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, 2017, 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 dentry *parent, struct ll_statahead_info *sai, __u64 index,
 183          const char *name, int len, 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 = ll_full_name_hash(parent, 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         /*
 329          * LU-9210: ll_statahead_interpet must be able to see this before
 330          * we wake it up
 331          */
 332         smp_store_release(&entry->se_state, ret < 0 ? SA_ENTRY_INVA : SA_ENTRY_SUCC);
 333
 334         return (index == sai->sai_index_wait);
 335 }
 336
 337 /* finish async stat RPC arguments */
 338 static void sa_fini_data(struct md_enqueue_info *minfo)
 339 {
 340         ll_unlock_md_op_lsm(&minfo->mi_data);
 341         iput(minfo->mi_dir);
 342         OBD_FREE_PTR(minfo);
 343 }
 344
 345 static int ll_statahead_interpret(struct ptlrpc_request *req,
 346                                   struct md_enqueue_info *minfo, int rc);
 347
 348 /*
 349  * prepare arguments for async stat RPC.
 350  */
 351 static struct md_enqueue_info *
 352 sa_prep_data(struct inode *dir, struct inode *child, struct sa_entry *entry)
 353 {
 354         struct md_enqueue_info   *minfo;
 355         struct ldlm_enqueue_info *einfo;
 356         struct md_op_data        *op_data;
 357
 358         OBD_ALLOC_PTR(minfo);
 359         if (minfo == NULL)
 360                 return ERR_PTR(-ENOMEM);
 361
 362         op_data = ll_prep_md_op_data(&minfo->mi_data, dir, child,
 363                                      entry->se_qstr.name, entry->se_qstr.len, 0,
 364                                      LUSTRE_OPC_ANY, NULL);
 365         if (IS_ERR(op_data)) {
 366                 OBD_FREE_PTR(minfo);
 367                 return (struct md_enqueue_info *)op_data;
 368         }
 369
 370         if (child == NULL)
 371                 op_data->op_fid2 = entry->se_fid;
 372
 373         minfo->mi_it.it_op = IT_GETATTR;
 374         minfo->mi_dir = igrab(dir);
 375         minfo->mi_cb = ll_statahead_interpret;
 376         minfo->mi_cbdata = entry;
 377
 378         einfo = &minfo->mi_einfo;
 379         einfo->ei_type   = LDLM_IBITS;
 380         einfo->ei_mode   = it_to_lock_mode(&minfo->mi_it);
 381         einfo->ei_cb_bl  = ll_md_blocking_ast;
 382         einfo->ei_cb_cp  = ldlm_completion_ast;
 383         einfo->ei_cb_gl  = NULL;
 384         einfo->ei_cbdata = NULL;
 385
 386         return minfo;
 387 }
 388
 389 /*
 390  * release resources used in async stat RPC, update entry state and wakeup if
 391  * scanner process it waiting on this entry.
 392  */
 393 static void
 394 sa_make_ready(struct ll_statahead_info *sai, struct sa_entry *entry, int ret)
 395 {
 396         struct ll_inode_info *lli = ll_i2info(sai->sai_dentry->d_inode);
 397         struct md_enqueue_info *minfo = entry->se_minfo;
 398         struct ptlrpc_request *req = entry->se_req;
 399         bool wakeup;
 400
 401         /* release resources used in RPC */
 402         if (minfo) {
 403                 entry->se_minfo = NULL;
 404                 ll_intent_release(&minfo->mi_it);
 405                 sa_fini_data(minfo);
 406         }
 407
 408         if (req) {
 409                 entry->se_req = NULL;
 410                 ptlrpc_req_finished(req);
 411         }
 412
 413         spin_lock(&lli->lli_sa_lock);
 414         wakeup = __sa_make_ready(sai, entry, ret);
 415         spin_unlock(&lli->lli_sa_lock);
 416
 417         if (wakeup)
 418                 wake_up(&sai->sai_waitq);
 419 }
 420
 421 /* insert inode into the list of sai_agls */
 422 static void ll_agl_add(struct ll_statahead_info *sai,
 423                        struct inode *inode, int index)
 424 {
 425         struct ll_inode_info *child  = ll_i2info(inode);
 426         struct ll_inode_info *parent = ll_i2info(sai->sai_dentry->d_inode);
 427         int                   added  = 0;
 428
 429         spin_lock(&child->lli_agl_lock);
 430         if (child->lli_agl_index == 0) {
 431                 child->lli_agl_index = index;
 432                 spin_unlock(&child->lli_agl_lock);
 433
 434                 LASSERT(list_empty(&child->lli_agl_list));
 435
 436                 igrab(inode);
 437                 spin_lock(&parent->lli_agl_lock);
 438                 if (agl_list_empty(sai))
 439                         added = 1;
 440                 list_add_tail(&child->lli_agl_list, &sai->sai_agls);
 441                 spin_unlock(&parent->lli_agl_lock);
 442         } else {
 443                 spin_unlock(&child->lli_agl_lock);
 444         }
 445
 446         if (added > 0)
 447                 wake_up(&sai->sai_agl_thread.t_ctl_waitq);
 448 }
 449
 450 /* allocate sai */
 451 static struct ll_statahead_info *ll_sai_alloc(struct dentry *dentry)
 452 {
 453         struct ll_statahead_info *sai;
 454         struct ll_inode_info *lli = ll_i2info(dentry->d_inode);
 455         int i;
 456         ENTRY;
 457
 458         OBD_ALLOC_PTR(sai);
 459         if (!sai)
 460                 RETURN(NULL);
 461
 462         sai->sai_dentry = dget(dentry);
 463         atomic_set(&sai->sai_refcount, 1);
 464         sai->sai_max = LL_SA_RPC_MIN;
 465         sai->sai_index = 1;
 466         init_waitqueue_head(&sai->sai_waitq);
 467         init_waitqueue_head(&sai->sai_thread.t_ctl_waitq);
 468         init_waitqueue_head(&sai->sai_agl_thread.t_ctl_waitq);
 469
 470         INIT_LIST_HEAD(&sai->sai_interim_entries);
 471         INIT_LIST_HEAD(&sai->sai_entries);
 472         INIT_LIST_HEAD(&sai->sai_agls);
 473
 474         for (i = 0; i < LL_SA_CACHE_SIZE; i++) {
 475                 INIT_LIST_HEAD(&sai->sai_cache[i]);
 476                 spin_lock_init(&sai->sai_cache_lock[i]);
 477         }
 478         atomic_set(&sai->sai_cache_count, 0);
 479
 480         spin_lock(&sai_generation_lock);
 481         lli->lli_sa_generation = ++sai_generation;
 482         if (unlikely(sai_generation == 0))
 483                 lli->lli_sa_generation = ++sai_generation;
 484         spin_unlock(&sai_generation_lock);
 485
 486         RETURN(sai);
 487 }
 488
 489 /* free sai */
 490 static inline void ll_sai_free(struct ll_statahead_info *sai)
 491 {
 492         LASSERT(sai->sai_dentry != NULL);
 493         dput(sai->sai_dentry);
 494         OBD_FREE_PTR(sai);
 495 }
 496
 497 /*
 498  * take refcount of sai if sai for @dir exists, which means statahead is on for
 499  * this directory.
