lock-free queue with RCU-based garbage collection
authorPaolo Bonzini <pbonzini@redhat.com>
Mon, 1 Mar 2010 19:57:10 +0000 (14:57 -0500)
committerMathieu Desnoyers <mathieu.desnoyers@polymtl.ca>
Mon, 1 Mar 2010 19:57:10 +0000 (14:57 -0500)
A lock-free queue example using RCU to avoid the need for double-word
compare-and-swap and, at the same time, to implement efficient garbage
collection.  I tested it lightly, and I don't think it's fully ready
for inclusion.  It does need the previous two patches to avoid deadlocks,
so I did test it somewhat. :-)

See individual files for detailed comments on how it works.

Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
tests/Makefile.am
tests/test_qsbr_lfq.c [new file with mode: 0644]
tests/test_urcu_lfq.c [new file with mode: 0644]

index 17b89db7bd9fa94754e8fcff8efd8c580e291aa6..2ebaae5b0f354a2dfed2748a973898210bc08050 100644 (file)
@@ -12,7 +12,12 @@ noinst_PROGRAMS = test_urcu test_urcu_dynamic_link test_urcu_timing \
         test_urcu_mb_gc test_qsbr_gc test_qsbr_lgc test_urcu_signal_lgc \
         test_urcu_mb_lgc test_qsbr_dynamic_link test_urcu_defer \
         test_uatomic test_urcu_assign test_urcu_assign_dynamic_link \
-        test_urcu_bp test_urcu_bp_dynamic_link
+        test_urcu_bp test_urcu_bp_dynamic_link \
+        test_urcu_lfq test_urcu_signal_lfq test_urcu_dynamic_link_lfq \
+        test_urcu_signal_dynamic_link_lfq test_urcu_mb_lfq \
+        test_urcu_yield_lfq test_urcu_signal_yield_lfq test_urcu_defer_lfq \
+       test_qsbr_lfq test_qsbr_dynamic_link_lfq
+
 noinst_HEADERS = rcutorture.h
 
 if COMPAT_ARCH
@@ -149,4 +154,33 @@ test_urcu_bp_SOURCES = test_urcu_bp.c $(URCU_BP)
 test_urcu_bp_dynamic_link_SOURCES = test_urcu_bp.c $(URCU_BP)
 test_urcu_bp_dynamic_link_CFLAGS = -DDYNAMIC_LINK_TEST $(AM_CFLAGS)
 
+test_urcu_lfq_SOURCES = test_urcu_lfq.c $(URCU)
+
+test_urcu_defer_lfq_SOURCES = test_urcu_defer_lfq.c $(URCU_DEFER)
+
+test_urcu_signal_lfq_SOURCES = test_urcu_lfq.c $(URCU_SIGNAL)
+test_urcu_signal_lfq_CFLAGS = -DRCU_SIGNAL $(AM_CFLAGS)
+
+test_urcu_dynamic_link_lfq_SOURCES = test_urcu_lfq.c $(URCU)
+test_urcu_dynamic_link_lfq_CFLAGS = -DDYNAMIC_LINK_TEST $(AM_CFLAGS)
+
+test_urcu_signal_dynamic_link_lfq_SOURCES = test_urcu_lfq.c $(URCU)
+test_urcu_signal_dynamic_link_lfq_CFLAGS = -DRCU_SIGNAL -DDYNAMIC_LINK_TEST $(AM_CFLAGS)
+
+test_urcu_mb_lfq_SOURCES = test_urcu_lfq.c $(URCU_MB)
+test_urcu_mb_lfq_CFLAGS = -DRCU_MB $(AM_CFLAGS)
+
+test_urcu_yield_lfq_SOURCES = test_urcu_lfq.c $(URCU)
+test_urcu_yield_lfq_CFLAGS = -DDEBUG_YIELD $(AM_CFLAGS)
+
+test_urcu_signal_yield_lfq_SOURCES = test_urcu_lfq.c $(URCU_SIGNAL)
+test_urcu_signal_yield_lfq_CFLAGS = -DRCU_SIGNAL -DDEBUG_YIELD $(AM_CFLAGS)
+
+test_qsbr_lfq_SOURCES = test_qsbr_lfq.c $(URCU_QSBR)
+
+test_qsbr_dynamic_link_lfq_SOURCES = test_qsbr_lfq.c $(URCU_QSBR)
+test_qsbr_dynamic_link_lfq_CFLAGS = -DDYNAMIC_LINK_TEST $(AM_CFLAGS)
+
+
+
 urcutorture.c: api.h
diff --git a/tests/test_qsbr_lfq.c b/tests/test_qsbr_lfq.c
new file mode 100644 (file)
index 0000000..2e3f130
--- /dev/null
@@ -0,0 +1,587 @@
+/*
+ * test_urcu.c
+ *
+ * Userspace RCU library - example RCU-based lock-free queue
+ *
+ * Copyright February 2010 - Paolo Bonzini <pbonzinI@redhat.com>
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License as published by
+ * the Free Software Foundation; either version 2 of the License, or
+ * (at your option) any later version.
