Whamcloud - gitweb
timeout and assert in invalidate import.
[fs/lustre-release.git] / lustre / ptlrpc / ptlrpcd.c
1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2  * vim:expandtab:shiftwidth=8:tabstop=8:
3  *
4  *  Copyright (C) 2001-2003 Cluster File Systems, Inc.
5  *   Author Peter Braam <braam@clusterfs.com>
6  *
7  *   This file is part of the Lustre file system, http://www.lustre.org
8  *   Lustre is a trademark of Cluster File Systems, Inc.
9  *
10  *   You may have signed or agreed to another license before downloading
11  *   this software.  If so, you are bound by the terms and conditions
12  *   of that agreement, and the following does not apply to you.  See the
13  *   LICENSE file included with this distribution for more information.
14  *
15  *   If you did not agree to a different license, then this copy of Lustre
16  *   is open source software; you can redistribute it and/or modify it
17  *   under the terms of version 2 of the GNU General Public License as
18  *   published by the Free Software Foundation.
19  *
20  *   In either case, Lustre is distributed in the hope that it will be
21  *   useful, but WITHOUT ANY WARRANTY; without even the implied warranty
22  *   of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
23  *   license text for more details.
24  *
25  */
26
27 #define DEBUG_SUBSYSTEM S_RPC
28
29 #ifdef __KERNEL__
30 # include <libcfs/libcfs.h>
31 #else /* __KERNEL__ */
32 # include <liblustre.h>
33 # include <ctype.h>
34 #endif
35
36 #include <libcfs/kp30.h>
37 #include <lustre_net.h>
38 # include <lustre_lib.h>
39
40 #include <lustre_ha.h>
41 #include <obd_class.h>   /* for obd_zombie */
42 #include <obd_support.h> /* for OBD_FAIL_CHECK */
43 #include <lprocfs_status.h>
44
45 #define LIOD_STOP 0
46 struct ptlrpcd_ctl {
47         unsigned long             pc_flags;
48         spinlock_t                pc_lock;
49         struct completion         pc_starting;
50         struct completion         pc_finishing;
51         struct ptlrpc_request_set *pc_set;
52         char                      pc_name[16];
53 #ifndef __KERNEL__
54         int                       pc_recurred;
55         void                     *pc_callback;
56         void                     *pc_wait_callback;
57         void                     *pc_idle_callback;
58 #endif
59 };
60
61 static struct ptlrpcd_ctl ptlrpcd_pc;
62 static struct ptlrpcd_ctl ptlrpcd_recovery_pc;
63
64 struct semaphore ptlrpcd_sem;
65 static int ptlrpcd_users = 0;
66
67 void ptlrpcd_wake(struct ptlrpc_request *req)
68 {
69         struct ptlrpc_request_set *rq_set = req->rq_set;
70
71         LASSERT(rq_set != NULL);
72
73         cfs_waitq_signal(&rq_set->set_waitq);
74 }
75
76 /* requests that are added to the ptlrpcd queue are sent via
77  * ptlrpcd_check->ptlrpc_check_set() */
78 void ptlrpcd_add_req(struct ptlrpc_request *req)
79 {
80         struct ptlrpcd_ctl *pc;
81
82         if (req->rq_send_state == LUSTRE_IMP_FULL)
83                 pc = &ptlrpcd_pc;
84         else
85                 pc = &ptlrpcd_recovery_pc;
86
87         ptlrpc_set_add_new_req(pc->pc_set, req);
88         cfs_waitq_signal(&pc->pc_set->set_waitq);
89 }
90
91 static int ptlrpcd_check(struct ptlrpcd_ctl *pc)
92 {
93         struct list_head *tmp, *pos;
94         struct ptlrpc_request *req;
95         int rc = 0;
96         ENTRY;
97
98         if (test_bit(LIOD_STOP, &pc->pc_flags))
99                 RETURN(1);
100
101         spin_lock(&pc->pc_set->set_new_req_lock);
102         list_for_each_safe(pos, tmp, &pc->pc_set->set_new_requests) {
103                 req = list_entry(pos, struct ptlrpc_request, rq_set_chain);
104                 list_del_init(&req->rq_set_chain);
105                 ptlrpc_set_add_req(pc->pc_set, req);
106                 rc = 1; /* need to calculate its timeout */
107         }
108         spin_unlock(&pc->pc_set->set_new_req_lock);
109
110         if (pc->pc_set->set_remaining) {
111                 rc = rc | ptlrpc_check_set(pc->pc_set);
112
113                 /* XXX our set never completes, so we prune the completed
114                  * reqs after each iteration. boy could this be smarter. */
115                 list_for_each_safe(pos, tmp, &pc->pc_set->set_requests) {
116                         req = list_entry(pos, struct ptlrpc_request,
117                                          rq_set_chain);
118                         if (req->rq_phase != RQ_PHASE_COMPLETE)
119                                 continue;
120
121                         list_del_init(&req->rq_set_chain);
122                         req->rq_set = NULL;
123                         ptlrpc_req_finished (req);
124                 }
125         }
126
127         if (rc == 0) {
128                 /* If new requests have been added, make sure to wake up */
129                 spin_lock(&pc->pc_set->set_new_req_lock);
130                 rc = !list_empty(&pc->pc_set->set_new_requests);
131                 spin_unlock(&pc->pc_set->set_new_req_lock);
132         }
133
134         RETURN(rc);
135 }
136
137 #ifdef __KERNEL__
