lustre/fld/fld_request.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  30  * Use is subject to license terms.
  31  *
  32  * Copyright (c) 2011, Whamcloud, Inc.
  33  */
  34 /*
  35  * This file is part of Lustre, http://www.lustre.org/
  36  * Lustre is a trademark of Sun Microsystems, Inc.
  37  *
  38  * lustre/fld/fld_request.c
  39  *
  40  * FLD (Fids Location Database)
  41  *
  42  * Author: Yury Umanets <umka@clusterfs.com>
  43  */
  44
  45 #ifndef EXPORT_SYMTAB
  46 # define EXPORT_SYMTAB
  47 #endif
  48 #define DEBUG_SUBSYSTEM S_FLD
  49
  50 #ifdef __KERNEL__
  51 # include <libcfs/libcfs.h>
  52 # include <linux/module.h>
  53 # include <linux/jbd.h>
  54 # include <asm/div64.h>
  55 #else /* __KERNEL__ */
  56 # include <liblustre.h>
  57 # include <libcfs/list.h>
  58 #endif
  59
  60 #include <obd.h>
  61 #include <obd_class.h>
  62 #include <lustre_ver.h>
  63 #include <obd_support.h>
  64 #include <lprocfs_status.h>
  65
  66 #include <dt_object.h>
  67 #include <md_object.h>
  68 #include <lustre_req_layout.h>
  69 #include <lustre_fld.h>
  70 #include <lustre_mdc.h>
  71 #include "fld_internal.h"
  72
  73 /* TODO: these 3 functions are copies of flow-control code from mdc_lib.c
  74  * It should be common thing. The same about mdc RPC lock */
  75 static int fld_req_avail(struct client_obd *cli, struct mdc_cache_waiter *mcw)
  76 {
  77         int rc;
  78         ENTRY;
  79         client_obd_list_lock(&cli->cl_loi_list_lock);
  80         rc = cfs_list_empty(&mcw->mcw_entry);
  81         client_obd_list_unlock(&cli->cl_loi_list_lock);
  82         RETURN(rc);
  83 };
  84
  85 static void fld_enter_request(struct client_obd *cli)
  86 {
  87         struct mdc_cache_waiter mcw;
  88         struct l_wait_info lwi = { 0 };
  89
  90         client_obd_list_lock(&cli->cl_loi_list_lock);
  91         if (cli->cl_r_in_flight >= cli->cl_max_rpcs_in_flight) {
  92                 cfs_list_add_tail(&mcw.mcw_entry, &cli->cl_cache_waiters);
  93                 cfs_waitq_init(&mcw.mcw_waitq);
  94                 client_obd_list_unlock(&cli->cl_loi_list_lock);
  95                 l_wait_event(mcw.mcw_waitq, fld_req_avail(cli, &mcw), &lwi);
  96         } else {
  97                 cli->cl_r_in_flight++;
  98                 client_obd_list_unlock(&cli->cl_loi_list_lock);
  99         }
 100 }
 101
 102 static void fld_exit_request(struct client_obd *cli)
 103 {
 104         cfs_list_t *l, *tmp;
 105         struct mdc_cache_waiter *mcw;
 106
 107         client_obd_list_lock(&cli->cl_loi_list_lock);
 108         cli->cl_r_in_flight--;
 109         cfs_list_for_each_safe(l, tmp, &cli->cl_cache_waiters) {
 110
 111                 if (cli->cl_r_in_flight >= cli->cl_max_rpcs_in_flight) {
 112                         /* No free request slots anymore */
 113                         break;
 114                 }
 115
 116                 mcw = cfs_list_entry(l, struct mdc_cache_waiter, mcw_entry);
 117                 cfs_list_del_init(&mcw->mcw_entry);
 118                 cli->cl_r_in_flight++;
 119                 cfs_waitq_signal(&mcw->mcw_waitq);
 120         }
 121         client_obd_list_unlock(&cli->cl_loi_list_lock);
 122 }
 123
 124 static int fld_rrb_hash(struct lu_client_fld *fld,
 125                         seqno_t seq)
 126 {
 127         LASSERT(fld->lcf_count > 0);
 128         return do_div(seq, fld->lcf_count);
 129 }
 130
 131 static struct lu_fld_target *
 132 fld_rrb_scan(struct lu_client_fld *fld, seqno_t seq)
 133 {
 134         struct lu_fld_target *target;
 135         int hash;
 136         ENTRY;
 137
 138         hash = fld_rrb_hash(fld, seq);
 139
 140         cfs_list_for_each_entry(target, &fld->lcf_targets, ft_chain) {
 141                 if (target->ft_idx == hash)
 142                         RETURN(target);
 143         }
 144
 145         CERROR("%s: Can't find target by hash %d (seq "LPX64"). "
 146                "Targets (%d):\n", fld->lcf_name, hash, seq,
 147                fld->lcf_count);
 148
 149         cfs_list_for_each_entry(target, &fld->lcf_targets, ft_chain) {
 150                 const char *srv_name = target->ft_srv != NULL  ?
