[lustre-devel] [PATCH 6/7] lustre: lnet: convert ni_refs to percpu_refcount.
James Simmons
jsimmons at infradead.org
Wed Aug 1 20:47:38 PDT 2018
> ni_refs is a per-cpt refcount.
> Linux already has a per-cpu refcount implementation
> which doesn't require anylocking.
>
> So convert ni_refs to percpu_refcount.
> As a bonus, we can get a wake-up when the refcount
> reaches zero, rather than having to wait a full second.
> The waiting in lnet_clear_zombies_nis_locked() is
> modified so that instead of waiting one second each
> time, and printing a warning on power-of-two seconds,
> we wait an increasing power-of-two seconds and print
> a warning if the wait ever timed out.
Reviewed-by: James Simmons <jsimmons at infradead.org>
> Signed-off-by: NeilBrown <neilb at suse.com>
> ---
> .../staging/lustre/include/linux/lnet/lib-lnet.h | 11 +-----
> .../staging/lustre/include/linux/lnet/lib-types.h | 2 +
> drivers/staging/lustre/lnet/lnet/api-ni.c | 34 +++++++++-----------
> drivers/staging/lustre/lnet/lnet/config.c | 13 +++++---
> drivers/staging/lustre/lnet/lnet/router_proc.c | 2 +
> 5 files changed, 28 insertions(+), 34 deletions(-)
>
> diff --git a/drivers/staging/lustre/include/linux/lnet/lib-lnet.h b/drivers/staging/lustre/include/linux/lnet/lib-lnet.h
> index 0fecf0d32c58..371002825a7d 100644
> --- a/drivers/staging/lustre/include/linux/lnet/lib-lnet.h
> +++ b/drivers/staging/lustre/include/linux/lnet/lib-lnet.h
> @@ -338,34 +338,27 @@ static inline void
> lnet_ni_addref_locked(struct lnet_ni *ni, int cpt)
> {
> LASSERT(cpt >= 0 && cpt < LNET_CPT_NUMBER);
> - LASSERT(*ni->ni_refs[cpt] >= 0);
> -
> - (*ni->ni_refs[cpt])++;
> + percpu_ref_get(&ni->ni_refs);
> }
>
> static inline void
> lnet_ni_addref(struct lnet_ni *ni)
> {
> - lnet_net_lock(0);
> lnet_ni_addref_locked(ni, 0);
> - lnet_net_unlock(0);
> }
>
> static inline void
> lnet_ni_decref_locked(struct lnet_ni *ni, int cpt)
> {
> LASSERT(cpt >= 0 && cpt < LNET_CPT_NUMBER);
> - LASSERT(*ni->ni_refs[cpt] > 0);
>
> - (*ni->ni_refs[cpt])--;
> + percpu_ref_put(&ni->ni_refs);
> }
>
> static inline void
> lnet_ni_decref(struct lnet_ni *ni)
> {
> - lnet_net_lock(0);
> lnet_ni_decref_locked(ni, 0);
> - lnet_net_unlock(0);
> }
>
> void lnet_ni_free(struct lnet_ni *ni);
> diff --git a/drivers/staging/lustre/include/linux/lnet/lib-types.h b/drivers/staging/lustre/include/linux/lnet/lib-types.h
> index 6d4106fd9039..7527fef90cac 100644
> --- a/drivers/staging/lustre/include/linux/lnet/lib-types.h
> +++ b/drivers/staging/lustre/include/linux/lnet/lib-types.h
> @@ -269,7 +269,7 @@ struct lnet_ni {
> void *ni_data; /* instance-specific data */
> struct lnet_lnd *ni_lnd; /* procedural interface */
> struct lnet_tx_queue **ni_tx_queues; /* percpt TX queues */
> - int **ni_refs; /* percpt reference count */
> + struct percpu_ref ni_refs;
> time64_t ni_last_alive;/* when I was last alive */
> struct lnet_ni_status *ni_status; /* my health status */
> /* per NI LND tunables */
> diff --git a/drivers/staging/lustre/lnet/lnet/api-ni.c b/drivers/staging/lustre/lnet/lnet/api-ni.c
> index cdbbe9cc8d95..fea03737439a 100644
> --- a/drivers/staging/lustre/lnet/lnet/api-ni.c
> +++ b/drivers/staging/lustre/lnet/lnet/api-ni.c
> @@ -1055,7 +1055,7 @@ lnet_ni_unlink_locked(struct lnet_ni *ni)
> /* move it to zombie list and nobody can find it anymore */
> LASSERT(!list_empty(&ni->ni_list));
> list_move(&ni->ni_list, &the_lnet.ln_nis_zombie);
> - lnet_ni_decref_locked(ni, 0); /* drop ln_nis' ref */
> + percpu_ref_kill_and_confirm(&ni->ni_refs, NULL); /* drop ln_nis' ref */
> }
>
> static void
