Whamcloud - gitweb
LU-9683 ptlrpc: fix argument misorder
[fs/lustre-release.git] / lustre / ptlrpc / service.c
index 7006a57..ecde6ab 100644 (file)
@@ -185,16 +185,12 @@ ptlrpc_save_lock(struct ptlrpc_request *req, struct lustre_handle *lock,
        LASSERT(rs != NULL);
        LASSERT(rs->rs_nlocks < RS_MAX_LOCKS);
 
-       if (req->rq_export->exp_disconnected) {
-               ldlm_lock_decref(lock, mode);
-       } else {
-               idx = rs->rs_nlocks++;
-               rs->rs_locks[idx] = *lock;
-               rs->rs_modes[idx] = mode;
-               rs->rs_difficult = 1;
-               rs->rs_no_ack = no_ack;
-               rs->rs_convert_lock = convert_lock;
-       }
+       idx = rs->rs_nlocks++;
+       rs->rs_locks[idx] = *lock;
+       rs->rs_modes[idx] = mode;
+       rs->rs_difficult = 1;
+       rs->rs_no_ack = no_ack;
+       rs->rs_convert_lock = convert_lock;
 }
 EXPORT_SYMBOL(ptlrpc_save_lock);
 
@@ -567,7 +563,6 @@ ptlrpc_server_nthreads_check(struct ptlrpc_service *svc,
                }
 
                weight = cfs_cpt_weight(svc->srv_cptable, 0);
-               LASSERT(weight > 0);
 
                for (; factor > 0 && weight > 0; factor--, weight -= fade)
                        nthrs += min(weight, fade) * factor;
@@ -688,6 +683,7 @@ ptlrpc_service_part_init(struct ptlrpc_service *svc,
  */
 struct ptlrpc_service *
 ptlrpc_register_service(struct ptlrpc_service_conf *conf,
+                       struct kset *parent,
                        struct proc_dir_entry *proc_entry)
 {
        struct ptlrpc_service_cpt_conf  *cconf = &conf->psc_cpt;
@@ -805,6 +801,12 @@ ptlrpc_register_service(struct ptlrpc_service_conf *conf,
        list_add(&service->srv_list, &ptlrpc_all_services);
        mutex_unlock(&ptlrpc_all_services_mutex);
 
+       if (parent) {
+               rc = ptlrpc_sysfs_register_service(parent, service);
+               if (rc)
+                       GOTO(failed, rc);
+       }
+
        if (proc_entry != NULL)
                ptlrpc_lprocfs_register_service(proc_entry, service);
 
@@ -939,10 +941,22 @@ void ptlrpc_server_drop_request(struct ptlrpc_request *req)
                        spin_lock(&svcpt->scp_lock);
                        /*
                         * now all reqs including the embedded req has been
-                        * disposed, schedule request buffer for re-use.
+                        * disposed, schedule request buffer for re-use
+                        * or free it to drain some in excess.
                         */
                        LASSERT(atomic_read(&rqbd->rqbd_req.rq_refcount) == 0);
-                       list_add_tail(&rqbd->rqbd_list, &svcpt->scp_rqbd_idle);
+                       if (svcpt->scp_nrqbds_posted >=
+                           svc->srv_nbuf_per_group &&
+                           !test_req_buffer_pressure) {
+                               /* like in ptlrpc_free_rqbd() */
+                               svcpt->scp_nrqbds_total--;
+                               OBD_FREE_LARGE(rqbd->rqbd_buffer,
+                                              svc->srv_buf_size);
+                               OBD_FREE_PTR(rqbd);
+                       } else {
+                               list_add_tail(&rqbd->rqbd_list,
+                                             &svcpt->scp_rqbd_idle);
+                       }
                }
 
                spin_unlock(&svcpt->scp_lock);
@@ -1092,7 +1106,7 @@ void ptlrpc_update_export_timer(struct obd_export *exp, long extra_delay)
                          * we better wait for 3. */
                         exp->exp_obd->obd_eviction_timer =
                                 cfs_time_current_sec() + 3 * PING_INTERVAL;
-                        CDEBUG(D_HA, "%s: Think about evicting %s from "CFS_TIME_T"\n",
+                       CDEBUG(D_HA, "%s: Think about evicting %s from %ld\n",
                                exp->exp_obd->obd_name,
                                obd_export_nid2str(oldest_exp), oldest_time);
                 }
@@ -2041,7 +2055,7 @@ ptlrpc_server_handle_request(struct ptlrpc_service_part *svcpt,
 
        work_start = ktime_get_real();
        arrived = timespec64_to_ktime(request->rq_arrival_time);
-       timediff_usecs = ktime_us_delta(arrived, work_start);
+       timediff_usecs = ktime_us_delta(work_start, arrived);
        if (likely(svc->srv_stats != NULL)) {
                 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQWAIT_CNTR,
                                    timediff_usecs);
@@ -2197,13 +2211,28 @@ ptlrpc_handle_rs(struct ptlrpc_reply_state *rs)
                if (rs->rs_convert_lock &&
                    rs->rs_transno > exp->exp_last_committed) {
                        struct ldlm_lock *lock;
+                       struct ldlm_lock *ack_locks[RS_MAX_LOCKS] = { NULL };
 
