lustre/fld/fld_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) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2013, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lustre/fld/fld_request.c
  37  *
  38  * FLD (Fids Location Database)
  39  *
  40  * Author: Yury Umanets <umka@clusterfs.com>
  41  */
  42
  43 #define DEBUG_SUBSYSTEM S_FLD
  44
  45 #ifdef __KERNEL__
  46 # include <libcfs/libcfs.h>
  47 # include <linux/module.h>
  48 # include <linux/jbd.h>
  49 # include <asm/div64.h>
  50 #else /* __KERNEL__ */
  51 # include <liblustre.h>
  52 # include <libcfs/list.h>
  53 #endif
  54
  55 #include <obd.h>
  56 #include <obd_class.h>
  57 #include <lustre_ver.h>
  58 #include <obd_support.h>
  59 #include <lprocfs_status.h>
  60
  61 #include <dt_object.h>
  62 #include <md_object.h>
  63 #include <lustre_req_layout.h>
  64 #include <lustre_fld.h>
  65 #include <lustre_mdc.h>
  66 #include "fld_internal.h"
  67
  68 /* TODO: these 3 functions are copies of flow-control code from mdc_lib.c
  69  * It should be common thing. The same about mdc RPC lock */
  70 static int fld_req_avail(struct client_obd *cli, struct mdc_cache_waiter *mcw)
  71 {
  72         int rc;
  73         ENTRY;
  74         client_obd_list_lock(&cli->cl_loi_list_lock);
  75         rc = cfs_list_empty(&mcw->mcw_entry);
  76         client_obd_list_unlock(&cli->cl_loi_list_lock);
  77         RETURN(rc);
  78 };
  79
  80 static void fld_enter_request(struct client_obd *cli)
  81 {
  82         struct mdc_cache_waiter mcw;
  83         struct l_wait_info lwi = { 0 };
  84
  85         client_obd_list_lock(&cli->cl_loi_list_lock);
  86         if (cli->cl_r_in_flight >= cli->cl_max_rpcs_in_flight) {
  87                 cfs_list_add_tail(&mcw.mcw_entry, &cli->cl_cache_waiters);
  88                 cfs_waitq_init(&mcw.mcw_waitq);
  89                 client_obd_list_unlock(&cli->cl_loi_list_lock);
  90                 l_wait_event(mcw.mcw_waitq, fld_req_avail(cli, &mcw), &lwi);
  91         } else {
  92                 cli->cl_r_in_flight++;
  93                 client_obd_list_unlock(&cli->cl_loi_list_lock);
  94         }
  95 }
  96
  97 static void fld_exit_request(struct client_obd *cli)
  98 {
  99         cfs_list_t *l, *tmp;
 100         struct mdc_cache_waiter *mcw;
 101
 102         client_obd_list_lock(&cli->cl_loi_list_lock);
 103         cli->cl_r_in_flight--;
 104         cfs_list_for_each_safe(l, tmp, &cli->cl_cache_waiters) {
 105
 106                 if (cli->cl_r_in_flight >= cli->cl_max_rpcs_in_flight) {
 107                         /* No free request slots anymore */
 108                         break;
 109                 }
 110
 111                 mcw = cfs_list_entry(l, struct mdc_cache_waiter, mcw_entry);
 112                 cfs_list_del_init(&mcw->mcw_entry);
 113                 cli->cl_r_in_flight++;
 114                 cfs_waitq_signal(&mcw->mcw_waitq);
 115         }
 116         client_obd_list_unlock(&cli->cl_loi_list_lock);
 117 }
 118
 119 static int fld_rrb_hash(struct lu_client_fld *fld,
 120                         seqno_t seq)
 121 {
 122         LASSERT(fld->lcf_count > 0);
 123         return do_div(seq, fld->lcf_count);
 124 }
 125
 126 static struct lu_fld_target *
 127 fld_rrb_scan(struct lu_client_fld *fld, seqno_t seq)
 128 {
 129         struct lu_fld_target *target;
 130         int hash;
 131         ENTRY;
 132
 133         hash = fld_rrb_hash(fld, seq);
 134
 135         cfs_list_for_each_entry(target, &fld->lcf_targets, ft_chain) {
 136                 if (target->ft_idx == hash)
 137                         RETURN(target);
 138         }
 139
 140         CERROR("%s: Can't find target by hash %d (seq "LPX64"). "
 141                "Targets (%d):\n", fld->lcf_name, hash, seq,
 142                fld->lcf_count);
 143
 144         cfs_list_for_each_entry(target, &fld->lcf_targets, ft_chain) {
 145                 const char *srv_name = target->ft_srv != NULL  ?
