rcu: Allow calling rcu_(un)register_thread() during synchronize_rcu()

If rcu_(un)register_thread() is called together with synchronize_rcu(), it will wait for the synchronize_rcu() to finish. But when synchronize_rcu() waits for some events, we can modify the list registry.
We also use the lock rcu_gp_lock to assume that synchronize_rcu() isn't executed in more than one thread at the same time. Add a new mutex lock rcu_sync_lock to assume it and rename rcu_gp_lock to rcu_registry_lock. Release rcu_registry_lock when synchronize_rcu() waits for some events.

Signed-off-by Wen Congyang <[email protected]>

---
 util/rcu.c | 48 +++++++++++++++++++++++++++++++++++-------------
 1 file changed, 35 insertions(+), 13 deletions(-)

util/rcu.c +33 -12

index cdcad67..4dd33df 100644
@@ -47,7 +47,8 @@
 unsigned long rcu_gp_ctr = RCU_GP_LOCKED;
 
 QemuEvent rcu_gp_event;
-static QemuMutex rcu_gp_lock;
+static QemuMutex rcu_registry_lock;
+static QemuMutex rcu_sync_lock;
 
 /*
  * Check whether a quiescent state was crossed between the beginning of

@@ -66,7 +67,7 @@ static inline int rcu_gp_ongoing(unsigned long *ctr)
  */
 __thread struct rcu_reader_data rcu_reader;
 
-/* Protected by rcu_gp_lock.  */
+/* Protected by rcu_registry_lock.  */
 typedef QLIST_HEAD(, rcu_reader_data) ThreadList;
 static ThreadList registry = QLIST_HEAD_INITIALIZER(registry);
 

@@ -114,10 +115,26 @@ static void wait_for_readers(void)
             break;
         }
 
-        /* Wait for one thread to report a quiescent state and
-         * try again.
+        /* Wait for one thread to report a quiescent state and try again.
+         * Release rcu_registry_lock, so rcu_(un)register_thread() doesn't
+         * wait too much time.
+         *
+         * rcu_register_thread() may add nodes to &registry; it will not
+         * wake up synchronize_rcu, but that is okay because at least another
+         * thread must exit its RCU read-side critical section before
+         * synchronize_rcu is done.  The next iteration of the loop will
+         * move the new thread's rcu_reader from &registry to &qsreaders,
+         * because rcu_gp_ongoing() will return false.
+         *
+         * rcu_unregister_thread() may remove nodes from &qsreaders instead
+         * of &registry if it runs during qemu_event_wait.  That's okay;
+         * the node then will not be added back to &registry by QLIST_SWAP
+         * below. The invariant is that the node is part of one list when
+         * rcu_registry_lock is released.
          */
+        qemu_mutex_unlock(&rcu_registry_lock);
         qemu_event_wait(&rcu_gp_event);
+        qemu_mutex_lock(&rcu_registry_lock);
     }
 
     /* put back the reader list in the registry */

@@ -126,7 +143,8 @@ static void wait_for_readers(void)
 
 void synchronize_rcu(void)
 {
-    qemu_mutex_lock(&rcu_gp_lock);
+    qemu_mutex_lock(&rcu_sync_lock);
+    qemu_mutex_lock(&rcu_registry_lock);
 
     if (!QLIST_EMPTY(&registry)) {
         /* In either case, the atomic_mb_set below blocks stores that free

@@ -149,7 +167,8 @@ void synchronize_rcu(void)
         wait_for_readers();
     }
 
-    qemu_mutex_unlock(&rcu_gp_lock);
+    qemu_mutex_unlock(&rcu_registry_lock);
+    qemu_mutex_unlock(&rcu_sync_lock);
 }
 
 

