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         /* Because almost all of special sequence located in MDT0,
 134          * it should go to index 0 directly, instead of calculating
 135          * hash again, and also if other MDTs is not being connected,
 136          * the fld lookup requests(for seq on MDT0) should not be
 137          * blocked because of other MDTs */
 138         if (fid_seq_is_norm(seq))
 139                 hash = fld_rrb_hash(fld, seq);
 140         else
 141                 hash = 0;
 142
 143         cfs_list_for_each_entry(target, &fld->lcf_targets, ft_chain) {
 144                 if (target->ft_idx == hash)
 145                         RETURN(target);
 146         }
 147
 148         CERROR("%s: Can't find target by hash %d (seq "LPX64"). "
 149                "Targets (%d):\n", fld->lcf_name, hash, seq,
 150                fld->lcf_count);
 151
 152         cfs_list_for_each_entry(target, &fld->lcf_targets, ft_chain) {
 153                 const char *srv_name = target->ft_srv != NULL  ?
 154                         target->ft_srv->lsf_name : "<null>";
 155                 const char *exp_name = target->ft_exp != NULL ?
 156                         (char *)target->ft_exp->exp_obd->obd_uuid.uuid :
 157                         "<null>";
 158
 159                 CERROR("  exp: 0x%p (%s), srv: 0x%p (%s), idx: "LPU64"\n",
 160                        target->ft_exp, exp_name, target->ft_srv,
 161                        srv_name, target->ft_idx);
 162         }
 163
 164         /*
 165          * If target is not found, there is logical error anyway, so here is
 166          * LBUG() to catch this situation.
 167          */
 168         LBUG();
 169         RETURN(NULL);
 170 }
 171
 172 struct lu_fld_hash fld_hash[] = {
 173         {
 174                 .fh_name = "RRB",
 175                 .fh_hash_func = fld_rrb_hash,
 176                 .fh_scan_func = fld_rrb_scan
 177         },
 178         {
 179                 0,
 180         }
 181 };
 182
 183 static struct lu_fld_target *
 184 fld_client_get_target(struct lu_client_fld *fld, seqno_t seq)
 185 {
 186         struct lu_fld_target *target;
 187         ENTRY;
 188
 189         LASSERT(fld->lcf_hash != NULL);
 190
 191         spin_lock(&fld->lcf_lock);
 192         target = fld->lcf_hash->fh_scan_func(fld, seq);
 193         spin_unlock(&fld->lcf_lock);
 194
 195         if (target != NULL) {
 196                 CDEBUG(D_INFO, "%s: Found target (idx "LPU64
 197                        ") by seq "LPX64"\n", fld->lcf_name,
 198                        target->ft_idx, seq);
 199         }
 200
 201         RETURN(target);
 202 }
 203
 204 /*
 205  * Add export to FLD. This is usually done by CMM and LMV as they are main users
 206  * of FLD module.
