drivers/staging/lustre/lustre/ldlm/ldlm_request.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) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2010, 2015, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  */
  32 /**
  33  * This file contains Asynchronous System Trap (AST) handlers and related
  34  * LDLM request-processing routines.
  35  *
  36  * An AST is a callback issued on a lock when its state is changed. There are
  37  * several different types of ASTs (callbacks) registered for each lock:
  38  *
  39  * - completion AST: when a lock is enqueued by some process, but cannot be
  40  *   granted immediately due to other conflicting locks on the same resource,
  41  *   the completion AST is sent to notify the caller when the lock is
  42  *   eventually granted
  43  *
  44  * - blocking AST: when a lock is granted to some process, if another process
  45  *   enqueues a conflicting (blocking) lock on a resource, a blocking AST is
  46  *   sent to notify the holder(s) of the lock(s) of the conflicting lock
  47  *   request. The lock holder(s) must release their lock(s) on that resource in
  48  *   a timely manner or be evicted by the server.
  49  *
  50  * - glimpse AST: this is used when a process wants information about a lock
  51  *   (i.e. the lock value block (LVB)) but does not necessarily require holding
  52  *   the lock. If the resource is locked, the lock holder(s) are sent glimpse
  53  *   ASTs and the LVB is returned to the caller, and lock holder(s) may CANCEL
  54  *   their lock(s) if they are idle. If the resource is not locked, the server
  55  *   may grant the lock.
  56  */
  57
  58 #define DEBUG_SUBSYSTEM S_LDLM
  59
  60 #include "../include/lustre_dlm.h"
  61 #include "../include/obd_class.h"
  62 #include "../include/obd.h"
  63
  64 #include "ldlm_internal.h"
  65
  66 unsigned int ldlm_enqueue_min = OBD_TIMEOUT_DEFAULT;
  67 module_param(ldlm_enqueue_min, uint, 0644);
  68 MODULE_PARM_DESC(ldlm_enqueue_min, "lock enqueue timeout minimum");
  69
  70 /* in client side, whether the cached locks will be canceled before replay */
  71 unsigned int ldlm_cancel_unused_locks_before_replay = 1;
  72
  73 static void interrupted_completion_wait(void *data)
  74 {
  75 }
  76
  77 struct lock_wait_data {
  78         struct ldlm_lock *lwd_lock;
  79         __u32        lwd_conn_cnt;
  80 };
  81
  82 struct ldlm_async_args {
  83         struct lustre_handle lock_handle;
  84 };
  85
  86 static int ldlm_expired_completion_wait(void *data)
  87 {
  88         struct lock_wait_data *lwd = data;
  89         struct ldlm_lock *lock = lwd->lwd_lock;
  90         struct obd_import *imp;
  91         struct obd_device *obd;
  92
  93         if (!lock->l_conn_export) {
  94                 static unsigned long next_dump, last_dump;
  95
  96                 LCONSOLE_WARN("lock timed out (enqueued at %lld, %llds ago)\n",
  97                               (s64)lock->l_last_activity,
  98                               (s64)(ktime_get_real_seconds() -
  99                                     lock->l_last_activity));
 100                 LDLM_DEBUG(lock, "lock timed out (enqueued at %lld, %llds ago); not entering recovery in server code, just going back to sleep",
 101                            (s64)lock->l_last_activity,
 102                            (s64)(ktime_get_real_seconds() -
 103                                  lock->l_last_activity));
 104                 if (cfs_time_after(cfs_time_current(), next_dump)) {
 105                         last_dump = next_dump;
 106                         next_dump = cfs_time_shift(300);
 107                         ldlm_namespace_dump(D_DLMTRACE,
 108                                             ldlm_lock_to_ns(lock));
 109                         if (last_dump == 0)
 110                                 libcfs_debug_dumplog();
 111                 }
 112                 return 0;
 113         }
 114
 115         obd = lock->l_conn_export->exp_obd;
 116         imp = obd->u.cli.cl_import;
 117         ptlrpc_fail_import(imp, lwd->lwd_conn_cnt);
 118         LDLM_ERROR(lock, "lock timed out (enqueued at %lld, %llds ago), entering recovery for %s@%s",
 119                    (s64)lock->l_last_activity,
 120                    (s64)(ktime_get_real_seconds() - lock->l_last_activity),
 121                    obd2cli_tgt(obd), imp->imp_connection->c_remote_uuid.uuid);
 122
 123         return 0;
 124 }
 125
 126 /**
 127  * Calculate the Completion timeout (covering enqueue, BL AST, data flush,
 128  * lock cancel, and their replies). Used for lock completion timeout on the
 129  * client side.
 130  *
 131  * \param[in] lock      lock which is waiting the completion callback
 132  *
 133  * \retval              timeout in seconds to wait for the server reply
 134  */
 135 /* We use the same basis for both server side and client side functions
 136  * from a single node.
 137  */
 138 static unsigned int ldlm_cp_timeout(struct ldlm_lock *lock)
 139 {
 140         unsigned int timeout;
 141
 142         if (AT_OFF)
 143                 return obd_timeout;
 144
 145         /*
 146          * Wait a long time for enqueue - server may have to callback a
 147          * lock from another client.  Server will evict the other client if it
 148          * doesn't respond reasonably, and then give us the lock.
 149          */
 150         timeout = at_get(ldlm_lock_to_ns_at(lock));
 151         return max(3 * timeout, ldlm_enqueue_min);
 152 }
 153
 154 /**
 155  * Helper function for ldlm_completion_ast(), updating timings when lock is
 156  * actually granted.
 157  */
 158 static int ldlm_completion_tail(struct ldlm_lock *lock, void *data)
 159 {
 160         long delay;
 161         int result = 0;
 162
 163         if (ldlm_is_destroyed(lock) || ldlm_is_failed(lock)) {
 164                 LDLM_DEBUG(lock, "client-side enqueue: destroyed");
 165                 result = -EIO;
 166         } else if (!data) {
 167                 LDLM_DEBUG(lock, "client-side enqueue: granted");
 168         } else {
 169                 /* Take into AT only CP RPC, not immediately granted locks */
 170                 delay = ktime_get_real_seconds() - lock->l_last_activity;
 171                 LDLM_DEBUG(lock, "client-side enqueue: granted after %lds",
 172                            delay);
 173
 174                 /* Update our time estimate */
 175                 at_measured(ldlm_lock_to_ns_at(lock), delay);
 176         }
 177         return result;
 178 }
 179
 180 /**
 181  * Implementation of ->l_completion_ast() for a client, that doesn't wait
 182  * until lock is granted. Suitable for locks enqueued through ptlrpcd, of
 183  * other threads that cannot block for long.
 184  */
 185 int ldlm_completion_ast_async(struct ldlm_lock *lock, __u64 flags, void *data)
 186 {
 187         if (flags == LDLM_FL_WAIT_NOREPROC) {
 188                 LDLM_DEBUG(lock, "client-side enqueue waiting on pending lock");
 189                 return 0;
 190         }
 191
 192         if (!(flags & LDLM_FL_BLOCKED_MASK)) {
 193                 wake_up(&lock->l_waitq);
 194                 return ldlm_completion_tail(lock, data);
 195         }
 196
 197         LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, going forward");
 198         return 0;
 199 }
 200 EXPORT_SYMBOL(ldlm_completion_ast_async);
 201
 202 /**
 203  * Generic LDLM "completion" AST. This is called in several cases:
 204  *
 205  *     - when a reply to an ENQUEUE RPC is received from the server
 206  *       (ldlm_cli_enqueue_fini()). Lock might be granted or not granted at
 207  *       this point (determined by flags);
 208  *
 209  *     - when LDLM_CP_CALLBACK RPC comes to client to notify it that lock has
 210  *       been granted;
 211  *
 212  *     - when ldlm_lock_match(LDLM_FL_LVB_READY) is about to wait until lock
 213  *       gets correct lvb;
 214  *
 215  *     - to force all locks when resource is destroyed (cleanup_resource());
 216  *
 217  *     - during lock conversion (not used currently).
 218  *
 219  * If lock is not granted in the first case, this function waits until second
 220  * or penultimate cases happen in some other thread.
 221  *
 222  */
 223 int ldlm_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data)
 224 {
 225         /* XXX ALLOCATE - 160 bytes */
 226         struct lock_wait_data lwd;
 227         struct obd_device *obd;
 228         struct obd_import *imp = NULL;
 229         struct l_wait_info lwi;
 230         __u32 timeout;
 231         int rc = 0;
 232
 233         if (flags == LDLM_FL_WAIT_NOREPROC) {
 234                 LDLM_DEBUG(lock, "client-side enqueue waiting on pending lock");
 235                 goto noreproc;
 236         }
 237
 238         if (!(flags & LDLM_FL_BLOCKED_MASK)) {
 239                 wake_up(&lock->l_waitq);
 240                 return 0;
 241         }
 242
 243         LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, sleeping");
 244
 245 noreproc:
 246
 247         obd = class_exp2obd(lock->l_conn_export);
 248
 249         /* if this is a local lock, then there is no import */
 250         if (obd)
 251                 imp = obd->u.cli.cl_import;
 252
 253         timeout = ldlm_cp_timeout(lock);
 254
 255         lwd.lwd_lock = lock;
 256         lock->l_last_activity = ktime_get_real_seconds();
 257
 258         if (ldlm_is_no_timeout(lock)) {
 259                 LDLM_DEBUG(lock, "waiting indefinitely because of NO_TIMEOUT");
 260                 lwi = LWI_INTR(interrupted_completion_wait, &lwd);
 261         } else {
 262                 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(timeout),
 263                                        ldlm_expired_completion_wait,
 264                                        interrupted_completion_wait, &lwd);
 265         }
 266
 267         if (imp) {
 268                 spin_lock(&imp->imp_lock);
 269                 lwd.lwd_conn_cnt = imp->imp_conn_cnt;
 270                 spin_unlock(&imp->imp_lock);
 271         }
 272
 273         if (OBD_FAIL_CHECK_RESET(OBD_FAIL_LDLM_INTR_CP_AST,
 274                                  OBD_FAIL_LDLM_CP_BL_RACE | OBD_FAIL_ONCE)) {
 275                 ldlm_set_fail_loc(lock);
 276                 rc = -EINTR;
 277         } else {
 278                 /* Go to sleep until the lock is granted or cancelled. */
 279                 rc = l_wait_event(lock->l_waitq,
 280                                   is_granted_or_cancelled(lock), &lwi);
 281         }
 282
 283         if (rc) {
 284                 LDLM_DEBUG(lock, "client-side enqueue waking up: failed (%d)",
 285                            rc);
 286                 return rc;
 287         }
 288
 289         return ldlm_completion_tail(lock, data);
 290 }
 291 EXPORT_SYMBOL(ldlm_completion_ast);
 292
 293 static void failed_lock_cleanup(struct ldlm_namespace *ns,
 294                                 struct ldlm_lock *lock, int mode)
 295 {
 296         int need_cancel = 0;
 297
 298         /* Set a flag to prevent us from sending a CANCEL (bug 407) */
 299         lock_res_and_lock(lock);
 300         /* Check that lock is not granted or failed, we might race. */
 301         if ((lock->l_req_mode != lock->l_granted_mode) &&
 302             !ldlm_is_failed(lock)) {
 303                 /* Make sure that this lock will not be found by raced
 304                  * bl_ast and -EINVAL reply is sent to server anyways.
