lustre/llite/statahead.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright  2008 Sun Microsystems, Inc. All rights reserved
  30  * Use is subject to license terms.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  */
  36
  37 #include <linux/fs.h>
  38 #include <linux/sched.h>
  39 #include <linux/mm.h>
  40 #include <linux/smp_lock.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_lite.h>
  48 #include <lustre_dlm.h>
  49 #include <linux/lustre_version.h>
  50 #include "llite_internal.h"
  51
  52 struct ll_sai_entry {
  53         cfs_list_t              se_list;
  54         unsigned int            se_index;
  55         int                     se_stat;
  56         struct ptlrpc_request  *se_req;
  57         struct md_enqueue_info *se_minfo;
  58 };
  59
  60 enum {
  61         SA_ENTRY_UNSTATED = 0,
  62         SA_ENTRY_STATED
  63 };
  64
  65 static unsigned int sai_generation = 0;
  66 static cfs_spinlock_t sai_generation_lock = CFS_SPIN_LOCK_UNLOCKED;
  67
  68 /**
  69  * Check whether first entry was stated already or not.
  70  * No need to hold lli_lock, for:
  71  * (1) it is me that remove entry from the list
  72  * (2) the statahead thread only add new entry to the list
  73  */
  74 static int ll_sai_entry_stated(struct ll_statahead_info *sai)
  75 {
  76         struct ll_sai_entry  *entry;
  77         int                   rc = 0;
  78
  79         if (!cfs_list_empty(&sai->sai_entries_stated)) {
  80                 entry = cfs_list_entry(sai->sai_entries_stated.next,
  81                                        struct ll_sai_entry, se_list);
  82                 if (entry->se_index == sai->sai_index_next)
  83                         rc = 1;
  84         }
  85         return rc;
  86 }
  87
  88 static inline int sa_received_empty(struct ll_statahead_info *sai)
  89 {
  90         return cfs_list_empty(&sai->sai_entries_received);
  91 }
  92
  93 static inline int sa_not_full(struct ll_statahead_info *sai)
  94 {
  95         return (sai->sai_index < sai->sai_hit + sai->sai_miss + sai->sai_max);
  96 }
  97
  98 static inline int sa_is_running(struct ll_statahead_info *sai)
  99 {
 100         return !!(sai->sai_thread.t_flags & SVC_RUNNING);
 101 }
 102
 103 static inline int sa_is_stopping(struct ll_statahead_info *sai)
 104 {
 105         return !!(sai->sai_thread.t_flags & SVC_STOPPING);
 106 }
 107
 108 static inline int sa_is_stopped(struct ll_statahead_info *sai)
 109 {
 110         return !!(sai->sai_thread.t_flags & SVC_STOPPED);
 111 }
 112
 113 /**
 114  * (1) hit ratio less than 80%
 115  * or
 116  * (2) consecutive miss more than 8
 117  */
 118 static inline int sa_low_hit(struct ll_statahead_info *sai)
 119 {
 120         return ((sai->sai_hit > 7 && sai->sai_hit < 4 * sai->sai_miss) ||
 121                 (sai->sai_consecutive_miss > 8));
 122 }
 123
 124 /**
 125  * process the deleted entry's member and free the entry.
 126  * (1) release intent
 127  * (2) free md_enqueue_info
 128  * (3) drop dentry's ref count
 129  * (4) release request's ref count
 130  */
 131 static void ll_sai_entry_cleanup(struct ll_sai_entry *entry, int free)
 132 {
 133         struct md_enqueue_info *minfo = entry->se_minfo;
 134         struct ptlrpc_request  *req = entry->se_req;
 135         ENTRY;
 136
 137         if (minfo) {
 138                 entry->se_minfo = NULL;
 139                 ll_intent_release(&minfo->mi_it);
 140                 dput(minfo->mi_dentry);
 141                 iput(minfo->mi_dir);
 142                 OBD_FREE_PTR(minfo);
 143         }
 144         if (req) {
 145                 entry->se_req = NULL;
 146                 ptlrpc_req_finished(req);
 147         }
 148         if (free) {
 149                 LASSERT(cfs_list_empty(&entry->se_list));
 150                 OBD_FREE_PTR(entry);
 151         }
 152
 153         EXIT;
 154 }
 155
 156 static struct ll_statahead_info *ll_sai_alloc(void)
 157 {
 158         struct ll_statahead_info *sai;
 159
 160         OBD_ALLOC_PTR(sai);
 161         if (!sai)
 162                 return NULL;
 163
 164         cfs_spin_lock(&sai_generation_lock);
 165         sai->sai_generation = ++sai_generation;
 166         if (unlikely(sai_generation == 0))
 167                 sai->sai_generation = ++sai_generation;
 168         cfs_spin_unlock(&sai_generation_lock);
 169         cfs_atomic_set(&sai->sai_refcount, 1);
 170         sai->sai_max = LL_SA_RPC_MIN;
 171         cfs_waitq_init(&sai->sai_waitq);
 172         cfs_waitq_init(&sai->sai_thread.t_ctl_waitq);
 173         CFS_INIT_LIST_HEAD(&sai->sai_entries_sent);
 174         CFS_INIT_LIST_HEAD(&sai->sai_entries_received);
 175         CFS_INIT_LIST_HEAD(&sai->sai_entries_stated);
 176         return sai;
 177 }
 178
 179 static inline
 180 struct ll_statahead_info *ll_sai_get(struct ll_statahead_info *sai)
 181 {
 182         LASSERT(sai);
 183         cfs_atomic_inc(&sai->sai_refcount);
 184         return sai;
 185 }
 186
 187 static void ll_sai_put(struct ll_statahead_info *sai)
 188 {
 189         struct inode         *inode = sai->sai_inode;
 190         struct ll_inode_info *lli;
 191         ENTRY;
 192
 193         LASSERT(inode != NULL);
 194         lli = ll_i2info(inode);
 195         LASSERT(lli->lli_sai == sai);
 196
 197         if (cfs_atomic_dec_and_test(&sai->sai_refcount)) {
 198                 struct ll_sai_entry *entry, *next;
 199
 200                 cfs_spin_lock(&lli->lli_lock);
 201                 if (unlikely(cfs_atomic_read(&sai->sai_refcount) > 0)) {
