Whamcloud - gitweb
LU-12616 obclass: fix MDS start/stop race
[fs/lustre-release.git] / 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.gnu.org/licenses/gpl-2.0.html
19  *
20  * GPL HEADER END
21  */
22 /*
23  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
24  * Use is subject to license terms.
25  *
26  * Copyright (c) 2011, 2017, Intel Corporation.
27  */
28 /*
29  * This file is part of Lustre, http://www.lustre.org/
30  * Lustre is a trademark of Sun Microsystems, Inc.
31  *
32  * lustre/fld/fld_request.c
33  *
34  * FLD (Fids Location Database)
35  *
36  * Author: Yury Umanets <umka@clusterfs.com>
37  */
38
39 #define DEBUG_SUBSYSTEM S_FLD
40
41 #include <libcfs/libcfs.h>
42 #include <linux/module.h>
43 #include <linux/math64.h>
44
45 #include <obd.h>
46 #include <obd_class.h>
47 #include <obd_support.h>
48 #include <lprocfs_status.h>
49 #include <lustre_req_layout.h>
50 #include <lustre_fld.h>
51 #include <lustre_mdc.h>
52 #include "fld_internal.h"
53
54 static int fld_rrb_hash(struct lu_client_fld *fld, u64 seq)
55 {
56         LASSERT(fld->lcf_count > 0);
57         return do_div(seq, fld->lcf_count);
58 }
59
60 static struct lu_fld_target *
61 fld_rrb_scan(struct lu_client_fld *fld, u64 seq)
62 {
63         struct lu_fld_target *target;
64         int hash;
65
66         ENTRY;
67
68         /*
69          * Because almost all of special sequence located in MDT0,
70          * it should go to index 0 directly, instead of calculating
71          * hash again, and also if other MDTs is not being connected,
72          * the fld lookup requests(for seq on MDT0) should not be
73          * blocked because of other MDTs
74          */
75         if (fid_seq_is_norm(seq))
76                 hash = fld_rrb_hash(fld, seq);
77         else
78                 hash = 0;
79
80 again:
81         list_for_each_entry(target, &fld->lcf_targets, ft_chain) {
82                 if (target->ft_idx == hash)
83                         RETURN(target);
84         }
85
86         if (hash != 0) {
87                 /*
88                  * It is possible the remote target(MDT) are not connected to
89                  * with client yet, so we will refer this to MDT0, which should
90                  * be connected during mount
91                  */
92                 hash = 0;
93                 goto again;
94         }
95
96         CERROR("%s: Can't find target by hash %d (seq %#llx). Targets (%d):\n",
97                fld->lcf_name, hash, seq, fld->lcf_count);
98
99         list_for_each_entry(target, &fld->lcf_targets, ft_chain) {
100                 const char *srv_name = target->ft_srv != NULL  ?
101                         target->ft_srv->lsf_name : "<null>";
102                 const char *exp_name = target->ft_exp != NULL ?
103                         (char *)target->ft_exp->exp_obd->obd_uuid.uuid :
104                         "<null>";
105
106                 CERROR("  exp: 0x%p (%s), srv: 0x%p (%s), idx: %llu\n",
107                        target->ft_exp, exp_name, target->ft_srv,
108                        srv_name, target->ft_idx);
109         }
110
111         /*
112          * If target is not found, there is logical error anyway, so here is
113          * LBUG() to catch this situation.