 305                  * bug 17645
 306                  */
 307                 lock->l_flags |= LDLM_FL_LOCAL_ONLY | LDLM_FL_FAILED |
 308                                  LDLM_FL_ATOMIC_CB | LDLM_FL_CBPENDING;
 309                 need_cancel = 1;
 310         }
 311         unlock_res_and_lock(lock);
 312
 313         if (need_cancel)
 314                 LDLM_DEBUG(lock,
 315                            "setting FL_LOCAL_ONLY | LDLM_FL_FAILED | LDLM_FL_ATOMIC_CB | LDLM_FL_CBPENDING");
 316         else
 317                 LDLM_DEBUG(lock, "lock was granted or failed in race");
 318
 319         /* XXX - HACK because we shouldn't call ldlm_lock_destroy()
 320          *       from llite/file.c/ll_file_flock().
 321          */
 322         /* This code makes for the fact that we do not have blocking handler on
 323          * a client for flock locks. As such this is the place where we must
 324          * completely kill failed locks. (interrupted and those that
 325          * were waiting to be granted when server evicted us.
 326          */
 327         if (lock->l_resource->lr_type == LDLM_FLOCK) {
 328                 lock_res_and_lock(lock);
 329                 if (!ldlm_is_destroyed(lock)) {
 330                         ldlm_resource_unlink_lock(lock);
 331                         ldlm_lock_decref_internal_nolock(lock, mode);
 332                         ldlm_lock_destroy_nolock(lock);
 333                 }
 334                 unlock_res_and_lock(lock);
 335         } else {
 336                 ldlm_lock_decref_internal(lock, mode);
 337         }
 338 }
 339
 340 /**
 341  * Finishing portion of client lock enqueue code.
 342  *
 343  * Called after receiving reply from server.
 344  */
 345 int ldlm_cli_enqueue_fini(struct obd_export *exp, struct ptlrpc_request *req,
 346                           enum ldlm_type type, __u8 with_policy,
 347                           enum ldlm_mode mode,
 348                           __u64 *flags, void *lvb, __u32 lvb_len,
 349                           const struct lustre_handle *lockh, int rc)
 350 {
 351         struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
 352         int is_replay = *flags & LDLM_FL_REPLAY;
 353         struct ldlm_lock *lock;
 354         struct ldlm_reply *reply;
 355         int cleanup_phase = 1;
 356
 357         lock = ldlm_handle2lock(lockh);
 358         /* ldlm_cli_enqueue is holding a reference on this lock. */
 359         if (!lock) {
 360                 LASSERT(type == LDLM_FLOCK);
 361                 return -ENOLCK;
 362         }
 363
 364         LASSERTF(ergo(lvb_len != 0, lvb_len == lock->l_lvb_len),
 365                  "lvb_len = %d, l_lvb_len = %d\n", lvb_len, lock->l_lvb_len);
 366
 367         if (rc != ELDLM_OK) {
 368                 LASSERT(!is_replay);
 369                 LDLM_DEBUG(lock, "client-side enqueue END (%s)",
 370                            rc == ELDLM_LOCK_ABORTED ? "ABORTED" : "FAILED");
 371
 372                 if (rc != ELDLM_LOCK_ABORTED)
 373                         goto cleanup;
 374         }
 375
 376         /* Before we return, swab the reply */
 377         reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
 378         if (!reply) {
 379                 rc = -EPROTO;
 380                 goto cleanup;
 381         }
 382
 383         if (lvb_len > 0) {
 384                 int size = 0;
 385
 386                 size = req_capsule_get_size(&req->rq_pill, &RMF_DLM_LVB,
 387                                             RCL_SERVER);
 388                 if (size < 0) {
 389                         LDLM_ERROR(lock, "Fail to get lvb_len, rc = %d", size);
 390                         rc = size;
 391                         goto cleanup;
 392                 } else if (unlikely(size > lvb_len)) {
 393                         LDLM_ERROR(lock, "Replied LVB is larger than expectation, expected = %d, replied = %d",
 394                                    lvb_len, size);
 395                         rc = -EINVAL;
 396                         goto cleanup;
 397                 }
 398                 lvb_len = size;
 399         }
 400
 401         if (rc == ELDLM_LOCK_ABORTED) {
 402                 if (lvb_len > 0 && lvb)
 403                         rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_SERVER,
 404                                            lvb, lvb_len);
 405                 if (rc == 0)
 406                         rc = ELDLM_LOCK_ABORTED;
 407                 goto cleanup;
 408         }
 409
 410         /* lock enqueued on the server */
 411         cleanup_phase = 0;
 412
 413         lock_res_and_lock(lock);
 414         /* Key change rehash lock in per-export hash with new key */
 415         if (exp->exp_lock_hash) {
 416                 /* In the function below, .hs_keycmp resolves to
 417                  * ldlm_export_lock_keycmp()
 418                  */
 419                 /* coverity[overrun-buffer-val] */
 420                 cfs_hash_rehash_key(exp->exp_lock_hash,
 421                                     &lock->l_remote_handle,
 422                                     &reply->lock_handle,
 423                                     &lock->l_exp_hash);
 424         } else {
 425                 lock->l_remote_handle = reply->lock_handle;
 426         }
 427
 428         *flags = ldlm_flags_from_wire(reply->lock_flags);
 429         lock->l_flags |= ldlm_flags_from_wire(reply->lock_flags &
 430                                               LDLM_FL_INHERIT_MASK);
 431         unlock_res_and_lock(lock);
 432
 433         CDEBUG(D_INFO, "local: %p, remote cookie: %#llx, flags: 0x%llx\n",
 434                lock, reply->lock_handle.cookie, *flags);
 435
 436         /* If enqueue returned a blocked lock but the completion handler has
 437          * already run, then it fixed up the resource and we don't need to do it
 438          * again.
 439          */
 440         if ((*flags) & LDLM_FL_LOCK_CHANGED) {
 441                 int newmode = reply->lock_desc.l_req_mode;
 442
 443                 LASSERT(!is_replay);
 444                 if (newmode && newmode != lock->l_req_mode) {
 445                         LDLM_DEBUG(lock, "server returned different mode %s",
 446                                    ldlm_lockname[newmode]);
 447                         lock->l_req_mode = newmode;
 448                 }
 449
 450                 if (!ldlm_res_eq(&reply->lock_desc.l_resource.lr_name,
 451                                  &lock->l_resource->lr_name)) {
 452                         CDEBUG(D_INFO, "remote intent success, locking "DLDLMRES
 453                                        " instead of "DLDLMRES"\n",
 454                                PLDLMRES(&reply->lock_desc.l_resource),
 455                                PLDLMRES(lock->l_resource));
 456
 457                         rc = ldlm_lock_change_resource(ns, lock,
 458                                         &reply->lock_desc.l_resource.lr_name);
 459                         if (rc || !lock->l_resource) {
 460                                 rc = -ENOMEM;
 461                                 goto cleanup;
 462                         }
 463                         LDLM_DEBUG(lock, "client-side enqueue, new resource");
 464                 }
 465                 if (with_policy)
 466                         if (!(type == LDLM_IBITS &&
 467                               !(exp_connect_flags(exp) & OBD_CONNECT_IBITS)))
 468                                 /* We assume lock type cannot change on server*/
 469                                 ldlm_convert_policy_to_local(exp,
 470                                                 lock->l_resource->lr_type,
 471                                                 &reply->lock_desc.l_policy_data,
 472                                                 &lock->l_policy_data);
 473                 if (type != LDLM_PLAIN)
 474                         LDLM_DEBUG(lock,
 475                                    "client-side enqueue, new policy data");
 476         }
 477
 478         if ((*flags) & LDLM_FL_AST_SENT ||
 479             /* Cancel extent locks as soon as possible on a liblustre client,
 480              * because it cannot handle asynchronous ASTs robustly (see
 481              * bug 7311).
 482              */
 483             (LIBLUSTRE_CLIENT && type == LDLM_EXTENT)) {
 484                 lock_res_and_lock(lock);
 485                 lock->l_flags |= LDLM_FL_CBPENDING |  LDLM_FL_BL_AST;
 486                 unlock_res_and_lock(lock);
 487                 LDLM_DEBUG(lock, "enqueue reply includes blocking AST");
 488         }
 489
 490         /* If the lock has already been granted by a completion AST, don't
 491          * clobber the LVB with an older one.
 492          */
 493         if (lvb_len > 0) {
 494                 /* We must lock or a racing completion might update lvb without
 495                  * letting us know and we'll clobber the correct value.
