RE: [RFC PATCH] net/tipc: reduce tipc_node lock holding time in tipc_rcv

xu posted 1 patch 2 years ago
RE: [RFC PATCH] net/tipc: reduce tipc_node lock holding time in tipc_rcv
Posted by xu 2 years ago
>>Happy to see your reply. But Why? 'delete' is false from tipc_node_timeout(). Refer to:
>>https://elixir.bootlin.com/linux/v6.7-rc2/source/net/tipc/node.c#L844
>I should have explained it clearly:
>1/ link status must be protected.
>tipc_node_timeout()
>   tipc_node_link_down()
>   {
>      struct tipc_link *l = le->link;
>   
>      ...
>     __tipc_node_link_down(); <-- link status is referred. 
>      ...
>     if (delete) {
>        kfree(l);
>        le->link = NULL;
>     }
>     ...
>   }
>
>__tipc_node_link_down()
>{
>    ...
>   if (!l || tipc_link_is_reset(l)) <-- read link status
>    ...
>    tipc_link_reset(l); <--- this function will reset all things related to link.
>}
>
>2/ le->link must be protected.
>bearer_disable()
>{
>   ...
>   tipc_node_delete_links(net, bearer_id); <--- this will delete all links.
>   ...
>}
>
>tipc_node_delete_links()
>{
>   ...
>  tipc_node_link_down(n, bearer_id, true);
>   ...
>}

Could we please solve the problem mentioned above by adding spinlock(&le->lock)?

For example:

(BTW, I have tested it, with this change, enabling RPS based on tipc port can improve 25% of general throughput)

diff --git a/net/tipc/node.c b/net/tipc/node.c
index 3105abe97bb9..470c272d798e 100644
--- a/net/tipc/node.c
+++ b/net/tipc/node.c
@@ -1079,12 +1079,16 @@ static void tipc_node_link_down(struct tipc_node *n, int bearer_id, bool delete)
                __tipc_node_link_down(n, &bearer_id, &xmitq, &maddr);
        } else {
                /* Defuse pending tipc_node_link_up() */
+               spin_lock_bh(&le->lock);
                tipc_link_reset(l);
+               spin_unlock_bh(&le->lock);
                tipc_link_fsm_evt(l, LINK_RESET_EVT);
        }
        if (delete) {
+               spin_lock_bh(&le->lock);
                kfree(l);
                le->link = NULL;
+               spin_unlock_bh(&le->lock);
                n->link_cnt--;
        }
        trace_tipc_node_link_down(n, true, "node link down or deleted!");
@@ -2154,14 +2158,15 @@ void tipc_rcv(struct net *net, struct sk_buff *skb, struct tipc_bearer *b)
        /* Receive packet directly if conditions permit */
        tipc_node_read_lock(n);
        if (likely((n->state == SELF_UP_PEER_UP) && (usr != TUNNEL_PROTOCOL))) {
+               tipc_node_read_unlock(n);
                spin_lock_bh(&le->lock);
                if (le->link) {
                        rc = tipc_link_rcv(le->link, skb, &xmitq);
                        skb = NULL;
                }
                spin_unlock_bh(&le->lock);
-       }
-       tipc_node_read_unlock(n);
+       } else
+               tipc_node_read_unlock(n);
 
        /* Check/update node state before receiving */
        if (unlikely(skb)) {
@@ -2169,12 +2174,13 @@ void tipc_rcv(struct net *net, struct sk_buff *skb, struct tipc_bearer *b)
                        goto out_node_put;
                tipc_node_write_lock(n);
                if (tipc_node_check_state(n, skb, bearer_id, &xmitq)) {
+                       tipc_node_write_unlock(n);
                        if (le->link) {
                                rc = tipc_link_rcv(le->link, skb, &xmitq);
                                skb = NULL;
                        }
-               }
-               tipc_node_write_unlock(n);
+               } else
+                       tipc_node_write_unlock(n);
        }
 
        if (unlikely(rc & TIPC_LINK_UP_EVT))
RE: [RFC PATCH] net/tipc: reduce tipc_node lock holding time in tipc_rcv
Posted by Tung Quang Nguyen 2 years ago
>Could we please solve the problem mentioned above by adding spinlock(&le->lock)?
>

No, you cannot do that. As I said before, the link status (including l->state) needs to be protected by node lock.
What I showed you were just 2 use cases (link reset/delete). There are more use cases (netlink, transmit path etc) that need proper locks.

