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  */
  31
  32 #include <linux/fs.h>
  33 #include <linux/sched.h>
  34 #include <linux/kthread.h>
  35 #include <linux/mm.h>
  36 #include <linux/highmem.h>
  37 #include <linux/pagemap.h>
  38 #include <linux/delay.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 /*
  56  * sa_entry is not refcounted: statahead thread allocates it and do async stat,
  57  * and in async stat callback ll_statahead_interpret() will prepare the inode
  58  * and set lock data in the ptlrpcd context. Then the scanner process will be
  59  * woken up if this entry is the waiting one, can access and free it.
  60  */
  61 struct sa_entry {
  62         /* link into sai_entries */
  63         struct list_head        se_list;
  64         /* link into sai hash table locally */
  65         struct list_head        se_hash;
  66         /* entry index in the sai */
  67         __u64                   se_index;
  68         /* low layer ldlm lock handle */
  69         __u64                   se_handle;
  70         /* entry status */
  71         se_state_t              se_state;
  72         /* entry size, contains name */
  73         int                     se_size;
  74         /* pointer to the target inode */
  75         struct inode           *se_inode;
  76         /* entry name */
  77         struct qstr             se_qstr;
  78         /* entry fid */
  79         struct lu_fid           se_fid;
  80 };
  81
  82 static unsigned int sai_generation;
  83 static DEFINE_SPINLOCK(sai_generation_lock);
  84
  85 static inline int sa_unhashed(struct sa_entry *entry)
  86 {
  87         return list_empty(&entry->se_hash);
  88 }
  89
  90 /* sa_entry is ready to use */
  91 static inline int sa_ready(struct sa_entry *entry)
  92 {
  93         /* Make sure sa_entry is updated and ready to use */
  94         smp_rmb();
  95         return (entry->se_state != SA_ENTRY_INIT);
  96 }
  97
  98 /* hash value to put in sai_cache */
  99 static inline int sa_hash(int val)
 100 {
 101         return val & LL_SA_CACHE_MASK;
 102 }
 103
 104 /* hash entry into sai_cache */
 105 static inline void
 106 sa_rehash(struct ll_statahead_info *sai, struct sa_entry *entry)
 107 {
 108         int i = sa_hash(entry->se_qstr.hash);
 109
 110         spin_lock(&sai->sai_cache_lock[i]);
 111         list_add_tail(&entry->se_hash, &sai->sai_cache[i]);
 112         spin_unlock(&sai->sai_cache_lock[i]);
 113 }
 114
 115 /* unhash entry from sai_cache */
 116 static inline void
 117 sa_unhash(struct ll_statahead_info *sai, struct sa_entry *entry)
 118 {
 119         int i = sa_hash(entry->se_qstr.hash);
 120
 121         spin_lock(&sai->sai_cache_lock[i]);
 122         list_del_init(&entry->se_hash);
 123         spin_unlock(&sai->sai_cache_lock[i]);
 124 }
 125
 126 static inline int agl_should_run(struct ll_statahead_info *sai,
 127                                  struct inode *inode)
 128 {
 129         return inode && S_ISREG(inode->i_mode) && sai->sai_agl_task;
 130 }
 131
 132 static inline struct ll_inode_info *
 133 agl_first_entry(struct ll_statahead_info *sai)
 134 {
 135         return list_first_entry(&sai->sai_agls, struct ll_inode_info,
 136                                 lli_agl_list);
 137 }
 138
 139 /* statahead window is full */
 140 static inline int sa_sent_full(struct ll_statahead_info *sai)
 141 {
 142         return atomic_read(&sai->sai_cache_count) >= sai->sai_max;
 143 }
 144
 145 /* Batch metadata handle */
 146 static inline bool sa_has_batch_handle(struct ll_statahead_info *sai)
 147 {
 148         return sai->sai_bh != NULL;
 149 }
 150
 151 static inline void ll_statahead_flush_nowait(struct ll_statahead_info *sai)
 152 {
 153         if (sa_has_batch_handle(sai)) {
 154                 sai->sai_index_end = sai->sai_index - 1;
 155                 (void) md_batch_flush(ll_i2mdexp(sai->sai_dentry->d_inode),
 156                                       sai->sai_bh, false);
 157         }
 158 }
 159
 160 static inline int agl_list_empty(struct ll_statahead_info *sai)
 161 {
 162         return list_empty(&sai->sai_agls);
 163 }
 164
 165 /**
 166  * (1) hit ratio less than 80%
 167  * or
 168  * (2) consecutive miss more than 8
 169  * then means low hit.
 170  */
 171 static inline int sa_low_hit(struct ll_statahead_info *sai)
 172 {
 173         return ((sai->sai_hit > 7 && sai->sai_hit < 4 * sai->sai_miss) ||
 174                 (sai->sai_consecutive_miss > 8));
 175 }
 176
 177 /*
 178  * if the given index is behind of statahead window more than
 179  * SA_OMITTED_ENTRY_MAX, then it is old.
 180  */
 181 static inline int is_omitted_entry(struct ll_statahead_info *sai, __u64 index)
 182 {
 183         return ((__u64)sai->sai_max + index + SA_OMITTED_ENTRY_MAX <
 184                 sai->sai_index);
 185 }
 186
 187 /* allocate sa_entry and hash it to allow scanner process to find it */
 188 static struct sa_entry *
 189 sa_alloc(struct dentry *parent, struct ll_statahead_info *sai, __u64 index,
 190          const char *name, int len, const struct lu_fid *fid)
 191 {
 192         struct ll_inode_info *lli;
 193         struct sa_entry *entry;
 194         int entry_size;
 195         char *dname;
 196
 197         ENTRY;
 198
 199         entry_size = sizeof(struct sa_entry) + (len & ~3) + 4;
 200         OBD_ALLOC(entry, entry_size);
 201         if (unlikely(!entry))
 202                 RETURN(ERR_PTR(-ENOMEM));
 203
 204         CDEBUG(D_READA, "alloc sa entry %.*s(%p) index %llu\n",
 205                len, name, entry, index);
 206
 207         entry->se_index = index;
 208
 209         entry->se_state = SA_ENTRY_INIT;
 210         entry->se_size = entry_size;
 211         dname = (char *)entry + sizeof(struct sa_entry);
 212         memcpy(dname, name, len);
 213         dname[len] = 0;
 214         entry->se_qstr.hash = ll_full_name_hash(parent, name, len);
 215         entry->se_qstr.len = len;
 216         entry->se_qstr.name = dname;
 217         entry->se_fid = *fid;
 218
 219         lli = ll_i2info(sai->sai_dentry->d_inode);
 220
 221         spin_lock(&lli->lli_sa_lock);
 222         INIT_LIST_HEAD(&entry->se_list);
 223         sa_rehash(sai, entry);
 224         spin_unlock(&lli->lli_sa_lock);
 225
 226         atomic_inc(&sai->sai_cache_count);
 227
 228         RETURN(entry);
 229 }
 230
 231 /* free sa_entry, which should have been unhashed and not in any list */
 232 static void sa_free(struct ll_statahead_info *sai, struct sa_entry *entry)
 233 {
 234         CDEBUG(D_READA, "free sa entry %.*s(%p) index %llu\n",
 235                entry->se_qstr.len, entry->se_qstr.name, entry,
 236                entry->se_index);
 237
 238         LASSERT(list_empty(&entry->se_list));
 239         LASSERT(sa_unhashed(entry));
 240
 241         OBD_FREE(entry, entry->se_size);
 242         atomic_dec(&sai->sai_cache_count);
 243 }
 244
 245 /*
 246  * find sa_entry by name, used by directory scanner, lock is not needed because
 247  * only scanner can remove the entry from cache.
