Add call_rcu() interface
authorPaul E. McKenney <paulmck@linux.vnet.ibm.com>
Wed, 9 Mar 2011 02:48:49 +0000 (21:48 -0500)
committerMathieu Desnoyers <mathieu.desnoyers@efficios.com>
Wed, 9 Mar 2011 02:48:49 +0000 (21:48 -0500)
Adds call_rcu(), with RCU threads to invoke the callbacks.  By default,
there will be one such RCU thread per process, created the first time
that call_rcu() is invoked.  On systems supporting sched_getcpu(), it
is possible to create one RCU thread per CPU by calling
create_all_cpu_call_rcu_data().

This version includes feedback from Mathieu Desnoyers.

Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Makefile.am
configure.ac
tests/Makefile.am
tests/rcutorture.h
urcu-call-rcu.c [new file with mode: 0644]
urcu-call-rcu.h [new file with mode: 0644]
urcu-defer.h
urcu/wfqueue-static.h
urcu/wfstack-static.h

index 79a7152c9adcd6586cb74f91560586f844419e20..7956e7e1d509a17eee3ca0c46da2d1612aa4a193 100644 (file)
@@ -1,6 +1,7 @@
 INCLUDES = -I$(top_builddir)/urcu
 
 AM_LDFLAGS=-lpthread
+AM_CFLAGS=-Wall
 
 SUBDIRS = . tests
 
@@ -29,8 +30,8 @@ COMPAT+=compat_futex.c
 endif
 
 lib_LTLIBRARIES = liburcu.la liburcu-qsbr.la liburcu-mb.la liburcu-signal.la \
-                 liburcu-bp.la liburcu-defer.la libwfqueue.la libwfstack.la \
-                 librculfqueue.la librculfstack.la
+                 liburcu-bp.la liburcu-defer.la liburcu-call.la \
+                 libwfqueue.la libwfstack.la librculfqueue.la librculfstack.la
 
 liburcu_la_SOURCES = urcu.c urcu-pointer.c $(COMPAT)
 
@@ -44,6 +45,7 @@ liburcu_signal_la_CFLAGS = -DRCU_SIGNAL
 
 liburcu_bp_la_SOURCES = urcu-bp.c urcu-pointer.c $(COMPAT)
 
+liburcu_call_la_SOURCES = urcu-call-rcu.c $(COMPAT)
 liburcu_defer_la_SOURCES = urcu-defer.c $(COMPAT)
 
 libwfqueue_la_SOURCES = wfqueue.c $(COMPAT)
index 187cf43696a609701edcaef65931998e348df6a5..3c61abc69519b158f52af779ca8583ad661f281d 100644 (file)
@@ -35,7 +35,7 @@ AC_TYPE_SIZE_T
 # Checks for library functions.
 AC_FUNC_MALLOC
 AC_FUNC_MMAP
-AC_CHECK_FUNCS([bzero gettimeofday munmap strtoul])
+AC_CHECK_FUNCS([bzero gettimeofday munmap sched_getcpu strtoul sysconf])
 
 # Find arch type
 case $host_cpu in
index a43dd75b73a1832737c8eef95f3fb60fd61f0670..3c025a4beef397b84f07b693ff753bf741833edc 100644 (file)
@@ -1,5 +1,5 @@
 AM_LDFLAGS=-lpthread
-AM_CFLAGS=-I$(top_srcdir) -I$(top_builddir)
+AM_CFLAGS=-I$(top_srcdir) -I$(top_builddir) -g
 
 noinst_PROGRAMS = test_urcu test_urcu_dynamic_link test_urcu_timing \
        test_urcu_signal test_urcu_signal_dynamic_link test_urcu_signal_timing \
@@ -28,20 +28,21 @@ if COMPAT_FUTEX
 COMPAT+=$(top_srcdir)/compat_futex.c
 endif
 
