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/ptlrpc/ptlrpcd.c
39 #define DEBUG_SUBSYSTEM S_RPC
42 # include <libcfs/libcfs.h>
43 #else /* __KERNEL__ */
44 # include <liblustre.h>
48 #include <lustre_net.h>
49 # include <lustre_lib.h>
51 #include <lustre_ha.h>
52 #include <obd_class.h> /* for obd_zombie */
53 #include <obd_support.h> /* for OBD_FAIL_CHECK */
54 #include <cl_object.h> /* cl_env_{get,put}() */
55 #include <lprocfs_status.h>
63 struct ptlrpcd_scope_ctl {
64 struct ptlrpcd_thread {
66 struct ptlrpcd_ctl pt_ctl;
67 } pscope_thread[PT_NR];
70 static struct ptlrpcd_scope_ctl ptlrpcd_scopes[PSCOPE_NR] = {
74 .pt_name = "ptlrpcd-brw"
77 .pt_name = "ptlrpcd-brw-rcv"
87 .pt_name = "ptlrpcd-rcv"
93 cfs_semaphore_t ptlrpcd_sem;
94 static int ptlrpcd_users = 0;
96 void ptlrpcd_wake(struct ptlrpc_request *req)
98 struct ptlrpc_request_set *rq_set = req->rq_set;
100 LASSERT(rq_set != NULL);
102 cfs_waitq_signal(&rq_set->set_waitq);
106 * Move all request from an existing request set to the ptlrpcd queue.
107 * All requests from the set must be in phase RQ_PHASE_NEW.
109 void ptlrpcd_add_rqset(struct ptlrpc_request_set *set)
111 cfs_list_t *tmp, *pos;
113 cfs_list_for_each_safe(pos, tmp, &set->set_requests) {
114 struct ptlrpc_request *req =
115 cfs_list_entry(pos, struct ptlrpc_request,
118 LASSERT(req->rq_phase == RQ_PHASE_NEW);
119 cfs_list_del_init(&req->rq_set_chain);
121 ptlrpcd_add_req(req, PSCOPE_OTHER);
122 cfs_atomic_dec(&set->set_remaining);
124 LASSERT(cfs_atomic_read(&set->set_remaining) == 0);
126 EXPORT_SYMBOL(ptlrpcd_add_rqset);
129 * Requests that are added to the ptlrpcd queue are sent via
130 * ptlrpcd_check->ptlrpc_check_set().
132 int ptlrpcd_add_req(struct ptlrpc_request *req, enum ptlrpcd_scope scope)
134 struct ptlrpcd_ctl *pc;
135 enum pscope_thread pt;
138 LASSERT(scope < PSCOPE_NR);
140 cfs_spin_lock(&req->rq_lock);
141 if (req->rq_invalid_rqset) {
142 cfs_duration_t timeout;
143 struct l_wait_info lwi;
145 req->rq_invalid_rqset = 0;
146 cfs_spin_unlock(&req->rq_lock);
148 timeout = cfs_time_seconds(5);
149 lwi = LWI_TIMEOUT(timeout, back_to_sleep, NULL);
150 l_wait_event(req->rq_reply_waitq, (req->rq_set == NULL), &lwi);
151 } else if (req->rq_set) {
152 LASSERT(req->rq_phase == RQ_PHASE_NEW);
153 LASSERT(req->rq_send_state == LUSTRE_IMP_REPLAY);
155 /* ptlrpc_check_set will decrease the count */
156 cfs_atomic_inc(&req->rq_set->set_remaining);
157 cfs_spin_unlock(&req->rq_lock);
159 cfs_waitq_signal(&req->rq_set->set_waitq);
161 cfs_spin_unlock(&req->rq_lock);
164 pt = req->rq_send_state == LUSTRE_IMP_FULL ? PT_NORMAL : PT_RECOVERY;
165 pc = &ptlrpcd_scopes[scope].pscope_thread[pt].pt_ctl;
166 rc = ptlrpc_set_add_new_req(pc, req);
168 * XXX disable this for CLIO: environment is needed for interpreter.
169 * add debug temporary to check rc.
171 LASSERTF(rc == 0, "ptlrpcd_add_req failed (rc = %d)\n", rc);
174 * Thread is probably in stop now so we need to
175 * kill this rpc as it was not added. Let's call
176 * interpret for it to let know we're killing it
177 * so that higher levels might free associated
180 ptlrpc_req_interpret(NULL, req, -EBADR);
182 ptlrpc_req_finished(req);
183 } else if (req->rq_send_state == LUSTRE_IMP_CONNECTING) {
185 * The request is for recovery, should be sent ASAP.
