lustre/llite/llite_capa.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2012, 2014, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lustre/llite/llite_capa.c
  37  *
  38  * Author: Lai Siyao <lsy@clusterfs.com>
  39  */
  40
  41 #define DEBUG_SUBSYSTEM S_LLITE
  42
  43 #include <linux/fs.h>
  44 #include <linux/version.h>
  45 #include <linux/kthread.h>
  46 #include <asm/uaccess.h>
  47 #include <linux/file.h>
  48 #include <linux/kmod.h>
  49
  50 #include "llite_internal.h"
  51
  52 /* for obd_capa.c_list, client capa might stay in three places:
  53  * 1. ll_capa_list.
  54  * 2. ll_idle_capas.
  55  * 3. stand alone: just allocated.
  56  */
  57
  58 /* capas for oss writeback and those failed to renew */
  59 static struct list_head ll_idle_capas = LIST_HEAD_INIT(ll_idle_capas);
  60 static struct ptlrpc_thread ll_capa_thread;
  61 static struct list_head *ll_capa_list = &capa_list[CAPA_SITE_CLIENT];
  62
  63 /* llite capa renewal timer */
  64 struct timer_list ll_capa_timer;
  65 /* for debug: indicate whether capa on llite is enabled or not */
  66 static atomic_t ll_capa_debug = ATOMIC_INIT(0);
  67 static unsigned long long ll_capa_renewed = 0;
  68 static unsigned long long ll_capa_renewal_noent = 0;
  69 static unsigned long long ll_capa_renewal_failed = 0;
  70 static unsigned long long ll_capa_renewal_retries = 0;
  71
  72 static int ll_update_capa(struct obd_capa *ocapa, struct lustre_capa *capa);
  73
  74 static inline void update_capa_timer(struct obd_capa *ocapa, cfs_time_t expiry)
  75 {
  76         if (cfs_time_before(expiry, ll_capa_timer.expires) ||
  77             !timer_pending(&ll_capa_timer)) {
  78                 mod_timer(&ll_capa_timer, expiry);
  79                 DEBUG_CAPA(D_SEC, &ocapa->c_capa,
  80                            "ll_capa_timer update: %lu/%lu by", expiry, jiffies);
  81         }
  82 }
  83
  84 static inline cfs_time_t capa_renewal_time(struct obd_capa *ocapa)
  85 {
  86         return cfs_time_sub(ocapa->c_expiry,
  87                             cfs_time_seconds(ocapa->c_capa.lc_timeout) / 2);
  88 }
  89
  90 static inline int capa_is_to_expire(struct obd_capa *ocapa)
  91 {
  92         return cfs_time_beforeq(capa_renewal_time(ocapa), cfs_time_current());
  93 }
  94
  95 static inline int have_expired_capa(void)
  96 {
  97         struct obd_capa *ocapa = NULL;
  98         int expired = 0;
  99
 100         /* if ll_capa_list has client capa to expire or ll_idle_capas has
 101          * expired capa, return 1.
