1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
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.
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).
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
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
29 * Copyright 2008 Sun Microsystems, Inc. All rights reserved
30 * Use is subject to license terms.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lustre/fld/fld_request.c
38 * FLD (Fids Location Database)
40 * Author: Yury Umanets <umka@clusterfs.com>
44 # define EXPORT_SYMTAB
46 #define DEBUG_SUBSYSTEM S_FLD
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>
59 #include <obd_class.h>
60 #include <lustre_ver.h>
61 #include <obd_support.h>
62 #include <lprocfs_status.h>
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"
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)
77 spin_lock(&cli->cl_loi_list_lock);
78 rc = list_empty(&mcw->mcw_entry);
79 spin_unlock(&cli->cl_loi_list_lock);
83 static void fld_enter_request(struct client_obd *cli)
85 struct mdc_cache_waiter mcw;
86 struct l_wait_info lwi = { 0 };
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);
95 cli->cl_r_in_flight++;
96 spin_unlock(&cli->cl_loi_list_lock);
100 static void fld_exit_request(struct client_obd *cli)
102 struct list_head *l, *tmp;
103 struct mdc_cache_waiter *mcw;
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) {
109 if (cli->cl_r_in_flight >= cli->cl_max_rpcs_in_flight) {
110 /* No free request slots anymore */
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);
119 spin_unlock(&cli->cl_loi_list_lock);
122 static int fld_rrb_hash(struct lu_client_fld *fld,
125 LASSERT(fld->lcf_count > 0);
126 return do_div(seq, fld->lcf_count);
129 static struct lu_fld_target *
130 fld_rrb_scan(struct lu_client_fld *fld, seqno_t seq)
132 struct lu_fld_target *target;
136 hash = fld_rrb_hash(fld, seq);
138 list_for_each_entry(target, &fld->lcf_targets, ft_chain) {
139 if (target->ft_idx == hash)
143 CERROR("%s: Can't find target by hash %d (seq "LPX64"). "
144 "Targets (%d):\n", fld->lcf_name, hash, seq,
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 :
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);
160 * If target is not found, there is logical error anyway, so here is
161 * LBUG() to catch this situation.
167 static int fld_dht_hash(struct lu_client_fld *fld,
170 /* XXX: here should be DHT hash */
171 return fld_rrb_hash(fld, seq);
174 static struct lu_fld_target *
175 fld_dht_scan(struct lu_client_fld *fld, seqno_t seq)
177 /* XXX: here should be DHT scan code */
178 return fld_rrb_scan(fld, seq);
181 struct lu_fld_hash fld_hash[3] = {
184 .fh_hash_func = fld_dht_hash,
185 .fh_scan_func = fld_dht_scan
189 .fh_hash_func = fld_rrb_hash,
190 .fh_scan_func = fld_rrb_scan
197 static struct lu_fld_target *
198 fld_client_get_target(struct lu_client_fld *fld,
201 struct lu_fld_target *target;
204 LASSERT(fld->lcf_hash != NULL);
206 spin_lock(&fld->lcf_lock);
207 target = fld->lcf_hash->fh_scan_func(fld, seq);
208 spin_unlock(&fld->lcf_lock);
210 if (target != NULL) {
211 CDEBUG(D_INFO, "%s: Found target (idx "LPU64
212 ") by seq "LPX64"\n", fld->lcf_name,
213 target->ft_idx, seq);
220 * Add export to FLD. This is usually done by CMM and LMV as they are main users
223 int fld_client_add_target(struct lu_client_fld *fld,
224 struct lu_fld_target *tar)
226 const char *name = fld_target_name(tar);
227 struct lu_fld_target *target, *tmp;
230 LASSERT(tar != NULL);
231 LASSERT(name != NULL);
232 LASSERT(tar->ft_srv != NULL || tar->ft_exp != NULL);
234 if (fld->lcf_flags != LUSTRE_FLD_INIT) {
235 CERROR("%s: Attempt to add target %s (idx "LPU64") "
236 "on fly - skip it\n", fld->lcf_name, name,
240 CDEBUG(D_INFO, "%s: Adding target %s (idx "
241 LPU64")\n", fld->lcf_name, name, tar->ft_idx);
244 OBD_ALLOC_PTR(target);
248 spin_lock(&fld->lcf_lock);