 151                         target->ft_srv->lsf_name : "<null>";
 152                 const char *exp_name = target->ft_exp != NULL ?
 153                         (char *)target->ft_exp->exp_obd->obd_uuid.uuid :
 154                         "<null>";
 155
 156                 CERROR("  exp: 0x%p (%s), srv: 0x%p (%s), idx: "LPU64"\n",
 157                        target->ft_exp, exp_name, target->ft_srv,
 158                        srv_name, target->ft_idx);
 159         }
 160
 161         /*
 162          * If target is not found, there is logical error anyway, so here is
 163          * LBUG() to catch this situation.
 164          */
 165         LBUG();
 166         RETURN(NULL);
 167 }
 168
 169 struct lu_fld_hash fld_hash[] = {
 170         {
 171                 .fh_name = "RRB",
 172                 .fh_hash_func = fld_rrb_hash,
 173                 .fh_scan_func = fld_rrb_scan
 174         },
 175         {
 176                 0,
 177         }
 178 };
 179
 180 static struct lu_fld_target *
 181 fld_client_get_target(struct lu_client_fld *fld,
 182                       seqno_t seq)
 183 {
 184         struct lu_fld_target *target;
 185         ENTRY;
 186
 187         LASSERT(fld->lcf_hash != NULL);
 188
 189         cfs_spin_lock(&fld->lcf_lock);
 190         target = fld->lcf_hash->fh_scan_func(fld, seq);
 191         cfs_spin_unlock(&fld->lcf_lock);
 192
 193         if (target != NULL) {
 194                 CDEBUG(D_INFO, "%s: Found target (idx "LPU64
 195                        ") by seq "LPX64"\n", fld->lcf_name,
 196                        target->ft_idx, seq);
 197         }
 198
 199         RETURN(target);
 200 }
 201
 202 /*
 203  * Add export to FLD. This is usually done by CMM and LMV as they are main users
 204  * of FLD module.