 102          */
 103         spin_lock(&capa_lock);
 104         if (!list_empty(ll_capa_list)) {
 105                 ocapa = list_entry(ll_capa_list->next, struct obd_capa,
 106                                        c_list);
 107                 expired = capa_is_to_expire(ocapa);
 108                 if (!expired)
 109                         update_capa_timer(ocapa, capa_renewal_time(ocapa));
 110         } else if (!list_empty(&ll_idle_capas)) {
 111                 ocapa = list_entry(ll_idle_capas.next, struct obd_capa,
 112                                        c_list);
 113                 expired = capa_is_expired(ocapa);
 114                 if (!expired)
 115                         update_capa_timer(ocapa, ocapa->c_expiry);
 116         }
 117         spin_unlock(&capa_lock);
 118
 119         if (expired)
 120                 DEBUG_CAPA(D_SEC, &ocapa->c_capa, "expired");
 121         return expired;
 122 }
 123
 124 static void sort_add_capa(struct obd_capa *ocapa, struct list_head *head)
 125 {
 126         struct obd_capa *tmp;
 127         struct list_head *before = NULL;
 128
 129         /* TODO: client capa is sorted by expiry, this could be optimized */
 130         list_for_each_entry_reverse(tmp, head, c_list) {
 131                 if (cfs_time_aftereq(ocapa->c_expiry, tmp->c_expiry)) {
 132                         before = &tmp->c_list;
 133                         break;
 134                 }
 135         }
 136
 137         LASSERT(&ocapa->c_list != before);
 138         list_add(&ocapa->c_list, before ?: head);
 139 }
 140
 141 static inline int obd_capa_open_count(struct obd_capa *oc)
 142 {
 143         struct ll_inode_info *lli = ll_i2info(oc->u.cli.inode);
 144         return atomic_read(&lli->lli_open_count);
 145 }
 146
 147 static void ll_delete_capa(struct obd_capa *ocapa)
 148 {
 149         struct ll_inode_info *lli = ll_i2info(ocapa->u.cli.inode);
 150
 151         if (capa_for_mds(&ocapa->c_capa)) {
 152                 LASSERT(lli->lli_mds_capa == ocapa);
 153                 lli->lli_mds_capa = NULL;
 154         } else if (capa_for_oss(&ocapa->c_capa)) {
 155                 list_del_init(&ocapa->u.cli.lli_list);
 156         }
 157
 158         DEBUG_CAPA(D_SEC, &ocapa->c_capa, "free client");
 159         list_del_init(&ocapa->c_list);
 160         capa_count[CAPA_SITE_CLIENT]--;
 161         /* release the ref when alloc */
 162         capa_put(ocapa);
 163 }
 164
 165 /* three places where client capa is deleted:
 166  * 1. capa_thread_main(), main place to delete expired capa.
 167  * 2. ll_clear_inode_capas() in ll_clear_inode().
 168  * 3. ll_truncate_free_capa() delete truncate capa explicitly in ll_setattr_ost().
 169  */
 170 static int capa_thread_main(void *unused)
 171 {
 172         struct obd_capa *ocapa, *tmp, *next;
 173         struct inode *inode = NULL;
 174         struct l_wait_info lwi = { 0 };
 175         int rc;
 176         ENTRY;
 177
 178         thread_set_flags(&ll_capa_thread, SVC_RUNNING);
 179         wake_up(&ll_capa_thread.t_ctl_waitq);
 180
 181         while (1) {
 182                 l_wait_event(ll_capa_thread.t_ctl_waitq,
 183                              !thread_is_running(&ll_capa_thread) ||
 184                              have_expired_capa(),
 185                              &lwi);
 186
 187                 if (!thread_is_running(&ll_capa_thread))
 188                         break;
 189
 190                 next = NULL;
 191
 192                 spin_lock(&capa_lock);
 193                 list_for_each_entry_safe(ocapa, tmp, ll_capa_list, c_list) {
 194                         __u64 ibits;
 195
 196                         LASSERT(ocapa->c_capa.lc_opc != CAPA_OPC_OSS_TRUNC);
 197
 198                         if (!capa_is_to_expire(ocapa)) {
 199                                 next = ocapa;
 200                                 break;
 201                         }
 202
 203                         list_del_init(&ocapa->c_list);
 204
 205                         /* for MDS capability, only renew those which belong to
 206                          * dir, or its inode is opened, or client holds LOOKUP
 207                          * lock.
 208                          */
 209                         /* ibits may be changed by ll_have_md_lock() so we have
 210                          * to set it each time */
 211                         ibits = MDS_INODELOCK_LOOKUP;
 212                         if (capa_for_mds(&ocapa->c_capa) &&
 213                             !S_ISDIR(ocapa->u.cli.inode->i_mode) &&
 214                             obd_capa_open_count(ocapa) == 0 &&
 215                             !ll_have_md_lock(ocapa->u.cli.inode,
 216                                              &ibits, LCK_MINMODE)) {
 217                                 DEBUG_CAPA(D_SEC, &ocapa->c_capa,
 218                                            "skip renewal for");
 219                                 sort_add_capa(ocapa, &ll_idle_capas);
 220                                 continue;
 221                         }
 222
 223                         /* for OSS capability, only renew those whose inode is
 224                          * opened.