+ *
+ * This program 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 General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License along
+ * with this program; if not, write to the Free Software Foundation, Inc.,
+ * 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+ */
+
+#define _GNU_SOURCE
+#include "../config.h"
+#include <stdio.h>
+#include <pthread.h>
+#include <stdlib.h>
+#include <stdint.h>
+#include <stdbool.h>
+#include <string.h>
+#include <sys/types.h>
+#include <sys/wait.h>
+#include <unistd.h>
+#include <stdio.h>
+#include <assert.h>
+#include <sys/syscall.h>
+#include <sched.h>
+#include <errno.h>
+
+#include <urcu/arch.h>
+
+/* hardcoded number of CPUs */
+#define NR_CPUS 16384
+
+#if defined(_syscall0)
+_syscall0(pid_t, gettid)
+#elif defined(__NR_gettid)
+static inline pid_t gettid(void)
+{
+       return syscall(__NR_gettid);
+}
+#else
+#warning "use pid as tid"
+static inline pid_t gettid(void)
+{
+       return getpid();
+}
+#endif
+
+#ifndef DYNAMIC_LINK_TEST
+#define _LGPL_SOURCE
+#endif
+#include "urcu-qsbr.h"
+
+
+static volatile int test_go, test_stop;
+
+static unsigned long rduration;
+
+static unsigned long duration;
+
+/* read-side C.S. duration, in loops */
+static unsigned long wdelay;
+
+static inline void loop_sleep(unsigned long l)
+{
+       while(l-- != 0)
+               cpu_relax();
+}
+
+static int verbose_mode;
+
+#define printf_verbose(fmt, args...)           \
+       do {                                    \
+               if (verbose_mode)               \
+                       printf(fmt, args);      \
+       } while (0)
+
+static unsigned int cpu_affinities[NR_CPUS];
+static unsigned int next_aff = 0;
+static int use_affinity = 0;
+
+pthread_mutex_t affinity_mutex = PTHREAD_MUTEX_INITIALIZER;
+
+#ifndef HAVE_CPU_SET_T
+typedef unsigned long cpu_set_t;
+# define CPU_ZERO(cpuset) do { *(cpuset) = 0; } while(0)
+# define CPU_SET(cpu, cpuset) do { *(cpuset) |= (1UL << (cpu)); } while(0)
+#endif
+
+static void set_affinity(void)
+{
+       cpu_set_t mask;
+       int cpu;
+       int ret;
+
+       if (!use_affinity)
+               return;
+
+#if HAVE_SCHED_SETAFFINITY
+       ret = pthread_mutex_lock(&affinity_mutex);
+       if (ret) {
+               perror("Error in pthread mutex lock");
+               exit(-1);
+       }
+       cpu = cpu_affinities[next_aff++];
+       ret = pthread_mutex_unlock(&affinity_mutex);
+       if (ret) {
+               perror("Error in pthread mutex unlock");
+               exit(-1);
+       }
+
+       CPU_ZERO(&mask);
+       CPU_SET(cpu, &mask);
+#if SCHED_SETAFFINITY_ARGS == 2
+       sched_setaffinity(0, &mask);
+#else
+       sched_setaffinity(0, sizeof(mask), &mask);
+#endif
+#endif /* HAVE_SCHED_SETAFFINITY */
+}
+
+/*
+ * returns 0 if test should end.
+ */
+static int test_duration_dequeue(void)
+{
+       return !test_stop;
+}
+
+static int test_duration_enqueue(void)
+{
+       return !test_stop;
+}
+
+static unsigned long long __thread nr_dequeues;
+static unsigned long long __thread nr_enqueues;
+
+static unsigned long nr_successful_dequeues;
+static unsigned int nr_enqueuers;
+static unsigned int nr_dequeuers;
+
+
+#define ARRAY_POISON   0xDEADBEEF
+#define PAGE_SIZE      4096
+#define PAGE_MASK      (PAGE_SIZE - 1)
+#define NODES_PER_PAGE (PAGE_SIZE - 16) / sizeof (struct node)
+
+/* Lock-free queue, using the RCU to avoid the ABA problem and (more
+   interestingly) to efficiently handle freeing memory.
+
+   We have to protect both the enqueuer and dequeuer's compare-and-
+   exchange operation from running across a free and a subsequent
+   reallocation of the same memory.  So, we protect the free with
+   synchronize_rcu; this is enough because all the allocations take
+   place before the compare-and-exchange ops.
+
+   Besides adding rcu_read_{,un}lock, the enqueue/dequeue are a standard
+   implementation of a lock-free-queue.  The first node in the queue is
+   always dummy: dequeuing returns the data from HEAD->NEXT, advances
+   HEAD to HEAD->NEXT (which will now serve as dummy node), and frees the
+   old HEAD.  Since RCU avoids the ABA problem, it doesn't use double-word
+   compare-and-exchange operations.  Node allocation and deallocation are
+   a "black box" and synchronize_rcu is hidden within node deallocation.
+
+   So, the tricky part is finding a good allocation strategy for nodes.
+   The allocator for nodes is shared by multiple threads, and since
+   malloc/free are not lock-free a layer above them is obviously
+   necessary: otherwise the whole exercise is useless.  In addition,
+   to avoid penalizing dequeues, the allocator should avoid frequent
+   synchronization (because synchronize_rcu is expensive).
+
+   The scheme that is used here uses a page as the allocation
+   unit for nodes.  A page is freed when no more nodes are in use.
+   Nodes from a page are never reused.