 248  */
 249 static struct sa_entry *
 250 sa_get(struct ll_statahead_info *sai, const struct qstr *qstr)
 251 {
 252         struct sa_entry *entry;
 253         int i = sa_hash(qstr->hash);
 254
 255         list_for_each_entry(entry, &sai->sai_cache[i], se_hash) {
 256                 if (entry->se_qstr.hash == qstr->hash &&
 257                     entry->se_qstr.len == qstr->len &&
 258                     memcmp(entry->se_qstr.name, qstr->name, qstr->len) == 0)
 259                         return entry;
 260         }
 261         return NULL;
 262 }
 263
 264 /* unhash and unlink sa_entry, and then free it */
 265 static inline void
 266 sa_kill(struct ll_statahead_info *sai, struct sa_entry *entry)
 267 {
 268         struct ll_inode_info *lli = ll_i2info(sai->sai_dentry->d_inode);
 269
 270         LASSERT(!sa_unhashed(entry));
 271         LASSERT(!list_empty(&entry->se_list));
 272         LASSERT(sa_ready(entry));
 273
 274         sa_unhash(sai, entry);
 275
 276         spin_lock(&lli->lli_sa_lock);
 277         list_del_init(&entry->se_list);
 278         spin_unlock(&lli->lli_sa_lock);
 279
 280         iput(entry->se_inode);
 281
 282         sa_free(sai, entry);
 283 }
 284
 285 /* called by scanner after use, sa_entry will be killed */
 286 static void
 287 sa_put(struct inode *dir, struct ll_statahead_info *sai, struct sa_entry *entry)
 288 {
 289         struct ll_inode_info *lli = ll_i2info(dir);
 290         struct sa_entry *tmp, *next;
 291         bool wakeup = false;
 292
 293         if (entry && entry->se_state == SA_ENTRY_SUCC) {
 294                 struct ll_sb_info *sbi = ll_i2sbi(sai->sai_dentry->d_inode);
 295
 296                 sai->sai_hit++;
 297                 sai->sai_consecutive_miss = 0;
 298                 if (sai->sai_max < sbi->ll_sa_max) {
 299                         sai->sai_max = min(2 * sai->sai_max, sbi->ll_sa_max);
 300                         wakeup = true;
 301                 } else if (sai->sai_max_batch_count > 0) {
 302                         if (sai->sai_max >= sai->sai_max_batch_count &&
 303                            (sai->sai_index_end - entry->se_index) %
 304                            sai->sai_max_batch_count == 0) {
 305                                 wakeup = true;
 306                         } else if (entry->se_index == sai->sai_index_end) {
 307                                 wakeup = true;
 308                         }
 309                 } else {
 310                         wakeup = true;
 311                 }
 312         } else {
 313                 sai->sai_miss++;
 314                 sai->sai_consecutive_miss++;
 315                 wakeup = true;
 316         }
 317
 318         if (entry)
 319                 sa_kill(sai, entry);
 320
 321         /*
 322          * kill old completed entries, only scanner process does this, no need
 323          * to lock
 324          */
 325         list_for_each_entry_safe(tmp, next, &sai->sai_entries, se_list) {
 326                 if (!is_omitted_entry(sai, tmp->se_index))
 327                         break;
 328                 sa_kill(sai, tmp);
 329         }
 330
 331         spin_lock(&lli->lli_sa_lock);
 332         if (wakeup && sai->sai_task)
 333                 wake_up_process(sai->sai_task);
 334         spin_unlock(&lli->lli_sa_lock);
 335 }
 336
 337 /*
 338  * update state and sort add entry to sai_entries by index, return true if
 339  * scanner is waiting on this entry.
 340  */
 341 static bool
 342 __sa_make_ready(struct ll_statahead_info *sai, struct sa_entry *entry, int ret)
 343 {
 344         struct sa_entry *se;
 345         struct list_head *pos = &sai->sai_entries;
 346         __u64 index = entry->se_index;
 347
 348         LASSERT(!sa_ready(entry));
 349         LASSERT(list_empty(&entry->se_list));
 350
 351         list_for_each_entry_reverse(se, &sai->sai_entries, se_list) {
 352                 if (se->se_index < entry->se_index) {
 353                         pos = &se->se_list;
 354                         break;
 355                 }
 356         }
 357         list_add(&entry->se_list, pos);
 358         /*
 359          * LU-9210: ll_statahead_interpet must be able to see this before
 360          * we wake it up
 361          */
 362         smp_store_release(&entry->se_state,
 363                           ret < 0 ? SA_ENTRY_INVA : SA_ENTRY_SUCC);
 364
 365         return (index == sai->sai_index_wait);
 366 }
 367
 368 /* finish async stat RPC arguments */
 369 static void sa_fini_data(struct md_op_item *item)
 370 {
 371         struct md_op_data *op_data = &item->mop_data;
 372
 373         if (op_data->op_flags & MF_OPNAME_KMALLOCED)
 374                 /* allocated via ll_setup_filename called from sa_prep_data */
 375                 kfree(op_data->op_name);
 376         ll_unlock_md_op_lsm(&item->mop_data);
 377         iput(item->mop_dir);
 378         if (item->mop_subpill_allocated)
 379                 OBD_FREE_PTR(item->mop_pill);
 380         OBD_FREE_PTR(item);
 381 }
 382
 383 static int ll_statahead_interpret(struct md_op_item *item, int rc);
 384
 385 /*
 386  * prepare arguments for async stat RPC.
 387  */
 388 static struct md_op_item *
 389 sa_prep_data(struct inode *dir, struct inode *child, struct sa_entry *entry)
 390 {
 391         struct md_op_item *item;
 392         struct ldlm_enqueue_info *einfo;
 393         struct md_op_data *op_data;
 394
 395         OBD_ALLOC_PTR(item);
 396         if (!item)
 397                 return ERR_PTR(-ENOMEM);
 398
 399         op_data = ll_prep_md_op_data(&item->mop_data, dir, child,
 400                                      entry->se_qstr.name, entry->se_qstr.len, 0,
 401                                      LUSTRE_OPC_ANY, NULL);
 402         if (IS_ERR(op_data)) {
 403                 OBD_FREE_PTR(item);
 404                 return (struct md_op_item *)op_data;
 405         }
 406
 407         if (!child)
 408                 op_data->op_fid2 = entry->se_fid;
 409
 410         item->mop_opc = MD_OP_GETATTR;
 411         item->mop_it.it_op = IT_GETATTR;
 412         item->mop_dir = igrab(dir);
 413         item->mop_cb = ll_statahead_interpret;
 414         item->mop_cbdata = entry;
 415
 416         einfo = &item->mop_einfo;
 417         einfo->ei_type = LDLM_IBITS;
 418         einfo->ei_mode = it_to_lock_mode(&item->mop_it);
 419         einfo->ei_cb_bl = ll_md_blocking_ast;
 420         einfo->ei_cb_cp = ldlm_completion_ast;
 421         einfo->ei_cb_gl = NULL;
 422         einfo->ei_cbdata = NULL;
 423         einfo->ei_req_slot = 1;
 424
 425         return item;
 426 }
 427
 428 /*
 429  * release resources used in async stat RPC, update entry state and wakeup if
 430  * scanner process it waiting on this entry.
 431  */
 432 static void
 433 sa_make_ready(struct ll_statahead_info *sai, struct sa_entry *entry, int ret)
 434 {
 435         struct ll_inode_info *lli = ll_i2info(sai->sai_dentry->d_inode);
 436         bool wakeup;
 437
 438         spin_lock(&lli->lli_sa_lock);
 439         wakeup = __sa_make_ready(sai, entry, ret);
 440         spin_unlock(&lli->lli_sa_lock);
 441
 442         if (wakeup)
 443                 wake_up(&sai->sai_waitq);
 444 }
 445
 446 /* insert inode into the list of sai_agls */
 447 static void ll_agl_add(struct ll_statahead_info *sai,
 448                        struct inode *inode, int index)
 449 {
 450         struct ll_inode_info *child  = ll_i2info(inode);
 451         struct ll_inode_info *parent = ll_i2info(sai->sai_dentry->d_inode);
 452
 453         spin_lock(&child->lli_agl_lock);
 454         if (child->lli_agl_index == 0) {
 455                 child->lli_agl_index = index;
 456                 spin_unlock(&child->lli_agl_lock);
 457
 458                 LASSERT(list_empty(&child->lli_agl_list));
 459
 460                 spin_lock(&parent->lli_agl_lock);
 461                 /* Re-check under the lock */
 462                 if (agl_should_run(sai, inode)) {
 463                         if (agl_list_empty(sai))
 464                                 wake_up_process(sai->sai_agl_task);
 465                         igrab(inode);
 466                         list_add_tail(&child->lli_agl_list, &sai->sai_agls);
 467                 } else
 468                         child->lli_agl_index = 0;
 469                 spin_unlock(&parent->lli_agl_lock);
 470         } else {
 471                 spin_unlock(&child->lli_agl_lock);
 472         }
 473 }
 474
 475 /* allocate sai */
 476 static struct ll_statahead_info *ll_sai_alloc(struct dentry *dentry)
 477 {
 478         struct ll_statahead_info *sai;
 479         struct ll_inode_info *lli = ll_i2info(dentry->d_inode);
 480         int i;
 481
 482         ENTRY;
 483
 484         OBD_ALLOC_PTR(sai);
 485         if (!sai)
 486                 RETURN(NULL);
 487
 488         sai->sai_dentry = dget(dentry);
 489         atomic_set(&sai->sai_refcount, 1);
 490         sai->sai_max = LL_SA_RPC_MIN;
 491         sai->sai_index = 1;
 492         init_waitqueue_head(&sai->sai_waitq);
 493
 494         INIT_LIST_HEAD(&sai->sai_entries);
 495         INIT_LIST_HEAD(&sai->sai_agls);
 496
 497         for (i = 0; i < LL_SA_CACHE_SIZE; i++) {
 498                 INIT_LIST_HEAD(&sai->sai_cache[i]);
 499                 spin_lock_init(&sai->sai_cache_lock[i]);
 500         }
 501         atomic_set(&sai->sai_cache_count, 0);
 502
 503         spin_lock(&sai_generation_lock);
 504         lli->lli_sa_generation = ++sai_generation;
 505         if (unlikely(sai_generation == 0))
 506                 lli->lli_sa_generation = ++sai_generation;
 507         spin_unlock(&sai_generation_lock);
 508
 509         RETURN(sai);
 510 }
 511
 512 /* free sai */
 513 static inline void ll_sai_free(struct ll_statahead_info *sai)
 514 {
 515         LASSERT(sai->sai_dentry != NULL);
 516         dput(sai->sai_dentry);
 517         OBD_FREE_PTR(sai);
 518 }
 519
 520 /*
 521  * take refcount of sai if sai for @dir exists, which means statahead is on for
 522  * this directory.
