Support kernels with broken signal delivery
authorPaul E. McKenney <paulmck@linux.vnet.ibm.com>
Sat, 18 Apr 2009 06:43:32 +0000 (02:43 -0400)
committerMathieu Desnoyers <mathieu.desnoyers@polymtl.ca>
Sat, 18 Apr 2009 06:43:32 +0000 (02:43 -0400)
Some Linux kernels seems to have problems sending _all_ the signals. Use a
workaround which re-sends the signal when we wait for it to be executed.

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@polymtl.ca>
urcu.c

diff --git a/urcu.c b/urcu.c
index 475b4d1c991a00832409ae7332e6317442b1f042..38abf7d6a92e045bb2aabbc306a4b7006ed108f8 100644 (file)
--- a/urcu.c
+++ b/urcu.c
@@ -14,6 +14,7 @@
 #include <assert.h>
 #include <stdlib.h>
 #include <string.h>
+#include <errno.h>
 
 #include "urcu.h"
 
@@ -39,6 +40,7 @@ long __thread urcu_active_readers;
 struct reader_registry {
        pthread_t tid;
        long *urcu_active_readers;
+       char *need_mb;
 };
 
 #ifdef DEBUG_YIELD
@@ -47,24 +49,34 @@ unsigned int __thread rand_yield;
 #endif
 
 static struct reader_registry *registry;
+static char __thread need_mb;
 static int num_readers, alloc_readers;
-#ifndef DEBUG_FULL_MB
-static int sig_done;
-#endif
 
 void internal_urcu_lock(void)
 {
-#if 0
        int ret;
-       /* Mutex sleeping does not play well with busy-waiting loop. */
+
+#ifndef DISTRUST_SIGNALS_EXTREME
        ret = pthread_mutex_lock(&urcu_mutex);
        if (ret) {
                perror("Error in pthread mutex lock");
                exit(-1);
        }
-#endif
-       while (pthread_mutex_trylock(&urcu_mutex) != 0)
-               cpu_relax();
+#else /* #ifndef DISTRUST_SIGNALS_EXTREME */
+       while ((ret = pthread_mutex_trylock(&urcu_mutex)) != 0) {
+               if (ret != EBUSY && ret != EINTR) {
+                       printf("ret = %d, errno = %d\n", ret, errno);
+                       perror("Error in pthread mutex lock");
+                       exit(-1);
+               }
+               if (need_mb) {
+                       smp_mb();
+                       need_mb = 0;
+                       smp_mb();
+               }
+               poll(NULL,0,10);
+       }
+#endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
 }
 
 void internal_urcu_unlock(void)
@@ -87,7 +99,7 @@ static void switch_next_urcu_qparity(void)
 }
 
 #ifdef DEBUG_FULL_MB
-static void force_mb_single_thread(pthread_t tid)
+static void force_mb_single_thread(struct reader_registry *index)
 {
        smp_mb();
 }
@@ -98,26 +110,27 @@ static void force_mb_all_threads(void)
 }
 #else
 
-static void force_mb_single_thread(pthread_t tid)
+static void force_mb_single_thread(struct reader_registry *index)
 {
        assert(registry);
-       sig_done = 0;
        /*
         * pthread_kill has a smp_mb(). But beware, we assume it performs
         * a cache flush on architectures with non-coherent cache. Let's play
         * safe and don't assume anything : we use smp_mc() to make sure the
         * cache flush is enforced.
-        * smp_mb();    write sig_done before sending the signals
         */
-       smp_mc();       /* write sig_done before sending the signals */
-       pthread_kill(tid, SIGURCU);
+       *index->need_mb = 1;
+       smp_mc();       /* write ->need_mb before sending the signals */
+       pthread_kill(index->tid, SIGURCU);
+       smp_mb();
        /*
         * Wait for sighandler (and thus mb()) to execute on every thread.
         * BUSY-LOOP.
         */
-       while (LOAD_SHARED(sig_done) < 1)
-               cpu_relax();
-       smp_mb();       /* read sig_done before ending the barrier */
+       while (*index->need_mb) {
+               poll(NULL, 0, 1);
+       }
+       smp_mb();       /* read ->need_mb before ending the barrier */
 }
 
 static void force_mb_all_threads(void)
@@ -129,24 +142,37 @@ static void force_mb_all_threads(void)
         */
        if (!registry)
                return;
-       sig_done = 0;
        /*
         * pthread_kill has a smp_mb(). But beware, we assume it performs
         * a cache flush on architectures with non-coherent cache. Let's play
         * safe and don't assume anything : we use smp_mc() to make sure the
         * cache flush is enforced.
-        * smp_mb();    write sig_done before sending the signals
         */
-       smp_mc();       /* write sig_done before sending the signals */
-       for (index = registry; index < registry + num_readers; index++)
+       for (index = registry; index < registry + num_readers; index++) {
+               *index->need_mb = 1;
+               smp_mc();       /* write need_mb before sending the signal */
                pthread_kill(index->tid, SIGURCU);
+       }
        /*
         * Wait for sighandler (and thus mb()) to execute on every thread.
-        * BUSY-LOOP.
+        *
+        * Note that the pthread_kill() will never be executed on systems
+        * that correctly deliver signals in a timely manner.  However, it
+        * is not uncommon for kernels to have bugs that can result in
+        * lost or unduly delayed signals.
+        *
+        * If you are seeing the below pthread_kill() executing much at
+        * all, we suggest testing the underlying kernel and filing the
+        * relevant bug report.  For Linux kernels, we recommend getting
+        * the Linux Test Project (LTP).
         */
-       while (LOAD_SHARED(sig_done) < num_readers)
-               cpu_relax();
-       smp_mb();       /* read sig_done before ending the barrier */
+       for (index = registry; index < registry + num_readers; index++) {
+               while (*index->need_mb) {
+                       pthread_kill(index->tid, SIGURCU);
+                       poll(NULL, 0, 1);
+               }
+       }
+       smp_mb();       /* read ->need_mb before ending the barrier */
 }
 #endif
 
@@ -167,7 +193,7 @@ void wait_for_quiescent_state(void)
                 */
                while (rcu_old_gp_ongoing(index->urcu_active_readers)) {
                        if (wait_loops++ == KICK_READER_LOOPS) {
-                               force_mb_single_thread(index->tid);
+                               force_mb_single_thread(index);
                                wait_loops = 0;
                        } else {
                                cpu_relax();
@@ -254,6 +280,7 @@ void urcu_add_reader(pthread_t id)
        registry[num_readers].tid = id;
        /* reference to the TLS of _this_ reader thread. */
        registry[num_readers].urcu_active_readers = &urcu_active_readers;
+       registry[num_readers].need_mb = &need_mb;
        num_readers++;
 }
 
@@ -303,7 +330,8 @@ void sigurcu_handler(int signo, siginfo_t *siginfo, void *context)
         * executed on.
         */
        smp_mb();
-       atomic_inc(&sig_done);
+       need_mb = 0;
+       smp_mb();
 }
 
 void __attribute__((constructor)) urcu_init(void)
This page took 0.02691 seconds and 4 git commands to generate.