 202                         /* It is race case, the interpret callback just hold
 203                          * a reference count */
 204                         cfs_spin_unlock(&lli->lli_lock);
 205                         EXIT;
 206                         return;
 207                 }
 208
 209                 LASSERT(lli->lli_opendir_key == NULL);
 210                 lli->lli_sai = NULL;
 211                 lli->lli_opendir_pid = 0;
 212                 cfs_spin_unlock(&lli->lli_lock);
 213
 214                 LASSERT(sa_is_stopped(sai));
 215
 216                 if (sai->sai_sent > sai->sai_replied)
 217                         CDEBUG(D_READA,"statahead for dir "DFID" does not "
 218                               "finish: [sent:%u] [replied:%u]\n",
 219                               PFID(&lli->lli_fid),
 220                               sai->sai_sent, sai->sai_replied);
 221
 222                 cfs_list_for_each_entry_safe(entry, next,
 223                                              &sai->sai_entries_sent, se_list) {
 224                         cfs_list_del_init(&entry->se_list);
 225                         ll_sai_entry_cleanup(entry, 1);
 226                 }
 227                 cfs_list_for_each_entry_safe(entry, next,
 228                                              &sai->sai_entries_received,
 229                                              se_list) {
 230                         cfs_list_del_init(&entry->se_list);
 231                         ll_sai_entry_cleanup(entry, 1);
 232                 }
 233                 cfs_list_for_each_entry_safe(entry, next,
 234                                              &sai->sai_entries_stated,
 235                                              se_list) {
 236                         cfs_list_del_init(&entry->se_list);
 237                         ll_sai_entry_cleanup(entry, 1);
 238                 }
 239                 iput(inode);
 240                 OBD_FREE_PTR(sai);
 241         }
 242         EXIT;
 243 }
 244
 245 /**
 246  * insert it into sai_entries_sent tail when init.
 247  */
 248 static struct ll_sai_entry *
 249 ll_sai_entry_init(struct ll_statahead_info *sai, unsigned int index)
 250 {
 251         struct ll_inode_info *lli = ll_i2info(sai->sai_inode);
 252         struct ll_sai_entry  *entry;
 253         ENTRY;
 254
 255         OBD_ALLOC_PTR(entry);
 256         if (entry == NULL)
 257                 RETURN(ERR_PTR(-ENOMEM));
 258
 259         CDEBUG(D_READA, "alloc sai entry %p index %u\n",
 260                entry, index);
 261         entry->se_index = index;
 262         entry->se_stat  = SA_ENTRY_UNSTATED;
 263
 264         cfs_spin_lock(&lli->lli_lock);
 265         cfs_list_add_tail(&entry->se_list, &sai->sai_entries_sent);
 266         cfs_spin_unlock(&lli->lli_lock);
 267
 268         RETURN(entry);
 269 }
 270
 271 /**
 272  * delete it from sai_entries_stated head when fini, it need not
 273  * to process entry's member.
 274  */
 275 static void ll_sai_entry_fini(struct ll_statahead_info *sai)
 276 {
 277         struct ll_inode_info *lli = ll_i2info(sai->sai_inode);
 278         struct ll_sai_entry  *entry;
 279         ENTRY;
 280
 281         cfs_spin_lock(&lli->lli_lock);
 282         sai->sai_index_next++;
 283         if (likely(!cfs_list_empty(&sai->sai_entries_stated))) {
 284                 entry = cfs_list_entry(sai->sai_entries_stated.next,
 285                                        struct ll_sai_entry, se_list);
 286                 if (entry->se_index < sai->sai_index_next) {
 287                         cfs_list_del(&entry->se_list);
 288                         OBD_FREE_PTR(entry);
 289                 }
 290         } else
 291                 LASSERT(sa_is_stopped(sai));
 292         cfs_spin_unlock(&lli->lli_lock);
 293
 294         EXIT;
 295 }
 296
 297 /**
 298  * inside lli_lock.
 299  * \retval NULL : can not find the entry in sai_entries_sent with the index
 300  * \retval entry: find the entry in sai_entries_sent with the index
 301  */
 302 static struct ll_sai_entry *
 303 ll_sai_entry_set(struct ll_statahead_info *sai, unsigned int index, int stat,
 304                  struct ptlrpc_request *req, struct md_enqueue_info *minfo)
 305 {
 306         struct ll_sai_entry *entry;
 307         ENTRY;
 308
 309         if (!cfs_list_empty(&sai->sai_entries_sent)) {
 310                 cfs_list_for_each_entry(entry, &sai->sai_entries_sent,
 311                                         se_list) {
 312                         if (entry->se_index == index) {
 313                                 entry->se_stat = stat;
 314                                 entry->se_req = ptlrpc_request_addref(req);
 315                                 entry->se_minfo = minfo;
 316                                 RETURN(entry);
 317                         } else if (entry->se_index > index)
 318                                 RETURN(NULL);
 319                 }
 320         }
 321         RETURN(NULL);
 322 }
 323
 324 /**
 325  * inside lli_lock.
 326  * Move entry to sai_entries_received and
 327  * insert it into sai_entries_received tail.
 328  */
 329 static inline void
 330 ll_sai_entry_to_received(struct ll_statahead_info *sai, struct ll_sai_entry *entry)
 331 {
 332         if (!cfs_list_empty(&entry->se_list))
 333                 cfs_list_del_init(&entry->se_list);
 334         cfs_list_add_tail(&entry->se_list, &sai->sai_entries_received);
 335 }
 336
 337 /**
 338  * Move entry to sai_entries_stated and
 339  * sort with the index.