@@ -273,23 +292,24 @@ void call_rcu1(struct rcu_head *node, void (*func)(struct rcu_head *node))
 void rcu_register_thread(void)
 {
     assert(rcu_reader.ctr == 0);
-    qemu_mutex_lock(&rcu_gp_lock);
+    qemu_mutex_lock(&rcu_registry_lock);
     QLIST_INSERT_HEAD(&registry, &rcu_reader, node);
-    qemu_mutex_unlock(&rcu_gp_lock);
+    qemu_mutex_unlock(&rcu_registry_lock);
 }
 
 void rcu_unregister_thread(void)
 {
-    qemu_mutex_lock(&rcu_gp_lock);
+    qemu_mutex_lock(&rcu_registry_lock);
     QLIST_REMOVE(&rcu_reader, node);
-    qemu_mutex_unlock(&rcu_gp_lock);
+    qemu_mutex_unlock(&rcu_registry_lock);
 }
 
 static void rcu_init_complete(void)
 {
     QemuThread thread;
 
-    qemu_mutex_init(&rcu_gp_lock);
+    qemu_mutex_init(&rcu_registry_lock);
+    qemu_mutex_init(&rcu_sync_lock);
     qemu_event_init(&rcu_gp_event, true);
 
     qemu_event_init(&rcu_call_ready_event, false);

@@ -306,12 +326,14 @@ static void rcu_init_complete(void)
 #ifdef CONFIG_POSIX
 static void rcu_init_lock(void)
 {
-    qemu_mutex_lock(&rcu_gp_lock);
+    qemu_mutex_lock(&rcu_sync_lock);
+    qemu_mutex_lock(&rcu_registry_lock);

Quoted 124 lines · Wen Congyang

If rcu_(un)register_thread() is called together with synchronize_rcu(), it will wait for the synchronize_rcu() to finish. But when synchronize_rcu() waits for some events, we can modify the list registry.
We also use the lock rcu_gp_lock to assume that synchronize_rcu() isn't executed in more than one thread at the same time. Add a new mutex lock rcu_sync_lock to assume it and rename rcu_gp_lock to rcu_registry_lock. Release rcu_registry_lock when synchronize_rcu() waits for some events.

Signed-off-by Wen Congyang <[email protected]>

---
 util/rcu.c | 48 +++++++++++++++++++++++++++++++++++-------------
 1 file changed, 35 insertions(+), 13 deletions(-)

util/rcu.c +33 -12

index cdcad67..4dd33df 100644
@@ -47,7 +47,8 @@
 unsigned long rcu_gp_ctr = RCU_GP_LOCKED;
 
 QemuEvent rcu_gp_event;
-static QemuMutex rcu_gp_lock;
+static QemuMutex rcu_registry_lock;
+static QemuMutex rcu_sync_lock;
 
 /*
  * Check whether a quiescent state was crossed between the beginning of

@@ -66,7 +67,7 @@ static inline int rcu_gp_ongoing(unsigned long *ctr)
  */
 __thread struct rcu_reader_data rcu_reader;
 
-/* Protected by rcu_gp_lock.  */
+/* Protected by rcu_registry_lock.  */
 typedef QLIST_HEAD(, rcu_reader_data) ThreadList;
 static ThreadList registry = QLIST_HEAD_INITIALIZER(registry);
 

@@ -114,10 +115,26 @@ static void wait_for_readers(void)
             break;
         }
 
-        /* Wait for one thread to report a quiescent state and
-         * try again.
+        /* Wait for one thread to report a quiescent state and try again.
+         * Release rcu_registry_lock, so rcu_(un)register_thread() doesn't
+         * wait too much time.
+         *
+         * rcu_register_thread() may add nodes to &registry; it will not
+         * wake up synchronize_rcu, but that is okay because at least another
+         * thread must exit its RCU read-side critical section before
+         * synchronize_rcu is done.  The next iteration of the loop will
+         * move the new thread's rcu_reader from &registry to &qsreaders,
+         * because rcu_gp_ongoing() will return false.
+         *
+         * rcu_unregister_thread() may remove nodes from &qsreaders instead
+         * of &registry if it runs during qemu_event_wait.  That's okay;
+         * the node then will not be added back to &registry by QLIST_SWAP
+         * below. The invariant is that the node is part of one list when
+         * rcu_registry_lock is released.
          */
+        qemu_mutex_unlock(&rcu_registry_lock);
         qemu_event_wait(&rcu_gp_event);
+        qemu_mutex_lock(&rcu_registry_lock);
     }
 
