Add QSBR RCU timing tests
authorMathieu Desnoyers <mathieu.desnoyers@polymtl.ca>
Mon, 8 Jun 2009 18:41:37 +0000 (14:41 -0400)
committerMathieu Desnoyers <mathieu.desnoyers@polymtl.ca>
Mon, 8 Jun 2009 18:41:37 +0000 (14:41 -0400)
urcu-qsbr can be used for QSBR-based RCU.

Currently not built as a full-blown library .so, but could.

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@polymtl.ca>
Makefile
test_qsbr_timing.c [new file with mode: 0644]
urcu-qsbr.h [new file with mode: 0644]

index c02252289411a959c811b95f0b7639a1e59c7cea..371c107d1300c99b4cfabd5b2ced189d13363abc 100644 (file)
--- a/Makefile
+++ b/Makefile
@@ -13,7 +13,7 @@ SRC_DEP=`echo $^ | sed 's/[^ ]*\.h//g'`
 
 all: arch-api test_urcu test_urcu_dynamic_link test_urcu_timing \
        test_rwlock_timing test_perthreadlock_timing test_urcu_yield urcu-asm.S \
-       urcu-asm.o urcutorture urcutorture-yield liburcu.so
+       test_qsbr_timing urcu-asm.o urcutorture urcutorture-yield liburcu.so
 
 arch-api: api.h arch.h
        # Run either make pthreads-x86 or make pthreads-ppc prior to build
@@ -42,6 +42,9 @@ test_urcu_yield: urcu-yield.o test_urcu.c urcu.h
 test_urcu_timing: urcu.o test_urcu_timing.c urcu.h
        $(CC) ${CFLAGS} $(LDFLAGS) -o $@ $(SRC_DEP)
 
+test_qsbr_timing: urcu-qsbr.o test_qsbr_timing.c urcu-qsbr.h
+       $(CC) ${CFLAGS} $(LDFLAGS) -o $@ $(SRC_DEP)
+
 test_rwlock_timing: urcu.o test_rwlock_timing.c urcu.h
        $(CC) ${CFLAGS} $(LDFLAGS) -o $@ $(SRC_DEP)
 
@@ -51,6 +54,9 @@ test_perthreadlock_timing: urcu.o test_perthreadlock_timing.c urcu.h
 urcu.o: urcu.c urcu.h
        $(CC) -fPIC ${CFLAGS} $(LDFLAGS) -c -o $@ $(SRC_DEP)
 
+urcu-qsbr.o: urcu-qsbr.c urcu-qsbr.h
+       $(CC) -fPIC ${CFLAGS} $(LDFLAGS) -c -o $@ $(SRC_DEP)
+
 liburcu.so: urcu.o
        $(CC) -fPIC -shared -o $@ $<
 