-URCU=$(top_srcdir)/urcu.c $(top_srcdir)/urcu-pointer.c $(COMPAT)
-URCU_QSBR=$(top_srcdir)/urcu-qsbr.c $(top_srcdir)/urcu-pointer.c $(COMPAT)
+URCU=$(top_srcdir)/urcu.c $(top_srcdir)/urcu-pointer.c $(top_srcdir)/urcu-call-rcu.c $(top_srcdir)/wfqueue.c $(COMPAT)
+URCU_QSBR=$(top_srcdir)/urcu-qsbr.c $(top_srcdir)/urcu-pointer.c $(top_srcdir)/urcu-call-rcu.c $(top_srcdir)/wfqueue.c $(COMPAT)
 # URCU_MB uses urcu.c but -DRCU_MB must be defined
-URCU_MB=$(top_srcdir)/urcu.c $(top_srcdir)/urcu-pointer.c $(COMPAT)
+URCU_MB=$(top_srcdir)/urcu.c $(top_srcdir)/urcu-pointer.c $(top_srcdir)/urcu-call-rcu.c $(top_srcdir)/wfqueue.c $(COMPAT)
 # URCU_SIGNAL uses urcu.c but -DRCU_SIGNAL must be defined
-URCU_SIGNAL=$(top_srcdir)/urcu.c $(top_srcdir)/urcu-pointer.c $(COMPAT)
-URCU_BP=$(top_srcdir)/urcu-bp.c $(top_srcdir)/urcu-pointer.c $(COMPAT)
-URCU_DEFER=$(top_srcdir)/urcu.c $(top_srcdir)/urcu-defer.c $(top_srcdir)/urcu-pointer.c $(COMPAT)
+URCU_SIGNAL=$(top_srcdir)/urcu.c $(top_srcdir)/urcu-pointer.c $(top_srcdir)/urcu-call-rcu.c $(top_srcdir)/wfqueue.c $(COMPAT)
+URCU_BP=$(top_srcdir)/urcu-bp.c $(top_srcdir)/urcu-pointer.c $(top_srcdir)/urcu-call-rcu.c $(top_srcdir)/wfqueue.c $(COMPAT)
+URCU_DEFER=$(top_srcdir)/urcu.c $(top_srcdir)/urcu-defer.c $(top_srcdir)/urcu-pointer.c $(top_srcdir)/urcu-call-rcu.c $(top_srcdir)/wfqueue.c $(COMPAT)
 
 URCU_LIB=$(top_builddir)/liburcu.la
 URCU_QSBR_LIB=$(top_builddir)/liburcu-qsbr.la
 URCU_MB_LIB=$(top_builddir)/liburcu-mb.la
 URCU_SIGNAL_LIB=$(top_builddir)/liburcu-signal.la
 URCU_BP_LIB=$(top_builddir)/liburcu-bp.la
+URCU_CALL_LIB=$(top_builddir)/liburcu-call.la
 WFQUEUE_LIB=$(top_builddir)/libwfqueue.la
 WFSTACK_LIB=$(top_builddir)/libwfstack.la
 RCULFQUEUE_LIB=$(top_builddir)/librculfqueue.la
@@ -95,23 +96,23 @@ test_perthreadlock_SOURCES = test_perthreadlock.c $(URCU_SIGNAL)
 
 rcutorture_urcu_SOURCES = urcutorture.c
 rcutorture_urcu_CFLAGS = -DTORTURE_URCU $(AM_CFLAGS)
-rcutorture_urcu_LDADD = $(URCU)
+rcutorture_urcu_LDADD = $(URCU) $(URCU_CALL_LIB) $(WFQUEUE_LIB)
 
 rcutorture_urcu_mb_SOURCES = urcutorture.c
 rcutorture_urcu_mb_CFLAGS = -DTORTURE_URCU_MB $(AM_CFLAGS)
-rcutorture_urcu_mb_LDADD = $(URCU_MB_LIB)
+rcutorture_urcu_mb_LDADD = $(URCU_MB_LIB) $(URCU_CALL_LIB) $(WFQUEUE_LIB)
 
 rcutorture_qsbr_SOURCES = urcutorture.c
 rcutorture_qsbr_CFLAGS = -DTORTURE_QSBR $(AM_CFLAGS)
-rcutorture_qsbr_LDADD = $(URCU_QSBR_LIB)
+rcutorture_qsbr_LDADD = $(URCU_QSBR_LIB) $(URCU_CALL_LIB) $(WFQUEUE_LIB)
 
 rcutorture_urcu_signal_SOURCES = urcutorture.c
 rcutorture_urcu_signal_CFLAGS = -DTORTURE_URCU_SIGNAL $(AM_CFLAGS)
-rcutorture_urcu_signal_LDADD = $(URCU_SIGNAL_LIB)
+rcutorture_urcu_signal_LDADD = $(URCU_SIGNAL_LIB) $(URCU_CALL_LIB) $(WFQUEUE_LIB)
 