+
+   The nodes are allocated from Q->CURRENT.  Since whoever finds a full
+   page has to busy wait, we use a trick to limit the duration of busy
+   waiting.  A free page Q->FREE is always kept ready, so that any thread
+   that allocates the last node in a page, or finds a full page can try
+   to update Q->CURRENT.  Whoever loses the race has to busy wait, OTOH
+   whoever wins the race has to allocate the new Q->FREE.  In other words,  
+   if the following sequence happens
+
+     Thread 1                    Thread 2                 other threads
+     -----------------------------------------------------------------------
+     Get last node from page
+                                q->current = q->free;
+                                                         fill up q->current
+     q->current = q->free fails
+
+   then thread 1 does not have anymore the duty of allocating q->current;
+   thread 2 will do that.  If a thread finds a full current page and
+   Q->CURRENT == Q->FREE, this means that another thread is going to
+   allocate Q->FREE soon, and it busy waits.  After the allocation
+   finishes, everything proceeds normally: some thread will take care
+   of setting Q->CURRENT and allocating a new Q->FREE.
+   One common scheme for allocation is to use a free list (implemented
+   as a lock-free stack), but this free list is potentially unbounded.
+   Instead, with the above scheme the number of live pages at any time
+   is equal to the number of enqueuing threads.  */
+
+struct node {
+       void *data;
+       void *next;
+};
+
+struct node_page {
+       int in;
+       int out;
+       char padding[16 - sizeof(int) * 2];
+       struct node nodes[NODES_PER_PAGE];
+};
+
+
+struct queue {
+       struct node_page *current, *free;
+       struct node *head, *tail;
+};
+
+static struct node_page *new_node_page()
+{
+       struct node_page *p = valloc (PAGE_SIZE);
+       rcu_quiescent_state();
+       p->in = p->out = 0;
+       return p;
+}
+
+static void free_node_page(struct node_page *p)
+{
+       /* Help making sure that accessing a dangling pointer is
+          adequately punished.  */
+       p->in = ARRAY_POISON;
+       free (p);
+}
+
+static struct node *new_node(struct queue *q)
+{
+       struct node *n;
+       struct node_page *p;
+       int i;
+
+       do {
+               p = q->current;
+               i = p->in;
+               if (i >= NODES_PER_PAGE - 1 &&
+                   q->free != p &&
+                   uatomic_cmpxchg(&q->current, p, q->free) == p)
+                       q->free = new_node_page();
+
+       } while (i == NODES_PER_PAGE || uatomic_cmpxchg(&p->in, i, i+1) != i);
+
+       assert (i >= 0 && i < NODES_PER_PAGE);
+       n = &p->nodes[i];
+       n->next = NULL;
+       return n;
+}
+
+void free_node(struct node *n)
+{
+       struct node_page *p = (struct node_page *) ((intptr_t) n & ~PAGE_MASK);
+
+       if (uatomic_add_return(&p->out, 1) == NODES_PER_PAGE) {
+               rcu_quiescent_state();
+               synchronize_rcu();
+               free_node_page(p);
+       }
+}
+
+void init_queue(struct queue *q)
+{
+       q->current = new_node_page();
+       q->free = new_node_page();
+       q->head = q->tail = new_node(q);
+}
+
+void enqueue(struct queue *q, void *value)
+{
+       struct node *n = new_node(q);
+       n->data = value;
+       rcu_read_lock();
+       for (;;) {
+               struct node *tail = rcu_dereference(q->tail);
+               struct node *next = rcu_dereference(tail->next);
+               if (tail != q->tail) {
+                       /* A change occurred while reading the values.  */
+                       continue;
+               }
+
+               if (next) {
+                       /* Help moving tail further.  */
+                       uatomic_cmpxchg(&q->tail, tail, next);
+                       continue;
+               }
+
+               if (uatomic_cmpxchg(&tail->next, NULL, n) == NULL) {
+                       /* Move tail (another operation might beat us to it,
+                          that's fine).  */
+                       uatomic_cmpxchg(&q->tail, tail, n);
+                       rcu_read_unlock();
+                       return;
+               }
+       }
+}
+
+void *dequeue(struct queue *q, bool *not_empty)
+{
+       bool dummy;
+       if (!not_empty)
+               not_empty = &dummy;
+
+       rcu_read_lock();
+       *not_empty = false;
+       for (;;) {
+               void *data;
+               struct node *head = rcu_dereference(q->head);
+               struct node *tail = rcu_dereference(q->tail);
+               struct node *next = rcu_dereference(head->next);
+
+               if (head != q->head) {
+                       /* A change occurred while reading the values.  */
+                       continue;
+               }
+
+               if (head == tail) {
+                       /* If all three are consistent, the queue is empty.  */
+                       if (!next)
+                               return NULL;
+
+                       /* Help moving tail further.  */
+                       uatomic_cmpxchg(&q->tail, tail, next);
+                       continue;
+               }
+
+               data = next->data;
+               if (uatomic_cmpxchg(&q->head, head, next) == head) {
+                       /* Next remains as a dummy node, head is freed.  */