 340  */
 341 static int
 342 ll_sai_entry_to_stated(struct ll_statahead_info *sai, struct ll_sai_entry *entry)
 343 {
 344         struct ll_inode_info *lli = ll_i2info(sai->sai_inode);
 345         struct ll_sai_entry  *se;
 346         ENTRY;
 347
 348         ll_sai_entry_cleanup(entry, 0);
 349
 350         cfs_spin_lock(&lli->lli_lock);
 351         if (!cfs_list_empty(&entry->se_list))
 352                 cfs_list_del_init(&entry->se_list);
 353
 354         if (unlikely(entry->se_index < sai->sai_index_next)) {
 355                 cfs_spin_unlock(&lli->lli_lock);
 356                 OBD_FREE_PTR(entry);
 357                 RETURN(0);
 358         }
 359
 360         cfs_list_for_each_entry_reverse(se, &sai->sai_entries_stated, se_list) {
 361                 if (se->se_index < entry->se_index) {
 362                         cfs_list_add(&entry->se_list, &se->se_list);
 363                         cfs_spin_unlock(&lli->lli_lock);
 364                         RETURN(1);
 365                 }
 366         }
 367
 368         /*
 369          * I am the first entry.
 370          */
 371         cfs_list_add(&entry->se_list, &sai->sai_entries_stated);
 372         cfs_spin_unlock(&lli->lli_lock);
 373         RETURN(1);
 374 }
 375
 376 /**
 377  * finish lookup/revalidate.
 378  */
 379 static int do_statahead_interpret(struct ll_statahead_info *sai)
 380 {
 381         struct ll_inode_info   *lli = ll_i2info(sai->sai_inode);
 382         struct ll_sai_entry    *entry;
 383         struct ptlrpc_request  *req;
 384         struct md_enqueue_info *minfo;
 385         struct lookup_intent   *it;
 386         struct dentry          *dentry;
 387         int                     rc = 0;
 388         struct mdt_body        *body;
 389         ENTRY;
 390
 391         cfs_spin_lock(&lli->lli_lock);
 392         LASSERT(!sa_received_empty(sai));
 393         entry = cfs_list_entry(sai->sai_entries_received.next,
 394                                struct ll_sai_entry, se_list);
 395         cfs_list_del_init(&entry->se_list);
 396         cfs_spin_unlock(&lli->lli_lock);
 397
 398         if (unlikely(entry->se_index < sai->sai_index_next)) {
 399                 CWARN("Found stale entry: [index %u] [next %u]\n",
 400                       entry->se_index, sai->sai_index_next);
 401                 ll_sai_entry_cleanup(entry, 1);
 402                 RETURN(0);
 403         }
 404
 405         if (entry->se_stat != SA_ENTRY_STATED)
 406                 GOTO(out, rc = entry->se_stat);
 407
 408         req = entry->se_req;
 409         minfo = entry->se_minfo;
 410         it = &minfo->mi_it;
 411         dentry = minfo->mi_dentry;
 412
 413         body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
 414         if (body == NULL)
 415                 GOTO(out, rc = -EFAULT);
 416
 417         if (dentry->d_inode == NULL) {
 418                 /*
 419                  * lookup.
 420                  */
 421                 struct dentry    *save = dentry;
 422                 struct it_cb_data icbd = {
 423                         .icbd_parent   = minfo->mi_dir,
 424                         .icbd_childp   = &dentry
 425                 };
 426
 427                 LASSERT(fid_is_zero(&minfo->mi_data.op_fid2));
 428
 429                 /*
 430                  * XXX: No fid in reply, this is probaly cross-ref case.
 431                  * SA can't handle it yet.
 432                  */
 433                 if (body->valid & OBD_MD_MDS)
 434                         GOTO(out, rc = -EAGAIN);
 435
 436                 rc = ll_lookup_it_finish(req, it, &icbd);
 437                 if (!rc)
 438                         /*
 439                          * Here dentry->d_inode might be NULL,
 440                          * because the entry may have been removed before
 441                          * we start doing stat ahead.
 442                          */
 443                         ll_lookup_finish_locks(it, dentry);
 444
 445                 if (dentry != save) {
 446                         minfo->mi_dentry = dentry;
 447                         dput(save);
 448                 }
 449         } else {
 450                 /*
 451                  * revalidate.
 452                  */
 453                 if (!lu_fid_eq(&minfo->mi_data.op_fid2, &body->fid1)) {
 454                         ll_unhash_aliases(dentry->d_inode);
 455                         GOTO(out, rc = -EAGAIN);
 456                 }
 457
 458                 rc = ll_revalidate_it_finish(req, it, dentry);
 459                 if (rc) {
 460                         ll_unhash_aliases(dentry->d_inode);
 461                         GOTO(out, rc);
 462                 }
 463
 464                 cfs_spin_lock(&ll_lookup_lock);
 465                 spin_lock(&dcache_lock);
 466                 lock_dentry(dentry);
 467                 __d_drop(dentry);
 468                 dentry->d_flags &= ~DCACHE_LUSTRE_INVALID;
 469                 unlock_dentry(dentry);
 470                 d_rehash_cond(dentry, 0);
 471                 spin_unlock(&dcache_lock);
 472                 cfs_spin_unlock(&ll_lookup_lock);
 473
 474                 ll_lookup_finish_locks(it, dentry);
 475         }
 476         EXIT;
 477
 478 out:
 479         if (likely(ll_sai_entry_to_stated(sai, entry)))
 480                 cfs_waitq_signal(&sai->sai_waitq);
 481         return rc;
 482 }
 483
 484 static int ll_statahead_interpret(struct ptlrpc_request *req,
 485                                   struct md_enqueue_info *minfo,
 486                                   int rc)
 487 {
 488         struct lookup_intent     *it = &minfo->mi_it;
