lustre/llite/statahead.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2014, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  */
  36
  37 #include <linux/fs.h>
  38 #include <linux/sched.h>
  39 #include <linux/kthread.h>
  40 #include <linux/mm.h>
  41 #include <linux/highmem.h>
  42 #include <linux/pagemap.h>
  43
  44 #define DEBUG_SUBSYSTEM S_LLITE
  45
  46 #include <obd_support.h>
  47 #include <lustre_dlm.h>
  48 #include "llite_internal.h"
  49
  50 #define SA_OMITTED_ENTRY_MAX 8ULL
  51
  52 typedef enum {
  53         /** negative values are for error cases */
  54         SA_ENTRY_INIT = 0,      /** init entry */
  55         SA_ENTRY_SUCC = 1,      /** stat succeed */
  56         SA_ENTRY_INVA = 2,      /** invalid entry */
  57 } se_state_t;
  58
  59 /* sa_entry is not refcounted: statahead thread allocates it and do async stat,
  60  * and in async stat callback ll_statahead_interpret() will add it into
  61  * sai_interim_entries, later statahead thread will call sa_handle_callback() to
  62  * instantiate entry and move it into sai_entries, and then only scanner process
  63  * can access and free it. */
  64 struct sa_entry {
  65         /* link into sai_interim_entries or sai_entries */
  66         struct list_head        se_list;
  67         /* link into sai hash table locally */
  68         struct list_head        se_hash;
  69         /* entry index in the sai */
  70         __u64                   se_index;
  71         /* low layer ldlm lock handle */
  72         __u64                   se_handle;
  73         /* entry status */
  74         se_state_t              se_state;
  75         /* entry size, contains name */
  76         int                     se_size;
  77         /* pointer to async getattr enqueue info */
  78         struct md_enqueue_info *se_minfo;
  79         /* pointer to the async getattr request */
  80         struct ptlrpc_request  *se_req;
  81         /* pointer to the target inode */
  82         struct inode           *se_inode;
  83         /* entry name */
  84         struct qstr             se_qstr;
  85 };
  86
  87 static unsigned int sai_generation = 0;
  88 static DEFINE_SPINLOCK(sai_generation_lock);
  89
  90 static inline int sa_unhashed(struct sa_entry *entry)
  91 {
  92         return list_empty(&entry->se_hash);
  93 }
  94
  95 /* sa_entry is ready to use */
  96 static inline int sa_ready(struct sa_entry *entry)
  97 {
  98         smp_rmb();
  99         return (entry->se_state != SA_ENTRY_INIT);
 100 }
 101
 102 /* hash value to put in sai_cache */
 103 static inline int sa_hash(int val)
 104 {
 105         return val & LL_SA_CACHE_MASK;
 106 }
 107
 108 /* hash entry into sai_cache */
 109 static inline void
 110 sa_rehash(struct ll_statahead_info *sai, struct sa_entry *entry)
 111 {
 112         int i = sa_hash(entry->se_qstr.hash);
 113
 114         spin_lock(&sai->sai_cache_lock[i]);
 115         list_add_tail(&entry->se_hash, &sai->sai_cache[i]);
 116         spin_unlock(&sai->sai_cache_lock[i]);
 117 }
 118
 119 /* unhash entry from sai_cache */
 120 static inline void
 121 sa_unhash(struct ll_statahead_info *sai, struct sa_entry *entry)
 122 {
 123         int i = sa_hash(entry->se_qstr.hash);
 124
 125         spin_lock(&sai->sai_cache_lock[i]);
 126         list_del_init(&entry->se_hash);
 127         spin_unlock(&sai->sai_cache_lock[i]);
 128 }
 129
 130 static inline int agl_should_run(struct ll_statahead_info *sai,
 131                                  struct inode *inode)
 132 {
 133         return (inode != NULL && S_ISREG(inode->i_mode) && sai->sai_agl_valid);
 134 }
 135
 136 static inline struct ll_inode_info *
 137 agl_first_entry(struct ll_statahead_info *sai)
 138 {
 139         return list_entry(sai->sai_agls.next, struct ll_inode_info,
 140                           lli_agl_list);
 141 }
 142
 143 /* statahead window is full */
 144 static inline int sa_sent_full(struct ll_statahead_info *sai)
 145 {
 146         return atomic_read(&sai->sai_cache_count) >= sai->sai_max;
 147 }
 148
 149 /* got async stat replies */
 150 static inline int sa_has_callback(struct ll_statahead_info *sai)
 151 {
 152         return !list_empty(&sai->sai_interim_entries);
 153 }
 154
 155 static inline int agl_list_empty(struct ll_statahead_info *sai)
 156 {
 157         return list_empty(&sai->sai_agls);
 158 }
 159
 160 /**
 161  * (1) hit ratio less than 80%
 162  * or
 163  * (2) consecutive miss more than 8
 164  * then means low hit.
 165  */
 166 static inline int sa_low_hit(struct ll_statahead_info *sai)
 167 {
 168         return ((sai->sai_hit > 7 && sai->sai_hit < 4 * sai->sai_miss) ||
 169                 (sai->sai_consecutive_miss > 8));
 170 }
 171
 172 /*
 173  * if the given index is behind of statahead window more than
 174  * SA_OMITTED_ENTRY_MAX, then it is old.
 175  */
 176 static inline int is_omitted_entry(struct ll_statahead_info *sai, __u64 index)
 177 {
 178         return ((__u64)sai->sai_max + index + SA_OMITTED_ENTRY_MAX <
 179                  sai->sai_index);
 180 }
 181
 182 /* allocate sa_entry and hash it to allow scanner process to find it */
 183 static struct sa_entry *
 184 sa_alloc(struct ll_statahead_info *sai, __u64 index, const char *name, int len)
 185 {
 186         struct ll_inode_info *lli;
 187         struct sa_entry *entry;
 188         int entry_size;
 189         char *dname;
 190         ENTRY;
 191
 192         entry_size = sizeof(struct sa_entry) + (len & ~3) + 4;
 193         OBD_ALLOC(entry, entry_size);
 194         if (unlikely(entry == NULL))
 195                 RETURN(ERR_PTR(-ENOMEM));
 196
 197         CDEBUG(D_READA, "alloc sa entry %.*s(%p) index "LPU64"\n",
 198                len, name, entry, index);
 199
 200         entry->se_index = index;
 201
 202         entry->se_state = SA_ENTRY_INIT;
 203         entry->se_size = entry_size;
 204         dname = (char *)entry + sizeof(struct sa_entry);
 205         memcpy(dname, name, len);
 206         dname[len] = 0;
 207         entry->se_qstr.hash = full_name_hash(name, len);
 208         entry->se_qstr.len = len;
 209         entry->se_qstr.name = dname;
 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 "LPU64"\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         /* Someone is in glimpse (sync or async), do nothing. */
 510         rc = down_write_trylock(&lli->lli_glimpse_sem);
 511         if (rc == 0) {
 512                 lli->lli_agl_index = 0;
 513                 iput(inode);
 514                 RETURN_EXIT;
 515         }
 516
 517         /*
 518          * Someone triggered glimpse within 1 sec before.