 523  */
 524 static inline struct ll_statahead_info *ll_sai_get(struct inode *dir)
 525 {
 526         struct ll_inode_info *lli = ll_i2info(dir);
 527         struct ll_statahead_info *sai = NULL;
 528
 529         spin_lock(&lli->lli_sa_lock);
 530         sai = lli->lli_sai;
 531         if (sai)
 532                 atomic_inc(&sai->sai_refcount);
 533         spin_unlock(&lli->lli_sa_lock);
 534
 535         return sai;
 536 }
 537
 538 /*
 539  * put sai refcount after use, if refcount reaches zero, free sai and sa_entries
 540  * attached to it.
 541  */
 542 static void ll_sai_put(struct ll_statahead_info *sai)
 543 {
 544         struct ll_inode_info *lli = ll_i2info(sai->sai_dentry->d_inode);
 545
 546         if (atomic_dec_and_lock(&sai->sai_refcount, &lli->lli_sa_lock)) {
 547                 struct sa_entry *entry, *next;
 548                 struct ll_sb_info *sbi = ll_i2sbi(sai->sai_dentry->d_inode);
 549
 550                 lli->lli_sai = NULL;
 551                 spin_unlock(&lli->lli_sa_lock);
 552
 553                 LASSERT(!sai->sai_task);
 554                 LASSERT(!sai->sai_agl_task);
 555                 LASSERT(sai->sai_sent == sai->sai_replied);
 556
 557                 list_for_each_entry_safe(entry, next, &sai->sai_entries,
 558                                          se_list)
 559                         sa_kill(sai, entry);
 560
 561                 LASSERT(atomic_read(&sai->sai_cache_count) == 0);
 562                 LASSERT(agl_list_empty(sai));
 563
 564                 ll_sai_free(sai);
 565                 atomic_dec(&sbi->ll_sa_running);
 566         }
 567 }
 568
 569 /* Do NOT forget to drop inode refcount when into sai_agls. */
 570 static void ll_agl_trigger(struct inode *inode, struct ll_statahead_info *sai)
 571 {
 572         struct ll_inode_info *lli = ll_i2info(inode);
 573         u64 index = lli->lli_agl_index;
 574         ktime_t expire;
 575         int rc;
 576
 577         ENTRY;
 578
 579         LASSERT(list_empty(&lli->lli_agl_list));
 580
 581         /* AGL maybe fall behind statahead with one entry */
 582         if (is_omitted_entry(sai, index + 1)) {
 583                 lli->lli_agl_index = 0;
 584                 iput(inode);
 585                 RETURN_EXIT;
 586         }
 587
 588         /*
 589          * In case of restore, the MDT has the right size and has already
 590          * sent it back without granting the layout lock, inode is up-to-date.
 591          * Then AGL (async glimpse lock) is useless.
 592          * Also to glimpse we need the layout, in case of a runninh restore
 593          * the MDT holds the layout lock so the glimpse will block up to the
 594          * end of restore (statahead/agl will block)
 595          */
 596         if (test_bit(LLIF_FILE_RESTORING, &lli->lli_flags)) {
 597                 lli->lli_agl_index = 0;
 598                 iput(inode);
 599                 RETURN_EXIT;
 600         }
 601
 602         /* Someone is in glimpse (sync or async), do nothing. */
 603         rc = down_write_trylock(&lli->lli_glimpse_sem);
 604         if (rc == 0) {
 605                 lli->lli_agl_index = 0;
 606                 iput(inode);
 607                 RETURN_EXIT;
 608         }
 609
 610         /*
 611          * Someone triggered glimpse within 1 sec before.
 612          * 1) The former glimpse succeeded with glimpse lock granted by OST, and
 613          *    if the lock is still cached on client, AGL needs to do nothing. If
 614          *    it is cancelled by other client, AGL maybe cannot obtaion new lock
 615          *    for no glimpse callback triggered by AGL.
 616          * 2) The former glimpse succeeded, but OST did not grant glimpse lock.
 617          *    Under such case, it is quite possible that the OST will not grant
 618          *    glimpse lock for AGL also.
 619          * 3) The former glimpse failed, compared with other two cases, it is
 620          *    relative rare. AGL can ignore such case, and it will not muchly
 621          *    affect the performance.
 622          */
 623         expire = ktime_sub_ns(ktime_get(), NSEC_PER_SEC);
 624         if (ktime_to_ns(lli->lli_glimpse_time) &&
 625             ktime_before(expire, lli->lli_glimpse_time)) {
 626                 up_write(&lli->lli_glimpse_sem);
 627                 lli->lli_agl_index = 0;
 628                 iput(inode);
 629                 RETURN_EXIT;
 630         }
 631
 632         CDEBUG(D_READA,
 633                "Handling (init) async glimpse: inode = " DFID", idx = %llu\n",
 634                PFID(&lli->lli_fid), index);
 635
 636         cl_agl(inode);
 637         lli->lli_agl_index = 0;
 638         lli->lli_glimpse_time = ktime_get();
 639         up_write(&lli->lli_glimpse_sem);
 640
 641         CDEBUG(D_READA,
 642                "Handled (init) async glimpse: inode= " DFID", idx = %llu, rc = %d\n",
 643                PFID(&lli->lli_fid), index, rc);
 644
 645         iput(inode);
 646
 647         EXIT;
 648 }
 649
 650 static void ll_statahead_interpret_fini(struct ll_inode_info *lli,
 651                                         struct ll_statahead_info *sai,
 652                                         struct md_op_item *item,
 653                                         struct sa_entry *entry,
 654                                         struct ptlrpc_request *req,
 655                                         int rc)
 656 {
 657         /*
 658          * First it will drop ldlm ibits lock refcount by calling
 659          * ll_intent_drop_lock() in spite of failures. Do not worry about
 660          * calling ll_intent_drop_lock() more than once.
 661          */
 662         ll_intent_release(&item->mop_it);
 663         sa_fini_data(item);
 664         if (req)
 665                 ptlrpc_req_finished(req);
 666         sa_make_ready(sai, entry, rc);
 667
 668         spin_lock(&lli->lli_sa_lock);
 669         sai->sai_replied++;
 670         spin_unlock(&lli->lli_sa_lock);
 671 }
 672
 673 static void ll_statahead_interpret_work(struct work_struct *work)
 674 {
 675         struct md_op_item *item = container_of(work, struct md_op_item,
 676                                                mop_work);
 677         struct req_capsule *pill = item->mop_pill;
 678         struct inode *dir = item->mop_dir;
 679         struct ll_inode_info *lli = ll_i2info(dir);
 680         struct ll_statahead_info *sai = lli->lli_sai;
 681         struct lookup_intent *it;
 682         struct sa_entry *entry;
 683         struct mdt_body *body;
 684         struct inode *child;
 685         int rc;
 686
 687         ENTRY;
 688
 689         entry = (struct sa_entry *)item->mop_cbdata;
 690         LASSERT(entry->se_handle != 0);
 691
 692         it = &item->mop_it;
 693         body = req_capsule_server_get(pill, &RMF_MDT_BODY);
 694         if (!body)
 695                 GOTO(out, rc = -EFAULT);
 696
 697         child = entry->se_inode;
 698         /* revalidate; unlinked and re-created with the same name */
 699         if (unlikely(!lu_fid_eq(&item->mop_data.op_fid2, &body->mbo_fid1))) {
 700                 if (child) {
 701                         entry->se_inode = NULL;
 702                         iput(child);
 703                 }
 704                 /* The mdt_body is invalid. Skip this entry */
 705                 GOTO(out, rc = -EAGAIN);
 706         }
 707
 708         it->it_lock_handle = entry->se_handle;
 709         rc = md_revalidate_lock(ll_i2mdexp(dir), it, ll_inode2fid(dir), NULL);
 710         if (rc != 1)
 711                 GOTO(out, rc = -EAGAIN);
 712
 713         rc = ll_prep_inode(&child, pill, dir->i_sb, it);
 714         if (rc) {
 715                 CERROR("%s: getattr callback for %.*s "DFID": rc = %d\n",
 716                        ll_i2sbi(dir)->ll_fsname, entry->se_qstr.len,
 717                        entry->se_qstr.name, PFID(&entry->se_fid), rc);
 718                 GOTO(out, rc);
 719         }
 720
 721         /* If encryption context was returned by MDT, put it in
 722          * inode now to save an extra getxattr.
