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