114          */
115         LBUG();
116         RETURN(NULL);
117 }
118
119 struct lu_fld_hash fld_hash[] = {
120         {
121                 .fh_name = "RRB",
122                 .fh_hash_func = fld_rrb_hash,
123                 .fh_scan_func = fld_rrb_scan
124         },
125         {
126                 NULL,
127         }
128 };
129
130 static struct lu_fld_target *
131 fld_client_get_target(struct lu_client_fld *fld, u64 seq)
132 {
133         struct lu_fld_target *target;
134
135         ENTRY;
136
137         LASSERT(fld->lcf_hash != NULL);
138
139         spin_lock(&fld->lcf_lock);
140         target = fld->lcf_hash->fh_scan_func(fld, seq);
141         spin_unlock(&fld->lcf_lock);
142
143         if (target) {
144                 CDEBUG(D_INFO, "%s: Found target (idx %llu) by seq %#llx\n",
145                        fld->lcf_name, target->ft_idx, seq);
146         }
147
148         RETURN(target);
149 }
150
151 /*
152  * Add export to FLD. This is usually done by CMM and LMV as they are main users
153  * of FLD module.
154  */
155 int fld_client_add_target(struct lu_client_fld *fld,
156                           struct lu_fld_target *tar)
157 {
158         const char *name;
159         struct lu_fld_target *target, *tmp;
160
161         ENTRY;
162
163         LASSERT(tar != NULL);
164         name = fld_target_name(tar);
165         LASSERT(name != NULL);
166         LASSERT(tar->ft_srv != NULL || tar->ft_exp != NULL);
167
168         CDEBUG(D_INFO, "%s: Adding target %s (idx %llu)\n", fld->lcf_name,
169                name, tar->ft_idx);
170
171         OBD_ALLOC_PTR(target);
172         if (!target)
173                 RETURN(-ENOMEM);
174
175         spin_lock(&fld->lcf_lock);
176         list_for_each_entry(tmp, &fld->lcf_targets, ft_chain) {
177                 if (tmp->ft_idx == tar->ft_idx) {
178                         spin_unlock(&fld->lcf_lock);
179                         OBD_FREE_PTR(target);
180                         CERROR("Target %s exists in FLD and known as %s:#%llu\n",
181                                name, fld_target_name(tmp), tmp->ft_idx);
182                         RETURN(-EEXIST);
183                 }
184         }
185
186         target->ft_exp = tar->ft_exp;
187         if (target->ft_exp)
188                 class_export_get(target->ft_exp);
189         target->ft_srv = tar->ft_srv;
190         target->ft_idx = tar->ft_idx;
191
192         list_add_tail(&target->ft_chain, &fld->lcf_targets);
193
194         fld->lcf_count++;
195         spin_unlock(&fld->lcf_lock);
196
197         RETURN(0);
198 }
199 EXPORT_SYMBOL(fld_client_add_target);
200
201 /* Remove export from FLD */
202 int fld_client_del_target(struct lu_client_fld *fld, u64 idx)
203 {
204         struct lu_fld_target *target, *tmp;
205
206         ENTRY;
207
208         spin_lock(&fld->lcf_lock);
209         list_for_each_entry_safe(target, tmp, &fld->lcf_targets, ft_chain) {
210                 if (target->ft_idx == idx) {
211                         fld->lcf_count--;
212                         list_del(&target->ft_chain);
213                         spin_unlock(&fld->lcf_lock);
214
215                         if (target->ft_exp)
216                                 class_export_put(target->ft_exp);
217
218                         OBD_FREE_PTR(target);
219                         RETURN(0);
220                 }
221         }
222         spin_unlock(&fld->lcf_lock);
223         RETURN(-ENOENT);
224 }
225
226 struct dentry *fld_debugfs_dir;
227
228 static int fld_client_debugfs_init(struct lu_client_fld *fld)
229 {
230         int rc;
231
232         ENTRY;
233         fld->lcf_debugfs_entry = ldebugfs_register(fld->lcf_name,
234                                                    fld_debugfs_dir,
235                                                    fld_client_debugfs_list,
236                                                    fld);
237         if (IS_ERR_OR_NULL(fld->lcf_debugfs_entry)) {
238                 CERROR("%s: LdebugFS failed in fld-init\n", fld->lcf_name);
239                 rc = fld->lcf_debugfs_entry ? PTR_ERR(fld->lcf_debugfs_entry)
240                                             : -ENOMEM;