 500  */
 501 static inline struct ll_statahead_info *ll_sai_get(struct inode *dir)
 502 {
 503         struct ll_inode_info *lli = ll_i2info(dir);
 504         struct ll_statahead_info *sai = NULL;
 505
 506         spin_lock(&lli->lli_sa_lock);
 507         sai = lli->lli_sai;
 508         if (sai != NULL)
 509                 atomic_inc(&sai->sai_refcount);
 510         spin_unlock(&lli->lli_sa_lock);
 511
 512         return sai;
 513 }
 514
 515 /*
 516  * put sai refcount after use, if refcount reaches zero, free sai and sa_entries
 517  * attached to it.
 518  */
 519 static void ll_sai_put(struct ll_statahead_info *sai)
 520 {
 521         struct ll_inode_info *lli = ll_i2info(sai->sai_dentry->d_inode);
 522
 523         if (atomic_dec_and_lock(&sai->sai_refcount, &lli->lli_sa_lock)) {
 524                 struct sa_entry *entry, *next;
 525                 struct ll_sb_info *sbi = ll_i2sbi(sai->sai_dentry->d_inode);
 526
 527                 lli->lli_sai = NULL;
 528                 spin_unlock(&lli->lli_sa_lock);
 529
 530                 LASSERT(thread_is_stopped(&sai->sai_thread));
 531                 LASSERT(thread_is_stopped(&sai->sai_agl_thread));
 532                 LASSERT(sai->sai_sent == sai->sai_replied);
 533                 LASSERT(!sa_has_callback(sai));
 534
 535                 list_for_each_entry_safe(entry, next, &sai->sai_entries,
 536                                          se_list)
 537                         sa_kill(sai, entry);
 538
 539                 LASSERT(atomic_read(&sai->sai_cache_count) == 0);
 540                 LASSERT(agl_list_empty(sai));
 541
 542                 ll_sai_free(sai);
 543                 atomic_dec(&sbi->ll_sa_running);
 544         }
 545 }
 546
 547 /* Do NOT forget to drop inode refcount when into sai_agls. */
 548 static void ll_agl_trigger(struct inode *inode, struct ll_statahead_info *sai)
 549 {
 550         struct ll_inode_info *lli = ll_i2info(inode);
 551         u64 index = lli->lli_agl_index;
 552         ktime_t expire;
 553         int rc;
 554
 555         ENTRY;
 556         LASSERT(list_empty(&lli->lli_agl_list));
 557
 558         /* AGL maybe fall behind statahead with one entry */
 559         if (is_omitted_entry(sai, index + 1)) {
 560                 lli->lli_agl_index = 0;
 561                 iput(inode);
 562                 RETURN_EXIT;
 563         }
 564
 565         /* In case of restore, the MDT has the right size and has already
 566          * sent it back without granting the layout lock, inode is up-to-date.
 567          * Then AGL (async glimpse lock) is useless.
 568          * Also to glimpse we need the layout, in case of a runninh restore
 569          * the MDT holds the layout lock so the glimpse will block up to the
 570          * end of restore (statahead/agl will block) */
 571         if (ll_file_test_flag(lli, LLIF_FILE_RESTORING)) {
 572                 lli->lli_agl_index = 0;
 573                 iput(inode);
 574                 RETURN_EXIT;
 575         }
 576
 577         /* Someone is in glimpse (sync or async), do nothing. */
 578         rc = down_write_trylock(&lli->lli_glimpse_sem);
 579         if (rc == 0) {
 580                 lli->lli_agl_index = 0;
 581                 iput(inode);
 582                 RETURN_EXIT;
 583         }
 584
 585         /*
 586          * Someone triggered glimpse within 1 sec before.
 587          * 1) The former glimpse succeeded with glimpse lock granted by OST, and
 588          *    if the lock is still cached on client, AGL needs to do nothing. If
 589          *    it is cancelled by other client, AGL maybe cannot obtaion new lock
 590          *    for no glimpse callback triggered by AGL.
 591          * 2) The former glimpse succeeded, but OST did not grant glimpse lock.
 592          *    Under such case, it is quite possible that the OST will not grant
 593          *    glimpse lock for AGL also.
 594          * 3) The former glimpse failed, compared with other two cases, it is
 595          *    relative rare. AGL can ignore such case, and it will not muchly
 596          *    affect the performance.
 597          */
 598         expire = ktime_sub_ns(ktime_get(), NSEC_PER_SEC);
 599         if (ktime_to_ns(lli->lli_glimpse_time) &&
 600             ktime_before(expire, lli->lli_glimpse_time)) {
 601                 up_write(&lli->lli_glimpse_sem);
 602                 lli->lli_agl_index = 0;
 603                 iput(inode);
 604                 RETURN_EXIT;
 605         }
 606
 607         CDEBUG(D_READA, "Handling (init) async glimpse: inode = "
 608                DFID", idx = %llu\n", PFID(&lli->lli_fid), index);
 609
 610         cl_agl(inode);
 611         lli->lli_agl_index = 0;
 612         lli->lli_glimpse_time = ktime_get();
 613         up_write(&lli->lli_glimpse_sem);
 614
 615         CDEBUG(D_READA, "Handled (init) async glimpse: inode= "
 616                DFID", idx = %llu, rc = %d\n",
 617                PFID(&lli->lli_fid), index, rc);
 618
 619         iput(inode);
 620
 621         EXIT;
 622 }
 623
 624 /*
 625  * prepare inode for sa entry, add it into agl list, now sa_entry is ready
 626  * to be used by scanner process.