 723          */
 724         if (body->mbo_valid & OBD_MD_ENCCTX) {
 725                 void *encctx = req_capsule_server_get(pill, &RMF_FILE_ENCCTX);
 726                 __u32 encctxlen = req_capsule_get_size(pill, &RMF_FILE_ENCCTX,
 727                                                        RCL_SERVER);
 728
 729                 if (encctxlen) {
 730                         CDEBUG(D_SEC,
 731                                "server returned encryption ctx for "DFID"\n",
 732                                PFID(ll_inode2fid(child)));
 733                         rc = ll_xattr_cache_insert(child,
 734                                                    xattr_for_enc(child),
 735                                                    encctx, encctxlen);
 736                         if (rc)
 737                                 CWARN("%s: cannot set enc ctx for "DFID": rc = %d\n",
 738                                       ll_i2sbi(child)->ll_fsname,
 739                                       PFID(ll_inode2fid(child)), rc);
 740                 }
 741         }
 742
 743         CDEBUG(D_READA, "%s: setting %.*s"DFID" l_data to inode %p\n",
 744                ll_i2sbi(dir)->ll_fsname, entry->se_qstr.len,
 745                entry->se_qstr.name, PFID(ll_inode2fid(child)), child);
 746         ll_set_lock_data(ll_i2sbi(dir)->ll_md_exp, child, it, NULL);
 747
 748         entry->se_inode = child;
 749
 750         if (agl_should_run(sai, child))
 751                 ll_agl_add(sai, child, entry->se_index);
 752 out:
 753         ll_statahead_interpret_fini(lli, sai, item, entry, pill->rc_req, rc);
 754 }
 755
 756 /*
 757  * Callback for async stat RPC, this is called in ptlrpcd context. It prepares
 758  * the inode and set lock data directly in the ptlrpcd context. It will wake up
 759  * the directory listing process if the dentry is the waiting one.
 760  */
 761 static int ll_statahead_interpret(struct md_op_item *item, int rc)
 762 {
 763         struct req_capsule *pill = item->mop_pill;
 764         struct lookup_intent *it = &item->mop_it;
 765         struct inode *dir = item->mop_dir;
 766         struct ll_inode_info *lli = ll_i2info(dir);
 767         struct ll_statahead_info *sai = lli->lli_sai;
 768         struct sa_entry *entry = (struct sa_entry *)item->mop_cbdata;
 769         struct work_struct *work = &item->mop_work;
 770         struct mdt_body *body;
 771         struct inode *child;
 772         __u64 handle = 0;
 773
 774         ENTRY;
 775
 776         if (it_disposition(it, DISP_LOOKUP_NEG))
 777                 rc = -ENOENT;
 778
 779         /*
 780          * because statahead thread will wait for all inflight RPC to finish,
 781          * sai should be always valid, no need to refcount
 782          */
 783         LASSERT(sai != NULL);
 784         LASSERT(entry != NULL);
 785
 786         CDEBUG(D_READA, "sa_entry %.*s rc %d\n",
 787                entry->se_qstr.len, entry->se_qstr.name, rc);
 788
 789         if (rc != 0)
 790                 GOTO(out, rc);
 791
 792         body = req_capsule_server_get(pill, &RMF_MDT_BODY);
 793         if (!body)
 794                 GOTO(out, rc = -EFAULT);
 795
 796         child = entry->se_inode;
 797         /* revalidate; unlinked and re-created with the same name */
 798         if (unlikely(!lu_fid_eq(&item->mop_data.op_fid2, &body->mbo_fid1))) {
 799                 if (child) {
 800                         entry->se_inode = NULL;
 801                         iput(child);
 802                 }
 803                 /* The mdt_body is invalid. Skip this entry */
 804                 GOTO(out, rc = -EAGAIN);
 805         }
 806
 807         entry->se_handle = it->it_lock_handle;
 808         /*
 809          * In ptlrpcd context, it is not allowed to generate new RPCs
 810          * especially for striped directories or regular files with layout
 811          * change.
 812          */
 813         /*
 814          * release ibits lock ASAP to avoid deadlock when statahead
 815          * thread enqueues lock on parent in readdir and another
 816          * process enqueues lock on child with parent lock held, eg.
 817          * unlink.
 818          */
 819         handle = it->it_lock_handle;
 820         ll_intent_drop_lock(it);
 821         ll_unlock_md_op_lsm(&item->mop_data);
 822
 823         /*
 824          * If the statahead entry is a striped directory or regular file with
 825          * layout change, it will generate a new RPC and long wait in the
 826          * ptlrpcd context.
 827          * However, it is dangerous of blocking in ptlrpcd thread.
 828          * Here we use work queue or the separate statahead thread to handle
 829          * the extra RPC and long wait:
 830          *      (@ll_prep_inode->@lmv_revalidate_slaves);
 831          *      (@ll_prep_inode->@lov_layout_change->osc_cache_wait_range);
 832          */
 833         INIT_WORK(work, ll_statahead_interpret_work);
 834         ptlrpc_request_addref(pill->rc_req);
 835         schedule_work(work);
 836         RETURN(0);
 837 out:
 838         ll_statahead_interpret_fini(lli, sai, item, entry, NULL, rc);
 839         RETURN(rc);
 840 }
 841
 842 static inline int sa_getattr(struct inode *dir, struct md_op_item *item)
 843 {
 844         struct ll_statahead_info *sai = ll_i2info(dir)->lli_sai;
 845         int rc;
 846
 847         if (sa_has_batch_handle(sai))
 848                 rc = md_batch_add(ll_i2mdexp(dir), sai->sai_bh, item);
 849         else
 850                 rc = md_intent_getattr_async(ll_i2mdexp(dir), item);
 851
 852         return rc;
 853 }
 854
 855 /* async stat for file not found in dcache */
 856 static int sa_lookup(struct inode *dir, struct sa_entry *entry)
 857 {
 858         struct md_op_item *item;
 859         int rc;
 860
 861         ENTRY;
 862
 863         item = sa_prep_data(dir, NULL, entry);
 864         if (IS_ERR(item))
 865                 RETURN(PTR_ERR(item));
 866
 867         rc = sa_getattr(dir, item);
 868         if (rc < 0)
 869                 sa_fini_data(item);
 870
 871         RETURN(rc);
 872 }
 873
 874 /**
 875  * async stat for file found in dcache, similar to .revalidate
 876  *
 877  * \retval      1 dentry valid, no RPC sent
 878  * \retval      0 dentry invalid, will send async stat RPC
 879  * \retval      negative number upon error
 880  */
 881 static int sa_revalidate(struct inode *dir, struct sa_entry *entry,
 882                          struct dentry *dentry)
 883 {
 884         struct inode *inode = dentry->d_inode;
 885         struct lookup_intent it = { .it_op = IT_GETATTR,
 886                                     .it_lock_handle = 0 };
 887         struct md_op_item *item;
 888         int rc;
 889
 890         ENTRY;
 891
 892         if (unlikely(!inode))
 893                 RETURN(1);
 894
 895         if (d_mountpoint(dentry))
 896                 RETURN(1);
 897
 898         item = sa_prep_data(dir, inode, entry);
 899         if (IS_ERR(item))
 900                 RETURN(PTR_ERR(item));
 901
 902         entry->se_inode = igrab(inode);
 903         rc = md_revalidate_lock(ll_i2mdexp(dir), &it, ll_inode2fid(inode),
 904                                 NULL);
 905         if (rc == 1) {
 906                 entry->se_handle = it.it_lock_handle;
 907                 ll_intent_release(&it);
 908                 sa_fini_data(item);
 909                 RETURN(1);
 910         }
 911
 912         rc = sa_getattr(dir, item);
 913         if (rc < 0) {
 914                 entry->se_inode = NULL;
 915                 iput(inode);
 916                 sa_fini_data(item);
 917         }
 918
 919         RETURN(rc);
 920 }
 921
 922 /* async stat for file with @name */
 923 static void sa_statahead(struct dentry *parent, const char *name, int len,
 924                          const struct lu_fid *fid)
 925 {
 926         struct inode *dir = parent->d_inode;
 927         struct ll_inode_info *lli = ll_i2info(dir);
 928         struct ll_statahead_info *sai = lli->lli_sai;
 929         struct dentry *dentry = NULL;
 930         struct sa_entry *entry;
 931         int rc;
 932
 933         ENTRY;
 934
 935         entry = sa_alloc(parent, sai, sai->sai_index, name, len, fid);
 936         if (IS_ERR(entry))
 937                 RETURN_EXIT;
 938
 939         dentry = d_lookup(parent, &entry->se_qstr);
 940         if (!dentry) {
 941                 rc = sa_lookup(dir, entry);
 942         } else {
 943                 rc = sa_revalidate(dir, entry, dentry);
 944                 if (rc == 1 && agl_should_run(sai, dentry->d_inode))
 945                         ll_agl_add(sai, dentry->d_inode, entry->se_index);
 946         }
 947
 948         if (dentry)
 949                 dput(dentry);
 950
 951         if (rc != 0)
 952                 sa_make_ready(sai, entry, rc);
 953         else
 954                 sai->sai_sent++;
 955
 956         sai->sai_index++;
 957
 958         if (sa_sent_full(sai))
 959                 ll_statahead_flush_nowait(sai);
 960
 961         EXIT;
 962 }
 963
 964 /* async glimpse (agl) thread main function */
 965 static int ll_agl_thread(void *arg)
 966 {
 967         struct dentry *parent = (struct dentry *)arg;
 968         struct inode *dir = parent->d_inode;
 969         struct ll_inode_info *plli = ll_i2info(dir);
 970         struct ll_inode_info *clli;
 971         /*
 972          * We already own this reference, so it is safe to take it
 973          * without a lock.