 519          * 1) The former glimpse succeeded with glimpse lock granted by OST, and
 520          *    if the lock is still cached on client, AGL needs to do nothing. If
 521          *    it is cancelled by other client, AGL maybe cannot obtaion new lock
 522          *    for no glimpse callback triggered by AGL.
 523          * 2) The former glimpse succeeded, but OST did not grant glimpse lock.
 524          *    Under such case, it is quite possible that the OST will not grant
 525          *    glimpse lock for AGL also.
 526          * 3) The former glimpse failed, compared with other two cases, it is
 527          *    relative rare. AGL can ignore such case, and it will not muchly
 528          *    affect the performance.
 529          */
 530         if (lli->lli_glimpse_time != 0 &&
 531             cfs_time_before(cfs_time_shift(-1), lli->lli_glimpse_time)) {
 532                 up_write(&lli->lli_glimpse_sem);
 533                 lli->lli_agl_index = 0;
 534                 iput(inode);
 535                 RETURN_EXIT;
 536         }
 537
 538         CDEBUG(D_READA, "Handling (init) async glimpse: inode = "
 539                DFID", idx = "LPU64"\n", PFID(&lli->lli_fid), index);
 540
 541         cl_agl(inode);
 542         lli->lli_agl_index = 0;
 543         lli->lli_glimpse_time = cfs_time_current();
 544         up_write(&lli->lli_glimpse_sem);
 545
 546         CDEBUG(D_READA, "Handled (init) async glimpse: inode= "
 547                DFID", idx = "LPU64", rc = %d\n",
 548                PFID(&lli->lli_fid), index, rc);
 549
 550         iput(inode);
 551
 552         EXIT;
 553 }
 554
 555 /*
 556  * prepare inode for sa entry, add it into agl list, now sa_entry is ready
 557  * to be used by scanner process.
 558  */
 559 static void sa_instantiate(struct ll_statahead_info *sai,
 560                                  struct sa_entry *entry)
 561 {
 562         struct inode *dir = sai->sai_dentry->d_inode;
 563         struct inode *child;
 564         struct md_enqueue_info *minfo;
 565         struct lookup_intent *it;
 566         struct ptlrpc_request *req;
 567         struct mdt_body *body;
 568         int rc = 0;
 569         ENTRY;
 570
 571         LASSERT(entry->se_handle != 0);
 572
 573         minfo = entry->se_minfo;
 574         it = &minfo->mi_it;
 575         req = entry->se_req;
 576         body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
 577         if (body == NULL)
 578                 GOTO(out, rc = -EFAULT);
 579
 580         child = entry->se_inode;
 581         if (child == NULL) {
 582                 /*
 583                  * lookup.
 584                  */
 585                 LASSERT(fid_is_zero(&minfo->mi_data.op_fid2));
 586
 587                 /* XXX: No fid in reply, this is probaly cross-ref case.
 588                  * SA can't handle it yet. */
 589                 if (body->mbo_valid & OBD_MD_MDS)
 590                         GOTO(out, rc = -EAGAIN);
 591         } else {
 592                 /*
 593                  * revalidate.
 594                  */
 595                 /* unlinked and re-created with the same name */
 596                 if (unlikely(!lu_fid_eq(&minfo->mi_data.op_fid2,
 597                                         &body->mbo_fid1))) {
 598                         entry->se_inode = NULL;
 599                         iput(child);
 600                         child = NULL;
 601                 }
 602         }
 603
 604         it->d.lustre.it_lock_handle = entry->se_handle;
 605         rc = md_revalidate_lock(ll_i2mdexp(dir), it, ll_inode2fid(dir), NULL);
 606         if (rc != 1)
 607                 GOTO(out, rc = -EAGAIN);
 608
 609         rc = ll_prep_inode(&child, req, dir->i_sb, it);
 610         if (rc)
 611                 GOTO(out, rc);
 612
 613         CDEBUG(D_READA, "%s: setting %.*s"DFID" l_data to inode %p\n",
 614                ll_get_fsname(child->i_sb, NULL, 0),
 615                entry->se_qstr.len, entry->se_qstr.name,
 616                PFID(ll_inode2fid(child)), child);
 617         ll_set_lock_data(ll_i2sbi(dir)->ll_md_exp, child, it, NULL);
 618
 619         entry->se_inode = child;
 620
 621         if (agl_should_run(sai, child))
 622                 ll_agl_add(sai, child, entry->se_index);
 623
 624         EXIT;
 625
 626 out:
 627         /* sa_make_ready() will drop ldlm ibits lock refcount by calling
 628          * ll_intent_drop_lock() in spite of failures. Do not worry about
 629          * calling ll_intent_drop_lock() more than once. */
 630         sa_make_ready(sai, entry, rc);
 631 }
 632
 633 /* once there are async stat replies, instantiate sa_entry from replies */
 634 static void sa_handle_callback(struct ll_statahead_info *sai)
 635 {
 636         struct ll_inode_info *lli;
 637
 638         lli = ll_i2info(sai->sai_dentry->d_inode);
 639
 640         while (sa_has_callback(sai)) {
 641                 struct sa_entry *entry;
 642
 643                 spin_lock(&lli->lli_sa_lock);
 644                 if (unlikely(!sa_has_callback(sai))) {
 645                         spin_unlock(&lli->lli_sa_lock);
 646                         break;
 647                 }
 648                 entry = list_entry(sai->sai_interim_entries.next,
 649                                    struct sa_entry, se_list);
 650                 list_del_init(&entry->se_list);
 651                 spin_unlock(&lli->lli_sa_lock);
 652
 653                 sa_instantiate(sai, entry);
 654         }
 655 }
 656
 657 /*
 658  * callback for async stat RPC, because this is called in ptlrpcd context, we
 659  * only put sa_entry in sai_interim_entries, and wake up statahead thread to
 660  * really prepare inode and instantiate sa_entry later.