 627  */
 628 static void sa_instantiate(struct ll_statahead_info *sai,
 629                                  struct sa_entry *entry)
 630 {
 631         struct inode *dir = sai->sai_dentry->d_inode;
 632         struct inode *child;
 633         struct md_enqueue_info *minfo;
 634         struct lookup_intent *it;
 635         struct ptlrpc_request *req;
 636         struct mdt_body *body;
 637         int rc = 0;
 638         ENTRY;
 639
 640         LASSERT(entry->se_handle != 0);
 641
 642         minfo = entry->se_minfo;
 643         it = &minfo->mi_it;
 644         req = entry->se_req;
 645         body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
 646         if (body == NULL)
 647                 GOTO(out, rc = -EFAULT);
 648
 649         child = entry->se_inode;
 650         if (child != NULL) {
 651                 /* revalidate; unlinked and re-created with the same name */
 652                 if (unlikely(!lu_fid_eq(&minfo->mi_data.op_fid2,
 653                                         &body->mbo_fid1))) {
 654                         entry->se_inode = NULL;
 655                         iput(child);
 656                         child = NULL;
 657                 }
 658         }
 659
 660         it->it_lock_handle = entry->se_handle;
 661         rc = md_revalidate_lock(ll_i2mdexp(dir), it, ll_inode2fid(dir), NULL);
 662         if (rc != 1)
 663                 GOTO(out, rc = -EAGAIN);
 664
 665         rc = ll_prep_inode(&child, req, dir->i_sb, it);
 666         if (rc)
 667                 GOTO(out, rc);
 668
 669         CDEBUG(D_READA, "%s: setting %.*s"DFID" l_data to inode %p\n",
 670                ll_i2sbi(dir)->ll_fsname, entry->se_qstr.len,
 671                entry->se_qstr.name, PFID(ll_inode2fid(child)), child);
 672         ll_set_lock_data(ll_i2sbi(dir)->ll_md_exp, child, it, NULL);
 673
 674         entry->se_inode = child;
 675
 676         if (agl_should_run(sai, child))
 677                 ll_agl_add(sai, child, entry->se_index);
 678
 679         EXIT;
 680
 681 out:
 682         /* sa_make_ready() will drop ldlm ibits lock refcount by calling
 683          * ll_intent_drop_lock() in spite of failures. Do not worry about
 684          * calling ll_intent_drop_lock() more than once. */
 685         sa_make_ready(sai, entry, rc);
 686 }
 687
 688 /* once there are async stat replies, instantiate sa_entry from replies */
 689 static void sa_handle_callback(struct ll_statahead_info *sai)
 690 {
 691         struct ll_inode_info *lli;
 692
 693         lli = ll_i2info(sai->sai_dentry->d_inode);
 694
 695         spin_lock(&lli->lli_sa_lock);
 696         while (sa_has_callback(sai)) {
 697                 struct sa_entry *entry;
 698
 699                 entry = list_entry(sai->sai_interim_entries.next,
 700                                    struct sa_entry, se_list);
 701                 list_del_init(&entry->se_list);
 702                 spin_unlock(&lli->lli_sa_lock);
 703
 704                 sa_instantiate(sai, entry);
 705                 spin_lock(&lli->lli_sa_lock);
 706         }
 707         spin_unlock(&lli->lli_sa_lock);
 708 }
 709
 710 /*
 711  * callback for async stat RPC, because this is called in ptlrpcd context, we
 712  * only put sa_entry in sai_interim_entries, and wake up statahead thread to
 713  * really prepare inode and instantiate sa_entry later.
 714  */
 715 static int ll_statahead_interpret(struct ptlrpc_request *req,
 716                                   struct md_enqueue_info *minfo, int rc)
 717 {
 718         struct lookup_intent *it = &minfo->mi_it;
 719         struct inode *dir = minfo->mi_dir;
 720         struct ll_inode_info *lli = ll_i2info(dir);
 721         struct ll_statahead_info *sai = lli->lli_sai;
 722         struct sa_entry *entry = (struct sa_entry *)minfo->mi_cbdata;
 723         __u64 handle = 0;
 724         wait_queue_head_t *waitq = NULL;
 725         ENTRY;
 726
 727         if (it_disposition(it, DISP_LOOKUP_NEG))
 728                 rc = -ENOENT;
 729
 730         /* because statahead thread will wait for all inflight RPC to finish,
 731          * sai should be always valid, no need to refcount */
 732         LASSERT(sai != NULL);
 733         LASSERT(!thread_is_stopped(&sai->sai_thread));
 734         LASSERT(entry != NULL);
 735
 736         CDEBUG(D_READA, "sa_entry %.*s rc %d\n",
 737                entry->se_qstr.len, entry->se_qstr.name, rc);
 738
 739         if (rc != 0) {
 740                 ll_intent_release(it);
 741                 sa_fini_data(minfo);
 742         } else {
 743                 /* release ibits lock ASAP to avoid deadlock when statahead
 744                  * thread enqueues lock on parent in readdir and another
 745                  * process enqueues lock on child with parent lock held, eg.
 746                  * unlink. */
 747                 handle = it->it_lock_handle;
 748                 ll_intent_drop_lock(it);
 749                 ll_unlock_md_op_lsm(&minfo->mi_data);
 750         }
 751
 752         spin_lock(&lli->lli_sa_lock);
 753         if (rc != 0) {
 754                 if (__sa_make_ready(sai, entry, rc))
 755                         waitq = &sai->sai_waitq;
 756         } else {
 757                 entry->se_minfo = minfo;
 758                 entry->se_req = ptlrpc_request_addref(req);
 759                 /* Release the async ibits lock ASAP to avoid deadlock
 760                  * when statahead thread tries to enqueue lock on parent
 761                  * for readpage and other tries to enqueue lock on child
 762                  * with parent's lock held, for example: unlink. */
 763                 entry->se_handle = handle;
 764                 if (!sa_has_callback(sai))
 765                         waitq = &sai->sai_thread.t_ctl_waitq;
 766
 767                 list_add_tail(&entry->se_list, &sai->sai_interim_entries);
 768         }
 769         sai->sai_replied++;
 770
 771         if (waitq != NULL)
 772                 wake_up(waitq);
 773         spin_unlock(&lli->lli_sa_lock);
 774
 775         RETURN(rc);
 776 }
 777
 778 /* async stat for file not found in dcache */
 779 static int sa_lookup(struct inode *dir, struct sa_entry *entry)
 780 {
 781         struct md_enqueue_info   *minfo;
 782         int                       rc;
 783         ENTRY;
 784
 785         minfo = sa_prep_data(dir, NULL, entry);
 786         if (IS_ERR(minfo))
 787                 RETURN(PTR_ERR(minfo));
 788
 789         rc = md_intent_getattr_async(ll_i2mdexp(dir), minfo);
 790         if (rc < 0)
 791                 sa_fini_data(minfo);
 792
 793         RETURN(rc);
 794 }
 795
 796 /**
 797  * async stat for file found in dcache, similar to .revalidate
 798  *
 799  * \retval      1 dentry valid, no RPC sent
 800  * \retval      0 dentry invalid, will send async stat RPC
 801  * \retval      negative number upon error
 802  */
 803 static int sa_revalidate(struct inode *dir, struct sa_entry *entry,
 804                          struct dentry *dentry)
 805 {
 806         struct inode *inode = dentry->d_inode;
 807         struct lookup_intent it = { .it_op = IT_GETATTR,
 808                                     .it_lock_handle = 0 };
 809         struct md_enqueue_info *minfo;
 810         int rc;
 811         ENTRY;
 812
 813         if (unlikely(inode == NULL))
 814                 RETURN(1);
 815
