Restartable sequences: self-tests
authorMathieu Desnoyers <mathieu.desnoyers@efficios.com>
Mon, 30 May 2016 07:46:53 +0000 (09:46 +0200)
committerMathieu Desnoyers <mathieu.desnoyers@efficios.com>
Fri, 21 Oct 2016 22:28:16 +0000 (18:28 -0400)
Implements two basic tests of RSEQ functionality, and one more
exhaustive parameterizable test.

The first, "basic_test" only asserts that RSEQ works moderately
correctly.
E.g. that:
- The CPUID pointer works
- Code infinitely looping within a critical section will eventually be
  interrupted.
- Critical sections are interrupted by signals.

"basic_percpu_ops_test" is a slightly more "realistic" variant,
implementing a few simple per-cpu operations and testing their
correctness.

"param_test" is a parametrizable restartable sequences test. See
the "--help" output for usage.

As part of those tests, a helper library "rseq" implements a user-space
API around restartable sequences. It takes care of ensuring progress in
case of debugger single-stepping with a fall-back to locking, and
exposes the instruction pointer addresses where the rseq assembly blocks
begin and end, as well as the associated abort instruction pointer, in
the __rseq_table section. This section allows debuggers may know where
to place breakpoints when single-stepping through assembly blocks which
may be aborted at any point by the kernel.

The following rseq APIs are implemented in this helper library:
- do_rseq(): Restartable sequence made of zero or more loads, completed
  by a word-sized store,
- do_rseq2(): Restartable sequence made of zero or more loads, one
  speculative word-sized store, completed by a word-sized store,
- do_rseq_memcpy(): Restartable sequence made of zero or more loads,
  a speculative copy of a variable length memory region, completed by a
  word-sized store.

PowerPC tests have been implemented by Boqun Feng.

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
CC: Russell King <linux@arm.linux.org.uk>
CC: Catalin Marinas <catalin.marinas@arm.com>
CC: Will Deacon <will.deacon@arm.com>
CC: Thomas Gleixner <tglx@linutronix.de>
CC: Paul Turner <pjt@google.com>
CC: Andrew Hunter <ahh@google.com>
CC: Peter Zijlstra <peterz@infradead.org>
CC: Andy Lutomirski <luto@amacapital.net>
CC: Andi Kleen <andi@firstfloor.org>
CC: Dave Watson <davejwatson@fb.com>
CC: Chris Lameter <cl@linux.com>
CC: Ingo Molnar <mingo@redhat.com>
CC: "H. Peter Anvin" <hpa@zytor.com>
CC: Ben Maurer <bmaurer@fb.com>
CC: Steven Rostedt <rostedt@goodmis.org>
CC: "Paul E. McKenney" <paulmck@linux.vnet.ibm.com>
CC: Josh Triplett <josh@joshtriplett.org>
CC: Linus Torvalds <torvalds@linux-foundation.org>
CC: Andrew Morton <akpm@linux-foundation.org>
CC: Boqun Feng <boqun.feng@gmail.com>
CC: linux-api@vger.kernel.org
MAINTAINERS
tools/testing/selftests/rseq/.gitignore [new file with mode: 0644]
tools/testing/selftests/rseq/Makefile [new file with mode: 0644]
tools/testing/selftests/rseq/basic_percpu_ops_test.c [new file with mode: 0644]
tools/testing/selftests/rseq/basic_test.c [new file with mode: 0644]
tools/testing/selftests/rseq/param_test.c [new file with mode: 0644]
tools/testing/selftests/rseq/rseq-arm.h [new file with mode: 0644]
tools/testing/selftests/rseq/rseq-ppc.h [new file with mode: 0644]
tools/testing/selftests/rseq/rseq-x86.h [new file with mode: 0644]
tools/testing/selftests/rseq/rseq.c [new file with mode: 0644]
tools/testing/selftests/rseq/rseq.h [new file with mode: 0644]

index 6219dafa4219877234971f68d44e2c3b2a7a0e56..7b116669f89605f3605dc541b9d983867948351f 100644 (file)
@@ -5230,6 +5230,7 @@ L:        linux-kernel@vger.kernel.org
 S:     Supported
 F:     kernel/rseq.c
 F:     include/uapi/linux/rseq.h
+F:     tools/testing/selftests/rseq/
 
 GFS2 FILE SYSTEM
 M:     Steven Whitehouse <swhiteho@redhat.com>
diff --git a/tools/testing/selftests/rseq/.gitignore b/tools/testing/selftests/rseq/.gitignore
new file mode 100644 (file)
index 0000000..2596e26
--- /dev/null
@@ -0,0 +1,3 @@
+basic_percpu_ops_test
+basic_test
+param_test
diff --git a/tools/testing/selftests/rseq/Makefile b/tools/testing/selftests/rseq/Makefile
new file mode 100644 (file)
index 0000000..0082d02
--- /dev/null
@@ -0,0 +1,13 @@
+CFLAGS += -O2 -Wall -g -I./ -I../../../../usr/include/
+LDFLAGS += -lpthread
+
+TESTS = basic_test basic_percpu_ops_test param_test
+
+all: $(TESTS)
+%: %.c rseq.h rseq-*.h rseq.c
+       $(CC) $(CFLAGS) -o $@ $^ $(LDFLAGS)
+
+include ../lib.mk
+
+clean:
+       $(RM) $(TESTS)
diff --git a/tools/testing/selftests/rseq/basic_percpu_ops_test.c b/tools/testing/selftests/rseq/basic_percpu_ops_test.c
new file mode 100644 (file)
index 0000000..8fe4a6b
--- /dev/null
@@ -0,0 +1,286 @@
+#define _GNU_SOURCE
+#include <assert.h>
+#include <pthread.h>
+#include <sched.h>
+#include <stdint.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+
+#include <rseq.h>
+
+static struct rseq_lock rseq_lock;
+
+struct percpu_lock_entry {
+       intptr_t v;
+} __attribute__((aligned(128)));
+
+struct percpu_lock {
+       struct percpu_lock_entry c[CPU_SETSIZE];
+};
+
+struct test_data_entry {
+       intptr_t count;
+} __attribute__((aligned(128)));
+
+struct spinlock_test_data {
+       struct percpu_lock lock;
+       struct test_data_entry c[CPU_SETSIZE];
+       int reps;
+};
+
+struct percpu_list_node {
+       intptr_t data;
+       struct percpu_list_node *next;
+};
+
+struct percpu_list_entry {
+       struct percpu_list_node *head;
+} __attribute__((aligned(128)));
+
+struct percpu_list {
+       struct percpu_list_entry c[CPU_SETSIZE];
+};
+
+/* A simple percpu spinlock.  Returns the cpu lock was acquired on. */
+int rseq_percpu_lock(struct percpu_lock *lock)
+{
+       struct rseq_state rseq_state;
+       intptr_t *targetptr, newval;
+       int cpu;
+       bool result;
+
+       for (;;) {
+               do_rseq(&rseq_lock, rseq_state, cpu, result, targetptr, newval,
+                       {
+                               if (unlikely(lock->c[cpu].v)) {
+                                       result = false;
+                               } else {
+                                       newval = 1;
+                                       targetptr = (intptr_t *)&lock->c[cpu].v;
+                               }
+                       });
+               if (likely(result))
+                       break;
+       }
+       /*
+        * Acquire semantic when taking lock after control dependency.
+        * Matches smp_store_release().
+        */
+       smp_acquire__after_ctrl_dep();
+       return cpu;
+}
+
+void rseq_percpu_unlock(struct percpu_lock *lock, int cpu)
+{
+       assert(lock->c[cpu].v == 1);
+       /*
+        * Release lock, with release semantic. Matches
+        * smp_acquire__after_ctrl_dep().
+        */
+       smp_store_release(&lock->c[cpu].v, 0);
+}
+
+void *test_percpu_spinlock_thread(void *arg)
+{
+       struct spinlock_test_data *data = arg;
+       int i, cpu;
+
+       if (rseq_register_current_thread())
+               abort();
+       for (i = 0; i < data->reps; i++) {
+               cpu = rseq_percpu_lock(&data->lock);
+               data->c[cpu].count++;
+               rseq_percpu_unlock(&data->lock, cpu);
+       }
+       if (rseq_unregister_current_thread())
+               abort();
+
+       return NULL;
+}
+
+/*
+ * A simple test which implements a sharded counter using a per-cpu
+ * lock.  Obviously real applications might prefer to simply use a
+ * per-cpu increment; however, this is reasonable for a test and the
+ * lock can be extended to synchronize more complicated operations.