 661  */
 662 static int ll_statahead_interpret(struct ptlrpc_request *req,
 663                                   struct md_enqueue_info *minfo, int rc)
 664 {
 665         struct lookup_intent *it = &minfo->mi_it;
 666         struct inode *dir = minfo->mi_dir;
 667         struct ll_inode_info *lli = ll_i2info(dir);
 668         struct ll_statahead_info *sai = lli->lli_sai;
 669         struct sa_entry *entry = (struct sa_entry *)minfo->mi_cbdata;
 670         __u64 handle = 0;
 671         bool wakeup;
 672         ENTRY;
 673
 674         if (it_disposition(it, DISP_LOOKUP_NEG))
 675                 rc = -ENOENT;
 676
 677         /* because statahead thread will wait for all inflight RPC to finish,
 678          * sai should be always valid, no need to refcount */
 679         LASSERT(sai != NULL);
 680         LASSERT(!thread_is_stopped(&sai->sai_thread));
 681         LASSERT(entry != NULL);
 682
 683         CDEBUG(D_READA, "sa_entry %.*s rc %d\n",
 684                entry->se_qstr.len, entry->se_qstr.name, rc);
 685
 686         if (rc != 0) {
 687                 ll_intent_release(it);
 688                 iput(dir);
 689                 OBD_FREE_PTR(minfo);
 690         } else {
 691                 /* release ibits lock ASAP to avoid deadlock when statahead
 692                  * thread enqueues lock on parent in readdir and another
 693                  * process enqueues lock on child with parent lock held, eg.
 694                  * unlink. */
 695                 handle = it->d.lustre.it_lock_handle;
 696                 ll_intent_drop_lock(it);
 697         }
 698
 699         spin_lock(&lli->lli_sa_lock);
 700         if (rc != 0) {
 701                 wakeup = __sa_make_ready(sai, entry, rc);
 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                 wakeup = !sa_has_callback(sai);
 711                 list_add_tail(&entry->se_list, &sai->sai_interim_entries);
 712         }
 713         sai->sai_replied++;
 714         if (wakeup)
 715                 wake_up(&sai->sai_thread.t_ctl_waitq);
 716         spin_unlock(&lli->lli_sa_lock);
 717
 718         RETURN(rc);
 719 }
 720
 721 /* finish async stat RPC arguments */
 722 static void sa_fini_data(struct md_enqueue_info *minfo,
 723                          struct ldlm_enqueue_info *einfo)
 724 {
 725         LASSERT(minfo && einfo);
 726         iput(minfo->mi_dir);
 727         OBD_FREE_PTR(minfo);
 728         OBD_FREE_PTR(einfo);
 729 }
 730
 731 /*
 732  * prepare arguments for async stat RPC.
 733  */
 734 static int sa_prep_data(struct inode *dir, struct inode *child,
 735                         struct sa_entry *entry, struct md_enqueue_info **pmi,
 736                         struct ldlm_enqueue_info **pei)
 737 {
 738         struct qstr              *qstr = &entry->se_qstr;
 739         struct md_enqueue_info   *minfo;
 740         struct ldlm_enqueue_info *einfo;
 741         struct md_op_data        *op_data;
 742
 743         OBD_ALLOC_PTR(einfo);
 744         if (einfo == NULL)
 745                 return -ENOMEM;
 746
 747         OBD_ALLOC_PTR(minfo);
 748         if (minfo == NULL) {
 749                 OBD_FREE_PTR(einfo);
 750                 return -ENOMEM;
 751         }
 752
 753         op_data = ll_prep_md_op_data(&minfo->mi_data, dir, child, qstr->name,
 754                                      qstr->len, 0, LUSTRE_OPC_ANY, NULL);
 755         if (IS_ERR(op_data)) {
 756                 OBD_FREE_PTR(einfo);
 757                 OBD_FREE_PTR(minfo);
 758                 return PTR_ERR(op_data);
 759         }
 760
 761         minfo->mi_it.it_op = IT_GETATTR;
 762         minfo->mi_dir = igrab(dir);
 763         minfo->mi_cb = ll_statahead_interpret;
 764         minfo->mi_cbdata = entry;
 765
 766         einfo->ei_type   = LDLM_IBITS;
 767         einfo->ei_mode   = it_to_lock_mode(&minfo->mi_it);
 768         einfo->ei_cb_bl  = ll_md_blocking_ast;
 769         einfo->ei_cb_cp  = ldlm_completion_ast;
 770         einfo->ei_cb_gl  = NULL;
 771         einfo->ei_cbdata = NULL;
 772
 773         *pmi = minfo;
 774         *pei = einfo;
 775
 776         return 0;
 777 }
 778
 779 /* async stat for file not found in dcache */
 780 static int sa_lookup(struct inode *dir, struct sa_entry *entry)
 781 {
 782         struct md_enqueue_info   *minfo;
 783         struct ldlm_enqueue_info *einfo;
 784         int                       rc;
 785         ENTRY;
 786
 787         rc = sa_prep_data(dir, NULL, entry, &minfo, &einfo);
 788         if (rc)
 789                 RETURN(rc);
 790
 791         rc = md_intent_getattr_async(ll_i2mdexp(dir), minfo, einfo);
 792         if (rc < 0)
 793                 sa_fini_data(minfo, einfo);
 794
 795         RETURN(rc);
 796 }
 797
 798 /**
 799  * async stat for file found in dcache, similar to .revalidate
 800  *
 801  * \retval      1 dentry valid, no RPC sent
 802  * \retval      0 dentry invalid, will send async stat RPC
 803  * \retval      negative number upon error
 804  */
 805 static int sa_revalidate(struct inode *dir, struct sa_entry *entry,
 806                          struct dentry *dentry)
 807 {
 808         struct inode *inode = dentry->d_inode;
 809         struct lookup_intent it = { .it_op = IT_GETATTR,
 810                                     .d.lustre.it_lock_handle = 0 };
 811         struct md_enqueue_info *minfo;
 812         struct ldlm_enqueue_info *einfo;
 813         int rc;
 814         ENTRY;
 815
 816         if (unlikely(inode == NULL))
 817                 RETURN(1);
 818
 819         if (d_mountpoint(dentry))
 820                 RETURN(1);
 821
 822         entry->se_inode = igrab(inode);
 823         rc = md_revalidate_lock(ll_i2mdexp(dir), &it, ll_inode2fid(inode),
 824                                 NULL);
 825         if (rc == 1) {
 826                 entry->se_handle = it.d.lustre.it_lock_handle;
 827                 ll_intent_release(&it);
 828                 RETURN(1);
 829         }
 830
 831         rc = sa_prep_data(dir, inode, entry, &minfo, &einfo);
 832         if (rc) {
 833                 entry->se_inode = NULL;
 834                 iput(inode);
 835                 RETURN(rc);
 836         }
 837
 838         rc = md_intent_getattr_async(ll_i2mdexp(dir), minfo, einfo);
 839         if (rc < 0) {
 840                 entry->se_inode = NULL;
 841                 iput(inode);
 842                 sa_fini_data(minfo, einfo);
 843         }
 844
 845         RETURN(rc);
 846 }
 847
 848 /* async stat for file with @name */
 849 static void sa_statahead(struct dentry *parent, const char *name, int len)
 850 {