 207  */
 208 int fld_client_add_target(struct lu_client_fld *fld,
 209                           struct lu_fld_target *tar)
 210 {
 211         const char *name;
 212         struct lu_fld_target *target, *tmp;
 213         ENTRY;
 214
 215         LASSERT(tar != NULL);
 216         name = fld_target_name(tar);
 217         LASSERT(name != NULL);
 218         LASSERT(tar->ft_srv != NULL || tar->ft_exp != NULL);
 219
 220         if (fld->lcf_flags != LUSTRE_FLD_INIT) {
 221                 CERROR("%s: Attempt to add target %s (idx "LPU64") "
 222                        "on fly - skip it\n", fld->lcf_name, name,
 223                        tar->ft_idx);
 224                 RETURN(0);
 225         } else {
 226                 CDEBUG(D_INFO, "%s: Adding target %s (idx "
 227                        LPU64")\n", fld->lcf_name, name, tar->ft_idx);
 228         }
 229
 230         OBD_ALLOC_PTR(target);
 231         if (target == NULL)
 232                 RETURN(-ENOMEM);
 233
 234         spin_lock(&fld->lcf_lock);
 235         cfs_list_for_each_entry(tmp, &fld->lcf_targets, ft_chain) {
 236                 if (tmp->ft_idx == tar->ft_idx) {
 237                         spin_unlock(&fld->lcf_lock);
 238                         OBD_FREE_PTR(target);
 239                         CERROR("Target %s exists in FLD and known as %s:#"LPU64"\n",
 240                                name, fld_target_name(tmp), tmp->ft_idx);
 241                         RETURN(-EEXIST);
 242                 }
 243         }
 244
 245         target->ft_exp = tar->ft_exp;
 246         if (target->ft_exp != NULL)
 247                 class_export_get(target->ft_exp);
 248         target->ft_srv = tar->ft_srv;
 249         target->ft_idx = tar->ft_idx;
 250
 251         cfs_list_add_tail(&target->ft_chain,
 252                           &fld->lcf_targets);
 253
 254         fld->lcf_count++;
 255         spin_unlock(&fld->lcf_lock);
 256
 257         RETURN(0);
 258 }
 259 EXPORT_SYMBOL(fld_client_add_target);
 260
 261 /* Remove export from FLD */
 262 int fld_client_del_target(struct lu_client_fld *fld, __u64 idx)
 263 {
 264         struct lu_fld_target *target, *tmp;
 265         ENTRY;
 266
 267         spin_lock(&fld->lcf_lock);
 268         cfs_list_for_each_entry_safe(target, tmp,
 269                                      &fld->lcf_targets, ft_chain) {
 270                 if (target->ft_idx == idx) {
 271                         fld->lcf_count--;
 272                         cfs_list_del(&target->ft_chain);
 273                         spin_unlock(&fld->lcf_lock);
 274
 275                         if (target->ft_exp != NULL)
 276                                 class_export_put(target->ft_exp);
 277
 278                         OBD_FREE_PTR(target);
 279                         RETURN(0);
 280                 }
 281         }
 282         spin_unlock(&fld->lcf_lock);
 283         RETURN(-ENOENT);
 284 }
 285 EXPORT_SYMBOL(fld_client_del_target);
 286
 287 #ifdef LPROCFS
 288 static int fld_client_proc_init(struct lu_client_fld *fld)
 289 {
 290         int rc;
 291         ENTRY;
 292
 293         fld->lcf_proc_dir = lprocfs_register(fld->lcf_name,
 294                                              fld_type_proc_dir,
 295                                              NULL, NULL);
 296
 297         if (IS_ERR(fld->lcf_proc_dir)) {
 298                 CERROR("%s: LProcFS failed in fld-init\n",
 299                        fld->lcf_name);
 300                 rc = PTR_ERR(fld->lcf_proc_dir);
 301                 RETURN(rc);
 302         }
 303
 304         rc = lprocfs_add_vars(fld->lcf_proc_dir,
 305                               fld_client_proc_list, fld);