 816         if (d_mountpoint(dentry))
 817                 RETURN(1);
 818
 819         minfo = sa_prep_data(dir, inode, entry);
 820         if (IS_ERR(minfo))
 821                 RETURN(PTR_ERR(minfo));
 822
 823         entry->se_inode = igrab(inode);
 824         rc = md_revalidate_lock(ll_i2mdexp(dir), &it, ll_inode2fid(inode),
 825                                 NULL);
 826         if (rc == 1) {
 827                 entry->se_handle = it.it_lock_handle;
 828                 ll_intent_release(&it);
 829                 sa_fini_data(minfo);
 830                 RETURN(1);
 831         }
 832
 833         rc = md_intent_getattr_async(ll_i2mdexp(dir), minfo);
 834         if (rc < 0) {
 835                 entry->se_inode = NULL;
 836                 iput(inode);
 837                 sa_fini_data(minfo);
 838         }
 839
 840         RETURN(rc);
 841 }
 842
 843 /* async stat for file with @name */
 844 static void sa_statahead(struct dentry *parent, const char *name, int len,
 845                          const struct lu_fid *fid)
 846 {
 847         struct inode *dir = parent->d_inode;
 848         struct ll_inode_info *lli = ll_i2info(dir);
 849         struct ll_statahead_info *sai = lli->lli_sai;
 850         struct dentry *dentry = NULL;
 851         struct sa_entry *entry;
 852         int rc;
 853         ENTRY;
 854
 855         entry = sa_alloc(parent, sai, sai->sai_index, name, len, fid);
 856         if (IS_ERR(entry))
 857                 RETURN_EXIT;
 858
 859         dentry = d_lookup(parent, &entry->se_qstr);
 860         if (!dentry) {
 861                 rc = sa_lookup(dir, entry);
 862         } else {
 863                 rc = sa_revalidate(dir, entry, dentry);
 864                 if (rc == 1 && agl_should_run(sai, dentry->d_inode))
 865                         ll_agl_add(sai, dentry->d_inode, entry->se_index);
 866         }
 867
 868         if (dentry != NULL)
 869                 dput(dentry);
 870
 871         if (rc != 0)
 872                 sa_make_ready(sai, entry, rc);
 873         else
 874                 sai->sai_sent++;
 875
 876         sai->sai_index++;
 877
 878         EXIT;
 879 }
 880
 881 /* async glimpse (agl) thread main function */
 882 static int ll_agl_thread(void *arg)
 883 {
 884         struct dentry *parent = (struct dentry *)arg;
 885         struct inode *dir = parent->d_inode;
 886         struct ll_inode_info *plli = ll_i2info(dir);
 887         struct ll_inode_info *clli;
 888         struct ll_sb_info *sbi = ll_i2sbi(dir);
 889         struct ll_statahead_info *sai;
 890         struct ptlrpc_thread *thread;
 891         struct l_wait_info lwi = { 0 };
 892         ENTRY;
 893
 894         sai = ll_sai_get(dir);
 895         thread = &sai->sai_agl_thread;
 896         thread->t_pid = current_pid();
 897         CDEBUG(D_READA, "agl thread started: sai %p, parent %.*s\n",
 898                sai, parent->d_name.len, parent->d_name.name);
 899
 900         atomic_inc(&sbi->ll_agl_total);
 901         spin_lock(&plli->lli_agl_lock);
 902         sai->sai_agl_valid = 1;
 903         if (thread_is_init(thread))
 904                 /* If someone else has changed the thread state
 905                  * (e.g. already changed to SVC_STOPPING), we can't just
 906                  * blindly overwrite that setting. */
 907                 thread_set_flags(thread, SVC_RUNNING);
 908         spin_unlock(&plli->lli_agl_lock);
 909         wake_up(&thread->t_ctl_waitq);
 910
 911         while (1) {
 912                 l_wait_event(thread->t_ctl_waitq,
 913                              !agl_list_empty(sai) ||
 914                              !thread_is_running(thread),
 915                              &lwi);
 916                 if (!thread_is_running(thread))
 917                         break;
 918
 919                 spin_lock(&plli->lli_agl_lock);
 920                 /* The statahead thread maybe help to process AGL entries,
 921                  * so check whether list empty again. */
 922                 if (!agl_list_empty(sai)) {
 923                         clli = agl_first_entry(sai);
 924                         list_del_init(&clli->lli_agl_list);
 925                         spin_unlock(&plli->lli_agl_lock);
 926                         ll_agl_trigger(&clli->lli_vfs_inode, sai);
 927                         cond_resched();
 928                 } else {
 929                         spin_unlock(&plli->lli_agl_lock);
 930                 }
 931         }
 932
 933         spin_lock(&plli->lli_agl_lock);
 934         sai->sai_agl_valid = 0;
 935         while (!agl_list_empty(sai)) {
 936                 clli = agl_first_entry(sai);
 937                 list_del_init(&clli->lli_agl_list);
 938                 spin_unlock(&plli->lli_agl_lock);
 939                 clli->lli_agl_index = 0;
 940                 iput(&clli->lli_vfs_inode);
 941                 spin_lock(&plli->lli_agl_lock);
 942         }
 943         thread_set_flags(thread, SVC_STOPPED);
 944         spin_unlock(&plli->lli_agl_lock);
 945         wake_up(&thread->t_ctl_waitq);
 946         ll_sai_put(sai);
 947         CDEBUG(D_READA, "agl thread stopped: sai %p, parent %.*s\n",
 948                sai, parent->d_name.len, parent->d_name.name);
 949         RETURN(0);
 950 }
 951
 952 /* start agl thread */
 953 static void ll_start_agl(struct dentry *parent, struct ll_statahead_info *sai)
 954 {
 955         struct ptlrpc_thread *thread = &sai->sai_agl_thread;
 956         struct l_wait_info    lwi    = { 0 };
 957         struct ll_inode_info  *plli;
 958         struct task_struct            *task;
 959         ENTRY;
 960
 961         CDEBUG(D_READA, "start agl thread: sai %p, parent %.*s\n",
 962                sai, parent->d_name.len, parent->d_name.name);
 963
 964         plli = ll_i2info(parent->d_inode);
 965         task = kthread_run(ll_agl_thread, parent,
 966                                "ll_agl_%u", plli->lli_opendir_pid);
 967         if (IS_ERR(task)) {
 968                 CERROR("can't start ll_agl thread, rc: %ld\n", PTR_ERR(task));
 969                 thread_set_flags(thread, SVC_STOPPED);
 970                 RETURN_EXIT;
 971         }
 972
 973         l_wait_event(thread->t_ctl_waitq,
 974                      thread_is_running(thread) || thread_is_stopped(thread),
 975                      &lwi);
 976         EXIT;
 977 }
 978
 979 /* statahead thread main function */
 980 static int ll_statahead_thread(void *arg)
 981 {
 982         struct dentry *parent = (struct dentry *)arg;
 983         struct inode *dir = parent->d_inode;
 984         struct ll_inode_info *lli = ll_i2info(dir);
 985         struct ll_sb_info *sbi = ll_i2sbi(dir);
 986         struct ll_statahead_info *sai;
 987         struct ptlrpc_thread *sa_thread;
 988         struct ptlrpc_thread *agl_thread;
 989         int first = 0;
 990         struct md_op_data *op_data;
 991         struct ll_dir_chain chain;
 992         struct l_wait_info lwi = { 0 };
 993         struct page *page = NULL;
 994         __u64 pos = 0;
 995         int rc = 0;
 996         ENTRY;
 997
 998         sai = ll_sai_get(dir);
 999         sa_thread = &sai->sai_thread;
1000         agl_thread = &sai->sai_agl_thread;
1001         sa_thread->t_pid = current_pid();
1002         CDEBUG(D_READA, "statahead thread starting: sai %p, parent %.*s\n",
