake_up(&sai->sai_waitq);
 364 }
 365
 366 /* insert inode into the list of sai_agls */
 367 static void ll_agl_add(struct ll_statahead_info *sai,
 368                        struct inode *inode, int index)
 369 {
 370         struct ll_inode_info *child  = ll_i2info(inode);
 371         struct ll_inode_info *parent = ll_i2info(sai->sai_dentry->d_inode);
 372         int                   added  = 0;
 373
 374         spin_lock(&child->lli_agl_lock);
 375         if (child->lli_agl_index == 0) {
 376                 child->lli_agl_index = index;
 377                 spin_unlock(&child->lli_agl_lock);
 378
 379                 LASSERT(list_empty(&child->lli_agl_list));
 380
 381                 igrab(inode);
 382                 spin_lock(&parent->lli_agl_lock);
 383                 if (agl_list_empty(sai))
 384                         added = 1;
 385                 list_add_tail(&child->lli_agl_list, &sai->sai_agls);
 386                 spin_unlock(&parent->lli_agl_lock);
 387         } else {
 388                 spin_unlock(&child->lli_agl_lock);
 389         }
 390
 391         if (added > 0)
 392                 wake_up(&sai->sai_agl_thread.t_ctl_waitq);
 393 }
 394
 395 /* allocate sai */
 396 static struct ll_statahead_info *ll_sai_alloc(struct dentry *dentry)
 397 {
 398         struct ll_statahead_info *sai;
 399         struct ll_inode_info *lli = ll_i2info(dentry->d_inode);
 400         int i;
 401         ENTRY;
 402
 403         OBD_ALLOC_PTR(sai);
 404         if (!sai)
 405                 RETURN(NULL);
 406
 407         sai->sai_dentry = dget(dentry);
 408         atomic_set(&sai->sai_refcount, 1);
 409         sai->sai_max = LL_SA_RPC_MIN;
 410         sai->sai_index = 1;
 411         init_waitqueue_head(&sai->sai_waitq);
 412         init_waitqueue_head(&sai->sai_thread.t_ctl_waitq);
 413         init_waitqueue_head(&sai->sai_agl_thread.t_ctl_waitq);
 414
 415         INIT_LIST_HEAD(&sai->sai_interim_entries);
 416         INIT_LIST_HEAD(&sai->sai_entries);
 417         INIT_LIST_HEAD(&sai->sai_agls);
 418
 419         for (i = 0; i < LL_SA_CACHE_SIZE; i++) {
 420                 INIT_LIST_HEAD(&sai->sai_cache[i]);
 421                 spin_lock_init(&sai->sai_cache_lock[i]);
 422         }
 423         atomic_set(&sai->sai_cache_count, 0);
 424
 425         spin_lock(&sai_generation_lock);
 426         lli->lli_sa_generation = ++sai_generation;
 427         if (unlikely(sai_generation == 0))
 428                 lli->lli_sa_generation = ++sai_generation;
 429         spin_unlock(&sai_generation_lock);
 430
 431         RETURN(sai);
 432 }
 433
 434 /* free sai */
 435 static inline void ll_sai_free(struct ll_statahead_info *sai)
 436 {
 437         LASSERT(sai->sai_dentry != NULL);
 438         dput(sai->sai_dentry);
 439         OBD_FREE_PTR(sai);
 440 }
 441
 442 /*
 443  * take refcount of sai if sai for @dir exists, which means statahead is on for
 444  * this directory.
 445  */
 446 static inline struct ll_statahead_info *ll_sai_get(struct inode *dir)
 447 {
 448         struct ll_inode_info *lli = ll_i2info(dir);
 449         struct ll_statahead_info *sai = NULL;
 450
 451         spin_lock(&lli->lli_sa_lock);
 452         sai = lli->lli_sai;
 453         if (sai != NULL)
 454                 atomic_inc(&sai->sai_refcount);
 455         spin_unlock(&lli->lli_sa_lock);
 456
 457         return sai;
 458 }
 459
 460 /*
 461  * put sai refcount after use, if refcount reaches zero, free sai and sa_entries
 462  * attached to it.
 463  */
 464 static void ll_sai_put(struct ll_statahead_info *sai)
 465 {
 466         struct ll_inode_info *lli = ll_i2info(sai->sai_dentry->d_inode);
 467
 468         if (atomic_dec_and_lock(&sai->sai_refcount, &lli->lli_sa_lock)) {
 469                 struct sa_entry *entry, *next;
 470                 struct ll_sb_info *sbi = ll_i2sbi(sai->sai_dentry->d_inode);
 471
 472                 lli->lli_sai = NULL;
 473                 spin_unlock(&lli->lli_sa_lock);
 474
 475                 LASSERT(thread_is_stopped(&sai->sai_thread));
 476                 LASSERT(thread_is_stopped(&sai->sai_agl_thread));
 477                 LASSERT(sai->sai_sent == sai->sai_replied);
 478                 LASSERT(!sa_has_callback(sai));
 479
 480                 list_for_each_entry_safe(entry, next, &sai->sai_entries,
 481                                          se_list)
 482                         sa_kill(sai, entry);
 483
 484                 LASSERT(atomic_read(&sai->sai_cache_count) == 0);
 485                 LASSERT(agl_list_empty(sai));
 486
 487                 ll_sai_free(sai);
 488                 atomic_dec(&sbi->ll_sa_running);
 489         }
 490 }
 491
 492 /* Do NOT forget to drop inode refcount when into sai_agls. */
 493 static void ll_agl_trigger(struct inode *inode, struct ll_statahead_info *sai)
 494 {
 495         struct ll_inode_info *lli = ll_i2info(inode);
 496         __u64 index = lli->lli_agl_index;
 497         int rc;
 498         ENTRY;
 499
 500         LASSERT(list_empty(&lli->lli_agl_list));
 501
 502         /* AGL maybe fall behind statahead with one entry */
 503         if (is_omitted_entry(sai, index + 1)) {
 504                 lli->lli_agl_index = 0;
 505                 iput(inode);
 506                 RETURN_EXIT;
 507         }
 508
 509         /* In case of restore, the MDT has the right size and has already
 510          * sent it back without granting the layout lock, inode is up-to-date.
 511          * Then AGL (async glimpse lock) is useless.