                        spin_lock(&rs->rs_lock);
                        if (rs->rs_convert_lock &&
                            rs->rs_transno > exp->exp_last_committed) {
                                nlocks = rs->rs_nlocks;
-                               while (nlocks-- > 0)
+                               while (nlocks-- > 0) {
+                                       /*
+                                        * NB don't assume rs is always handled
+                                        * by the same service thread (see
+                                        * ptlrpc_hr_select, so REP-ACK hr may
+                                        * race with trans commit, while the
+                                        * latter will release locks, get locks
+                                        * here early to convert to COS mode
+                                        * safely.
+                                        */
+                                       lock = ldlm_handle2lock(
+                                                       &rs->rs_locks[nlocks]);
+                                       LASSERT(lock);
+                                       ack_locks[nlocks] = lock;
                                        rs->rs_modes[nlocks] = LCK_COS;
+                               }
                                nlocks = rs->rs_nlocks;
                                rs->rs_convert_lock = 0;
                                /* clear rs_scheduled so that commit callback
@@ -2212,9 +2241,7 @@ ptlrpc_handle_rs(struct ptlrpc_reply_state *rs)
                                spin_unlock(&rs->rs_lock);
 
                                while (nlocks-- > 0) {
-                                       lock = ldlm_handle2lock(
-                                                       &rs->rs_locks[nlocks]);
-                                       LASSERT(lock != NULL);
+                                       lock = ack_locks[nlocks];
                                        ldlm_lock_downgrade(lock, LCK_COS);
                                        LDLM_LOCK_PUT(lock);
                                }
@@ -2931,8 +2958,8 @@ int ptlrpc_hr_init(void)
        struct ptlrpc_hr_partition      *hrp;
        struct ptlrpc_hr_thread         *hrt;
        int                             rc;
+       int                             cpt;
        int                             i;
-       int                             j;
        int                             weight;
        ENTRY;
 
@@ -2948,27 +2975,26 @@ int ptlrpc_hr_init(void)
 
        weight = cpumask_weight(topology_sibling_cpumask(smp_processor_id()));
 
-       cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
-               hrp->hrp_cpt = i;
+       cfs_percpt_for_each(hrp, cpt, ptlrpc_hr.hr_partitions) {
+               hrp->hrp_cpt = cpt;
 
                atomic_set(&hrp->hrp_nstarted, 0);
                atomic_set(&hrp->hrp_nstopped, 0);
 
-               hrp->hrp_nthrs = cfs_cpt_weight(ptlrpc_hr.hr_cpt_table, i);
-
+               hrp->hrp_nthrs = cfs_cpt_weight(ptlrpc_hr.hr_cpt_table, cpt);
                hrp->hrp_nthrs /= weight;
                if (hrp->hrp_nthrs == 0)
                        hrp->hrp_nthrs = 1;
 
-               OBD_CPT_ALLOC(hrp->hrp_thrs, ptlrpc_hr.hr_cpt_table, i,
+               OBD_CPT_ALLOC(hrp->hrp_thrs, ptlrpc_hr.hr_cpt_table, cpt,
                              hrp->hrp_nthrs * sizeof(*hrt));
                if (hrp->hrp_thrs == NULL)
                        GOTO(out, rc = -ENOMEM);
 
-               for (j = 0; j < hrp->hrp_nthrs; j++) {
-                       hrt = &hrp->hrp_thrs[j];
+               for (i = 0; i < hrp->hrp_nthrs; i++) {
+                       hrt = &hrp->hrp_thrs[i];
 
-                       hrt->hrt_id = j;
+                       hrt->hrt_id = i;
                        hrt->hrt_partition = hrp;
                        init_waitqueue_head(&hrt->hrt_waitq);
                        spin_lock_init(&hrt->hrt_lock);
@@ -2986,14 +3012,14 @@ out:
 void ptlrpc_hr_fini(void)
 {
        struct ptlrpc_hr_partition      *hrp;
-       int                             i;
+       int                             cpt;
 
        if (ptlrpc_hr.hr_partitions == NULL)
                return;
 
        ptlrpc_stop_hr_threads();
 
-       cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
+       cfs_percpt_for_each(hrp, cpt, ptlrpc_hr.hr_partitions) {
                if (hrp->hrp_thrs != NULL) {
                        OBD_FREE(hrp->hrp_thrs,
                                 hrp->hrp_nthrs * sizeof(hrp->hrp_thrs[0]));
@@ -3220,6 +3246,7 @@ int ptlrpc_unregister_service(struct ptlrpc_service *service)
        ptlrpc_service_nrs_cleanup(service);
 
        ptlrpc_lprocfs_unregister_service(service);
+       ptlrpc_sysfs_unregister_service(service);
 
        ptlrpc_service_free(service);