 489         struct dentry            *dentry = minfo->mi_dentry;
 490         struct inode             *dir = minfo->mi_dir;
 491         struct ll_inode_info     *lli = ll_i2info(dir);
 492         struct ll_statahead_info *sai;
 493         struct ll_sai_entry      *entry;
 494         ENTRY;
 495
 496         CDEBUG(D_READA, "interpret statahead %.*s rc %d\n",
 497                dentry->d_name.len, dentry->d_name.name, rc);
 498
 499         cfs_spin_lock(&lli->lli_lock);
 500         if (unlikely(lli->lli_sai == NULL ||
 501             lli->lli_sai->sai_generation != minfo->mi_generation)) {
 502                 cfs_spin_unlock(&lli->lli_lock);
 503                 ll_intent_release(it);
 504                 dput(dentry);
 505                 iput(dir);
 506                 OBD_FREE_PTR(minfo);
 507                 RETURN(-ESTALE);
 508         } else {
 509                 sai = ll_sai_get(lli->lli_sai);
 510                 entry = ll_sai_entry_set(sai,
 511                                          (unsigned int)(long)minfo->mi_cbdata,
 512                                          rc ? SA_ENTRY_UNSTATED :
 513                                          SA_ENTRY_STATED, req, minfo);
 514                 LASSERT(entry != NULL);
 515                 if (likely(sa_is_running(sai))) {
 516                         ll_sai_entry_to_received(sai, entry);
 517                         sai->sai_replied++;
 518                         cfs_spin_unlock(&lli->lli_lock);
 519                         cfs_waitq_signal(&sai->sai_thread.t_ctl_waitq);
 520                 } else {
 521                         if (!cfs_list_empty(&entry->se_list))
 522                                 cfs_list_del_init(&entry->se_list);
 523                         sai->sai_replied++;
 524                         cfs_spin_unlock(&lli->lli_lock);
 525                         ll_sai_entry_cleanup(entry, 1);
 526                 }
 527                 ll_sai_put(sai);
 528                 RETURN(rc);
 529         }
 530 }
 531
 532 static void sa_args_fini(struct md_enqueue_info *minfo,
 533                          struct ldlm_enqueue_info *einfo)
 534 {
 535         LASSERT(minfo && einfo);
 536         iput(minfo->mi_dir);
 537         capa_put(minfo->mi_data.op_capa1);
 538         capa_put(minfo->mi_data.op_capa2);
 539         OBD_FREE_PTR(minfo);
 540         OBD_FREE_PTR(einfo);
 541 }
 542
 543 /**
 544  * There is race condition between "capa_put" and "ll_statahead_interpret" for
 545  * accessing "op_data.op_capa[1,2]" as following:
 546  * "capa_put" releases "op_data.op_capa[1,2]"'s reference count after calling
 547  * "md_intent_getattr_async". But "ll_statahead_interpret" maybe run first, and
 548  * fill "op_data.op_capa[1,2]" as POISON, then cause "capa_put" access invalid
 549  * "ocapa". So here reserve "op_data.op_capa[1,2]" in "pcapa" before calling
 550  * "md_intent_getattr_async".
 551  */
 552 static int sa_args_init(struct inode *dir, struct dentry *dentry,
 553                         struct md_enqueue_info **pmi,
 554                         struct ldlm_enqueue_info **pei,
 555                         struct obd_capa **pcapa)
 556 {
 557         struct ll_inode_info     *lli = ll_i2info(dir);
 558         struct md_enqueue_info   *minfo;
 559         struct ldlm_enqueue_info *einfo;
 560         struct md_op_data        *op_data;
 561
 562         OBD_ALLOC_PTR(einfo);
 563         if (einfo == NULL)
 564                 return -ENOMEM;
 565
 566         OBD_ALLOC_PTR(minfo);
 567         if (minfo == NULL) {
 568                 OBD_FREE_PTR(einfo);
 569                 return -ENOMEM;
 570         }
 571
 572         op_data = ll_prep_md_op_data(&minfo->mi_data, dir, dentry->d_inode,
 573                                      dentry->d_name.name, dentry->d_name.len,
 574                                      0, LUSTRE_OPC_ANY, NULL);
 575         if (IS_ERR(op_data)) {
 576                 OBD_FREE_PTR(einfo);
 577                 OBD_FREE_PTR(minfo);
 578                 return PTR_ERR(op_data);
 579         }
 580
 581         minfo->mi_it.it_op = IT_GETATTR;
 582         minfo->mi_dentry = dentry;
 583         minfo->mi_dir = igrab(dir);
 584         minfo->mi_cb = ll_statahead_interpret;
 585         minfo->mi_generation = lli->lli_sai->sai_generation;
 586         minfo->mi_cbdata = (void *)(long)lli->lli_sai->sai_index;
 587
 588         einfo->ei_type   = LDLM_IBITS;
 589         einfo->ei_mode   = it_to_lock_mode(&minfo->mi_it);
 590         einfo->ei_cb_bl  = ll_md_blocking_ast;
 591         einfo->ei_cb_cp  = ldlm_completion_ast;
 592         einfo->ei_cb_gl  = NULL;
 593         einfo->ei_cbdata = NULL;
 594
 595         *pmi = minfo;
 596         *pei = einfo;
 597         pcapa[0] = op_data->op_capa1;
 598         pcapa[1] = op_data->op_capa2;
 599
 600         return 0;
 601 }
 602
 603 /**
 604  * similar to ll_lookup_it().
 605  */
 606 static int do_sa_lookup(struct inode *dir, struct dentry *dentry)
 607 {
 608         struct md_enqueue_info   *minfo;
 609         struct ldlm_enqueue_info *einfo;
 610         struct obd_capa          *capas[2];
 611         int                       rc;
 612         ENTRY;
 613
 614         rc = sa_args_init(dir, dentry, &minfo, &einfo, capas);
 615         if (rc)
 616                 RETURN(rc);
 617
 618         rc = md_intent_getattr_async(ll_i2mdexp(dir), minfo, einfo);
 619         if (!rc) {
 620                 capa_put(capas[0]);
 621                 capa_put(capas[1]);
 622         } else {
 623                 sa_args_fini(minfo, einfo);
 624         }
 625
 626         RETURN(rc);
 627 }
 628
 629 /**
 630  * similar to ll_revalidate_it().