 306         if (rc) {
 307                 CERROR("%s: Can't init FLD proc, rc %d\n",
 308                        fld->lcf_name, rc);
 309                 GOTO(out_cleanup, rc);
 310         }
 311
 312         RETURN(0);
 313
 314 out_cleanup:
 315         fld_client_proc_fini(fld);
 316         return rc;
 317 }
 318
 319 void fld_client_proc_fini(struct lu_client_fld *fld)
 320 {
 321         ENTRY;
 322         if (fld->lcf_proc_dir) {
 323                 if (!IS_ERR(fld->lcf_proc_dir))
 324                         lprocfs_remove(&fld->lcf_proc_dir);
 325                 fld->lcf_proc_dir = NULL;
 326         }
 327         EXIT;
 328 }
 329 #else
 330 static int fld_client_proc_init(struct lu_client_fld *fld)
 331 {
 332         return 0;
 333 }
 334
 335 void fld_client_proc_fini(struct lu_client_fld *fld)
 336 {
 337         return;
 338 }
 339 #endif
 340
 341 EXPORT_SYMBOL(fld_client_proc_fini);
 342
 343 static inline int hash_is_sane(int hash)
 344 {
 345         return (hash >= 0 && hash < ARRAY_SIZE(fld_hash));
 346 }
 347
 348 int fld_client_init(struct lu_client_fld *fld,
 349                     const char *prefix, int hash)
 350 {
 351         int cache_size, cache_threshold;
 352         int rc;
 353         ENTRY;
 354
 355         LASSERT(fld != NULL);
 356
 357         snprintf(fld->lcf_name, sizeof(fld->lcf_name),
 358                  "cli-%s", prefix);
 359
 360         if (!hash_is_sane(hash)) {
 361                 CERROR("%s: Wrong hash function %#x\n",
 362                        fld->lcf_name, hash);
 363                 RETURN(-EINVAL);
 364         }
 365
 366         fld->lcf_count = 0;
 367         spin_lock_init(&fld->lcf_lock);
 368         fld->lcf_hash = &fld_hash[hash];
 369         fld->lcf_flags = LUSTRE_FLD_INIT;
 370         CFS_INIT_LIST_HEAD(&fld->lcf_targets);
 371
 372         cache_size = FLD_CLIENT_CACHE_SIZE /
 373                 sizeof(struct fld_cache_entry);
 374
 375         cache_threshold = cache_size *
 376                 FLD_CLIENT_CACHE_THRESHOLD / 100;
 377
 378         fld->lcf_cache = fld_cache_init(fld->lcf_name,
 379                                         cache_size, cache_threshold);
 380         if (IS_ERR(fld->lcf_cache)) {
 381                 rc = PTR_ERR(fld->lcf_cache);
 382                 fld->lcf_cache = NULL;
 383                 GOTO(out, rc);
 384         }
 385
 386         rc = fld_client_proc_init(fld);
 387         if (rc)
 388                 GOTO(out, rc);
 389         EXIT;
 390 out:
 391         if (rc)
 392                 fld_client_fini(fld);
 393         else
 394                 CDEBUG(D_INFO, "%s: Using \"%s\" hash\n",
 395                        fld->lcf_name, fld->lcf_hash->fh_name);
 396         return rc;
 397 }
 398 EXPORT_SYMBOL(fld_client_init);
 399
 400 void fld_client_fini(struct lu_client_fld *fld)
 401 {
 402         struct lu_fld_target *target, *tmp;
 403         ENTRY;
 404
 405         spin_lock(&fld->lcf_lock);
 406         cfs_list_for_each_entry_safe(target, tmp,
 407                                      &fld->lcf_targets, ft_chain) {
 408                 fld->lcf_count--;
 409                 cfs_list_del(&target->ft_chain);
 410                 if (target->ft_exp != NULL)
 411                         class_export_put(target->ft_exp);
 412                 OBD_FREE_PTR(target);
 413         }
 414         spin_unlock(&fld->lcf_lock);
 415
 416         if (fld->lcf_cache != NULL) {
 417                 if (!IS_ERR(fld->lcf_cache))
 418                         fld_cache_fini(fld->lcf_cache);
 419                 fld->lcf_cache = NULL;
 420         }
 421
 422         EXIT;
