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, Whamcloud, Inc.
  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 = fld_target_name(tar);
 204         struct lu_fld_target *target, *tmp;
 205         ENTRY;
 206
 207         LASSERT(tar != NULL);
 208         LASSERT(name != NULL);
 209         LASSERT(tar->ft_srv != NULL || tar->ft_exp != NULL);
 210
 211         if (fld->lcf_flags != LUSTRE_FLD_INIT) {
 212                 CERROR("%s: Attempt to add target %s (idx "LPU64") "
 213                        "on fly - skip it\n", fld->lcf_name, name,
 214                        tar->ft_idx);
 215                 RETURN(0);
 216         } else {
 217                 CDEBUG(D_INFO, "%s: Adding target %s (idx "
 218                        LPU64")\n", fld->lcf_name, name, tar->ft_idx);
 219         }
 220
 221         OBD_ALLOC_PTR(target);
 222         if (target == NULL)
 223                 RETURN(-ENOMEM);
 224
 225         spin_lock(&fld->lcf_lock);
 226         cfs_list_for_each_entry(tmp, &fld->lcf_targets, ft_chain) {
 227                 if (tmp->ft_idx == tar->ft_idx) {
 228                         spin_unlock(&fld->lcf_lock);
 229                         OBD_FREE_PTR(target);
 230                         CERROR("Target %s exists in FLD and known as %s:#"LPU64"\n",
 231                                name, fld_target_name(tmp), tmp->ft_idx);
 232                         RETURN(-EEXIST);
 233                 }
 234         }
 235
 236         target->ft_exp = tar->ft_exp;
 237         if (target->ft_exp != NULL)
 238                 class_export_get(target->ft_exp);
 239         target->ft_srv = tar->ft_srv;
 240         target->ft_idx = tar->ft_idx;
 241
 242         cfs_list_add_tail(&target->ft_chain,
 243                           &fld->lcf_targets);
 244
 245         fld->lcf_count++;
 246         spin_unlock(&fld->lcf_lock);
 247
 248         RETURN(0);
 249 }
 250 EXPORT_SYMBOL(fld_client_add_target);
 251
 252 /* Remove export from FLD */
 253 int fld_client_del_target(struct lu_client_fld *fld, __u64 idx)
 254 {
 255         struct lu_fld_target *target, *tmp;
 256         ENTRY;
 257
 258         spin_lock(&fld->lcf_lock);
 259         cfs_list_for_each_entry_safe(target, tmp,
 260                                      &fld->lcf_targets, ft_chain) {
 261                 if (target->ft_idx == idx) {
 262                         fld->lcf_count--;
 263                         cfs_list_del(&target->ft_chain);
 264                         spin_unlock(&fld->lcf_lock);
 265
 266                         if (target->ft_exp != NULL)
 267                                 class_export_put(target->ft_exp);
 268
 269                         OBD_FREE_PTR(target);
 270                         RETURN(0);
 271                 }
 272         }
 273         spin_unlock(&fld->lcf_lock);
 274         RETURN(-ENOENT);
 275 }
 276 EXPORT_SYMBOL(fld_client_del_target);
 277
 278 #ifdef LPROCFS
 279 static int fld_client_proc_init(struct lu_client_fld *fld)
 280 {
 281         int rc;
 282         ENTRY;
 283
 284         fld->lcf_proc_dir = lprocfs_register(fld->lcf_name,
 285                                              fld_type_proc_dir,
 286                                              NULL, NULL);
 287
 288         if (IS_ERR(fld->lcf_proc_dir)) {
 289                 CERROR("%s: LProcFS failed in fld-init\n",
 290                        fld->lcf_name);
 291                 rc = PTR_ERR(fld->lcf_proc_dir);
 292                 RETURN(rc);
 293         }
 294
 295         rc = lprocfs_add_vars(fld->lcf_proc_dir,