+                       rcu_read_unlock();
+                       *not_empty = true;
+                       free_node (head);
+                       return data;
+               }
+       }
+}
+
+\f
+static struct queue q;
+
+void *thr_enqueuer(void *_count)
+{
+       unsigned long long *count = _count;
+
+       printf_verbose("thread_begin %s, thread id : %lx, tid %lu\n",
+                       "enqueuer", pthread_self(), (unsigned long)gettid());
+
+       set_affinity();
+
+       rcu_register_thread();
+
+       while (!test_go)
+       {
+       }
+       smp_mb();
+
+       for (;;) {
+               enqueue (&q, NULL);
+
+               if (unlikely(wdelay))
+                       loop_sleep(wdelay);
+               nr_enqueues++;
+               if (unlikely(!test_duration_enqueue()))
+                       break;
+       }
+
+       rcu_unregister_thread();
+
+       *count = nr_enqueues;
+       printf_verbose("thread_end %s, thread id : %lx, tid %lu - count %d\n",
+                      "enqueuer", pthread_self(), (unsigned long)gettid(), nr_enqueues);
+       return ((void*)1);
+
+}
+
+void *thr_dequeuer(void *_count)
+{
+       unsigned long long *count = _count;
+
+       printf_verbose("thread_begin %s, thread id : %lx, tid %lu\n",
+                       "dequeuer", pthread_self(), (unsigned long)gettid());
+
+       set_affinity();
+
+       rcu_register_thread();
+
+       while (!test_go)
+       {
+       }
+       smp_mb();
+
+       for (;;) {
+               bool not_empty;
+               dequeue (&q, &not_empty);
+               if (not_empty)
+                       uatomic_inc (&nr_successful_dequeues);
+
+               nr_dequeues++;
+               if (unlikely(!test_duration_dequeue()))
+                       break;
+               if (unlikely(rduration))
+                       loop_sleep(rduration);
+       }
+
+       rcu_unregister_thread();
+
+       printf_verbose("thread_end %s, thread id : %lx, tid %lu - count %d\n",
+                       "dequeuer", pthread_self(), (unsigned long)gettid(), nr_dequeues);
+       *count = nr_dequeues;
+       return ((void*)2);
+}
+
+void test_end(struct queue *q)
+{
+       bool not_empty;
+       do
+               dequeue (q, &not_empty);
+       while (!not_empty);
+       if (q->current != q->free)
+               free_node_page(q->free);
+       free_node_page(q->current);
+}
+
+void show_usage(int argc, char **argv)
+{
+       printf("Usage : %s nr_dequeuers nr_enqueuers duration (s)", argv[0]);
+       printf(" [-d delay] (enqueuer period (us))");
+       printf(" [-c duration] (dequeuer C.S. duration (in loops))");
+       printf(" [-v] (verbose output)");
+       printf(" [-a cpu#] [-a cpu#]... (affinity)");
+       printf("\n");
+}
+
+int main(int argc, char **argv)
+{
+       int err;
+       pthread_t *tid_enqueuer, *tid_dequeuer;
+       void *tret;
+       unsigned long long *count_enqueuer, *count_dequeuer;
+       unsigned long long tot_enqueues = 0, tot_dequeues = 0;
+       int i, a;
+
+       if (argc < 4) {
+               show_usage(argc, argv);
+               return -1;
+       }
+
+       err = sscanf(argv[1], "%u", &nr_dequeuers);
+       if (err != 1) {
+               show_usage(argc, argv);
+               return -1;
+       }
+
+       err = sscanf(argv[2], "%u", &nr_enqueuers);
+       if (err != 1) {
+               show_usage(argc, argv);
+               return -1;
+       }
+       
+       err = sscanf(argv[3], "%lu", &duration);
+       if (err != 1) {
+               show_usage(argc, argv);
+               return -1;
+       }
+
+       for (i = 4; i < argc; i++) {
+               if (argv[i][0] != '-')
+                       continue;
+               switch (argv[i][1]) {
+               case 'a':
+                       if (argc < i + 2) {
+                               show_usage(argc, argv);
+                               return -1;
+                       }
+                       a = atoi(argv[++i]);
+                       cpu_affinities[next_aff++] = a;
+                       use_affinity = 1;
+                       printf_verbose("Adding CPU %d affinity\n", a);
+                       break;
+               case 'c':
+                       if (argc < i + 2) {
+                               show_usage(argc, argv);
+                               return -1;
+                       }
+                       rduration = atol(argv[++i]);
+                       break;
+               case 'd':
+                       if (argc < i + 2) {
+                               show_usage(argc, argv);
+                               return -1;
+                       }
+                       wdelay = atol(argv[++i]);
+                       break;
+               case 'v':
+                       verbose_mode = 1;
+                       break;
+               }
+       }
+
+       printf_verbose("running test for %lu seconds, %u enqueuers, %u dequeuers.\n",
+               duration, nr_enqueuers, nr_dequeuers);
+       printf_verbose("Writer delay : %lu loops.\n", rduration);
+       printf_verbose("Reader duration : %lu loops.\n", wdelay);
+       printf_verbose("thread %-6s, thread id : %lx, tid %lu\n",
+                       "main", pthread_self(), (unsigned long)gettid());
+
+       tid_enqueuer = malloc(sizeof(*tid_enqueuer) * nr_enqueuers);