 205  */
 206 int fld_client_add_target(struct lu_client_fld *fld,
 207                           struct lu_fld_target *tar)
 208 {
 209         const char *name = fld_target_name(tar);
 210         struct lu_fld_target *target, *tmp;
 211         ENTRY;
 212
 213         LASSERT(tar != NULL);
 214         LASSERT(name != NULL);
 215         LASSERT(tar->ft_srv != NULL || tar->ft_exp != NULL);
 216
 217         if (fld->lcf_flags != LUSTRE_FLD_INIT) {
 218                 CERROR("%s: Attempt to add target %s (idx "LPU64") "
 219                        "on fly - skip it\n", fld->lcf_name, name,
 220                        tar->ft_idx);
 221                 RETURN(0);
 222         } else {
 223                 CDEBUG(D_INFO, "%s: Adding target %s (idx "
 224                        LPU64")\n", fld->lcf_name, name, tar->ft_idx);
 225         }
 226
 227         OBD_ALLOC_PTR(target);
 228         if (target == NULL)
 229                 RETURN(-ENOMEM);
 230
 231         cfs_spin_lock(&fld->lcf_lock);
 232         cfs_list_for_each_entry(tmp, &fld->lcf_targets, ft_chain) {
 233                 if (tmp->ft_idx == tar->ft_idx) {
 234                         cfs_spin_unlock(&fld->lcf_lock);
 235                         OBD_FREE_PTR(target);
 236                         CERROR("Target %s exists in FLD and known as %s:#"LPU64"\n",
 237                                name, fld_target_name(tmp), tmp->ft_idx);
 238                         RETURN(-EEXIST);
 239                 }
 240         }
 241
 242         target->ft_exp = tar->ft_exp;
 243         if (target->ft_exp != NULL)
 244                 class_export_get(target->ft_exp);
 245         target->ft_srv = tar->ft_srv;
 246         target->ft_idx = tar->ft_idx;
 247
 248         cfs_list_add_tail(&target->ft_chain,
 249                           &fld->lcf_targets);
 250
 251         fld->lcf_count++;
 252         cfs_spin_unlock(&fld->lcf_lock);
 253
 254         RETURN(0);
 255 }
 256 EXPORT_SYMBOL(fld_client_add_target);
 257
 258 /* Remove export from FLD */
 259 int fld_client_del_target(struct lu_client_fld *fld,
 260                           __u64 idx)
 261 {
 262         struct lu_fld_target *target, *tmp;
 263         ENTRY;
 264
 265         cfs_spin_lock(&fld->lcf_lock);
 266         cfs_list_for_each_entry_safe(target, tmp,
 267                                      &fld->lcf_targets, ft_chain) {
 268                 if (target->ft_idx == idx) {
 269                         fld->lcf_count--;
 270                         cfs_list_del(&target->ft_chain);
 271                         cfs_spin_unlock(&fld->lcf_lock);
 272
 273                         if (target->ft_exp != NULL)
 274                                 class_export_put(target->ft_exp);
 275
 276                         OBD_FREE_PTR(target);
 277                         RETURN(0);
 278                 }
 279         }
 280         cfs_spin_unlock(&fld->lcf_lock);
 281         RETURN(-ENOENT);
 282 }
 283 EXPORT_SYMBOL(fld_client_del_target);
 284
 285 #ifdef LPROCFS
 286 static int fld_client_proc_init(struct lu_client_fld *fld)
 287 {
 288         int rc;
 289         ENTRY;
 290
 291         fld->lcf_proc_dir = lprocfs_register(fld->lcf_name,
 292                                              fld_type_proc_dir,
 293                                              NULL, NULL);
 294
 295         if (IS_ERR(fld->lcf_proc_dir)) {
 296                 CERROR("%s: LProcFS failed in fld-init\n",
 297                        fld->lcf_name);
 298                 rc = PTR_ERR(fld->lcf_proc_dir);
 299                 RETURN(rc);
 300         }
 301