 512          * Also to glimpse we need the layout, in case of a runninh restore
 513          * the MDT holds the layout lock so the glimpse will block up to the
 514          * end of restore (statahead/agl will block) */
 515         if (ll_file_test_flag(lli, LLIF_FILE_RESTORING)) {
 516                 lli->lli_agl_index = 0;
 517                 iput(inode);
 518                 RETURN_EXIT;
 519         }
 520
 521         /* Someone is in glimpse (sync or async), do nothing. */
 522         rc = down_write_trylock(&lli->lli_glimpse_sem);
 523         if (rc == 0) {
 524                 lli->lli_agl_index = 0;
 525                 iput(inode);
 526                 RETURN_EXIT;
 527         }
 528
 529         /*
 530          * Someone triggered glimpse within 1 sec before.
 531          * 1) The former glimpse succeeded with glimpse lock granted by OST, and
 532          *    if the lock is still cached on client, AGL needs to do nothing. If
 533          *    it is cancelled by other client, AGL maybe cannot obtaion new lock
 534          *    for no glimpse callback triggered by AGL.
 535          * 2) The former glimpse succeeded, but OST did not grant glimpse lock.
 536          *    Under such case, it is quite possible that the OST will not grant
 537          *    glimpse lock for AGL also.
 538          * 3) The former glimpse failed, compared with other two cases, it is
 539          *    relative rare. AGL can ignore such case, and it will not muchly
 540          *    affect the performance.
 541          */
 542         if (lli->lli_glimpse_time != 0 &&
 543             cfs_time_before(cfs_time_shift(-1), lli->lli_glimpse_time)) {
 544                 up_write(&lli->lli_glimpse_sem);
 545                 lli->lli_agl_index = 0;
 546                 iput(inode);
 547                 RETURN_EXIT;
 548         }
 549
 550         CDEBUG(D_READA, "Handling (init) async glimpse: inode = "
 551                DFID", idx = %llu\n", PFID(&lli->lli_fid), index);
 552
 553         cl_agl(inode);
 554         lli->lli_agl_index = 0;
 555         lli->lli_glimpse_time = cfs_time_current();
 556         up_write(&lli->lli_glimpse_sem);
 557
 558         CDEBUG(D_READA, "Handled (init) async glimpse: inode= "
 559                DFID", idx = %llu, rc = %d\n",
 560                PFID(&lli->lli_fid), index, rc);
 561
 562         iput(inode);
 563
 564         EXIT;
 565 }
 566
 567 /*
 568  * prepare inode for sa entry, add it into agl list, now sa_entry is ready
 569  * to be used by scanner process.
 570  */
 571 static void sa_instantiate(struct ll_statahead_info *sai,
 572                                  struct sa_entry *entry)
 573 {
 574         struct inode *dir = sai->sai_dentry->d_inode;
 575         struct inode *child;
 576         struct md_enqueue_info *minfo;
 577         struct lookup_intent *it;
 578         struct ptlrpc_request *req;
 579         struct mdt_body *body;
 580         int rc = 0;
 581         ENTRY;
 582
 583         LASSERT(entry->se_handle != 0);
 584
 585         minfo = entry->se_minfo;
 586         it = &minfo->mi_it;
 587         req = entry->se_req;
 588         body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
 589         if (body == NULL)
 590                 GOTO(out, rc = -EFAULT);
 591
 592         child = entry->se_inode;
 593         if (child != NULL) {
 594                 /* revalidate; unlinked and re-created with the same name */
 595                 if (unlikely(!lu_fid_eq(&minfo->mi_data.op_fid2,
 596                                         &body->mbo_fid1))) {
 597                         entry->se_inode = NULL;
 598                         iput(child);
 599                         child = NULL;
 600                 }
 601         }
 602
 603         it->it_lock_handle = entry->se_handle;
 604         rc = md_revalidate_lock(ll_i2mdexp(dir), it, ll_inode2fid(dir), NULL);
 605         if (rc != 1)
 606                 GOTO(out, rc = -EAGAIN);
 607
 608         rc = ll_prep_inode(&child, req, dir->i_sb, it);
 609         if (rc)
 610                 GOTO(out, rc);
 611
 612         CDEBUG(D_READA, "%s: setting %.*s"DFID" l_data to inode %p\n",
 613                ll_get_fsname(child->i_sb, NULL, 0),
 614                entry->se_qstr.len, entry->se_qstr.name,
 615                PFID(ll_inode2fid(child)), child);
 616         ll_set_lock_data(ll_i2sbi(dir)->ll_md_exp, child, it, NULL);
 617
 618         entry->se_inode = child;
 619
 620         if (agl_should_run(sai, child))
 621                 ll_agl_add(sai, child, entry->se_index);
 622
 623         EXIT;
 624
 625 out:
 626         /* sa_make_ready() will drop ldlm ibits lock refcount by calling
 627          * ll_intent_drop_lock() in spite of failures. Do not worry about
 628          * calling ll_intent_drop_lock() more than once. */
 629         sa_make_ready(sai, entry, rc);
 630 }
 631
 632 /* once there are async stat replies, instantiate sa_entry from replies */
 633 static void sa_handle_callback(struct ll_statahead_info *sai)
 634 {
 635         struct ll_inode_info *lli;
 636
 637         lli = ll_i2info(sai->sai_dentry->d_inode);
 638
 639         while (sa_has_callback(sai)) {
 640                 struct sa_entry *entry;
 641
 642                 spin_lock(&lli->lli_sa_lock);
 643                 if (unlikely(!sa_has_callback(sai))) {
 644                         spin_unlock(&lli->lli_sa_lock);
 645                         break;
 646                 }
 647                 entry = list_entry(sai->sai_interim_entries.next,
 648                                    struct sa_entry, se_list);
 649                 list_del_init(&entry->se_list);
 650                 spin_unlock(&lli->lli_sa_lock);
 651
 652                 sa_instantiate(sai, entry);
 653         }
 654 }
 655
 656 /*
 657  * callback for async stat RPC, because this is called in ptlrpcd context, we
 658  * only put sa_entry in sai_interim_entries, and wake up statahead thread to
 659  * really prepare inode and instantiate sa_entry later.
 660  */
 661 static int ll_statahead_interpret(struct ptlrpc_request *req,
 662                                   struct md_enqueue_info *minfo, int rc)
 663 {
 664         struct lookup_intent *it = &minfo->mi_it;
 665         struct inode *dir = minfo->mi_dir;
 666         struct ll_inode_info *lli = ll_i2info(dir);
 667         struct ll_statahead_info *sai = lli->lli_sai;
 668         struct sa_entry *entry = (struct sa_entry *)minfo->mi_cbdata;
 669         __u64 handle = 0;
 670         wait_queue_head_t *waitq = NULL;
 671         ENTRY;
 672
 673         if (it_disposition(it, DISP_LOOKUP_NEG))
 674                 rc = -ENOENT;
 675
 676         /* because statahead thread will wait for all inflight RPC to finish,
 677          * sai should be always valid, no need to refcount */
 678         LASSERT(sai != NULL);
 679         LASSERT(!thread_is_stopped(&sai->sai_thread));
 680         LASSERT(entry != NULL);
 681
 682         CDEBUG(D_READA, "sa_entry %.*s rc %d\n",
 683                entry->se_qstr.len, entry->se_qstr.name, rc);
 684
 685         if (rc != 0) {
 686                 ll_intent_release(it);
 687                 iput(dir);
 688                 OBD_FREE_PTR(minfo);
 689         } else {
 690                 /* release ibits lock ASAP to avoid deadlock when statahead
 691                  * thread enqueues lock on parent in readdir and another
 692                  * process enqueues lock on child with parent lock held, eg.