 974          */
 975         struct ll_statahead_info *sai = plli->lli_sai;
 976
 977         ENTRY;
 978
 979         CDEBUG(D_READA, "agl thread started: sai %p, parent %pd\n",
 980                sai, parent);
 981
 982         while (({set_current_state(TASK_IDLE);
 983                  !kthread_should_stop(); })) {
 984                 spin_lock(&plli->lli_agl_lock);
 985                 clli = list_first_entry_or_null(&sai->sai_agls,
 986                                                 struct ll_inode_info,
 987                                                 lli_agl_list);
 988                 if (clli) {
 989                         __set_current_state(TASK_RUNNING);
 990                         list_del_init(&clli->lli_agl_list);
 991                         spin_unlock(&plli->lli_agl_lock);
 992                         ll_agl_trigger(&clli->lli_vfs_inode, sai);
 993                         cond_resched();
 994                 } else {
 995                         spin_unlock(&plli->lli_agl_lock);
 996                         schedule();
 997                 }
 998         }
 999         __set_current_state(TASK_RUNNING);
1000         RETURN(0);
1001 }
1002
1003 static void ll_stop_agl(struct ll_statahead_info *sai)
1004 {
1005         struct dentry *parent = sai->sai_dentry;
1006         struct ll_inode_info *plli = ll_i2info(parent->d_inode);
1007         struct ll_inode_info *clli;
1008         struct task_struct *agl_task;
1009
1010         spin_lock(&plli->lli_agl_lock);
1011         agl_task = sai->sai_agl_task;
1012         sai->sai_agl_task = NULL;
1013         spin_unlock(&plli->lli_agl_lock);
1014         if (!agl_task)
1015                 return;
1016
1017         CDEBUG(D_READA, "stop agl thread: sai %p pid %u\n",
1018                sai, (unsigned int)agl_task->pid);
1019         kthread_stop(agl_task);
1020
1021         spin_lock(&plli->lli_agl_lock);
1022         while ((clli = list_first_entry_or_null(&sai->sai_agls,
1023                                                 struct ll_inode_info,
1024                                                 lli_agl_list)) != NULL) {
1025                 list_del_init(&clli->lli_agl_list);
1026                 spin_unlock(&plli->lli_agl_lock);
1027                 clli->lli_agl_index = 0;
1028                 iput(&clli->lli_vfs_inode);
1029                 spin_lock(&plli->lli_agl_lock);
1030         }
1031         spin_unlock(&plli->lli_agl_lock);
1032         CDEBUG(D_READA, "agl thread stopped: sai %p, parent %pd\n",
1033                sai, parent);
1034         ll_sai_put(sai);
1035 }
1036
1037 /* start agl thread */
1038 static void ll_start_agl(struct dentry *parent, struct ll_statahead_info *sai)
1039 {
1040         int node = cfs_cpt_spread_node(cfs_cpt_tab, CFS_CPT_ANY);
1041         struct ll_inode_info *plli;
1042         struct task_struct *task;
1043
1044         ENTRY;
1045
1046         CDEBUG(D_READA, "start agl thread: sai %p, parent %pd\n",
1047                sai, parent);
1048
1049         plli = ll_i2info(parent->d_inode);
1050         task = kthread_create_on_node(ll_agl_thread, parent, node, "ll_agl_%d",
1051                                       plli->lli_opendir_pid);
1052         if (IS_ERR(task)) {
1053                 CERROR("can't start ll_agl thread, rc: %ld\n", PTR_ERR(task));
1054                 RETURN_EXIT;
1055         }
1056         sai->sai_agl_task = task;
1057         atomic_inc(&ll_i2sbi(d_inode(parent))->ll_agl_total);
1058         /* Get an extra reference that the thread holds */
1059         ll_sai_get(d_inode(parent));
1060
1061         wake_up_process(task);
1062
1063         EXIT;
1064 }
1065
1066 /* statahead thread main function */
1067 static int ll_statahead_thread(void *arg)
1068 {
1069         struct dentry *parent = (struct dentry *)arg;
1070         struct inode *dir = parent->d_inode;
1071         struct ll_inode_info *lli = ll_i2info(dir);
1072         struct ll_sb_info *sbi = ll_i2sbi(dir);
1073         struct ll_statahead_info *sai = lli->lli_sai;
1074         int first = 0;
1075         struct md_op_data *op_data;
1076         struct page *page = NULL;
1077         struct lu_batch *bh = NULL;
1078         __u64 pos = 0;
1079         int rc = 0;
1080
1081         ENTRY;
1082
1083         CDEBUG(D_READA, "statahead thread starting: sai %p, parent %pd\n",
1084                sai, parent);
1085
1086         sai->sai_max_batch_count = sbi->ll_sa_batch_max;
1087         if (sai->sai_max_batch_count) {
1088                 bh = md_batch_create(ll_i2mdexp(dir), BATCH_FL_RDONLY,
1089                                      sai->sai_max_batch_count);
1090                 if (IS_ERR(bh))
1091                         GOTO(out_stop_agl, rc = PTR_ERR(bh));
1092         }
1093
1094         sai->sai_bh = bh;
1095         OBD_ALLOC_PTR(op_data);
1096         if (!op_data)
1097                 GOTO(out, rc = -ENOMEM);
1098
1099         /* matches smp_store_release() in ll_deauthorize_statahead() */
1100         while (pos != MDS_DIR_END_OFF && smp_load_acquire(&sai->sai_task)) {
1101                 struct lu_dirpage *dp;
1102                 struct lu_dirent  *ent;
1103
1104                 op_data = ll_prep_md_op_data(op_data, dir, dir, NULL, 0, 0,
1105                                              LUSTRE_OPC_ANY, dir);
1106                 if (IS_ERR(op_data)) {
1107                         rc = PTR_ERR(op_data);
1108                         break;
1109                 }
1110
1111                 page = ll_get_dir_page(dir, op_data, pos, NULL);
1112                 ll_unlock_md_op_lsm(op_data);
1113                 if (IS_ERR(page)) {
1114                         rc = PTR_ERR(page);
1115                         CDEBUG(D_READA,
1116                                "error reading dir "DFID" at %llu /%llu opendir_pid = %u: rc = %d\n",
1117                                PFID(ll_inode2fid(dir)), pos, sai->sai_index,
1118                                lli->lli_opendir_pid, rc);
1119                         break;
1120                 }
1121
1122                 dp = page_address(page);
1123                 for (ent = lu_dirent_start(dp);
1124                      /* matches smp_store_release() in ll_deauthorize_statahead() */
1125                      ent != NULL && smp_load_acquire(&sai->sai_task) &&
1126                      !sa_low_hit(sai);
1127                      ent = lu_dirent_next(ent)) {
1128                         __u64 hash;
1129                         int namelen;
1130                         char *name;
1131                         struct lu_fid fid;
1132                         struct llcrypt_str lltr = LLTR_INIT(NULL, 0);
1133
1134                         hash = le64_to_cpu(ent->lde_hash);
1135                         if (unlikely(hash < pos))
1136                                 /*
1137                                  * Skip until we find target hash value.
1138                                  */
1139                                 continue;
1140
1141                         namelen = le16_to_cpu(ent->lde_namelen);
1142                         if (unlikely(namelen == 0))
1143                                 /*
1144                                  * Skip dummy record.