 296                               fld_client_proc_list, fld);
 297         if (rc) {
 298                 CERROR("%s: Can't init FLD proc, rc %d\n",
 299                        fld->lcf_name, rc);
 300                 GOTO(out_cleanup, rc);
 301         }
 302
 303         RETURN(0);
 304
 305 out_cleanup:
 306         fld_client_proc_fini(fld);
 307         return rc;
 308 }
 309
 310 void fld_client_proc_fini(struct lu_client_fld *fld)
 311 {
 312         ENTRY;
 313         if (fld->lcf_proc_dir) {
 314                 if (!IS_ERR(fld->lcf_proc_dir))
 315                         lprocfs_remove(&fld->lcf_proc_dir);
 316                 fld->lcf_proc_dir = NULL;
 317         }
 318         EXIT;
 319 }
 320 #else
 321 static int fld_client_proc_init(struct lu_client_fld *fld)
 322 {
 323         return 0;
 324 }
 325
 326 void fld_client_proc_fini(struct lu_client_fld *fld)
 327 {
 328         return;
 329 }
 330 #endif
 331
 332 EXPORT_SYMBOL(fld_client_proc_fini);
 333
 334 static inline int hash_is_sane(int hash)
 335 {
 336         return (hash >= 0 && hash < ARRAY_SIZE(fld_hash));
 337 }
 338
 339 int fld_client_init(struct lu_client_fld *fld,
 340                     const char *prefix, int hash)
 341 {
 342         int cache_size, cache_threshold;
 343         int rc;
 344         ENTRY;
 345
 346         LASSERT(fld != NULL);
 347
 348         snprintf(fld->lcf_name, sizeof(fld->lcf_name),
 349                  "cli-%s", prefix);
 350
 351         if (!hash_is_sane(hash)) {
 352                 CERROR("%s: Wrong hash function %#x\n",
 353                        fld->lcf_name, hash);
 354                 RETURN(-EINVAL);
 355         }
 356
 357         fld->lcf_count = 0;
 358         spin_lock_init(&fld->lcf_lock);
 359         fld->lcf_hash = &fld_hash[hash];
 360         fld->lcf_flags = LUSTRE_FLD_INIT;
 361         CFS_INIT_LIST_HEAD(&fld->lcf_targets);
 362
 363         cache_size = FLD_CLIENT_CACHE_SIZE /
 364                 sizeof(struct fld_cache_entry);
 365
 366         cache_threshold = cache_size *
 367                 FLD_CLIENT_CACHE_THRESHOLD / 100;
 368
 369         fld->lcf_cache = fld_cache_init(fld->lcf_name,
 370                                         cache_size, cache_threshold);
 371         if (IS_ERR(fld->lcf_cache)) {
 372                 rc = PTR_ERR(fld->lcf_cache);
 373                 fld->lcf_cache = NULL;
 374                 GOTO(out, rc);
 375         }
 376
 377         rc = fld_client_proc_init(fld);
 378         if (rc)
 379                 GOTO(out, rc);
 380         EXIT;
 381 out:
 382         if (rc)
 383                 fld_client_fini(fld);
 384         else
 385                 CDEBUG(D_INFO, "%s: Using \"%s\" hash\n",
 386                        fld->lcf_name, fld->lcf_hash->fh_name);
 387         return rc;
 388 }
 389 EXPORT_SYMBOL(fld_client_init);
 390
 391 void fld_client_fini(struct lu_client_fld *fld)
 392 {
 393         struct lu_fld_target *target, *tmp;
 394         ENTRY;
 395
 396         spin_lock(&fld->lcf_lock);
 397         cfs_list_for_each_entry_safe(target, tmp,
 398                                      &fld->lcf_targets, ft_chain) {
 399                 fld->lcf_count--;
 400                 cfs_list_del(&target->ft_chain);
 401                 if (target->ft_exp != NULL)
 402                         class_export_put(target->ft_exp);
 403                 OBD_FREE_PTR(target);
 404         }
 405         spin_unlock(&fld->lcf_lock);
 406
 407         if (fld->lcf_cache != NULL) {
 408                 if (!IS_ERR(fld->lcf_cache))