 rcutorture_urcu_bp_SOURCES = urcutorture.c
 rcutorture_urcu_bp_CFLAGS = -DTORTURE_URCU_BP $(AM_CFLAGS)
-rcutorture_urcu_bp_LDADD = $(URCU_BP_LIB)
+rcutorture_urcu_bp_LDADD = $(URCU_BP_LIB) $(URCU_CALL_LIB) $(WFQUEUE_LIB)
 
 test_mutex_SOURCES = test_mutex.c $(URCU)
 
index 4dac2f224faa9605f309195bea5e62e818080da0..b42b8ab7c828a52fd06ed9ddc187e498b3589d30 100644 (file)
@@ -65,6 +65,9 @@
  * Test variables.
  */
 
+#include <stdlib.h>
+#include "../urcu-call-rcu.h"
+
 DEFINE_PER_THREAD(long long, n_reads_pt);
 DEFINE_PER_THREAD(long long, n_updates_pt);
 
@@ -147,6 +150,16 @@ void *rcu_update_perf_test(void *arg)
 {
        long long n_updates_local = 0;
 
+       if ((random() & 0xf00) == 0) {
+               struct call_rcu_data *crdp;
+
+               crdp = create_call_rcu_data(0);
+               if (crdp != NULL) {
+                       fprintf(stderr,
+                               "Using per-thread call_rcu() worker.\n");
+                       set_thread_call_rcu_data(crdp);
+               }
+       }
        uatomic_inc(&nthreadsrunning);
        while (goflag == GOFLAG_INIT)
                poll(NULL, 0, 1);
@@ -296,10 +309,30 @@ void *rcu_read_stress_test(void *arg)
        return (NULL);
 }
 
+static pthread_mutex_t call_rcu_test_mutex = PTHREAD_MUTEX_INITIALIZER;
+static pthread_cond_t call_rcu_test_cond = PTHREAD_COND_INITIALIZER;
+
+void rcu_update_stress_test_rcu(struct rcu_head *head)
+{
+       if (pthread_mutex_lock(&call_rcu_test_mutex) != 0) {
+               perror("pthread_mutex_lock");
+               exit(-1);
+       }
+       if (pthread_cond_signal(&call_rcu_test_cond) != 0) {
+               perror("pthread_cond_signal");
+               exit(-1);
+       }
+       if (pthread_mutex_unlock(&call_rcu_test_mutex) != 0) {
+               perror("pthread_mutex_unlock");
+               exit(-1);
+       }
+}
+
 void *rcu_update_stress_test(void *arg)
 {
        int i;
        struct rcu_stress *p;
+       struct rcu_head rh;
 
        while (goflag == GOFLAG_INIT)
                poll(NULL, 0, 1);
@@ -317,7 +350,24 @@ void *rcu_update_stress_test(void *arg)
                for (i = 0; i < RCU_STRESS_PIPE_LEN; i++)
                        if (i != rcu_stress_idx)
                                rcu_stress_array[i].pipe_count++;
-               synchronize_rcu();
+               if (n_updates & 0x1)
+                       synchronize_rcu();
+               else {
+                       if (pthread_mutex_lock(&call_rcu_test_mutex) != 0) {
+                               perror("pthread_mutex_lock");
+                               exit(-1);
+                       }
+                       call_rcu(&rh, rcu_update_stress_test_rcu);
+                       if (pthread_cond_wait(&call_rcu_test_cond,
+                                             &call_rcu_test_mutex) != 0) {
+                               perror("pthread_cond_wait");
+                               exit(-1);
+                       }
+                       if (pthread_mutex_unlock(&call_rcu_test_mutex) != 0) {
+                               perror("pthread_mutex_unlock");
+                               exit(-1);
+                       }
+               }
                n_updates++;
        }
        return NULL;
@@ -325,6 +375,16 @@ void *rcu_update_stress_test(void *arg)
 
 void *rcu_fake_update_stress_test(void *arg)
 {
+       if ((random() & 0xf00) == 0) {
+               struct call_rcu_data *crdp;
+
+               crdp = create_call_rcu_data(0);
+               if (crdp != NULL) {
+                       fprintf(stderr,
+                               "Using per-thread call_rcu() worker.\n");
+                       set_thread_call_rcu_data(crdp);
+               }
+       }
        while (goflag == GOFLAG_INIT)
                poll(NULL, 0, 1);
        while (goflag == GOFLAG_RUN) {
@@ -396,6 +456,12 @@ int main(int argc, char *argv[])
 
        smp_init();
        //rcu_init();
+       srandom(time(NULL));
+       if (random() & 0x100) {
+               fprintf(stderr, "Allocating per-CPU call_rcu threads.\n");
+               if (create_all_cpu_call_rcu_data(0))
+                       perror("create_all_cpu_call_rcu_data");
+       }
 