1145                                  */
1146                                 continue;
1147
1148                         name = ent->lde_name;
1149                         if (name[0] == '.') {
1150                                 if (namelen == 1) {
1151                                         /*
1152                                          * skip "."
1153                                          */
1154                                         continue;
1155                                 } else if (name[1] == '.' && namelen == 2) {
1156                                         /*
1157                                          * skip ".."
1158                                          */
1159                                         continue;
1160                                 } else if (!sai->sai_ls_all) {
1161                                         /*
1162                                          * skip hidden files.
1163                                          */
1164                                         sai->sai_skip_hidden++;
1165                                         continue;
1166                                 }
1167                         }
1168
1169                         /*
1170                          * don't stat-ahead first entry.
1171                          */
1172                         if (unlikely(++first == 1))
1173                                 continue;
1174
1175                         fid_le_to_cpu(&fid, &ent->lde_fid);
1176
1177                         while (({set_current_state(TASK_IDLE);
1178                                  /* matches smp_store_release() in
1179                                   * ll_deauthorize_statahead() */
1180                                  smp_load_acquire(&sai->sai_task); })) {
1181                                 spin_lock(&lli->lli_agl_lock);
1182                                 while (sa_sent_full(sai) &&
1183                                        !agl_list_empty(sai)) {
1184                                         struct ll_inode_info *clli;
1185
1186                                         __set_current_state(TASK_RUNNING);
1187                                         clli = agl_first_entry(sai);
1188                                         list_del_init(&clli->lli_agl_list);
1189                                         spin_unlock(&lli->lli_agl_lock);
1190
1191                                         ll_agl_trigger(&clli->lli_vfs_inode,
1192                                                        sai);
1193                                         cond_resched();
1194                                         spin_lock(&lli->lli_agl_lock);
1195                                 }
1196                                 spin_unlock(&lli->lli_agl_lock);
1197
1198                                 if (!sa_sent_full(sai))
1199                                         break;
1200                                 schedule();
1201                         }
1202                         __set_current_state(TASK_RUNNING);
1203
1204                         if (IS_ENCRYPTED(dir)) {
1205                                 struct llcrypt_str de_name =
1206                                         LLTR_INIT(ent->lde_name, namelen);
1207                                 struct lu_fid fid;
1208
1209                                 rc = llcrypt_fname_alloc_buffer(dir, NAME_MAX,
1210                                                                 &lltr);
1211                                 if (rc < 0)
1212                                         continue;
1213
1214                                 fid_le_to_cpu(&fid, &ent->lde_fid);
1215                                 if (ll_fname_disk_to_usr(dir, 0, 0, &de_name,
1216                                                          &lltr, &fid)) {
1217                                         llcrypt_fname_free_buffer(&lltr);
1218                                         continue;
1219                                 }
1220
1221                                 name = lltr.name;
1222                                 namelen = lltr.len;
1223                         }
1224
1225                         sa_statahead(parent, name, namelen, &fid);
1226                         llcrypt_fname_free_buffer(&lltr);
1227                 }
1228
1229                 pos = le64_to_cpu(dp->ldp_hash_end);
1230                 down_read(&lli->lli_lsm_sem);
1231                 ll_release_page(dir, page,
1232                                 le32_to_cpu(dp->ldp_flags) & LDF_COLLIDE);
1233                 up_read(&lli->lli_lsm_sem);
1234
1235                 if (sa_low_hit(sai)) {
1236                         rc = -EFAULT;
1237                         atomic_inc(&sbi->ll_sa_wrong);
1238                         CDEBUG(D_READA,
1239                                "Statahead for dir "DFID" hit ratio too low: hit/miss %llu/%llu, sent/replied %llu/%llu, stoppingstatahead thread: pid %d\n",
1240                                PFID(&lli->lli_fid), sai->sai_hit,
1241                                sai->sai_miss, sai->sai_sent,
1242                                sai->sai_replied, current->pid);
1243                         break;
1244                 }
1245         }
1246         ll_finish_md_op_data(op_data);
1247
1248         if (rc < 0) {
1249                 spin_lock(&lli->lli_sa_lock);
1250                 sai->sai_task = NULL;
1251                 lli->lli_sa_enabled = 0;
1252                 spin_unlock(&lli->lli_sa_lock);
1253         }
1254
1255         ll_statahead_flush_nowait(sai);
1256
1257         /*
1258          * statahead is finished, but statahead entries need to be cached, wait
1259          * for file release closedir() call to stop me.
1260          */
1261         while (({set_current_state(TASK_IDLE);
1262                 /* matches smp_store_release() in ll_deauthorize_statahead() */
1263                 smp_load_acquire(&sai->sai_task); })) {
1264                 schedule();
1265         }
1266         __set_current_state(TASK_RUNNING);
1267
1268         EXIT;
1269 out:
1270         if (bh) {
1271                 rc = md_batch_stop(ll_i2mdexp(dir), sai->sai_bh);
1272                 sai->sai_bh = NULL;
1273         }
1274
1275 out_stop_agl:
1276         ll_stop_agl(sai);
1277
1278         /*
1279          * wait for inflight statahead RPCs to finish, and then we can free sai
1280          * safely because statahead RPC will access sai data
1281          */
1282         while (sai->sai_sent != sai->sai_replied)
1283                 /* in case we're not woken up, timeout wait */
1284                 msleep(125);
1285
1286         CDEBUG(D_READA, "%s: statahead thread stopped: sai %p, parent %pd\n",
1287                sbi->ll_fsname, sai, parent);
1288
1289         spin_lock(&lli->lli_sa_lock);
1290         sai->sai_task = NULL;
1291         spin_unlock(&lli->lli_sa_lock);
1292         wake_up(&sai->sai_waitq);
1293
1294         atomic_add(sai->sai_hit, &sbi->ll_sa_hit_total);
1295         atomic_add(sai->sai_miss, &sbi->ll_sa_miss_total);
1296
1297         ll_sai_put(sai);
1298
1299         return rc;
1300 }
1301
1302 /* authorize opened dir handle @key to statahead */
1303 void ll_authorize_statahead(struct inode *dir, void *key)
1304 {
1305         struct ll_inode_info *lli = ll_i2info(dir);
1306
1307         spin_lock(&lli->lli_sa_lock);
1308         if (!lli->lli_opendir_key && !lli->lli_sai) {
1309                 /*
1310                  * if lli_sai is not NULL, it means previous statahead is not
1311                  * finished yet, we'd better not start a new statahead for now.
1312                  */
1313                 LASSERT(lli->lli_opendir_pid == 0);
1314                 lli->lli_opendir_key = key;
1315                 lli->lli_opendir_pid = current->pid;
1316                 lli->lli_sa_enabled = 1;
1317         }
1318         spin_unlock(&lli->lli_sa_lock);
1319 }
1320
1321 /*
1322  * deauthorize opened dir handle @key to statahead, and notify statahead thread
1323  * to quit if it's running.
1324  */
1325 void ll_deauthorize_statahead(struct inode *dir, void *key)
1326 {
1327         struct ll_inode_info *lli = ll_i2info(dir);
1328         struct ll_statahead_info *sai;
1329
1330         LASSERT(lli->lli_opendir_key == key);
1331         LASSERT(lli->lli_opendir_pid != 0);
1332
1333         CDEBUG(D_READA, "deauthorize statahead for "DFID"\n",
1334                PFID(&lli->lli_fid));
1335
1336         spin_lock(&lli->lli_sa_lock);
1337         lli->lli_opendir_key = NULL;
1338         lli->lli_opendir_pid = 0;
1339         lli->lli_sa_enabled = 0;
1340         sai = lli->lli_sai;
1341         if (sai && sai->sai_task) {
1342                 /*
1343                  * statahead thread may not have quit yet because it needs to
1344                  * cache entries, now it's time to tell it to quit.
1345                  *
1346                  * wake_up_process() provides the necessary barriers
1347                  * to pair with set_current_state().
1348                  */
1349                 struct task_struct *task = sai->sai_task;
1350
1351                 /* matches smp_load_acquire() in ll_statahead_thread() */
1352                 smp_store_release(&sai->sai_task, NULL);
1353                 wake_up_process(task);
1354         }
1355         spin_unlock(&lli->lli_sa_lock);
1356 }
1357
1358 enum {
1359         /**
1360          * not first dirent, or is "."
1361          */
1362         LS_NOT_FIRST_DE = 0,
1363         /**
1364          * the first non-hidden dirent
1365          */
1366         LS_FIRST_DE,
1367         /**
1368          * the first hidden dirent, that is "."