 302         rc = lprocfs_add_vars(fld->lcf_proc_dir,
 303                               fld_client_proc_list, fld);
 304         if (rc) {
 305                 CERROR("%s: Can't init FLD proc, rc %d\n",
 306                        fld->lcf_name, rc);
 307                 GOTO(out_cleanup, rc);
 308         }
 309
 310         RETURN(0);
 311
 312 out_cleanup:
 313         fld_client_proc_fini(fld);
 314         return rc;
 315 }
 316
 317 void fld_client_proc_fini(struct lu_client_fld *fld)
 318 {
 319         ENTRY;
 320         if (fld->lcf_proc_dir) {
 321                 if (!IS_ERR(fld->lcf_proc_dir))
 322                         lprocfs_remove(&fld->lcf_proc_dir);
 323                 fld->lcf_proc_dir = NULL;
 324         }
 325         EXIT;
 326 }
 327 #else
 328 static int fld_client_proc_init(struct lu_client_fld *fld)
 329 {
 330         return 0;
 331 }
 332
 333 void fld_client_proc_fini(struct lu_client_fld *fld)
 334 {
 335         return;
 336 }
 337 #endif
 338
 339 EXPORT_SYMBOL(fld_client_proc_fini);
 340
 341 static inline int hash_is_sane(int hash)
 342 {
 343         return (hash >= 0 && hash < ARRAY_SIZE(fld_hash));
 344 }
 345
 346 int fld_client_init(struct lu_client_fld *fld,
 347                     const char *prefix, int hash)
 348 {
 349         int cache_size, cache_threshold;
 350         int rc;
 351         ENTRY;
 352
 353         LASSERT(fld != NULL);
 354
 355         snprintf(fld->lcf_name, sizeof(fld->lcf_name),
 356                  "cli-%s", prefix);
 357
 358         if (!hash_is_sane(hash)) {
 359                 CERROR("%s: Wrong hash function %#x\n",
 360                        fld->lcf_name, hash);
 361                 RETURN(-EINVAL);
 362         }
 363
 364         fld->lcf_count = 0;
 365         cfs_spin_lock_init(&fld->lcf_lock);
 366         fld->lcf_hash = &fld_hash[hash];
 367         fld->lcf_flags = LUSTRE_FLD_INIT;
 368         CFS_INIT_LIST_HEAD(&fld->lcf_targets);
 369
 370         cache_size = FLD_CLIENT_CACHE_SIZE /
 371                 sizeof(struct fld_cache_entry);
 372
 373         cache_threshold = cache_size *
 374                 FLD_CLIENT_CACHE_THRESHOLD / 100;
 375
 376         fld->lcf_cache = fld_cache_init(fld->lcf_name,
 377                                         cache_size, cache_threshold);
 378         if (IS_ERR(fld->lcf_cache)) {
 379                 rc = PTR_ERR(fld->lcf_cache);
 380                 fld->lcf_cache = NULL;
 381                 GOTO(out, rc);
 382         }
 383
 384         rc = fld_client_proc_init(fld);
 385         if (rc)
 386                 GOTO(out, rc);
 387         EXIT;
 388 out:
 389         if (rc)
 390                 fld_client_fini(fld);
 391         else
 392                 CDEBUG(D_INFO, "%s: Using \"%s\" hash\n",
 393                        fld->lcf_name, fld->lcf_hash->fh_name);
 394         return rc;
 395 }
 396 EXPORT_SYMBOL(fld_client_init);
 397
 398 void fld_client_fini(struct lu_client_fld *fld)
 399 {
 400         struct lu_fld_target *target, *tmp;
 401         ENTRY;
 402
 403         cfs_spin_lock(&fld->lcf_lock);
 404         cfs_list_for_each_entry_safe(target, tmp,
 405                                      &fld->lcf_targets, ft_chain) {
 406                 fld->lcf_count--;
 407                 cfs_list_del(&target->ft_chain);
 408                 if (target->ft_exp != NULL)
 409                         class_export_put(target->ft_exp);
 410                 OBD_FREE_PTR(target);
 411         }
 412         cfs_spin_unlock(&fld->lcf_lock);
 413
 414         if (fld->lcf_cache != NULL) {
 415                 if (!IS_ERR(fld->lcf_cache))