 225                          */
 226                         if (capa_for_oss(&ocapa->c_capa) &&
 227                             obd_capa_open_count(ocapa) == 0) {
 228                                 /* oss capa with open count == 0 won't renew,
 229                                  * move to idle list */
 230                                 sort_add_capa(ocapa, &ll_idle_capas);
 231                                 continue;
 232                         }
 233
 234                         /* NB iput() is in ll_update_capa() */
 235                         inode = igrab(ocapa->u.cli.inode);
 236                         if (inode == NULL) {
 237                                 DEBUG_CAPA(D_ERROR, &ocapa->c_capa,
 238                                            "igrab failed for");
 239                                 continue;
 240                         }
 241
 242                         capa_get(ocapa);
 243                         ll_capa_renewed++;
 244                         spin_unlock(&capa_lock);
 245                         rc = md_renew_capa(ll_i2mdexp(inode), ocapa,
 246                                            ll_update_capa);
 247                         spin_lock(&capa_lock);
 248                         if (rc) {
 249                                 DEBUG_CAPA(D_ERROR, &ocapa->c_capa,
 250                                            "renew failed: %d", rc);
 251                                 ll_capa_renewal_failed++;
 252                         }
 253                 }
 254
 255                 if (next)
 256                         update_capa_timer(next, capa_renewal_time(next));
 257
 258                 list_for_each_entry_safe(ocapa, tmp, &ll_idle_capas,
 259                                              c_list) {
 260                         if (!capa_is_expired(ocapa)) {
 261                                 if (!next)
 262                                         update_capa_timer(ocapa,
 263                                                           ocapa->c_expiry);
 264                                 break;
 265                         }
 266
 267                         if (atomic_read(&ocapa->c_refc) > 1) {
 268                                 DEBUG_CAPA(D_SEC, &ocapa->c_capa,
 269                                            "expired(c_refc %d), don't release",
 270                                            atomic_read(&ocapa->c_refc));
 271                                 /* don't try to renew any more */
 272                                 list_del_init(&ocapa->c_list);
 273                                 continue;
 274                         }
 275
 276                         /* expired capa is released. */
 277                         DEBUG_CAPA(D_SEC, &ocapa->c_capa, "release expired");
 278                         ll_delete_capa(ocapa);
 279                 }
 280
 281                 spin_unlock(&capa_lock);
 282         }
 283
 284         thread_set_flags(&ll_capa_thread, SVC_STOPPED);
 285         wake_up(&ll_capa_thread.t_ctl_waitq);
 286         RETURN(0);
 287 }
 288
 289 void ll_capa_timer_callback(unsigned long unused)
 290 {
 291         wake_up(&ll_capa_thread.t_ctl_waitq);
 292 }
 293
 294 int ll_capa_thread_start(void)
 295 {
 296         struct task_struct *task;
 297         ENTRY;
 298
 299         init_waitqueue_head(&ll_capa_thread.t_ctl_waitq);
 300
 301         task = kthread_run(capa_thread_main, NULL, "ll_capa");
 302         if (IS_ERR(task)) {
 303                 CERROR("cannot start expired capa thread: rc %ld\n",
 304                         PTR_ERR(task));
 305                 RETURN(PTR_ERR(task));
 306         }
 307         wait_event(ll_capa_thread.t_ctl_waitq,