187 cfs_waitq_signal(&pc->pc_set->set_waitq);
193 static int ptlrpcd_check(const struct lu_env *env, struct ptlrpcd_ctl *pc)
195 cfs_list_t *tmp, *pos;
196 struct ptlrpc_request *req;
200 cfs_spin_lock(&pc->pc_set->set_new_req_lock);
201 cfs_list_for_each_safe(pos, tmp, &pc->pc_set->set_new_requests) {
202 req = cfs_list_entry(pos, struct ptlrpc_request, rq_set_chain);
203 cfs_list_del_init(&req->rq_set_chain);
204 ptlrpc_set_add_req(pc->pc_set, req);
206 * Need to calculate its timeout.
210 cfs_spin_unlock(&pc->pc_set->set_new_req_lock);
212 if (cfs_atomic_read(&pc->pc_set->set_remaining)) {
213 rc = rc | ptlrpc_check_set(env, pc->pc_set);
216 * XXX: our set never completes, so we prune the completed
217 * reqs after each iteration. boy could this be smarter.
219 cfs_list_for_each_safe(pos, tmp, &pc->pc_set->set_requests) {
220 req = cfs_list_entry(pos, struct ptlrpc_request,
222 if (req->rq_phase != RQ_PHASE_COMPLETE)
225 cfs_list_del_init(&req->rq_set_chain);
227 ptlrpc_req_finished (req);
233 * If new requests have been added, make sure to wake up.
235 cfs_spin_lock(&pc->pc_set->set_new_req_lock);
236 rc = !cfs_list_empty(&pc->pc_set->set_new_requests);
237 cfs_spin_unlock(&pc->pc_set->set_new_req_lock);
245 * ptlrpc's code paths like to execute in process context, so we have this
246 * thread which spins on a set which contains the io rpcs. llite specifies
247 * ptlrpcd's set when it pushes pages down into the oscs.
249 static int ptlrpcd(void *arg)
251 struct ptlrpcd_ctl *pc = arg;
252 struct lu_env env = { .le_ses = NULL };
256 rc = cfs_daemonize_ctxt(pc->pc_name);
259 * XXX So far only "client" ptlrpcd uses an environment. In
260 * the future, ptlrpcd thread (or a thread-set) has to given
261 * an argument, describing its "scope".
263 rc = lu_context_init(&env.le_ctx,
264 LCT_CL_THREAD|LCT_REMEMBER|LCT_NOREF);
267 cfs_complete(&pc->pc_starting);
271 env.le_ctx.lc_cookie = 0x7;
274 * This mainloop strongly resembles ptlrpc_set_wait() except that our
275 * set never completes. ptlrpcd_check() calls ptlrpc_check_set() when
276 * there are requests in the set. New requests come in on the set's
277 * new_req_list and ptlrpcd_check() moves them into the set.
280 struct l_wait_info lwi;
283 rc = lu_env_refill(&env);
286 * XXX This is very awkward situation, because
287 * execution can neither continue (request
288 * interpreters assume that env is set up), nor repeat
289 * the loop (as this potentially results in a tight
290 * loop of -ENOMEM's).
292 * Fortunately, refill only ever does something when
293 * new modules are loaded, i.e., early during boot up.
295 CERROR("Failure to refill session: %d\n", rc);
299 timeout = ptlrpc_set_next_timeout(pc->pc_set);
300 lwi = LWI_TIMEOUT(cfs_time_seconds(timeout ? timeout : 1),
301 ptlrpc_expired_set, pc->pc_set);
303 lu_context_enter(&env.le_ctx);
304 l_wait_event(pc->pc_set->set_waitq,
305 ptlrpcd_check(&env, pc), &lwi);
306 lu_context_exit(&env.le_ctx);
309 * Abort inflight rpcs for forced stop case.
311 if (cfs_test_bit(LIOD_STOP, &pc->pc_flags)) {
312 if (cfs_test_bit(LIOD_FORCE, &pc->pc_flags))
313 ptlrpc_abort_set(pc->pc_set);
318 * Let's make one more loop to make sure that ptlrpcd_check()
319 * copied all raced new rpcs into the set so we can kill them.
324 * Wait for inflight requests to drain.
326 if (!cfs_list_empty(&pc->pc_set->set_requests))
327 ptlrpc_set_wait(pc->pc_set);
328 lu_context_fini(&env.le_ctx);
329 cfs_complete(&pc->pc_finishing);
331 cfs_clear_bit(LIOD_START, &pc->pc_flags);
332 cfs_clear_bit(LIOD_STOP, &pc->pc_flags);
333 cfs_clear_bit(LIOD_FORCE, &pc->pc_flags);
337 #else /* !__KERNEL__ */
339 int ptlrpcd_check_async_rpcs(void *arg)
341 struct ptlrpcd_ctl *pc = arg;
349 if (pc->pc_recurred == 1) {
350 rc = lu_env_refill(&pc->pc_env);
352 lu_context_enter(&pc->pc_env.le_ctx);
353 rc = ptlrpcd_check(&pc->pc_env, pc);
354 lu_context_exit(&pc->pc_env.le_ctx);
356 ptlrpc_expired_set(pc->pc_set);
358 * XXX: send replay requests.