+ */
+void test_percpu_spinlock(void)
+{
+       const int num_threads = 200;
+       int i;
+       uint64_t sum;
+       pthread_t test_threads[num_threads];
+       struct spinlock_test_data data;
+
+       memset(&data, 0, sizeof(data));
+       data.reps = 5000;
+
+       for (i = 0; i < num_threads; i++)
+               pthread_create(&test_threads[i], NULL,
+                       test_percpu_spinlock_thread, &data);
+
+       for (i = 0; i < num_threads; i++)
+               pthread_join(test_threads[i], NULL);
+
+       sum = 0;
+       for (i = 0; i < CPU_SETSIZE; i++)
+               sum += data.c[i].count;
+
+       assert(sum == (uint64_t)data.reps * num_threads);
+}
+
+int percpu_list_push(struct percpu_list *list, struct percpu_list_node *node)
+{
+       struct rseq_state rseq_state;
+       intptr_t *targetptr, newval;
+       int cpu;
+       bool result;
+
+       do_rseq(&rseq_lock, rseq_state, cpu, result, targetptr, newval,
+               {
+                       newval = (intptr_t)node;
+                       targetptr = (intptr_t *)&list->c[cpu].head;
+                       node->next = list->c[cpu].head;
+               });
+
+       return cpu;
+}
+
+/*
+ * Unlike a traditional lock-less linked list; the availability of a
+ * rseq primitive allows us to implement pop without concerns over
+ * ABA-type races.
+ */
+struct percpu_list_node *percpu_list_pop(struct percpu_list *list)
+{
+       struct percpu_list_node *head, *next;
+       struct rseq_state rseq_state;
+       intptr_t *targetptr, newval;
+       int cpu;
+       bool result;
+
+       do_rseq(&rseq_lock, rseq_state, cpu, result, targetptr, newval,
+               {
+                       head = list->c[cpu].head;
+                       if (!head) {
+                               result = false;
+                       } else {
+                               next = head->next;
+                               newval = (intptr_t) next;
+                               targetptr = (intptr_t *)&list->c[cpu].head;
+                       }
+               });
+
+       return head;
+}
+
+void *test_percpu_list_thread(void *arg)
+{
+       int i;
+       struct percpu_list *list = (struct percpu_list *)arg;
+
+       if (rseq_register_current_thread())
+               abort();
+
+       for (i = 0; i < 100000; i++) {
+               struct percpu_list_node *node = percpu_list_pop(list);
+
+               sched_yield();  /* encourage shuffling */
+               if (node)
+                       percpu_list_push(list, node);
+       }
+
+       if (rseq_unregister_current_thread())
+               abort();
+
+       return NULL;
+}
+
+/* Simultaneous modification to a per-cpu linked list from many threads.  */
+void test_percpu_list(void)
+{
+       int i, j;
+       uint64_t sum = 0, expected_sum = 0;
+       struct percpu_list list;
+       pthread_t test_threads[200];
+       cpu_set_t allowed_cpus;
+
+       memset(&list, 0, sizeof(list));
+
+       /* Generate list entries for every usable cpu. */
+       sched_getaffinity(0, sizeof(allowed_cpus), &allowed_cpus);
+       for (i = 0; i < CPU_SETSIZE; i++) {
+               if (!CPU_ISSET(i, &allowed_cpus))
+                       continue;
+               for (j = 1; j <= 100; j++) {
+                       struct percpu_list_node *node;
+
+                       expected_sum += j;
+
+                       node = malloc(sizeof(*node));
+                       assert(node);
+                       node->data = j;
+                       node->next = list.c[i].head;
+                       list.c[i].head = node;
+               }
+       }
+
+       for (i = 0; i < 200; i++)
+               assert(pthread_create(&test_threads[i], NULL,
+                       test_percpu_list_thread, &list) == 0);
+
+       for (i = 0; i < 200; i++)
+               pthread_join(test_threads[i], NULL);
+
+       for (i = 0; i < CPU_SETSIZE; i++) {
+               cpu_set_t pin_mask;
+               struct percpu_list_node *node;
+
+               if (!CPU_ISSET(i, &allowed_cpus))
+                       continue;
+
+               CPU_ZERO(&pin_mask);
+               CPU_SET(i, &pin_mask);
+               sched_setaffinity(0, sizeof(pin_mask), &pin_mask);
+
+               while ((node = percpu_list_pop(&list))) {
+                       sum += node->data;
+                       free(node);
+               }
+       }
+
+       /*
+        * All entries should now be accounted for (unless some external
+        * actor is interfering with our allowed affinity while this
+        * test is running).
+        */
+       assert(sum == expected_sum);
+}
+
+int main(int argc, char **argv)
+{
+       if (rseq_init_lock(&rseq_lock)) {
+               perror("rseq_init_lock");
+               return -1;
+       }
+       if (rseq_register_current_thread())
+               goto error;
+       printf("spinlock\n");
+       test_percpu_spinlock();
+       printf("percpu_list\n");
+       test_percpu_list();
+       if (rseq_unregister_current_thread())
+               goto error;
+       if (rseq_destroy_lock(&rseq_lock)) {
+               perror("rseq_destroy_lock");
+               return -1;
+       }
+       return 0;
+
+error:
+       if (rseq_destroy_lock(&rseq_lock))
+               perror("rseq_destroy_lock");
+       return -1;
+}
+
diff --git a/tools/testing/selftests/rseq/basic_test.c b/tools/testing/selftests/rseq/basic_test.c
new file mode 100644 (file)
index 0000000..dad78f6
--- /dev/null
@@ -0,0 +1,107 @@
+/*
+ * Basic test coverage for critical regions and rseq_current_cpu().
+ */
+
+#define _GNU_SOURCE
+#include <assert.h>
+#include <sched.h>
+#include <signal.h>
+#include <stdio.h>
+#include <string.h>
+#include <sys/time.h>
+
+#include <rseq.h>
+
+volatile int signals_delivered;
+volatile __thread struct rseq_state sigtest_start;
+static struct rseq_lock rseq_lock;
+
+void test_cpu_pointer(void)
+{
+       cpu_set_t affinity, test_affinity;
+       int i;
+
+       sched_getaffinity(0, sizeof(affinity), &affinity);
+       CPU_ZERO(&test_affinity);
+       for (i = 0; i < CPU_SETSIZE; i++) {
+               if (CPU_ISSET(i, &affinity)) {
+                       CPU_SET(i, &test_affinity);
+                       sched_setaffinity(0, sizeof(test_affinity),
+                                       &test_affinity);
+                       assert(rseq_current_cpu() == sched_getcpu());
+                       assert(rseq_current_cpu() == i);
+                       CPU_CLR(i, &test_affinity);
+               }
+       }
+       sched_setaffinity(0, sizeof(affinity), &affinity);
+}
+
+/*
+ * This depends solely on some environmental event triggering a counter
+ * increase.
+ */
+void test_critical_section(void)
+{
+       struct rseq_state start;
+       uint32_t event_counter;
+
+       start = rseq_start(&rseq_lock);
+       event_counter = start.event_counter;
+       do {
+               start = rseq_start(&rseq_lock);
+       } while (start.event_counter == event_counter);
+}
+
+void test_signal_interrupt_handler(int signo)
+{
+       struct rseq_state current;
+
+       current = rseq_start(&rseq_lock);
+       /*
+        * The potential critical section bordered by 'start' must be
+        * invalid.