 693                  * unlink. */
 694                 handle = it->it_lock_handle;
 695                 ll_intent_drop_lock(it);
 696         }
 697
 698         spin_lock(&lli->lli_sa_lock);
 699         if (rc != 0) {
 700                 if (__sa_make_ready(sai, entry, rc))
 701                         waitq = &sai->sai_waitq;
 702         } else {
 703                 entry->se_minfo = minfo;
 704                 entry->se_req = ptlrpc_request_addref(req);
 705                 /* Release the async ibits lock ASAP to avoid deadlock
 706                  * when statahead thread tries to enqueue lock on parent
 707                  * for readpage and other tries to enqueue lock on child
 708                  * with parent's lock held, for example: unlink. */
 709                 entry->se_handle = handle;
 710                 if (!sa_has_callback(sai))
 711                         waitq = &sai->sai_thread.t_ctl_waitq;
 712
 713                 list_add_tail(&entry->se_list, &sai->sai_interim_entries);
 714         }
 715         sai->sai_replied++;
 716         if (waitq != NULL)
 717                 wake_up(waitq);
 718         spin_unlock(&lli->lli_sa_lock);
 719
 720         RETURN(rc);
 721 }
 722
 723 /* finish async stat RPC arguments */
 724 static void sa_fini_data(struct md_enqueue_info *minfo)
 725 {
 726         iput(minfo->mi_dir);
 727         OBD_FREE_PTR(minfo);
 728 }
 729
 730 /*
 731  * prepare arguments for async stat RPC.
 732  */
 733 static struct md_enqueue_info *
 734 sa_prep_data(struct inode *dir, struct inode *child, struct sa_entry *entry)
 735 {
 736         struct md_enqueue_info   *minfo;
 737         struct ldlm_enqueue_info *einfo;
 738         struct md_op_data        *op_data;
 739
 740         OBD_ALLOC_PTR(minfo);
 741         if (minfo == NULL)
 742                 return ERR_PTR(-ENOMEM);
 743
 744         op_data = ll_prep_md_op_data(&minfo->mi_data, dir, child, NULL, 0, 0,
 745                                      LUSTRE_OPC_ANY, NULL);
 746         if (IS_ERR(op_data)) {
 747                 OBD_FREE_PTR(minfo);
 748                 return (struct md_enqueue_info *)op_data;
 749         }
 750
 751         if (child == NULL)
 752                 op_data->op_fid2 = entry->se_fid;
 753
 754         minfo->mi_it.it_op = IT_GETATTR;
 755         minfo->mi_dir = igrab(dir);
 756         minfo->mi_cb = ll_statahead_interpret;
 757         minfo->mi_cbdata = entry;
 758
 759         einfo = &minfo->mi_einfo;
 760         einfo->ei_type   = LDLM_IBITS;
 761         einfo->ei_mode   = it_to_lock_mode(&minfo->mi_it);
 762         einfo->ei_cb_bl  = ll_md_blocking_ast;
 763         einfo->ei_cb_cp  = ldlm_completion_ast;
 764         einfo->ei_cb_gl  = NULL;
 765         einfo->ei_cbdata = NULL;
 766
 767         return minfo;
 768 }
 769
 770 /* async stat for file not found in dcache */
 771 static int sa_lookup(struct inode *dir, struct sa_entry *entry)
 772 {
 773         struct md_enqueue_info   *minfo;
 774         int                       rc;
 775         ENTRY;
 776
 777         minfo = sa_prep_data(dir, NULL, entry);
 778         if (IS_ERR(minfo))
 779                 RETURN(PTR_ERR(minfo));
 780
 781         rc = md_intent_getattr_async(ll_i2mdexp(dir), minfo);
 782         if (rc < 0)
 783                 sa_fini_data(minfo);
 784
 785         RETURN(rc);
 786 }
 787
 788 /**
 789  * async stat for file found in dcache, similar to .revalidate
 790  *
 791  * \retval      1 dentry valid, no RPC sent
 792  * \retval      0 dentry invalid, will send async stat RPC
 793  * \retval      negative number upon error
 794  */
 795 static int sa_revalidate(struct inode *dir, struct sa_entry *entry,
 796                          struct dentry *dentry)
 797 {
 798         struct inode *inode = dentry->d_inode;
 799         struct lookup_intent it = { .it_op = IT_GETATTR,
 800                                     .it_lock_handle = 0 };
 801         struct md_enqueue_info *minfo;
 802         int rc;
 803         ENTRY;
 804
 805         if (unlikely(inode == NULL))
 806                 RETURN(1);
 807
 808         if (d_mountpoint(dentry))
 809                 RETURN(1);
 810
 811         entry->se_inode = igrab(inode);
 812         rc = md_revalidate_lock(ll_i2mdexp(dir), &it, ll_inode2fid(inode),
 813                                 NULL);
 814         if (rc == 1) {
 815                 entry->se_handle = it.it_lock_handle;
 816                 ll_intent_release(&it);
 817                 RETURN(1);
 818         }
 819
 820         minfo = sa_prep_data(dir, inode, entry);
 821         if (IS_ERR(minfo)) {
 822                 entry->se_inode = NULL;
 823                 iput(inode);
 824                 RETURN(PTR_ERR(minfo));
 825         }
 826
 827         rc = md_intent_getattr_async(ll_i2mdexp(dir), minfo);
 828         if (rc < 0) {
 829                 entry->se_inode = NULL;
 830                 iput(inode);
 831                 sa_fini_data(minfo);
 832         }
 833
 834         RETURN(rc);
 835 }
 836
 837 /* async stat for file with @name */
 838 static void sa_statahead(struct dentry *parent, const char *name, int len,
 839                          const struct lu_fid *fid)
 840 {
 841         struct inode *dir = parent->d_inode;
 842         struct ll_inode_info *lli = ll_i2info(dir);
 843         struct ll_statahead_info *sai = lli->lli_sai;
 844         struct dentry *dentry = NULL;
 845         struct sa_entry *entry;
 846         int rc;
 847         ENTRY;
 848
 849         entry = sa_alloc(sai, sai->sai_index, name, len, fid);
 850         if (IS_ERR(entry))
 851                 RETURN_EXIT;
 852
 853         dentry = d_lookup(parent, &entry->se_qstr);
 854         if (!dentry) {
 855                 rc = sa_lookup(dir, entry);
 856         } else {
 857                 rc = sa_revalidate(dir, entry, dentry);
 858                 if (rc == 1 && agl_should_run(sai, dentry->d_inode))
 859                         ll_agl_add(sai, dentry->d_inode, entry->se_index);
 860         }
 861
 862         if (dentry != NULL)
 863                 dput(dentry);
 864
 865         if (rc != 0)
 866                 sa_make_ready(sai, entry, rc);
 867         else
 868                 sai->sai_sent++;
 869
 870         sai->sai_index++;
 871
 872         EXIT;
 873 }
 874
 875 /* async glimpse (agl) thread main function */
 876 static int ll_agl_thread(void *arg)
 877 {
 878         struct dentry *parent = (struct dentry *)arg;
 879         struct inode *dir = parent->d_inode;