 496                  * Cannot unlock after the check either, as that still leaves
 497                  * a tiny window for completion to get in
 498                  */
 499                 lock_res_and_lock(lock);
 500                 if (lock->l_req_mode != lock->l_granted_mode)
 501                         rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_SERVER,
 502                                            lock->l_lvb_data, lvb_len);
 503                 unlock_res_and_lock(lock);
 504                 if (rc < 0) {
 505                         cleanup_phase = 1;
 506                         goto cleanup;
 507                 }
 508         }
 509
 510         if (!is_replay) {
 511                 rc = ldlm_lock_enqueue(ns, &lock, NULL, flags);
 512                 if (lock->l_completion_ast) {
 513                         int err = lock->l_completion_ast(lock, *flags, NULL);
 514
 515                         if (!rc)
 516                                 rc = err;
 517                         if (rc)
 518                                 cleanup_phase = 1;
 519                 }
 520         }
 521
 522         if (lvb_len > 0 && lvb) {
 523                 /* Copy the LVB here, and not earlier, because the completion
 524                  * AST (if any) can override what we got in the reply
 525                  */
 526                 memcpy(lvb, lock->l_lvb_data, lvb_len);
 527         }
 528
 529         LDLM_DEBUG(lock, "client-side enqueue END");
 530 cleanup:
 531         if (cleanup_phase == 1 && rc)
 532                 failed_lock_cleanup(ns, lock, mode);
 533         /* Put lock 2 times, the second reference is held by ldlm_cli_enqueue */
 534         LDLM_LOCK_PUT(lock);
 535         LDLM_LOCK_RELEASE(lock);
 536         return rc;
 537 }
 538 EXPORT_SYMBOL(ldlm_cli_enqueue_fini);
 539
 540 /**
 541  * Estimate number of lock handles that would fit into request of given
 542  * size.  PAGE_SIZE-512 is to allow TCP/IP and LNET headers to fit into
 543  * a single page on the send/receive side. XXX: 512 should be changed to
 544  * more adequate value.
 545  */
 546 static inline int ldlm_req_handles_avail(int req_size, int off)
 547 {
 548         int avail;
 549
 550         avail = min_t(int, LDLM_MAXREQSIZE, PAGE_SIZE - 512) - req_size;
 551         if (likely(avail >= 0))
 552                 avail /= (int)sizeof(struct lustre_handle);
 553         else
 554                 avail = 0;
 555         avail += LDLM_LOCKREQ_HANDLES - off;
 556
 557         return avail;
 558 }
 559
 560 static inline int ldlm_capsule_handles_avail(struct req_capsule *pill,
 561                                              enum req_location loc,
 562                                              int off)
 563 {
 564         u32 size = req_capsule_msg_size(pill, loc);
 565
 566         return ldlm_req_handles_avail(size, off);
 567 }
 568
 569 static inline int ldlm_format_handles_avail(struct obd_import *imp,
 570                                             const struct req_format *fmt,
 571                                             enum req_location loc, int off)
 572 {
 573         u32 size = req_capsule_fmt_size(imp->imp_msg_magic, fmt, loc);
 574
 575         return ldlm_req_handles_avail(size, off);
 576 }
 577
 578 /**
 579  * Cancel LRU locks and pack them into the enqueue request. Pack there the given
 580  * \a count locks in \a cancels.
 581  *
 582  * This is to be called by functions preparing their own requests that
 583  * might contain lists of locks to cancel in addition to actual operation
 584  * that needs to be performed.
 585  */
 586 int ldlm_prep_elc_req(struct obd_export *exp, struct ptlrpc_request *req,
 587                       int version, int opc, int canceloff,
 588                       struct list_head *cancels, int count)
 589 {
 590         struct ldlm_namespace   *ns = exp->exp_obd->obd_namespace;
 591         struct req_capsule      *pill = &req->rq_pill;
 592         struct ldlm_request     *dlm = NULL;
 593         int flags, avail, to_free, pack = 0;
 594         LIST_HEAD(head);
 595         int rc;
 596
 597         if (!cancels)
 598                 cancels = &head;
 599         if (ns_connect_cancelset(ns)) {
 600                 /* Estimate the amount of available space in the request. */
 601                 req_capsule_filled_sizes(pill, RCL_CLIENT);
 602                 avail = ldlm_capsule_handles_avail(pill, RCL_CLIENT, canceloff);
 603
 604                 flags = ns_connect_lru_resize(ns) ?
 605                         LDLM_CANCEL_LRUR_NO_WAIT : LDLM_CANCEL_AGED;
 606                 to_free = !ns_connect_lru_resize(ns) &&
 607                           opc == LDLM_ENQUEUE ? 1 : 0;
 608
 609                 /* Cancel LRU locks here _only_ if the server supports
 610                  * EARLY_CANCEL. Otherwise we have to send extra CANCEL
 611                  * RPC, which will make us slower.
 612                  */
 613                 if (avail > count)
 614                         count += ldlm_cancel_lru_local(ns, cancels, to_free,
 615                                                        avail - count, 0, flags);
 616                 if (avail > count)
 617                         pack = count;
 618                 else
 619                         pack = avail;
 620                 req_capsule_set_size(pill, &RMF_DLM_REQ, RCL_CLIENT,
 621                                      ldlm_request_bufsize(pack, opc));
 622         }
 623
 624         rc = ptlrpc_request_pack(req, version, opc);
 625         if (rc) {
 626                 ldlm_lock_list_put(cancels, l_bl_ast, count);
 627                 return rc;
 628         }
 629
 630         if (ns_connect_cancelset(ns)) {
 631                 if (canceloff) {
 632                         dlm = req_capsule_client_get(pill, &RMF_DLM_REQ);
 633                         LASSERT(dlm);
 634                         /* Skip first lock handler in ldlm_request_pack(),
 635                          * this method will increment @lock_count according
 636                          * to the lock handle amount actually written to
 637                          * the buffer.
 638                          */
 639                         dlm->lock_count = canceloff;
 640                 }
 641                 /* Pack into the request @pack lock handles. */
 642                 ldlm_cli_cancel_list(cancels, pack, req, 0);
 643                 /* Prepare and send separate cancel RPC for others. */
 644                 ldlm_cli_cancel_list(cancels, count - pack, NULL, 0);
 645         } else {
 646                 ldlm_lock_list_put(cancels, l_bl_ast, count);
 647         }
 648         return 0;
 649 }
 650 EXPORT_SYMBOL(ldlm_prep_elc_req);
 651
 652 int ldlm_prep_enqueue_req(struct obd_export *exp, struct ptlrpc_request *req,
 653                           struct list_head *cancels, int count)
 654 {
 655         return ldlm_prep_elc_req(exp, req, LUSTRE_DLM_VERSION, LDLM_ENQUEUE,
 656                                  LDLM_ENQUEUE_CANCEL_OFF, cancels, count);
 657 }
 658 EXPORT_SYMBOL(ldlm_prep_enqueue_req);
 659
 660 /**
 661  * Client-side lock enqueue.
 662  *
 663  * If a request has some specific initialisation it is passed in \a reqp,
 664  * otherwise it is created in ldlm_cli_enqueue.
 665  *
 666  * Supports sync and async requests, pass \a async flag accordingly. If a
 667  * request was created in ldlm_cli_enqueue and it is the async request,
 668  * pass it to the caller in \a reqp.
 669  */
 670 int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp,
 671                      struct ldlm_enqueue_info *einfo,
 672                      const struct ldlm_res_id *res_id,
 673                      ldlm_policy_data_t const *policy, __u64 *flags,
 674                      void *lvb, __u32 lvb_len, enum lvb_type lvb_type,
 675                      struct lustre_handle *lockh, int async)
 676 {
 677         struct ldlm_namespace *ns;
 678         struct ldlm_lock      *lock;
 679         struct ldlm_request   *body;
 680         int                 is_replay = *flags & LDLM_FL_REPLAY;
 681         int                 req_passed_in = 1;
 682         int                 rc, err;
 683         struct ptlrpc_request *req;
 684
 685         ns = exp->exp_obd->obd_namespace;
 686
 687         /* If we're replaying this lock, just check some invariants.
 688          * If we're creating a new lock, get everything all setup nicely.
 689          */
 690         if (is_replay) {
 691                 lock = ldlm_handle2lock_long(lockh, 0);
 692                 LASSERT(lock);
 693                 LDLM_DEBUG(lock, "client-side enqueue START");
 694                 LASSERT(exp == lock->l_conn_export);
 695         } else {
 696                 const struct ldlm_callback_suite cbs = {
 697                         .lcs_completion = einfo->ei_cb_cp,
 698                         .lcs_blocking   = einfo->ei_cb_bl,
 699                         .lcs_glimpse    = einfo->ei_cb_gl
 700                 };
 701                 lock = ldlm_lock_create(ns, res_id, einfo->ei_type,
 702                                         einfo->ei_mode, &cbs, einfo->ei_cbdata,
 703                                         lvb_len, lvb_type);
 704                 if (IS_ERR(lock))
 705                         return PTR_ERR(lock);
 706                 /* for the local lock, add the reference */
 707                 ldlm_lock_addref_internal(lock, einfo->ei_mode);
 708                 ldlm_lock2handle(lock, lockh);
 709                 if (policy)
 710                         lock->l_policy_data = *policy;
 711
 712                 if (einfo->ei_type == LDLM_EXTENT) {
 713                         /* extent lock without policy is a bug */
 714                         if (!policy)
 715                                 LBUG();
 716
 717                         lock->l_req_extent = policy->l_extent;
 718                 }
 719                 LDLM_DEBUG(lock, "client-side enqueue START, flags %llx",
 720                            *flags);
 721         }
 722
 723         lock->l_conn_export = exp;
 724         lock->l_export = NULL;
 725         lock->l_blocking_ast = einfo->ei_cb_bl;
 726         lock->l_flags |= (*flags & (LDLM_FL_NO_LRU | LDLM_FL_EXCL));
 727         lock->l_last_activity = ktime_get_real_seconds();
 728
 729         /* lock not sent to server yet */
 730
 731         if (!reqp || !*reqp) {
 732                 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
 733                                                 &RQF_LDLM_ENQUEUE,
 734                                                 LUSTRE_DLM_VERSION,
 735                                                 LDLM_ENQUEUE);
 736                 if (!req) {
 737                         failed_lock_cleanup(ns, lock, einfo->ei_mode);
 738                         LDLM_LOCK_RELEASE(lock);
 739                         return -ENOMEM;
 740                 }
 741                 req_passed_in = 0;
 742                 if (reqp)
 743                         *reqp = req;
 744         } else {
 745                 int len;
 746
 747                 req = *reqp;
 748                 len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ,
 749                                            RCL_CLIENT);
 750                 LASSERTF(len >= sizeof(*body), "buflen[%d] = %d, not %d\n",
 751                          DLM_LOCKREQ_OFF, len, (int)sizeof(*body));
 752         }
 753
 754         /* Dump lock data into the request buffer */
 755         body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
 756         ldlm_lock2desc(lock, &body->lock_desc);
 757         body->lock_flags = ldlm_flags_to_wire(*flags);
 758         body->lock_handle[0] = *lockh;
 759
 760         /* Continue as normal. */
 761         if (!req_passed_in) {
 762                 if (lvb_len > 0)
 763                         req_capsule_extend(&req->rq_pill,
 764                                            &RQF_LDLM_ENQUEUE_LVB);
 765                 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
 766                                      lvb_len);
 767                 ptlrpc_request_set_replen(req);
 768         }
 769
 770         /*
 771          * Liblustre client doesn't get extent locks, except for O_APPEND case
 772          * where [0, OBD_OBJECT_EOF] lock is taken, or truncate, where
 773          * [i_size, OBD_OBJECT_EOF] lock is taken.