 409                         fld_cache_fini(fld->lcf_cache);
 410                 fld->lcf_cache = NULL;
 411         }
 412
 413         EXIT;
 414 }
 415 EXPORT_SYMBOL(fld_client_fini);
 416
 417 int fld_client_rpc(struct obd_export *exp,
 418                    struct lu_seq_range *range, __u32 fld_op)
 419 {
 420         struct ptlrpc_request *req;
 421         struct lu_seq_range      *prange;
 422         __u32                 *op;
 423         int                    rc;
 424         ENTRY;
 425
 426         LASSERT(exp != NULL);
 427
 428         req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp), &RQF_FLD_QUERY,
 429                                         LUSTRE_MDS_VERSION, FLD_QUERY);
 430         if (req == NULL)
 431                 RETURN(-ENOMEM);
 432
 433         op = req_capsule_client_get(&req->rq_pill, &RMF_FLD_OPC);
 434         *op = fld_op;
 435
 436         prange = req_capsule_client_get(&req->rq_pill, &RMF_FLD_MDFLD);
 437         *prange = *range;
 438
 439         ptlrpc_request_set_replen(req);
 440         req->rq_request_portal = FLD_REQUEST_PORTAL;
 441         ptlrpc_at_set_req_timeout(req);
 442
 443         if (fld_op != FLD_LOOKUP)
 444                 mdc_get_rpc_lock(exp->exp_obd->u.cli.cl_rpc_lock, NULL);
 445         fld_enter_request(&exp->exp_obd->u.cli);
 446         rc = ptlrpc_queue_wait(req);
 447         fld_exit_request(&exp->exp_obd->u.cli);
 448         if (fld_op != FLD_LOOKUP)
 449                 mdc_put_rpc_lock(exp->exp_obd->u.cli.cl_rpc_lock, NULL);
 450         if (rc)
 451                 GOTO(out_req, rc);
 452
 453         prange = req_capsule_server_get(&req->rq_pill, &RMF_FLD_MDFLD);
 454         if (prange == NULL)
 455                 GOTO(out_req, rc = -EFAULT);
 456         *range = *prange;
 457         EXIT;
 458 out_req:
 459         ptlrpc_req_finished(req);
 460         return rc;
 461 }
 462
 463 int fld_client_lookup(struct lu_client_fld *fld, seqno_t seq, mdsno_t *mds,
 464                       __u32 flags, const struct lu_env *env)
 465 {
 466         struct lu_seq_range res;
 467         struct lu_fld_target *target;
 468         int rc;
 469         ENTRY;
 470
 471         fld->lcf_flags |= LUSTRE_FLD_RUN;
 472
 473         rc = fld_cache_lookup(fld->lcf_cache, seq, &res);
 474         if (rc == 0) {
 475                 *mds = res.lsr_index;
 476                 RETURN(0);
 477         }
 478
 479         /* Can not find it in the cache */
 480         target = fld_client_get_target(fld, seq);
 481         LASSERT(target != NULL);
 482
 483         CDEBUG(D_INFO, "%s: Lookup fld entry (seq: "LPX64") on "
 484                "target %s (idx "LPU64")\n", fld->lcf_name, seq,
 485                fld_target_name(target), target->ft_idx);
 486
 487         res.lsr_start = seq;
 488         res.lsr_flags = flags;
 489 #ifdef __KERNEL__
 490         if (target->ft_srv != NULL) {
 491                 LASSERT(env != NULL);
 492                 rc = fld_server_lookup(env, target->ft_srv, seq, &res);
 493         } else {
 494 #endif
 495                 rc = fld_client_rpc(target->ft_exp,
 496                                     &res, FLD_LOOKUP);
 497 #ifdef __KERNEL__
 498         }
 499 #endif
 500
 501         if (rc == 0) {
 502                 *mds = res.lsr_index;
 503
 504                 fld_cache_insert(fld->lcf_cache, &res);
 505         }
 506         RETURN(rc);
 507 }
 508 EXPORT_SYMBOL(fld_client_lookup);
 509
 510 void fld_client_flush(struct lu_client_fld *fld)
 511 {
 512         fld_cache_flush(fld->lcf_cache);
 513 }
 514 EXPORT_SYMBOL(fld_client_flush);