 880         struct ll_inode_info *plli = ll_i2info(dir);
 881         struct ll_inode_info *clli;
 882         struct ll_sb_info *sbi = ll_i2sbi(dir);
 883         struct ll_statahead_info *sai;
 884         struct ptlrpc_thread *thread;
 885         struct l_wait_info lwi = { 0 };
 886         ENTRY;
 887
 888
 889         sai = ll_sai_get(dir);
 890         thread = &sai->sai_agl_thread;
 891         thread->t_pid = current_pid();
 892         CDEBUG(D_READA, "agl thread started: sai %p, parent %.*s\n",
 893                sai, parent->d_name.len, parent->d_name.name);
 894
 895         atomic_inc(&sbi->ll_agl_total);
 896         spin_lock(&plli->lli_agl_lock);
 897         sai->sai_agl_valid = 1;
 898         if (thread_is_init(thread))
 899                 /* If someone else has changed the thread state
 900                  * (e.g. already changed to SVC_STOPPING), we can't just
 901                  * blindly overwrite that setting. */
 902                 thread_set_flags(thread, SVC_RUNNING);
 903         spin_unlock(&plli->lli_agl_lock);
 904         wake_up(&thread->t_ctl_waitq);
 905
 906         while (1) {
 907                 l_wait_event(thread->t_ctl_waitq,
 908                              !agl_list_empty(sai) ||
 909                              !thread_is_running(thread),
 910                              &lwi);
 911
 912                 if (!thread_is_running(thread))
 913                         break;
 914
 915                 spin_lock(&plli->lli_agl_lock);
 916                 /* The statahead thread maybe help to process AGL entries,
 917                  * so check whether list empty again. */
 918                 if (!agl_list_empty(sai)) {
 919                         clli = agl_first_entry(sai);
 920                         list_del_init(&clli->lli_agl_list);
 921                         spin_unlock(&plli->lli_agl_lock);
 922                         ll_agl_trigger(&clli->lli_vfs_inode, sai);
 923                 } else {
 924                         spin_unlock(&plli->lli_agl_lock);
 925                 }
 926         }
 927
 928         spin_lock(&plli->lli_agl_lock);
 929         sai->sai_agl_valid = 0;
 930         while (!agl_list_empty(sai)) {
 931                 clli = agl_first_entry(sai);
 932                 list_del_init(&clli->lli_agl_list);
 933                 spin_unlock(&plli->lli_agl_lock);
 934                 clli->lli_agl_index = 0;
 935                 iput(&clli->lli_vfs_inode);
 936                 spin_lock(&plli->lli_agl_lock);
 937         }
 938         thread_set_flags(thread, SVC_STOPPED);
 939         spin_unlock(&plli->lli_agl_lock);
 940         wake_up(&thread->t_ctl_waitq);
 941         ll_sai_put(sai);
 942         CDEBUG(D_READA, "agl thread stopped: sai %p, parent %.*s\n",
 943                sai, parent->d_name.len, parent->d_name.name);
 944         RETURN(0);
 945 }
 946
 947 /* start agl thread */
 948 static void ll_start_agl(struct dentry *parent, struct ll_statahead_info *sai)
 949 {
 950         struct ptlrpc_thread *thread = &sai->sai_agl_thread;
 951         struct l_wait_info    lwi    = { 0 };
 952         struct ll_inode_info  *plli;
 953         struct task_struct            *task;
 954         ENTRY;
 955
 956         CDEBUG(D_READA, "start agl thread: sai %p, parent %.*s\n",
 957                sai, parent->d_name.len, parent->d_name.name);
 958
 959         plli = ll_i2info(parent->d_inode);
 960         task = kthread_run(ll_agl_thread, parent,
 961                                "ll_agl_%u", plli->lli_opendir_pid);
 962         if (IS_ERR(task)) {
 963                 CERROR("can't start ll_agl thread, rc: %ld\n", PTR_ERR(task));
 964                 thread_set_flags(thread, SVC_STOPPED);
 965                 RETURN_EXIT;
 966         }
 967
 968         l_wait_event(thread->t_ctl_waitq,
 969                      thread_is_running(thread) || thread_is_stopped(thread),
 970                      &lwi);
 971         EXIT;
 972 }
 973
 974 /* statahead thread main function */
 975 static int ll_statahead_thread(void *arg)
 976 {
 977         struct dentry *parent = (struct dentry *)arg;
 978         struct inode *dir = parent->d_inode;
 979         struct ll_inode_info *lli = ll_i2info(dir);
 980         struct ll_sb_info *sbi = ll_i2sbi(dir);
 981         struct ll_statahead_info *sai;
 982         struct ptlrpc_thread *sa_thread;
 983         struct ptlrpc_thread *agl_thread;
 984         int first = 0;
 985         struct md_op_data *op_data;
 986         struct ll_dir_chain chain;
 987         struct l_wait_info lwi = { 0 };
 988         struct page *page = NULL;
 989         __u64 pos = 0;
 990         int rc = 0;
 991         ENTRY;
 992
 993         sai = ll_sai_get(dir);
 994         sa_thread = &sai->sai_thread;
 995         agl_thread = &sai->sai_agl_thread;
 996         sa_thread->t_pid = current_pid();
 997         CDEBUG(D_READA, "statahead thread starting: sai %p, parent %.*s\n",
 998                sai, parent->d_name.len, parent->d_name.name);
 999
1000         op_data = ll_prep_md_op_data(NULL, dir, dir, NULL, 0, 0,
1001                                      LUSTRE_OPC_ANY, dir);
1002         if (IS_ERR(op_data))
1003                 GOTO(out, rc = PTR_ERR(op_data));
1004
1005         if (sbi->ll_flags & LL_SBI_AGL_ENABLED)
1006                 ll_start_agl(parent, sai);
1007
1008         atomic_inc(&sbi->ll_sa_total);
1009         spin_lock(&lli->lli_sa_lock);
1010         if (thread_is_init(sa_thread))
1011                 /* If someone else has changed the thread state
1012                  * (e.g. already changed to SVC_STOPPING), we can't just
1013                  * blindly overwrite that setting. */
1014                 thread_set_flags(sa_thread, SVC_RUNNING);
1015         spin_unlock(&lli->lli_sa_lock);
1016         wake_up(&sa_thread->t_ctl_waitq);
1017
1018         ll_dir_chain_init(&chain);
1019         while (pos != MDS_DIR_END_OFF && thread_is_running(sa_thread)) {
1020                 struct lu_dirpage *dp;
1021                 struct lu_dirent  *ent;
1022
1023                 sai->sai_in_readpage = 1;
1024                 page = ll_get_dir_page(dir, op_data, pos, &chain);
1025                 sai->sai_in_readpage = 0;
1026                 if (IS_ERR(page)) {
1027                         rc = PTR_ERR(page);
1028                         CDEBUG(D_READA, "error reading dir "DFID" at %llu"
1029                                "/%llu opendir_pid = %u: rc = %d\n",
