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  *  lustre/fld/fld_request.c
   5  *  FLD (Fids Location Database)
   6  *
   7  *  Copyright (C) 2006 Cluster File Systems, Inc.
   8  *   Author: Yury Umanets <umka@clusterfs.com>
   9  *
  10  *   This file is part of the Lustre file system, http://www.lustre.org
  11  *   Lustre is a trademark of Cluster File Systems, Inc.
  12  *
  13  *   You may have signed or agreed to another license before downloading
  14  *   this software.  If so, you are bound by the terms and conditions
  15  *   of that agreement, and the following does not apply to you.  See the
  16  *   LICENSE file included with this distribution for more information.
  17  *
  18  *   If you did not agree to a different license, then this copy of Lustre
  19  *   is open source software; you can redistribute it and/or modify it
  20  *   under the terms of version 2 of the GNU General Public License as
  21  *   published by the Free Software Foundation.
  22  *
  23  *   In either case, Lustre is distributed in the hope that it will be
  24  *   useful, but WITHOUT ANY WARRANTY; without even the implied warranty
  25  *   of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  26  *   license text for more details.
  27  */
  28 #ifndef EXPORT_SYMTAB
  29 # define EXPORT_SYMTAB
  30 #endif
  31 #define DEBUG_SUBSYSTEM S_FLD
  32
  33 #ifdef __KERNEL__
  34 # include <libcfs/libcfs.h>
  35 # include <linux/module.h>
  36 # include <linux/jbd.h>
  37 # include <asm/div64.h>
  38 #else /* __KERNEL__ */
  39 # include <liblustre.h>
  40 # include <libcfs/list.h>
  41 #endif
  42
  43 #include <obd.h>
  44 #include <obd_class.h>
  45 #include <lustre_ver.h>
  46 #include <obd_support.h>
  47 #include <lprocfs_status.h>
  48
  49 #include <dt_object.h>
  50 #include <md_object.h>
  51 #include <lustre_req_layout.h>
  52 #include <lustre_fld.h>
  53 #include "fld_internal.h"
  54
  55 static int fld_rrb_hash(struct lu_client_fld *fld,
  56                         seqno_t seq)
  57 {
  58         if (fld->fld_count == 0)
  59                 return 0;
  60
  61         return do_div(seq, fld->fld_count);
  62 }
  63
  64 static int fld_dht_hash(struct lu_client_fld *fld,
  65                         seqno_t seq)
  66 {
  67         /* XXX: here should be DHT hash */
  68         return fld_rrb_hash(fld, seq);
  69 }
  70
  71 struct lu_fld_hash fld_hash[3] = {
  72         {
  73                 .fh_name = "DHT",
  74                 .fh_func = fld_dht_hash
  75         },
  76         {
  77                 .fh_name = "Round Robin",
  78                 .fh_func = fld_rrb_hash
  79         },
  80         {
  81                 0,
  82         }
  83 };
  84
  85 /* this function makes decision if passed @target appropriate acoordingly to
  86  * passed @hash. In case of usual round-robin hash, this is decided by comparing
  87  * hash and target's index. In the case of DHT, algorithm is a bit more
  88  * complicated. */
  89 static int fld_client_apt_target(struct fld_target *target,
  90                                  int hash)
  91 {
  92         /* XXX: DHT case should be worked out. */
  93         return (target->fldt_idx == hash);
  94 }
  95
  96 static struct fld_target *
  97 fld_client_get_target(struct lu_client_fld *fld,
  98                       seqno_t seq)
  99 {
 100         struct fld_target *target;
 101         int hash;
 102         ENTRY;
 103
 104         LASSERT(fld->fld_hash != NULL);
 105
 106         spin_lock(&fld->fld_lock);
 107         hash = fld->fld_hash->fh_func(fld, seq);
 108
 109         list_for_each_entry(target,
 110                             &fld->fld_targets, fldt_chain) {
 111                 if (fld_client_apt_target(target, hash)) {
 112                         spin_unlock(&fld->fld_lock);
 113                         RETURN(target);
 114                 }
 115         }
 116         spin_unlock(&fld->fld_lock);
