Whamcloud - gitweb
LU-11415 socklnd: improve scheduling algorithm
[fs/lustre-release.git] / lnet / klnds / socklnd / socklnd.c
index dd59e93..ef20ef4 100644 (file)
@@ -665,33 +665,20 @@ ksocknal_get_conn_by_idx(struct lnet_ni *ni, int index)
 static struct ksock_sched *
 ksocknal_choose_scheduler_locked(unsigned int cpt)
 {
-       struct ksock_sched_info *info = ksocknal_data.ksnd_sched_info[cpt];
-       struct ksock_sched *sched;
+       struct ksock_sched *sched = ksocknal_data.ksnd_schedulers[cpt];
        int i;
 
-       if (info->ksi_nthreads == 0) {
-               cfs_percpt_for_each(info, i, ksocknal_data.ksnd_sched_info) {
-                       if (info->ksi_nthreads > 0) {
+       if (sched->kss_nthreads == 0) {
+               cfs_percpt_for_each(sched, i, ksocknal_data.ksnd_schedulers) {
+                       if (sched->kss_nthreads > 0) {
                                CDEBUG(D_NET, "scheduler[%d] has no threads. selected scheduler[%d]\n",
-                                      cpt, info->ksi_cpt);
-                               goto select_sched;
+                                      cpt, sched->kss_cpt);
+                               return sched;
                        }
                }
                return NULL;
        }
 
-select_sched:
-       sched = &info->ksi_scheds[0];
-       /*
-        * NB: it's safe so far, but info->ksi_nthreads could be changed
-        * at runtime when we have dynamic LNet configuration, then we
-        * need to take care of this.
-        */
-       for (i = 1; i < info->ksi_nthreads; i++) {
-               if (sched->kss_nconns > info->ksi_scheds[i].kss_nconns)
-                       sched = &info->ksi_scheds[i];
-       }
-
        return sched;
 }
 
@@ -1280,7 +1267,7 @@ ksocknal_create_conn(struct lnet_ni *ni, struct ksock_route *route,
         * The cpt might have changed if we ended up selecting a non cpt
         * native scheduler. So use the scheduler's cpt instead.
         */
-       cpt = sched->kss_info->ksi_cpt;
+       cpt = sched->kss_cpt;
         sched->kss_nconns++;
         conn->ksnc_scheduler = sched;
 
@@ -1319,11 +1306,10 @@ ksocknal_create_conn(struct lnet_ni *ni, struct ksock_route *route,
          */
 
        CDEBUG(D_NET, "New conn %s p %d.x %pI4h -> %pI4h/%d"
-              " incarnation:%lld sched[%d:%d]\n",
+              " incarnation:%lld sched[%d]\n",
               libcfs_id2str(peerid), conn->ksnc_proto->pro_version,
               &conn->ksnc_myipaddr, &conn->ksnc_ipaddr,
-              conn->ksnc_port, incarnation, cpt,
-              (int)(sched - &sched->kss_info->ksi_scheds[0]));
+              conn->ksnc_port, incarnation, cpt);
 
         if (active) {
                 /* additional routes after interface exchange? */
@@ -2208,7 +2194,7 @@ ksocknal_ctl(struct lnet_ni *ni, unsigned int cmd, void *arg)
                 data->ioc_u32[1] = conn->ksnc_port;
                 data->ioc_u32[2] = conn->ksnc_myipaddr;
                 data->ioc_u32[3] = conn->ksnc_type;
-               data->ioc_u32[4] = conn->ksnc_scheduler->kss_info->ksi_cpt;
+               data->ioc_u32[4] = conn->ksnc_scheduler->kss_cpt;
                 data->ioc_u32[5] = rxmem;
                 data->ioc_u32[6] = conn->ksnc_peer->ksnp_id.pid;
                 ksocknal_conn_decref(conn);
@@ -2247,19 +2233,8 @@ ksocknal_free_buffers (void)
 {
        LASSERT (atomic_read(&ksocknal_data.ksnd_nactive_txs) == 0);
 
-       if (ksocknal_data.ksnd_sched_info != NULL) {
-               struct ksock_sched_info *info;
-               int                     i;
-
-               cfs_percpt_for_each(info, i, ksocknal_data.ksnd_sched_info) {
-                       if (info->ksi_scheds != NULL) {
-                               LIBCFS_FREE(info->ksi_scheds,
-                                           info->ksi_nthreads_max *
-                                           sizeof(info->ksi_scheds[0]));
-                       }
-               }
-               cfs_percpt_free(ksocknal_data.ksnd_sched_info);
-       }
+       if (ksocknal_data.ksnd_schedulers != NULL)
+               cfs_percpt_free(ksocknal_data.ksnd_schedulers);
 
         LIBCFS_FREE (ksocknal_data.ksnd_peers,
                     sizeof(struct list_head) *
@@ -2288,10 +2263,8 @@ ksocknal_free_buffers (void)
 static void
 ksocknal_base_shutdown(void)
 {
-       struct ksock_sched_info *info;
        struct ksock_sched *sched;
        int i;
-       int j;
 