+        */
+       assert(current.event_counter != sigtest_start.event_counter);
+       signals_delivered++;
+}
+
+void test_signal_interrupts(void)
+{
+       struct itimerval it = { { 0, 1 }, { 0, 1 } };
+
+       setitimer(ITIMER_PROF, &it, NULL);
+       signal(SIGPROF, test_signal_interrupt_handler);
+
+       do {
+               sigtest_start = rseq_start(&rseq_lock);
+       } while (signals_delivered < 10);
+       setitimer(ITIMER_PROF, NULL, NULL);
+}
+
+int main(int argc, char **argv)
+{
+       if (rseq_init_lock(&rseq_lock)) {
+               perror("rseq_init_lock");
+               return -1;
+       }
+       if (rseq_register_current_thread())
+               goto init_thread_error;
+       printf("testing current cpu\n");
+       test_cpu_pointer();
+       printf("testing critical section\n");
+       test_critical_section();
+       printf("testing critical section is interrupted by signal\n");
+       test_signal_interrupts();
+       if (rseq_unregister_current_thread())
+               goto init_thread_error;
+       if (rseq_destroy_lock(&rseq_lock)) {
+               perror("rseq_destroy_lock");
+               return -1;
+       }
+       return 0;
+
+init_thread_error:
+       if (rseq_destroy_lock(&rseq_lock))
+               perror("rseq_destroy_lock");
+       return -1;
+}
diff --git a/tools/testing/selftests/rseq/param_test.c b/tools/testing/selftests/rseq/param_test.c
new file mode 100644 (file)
index 0000000..d145d85
--- /dev/null
@@ -0,0 +1,1116 @@
+#define _GNU_SOURCE
+#include <assert.h>
+#include <pthread.h>
+#include <sched.h>
+#include <stdint.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <syscall.h>
+#include <unistd.h>
+#include <poll.h>
+#include <sys/types.h>
+#include <signal.h>
+#include <errno.h>
+
+static inline pid_t gettid(void)
+{
+       return syscall(__NR_gettid);
+}
+
+#define NR_INJECT      9
+static int loop_cnt[NR_INJECT + 1];
+
+static int opt_modulo;
+
+static int opt_yield, opt_signal, opt_sleep, opt_fallback_cnt = 3,
+               opt_disable_rseq, opt_threads = 200,
+               opt_reps = 5000, opt_disable_mod = 0, opt_test = 's';
+
+static __thread unsigned int signals_delivered;
+
+static struct rseq_lock rseq_lock;
+
+#ifndef BENCHMARK
+
+static __thread unsigned int yield_mod_cnt, nr_retry;
+
+#define printf_nobench(fmt, ...)       printf(fmt, ## __VA_ARGS__)
+
+#define RSEQ_INJECT_INPUT \
+       , [loop_cnt_1]"m"(loop_cnt[1]) \
+       , [loop_cnt_2]"m"(loop_cnt[2]) \
+       , [loop_cnt_3]"m"(loop_cnt[3]) \
+       , [loop_cnt_4]"m"(loop_cnt[4]) \
+       , [loop_cnt_5]"m"(loop_cnt[5])
+
+#if defined(__x86_64__) || defined(__i386__)
+
+#define INJECT_ASM_REG "eax"
+
+#define RSEQ_INJECT_CLOBBER \
+       , INJECT_ASM_REG
+
+#define RSEQ_INJECT_ASM(n) \
+       "mov %[loop_cnt_" #n "], %%" INJECT_ASM_REG "\n\t" \
+       "test %%" INJECT_ASM_REG ",%%" INJECT_ASM_REG "\n\t" \
+       "jz 333f\n\t" \
+       "222:\n\t" \
+       "dec %%" INJECT_ASM_REG "\n\t" \
+       "jnz 222b\n\t" \
+       "333:\n\t"
+
+#elif defined(__ARMEL__)
+
+#define INJECT_ASM_REG "r4"
+
+#define RSEQ_INJECT_CLOBBER \
+       , INJECT_ASM_REG
+
+#define RSEQ_INJECT_ASM(n) \
+       "ldr " INJECT_ASM_REG ", %[loop_cnt_" #n "]\n\t" \
+       "cmp " INJECT_ASM_REG ", #0\n\t" \
+       "beq 333f\n\t" \
+       "222:\n\t" \
+       "subs " INJECT_ASM_REG ", #1\n\t" \
+       "bne 222b\n\t" \
+       "333:\n\t"
+
+#elif __PPC__
+#define INJECT_ASM_REG "r18"
+
+#define RSEQ_INJECT_CLOBBER \
+       , INJECT_ASM_REG
+
+#define RSEQ_INJECT_ASM(n) \
+       "lwz %%" INJECT_ASM_REG ", %[loop_cnt_" #n "]\n\t" \
+       "cmpwi %%" INJECT_ASM_REG ", 0\n\t" \
+       "beq 333f\n\t" \
+       "222:\n\t" \
+       "subic. %%" INJECT_ASM_REG ", %%" INJECT_ASM_REG ", 1\n\t" \
+       "bne 222b\n\t" \
+       "333:\n\t"
+#else
+#error unsupported target
+#endif
+
+#define RSEQ_INJECT_FAILED \
+       nr_retry++;
+
+#define RSEQ_INJECT_C(n) \
+{ \
+       int loc_i, loc_nr_loops = loop_cnt[n]; \
+       \
+       for (loc_i = 0; loc_i < loc_nr_loops; loc_i++) { \
+               barrier(); \
+       } \
+       if (loc_nr_loops == -1 && opt_modulo) { \
+               if (yield_mod_cnt == opt_modulo - 1) { \
+                       if (opt_sleep > 0) \
+                               poll(NULL, 0, opt_sleep); \
+                       if (opt_yield) \
+                               sched_yield(); \
+                       if (opt_signal) \
+                               raise(SIGUSR1); \
+                       yield_mod_cnt = 0; \
+               } else { \
+                       yield_mod_cnt++; \
+               } \
+       } \
+}
+
+#define RSEQ_FALLBACK_CNT      \
+       opt_fallback_cnt
+
+#else
+
+#define printf_nobench(fmt, ...)
+
+#endif /* BENCHMARK */
+
+#include <rseq.h>
+
+struct percpu_lock_entry {
+       intptr_t v;
+} __attribute__((aligned(128)));
+
+struct percpu_lock {
+       struct percpu_lock_entry c[CPU_SETSIZE];
+};
+
+struct test_data_entry {
+       intptr_t count;
+} __attribute__((aligned(128)));
+
+struct spinlock_test_data {
+       struct percpu_lock lock;
+       struct test_data_entry c[CPU_SETSIZE];
+};
+
+struct spinlock_thread_test_data {
+       struct spinlock_test_data *data;
+       int reps;
+       int reg;
+};
+
+struct inc_test_data {
+       struct test_data_entry c[CPU_SETSIZE];
+};
+
+struct inc_thread_test_data {
+       struct inc_test_data *data;
+       int reps;
+       int reg;
+};
+
+struct percpu_list_node {
+       intptr_t data;
+       struct percpu_list_node *next;
+};
+
+struct percpu_list_entry {
+       struct percpu_list_node *head;
+} __attribute__((aligned(128)));
+
+struct percpu_list {
+       struct percpu_list_entry c[CPU_SETSIZE];
+};
+
+#define BUFFER_ITEM_PER_CPU    100
+
+struct percpu_buffer_node {
+       intptr_t data;
+};
+
+struct percpu_buffer_entry {
+       intptr_t offset;
+       intptr_t buflen;
+       struct percpu_buffer_node **array;
+} __attribute__((aligned(128)));
+
+struct percpu_buffer {
+       struct percpu_buffer_entry c[CPU_SETSIZE];
+};
+
+#define MEMCPY_BUFFER_ITEM_PER_CPU     100
+
+struct percpu_memcpy_buffer_node {
+       intptr_t data1;
+       uint64_t data2;
+};
+
+struct percpu_memcpy_buffer_entry {
+       intptr_t offset;
+       intptr_t buflen;
+       struct percpu_memcpy_buffer_node *array;
+} __attribute__((aligned(128)));
+
+struct percpu_memcpy_buffer {
+       struct percpu_memcpy_buffer_entry c[CPU_SETSIZE];
+};
+
+/* A simple percpu spinlock.  Returns the cpu lock was acquired on. */
+static int rseq_percpu_lock(struct percpu_lock *lock)
+{
+       struct rseq_state rseq_state;
+       intptr_t *targetptr, newval;
+       int cpu;
+       bool result;
+
+       for (;;) {
+               do_rseq(&rseq_lock, rseq_state, cpu, result, targetptr, newval,
+                       {
+                               if (unlikely(lock->c[cpu].v)) {
+                                       result = false;
+                               } else {
+                                       newval = 1;
+                                       targetptr = (intptr_t *)&lock->c[cpu].v;
+                               }
+                       });
+               if (likely(result))
+                       break;
+       }
+       /*
+        * Acquire semantic when taking lock after control dependency.
+        * Matches smp_store_release().
+        */
+       smp_acquire__after_ctrl_dep();
+       return cpu;
+}
+
+static void rseq_percpu_unlock(struct percpu_lock *lock, int cpu)
+{
+       assert(lock->c[cpu].v == 1);
+       /*
+        * Release lock, with release semantic. Matches
+        * smp_acquire__after_ctrl_dep().