+       tid_dequeuer = malloc(sizeof(*tid_dequeuer) * nr_dequeuers);
+       count_enqueuer = malloc(sizeof(*count_enqueuer) * nr_enqueuers);
+       count_dequeuer = malloc(sizeof(*count_dequeuer) * nr_dequeuers);
+       init_queue (&q);
+
+       next_aff = 0;
+
+       for (i = 0; i < nr_enqueuers; i++) {
+               err = pthread_create(&tid_enqueuer[i], NULL, thr_enqueuer,
+                                    &count_enqueuer[i]);
+               if (err != 0)
+                       exit(1);
+       }
+       for (i = 0; i < nr_dequeuers; i++) {
+               err = pthread_create(&tid_dequeuer[i], NULL, thr_dequeuer,
+                                    &count_dequeuer[i]);
+               if (err != 0)
+                       exit(1);
+       }
+
+       smp_mb();
+
+       test_go = 1;
+
+       for (i = 0; i < duration; i++) {
+               sleep(1);
+               if (verbose_mode)
+                       write (1, ".", 1);
+       }
+
+       test_stop = 1;
+
+       for (i = 0; i < nr_enqueuers; i++) {
+               err = pthread_join(tid_enqueuer[i], &tret);
+               if (err != 0)
+                       exit(1);
+               tot_enqueues += count_enqueuer[i];
+       }
+       for (i = 0; i < nr_dequeuers; i++) {
+               err = pthread_join(tid_dequeuer[i], &tret);
+               if (err != 0)
+                       exit(1);
+               tot_dequeues += count_dequeuer[i];
+       }
+       
+       printf_verbose("total number of enqueues : %llu, dequeues %llu\n", tot_enqueues,
+              tot_dequeues);
+       printf("SUMMARY %-25s testdur %4lu nr_enqueuers %3u wdelay %6lu "
+               "nr_dequeuers %3u "
+               "rdur %6lu nr_enqueues %12llu nr_dequeues %12llu successful %12lu nr_ops %12llu\n",
+               argv[0], duration, nr_enqueuers, wdelay,
+               nr_dequeuers, rduration, tot_enqueues, tot_dequeues,
+               nr_successful_dequeues, tot_enqueues + tot_dequeues);
+
+       test_end(&q);
+       free(tid_enqueuer);
+       free(tid_dequeuer);
+       free(count_enqueuer);
+       free(count_dequeuer);
+       return 0;
+}
diff --git a/tests/test_urcu_lfq.c b/tests/test_urcu_lfq.c
new file mode 100644 (file)
index 0000000..90587b7
--- /dev/null
@@ -0,0 +1,585 @@
+/*
+ * test_urcu.c
+ *
+ * Userspace RCU library - example RCU-based lock-free queue
+ *
+ * Copyright February 2010 - Paolo Bonzini <pbonzinI@redhat.com>
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License as published by
+ * the Free Software Foundation; either version 2 of the License, or
+ * (at your option) any later version.
+ *
+ * This program 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 General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License along
+ * with this program; if not, write to the Free Software Foundation, Inc.,
+ * 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+ */
+
+#define _GNU_SOURCE
+#include "../config.h"
+#include <stdio.h>
+#include <pthread.h>
+#include <stdlib.h>
+#include <stdint.h>
+#include <stdbool.h>
+#include <string.h>
+#include <sys/types.h>
+#include <sys/wait.h>
+#include <unistd.h>
+#include <stdio.h>
+#include <assert.h>
+#include <sys/syscall.h>
+#include <sched.h>
+#include <errno.h>
+
+#include <urcu/arch.h>
+
+/* hardcoded number of CPUs */
+#define NR_CPUS 16384
+
+#if defined(_syscall0)
+_syscall0(pid_t, gettid)
+#elif defined(__NR_gettid)
+static inline pid_t gettid(void)
+{
+       return syscall(__NR_gettid);
+}
+#else
+#warning "use pid as tid"
+static inline pid_t gettid(void)
+{
+       return getpid();
+}
+#endif
+
+#ifndef DYNAMIC_LINK_TEST
+#define _LGPL_SOURCE
+#endif
+#include <urcu.h>
+
+
+static volatile int test_go, test_stop;
+
+static unsigned long rduration;
+
+static unsigned long duration;
+
+/* read-side C.S. duration, in loops */
+static unsigned long wdelay;
+
+static inline void loop_sleep(unsigned long l)
+{
+       while(l-- != 0)
+               cpu_relax();
+}
+
+static int verbose_mode;
+
+#define printf_verbose(fmt, args...)           \
+       do {                                    \
+               if (verbose_mode)               \
+                       printf(fmt, args);      \
+       } while (0)
+
+static unsigned int cpu_affinities[NR_CPUS];
+static unsigned int next_aff = 0;
+static int use_affinity = 0;
+
+pthread_mutex_t affinity_mutex = PTHREAD_MUTEX_INITIALIZER;
+
+#ifndef HAVE_CPU_SET_T
+typedef unsigned long cpu_set_t;
+# define CPU_ZERO(cpuset) do { *(cpuset) = 0; } while(0)
+# define CPU_SET(cpu, cpuset) do { *(cpuset) |= (1UL << (cpu)); } while(0)
+#endif
+
+static void set_affinity(void)
+{
+       cpu_set_t mask;
+       int cpu;
+       int ret;
+
+       if (!use_affinity)
+               return;
+
+#if HAVE_SCHED_SETAFFINITY
+       ret = pthread_mutex_lock(&affinity_mutex);
+       if (ret) {
+               perror("Error in pthread mutex lock");
+               exit(-1);
+       }
+       cpu = cpu_affinities[next_aff++];
+       ret = pthread_mutex_unlock(&affinity_mutex);
+       if (ret) {
+               perror("Error in pthread mutex unlock");
+               exit(-1);
+       }
+
+       CPU_ZERO(&mask);
+       CPU_SET(cpu, &mask);
+#if SCHED_SETAFFINITY_ARGS == 2
+       sched_setaffinity(0, &mask);
+#else
+       sched_setaffinity(0, sizeof(mask), &mask);
+#endif
+#endif /* HAVE_SCHED_SETAFFINITY */
+}
+
+/*
+ * returns 0 if test should end.
