LVS
lvs-devel
Google
 
Web LinuxVirtualServer.org

[PATCH 28/34] ipvs: convert wrr scheduler to rcu

To: Pablo Neira Ayuso <pablo@xxxxxxxxxxxxx>
Subject: [PATCH 28/34] ipvs: convert wrr scheduler to rcu
Cc: lvs-devel@xxxxxxxxxxxxxxx, netdev@xxxxxxxxxxxxxxx, netfilter-devel@xxxxxxxxxxxxxxx, Wensong Zhang <wensong@xxxxxxxxxxxx>, Julian Anastasov <ja@xxxxxx>, Simon Horman <horms@xxxxxxxxxxxx>
From: Simon Horman <horms@xxxxxxxxxxxx>
Date: Fri, 29 Mar 2013 13:11:45 +0900
From: Julian Anastasov <ja@xxxxxx>

The schedule method now needs _rcu list-traversal
primitive for svc->destinations. As the weight for some
dest can be reduced during dest selection, change the
algorithm to check weights by using minimum weights in the
1 .. max_weight-(di-1) range, with the same step (di). By this
way we ensure that there will be always a weight >= 1 check
before claiming that all destinations are overloaded.

Signed-off-by: Julian Anastasov <ja@xxxxxx>
Signed-off-by: Simon Horman <horms@xxxxxxxxxxxx>
---
 net/netfilter/ipvs/ip_vs_wrr.c |  167 +++++++++++++++++++++++++---------------
 1 file changed, 103 insertions(+), 64 deletions(-)

diff --git a/net/netfilter/ipvs/ip_vs_wrr.c b/net/netfilter/ipvs/ip_vs_wrr.c
index 231be7d..98cb05e 100644
--- a/net/netfilter/ipvs/ip_vs_wrr.c
+++ b/net/netfilter/ipvs/ip_vs_wrr.c
@@ -29,14 +29,45 @@
 
 #include <net/ip_vs.h>
 
+/* The WRR algorithm depends on some caclulations:
+ * - mw: maximum weight
+ * - di: weight step, greatest common divisor from all weights
+ * - cw: current required weight
+ * As result, all weights are in the [di..mw] range with a step=di.
+ *
+ * First, we start with cw = mw and select dests with weight >= cw.
+ * Then cw is reduced with di and all dests are checked again.
+ * Last pass should be with cw = di. We have mw/di passes in total:
+ *
+ * pass 1: cw = max weight
+ * pass 2: cw = max weight - di
+ * pass 3: cw = max weight - 2 * di
+ * ...
+ * last pass: cw = di
+ *
+ * Weights are supposed to be >= di but we run in parallel with
+ * weight changes, it is possible some dest weight to be reduced
+ * below di, bad if it is the only available dest.
+ *
+ * So, we modify how mw is calculated, now it is reduced with (di - 1),
+ * so that last cw is 1 to catch such dests with weight below di:
+ * pass 1: cw = max weight - (di - 1)
+ * pass 2: cw = max weight - di - (di - 1)
+ * pass 3: cw = max weight - 2 * di - (di - 1)
+ * ...
+ * last pass: cw = 1
+ *
+ */
+
 /*
  * current destination pointer for weighted round-robin scheduling
  */
 struct ip_vs_wrr_mark {
-       struct list_head *cl;   /* current list head */
+       struct ip_vs_dest *cl;  /* current dest or head */
        int cw;                 /* current weight */
        int mw;                 /* maximum weight */
        int di;                 /* decreasing interval */
+       struct rcu_head         rcu_head;
 };
 
 
@@ -88,10 +119,10 @@ static int ip_vs_wrr_init_svc(struct ip_vs_service *svc)
        if (mark == NULL)
                return -ENOMEM;
 
-       mark->cl = &svc->destinations;
-       mark->cw = 0;
-       mark->mw = ip_vs_wrr_max_weight(svc);
+       mark->cl = list_entry(&svc->destinations, struct ip_vs_dest, n_list);
        mark->di = ip_vs_wrr_gcd_weight(svc);
+       mark->mw = ip_vs_wrr_max_weight(svc) - (mark->di - 1);
+       mark->cw = mark->mw;
        svc->sched_data = mark;
 
        return 0;
@@ -100,24 +131,31 @@ static int ip_vs_wrr_init_svc(struct ip_vs_service *svc)
 
 static int ip_vs_wrr_done_svc(struct ip_vs_service *svc)
 {
+       struct ip_vs_wrr_mark *mark = svc->sched_data;
+
        /*
         *    Release the mark variable
         */
-       kfree(svc->sched_data);
+       kfree_rcu(mark, rcu_head);
 
        return 0;
 }
 
 
-static int ip_vs_wrr_update_svc(struct ip_vs_service *svc)
+static int ip_vs_wrr_dest_changed(struct ip_vs_service *svc,
+                                 struct ip_vs_dest *dest)
 {
        struct ip_vs_wrr_mark *mark = svc->sched_data;
 
-       mark->cl = &svc->destinations;
-       mark->mw = ip_vs_wrr_max_weight(svc);
+       write_lock_bh(&svc->sched_lock);
+       mark->cl = list_entry(&svc->destinations, struct ip_vs_dest, n_list);
        mark->di = ip_vs_wrr_gcd_weight(svc);
-       if (mark->cw > mark->mw)
-               mark->cw = 0;
+       mark->mw = ip_vs_wrr_max_weight(svc) - (mark->di - 1);
+       if (mark->cw > mark->mw || !mark->cw)
+               mark->cw = mark->mw;
+       else if (mark->di > 1)
+               mark->cw = (mark->cw / mark->di) * mark->di + 1;
+       write_unlock_bh(&svc->sched_lock);
        return 0;
 }
 