 #ifdef DEBUG_YIELD
        yield_active |= YIELD_READ;
diff --git a/urcu-call-rcu.c b/urcu-call-rcu.c
new file mode 100644 (file)
index 0000000..5c003aa
--- /dev/null
@@ -0,0 +1,450 @@
+/*
+ * urcu-call-rcu.c
+ *
+ * Userspace RCU library - batch memory reclamation with kernel API
+ *
+ * Copyright (c) 2010 Paul E. McKenney <paulmck@linux.vnet.ibm.com>
+ *
+ * This library is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2.1 of the License, or (at your option) any later version.
+ *
+ * This library is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with this library; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+ */
+
+#include <stdio.h>
+#include <pthread.h>
+#include <signal.h>
+#include <assert.h>
+#include <stdlib.h>
+#include <string.h>
+#include <errno.h>
+#include <poll.h>
+#include <sys/time.h>
+#include <syscall.h>
+#include <unistd.h>
+
+#include "config.h"
+#include "urcu/wfqueue.h"
+#include "urcu-call-rcu.h"
+#include "urcu-pointer.h"
+
+/* Data structure that identifies a call_rcu thread. */
+
+struct call_rcu_data {
+       struct cds_wfq_queue cbs;
+       unsigned long flags;
+       pthread_mutex_t mtx;
+       pthread_cond_t cond;
+       unsigned long qlen;
+       pthread_t tid;
+} __attribute__((aligned(CAA_CACHE_LINE_SIZE)));
+
+/* Link a thread using call_rcu() to its call_rcu thread. */
+
+static __thread struct call_rcu_data *thread_call_rcu_data;
+
+/* Guard call_rcu thread creation. */
+
+static pthread_mutex_t call_rcu_mutex = PTHREAD_MUTEX_INITIALIZER;
+
+/* If a given thread does not have its own call_rcu thread, this is default. */
+
+static struct call_rcu_data *default_call_rcu_data;
+
+extern void synchronize_rcu(void);
+
+/*
+ * If the sched_getcpu() and sysconf(_SC_NPROCESSORS_CONF) calls are
+ * available, then we can have call_rcu threads assigned to individual
+ * CPUs rather than only to specific threads.
+ */
+
+#if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF)
+
+/*
+ * Pointer to array of pointers to per-CPU call_rcu_data structures
+ * and # CPUs.
+ */
+
+static struct call_rcu_data **per_cpu_call_rcu_data;
+static long maxcpus;
+
+/* Allocate the array if it has not already been allocated. */
+
+static void alloc_cpu_call_rcu_data(void)
+{
+       struct call_rcu_data **p;
+       static int warned = 0;
+
+       if (maxcpus != 0)
+               return;
+       maxcpus = sysconf(_SC_NPROCESSORS_CONF);
+       if (maxcpus <= 0) {
+               return;
+       }
+       p = malloc(maxcpus * sizeof(*per_cpu_call_rcu_data));
+       if (p != NULL) {
+               memset(p, '\0', maxcpus * sizeof(*per_cpu_call_rcu_data));
+               per_cpu_call_rcu_data = p;
+       } else {
+               if (!warned) {
+                       fprintf(stderr, "[error] liburcu: unable to allocate per-CPU pointer array\n");
+               }
+               warned = 1;
+       }
+}
+
+#else /* #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
+
+static const struct call_rcu_data **per_cpu_call_rcu_data = NULL;
+static const long maxcpus = -1;
+
+static void alloc_cpu_call_rcu_data(void)
+{
+}
+
+static int sched_getcpu(void)
+{
+       return -1;
+}
+
+#endif /* #else #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
+
+/* Acquire the specified pthread mutex. */
+
+static void call_rcu_lock(pthread_mutex_t *pmp)
+{
+       if (pthread_mutex_lock(pmp) != 0) {
+               perror("pthread_mutex_lock");
+               exit(-1);
+       }
+}
+
+/* Release the specified pthread mutex. */
+
+static void call_rcu_unlock(pthread_mutex_t *pmp)
+{
+       if (pthread_mutex_unlock(pmp) != 0) {
+               perror("pthread_mutex_unlock");
+               exit(-1);
+       }
+}
+
+/* This is the code run by each call_rcu thread. */
+
+static void *call_rcu_thread(void *arg)
+{
+       unsigned long cbcount;
+       struct cds_wfq_node *cbs;
+       struct cds_wfq_node **cbs_tail;
+       struct call_rcu_data *crdp = (struct call_rcu_data *)arg;
+       struct rcu_head *rhp;
+
+       thread_call_rcu_data = crdp;
+       for (;;) {
+               if (&crdp->cbs.head != _CMM_LOAD_SHARED(crdp->cbs.tail)) {