 423 }
 424 EXPORT_SYMBOL(fld_client_fini);
 425
 426 int fld_client_rpc(struct obd_export *exp,
 427                    struct lu_seq_range *range, __u32 fld_op)
 428 {
 429         struct ptlrpc_request *req;
 430         struct lu_seq_range   *prange;
 431         __u32                 *op;
 432         int                    rc;
 433         struct obd_import     *imp;
 434         ENTRY;
 435
 436         LASSERT(exp != NULL);
 437
 438         imp = class_exp2cliimp(exp);
 439         req = ptlrpc_request_alloc_pack(imp, &RQF_FLD_QUERY, LUSTRE_MDS_VERSION,
 440                                         FLD_QUERY);
 441         if (req == NULL)
 442                 RETURN(-ENOMEM);
 443
 444         op = req_capsule_client_get(&req->rq_pill, &RMF_FLD_OPC);
 445         *op = fld_op;
 446
 447         prange = req_capsule_client_get(&req->rq_pill, &RMF_FLD_MDFLD);
 448         *prange = *range;
 449
 450         ptlrpc_request_set_replen(req);
 451         req->rq_request_portal = FLD_REQUEST_PORTAL;
 452         ptlrpc_at_set_req_timeout(req);
 453
 454         if (fld_op == FLD_LOOKUP &&
 455             imp->imp_connect_flags_orig & OBD_CONNECT_MDS_MDS)
 456                 req->rq_allow_replay = 1;
 457
 458         if (fld_op != FLD_LOOKUP)
 459                 mdc_get_rpc_lock(exp->exp_obd->u.cli.cl_rpc_lock, NULL);
 460         fld_enter_request(&exp->exp_obd->u.cli);
 461         rc = ptlrpc_queue_wait(req);
 462         fld_exit_request(&exp->exp_obd->u.cli);
 463         if (fld_op != FLD_LOOKUP)
 464                 mdc_put_rpc_lock(exp->exp_obd->u.cli.cl_rpc_lock, NULL);
 465         if (rc)
 466                 GOTO(out_req, rc);
 467
 468         prange = req_capsule_server_get(&req->rq_pill, &RMF_FLD_MDFLD);
 469         if (prange == NULL)
 470                 GOTO(out_req, rc = -EFAULT);
 471         *range = *prange;
 472         EXIT;
 473 out_req:
 474         ptlrpc_req_finished(req);
 475         return rc;
 476 }
 477
 478 int fld_client_lookup(struct lu_client_fld *fld, seqno_t seq, mdsno_t *mds,
 479                       __u32 flags, const struct lu_env *env)
 480 {
 481         struct lu_seq_range res = { 0 };
 482         struct lu_fld_target *target;
 483         int rc;
 484         ENTRY;
 485
 486         fld->lcf_flags |= LUSTRE_FLD_RUN;
 487
 488         rc = fld_cache_lookup(fld->lcf_cache, seq, &res);
 489         if (rc == 0) {
 490                 *mds = res.lsr_index;
 491                 RETURN(0);
 492         }
 493
 494         /* Can not find it in the cache */
 495         target = fld_client_get_target(fld, seq);
 496         LASSERT(target != NULL);
 497
 498         CDEBUG(D_INFO, "%s: Lookup fld entry (seq: "LPX64") on "
 499                "target %s (idx "LPU64")\n", fld->lcf_name, seq,
 500                fld_target_name(target), target->ft_idx);
 501
 502         res.lsr_start = seq;
 503         fld_range_set_type(&res, flags);
 504 #ifdef __KERNEL__
 505         if (target->ft_srv != NULL) {
 506                 LASSERT(env != NULL);
 507                 rc = fld_server_lookup(env, target->ft_srv, seq, &res);
 508         } else {
 509 #endif
 510                 rc = fld_client_rpc(target->ft_exp, &res, FLD_LOOKUP);
 511 #ifdef __KERNEL__
 512         }
 513 #endif
 514
 515         if (rc == 0) {
 516                 *mds = res.lsr_index;
 517
 518                 fld_cache_insert(fld->lcf_cache, &res);
 519         }
 520         RETURN(rc);
 521 }
 522 EXPORT_SYMBOL(fld_client_lookup);
 523
 524 void fld_client_flush(struct lu_client_fld *fld)
 525 {
 526         fld_cache_flush(fld->lcf_cache);
 527 }
 528 EXPORT_SYMBOL(fld_client_flush);