>For example:
>
>(BTW, I have tested it, with this change, enabling RPS based on tipc port can improve 25% of general throughput)
>
>diff --git a/net/tipc/node.c b/net/tipc/node.c index 3105abe97bb9..470c272d798e 100644
>--- a/net/tipc/node.c
>+++ b/net/tipc/node.c
>@@ -1079,12 +1079,16 @@ static void tipc_node_link_down(struct tipc_node *n, int bearer_id, bool delete)
>                __tipc_node_link_down(n, &bearer_id, &xmitq, &maddr);
>        } else {
>                /* Defuse pending tipc_node_link_up() */
>+               spin_lock_bh(&le->lock);
>                tipc_link_reset(l);
>+               spin_unlock_bh(&le->lock);
>                tipc_link_fsm_evt(l, LINK_RESET_EVT);
>        }
>        if (delete) {
>+               spin_lock_bh(&le->lock);
>                kfree(l);
>                le->link = NULL;
>+               spin_unlock_bh(&le->lock);
>                n->link_cnt--;
>        }
>        trace_tipc_node_link_down(n, true, "node link down or deleted!"); @@ -2154,14 +2158,15 @@ void tipc_rcv(struct net *net,
>struct sk_buff *skb, struct tipc_bearer *b)
>        /* Receive packet directly if conditions permit */
>        tipc_node_read_lock(n);
>        if (likely((n->state == SELF_UP_PEER_UP) && (usr != TUNNEL_PROTOCOL))) {
>+               tipc_node_read_unlock(n);
>                spin_lock_bh(&le->lock);
>                if (le->link) {
>                        rc = tipc_link_rcv(le->link, skb, &xmitq);
>                        skb = NULL;
>                }
>                spin_unlock_bh(&le->lock);
>-       }
>-       tipc_node_read_unlock(n);
>+       } else
>+               tipc_node_read_unlock(n);
>
>        /* Check/update node state before receiving */
>        if (unlikely(skb)) {
>@@ -2169,12 +2174,13 @@ void tipc_rcv(struct net *net, struct sk_buff *skb, struct tipc_bearer *b)
>                        goto out_node_put;
>                tipc_node_write_lock(n);
>                if (tipc_node_check_state(n, skb, bearer_id, &xmitq)) {
>+                       tipc_node_write_unlock(n);
>                        if (le->link) {
>                                rc = tipc_link_rcv(le->link, skb, &xmitq);
>                                skb = NULL;
>                        }
>-               }
>-               tipc_node_write_unlock(n);
>+               } else
>+                       tipc_node_write_unlock(n);
>        }
>
>        if (unlikely(rc & TIPC_LINK_UP_EVT))
RE: [RFC PATCH] net/tipc: reduce tipc_node lock holding time in tipc_rcv
Posted by xu 2 years ago
>>Could we please solve the problem mentioned above by adding spinlock(&le->lock)?
>>
>
>No, you cannot do that. As I said before, the link status (including l->state) needs to be protected by node lock.

Why can't use le->lock instead of node's lock to protect it in tipc_link_rcv.

>What I showed you were just 2 use cases (link reset/delete). There are more use cases (netlink, transmit path etc) that need proper locks.

The same. We can also add spin_lock_bh(&le->lock) to protect the link in other places where it changes the
link status in addition to 'reset/delete'. Because using node lock to protect the link in tipc_link_rcv is
really wasting CPU performance.

>
>>For example:
>>
>>(BTW, I have tested it, with this change, enabling RPS based on tipc port can improve 25% of general throughput)
>>
>>diff --git a/net/tipc/node.c b/net/tipc/node.c index 3105abe97bb9..470c272d798e 100644
>>--- a/net/tipc/node.c
>>+++ b/net/tipc/node.c
>>@@ -1079,12 +1079,16 @@ static void tipc_node_link_down(struct tipc_node *n, int bearer_id, bool delete)
>>                __tipc_node_link_down(n, &bearer_id, &xmitq, &maddr);
>>        } else {
>>                /* Defuse pending tipc_node_link_up() */
>>+               spin_lock_bh(&le->lock);
>>                tipc_link_reset(l);
>>+               spin_unlock_bh(&le->lock);
>>                tipc_link_fsm_evt(l, LINK_RESET_EVT);
>>        }
>>        if (delete) {
>>+               spin_lock_bh(&le->lock);
>>                kfree(l);
>>                le->link = NULL;
>>+               spin_unlock_bh(&le->lock);
>>                n->link_cnt--;
>>        }
>>        trace_tipc_node_link_down(n, true, "node link down or deleted!"); @@ -2154,14 +2158,15 @@ void tipc_rcv(struct net *net,
>>struct sk_buff *skb, struct tipc_bearer *b)
>>        /* Receive packet directly if conditions permit */
>>        tipc_node_read_lock(n);
>>        if (likely((n->state == SELF_UP_PEER_UP) && (usr != TUNNEL_PROTOCOL))) {
>>+               tipc_node_read_unlock(n);
>>                spin_lock_bh(&le->lock);
>>                if (le->link) {
>>                        rc = tipc_link_rcv(le->link, skb, &xmitq);
>>                        skb = NULL;
>>                }
>>                spin_unlock_bh(&le->lock);
>>-       }
>>-       tipc_node_read_unlock(n);
>>+       } else
>>+               tipc_node_read_unlock(n);
>>
>>        /* Check/update node state before receiving */
>>        if (unlikely(skb)) {
>>@@ -2169,12 +2174,13 @@ void tipc_rcv(struct net *net, struct sk_buff *skb, struct tipc_bearer *b)
>>                        goto out_node_put;
>>                tipc_node_write_lock(n);
>>                if (tipc_node_check_state(n, skb, bearer_id, &xmitq)) {
>>+                       tipc_node_write_unlock(n);
>>                        if (le->link) {
>>                                rc = tipc_link_rcv(le->link, skb, &xmitq);
>>                                skb = NULL;
>>                        }
>>-               }
>>-               tipc_node_write_unlock(n);
>>+               } else
>>+                       tipc_node_write_unlock(n);
>>        }
>>
>>        if (unlikely(rc & TIPC_LINK_UP_EVT))
RE: [RFC PATCH] net/tipc: reduce tipc_node lock holding time in tipc_rcv
Posted by Tung Quang Nguyen 2 years ago
>Why can't use le->lock instead of node's lock to protect it in tipc_link_rcv.
>
I have already explained:
__tipc_node_link_down()
{
    ...
   if (!l || tipc_link_is_reset(l)) <-- read link status
    ...
}
>>What I showed you were just 2 use cases (link reset/delete). There are more use cases (netlink, transmit path etc) that need proper
>locks.
>
>The same. We can also add spin_lock_bh(&le->lock) to protect the link in other places where it changes the link status in addition to
>'reset/delete'. Because using node lock to protect the link in tipc_link_rcv is really wasting CPU performance.
>
If you want to change current lock policy, you need to submit a complete/correct patch. I will acknowledge this patch if I can see a significant improvement in my test.