 851         struct inode *dir = parent->d_inode;
 852         struct ll_inode_info *lli = ll_i2info(dir);
 853         struct ll_statahead_info *sai = lli->lli_sai;
 854         struct dentry *dentry = NULL;
 855         struct sa_entry *entry;
 856         int rc;
 857         ENTRY;
 858
 859         entry = sa_alloc(sai, sai->sai_index, name, len);
 860         if (IS_ERR(entry))
 861                 RETURN_EXIT;
 862
 863         dentry = d_lookup(parent, &entry->se_qstr);
 864         if (!dentry) {
 865                 rc = sa_lookup(dir, entry);
 866         } else {
 867                 rc = sa_revalidate(dir, entry, dentry);
 868                 if (rc == 1 && agl_should_run(sai, dentry->d_inode))
 869                         ll_agl_add(sai, dentry->d_inode, entry->se_index);
 870         }
 871
 872         if (dentry != NULL)
 873                 dput(dentry);
 874
 875         if (rc != 0)
 876                 sa_make_ready(sai, entry, rc);
 877         else
 878                 sai->sai_sent++;
 879
 880         sai->sai_index++;
 881
 882         EXIT;
 883 }
 884
 885 /* async glimpse (agl) thread main function */
 886 static int ll_agl_thread(void *arg)
 887 {
 888         struct dentry *parent = (struct dentry *)arg;
 889         struct inode *dir = parent->d_inode;
 890         struct ll_inode_info *plli = ll_i2info(dir);
 891         struct ll_inode_info *clli;
 892         struct ll_sb_info *sbi = ll_i2sbi(dir);
 893         struct ll_statahead_info *sai;
 894         struct ptlrpc_thread *thread;
 895         struct l_wait_info lwi = { 0 };
 896         ENTRY;
 897
 898
 899         sai = ll_sai_get(dir);
 900         thread = &sai->sai_agl_thread;
 901         thread->t_pid = current_pid();
 902         CDEBUG(D_READA, "agl thread started: sai %p, parent %.*s\n",
 903                sai, parent->d_name.len, parent->d_name.name);
 904
 905         atomic_inc(&sbi->ll_agl_total);
 906         spin_lock(&plli->lli_agl_lock);
 907         sai->sai_agl_valid = 1;
 908         if (thread_is_init(thread))
 909                 /* If someone else has changed the thread state
 910                  * (e.g. already changed to SVC_STOPPING), we can't just
 911                  * blindly overwrite that setting. */
 912                 thread_set_flags(thread, SVC_RUNNING);
 913         spin_unlock(&plli->lli_agl_lock);
 914         wake_up(&thread->t_ctl_waitq);
 915
 916         while (1) {
 917                 l_wait_event(thread->t_ctl_waitq,
 918                              !agl_list_empty(sai) ||
 919                              !thread_is_running(thread),
 920                              &lwi);
 921
 922                 if (!thread_is_running(thread))
 923                         break;
 924
 925                 spin_lock(&plli->lli_agl_lock);
 926                 /* The statahead thread maybe help to process AGL entries,
 927                  * so check whether list empty again. */
 928                 if (!agl_list_empty(sai)) {
 929                         clli = agl_first_entry(sai);
 930                         list_del_init(&clli->lli_agl_list);
 931                         spin_unlock(&plli->lli_agl_lock);
 932                         ll_agl_trigger(&clli->lli_vfs_inode, sai);
 933                 } else {
 934                         spin_unlock(&plli->lli_agl_lock);
 935                 }
 936         }
 937
 938         spin_lock(&plli->lli_agl_lock);
 939         sai->sai_agl_valid = 0;
 940         while (!agl_list_empty(sai)) {
 941                 clli = agl_first_entry(sai);
 942                 list_del_init(&clli->lli_agl_list);
 943                 spin_unlock(&plli->lli_agl_lock);
 944                 clli->lli_agl_index = 0;
 945                 iput(&clli->lli_vfs_inode);
 946                 spin_lock(&plli->lli_agl_lock);
 947         }
 948         thread_set_flags(thread, SVC_STOPPED);
 949         spin_unlock(&plli->lli_agl_lock);
 950         wake_up(&thread->t_ctl_waitq);
 951         ll_sai_put(sai);
 952         CDEBUG(D_READA, "agl thread stopped: sai %p, parent %.*s\n",
 953                sai, parent->d_name.len, parent->d_name.name);
 954         RETURN(0);
 955 }
 956
 957 /* start agl thread */
 958 static void ll_start_agl(struct dentry *parent, struct ll_statahead_info *sai)
 959 {
 960         struct ptlrpc_thread *thread = &sai->sai_agl_thread;
 961         struct l_wait_info    lwi    = { 0 };
 962         struct ll_inode_info  *plli;
 963         struct task_struct            *task;
 964         ENTRY;
 965
 966         CDEBUG(D_READA, "start agl thread: sai %p, parent %.*s\n",
 967                sai, parent->d_name.len, parent->d_name.name);
 968
 969         plli = ll_i2info(parent->d_inode);
 970         task = kthread_run(ll_agl_thread, parent,
 971                                "ll_agl_%u", plli->lli_opendir_pid);
 972         if (IS_ERR(task)) {
 973                 CERROR("can't start ll_agl thread, rc: %ld\n", PTR_ERR(task));
 974                 thread_set_flags(thread, SVC_STOPPED);
 975                 RETURN_EXIT;
 976         }
 977
 978         l_wait_event(thread->t_ctl_waitq,
 979                      thread_is_running(thread) || thread_is_stopped(thread),
 980                      &lwi);
 981         EXIT;
 982 }
 983
 984 /* statahead thread main function */
 985 static int ll_statahead_thread(void *arg)
 986 {
 987         struct dentry *parent = (struct dentry *)arg;
 988         struct inode *dir = parent->d_inode;
 989         struct ll_inode_info *lli = ll_i2info(dir);
 990         struct ll_sb_info *sbi = ll_i2sbi(dir);
 991         struct ll_statahead_info *sai;
 992         struct ptlrpc_thread *sa_thread;
 993         struct ptlrpc_thread *agl_thread;
 994         int first = 0;
 995         struct md_op_data *op_data;
 996         struct ll_dir_chain chain;
 997         struct l_wait_info lwi = { 0 };
 998         struct page *page = NULL;
 999         __u64 pos = 0;
1000         int rc = 0;
1001         ENTRY;
1002
1003         sai = ll_sai_get(dir);
1004         sa_thread = &sai->sai_thread;
1005         agl_thread = &sai->sai_agl_thread;
1006         sa_thread->t_pid = current_pid();
1007         CDEBUG(D_READA, "statahead thread starting: sai %p, parent %.*s\n",
1008                sai, parent->d_name.len, parent->d_name.name);
1009
1010         op_data = ll_prep_md_op_data(NULL, dir, dir, NULL, 0, 0,
1011                                      LUSTRE_OPC_ANY, dir);
1012         if (IS_ERR(op_data))
1013                 GOTO(out, rc = PTR_ERR(op_data));
1014
1015         op_data->op_max_pages = ll_i2sbi(dir)->ll_md_brw_pages;
1016