 117
 118         /* if target is not found, there is logical error anyway, so here is
 119          * LBUG() to catch that situation. */
 120         LBUG();
 121         RETURN(NULL);
 122 }
 123
 124 /* add export to FLD. This is usually done by CMM and LMV as they are main users
 125  * of FLD module. */
 126 int fld_client_add_target(struct lu_client_fld *fld,
 127                           struct obd_export *exp)
 128 {
 129         struct client_obd *cli = &exp->exp_obd->u.cli;
 130         struct fld_target *target, *tmp;
 131         ENTRY;
 132
 133         LASSERT(exp != NULL);
 134
 135         CDEBUG(D_INFO|D_WARNING, "%s: adding export %s\n",
 136                fld->fld_name, cli->cl_target_uuid.uuid);
 137
 138         OBD_ALLOC_PTR(target);
 139         if (target == NULL)
 140                 RETURN(-ENOMEM);
 141
 142         spin_lock(&fld->fld_lock);
 143         list_for_each_entry(tmp, &fld->fld_targets, fldt_chain) {
 144                 if (obd_uuid_equals(&tmp->fldt_exp->exp_client_uuid,
 145                                     &exp->exp_client_uuid))
 146                 {
 147                         spin_unlock(&fld->fld_lock);
 148                         OBD_FREE_PTR(target);
 149                         RETURN(-EEXIST);
 150                 }
 151         }
 152
 153         target->fldt_exp = class_export_get(exp);
 154         target->fldt_idx = fld->fld_count;
 155
 156         list_add_tail(&target->fldt_chain,
 157                       &fld->fld_targets);
 158         fld->fld_count++;
 159         spin_unlock(&fld->fld_lock);
 160
 161         RETURN(0);
 162 }
 163 EXPORT_SYMBOL(fld_client_add_target);
 164
 165 /* remove export from FLD */
 166 int fld_client_del_target(struct lu_client_fld *fld,
 167                           struct obd_export *exp)
 168 {
 169         struct fld_target *target, *tmp;
 170         ENTRY;
 171
 172         spin_lock(&fld->fld_lock);
 173         list_for_each_entry_safe(target, tmp,
 174                                  &fld->fld_targets, fldt_chain) {
 175                 if (obd_uuid_equals(&target->fldt_exp->exp_client_uuid,
 176                                     &exp->exp_client_uuid))
 177                 {
 178                         fld->fld_count--;
 179                         list_del(&target->fldt_chain);
 180                         spin_unlock(&fld->fld_lock);
 181                         class_export_put(target->fldt_exp);
 182                         OBD_FREE_PTR(target);
 183                         RETURN(0);
 184                 }
 185         }
 186         spin_unlock(&fld->fld_lock);
 187         RETURN(-ENOENT);
 188 }
 189 EXPORT_SYMBOL(fld_client_del_target);
 190
 191 #ifdef LPROCFS
 192 static int fld_client_proc_init(struct lu_client_fld *fld)
 193 {
 194         int rc;
 195         ENTRY;
 196
 197         fld->fld_proc_dir = lprocfs_register(fld->fld_name,
 198                                              proc_lustre_root,
 199                                              NULL, NULL);
 200
 201         if (IS_ERR(fld->fld_proc_dir)) {
 202                 CERROR("LProcFS failed in fld-init\n");
 203                 rc = PTR_ERR(fld->fld_proc_dir);
 204                 GOTO(err, rc);
 205         }
 206
 207         rc = lprocfs_add_vars(fld->fld_proc_dir,
 208                               fld_client_proc_list, fld);
 209         if (rc) {
 210                 CERROR("can't init FLD "
 211                        "proc, rc %d\n", rc);
 212                 GOTO(err_dir, rc);
 213         }
 214
 215         RETURN(0);
 216
 217 err_dir:
 218         lprocfs_remove(fld->fld_proc_dir);
 219 err:
 220         fld->fld_proc_dir = NULL;
 221         return rc;
 222 }
 223
 224 static void fld_client_proc_fini(struct lu_client_fld *fld)
 225 {
 226         ENTRY;
 227         if (fld->fld_proc_dir) {
 228                 lprocfs_remove(fld->fld_proc_dir);