 308                        thread_is_running(&ll_capa_thread));
 309
 310         RETURN(0);
 311 }
 312
 313 void ll_capa_thread_stop(void)
 314 {
 315         thread_set_flags(&ll_capa_thread, SVC_STOPPING);
 316         wake_up(&ll_capa_thread.t_ctl_waitq);
 317         wait_event(ll_capa_thread.t_ctl_waitq,
 318                        thread_is_stopped(&ll_capa_thread));
 319 }
 320
 321 struct obd_capa *ll_mdscapa_get(struct inode *inode)
 322 {
 323         struct ll_inode_info *lli = ll_i2info(inode);
 324         struct obd_capa *ocapa;
 325         ENTRY;
 326
 327         LASSERT(inode != NULL);
 328
 329         if ((ll_i2sbi(inode)->ll_flags & LL_SBI_MDS_CAPA) == 0)
 330                 RETURN(NULL);
 331
 332         spin_lock(&capa_lock);
 333         ocapa = capa_get(lli->lli_mds_capa);
 334         spin_unlock(&capa_lock);
 335         if (!ocapa && atomic_read(&ll_capa_debug)) {
 336                 CERROR("no mds capability for "DFID"\n", PFID(&lli->lli_fid));
 337                 atomic_set(&ll_capa_debug, 0);
 338         }
 339
 340         RETURN(ocapa);
 341 }
 342
 343 static struct obd_capa *do_add_mds_capa(struct inode *inode,
 344                                         struct obd_capa *ocapa)
 345 {
 346         struct ll_inode_info *lli = ll_i2info(inode);
 347         struct obd_capa *old = lli->lli_mds_capa;
 348         struct lustre_capa *capa = &ocapa->c_capa;
 349
 350         if (!old) {
 351                 ocapa->u.cli.inode = inode;
 352                 lli->lli_mds_capa = ocapa;
 353                 capa_count[CAPA_SITE_CLIENT]++;
 354
 355                 DEBUG_CAPA(D_SEC, capa, "add MDS");
 356         } else {
 357                 spin_lock(&old->c_lock);
 358                 old->c_capa = *capa;
 359                 spin_unlock(&old->c_lock);
 360
 361                 DEBUG_CAPA(D_SEC, capa, "update MDS");
 362
 363                 capa_put(ocapa);
 364                 ocapa = old;
 365         }
 366         return ocapa;
 367 }
 368
 369 static struct obd_capa *do_lookup_oss_capa(struct inode *inode, int opc)
 370 {
 371         struct ll_inode_info *lli = ll_i2info(inode);
 372         struct obd_capa *ocapa;
 373
 374         /* inside capa_lock */
 375         list_for_each_entry(ocapa, &lli->lli_oss_capas, u.cli.lli_list) {
 376                 if ((capa_opc(&ocapa->c_capa) & opc) != opc)
 377                         continue;
 378
 379                 LASSERT(lu_fid_eq(capa_fid(&ocapa->c_capa),
 380                                   ll_inode2fid(inode)));
 381                 LASSERT(ocapa->c_site == CAPA_SITE_CLIENT);
 382
 383                 DEBUG_CAPA(D_SEC, &ocapa->c_capa, "found client");
 384                 return ocapa;
 385         }
 386
 387         return NULL;
 388 }
 389
 390 static inline void inode_add_oss_capa(struct inode *inode,
 391                                       struct obd_capa *ocapa)
 392 {
 393         struct ll_inode_info *lli = ll_i2info(inode);
 394         struct obd_capa *tmp;
 395         struct list_head *next = NULL;
 396
 397         /* capa is sorted in lli_oss_capas so lookup can always find the
 398          * latest one */
 399         list_for_each_entry(tmp, &lli->lli_oss_capas, u.cli.lli_list) {
 400                 if (cfs_time_after(ocapa->c_expiry, tmp->c_expiry)) {
 401                         next = &tmp->u.cli.lli_list;
 402                         break;
 403                 }
 404         }
 405         LASSERT(&ocapa->u.cli.lli_list != next);
 406         list_move_tail(&ocapa->u.cli.lli_list, next ?: &lli->lli_oss_capas);
 407 }
 408
 409 static struct obd_capa *do_add_oss_capa(struct inode *inode,