diff --git a/test_qsbr_timing.c b/test_qsbr_timing.c
new file mode 100644 (file)
index 0000000..bdbb581
--- /dev/null
@@ -0,0 +1,217 @@
+/*
+ * test_qsbr_timing.c
+ *
+ * Userspace QSBR - test program
+ *
+ * Copyright February 2009 - Mathieu Desnoyers <mathieu.desnoyers@polymtl.ca>
+ *
+ * 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.
+ */
+
+#include <stdio.h>
+#include <pthread.h>
+#include <stdlib.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 <arch.h>
+
+#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
+
+#define _LGPL_SOURCE
+#include "urcu-qsbr.h"
+
+pthread_mutex_t rcu_copy_mutex = PTHREAD_MUTEX_INITIALIZER;
+
+void rcu_copy_mutex_lock(void)
+{
+       int ret;
+       ret = pthread_mutex_lock(&rcu_copy_mutex);
+       if (ret) {
+               perror("Error in pthread mutex lock");
+               exit(-1);
+       }
+}
+
+void rcu_copy_mutex_unlock(void)
+{
+       int ret;
+
+       ret = pthread_mutex_unlock(&rcu_copy_mutex);
+       if (ret) {
+               perror("Error in pthread mutex unlock");
+               exit(-1);
+       }
+}
+
+struct test_array {
+       int a;
+};
+
+static struct test_array *test_rcu_pointer;
+
+#define OUTER_READ_LOOP        2000U
+#define INNER_READ_LOOP        100000U
+#define READ_LOOP ((unsigned long long)OUTER_READ_LOOP * INNER_READ_LOOP)
+
+#define OUTER_WRITE_LOOP 10U
+#define INNER_WRITE_LOOP 200U
+#define WRITE_LOOP ((unsigned long long)OUTER_WRITE_LOOP * INNER_WRITE_LOOP)
+
+#define NR_READ 10
+#define NR_WRITE 9
+
+static cycles_t reader_time[NR_READ] __attribute__((aligned(128)));
+static cycles_t writer_time[NR_WRITE] __attribute__((aligned(128)));
+
+void *thr_reader(void *arg)
+{
+       int i, j;
+       struct test_array *local_ptr;
+       cycles_t time1, time2;
+
+       printf("thread_begin %s, thread id : %lx, tid %lu\n",
+                       "reader", pthread_self(), (unsigned long)gettid());
+       sleep(2);
+
+       rcu_register_thread();
+
+       time1 = get_cycles();
+       for (i = 0; i < OUTER_READ_LOOP; i++) {
+               for (j = 0; j < INNER_READ_LOOP; j++) {
+                       _rcu_read_lock();
+                       local_ptr = _rcu_dereference(test_rcu_pointer);
+                       if (local_ptr) {
+                               assert(local_ptr->a == 8);
+                       }
+                       _rcu_read_unlock();
+               }
+               _rcu_quiescent_state();
+       }
+       time2 = get_cycles();
+
+       rcu_unregister_thread();
+
+       reader_time[(unsigned long)arg] = time2 - time1;
+
+       sleep(2);
+       printf("thread_end %s, thread id : %lx, tid %lu\n",
+                       "reader", pthread_self(), (unsigned long)gettid());
+       return ((void*)1);
+
+}
+
+void *thr_writer(void *arg)
+{
+       int i, j;
+       struct test_array *new, *old;
+       cycles_t time1, time2;
+
+       printf("thread_begin %s, thread id : %lx, tid %lu\n",
+                       "writer", pthread_self(), (unsigned long)gettid());
+       sleep(2);
+
+       for (i = 0; i < OUTER_WRITE_LOOP; i++) {
+               for (j = 0; j < INNER_WRITE_LOOP; j++) {
+                       time1 = get_cycles();
+                       new = malloc(sizeof(struct test_array));
+                       rcu_copy_mutex_lock();
+                       old = test_rcu_pointer;
+                       if (old) {
+                               assert(old->a == 8);
+                       }
+                       new->a = 8;
+                       old = _rcu_publish_content(&test_rcu_pointer, new);
+                       rcu_copy_mutex_unlock();
+                       /* can be done after unlock */
+                       if (old) {
+                               old->a = 0;
+                       }
+                       free(old);
+                       time2 = get_cycles();
+                       writer_time[(unsigned long)arg] += time2 - time1;
+                       usleep(1);
+               }
+       }
+
+       printf("thread_end %s, thread id : %lx, tid %lu\n",
+                       "writer", pthread_self(), (unsigned long)gettid());
+       return ((void*)2);
+}
+
+int main()
+{
+       int err;
+       pthread_t tid_reader[NR_READ], tid_writer[NR_WRITE];
+       void *tret;
+       int i;
+       cycles_t tot_rtime = 0;
+       cycles_t tot_wtime = 0;
+
+       printf("thread %-6s, thread id : %lx, tid %lu\n",
+                       "main", pthread_self(), (unsigned long)gettid());
+
+       for (i = 0; i < NR_READ; i++) {
+               err = pthread_create(&tid_reader[i], NULL, thr_reader,