 631  * \retval      1 -- dentry valid
 632  * \retval      0 -- will send stat-ahead request
 633  * \retval others -- prepare stat-ahead request failed
 634  */
 635 static int do_sa_revalidate(struct inode *dir, struct dentry *dentry)
 636 {
 637         struct inode             *inode = dentry->d_inode;
 638         struct lookup_intent      it = { .it_op = IT_GETATTR };
 639         struct md_enqueue_info   *minfo;
 640         struct ldlm_enqueue_info *einfo;
 641         struct obd_capa          *capas[2];
 642         int rc;
 643         ENTRY;
 644
 645         if (inode == NULL)
 646                 RETURN(1);
 647
 648         if (d_mountpoint(dentry))
 649                 RETURN(1);
 650
 651         if (dentry == dentry->d_sb->s_root)
 652                 RETURN(1);
 653
 654         rc = md_revalidate_lock(ll_i2mdexp(dir), &it, ll_inode2fid(inode));
 655         if (rc == 1) {
 656                 ll_intent_release(&it);
 657                 RETURN(1);
 658         }
 659
 660         rc = sa_args_init(dir, dentry, &minfo, &einfo, capas);
 661         if (rc)
 662                 RETURN(rc);
 663
 664         rc = md_intent_getattr_async(ll_i2mdexp(dir), minfo, einfo);
 665         if (!rc) {
 666                 capa_put(capas[0]);
 667                 capa_put(capas[1]);
 668         } else {
 669                 sa_args_fini(minfo, einfo);
 670         }
 671
 672         RETURN(rc);
 673 }
 674
 675 static inline void ll_name2qstr(struct qstr *q, const char *name, int namelen)
 676 {
 677         q->name = name;
 678         q->len  = namelen;
 679         q->hash = full_name_hash(name, namelen);
 680 }
 681
 682 static int ll_statahead_one(struct dentry *parent, const char* entry_name,
 683                             int entry_name_len)
 684 {
 685         struct inode             *dir = parent->d_inode;
 686         struct ll_inode_info     *lli = ll_i2info(dir);
 687         struct ll_statahead_info *sai = lli->lli_sai;
 688         struct qstr               name;
 689         struct dentry            *dentry;
 690         struct ll_sai_entry      *se;
 691         int                       rc;
 692         ENTRY;
 693
 694         if (parent->d_flags & DCACHE_LUSTRE_INVALID) {
 695                 CDEBUG(D_READA, "parent dentry@%p %.*s is "
 696                        "invalid, skip statahead\n",
 697                        parent, parent->d_name.len, parent->d_name.name);
 698                 RETURN(-EINVAL);
 699         }
 700
 701         se = ll_sai_entry_init(sai, sai->sai_index);
 702         if (IS_ERR(se))
 703                 RETURN(PTR_ERR(se));
 704
 705         ll_name2qstr(&name, entry_name, entry_name_len);
 706         dentry = d_lookup(parent, &name);
 707         if (!dentry) {
 708                 dentry = d_alloc(parent, &name);
 709                 if (dentry) {
 710                         rc = do_sa_lookup(dir, dentry);
 711                         if (rc)
 712                                 dput(dentry);
 713                 } else {
 714                         GOTO(out, rc = -ENOMEM);
 715                 }
 716         } else {
 717                 rc = do_sa_revalidate(dir, dentry);
 718                 if (rc)
 719                         dput(dentry);
 720         }
 721
 722         EXIT;
 723
 724 out:
 725         if (rc) {
 726                 CDEBUG(D_READA, "set sai entry %p index %u stat %d rc %d\n",
 727                        se, se->se_index, se->se_stat, rc);
 728                 se->se_stat = rc;
 729                 if (ll_sai_entry_to_stated(sai, se))
 730                         cfs_waitq_signal(&sai->sai_waitq);
 731         } else {
 732                 sai->sai_sent++;
 733         }
 734
 735         sai->sai_index++;
 736         return rc;
 737 }
 738
 739 static int ll_statahead_thread(void *arg)
 740 {
 741         struct dentry            *parent = (struct dentry *)arg;
 742         struct inode             *dir = parent->d_inode;
 743         struct ll_inode_info     *lli = ll_i2info(dir);
 744         struct ll_sb_info        *sbi = ll_i2sbi(dir);
 745         struct ll_statahead_info *sai = ll_sai_get(lli->lli_sai);
 746         struct ptlrpc_thread     *thread = &sai->sai_thread;
 747         struct page              *page;
 748         __u64                     pos = 0;
 749         int                       first = 0;
 750         int                       rc = 0;
 751         struct ll_dir_chain       chain;
 752         ENTRY;
 753
 754         {
 755                 char pname[16];
 756                 snprintf(pname, 15, "ll_sa_%u", lli->lli_opendir_pid);
 757                 cfs_daemonize(pname);
 758         }
 759
 760         sbi->ll_sa_total++;
 761         cfs_spin_lock(&lli->lli_lock);
 762         thread->t_flags = SVC_RUNNING;
 763         cfs_spin_unlock(&lli->lli_lock);
 764         cfs_waitq_signal(&thread->t_ctl_waitq);
 765         CDEBUG(D_READA, "start doing statahead for %s\n", parent->d_name.name);
 766
 767         ll_dir_chain_init(&chain);
 768         page = ll_get_dir_page(dir, pos, 0, &chain);
 769
 770         while (1) {
 771                 struct l_wait_info lwi = { 0 };
 772                 struct lu_dirpage *dp;
 773                 struct lu_dirent  *ent;
 774
 775                 if (IS_ERR(page)) {
 776                         rc = PTR_ERR(page);
 777                         CDEBUG(D_READA, "error reading dir "DFID" at "LPU64"/%u: rc %d\n",
 778                                PFID(ll_inode2fid(dir)), pos,
 779                                sai->sai_index, rc);
 780                         break;
 781                 }
 782
 783                 dp = page_address(page);
 784                 for (ent = lu_dirent_start(dp); ent != NULL;
 785                      ent = lu_dirent_next(ent)) {
 786                         char *name = ent->lde_name;
 787                         int namelen = le16_to_cpu(ent->lde_namelen);
 788
 789                         if (namelen == 0)
 790                                 /*
 791                                  * Skip dummy record.