        CDEBUG(D_MALLOC, "before NAL cleanup: kmem %d\n",
               atomic_read (&libcfs_kmemory));
@@ -2314,23 +2287,14 @@ ksocknal_base_shutdown(void)
                LASSERT(list_empty(&ksocknal_data.ksnd_connd_connreqs));
                LASSERT(list_empty(&ksocknal_data.ksnd_connd_routes));
 
-               if (ksocknal_data.ksnd_sched_info != NULL) {
-                       cfs_percpt_for_each(info, i,
-                                           ksocknal_data.ksnd_sched_info) {
-                               if (info->ksi_scheds == NULL)
-                                       continue;
-
-                               for (j = 0; j < info->ksi_nthreads_max; j++) {
+               if (ksocknal_data.ksnd_schedulers != NULL) {
+                       cfs_percpt_for_each(sched, i,
+                                           ksocknal_data.ksnd_schedulers) {
 
-                                       sched = &info->ksi_scheds[j];
-                                       LASSERT(list_empty(&sched->\
-                                                              kss_tx_conns));
-                                       LASSERT(list_empty(&sched->\
-                                                              kss_rx_conns));
-                                       LASSERT(list_empty(&sched-> \
-                                                 kss_zombie_noop_txs));
-                                       LASSERT(sched->kss_nconns == 0);
-                               }
+                               LASSERT(list_empty(&sched->kss_tx_conns));
+                               LASSERT(list_empty(&sched->kss_rx_conns));
+                               LASSERT(list_empty(&sched->kss_zombie_noop_txs));
+                               LASSERT(sched->kss_nconns == 0);
                        }
                }
 
@@ -2339,17 +2303,10 @@ ksocknal_base_shutdown(void)
                wake_up_all(&ksocknal_data.ksnd_connd_waitq);
                wake_up_all(&ksocknal_data.ksnd_reaper_waitq);
 
-               if (ksocknal_data.ksnd_sched_info != NULL) {
-                       cfs_percpt_for_each(info, i,
-                                           ksocknal_data.ksnd_sched_info) {
-                               if (info->ksi_scheds == NULL)
-                                       continue;
-
-                               for (j = 0; j < info->ksi_nthreads_max; j++) {
-                                       sched = &info->ksi_scheds[j];
+               if (ksocknal_data.ksnd_schedulers != NULL) {
+                       cfs_percpt_for_each(sched, i,
+                                           ksocknal_data.ksnd_schedulers)
                                        wake_up_all(&sched->kss_waitq);
-                               }
-                       }
                }
 
                i = 4;
@@ -2382,9 +2339,9 @@ ksocknal_base_shutdown(void)
 static int
 ksocknal_base_startup(void)
 {
-       struct ksock_sched_info *info;
-       int                     rc;
-       int                     i;
+       struct ksock_sched *sched;
+       int rc;
+       int i;
 
         LASSERT (ksocknal_data.ksnd_init == SOCKNAL_INIT_NOTHING);
         LASSERT (ksocknal_data.ksnd_nnets == 0);
@@ -2424,45 +2381,38 @@ ksocknal_base_startup(void)
        ksocknal_data.ksnd_init = SOCKNAL_INIT_DATA;
        try_module_get(THIS_MODULE);
 
-       ksocknal_data.ksnd_sched_info = cfs_percpt_alloc(lnet_cpt_table(),
-                                                        sizeof(*info));
-       if (ksocknal_data.ksnd_sched_info == NULL)
+       /* Create a scheduler block per available CPT */
+       ksocknal_data.ksnd_schedulers = cfs_percpt_alloc(lnet_cpt_table(),
+                                                        sizeof(*sched));
+       if (ksocknal_data.ksnd_schedulers == NULL)
                goto failed;
 
-       cfs_percpt_for_each(info, i, ksocknal_data.ksnd_sched_info) {
-               struct ksock_sched *sched;
+       cfs_percpt_for_each(sched, i, ksocknal_data.ksnd_schedulers) {
                int nthrs;
 
+               /*
+                * make sure not to allocate more threads than there are
+                * cores/CPUs in teh CPT
+                */
                nthrs = cfs_cpt_weight(lnet_cpt_table(), i);
                if (*ksocknal_tunables.ksnd_nscheds > 0) {
                        nthrs = min(nthrs, *ksocknal_tunables.ksnd_nscheds);
                } else {
-                       /* max to half of CPUs, assume another half should be
-                        * reserved for upper layer modules */
+                       /*
+                        * max to half of CPUs, assume another half should be
+                        * reserved for upper layer modules
+                        */
                        nthrs = min(max(SOCKNAL_NSCHEDS, nthrs >> 1), nthrs);
                }
 