> @@ -1069,34 +1069,32 @@ lnet_clear_zombies_nis_locked(void)
> * Now wait for the NI's I just nuked to show up on ln_zombie_nis
> * and shut them down in guaranteed thread context
> */
> - i = 2;
> + i = 1;
> while (!list_empty(&the_lnet.ln_nis_zombie)) {
> - int *ref;
> - int j;
>
> ni = list_entry(the_lnet.ln_nis_zombie.next,
> struct lnet_ni, ni_list);
> - list_del_init(&ni->ni_list);
> - cfs_percpt_for_each(ref, j, ni->ni_refs) {
> - if (!*ref)
> - continue;
> - /* still busy, add it back to zombie list */
> - list_add(&ni->ni_list, &the_lnet.ln_nis_zombie);
> - break;
> - }
>
> - if (!list_empty(&ni->ni_list)) {
> + if (!percpu_ref_is_zero(&ni->ni_refs)) {
> + /* still busy, wait a while */
> +
> lnet_net_unlock(LNET_LOCK_EX);
> ++i;
> - if ((i & (-i)) == i) {
> +
> + if (wait_var_event_timeout(
> + &ni->ni_refs,
> + percpu_ref_is_zero(&ni->ni_refs),
> + HZ << i) == 0)
> CDEBUG(D_WARNING, "Waiting for zombie LNI %s\n",
> libcfs_nid2str(ni->ni_nid));
> - }
> +
> schedule_timeout_uninterruptible(HZ);
> lnet_net_lock(LNET_LOCK_EX);
> continue;
> }
>
> + list_del_init(&ni->ni_list);
> +
> ni->ni_lnd->lnd_refcount--;
> lnet_net_unlock(LNET_LOCK_EX);
>
> @@ -1114,7 +1112,7 @@ lnet_clear_zombies_nis_locked(void)
> libcfs_nid2str(ni->ni_nid));
>
> lnet_ni_free(ni);
> - i = 2;
> + i = 1;
>
> lnet_net_lock(LNET_LOCK_EX);
> }
> @@ -1305,8 +1303,8 @@ lnet_startup_lndni(struct lnet_ni *ni, struct lnet_ioctl_config_data *conf)
> LASSERT(ni->ni_peertimeout <= 0 || lnd->lnd_query);
>
> lnet_net_lock(LNET_LOCK_EX);
> - /* refcount for ln_nis */
> - lnet_ni_addref_locked(ni, 0);
> + /* Initialise refcount for ln_nis to 1 */
> + percpu_ref_reinit(&ni->ni_refs);
> list_add_tail(&ni->ni_list, &the_lnet.ln_nis);
> if (ni->ni_cpts) {
> lnet_ni_addref_locked(ni, 0);
> diff --git a/drivers/staging/lustre/lnet/lnet/config.c b/drivers/staging/lustre/lnet/lnet/config.c
> index 091c4f714e84..4145c7431576 100644
> --- a/drivers/staging/lustre/lnet/lnet/config.c
> +++ b/drivers/staging/lustre/lnet/lnet/config.c
> @@ -96,8 +96,7 @@ lnet_ni_free(struct lnet_ni *ni)
> {
> int i;
>
> - if (ni->ni_refs)
> - cfs_percpt_free(ni->ni_refs);
> + percpu_ref_exit(&ni->ni_refs);
>
> if (ni->ni_tx_queues)
> cfs_percpt_free(ni->ni_tx_queues);
> @@ -117,6 +116,11 @@ lnet_ni_free(struct lnet_ni *ni)
> kfree(ni);
> }
>
> +static void ref_release(struct percpu_ref *ref)
> +{
> + wake_up_var(ref);
> +}
> +
> struct lnet_ni *
> lnet_ni_alloc(__u32 net, struct cfs_expr_list *el, struct list_head *nilist)
> {
> @@ -140,9 +144,8 @@ lnet_ni_alloc(__u32 net, struct cfs_expr_list *el, struct list_head *nilist)
>
> spin_lock_init(&ni->ni_lock);
> INIT_LIST_HEAD(&ni->ni_cptlist);
> - ni->ni_refs = cfs_percpt_alloc(lnet_cpt_table(),
> - sizeof(*ni->ni_refs[0]));
> - if (!ni->ni_refs)
> + if (percpu_ref_init(&ni->ni_refs, ref_release,
> + PERCPU_REF_INIT_DEAD, GFP_KERNEL) < 0)
> goto failed;
>
> ni->ni_tx_queues = cfs_percpt_alloc(lnet_cpt_table(),
> diff --git a/drivers/staging/lustre/lnet/lnet/router_proc.c b/drivers/staging/lustre/lnet/lnet/router_proc.c
> index d779445fefb5..8856798d263f 100644
> --- a/drivers/staging/lustre/lnet/lnet/router_proc.c
> +++ b/drivers/staging/lustre/lnet/lnet/router_proc.c
> @@ -703,7 +703,7 @@ static int proc_lnet_nis(struct ctl_table *table, int write,
> s += snprintf(s, tmpstr + tmpsiz - s,
> "%-24s %6s %5lld %4d %4d %4d %5d %5d %5d\n",
> libcfs_nid2str(ni->ni_nid), stat,
> - last_alive, *ni->ni_refs[i],
> + last_alive, 0/* No per-cpt refcount */,
> ni->ni_peertxcredits,
> ni->ni_peerrtrcredits,
> tq->tq_credits_max,
>
>
>
More information about the lustre-devel
mailing list