 792                                  */
 793                                 continue;
 794
 795                         if (name[0] == '.') {
 796                                 if (namelen == 1) {
 797                                         /*
 798                                          * skip "."
 799                                          */
 800                                         continue;
 801                                 } else if (name[1] == '.' && namelen == 2) {
 802                                         /*
 803                                          * skip ".."
 804                                          */
 805                                         continue;
 806                                 } else if (!sai->sai_ls_all) {
 807                                         /*
 808                                          * skip hidden files.
 809                                          */
 810                                         sai->sai_skip_hidden++;
 811                                         continue;
 812                                 }
 813                         }
 814
 815                         /*
 816                          * don't stat-ahead first entry.
 817                          */
 818                         if (unlikely(!first)) {
 819                                 first++;
 820                                 continue;
 821                         }
 822
 823 keep_de:
 824                         l_wait_event(thread->t_ctl_waitq,
 825                                      !sa_is_running(sai) || sa_not_full(sai) ||
 826                                      !sa_received_empty(sai),
 827                                      &lwi);
 828
 829                         while (!sa_received_empty(sai) && sa_is_running(sai))
 830                                 do_statahead_interpret(sai);
 831
 832                         if (unlikely(!sa_is_running(sai))) {
 833                                 ll_put_page(page);
 834                                 GOTO(out, rc);
 835                         }
 836
 837                         if (!sa_not_full(sai))
 838                                 /*
 839                                  * do not skip the current de.
 840                                  */
 841                                 goto keep_de;
 842
 843                         rc = ll_statahead_one(parent, name, namelen);
 844                         if (rc < 0) {
 845                                 ll_put_page(page);
 846                                 GOTO(out, rc);
 847                         }
 848                 }
 849                 pos = le64_to_cpu(dp->ldp_hash_end);
 850                 ll_put_page(page);
 851                 if (pos == DIR_END_OFF) {
 852                         /*
 853                          * End of directory reached.
 854                          */
 855                         while (1) {
 856                                 l_wait_event(thread->t_ctl_waitq,
 857                                              !sa_is_running(sai) ||
 858                                              !sa_received_empty(sai) ||
 859                                              sai->sai_sent == sai->sai_replied,
 860                                              &lwi);
 861                                 if (!sa_received_empty(sai) &&
 862                                     sa_is_running(sai))
 863                                         do_statahead_interpret(sai);
 864                                 else
 865                                         GOTO(out, rc);
 866                         }
 867                 } else if (1) {
 868                         /*
 869                          * chain is exhausted.
 870                          * Normal case: continue to the next page.
 871                          */
 872                         page = ll_get_dir_page(dir, pos, 1, &chain);
 873                 } else {
 874                         /*
 875                          * go into overflow page.
 876                          */
 877                 }
 878         }
 879         EXIT;
 880
 881 out:
 882         ll_dir_chain_fini(&chain);
 883         cfs_spin_lock(&lli->lli_lock);
 884         thread->t_flags = SVC_STOPPED;
 885         cfs_spin_unlock(&lli->lli_lock);
 886         cfs_waitq_signal(&sai->sai_waitq);
 887         cfs_waitq_signal(&thread->t_ctl_waitq);
 888         ll_sai_put(sai);
 889         dput(parent);
 890         CDEBUG(D_READA, "statahead thread stopped, pid %d\n",
 891                cfs_curproc_pid());
 892         return rc;
 893 }
 894
 895 /**
 896  * called in ll_file_release().
 897  */
 898 void ll_stop_statahead(struct inode *inode, void *key)
 899 {
 900         struct ll_inode_info *lli = ll_i2info(inode);
 901
 902         if (unlikely(key == NULL))
 903                 return;
 904
 905         cfs_spin_lock(&lli->lli_lock);
 906         if (lli->lli_opendir_key != key || lli->lli_opendir_pid == 0) {
 907                 cfs_spin_unlock(&lli->lli_lock);
 908                 return;
 909         }
 910
 911         lli->lli_opendir_key = NULL;
 912
 913         if (lli->lli_sai) {
 914                 struct l_wait_info lwi = { 0 };
 915                 struct ptlrpc_thread *thread = &lli->lli_sai->sai_thread;
 916
 917                 if (!sa_is_stopped(lli->lli_sai)) {
 918                         thread->t_flags = SVC_STOPPING;
 919                         cfs_spin_unlock(&lli->lli_lock);
 920                         cfs_waitq_signal(&thread->t_ctl_waitq);
 921
 922                         CDEBUG(D_READA, "stopping statahead thread, pid %d\n",
 923                                cfs_curproc_pid());
 924                         l_wait_event(thread->t_ctl_waitq,
 925                                      sa_is_stopped(lli->lli_sai),
 926                                      &lwi);
 927                 } else {
 928                         cfs_spin_unlock(&lli->lli_lock);
 929                 }
 930
 931                 /*
 932                  * Put the ref which was held when first statahead_enter.
 933                  * It maybe not the last ref for some statahead requests
 934                  * maybe inflight.
 935                  */
 936                 ll_sai_put(lli->lli_sai);
 937         } else {
 938                 lli->lli_opendir_pid = 0;
 939                 cfs_spin_unlock(&lli->lli_lock);
 940         }
 941 }
 942
 943 enum {
 944         /**
 945          * not first dirent, or is "."
 946          */
 947         LS_NONE_FIRST_DE = 0,
 948         /**
 949          * the first non-hidden dirent
 950          */
 951         LS_FIRST_DE,
 952         /**
 953          * the first hidden dirent, that is "."
