X-Git-Url: https://git.whamcloud.com/?a=blobdiff_plain;f=lustre%2Fptlrpc%2Fptlrpcd.c;h=78c054606ebb8c41ccd49589c24afb1f910568dc;hb=06a67987e06f7fa9c0bfde549d54e2bed0a78ade;hp=30c282380501a943b1e5daf8ac72dcdcc49affd5;hpb=fbf5870b9848929d352460f1f005b79c0b5ccc5a;p=fs%2Flustre-release.git diff --git a/lustre/ptlrpc/ptlrpcd.c b/lustre/ptlrpc/ptlrpcd.c index 30c2823..78c0546 100644 --- a/lustre/ptlrpc/ptlrpcd.c +++ b/lustre/ptlrpc/ptlrpcd.c @@ -26,7 +26,7 @@ * GPL HEADER END */ /* - * Copyright 2008 Sun Microsystems, Inc. All rights reserved + * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved. * Use is subject to license terms. */ /* @@ -36,6 +36,23 @@ * lustre/ptlrpc/ptlrpcd.c */ +/** \defgroup ptlrpcd PortalRPC daemon + * + * ptlrpcd is a special thread with its own set where other user might add + * requests when they don't want to wait for their completion. + * PtlRPCD will take care of sending such requests and then processing their + * replies and calling completion callbacks as necessary. + * The callbacks are called directly from ptlrpcd context. + * It is important to never significantly block (esp. on RPCs!) within such + * completion handler or a deadlock might occur where ptlrpcd enters some + * callback that attempts to send another RPC and wait for it to return, + * during which time ptlrpcd is completely blocked, so e.g. if import + * fails, recovery cannot progress because connection requests are also + * sent by ptlrpcd. + * + * @{ + */ + #define DEBUG_SUBSYSTEM S_RPC #ifdef __KERNEL__ @@ -90,7 +107,7 @@ static struct ptlrpcd_scope_ctl ptlrpcd_scopes[PSCOPE_NR] = { } }; -struct semaphore ptlrpcd_sem; +cfs_semaphore_t ptlrpcd_sem; static int ptlrpcd_users = 0; void ptlrpcd_wake(struct ptlrpc_request *req) @@ -102,79 +119,131 @@ void ptlrpcd_wake(struct ptlrpc_request *req) cfs_waitq_signal(&rq_set->set_waitq); } -/* +/** + * Move all request from an existing request set to the ptlrpcd queue. + * All requests from the set must be in phase RQ_PHASE_NEW. + */ +void ptlrpcd_add_rqset(struct ptlrpc_request_set *set) +{ + cfs_list_t *tmp, *pos; + + cfs_list_for_each_safe(pos, tmp, &set->set_requests) { + struct ptlrpc_request *req = + cfs_list_entry(pos, struct ptlrpc_request, + rq_set_chain); + + LASSERT(req->rq_phase == RQ_PHASE_NEW); + cfs_list_del_init(&req->rq_set_chain); + req->rq_set = NULL; + ptlrpcd_add_req(req, PSCOPE_OTHER); + cfs_atomic_dec(&set->set_remaining); + } + LASSERT(cfs_atomic_read(&set->set_remaining) == 0); +} +EXPORT_SYMBOL(ptlrpcd_add_rqset); + +/** * Requests that are added to the ptlrpcd queue are sent via * ptlrpcd_check->ptlrpc_check_set(). */ -void ptlrpcd_add_req(struct ptlrpc_request *req, enum ptlrpcd_scope scope) +int ptlrpcd_add_req(struct ptlrpc_request *req, enum ptlrpcd_scope scope) { struct ptlrpcd_ctl *pc; enum pscope_thread pt; int rc; LASSERT(scope < PSCOPE_NR); + + cfs_spin_lock(&req->rq_lock); + if (req->rq_invalid_rqset) { + cfs_duration_t timeout; + struct l_wait_info lwi; + + req->rq_invalid_rqset = 0; + cfs_spin_unlock(&req->rq_lock); + + timeout = cfs_time_seconds(5); + lwi = LWI_TIMEOUT(timeout, back_to_sleep, NULL); + l_wait_event(req->rq_set_waitq, (req->rq_set == NULL), &lwi); + } else if (req->rq_set) { + LASSERT(req->rq_phase == RQ_PHASE_NEW); + LASSERT(req->rq_send_state == LUSTRE_IMP_REPLAY); + + /* ptlrpc_check_set will decrease the count */ + cfs_atomic_inc(&req->rq_set->set_remaining); + cfs_spin_unlock(&req->rq_lock); + + cfs_waitq_signal(&req->rq_set->set_waitq); + } else { + cfs_spin_unlock(&req->rq_lock); + } + pt = req->rq_send_state == LUSTRE_IMP_FULL ? PT_NORMAL : PT_RECOVERY; pc = &ptlrpcd_scopes[scope].pscope_thread[pt].pt_ctl; rc = ptlrpc_set_add_new_req(pc, req); /* * XXX disable this for CLIO: environment is needed for interpreter. + * add debug temporary to check rc. */ + LASSERTF(rc == 0, "ptlrpcd_add_req failed (rc = %d)\n", rc); if (rc && 0) { - ptlrpc_interpterer_t interpreter; - - interpreter = req->rq_interpret_reply; - /* * Thread is probably in stop now so we need to * kill this rpc as it was not added. Let's call * interpret for it to let know we're killing it - * so that higher levels might free assosiated + * so that higher levels might free associated * resources. */ - req->rq_status = -EBADR; - interpreter(NULL, req, &req->rq_async_args, - req->rq_status); + ptlrpc_req_interpret(NULL, req, -EBADR); req->rq_set = NULL; ptlrpc_req_finished(req); + } else if (req->rq_send_state == LUSTRE_IMP_CONNECTING) { + /* + * The request is for recovery, should be sent ASAP. + */ + cfs_waitq_signal(&pc->pc_set->set_waitq); } + + return rc; } +/** + * Check if there is more work to do on ptlrpcd set. + * Returns 1 if yes. + */ static int ptlrpcd_check(const struct lu_env *env, struct ptlrpcd_ctl *pc) { - struct list_head *tmp, *pos; + cfs_list_t *tmp, *pos; struct ptlrpc_request *req; int rc = 0; ENTRY; - if (test_bit(LIOD_STOP, &pc->pc_flags)) - RETURN(1); - - spin_lock(&pc->pc_set->set_new_req_lock); - list_for_each_safe(pos, tmp, &pc->pc_set->set_new_requests) { - req = list_entry(pos, struct ptlrpc_request, rq_set_chain); - list_del_init(&req->rq_set_chain); + cfs_spin_lock(&pc->pc_set->set_new_req_lock); + cfs_list_for_each_safe(pos, tmp, &pc->pc_set->set_new_requests) { + req = cfs_list_entry(pos, struct ptlrpc_request, rq_set_chain); + cfs_list_del_init(&req->rq_set_chain); ptlrpc_set_add_req(pc->pc_set, req); /* * Need to calculate its timeout. */ rc = 1; } - spin_unlock(&pc->pc_set->set_new_req_lock); + cfs_spin_unlock(&pc->pc_set->set_new_req_lock); - if (pc->pc_set->set_remaining) { + if (cfs_atomic_read(&pc->pc_set->set_remaining)) { rc = rc | ptlrpc_check_set(env, pc->pc_set); /* * XXX: our set never completes, so we prune the completed * reqs after each iteration. boy could this be smarter. */ - list_for_each_safe(pos, tmp, &pc->pc_set->set_requests) { - req = list_entry(pos, struct ptlrpc_request, + cfs_list_for_each_safe(pos, tmp, &pc->pc_set->set_requests) { + req = cfs_list_entry(pos, struct ptlrpc_request, rq_set_chain); if (req->rq_phase != RQ_PHASE_COMPLETE) continue; - list_del_init(&req->rq_set_chain); + cfs_list_del_init(&req->rq_set_chain); req->rq_set = NULL; ptlrpc_req_finished (req); } @@ -184,25 +253,26 @@ static int ptlrpcd_check(const struct lu_env *env, struct ptlrpcd_ctl *pc) /* * If new requests have been added, make sure to wake up. */ - spin_lock(&pc->pc_set->set_new_req_lock); - rc = !list_empty(&pc->pc_set->set_new_requests); - spin_unlock(&pc->pc_set->set_new_req_lock); + cfs_spin_lock(&pc->pc_set->set_new_req_lock); + rc = !cfs_list_empty(&pc->pc_set->set_new_requests); + cfs_spin_unlock(&pc->pc_set->set_new_req_lock); } RETURN(rc); } #ifdef __KERNEL__ -/* +/** + * Main ptlrpcd thread. * ptlrpc's code paths like to execute in process context, so we have this - * thread which spins on a set which contains the io rpcs. llite specifies - * ptlrpcd's set when it pushes pages down into the oscs. + * thread which spins on a set which contains the rpcs and sends them. + * */ static int ptlrpcd(void *arg) { struct ptlrpcd_ctl *pc = arg; struct lu_env env = { .le_ses = NULL }; - int rc; + int rc, exit = 0; ENTRY; rc = cfs_daemonize_ctxt(pc->pc_name); @@ -216,18 +286,19 @@ static int ptlrpcd(void *arg) LCT_CL_THREAD|LCT_REMEMBER|LCT_NOREF); } - complete(&pc->pc_starting); + cfs_complete(&pc->pc_starting); if (rc != 0) RETURN(rc); env.le_ctx.lc_cookie = 0x7; + /* * This mainloop strongly resembles ptlrpc_set_wait() except that our * set never completes. ptlrpcd_check() calls ptlrpc_check_set() when * there are requests in the set. New requests come in on the set's * new_req_list and ptlrpcd_check() moves them into the set. */ - while (1) { + do { struct l_wait_info lwi; int timeout; @@ -259,28 +330,39 @@ static int ptlrpcd(void *arg) /* * Abort inflight rpcs for forced stop case. */ - if (test_bit(LIOD_STOP_FORCE, &pc->pc_flags)) - ptlrpc_abort_set(pc->pc_set); + if (cfs_test_bit(LIOD_STOP, &pc->pc_flags)) { + if (cfs_test_bit(LIOD_FORCE, &pc->pc_flags)) + ptlrpc_abort_set(pc->pc_set); + exit++; + } - if (test_bit(LIOD_STOP, &pc->pc_flags)) - break; - } + /* + * Let's make one more loop to make sure that ptlrpcd_check() + * copied all raced new rpcs into the set so we can kill them. + */ + } while (exit < 2); /* * Wait for inflight requests to drain. */ - if (!list_empty(&pc->pc_set->set_requests)) + if (!cfs_list_empty(&pc->pc_set->set_requests)) ptlrpc_set_wait(pc->pc_set); lu_context_fini(&env.le_ctx); - complete(&pc->pc_finishing); + cfs_complete(&pc->pc_finishing); - clear_bit(LIOD_START, &pc->pc_flags); - clear_bit(LIOD_STOP, &pc->pc_flags); + cfs_clear_bit(LIOD_START, &pc->pc_flags); + cfs_clear_bit(LIOD_STOP, &pc->pc_flags); + cfs_clear_bit(LIOD_FORCE, &pc->pc_flags); return 0; } #else /* !__KERNEL__ */ +/** + * In liblustre we do not have separate threads, so this function + * is called from time to time all across common code to see + * if something needs to be processed on ptlrpcd set. + */ int ptlrpcd_check_async_rpcs(void *arg) { struct ptlrpcd_ctl *pc = arg; @@ -292,16 +374,19 @@ int ptlrpcd_check_async_rpcs(void *arg) pc->pc_recurred++; if (pc->pc_recurred == 1) { - lu_context_enter(&pc->pc_env.le_ctx); - rc = ptlrpcd_check(&pc->pc_env, pc); - lu_context_exit(&pc->pc_env.le_ctx); - if (!rc) - ptlrpc_expired_set(pc->pc_set); - /* - * XXX: send replay requests. - */ - if (test_bit(LIOD_RECOVERY, &pc->pc_flags)) + rc = lu_env_refill(&pc->pc_env); + if (rc == 0) { + lu_context_enter(&pc->pc_env.le_ctx); rc = ptlrpcd_check(&pc->pc_env, pc); + lu_context_exit(&pc->pc_env.le_ctx); + if (!rc) + ptlrpc_expired_set(pc->pc_set); + /* + * XXX: send replay requests. + */ + if (cfs_test_bit(LIOD_RECOVERY, &pc->pc_flags)) + rc = ptlrpcd_check(&pc->pc_env, pc); + } } pc->pc_recurred--; @@ -312,8 +397,8 @@ int ptlrpcd_idle(void *arg) { struct ptlrpcd_ctl *pc = arg; - return (list_empty(&pc->pc_set->set_new_requests) && - pc->pc_set->set_remaining == 0); + return (cfs_list_empty(&pc->pc_set->set_new_requests) && + cfs_atomic_read(&pc->pc_set->set_remaining) == 0); } #endif @@ -326,16 +411,16 @@ int ptlrpcd_start(const char *name, struct ptlrpcd_ctl *pc) /* * Do not allow start second thread for one pc. */ - if (test_and_set_bit(LIOD_START, &pc->pc_flags)) { + if (cfs_test_and_set_bit(LIOD_START, &pc->pc_flags)) { CERROR("Starting second thread (%s) for same pc %p\n", name, pc); RETURN(-EALREADY); } - init_completion(&pc->pc_starting); - init_completion(&pc->pc_finishing); - spin_lock_init(&pc->pc_lock); - snprintf (pc->pc_name, sizeof (pc->pc_name), name); + cfs_init_completion(&pc->pc_starting); + cfs_init_completion(&pc->pc_finishing); + cfs_spin_lock_init(&pc->pc_lock); + strncpy(pc->pc_name, name, sizeof(pc->pc_name) - 1); pc->pc_set = ptlrpc_prep_set(); if (pc->pc_set == NULL) GOTO(out, rc = -ENOMEM); @@ -358,7 +443,7 @@ int ptlrpcd_start(const char *name, struct ptlrpcd_ctl *pc) GOTO(out, rc); } rc = 0; - wait_for_completion(&pc->pc_starting); + cfs_wait_for_completion(&pc->pc_starting); #else pc->pc_wait_callback = liblustre_register_wait_callback("ptlrpcd_check_async_rpcs", @@ -369,23 +454,23 @@ int ptlrpcd_start(const char *name, struct ptlrpcd_ctl *pc) #endif out: if (rc) - clear_bit(LIOD_START, &pc->pc_flags); + cfs_clear_bit(LIOD_START, &pc->pc_flags); RETURN(rc); } void ptlrpcd_stop(struct ptlrpcd_ctl *pc, int force) { - if (!test_bit(LIOD_START, &pc->pc_flags)) { + if (!cfs_test_bit(LIOD_START, &pc->pc_flags)) { CERROR("Thread for pc %p was not started\n", pc); return; } - set_bit(LIOD_STOP, &pc->pc_flags); + cfs_set_bit(LIOD_STOP, &pc->pc_flags); if (force) - set_bit(LIOD_STOP_FORCE, &pc->pc_flags); + cfs_set_bit(LIOD_FORCE, &pc->pc_flags); cfs_waitq_signal(&pc->pc_set->set_waitq); #ifdef __KERNEL__ - wait_for_completion(&pc->pc_finishing); + cfs_wait_for_completion(&pc->pc_finishing); #else liblustre_deregister_wait_callback(pc->pc_wait_callback); liblustre_deregister_idle_callback(pc->pc_idle_callback); @@ -407,7 +492,7 @@ void ptlrpcd_fini(void) pc = &ptlrpcd_scopes[i].pscope_thread[j].pt_ctl; - if (test_bit(LIOD_START, &pc->pc_flags)) + if (cfs_test_bit(LIOD_START, &pc->pc_flags)) ptlrpcd_stop(pc, 0); } } @@ -421,7 +506,7 @@ int ptlrpcd_addref(void) int j; ENTRY; - mutex_down(&ptlrpcd_sem); + cfs_mutex_down(&ptlrpcd_sem); if (++ptlrpcd_users == 1) { for (i = 0; rc == 0 && i < PSCOPE_NR; ++i) { for (j = 0; rc == 0 && j < PT_NR; ++j) { @@ -431,7 +516,7 @@ int ptlrpcd_addref(void) pt = &ptlrpcd_scopes[i].pscope_thread[j]; pc = &pt->pt_ctl; if (j == PT_RECOVERY) - set_bit(LIOD_RECOVERY, &pc->pc_flags); + cfs_set_bit(LIOD_RECOVERY, &pc->pc_flags); rc = ptlrpcd_start(pt->pt_name, pc); } } @@ -440,14 +525,15 @@ int ptlrpcd_addref(void) ptlrpcd_fini(); } } - mutex_up(&ptlrpcd_sem); + cfs_mutex_up(&ptlrpcd_sem); RETURN(rc); } void ptlrpcd_decref(void) { - mutex_down(&ptlrpcd_sem); + cfs_mutex_down(&ptlrpcd_sem); if (--ptlrpcd_users == 0) ptlrpcd_fini(); - mutex_up(&ptlrpcd_sem); + cfs_mutex_up(&ptlrpcd_sem); } +/** @} ptlrpcd */