Whamcloud - gitweb
LU-15785 tests: do not detect versions for RPC_MODE mode
[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  *
31  * lustre/fld/fld_request.c
32  *
33  * FLD (Fids Location Database)
34  *
35  * Author: Yury Umanets <umka@clusterfs.com>
36  */
37
38 #define DEBUG_SUBSYSTEM S_FLD
39
40 #include <libcfs/libcfs.h>
41 #include <linux/module.h>
42 #include <linux/math64.h>
43 #include <linux/delay.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 void fld_client_debugfs_init(struct lu_client_fld *fld)
229 {
230         ENTRY;
231         fld->lcf_debugfs_entry = debugfs_create_dir(fld->lcf_name,
232                                                    fld_debugfs_dir);
233         ldebugfs_add_vars(fld->lcf_debugfs_entry,
234                           fld_client_debugfs_list,
235                           fld);
236 }
237
238 void fld_client_debugfs_fini(struct lu_client_fld *fld)
239 {
240         debugfs_remove_recursive(fld->lcf_debugfs_entry);
241 }
242 EXPORT_SYMBOL(fld_client_debugfs_fini);
243
244 static inline int hash_is_sane(int hash)
245 {
246         return (hash >= 0 && hash < ARRAY_SIZE(fld_hash));
247 }
248
249 int fld_client_init(struct lu_client_fld *fld,
250                     const char *prefix, int hash)
251 {
252         int cache_size, cache_threshold;
253         int rc = 0;
254
255         ENTRY;
256         snprintf(fld->lcf_name, sizeof(fld->lcf_name),
257                  "cli-%s", prefix);
258
259         if (!hash_is_sane(hash)) {
260                 CERROR("%s: Wrong hash function %#x\n",
261                        fld->lcf_name, hash);
262                 RETURN(-EINVAL);
263         }
264
265         fld->lcf_count = 0;
266         spin_lock_init(&fld->lcf_lock);
267         fld->lcf_hash = &fld_hash[hash];
268         INIT_LIST_HEAD(&fld->lcf_targets);
269
270         cache_size = FLD_CLIENT_CACHE_SIZE /
271                 sizeof(struct fld_cache_entry);
272
273         cache_threshold = cache_size *
274                 FLD_CLIENT_CACHE_THRESHOLD / 100;
275
276         fld->lcf_cache = fld_cache_init(fld->lcf_name,
277                                         cache_size, cache_threshold);
278         if (IS_ERR(fld->lcf_cache)) {
279                 rc = PTR_ERR(fld->lcf_cache);
280                 fld->lcf_cache = NULL;
281                 GOTO(out, rc);
282         }
283
284         fld_client_debugfs_init(fld);
285         EXIT;
286 out:
287         if (rc)
288                 fld_client_fini(fld);
289         else
290                 CDEBUG(D_INFO, "%s: Using \"%s\" hash\n",
291                        fld->lcf_name, fld->lcf_hash->fh_name);
292         return rc;
293 }
294 EXPORT_SYMBOL(fld_client_init);
295
296 void fld_client_fini(struct lu_client_fld *fld)
297 {
298         struct lu_fld_target *target, *tmp;
299
300         ENTRY;
301
302         spin_lock(&fld->lcf_lock);
303         list_for_each_entry_safe(target, tmp, &fld->lcf_targets, ft_chain) {
304                 fld->lcf_count--;
305                 list_del(&target->ft_chain);
306                 if (target->ft_exp)
307                         class_export_put(target->ft_exp);
308                 OBD_FREE_PTR(target);
309         }
310         spin_unlock(&fld->lcf_lock);
311
312         if (fld->lcf_cache) {
313                 if (!IS_ERR(fld->lcf_cache))
314                         fld_cache_fini(fld->lcf_cache);
315                 fld->lcf_cache = NULL;
316         }
317
318         EXIT;
319 }
320 EXPORT_SYMBOL(fld_client_fini);
321
322 int fld_client_rpc(struct obd_export *exp,
323                    struct lu_seq_range *range, u32 fld_op,
324                    struct ptlrpc_request **reqp)
325 {
326         struct ptlrpc_request *req = NULL;
327         struct lu_seq_range *prange;
328         u32 *op;
329         int rc = 0;
330         struct obd_import *imp;
331
332         ENTRY;
333
334         LASSERT(exp != NULL);
335
336         imp = class_exp2cliimp(exp);
337 again:
338         switch (fld_op) {
339         case FLD_QUERY:
340                 req = ptlrpc_request_alloc_pack(imp, &RQF_FLD_QUERY,
341                                                 LUSTRE_MDS_VERSION, FLD_QUERY);
342                 if (!req)
343                         RETURN(-ENOMEM);
344
345                 /*
346                  * XXX: only needed when talking to old server(< 2.6), it should
347                  * be removed when < 2.6 server is not supported
348                  */
349                 op = req_capsule_client_get(&req->rq_pill, &RMF_FLD_OPC);
350                 *op = FLD_LOOKUP;
351
352                 /*
353                  * For MDS_MDS seq lookup, it will always use LWP connection,
354                  * but LWP will be evicted after restart, so cause the error.