 774          */
 775         LASSERT(ergo(LIBLUSTRE_CLIENT, einfo->ei_type != LDLM_EXTENT ||
 776                      policy->l_extent.end == OBD_OBJECT_EOF));
 777
 778         if (async) {
 779                 LASSERT(reqp);
 780                 return 0;
 781         }
 782
 783         LDLM_DEBUG(lock, "sending request");
 784
 785         rc = ptlrpc_queue_wait(req);
 786
 787         err = ldlm_cli_enqueue_fini(exp, req, einfo->ei_type, policy ? 1 : 0,
 788                                     einfo->ei_mode, flags, lvb, lvb_len,
 789                                     lockh, rc);
 790
 791         /* If ldlm_cli_enqueue_fini did not find the lock, we need to free
 792          * one reference that we took
 793          */
 794         if (err == -ENOLCK)
 795                 LDLM_LOCK_RELEASE(lock);
 796         else
 797                 rc = err;
 798
 799         if (!req_passed_in && req) {
 800                 ptlrpc_req_finished(req);
 801                 if (reqp)
 802                         *reqp = NULL;
 803         }
 804
 805         return rc;
 806 }
 807 EXPORT_SYMBOL(ldlm_cli_enqueue);
 808
 809 /**
 810  * Cancel locks locally.
 811  * Returns:
 812  * \retval LDLM_FL_LOCAL_ONLY if there is no need for a CANCEL RPC to the server
 813  * \retval LDLM_FL_CANCELING otherwise;
 814  * \retval LDLM_FL_BL_AST if there is a need for a separate CANCEL RPC.
 815  */
 816 static __u64 ldlm_cli_cancel_local(struct ldlm_lock *lock)
 817 {
 818         __u64 rc = LDLM_FL_LOCAL_ONLY;
 819
 820         if (lock->l_conn_export) {
 821                 bool local_only;
 822
 823                 LDLM_DEBUG(lock, "client-side cancel");
 824                 /* Set this flag to prevent others from getting new references*/
 825                 lock_res_and_lock(lock);
 826                 ldlm_set_cbpending(lock);
 827                 local_only = !!(lock->l_flags &
 828                                 (LDLM_FL_LOCAL_ONLY | LDLM_FL_CANCEL_ON_BLOCK));
 829                 ldlm_cancel_callback(lock);
 830                 rc = ldlm_is_bl_ast(lock) ? LDLM_FL_BL_AST : LDLM_FL_CANCELING;
 831                 unlock_res_and_lock(lock);
 832
 833                 if (local_only) {
 834                         CDEBUG(D_DLMTRACE, "not sending request (at caller's instruction)\n");
 835                         rc = LDLM_FL_LOCAL_ONLY;
 836                 }
 837                 ldlm_lock_cancel(lock);
 838         } else {
 839                 LDLM_ERROR(lock, "Trying to cancel local lock");
 840                 LBUG();
 841         }
 842
 843         return rc;
 844 }
 845
 846 /**
 847  * Pack \a count locks in \a head into ldlm_request buffer of request \a req.
 848  */
 849 static void ldlm_cancel_pack(struct ptlrpc_request *req,
 850                              struct list_head *head, int count)
 851 {
 852         struct ldlm_request *dlm;
 853         struct ldlm_lock *lock;
 854         int max, packed = 0;
 855
 856         dlm = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
 857         LASSERT(dlm);
 858
 859         /* Check the room in the request buffer. */
 860         max = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT) -
 861                 sizeof(struct ldlm_request);
 862         max /= sizeof(struct lustre_handle);
 863         max += LDLM_LOCKREQ_HANDLES;
 864         LASSERT(max >= dlm->lock_count + count);
 865
 866         /* XXX: it would be better to pack lock handles grouped by resource.
 867          * so that the server cancel would call filter_lvbo_update() less
 868          * frequently.
 869          */
 870         list_for_each_entry(lock, head, l_bl_ast) {
 871                 if (!count--)
 872                         break;
 873                 LASSERT(lock->l_conn_export);
 874                 /* Pack the lock handle to the given request buffer. */
 875                 LDLM_DEBUG(lock, "packing");
 876                 dlm->lock_handle[dlm->lock_count++] = lock->l_remote_handle;
 877                 packed++;
 878         }
 879         CDEBUG(D_DLMTRACE, "%d locks packed\n", packed);
 880 }
 881
 882 /**
 883  * Prepare and send a batched cancel RPC. It will include \a count lock
 884  * handles of locks given in \a cancels list.
 885  */
 886 static int ldlm_cli_cancel_req(struct obd_export *exp,
 887                                struct list_head *cancels,
 888                                int count, enum ldlm_cancel_flags flags)
 889 {
 890         struct ptlrpc_request *req = NULL;
 891         struct obd_import *imp;
 892         int free, sent = 0;
 893         int rc = 0;
 894
 895         LASSERT(exp);
 896         LASSERT(count > 0);
 897
 898         CFS_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL, cfs_fail_val);
 899
 900         if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_RACE))
 901                 return count;
 902
 903         free = ldlm_format_handles_avail(class_exp2cliimp(exp),
 904                                          &RQF_LDLM_CANCEL, RCL_CLIENT, 0);
 905         if (count > free)
 906                 count = free;
 907
 908         while (1) {
 909                 imp = class_exp2cliimp(exp);
 910                 if (!imp || imp->imp_invalid) {
 911                         CDEBUG(D_DLMTRACE,
 912                                "skipping cancel on invalid import %p\n", imp);
 913                         return count;
 914                 }
 915
 916                 req = ptlrpc_request_alloc(imp, &RQF_LDLM_CANCEL);
 917                 if (!req) {
 918                         rc = -ENOMEM;
 919                         goto out;
 920                 }
 921
 922                 req_capsule_filled_sizes(&req->rq_pill, RCL_CLIENT);
 923                 req_capsule_set_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT,
 924                                      ldlm_request_bufsize(count, LDLM_CANCEL));
 925
 926                 rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CANCEL);
 927                 if (rc) {
 928                         ptlrpc_request_free(req);
 929                         goto out;
 930                 }
 931
 932                 req->rq_request_portal = LDLM_CANCEL_REQUEST_PORTAL;
 933                 req->rq_reply_portal = LDLM_CANCEL_REPLY_PORTAL;
 934                 ptlrpc_at_set_req_timeout(req);
 935
 936                 ldlm_cancel_pack(req, cancels, count);
 937
 938                 ptlrpc_request_set_replen(req);
 939                 if (flags & LCF_ASYNC) {
 940                         ptlrpcd_add_req(req);
 941                         sent = count;
 942                         goto out;
 943                 }
 944
 945                 rc = ptlrpc_queue_wait(req);
 946                 if (rc == LUSTRE_ESTALE) {
 947                         CDEBUG(D_DLMTRACE, "client/server (nid %s) out of sync -- not fatal\n",
 948                                libcfs_nid2str(req->rq_import->
 949                                               imp_connection->c_peer.nid));
 950                         rc = 0;
 951                 } else if (rc == -ETIMEDOUT && /* check there was no reconnect*/
 952                            req->rq_import_generation == imp->imp_generation) {
 953                         ptlrpc_req_finished(req);
 954                         continue;
 955                 } else if (rc != ELDLM_OK) {
 956                         /* -ESHUTDOWN is common on umount */
 957                         CDEBUG_LIMIT(rc == -ESHUTDOWN ? D_DLMTRACE : D_ERROR,
 958                                      "Got rc %d from cancel RPC: canceling anyway\n",
 959                                      rc);
 960                         break;
 961                 }
 962                 sent = count;
 963                 break;
 964         }
 965
 966         ptlrpc_req_finished(req);
 967 out:
 968         return sent ? sent : rc;
 969 }
 970
 971 static inline struct ldlm_pool *ldlm_imp2pl(struct obd_import *imp)
 972 {
 973         return &imp->imp_obd->obd_namespace->ns_pool;
 974 }
 975
 976 /**
 977  * Update client's OBD pool related fields with new SLV and Limit from \a req.
 978  */
 979 int ldlm_cli_update_pool(struct ptlrpc_request *req)
 980 {
 981         struct obd_device *obd;
 982         __u64 new_slv;
 983         __u32 new_limit;
 984
 985         if (unlikely(!req->rq_import || !req->rq_import->imp_obd ||
 986                      !imp_connect_lru_resize(req->rq_import))) {
 987                 /*
 988                  * Do nothing for corner cases.
 989                  */
 990                 return 0;
 991         }
 992
 993         /* In some cases RPC may contain SLV and limit zeroed out. This
 994          * is the case when server does not support LRU resize feature.
 995          * This is also possible in some recovery cases when server-side
 996          * reqs have no reference to the OBD export and thus access to
 997          * server-side namespace is not possible.
 998          */
 999         if (lustre_msg_get_slv(req->rq_repmsg) == 0 ||
1000             lustre_msg_get_limit(req->rq_repmsg) == 0) {
1001                 DEBUG_REQ(D_HA, req,
1002                           "Zero SLV or Limit found (SLV: %llu, Limit: %u)",
1003                           lustre_msg_get_slv(req->rq_repmsg),
1004                           lustre_msg_get_limit(req->rq_repmsg));
1005                 return 0;
1006         }
1007
1008         new_limit = lustre_msg_get_limit(req->rq_repmsg);
1009         new_slv = lustre_msg_get_slv(req->rq_repmsg);
1010         obd = req->rq_import->imp_obd;
1011
1012         /* Set new SLV and limit in OBD fields to make them accessible
1013          * to the pool thread. We do not access obd_namespace and pool
1014          * directly here as there is no reliable way to make sure that
1015          * they are still alive at cleanup time. Evil races are possible
1016          * which may cause Oops at that time.
