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_handler.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
  56 fld_rrb_hash(struct lu_client_fld *fld, 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
  65 fld_dht_hash(struct lu_client_fld *fld, seqno_t seq)
  66 {
  67         /* XXX: here should 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
  90 fld_client_apt_target(struct fld_target *target, 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, seqno_t seq)
  98 {
  99         struct fld_target *target;
 100         int hash;
 101         ENTRY;
 102
 103         LASSERT(fld->fld_hash != NULL);
 104
 105         spin_lock(&fld->fld_lock);
 106         hash = fld->fld_hash->fh_func(fld, seq);
 107
 108         list_for_each_entry(target,
 109                             &fld->fld_targets, fldt_chain) {
 110                 if (fld_client_apt_target(target, hash)) {
 111                         spin_unlock(&fld->fld_lock);
 112                         RETURN(target);
 113                 }
 114         }
 115         spin_unlock(&fld->fld_lock);
 116
 117         /* if target is not found, there is logical error anyway, so here is
 118          * LBUG() to catch that situation. */
 119         LBUG();
 120         RETURN(NULL);
 121 }
 122
 123 /* add export to FLD. This is usually done by CMM and LMV as they are main users
 124  * of FLD module. */
 125 int
 126 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, "FLD(cli): adding export %s\n",
 136                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
 167 fld_client_del_target(struct lu_client_fld *fld,
 168                       struct obd_export *exp)
 169 {
 170         struct fld_target *target, *tmp;
 171         ENTRY;
 172
 173         spin_lock(&fld->fld_lock);
 174         list_for_each_entry_safe(target, tmp,
 175                                  &fld->fld_targets, fldt_chain) {
 176                 if (obd_uuid_equals(&target->fldt_exp->exp_client_uuid,
 177                                     &exp->exp_client_uuid))
 178                 {
 179                         fld->fld_count--;
 180                         list_del(&target->fldt_chain);
 181                         spin_unlock(&fld->fld_lock);
 182                         class_export_put(target->fldt_exp);
 183                         OBD_FREE_PTR(target);
 184                         RETURN(0);
 185                 }
 186         }
 187         spin_unlock(&fld->fld_lock);
 188         RETURN(-ENOENT);
 189 }
 190 EXPORT_SYMBOL(fld_client_del_target);
 191
 192 #ifdef LPROCFS
 193 static int
 194 fld_client_proc_init(struct lu_client_fld *fld)
 195 {
 196         int rc;
 197         ENTRY;
 198
 199         fld->fld_proc_dir = lprocfs_register(fld->fld_name,
 200                                              proc_lustre_root,
 201                                              NULL, NULL);
 202
 203         if (IS_ERR(fld->fld_proc_dir)) {
 204                 CERROR("LProcFS failed in fld-init\n");
 205                 rc = PTR_ERR(fld->fld_proc_dir);
 206                 GOTO(err, rc);
 207         }
 208
 209         rc = lprocfs_add_vars(fld->fld_proc_dir,
 210                               fld_client_proc_list, fld);
 211         if (rc) {
 212                 CERROR("can't init FLD "
 213                        "proc, rc %d\n", rc);
 214                 GOTO(err, rc);
 215         }
 216
 217         RETURN(0);
 218
 219 err:
 220         fld->fld_proc_dir = NULL;
 221         return rc;
 222 }
 223
 224 static void
 225 fld_client_proc_fini(struct lu_client_fld *fld)
 226 {
 227         ENTRY;