1003                sai, parent->d_name.len, parent->d_name.name);
1004
1005         OBD_ALLOC_PTR(op_data);
1006         if (!op_data)
1007                 GOTO(out, rc = -ENOMEM);
1008
1009         if (sbi->ll_flags & LL_SBI_AGL_ENABLED)
1010                 ll_start_agl(parent, sai);
1011
1012         atomic_inc(&sbi->ll_sa_total);
1013         spin_lock(&lli->lli_sa_lock);
1014         if (thread_is_init(sa_thread))
1015                 /* If someone else has changed the thread state
1016                  * (e.g. already changed to SVC_STOPPING), we can't just
1017                  * blindly overwrite that setting. */
1018                 thread_set_flags(sa_thread, SVC_RUNNING);
1019         spin_unlock(&lli->lli_sa_lock);
1020         wake_up(&sa_thread->t_ctl_waitq);
1021
1022         ll_dir_chain_init(&chain);
1023         while (pos != MDS_DIR_END_OFF && thread_is_running(sa_thread)) {
1024                 struct lu_dirpage *dp;
1025                 struct lu_dirent  *ent;
1026
1027                 op_data = ll_prep_md_op_data(op_data, dir, dir, NULL, 0, 0,
1028                                      LUSTRE_OPC_ANY, dir);
1029                 if (IS_ERR(op_data)) {
1030                         rc = PTR_ERR(op_data);
1031                         break;
1032                 }
1033
1034                 sai->sai_in_readpage = 1;
1035                 page = ll_get_dir_page(dir, op_data, pos, &chain);
1036                 ll_unlock_md_op_lsm(op_data);
1037                 sai->sai_in_readpage = 0;
1038                 if (IS_ERR(page)) {
1039                         rc = PTR_ERR(page);
1040                         CDEBUG(D_READA, "error reading dir "DFID" at %llu"
1041                                "/%llu opendir_pid = %u: rc = %d\n",
1042                                PFID(ll_inode2fid(dir)), pos, sai->sai_index,
1043                                lli->lli_opendir_pid, rc);
1044                         break;
1045                 }
1046
1047                 dp = page_address(page);
1048                 for (ent = lu_dirent_start(dp);
1049                      ent != NULL && thread_is_running(sa_thread) &&
1050                      !sa_low_hit(sai);
1051                      ent = lu_dirent_next(ent)) {
1052                         __u64 hash;
1053                         int namelen;
1054                         char *name;
1055                         struct lu_fid fid;
1056
1057                         hash = le64_to_cpu(ent->lde_hash);
1058                         if (unlikely(hash < pos))
1059                                 /*
1060                                  * Skip until we find target hash value.
1061                                  */
1062                                 continue;
1063
1064                         namelen = le16_to_cpu(ent->lde_namelen);
1065                         if (unlikely(namelen == 0))
1066                                 /*
1067                                  * Skip dummy record.
1068                                  */
1069                                 continue;
1070
1071                         name = ent->lde_name;
1072                         if (name[0] == '.') {
1073                                 if (namelen == 1) {
1074                                         /*
1075                                          * skip "."
1076                                          */
1077                                         continue;
1078                                 } else if (name[1] == '.' && namelen == 2) {
1079                                         /*
1080                                          * skip ".."
1081                                          */
1082                                         continue;
1083                                 } else if (!sai->sai_ls_all) {
1084                                         /*
1085                                          * skip hidden files.
1086                                          */
1087                                         sai->sai_skip_hidden++;
1088                                         continue;
1089                                 }
1090                         }
1091
1092                         /*
1093                          * don't stat-ahead first entry.
1094                          */
1095                         if (unlikely(++first == 1))
1096                                 continue;
1097
1098                         fid_le_to_cpu(&fid, &ent->lde_fid);
1099
1100                         /* wait for spare statahead window */
1101                         do {
1102                                 l_wait_event(sa_thread->t_ctl_waitq,
1103                                              !sa_sent_full(sai) ||
1104                                              sa_has_callback(sai) ||
1105                                              !agl_list_empty(sai) ||
1106                                              !thread_is_running(sa_thread),
1107                                              &lwi);
1108
1109                                 sa_handle_callback(sai);
1110
1111                                 spin_lock(&lli->lli_agl_lock);
1112                                 while (sa_sent_full(sai) &&
1113                                        !agl_list_empty(sai)) {
1114                                         struct ll_inode_info *clli;
1115
1116                                         clli = agl_first_entry(sai);
1117                                         list_del_init(&clli->lli_agl_list);
1118                                         spin_unlock(&lli->lli_agl_lock);
1119
1120                                         ll_agl_trigger(&clli->lli_vfs_inode,
1121                                                         sai);
1122                                         cond_resched();
1123                                         spin_lock(&lli->lli_agl_lock);
1124                                 }
1125                                 spin_unlock(&lli->lli_agl_lock);
1126                         } while (sa_sent_full(sai) &&
1127                                  thread_is_running(sa_thread));
1128
1129                         sa_statahead(parent, name, namelen, &fid);
1130                 }
1131
1132                 pos = le64_to_cpu(dp->ldp_hash_end);
1133                 ll_release_page(dir, page,
1134                                 le32_to_cpu(dp->ldp_flags) & LDF_COLLIDE);
1135
1136                 if (sa_low_hit(sai)) {
1137                         rc = -EFAULT;
1138                         atomic_inc(&sbi->ll_sa_wrong);
1139                         CDEBUG(D_READA, "Statahead for dir "DFID" hit "
1140                                "ratio too low: hit/miss %llu/%llu"
1141                                ", sent/replied %llu/%llu, stopping "
1142                                "statahead thread: pid %d\n",
1143                                PFID(&lli->lli_fid), sai->sai_hit,
1144                                sai->sai_miss, sai->sai_sent,
1145                                sai->sai_replied, current_pid());
1146                         break;
1147                 }
1148         }
1149         ll_dir_chain_fini(&chain);
1150         ll_finish_md_op_data(op_data);
1151
1152         if (rc < 0) {
1153                 spin_lock(&lli->lli_sa_lock);
1154                 thread_set_flags(sa_thread, SVC_STOPPING);
1155                 lli->lli_sa_enabled = 0;