1017         if (sbi->ll_flags & LL_SBI_AGL_ENABLED)
1018                 ll_start_agl(parent, sai);
1019
1020         atomic_inc(&sbi->ll_sa_total);
1021         spin_lock(&lli->lli_sa_lock);
1022         if (thread_is_init(sa_thread))
1023                 /* If someone else has changed the thread state
1024                  * (e.g. already changed to SVC_STOPPING), we can't just
1025                  * blindly overwrite that setting. */
1026                 thread_set_flags(sa_thread, SVC_RUNNING);
1027         spin_unlock(&lli->lli_sa_lock);
1028         wake_up(&sa_thread->t_ctl_waitq);
1029
1030         ll_dir_chain_init(&chain);
1031         while (pos != MDS_DIR_END_OFF && thread_is_running(sa_thread)) {
1032                 struct lu_dirpage *dp;
1033                 struct lu_dirent  *ent;
1034
1035                 sai->sai_in_readpage = 1;
1036                 page = ll_get_dir_page(dir, op_data, pos, &chain);
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 "LPU64
1041                                "/"LPU64" 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
1056                         hash = le64_to_cpu(ent->lde_hash);
1057                         if (unlikely(hash < pos))
1058                                 /*
1059                                  * Skip until we find target hash value.
1060                                  */
1061                                 continue;
1062
1063                         namelen = le16_to_cpu(ent->lde_namelen);
1064                         if (unlikely(namelen == 0))
1065                                 /*
1066                                  * Skip dummy record.
1067                                  */
1068                                 continue;
1069
1070                         name = ent->lde_name;
1071                         if (name[0] == '.') {
1072                                 if (namelen == 1) {
1073                                         /*
1074                                          * skip "."
1075                                          */
1076                                         continue;
1077                                 } else if (name[1] == '.' && namelen == 2) {
1078                                         /*
1079                                          * skip ".."
1080                                          */
1081                                         continue;
1082                                 } else if (!sai->sai_ls_all) {
1083                                         /*
1084                                          * skip hidden files.
1085                                          */
1086                                         sai->sai_skip_hidden++;
1087                                         continue;
1088                                 }
1089                         }
1090
1091                         /*
1092                          * don't stat-ahead first entry.
1093                          */
1094                         if (unlikely(++first == 1))
1095                                 continue;
1096
1097                         /* wait for spare statahead window */
1098                         do {
1099                                 l_wait_event(sa_thread->t_ctl_waitq,
1100                                              !sa_sent_full(sai) ||
1101                                              sa_has_callback(sai) ||
1102                                              !agl_list_empty(sai) ||
1103                                              !thread_is_running(sa_thread),
1104                                              &lwi);
1105
1106                                 sa_handle_callback(sai);
1107
1108                                 spin_lock(&lli->lli_agl_lock);
1109                                 while (sa_sent_full(sai) &&
1110                                        !agl_list_empty(sai)) {
1111                                         struct ll_inode_info *clli;
1112
1113                                         clli = agl_first_entry(sai);
1114                                         list_del_init(&clli->lli_agl_list);
1115                                         spin_unlock(&lli->lli_agl_lock);
1116
1117                                         ll_agl_trigger(&clli->lli_vfs_inode,
1118                                                         sai);
1119
1120                                         spin_lock(&lli->lli_agl_lock);
1121                                 }
1122                                 spin_unlock(&lli->lli_agl_lock);
1123                         } while (sa_sent_full(sai) &&
1124                                  thread_is_running(sa_thread));
1125
1126                         sa_statahead(parent, name, namelen);
1127                 }
1128
1129                 pos = le64_to_cpu(dp->ldp_hash_end);
1130                 ll_release_page(dir, page,
1131                                 le32_to_cpu(dp->ldp_flags) & LDF_COLLIDE);
1132
1133                 if (sa_low_hit(sai)) {
1134                         rc = -EFAULT;
1135                         atomic_inc(&sbi->ll_sa_wrong);
1136                         CDEBUG(D_READA, "Statahead for dir "DFID" hit "
1137                                "ratio too low: hit/miss "LPU64"/"LPU64
1138                                ", sent/replied "LPU64"/"LPU64", stopping "
1139                                "statahead thread: pid %d\n",
1140                                PFID(&lli->lli_fid), sai->sai_hit,
1141                                sai->sai_miss, sai->sai_sent,
1142                                sai->sai_replied, current_pid());
1143                         break;
1144                 }
1145         }
1146         ll_dir_chain_fini(&chain);
1147         ll_finish_md_op_data(op_data);
1148
1149         if (rc < 0) {
1150                 spin_lock(&lli->lli_sa_lock);
1151                 thread_set_flags(sa_thread, SVC_STOPPING);
1152                 lli->lli_sa_enabled = 0;
1153                 spin_unlock(&lli->lli_sa_lock);
1154         }
1155
1156         /* statahead is finished, but statahead entries need to be cached, wait
1157          * for file release to stop me. */
1158         while (thread_is_running(sa_thread)) {
1159                 l_wait_event(sa_thread->t_ctl_waitq,
1160                              sa_has_callback(sai) ||
1161                              !thread_is_running(sa_thread),
1162                              &lwi);
1163
1164                 sa_handle_callback(sai);
1165         }
1166
1167         EXIT;
1168 out:
1169         if (sai->sai_agl_valid) {
1170                 spin_lock(&lli->lli_agl_lock);
