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  2008 Sun Microsystems, Inc. All rights reserved
  30  * Use is subject to license terms.
  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 #ifndef EXPORT_SYMTAB
  44 # define EXPORT_SYMTAB
  45 #endif
  46 #define DEBUG_SUBSYSTEM S_FLD
  47
  48 #ifdef __KERNEL__
  49 # include <libcfs/libcfs.h>
  50 # include <linux/module.h>
  51 # include <linux/jbd.h>
  52 # include <asm/div64.h>
  53 #else /* __KERNEL__ */
  54 # include <liblustre.h>
  55 # include <libcfs/list.h>
  56 #endif
  57
  58 #include <obd.h>
  59 #include <obd_class.h>
  60 #include <lustre_ver.h>
  61 #include <obd_support.h>
  62 #include <lprocfs_status.h>
  63
  64 #include <dt_object.h>
  65 #include <md_object.h>
  66 #include <lustre_req_layout.h>
  67 #include <lustre_fld.h>
  68 #include <lustre_mdc.h>
  69 #include "fld_internal.h"
  70
  71 /* TODO: these 3 functions are copies of flow-control code from mdc_lib.c
  72  * It should be common thing. The same about mdc RPC lock */
  73 static int fld_req_avail(struct client_obd *cli, struct mdc_cache_waiter *mcw)
  74 {
  75         int rc;
  76         ENTRY;
  77         spin_lock(&cli->cl_loi_list_lock);
  78         rc = list_empty(&mcw->mcw_entry);
  79         spin_unlock(&cli->cl_loi_list_lock);
  80         RETURN(rc);
  81 };
  82
  83 static void fld_enter_request(struct client_obd *cli)
  84 {
  85         struct mdc_cache_waiter mcw;
  86         struct l_wait_info lwi = { 0 };
  87
  88         spin_lock(&cli->cl_loi_list_lock);
  89         if (cli->cl_r_in_flight >= cli->cl_max_rpcs_in_flight) {
  90                 list_add_tail(&mcw.mcw_entry, &cli->cl_cache_waiters);
  91                 cfs_waitq_init(&mcw.mcw_waitq);
  92                 spin_unlock(&cli->cl_loi_list_lock);
  93                 l_wait_event(mcw.mcw_waitq, fld_req_avail(cli, &mcw), &lwi);
  94         } else {
  95                 cli->cl_r_in_flight++;
  96                 spin_unlock(&cli->cl_loi_list_lock);
  97         }
  98 }
  99
 100 static void fld_exit_request(struct client_obd *cli)
 101 {
 102         struct list_head *l, *tmp;
 103         struct mdc_cache_waiter *mcw;
 104
 105         spin_lock(&cli->cl_loi_list_lock);
 106         cli->cl_r_in_flight--;
 107         list_for_each_safe(l, tmp, &cli->cl_cache_waiters) {
 108
 109                 if (cli->cl_r_in_flight >= cli->cl_max_rpcs_in_flight) {
 110                         /* No free request slots anymore */
 111                         break;
 112                 }
 113
 114                 mcw = list_entry(l, struct mdc_cache_waiter, mcw_entry);
 115                 list_del_init(&mcw->mcw_entry);
 116                 cli->cl_r_in_flight++;
 117                 cfs_waitq_signal(&mcw->mcw_waitq);
 118         }
 119         spin_unlock(&cli->cl_loi_list_lock);
 120 }
 121
 122 static int fld_rrb_hash(struct lu_client_fld *fld,
 123                         seqno_t seq)
 124 {
 125         LASSERT(fld->lcf_count > 0);
 126         return do_div(seq, fld->lcf_count);
 127 }
 128
 129 static struct lu_fld_target *
 130 fld_rrb_scan(struct lu_client_fld *fld, seqno_t seq)
 131 {
 132         struct lu_fld_target *target;
 133         int hash;
 134         ENTRY;
 135
 136         hash = fld_rrb_hash(fld, seq);
 137
 138         list_for_each_entry(target, &fld->lcf_targets, ft_chain) {
 139                 if (target->ft_idx == hash)
 140                         RETURN(target);
 141         }
 142
 143         CERROR("%s: Can't find target by hash %d (seq "LPX64"). "
 144                "Targets (%d):\n", fld->lcf_name, hash, seq,
 145                fld->lcf_count);
 146
 147         list_for_each_entry(target, &fld->lcf_targets, ft_chain) {
 148                 const char *srv_name = target->ft_srv != NULL  ?