+        */
+       smp_store_release(&lock->c[cpu].v, 0);
+}
+
+void *test_percpu_spinlock_thread(void *arg)
+{
+       struct spinlock_thread_test_data *thread_data = arg;
+       struct spinlock_test_data *data = thread_data->data;
+       int i, cpu;
+
+       if (!opt_disable_rseq && thread_data->reg
+                       && rseq_register_current_thread())
+               abort();
+       for (i = 0; i < thread_data->reps; i++) {
+               cpu = rseq_percpu_lock(&data->lock);
+               data->c[cpu].count++;
+               rseq_percpu_unlock(&data->lock, cpu);
+#ifndef BENCHMARK
+               if (i != 0 && !(i % (thread_data->reps / 10)))
+                       printf("tid %d: count %d\n", (int) gettid(), i);
+#endif
+       }
+       printf_nobench("tid %d: number of retry: %d, signals delivered: %u, nr_fallback %u, nr_fallback_wait %u\n",
+               (int) gettid(), nr_retry, signals_delivered,
+               rseq_get_fallback_cnt(),
+               rseq_get_fallback_wait_cnt());
+       if (rseq_unregister_current_thread())
+               abort();
+       return NULL;
+}
+
+/*
+ * A simple test which implements a sharded counter using a per-cpu
+ * lock.  Obviously real applications might prefer to simply use a
+ * per-cpu increment; however, this is reasonable for a test and the
+ * lock can be extended to synchronize more complicated operations.
+ */
+void test_percpu_spinlock(void)
+{
+       const int num_threads = opt_threads;
+       int i, ret;
+       uint64_t sum;
+       pthread_t test_threads[num_threads];
+       struct spinlock_test_data data;
+       struct spinlock_thread_test_data thread_data[num_threads];
+
+       memset(&data, 0, sizeof(data));
+       for (i = 0; i < num_threads; i++) {
+               thread_data[i].reps = opt_reps;
+               if (opt_disable_mod <= 0 || (i % opt_disable_mod))
+                       thread_data[i].reg = 1;
+               else
+                       thread_data[i].reg = 0;
+               thread_data[i].data = &data;
+               ret = pthread_create(&test_threads[i], NULL,
+                       test_percpu_spinlock_thread, &thread_data[i]);
+               if (ret) {
+                       errno = ret;
+                       perror("pthread_create");
+                       abort();
+               }
+       }
+
+       for (i = 0; i < num_threads; i++) {
+               pthread_join(test_threads[i], NULL);
+               if (ret) {
+                       errno = ret;
+                       perror("pthread_join");
+                       abort();
+               }
+       }
+
+       sum = 0;
+       for (i = 0; i < CPU_SETSIZE; i++)
+               sum += data.c[i].count;
+
+       assert(sum == (uint64_t)opt_reps * num_threads);
+}
+
+void *test_percpu_inc_thread(void *arg)
+{
+       struct inc_thread_test_data *thread_data = arg;
+       struct inc_test_data *data = thread_data->data;
+       int i;
+
+       if (!opt_disable_rseq && thread_data->reg
+                       && rseq_register_current_thread())
+               abort();
+       for (i = 0; i < thread_data->reps; i++) {
+               struct rseq_state rseq_state;
+               intptr_t *targetptr, newval;
+               int cpu;
+               bool result;
+
+               do_rseq(&rseq_lock, rseq_state, cpu, result, targetptr, newval,
+                       {
+                               newval = (intptr_t)data->c[cpu].count + 1;
+                               targetptr = (intptr_t *)&data->c[cpu].count;
+                       });
+
+#ifndef BENCHMARK
+               if (i != 0 && !(i % (thread_data->reps / 10)))
+                       printf("tid %d: count %d\n", (int) gettid(), i);
+#endif
+       }
+       printf_nobench("tid %d: number of retry: %d, signals delivered: %u, nr_fallback %u, nr_fallback_wait %u\n",
+               (int) gettid(), nr_retry, signals_delivered,
+               rseq_get_fallback_cnt(),
+               rseq_get_fallback_wait_cnt());
+       if (rseq_unregister_current_thread())
+               abort();
+       return NULL;
+}
+
+void test_percpu_inc(void)
+{
+       const int num_threads = opt_threads;
+       int i, ret;
+       uint64_t sum;
+       pthread_t test_threads[num_threads];
+       struct inc_test_data data;
+       struct inc_thread_test_data thread_data[num_threads];
+
+       memset(&data, 0, sizeof(data));
+       for (i = 0; i < num_threads; i++) {
+               thread_data[i].reps = opt_reps;
+               if (opt_disable_mod <= 0 || (i % opt_disable_mod))
+                       thread_data[i].reg = 1;
+               else
+                       thread_data[i].reg = 0;
+               thread_data[i].data = &data;
+               ret = pthread_create(&test_threads[i], NULL,
+                       test_percpu_inc_thread, &thread_data[i]);
+               if (ret) {
+                       errno = ret;
+                       perror("pthread_create");
+                       abort();
+               }
+       }
+
+       for (i = 0; i < num_threads; i++) {
+               pthread_join(test_threads[i], NULL);
+               if (ret) {
+                       errno = ret;
+                       perror("pthread_join");
+                       abort();
+               }
+       }
+
+       sum = 0;
+       for (i = 0; i < CPU_SETSIZE; i++)
+               sum += data.c[i].count;
+
+       assert(sum == (uint64_t)opt_reps * num_threads);
+}
+
+int percpu_list_push(struct percpu_list *list, struct percpu_list_node *node)
+{
+       struct rseq_state rseq_state;
+       intptr_t *targetptr, newval;
+       int cpu;
+       bool result;
+
+       do_rseq(&rseq_lock, rseq_state, cpu, result, targetptr, newval,
+               {
+                       newval = (intptr_t)node;
+                       targetptr = (intptr_t *)&list->c[cpu].head;
+                       node->next = list->c[cpu].head;
+               });
+
+       return cpu;
+}
+
+/*
+ * Unlike a traditional lock-less linked list; the availability of a
+ * rseq primitive allows us to implement pop without concerns over
+ * ABA-type races.
+ */
+struct percpu_list_node *percpu_list_pop(struct percpu_list *list)
+{
+       struct percpu_list_node *head, *next;
+       struct rseq_state rseq_state;
+       intptr_t *targetptr, newval;
+       int cpu;
+       bool result;
+
+       do_rseq(&rseq_lock, rseq_state, cpu, result, targetptr, newval,
+               {
+                       head = list->c[cpu].head;
+                       if (!head) {
+                               result = false;
+                       } else {
+                               next = head->next;
+                               newval = (intptr_t) next;
+                               targetptr = (intptr_t *) &list->c[cpu].head;
+                       }
+               });
+
+       return head;
+}
+
+void *test_percpu_list_thread(void *arg)
+{
+       int i;
+       struct percpu_list *list = (struct percpu_list *)arg;
+
+       if (rseq_register_current_thread())
+               abort();
+
+       for (i = 0; i < opt_reps; i++) {
+               struct percpu_list_node *node = percpu_list_pop(list);
+
+               if (opt_yield)
+                       sched_yield();  /* encourage shuffling */
+               if (node)
+                       percpu_list_push(list, node);
+       }
+
+       if (rseq_unregister_current_thread())
+               abort();
+
+       return NULL;
+}
+
+/* Simultaneous modification to a per-cpu linked list from many threads.  */
+void test_percpu_list(void)
+{
+       const int num_threads = opt_threads;
+       int i, j, ret;
+       uint64_t sum = 0, expected_sum = 0;
+       struct percpu_list list;
+       pthread_t test_threads[num_threads];
+       cpu_set_t allowed_cpus;
+
+       memset(&list, 0, sizeof(list));
+
+       /* Generate list entries for every usable cpu. */
+       sched_getaffinity(0, sizeof(allowed_cpus), &allowed_cpus);
+       for (i = 0; i < CPU_SETSIZE; i++) {
+               if (!CPU_ISSET(i, &allowed_cpus))
+                       continue;
+               for (j = 1; j <= 100; j++) {
+                       struct percpu_list_node *node;
+
+                       expected_sum += j;
+
+                       node = malloc(sizeof(*node));
+                       assert(node);
+                       node->data = j;
+                       node->next = list.c[i].head;
+                       list.c[i].head = node;
+               }
+       }
+
+       for (i = 0; i < num_threads; i++) {
+               ret = pthread_create(&test_threads[i], NULL,
+                       test_percpu_list_thread, &list);
+               if (ret) {
+                       errno = ret;
+                       perror("pthread_create");
+                       abort();
+               }
+       }
+
+       for (i = 0; i < num_threads; i++) {
+               pthread_join(test_threads[i], NULL);
+               if (ret) {
+                       errno = ret;
+                       perror("pthread_join");
+                       abort();
+               }
+       }
+
+       for (i = 0; i < CPU_SETSIZE; i++) {
+               cpu_set_t pin_mask;
+               struct percpu_list_node *node;
+
+               if (!CPU_ISSET(i, &allowed_cpus))
+                       continue;
+
+               CPU_ZERO(&pin_mask);
+               CPU_SET(i, &pin_mask);
+               sched_setaffinity(0, sizeof(pin_mask), &pin_mask);
+
+               while ((node = percpu_list_pop(&list))) {
+                       sum += node->data;
+                       free(node);
+               }
+       }
+
+       /*
+        * All entries should now be accounted for (unless some external
+        * actor is interfering with our allowed affinity while this
+        * test is running).