1030                                PFID(ll_inode2fid(dir)), pos, sai->sai_index,
1031                                lli->lli_opendir_pid, rc);
1032                         break;
1033                 }
1034
1035                 dp = page_address(page);
1036                 for (ent = lu_dirent_start(dp);
1037                      ent != NULL && thread_is_running(sa_thread) &&
1038                      !sa_low_hit(sai);
1039                      ent = lu_dirent_next(ent)) {
1040                         __u64 hash;
1041                         int namelen;
1042                         char *name;
1043                         struct lu_fid fid;
1044
1045                         hash = le64_to_cpu(ent->lde_hash);
1046                         if (unlikely(hash < pos))
1047                                 /*
1048                                  * Skip until we find target hash value.
1049                                  */
1050                                 continue;
1051
1052                         namelen = le16_to_cpu(ent->lde_namelen);
1053                         if (unlikely(namelen == 0))
1054                                 /*
1055                                  * Skip dummy record.
1056                                  */
1057                                 continue;
1058
1059                         name = ent->lde_name;
1060                         if (name[0] == '.') {
1061                                 if (namelen == 1) {
1062                                         /*
1063                                          * skip "."
1064                                          */
1065                                         continue;
1066                                 } else if (name[1] == '.' && namelen == 2) {
1067                                         /*
1068                                          * skip ".."
1069                                          */
1070                                         continue;
1071                                 } else if (!sai->sai_ls_all) {
1072                                         /*
1073                                          * skip hidden files.
1074                                          */
1075                                         sai->sai_skip_hidden++;
1076                                         continue;
1077                                 }
1078                         }
1079
1080                         /*
1081                          * don't stat-ahead first entry.
1082                          */
1083                         if (unlikely(++first == 1))
1084                                 continue;
1085
1086                         fid_le_to_cpu(&fid, &ent->lde_fid);
1087
1088                         /* wait for spare statahead window */
1089                         do {
1090                                 l_wait_event(sa_thread->t_ctl_waitq,
1091                                              !sa_sent_full(sai) ||
1092                                              sa_has_callback(sai) ||
1093                                              !agl_list_empty(sai) ||
1094                                              !thread_is_running(sa_thread),
1095                                              &lwi);
1096
1097                                 sa_handle_callback(sai);
1098
1099                                 spin_lock(&lli->lli_agl_lock);
1100                                 while (sa_sent_full(sai) &&
1101                                        !agl_list_empty(sai)) {
1102                                         struct ll_inode_info *clli;
1103
1104                                         clli = agl_first_entry(sai);
1105                                         list_del_init(&clli->lli_agl_list);
1106                                         spin_unlock(&lli->lli_agl_lock);
1107
1108                                         ll_agl_trigger(&clli->lli_vfs_inode,
1109                                                         sai);
1110
1111                                         spin_lock(&lli->lli_agl_lock);
1112                                 }
1113                                 spin_unlock(&lli->lli_agl_lock);
1114                         } while (sa_sent_full(sai) &&
1115                                  thread_is_running(sa_thread));
1116
1117                         sa_statahead(parent, name, namelen, &fid);
1118                 }
1119
1120                 pos = le64_to_cpu(dp->ldp_hash_end);
1121                 ll_release_page(dir, page,
1122                                 le32_to_cpu(dp->ldp_flags) & LDF_COLLIDE);
1123
1124                 if (sa_low_hit(sai)) {
1125                         rc = -EFAULT;
1126                         atomic_inc(&sbi->ll_sa_wrong);
1127                         CDEBUG(D_READA, "Statahead for dir "DFID" hit "
1128                                "ratio too low: hit/miss %llu/%llu"
1129                                ", sent/replied %llu/%llu, stopping "
1130                                "statahead thread: pid %d\n",
1131                                PFID(&lli->lli_fid), sai->sai_hit,
1132                                sai->sai_miss, sai->sai_sent,
1133                                sai->sai_replied, current_pid());
1134                         break;
1135                 }
1136         }
1137         ll_dir_chain_fini(&chain);
1138         ll_finish_md_op_data(op_data);
1139
1140         if (rc < 0) {
1141                 spin_lock(&lli->lli_sa_lock);
1142                 thread_set_flags(sa_thread, SVC_STOPPING);
1143                 lli->lli_sa_enabled = 0;
1144                 spin_unlock(&lli->lli_sa_lock);
1145         }
1146
1147         /* statahead is finished, but statahead entries need to be cached, wait
1148          * for file release to stop me. */
1149         while (thread_is_running(sa_thread)) {
1150                 l_wait_event(sa_thread->t_ctl_waitq,
1151                              sa_has_callback(sai) ||
1152                              !thread_is_running(sa_thread),
1153                              &lwi);
1154
1155                 sa_handle_callback(sai);
1156         }
1157
1158         EXIT;
1159 out:
1160         if (sai->sai_agl_valid) {
1161                 spin_lock(&lli->lli_agl_lock);
1162                 thread_set_flags(agl_thread, SVC_STOPPING);
1163                 spin_unlock(&lli->lli_agl_lock);
1164                 wake_up(&agl_thread->t_ctl_waitq);
1165
1166                 CDEBUG(D_READA, "stop agl thread: sai %p pid %u\n",