1171                 thread_set_flags(agl_thread, SVC_STOPPING);
1172                 spin_unlock(&lli->lli_agl_lock);
1173                 wake_up(&agl_thread->t_ctl_waitq);
1174
1175                 CDEBUG(D_READA, "stop agl thread: sai %p pid %u\n",
1176                        sai, (unsigned int)agl_thread->t_pid);
1177                 l_wait_event(agl_thread->t_ctl_waitq,
1178                              thread_is_stopped(agl_thread),
1179                              &lwi);
1180         } else {
1181                 /* Set agl_thread flags anyway. */
1182                 thread_set_flags(agl_thread, SVC_STOPPED);
1183         }
1184
1185         /* wait for inflight statahead RPCs to finish, and then we can free sai
1186          * safely because statahead RPC will access sai data */
1187         while (sai->sai_sent != sai->sai_replied) {
1188                 /* in case we're not woken up, timeout wait */
1189                 lwi = LWI_TIMEOUT(msecs_to_jiffies(MSEC_PER_SEC >> 3),
1190                                   NULL, NULL);
1191                 l_wait_event(sa_thread->t_ctl_waitq,
1192                         sai->sai_sent == sai->sai_replied, &lwi);
1193         }
1194
1195         /* release resources held by statahead RPCs */
1196         sa_handle_callback(sai);
1197
1198         spin_lock(&lli->lli_sa_lock);
1199         thread_set_flags(sa_thread, SVC_STOPPED);
1200         spin_unlock(&lli->lli_sa_lock);
1201
1202         CDEBUG(D_READA, "statahead thread stopped: sai %p, parent %.*s\n",
1203                sai, parent->d_name.len, parent->d_name.name);
1204
1205         wake_up(&sai->sai_waitq);
1206         wake_up(&sa_thread->t_ctl_waitq);
1207         ll_sai_put(sai);
1208
1209         return rc;
1210 }
1211
1212 /* authorize opened dir handle @key to statahead */
1213 void ll_authorize_statahead(struct inode *dir, void *key)
1214 {
1215         struct ll_inode_info *lli = ll_i2info(dir);
1216
1217         spin_lock(&lli->lli_sa_lock);
1218         if (lli->lli_opendir_key == NULL && lli->lli_sai == NULL) {
1219                 /*
1220                  * if lli_sai is not NULL, it means previous statahead is not
1221                  * finished yet, we'd better not start a new statahead for now.
1222                  */
1223                 LASSERT(lli->lli_opendir_pid == 0);
1224                 lli->lli_opendir_key = key;
1225                 lli->lli_opendir_pid = current_pid();
1226                 lli->lli_sa_enabled = 1;
1227         }
1228         spin_unlock(&lli->lli_sa_lock);
1229 }
1230
1231 /*
1232  * deauthorize opened dir handle @key to statahead, and notify statahead thread
1233  * to quit if it's running.
1234  */
1235 void ll_deauthorize_statahead(struct inode *dir, void *key)
1236 {
1237         struct ll_inode_info *lli = ll_i2info(dir);
1238         struct ll_statahead_info *sai;
1239
1240         LASSERT(lli->lli_opendir_key == key);
1241         LASSERT(lli->lli_opendir_pid != 0);
1242
1243         CDEBUG(D_READA, "deauthorize statahead for "DFID"\n",
1244                 PFID(&lli->lli_fid));
1245
1246         spin_lock(&lli->lli_sa_lock);
1247         lli->lli_opendir_key = NULL;
1248         lli->lli_opendir_pid = 0;
1249         lli->lli_sa_enabled = 0;
1250         sai = lli->lli_sai;
1251         if (sai != NULL && thread_is_running(&sai->sai_thread)) {
1252                 /*
1253                  * statahead thread may not quit yet because it needs to cache
1254                  * entries, now it's time to tell it to quit.
1255                  */
1256                 thread_set_flags(&sai->sai_thread, SVC_STOPPING);
1257                 wake_up(&sai->sai_thread.t_ctl_waitq);
1258         }
1259         spin_unlock(&lli->lli_sa_lock);
1260 }
1261
1262 enum {
1263         /**
1264          * not first dirent, or is "."
1265          */
1266         LS_NOT_FIRST_DE = 0,
1267         /**
1268          * the first non-hidden dirent
1269          */
1270         LS_FIRST_DE,
1271         /**
1272          * the first hidden dirent, that is "."
1273          */
1274         LS_FIRST_DOT_DE
1275 };
1276
1277 /* file is first dirent under @dir */
1278 static int is_first_dirent(struct inode *dir, struct dentry *dentry)
1279 {
1280         struct ll_dir_chain   chain;
1281         struct qstr          *target = &dentry->d_name;
1282         struct md_op_data    *op_data;
1283         int                   dot_de;
1284         struct page          *page = NULL;
1285         int                   rc = LS_NOT_FIRST_DE;
1286         __u64                 pos = 0;
1287         ENTRY;
1288
1289         op_data = ll_prep_md_op_data(NULL, dir, dir, NULL, 0, 0,
1290                                      LUSTRE_OPC_ANY, dir);
1291         if (IS_ERR(op_data))
1292                 RETURN(PTR_ERR(op_data));
1293         /**
1294          *FIXME choose the start offset of the readdir
1295          */
1296         op_data->op_max_pages = ll_i2sbi(dir)->ll_md_brw_pages;
1297
1298         ll_dir_chain_init(&chain);
1299         page = ll_get_dir_page(dir, op_data, 0, &chain);
1300
1301         while (1) {
1302                 struct lu_dirpage *dp;
1303                 struct lu_dirent  *ent;
1304
1305                 if (IS_ERR(page)) {
1306                         struct ll_inode_info *lli = ll_i2info(dir);
1307
1308                         rc = PTR_ERR(page);
1309                         CERROR("%s: reading dir "DFID" at "LPU64
1310                                "opendir_pid = %u : rc = %d\n",
1311                                ll_get_fsname(dir->i_sb, NULL, 0),
1312                                PFID(ll_inode2fid(dir)), pos,
1313                                lli->lli_opendir_pid, rc);
1314                         break;
1315                 }
1316
1317                 dp = page_address(page);
1318                 for (ent = lu_dirent_start(dp); ent != NULL;
1319                      ent = lu_dirent_next(ent)) {
1320                         __u64 hash;
1321                         int namelen;
1322                         char *name;
1323
1324                         hash = le64_to_cpu(ent->lde_hash);
1325                         /* The ll_get_dir_page() can return any page containing
1326                          * the given hash which may be not the start hash. */
1327                         if (unlikely(hash < pos))
1328                                 continue;
1329
1330                         namelen = le16_to_cpu(ent->lde_namelen);
1331                         if (unlikely(namelen == 0))
1332                                 /*
1333                                  * skip dummy record.