138 /* ptlrpc's code paths like to execute in process context, so we have this
139  * thread which spins on a set which contains the io rpcs.  llite specifies
140  * ptlrpcd's set when it pushes pages down into the oscs */
141 static int ptlrpcd(void *arg)
142 {
143         struct ptlrpcd_ctl *pc = arg;
144         int rc;
145         ENTRY;
146
147         if ((rc = cfs_daemonize_ctxt(pc->pc_name))) {
148                 complete(&pc->pc_starting);
149                 return rc;
150         }
151
152         complete(&pc->pc_starting);
153
154         /* this mainloop strongly resembles ptlrpc_set_wait except
155          * that our set never completes.  ptlrpcd_check calls ptlrpc_check_set
156          * when there are requests in the set.  new requests come in
157          * on the set's new_req_list and ptlrpcd_check moves them into
158          * the set. */
159         while (1) {
160                 struct l_wait_info lwi;
161                 cfs_duration_t timeout;
162
163                 timeout = cfs_time_seconds(ptlrpc_set_next_timeout(pc->pc_set));
164                 lwi = LWI_TIMEOUT(timeout, ptlrpc_expired_set, pc->pc_set);
165
166                 l_wait_event(pc->pc_set->set_waitq, ptlrpcd_check(pc), &lwi);
167
168                 if (test_bit(LIOD_STOP, &pc->pc_flags))
169                         break;
170         }
171         /* wait for inflight requests to drain */
172         if (!list_empty(&pc->pc_set->set_requests))
173                 ptlrpc_set_wait(pc->pc_set);
174         complete(&pc->pc_finishing);
175         return 0;
176 }
177
178 #else
179
180 int ptlrpcd_check_async_rpcs(void *arg)
181 {
182         struct ptlrpcd_ctl *pc = arg;
183         int                  rc = 0;
184
185         /* single threaded!! */
186         pc->pc_recurred++;
187
188         if (pc->pc_recurred == 1) {
189                 rc = ptlrpcd_check(pc);
190                 if (!rc)
191                         ptlrpc_expired_set(pc->pc_set);
192                 /*XXX send replay requests */
193                 if (pc == &ptlrpcd_recovery_pc)
194                         rc = ptlrpcd_check(pc);
195         }
196
197         pc->pc_recurred--;
198         return rc;
199 }
200
201 int ptlrpcd_idle(void *arg)
202 {
203         struct ptlrpcd_ctl *pc = arg;
204
205         return (list_empty(&pc->pc_set->set_new_requests) &&
206                 pc->pc_set->set_remaining == 0);
207 }
208
209 #endif
210
211 static int ptlrpcd_start(char *name, struct ptlrpcd_ctl *pc)
212 {
213         int rc;
214
215         ENTRY;
216         memset(pc, 0, sizeof(*pc));
217         init_completion(&pc->pc_starting);
218         init_completion(&pc->pc_finishing);
219         pc->pc_flags = 0;
220         spin_lock_init(&pc->pc_lock);
221         snprintf (pc->pc_name, sizeof (pc->pc_name), name);
222
223         pc->pc_set = ptlrpc_prep_set();
224         if (pc->pc_set == NULL)
225                 RETURN(-ENOMEM);
226
227 #ifdef __KERNEL__
228         rc = cfs_kernel_thread(ptlrpcd, pc, 0);
229         if (rc < 0)  {
230                 ptlrpc_set_destroy(pc->pc_set);
231                 RETURN(rc);
232         }
233
234         wait_for_completion(&pc->pc_starting);
235 #else
236         pc->pc_wait_callback =
237                 liblustre_register_wait_callback("ptlrpcd_check_async_rpcs",
238                                                  &ptlrpcd_check_async_rpcs, pc);
239         pc->pc_idle_callback =
240                 liblustre_register_idle_callback("ptlrpcd_check_idle_rpcs",
241                                                  &ptlrpcd_idle, pc);
242         (void)rc;
243 #endif
244         RETURN(0);
245 }
246
247 static void ptlrpcd_stop(struct ptlrpcd_ctl *pc)
248 {
249         set_bit(LIOD_STOP, &pc->pc_flags);
250         cfs_waitq_signal(&pc->pc_set->set_waitq);
251 #ifdef __KERNEL__
252         wait_for_completion(&pc->pc_finishing);
253 #else
254         liblustre_deregister_wait_callback(pc->pc_wait_callback);
255         liblustre_deregister_idle_callback(pc->pc_idle_callback);
256 #endif
257         ptlrpc_set_destroy(pc->pc_set);
258 }
259
260 int ptlrpcd_addref(void)
261 {
262         int rc = 0;
263         ENTRY;
264
265         mutex_down(&ptlrpcd_sem);
266         if (++ptlrpcd_users != 1)
267                 GOTO(out, rc);
268
269         rc = ptlrpcd_start("ptlrpcd", &ptlrpcd_pc);
270         if (rc) {
271                 --ptlrpcd_users;
272                 GOTO(out, rc);
273         }
274
275         rc = ptlrpcd_start("ptlrpcd-recov", &ptlrpcd_recovery_pc);
276         if (rc) {
277                 ptlrpcd_stop(&ptlrpcd_pc);
278                 --ptlrpcd_users;
279                 GOTO(out, rc);
280         }
281 out:
282         mutex_up(&ptlrpcd_sem);
283         RETURN(rc);
284 }
285
286 void ptlrpcd_decref(void)
287 {
288         mutex_down(&ptlrpcd_sem);
289         if (--ptlrpcd_users == 0) {
290                 ptlrpcd_stop(&ptlrpcd_pc);
291                 ptlrpcd_stop(&ptlrpcd_recovery_pc);
292         }
293         mutex_up(&ptlrpcd_sem);
294 }