@@ -128,80 +166,79 @@ static int ip_vs_wrr_update_svc(struct ip_vs_service *svc)
 static struct ip_vs_dest *
 ip_vs_wrr_schedule(struct ip_vs_service *svc, const struct sk_buff *skb)
 {
-       struct ip_vs_dest *dest;
+       struct ip_vs_dest *dest, *last, *stop = NULL;
        struct ip_vs_wrr_mark *mark = svc->sched_data;
-       struct list_head *p;
+       bool last_pass = false, restarted = false;
 
        IP_VS_DBG(6, "%s(): Scheduling...\n", __func__);
 
-       /*
-        * This loop will always terminate, because mark->cw in (0, max_weight]
-        * and at least one server has its weight equal to max_weight.
-        */
        write_lock(&svc->sched_lock);
-       p = mark->cl;
+       dest = mark->cl;
+       /* No available dests? */
+       if (mark->mw == 0)
+               goto err_noavail;
+       last = dest;
+       /* Stop only after all dests were checked for weight >= 1 (last pass) */
        while (1) {
-               if (mark->cl == &svc->destinations) {
-                       /* it is at the head of the destination list */
-
-                       if (mark->cl == mark->cl->next) {
-                               /* no dest entry */
-                               ip_vs_scheduler_err(svc,
-                                       "no destination available: "
-                                       "no destinations present");
-                               dest = NULL;
-                               goto out;
-                       }
-
-                       mark->cl = svc->destinations.next;
-                       mark->cw -= mark->di;
-                       if (mark->cw <= 0) {
-                               mark->cw = mark->mw;
-                               /*
-                                * Still zero, which means no available servers.
-                                */
-                               if (mark->cw == 0) {
-                                       mark->cl = &svc->destinations;
-                                       ip_vs_scheduler_err(svc,
-                                               "no destination available");
-                                       dest = NULL;
-                                       goto out;
-                               }
-                       }
-               } else
-                       mark->cl = mark->cl->next;
-
-               if (mark->cl != &svc->destinations) {
-                       /* not at the head of the list */
-                       dest = list_entry(mark->cl, struct ip_vs_dest, n_list);
+               list_for_each_entry_continue_rcu(dest,
+                                                &svc->destinations,
+                                                n_list) {
                        if (!(dest->flags & IP_VS_DEST_F_OVERLOAD) &&
-                           atomic_read(&dest->weight) >= mark->cw) {
-                               /* got it */
-                               break;
-                       }
+                           atomic_read(&dest->weight) >= mark->cw)
+                               goto found;
+                       if (dest == stop)
+                               goto err_over;
                }
-
-               if (mark->cl == p && mark->cw == mark->di) {
-                       /* back to the start, and no dest is found.
-                          It is only possible when all dests are OVERLOADED */
-                       dest = NULL;
-                       ip_vs_scheduler_err(svc,
-                               "no destination available: "
-                               "all destinations are overloaded");
-                       goto out;
+               mark->cw -= mark->di;
+               if (mark->cw <= 0) {
+                       mark->cw = mark->mw;
+                       /* Stop if we tried last pass from first dest:
+                        * 1. last_pass: we started checks when cw > di but
+                        *      then all dests were checked for w >= 1
+                        * 2. last was head: the first and only traversal
+                        *      was for weight >= 1, for all dests.
+                        */
+                       if (last_pass ||
+                           &last->n_list == &svc->destinations)
+                               goto err_over;
+                       restarted = true;
+               }
+               last_pass = mark->cw <= mark->di;
+               if (last_pass && restarted &&
+                   &last->n_list != &svc->destinations) {
+                       /* First traversal was for w >= 1 but only
+                        * for dests after 'last', now do the same
+                        * for all dests up to 'last'.
+                        */
+                       stop = last;
                }
        }
 
+found:
        IP_VS_DBG_BUF(6, "WRR: server %s:%u "
                      "activeconns %d refcnt %d weight %d\n",
                      IP_VS_DBG_ADDR(svc->af, &dest->addr), ntohs(dest->port),
                      atomic_read(&dest->activeconns),
                      atomic_read(&dest->refcnt),
                      atomic_read(&dest->weight));
+       mark->cl = dest;
 
   out:
        write_unlock(&svc->sched_lock);
        return dest;
+
+err_noavail:
+       mark->cl = dest;
+       dest = NULL;
+       ip_vs_scheduler_err(svc, "no destination available");
+       goto out;
+
+err_over:
+       mark->cl = dest;
+       dest = NULL;
+       ip_vs_scheduler_err(svc, "no destination available: "
+                           "all destinations are overloaded");
+       goto out;
 }
 
 
@@ -212,7 +249,9 @@ static struct ip_vs_scheduler ip_vs_wrr_scheduler = {
        .n_list =               LIST_HEAD_INIT(ip_vs_wrr_scheduler.n_list),
        .init_service =         ip_vs_wrr_init_svc,
        .done_service =         ip_vs_wrr_done_svc,
-       .update_service =       ip_vs_wrr_update_svc,
+       .add_dest =             ip_vs_wrr_dest_changed,
+       .del_dest =             ip_vs_wrr_dest_changed,
+       .upd_dest =             ip_vs_wrr_dest_changed,
        .schedule =             ip_vs_wrr_schedule,
 };
 
-- 
1.7.10.4

--
To unsubscribe from this list: send the line "unsubscribe lvs-devel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html

<Prev in Thread] Current Thread [Next in Thread>