1017          */
1018         write_lock(&obd->obd_pool_lock);
1019         obd->obd_pool_slv = new_slv;
1020         obd->obd_pool_limit = new_limit;
1021         write_unlock(&obd->obd_pool_lock);
1022
1023         return 0;
1024 }
1025 EXPORT_SYMBOL(ldlm_cli_update_pool);
1026
1027 /**
1028  * Client side lock cancel.
1029  *
1030  * Lock must not have any readers or writers by this time.
1031  */
1032 int ldlm_cli_cancel(const struct lustre_handle *lockh,
1033                     enum ldlm_cancel_flags cancel_flags)
1034 {
1035         struct obd_export *exp;
1036         int avail, flags, count = 1;
1037         __u64 rc = 0;
1038         struct ldlm_namespace *ns;
1039         struct ldlm_lock *lock;
1040         LIST_HEAD(cancels);
1041
1042         /* concurrent cancels on the same handle can happen */
1043         lock = ldlm_handle2lock_long(lockh, LDLM_FL_CANCELING);
1044         if (!lock) {
1045                 LDLM_DEBUG_NOLOCK("lock is already being destroyed");
1046                 return 0;
1047         }
1048
1049         rc = ldlm_cli_cancel_local(lock);
1050         if (rc == LDLM_FL_LOCAL_ONLY || cancel_flags & LCF_LOCAL) {
1051                 LDLM_LOCK_RELEASE(lock);
1052                 return 0;
1053         }
1054         /* Even if the lock is marked as LDLM_FL_BL_AST, this is a LDLM_CANCEL
1055          * RPC which goes to canceld portal, so we can cancel other LRU locks
1056          * here and send them all as one LDLM_CANCEL RPC.
1057          */
1058         LASSERT(list_empty(&lock->l_bl_ast));
1059         list_add(&lock->l_bl_ast, &cancels);
1060
1061         exp = lock->l_conn_export;
1062         if (exp_connect_cancelset(exp)) {
1063                 avail = ldlm_format_handles_avail(class_exp2cliimp(exp),
1064                                                   &RQF_LDLM_CANCEL,
1065                                                   RCL_CLIENT, 0);
1066                 LASSERT(avail > 0);
1067
1068                 ns = ldlm_lock_to_ns(lock);
1069                 flags = ns_connect_lru_resize(ns) ?
1070                         LDLM_CANCEL_LRUR : LDLM_CANCEL_AGED;
1071                 count += ldlm_cancel_lru_local(ns, &cancels, 0, avail - 1,
1072                                                LCF_BL_AST, flags);
1073         }
1074         ldlm_cli_cancel_list(&cancels, count, NULL, cancel_flags);
1075         return 0;
1076 }
1077 EXPORT_SYMBOL(ldlm_cli_cancel);
1078
1079 /**
1080  * Locally cancel up to \a count locks in list \a cancels.
1081  * Return the number of cancelled locks.
1082  */
1083 int ldlm_cli_cancel_list_local(struct list_head *cancels, int count,
1084                                enum ldlm_cancel_flags flags)
1085 {
1086         LIST_HEAD(head);
1087         struct ldlm_lock *lock, *next;
1088         int left = 0, bl_ast = 0;
1089         __u64 rc;
1090
1091         left = count;
1092         list_for_each_entry_safe(lock, next, cancels, l_bl_ast) {
1093                 if (left-- == 0)
1094                         break;
1095
1096                 if (flags & LCF_LOCAL) {
1097                         rc = LDLM_FL_LOCAL_ONLY;
1098                         ldlm_lock_cancel(lock);
1099                 } else {
1100                         rc = ldlm_cli_cancel_local(lock);
1101                 }
1102                 /* Until we have compound requests and can send LDLM_CANCEL
1103                  * requests batched with generic RPCs, we need to send cancels
1104                  * with the LDLM_FL_BL_AST flag in a separate RPC from
1105                  * the one being generated now.
1106                  */
1107                 if (!(flags & LCF_BL_AST) && (rc == LDLM_FL_BL_AST)) {
1108                         LDLM_DEBUG(lock, "Cancel lock separately");
1109                         list_del_init(&lock->l_bl_ast);
1110                         list_add(&lock->l_bl_ast, &head);
1111                         bl_ast++;
1112                         continue;
1113                 }
1114                 if (rc == LDLM_FL_LOCAL_ONLY) {
1115                         /* CANCEL RPC should not be sent to server. */
1116                         list_del_init(&lock->l_bl_ast);
1117                         LDLM_LOCK_RELEASE(lock);
1118                         count--;
1119                 }
1120         }
1121         if (bl_ast > 0) {
1122                 count -= bl_ast;
1123                 ldlm_cli_cancel_list(&head, bl_ast, NULL, 0);
1124         }
1125
1126         return count;
1127 }
1128 EXPORT_SYMBOL(ldlm_cli_cancel_list_local);
1129
1130 /**
1131  * Cancel as many locks as possible w/o sending any RPCs (e.g. to write back
1132  * dirty data, to close a file, ...) or waiting for any RPCs in-flight (e.g.
1133  * readahead requests, ...)
1134  */
1135 static enum ldlm_policy_res
1136 ldlm_cancel_no_wait_policy(struct ldlm_namespace *ns, struct ldlm_lock *lock,
1137                            int unused, int added, int count)
1138 {
1139         enum ldlm_policy_res result = LDLM_POLICY_CANCEL_LOCK;
1140
1141         /* don't check added & count since we want to process all locks
1142          * from unused list.
1143          * It's fine to not take lock to access lock->l_resource since
1144          * the lock has already been granted so it won't change.
1145          */
1146         switch (lock->l_resource->lr_type) {
1147         case LDLM_EXTENT:
1148         case LDLM_IBITS:
1149                 if (ns->ns_cancel && ns->ns_cancel(lock) != 0)
1150                         break;
1151         default:
1152                 result = LDLM_POLICY_SKIP_LOCK;
1153                 lock_res_and_lock(lock);
1154                 ldlm_set_skipped(lock);
1155                 unlock_res_and_lock(lock);
1156                 break;
1157         }
1158
1159         return result;
1160 }
1161
1162 /**
1163  * Callback function for LRU-resize policy. Decides whether to keep
1164  * \a lock in LRU for current \a LRU size \a unused, added in current
1165  * scan \a added and number of locks to be preferably canceled \a count.
1166  *
1167  * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1168  *
1169  * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1170  */
1171 static enum ldlm_policy_res ldlm_cancel_lrur_policy(struct ldlm_namespace *ns,
1172                                                     struct ldlm_lock *lock,
1173                                                     int unused, int added,
1174                                                     int count)
1175 {
1176         unsigned long cur = cfs_time_current();
1177         struct ldlm_pool *pl = &ns->ns_pool;
1178         __u64 slv, lvf, lv;
1179         unsigned long la;
1180
1181         /* Stop LRU processing when we reach past @count or have checked all
1182          * locks in LRU.
1183          */
1184         if (count && added >= count)
1185                 return LDLM_POLICY_KEEP_LOCK;
1186
1187         slv = ldlm_pool_get_slv(pl);
1188         lvf = ldlm_pool_get_lvf(pl);
1189         la = cfs_duration_sec(cfs_time_sub(cur, lock->l_last_used));
1190         lv = lvf * la * unused;
1191
1192         /* Inform pool about current CLV to see it via debugfs. */
1193         ldlm_pool_set_clv(pl, lv);
1194
1195         /* Stop when SLV is not yet come from server or lv is smaller than
1196          * it is.
1197          */
1198         if (slv == 0 || lv < slv)
1199                 return LDLM_POLICY_KEEP_LOCK;
1200
1201         return LDLM_POLICY_CANCEL_LOCK;
1202 }
1203
1204 /**
1205  * Callback function for debugfs used policy. Makes decision whether to keep
1206  * \a lock in LRU for current \a LRU size \a unused, added in current scan \a
1207  * added and number of locks to be preferably canceled \a count.
1208  *
1209  * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1210  *
1211  * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1212  */
1213 static enum ldlm_policy_res ldlm_cancel_passed_policy(struct ldlm_namespace *ns,
1214                                                       struct ldlm_lock *lock,
1215                                                       int unused, int added,
1216                                                       int count)
1217 {
1218         /* Stop LRU processing when we reach past @count or have checked all
1219          * locks in LRU.
1220          */
1221         return (added >= count) ?
1222                 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1223 }
1224
1225 /**
1226  * Callback function for aged policy. Makes decision whether to keep \a lock in
1227  * LRU for current LRU size \a unused, added in current scan \a added and
1228  * number of locks to be preferably canceled \a count.
1229  *
1230  * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1231  *
1232  * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1233  */
1234 static enum ldlm_policy_res ldlm_cancel_aged_policy(struct ldlm_namespace *ns,
1235                                                     struct ldlm_lock *lock,
1236                                                     int unused, int added,
1237                                                     int count)
1238 {
1239         if ((added >= count) &&
1240             time_before(cfs_time_current(),
1241                         cfs_time_add(lock->l_last_used, ns->ns_max_age)))
1242                 return LDLM_POLICY_KEEP_LOCK;
1243
1244         return LDLM_POLICY_CANCEL_LOCK;
1245 }
1246
1247 static enum ldlm_policy_res
1248 ldlm_cancel_lrur_no_wait_policy(struct ldlm_namespace *ns,
1249                                 struct ldlm_lock *lock,
1250                                 int unused, int added,
1251                                 int count)
1252 {
1253         enum ldlm_policy_res result;
1254
1255         result = ldlm_cancel_lrur_policy(ns, lock, unused, added, count);
1256         if (result == LDLM_POLICY_KEEP_LOCK)
1257                 return result;
1258
1259         return ldlm_cancel_no_wait_policy(ns, lock, unused, added, count);
1260 }
1261
1262 /**
1263  * Callback function for default policy. Makes decision whether to keep \a lock
1264  * in LRU for current LRU size \a unused, added in current scan \a added and
1265  * number of locks to be preferably canceled \a count.
1266  *
1267  * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1268  *
1269  * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1270  */
1271 static enum ldlm_policy_res
1272 ldlm_cancel_default_policy(struct ldlm_namespace *ns, struct ldlm_lock *lock,
1273                            int unused, int added, int count)
1274 {
1275         /* Stop LRU processing when we reach past count or have checked all
1276          * locks in LRU.
1277          */
1278         return (added >= count) ?