 228         if (fld->fld_proc_dir) {
 229                 lprocfs_remove(fld->fld_proc_dir);
 230                 fld->fld_proc_dir = NULL;
 231         }
 232         EXIT;
 233 }
 234 #endif
 235
 236 static inline int hash_is_sane(int hash)
 237 {
 238         return (hash >= 0 && hash < ARRAY_SIZE(fld_hash));
 239 }
 240
 241 int
 242 fld_client_init(struct lu_client_fld *fld,
 243                 const char *uuid, int hash)
 244 {
 245         int rc = 0;
 246         ENTRY;
 247
 248         LASSERT(fld != NULL);
 249
 250         if (!hash_is_sane(hash)) {
 251                 CERROR("wrong hash function 0x%x\n", hash);
 252                 RETURN(-EINVAL);
 253         }
 254
 255         INIT_LIST_HEAD(&fld->fld_targets);
 256         spin_lock_init(&fld->fld_lock);
 257         fld->fld_hash = &fld_hash[hash];
 258         fld->fld_count = 0;
 259
 260         snprintf(fld->fld_name, sizeof(fld->fld_name),
 261                  "%s-%s", LUSTRE_FLD_NAME, uuid);
 262
 263 #ifdef __KERNEL__
 264         fld->fld_cache = fld_cache_init(FLD_HTABLE_SIZE);
 265         if (IS_ERR(fld->fld_cache)) {
 266                 rc = PTR_ERR(fld->fld_cache);
 267                 fld->fld_cache = NULL;
 268                 GOTO(out, rc);
 269         }
 270 #endif
 271
 272 #ifdef LPROCFS
 273         rc = fld_client_proc_init(fld);
 274         if (rc)
 275                 GOTO(out, rc);
 276 #endif
 277         EXIT;
 278 out:
 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
 290 fld_client_fini(struct lu_client_fld *fld)
 291 {
 292         struct fld_target *target, *tmp;
 293         ENTRY;
 294
 295 #ifdef LPROCFS
 296         fld_client_proc_fini(fld);
 297 #endif
 298
 299         spin_lock(&fld->fld_lock);
 300         list_for_each_entry_safe(target, tmp,
 301                                  &fld->fld_targets, fldt_chain) {
 302                 fld->fld_count--;
 303                 list_del(&target->fldt_chain);
 304                 class_export_put(target->fldt_exp);
 305                 OBD_FREE_PTR(target);
 306         }
 307         spin_unlock(&fld->fld_lock);
 308
 309 #ifdef __KERNEL__
 310         if (fld->fld_cache != NULL) {
 311                 fld_cache_fini(fld->fld_cache);
 312                 fld->fld_cache = NULL;
 313         }
 314 #endif
 315
 316         CDEBUG(D_INFO|D_WARNING, "Client FLD finalized\n");
 317         EXIT;
 318 }
 319 EXPORT_SYMBOL(fld_client_fini);
 320
 321 static int
 322 fld_client_rpc(struct obd_export *exp,
 323                struct md_fld *mf, __u32 fld_op)
 324 {
 325         int size[2] = {sizeof(__u32), sizeof(struct md_fld)}, rc;
 326         int mf_size = sizeof(struct md_fld);
 327         struct ptlrpc_request *req;
 328         struct md_fld *pmf;
 329         __u32 *op;
 330         ENTRY;
 331
 332         LASSERT(exp != NULL);
 333
 334         req = ptlrpc_prep_req(class_exp2cliimp(exp),
 335                               LUSTRE_MDS_VERSION, FLD_QUERY,
 336                               2, size, NULL);
 337         if (req == NULL)
 338                 RETURN(-ENOMEM);
 339
 340         op = lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*op));
 341         *op = fld_op;
 342
 343         pmf = lustre_msg_buf(req->rq_reqmsg, 1, sizeof (*pmf));
 344         *pmf = *mf;
 345
 346         req->rq_replen = lustre_msg_size(1, &mf_size);
 347         req->rq_request_portal = FLD_REQUEST_PORTAL;
 348
 349         rc = ptlrpc_queue_wait(req);
 350         if (rc)
 351                 GOTO(out_req, rc);
 352
 353         pmf = lustre_swab_repbuf(req, 0, sizeof(*pmf),
 354                                  lustre_swab_md_fld);