+ */
+static int test_duration_dequeue(void)
+{
+       return !test_stop;
+}
+
+static int test_duration_enqueue(void)
+{
+       return !test_stop;
+}
+
+static unsigned long long __thread nr_dequeues;
+static unsigned long long __thread nr_enqueues;
+
+static unsigned long nr_successful_dequeues;
+static unsigned int nr_enqueuers;
+static unsigned int nr_dequeuers;
+
+
+#define ARRAY_POISON   0xDEADBEEF
+#define PAGE_SIZE      4096
+#define PAGE_MASK      (PAGE_SIZE - 1)
+#define NODES_PER_PAGE (PAGE_SIZE - 16) / sizeof (struct node)
+
+/* Lock-free queue, using the RCU to avoid the ABA problem and (more
+   interestingly) to efficiently handle freeing memory.
+
+   We have to protect both the enqueuer and dequeuer's compare-and-
+   exchange operation from running across a free and a subsequent
+   reallocation of the same memory.  So, we protect the free with
+   synchronize_rcu; this is enough because all the allocations take
+   place before the compare-and-exchange ops.
+
+   Besides adding rcu_read_{,un}lock, the enqueue/dequeue are a standard
+   implementation of a lock-free-queue.  The first node in the queue is
+   always dummy: dequeuing returns the data from HEAD->NEXT, advances
+   HEAD to HEAD->NEXT (which will now serve as dummy node), and frees the
+   old HEAD.  Since RCU avoids the ABA problem, it doesn't use double-word
+   compare-and-exchange operations.  Node allocation and deallocation are
+   a "black box" and synchronize_rcu is hidden within node deallocation.
+
+   So, the tricky part is finding a good allocation strategy for nodes.
+   The allocator for nodes is shared by multiple threads, and since
+   malloc/free are not lock-free a layer above them is obviously
+   necessary: otherwise the whole exercise is useless.  In addition,
+   to avoid penalizing dequeues, the allocator should avoid frequent
+   synchronization (because synchronize_rcu is expensive).
+
+   The scheme that is used here uses a page as the allocation
+   unit for nodes.  A page is freed when no more nodes are in use.
+   Nodes from a page are never reused.
+
+   The nodes are allocated from Q->CURRENT.  Since whoever finds a full
+   page has to busy wait, we use a trick to limit the duration of busy
+   waiting.  A free page Q->FREE is always kept ready, so that any thread
+   that allocates the last node in a page, or finds a full page can try
+   to update Q->CURRENT.  Whoever loses the race has to busy wait, OTOH
+   whoever wins the race has to allocate the new Q->FREE.  In other words,  
+   if the following sequence happens
+
+     Thread 1                    Thread 2                 other threads
+     -----------------------------------------------------------------------
+     Get last node from page
+                                q->current = q->free;
+                                                         fill up q->current
+     q->current = q->free fails
+
+   then thread 1 does not have anymore the duty of allocating q->current;
+   thread 2 will do that.  If a thread finds a full current page and
+   Q->CURRENT == Q->FREE, this means that another thread is going to
+   allocate Q->FREE soon, and it busy waits.  After the allocation
+   finishes, everything proceeds normally: some thread will take care
+   of setting Q->CURRENT and allocating a new Q->FREE.
+   One common scheme for allocation is to use a free list (implemented
+   as a lock-free stack), but this free list is potentially unbounded.