241                 fld->lcf_debugfs_entry = NULL;
242                 RETURN(rc);
243         }
244
245         return 0;
246 }
247
248 void fld_client_debugfs_fini(struct lu_client_fld *fld)
249 {
250         if (!IS_ERR_OR_NULL(fld->lcf_debugfs_entry))
251                 ldebugfs_remove(&fld->lcf_debugfs_entry);
252 }
253 EXPORT_SYMBOL(fld_client_debugfs_fini);
254
255 static inline int hash_is_sane(int hash)
256 {
257         return (hash >= 0 && hash < ARRAY_SIZE(fld_hash));
258 }
259
260 int fld_client_init(struct lu_client_fld *fld,
261                     const char *prefix, int hash)
262 {
263         int cache_size, cache_threshold;
264         int rc;
265
266         ENTRY;
267         snprintf(fld->lcf_name, sizeof(fld->lcf_name),
268                  "cli-%s", prefix);
269
270         if (!hash_is_sane(hash)) {
271                 CERROR("%s: Wrong hash function %#x\n",
272                        fld->lcf_name, hash);
273                 RETURN(-EINVAL);
274         }
275
276         fld->lcf_count = 0;
277         spin_lock_init(&fld->lcf_lock);
278         fld->lcf_hash = &fld_hash[hash];
279         INIT_LIST_HEAD(&fld->lcf_targets);
280
281         cache_size = FLD_CLIENT_CACHE_SIZE /
282                 sizeof(struct fld_cache_entry);
283
284         cache_threshold = cache_size *
285                 FLD_CLIENT_CACHE_THRESHOLD / 100;
286
287         fld->lcf_cache = fld_cache_init(fld->lcf_name,
288                                         cache_size, cache_threshold);
289         if (IS_ERR(fld->lcf_cache)) {
290                 rc = PTR_ERR(fld->lcf_cache);
291                 fld->lcf_cache = NULL;
292                 GOTO(out, rc);
293         }
294
295         rc = fld_client_debugfs_init(fld);
296         if (rc)
297                 GOTO(out, rc);
298         EXIT;
299 out:
300         if (rc)
301                 fld_client_fini(fld);
302         else
303                 CDEBUG(D_INFO, "%s: Using \"%s\" hash\n",
304                        fld->lcf_name, fld->lcf_hash->fh_name);
305         return rc;
306 }
307 EXPORT_SYMBOL(fld_client_init);
308
309 void fld_client_fini(struct lu_client_fld *fld)
310 {
311         struct lu_fld_target *target, *tmp;
312
313         ENTRY;
314
315         spin_lock(&fld->lcf_lock);
316         list_for_each_entry_safe(target, tmp, &fld->lcf_targets, ft_chain) {
317                 fld->lcf_count--;
318                 list_del(&target->ft_chain);
319                 if (target->ft_exp)
320                         class_export_put(target->ft_exp);
321                 OBD_FREE_PTR(target);
322         }
323         spin_unlock(&fld->lcf_lock);
324
325         if (fld->lcf_cache) {
326                 if (!IS_ERR(fld->lcf_cache))
327                         fld_cache_fini(fld->lcf_cache);
328                 fld->lcf_cache = NULL;
329         }
330
331         EXIT;
332 }
333 EXPORT_SYMBOL(fld_client_fini);
334
335 int fld_client_rpc(struct obd_export *exp,
336                    struct lu_seq_range *range, u32 fld_op,
337                    struct ptlrpc_request **reqp)
338 {
339         struct ptlrpc_request *req = NULL;
340         struct lu_seq_range *prange;
341         u32 *op;
342         int rc = 0;
343         struct obd_import *imp;
344
345         ENTRY;
346
347         LASSERT(exp != NULL);
348
349         imp = class_exp2cliimp(exp);
350         switch (fld_op) {
351         case FLD_QUERY:
352                 req = ptlrpc_request_alloc_pack(imp, &RQF_FLD_QUERY,
353                                                 LUSTRE_MDS_VERSION, FLD_QUERY);
354                 if (!req)
355                         RETURN(-ENOMEM);
356
357                 /*
358                  * XXX: only needed when talking to old server(< 2.6), it should
359                  * be removed when < 2.6 server is not supported
360                  */
361                 op = req_capsule_client_get(&req->rq_pill, &RMF_FLD_OPC);
362                 *op = FLD_LOOKUP;
363
364                 /*
365                  * For MDS_MDS seq lookup, it will always use LWP connection,
366                  * but LWP will be evicted after restart, so cause the error.