 146                         target->ft_srv->lsf_name : "<null>";
 147                 const char *exp_name = target->ft_exp != NULL ?
 148                         (char *)target->ft_exp->exp_obd->obd_uuid.uuid :
 149                         "<null>";
 150
 151                 CERROR("  exp: 0x%p (%s), srv: 0x%p (%s), idx: "LPU64"\n",
 152                        target->ft_exp, exp_name, target->ft_srv,
 153                        srv_name, target->ft_idx);
 154         }
 155
 156         /*
 157          * If target is not found, there is logical error anyway, so here is
 158          * LBUG() to catch this situation.
 159          */
 160         LBUG();
 161         RETURN(NULL);
 162 }
 163
 164 struct lu_fld_hash fld_hash[] = {
 165         {
 166                 .fh_name = "RRB",
 167                 .fh_hash_func = fld_rrb_hash,
 168                 .fh_scan_func = fld_rrb_scan
 169         },
 170         {
 171                 0,
 172         }
 173 };
 174
 175 static struct lu_fld_target *
 176 fld_client_get_target(struct lu_client_fld *fld, seqno_t seq)
 177 {
 178         struct lu_fld_target *target;
 179         ENTRY;
 180
 181         LASSERT(fld->lcf_hash != NULL);
 182
 183         spin_lock(&fld->lcf_lock);
 184         target = fld->lcf_hash->fh_scan_func(fld, seq);
 185         spin_unlock(&fld->lcf_lock);
 186
 187         if (target != NULL) {
 188                 CDEBUG(D_INFO, "%s: Found target (idx "LPU64
 189                        ") by seq "LPX64"\n", fld->lcf_name,
 190                        target->ft_idx, seq);
 191         }
 192
 193         RETURN(target);
 194 }
 195
 196 /*
 197  * Add export to FLD. This is usually done by CMM and LMV as they are main users
 198  * of FLD module.
 199  */
 200 int fld_client_add_target(struct lu_client_fld *fld,
 201                           struct lu_fld_target *tar)
 202 {
 203         const char *name;
 204         struct lu_fld_target *target, *tmp;
 205         ENTRY;
 206
 207         LASSERT(tar != NULL);
 208         name = fld_target_name(tar);
 209         LASSERT(name != NULL);
 210         LASSERT(tar->ft_srv != NULL || tar->ft_exp != NULL);
 211
 212         if (fld->lcf_flags != LUSTRE_FLD_INIT) {
 213                 CERROR("%s: Attempt to add target %s (idx "LPU64") "
 214                        "on fly - skip it\n", fld->lcf_name, name,
 215                        tar->ft_idx);
 216                 RETURN(0);
 217         } else {
 218                 CDEBUG(D_INFO, "%s: Adding target %s (idx "
 219                        LPU64")\n", fld->lcf_name, name, tar->ft_idx);
 220         }
 221
 222         OBD_ALLOC_PTR(target);
 223         if (target == NULL)
 224                 RETURN(-ENOMEM);
 225
 226         spin_lock(&fld->lcf_lock);
 227         cfs_list_for_each_entry(tmp, &fld->lcf_targets, ft_chain) {
 228                 if (tmp->ft_idx == tar->ft_idx) {
 229                         spin_unlock(&fld->lcf_lock);
 230                         OBD_FREE_PTR(target);
 231                         CERROR("Target %s exists in FLD and known as %s:#"LPU64"\n",
 232                                name, fld_target_name(tmp), tmp->ft_idx);
 233                         RETURN(-EEXIST);
 234                 }
 235         }
 236
 237         target->ft_exp = tar->ft_exp;
 238         if (target->ft_exp != NULL)
 239                 class_export_get(target->ft_exp);
 240         target->ft_srv = tar->ft_srv;
 241         target->ft_idx = tar->ft_idx;
 242
 243         cfs_list_add_tail(&target->ft_chain,
 244                           &fld->lcf_targets);
 245
 246         fld->lcf_count++;
 247         spin_unlock(&fld->lcf_lock);