1334                                  */
1335                                 continue;
1336
1337                         name = ent->lde_name;
1338                         if (name[0] == '.') {
1339                                 if (namelen == 1)
1340                                         /*
1341                                          * skip "."
1342                                          */
1343                                         continue;
1344                                 else if (name[1] == '.' && namelen == 2)
1345                                         /*
1346                                          * skip ".."
1347                                          */
1348                                         continue;
1349                                 else
1350                                         dot_de = 1;
1351                         } else {
1352                                 dot_de = 0;
1353                         }
1354
1355                         if (dot_de && target->name[0] != '.') {
1356                                 CDEBUG(D_READA, "%.*s skip hidden file %.*s\n",
1357                                        target->len, target->name,
1358                                        namelen, name);
1359                                 continue;
1360                         }
1361
1362                         if (target->len != namelen ||
1363                             memcmp(target->name, name, namelen) != 0)
1364                                 rc = LS_NOT_FIRST_DE;
1365                         else if (!dot_de)
1366                                 rc = LS_FIRST_DE;
1367                         else
1368                                 rc = LS_FIRST_DOT_DE;
1369
1370                         ll_release_page(dir, page, false);
1371                         GOTO(out, rc);
1372                 }
1373                 pos = le64_to_cpu(dp->ldp_hash_end);
1374                 if (pos == MDS_DIR_END_OFF) {
1375                         /*
1376                          * End of directory reached.
1377                          */
1378                         ll_release_page(dir, page, false);
1379                         GOTO(out, rc);
1380                 } else {
1381                         /*
1382                          * chain is exhausted
1383                          * Normal case: continue to the next page.
1384                          */
1385                         ll_release_page(dir, page, le32_to_cpu(dp->ldp_flags) &
1386                                               LDF_COLLIDE);
1387                         page = ll_get_dir_page(dir, op_data, pos, &chain);
1388                 }
1389         }
1390         EXIT;
1391 out:
1392         ll_dir_chain_fini(&chain);
1393         ll_finish_md_op_data(op_data);
1394         return rc;
1395 }
1396
1397 /**
1398  * revalidate @dentryp from statahead cache
1399  *
1400  * \param[in] dir       parent directory
1401  * \param[in] sai       sai structure
1402  * \param[out] dentryp  pointer to dentry which will be revalidated
1403  * \param[in] unplug    unplug statahead window only (normally for negative
1404  *                      dentry)
1405  * \retval              1 on success, dentry is saved in @dentryp
1406  * \retval              0 if revalidation failed (no proper lock on client)
1407  * \retval              negative number upon error
1408  */
1409 static int revalidate_statahead_dentry(struct inode *dir,
1410                                         struct ll_statahead_info *sai,
1411                                         struct dentry **dentryp,
1412                                         bool unplug)
1413 {
1414         struct sa_entry *entry = NULL;
1415         struct l_wait_info lwi = { 0 };
1416         struct ll_dentry_data *ldd;
1417         struct ll_inode_info *lli;
1418         int rc = 0;
1419         ENTRY;
1420
1421         if ((*dentryp)->d_name.name[0] == '.') {
1422                 if (sai->sai_ls_all ||
1423                     sai->sai_miss_hidden >= sai->sai_skip_hidden) {
1424                         /*
1425                          * Hidden dentry is the first one, or statahead
1426                          * thread does not skip so many hidden dentries
1427                          * before "sai_ls_all" enabled as below.
1428                          */
1429                 } else {
1430                         if (!sai->sai_ls_all)
1431                                 /*
1432                                  * It maybe because hidden dentry is not
1433                                  * the first one, "sai_ls_all" was not
1434                                  * set, then "ls -al" missed. Enable
1435                                  * "sai_ls_all" for such case.
1436                                  */
1437                                 sai->sai_ls_all = 1;
1438
1439                         /*
1440                          * Such "getattr" has been skipped before
1441                          * "sai_ls_all" enabled as above.
1442                          */
1443                         sai->sai_miss_hidden++;
1444                         RETURN(-EAGAIN);
1445                 }
1446         }
1447
1448         if (unplug)
1449                 GOTO(out, rc = 1);
1450
1451         entry = sa_get(sai, &(*dentryp)->d_name);
1452         if (entry == NULL)
1453                 GOTO(out, rc = -EAGAIN);
1454
1455         /* if statahead is busy in readdir, help it do post-work */
1456         if (!sa_ready(entry) && sai->sai_in_readpage)
1457                 sa_handle_callback(sai);
1458
1459         if (!sa_ready(entry)) {
1460                 sai->sai_index_wait = entry->se_index;
1461                 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(30), NULL,
1462                                         LWI_ON_SIGNAL_NOOP, NULL);
1463                 rc = l_wait_event(sai->sai_waitq, sa_ready(entry), &lwi);
1464                 if (rc < 0) {
1465                         /*
1466                          * entry may not be ready, so it may be used by inflight
1467                          * statahead RPC, don't free it.
1468                          */
1469                         entry = NULL;
1470                         GOTO(out, rc = -EAGAIN);
1471                 }
1472         }
1473
1474         if (entry->se_state == SA_ENTRY_SUCC && entry->se_inode != NULL) {
1475                 struct inode *inode = entry->se_inode;
1476                 struct lookup_intent it = { .it_op = IT_GETATTR,
1477                                             .d.lustre.it_lock_handle =
1478                                                 entry->se_handle };
1479                 __u64 bits;
1480
1481                 rc = md_revalidate_lock(ll_i2mdexp(dir), &it,
1482                                         ll_inode2fid(inode), &bits);
1483                 if (rc == 1) {
1484                         if ((*dentryp)->d_inode == NULL) {
1485                                 struct dentry *alias;
1486
1487                                 alias = ll_splice_alias(inode, *dentryp);
1488                                 if (IS_ERR(alias))
1489                                         GOTO(out, rc = PTR_ERR(alias));
1490                                 *dentryp = alias;
1491                                 /* statahead prepared this inode, transfer inode
1492                                  * refcount from sa_entry to dentry */
1493                                 entry->se_inode = NULL;
1494                         } else if ((*dentryp)->d_inode != inode) {
1495                                 /* revalidate, but inode is recreated */
1496                                 CDEBUG(D_READA,
1497                                         "%s: stale dentry %.*s inode "
1498                                         DFID", statahead inode "DFID
1499                                         "\n",
1500                                         ll_get_fsname((*dentryp)->d_inode->i_sb,
1501                                                       NULL, 0),
1502                                         (*dentryp)->d_name.len,
1503                                         (*dentryp)->d_name.name,
1504                                         PFID(ll_inode2fid((*dentryp)->d_inode)),
1505                                         PFID(ll_inode2fid(inode)));
1506                                 GOTO(out, rc = -ESTALE);
1507                         }
1508
1509                         if ((bits & MDS_INODELOCK_LOOKUP) &&
1510                             d_lustre_invalid(*dentryp))
1511                                 d_lustre_revalidate(*dentryp);
1512                         ll_intent_release(&it);
1513                 }
1514         }
1515 out:
1516         /*
1517          * statahead cached sa_entry can be used only once, and will be killed
1518          * right after use, so if lookup/revalidate accessed statahead cache,
1519          * set dentry ldd_sa_generation to parent lli_sa_generation, later if we
1520          * stat this file again, we know we've done statahead before, see
1521          * dentry_may_statahead().