+                                    (void *)(long)i);
+               if (err != 0)
+                       exit(1);
+       }
+       for (i = 0; i < NR_WRITE; i++) {
+               err = pthread_create(&tid_writer[i], NULL, thr_writer,
+                                    (void *)(long)i);
+               if (err != 0)
+                       exit(1);
+       }
+
+       sleep(10);
+
+       for (i = 0; i < NR_READ; i++) {
+               err = pthread_join(tid_reader[i], &tret);
+               if (err != 0)
+                       exit(1);
+               tot_rtime += reader_time[i];
+       }
+       for (i = 0; i < NR_WRITE; i++) {
+               err = pthread_join(tid_writer[i], &tret);
+               if (err != 0)
+                       exit(1);
+               tot_wtime += writer_time[i];
+       }
+       free(test_rcu_pointer);
+       printf("Time per read : %g cycles\n",
+              (double)tot_rtime / ((double)NR_READ * (double)READ_LOOP));
+       printf("Time per write : %g cycles\n",
+              (double)tot_wtime / ((double)NR_WRITE * (double)WRITE_LOOP));
+
+       return 0;
+}
diff --git a/urcu-qsbr.h b/urcu-qsbr.h
new file mode 100644 (file)
index 0000000..101b94d
--- /dev/null
@@ -0,0 +1,284 @@
+#ifndef _URCU_QSBR_H
+#define _URCU_QSBR_H
+
+/*
+ * urcu-static.h
+ *
+ * Userspace RCU header.
+ *
+ * TO BE INCLUDED ONLY IN LGPL-COMPATIBLE CODE. See urcu.h for linking
+ * dynamically with the userspace rcu library.
+ *
+ * Copyright (c) 2009 Mathieu Desnoyers <mathieu.desnoyers@polymtl.ca>
+ * Copyright (c) 2009 Paul E. McKenney, IBM Corporation.
+ *
+ * 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
+ *
+ * IBM's contributions to this file may be relicensed under LGPLv2 or later.
+ */
+
+#include <stdlib.h>
+#include <pthread.h>
+
+#include <compiler.h>
+#include <arch.h>
+
+/*
+ * Identify a shared load. A smp_rmc() or smp_mc() should come before the load.
+ */
+#define _LOAD_SHARED(p)               ACCESS_ONCE(p)
+
+/*
+ * Load a data from shared memory, doing a cache flush if required.
+ */
+#define LOAD_SHARED(p)                 \
+       ({                              \
+               smp_rmc();              \
+               _LOAD_SHARED(p);        \
+       })
+
+/*
+ * Identify a shared store. A smp_wmc() or smp_mc() should follow the store.
+ */
+#define _STORE_SHARED(x, v)    ({ ACCESS_ONCE(x) = (v); })
+
+/*
+ * Store v into x, where x is located in shared memory. Performs the required
+ * cache flush after writing. Returns v.
+ */
+#define STORE_SHARED(x, v)             \
+       ({                              \
+               _STORE_SHARED(x, v);    \
+               smp_wmc();              \
+               (v);                    \
+       })
+
+/**
+ * _rcu_dereference - reads (copy) a RCU-protected pointer to a local variable
+ * into a RCU read-side critical section. The pointer can later be safely
+ * dereferenced within the critical section.
+ *
+ * This ensures that the pointer copy is invariant thorough the whole critical
+ * section.
+ *
+ * Inserts memory barriers on architectures that require them (currently only
+ * Alpha) and documents which pointers are protected by RCU.
+ *
+ * Should match rcu_assign_pointer() or rcu_xchg_pointer().
+ */
+
+#define _rcu_dereference(p)     ({                                     \
+                               typeof(p) _________p1 = LOAD_SHARED(p); \
+                               smp_read_barrier_depends();             \
+                               (_________p1);                          \
+                               })
+
+/*
+ * This code section can only be included in LGPL 2.1 compatible source code.
+ * See below for the function call wrappers which can be used in code meant to
+ * be only linked with the Userspace RCU library. This comes with a small
+ * performance degradation on the read-side due to the added function calls.
+ * This is required to permit relinking with newer versions of the library.
+ */
+
+/*
+ * The signal number used by the RCU library can be overridden with
+ * -DSIGURCU= when compiling the library.
+ */
+#ifndef SIGURCU
+#define SIGURCU SIGUSR1
+#endif
+
+/*
+ * If a reader is really non-cooperative and refuses to commit its
+ * urcu_active_readers count to memory (there is no barrier in the reader
+ * per-se), kick it after a few loops waiting for it.
+ */
+#define KICK_READER_LOOPS 10000
+
+#ifdef DEBUG_YIELD
+#include <sched.h>
+#include <time.h>
+#include <pthread.h>
+#include <unistd.h>
+
+#define YIELD_READ     (1 << 0)
+#define YIELD_WRITE    (1 << 1)
+