 248
 249         RETURN(0);
 250 }
 251 EXPORT_SYMBOL(fld_client_add_target);
 252
 253 /* Remove export from FLD */
 254 int fld_client_del_target(struct lu_client_fld *fld, __u64 idx)
 255 {
 256         struct lu_fld_target *target, *tmp;
 257         ENTRY;
 258
 259         spin_lock(&fld->lcf_lock);
 260         cfs_list_for_each_entry_safe(target, tmp,
 261                                      &fld->lcf_targets, ft_chain) {
 262                 if (target->ft_idx == idx) {
 263                         fld->lcf_count--;
 264                         cfs_list_del(&target->ft_chain);
 265                         spin_unlock(&fld->lcf_lock);
 266
 267                         if (target->ft_exp != NULL)
 268                                 class_export_put(target->ft_exp);
 269
 270                         OBD_FREE_PTR(target);
 271                         RETURN(0);
 272                 }
 273         }
 274         spin_unlock(&fld->lcf_lock);
 275         RETURN(-ENOENT);
 276 }
 277 EXPORT_SYMBOL(fld_client_del_target);
 278
 279 #ifdef LPROCFS
 280 static int fld_client_proc_init(struct lu_client_fld *fld)
 281 {
 282         int rc;
 283         ENTRY;
 284
 285         fld->lcf_proc_dir = lprocfs_register(fld->lcf_name,
 286                                              fld_type_proc_dir,
 287                                              NULL, NULL);
 288
 289         if (IS_ERR(fld->lcf_proc_dir)) {
 290                 CERROR("%s: LProcFS failed in fld-init\n",
 291                        fld->lcf_name);
 292                 rc = PTR_ERR(fld->lcf_proc_dir);
 293                 RETURN(rc);
 294         }
 295
 296         rc = lprocfs_add_vars(fld->lcf_proc_dir,
 297                               fld_client_proc_list, fld);
 298         if (rc) {
 299                 CERROR("%s: Can't init FLD proc, rc %d\n",
 300                        fld->lcf_name, rc);
 301                 GOTO(out_cleanup, rc);
 302         }
 303
 304         RETURN(0);
 305
 306 out_cleanup:
 307         fld_client_proc_fini(fld);
 308         return rc;
 309 }
 310
 311 void fld_client_proc_fini(struct lu_client_fld *fld)
 312 {
 313         ENTRY;
 314         if (fld->lcf_proc_dir) {
 315                 if (!IS_ERR(fld->lcf_proc_dir))
 316                         lprocfs_remove(&fld->lcf_proc_dir);
 317                 fld->lcf_proc_dir = NULL;
 318         }
 319         EXIT;
 320 }
 321 #else
 322 static int fld_client_proc_init(struct lu_client_fld *fld)
 323 {
 324         return 0;
 325 }
 326
 327 void fld_client_proc_fini(struct lu_client_fld *fld)
 328 {
 329         return;
 330 }
 331 #endif
 332
 333 EXPORT_SYMBOL(fld_client_proc_fini);
 334
 335 static inline int hash_is_sane(int hash)
 336 {
 337         return (hash >= 0 && hash < ARRAY_SIZE(fld_hash));
 338 }
 339
 340 int fld_client_init(struct lu_client_fld *fld,
 341                     const char *prefix, int hash)
 342 {
 343         int cache_size, cache_threshold;
 344         int rc;
 345         ENTRY;
 346
 347         LASSERT(fld != NULL);
 348
 349         snprintf(fld->lcf_name, sizeof(fld->lcf_name),
 350                  "cli-%s", prefix);
 351
 352         if (!hash_is_sane(hash)) {
 353                 CERROR("%s: Wrong hash function %#x\n",
 354                        fld->lcf_name, hash);
 355                 RETURN(-EINVAL);
 356         }
 357
 358         fld->lcf_count = 0;
 359         spin_lock_init(&fld->lcf_lock);
 360         fld->lcf_hash = &fld_hash[hash];
 361         fld->lcf_flags = LUSTRE_FLD_INIT;
 362         CFS_INIT_LIST_HEAD(&fld->lcf_targets);
 363
 364         cache_size = FLD_CLIENT_CACHE_SIZE /