 229                 fld->fld_proc_dir = NULL;
 230         }
 231         EXIT;
 232 }
 233 #endif
 234
 235 static inline int hash_is_sane(int hash)
 236 {
 237         return (hash >= 0 && hash < ARRAY_SIZE(fld_hash));
 238 }
 239
 240 int fld_client_init(struct lu_client_fld *fld,
 241                     const char *uuid, int hash)
 242 {
 243         int rc = 0;
 244         ENTRY;
 245
 246         LASSERT(fld != NULL);
 247
 248         if (!hash_is_sane(hash)) {
 249                 CERROR("wrong hash function 0x%x\n", hash);
 250                 RETURN(-EINVAL);
 251         }
 252
 253         INIT_LIST_HEAD(&fld->fld_targets);
 254         spin_lock_init(&fld->fld_lock);
 255         fld->fld_hash = &fld_hash[hash];
 256         fld->fld_count = 0;
 257
 258         snprintf(fld->fld_name, sizeof(fld->fld_name),
 259                  "%s-cli-%s", LUSTRE_FLD_NAME, uuid);
 260
 261 #ifdef __KERNEL__
 262         fld->fld_cache = fld_cache_init(FLD_HTABLE_SIZE);
 263         if (IS_ERR(fld->fld_cache)) {
 264                 rc = PTR_ERR(fld->fld_cache);
 265                 fld->fld_cache = NULL;
 266                 GOTO(out, rc);
 267         }
 268 #endif
 269
 270 #ifdef LPROCFS
 271         rc = fld_client_proc_init(fld);
 272         if (rc)
 273                 GOTO(out, rc);
 274 #endif
 275         EXIT;
 276 #ifdef __KERNEL__
 277 out:
 278 #endif
 279         if (rc)
 280                 fld_client_fini(fld);
 281         else
 282                 CDEBUG(D_INFO|D_WARNING,
 283                        "Client FLD, using \"%s\" hash\n",
 284                        fld->fld_hash->fh_name);
 285         return rc;
 286 }
 287 EXPORT_SYMBOL(fld_client_init);
 288
 289 void fld_client_fini(struct lu_client_fld *fld)
 290 {
 291         struct fld_target *target, *tmp;
 292         ENTRY;
 293
 294 #ifdef LPROCFS
 295         fld_client_proc_fini(fld);
 296 #endif
 297
 298         spin_lock(&fld->fld_lock);
 299         list_for_each_entry_safe(target, tmp,
 300                                  &fld->fld_targets, fldt_chain) {
 301                 fld->fld_count--;
 302                 list_del(&target->fldt_chain);
 303                 class_export_put(target->fldt_exp);
 304                 OBD_FREE_PTR(target);
 305         }
 306         spin_unlock(&fld->fld_lock);
 307
 308 #ifdef __KERNEL__
 309         if (fld->fld_cache != NULL) {
 310                 fld_cache_fini(fld->fld_cache);
 311                 fld->fld_cache = NULL;
 312         }
 313 #endif
 314
 315         CDEBUG(D_INFO|D_WARNING, "Client FLD finalized\n");
 316         EXIT;
 317 }
 318 EXPORT_SYMBOL(fld_client_fini);
 319
 320 static int fld_client_rpc(struct obd_export *exp,
 321                           struct md_fld *mf, __u32 fld_op)
 322 {
 323         int size[2] = {sizeof(__u32), sizeof(struct md_fld)}, rc;
 324         int mf_size = sizeof(struct md_fld);
 325         struct ptlrpc_request *req;
 326         struct req_capsule pill;
 327         struct md_fld *pmf;
 328         __u32 *op;
 329         ENTRY;
 330
 331         LASSERT(exp != NULL);
 332
 333         req = ptlrpc_prep_req(class_exp2cliimp(exp),
 334                               LUSTRE_MDS_VERSION, FLD_QUERY,
 335                               2, size, NULL);
 336         if (req == NULL)
 337                 RETURN(-ENOMEM);
 338
 339         req_capsule_init(&pill, req, RCL_CLIENT,
 340                          &mf_size);
 341
 342         req_capsule_set(&pill, &RQF_FLD_QUERY);
 343
 344         op = req_capsule_client_get(&pill, &RMF_FLD_OPC);
 345         *op = fld_op;
 346
 347         pmf = req_capsule_client_get(&pill, &RMF_FLD_MDFLD);
 348         *pmf = *mf;
 349
 350         req->rq_replen = lustre_msg_size(1, &mf_size);
 351         req->rq_request_portal = FLD_REQUEST_PORTAL;
 352
 353         rc = ptlrpc_queue_wait(req);
 354         if (rc)
 355                 GOTO(out_req, rc);
 356