1167                        sai, (unsigned int)agl_thread->t_pid);
1168                 l_wait_event(agl_thread->t_ctl_waitq,
1169                              thread_is_stopped(agl_thread),
1170                              &lwi);
1171         } else {
1172                 /* Set agl_thread flags anyway. */
1173                 thread_set_flags(agl_thread, SVC_STOPPED);
1174         }
1175
1176         /* wait for inflight statahead RPCs to finish, and then we can free sai
1177          * safely because statahead RPC will access sai data */
1178         while (sai->sai_sent != sai->sai_replied) {
1179                 /* in case we're not woken up, timeout wait */
1180                 lwi = LWI_TIMEOUT(msecs_to_jiffies(MSEC_PER_SEC >> 3),
1181                                   NULL, NULL);
1182                 l_wait_event(sa_thread->t_ctl_waitq,
1183                         sai->sai_sent == sai->sai_replied, &lwi);
1184         }
1185
1186         /* release resources held by statahead RPCs */
1187         sa_handle_callback(sai);
1188
1189         spin_lock(&lli->lli_sa_lock);
1190         thread_set_flags(sa_thread, SVC_STOPPED);
1191         spin_unlock(&lli->lli_sa_lock);
1192
1193         CDEBUG(D_READA, "statahead thread stopped: sai %p, parent %.*s\n",
1194                sai, parent->d_name.len, parent->d_name.name);
1195
1196         wake_up(&sai->sai_waitq);
1197         wake_up(&sa_thread->t_ctl_waitq);
1198         ll_sai_put(sai);
1199
1200         return rc;
1201 }
1202
1203 /* authorize opened dir handle @key to statahead */
1204 void ll_authorize_statahead(struct inode *dir, void *key)
1205 {
1206         struct ll_inode_info *lli = ll_i2info(dir);
1207
1208         spin_lock(&lli->lli_sa_lock);
1209         if (lli->lli_opendir_key == NULL && lli->lli_sai == NULL) {
1210                 /*
1211                  * if lli_sai is not NULL, it means previous statahead is not
1212                  * finished yet, we'd better not start a new statahead for now.
1213                  */
1214                 LASSERT(lli->lli_opendir_pid == 0);
1215                 lli->lli_opendir_key = key;
1216                 lli->lli_opendir_pid = current_pid();
1217                 lli->lli_sa_enabled = 1;
1218         }
1219         spin_unlock(&lli->lli_sa_lock);
1220 }
1221
1222 /*
1223  * deauthorize opened dir handle @key to statahead, and notify statahead thread
1224  * to quit if it's running.
1225  */
1226 void ll_deauthorize_statahead(struct inode *dir, void *key)
1227 {
1228         struct ll_inode_info *lli = ll_i2info(dir);
1229         struct ll_statahead_info *sai;
1230
1231         LASSERT(lli->lli_opendir_key == key);
1232         LASSERT(lli->lli_opendir_pid != 0);
1233
1234         CDEBUG(D_READA, "deauthorize statahead for "DFID"\n",
1235                 PFID(&lli->lli_fid));
1236
1237         spin_lock(&lli->lli_sa_lock);
1238         lli->lli_opendir_key = NULL;
1239         lli->lli_opendir_pid = 0;
1240         lli->lli_sa_enabled = 0;
1241         sai = lli->lli_sai;
1242         if (sai != NULL && thread_is_running(&sai->sai_thread)) {
1243                 /*
1244                  * statahead thread may not quit yet because it needs to cache
1245                  * entries, now it's time to tell it to quit.
1246                  *
1247                  * In case sai is released, wake_up() is called inside spinlock,
1248                  * so we have to call smp_mb() explicitely to serialize ops.
1249                  */
1250                 thread_set_flags(&sai->sai_thread, SVC_STOPPING);
1251                 smp_mb();
1252                 wake_up(&sai->sai_thread.t_ctl_waitq);
1253         }
1254         spin_unlock(&lli->lli_sa_lock);
1255 }
1256
1257 enum {
1258         /**
1259          * not first dirent, or is "."
1260          */
1261         LS_NOT_FIRST_DE = 0,
1262         /**
1263          * the first non-hidden dirent
1264          */
1265         LS_FIRST_DE,
1266         /**
1267          * the first hidden dirent, that is "."
1268          */
1269         LS_FIRST_DOT_DE
1270 };
1271
1272 /* file is first dirent under @dir */
1273 static int is_first_dirent(struct inode *dir, struct dentry *dentry)
1274 {
1275         struct ll_dir_chain   chain;
1276         struct qstr          *target = &dentry->d_name;
1277         struct md_op_data    *op_data;
1278         int                   dot_de;
1279         struct page          *page = NULL;
1280         int                   rc = LS_NOT_FIRST_DE;
1281         __u64                 pos = 0;
1282         ENTRY;
1283
1284         op_data = ll_prep_md_op_data(NULL, dir, dir, NULL, 0, 0,
1285                                      LUSTRE_OPC_ANY, dir);
1286         if (IS_ERR(op_data))
1287                 RETURN(PTR_ERR(op_data));
1288         /**
1289          *FIXME choose the start offset of the readdir
1290          */
1291
1292         ll_dir_chain_init(&chain);
1293         page = ll_get_dir_page(dir, op_data, 0, &chain);
1294
1295         while (1) {
1296                 struct lu_dirpage *dp;
1297                 struct lu_dirent  *ent;
1298
1299                 if (IS_ERR(page)) {
1300                         struct ll_inode_info *lli = ll_i2info(dir);
1301
1302                         rc = PTR_ERR(page);
1303                         CERROR("%s: reading dir "DFID" at %llu"
1304                                "opendir_pid = %u : rc = %d\n",
1305                                ll_get_fsname(dir->i_sb, NULL, 0),
1306                                PFID(ll_inode2fid(dir)), pos,
1307                                lli->lli_opendir_pid, rc);
1308                         break;
1309                 }
1310
1311                 dp = page_address(page);
1312                 for (ent = lu_dirent_start(dp); ent != NULL;
1313                      ent = lu_dirent_next(ent)) {
1314                         __u64 hash;
1315                         int namelen;
1316                         char *name;
1317
1318                         hash = le64_to_cpu(ent->lde_hash);
1319                         /* The ll_get_dir_page() can return any page containing
1320                          * the given hash which may be not the start hash. */
1321                         if (unlikely(hash < pos))
1322                                 continue;
1323
1324                         namelen = le16_to_cpu(ent->lde_namelen);
1325                         if (unlikely(namelen == 0))
1326                                 /*
1327                                  * skip dummy record.