+        */
+       assert(sum == expected_sum);
+}
+
+bool percpu_buffer_push(struct percpu_buffer *buffer,
+               struct percpu_buffer_node *node)
+{
+       struct rseq_state rseq_state;
+       intptr_t *targetptr_spec, newval_spec;
+       intptr_t *targetptr_final, newval_final;
+       int cpu;
+       bool result;
+
+       do_rseq2(&rseq_lock, rseq_state, cpu, result,
+               targetptr_spec, newval_spec, targetptr_final, newval_final,
+               {
+                       intptr_t offset = buffer->c[cpu].offset;
+
+                       if (offset == buffer->c[cpu].buflen) {
+                               result = false;
+                       } else {
+                               newval_spec = (intptr_t)node;
+                               targetptr_spec = (intptr_t *)&buffer->c[cpu].array[offset];
+                               newval_final = offset + 1;
+                               targetptr_final = &buffer->c[cpu].offset;
+                       }
+               });
+
+       return result;
+}
+
+struct percpu_buffer_node *percpu_buffer_pop(struct percpu_buffer *buffer)
+{
+       struct percpu_buffer_node *head;
+       struct rseq_state rseq_state;
+       intptr_t *targetptr, newval;
+       int cpu;
+       bool result;
+
+       do_rseq(&rseq_lock, rseq_state, cpu, result, targetptr, newval,
+               {
+                       intptr_t offset = buffer->c[cpu].offset;
+
+                       if (offset == 0) {
+                               result = false;
+                       } else {
+                               head = buffer->c[cpu].array[offset - 1];
+                               newval = offset - 1;
+                               targetptr = (intptr_t *)&buffer->c[cpu].offset;
+                       }
+               });
+
+       if (result)
+               return head;
+       else
+               return NULL;
+}
+
+void *test_percpu_buffer_thread(void *arg)
+{
+       int i;
+       struct percpu_buffer *buffer = (struct percpu_buffer *)arg;
+
+       if (rseq_register_current_thread())
+               abort();
+
+       for (i = 0; i < opt_reps; i++) {
+               struct percpu_buffer_node *node = percpu_buffer_pop(buffer);
+
+               if (opt_yield)
+                       sched_yield();  /* encourage shuffling */
+               if (node) {
+                       if (!percpu_buffer_push(buffer, node)) {
+                               /* Should increase buffer size. */
+                               abort();
+                       }
+               }
+       }
+
+       if (rseq_unregister_current_thread())
+               abort();
+
+       return NULL;
+}
+
+/* Simultaneous modification to a per-cpu buffer from many threads.  */
+void test_percpu_buffer(void)
+{
+       const int num_threads = opt_threads;
+       int i, j, ret;
+       uint64_t sum = 0, expected_sum = 0;
+       struct percpu_buffer buffer;
+       pthread_t test_threads[num_threads];
+       cpu_set_t allowed_cpus;
+
+       memset(&buffer, 0, sizeof(buffer));
+
+       /* Generate list entries for every usable cpu. */
+       sched_getaffinity(0, sizeof(allowed_cpus), &allowed_cpus);
+       for (i = 0; i < CPU_SETSIZE; i++) {
+               if (!CPU_ISSET(i, &allowed_cpus))
+                       continue;
+               /* Worse-case is every item in same CPU. */
+               buffer.c[i].array =
+                       malloc(sizeof(*buffer.c[i].array) * CPU_SETSIZE
+                               * BUFFER_ITEM_PER_CPU);
+               assert(buffer.c[i].array);
+               buffer.c[i].buflen = CPU_SETSIZE * BUFFER_ITEM_PER_CPU;
+               for (j = 1; j <= BUFFER_ITEM_PER_CPU; j++) {
+                       struct percpu_buffer_node *node;
+
+                       expected_sum += j;
+
+                       /*
+                        * We could theoretically put the word-sized
+                        * "data" directly in the buffer. However, we
+                        * want to model objects that would not fit
+                        * within a single word, so allocate an object
+                        * for each node.
+                        */
+                       node = malloc(sizeof(*node));
+                       assert(node);
+                       node->data = j;
+                       buffer.c[i].array[j - 1] = node;
+                       buffer.c[i].offset++;
+               }
+       }
+
+       for (i = 0; i < num_threads; i++) {
+               ret = pthread_create(&test_threads[i], NULL,
+                       test_percpu_buffer_thread, &buffer);
+               if (ret) {
+                       errno = ret;
+                       perror("pthread_create");
+                       abort();
+               }
+       }
+
+       for (i = 0; i < num_threads; i++) {
+               pthread_join(test_threads[i], NULL);
+               if (ret) {
+                       errno = ret;
+                       perror("pthread_join");
+                       abort();
+               }
+       }
+
+       for (i = 0; i < CPU_SETSIZE; i++) {
+               cpu_set_t pin_mask;
+               struct percpu_buffer_node *node;
+
+               if (!CPU_ISSET(i, &allowed_cpus))
+                       continue;
+
+               CPU_ZERO(&pin_mask);
+               CPU_SET(i, &pin_mask);
+               sched_setaffinity(0, sizeof(pin_mask), &pin_mask);
+
+               while ((node = percpu_buffer_pop(&buffer))) {
+                       sum += node->data;
+                       free(node);
+               }
+               free(buffer.c[i].array);
+       }
+
+       /*
+        * All entries should now be accounted for (unless some external
+        * actor is interfering with our allowed affinity while this
+        * test is running).