 355         *mf = *pmf;
 356 out_req:
 357         ptlrpc_req_finished(req);
 358         RETURN(rc);
 359 }
 360
 361 static int
 362 __fld_client_create(struct lu_client_fld *fld,
 363                     seqno_t seq, mdsno_t mds,
 364                     struct md_fld *md_fld)
 365 {
 366         struct fld_target *target;
 367         __u32 rc;
 368         ENTRY;
 369
 370         target = fld_client_get_target(fld, seq);
 371         if (!target)
 372                 RETURN(-EINVAL);
 373
 374         rc = fld_client_rpc(target->fldt_exp, md_fld, FLD_CREATE);
 375
 376         if (rc  == 0) {
 377                 /* do not return result of calling fld_cache_insert()
 378                  * here. First of all because it may return -EEXISTS. Another
 379                  * reason is that, we do not want to stop proceeding because of
 380                  * cache errors. --umka */
 381                 fld_cache_insert(fld->fld_cache, seq, mds);
 382         }
 383
 384         RETURN(rc);
 385 }
 386
 387 int
 388 fld_client_create(struct lu_client_fld *fld,
 389                   seqno_t seq, mdsno_t mds)
 390 {
 391         struct md_fld md_fld = { .mf_seq = seq, .mf_mds = mds };
 392         __u32 rc;
 393         ENTRY;
 394
 395         rc = __fld_client_create(fld, seq, mds, &md_fld);
 396         RETURN(rc);
 397 }
 398 EXPORT_SYMBOL(fld_client_create);
 399
 400 static int
 401 __fld_client_delete(struct lu_client_fld *fld,
 402                     seqno_t seq, struct md_fld *md_fld)
 403 {
 404         struct fld_target *target;
 405         __u32 rc;
 406
 407         fld_cache_delete(fld->fld_cache, seq);
 408
 409         target = fld_client_get_target(fld, seq);
 410         if (!target)
 411                 RETURN(-EINVAL);
 412
 413         rc = fld_client_rpc(target->fldt_exp,
 414                             md_fld, FLD_DELETE);
 415         RETURN(rc);
 416 }
 417
 418 int
 419 fld_client_delete(struct lu_client_fld *fld,
 420                   seqno_t seq)
 421 {
 422         struct md_fld md_fld = { .mf_seq = seq, .mf_mds = 0 };
 423         __u32 rc;
 424
 425         rc = __fld_client_delete(fld, seq, &md_fld);
 426         RETURN(rc);
 427 }
 428 EXPORT_SYMBOL(fld_client_delete);
 429
 430 static int
 431 __fld_client_lookup(struct lu_client_fld *fld,
 432                     seqno_t seq, mdsno_t *mds,
 433                     struct md_fld *md_fld)
 434 {
 435         struct fld_target *target;
 436         int rc;
 437         ENTRY;
 438
 439         /* lookup it in the cache */
 440         rc = fld_cache_lookup(fld->fld_cache, seq, mds);
 441         if (rc == 0)
 442                 RETURN(0);
 443
 444         /* can not find it in the cache */
 445         target = fld_client_get_target(fld, seq);
 446         if (!target)
 447                 RETURN(-EINVAL);
 448
 449         rc = fld_client_rpc(target->fldt_exp,
 450                             md_fld, FLD_LOOKUP);
 451         if (rc == 0)
 452                 *mds = md_fld->mf_mds;
 453
 454         /* do not return error here as well. See previous comment in same
 455          * situation in function fld_client_create(). --umka */
 456         fld_cache_insert(fld->fld_cache, seq, *mds);
 457
 458         RETURN(rc);
 459 }
 460
 461 int
 462 fld_client_lookup(struct lu_client_fld *fld,
 463                   seqno_t seq, mdsno_t *mds)
 464 {
 465         struct md_fld md_fld = { .mf_seq = seq, .mf_mds = 0 };
 466         int rc;
 467         ENTRY;
 468
 469         rc = __fld_client_lookup(fld, seq, mds, &md_fld);
 470         RETURN(rc);
 471 }
 472 EXPORT_SYMBOL(fld_client_lookup);