1522          */
1523         ldd = ll_d2d(*dentryp);
1524         lli = ll_i2info(dir);
1525         /* ldd can be NULL if llite lookup failed. */
1526         if (ldd != NULL)
1527                 ldd->lld_sa_generation = lli->lli_sa_generation;
1528         sa_put(sai, entry);
1529
1530         RETURN(rc);
1531 }
1532
1533 /**
1534  * start statahead thread
1535  *
1536  * \param[in] dir       parent directory
1537  * \param[in] dentry    dentry that triggers statahead, normally the first
1538  *                      dirent under @dir
1539  * \retval              -EAGAIN on success, because when this function is
1540  *                      called, it's already in lookup call, so client should
1541  *                      do it itself instead of waiting for statahead thread
1542  *                      to do it asynchronously.
1543  * \retval              negative number upon error
1544  */
1545 static int start_statahead_thread(struct inode *dir, struct dentry *dentry)
1546 {
1547         struct ll_inode_info *lli = ll_i2info(dir);
1548         struct ll_statahead_info *sai = NULL;
1549         struct dentry *parent = dentry->d_parent;
1550         struct ptlrpc_thread *thread;
1551         struct l_wait_info lwi = { 0 };
1552         struct task_struct *task;
1553         int rc;
1554         ENTRY;
1555
1556         /* I am the "lli_opendir_pid" owner, only me can set "lli_sai". */
1557         rc = is_first_dirent(dir, dentry);
1558         if (rc == LS_NOT_FIRST_DE)
1559                 /* It is not "ls -{a}l" operation, no need statahead for it. */
1560                 GOTO(out, rc = -EFAULT);
1561
1562         sai = ll_sai_alloc(parent);
1563         if (sai == NULL)
1564                 GOTO(out, rc = -ENOMEM);
1565
1566         sai->sai_ls_all = (rc == LS_FIRST_DOT_DE);
1567
1568         /* if current lli_opendir_key was deauthorized, or dir re-opened by
1569          * another process, don't start statahead, otherwise the newly spawned
1570          * statahead thread won't be notified to quit. */
1571         spin_lock(&lli->lli_sa_lock);
1572         if (unlikely(lli->lli_sai != NULL ||
1573                      lli->lli_opendir_key == NULL ||
1574                      lli->lli_opendir_pid != current->pid)) {
1575                 spin_unlock(&lli->lli_sa_lock);
1576                 GOTO(out, rc = -EPERM);
1577         }
1578         lli->lli_sai = sai;
1579         spin_unlock(&lli->lli_sa_lock);
1580
1581         atomic_inc(&ll_i2sbi(parent->d_inode)->ll_sa_running);
1582
1583         CDEBUG(D_READA, "start statahead thread: [pid %d] [parent %.*s]\n",
1584                current_pid(), parent->d_name.len, parent->d_name.name);
1585
1586         task = kthread_run(ll_statahead_thread, parent, "ll_sa_%u",
1587                            lli->lli_opendir_pid);
1588         thread = &sai->sai_thread;
1589         if (IS_ERR(task)) {
1590                 rc = PTR_ERR(task);
1591                 CERROR("can't start ll_sa thread, rc: %d\n", rc);
1592                 GOTO(out, rc);
1593         }
1594
1595         l_wait_event(thread->t_ctl_waitq,
1596                      thread_is_running(thread) || thread_is_stopped(thread),
1597                      &lwi);
1598         ll_sai_put(sai);
1599
1600         /*
1601          * We don't stat-ahead for the first dirent since we are already in
1602          * lookup.
1603          */
1604         RETURN(-EAGAIN);
1605
1606 out:
1607         /* once we start statahead thread failed, disable statahead so that
1608          * subsequent stat won't waste time to try it. */
1609         spin_lock(&lli->lli_sa_lock);
1610         lli->lli_sa_enabled = 0;
1611         lli->lli_sai = NULL;
1612         spin_unlock(&lli->lli_sa_lock);
1613
1614         if (sai != NULL)
1615                 ll_sai_free(sai);
1616
1617         RETURN(rc);
1618 }
1619
1620 /**
1621  * statahead entry function, this is called when client getattr on a file, it
1622  * will start statahead thread if this is the first dir entry, else revalidate
1623  * dentry from statahead cache.
1624  *
1625  * \param[in]  dir      parent directory
1626  * \param[out] dentryp  dentry to getattr
1627  * \param[in]  unplug   unplug statahead window only (normally for negative
1628  *                      dentry)
1629  * \retval              1 on success
1630  * \retval              0 revalidation from statahead cache failed, caller needs
1631  *                      to getattr from server directly
1632  * \retval              negative number on error, caller often ignores this and
1633  *                      then getattr from server
1634  */
1635 int ll_statahead(struct inode *dir, struct dentry **dentryp, bool unplug)
1636 {
1637         struct ll_statahead_info *sai;
1638
1639         sai = ll_sai_get(dir);
1640         if (sai != NULL) {
1641                 int rc;
1642
1643                 rc = revalidate_statahead_dentry(dir, sai, dentryp, unplug);
1644                 CDEBUG(D_READA, "revalidate statahead %.*s: %d.\n",
1645                         (*dentryp)->d_name.len, (*dentryp)->d_name.name, rc);
1646                 ll_sai_put(sai);
1647                 return rc;
1648         }
1649         return start_statahead_thread(dir, *dentryp);
1650 }