+/* Updates without DEBUG_FULL_MB are much slower. Account this in the delay */
+#ifdef DEBUG_FULL_MB
+/* maximum sleep delay, in us */
+#define MAX_SLEEP 50
+#else
+#define MAX_SLEEP 30000
+#endif
+
+extern unsigned int yield_active;
+extern unsigned int __thread rand_yield;
+
+static inline void debug_yield_read(void)
+{
+       if (yield_active & YIELD_READ)
+               if (rand_r(&rand_yield) & 0x1)
+                       usleep(rand_r(&rand_yield) % MAX_SLEEP);
+}
+
+static inline void debug_yield_write(void)
+{
+       if (yield_active & YIELD_WRITE)
+               if (rand_r(&rand_yield) & 0x1)
+                       usleep(rand_r(&rand_yield) % MAX_SLEEP);
+}
+
+static inline void debug_yield_init(void)
+{
+       rand_yield = time(NULL) ^ pthread_self();
+}
+#else
+static inline void debug_yield_read(void)
+{
+}
+
+static inline void debug_yield_write(void)
+{
+}
+
+static inline void debug_yield_init(void)
+{
+
+}
+#endif
+
+#ifdef DEBUG_FULL_MB
+static inline void reader_barrier()
+{
+       smp_mb();
+}
+#else
+static inline void reader_barrier()
+{
+       barrier();
+}
+#endif
+
+/*
+ * The trick here is that RCU_GP_CTR_BIT must be a multiple of 8 so we can use a
+ * full 8-bits, 16-bits or 32-bits bitmask for the lower order bits.
+ */
+#define RCU_GP_COUNT           (1UL << 0)
+/* Use the amount of bits equal to half of the architecture long size */
+#define RCU_GP_CTR_BIT         (1UL << (sizeof(long) << 2))
+#define RCU_GP_CTR_NEST_MASK   (RCU_GP_CTR_BIT - 1)
+
+/*
+ * Global quiescent period counter with low-order bits unused.
+ * Using a int rather than a char to eliminate false register dependencies
+ * causing stalls on some architectures.
+ */
+extern long urcu_gp_ctr;
+
+extern long __thread urcu_active_readers;
+
+static inline int rcu_old_gp_ongoing(long *value)
+{
+       if (value == NULL)
+               return 0;
+
+       return LOAD_SHARED(*value) & 1;
+}
+
+static inline void _rcu_read_lock(void)
+{
+}
+
+static inline void _rcu_read_unlock(void)
+{
+}
+
+static inline void _rcu_quiescent_state(void)
+{
+       smp_mb();       
+       urcu_active_readers = ACCESS_ONCE(urcu_gp_ctr) + 1;
+       smp_mb();
+}
+
+static inline void _rcu_thread_offline(void)
+{
+       smp_mb();
+       urcu_active_readers = ACCESS_ONCE(urcu_gp_ctr);
+}
+
+static inline void _rcu_thread_online(void)
+{
+       urcu_active_readers = ACCESS_ONCE(urcu_gp_ctr) + 1;
+       smp_mb();
+}
+
+/**
+ * _rcu_assign_pointer - assign (publicize) a pointer to a new data structure
+ * meant to be read by RCU read-side critical sections. Returns the assigned
+ * value.
+ *
+ * Documents which pointers will be dereferenced by RCU read-side critical
+ * sections and adds the required memory barriers on architectures requiring
+ * them. It also makes sure the compiler does not reorder code initializing the
+ * data structure before its publication.
+ *
+ * Should match rcu_dereference_pointer().
+ */
+
+#define _rcu_assign_pointer(p, v)                      \
+       ({                                              \
+               if (!__builtin_constant_p(v) ||         \
+                   ((v) != NULL))                      \
+                       wmb();                          \
+               STORE_SHARED(p, v);                     \
+       })
+
+/**
+ * _rcu_xchg_pointer - same as rcu_assign_pointer, but returns the previous
+ * pointer to the data structure, which can be safely freed after waiting for a
+ * quiescent state using synchronize_rcu().
+ */
+
+#define _rcu_xchg_pointer(p, v)                                \
+       ({                                              \
+               if (!__builtin_constant_p(v) ||         \
+                   ((v) != NULL))                      \
+                       wmb();                          \
+               xchg(p, v);                             \
+       })
+
+/*
+ * Exchanges the pointer and waits for quiescent state.
+ * The pointer returned can be freed.
+ */
+#define _rcu_publish_content(p, v)                     \
+       ({                                              \
+               void *oldptr;                           \
+               oldptr = _rcu_xchg_pointer(p, v);       \
+               synchronize_rcu();                      \
+               oldptr;                                 \
+       })
+
+extern void synchronize_rcu(void);
+
+/*
+ * Reader thread registration.
+ */
+extern void rcu_register_thread(void);
+extern void rcu_unregister_thread(void);
+
+#endif /* _URCU_QSBR_H */
This page took 0.030903 seconds and 4 git commands to generate.