1156                 spin_unlock(&lli->lli_sa_lock);
1157         }
1158
1159         /* statahead is finished, but statahead entries need to be cached, wait
1160          * for file release to stop me. */
1161         while (thread_is_running(sa_thread)) {
1162                 l_wait_event(sa_thread->t_ctl_waitq,
1163                              sa_has_callback(sai) ||
1164                              !thread_is_running(sa_thread),
1165                              &lwi);
1166
1167                 sa_handle_callback(sai);
1168         }
1169
1170         EXIT;
1171 out:
1172         if (sai->sai_agl_valid) {
1173                 spin_lock(&lli->lli_agl_lock);
1174                 thread_set_flags(agl_thread, SVC_STOPPING);
1175                 spin_unlock(&lli->lli_agl_lock);
1176                 wake_up(&agl_thread->t_ctl_waitq);
1177
1178                 CDEBUG(D_READA, "stop agl thread: sai %p pid %u\n",
1179                        sai, (unsigned int)agl_thread->t_pid);
1180                 l_wait_event(agl_thread->t_ctl_waitq,
1181                              thread_is_stopped(agl_thread),
1182                              &lwi);
1183         } else {
1184                 /* Set agl_thread flags anyway. */
1185                 thread_set_flags(agl_thread, SVC_STOPPED);
1186         }
1187
1188         /* wait for inflight statahead RPCs to finish, and then we can free sai
1189          * safely because statahead RPC will access sai data */
1190         while (sai->sai_sent != sai->sai_replied) {
1191                 /* in case we're not woken up, timeout wait */
1192                 lwi = LWI_TIMEOUT(cfs_time_seconds(1) >> 3, NULL, NULL);
1193                 l_wait_event(sa_thread->t_ctl_waitq,
1194                         sai->sai_sent == sai->sai_replied, &lwi);
1195         }
1196
1197         /* release resources held by statahead RPCs */
1198         sa_handle_callback(sai);
1199
1200         spin_lock(&lli->lli_sa_lock);
1201         thread_set_flags(sa_thread, SVC_STOPPED);
1202         spin_unlock(&lli->lli_sa_lock);
1203
1204         CDEBUG(D_READA, "statahead thread stopped: sai %p, parent %.*s\n",
1205                sai, parent->d_name.len, parent->d_name.name);
1206
1207         wake_up(&sai->sai_waitq);
1208         wake_up(&sa_thread->t_ctl_waitq);
1209         ll_sai_put(sai);
1210
1211         return rc;
1212 }
1213
1214 /* authorize opened dir handle @key to statahead */
1215 void ll_authorize_statahead(struct inode *dir, void *key)
1216 {
1217         struct ll_inode_info *lli = ll_i2info(dir);
1218
1219         spin_lock(&lli->lli_sa_lock);
1220         if (lli->lli_opendir_key == NULL && lli->lli_sai == NULL) {
1221                 /*
1222                  * if lli_sai is not NULL, it means previous statahead is not
1223                  * finished yet, we'd better not start a new statahead for now.
1224                  */
1225                 LASSERT(lli->lli_opendir_pid == 0);
1226                 lli->lli_opendir_key = key;
1227                 lli->lli_opendir_pid = current_pid();
1228                 lli->lli_sa_enabled = 1;
1229         }
1230         spin_unlock(&lli->lli_sa_lock);
1231 }
1232
1233 /*
1234  * deauthorize opened dir handle @key to statahead, and notify statahead thread
1235  * to quit if it's running.
1236  */
1237 void ll_deauthorize_statahead(struct inode *dir, void *key)
1238 {
1239         struct ll_inode_info *lli = ll_i2info(dir);
1240         struct ll_statahead_info *sai;
1241
1242         LASSERT(lli->lli_opendir_key == key);
1243         LASSERT(lli->lli_opendir_pid != 0);
1244
1245         CDEBUG(D_READA, "deauthorize statahead for "DFID"\n",
1246                 PFID(&lli->lli_fid));
1247
1248         spin_lock(&lli->lli_sa_lock);
1249         lli->lli_opendir_key = NULL;
1250         lli->lli_opendir_pid = 0;
1251         lli->lli_sa_enabled = 0;
1252         sai = lli->lli_sai;
1253         if (sai != NULL && thread_is_running(&sai->sai_thread)) {
1254                 /*
1255                  * statahead thread may not quit yet because it needs to cache
1256                  * entries, now it's time to tell it to quit.
1257                  *
1258                  * In case sai is released, wake_up() is called inside spinlock,
1259                  * so we have to call smp_mb() explicitely to serialize ops.
1260                  */
1261                 thread_set_flags(&sai->sai_thread, SVC_STOPPING);
1262                 smp_mb();
1263                 wake_up(&sai->sai_thread.t_ctl_waitq);
1264         }
1265         spin_unlock(&lli->lli_sa_lock);
1266 }
1267
1268 enum {
1269         /**
1270          * not first dirent, or is "."
1271          */
1272         LS_NOT_FIRST_DE = 0,
1273         /**
1274          * the first non-hidden dirent
1275          */
1276         LS_FIRST_DE,
1277         /**
1278          * the first hidden dirent, that is "."
1279          */
1280         LS_FIRST_DOT_DE
1281 };
1282
1283 /* file is first dirent under @dir */
1284 static int is_first_dirent(struct inode *dir, struct dentry *dentry)
1285 {
1286         struct ll_dir_chain   chain;
1287         struct qstr          *target = &dentry->d_name;
1288         struct md_op_data    *op_data;
1289         int                   dot_de;
1290         struct page          *page = NULL;
1291         int                   rc = LS_NOT_FIRST_DE;
1292         __u64                 pos = 0;
1293         ENTRY;
1294
1295         op_data = ll_prep_md_op_data(NULL, dir, dir, NULL, 0, 0,
1296                                      LUSTRE_OPC_ANY, dir);
1297         if (IS_ERR(op_data))
1298                 RETURN(PTR_ERR(op_data));
1299         /**
1300          *FIXME choose the start offset of the readdir
1301          */
1302
1303         ll_dir_chain_init(&chain);
1304         page = ll_get_dir_page(dir, op_data, 0, &chain);
1305
1306         while (1) {
1307                 struct lu_dirpage *dp;
1308                 struct lu_dirent  *ent;
1309
1310                 if (IS_ERR(page)) {
1311                         struct ll_inode_info *lli = ll_i2info(dir);
1312
1313                         rc = PTR_ERR(page);
1314                         CERROR("%s: reading dir "DFID" at %llu"
1315                                "opendir_pid = %u : rc = %d\n",
1316                                ll_i2sbi(dir)->ll_fsname,
1317                                PFID(ll_inode2fid(dir)), pos,
1318                                lli->lli_opendir_pid, rc);
1319                         break;
1320                 }
1321
1322                 dp = page_address(page);
1323                 for (ent = lu_dirent_start(dp); ent != NULL;
1324                      ent = lu_dirent_next(ent)) {
1325                         __u64 hash;
1326                         int namelen;
1327                         char *name;
1328
1329                         hash = le64_to_cpu(ent->lde_hash);
1330                         /* The ll_get_dir_page() can return any page containing
1331                          * the given hash which may be not the start hash. */
1332                         if (unlikely(hash < pos))
1333                                 continue;
1334
1335                         namelen = le16_to_cpu(ent->lde_namelen);
1336                         if (unlikely(namelen == 0))
1337                                 /*
1338                                  * skip dummy record.
