1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * Copyright (C) 2001-2003 Cluster File Systems, Inc.
5 * Author Peter Braam <braam@clusterfs.com>
7 * This file is part of the Lustre file system, http://www.lustre.org
8 * Lustre is a trademark of Cluster File Systems, Inc.
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.
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.
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.
27 #define DEBUG_SUBSYSTEM S_RPC
30 # include <libcfs/libcfs.h>
31 #else /* __KERNEL__ */
32 # include <liblustre.h>
36 #include <libcfs/kp30.h>
37 #include <lustre_net.h>
38 # include <lustre_lib.h>
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>
47 unsigned long pc_flags;
49 struct completion pc_starting;
50 struct completion pc_finishing;
51 struct list_head pc_req_list;
53 struct ptlrpc_request_set *pc_set;
57 void *pc_wait_callback;
58 void *pc_idle_callback;
62 static struct ptlrpcd_ctl ptlrpcd_pc;
63 static struct ptlrpcd_ctl ptlrpcd_recovery_pc;
65 struct semaphore ptlrpcd_sem;
66 static int ptlrpcd_users = 0;
68 void ptlrpcd_wake(struct ptlrpc_request *req)
70 struct ptlrpcd_ctl *pc = req->rq_ptlrpcd_data;
74 cfs_waitq_signal(&pc->pc_waitq);
77 /* requests that are added to the ptlrpcd queue are sent via
78 * ptlrpcd_check->ptlrpc_check_set() */
79 void ptlrpcd_add_req(struct ptlrpc_request *req)
81 struct ptlrpcd_ctl *pc;
83 if (req->rq_send_state == LUSTRE_IMP_FULL)
86 pc = &ptlrpcd_recovery_pc;
88 req->rq_ptlrpcd_data = pc;
89 ptlrpc_set_add_new_req(pc->pc_set, req);
90 wake_up(&pc->pc_waitq);
93 static int ptlrpcd_check(struct ptlrpcd_ctl *pc)
95 struct list_head *tmp, *pos;
96 struct ptlrpc_request *req;
100 if (test_bit(LIOD_STOP, &pc->pc_flags))
103 obd_zombie_impexp_cull();
105 spin_lock(&pc->pc_set->set_new_req_lock);
106 list_for_each_safe(pos, tmp, &pc->pc_set->set_new_requests) {
107 req = list_entry(pos, struct ptlrpc_request, rq_set_chain);
108 list_del_init(&req->rq_set_chain);
109 ptlrpc_set_add_req(pc->pc_set, req);
110 rc = 1; /* need to calculate its timeout */
112 spin_unlock(&pc->pc_set->set_new_req_lock);
114 if (pc->pc_set->set_remaining) {
115 rc = rc | ptlrpc_check_set(pc->pc_set);
117 /* XXX our set never completes, so we prune the completed
118 * reqs after each iteration. boy could this be smarter. */
119 list_for_each_safe(pos, tmp, &pc->pc_set->set_requests) {
120 req = list_entry(pos, struct ptlrpc_request,
122 if (req->rq_phase != RQ_PHASE_COMPLETE)
125 list_del_init(&req->rq_set_chain);
127 ptlrpc_req_finished (req);
132 /* If new requests have been added, make sure to wake up */
133 spin_lock(&pc->pc_set->set_new_req_lock);
134 rc = !list_empty(&pc->pc_set->set_new_requests);
135 spin_unlock(&pc->pc_set->set_new_req_lock);
142 /* ptlrpc's code paths like to execute in process context, so we have this
143 * thread which spins on a set which contains the io rpcs. llite specifies
144 * ptlrpcd's set when it pushes pages down into the oscs */
145 static int ptlrpcd(void *arg)
147 struct ptlrpcd_ctl *pc = arg;
151 if ((rc = cfs_daemonize_ctxt(pc->pc_name))) {
152 complete(&pc->pc_starting);
156 complete(&pc->pc_starting);
158 /* this mainloop strongly resembles ptlrpc_set_wait except