 365                 sizeof(struct fld_cache_entry);
 366
 367         cache_threshold = cache_size *
 368                 FLD_CLIENT_CACHE_THRESHOLD / 100;
 369
 370         fld->lcf_cache = fld_cache_init(fld->lcf_name,
 371                                         cache_size, cache_threshold);
 372         if (IS_ERR(fld->lcf_cache)) {
 373                 rc = PTR_ERR(fld->lcf_cache);
 374                 fld->lcf_cache = NULL;
 375                 GOTO(out, rc);
 376         }
 377
 378         rc = fld_client_proc_init(fld);
 379         if (rc)
 380                 GOTO(out, rc);
 381         EXIT;
 382 out:
 383         if (rc)
 384                 fld_client_fini(fld);
 385         else
 386                 CDEBUG(D_INFO, "%s: Using \"%s\" hash\n",
 387                        fld->lcf_name, fld->lcf_hash->fh_name);
 388         return rc;
 389 }
 390 EXPORT_SYMBOL(fld_client_init);
 391
 392 void fld_client_fini(struct lu_client_fld *fld)
 393 {
 394         struct lu_fld_target *target, *tmp;
 395         ENTRY;
 396
 397         spin_lock(&fld->lcf_lock);
 398         cfs_list_for_each_entry_safe(target, tmp,
 399                                      &fld->lcf_targets, ft_chain) {
 400                 fld->lcf_count--;
 401                 cfs_list_del(&target->ft_chain);
 402                 if (target->ft_exp != NULL)
 403                         class_export_put(target->ft_exp);
 404                 OBD_FREE_PTR(target);
 405         }
 406         spin_unlock(&fld->lcf_lock);
 407
 408         if (fld->lcf_cache != NULL) {
 409                 if (!IS_ERR(fld->lcf_cache))
 410                         fld_cache_fini(fld->lcf_cache);
 411                 fld->lcf_cache = NULL;
 412         }
 413
 414         EXIT;
 415 }
 416 EXPORT_SYMBOL(fld_client_fini);
 417
 418 int fld_client_rpc(struct obd_export *exp,
 419                    struct lu_seq_range *range, __u32 fld_op)
 420 {
 421         struct ptlrpc_request *req;
 422         struct lu_seq_range   *prange;
 423         __u32                 *op;
 424         int                    rc;
 425         struct obd_import     *imp;
 426         ENTRY;
 427
 428         LASSERT(exp != NULL);
 429
 430         imp = class_exp2cliimp(exp);
 431         req = ptlrpc_request_alloc_pack(imp, &RQF_FLD_QUERY, LUSTRE_MDS_VERSION,
 432                                         FLD_QUERY);
 433         if (req == NULL)
 434                 RETURN(-ENOMEM);
 435
 436         op = req_capsule_client_get(&req->rq_pill, &RMF_FLD_OPC);
 437         *op = fld_op;
 438
 439         prange = req_capsule_client_get(&req->rq_pill, &RMF_FLD_MDFLD);
 440         *prange = *range;
 441
 442         ptlrpc_request_set_replen(req);
 443         req->rq_request_portal = FLD_REQUEST_PORTAL;
 444         ptlrpc_at_set_req_timeout(req);
 445
 446         if (fld_op == FLD_LOOKUP &&
 447             imp->imp_connect_flags_orig & OBD_CONNECT_MDS_MDS)
 448                 req->rq_allow_replay = 1;
 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(env, target->ft_srv, seq, &res);
 500         } else {
 501 #endif
 502                 rc = fld_client_rpc(target->ft_exp,
 503                                     &res, FLD_LOOKUP);
 504 #ifdef __KERNEL__
 505         }
 506 #endif
 507
 508         if (rc == 0) {
 509                 *mds = res.lsr_index;
 510
 511                 fld_cache_insert(fld->lcf_cache, &res);
 512         }
 513         RETURN(rc);
 514 }
 515 EXPORT_SYMBOL(fld_client_lookup);
 516
 517 void fld_client_flush(struct lu_client_fld *fld)
 518 {
 519         fld_cache_flush(fld->lcf_cache);
 520 }
 521 EXPORT_SYMBOL(fld_client_flush);