249 list_for_each_entry(tmp, &fld->lcf_targets, ft_chain) {
250 if (tmp->ft_idx == tar->ft_idx) {
251 spin_unlock(&fld->lcf_lock);
252 OBD_FREE_PTR(target);
253 CERROR("Target %s exists in FLD and known as %s:#"LPU64"\n",
254 name, fld_target_name(tmp), tmp->ft_idx);
259 target->ft_exp = tar->ft_exp;
260 if (target->ft_exp != NULL)
261 class_export_get(target->ft_exp);
262 target->ft_srv = tar->ft_srv;
263 target->ft_idx = tar->ft_idx;
265 list_add_tail(&target->ft_chain,
269 spin_unlock(&fld->lcf_lock);
273 EXPORT_SYMBOL(fld_client_add_target);
275 /* Remove export from FLD */
276 int fld_client_del_target(struct lu_client_fld *fld,
279 struct lu_fld_target *target, *tmp;
282 spin_lock(&fld->lcf_lock);
283 list_for_each_entry_safe(target, tmp,
284 &fld->lcf_targets, ft_chain) {
285 if (target->ft_idx == idx) {
287 list_del(&target->ft_chain);
288 spin_unlock(&fld->lcf_lock);
290 if (target->ft_exp != NULL)
291 class_export_put(target->ft_exp);
293 OBD_FREE_PTR(target);
297 spin_unlock(&fld->lcf_lock);
300 EXPORT_SYMBOL(fld_client_del_target);
302 static void fld_client_proc_fini(struct lu_client_fld *fld);
305 static int fld_client_proc_init(struct lu_client_fld *fld)
310 fld->lcf_proc_dir = lprocfs_register(fld->lcf_name,
314 if (IS_ERR(fld->lcf_proc_dir)) {
315 CERROR("%s: LProcFS failed in fld-init\n",
317 rc = PTR_ERR(fld->lcf_proc_dir);
321 rc = lprocfs_add_vars(fld->lcf_proc_dir,
322 fld_client_proc_list, fld);
324 CERROR("%s: Can't init FLD proc, rc %d\n",
326 GOTO(out_cleanup, rc);
332 fld_client_proc_fini(fld);
336 static void fld_client_proc_fini(struct lu_client_fld *fld)
339 if (fld->lcf_proc_dir) {
340 if (!IS_ERR(fld->lcf_proc_dir))
341 lprocfs_remove(&fld->lcf_proc_dir);
342 fld->lcf_proc_dir = NULL;
347 static int fld_client_proc_init(struct lu_client_fld *fld)
352 static void fld_client_proc_fini(struct lu_client_fld *fld)
358 static inline int hash_is_sane(int hash)
360 return (hash >= 0 && hash < ARRAY_SIZE(fld_hash));
363 int fld_client_init(struct lu_client_fld *fld,
364 const char *prefix, int hash)
367 int cache_size, cache_threshold;
372 LASSERT(fld != NULL);
374 snprintf(fld->lcf_name, sizeof(fld->lcf_name),
377 if (!hash_is_sane(hash)) {
378 CERROR("%s: Wrong hash function %#x\n",
379 fld->lcf_name, hash);
384 spin_lock_init(&fld->lcf_lock);
385 fld->lcf_hash = &fld_hash[hash];
386 fld->lcf_flags = LUSTRE_FLD_INIT;
387 CFS_INIT_LIST_HEAD(&fld->lcf_targets);
390 cache_size = FLD_CLIENT_CACHE_SIZE /
391 sizeof(struct fld_cache_entry);
393 cache_threshold = cache_size *
394 FLD_CLIENT_CACHE_THRESHOLD / 100;
396 fld->lcf_cache = fld_cache_init(fld->lcf_name,
397 FLD_CLIENT_HTABLE_SIZE,
398 cache_size, cache_threshold);
399 if (IS_ERR(fld->lcf_cache)) {
400 rc = PTR_ERR(fld->lcf_cache);
401 fld->lcf_cache = NULL;
406 rc = fld_client_proc_init(fld);
412 fld_client_fini(fld);
414 CDEBUG(D_INFO, "%s: Using \"%s\" hash\n",
415 fld->lcf_name, fld->lcf_hash->fh_name);
418 EXPORT_SYMBOL(fld_client_init);
420 void fld_client_fini(struct lu_client_fld *fld)
422 struct lu_fld_target *target, *tmp;
425 fld_client_proc_fini(fld);
427 spin_lock(&fld->lcf_lock);
428 list_for_each_entry_safe(target, tmp,
429 &fld->lcf_targets, ft_chain) {
431 list_del(&target->ft_chain);
432 if (target->ft_exp != NULL)
433 class_export_put(target->ft_exp);
434 OBD_FREE_PTR(target);
436 spin_unlock(&fld->lcf_lock);
439 if (fld->lcf_cache != NULL) {
440 if (!IS_ERR(fld->lcf_cache))
441 fld_cache_fini(fld->lcf_cache);
442 fld->lcf_cache = NULL;
448 EXPORT_SYMBOL(fld_client_fini);
450 static int fld_client_rpc(struct obd_export *exp,
451 struct md_fld *mf, __u32 fld_op)
453 struct ptlrpc_request *req;
459 LASSERT(exp != NULL);
461 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp), &RQF_FLD_QUERY,
462 LUSTRE_MDS_VERSION, FLD_QUERY);
466 op = req_capsule_client_get(&req->rq_pill, &RMF_FLD_OPC);