 149                         target->ft_srv->lsf_name : "<null>";
 150                 const char *exp_name = target->ft_exp != NULL ?
 151                         (char *)target->ft_exp->exp_obd->obd_uuid.uuid :
 152                         "<null>";
 153
 154                 CERROR("  exp: 0x%p (%s), srv: 0x%p (%s), idx: "LPU64"\n",
 155                        target->ft_exp, exp_name, target->ft_srv,
 156                        srv_name, target->ft_idx);
 157         }
 158
 159         /*
 160          * If target is not found, there is logical error anyway, so here is
 161          * LBUG() to catch this situation.
 162          */
 163         LBUG();
 164         RETURN(NULL);
 165 }
 166
 167 struct lu_fld_hash fld_hash[] = {
 168         {
 169                 .fh_name = "RRB",
 170                 .fh_hash_func = fld_rrb_hash,
 171                 .fh_scan_func = fld_rrb_scan
 172         },
 173         {
 174                 0,
 175         }
 176 };
 177
 178 static struct lu_fld_target *
 179 fld_client_get_target(struct lu_client_fld *fld,
 180                       seqno_t seq)
 181 {
 182         struct lu_fld_target *target;
 183         ENTRY;
 184
 185         LASSERT(fld->lcf_hash != NULL);
 186
 187         spin_lock(&fld->lcf_lock);
 188         target = fld->lcf_hash->fh_scan_func(fld, seq);
 189         spin_unlock(&fld->lcf_lock);
 190
 191         if (target != NULL) {
 192                 CDEBUG(D_INFO, "%s: Found target (idx "LPU64
 193                        ") by seq "LPX64"\n", fld->lcf_name,
 194                        target->ft_idx, seq);
 195         }
 196
 197         RETURN(target);
 198 }
 199
 200 /*
 201  * Add export to FLD. This is usually done by CMM and LMV as they are main users
 202  * of FLD module.
 203  */
 204 int fld_client_add_target(struct lu_client_fld *fld,
 205                           struct lu_fld_target *tar)
 206 {
 207         const char *name = fld_target_name(tar);
 208         struct lu_fld_target *target, *tmp;
 209         ENTRY;
 210
 211         LASSERT(tar != NULL);
 212         LASSERT(name != NULL);
 213         LASSERT(tar->ft_srv != NULL || tar->ft_exp != NULL);
 214
 215         if (fld->lcf_flags != LUSTRE_FLD_INIT) {
 216                 CERROR("%s: Attempt to add target %s (idx "LPU64") "
 217                        "on fly - skip it\n", fld->lcf_name, name,
 218                        tar->ft_idx);
 219                 RETURN(0);
 220         } else {
 221                 CDEBUG(D_INFO, "%s: Adding target %s (idx "
 222                        LPU64")\n", fld->lcf_name, name, tar->ft_idx);
 223         }
 224
 225         OBD_ALLOC_PTR(target);
 226         if (target == NULL)
 227                 RETURN(-ENOMEM);
 228
 229         spin_lock(&fld->lcf_lock);
 230         list_for_each_entry(tmp, &fld->lcf_targets, ft_chain) {
 231                 if (tmp->ft_idx == tar->ft_idx) {
 232                         spin_unlock(&fld->lcf_lock);
 233                         OBD_FREE_PTR(target);
 234                         CERROR("Target %s exists in FLD and known as %s:#"LPU64"\n",
 235                                name, fld_target_name(tmp), tmp->ft_idx);
 236                         RETURN(-EEXIST);
 237                 }
 238         }
 239
 240         target->ft_exp = tar->ft_exp;
 241         if (target->ft_exp != NULL)
 242                 class_export_get(target->ft_exp);
 243         target->ft_srv = tar->ft_srv;
 244         target->ft_idx = tar->ft_idx;
 245
 246         list_add_tail(&target->ft_chain,
 247                       &fld->lcf_targets);
 248
 249         fld->lcf_count++;
 250         spin_unlock(&fld->lcf_lock);
 251
 252         RETURN(0);
 253 }
 254 EXPORT_SYMBOL(fld_client_add_target);
 255
 256 /* Remove export from FLD */
 257 int fld_client_del_target(struct lu_client_fld *fld,
 258                           __u64 idx)
 259 {
 260         struct lu_fld_target *target, *tmp;
 261         ENTRY;
 262
 263         spin_lock(&fld->lcf_lock);
 264         list_for_each_entry_safe(target, tmp,
 265                                  &fld->lcf_targets, ft_chain) {
 266                 if (target->ft_idx == idx) {