360 if (cfs_test_bit(LIOD_RECOVERY, &pc->pc_flags))
361 rc = ptlrpcd_check(&pc->pc_env, pc);
369 int ptlrpcd_idle(void *arg)
371 struct ptlrpcd_ctl *pc = arg;
373 return (cfs_list_empty(&pc->pc_set->set_new_requests) &&
374 cfs_atomic_read(&pc->pc_set->set_remaining) == 0);
379 int ptlrpcd_start(const char *name, struct ptlrpcd_ctl *pc)
385 * Do not allow start second thread for one pc.
387 if (cfs_test_and_set_bit(LIOD_START, &pc->pc_flags)) {
388 CERROR("Starting second thread (%s) for same pc %p\n",
393 cfs_init_completion(&pc->pc_starting);
394 cfs_init_completion(&pc->pc_finishing);
395 cfs_spin_lock_init(&pc->pc_lock);
396 strncpy(pc->pc_name, name, sizeof(pc->pc_name) - 1);
397 pc->pc_set = ptlrpc_prep_set();
398 if (pc->pc_set == NULL)
399 GOTO(out, rc = -ENOMEM);
401 * So far only "client" ptlrpcd uses an environment. In the future,
402 * ptlrpcd thread (or a thread-set) has to be given an argument,
403 * describing its "scope".
405 rc = lu_context_init(&pc->pc_env.le_ctx, LCT_CL_THREAD|LCT_REMEMBER);
407 ptlrpc_set_destroy(pc->pc_set);
412 rc = cfs_kernel_thread(ptlrpcd, pc, 0);
414 lu_context_fini(&pc->pc_env.le_ctx);
415 ptlrpc_set_destroy(pc->pc_set);
419 cfs_wait_for_completion(&pc->pc_starting);
421 pc->pc_wait_callback =
422 liblustre_register_wait_callback("ptlrpcd_check_async_rpcs",
423 &ptlrpcd_check_async_rpcs, pc);
424 pc->pc_idle_callback =
425 liblustre_register_idle_callback("ptlrpcd_check_idle_rpcs",
430 cfs_clear_bit(LIOD_START, &pc->pc_flags);
434 void ptlrpcd_stop(struct ptlrpcd_ctl *pc, int force)
436 if (!cfs_test_bit(LIOD_START, &pc->pc_flags)) {
437 CERROR("Thread for pc %p was not started\n", pc);
441 cfs_set_bit(LIOD_STOP, &pc->pc_flags);
443 cfs_set_bit(LIOD_FORCE, &pc->pc_flags);
444 cfs_waitq_signal(&pc->pc_set->set_waitq);
446 cfs_wait_for_completion(&pc->pc_finishing);
448 liblustre_deregister_wait_callback(pc->pc_wait_callback);
449 liblustre_deregister_idle_callback(pc->pc_idle_callback);
451 lu_context_fini(&pc->pc_env.le_ctx);
452 ptlrpc_set_destroy(pc->pc_set);
455 void ptlrpcd_fini(void)
462 for (i = 0; i < PSCOPE_NR; ++i) {
463 for (j = 0; j < PT_NR; ++j) {
464 struct ptlrpcd_ctl *pc;
466 pc = &ptlrpcd_scopes[i].pscope_thread[j].pt_ctl;
468 if (cfs_test_bit(LIOD_START, &pc->pc_flags))
475 int ptlrpcd_addref(void)
482 cfs_mutex_down(&ptlrpcd_sem);
483 if (++ptlrpcd_users == 1) {
484 for (i = 0; rc == 0 && i < PSCOPE_NR; ++i) {
485 for (j = 0; rc == 0 && j < PT_NR; ++j) {
486 struct ptlrpcd_thread *pt;
487 struct ptlrpcd_ctl *pc;
489 pt = &ptlrpcd_scopes[i].pscope_thread[j];
491 if (j == PT_RECOVERY)
492 cfs_set_bit(LIOD_RECOVERY, &pc->pc_flags);
493 rc = ptlrpcd_start(pt->pt_name, pc);
501 cfs_mutex_up(&ptlrpcd_sem);
505 void ptlrpcd_decref(void)
507 cfs_mutex_down(&ptlrpcd_sem);
508 if (--ptlrpcd_users == 0)
510 cfs_mutex_up(&ptlrpcd_sem);