1279                 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1280 }
1281
1282 typedef enum ldlm_policy_res (*ldlm_cancel_lru_policy_t)(
1283                                                       struct ldlm_namespace *,
1284                                                       struct ldlm_lock *, int,
1285                                                       int, int);
1286
1287 static ldlm_cancel_lru_policy_t
1288 ldlm_cancel_lru_policy(struct ldlm_namespace *ns, int flags)
1289 {
1290         if (flags & LDLM_CANCEL_NO_WAIT)
1291                 return ldlm_cancel_no_wait_policy;
1292
1293         if (ns_connect_lru_resize(ns)) {
1294                 if (flags & LDLM_CANCEL_SHRINK)
1295                         /* We kill passed number of old locks. */
1296                         return ldlm_cancel_passed_policy;
1297                 else if (flags & LDLM_CANCEL_LRUR)
1298                         return ldlm_cancel_lrur_policy;
1299                 else if (flags & LDLM_CANCEL_PASSED)
1300                         return ldlm_cancel_passed_policy;
1301                 else if (flags & LDLM_CANCEL_LRUR_NO_WAIT)
1302                         return ldlm_cancel_lrur_no_wait_policy;
1303         } else {
1304                 if (flags & LDLM_CANCEL_AGED)
1305                         return ldlm_cancel_aged_policy;
1306         }
1307
1308         return ldlm_cancel_default_policy;
1309 }
1310
1311 /**
1312  * - Free space in LRU for \a count new locks,
1313  *   redundant unused locks are canceled locally;
1314  * - also cancel locally unused aged locks;
1315  * - do not cancel more than \a max locks;
1316  * - GET the found locks and add them into the \a cancels list.
1317  *
1318  * A client lock can be added to the l_bl_ast list only when it is
1319  * marked LDLM_FL_CANCELING. Otherwise, somebody is already doing
1320  * CANCEL.  There are the following use cases:
1321  * ldlm_cancel_resource_local(), ldlm_cancel_lru_local() and
1322  * ldlm_cli_cancel(), which check and set this flag properly. As any
1323  * attempt to cancel a lock rely on this flag, l_bl_ast list is accessed
1324  * later without any special locking.
1325  *
1326  * Calling policies for enabled LRU resize:
1327  * ----------------------------------------
1328  * flags & LDLM_CANCEL_LRUR - use LRU resize policy (SLV from server) to
1329  *                          cancel not more than \a count locks;
1330  *
1331  * flags & LDLM_CANCEL_PASSED - cancel \a count number of old locks (located at
1332  *                            the beginning of LRU list);
1333  *
1334  * flags & LDLM_CANCEL_SHRINK - cancel not more than \a count locks according to
1335  *                            memory pressure policy function;
1336  *
1337  * flags & LDLM_CANCEL_AGED - cancel \a count locks according to "aged policy".
1338  *
1339  * flags & LDLM_CANCEL_NO_WAIT - cancel as many unused locks as possible
1340  *                             (typically before replaying locks) w/o
1341  *                             sending any RPCs or waiting for any
1342  *                             outstanding RPC to complete.
1343  */
1344 static int ldlm_prepare_lru_list(struct ldlm_namespace *ns,
1345                                  struct list_head *cancels, int count, int max,
1346                                  int flags)
1347 {
1348         ldlm_cancel_lru_policy_t pf;
1349         struct ldlm_lock *lock, *next;
1350         int added = 0, unused, remained;
1351         int no_wait = flags & (LDLM_CANCEL_NO_WAIT | LDLM_CANCEL_LRUR_NO_WAIT);
1352
1353         spin_lock(&ns->ns_lock);
1354         unused = ns->ns_nr_unused;
1355         remained = unused;
1356
1357         if (!ns_connect_lru_resize(ns))
1358                 count += unused - ns->ns_max_unused;
1359
1360         pf = ldlm_cancel_lru_policy(ns, flags);
1361         LASSERT(pf);
1362
1363         while (!list_empty(&ns->ns_unused_list)) {
1364                 enum ldlm_policy_res result;
1365                 time_t last_use = 0;
1366
1367                 /* all unused locks */
1368                 if (remained-- <= 0)
1369                         break;
1370
1371                 /* For any flags, stop scanning if @max is reached. */
1372                 if (max && added >= max)
1373                         break;
1374
1375                 list_for_each_entry_safe(lock, next, &ns->ns_unused_list,
1376                                          l_lru) {
1377                         /* No locks which got blocking requests. */
1378                         LASSERT(!ldlm_is_bl_ast(lock));
1379
1380                         if (no_wait && ldlm_is_skipped(lock))
1381                                 /* already processed */
1382                                 continue;
1383
1384                         last_use = lock->l_last_used;
1385                         if (last_use == cfs_time_current())
1386                                 continue;
1387
1388                         /* Somebody is already doing CANCEL. No need for this
1389                          * lock in LRU, do not traverse it again.
1390                          */
1391                         if (!ldlm_is_canceling(lock))
1392                                 break;
1393
1394                         ldlm_lock_remove_from_lru_nolock(lock);
1395                 }
1396                 if (&lock->l_lru == &ns->ns_unused_list)
1397                         break;
1398
1399                 LDLM_LOCK_GET(lock);
1400                 spin_unlock(&ns->ns_lock);
1401                 lu_ref_add(&lock->l_reference, __func__, current);
1402
1403                 /* Pass the lock through the policy filter and see if it
1404                  * should stay in LRU.
1405                  *
1406                  * Even for shrinker policy we stop scanning if
1407                  * we find a lock that should stay in the cache.
1408                  * We should take into account lock age anyway
1409                  * as a new lock is a valuable resource even if
1410                  * it has a low weight.
1411                  *
1412                  * That is, for shrinker policy we drop only
1413                  * old locks, but additionally choose them by
1414                  * their weight. Big extent locks will stay in
1415                  * the cache.
1416                  */
1417                 result = pf(ns, lock, unused, added, count);
1418                 if (result == LDLM_POLICY_KEEP_LOCK) {
1419                         lu_ref_del(&lock->l_reference,
1420                                    __func__, current);
1421                         LDLM_LOCK_RELEASE(lock);
1422                         spin_lock(&ns->ns_lock);
1423                         break;
1424                 }
1425                 if (result == LDLM_POLICY_SKIP_LOCK) {
1426                         lu_ref_del(&lock->l_reference,
1427                                    __func__, current);
1428                         LDLM_LOCK_RELEASE(lock);
1429                         spin_lock(&ns->ns_lock);
1430                         continue;
1431                 }
1432
1433                 lock_res_and_lock(lock);
1434                 /* Check flags again under the lock. */
1435                 if (ldlm_is_canceling(lock) ||
1436                     (ldlm_lock_remove_from_lru_check(lock, last_use) == 0)) {
1437                         /* Another thread is removing lock from LRU, or
1438                          * somebody is already doing CANCEL, or there
1439                          * is a blocking request which will send cancel
1440                          * by itself, or the lock is no longer unused or
1441                          * the lock has been used since the pf() call and
1442                          * pages could be put under it.
1443                          */
1444                         unlock_res_and_lock(lock);
1445                         lu_ref_del(&lock->l_reference,
1446                                    __func__, current);
1447                         LDLM_LOCK_RELEASE(lock);
1448                         spin_lock(&ns->ns_lock);
1449                         continue;
1450                 }
1451                 LASSERT(!lock->l_readers && !lock->l_writers);
1452
1453                 /* If we have chosen to cancel this lock voluntarily, we
1454                  * better send cancel notification to server, so that it
1455                  * frees appropriate state. This might lead to a race
1456                  * where while we are doing cancel here, server is also
1457                  * silently cancelling this lock.
1458                  */
1459                 ldlm_clear_cancel_on_block(lock);
1460
1461                 /* Setting the CBPENDING flag is a little misleading,
1462                  * but prevents an important race; namely, once
1463                  * CBPENDING is set, the lock can accumulate no more
1464                  * readers/writers. Since readers and writers are
1465                  * already zero here, ldlm_lock_decref() won't see
1466                  * this flag and call l_blocking_ast
1467                  */
1468                 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING;
1469
1470                 /* We can't re-add to l_lru as it confuses the
1471                  * refcounting in ldlm_lock_remove_from_lru() if an AST
1472                  * arrives after we drop lr_lock below. We use l_bl_ast
1473                  * and can't use l_pending_chain as it is used both on
1474                  * server and client nevertheless bug 5666 says it is
1475                  * used only on server
1476                  */
1477                 LASSERT(list_empty(&lock->l_bl_ast));
1478                 list_add(&lock->l_bl_ast, cancels);
1479                 unlock_res_and_lock(lock);
1480                 lu_ref_del(&lock->l_reference, __func__, current);
1481                 spin_lock(&ns->ns_lock);
1482                 added++;
1483                 unused--;
1484         }
1485         spin_unlock(&ns->ns_lock);
1486         return added;
1487 }
1488
1489 int ldlm_cancel_lru_local(struct ldlm_namespace *ns,
1490                           struct list_head *cancels, int count, int max,
1491                           enum ldlm_cancel_flags cancel_flags, int flags)
1492 {
1493         int added;
1494
1495         added = ldlm_prepare_lru_list(ns, cancels, count, max, flags);
1496         if (added <= 0)
1497                 return added;
1498         return ldlm_cli_cancel_list_local(cancels, added, cancel_flags);
1499 }
1500
1501 /**
1502  * Cancel at least \a nr locks from given namespace LRU.
1503  *
1504  * When called with LCF_ASYNC the blocking callback will be handled
1505  * in a thread and this function will return after the thread has been
1506  * asked to call the callback.  When called with LCF_ASYNC the blocking
1507  * callback will be performed in this function.
1508  */
1509 int ldlm_cancel_lru(struct ldlm_namespace *ns, int nr,
1510                     enum ldlm_cancel_flags cancel_flags,
1511                     int flags)
1512 {
1513         LIST_HEAD(cancels);
1514         int count, rc;
1515
1516         /* Just prepare the list of locks, do not actually cancel them yet.
1517          * Locks are cancelled later in a separate thread.
1518          */
1519         count = ldlm_prepare_lru_list(ns, &cancels, nr, 0, flags);
1520         rc = ldlm_bl_to_thread_list(ns, NULL, &cancels, count, cancel_flags);
1521         if (rc == 0)
1522                 return count;
1523
1524         return 0;
1525 }
1526
1527 /**
1528  * Find and cancel locally unused locks found on resource, matched to the
1529  * given policy, mode. GET the found locks and add them into the \a cancels
1530  * list.