 267                         fld->lcf_count--;
 268                         list_del(&target->ft_chain);
 269                         spin_unlock(&fld->lcf_lock);
 270
 271                         if (target->ft_exp != NULL)
 272                                 class_export_put(target->ft_exp);
 273
 274                         OBD_FREE_PTR(target);
 275                         RETURN(0);
 276                 }
 277         }
 278         spin_unlock(&fld->lcf_lock);
 279         RETURN(-ENOENT);
 280 }
 281 EXPORT_SYMBOL(fld_client_del_target);
 282
 283 static void fld_client_proc_fini(struct lu_client_fld *fld);
 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 static 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 static void fld_client_proc_fini(struct lu_client_fld *fld)
 334 {
 335         return;
 336 }
 337 #endif
 338
 339 static inline int hash_is_sane(int hash)
 340 {
 341         return (hash >= 0 && hash < ARRAY_SIZE(fld_hash));
 342 }
 343
 344 int fld_client_init(struct lu_client_fld *fld,
 345                     const char *prefix, int hash)
 346 {
 347         int cache_size, cache_threshold;
 348         int rc;
 349         ENTRY;
 350
 351         LASSERT(fld != NULL);
 352
 353         snprintf(fld->lcf_name, sizeof(fld->lcf_name),
 354                  "cli-%s", prefix);
 355
 356         if (!hash_is_sane(hash)) {
 357                 CERROR("%s: Wrong hash function %#x\n",
 358                        fld->lcf_name, hash);
 359                 RETURN(-EINVAL);
 360         }
 361
 362         fld->lcf_count = 0;
 363         spin_lock_init(&fld->lcf_lock);
 364         fld->lcf_hash = &fld_hash[hash];
 365         fld->lcf_flags = LUSTRE_FLD_INIT;
 366         CFS_INIT_LIST_HEAD(&fld->lcf_targets);
 367
 368         cache_size = FLD_CLIENT_CACHE_SIZE /
 369                 sizeof(struct fld_cache_entry);
 370
 371         cache_threshold = cache_size *
 372                 FLD_CLIENT_CACHE_THRESHOLD / 100;
 373
 374         fld->lcf_cache = fld_cache_init(fld->lcf_name,
 375                                         cache_size, cache_threshold);
 376         if (IS_ERR(fld->lcf_cache)) {
 377                 rc = PTR_ERR(fld->lcf_cache);
 378                 fld->lcf_cache = NULL;
 379                 GOTO(out, rc);
 380         }
 381
 382         rc = fld_client_proc_init(fld);
 383         if (rc)
 384                 GOTO(out, rc);
 385         EXIT;
 386 out:
 387         if (rc)
 388                 fld_client_fini(fld);
 389         else
 390                 CDEBUG(D_INFO, "%s: Using \"%s\" hash\n",
 391                        fld->lcf_name, fld->lcf_hash->fh_name);
 392         return rc;
 393 }
 394 EXPORT_SYMBOL(fld_client_init);
 395
 396 void fld_client_fini(struct lu_client_fld *fld)
 397 {
 398         struct lu_fld_target *target, *tmp;
 399         ENTRY;
 400
 401         fld_client_proc_fini(fld);
 402
 403         spin_lock(&fld->lcf_lock);
 404         list_for_each_entry_safe(target, tmp,
 405                                  &fld->lcf_targets, ft_chain) {
 406                 fld->lcf_count--;
 407                 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         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,
 471                       seqno_t seq, mdsno_t *mds,
 472                       const struct lu_env *env)
 473 {
 474         struct lu_seq_range res;
 475         struct lu_fld_target *target;
 476         int rc;
 477         ENTRY;
 478
 479         fld->lcf_flags |= LUSTRE_FLD_RUN;
 480
 481         rc = fld_cache_lookup(fld->lcf_cache, seq, &res);
 482         if (rc == 0) {
 483                 *mds = res.lsr_mdt;
 484                 RETURN(0);
 485         }
 486
 487         /* Can not find it in the cache */
 488         target = fld_client_get_target(fld, seq);
 489         LASSERT(target != NULL);
 490
 491         CDEBUG(D_INFO, "%s: Lookup fld entry (seq: "LPX64") on "
 492                "target %s (idx "LPU64")\n", fld->lcf_name, seq,
 493                fld_target_name(target), target->ft_idx);
 494
 495         res.lsr_start = seq;
 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_mdt;
 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);