355                  * so we will set no_delay for seq lookup request, once the
356                  * request fails because of the eviction. always retry here
357                  */
358                 if (imp->imp_connect_flags_orig & OBD_CONNECT_MDS_MDS) {
359                         req->rq_allow_replay = 1;
360                         req->rq_no_delay = 1;
361                 }
362                 break;
363         case FLD_READ:
364                 req = ptlrpc_request_alloc_pack(imp, &RQF_FLD_READ,
365                                                 LUSTRE_MDS_VERSION, FLD_READ);
366                 if (!req)
367                         RETURN(-ENOMEM);
368
369                 req_capsule_set_size(&req->rq_pill, &RMF_GENERIC_DATA,
370                                      RCL_SERVER, PAGE_SIZE);
371                 break;
372         default:
373                 rc = -EINVAL;
374                 break;
375         }
376
377         if (rc != 0)
378                 RETURN(rc);
379
380         prange = req_capsule_client_get(&req->rq_pill, &RMF_FLD_MDFLD);
381         *prange = *range;
382         ptlrpc_request_set_replen(req);
383         req->rq_request_portal = FLD_REQUEST_PORTAL;
384         req->rq_reply_portal = MDC_REPLY_PORTAL;
385         ptlrpc_at_set_req_timeout(req);
386
387         if (OBD_FAIL_CHECK(OBD_FAIL_FLD_QUERY_REQ) && req->rq_no_delay) {
388                 /* the same error returned by ptlrpc_import_delay_req */
389                 rc = -EAGAIN;
390                 req->rq_status = rc;
391         } else {
392                 rc = ptlrpc_queue_wait(req);
393         }
394
395         if (rc == -ENOENT) {
396                 /* Don't loop forever on non-existing FID sequences. */
397                 GOTO(out_req, rc);
398         }
399
400         if (rc != 0) {
401                 if (imp->imp_state != LUSTRE_IMP_CLOSED &&
402                     !imp->imp_deactive &&
403                     imp->imp_connect_flags_orig & OBD_CONNECT_MDS_MDS &&
404                     OCD_HAS_FLAG(&imp->imp_connect_data, LIGHTWEIGHT) &&
405                     rc != -ENOTSUPP) {
406                         /* LWP is not replayable, retry after a while */
407                         rc = -EAGAIN;
408                 }
409                 if (rc == -EAGAIN) {
410                         ptlrpc_req_finished(req);
411                         if (msleep_interruptible(2 * MSEC_PER_SEC))
412                                 GOTO(out_req, rc = -EINTR);
413                         rc = 0;
414                         goto again;
415                 }
416                 GOTO(out_req, rc);
417         }
418
419         if (fld_op == FLD_QUERY) {
420                 prange = req_capsule_server_get(&req->rq_pill,
421                                                 &RMF_FLD_MDFLD);
422                 if (!prange)
423                         GOTO(out_req, rc = -EFAULT);
424                 *range = *prange;
425         }
426
427         EXIT;
428 out_req:
429         if (rc != 0 || !reqp) {
430                 ptlrpc_req_finished(req);
431                 req = NULL;
432         }
433
434         if (reqp)
435                 *reqp = req;
436
437         return rc;
438 }
439
440 int fld_client_lookup(struct lu_client_fld *fld, u64 seq, u32 *mds,
441                       u32 flags, const struct lu_env *env)
442 {
443         struct lu_seq_range res = { 0 };
444         struct lu_fld_target *target;