 410                                         struct obd_capa *ocapa)
 411 {
 412         struct obd_capa *old;
 413         struct lustre_capa *capa = &ocapa->c_capa;
 414
 415         LASSERTF(S_ISREG(inode->i_mode),
 416                  "inode has oss capa, but not regular file, mode: %d\n",
 417                  inode->i_mode);
 418
 419         /* FIXME: can't replace it so easily with fine-grained opc */
 420         old = do_lookup_oss_capa(inode, capa_opc(capa) & CAPA_OPC_OSS_ONLY);
 421         if (!old) {
 422                 ocapa->u.cli.inode = inode;
 423                 INIT_LIST_HEAD(&ocapa->u.cli.lli_list);
 424                 capa_count[CAPA_SITE_CLIENT]++;
 425
 426                 DEBUG_CAPA(D_SEC, capa, "add OSS");
 427         } else {
 428                 spin_lock(&old->c_lock);
 429                 old->c_capa = *capa;
 430                 spin_unlock(&old->c_lock);
 431
 432                 DEBUG_CAPA(D_SEC, capa, "update OSS");
 433
 434                 capa_put(ocapa);
 435                 ocapa = old;
 436         }
 437
 438         inode_add_oss_capa(inode, ocapa);
 439         return ocapa;
 440 }
 441
 442 struct obd_capa *ll_add_capa(struct inode *inode, struct obd_capa *ocapa)
 443 {
 444         spin_lock(&capa_lock);
 445         ocapa = capa_for_mds(&ocapa->c_capa) ? do_add_mds_capa(inode, ocapa) :
 446                                                do_add_oss_capa(inode, ocapa);
 447
 448         /* truncate capa won't renew */
 449         if (ocapa->c_capa.lc_opc != CAPA_OPC_OSS_TRUNC) {
 450                 set_capa_expiry(ocapa);
 451                 list_del_init(&ocapa->c_list);
 452                 sort_add_capa(ocapa, ll_capa_list);
 453
 454                 update_capa_timer(ocapa, capa_renewal_time(ocapa));
 455         }
 456
 457         spin_unlock(&capa_lock);
 458
 459         atomic_set(&ll_capa_debug, 1);
 460         return ocapa;
 461 }
 462
 463 static inline void delay_capa_renew(struct obd_capa *oc, cfs_time_t delay)
 464 {
 465         /* NB: set a fake expiry for this capa to prevent it renew too soon */
 466         oc->c_expiry = cfs_time_add(oc->c_expiry, cfs_time_seconds(delay));
 467 }
 468
 469 static int ll_update_capa(struct obd_capa *ocapa, struct lustre_capa *capa)
 470 {
 471         struct inode *inode = ocapa->u.cli.inode;
 472         int rc = 0;
 473         ENTRY;
 474
 475         LASSERT(ocapa);
 476
 477         if (IS_ERR(capa)) {
 478                 /* set error code */
 479                 rc = PTR_ERR(capa);
 480                 spin_lock(&capa_lock);
 481                 if (rc == -ENOENT) {
 482                         DEBUG_CAPA(D_SEC, &ocapa->c_capa,
 483                                    "renewal canceled because object removed");
 484                         ll_capa_renewal_noent++;
 485                 } else {
 486                         ll_capa_renewal_failed++;
 487
 488                         /* failed capa won't be renewed any longer, but if -EIO,
 489                          * client might be doing recovery, retry in 2 min. */
 490                         if (rc == -EIO && !capa_is_expired(ocapa)) {
 491                                 delay_capa_renew(ocapa, 120);
 492                                 DEBUG_CAPA(D_ERROR, &ocapa->c_capa,
 493                                            "renewal failed: -EIO, "
 494                                            "retry in 2 mins");
 495                                 ll_capa_renewal_retries++;
 496                                 GOTO(retry, rc);
 497                         } else {
 498                                 DEBUG_CAPA(D_ERROR, &ocapa->c_capa,