+                       while ((cbs = _CMM_LOAD_SHARED(crdp->cbs.head)) == NULL)
+                               poll(NULL, 0, 1);
+                       _CMM_STORE_SHARED(crdp->cbs.head, NULL);
+                       cbs_tail = (struct cds_wfq_node **)
+                               uatomic_xchg(&crdp->cbs.tail, &crdp->cbs.head);
+                       synchronize_rcu();
+                       cbcount = 0;
+                       do {
+                               while (cbs->next == NULL &&
+                                      &cbs->next != cbs_tail)
+                                       poll(NULL, 0, 1);
+                               if (cbs == &crdp->cbs.dummy) {
+                                       cbs = cbs->next;
+                                       continue;
+                               }
+                               rhp = (struct rcu_head *)cbs;
+                               cbs = cbs->next;
+                               rhp->func(rhp);
+                               cbcount++;
+                       } while (cbs != NULL);
+                       uatomic_sub(&crdp->qlen, cbcount);
+               }
+               if (crdp->flags & URCU_CALL_RCU_RT)
+                       poll(NULL, 0, 10);
+               else {
+                       call_rcu_lock(&crdp->mtx);
+                       _CMM_STORE_SHARED(crdp->flags,
+                                    crdp->flags & ~URCU_CALL_RCU_RUNNING);
+                       if (&crdp->cbs.head ==
+                           _CMM_LOAD_SHARED(crdp->cbs.tail) &&
+                           pthread_cond_wait(&crdp->cond, &crdp->mtx) != 0) {
+                               perror("pthread_cond_wait");
+                               exit(-1);
+                       }
+                       _CMM_STORE_SHARED(crdp->flags,
+                                    crdp->flags | URCU_CALL_RCU_RUNNING);
+                       poll(NULL, 0, 10);
+                       call_rcu_unlock(&crdp->mtx);
+               }
+       }
+       return NULL;  /* NOTREACHED */
+}
+
+/*
+ * Create both a call_rcu thread and the corresponding call_rcu_data
+ * structure, linking the structure in as specified.
+ */
+
+void call_rcu_data_init(struct call_rcu_data **crdpp, unsigned long flags)
+{
+       struct call_rcu_data *crdp;
+
+       crdp = malloc(sizeof(*crdp));
+       if (crdp == NULL) {
+               fprintf(stderr, "Out of memory.\n");
+               exit(-1);
+       }
+       memset(crdp, '\0', sizeof(*crdp));
+       cds_wfq_init(&crdp->cbs);
+       crdp->qlen = 0;
+       if (pthread_mutex_init(&crdp->mtx, NULL) != 0) {
+               perror("pthread_mutex_init");
+               exit(-1);
+       }
+       if (pthread_cond_init(&crdp->cond, NULL) != 0) {
+               perror("pthread_cond_init");
+               exit(-1);
+       }
+       crdp->flags = flags | URCU_CALL_RCU_RUNNING;
+       cmm_smp_mb();  /* Structure initialized before pointer is planted. */
+       *crdpp = crdp;
+       if (pthread_create(&crdp->tid, NULL, call_rcu_thread, crdp) != 0) {
+               perror("pthread_create");
+               exit(-1);
+       }
+}
+
+/*
+ * Return a pointer to the call_rcu_data structure for the specified
+ * CPU, returning NULL if there is none.  We cannot automatically
+ * created it because the platform we are running on might not define
+ * sched_getcpu().
+ */
+
+struct call_rcu_data *get_cpu_call_rcu_data(int cpu)
+{
+       static int warned = 0;
+
+       if (per_cpu_call_rcu_data == NULL)
+               return NULL;
+       if (!warned && maxcpus > 0 && (cpu < 0 || maxcpus <= cpu)) {
+               fprintf(stderr, "[error] liburcu: get CPU # out of range\n");
+               warned = 1;
+       }
+       if (cpu < 0 || maxcpus <= cpu)
+               return NULL;
+       return per_cpu_call_rcu_data[cpu];
+}
+
+/*
+ * Return the tid corresponding to the call_rcu thread whose
+ * call_rcu_data structure is specified.
+ */
+
+pthread_t get_call_rcu_thread(struct call_rcu_data *crdp)
+{
+       return crdp->tid;
+}
+
+/*
+ * Create a call_rcu_data structure (with thread) and return a pointer.
+ */
+
+struct call_rcu_data *create_call_rcu_data(unsigned long flags)