 954          */
 955         LS_FIRST_DOT_DE
 956 };
 957
 958 static int is_first_dirent(struct inode *dir, struct dentry *dentry)
 959 {
 960         struct ll_dir_chain chain;
 961         struct qstr        *target = &dentry->d_name;
 962         struct page        *page;
 963         __u64               pos = 0;
 964         int                 dot_de;
 965         int                 rc = LS_NONE_FIRST_DE;
 966         ENTRY;
 967
 968         ll_dir_chain_init(&chain);
 969         page = ll_get_dir_page(dir, pos, 0, &chain);
 970
 971         while (1) {
 972                 struct lu_dirpage *dp;
 973                 struct lu_dirent  *ent;
 974
 975                 if (IS_ERR(page)) {
 976                         rc = PTR_ERR(page);
 977                         CERROR("error reading dir "DFID" at "LPU64": rc %d\n",
 978                                PFID(ll_inode2fid(dir)), pos, rc);
 979                         break;
 980                 }
 981
 982                 dp = page_address(page);
 983                 for (ent = lu_dirent_start(dp); ent != NULL;
 984                      ent = lu_dirent_next(ent)) {
 985                         char *name = ent->lde_name;
 986                         int namelen = le16_to_cpu(ent->lde_namelen);
 987
 988                         if (namelen == 0)
 989                                 /*
 990                                  * skip dummy record.
 991                                  */
 992                                 continue;
 993
 994                         if (name[0] == '.') {
 995                                 if (namelen == 1)
 996                                         /*
 997                                          * skip "."
 998                                          */
 999                                         continue;
1000                                 else if (name[1] == '.' && namelen == 2)
1001                                         /*
1002                                          * skip ".."
1003                                          */
1004                                         continue;
1005                                 else
1006                                         dot_de = 1;
1007                         } else {
1008                                 dot_de = 0;
1009                         }
1010
1011                         if (dot_de && target->name[0] != '.') {
1012                                 CDEBUG(D_READA, "%.*s skip hidden file %.*s\n",
1013                                        target->len, target->name,
1014                                        namelen, name);
1015                                 continue;
1016                         }
1017
1018                         if (target->len != namelen ||
1019                             memcmp(target->name, name, namelen) != 0)
1020                                 rc = LS_NONE_FIRST_DE;
1021                         else if (!dot_de)
1022                                 rc = LS_FIRST_DE;
1023                         else
1024                                 rc = LS_FIRST_DOT_DE;
1025
1026                         ll_put_page(page);
1027                         GOTO(out, rc);
1028                 }
1029                 pos = le64_to_cpu(dp->ldp_hash_end);
1030                 ll_put_page(page);
1031                 if (pos == DIR_END_OFF) {
1032                         /*
1033                          * End of directory reached.
1034                          */
1035                         break;
1036                 } else if (1) {
1037                         /*
1038                          * chain is exhausted
1039                          * Normal case: continue to the next page.
1040                          */
1041                         page = ll_get_dir_page(dir, pos, 1, &chain);
1042                 } else {
1043                         /*
1044                          * go into overflow page.
1045                          */
1046                 }
1047         }
1048         EXIT;
1049
1050 out:
1051         ll_dir_chain_fini(&chain);
1052         return rc;
1053 }
1054
1055 /**
1056  * Start statahead thread if this is the first dir entry.
1057  * Otherwise if a thread is started already, wait it until it is ahead of me.
1058  * \retval 0       -- stat ahead thread process such dentry, for lookup, it miss
1059  * \retval 1       -- stat ahead thread process such dentry, for lookup, it hit
1060  * \retval -EEXIST -- stat ahead thread started, and this is the first dentry
1061  * \retval -EBADFD -- statahead thread exit and not dentry available
1062  * \retval -EAGAIN -- try to stat by caller
1063  * \retval others  -- error
1064  */
1065 int do_statahead_enter(struct inode *dir, struct dentry **dentryp, int lookup)
1066 {
1067         struct ll_inode_info     *lli;
1068         struct ll_statahead_info *sai;
1069         struct dentry            *parent;
1070         struct l_wait_info        lwi = { 0 };
1071         int                       rc = 0;
1072         ENTRY;
1073
1074         LASSERT(dir != NULL);
1075         lli = ll_i2info(dir);
1076         LASSERT(lli->lli_opendir_pid == cfs_curproc_pid());
1077         sai = lli->lli_sai;
1078
1079         if (sai) {
1080                 struct ll_sb_info *sbi;
1081
1082                 if (unlikely(sa_is_stopped(sai) &&
1083                              cfs_list_empty(&sai->sai_entries_stated)))
1084                         RETURN(-EBADFD);
1085
1086                 if ((*dentryp)->d_name.name[0] == '.') {
1087                         if (likely(sai->sai_ls_all ||
1088                             sai->sai_miss_hidden >= sai->sai_skip_hidden)) {
1089                                 /*
1090                                  * Hidden dentry is the first one, or statahead
1091                                  * thread does not skip so many hidden dentries
1092                                  * before "sai_ls_all" enabled as below.
1093                                  */
1094                         } else {
1095                                 if (!sai->sai_ls_all)
1096                                         /*
1097                                          * It maybe because hidden dentry is not
1098                                          * the first one, "sai_ls_all" was not
1099                                          * set, then "ls -al" missed. Enable
1100                                          * "sai_ls_all" for such case.
1101                                          */
1102                                         sai->sai_ls_all = 1;
1103
1104                                 /*
1105                                  * Such "getattr" has been skipped before
1106                                  * "sai_ls_all" enabled as above.
1107                                  */
1108                                 sai->sai_miss_hidden++;
1109                                 RETURN(-ENOENT);
1110                         }
1111                 }
1112
1113                 sbi = ll_i2sbi(dir);
1114                 if (ll_sai_entry_stated(sai)) {
1115                         sbi->ll_sa_cached++;
1116                 } else {
1117                         sbi->ll_sa_blocked++;
1118                         /*
1119                          * thread started already, avoid double-stat.
