urcu-mb/signal/membarrier: improve 2-phase wait scheme

In the single-bit, 2-phase grace period scheme, all we need to do is to
observe each reader going through a quiescent state while we are in the
grace period.

We therefore only need to perform one global counter update, surrounded
by 2 iterations on readers to observe change in their snapshot.

We can therefore remove the first counter update (prior to the first
iteration on readers): it was useless and was only slowing down the
grace period.

CC: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
CC: Lai Jiangshan <laijs@cn.fujitsu.com>
CC: Alan Stern <stern@rowland.harvard.edu>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

diff --git a/urcu.c b/urcu.c
index c42184656a3a0381b518aa15372b84a365400dea..2d5c510aa6744418c5557d642b451e0afae415cd 100644 (file)
--- a/urcu.c
+++ b/urcu.c
@@ -215,33 +215,16 @@ static void wait_gp(void)
                      NULL, NULL, 0);
 }
 
-static void update_counter_and_wait(void)
+static void wait_for_readers(void)
 {
        CDS_LIST_HEAD(qsreaders);
        int wait_loops = 0;
        struct rcu_reader *index, *tmp;
 
-       /* Switch parity: 0 -> 1, 1 -> 0 */
-       CMM_STORE_SHARED(rcu_gp_ctr, rcu_gp_ctr ^ RCU_GP_CTR_PHASE);
-
-       /*
-        * Must commit rcu_gp_ctr update to memory before waiting for quiescent
-        * state. Failure to do so could result in the writer waiting forever
-        * while new readers are always accessing data (no progress). Enforce
-        * compiler-order of store to rcu_gp_ctr before load rcu_reader ctr.
-        */
-       cmm_barrier();
-
-       /*
-        *
-        * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
-        * model easier to understand. It does not have a big performance impact
-        * anyway, given this is the write-side.
-        */
-       cmm_smp_mb();
-
        /*
-        * Wait for each thread URCU_TLS(rcu_reader).ctr count to become 0.
+        * Wait for each thread URCU_TLS(rcu_reader).ctr to either
+        * indicate quiescence (not nested), or observe the current
+        * rcu_gp_ctr value.
         */
        for (;;) {
                wait_loops++;
@@ -316,12 +299,12 @@ void synchronize_rcu(void)
        smp_mb_master(RCU_MB_GROUP);
 
        /*
-        * Wait for previous parity to be empty of readers.
+        * Wait for readers to observe original parity or be quiescent.
         */
-       update_counter_and_wait();      /* 0 -> 1, wait readers in parity 0 */
+       wait_for_readers();
 
        /*
-        * Must finish waiting for quiescent state for parity 0 before
+        * Must finish waiting for quiescent state for original parity before
         * committing next rcu_gp_ctr update to memory. Failure to do so could
         * result in the writer waiting forever while new readers are always
         * accessing data (no progress).  Enforce compiler-order of load
@@ -336,10 +319,29 @@ void synchronize_rcu(void)
         */
        cmm_smp_mb();
 
+       /* Switch parity: 0 -> 1, 1 -> 0 */
+       CMM_STORE_SHARED(rcu_gp_ctr, rcu_gp_ctr ^ RCU_GP_CTR_PHASE);
+
+       /*
+        * Must commit rcu_gp_ctr update to memory before waiting for quiescent
+        * state. Failure to do so could result in the writer waiting forever
+        * while new readers are always accessing data (no progress). Enforce
+        * compiler-order of store to rcu_gp_ctr before load rcu_reader ctr.
+        */
+       cmm_barrier();
+
+       /*
+        *
+        * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
+        * model easier to understand. It does not have a big performance impact
+        * anyway, given this is the write-side.
+        */
+       cmm_smp_mb();
+
        /*
-        * Wait for previous parity to be empty of readers.
+        * Wait for readers to observe new parity or be quiescent.
         */
-       update_counter_and_wait();      /* 1 -> 0, wait readers in parity 1 */
+       wait_for_readers();
 
        /* Finish waiting for reader threads before letting the old ptr being
         * freed. Must be done within rcu_gp_lock because it iterates on reader