+{
+       struct call_rcu_data *crdp;
+
+       call_rcu_data_init(&crdp, flags);
+       return crdp;
+}
+
+/*
+ * Set the specified CPU to use the specified call_rcu_data structure.
+ */
+
+int set_cpu_call_rcu_data(int cpu, struct call_rcu_data *crdp)
+{
+       int warned = 0;
+
+       call_rcu_lock(&call_rcu_mutex);
+       if (cpu < 0 || maxcpus <= cpu) {
+               if (!warned) {
+                       fprintf(stderr, "[error] liburcu: set CPU # out of range\n");
+                       warned = 1;
+               }
+               call_rcu_unlock(&call_rcu_mutex);
+               errno = EINVAL;
+               return -EINVAL;
+       }
+       alloc_cpu_call_rcu_data();
+       call_rcu_unlock(&call_rcu_mutex);
+       if (per_cpu_call_rcu_data == NULL) {
+               errno = ENOMEM;
+               return -ENOMEM;
+       }
+       per_cpu_call_rcu_data[cpu] = crdp;
+       return 0;
+}
+
+/*
+ * Return a pointer to the default call_rcu_data structure, creating
+ * one if need be.  Because we never free call_rcu_data structures,
+ * we don't need to be in an RCU read-side critical section.
+ */
+
+struct call_rcu_data *get_default_call_rcu_data(void)
+{
+       if (default_call_rcu_data != NULL)
+               return rcu_dereference(default_call_rcu_data);
+       call_rcu_lock(&call_rcu_mutex);
+       if (default_call_rcu_data != NULL) {
+               call_rcu_unlock(&call_rcu_mutex);
+               return default_call_rcu_data;
+       }
+       call_rcu_data_init(&default_call_rcu_data, 0);
+       call_rcu_unlock(&call_rcu_mutex);
+       return default_call_rcu_data;
+}
+
+/*
+ * Return the call_rcu_data structure that applies to the currently
+ * running thread.  Any call_rcu_data structure assigned specifically
+ * to this thread has first priority, followed by any call_rcu_data
+ * structure assigned to the CPU on which the thread is running,
+ * followed by the default call_rcu_data structure.  If there is not
+ * yet a default call_rcu_data structure, one will be created.
+ */
+struct call_rcu_data *get_call_rcu_data(void)
+{
+       int curcpu;
+       static int warned = 0;
+
+       if (thread_call_rcu_data != NULL)
+               return thread_call_rcu_data;
+       if (maxcpus <= 0)
+               return get_default_call_rcu_data();
+       curcpu = sched_getcpu();
+       if (!warned && (curcpu < 0 || maxcpus <= curcpu)) {
+               fprintf(stderr, "[error] liburcu: gcrd CPU # out of range\n");
+               warned = 1;
+       }
+       if (curcpu >= 0 && maxcpus > curcpu &&
+           per_cpu_call_rcu_data != NULL &&
+           per_cpu_call_rcu_data[curcpu] != NULL)
+               return per_cpu_call_rcu_data[curcpu];
+       return get_default_call_rcu_data();
+}
+
+/*
+ * Return a pointer to this task's call_rcu_data if there is one.
+ */
+
+struct call_rcu_data *get_thread_call_rcu_data(void)
+{
+       return thread_call_rcu_data;
+}
+
+/*
+ * Set this task's call_rcu_data structure as specified, regardless
+ * of whether or not this task already had one.  (This allows switching
+ * to and from real-time call_rcu threads, for example.)
+ */
+
+void set_thread_call_rcu_data(struct call_rcu_data *crdp)
+{
+       thread_call_rcu_data = crdp;
+}
+
+/*
+ * Create a separate call_rcu thread for each CPU.  This does not
+ * replace a pre-existing call_rcu thread -- use the set_cpu_call_rcu_data()
+ * function if you want that behavior.
+ */
+
+int create_all_cpu_call_rcu_data(unsigned long flags)
+{
+       int i;
+       struct call_rcu_data *crdp;
+       int ret;
+
+       call_rcu_lock(&call_rcu_mutex);
+       alloc_cpu_call_rcu_data();
+       call_rcu_unlock(&call_rcu_mutex);
+       if (maxcpus <= 0) {
+               errno = EINVAL;
+               return -EINVAL;
+       }
+       if (per_cpu_call_rcu_data == NULL) {
+               errno = ENOMEM;
+               return -ENOMEM;
+       }
+       for (i = 0; i < maxcpus; i++) {
+               call_rcu_lock(&call_rcu_mutex);
+               if (get_cpu_call_rcu_data(i)) {
+                       call_rcu_unlock(&call_rcu_mutex);
+                       continue;