1369          */
1370         LS_FIRST_DOT_DE
1371 };
1372
1373 /* file is first dirent under @dir */
1374 static int is_first_dirent(struct inode *dir, struct dentry *dentry)
1375 {
1376         struct qstr *target = &dentry->d_name;
1377         struct md_op_data *op_data;
1378         int dot_de;
1379         struct page *page = NULL;
1380         int rc = LS_NOT_FIRST_DE;
1381         __u64 pos = 0;
1382         struct llcrypt_str lltr = LLTR_INIT(NULL, 0);
1383
1384         ENTRY;
1385
1386         op_data = ll_prep_md_op_data(NULL, dir, dir, NULL, 0, 0,
1387                                      LUSTRE_OPC_ANY, dir);
1388         if (IS_ERR(op_data))
1389                 RETURN(PTR_ERR(op_data));
1390
1391         if (IS_ENCRYPTED(dir)) {
1392                 int rc2 = llcrypt_fname_alloc_buffer(dir, NAME_MAX, &lltr);
1393
1394                 if (rc2 < 0)
1395                         RETURN(rc2);
1396         }
1397
1398         /**
1399          *FIXME choose the start offset of the readdir
1400          */
1401
1402         page = ll_get_dir_page(dir, op_data, 0, NULL);
1403
1404         while (1) {
1405                 struct lu_dirpage *dp;
1406                 struct lu_dirent  *ent;
1407
1408                 if (IS_ERR(page)) {
1409                         struct ll_inode_info *lli = ll_i2info(dir);
1410
1411                         rc = PTR_ERR(page);
1412                         CERROR("%s: reading dir "DFID" at %llu opendir_pid = %u : rc = %d\n",
1413                                ll_i2sbi(dir)->ll_fsname,
1414                                PFID(ll_inode2fid(dir)), pos,
1415                                lli->lli_opendir_pid, rc);
1416                         break;
1417                 }
1418
1419                 dp = page_address(page);
1420                 for (ent = lu_dirent_start(dp); ent != NULL;
1421                      ent = lu_dirent_next(ent)) {
1422                         __u64 hash;
1423                         int namelen;
1424                         char *name;
1425
1426                         hash = le64_to_cpu(ent->lde_hash);
1427                         /*
1428                          * The ll_get_dir_page() can return any page containing
1429                          * the given hash which may be not the start hash.
1430                          */
1431                         if (unlikely(hash < pos))
1432                                 continue;
1433
1434                         namelen = le16_to_cpu(ent->lde_namelen);
1435                         if (unlikely(namelen == 0))
1436                                 /*
1437                                  * skip dummy record.
1438                                  */
1439                                 continue;
1440
1441                         name = ent->lde_name;
1442                         if (name[0] == '.') {
1443                                 if (namelen == 1)
1444                                         /*
1445                                          * skip "."
1446                                          */
1447                                         continue;
1448                                 else if (name[1] == '.' && namelen == 2)
1449                                         /*
1450                                          * skip ".."
1451                                          */
1452                                         continue;
1453                                 else
1454                                         dot_de = 1;
1455                         } else {
1456                                 dot_de = 0;
1457                         }
1458
1459                         if (dot_de && target->name[0] != '.') {
1460                                 CDEBUG(D_READA, "%.*s skip hidden file %.*s\n",
1461                                        target->len, target->name,
1462                                        namelen, name);
1463                                 continue;
1464                         }
1465
1466                         if (IS_ENCRYPTED(dir)) {
1467                                 struct llcrypt_str de_name =
1468                                         LLTR_INIT(ent->lde_name, namelen);
1469                                 struct lu_fid fid;
1470
1471                                 fid_le_to_cpu(&fid, &ent->lde_fid);
1472                                 if (ll_fname_disk_to_usr(dir, 0, 0, &de_name,
1473                                                          &lltr, &fid))
1474                                         continue;
1475                                 name = lltr.name;
1476                                 namelen = lltr.len;
1477                         }
1478
1479                         if (target->len != namelen ||
1480                             memcmp(target->name, name, namelen) != 0)
1481                                 rc = LS_NOT_FIRST_DE;
1482                         else if (!dot_de)
1483                                 rc = LS_FIRST_DE;
1484                         else
1485                                 rc = LS_FIRST_DOT_DE;
1486
1487                         ll_release_page(dir, page, false);
1488                         GOTO(out, rc);
1489                 }
1490                 pos = le64_to_cpu(dp->ldp_hash_end);
1491                 if (pos == MDS_DIR_END_OFF) {
1492                         /*
1493                          * End of directory reached.
1494                          */
1495                         ll_release_page(dir, page, false);
1496                         GOTO(out, rc);
1497                 } else {
1498                         /*
1499                          * chain is exhausted
1500                          * Normal case: continue to the next page.
1501                          */
1502                         ll_release_page(dir, page, le32_to_cpu(dp->ldp_flags) &
1503                                               LDF_COLLIDE);
1504                         page = ll_get_dir_page(dir, op_data, pos, NULL);
1505                 }
1506         }
1507         EXIT;
1508 out:
1509         llcrypt_fname_free_buffer(&lltr);
1510         ll_finish_md_op_data(op_data);
1511
1512         return rc;
1513 }
1514
1515 /**
1516  * revalidate @dentryp from statahead cache
1517  *
1518  * \param[in] dir       parent directory
1519  * \param[in] sai       sai structure
1520  * \param[out] dentryp  pointer to dentry which will be revalidated
1521  * \param[in] unplug    unplug statahead window only (normally for negative
1522  *                      dentry)
1523  * \retval              1 on success, dentry is saved in @dentryp
1524  * \retval              0 if revalidation failed (no proper lock on client)
1525  * \retval              negative number upon error
1526  */
1527 static int revalidate_statahead_dentry(struct inode *dir,
1528                                        struct ll_statahead_info *sai,
1529                                        struct dentry **dentryp,
1530                                        bool unplug)
1531 {
1532         struct sa_entry *entry = NULL;
1533         struct ll_inode_info *lli = ll_i2info(dir);
1534         int rc = 0;
1535
1536         ENTRY;
1537
1538         if ((*dentryp)->d_name.name[0] == '.') {
1539                 if (sai->sai_ls_all ||
1540                     sai->sai_miss_hidden >= sai->sai_skip_hidden) {
1541                         /*
1542                          * Hidden dentry is the first one, or statahead
1543                          * thread does not skip so many hidden dentries
1544                          * before "sai_ls_all" enabled as below.
1545                          */
1546                 } else {
1547                         if (!sai->sai_ls_all)
1548                                 /*
1549                                  * It maybe because hidden dentry is not
1550                                  * the first one, "sai_ls_all" was not
1551                                  * set, then "ls -al" missed. Enable
1552                                  * "sai_ls_all" for such case.
1553                                  */
1554                                 sai->sai_ls_all = 1;
1555
1556                         /*
1557                          * Such "getattr" has been skipped before
1558                          * "sai_ls_all" enabled as above.
1559                          */
1560                         sai->sai_miss_hidden++;
1561                         RETURN(-EAGAIN);
1562                 }
1563         }
1564
1565         if (unplug)
1566                 GOTO(out, rc = 1);
1567
1568         entry = sa_get(sai, &(*dentryp)->d_name);
1569         if (!entry)
1570                 GOTO(out, rc = -EAGAIN);
1571
1572         if (!sa_ready(entry)) {
1573                 spin_lock(&lli->lli_sa_lock);
1574                 sai->sai_index_wait = entry->se_index;
1575                 spin_unlock(&lli->lli_sa_lock);
1576                 rc = wait_event_idle_timeout(sai->sai_waitq, sa_ready(entry),
1577                                              cfs_time_seconds(30));
1578                 if (rc == 0) {
1579                         /*
1580                          * entry may not be ready, so it may be used by inflight
1581                          * statahead RPC, don't free it.
1582                          */
1583                         entry = NULL;
1584                         GOTO(out, rc = -EAGAIN);
1585                 }
1586         }
1587
1588         /*
1589          * We need to see the value that was set immediately before we
1590          * were woken up.