1339                                  */
1340                                 continue;
1341
1342                         name = ent->lde_name;
1343                         if (name[0] == '.') {
1344                                 if (namelen == 1)
1345                                         /*
1346                                          * skip "."
1347                                          */
1348                                         continue;
1349                                 else if (name[1] == '.' && namelen == 2)
1350                                         /*
1351                                          * skip ".."
1352                                          */
1353                                         continue;
1354                                 else
1355                                         dot_de = 1;
1356                         } else {
1357                                 dot_de = 0;
1358                         }
1359
1360                         if (dot_de && target->name[0] != '.') {
1361                                 CDEBUG(D_READA, "%.*s skip hidden file %.*s\n",
1362                                        target->len, target->name,
1363                                        namelen, name);
1364                                 continue;
1365                         }
1366
1367                         if (target->len != namelen ||
1368                             memcmp(target->name, name, namelen) != 0)
1369                                 rc = LS_NOT_FIRST_DE;
1370                         else if (!dot_de)
1371                                 rc = LS_FIRST_DE;
1372                         else
1373                                 rc = LS_FIRST_DOT_DE;
1374
1375                         ll_release_page(dir, page, false);
1376                         GOTO(out, rc);
1377                 }
1378                 pos = le64_to_cpu(dp->ldp_hash_end);
1379                 if (pos == MDS_DIR_END_OFF) {
1380                         /*
1381                          * End of directory reached.
1382                          */
1383                         ll_release_page(dir, page, false);
1384                         GOTO(out, rc);
1385                 } else {
1386                         /*
1387                          * chain is exhausted
1388                          * Normal case: continue to the next page.
1389                          */
1390                         ll_release_page(dir, page, le32_to_cpu(dp->ldp_flags) &
1391                                               LDF_COLLIDE);
1392                         page = ll_get_dir_page(dir, op_data, pos, &chain);
1393                 }
1394         }
1395         EXIT;
1396 out:
1397         ll_dir_chain_fini(&chain);
1398         ll_finish_md_op_data(op_data);
1399         return rc;
1400 }
1401
1402 /**
1403  * revalidate @dentryp from statahead cache
1404  *
1405  * \param[in] dir       parent directory
1406  * \param[in] sai       sai structure
1407  * \param[out] dentryp  pointer to dentry which will be revalidated
1408  * \param[in] unplug    unplug statahead window only (normally for negative
1409  *                      dentry)
1410  * \retval              1 on success, dentry is saved in @dentryp
1411  * \retval              0 if revalidation failed (no proper lock on client)
1412  * \retval              negative number upon error
1413  */
1414 static int revalidate_statahead_dentry(struct inode *dir,
1415                                         struct ll_statahead_info *sai,
1416                                         struct dentry **dentryp,
1417                                         bool unplug)
1418 {
1419         struct sa_entry *entry = NULL;
1420         struct l_wait_info lwi = { 0 };
1421         struct ll_dentry_data *ldd;
1422         struct ll_inode_info *lli = ll_i2info(dir);
1423         int rc = 0;
1424         ENTRY;
1425
1426         if ((*dentryp)->d_name.name[0] == '.') {
1427                 if (sai->sai_ls_all ||
1428                     sai->sai_miss_hidden >= sai->sai_skip_hidden) {
1429                         /*
1430                          * Hidden dentry is the first one, or statahead
1431                          * thread does not skip so many hidden dentries
1432                          * before "sai_ls_all" enabled as below.
1433                          */
1434                 } else {
1435                         if (!sai->sai_ls_all)
1436                                 /*
1437                                  * It maybe because hidden dentry is not
1438                                  * the first one, "sai_ls_all" was not
1439                                  * set, then "ls -al" missed. Enable
1440                                  * "sai_ls_all" for such case.
1441                                  */
1442                                 sai->sai_ls_all = 1;
1443
1444                         /*
1445                          * Such "getattr" has been skipped before
1446                          * "sai_ls_all" enabled as above.
1447                          */
1448                         sai->sai_miss_hidden++;
1449                         RETURN(-EAGAIN);
1450                 }
1451         }
1452
1453         if (unplug)
1454                 GOTO(out, rc = 1);
1455
1456         entry = sa_get(sai, &(*dentryp)->d_name);
1457         if (entry == NULL)
1458                 GOTO(out, rc = -EAGAIN);
1459
1460         /* if statahead is busy in readdir, help it do post-work */
1461         if (!sa_ready(entry) && sai->sai_in_readpage)
1462                 sa_handle_callback(sai);
1463
1464         if (!sa_ready(entry)) {
1465                 spin_lock(&lli->lli_sa_lock);
1466                 sai->sai_index_wait = entry->se_index;
1467                 spin_unlock(&lli->lli_sa_lock);
1468                 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(30), NULL,
1469                                        LWI_ON_SIGNAL_NOOP, NULL);
1470                 rc = l_wait_event(sai->sai_waitq, sa_ready(entry), &lwi);
1471                 if (rc < 0) {
1472                         /*
1473                          * entry may not be ready, so it may be used by inflight
1474                          * statahead RPC, don't free it.
1475                          */
1476                         entry = NULL;
1477                         GOTO(out, rc = -EAGAIN);
1478                 }
1479         }
1480
1481         /*
1482          * We need to see the value that was set immediately before we
1483          * were woken up.
1484          */
1485         if (smp_load_acquire(&entry->se_state) == SA_ENTRY_SUCC &&
1486             entry->se_inode) {
1487                 struct inode *inode = entry->se_inode;
1488                 struct lookup_intent it = { .it_op = IT_GETATTR,
1489                                             .it_lock_handle =
1490                                                 entry->se_handle };
1491                 __u64 bits;
1492
1493                 rc = md_revalidate_lock(ll_i2mdexp(dir), &it,
1494                                         ll_inode2fid(inode), &bits);
1495                 if (rc == 1) {
1496                         if ((*dentryp)->d_inode == NULL) {
1497                                 struct dentry *alias;
1498
1499                                 alias = ll_splice_alias(inode, *dentryp);
1500                                 if (IS_ERR(alias)) {
1501                                         ll_intent_release(&it);
1502                                         GOTO(out, rc = PTR_ERR(alias));
1503                                 }
1504                                 *dentryp = alias;
1505                                 /* statahead prepared this inode, transfer inode
1506                                  * refcount from sa_entry to dentry */
1507                                 entry->se_inode = NULL;
1508                         } else if ((*dentryp)->d_inode != inode) {
1509                                 /* revalidate, but inode is recreated */
1510                                 CDEBUG(D_READA,
1511                                         "%s: stale dentry %.*s inode "
1512                                         DFID", statahead inode "DFID
1513                                         "\n",
1514                                         ll_i2sbi(inode)->ll_fsname,
1515                                         (*dentryp)->d_name.len,
1516                                         (*dentryp)->d_name.name,
1517                                         PFID(ll_inode2fid((*dentryp)->d_inode)),
1518                                         PFID(ll_inode2fid(inode)));
1519                                 ll_intent_release(&it);
1520                                 GOTO(out, rc = -ESTALE);
1521                         }
1522
1523                         if ((bits & MDS_INODELOCK_LOOKUP) &&
1524                             d_lustre_invalid(*dentryp))
1525                                 d_lustre_revalidate(*dentryp);
1526                         ll_intent_release(&it);
1527                 }
1528         }
1529 out:
1530         /*
1531          * statahead cached sa_entry can be used only once, and will be killed
1532          * right after use, so if lookup/revalidate accessed statahead cache,
1533          * set dentry ldd_sa_generation to parent lli_sa_generation, later if we
1534          * stat this file again, we know we've done statahead before, see
1535          * dentry_may_statahead().