+               }
+               crdp = create_call_rcu_data(flags);
+               if (crdp == NULL) {
+                       call_rcu_unlock(&call_rcu_mutex);
+                       errno = ENOMEM;
+                       return -ENOMEM;
+               }
+               call_rcu_unlock(&call_rcu_mutex);
+               if ((ret = set_cpu_call_rcu_data(i, crdp)) != 0) {
+                       /* FIXME: Leaks crdp for now. */
+                       return ret; /* Can happen on race. */
+               }
+       }
+       return 0;
+}
+
+/*
+ * Schedule a function to be invoked after a following grace period.
+ * This is the only function that must be called -- the others are
+ * only present to allow applications to tune their use of RCU for
+ * maximum performance.
+ *
+ * Note that unless a call_rcu thread has not already been created,
+ * the first invocation of call_rcu() will create one.  So, if you
+ * need the first invocation of call_rcu() to be fast, make sure
+ * to create a call_rcu thread first.  One way to accomplish this is
+ * "get_call_rcu_data();", and another is create_all_cpu_call_rcu_data().
+ */
+
+void call_rcu(struct rcu_head *head,
+             void (*func)(struct rcu_head *head))
+{
+       struct call_rcu_data *crdp;
+
+       cds_wfq_node_init(&head->next);
+       head->func = func;
+       crdp = get_call_rcu_data();
+       cds_wfq_enqueue(&crdp->cbs, &head->next);
+       uatomic_inc(&crdp->qlen);
+       if (!(_CMM_LOAD_SHARED(crdp->flags) & URCU_CALL_RCU_RT)) {
+               call_rcu_lock(&crdp->mtx);
+               if (!(_CMM_LOAD_SHARED(crdp->flags) & URCU_CALL_RCU_RUNNING)) {
+                       if (pthread_cond_signal(&crdp->cond) != 0) {
+                               perror("pthread_cond_signal");
+                               exit(-1);
+                       }
+               }
+               call_rcu_unlock(&crdp->mtx);
+       }
+}
diff --git a/urcu-call-rcu.h b/urcu-call-rcu.h
new file mode 100644 (file)
index 0000000..2c13388
--- /dev/null
@@ -0,0 +1,80 @@
+#ifndef _URCU_CALL_RCU_H
+#define _URCU_CALL_RCU_H
+
+/*
+ * urcu-call-rcu.h
+ *
+ * Userspace RCU header - deferred execution
+ *
+ * Copyright (c) 2009 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
+ * Copyright (c) 2009 Paul E. McKenney, IBM Corporation.
+ *
+ * LGPL-compatible code should include this header with :
+ *
+ * #define _LGPL_SOURCE
+ * #include <urcu-defer.h>
+ *
+ * This library is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2.1 of the License, or (at your option) any later version.
+ *
+ * This library is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with this library; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+ */
+
+#include <stdlib.h>
+#include <pthread.h>
+
+#include <urcu/wfqueue.h>
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+/* Note that struct call_rcu_data is opaque to callers. */
+
+struct call_rcu_data;
+
+/* Flag values. */
+
+#define URCU_CALL_RCU_RT       0x1
+#define URCU_CALL_RCU_RUNNING  0x2
+
+/*
+ * The rcu_head data structure is placed in the structure to be freed
+ * via call_rcu().
+ */
+
+struct rcu_head {
+       struct cds_wfq_node next;
+       void (*func)(struct rcu_head *head);
+};
+
+/*
+ * Exported functions
+ */
+void call_rcu_data_init(struct call_rcu_data **crdpp, unsigned long flags);
+struct call_rcu_data *get_cpu_call_rcu_data(int cpu);
+pthread_t get_call_rcu_thread(struct call_rcu_data *crdp);
+struct call_rcu_data *create_call_rcu_data(unsigned long flags);
+int set_cpu_call_rcu_data(int cpu, struct call_rcu_data *crdp);
+struct call_rcu_data *get_default_call_rcu_data(void);
+struct call_rcu_data *get_call_rcu_data(void);
+struct call_rcu_data *get_thread_call_rcu_data(void);
+void set_thread_call_rcu_data(struct call_rcu_data *crdp);
+int create_all_cpu_call_rcu_data(unsigned long flags);
+void call_rcu(struct rcu_head *head,
+             void (*func)(struct rcu_head *head));
+
+#ifdef __cplusplus 
+}
+#endif
+
+#endif /* _URCU_CALL_RCU_H */
index 2019e133f2bbe0552746b00450a074a1ba9cfe2c..3710081cd06a20d796eee3d5cfb8deea83204b27 100644 (file)
@@ -52,14 +52,6 @@ extern "C" {
 