+        */
+       assert(sum == expected_sum);
+}
+
+bool percpu_memcpy_buffer_push(struct percpu_memcpy_buffer *buffer,
+               struct percpu_memcpy_buffer_node item)
+{
+       struct rseq_state rseq_state;
+       char *destptr, *srcptr;
+       size_t copylen;
+       intptr_t *targetptr_final, newval_final;
+       int cpu;
+       bool result;
+
+       do_rseq_memcpy(&rseq_lock, rseq_state, cpu, result,
+               destptr, srcptr, copylen, targetptr_final, newval_final,
+               {
+                       intptr_t offset = buffer->c[cpu].offset;
+
+                       if (offset == buffer->c[cpu].buflen) {
+                               result = false;
+                       } else {
+                               destptr = (char *)&buffer->c[cpu].array[offset];
+                               srcptr = (char *)&item;
+                               copylen = sizeof(item);
+                               newval_final = offset + 1;
+                               targetptr_final = &buffer->c[cpu].offset;
+                       }
+               });
+
+       return result;
+}
+
+bool percpu_memcpy_buffer_pop(struct percpu_memcpy_buffer *buffer,
+               struct percpu_memcpy_buffer_node *item)
+{
+       struct rseq_state rseq_state;
+       char *destptr, *srcptr;
+       size_t copylen;
+       intptr_t *targetptr_final, newval_final;
+       int cpu;
+       bool result;
+
+       do_rseq_memcpy(&rseq_lock, rseq_state, cpu, result,
+               destptr, srcptr, copylen, targetptr_final, newval_final,
+               {
+                       intptr_t offset = buffer->c[cpu].offset;
+
+                       if (offset == 0) {
+                               result = false;
+                       } else {
+                               destptr = (char *)item;
+                               srcptr = (char *)&buffer->c[cpu].array[offset - 1];
+                               copylen = sizeof(*item);
+                               newval_final = offset - 1;
+                               targetptr_final = &buffer->c[cpu].offset;
+                       }
+               });
+
+       return result;
+}
+
+void *test_percpu_memcpy_buffer_thread(void *arg)
+{
+       int i;
+       struct percpu_memcpy_buffer *buffer = (struct percpu_memcpy_buffer *)arg;
+
+       if (rseq_register_current_thread())
+               abort();
+
+       for (i = 0; i < opt_reps; i++) {
+               struct percpu_memcpy_buffer_node item;
+               bool result;
+
+               result = percpu_memcpy_buffer_pop(buffer, &item);
+               if (opt_yield)
+                       sched_yield();  /* encourage shuffling */
+               if (result) {
+                       if (!percpu_memcpy_buffer_push(buffer, item)) {
+                               /* Should increase buffer size. */
+                               abort();
+                       }
+               }
+       }
+
+       if (rseq_unregister_current_thread())
+               abort();
+
+       return NULL;
+}
+
+/* Simultaneous modification to a per-cpu buffer from many threads.  */
+void test_percpu_memcpy_buffer(void)
+{
+       const int num_threads = opt_threads;
+       int i, j, ret;
+       uint64_t sum = 0, expected_sum = 0;
+       struct percpu_memcpy_buffer buffer;
+       pthread_t test_threads[num_threads];
+       cpu_set_t allowed_cpus;
+
+       memset(&buffer, 0, sizeof(buffer));
+
+       /* Generate list entries for every usable cpu. */
+       sched_getaffinity(0, sizeof(allowed_cpus), &allowed_cpus);
+       for (i = 0; i < CPU_SETSIZE; i++) {
+               if (!CPU_ISSET(i, &allowed_cpus))
+                       continue;
+               /* Worse-case is every item in same CPU. */
+               buffer.c[i].array =
+                       malloc(sizeof(*buffer.c[i].array) * CPU_SETSIZE
+                               * MEMCPY_BUFFER_ITEM_PER_CPU);
+               assert(buffer.c[i].array);
+               buffer.c[i].buflen = CPU_SETSIZE * MEMCPY_BUFFER_ITEM_PER_CPU;
+               for (j = 1; j <= MEMCPY_BUFFER_ITEM_PER_CPU; j++) {
+                       expected_sum += 2 * j + 1;
+
+                       /*
+                        * We could theoretically put the word-sized
+                        * "data" directly in the buffer. However, we
+                        * want to model objects that would not fit
+                        * within a single word, so allocate an object
+                        * for each node.
+                        */
+                       buffer.c[i].array[j - 1].data1 = j;
+                       buffer.c[i].array[j - 1].data2 = j + 1;
+                       buffer.c[i].offset++;
+               }
+       }
+
+       for (i = 0; i < num_threads; i++) {
+               ret = pthread_create(&test_threads[i], NULL,
+                       test_percpu_memcpy_buffer_thread, &buffer);
+               if (ret) {
+                       errno = ret;
+                       perror("pthread_create");
+                       abort();
+               }
+       }
+
+       for (i = 0; i < num_threads; i++) {
+               pthread_join(test_threads[i], NULL);
+               if (ret) {
+                       errno = ret;
+                       perror("pthread_join");
+                       abort();
+               }
+       }
+
+       for (i = 0; i < CPU_SETSIZE; i++) {
+               cpu_set_t pin_mask;
+               struct percpu_memcpy_buffer_node item;
+
+               if (!CPU_ISSET(i, &allowed_cpus))
+                       continue;
+
+               CPU_ZERO(&pin_mask);
+               CPU_SET(i, &pin_mask);
+               sched_setaffinity(0, sizeof(pin_mask), &pin_mask);
+
+               while (percpu_memcpy_buffer_pop(&buffer, &item)) {
+                       sum += item.data1;
+                       sum += item.data2;
+               }
+               free(buffer.c[i].array);
+       }
+
+       /*
+        * All entries should now be accounted for (unless some external
+        * actor is interfering with our allowed affinity while this
+        * test is running).
+        */
+       assert(sum == expected_sum);
+}
+
+static void test_signal_interrupt_handler(int signo)
+{
+       signals_delivered++;
+}
+
+static int set_signal_handler(void)
+{
+       int ret = 0;
+       struct sigaction sa;
+       sigset_t sigset;
+
+       ret = sigemptyset(&sigset);
+       if (ret < 0) {
+               perror("sigemptyset");
+               return ret;
+       }
+
+       sa.sa_handler = test_signal_interrupt_handler;
+       sa.sa_mask = sigset;
+       sa.sa_flags = 0;
+       ret = sigaction(SIGUSR1, &sa, NULL);
+       if (ret < 0) {
+               perror("sigaction");
+               return ret;
+       }
+
+       printf_nobench("Signal handler set for SIGUSR1\n");
+
+       return ret;
+}
+
+static void show_usage(int argc, char **argv)
+{
+       printf("Usage : %s <OPTIONS>\n",
+               argv[0]);
+       printf("OPTIONS:\n");
+       printf("        [-1 loops] Number of loops for delay injection 1\n");
+       printf("        [-2 loops] Number of loops for delay injection 2\n");
+       printf("        [-3 loops] Number of loops for delay injection 3\n");
+       printf("        [-4 loops] Number of loops for delay injection 4\n");
+       printf("        [-5 loops] Number of loops for delay injection 5\n");
+       printf("        [-6 loops] Number of loops for delay injection 6 (-1 to enable -m)\n");
+       printf("        [-7 loops] Number of loops for delay injection 7 (-1 to enable -m)\n");
+       printf("        [-8 loops] Number of loops for delay injection 8 (-1 to enable -m)\n");
+       printf("        [-9 loops] Number of loops for delay injection 9 (-1 to enable -m)\n");
+       printf("        [-m N] Yield/sleep/kill every modulo N (default 0: disabled) (>= 0)\n");
+       printf("        [-y] Yield\n");
+       printf("        [-k] Kill thread with signal\n");
+       printf("        [-s S] S: =0: disabled (default), >0: sleep time (ms)\n");
+       printf("        [-f N] Use fallback every N failure (>= 1)\n");
+       printf("        [-t N] Number of threads (default 200)\n");
+       printf("        [-r N] Number of repetitions per thread (default 5000)\n");
+       printf("        [-d] Disable rseq system call (no initialization)\n");
+       printf("        [-D M] Disable rseq for each M threads\n");
+       printf("        [-T test] Choose test: (s)pinlock, (l)ist, (b)uffer, (m)emcpy, (i)ncrement\n");
+       printf("        [-h] Show this help.\n");
+       printf("\n");
+}
+
+int main(int argc, char **argv)
+{
+       int i;
+
+       if (rseq_init_lock(&rseq_lock)) {
+               perror("rseq_init_lock");
+               return -1;
+       }
+       if (set_signal_handler())
+               goto error;
+       for (i = 1; i < argc; i++) {
+               if (argv[i][0] != '-')
+                       continue;
+               switch (argv[i][1]) {
+               case '1':
+               case '2':
+               case '3':
+               case '4':
+               case '5':
+               case '6':
+               case '7':
+               case '8':
+               case '9':
+                       if (argc < i + 2) {
+                               show_usage(argc, argv);
+                               goto error;
+                       }
+                       loop_cnt[argv[i][1] - '0'] = atol(argv[i + 1]);
+                       i++;
+                       break;
+               case 'm':
+                       if (argc < i + 2) {
+                               show_usage(argc, argv);
+                               goto error;
+                       }
+                       opt_modulo = atol(argv[i + 1]);
+                       if (opt_modulo < 0) {
+                               show_usage(argc, argv);
+                               goto error;
+                       }
+                       i++;
+                       break;
+               case 's':
+                       if (argc < i + 2) {
+                               show_usage(argc, argv);
+                               goto error;
+                       }