1120                          */
1121                         lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
1122                         rc = l_wait_event(sai->sai_waitq,
1123                                           ll_sai_entry_stated(sai) ||
1124                                           sa_is_stopped(sai),
1125                                           &lwi);
1126                 }
1127
1128                 if (lookup) {
1129                         struct dentry *result;
1130
1131                         result = d_lookup((*dentryp)->d_parent,
1132                                           &(*dentryp)->d_name);
1133                         if (result) {
1134                                 LASSERT(result != *dentryp);
1135                                 /* BUG 16303: do not drop reference count for
1136                                  * "*dentryp", VFS will do that by itself. */
1137                                 *dentryp = result;
1138                                 RETURN(1);
1139                         }
1140                 }
1141                 /*
1142                  * do nothing for revalidate.
1143                  */
1144                 RETURN(rc);
1145         }
1146
1147         /* I am the "lli_opendir_pid" owner, only me can set "lli_sai". */
1148         rc = is_first_dirent(dir, *dentryp);
1149         if (rc == LS_NONE_FIRST_DE)
1150                 /* It is not "ls -{a}l" operation, no need statahead for it. */
1151                 GOTO(out, rc = -EAGAIN);
1152
1153         sai = ll_sai_alloc();
1154         if (sai == NULL)
1155                 GOTO(out, rc = -ENOMEM);
1156
1157         sai->sai_ls_all = (rc == LS_FIRST_DOT_DE);
1158         sai->sai_inode = igrab(dir);
1159         if (unlikely(sai->sai_inode == NULL)) {
1160                 CWARN("Do not start stat ahead on dying inode "DFID" .\n",
1161                       PFID(&lli->lli_fid));
1162                 OBD_FREE_PTR(sai);
1163                 GOTO(out, rc = -ESTALE);
1164         }
1165
1166         /* get parent reference count here, and put it in ll_statahead_thread */
1167         parent = dget((*dentryp)->d_parent);
1168         if (unlikely(sai->sai_inode != parent->d_inode)) {
1169                 struct ll_inode_info *nlli = ll_i2info(parent->d_inode);
1170
1171                 CWARN("Race condition, someone changed %.*s just now: "
1172                       "old parent "DFID", new parent "DFID" .\n",
1173                       (*dentryp)->d_name.len, (*dentryp)->d_name.name,
1174                       PFID(&lli->lli_fid), PFID(&nlli->lli_fid));
1175                 dput(parent);
1176                 iput(sai->sai_inode);
1177                 OBD_FREE_PTR(sai);
1178                 RETURN(-EAGAIN);
1179         }
1180
1181         lli->lli_sai = sai;
1182         rc = cfs_kernel_thread(ll_statahead_thread, parent, 0);
1183         if (rc < 0) {
1184                 CERROR("can't start ll_sa thread, rc: %d\n", rc);
1185                 dput(parent);
1186                 lli->lli_opendir_key = NULL;
1187                 sai->sai_thread.t_flags = SVC_STOPPED;
1188                 ll_sai_put(sai);
1189                 LASSERT(lli->lli_sai == NULL);
1190                 RETURN(-EAGAIN);
1191         }
1192
1193         l_wait_event(sai->sai_thread.t_ctl_waitq,
1194                      sa_is_running(sai) || sa_is_stopped(sai),
1195                      &lwi);
1196
1197         /*
1198          * We don't stat-ahead for the first dirent since we are already in
1199          * lookup, and -EEXIST also indicates that this is the first dirent.
1200          */
1201         RETURN(-EEXIST);
1202
1203 out:
1204         cfs_spin_lock(&lli->lli_lock);
1205         lli->lli_opendir_key = NULL;
1206         lli->lli_opendir_pid = 0;
1207         cfs_spin_unlock(&lli->lli_lock);
1208         return rc;
1209 }
1210
1211 /**
1212  * update hit/miss count.
1213  */
1214 void ll_statahead_exit(struct inode *dir, struct dentry *dentry, int result)
1215 {
1216         struct ll_inode_info     *lli;
1217         struct ll_statahead_info *sai;
1218         struct ll_sb_info        *sbi;
1219         struct ll_dentry_data    *ldd = ll_d2d(dentry);
1220         ENTRY;
1221
1222         LASSERT(dir != NULL);
1223         lli = ll_i2info(dir);
1224         LASSERT(lli->lli_opendir_pid == cfs_curproc_pid());
1225         sai = lli->lli_sai;
1226         LASSERT(sai != NULL);
1227         sbi = ll_i2sbi(dir);
1228
1229         if (result >= 1) {
1230                 sbi->ll_sa_hit++;
1231                 sai->sai_hit++;
1232                 sai->sai_consecutive_miss = 0;
1233                 sai->sai_max = min(2 * sai->sai_max, sbi->ll_sa_max);
1234         } else {
1235                 sbi->ll_sa_miss++;
1236                 sai->sai_miss++;
1237                 sai->sai_consecutive_miss++;
1238                 if (sa_low_hit(sai) && sa_is_running(sai)) {
1239                         sbi->ll_sa_wrong++;
1240                         CDEBUG(D_READA, "Statahead for dir "DFID" hit ratio "
1241                                "too low: hit/miss %u/%u, sent/replied %u/%u, "
1242                                "stopping statahead thread: pid %d\n",
1243                                PFID(&lli->lli_fid), sai->sai_hit,
1244                                sai->sai_miss, sai->sai_sent,
1245                                sai->sai_replied, cfs_curproc_pid());
1246                         cfs_spin_lock(&lli->lli_lock);
1247                         if (!sa_is_stopped(sai))
1248                                 sai->sai_thread.t_flags = SVC_STOPPING;
1249                         cfs_spin_unlock(&lli->lli_lock);
1250                 }
1251         }
1252
1253         if (!sa_is_stopped(sai))
1254                 cfs_waitq_signal(&sai->sai_thread.t_ctl_waitq);
1255         ll_sai_entry_fini(sai);
1256         if (likely(ldd != NULL))
1257                 ldd->lld_sa_generation = sai->sai_generation;
1258
1259         EXIT;
1260 }