 extern void defer_rcu(void (*fct)(void *p), void *p);
 
-/*
- * call_rcu will eventually be implemented with an API similar to the Linux
- * kernel call_rcu(), which will allow its use within RCU read-side C.S.
- * Generate an error if used for now.
- */
-
-#define call_rcu       __error_call_rcu_not_implemented_please_use_defer_rcu
-
 /*
  * Thread registration for reclamation.
  */
index 30d6e963f321f9c7f836e8ae2f9a0b2d5a7e857d..790931bef25db04d2156fe0cd39dd260b31a0474 100644 (file)
@@ -28,6 +28,7 @@
 
 #include <pthread.h>
 #include <assert.h>
+#include <poll.h>
 #include <urcu/compiler.h>
 #include <urcu/uatomic_arch.h>
 
@@ -47,12 +48,12 @@ extern "C" {
 #define WFQ_ADAPT_ATTEMPTS             10      /* Retry if being set */
 #define WFQ_WAIT                       10      /* Wait 10 ms if being set */
 
-void _cds_wfq_node_init(struct cds_wfq_node *node)
+static inline void _cds_wfq_node_init(struct cds_wfq_node *node)
 {
        node->next = NULL;
 }
 
-void _cds_wfq_init(struct cds_wfq_queue *q)
+static inline void _cds_wfq_init(struct cds_wfq_queue *q)
 {
        int ret;
 
@@ -64,7 +65,8 @@ void _cds_wfq_init(struct cds_wfq_queue *q)
        assert(!ret);
 }
 
-void _cds_wfq_enqueue(struct cds_wfq_queue *q, struct cds_wfq_node *node)
+static inline void _cds_wfq_enqueue(struct cds_wfq_queue *q,
+                                   struct cds_wfq_node *node)
 {
        struct cds_wfq_node **old_tail;
 
@@ -90,7 +92,7 @@ void _cds_wfq_enqueue(struct cds_wfq_queue *q, struct cds_wfq_node *node)
  * thread to be scheduled. The queue appears empty until tail->next is set by
  * enqueue.
  */
-struct cds_wfq_node *
+static inline struct cds_wfq_node *
 ___cds_wfq_dequeue_blocking(struct cds_wfq_queue *q)
 {
        struct cds_wfq_node *node, *next;
@@ -128,7 +130,7 @@ ___cds_wfq_dequeue_blocking(struct cds_wfq_queue *q)
        return node;
 }
 
-struct cds_wfq_node *
+static inline struct cds_wfq_node *
 _cds_wfq_dequeue_blocking(struct cds_wfq_queue *q)
 {
        struct cds_wfq_node *retnode;
index eed83da334a173fa4e3a0d062c89463473ebfb94..ff18c4a3f213214188ef6bfe6f6c98a795132bab 100644 (file)
@@ -28,6 +28,7 @@
 
 #include <pthread.h>
 #include <assert.h>
+#include <poll.h>
 #include <urcu/compiler.h>
 #include <urcu/uatomic_arch.h>
 
This page took 0.036022 seconds and 4 git commands to generate.