+                       opt_sleep = atol(argv[i + 1]);
+                       if (opt_sleep < 0) {
+                               show_usage(argc, argv);
+                               goto error;
+                       }
+                       i++;
+                       break;
+               case 'y':
+                       opt_yield = 1;
+                       break;
+               case 'k':
+                       opt_signal = 1;
+                       break;
+               case 'd':
+                       opt_disable_rseq = 1;
+                       break;
+               case 'D':
+                       if (argc < i + 2) {
+                               show_usage(argc, argv);
+                               goto error;
+                       }
+                       opt_disable_mod = atol(argv[i + 1]);
+                       if (opt_disable_mod < 0) {
+                               show_usage(argc, argv);
+                               goto error;
+                       }
+                       i++;
+                       break;
+               case 'f':
+                       if (argc < i + 2) {
+                               show_usage(argc, argv);
+                               goto error;
+                       }
+                       opt_fallback_cnt = atol(argv[i + 1]);
+                       if (opt_fallback_cnt < 1) {
+                               show_usage(argc, argv);
+                               goto error;
+                       }
+                       i++;
+                       break;
+               case 't':
+                       if (argc < i + 2) {
+                               show_usage(argc, argv);
+                               goto error;
+                       }
+                       opt_threads = atol(argv[i + 1]);
+                       if (opt_threads < 0) {
+                               show_usage(argc, argv);
+                               goto error;
+                       }
+                       i++;
+                       break;
+               case 'r':
+                       if (argc < i + 2) {
+                               show_usage(argc, argv);
+                               goto error;
+                       }
+                       opt_reps = atol(argv[i + 1]);
+                       if (opt_reps < 0) {
+                               show_usage(argc, argv);
+                               goto error;
+                       }
+                       i++;
+                       break;
+               case 'h':
+                       show_usage(argc, argv);
+                       goto end;
+               case 'T':
+                       if (argc < i + 2) {
+                               show_usage(argc, argv);
+                               goto error;
+                       }
+                       opt_test = *argv[i + 1];
+                       switch (opt_test) {
+                       case 's':
+                       case 'l':
+                       case 'i':
+                       case 'b':
+                       case 'm':
+                               break;
+                       default:
+                               show_usage(argc, argv);
+                               goto error;
+                       }
+                       i++;
+                       break;
+               default:
+                       show_usage(argc, argv);
+                       goto error;
+               }
+       }
+
+       if (!opt_disable_rseq && rseq_register_current_thread())
+               goto error;
+       switch (opt_test) {
+       case 's':
+               printf_nobench("spinlock\n");
+               test_percpu_spinlock();
+               break;
+       case 'l':
+               printf_nobench("linked list\n");
+               test_percpu_list();
+               break;
+       case 'b':
+               printf_nobench("buffer\n");
+               test_percpu_buffer();
+               break;
+       case 'm':
+               printf_nobench("memcpy buffer\n");
+               test_percpu_memcpy_buffer();
+               break;
+       case 'i':
+               printf_nobench("counter increment\n");
+               test_percpu_inc();
+               break;
+       }
+       if (rseq_unregister_current_thread())
+               abort();
+end:
+       return 0;
+
+error:
+       if (rseq_destroy_lock(&rseq_lock))
+               perror("rseq_destroy_lock");
+       return -1;
+}
diff --git a/tools/testing/selftests/rseq/rseq-arm.h b/tools/testing/selftests/rseq/rseq-arm.h
new file mode 100644 (file)
index 0000000..9966df3
--- /dev/null
@@ -0,0 +1,168 @@
+/*
+ * rseq-arm.h
+ *
+ * (C) Copyright 2016 - Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#define smp_mb()       __asm__ __volatile__ ("dmb" : : : "memory")
+#define smp_rmb()      __asm__ __volatile__ ("dmb" : : : "memory")
+#define smp_wmb()      __asm__ __volatile__ ("dmb" : : : "memory")
+
+#define smp_load_acquire(p)                                            \
+__extension__ ({                                                       \
+       __typeof(*p) ____p1 = READ_ONCE(*p);                            \
+       smp_mb();                                                       \
+       ____p1;                                                         \
+})
+
+#define smp_acquire__after_ctrl_dep()  smp_rmb()
+
+#define smp_store_release(p, v)                                                \
+do {                                                                   \
+       smp_mb();                                                       \
+       WRITE_ONCE(*p, v);                                              \
+} while (0)
+
+#define has_fast_acquire_release()     0
+#define has_single_copy_load_64()      1
+
+/*
+ * The __rseq_table section can be used by debuggers to better handle
+ * single-stepping through the restartable critical sections.
+ *
+ * Load the immediate value 0 into register r1 right after the ldr
+ * instruction to improve instruction-level parallelism: load the
+ * constant while the processor is stalled waiting for the load to
+ * complete, which is required by the following comparison and branch.
+ */
+
+#define RSEQ_FINISH_ASM(_target_final, _to_write_final, _start_value, \
+               _failure, _spec_store, _spec_input, \
+               _final_store, _final_input, _extra_clobber, \
+               _setup, _teardown, _scratch) \
+do { \
+       _scratch \
+       __asm__ __volatile__ goto ( \
+               ".pushsection __rseq_table, \"aw\"\n\t" \
+               ".balign 32\n\t" \
+               ".word 1f, 0x0, 2f, 0x0, 5f, 0x0, 0x0, 0x0\n\t" \
+               ".popsection\n\t" \
+               "1:\n\t" \
+               _setup \
+               RSEQ_INJECT_ASM(1) \
+               "adr r0, 3f\n\t" \
+               "str r0, [%[rseq_cs]]\n\t" \
+               RSEQ_INJECT_ASM(2) \
+               "ldr r0, %[current_event_counter]\n\t" \
+               "mov r1, #0\n\t" \
+               "cmp %[start_event_counter], r0\n\t" \
+               "bne 5f\n\t" \
+               RSEQ_INJECT_ASM(3) \
+               _spec_store \
+               _final_store \
+               "2:\n\t" \
+               RSEQ_INJECT_ASM(5) \
+               "str r1, [%[rseq_cs]]\n\t" \
+               _teardown \
+               "b 4f\n\t" \
+               ".balign 32\n\t" \
+               "3:\n\t" \
+               ".word 1b, 0x0, 2b, 0x0, 5f, 0x0, 0x0, 0x0\n\t" \
+               "5:\n\t" \
+               "mov r1, #0\n\t" \
+               "str r1, [%[rseq_cs]]\n\t" \
+               _teardown \
+               "ldr pc, %l[failure]\n\t" \
+               "4:\n\t" \
+               : /* gcc asm goto does not allow outputs */ \
+               : [start_event_counter]"r"((_start_value).event_counter), \
+                 [current_event_counter]"m"((_start_value).rseqp->u.e.event_counter), \
+                 [rseq_cs]"r"(&(_start_value).rseqp->rseq_cs) \
+                 _spec_input \
+                 _final_input \
+                 RSEQ_INJECT_INPUT \
+               : "r0", "r1", "memory", "cc" \
+                 _extra_clobber \
+                 RSEQ_INJECT_CLOBBER \
+               : _failure \
+       ); \
+} while (0)
+
+#define RSEQ_FINISH_FINAL_STORE_ASM() \
+               "str %[to_write_final], [%[target_final]]\n\t"
+
+#define RSEQ_FINISH_FINAL_STORE_RELEASE_ASM() \
+               "dmb\n\t" \
+               RSEQ_FINISH_FINAL_STORE_ASM()
+
+#define RSEQ_FINISH_FINAL_STORE_INPUT(_target_final, _to_write_final) \
+               , [to_write_final]"r"(_to_write_final), \
+               [target_final]"r"(_target_final)
+
+#define RSEQ_FINISH_SPECULATIVE_STORE_ASM() \
+               "str %[to_write_spec], [%[target_spec]]\n\t" \
+               RSEQ_INJECT_ASM(4)
+
+#define RSEQ_FINISH_SPECULATIVE_STORE_INPUT(_target_spec, _to_write_spec) \
+               , [to_write_spec]"r"(_to_write_spec), \
+               [target_spec]"r"(_target_spec)
+
+/* TODO: implement a faster memcpy. */
+#define RSEQ_FINISH_MEMCPY_STORE_ASM() \
+               "cmp %[len_memcpy], #0\n\t" \
+               "beq 333f\n\t" \
+               "222:\n\t" \
+               "ldrb %%r0, [%[to_write_memcpy]]\n\t" \
+               "strb %%r0, [%[target_memcpy]]\n\t" \
+               "adds %[to_write_memcpy], #1\n\t" \
+               "adds %[target_memcpy], #1\n\t" \
+               "subs %[len_memcpy], #1\n\t" \
+               "bne 222b\n\t" \
+               "333:\n\t" \
+               RSEQ_INJECT_ASM(4)
+
+#define RSEQ_FINISH_MEMCPY_STORE_INPUT(_target_memcpy, _to_write_memcpy, _len_memcpy) \
+               , [to_write_memcpy]"r"(_to_write_memcpy), \
+               [target_memcpy]"r"(_target_memcpy), \
+               [len_memcpy]"r"(_len_memcpy), \
+               [rseq_scratch0]"m"(rseq_scratch[0]), \
+               [rseq_scratch1]"m"(rseq_scratch[1]), \
+               [rseq_scratch2]"m"(rseq_scratch[2])
+
+/* We can use r0. */
+#define RSEQ_FINISH_MEMCPY_CLOBBER()
+
+#define RSEQ_FINISH_MEMCPY_SCRATCH() \
+               uint32_t rseq_scratch[3];
+
+/*
+ * We need to save and restore those input registers so they can be
+ * modified within the assembly.