 357         pmf = req_capsule_server_get(&pill, &RMF_FLD_MDFLD);
 358         if (pmf == NULL) {
 359                 CERROR("Can't unpack FLD response\n");
 360                 GOTO(out_req, rc = -EFAULT);
 361         }
 362         *mf = *pmf;
 363         EXIT;
 364 out_req:
 365         req_capsule_fini(&pill);
 366         ptlrpc_req_finished(req);
 367         return rc;
 368 }
 369
 370 static int __fld_client_create(struct lu_client_fld *fld,
 371                                seqno_t seq, mdsno_t mds,
 372                     struct md_fld *md_fld)
 373 {
 374         struct fld_target *target;
 375         __u32 rc;
 376         ENTRY;
 377
 378         target = fld_client_get_target(fld, seq);
 379         if (!target)
 380                 RETURN(-EINVAL);
 381
 382         rc = fld_client_rpc(target->fldt_exp, md_fld, FLD_CREATE);
 383
 384         if (rc  == 0) {
 385                 /* do not return result of calling fld_cache_insert()
 386                  * here. First of all because it may return -EEXISTS. Another
 387                  * reason is that, we do not want to stop proceeding because of
 388                  * cache errors. --umka */
 389                 fld_cache_insert(fld->fld_cache, seq, mds);
 390         }
 391
 392         RETURN(rc);
 393 }
 394
 395 int fld_client_create(struct lu_client_fld *fld,
 396                       seqno_t seq, mdsno_t mds)
 397 {
 398         struct md_fld md_fld = { .mf_seq = seq, .mf_mds = mds };
 399         __u32 rc;
 400         ENTRY;
 401
 402         rc = __fld_client_create(fld, seq, mds, &md_fld);
 403         RETURN(rc);
 404 }
 405 EXPORT_SYMBOL(fld_client_create);
 406
 407 static int __fld_client_delete(struct lu_client_fld *fld,
 408                                seqno_t seq, struct md_fld *md_fld)
 409 {
 410         struct fld_target *target;
 411         __u32 rc;
 412
 413         fld_cache_delete(fld->fld_cache, seq);
 414
 415         target = fld_client_get_target(fld, seq);
 416         if (!target)
 417                 RETURN(-EINVAL);
 418
 419         rc = fld_client_rpc(target->fldt_exp,
 420                             md_fld, FLD_DELETE);
 421         RETURN(rc);
 422 }
 423
 424 int fld_client_delete(struct lu_client_fld *fld,
 425                       seqno_t seq)
 426 {
 427         struct md_fld md_fld = { .mf_seq = seq, .mf_mds = 0 };
 428         __u32 rc;
 429
 430         rc = __fld_client_delete(fld, seq, &md_fld);
 431         RETURN(rc);
 432 }
 433 EXPORT_SYMBOL(fld_client_delete);
 434
 435 static int __fld_client_lookup(struct lu_client_fld *fld,
 436                                seqno_t seq, mdsno_t *mds,
 437                                struct md_fld *md_fld)
 438 {
 439         struct fld_target *target;
 440         int rc;
 441         ENTRY;
 442
 443         /* lookup it in the cache */
 444         rc = fld_cache_lookup(fld->fld_cache, seq, mds);
 445         if (rc == 0)
 446                 RETURN(0);
 447
 448         /* can not find it in the cache */
 449         target = fld_client_get_target(fld, seq);
 450         if (!target)
 451                 RETURN(-EINVAL);
 452
 453         rc = fld_client_rpc(target->fldt_exp,
 454                             md_fld, FLD_LOOKUP);
 455         if (rc == 0)
 456                 *mds = md_fld->mf_mds;
 457
 458         /* do not return error here as well. See previous comment in same
 459          * situation in function fld_client_create(). --umka */
 460         fld_cache_insert(fld->fld_cache, seq, *mds);
 461
 462         RETURN(rc);
 463 }
 464
 465 int fld_client_lookup(struct lu_client_fld *fld,
 466                       seqno_t seq, mdsno_t *mds)
 467 {
 468         struct md_fld md_fld = { .mf_seq = seq, .mf_mds = 0 };
 469         int rc;
 470         ENTRY;
 471
 472         rc = __fld_client_lookup(fld, seq, mds, &md_fld);
 473         RETURN(rc);
 474 }
 475 EXPORT_SYMBOL(fld_client_lookup);