     /* put back the reader list in the registry */

@@ -126,7 +143,8 @@ static void wait_for_readers(void)
 
 void synchronize_rcu(void)
 {
-    qemu_mutex_lock(&rcu_gp_lock);
+    qemu_mutex_lock(&rcu_sync_lock);
+    qemu_mutex_lock(&rcu_registry_lock);
 
     if (!QLIST_EMPTY(&registry)) {
         /* In either case, the atomic_mb_set below blocks stores that free

@@ -149,7 +167,8 @@ void synchronize_rcu(void)
         wait_for_readers();
     }
 
-    qemu_mutex_unlock(&rcu_gp_lock);
+    qemu_mutex_unlock(&rcu_registry_lock);
+    qemu_mutex_unlock(&rcu_sync_lock);
 }
 
 

@@ -273,23 +292,24 @@ void call_rcu1(struct rcu_head *node, void (*func)(struct rcu_head *node))
 void rcu_register_thread(void)
 {
     assert(rcu_reader.ctr == 0);
-    qemu_mutex_lock(&rcu_gp_lock);
+    qemu_mutex_lock(&rcu_registry_lock);
     QLIST_INSERT_HEAD(&registry, &rcu_reader, node);
-    qemu_mutex_unlock(&rcu_gp_lock);
+    qemu_mutex_unlock(&rcu_registry_lock);
 }
 
 void rcu_unregister_thread(void)
 {
-    qemu_mutex_lock(&rcu_gp_lock);
+    qemu_mutex_lock(&rcu_registry_lock);
     QLIST_REMOVE(&rcu_reader, node);
-    qemu_mutex_unlock(&rcu_gp_lock);
+    qemu_mutex_unlock(&rcu_registry_lock);
 }
 
 static void rcu_init_complete(void)
 {
     QemuThread thread;
 
-    qemu_mutex_init(&rcu_gp_lock);
+    qemu_mutex_init(&rcu_registry_lock);
+    qemu_mutex_init(&rcu_sync_lock);
     qemu_event_init(&rcu_gp_event, true);
 
     qemu_event_init(&rcu_call_ready_event, false);

@@ -306,12 +326,14 @@ static void rcu_init_complete(void)
 #ifdef CONFIG_POSIX
 static void rcu_init_lock(void)
 {
-    qemu_mutex_lock(&rcu_gp_lock);
+    qemu_mutex_lock(&rcu_sync_lock);
+    qemu_mutex_lock(&rcu_registry_lock);
 }
 

util/rcu.c static void rcu_init_complete(void) · line 339

Quoted 9 lines

util/rcu.c +2 -1

 static void rcu_init_unlock(void)
 {
-    qemu_mutex_unlock(&rcu_gp_lock);
+    qemu_mutex_unlock(&rcu_registry_lock);
+    qemu_mutex_unlock(&rcu_sync_lock);
 }
 #endif
 

Ok, thanks.

Paolo
.

I have a question about rcu: while do we call wait_for_readers() twice for 32-bit host?

Quoted 2 lines · Wen Congyang

I have a question about rcu: while do we call wait_for_readers() twice for 32-bit host?

Because there is a very small but non-zero probability of the counter going up by exactly 2^31 periods (periods are stored in bits 1-31 so you lose one bit) while the thread is sleeping. This detail of the implementation comes from URCU.

Paolo

Quoted 10 lines · Paolo Bonzini

Quoted 2 lines · Wen Congyang

I have a question about rcu: while do we call wait_for_readers() twice for 32-bit host?

Because there is a very small but non-zero probability of the counter going up by exactly 2^31 periods (periods are stored in bits 1-31 so you lose one bit) while the thread is sleeping. This detail of the implementation comes from URCU.