367                  * so we will set no_delay for seq lookup request, once the
368                  * request fails because of the eviction. always retry here
369                  */
370                 if (imp->imp_connect_flags_orig & OBD_CONNECT_MDS_MDS) {
371                         req->rq_allow_replay = 1;
372                         req->rq_no_delay = 1;
373                 }
374                 break;
375         case FLD_READ:
376                 req = ptlrpc_request_alloc_pack(imp, &RQF_FLD_READ,
377                                                 LUSTRE_MDS_VERSION, FLD_READ);
378                 if (!req)
379                         RETURN(-ENOMEM);
380
381                 req_capsule_set_size(&req->rq_pill, &RMF_GENERIC_DATA,
382                                      RCL_SERVER, PAGE_SIZE);
383                 break;
384         default:
385                 rc = -EINVAL;
386                 break;
387         }
388
389         if (rc != 0)
390                 RETURN(rc);
391
392         prange = req_capsule_client_get(&req->rq_pill, &RMF_FLD_MDFLD);
393         *prange = *range;
394         ptlrpc_request_set_replen(req);
395         req->rq_request_portal = FLD_REQUEST_PORTAL;
396         req->rq_reply_portal = MDC_REPLY_PORTAL;
397         ptlrpc_at_set_req_timeout(req);
398
399         if (OBD_FAIL_CHECK(OBD_FAIL_FLD_QUERY_REQ && req->rq_no_delay)) {
400                 /* the same error returned by ptlrpc_import_delay_req */
401                 rc = -EWOULDBLOCK;
402                 req->rq_status = rc;
403         } else {
404                 obd_get_request_slot(&exp->exp_obd->u.cli);
405                 rc = ptlrpc_queue_wait(req);
406                 obd_put_request_slot(&exp->exp_obd->u.cli);
407         }
408
409         if (rc == -ENOENT) {
410                 /* Don't loop forever on non-existing FID sequences. */
411                 GOTO(out_req, rc);
412         }
413
414         if (rc != 0) {
415                 if (imp->imp_state != LUSTRE_IMP_CLOSED &&
416                     !imp->imp_deactive &&
417                     imp->imp_connect_flags_orig & OBD_CONNECT_MDS_MDS &&
418                     OCD_HAS_FLAG(&imp->imp_connect_data, LIGHTWEIGHT) &&
419                     rc != -ENOTSUPP) {
420                         /*
421                          * Since LWP is not replayable, so notify the caller
422                          * to retry if needed after a while.