 416                         fld_cache_fini(fld->lcf_cache);
 417                 fld->lcf_cache = NULL;
 418         }
 419
 420         EXIT;
 421 }
 422 EXPORT_SYMBOL(fld_client_fini);
 423
 424 int fld_client_rpc(struct obd_export *exp,
 425                    struct lu_seq_range *range, __u32 fld_op)
 426 {
 427         struct ptlrpc_request *req;
 428         struct lu_seq_range      *prange;
 429         __u32                 *op;
 430         int                    rc;
 431         ENTRY;
 432
 433         LASSERT(exp != NULL);
 434
 435         req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp), &RQF_FLD_QUERY,
 436                                         LUSTRE_MDS_VERSION, FLD_QUERY);
 437         if (req == NULL)
 438                 RETURN(-ENOMEM);
 439
 440         op = req_capsule_client_get(&req->rq_pill, &RMF_FLD_OPC);
 441         *op = fld_op;
 442
 443         prange = req_capsule_client_get(&req->rq_pill, &RMF_FLD_MDFLD);
 444         *prange = *range;
 445
 446         ptlrpc_request_set_replen(req);
 447         req->rq_request_portal = FLD_REQUEST_PORTAL;
 448         ptlrpc_at_set_req_timeout(req);
 449
 450         if (fld_op != FLD_LOOKUP)
 451                 mdc_get_rpc_lock(exp->exp_obd->u.cli.cl_rpc_lock, NULL);
 452         fld_enter_request(&exp->exp_obd->u.cli);
 453         rc = ptlrpc_queue_wait(req);
 454         fld_exit_request(&exp->exp_obd->u.cli);
 455         if (fld_op != FLD_LOOKUP)
 456                 mdc_put_rpc_lock(exp->exp_obd->u.cli.cl_rpc_lock, NULL);
 457         if (rc)
 458                 GOTO(out_req, rc);
 459
 460         prange = req_capsule_server_get(&req->rq_pill, &RMF_FLD_MDFLD);
 461         if (prange == NULL)
 462                 GOTO(out_req, rc = -EFAULT);
 463         *range = *prange;
 464         EXIT;
 465 out_req:
 466         ptlrpc_req_finished(req);
 467         return rc;
 468 }
 469
 470 int fld_client_lookup(struct lu_client_fld *fld, seqno_t seq, mdsno_t *mds,
 471                       __u32 flags, const struct lu_env *env)
 472 {
 473         struct lu_seq_range res;
 474         struct lu_fld_target *target;
 475         int rc;
 476         ENTRY;
 477
 478         fld->lcf_flags |= LUSTRE_FLD_RUN;
 479
 480         rc = fld_cache_lookup(fld->lcf_cache, seq, &res);
 481         if (rc == 0) {
 482                 *mds = res.lsr_index;
 483                 RETURN(0);
 484         }
 485
 486         /* Can not find it in the cache */
 487         target = fld_client_get_target(fld, seq);
 488         LASSERT(target != NULL);
 489
 490         CDEBUG(D_INFO, "%s: Lookup fld entry (seq: "LPX64") on "
 491                "target %s (idx "LPU64")\n", fld->lcf_name, seq,
 492                fld_target_name(target), target->ft_idx);
 493
 494         res.lsr_start = seq;
 495         res.lsr_flags = flags;
 496 #ifdef __KERNEL__
 497         if (target->ft_srv != NULL) {
 498                 LASSERT(env != NULL);
 499                 rc = fld_server_lookup(target->ft_srv,
 500                                        env, seq, &res);
 501         } else {
 502 #endif
 503                 rc = fld_client_rpc(target->ft_exp,
 504                                     &res, FLD_LOOKUP);
 505 #ifdef __KERNEL__
 506         }
 507 #endif
 508
 509         if (rc == 0) {
 510                 *mds = res.lsr_index;
 511
 512                 fld_cache_insert(fld->lcf_cache, &res);
 513         }
 514         RETURN(rc);
 515 }
 516 EXPORT_SYMBOL(fld_client_lookup);
 517
 518 void fld_client_flush(struct lu_client_fld *fld)
 519 {
 520         fld_cache_flush(fld->lcf_cache);
 521 }
 522 EXPORT_SYMBOL(fld_client_flush);