Yes, so you use rcu_gp_ctr ^ RCU_GP_CTR to instead of rcu_gp_ctr + RCU_GP_CTR. The initial value is 1, so rcu_gp_ctr is: 1, 3, 1, 3, ...
The rcu_gp_ctr will never be 0. I think calling wait_for_readers() once is enough.

Do I miss something?

Quoted 17 lines · Paolo Bonzini

Quoted 10 lines · Wen Congyang

Quoted 2 lines

I have a question about rcu: while do we call wait_for_readers() twice for 32-bit host?

Because there is a very small but non-zero probability of the counter going up by exactly 2^31 periods (periods are stored in bits 1-31 so you lose one bit) while the thread is sleeping. This detail of the implementation comes from URCU.

Yes, so you use rcu_gp_ctr ^ RCU_GP_CTR to instead of rcu_gp_ctr + RCU_GP_CTR. The initial value is 1, so rcu_gp_ctr is: 1, 3, 1, 3, ...
The rcu_gp_ctr will never be 0. I think calling wait_for_readers() once is enough.

Quoted 1 line

Do I miss something?

If you call it just once, you have the same problem as before. In fact, it's worse because instead of having an overflow every 2^31 periods, you have one every 2 periods. Instead, by checking that rcu_reader went through 1 _and_ 3 (or that it was at least once 0, i.e. the thread was quiescent), you are sure that the thread went through _at least one_ grace period.

Paolo

Quoted 8 lines

Quoted 1 line

If you call it just once, you have the same problem as before. In fact, it's worse because instead of having an overflow every 2^31 periods, you have one every 2 periods. Instead, by checking that rcu_reader went through 1 _and_ 3 (or that it was at least once 0, i.e. the thread was quiescent), you are sure that the thread went through _at least one_ grace period.

The overflow is acceptable. We only compare if rcu_reader.ctr is equal than rcu_gp_ctr. If not, we should wait that thread to call rcu_read_unlock(). We don't care which is bigger. If no threads calls sync_rcu(), all threads rcu_read.ctr is 0 or rcu_gp_ctr. If one thread calls sync_rcu(), all threads rcu_read.ctr is 0, old_rcu_gp_ctr, or new_rcu_gp_ctr. We only wait the thread that's rcu_read.ctr is old_rcu_gp_ctr.

Quoted 9 lines · Wen Congyang

Quoted 8 lines

Quoted 1 line

If you call it just once, you have the same problem as before. In fact, it's worse because instead of having an overflow every 2^31 periods, you have one every 2 periods. Instead, by checking that rcu_reader went through 1 _and_ 3 (or that it was at least once 0, i.e. the thread was quiescent), you are sure that the thread went through _at least one_ grace period.

Quoted 6 lines

The overflow is acceptable. We only compare if rcu_reader.ctr is equal than rcu_gp_ctr. If not, we should wait that thread to call rcu_read_unlock(). We don't care which is bigger. If no threads calls sync_rcu(), all threads rcu_read.ctr is 0 or rcu_gp_ctr. If one thread calls sync_rcu(), all threads rcu_read.ctr is 0, old_rcu_gp_ctr, or new_rcu_gp_ctr. We only wait the thread that's rcu_read.ctr is old_rcu_gp_ctr.

Suppose you have

reader writer
---------------------------------------------------------------

    rcu_read_lock()
      ctr = atomic_read(&rcu_gp_ctr);
        ctr = 1
                                                  synchronize_rcu()
                                                     rcu_gp_ctr = 3
      atomic_xchg(&p_rcu_reader->ctr, ctr);
        rcu_reader.ctr = 1

We're in the critical section.

                                                  synchronize_rcu()
                                                     rcu_gp_ctr = 1

rcu_gp_ongoing(&p_rcu_reader->ctr) returns false, so synchronize_rcu() exits. But we're still in the critical section.

    p = atomic_rcu_read(&foo);
                                                  g_free(p);

Boom. :) For more information see
http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.220.8810

Paolo