423                          */
424                         rc = -EAGAIN;
425                 }
426                 GOTO(out_req, rc);
427         }
428
429         if (fld_op == FLD_QUERY) {
430                 prange = req_capsule_server_get(&req->rq_pill,
431                                                 &RMF_FLD_MDFLD);
432                 if (!prange)
433                         GOTO(out_req, rc = -EFAULT);
434                 *range = *prange;
435         }
436
437         EXIT;
438 out_req:
439         if (rc != 0 || !reqp) {
440                 ptlrpc_req_finished(req);
441                 req = NULL;
442         }
443
444         if (reqp)
445                 *reqp = req;
446
447         return rc;
448 }
449
450 int fld_client_lookup(struct lu_client_fld *fld, u64 seq, u32 *mds,
451                       u32 flags, const struct lu_env *env)
452 {
453         struct lu_seq_range res = { 0 };
454         struct lu_fld_target *target;
455         struct lu_fld_target *origin;
456         int rc;
457
458         ENTRY;
459
460         rc = fld_cache_lookup(fld->lcf_cache, seq, &res);
461         if (rc == 0) {
462                 *mds = res.lsr_index;
463                 RETURN(0);
464         }
465
466         /* Can not find it in the cache */
467         target = fld_client_get_target(fld, seq);
468         LASSERT(target != NULL);
469         origin = target;
470 again:
471         CDEBUG(D_INFO, "%s: Lookup fld entry (seq: %#llx) on target %s (idx %llu)\n",
472                fld->lcf_name, seq, fld_target_name(target), target->ft_idx);
473
474         res.lsr_start = seq;
475         fld_range_set_type(&res, flags);
476
477 #ifdef HAVE_SERVER_SUPPORT
478         if (target->ft_srv) {
479                 LASSERT(env != NULL);
480                 rc = fld_server_lookup(env, target->ft_srv, seq, &res);
481         } else
482 #endif /* HAVE_SERVER_SUPPORT */
483         {
484                 rc = fld_client_rpc(target->ft_exp, &res, FLD_QUERY, NULL);
485         }
486
487         if (rc == -ESHUTDOWN) {
488                 /*
489                  * If fld lookup failed because the target has been shutdown,
490                  * then try next target in the list, until trying all targets
491                  * or fld lookup succeeds
492                  */
493                 spin_lock(&fld->lcf_lock);
494                 /*
495                  * If the next entry in the list is the head of the list,
496                  * move to the next entry after the head and retrieve
497                  * the target. Else retreive the next target entry.
498                  */
499                 if (target->ft_chain.next == &fld->lcf_targets)
500                         target = list_entry(target->ft_chain.next->next,
501                                             struct lu_fld_target, ft_chain);
502                 else
503                         target = list_entry(target->ft_chain.next,
504                                                  struct lu_fld_target,
505                                                  ft_chain);
506                 spin_unlock(&fld->lcf_lock);
507                 if (target != origin)
508                         goto again;
509         }
510         if (rc == 0) {
511                 *mds = res.lsr_index;
512                 fld_cache_insert(fld->lcf_cache, &res);
513         }
514
515         RETURN(rc);
516 }
517 EXPORT_SYMBOL(fld_client_lookup);
518
519 void fld_client_flush(struct lu_client_fld *fld)
520 {
521         fld_cache_flush(fld->lcf_cache);
522 }
523
524 static int __init fld_init(void)
525 {
526 #ifdef HAVE_SERVER_SUPPORT
527         int rc;
528
529         rc = fld_server_mod_init();
530         if (rc)
531                 return rc;
532 #endif /* HAVE_SERVER_SUPPORT */
533
534         fld_debugfs_dir = ldebugfs_register(LUSTRE_FLD_NAME,
535                                             debugfs_lustre_root,
536                                             NULL, NULL);
537         return PTR_ERR_OR_ZERO(fld_debugfs_dir);
538 }
539
540 static void __exit fld_exit(void)
541 {
542 #ifdef HAVE_SERVER_SUPPORT
543         fld_server_mod_exit();
544 #endif /* HAVE_SERVER_SUPPORT */
545
546         if (!IS_ERR_OR_NULL(fld_debugfs_dir))
547                 ldebugfs_remove(&fld_debugfs_dir);
548 }
549
550 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
551 MODULE_DESCRIPTION("Lustre FID Location Database");
552 MODULE_VERSION(LUSTRE_VERSION_STRING);
553 MODULE_LICENSE("GPL");
554
555 module_init(fld_init);
556 module_exit(fld_exit);