+   Instead, with the above scheme the number of live pages at any time
+   is equal to the number of enqueuing threads.  */
+
+struct node {
+       void *data;
+       void *next;
+};
+
+struct node_page {
+       int in;
+       int out;
+       char padding[16 - sizeof(int) * 2];
+       struct node nodes[NODES_PER_PAGE];
+};
+
+
+struct queue {
+       struct node_page *current, *free;
+       struct node *head, *tail;
+};
+
+static struct node_page *new_node_page()
+{
+       struct node_page *p = valloc (PAGE_SIZE);
+       p->in = p->out = 0;
+       return p;
+}
+
+static void free_node_page(struct node_page *p)
+{
+       /* Help making sure that accessing a dangling pointer is
+          adequately punished.  */
+       p->in = ARRAY_POISON;
+       free (p);
+}
+
+static struct node *new_node(struct queue *q)
+{
+       struct node *n;
+       struct node_page *p;
+       int i;
+
+       do {
+               p = q->current;
+               i = p->in;
+               if (i >= NODES_PER_PAGE - 1 &&
+                   q->free != p &&
+                   uatomic_cmpxchg(&q->current, p, q->free) == p)
+                       q->free = new_node_page();
+
+       } while (i == NODES_PER_PAGE || uatomic_cmpxchg(&p->in, i, i+1) != i);
+
+       assert (i >= 0 && i < NODES_PER_PAGE);
+       n = &p->nodes[i];
+       n->next = NULL;
+       return n;
+}
+
+void free_node(struct node *n)
+{
+       struct node_page *p = (struct node_page *) ((intptr_t) n & ~PAGE_MASK);
+
+       if (uatomic_add_return(&p->out, 1) == NODES_PER_PAGE) {
+               synchronize_rcu();
+               free_node_page(p);
+       }
+}
+
+void init_queue(struct queue *q)
+{
+       q->current = new_node_page();
+       q->free = new_node_page();
+       q->head = q->tail = new_node(q);
+}
+
+void enqueue(struct queue *q, void *value)
+{
+       struct node *n = new_node(q);
+       n->data = value;
+       rcu_read_lock();
+       for (;;) {
+               struct node *tail = rcu_dereference(q->tail);
+               struct node *next = rcu_dereference(tail->next);
+               if (tail != q->tail) {
+                       /* A change occurred while reading the values.  */
+                       continue;
+               }
+
+               if (next) {
+                       /* Help moving tail further.  */
+                       uatomic_cmpxchg(&q->tail, tail, next);
+                       continue;
+               }
+
+               if (uatomic_cmpxchg(&tail->next, NULL, n) == NULL) {
+                       /* Move tail (another operation might beat us to it,
+                          that's fine).  */
+                       uatomic_cmpxchg(&q->tail, tail, n);
+                       rcu_read_unlock();
+                       return;
+               }
+       }
+}
+
+void *dequeue(struct queue *q, bool *not_empty)
+{
+       bool dummy;
+       if (!not_empty)
+               not_empty = &dummy;
+
+       rcu_read_lock();
+       *not_empty = false;
+       for (;;) {
+               void *data;
+               struct node *head = rcu_dereference(q->head);
+               struct node *tail = rcu_dereference(q->tail);
+               struct node *next = rcu_dereference(head->next);
+
+               if (head != q->head) {
+                       /* A change occurred while reading the values.  */
+                       continue;
+               }
+
+               if (head == tail) {
+                       /* If all three are consistent, the queue is empty.  */
+                       if (!next)
+                               return NULL;
+
+                       /* Help moving tail further.  */
+                       uatomic_cmpxchg(&q->tail, tail, next);
+                       continue;
+               }
+
+               data = next->data;
+               if (uatomic_cmpxchg(&q->head, head, next) == head) {
+                       /* Next remains as a dummy node, head is freed.  */
+                       rcu_read_unlock();
+                       *not_empty = true;
+                       free_node (head);
+                       return data;
+               }
+       }
+}
+
+\f
+static struct queue q;
+
+void *thr_enqueuer(void *_count)
+{
+       unsigned long long *count = _count;
+
+       printf_verbose("thread_begin %s, thread id : %lx, tid %lu\n",
+                       "enqueuer", pthread_self(), (unsigned long)gettid());
+
+       set_affinity();
+
+       rcu_register_thread();
+
+       while (!test_go)
+       {
+       }
+       smp_mb();
+
+       for (;;) {
+               enqueue (&q, NULL);
+
+               if (unlikely(wdelay))
+                       loop_sleep(wdelay);
+               nr_enqueues++;
+               if (unlikely(!test_duration_enqueue()))
+                       break;
+       }
+
+       rcu_unregister_thread();
+
+       *count = nr_enqueues;
+       printf_verbose("thread_end %s, thread id : %lx, tid %lu - count %d\n",
+                      "enqueuer", pthread_self(), (unsigned long)gettid(), nr_enqueues);
+       return ((void*)1);
+
+}
+
+void *thr_dequeuer(void *_count)
+{
+       unsigned long long *count = _count;
+
+       printf_verbose("thread_begin %s, thread id : %lx, tid %lu\n",
+                       "dequeuer", pthread_self(), (unsigned long)gettid());
+
+       set_affinity();
+
+       rcu_register_thread();
+
+       while (!test_go)
+       {
+       }
+       smp_mb();
+
+       for (;;) {
+               bool not_empty;
+               dequeue (&q, &not_empty);
+               if (not_empty)
+                       uatomic_inc (&nr_successful_dequeues);
+
+               nr_dequeues++;
+               if (unlikely(!test_duration_dequeue()))
+                       break;
+               if (unlikely(rduration))
