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