 499                                            "renewal failed(rc: %d) for", rc);
 500                         }
 501                 }
 502
 503                 list_del_init(&ocapa->c_list);
 504                 sort_add_capa(ocapa, &ll_idle_capas);
 505                 spin_unlock(&capa_lock);
 506
 507                 capa_put(ocapa);
 508                 iput(inode);
 509                 RETURN(rc);
 510         }
 511
 512         spin_lock(&ocapa->c_lock);
 513         LASSERT(!memcmp(&ocapa->c_capa, capa,
 514                         offsetof(struct lustre_capa, lc_opc)));
 515         ocapa->c_capa = *capa;
 516         set_capa_expiry(ocapa);
 517         spin_unlock(&ocapa->c_lock);
 518
 519         spin_lock(&capa_lock);
 520         if (capa_for_oss(capa))
 521                 inode_add_oss_capa(inode, ocapa);
 522         DEBUG_CAPA(D_SEC, capa, "renew");
 523         EXIT;
 524 retry:
 525         list_del_init(&ocapa->c_list);
 526         sort_add_capa(ocapa, ll_capa_list);
 527         update_capa_timer(ocapa, capa_renewal_time(ocapa));
 528         spin_unlock(&capa_lock);
 529
 530         capa_put(ocapa);
 531         iput(inode);
 532         return rc;
 533 }
 534
 535 void ll_capa_open(struct inode *inode)
 536 {
 537         struct ll_inode_info *lli = ll_i2info(inode);
 538
 539         if ((ll_i2sbi(inode)->ll_flags & (LL_SBI_MDS_CAPA | LL_SBI_OSS_CAPA))
 540             == 0)
 541                 return;
 542
 543         if (!S_ISREG(inode->i_mode))
 544                 return;
 545
 546         atomic_inc(&lli->lli_open_count);
 547 }
 548
 549 void ll_capa_close(struct inode *inode)
 550 {
 551         struct ll_inode_info *lli = ll_i2info(inode);
 552
 553         if ((ll_i2sbi(inode)->ll_flags & (LL_SBI_MDS_CAPA | LL_SBI_OSS_CAPA))
 554             == 0)
 555                 return;
 556
 557         if (!S_ISREG(inode->i_mode))
 558                 return;
 559
 560         atomic_dec(&lli->lli_open_count);
 561 }
 562
 563 /* delete CAPA_OPC_OSS_TRUNC only */
 564 void ll_truncate_free_capa(struct obd_capa *ocapa)
 565 {
 566         if (!ocapa)
 567                 return;
 568
 569         LASSERT(ocapa->c_capa.lc_opc & CAPA_OPC_OSS_TRUNC);
 570         DEBUG_CAPA(D_SEC, &ocapa->c_capa, "free truncate");
 571
 572         /* release ref when find */
 573         capa_put(ocapa);
 574         if (likely(ocapa->c_capa.lc_opc == CAPA_OPC_OSS_TRUNC)) {
 575                 spin_lock(&capa_lock);
 576                 ll_delete_capa(ocapa);
 577                 spin_unlock(&capa_lock);
 578         }
 579 }
 580
 581 void ll_clear_inode_capas(struct inode *inode)
 582 {
 583         struct ll_inode_info *lli = ll_i2info(inode);
 584         struct obd_capa *ocapa, *tmp;
 585
 586         spin_lock(&capa_lock);
 587         ocapa = lli->lli_mds_capa;
 588         if (ocapa)
 589                 ll_delete_capa(ocapa);
 590
 591         list_for_each_entry_safe(ocapa, tmp, &lli->lli_oss_capas,
 592                                      u.cli.lli_list)
 593                 ll_delete_capa(ocapa);
 594         spin_unlock(&capa_lock);
 595 }
 596
 597 void ll_print_capa_stat(struct ll_sb_info *sbi)
 598 {
 599         if (sbi->ll_flags & (LL_SBI_MDS_CAPA | LL_SBI_OSS_CAPA))
 600                 LCONSOLE_INFO("Fid capabilities renewed: %llu\n"
 601                               "Fid capabilities renewal ENOENT: %llu\n"
 602                               "Fid capabilities failed to renew: %llu\n"
 603                               "Fid capabilities renewal retries: %llu\n",
 604                               ll_capa_renewed, ll_capa_renewal_noent,
 605                               ll_capa_renewal_failed, ll_capa_renewal_retries);
 606 }