1328                                  */
1329                                 continue;
1330
1331                         name = ent->lde_name;
1332                         if (name[0] == '.') {
1333                                 if (namelen == 1)
1334                                         /*
1335                                          * skip "."
1336                                          */
1337                                         continue;
1338                                 else if (name[1] == '.' && namelen == 2)
1339                                         /*
1340                                          * skip ".."
1341                                          */
1342                                         continue;
1343                                 else
1344                                         dot_de = 1;
1345                         } else {
1346                                 dot_de = 0;
1347                         }
1348
1349                         if (dot_de && target->name[0] != '.') {
1350                                 CDEBUG(D_READA, "%.*s skip hidden file %.*s\n",
1351                                        target->len, target->name,
1352                                        namelen, name);
1353                                 continue;
1354                         }
1355
1356                         if (target->len != namelen ||
1357                             memcmp(target->name, name, namelen) != 0)
1358                                 rc = LS_NOT_FIRST_DE;
1359                         else if (!dot_de)
1360                                 rc = LS_FIRST_DE;
1361                         else
1362                                 rc = LS_FIRST_DOT_DE;
1363
1364                         ll_release_page(dir, page, false);
1365                         GOTO(out, rc);
1366                 }
1367                 pos = le64_to_cpu(dp->ldp_hash_end);
1368                 if (pos == MDS_DIR_END_OFF) {
1369                         /*
1370                          * End of directory reached.
1371                          */
1372                         ll_release_page(dir, page, false);
1373                         GOTO(out, rc);
1374                 } else {
1375                         /*
1376                          * chain is exhausted
1377                          * Normal case: continue to the next page.
1378                          */
1379                         ll_release_page(dir, page, le32_to_cpu(dp->ldp_flags) &
1380                                               LDF_COLLIDE);
1381                         page = ll_get_dir_page(dir, op_data, pos, &chain);
1382                 }
1383         }
1384         EXIT;
1385 out:
1386         ll_dir_chain_fini(&chain);
1387         ll_finish_md_op_data(op_data);
1388         return rc;
1389 }
1390
1391 /**
1392  * revalidate @dentryp from statahead cache
1393  *
1394  * \param[in] dir       parent directory
1395  * \param[in] sai       sai structure
1396  * \param[out] dentryp  pointer to dentry which will be revalidated
1397  * \param[in] unplug    unplug statahead window only (normally for negative
1398  *                      dentry)
1399  * \retval              1 on success, dentry is saved in @dentryp
1400  * \retval              0 if revalidation failed (no proper lock on client)
1401  * \retval              negative number upon error
1402  */
1403 static int revalidate_statahead_dentry(struct inode *dir,
1404                                         struct ll_statahead_info *sai,
1405                                         struct dentry **dentryp,
1406                                         bool unplug)
1407 {
1408         struct sa_entry *entry = NULL;
1409         struct l_wait_info lwi = { 0 };
1410         struct ll_dentry_data *ldd;
1411         struct ll_inode_info *lli = ll_i2info(dir);
1412         int rc = 0;
1413         ENTRY;
1414
1415         if ((*dentryp)->d_name.name[0] == '.') {
1416                 if (sai->sai_ls_all ||
1417                     sai->sai_miss_hidden >= sai->sai_skip_hidden) {
1418                         /*
1419                          * Hidden dentry is the first one, or statahead
1420                          * thread does not skip so many hidden dentries
1421                          * before "sai_ls_all" enabled as below.
1422                          */
1423                 } else {
1424                         if (!sai->sai_ls_all)
1425                                 /*
1426                                  * It maybe because hidden dentry is not
1427                                  * the first one, "sai_ls_all" was not
1428                                  * set, then "ls -al" missed. Enable
1429                                  * "sai_ls_all" for such case.
1430                                  */
1431                                 sai->sai_ls_all = 1;
1432
1433                         /*
1434                          * Such "getattr" has been skipped before
1435                          * "sai_ls_all" enabled as above.
1436                          */
1437                         sai->sai_miss_hidden++;
1438                         RETURN(-EAGAIN);
1439                 }
1440         }
1441
1442         if (unplug)
1443                 GOTO(out, rc = 1);
1444
1445         entry = sa_get(sai, &(*dentryp)->d_name);
1446         if (entry == NULL)
1447                 GOTO(out, rc = -EAGAIN);
1448
1449         /* if statahead is busy in readdir, help it do post-work */
1450         if (!sa_ready(entry) && sai->sai_in_readpage)
1451                 sa_handle_callback(sai);
1452
1453         if (!sa_ready(entry)) {
1454                 spin_lock(&lli->lli_sa_lock);
1455                 sai->sai_index_wait = entry->se_index;
1456                 spin_unlock(&lli->lli_sa_lock);
1457                 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(30), NULL,
1458                                        LWI_ON_SIGNAL_NOOP, NULL);
1459                 rc = l_wait_event(sai->sai_waitq, sa_ready(entry), &lwi);
1460                 if (rc < 0) {
1461                         /*
1462                          * entry may not be ready, so it may be used by inflight
1463                          * statahead RPC, don't free it.
1464                          */
1465                         entry = NULL;
1466                         GOTO(out, rc = -EAGAIN);
1467                 }
1468         }
1469
1470         if (entry->se_state == SA_ENTRY_SUCC && entry->se_inode != NULL) {
1471                 struct inode *inode = entry->se_inode;
1472                 struct lookup_intent it = { .it_op = IT_GETATTR,
1473                                             .it_lock_handle =
1474                                                 entry->se_handle };
1475                 __u64 bits;
1476
1477                 rc = md_revalidate_lock(ll_i2mdexp(dir), &it,
1478                                         ll_inode2fid(inode), &bits);
1479                 if (rc == 1) {
1480                         if ((*dentryp)->d_inode == NULL) {
1481                                 struct dentry *alias;
1482
1483                                 alias = ll_splice_alias(inode, *dentryp);
1484                                 if (IS_ERR(alias)) {
1485                                         ll_intent_release(&it);
1486                                         GOTO(out, rc = PTR_ERR(alias));
1487                                 }
1488                                 *dentryp = alias;
1489                                 /* statahead prepared this inode, transfer inode
1490                                  * refcount from sa_entry to dentry */
1491                                 entry->se_inode = NULL;
1492                         } else if ((*dentryp)->d_inode != inode) {
1493                                 /* revalidate, but inode is recreated */
1494                                 CDEBUG(D_READA,
1495                                         "%s: stale dentry %.*s inode "
1496                                         DFID", statahead inode "DFID
1497                                         "\n",
1498                                         ll_get_fsname((*dentryp)->d_inode->i_sb,
1499                                                       NULL, 0),
1500                                         (*dentryp)->d_name.len,
1501                                         (*dentryp)->d_name.name,
1502                                         PFID(ll_inode2fid((*dentryp)->d_inode)),
1503                                         PFID(ll_inode2fid(inode)));
1504                                 ll_intent_release(&it);
1505                                 GOTO(out, rc = -ESTALE);
1506                         }
1507
1508                         if ((bits & MDS_INODELOCK_LOOKUP) &&
1509                             d_lustre_invalid(*dentryp))
1510                                 d_lustre_revalidate(*dentryp);
1511                         ll_intent_release(&it);
1512                 }
1513         }
1514 out:
1515         /*
1516          * statahead cached sa_entry can be used only once, and will be killed
1517          * right after use, so if lookup/revalidate accessed statahead cache,
1518          * set dentry ldd_sa_generation to parent lli_sa_generation, later if we
1519          * stat this file again, we know we've done statahead before, see
1520          * dentry_may_statahead().