-               info->ksi_nthreads_max = nthrs;
-               info->ksi_cpt = i;
-
-               if (nthrs != 0) {
-                       LIBCFS_CPT_ALLOC(info->ksi_scheds, lnet_cpt_table(), i,
-                                        info->ksi_nthreads_max *
-                                               sizeof(*sched));
-                       if (info->ksi_scheds == NULL)
-                               goto failed;
-
-                       for (; nthrs > 0; nthrs--) {
-                               sched = &info->ksi_scheds[nthrs - 1];
-
-                               sched->kss_info = info;
-                               spin_lock_init(&sched->kss_lock);
-                               INIT_LIST_HEAD(&sched->kss_rx_conns);
-                               INIT_LIST_HEAD(&sched->kss_tx_conns);
-                               INIT_LIST_HEAD(&sched->kss_zombie_noop_txs);
-                               init_waitqueue_head(&sched->kss_waitq);
-                       }
-               }
+               sched->kss_nthreads_max = nthrs;
+               sched->kss_cpt = i;
+
+               spin_lock_init(&sched->kss_lock);
+               INIT_LIST_HEAD(&sched->kss_rx_conns);
+               INIT_LIST_HEAD(&sched->kss_tx_conns);
+               INIT_LIST_HEAD(&sched->kss_zombie_noop_txs);
+               init_waitqueue_head(&sched->kss_waitq);
         }
 
         ksocknal_data.ksnd_connd_starting         = 0;
@@ -2720,37 +2670,35 @@ ksocknal_search_new_ipif(struct ksock_net *net)
 }
 
 static int
-ksocknal_start_schedulers(struct ksock_sched_info *info)
+ksocknal_start_schedulers(struct ksock_sched *sched)
 {
        int     nthrs;
        int     rc = 0;
        int     i;
 
-       if (info->ksi_nthreads == 0) {
+       if (sched->kss_nthreads == 0) {
                if (*ksocknal_tunables.ksnd_nscheds > 0) {
-                       nthrs = info->ksi_nthreads_max;
+                       nthrs = sched->kss_nthreads_max;
                } else {
                        nthrs = cfs_cpt_weight(lnet_cpt_table(),
-                                              info->ksi_cpt);
+                                              sched->kss_cpt);
                        nthrs = min(max(SOCKNAL_NSCHEDS, nthrs >> 1), nthrs);
                        nthrs = min(SOCKNAL_NSCHEDS_HIGH, nthrs);
                }
-               nthrs = min(nthrs, info->ksi_nthreads_max);
+               nthrs = min(nthrs, sched->kss_nthreads_max);
        } else {
-               LASSERT(info->ksi_nthreads <= info->ksi_nthreads_max);
+               LASSERT(sched->kss_nthreads <= sched->kss_nthreads_max);
                /* increase two threads if there is new interface */
-               nthrs = min(2, info->ksi_nthreads_max - info->ksi_nthreads);
+               nthrs = min(2, sched->kss_nthreads_max - sched->kss_nthreads);
        }
 
        for (i = 0; i < nthrs; i++) {
                long id;
                char name[20];
-               struct ksock_sched *sched;
 
-               id = KSOCK_THREAD_ID(info->ksi_cpt, info->ksi_nthreads + i);
-               sched = &info->ksi_scheds[KSOCK_THREAD_SID(id)];
+               id = KSOCK_THREAD_ID(sched->kss_cpt, sched->kss_nthreads + i);
                snprintf(name, sizeof(name), "socknal_sd%02d_%02d",
-                        info->ksi_cpt, (int)(sched - &info->ksi_scheds[0]));
+                        sched->kss_cpt, (int)KSOCK_THREAD_SID(id));
 
                rc = ksocknal_thread_start(ksocknal_scheduler,
                                           (void *)id, name);
@@ -2758,11 +2706,11 @@ ksocknal_start_schedulers(struct ksock_sched_info *info)
                        continue;
 
                CERROR("Can't spawn thread %d for scheduler[%d]: %d\n",
-                      info->ksi_cpt, info->ksi_nthreads + i, rc);
+                      sched->kss_cpt, (int) KSOCK_THREAD_SID(id), rc);
                break;
        }
 
-       info->ksi_nthreads += i;
+       sched->kss_nthreads += i;
        return rc;
 }
 
@@ -2777,16 +2725,16 @@ ksocknal_net_start_threads(struct ksock_net *net, __u32 *cpts, int ncpts)
                return -EINVAL;
 
        for (i = 0; i < ncpts; i++) {
-               struct ksock_sched_info *info;
+               struct ksock_sched *sched;
                int cpt = (cpts == NULL) ? i : cpts[i];
 
                LASSERT(cpt < cfs_cpt_number(lnet_cpt_table()));
-               info = ksocknal_data.ksnd_sched_info[cpt];
+               sched = ksocknal_data.ksnd_schedulers[cpt];
 
-               if (!newif && info->ksi_nthreads > 0)
+               if (!newif && sched->kss_nthreads > 0)
                        continue;
 
-               rc = ksocknal_start_schedulers(info);
+               rc = ksocknal_start_schedulers(sched);
                if (rc != 0)
                        return rc;
        }