159 * that our set never completes. ptlrpcd_check calls ptlrpc_check_set
160 * when there are requests in the set. new requests come in
161 * on the set's new_req_list and ptlrpcd_check moves them into
164 cfs_waitlink_t set_wait;
165 struct l_wait_info lwi;
166 cfs_duration_t timeout;
168 timeout = cfs_time_seconds(ptlrpc_set_next_timeout(pc->pc_set));
169 lwi = LWI_TIMEOUT(timeout, ptlrpc_expired_set, pc->pc_set);
171 /* ala the pinger, wait on pc's waitqueue and the set's */
172 cfs_waitlink_init(&set_wait);
173 cfs_waitq_add(&pc->pc_set->set_waitq, &set_wait);
174 cfs_waitq_forward(&set_wait, &pc->pc_waitq);
175 l_wait_event(pc->pc_waitq, ptlrpcd_check(pc), &lwi);
176 cfs_waitq_del(&pc->pc_set->set_waitq, &set_wait);
178 if (test_bit(LIOD_STOP, &pc->pc_flags))
181 /* wait for inflight requests to drain */
182 if (!list_empty(&pc->pc_set->set_requests))
183 ptlrpc_set_wait(pc->pc_set);
184 complete(&pc->pc_finishing);
188 static void ptlrpcd_zombie_impexp_notify(void)
190 cfs_waitq_signal(&ptlrpcd_pc.pc_waitq);
194 int ptlrpcd_check_async_rpcs(void *arg)
196 struct ptlrpcd_ctl *pc = arg;
199 /* single threaded!! */
202 if (pc->pc_recurred == 1) {
203 rc = ptlrpcd_check(pc);
205 ptlrpc_expired_set(pc->pc_set);
206 /*XXX send replay requests */
207 if (pc == &ptlrpcd_recovery_pc)
208 rc = ptlrpcd_check(pc);
215 int ptlrpcd_idle(void *arg)
217 struct ptlrpcd_ctl *pc = arg;
219 return (list_empty(&pc->pc_set->set_new_requests) &&
220 pc->pc_set->set_remaining == 0);
225 static int ptlrpcd_start(char *name, struct ptlrpcd_ctl *pc)
230 memset(pc, 0, sizeof(*pc));
231 init_completion(&pc->pc_starting);
232 init_completion(&pc->pc_finishing);
233 cfs_waitq_init(&pc->pc_waitq);
235 spin_lock_init(&pc->pc_lock);
236 CFS_INIT_LIST_HEAD(&pc->pc_req_list);
237 snprintf (pc->pc_name, sizeof (pc->pc_name), name);
239 pc->pc_set = ptlrpc_prep_set();
240 if (pc->pc_set == NULL)
244 /* wake ptlrpcd when zombie imports or exports exist */
245 obd_zombie_impexp_notify = ptlrpcd_zombie_impexp_notify;
247 rc = cfs_kernel_thread(ptlrpcd, pc, 0);
249 ptlrpc_set_destroy(pc->pc_set);
253 wait_for_completion(&pc->pc_starting);
255 pc->pc_wait_callback =
256 liblustre_register_wait_callback("ptlrpcd_check_async_rpcs",
257 &ptlrpcd_check_async_rpcs, pc);
258 pc->pc_idle_callback =
259 liblustre_register_wait_callback("ptlrpcd_check_idle_rpcs",
266 static void ptlrpcd_stop(struct ptlrpcd_ctl *pc)
268 set_bit(LIOD_STOP, &pc->pc_flags);
269 cfs_waitq_signal(&pc->pc_waitq);
271 obd_zombie_impexp_notify = NULL;
272 wait_for_completion(&pc->pc_finishing);
274 liblustre_deregister_wait_callback(pc->pc_wait_callback);
275 liblustre_deregister_wait_callback(pc->pc_idle_callback);
277 ptlrpc_set_destroy(pc->pc_set);
280 int ptlrpcd_addref(void)
285 mutex_down(&ptlrpcd_sem);
286 if (++ptlrpcd_users != 1)
289 rc = ptlrpcd_start("ptlrpcd", &ptlrpcd_pc);
295 rc = ptlrpcd_start("ptlrpcd-recov", &ptlrpcd_recovery_pc);
297 ptlrpcd_stop(&ptlrpcd_pc);
302 mutex_up(&ptlrpcd_sem);
306 void ptlrpcd_decref(void)
308 mutex_down(&ptlrpcd_sem);
309 if (--ptlrpcd_users == 0) {
310 ptlrpcd_stop(&ptlrpcd_pc);
311 ptlrpcd_stop(&ptlrpcd_recovery_pc);
313 mutex_up(&ptlrpcd_sem);