1591          */
1592         if (smp_load_acquire(&entry->se_state) == SA_ENTRY_SUCC &&
1593             entry->se_inode) {
1594                 struct inode *inode = entry->se_inode;
1595                 struct lookup_intent it = { .it_op = IT_GETATTR,
1596                                             .it_lock_handle =
1597                                                 entry->se_handle };
1598                 __u64 bits;
1599
1600                 rc = md_revalidate_lock(ll_i2mdexp(dir), &it,
1601                                         ll_inode2fid(inode), &bits);
1602                 if (rc == 1) {
1603                         if (!(*dentryp)->d_inode) {
1604                                 struct dentry *alias;
1605
1606                                 alias = ll_splice_alias(inode, *dentryp);
1607                                 if (IS_ERR(alias)) {
1608                                         ll_intent_release(&it);
1609                                         GOTO(out, rc = PTR_ERR(alias));
1610                                 }
1611                                 *dentryp = alias;
1612                                 /*
1613                                  * statahead prepared this inode, transfer inode
1614                                  * refcount from sa_entry to dentry
1615                                  */
1616                                 entry->se_inode = NULL;
1617                         } else if ((*dentryp)->d_inode != inode) {
1618                                 /* revalidate, but inode is recreated */
1619                                 CDEBUG(D_READA,
1620                                        "%s: stale dentry %pd inode " DFID", statahead inode "DFID "\n",
1621                                        ll_i2sbi(inode)->ll_fsname, *dentryp,
1622                                        PFID(ll_inode2fid((*dentryp)->d_inode)),
1623                                        PFID(ll_inode2fid(inode)));
1624                                 ll_intent_release(&it);
1625                                 GOTO(out, rc = -ESTALE);
1626                         }
1627
1628                         if (bits & MDS_INODELOCK_LOOKUP) {
1629                                 d_lustre_revalidate(*dentryp);
1630                                 if (S_ISDIR(inode->i_mode))
1631                                         ll_update_dir_depth_dmv(dir, *dentryp);
1632                         }
1633
1634                         ll_intent_release(&it);
1635                 }
1636         }
1637 out:
1638         /*
1639          * statahead cached sa_entry can be used only once, and will be killed
1640          * right after use, so if lookup/revalidate accessed statahead cache,
1641          * set dentry ldd_sa_generation to parent lli_sa_generation, later if we
1642          * stat this file again, we know we've done statahead before, see
1643          * dentry_may_statahead().
1644          */
1645         if (lld_is_init(*dentryp))
1646                 ll_d2d(*dentryp)->lld_sa_generation = lli->lli_sa_generation;
1647         sa_put(dir, sai, entry);
1648
1649         RETURN(rc);
1650 }
1651
1652 /**
1653  * start statahead thread
1654  *
1655  * \param[in] dir       parent directory
1656  * \param[in] dentry    dentry that triggers statahead, normally the first
1657  *                      dirent under @dir
1658  * \param[in] agl       indicate whether AGL is needed
1659  * \retval              -EAGAIN on success, because when this function is
1660  *                      called, it's already in lookup call, so client should
1661  *                      do it itself instead of waiting for statahead thread
1662  *                      to do it asynchronously.
1663  * \retval              negative number upon error
1664  */
1665 static int start_statahead_thread(struct inode *dir, struct dentry *dentry,
1666                                   bool agl)
1667 {
1668         int node = cfs_cpt_spread_node(cfs_cpt_tab, CFS_CPT_ANY);
1669         struct ll_inode_info *lli = ll_i2info(dir);
1670         struct ll_statahead_info *sai = NULL;
1671         struct dentry *parent = dentry->d_parent;
1672         struct task_struct *task;
1673         struct ll_sb_info *sbi = ll_i2sbi(parent->d_inode);
1674         int first = LS_FIRST_DE;
1675         int rc = 0;
1676
1677         ENTRY;
1678
1679         /* I am the "lli_opendir_pid" owner, only me can set "lli_sai". */
1680         first = is_first_dirent(dir, dentry);
1681         if (first == LS_NOT_FIRST_DE)
1682                 /* It is not "ls -{a}l" operation, no need statahead for it. */
1683                 GOTO(out, rc = -EFAULT);
1684
1685         if (unlikely(atomic_inc_return(&sbi->ll_sa_running) >
1686                                        sbi->ll_sa_running_max)) {
1687                 CDEBUG(D_READA,
1688                        "Too many concurrent statahead instances, avoid new statahead instance temporarily.\n");
1689                 GOTO(out, rc = -EMFILE);
1690         }
1691
1692         sai = ll_sai_alloc(parent);
1693         if (!sai)
1694                 GOTO(out, rc = -ENOMEM);
1695
1696         sai->sai_ls_all = (first == LS_FIRST_DOT_DE);
1697
1698         /*
1699          * if current lli_opendir_key was deauthorized, or dir re-opened by
1700          * another process, don't start statahead, otherwise the newly spawned
1701          * statahead thread won't be notified to quit.
1702          */
1703         spin_lock(&lli->lli_sa_lock);
1704         if (unlikely(lli->lli_sai || !lli->lli_opendir_key ||
1705                      lli->lli_opendir_pid != current->pid)) {
1706                 spin_unlock(&lli->lli_sa_lock);
1707                 GOTO(out, rc = -EPERM);
1708         }
1709         lli->lli_sai = sai;
1710         spin_unlock(&lli->lli_sa_lock);
1711
1712         CDEBUG(D_READA, "start statahead thread: [pid %d] [parent %pd]\n",
1713                current->pid, parent);
1714
1715         task = kthread_create_on_node(ll_statahead_thread, parent, node,
1716                                       "ll_sa_%u", lli->lli_opendir_pid);
1717         if (IS_ERR(task)) {
1718                 spin_lock(&lli->lli_sa_lock);
1719                 lli->lli_sai = NULL;
1720                 spin_unlock(&lli->lli_sa_lock);
1721                 rc = PTR_ERR(task);
1722                 CERROR("can't start ll_sa thread, rc: %d\n", rc);
1723                 GOTO(out, rc);
1724         }
1725
1726         if (test_bit(LL_SBI_AGL_ENABLED, sbi->ll_flags) && agl)
1727                 ll_start_agl(parent, sai);
1728
1729         atomic_inc(&sbi->ll_sa_total);
1730         sai->sai_task = task;
1731
1732         wake_up_process(task);
1733         /*
1734          * We don't stat-ahead for the first dirent since we are already in
1735          * lookup.
1736          */
1737         RETURN(-EAGAIN);
1738
1739 out:
1740         /*
1741          * once we start statahead thread failed, disable statahead so that
1742          * subsequent stat won't waste time to try it.
1743          */
1744         spin_lock(&lli->lli_sa_lock);
1745         if (lli->lli_opendir_pid == current->pid)
1746                 lli->lli_sa_enabled = 0;
1747         spin_unlock(&lli->lli_sa_lock);
1748
1749         if (sai)
1750                 ll_sai_free(sai);
1751         if (first != LS_NOT_FIRST_DE)
1752                 atomic_dec(&sbi->ll_sa_running);
1753
1754         RETURN(rc);
1755 }
1756
1757 /*
1758  * Check whether statahead for @dir was started.
1759  */
1760 static inline bool ll_statahead_started(struct inode *dir, bool agl)
1761 {
1762         struct ll_inode_info *lli = ll_i2info(dir);
1763         struct ll_statahead_info *sai;
1764
1765         spin_lock(&lli->lli_sa_lock);
1766         sai = lli->lli_sai;
1767         if (sai && (sai->sai_agl_task != NULL) != agl)
1768                 CDEBUG(D_READA,
1769                        "%s: Statahead AGL hint changed from %d to %d\n",
1770                        ll_i2sbi(dir)->ll_fsname,
1771                        sai->sai_agl_task != NULL, agl);
1772         spin_unlock(&lli->lli_sa_lock);
1773
1774         return !!sai;
1775 }
1776
1777 /**
1778  * statahead entry function, this is called when client getattr on a file, it
1779  * will start statahead thread if this is the first dir entry, else revalidate
1780  * dentry from statahead cache.
1781  *
1782  * \param[in]  dir      parent directory
1783  * \param[out] dentryp  dentry to getattr
1784  * \param[in]  agl      whether start the agl thread
1785  *
1786  * \retval              1 on success
1787  * \retval              0 revalidation from statahead cache failed, caller needs
1788  *                      to getattr from server directly
1789  * \retval              negative number on error, caller often ignores this and
1790  *                      then getattr from server
1791  */
1792 int ll_start_statahead(struct inode *dir, struct dentry *dentry, bool agl)
1793 {
1794         if (!ll_statahead_started(dir, agl))
1795                 return start_statahead_thread(dir, dentry, agl);
1796         return 0;
1797 }
1798
1799 /**
1800  * revalidate dentry from statahead cache.
1801  *
1802  * \param[in]  dir      parent directory
1803  * \param[out] dentryp  dentry to getattr
1804  * \param[in]  unplug   unplug statahead window only (normally for negative
1805  *                      dentry)
1806  * \retval              1 on success
1807  * \retval              0 revalidation from statahead cache failed, caller needs
1808  *                      to getattr from server directly
1809  * \retval              negative number on error, caller often ignores this and
1810  *                      then getattr from server
1811  */
1812 int ll_revalidate_statahead(struct inode *dir, struct dentry **dentryp,
1813                             bool unplug)
1814 {
1815         struct ll_statahead_info *sai;
1816         int rc = 0;
1817
1818         sai = ll_sai_get(dir);
1819         if (sai) {
1820                 rc = revalidate_statahead_dentry(dir, sai, dentryp, unplug);
1821                 CDEBUG(D_READA, "revalidate statahead %pd: rc = %d.\n",
1822                        *dentryp, rc);
1823                 ll_sai_put(sai);
1824         }
1825         return rc;
1826 }