1531  */
1532 int ldlm_cancel_resource_local(struct ldlm_resource *res,
1533                                struct list_head *cancels,
1534                                ldlm_policy_data_t *policy,
1535                                enum ldlm_mode mode, __u64 lock_flags,
1536                                enum ldlm_cancel_flags cancel_flags,
1537                                void *opaque)
1538 {
1539         struct ldlm_lock *lock;
1540         int count = 0;
1541
1542         lock_res(res);
1543         list_for_each_entry(lock, &res->lr_granted, l_res_link) {
1544                 if (opaque && lock->l_ast_data != opaque) {
1545                         LDLM_ERROR(lock, "data %p doesn't match opaque %p",
1546                                    lock->l_ast_data, opaque);
1547                         continue;
1548                 }
1549
1550                 if (lock->l_readers || lock->l_writers)
1551                         continue;
1552
1553                 /* If somebody is already doing CANCEL, or blocking AST came,
1554                  * skip this lock.
1555                  */
1556                 if (ldlm_is_bl_ast(lock) || ldlm_is_canceling(lock))
1557                         continue;
1558
1559                 if (lockmode_compat(lock->l_granted_mode, mode))
1560                         continue;
1561
1562                 /* If policy is given and this is IBITS lock, add to list only
1563                  * those locks that match by policy.
1564                  */
1565                 if (policy && (lock->l_resource->lr_type == LDLM_IBITS) &&
1566                     !(lock->l_policy_data.l_inodebits.bits &
1567                       policy->l_inodebits.bits))
1568                         continue;
1569
1570                 /* See CBPENDING comment in ldlm_cancel_lru */
1571                 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING |
1572                                  lock_flags;
1573
1574                 LASSERT(list_empty(&lock->l_bl_ast));
1575                 list_add(&lock->l_bl_ast, cancels);
1576                 LDLM_LOCK_GET(lock);
1577                 count++;
1578         }
1579         unlock_res(res);
1580
1581         return ldlm_cli_cancel_list_local(cancels, count, cancel_flags);
1582 }
1583 EXPORT_SYMBOL(ldlm_cancel_resource_local);
1584
1585 /**
1586  * Cancel client-side locks from a list and send/prepare cancel RPCs to the
1587  * server.
1588  * If \a req is NULL, send CANCEL request to server with handles of locks
1589  * in the \a cancels. If EARLY_CANCEL is not supported, send CANCEL requests
1590  * separately per lock.
1591  * If \a req is not NULL, put handles of locks in \a cancels into the request
1592  * buffer at the offset \a off.
1593  * Destroy \a cancels at the end.
1594  */
1595 int ldlm_cli_cancel_list(struct list_head *cancels, int count,
1596                          struct ptlrpc_request *req,
1597                          enum ldlm_cancel_flags flags)
1598 {
1599         struct ldlm_lock *lock;
1600         int res = 0;
1601
1602         if (list_empty(cancels) || count == 0)
1603                 return 0;
1604
1605         /* XXX: requests (both batched and not) could be sent in parallel.
1606          * Usually it is enough to have just 1 RPC, but it is possible that
1607          * there are too many locks to be cancelled in LRU or on a resource.
1608          * It would also speed up the case when the server does not support
1609          * the feature.
1610          */
1611         while (count > 0) {
1612                 LASSERT(!list_empty(cancels));
1613                 lock = list_entry(cancels->next, struct ldlm_lock, l_bl_ast);
1614                 LASSERT(lock->l_conn_export);
1615
1616                 if (exp_connect_cancelset(lock->l_conn_export)) {
1617                         res = count;
1618                         if (req)
1619                                 ldlm_cancel_pack(req, cancels, count);
1620                         else
1621                                 res = ldlm_cli_cancel_req(lock->l_conn_export,
1622                                                           cancels, count,
1623                                                           flags);
1624                 } else {
1625                         res = ldlm_cli_cancel_req(lock->l_conn_export,
1626                                                   cancels, 1, flags);
1627                 }
1628
1629                 if (res < 0) {
1630                         CDEBUG_LIMIT(res == -ESHUTDOWN ? D_DLMTRACE : D_ERROR,
1631                                      "ldlm_cli_cancel_list: %d\n", res);
1632                         res = count;
1633                 }
1634
1635                 count -= res;
1636                 ldlm_lock_list_put(cancels, l_bl_ast, res);
1637         }
1638         LASSERT(count == 0);
1639         return 0;
1640 }
1641 EXPORT_SYMBOL(ldlm_cli_cancel_list);
1642
1643 /**
1644  * Cancel all locks on a resource that have 0 readers/writers.
1645  *
1646  * If flags & LDLM_FL_LOCAL_ONLY, throw the locks away without trying
1647  * to notify the server.
1648  */
1649 int ldlm_cli_cancel_unused_resource(struct ldlm_namespace *ns,
1650                                     const struct ldlm_res_id *res_id,
1651                                     ldlm_policy_data_t *policy,
1652                                     enum ldlm_mode mode,
1653                                     enum ldlm_cancel_flags flags,
1654                                     void *opaque)
1655 {
1656         struct ldlm_resource *res;
1657         LIST_HEAD(cancels);
1658         int count;
1659         int rc;
1660
1661         res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
1662         if (IS_ERR(res)) {
1663                 /* This is not a problem. */
1664                 CDEBUG(D_INFO, "No resource %llu\n", res_id->name[0]);
1665                 return 0;
1666         }
1667
1668         LDLM_RESOURCE_ADDREF(res);
1669         count = ldlm_cancel_resource_local(res, &cancels, policy, mode,
1670                                            0, flags | LCF_BL_AST, opaque);
1671         rc = ldlm_cli_cancel_list(&cancels, count, NULL, flags);
1672         if (rc != ELDLM_OK)
1673                 CERROR("canceling unused lock "DLDLMRES": rc = %d\n",
1674                        PLDLMRES(res), rc);
1675
1676         LDLM_RESOURCE_DELREF(res);
1677         ldlm_resource_putref(res);
1678         return 0;
1679 }
1680 EXPORT_SYMBOL(ldlm_cli_cancel_unused_resource);
1681
1682 struct ldlm_cli_cancel_arg {
1683         int     lc_flags;
1684         void   *lc_opaque;
1685 };
1686
1687 static int ldlm_cli_hash_cancel_unused(struct cfs_hash *hs,
1688                                        struct cfs_hash_bd *bd,
1689                                        struct hlist_node *hnode, void *arg)
1690 {
1691         struct ldlm_resource       *res = cfs_hash_object(hs, hnode);
1692         struct ldlm_cli_cancel_arg     *lc = arg;
1693
1694         ldlm_cli_cancel_unused_resource(ldlm_res_to_ns(res), &res->lr_name,
1695                                         NULL, LCK_MINMODE,
1696                                         lc->lc_flags, lc->lc_opaque);
1697         /* must return 0 for hash iteration */
1698         return 0;
1699 }
1700
1701 /**
1702  * Cancel all locks on a namespace (or a specific resource, if given)
1703  * that have 0 readers/writers.
1704  *
1705  * If flags & LCF_LOCAL, throw the locks away without trying
1706  * to notify the server.
1707  */
1708 int ldlm_cli_cancel_unused(struct ldlm_namespace *ns,
1709                            const struct ldlm_res_id *res_id,
1710                            enum ldlm_cancel_flags flags, void *opaque)
1711 {
1712         struct ldlm_cli_cancel_arg arg = {
1713                 .lc_flags       = flags,
1714                 .lc_opaque      = opaque,
1715         };
1716
1717         if (!ns)
1718                 return ELDLM_OK;
1719
1720         if (res_id) {
1721                 return ldlm_cli_cancel_unused_resource(ns, res_id, NULL,
1722                                                        LCK_MINMODE, flags,
1723                                                        opaque);
1724         } else {
1725                 cfs_hash_for_each_nolock(ns->ns_rs_hash,
1726                                          ldlm_cli_hash_cancel_unused, &arg);
1727                 return ELDLM_OK;
1728         }
1729 }
1730 EXPORT_SYMBOL(ldlm_cli_cancel_unused);
1731
1732 /* Lock iterators. */
1733
1734 static int ldlm_resource_foreach(struct ldlm_resource *res,
1735                                  ldlm_iterator_t iter, void *closure)
1736 {
1737         struct list_head *tmp, *next;
1738         struct ldlm_lock *lock;
1739         int rc = LDLM_ITER_CONTINUE;
1740
1741         if (!res)
1742                 return LDLM_ITER_CONTINUE;
1743
1744         lock_res(res);
1745         list_for_each_safe(tmp, next, &res->lr_granted) {
1746                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1747
1748                 if (iter(lock, closure) == LDLM_ITER_STOP) {
1749                         rc = LDLM_ITER_STOP;
1750                         goto out;
1751                 }
1752         }
1753
1754         list_for_each_safe(tmp, next, &res->lr_waiting) {
1755                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1756
1757                 if (iter(lock, closure) == LDLM_ITER_STOP) {
1758                         rc = LDLM_ITER_STOP;
1759                         goto out;
1760                 }
1761         }
1762  out:
1763         unlock_res(res);
1764         return rc;
1765 }
1766
1767 struct iter_helper_data {
1768         ldlm_iterator_t iter;
1769         void *closure;
1770 };
1771
1772 static int ldlm_iter_helper(struct ldlm_lock *lock, void *closure)
1773 {
1774         struct iter_helper_data *helper = closure;
1775
1776         return helper->iter(lock, helper->closure);
1777 }
1778
1779 static int ldlm_res_iter_helper(struct cfs_hash *hs, struct cfs_hash_bd *bd,
1780                                 struct hlist_node *hnode, void *arg)
1781
1782 {
1783         struct ldlm_resource *res = cfs_hash_object(hs, hnode);
1784
1785         return ldlm_resource_foreach(res, ldlm_iter_helper, arg) ==
1786                LDLM_ITER_STOP;
1787 }
1788
1789 static void ldlm_namespace_foreach(struct ldlm_namespace *ns,
1790                                    ldlm_iterator_t iter, void *closure)
1791
1792 {
1793         struct iter_helper_data helper = {
1794                 .iter           = iter,
1795                 .closure        = closure,
1796         };
1797
1798         cfs_hash_for_each_nolock(ns->ns_rs_hash,
1799                                  ldlm_res_iter_helper, &helper);
1800 }
1801
1802 /* non-blocking function to manipulate a lock whose cb_data is being put away.