1536          */
1537         ldd = ll_d2d(*dentryp);
1538         /* ldd can be NULL if llite lookup failed. */
1539         if (ldd != NULL)
1540                 ldd->lld_sa_generation = lli->lli_sa_generation;
1541         sa_put(sai, entry);
1542
1543         RETURN(rc);
1544 }
1545
1546 /**
1547  * start statahead thread
1548  *
1549  * \param[in] dir       parent directory
1550  * \param[in] dentry    dentry that triggers statahead, normally the first
1551  *                      dirent under @dir
1552  * \retval              -EAGAIN on success, because when this function is
1553  *                      called, it's already in lookup call, so client should
1554  *                      do it itself instead of waiting for statahead thread
1555  *                      to do it asynchronously.
1556  * \retval              negative number upon error
1557  */
1558 static int start_statahead_thread(struct inode *dir, struct dentry *dentry)
1559 {
1560         struct ll_inode_info *lli = ll_i2info(dir);
1561         struct ll_statahead_info *sai = NULL;
1562         struct dentry *parent = dentry->d_parent;
1563         struct ptlrpc_thread *thread;
1564         struct l_wait_info lwi = { 0 };
1565         struct task_struct *task;
1566         struct ll_sb_info *sbi = ll_i2sbi(parent->d_inode);
1567         int first = LS_FIRST_DE;
1568         int rc = 0;
1569         ENTRY;
1570
1571         /* I am the "lli_opendir_pid" owner, only me can set "lli_sai". */
1572         first = is_first_dirent(dir, dentry);
1573         if (first == LS_NOT_FIRST_DE)
1574                 /* It is not "ls -{a}l" operation, no need statahead for it. */
1575                 GOTO(out, rc = -EFAULT);
1576
1577         if (unlikely(atomic_inc_return(&sbi->ll_sa_running) >
1578                                        sbi->ll_sa_running_max)) {
1579                 CDEBUG(D_READA,
1580                        "Too many concurrent statahead instances, "
1581                        "avoid new statahead instance temporarily.\n");
1582                 GOTO(out, rc = -EMFILE);
1583         }
1584
1585         sai = ll_sai_alloc(parent);
1586         if (sai == NULL)
1587                 GOTO(out, rc = -ENOMEM);
1588
1589         sai->sai_ls_all = (first == LS_FIRST_DOT_DE);
1590
1591         /* if current lli_opendir_key was deauthorized, or dir re-opened by
1592          * another process, don't start statahead, otherwise the newly spawned
1593          * statahead thread won't be notified to quit. */
1594         spin_lock(&lli->lli_sa_lock);
1595         if (unlikely(lli->lli_sai != NULL ||
1596                      lli->lli_opendir_key == NULL ||
1597                      lli->lli_opendir_pid != current->pid)) {
1598                 spin_unlock(&lli->lli_sa_lock);
1599                 GOTO(out, rc = -EPERM);
1600         }
1601         lli->lli_sai = sai;
1602         spin_unlock(&lli->lli_sa_lock);
1603
1604         CDEBUG(D_READA, "start statahead thread: [pid %d] [parent %.*s]\n",
1605                current_pid(), parent->d_name.len, parent->d_name.name);
1606
1607         task = kthread_run(ll_statahead_thread, parent, "ll_sa_%u",
1608                            lli->lli_opendir_pid);
1609         thread = &sai->sai_thread;
1610         if (IS_ERR(task)) {
1611                 spin_lock(&lli->lli_sa_lock);
1612                 lli->lli_sai = NULL;
1613                 spin_unlock(&lli->lli_sa_lock);
1614                 rc = PTR_ERR(task);
1615                 CERROR("can't start ll_sa thread, rc: %d\n", rc);
1616                 GOTO(out, rc);
1617         }
1618
1619         l_wait_event(thread->t_ctl_waitq,
1620                      thread_is_running(thread) || thread_is_stopped(thread),
1621                      &lwi);
1622         ll_sai_put(sai);
1623
1624         /*
1625          * We don't stat-ahead for the first dirent since we are already in
1626          * lookup.
1627          */
1628         RETURN(-EAGAIN);
1629
1630 out:
1631         /* once we start statahead thread failed, disable statahead so that
1632          * subsequent stat won't waste time to try it. */
1633         spin_lock(&lli->lli_sa_lock);
1634         if (lli->lli_opendir_pid == current->pid)
1635                 lli->lli_sa_enabled = 0;
1636         spin_unlock(&lli->lli_sa_lock);
1637
1638         if (sai != NULL)
1639                 ll_sai_free(sai);
1640         if (first != LS_NOT_FIRST_DE)
1641                 atomic_dec(&sbi->ll_sa_running);
1642
1643         RETURN(rc);
1644 }
1645
1646 /**
1647  * statahead entry function, this is called when client getattr on a file, it
1648  * will start statahead thread if this is the first dir entry, else revalidate
1649  * dentry from statahead cache.
1650  *
1651  * \param[in]  dir      parent directory
1652  * \param[out] dentryp  dentry to getattr
1653  * \param[in]  unplug   unplug statahead window only (normally for negative
1654  *                      dentry)
1655  * \retval              1 on success
1656  * \retval              0 revalidation from statahead cache failed, caller needs
1657  *                      to getattr from server directly
1658  * \retval              negative number on error, caller often ignores this and
1659  *                      then getattr from server
1660  */
1661 int ll_statahead(struct inode *dir, struct dentry **dentryp, bool unplug)
1662 {
1663         struct ll_statahead_info *sai;
1664
1665         sai = ll_sai_get(dir);
1666         if (sai != NULL) {
1667                 int rc;
1668
1669                 rc = revalidate_statahead_dentry(dir, sai, dentryp, unplug);
1670                 CDEBUG(D_READA, "revalidate statahead %.*s: %d.\n",
1671                         (*dentryp)->d_name.len, (*dentryp)->d_name.name, rc);
1672                 ll_sai_put(sai);
1673                 return rc;
1674         }
1675         return start_statahead_thread(dir, *dentryp);
1676 }