1521          */
1522         ldd = ll_d2d(*dentryp);
1523         /* ldd can be NULL if llite lookup failed. */
1524         if (ldd != NULL)
1525                 ldd->lld_sa_generation = lli->lli_sa_generation;
1526         sa_put(sai, entry);
1527
1528         RETURN(rc);
1529 }
1530
1531 /**
1532  * start statahead thread
1533  *
1534  * \param[in] dir       parent directory
1535  * \param[in] dentry    dentry that triggers statahead, normally the first
1536  *                      dirent under @dir
1537  * \retval              -EAGAIN on success, because when this function is
1538  *                      called, it's already in lookup call, so client should
1539  *                      do it itself instead of waiting for statahead thread
1540  *                      to do it asynchronously.
1541  * \retval              negative number upon error
1542  */
1543 static int start_statahead_thread(struct inode *dir, struct dentry *dentry)
1544 {
1545         struct ll_inode_info *lli = ll_i2info(dir);
1546         struct ll_statahead_info *sai = NULL;
1547         struct dentry *parent = dentry->d_parent;
1548         struct ptlrpc_thread *thread;
1549         struct l_wait_info lwi = { 0 };
1550         struct task_struct *task;
1551         int rc;
1552         ENTRY;
1553
1554         /* I am the "lli_opendir_pid" owner, only me can set "lli_sai". */
1555         rc = is_first_dirent(dir, dentry);
1556         if (rc == LS_NOT_FIRST_DE)
1557                 /* It is not "ls -{a}l" operation, no need statahead for it. */
1558                 GOTO(out, rc = -EFAULT);
1559
1560         sai = ll_sai_alloc(parent);
1561         if (sai == NULL)
1562                 GOTO(out, rc = -ENOMEM);
1563
1564         sai->sai_ls_all = (rc == LS_FIRST_DOT_DE);
1565
1566         /* if current lli_opendir_key was deauthorized, or dir re-opened by
1567          * another process, don't start statahead, otherwise the newly spawned
1568          * statahead thread won't be notified to quit. */
1569         spin_lock(&lli->lli_sa_lock);
1570         if (unlikely(lli->lli_sai != NULL ||
1571                      lli->lli_opendir_key == NULL ||
1572                      lli->lli_opendir_pid != current->pid)) {
1573                 spin_unlock(&lli->lli_sa_lock);
1574                 GOTO(out, rc = -EPERM);
1575         }
1576         lli->lli_sai = sai;
1577         spin_unlock(&lli->lli_sa_lock);
1578
1579         atomic_inc(&ll_i2sbi(parent->d_inode)->ll_sa_running);
1580
1581         CDEBUG(D_READA, "start statahead thread: [pid %d] [parent %.*s]\n",
1582                current_pid(), parent->d_name.len, parent->d_name.name);
1583
1584         task = kthread_run(ll_statahead_thread, parent, "ll_sa_%u",
1585                            lli->lli_opendir_pid);
1586         thread = &sai->sai_thread;
1587         if (IS_ERR(task)) {
1588                 rc = PTR_ERR(task);
1589                 CERROR("can't start ll_sa thread, rc: %d\n", rc);
1590                 GOTO(out, rc);
1591         }
1592
1593         l_wait_event(thread->t_ctl_waitq,
1594                      thread_is_running(thread) || thread_is_stopped(thread),
1595                      &lwi);
1596         ll_sai_put(sai);
1597
1598         /*
1599          * We don't stat-ahead for the first dirent since we are already in
1600          * lookup.
1601          */
1602         RETURN(-EAGAIN);
1603
1604 out:
1605         /* once we start statahead thread failed, disable statahead so that
1606          * subsequent stat won't waste time to try it. */
1607         spin_lock(&lli->lli_sa_lock);
1608         lli->lli_sa_enabled = 0;
1609         lli->lli_sai = NULL;
1610         spin_unlock(&lli->lli_sa_lock);
1611
1612         if (sai != NULL)
1613                 ll_sai_free(sai);
1614
1615         RETURN(rc);
1616 }
1617
1618 /**
1619  * statahead entry function, this is called when client getattr on a file, it
1620  * will start statahead thread if this is the first dir entry, else revalidate
1621  * dentry from statahead cache.
1622  *
1623  * \param[in]  dir      parent directory
1624  * \param[out] dentryp  dentry to getattr
1625  * \param[in]  unplug   unplug statahead window only (normally for negative
1626  *                      dentry)
1627  * \retval              1 on success
1628  * \retval              0 revalidation from statahead cache failed, caller needs
1629  *                      to getattr from server directly
1630  * \retval              negative number on error, caller often ignores this and
1631  *                      then getattr from server
1632  */
1633 int ll_statahead(struct inode *dir, struct dentry **dentryp, bool unplug)
1634 {
1635         struct ll_statahead_info *sai;
1636
1637         sai = ll_sai_get(dir);
1638         if (sai != NULL) {
1639                 int rc;
1640
1641                 rc = revalidate_statahead_dentry(dir, sai, dentryp, unplug);
1642                 CDEBUG(D_READA, "revalidate statahead %.*s: %d.\n",
1643                         (*dentryp)->d_name.len, (*dentryp)->d_name.name, rc);
1644                 ll_sai_put(sai);
1645                 return rc;
1646         }
1647         return start_statahead_thread(dir, *dentryp);
1648 }