+                       loop_sleep(rduration);
+       }
+
+       rcu_unregister_thread();
+
+       printf_verbose("thread_end %s, thread id : %lx, tid %lu - count %d\n",
+                       "dequeuer", pthread_self(), (unsigned long)gettid(), nr_dequeues);
+       *count = nr_dequeues;
+       return ((void*)2);
+}
+
+void test_end(struct queue *q)
+{
+       bool not_empty;
+       do
+               dequeue (q, &not_empty);
+       while (!not_empty);
+       if (q->current != q->free)
+               free_node_page(q->free);
+       free_node_page(q->current);
+}
+
+void show_usage(int argc, char **argv)
+{
+       printf("Usage : %s nr_dequeuers nr_enqueuers duration (s)", argv[0]);
+       printf(" [-d delay] (enqueuer period (in loops))");
+       printf(" [-c duration] (dequeuer period (in loops))");
+       printf(" [-v] (verbose output)");
+       printf(" [-a cpu#] [-a cpu#]... (affinity)");
+       printf("\n");
+}
+
+int main(int argc, char **argv)
+{
+       int err;
+       pthread_t *tid_enqueuer, *tid_dequeuer;
+       void *tret;
+       unsigned long long *count_enqueuer, *count_dequeuer;
+       unsigned long long tot_enqueues = 0, tot_dequeues = 0;
+       int i, a;
+
+       if (argc < 4) {
+               show_usage(argc, argv);
+               return -1;
+       }
+
+       err = sscanf(argv[1], "%u", &nr_dequeuers);
+       if (err != 1) {
+               show_usage(argc, argv);
+               return -1;
+       }
+
+       err = sscanf(argv[2], "%u", &nr_enqueuers);
+       if (err != 1) {
+               show_usage(argc, argv);
+               return -1;
+       }
+       
+       err = sscanf(argv[3], "%lu", &duration);
+       if (err != 1) {
+               show_usage(argc, argv);
+               return -1;
+       }
+
+       for (i = 4; i < argc; i++) {
+               if (argv[i][0] != '-')
+                       continue;
+               switch (argv[i][1]) {
+               case 'a':
+                       if (argc < i + 2) {
+                               show_usage(argc, argv);
+                               return -1;
+                       }
+                       a = atoi(argv[++i]);
+                       cpu_affinities[next_aff++] = a;
+                       use_affinity = 1;
+                       printf_verbose("Adding CPU %d affinity\n", a);
+                       break;
+               case 'c':
+                       if (argc < i + 2) {
+                               show_usage(argc, argv);
+                               return -1;
+                       }
+                       rduration = atol(argv[++i]);
+                       break;
+               case 'd':
+                       if (argc < i + 2) {
+                               show_usage(argc, argv);
+                               return -1;
+                       }
+                       wdelay = atol(argv[++i]);
+                       break;
+               case 'v':
+                       verbose_mode = 1;
+                       break;
+               }
+       }
+
+       printf_verbose("running test for %lu seconds, %u enqueuers, %u dequeuers.\n",
+               duration, nr_enqueuers, nr_dequeuers);
+       printf_verbose("Writer delay : %lu loops.\n", rduration);
+       printf_verbose("Reader duration : %lu loops.\n", wdelay);
+       printf_verbose("thread %-6s, thread id : %lx, tid %lu\n",
+                       "main", pthread_self(), (unsigned long)gettid());
+
+       tid_enqueuer = malloc(sizeof(*tid_enqueuer) * nr_enqueuers);
+       tid_dequeuer = malloc(sizeof(*tid_dequeuer) * nr_dequeuers);
+       count_enqueuer = malloc(sizeof(*count_enqueuer) * nr_enqueuers);
+       count_dequeuer = malloc(sizeof(*count_dequeuer) * nr_dequeuers);
+       init_queue (&q);
+
+       next_aff = 0;
+
+       for (i = 0; i < nr_enqueuers; i++) {
+               err = pthread_create(&tid_enqueuer[i], NULL, thr_enqueuer,
+                                    &count_enqueuer[i]);
+               if (err != 0)
+                       exit(1);
+       }
+       for (i = 0; i < nr_dequeuers; i++) {
+               err = pthread_create(&tid_dequeuer[i], NULL, thr_dequeuer,
+                                    &count_dequeuer[i]);
+               if (err != 0)
+                       exit(1);
+       }
+
+       smp_mb();
+
+       test_go = 1;
+
+       for (i = 0; i < duration; i++) {
+               sleep(1);
+               if (verbose_mode)
+                       write (1, ".", 1);
+       }
+
+       test_stop = 1;
+
+       for (i = 0; i < nr_enqueuers; i++) {
+               err = pthread_join(tid_enqueuer[i], &tret);
+               if (err != 0)
+                       exit(1);
+               tot_enqueues += count_enqueuer[i];
+       }
+       for (i = 0; i < nr_dequeuers; i++) {
+               err = pthread_join(tid_dequeuer[i], &tret);
+               if (err != 0)
+                       exit(1);
+               tot_dequeues += count_dequeuer[i];
+       }
+       
+       printf_verbose("total number of enqueues : %llu, dequeues %llu\n", tot_enqueues,
+              tot_dequeues);
+       printf("SUMMARY %-25s testdur %4lu nr_enqueuers %3u wdelay %6lu "
+               "nr_dequeuers %3u "
+               "rdur %6lu nr_enqueues %12llu nr_dequeues %12llu successful %12lu nr_ops %12llu\n",
+               argv[0], duration, nr_enqueuers, wdelay,
+               nr_dequeuers, rduration, tot_enqueues, tot_dequeues,
+               nr_successful_dequeues, tot_enqueues + tot_dequeues);
+
+       test_end(&q);
+       free(tid_enqueuer);
+       free(tid_dequeuer);
+       free(count_enqueuer);
+       free(count_dequeuer);
+       return 0;
+}
This page took 0.067941 seconds and 4 git commands to generate.