469 pmf = req_capsule_client_get(&req->rq_pill, &RMF_FLD_MDFLD);
472 ptlrpc_request_set_replen(req);
473 req->rq_request_portal = FLD_REQUEST_PORTAL;
474 ptlrpc_at_set_req_timeout(req);
476 if (fld_op != FLD_LOOKUP)
477 mdc_get_rpc_lock(exp->exp_obd->u.cli.cl_rpc_lock, NULL);
478 fld_enter_request(&exp->exp_obd->u.cli);
479 rc = ptlrpc_queue_wait(req);
480 fld_exit_request(&exp->exp_obd->u.cli);
481 if (fld_op != FLD_LOOKUP)
482 mdc_put_rpc_lock(exp->exp_obd->u.cli.cl_rpc_lock, NULL);
486 pmf = req_capsule_server_get(&req->rq_pill, &RMF_FLD_MDFLD);
488 GOTO(out_req, rc = -EFAULT);
492 ptlrpc_req_finished(req);
496 int fld_client_create(struct lu_client_fld *fld,
497 seqno_t seq, mdsno_t mds,
498 const struct lu_env *env)
500 struct md_fld md_fld = { .mf_seq = seq, .mf_mds = mds };
501 struct lu_fld_target *target;
505 fld->lcf_flags |= LUSTRE_FLD_RUN;
506 target = fld_client_get_target(fld, seq);
507 LASSERT(target != NULL);
509 CDEBUG(D_INFO, "%s: Create fld entry (seq: "LPX64"; mds: "
510 LPU64") on target %s (idx "LPU64")\n", fld->lcf_name,
511 seq, mds, fld_target_name(target), target->ft_idx);
514 if (target->ft_srv != NULL) {
515 LASSERT(env != NULL);
516 rc = fld_server_create(target->ft_srv, env, seq, mds);
519 rc = fld_client_rpc(target->ft_exp, &md_fld, FLD_CREATE);
526 * Do not return result of calling fld_cache_insert()
527 * here. First of all because it may return -EEXISTS. Another
528 * reason is that, we do not want to stop proceeding because of
531 fld_cache_insert(fld->lcf_cache, seq, mds);
533 CERROR("%s: Can't create FLD entry, rc %d\n",
539 EXPORT_SYMBOL(fld_client_create);
541 int fld_client_delete(struct lu_client_fld *fld, seqno_t seq,
542 const struct lu_env *env)
544 struct md_fld md_fld = { .mf_seq = seq, .mf_mds = 0 };
545 struct lu_fld_target *target;
549 fld->lcf_flags |= LUSTRE_FLD_RUN;
550 fld_cache_delete(fld->lcf_cache, seq);
552 target = fld_client_get_target(fld, seq);
553 LASSERT(target != NULL);
555 CDEBUG(D_INFO, "%s: Delete fld entry (seq: "LPX64") on "
556 "target %s (idx "LPU64")\n", fld->lcf_name, seq,
557 fld_target_name(target), target->ft_idx);
560 if (target->ft_srv != NULL) {
561 LASSERT(env != NULL);
562 rc = fld_server_delete(target->ft_srv,
566 rc = fld_client_rpc(target->ft_exp,
567 &md_fld, FLD_DELETE);
574 EXPORT_SYMBOL(fld_client_delete);
576 int fld_client_lookup(struct lu_client_fld *fld,
577 seqno_t seq, mdsno_t *mds,
578 const struct lu_env *env)
580 struct md_fld md_fld = { .mf_seq = seq, .mf_mds = 0 };
581 struct lu_fld_target *target;
585 fld->lcf_flags |= LUSTRE_FLD_RUN;
587 rc = fld_cache_lookup(fld->lcf_cache, seq, mds);
591 /* Can not find it in the cache */
592 target = fld_client_get_target(fld, seq);
593 LASSERT(target != NULL);
595 CDEBUG(D_INFO, "%s: Lookup fld entry (seq: "LPX64") on "
596 "target %s (idx "LPU64")\n", fld->lcf_name, seq,
597 fld_target_name(target), target->ft_idx);
600 if (target->ft_srv != NULL) {
601 LASSERT(env != NULL);
602 rc = fld_server_lookup(target->ft_srv,
603 env, seq, &md_fld.mf_mds);
607 * insert the 'inflight' sequence. No need to protect that,
608 * we are trying to reduce numbers of RPC but not restrict
609 * to them exactly one
611 fld_cache_insert_inflight(fld->lcf_cache, seq);
612 rc = fld_client_rpc(target->ft_exp,
613 &md_fld, FLD_LOOKUP);
617 if (seq < FID_SEQ_START) {
619 * The current solution for IGIF is to bind it to mds0.
620 * In the future, this should be fixed once IGIF can be found
628 *mds = md_fld.mf_mds;
631 * Do not return error here as well. See previous comment in
632 * same situation in function fld_client_create().
634 fld_cache_insert(fld->lcf_cache, seq, *mds);
636 /* remove 'inflight' seq if it exists */
637 fld_cache_delete(fld->lcf_cache, seq);
641 EXPORT_SYMBOL(fld_client_lookup);
643 void fld_client_flush(struct lu_client_fld *fld)
646 fld_cache_flush(fld->lcf_cache);
649 EXPORT_SYMBOL(fld_client_flush);