1803  * return  0:  find no resource
1804  *       > 0:  must be LDLM_ITER_STOP/LDLM_ITER_CONTINUE.
1805  *       < 0:  errors
1806  */
1807 int ldlm_resource_iterate(struct ldlm_namespace *ns,
1808                           const struct ldlm_res_id *res_id,
1809                           ldlm_iterator_t iter, void *data)
1810 {
1811         struct ldlm_resource *res;
1812         int rc;
1813
1814         LASSERTF(ns, "must pass in namespace\n");
1815
1816         res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
1817         if (IS_ERR(res))
1818                 return 0;
1819
1820         LDLM_RESOURCE_ADDREF(res);
1821         rc = ldlm_resource_foreach(res, iter, data);
1822         LDLM_RESOURCE_DELREF(res);
1823         ldlm_resource_putref(res);
1824         return rc;
1825 }
1826 EXPORT_SYMBOL(ldlm_resource_iterate);
1827
1828 /* Lock replay */
1829
1830 static int ldlm_chain_lock_for_replay(struct ldlm_lock *lock, void *closure)
1831 {
1832         struct list_head *list = closure;
1833
1834         /* we use l_pending_chain here, because it's unused on clients. */
1835         LASSERTF(list_empty(&lock->l_pending_chain),
1836                  "lock %p next %p prev %p\n",
1837                  lock, &lock->l_pending_chain.next,
1838                  &lock->l_pending_chain.prev);
1839         /* bug 9573: don't replay locks left after eviction, or
1840          * bug 17614: locks being actively cancelled. Get a reference
1841          * on a lock so that it does not disappear under us (e.g. due to cancel)
1842          */
1843         if (!(lock->l_flags & (LDLM_FL_FAILED | LDLM_FL_CANCELING))) {
1844                 list_add(&lock->l_pending_chain, list);
1845                 LDLM_LOCK_GET(lock);
1846         }
1847
1848         return LDLM_ITER_CONTINUE;
1849 }
1850
1851 static int replay_lock_interpret(const struct lu_env *env,
1852                                  struct ptlrpc_request *req,
1853                                  struct ldlm_async_args *aa, int rc)
1854 {
1855         struct ldlm_lock     *lock;
1856         struct ldlm_reply    *reply;
1857         struct obd_export    *exp;
1858
1859         atomic_dec(&req->rq_import->imp_replay_inflight);
1860         if (rc != ELDLM_OK)
1861                 goto out;
1862
1863         reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1864         if (!reply) {
1865                 rc = -EPROTO;
1866                 goto out;
1867         }
1868
1869         lock = ldlm_handle2lock(&aa->lock_handle);
1870         if (!lock) {
1871                 CERROR("received replay ack for unknown local cookie %#llx remote cookie %#llx from server %s id %s\n",
1872                        aa->lock_handle.cookie, reply->lock_handle.cookie,
1873                        req->rq_export->exp_client_uuid.uuid,
1874                        libcfs_id2str(req->rq_peer));
1875                 rc = -ESTALE;
1876                 goto out;
1877         }
1878
1879         /* Key change rehash lock in per-export hash with new key */
1880         exp = req->rq_export;
1881         if (exp && exp->exp_lock_hash) {
1882                 /* In the function below, .hs_keycmp resolves to
1883                  * ldlm_export_lock_keycmp()
1884                  */
1885                 /* coverity[overrun-buffer-val] */
1886                 cfs_hash_rehash_key(exp->exp_lock_hash,
1887                                     &lock->l_remote_handle,
1888                                     &reply->lock_handle,
1889                                     &lock->l_exp_hash);
1890         } else {
1891                 lock->l_remote_handle = reply->lock_handle;
1892         }
1893
1894         LDLM_DEBUG(lock, "replayed lock:");
1895         ptlrpc_import_recovery_state_machine(req->rq_import);
1896         LDLM_LOCK_PUT(lock);
1897 out:
1898         if (rc != ELDLM_OK)
1899                 ptlrpc_connect_import(req->rq_import);
1900
1901         return rc;
1902 }
1903
1904 static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock)
1905 {
1906         struct ptlrpc_request *req;
1907         struct ldlm_async_args *aa;
1908         struct ldlm_request   *body;
1909         int flags;
1910
1911         /* Bug 11974: Do not replay a lock which is actively being canceled */
1912         if (ldlm_is_canceling(lock)) {
1913                 LDLM_DEBUG(lock, "Not replaying canceled lock:");
1914                 return 0;
1915         }
1916
1917         /* If this is reply-less callback lock, we cannot replay it, since
1918          * server might have long dropped it, but notification of that event was
1919          * lost by network. (and server granted conflicting lock already)
1920          */
1921         if (ldlm_is_cancel_on_block(lock)) {
1922                 LDLM_DEBUG(lock, "Not replaying reply-less lock:");
1923                 ldlm_lock_cancel(lock);
1924                 return 0;
1925         }
1926
1927         /*
1928          * If granted mode matches the requested mode, this lock is granted.
1929          *
1930          * If they differ, but we have a granted mode, then we were granted
1931          * one mode and now want another: ergo, converting.
1932          *
1933          * If we haven't been granted anything and are on a resource list,
1934          * then we're blocked/waiting.
1935          *
1936          * If we haven't been granted anything and we're NOT on a resource list,
1937          * then we haven't got a reply yet and don't have a known disposition.
1938          * This happens whenever a lock enqueue is the request that triggers
1939          * recovery.
1940          */
1941         if (lock->l_granted_mode == lock->l_req_mode)
1942                 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_GRANTED;
1943         else if (lock->l_granted_mode)
1944                 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_CONV;
1945         else if (!list_empty(&lock->l_res_link))
1946                 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_WAIT;
1947         else
1948                 flags = LDLM_FL_REPLAY;
1949
1950         req = ptlrpc_request_alloc_pack(imp, &RQF_LDLM_ENQUEUE,
1951                                         LUSTRE_DLM_VERSION, LDLM_ENQUEUE);
1952         if (!req)
1953                 return -ENOMEM;
1954
1955         /* We're part of recovery, so don't wait for it. */
1956         req->rq_send_state = LUSTRE_IMP_REPLAY_LOCKS;
1957
1958         body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1959         ldlm_lock2desc(lock, &body->lock_desc);
1960         body->lock_flags = ldlm_flags_to_wire(flags);
1961
1962         ldlm_lock2handle(lock, &body->lock_handle[0]);
1963         if (lock->l_lvb_len > 0)
1964                 req_capsule_extend(&req->rq_pill, &RQF_LDLM_ENQUEUE_LVB);
1965         req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
1966                              lock->l_lvb_len);
1967         ptlrpc_request_set_replen(req);
1968         /* notify the server we've replayed all requests.
1969          * also, we mark the request to be put on a dedicated
1970          * queue to be processed after all request replayes.
1971          * bug 6063
1972          */
1973         lustre_msg_set_flags(req->rq_reqmsg, MSG_REQ_REPLAY_DONE);
1974
1975         LDLM_DEBUG(lock, "replaying lock:");
1976
1977         atomic_inc(&req->rq_import->imp_replay_inflight);
1978         CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
1979         aa = ptlrpc_req_async_args(req);
1980         aa->lock_handle = body->lock_handle[0];
1981         req->rq_interpret_reply = (ptlrpc_interpterer_t)replay_lock_interpret;
1982         ptlrpcd_add_req(req);
1983
1984         return 0;
1985 }
1986
1987 /**
1988  * Cancel as many unused locks as possible before replay. since we are
1989  * in recovery, we can't wait for any outstanding RPCs to send any RPC
1990  * to the server.
1991  *
1992  * Called only in recovery before replaying locks. there is no need to
1993  * replay locks that are unused. since the clients may hold thousands of
1994  * cached unused locks, dropping the unused locks can greatly reduce the
1995  * load on the servers at recovery time.
1996  */
1997 static void ldlm_cancel_unused_locks_for_replay(struct ldlm_namespace *ns)
1998 {
1999         int canceled;
2000         LIST_HEAD(cancels);
2001
2002         CDEBUG(D_DLMTRACE, "Dropping as many unused locks as possible before replay for namespace %s (%d)\n",
2003                ldlm_ns_name(ns), ns->ns_nr_unused);
2004
2005         /* We don't need to care whether or not LRU resize is enabled
2006          * because the LDLM_CANCEL_NO_WAIT policy doesn't use the
2007          * count parameter
2008          */
2009         canceled = ldlm_cancel_lru_local(ns, &cancels, ns->ns_nr_unused, 0,
2010                                          LCF_LOCAL, LDLM_CANCEL_NO_WAIT);
2011
2012         CDEBUG(D_DLMTRACE, "Canceled %d unused locks from namespace %s\n",
2013                canceled, ldlm_ns_name(ns));
2014 }
2015
2016 int ldlm_replay_locks(struct obd_import *imp)
2017 {
2018         struct ldlm_namespace *ns = imp->imp_obd->obd_namespace;
2019         LIST_HEAD(list);
2020         struct ldlm_lock *lock, *next;
2021         int rc = 0;
2022
2023         LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
2024
2025         /* don't replay locks if import failed recovery */
2026         if (imp->imp_vbr_failed)
2027                 return 0;
2028
2029         /* ensure this doesn't fall to 0 before all have been queued */
2030         atomic_inc(&imp->imp_replay_inflight);
2031
2032         if (ldlm_cancel_unused_locks_before_replay)
2033                 ldlm_cancel_unused_locks_for_replay(ns);
2034
2035         ldlm_namespace_foreach(ns, ldlm_chain_lock_for_replay, &list);
2036
2037         list_for_each_entry_safe(lock, next, &list, l_pending_chain) {
2038                 list_del_init(&lock->l_pending_chain);
2039                 if (rc) {
2040                         LDLM_LOCK_RELEASE(lock);
2041                         continue; /* or try to do the rest? */
2042                 }
2043                 rc = replay_one_lock(imp, lock);
2044                 LDLM_LOCK_RELEASE(lock);
2045         }
2046
2047         atomic_dec(&imp->imp_replay_inflight);
2048
2049         return rc;
2050 }
2051 EXPORT_SYMBOL(ldlm_replay_locks);