445         struct lu_fld_target *origin;
446         int rc;
447
448         ENTRY;
449
450         rc = fld_cache_lookup(fld->lcf_cache, seq, &res);
451         if (rc == 0) {
452                 *mds = res.lsr_index;
453                 RETURN(0);
454         }
455
456         /* Can not find it in the cache */
457         target = fld_client_get_target(fld, seq);
458         LASSERT(target != NULL);
459         origin = target;
460 again:
461         CDEBUG(D_INFO, "%s: Lookup fld entry (seq: %#llx) on target %s (idx %llu)\n",
462                fld->lcf_name, seq, fld_target_name(target), target->ft_idx);
463
464         res.lsr_start = seq;
465         fld_range_set_type(&res, flags);
466
467 #ifdef HAVE_SERVER_SUPPORT
468         if (target->ft_srv) {
469                 LASSERT(env != NULL);
470                 rc = fld_server_lookup(env, target->ft_srv, seq, &res);
471         } else
472 #endif /* HAVE_SERVER_SUPPORT */
473         {
474                 rc = fld_client_rpc(target->ft_exp, &res, FLD_QUERY, NULL);
475         }
476
477         if (rc == -ESHUTDOWN) {
478                 /*
479                  * If fld lookup failed because the target has been shutdown,
480                  * then try next target in the list, until trying all targets
481                  * or fld lookup succeeds
482                  */
483                 spin_lock(&fld->lcf_lock);
484                 /*
485                  * If the next entry in the list is the head of the list,
486                  * move to the next entry after the head and retrieve
487                  * the target. Else retreive the next target entry.
488                  */
489                 if (target->ft_chain.next == &fld->lcf_targets)
490                         target = list_entry(target->ft_chain.next->next,
491                                             struct lu_fld_target, ft_chain);
492                 else
493                         target = list_entry(target->ft_chain.next,
494                                                  struct lu_fld_target,
495                                                  ft_chain);
496                 spin_unlock(&fld->lcf_lock);
497                 if (target != origin)
498                         goto again;
499         }
500         if (rc == 0) {
501                 *mds = res.lsr_index;
502                 fld_cache_insert(fld->lcf_cache, &res);
503         }
504
505         RETURN(rc);
506 }
507 EXPORT_SYMBOL(fld_client_lookup);
508
509 void fld_client_flush(struct lu_client_fld *fld)
510 {
511         fld_cache_flush(fld->lcf_cache);
512 }
513
514 static int __init fld_init(void)
515 {
516 #ifdef HAVE_SERVER_SUPPORT
517         int rc;
518
519         rc = fld_server_mod_init();
520         if (rc)
521                 return rc;
522 #endif /* HAVE_SERVER_SUPPORT */
523
524         fld_debugfs_dir = debugfs_create_dir(LUSTRE_FLD_NAME,
525                                              debugfs_lustre_root);
526         return PTR_ERR_OR_ZERO(fld_debugfs_dir);
527 }
528
529 static void __exit fld_exit(void)
530 {
531 #ifdef HAVE_SERVER_SUPPORT
532         fld_server_mod_exit();
533 #endif /* HAVE_SERVER_SUPPORT */
534
535         debugfs_remove_recursive(fld_debugfs_dir);
536 }
537
538 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
539 MODULE_DESCRIPTION("Lustre FID Location Database");
540 MODULE_VERSION(LUSTRE_VERSION_STRING);
541 MODULE_LICENSE("GPL");
542
543 module_init(fld_init);
544 module_exit(fld_exit);