+ */
+#define RSEQ_FINISH_MEMCPY_SETUP() \
+               "str %[to_write_memcpy], %[rseq_scratch0]\n\t" \
+               "str %[target_memcpy], %[rseq_scratch1]\n\t" \
+               "str %[len_memcpy], %[rseq_scratch2]\n\t"
+
+#define RSEQ_FINISH_MEMCPY_TEARDOWN() \
+               "ldr %[len_memcpy], %[rseq_scratch2]\n\t" \
+               "ldr %[target_memcpy], %[rseq_scratch1]\n\t" \
+               "ldr %[to_write_memcpy], %[rseq_scratch0]\n\t"
diff --git a/tools/testing/selftests/rseq/rseq-ppc.h b/tools/testing/selftests/rseq/rseq-ppc.h
new file mode 100644 (file)
index 0000000..04dac92
--- /dev/null
@@ -0,0 +1,273 @@
+/*
+ * rseq-ppc.h
+ *
+ * (C) Copyright 2016 - Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
+ * (C) Copyright 2016 - Boqun Feng <boqun.feng@gmail.com>
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#define smp_mb()       __asm__ __volatile__ ("sync" : : : "memory")
+#define smp_lwsync()   __asm__ __volatile__ ("lwsync" : : : "memory")
+#define smp_rmb()      smp_lwsync()
+#define smp_wmb()      smp_lwsync()
+
+#define smp_load_acquire(p)                                            \
+__extension__ ({                                                       \
+       __typeof(*p) ____p1 = READ_ONCE(*p);                            \
+       smp_lwsync();                                                   \
+       ____p1;                                                         \
+})
+
+#define smp_acquire__after_ctrl_dep()  smp_lwsync()
+
+#define smp_store_release(p, v)                                                \
+do {                                                                   \
+       smp_lwsync();                                                   \
+       WRITE_ONCE(*p, v);                                              \
+} while (0)
+
+#define has_fast_acquire_release()     0
+
+#ifdef __PPC64__
+#define has_single_copy_load_64()      1
+#else
+#define has_single_copy_load_64()      0
+#endif
+
+/*
+ * The __rseq_table section can be used by debuggers to better handle
+ * single-stepping through the restartable critical sections.
+ */
+
+#ifdef __PPC64__
+
+#define RSEQ_FINISH_ASM(_target_final, _to_write_final, _start_value, \
+               _failure, _spec_store, _spec_input, \
+               _final_store, _final_input, _extra_clobber, \
+               _setup, _teardown, _scratch) \
+       __asm__ __volatile__ goto ( \
+               ".pushsection __rseq_table, \"aw\"\n\t" \
+               ".balign 32\n\t" \
+               "3:\n\t" \
+               ".quad 1f, 2f, 4f, 0x0\n\t" \
+               ".popsection\n\t" \
+               "1:\n\t" \
+               _setup \
+               RSEQ_INJECT_ASM(1) \
+               "lis %%r17, (3b)@highest\n\t" \
+               "ori %%r17, %%r17, (3b)@higher\n\t" \
+               "rldicr %%r17, %%r17, 32, 31\n\t" \
+               "oris %%r17, %%r17, (3b)@h\n\t" \
+               "ori %%r17, %%r17, (3b)@l\n\t" \
+               "std %%r17, 0(%[rseq_cs])\n\t" \
+               RSEQ_INJECT_ASM(2) \
+               "lwz %%r17, %[current_event_counter]\n\t" \
+               "cmpw cr7, %[start_event_counter], %%r17\n\t" \
+               "bne- cr7, 4f\n\t" \
+               RSEQ_INJECT_ASM(3) \
+               _spec_store \
+               _final_store \
+               "2:\n\t" \
+               RSEQ_INJECT_ASM(5) \
+               "li %%r17, 0\n\t" \
+               "std %%r17, 0(%[rseq_cs])\n\t" \
+               _teardown \
+               "b 5f\n\t" \
+               "4:\n\t" \
+               "li %%r17, 0\n\t" \
+               "std %%r17, 0(%[rseq_cs])\n\t" \
+               _teardown \
+               "b %l[failure]\n\t" \
+               "5:\n\t" \
+               : /* gcc asm goto does not allow outputs */ \
+               : [start_event_counter]"r"((_start_value).event_counter), \
+                 [current_event_counter]"m"((_start_value).rseqp->u.e.event_counter), \
+                 [rseq_cs]"b"(&(_start_value).rseqp->rseq_cs) \
+                 _spec_input \
+                 _final_input \
+                 RSEQ_INJECT_INPUT \
+               : "r17", "memory", "cc" \
+                 _extra_clobber \
+                 RSEQ_INJECT_CLOBBER \
+               : _failure \
+       )
+
+#define RSEQ_FINISH_FINAL_STORE_ASM() \
+               "std %[to_write_final], 0(%[target_final])\n\t"
+
+#define RSEQ_FINISH_FINAL_STORE_RELEASE_ASM() \
+               "lwsync\n\t" \
+               RSEQ_FINISH_FINAL_STORE_ASM()
+
+#define RSEQ_FINISH_FINAL_STORE_INPUT(_target_final, _to_write_final) \
+               , [to_write_final]"r"(_to_write_final), \
+               [target_final]"b"(_target_final)
+
+#define RSEQ_FINISH_SPECULATIVE_STORE_ASM() \
+               "std %[to_write_spec], 0(%[target_spec])\n\t" \
+               RSEQ_INJECT_ASM(4)
+
+#define RSEQ_FINISH_SPECULATIVE_STORE_INPUT(_target_spec, _to_write_spec) \
+               , [to_write_spec]"r"(_to_write_spec), \
+               [target_spec]"b"(_target_spec)
+
+/* TODO: implement a faster memcpy. */
+#define RSEQ_FINISH_MEMCPY_STORE_ASM() \
+               "cmpdi %%r19, 0\n\t" \
+               "beq 333f\n\t" \
+               "addi %%r20, %%r20, -1\n\t" \
+               "addi %%r21, %%r21, -1\n\t" \
+               "222:\n\t" \
+               "lbzu %%r18, 1(%%r20)\n\t" \
+               "stbu %%r18, 1(%%r21)\n\t" \
+               "addi %%r19, %%r19, -1\n\t" \
+               "cmpdi %%r19, 0\n\t" \
+               "bne 222b\n\t" \
+               "333:\n\t" \
+               RSEQ_INJECT_ASM(4)
+
+#define RSEQ_FINISH_MEMCPY_STORE_INPUT(_target_memcpy, _to_write_memcpy, _len_memcpy) \
+               , [to_write_memcpy]"r"(_to_write_memcpy), \
+               [target_memcpy]"r"(_target_memcpy), \
+               [len_memcpy]"r"(_len_memcpy)
+
+#define RSEQ_FINISH_MEMCPY_CLOBBER() \
+               , "r18", "r19", "r20", "r21"
+
+#define RSEQ_FINISH_MEMCPY_SCRATCH()
+
+/*
+ * We use extra registers to hold the input registers, and we don't need to
+ * save and restore the input registers.
+ */
+#define RSEQ_FINISH_MEMCPY_SETUP() \
+               "mr %%r19, %[len_memcpy]\n\t" \
+               "mr %%r20, %[to_write_memcpy]\n\t" \
+               "mr %%r21, %[target_memcpy]\n\t" \
+
+#define RSEQ_FINISH_MEMCPY_TEARDOWN()
+
+#else  /* #ifdef __PPC64__ */
+
+#define RSEQ_FINISH_ASM(_target_final, _to_write_final, _start_value, \
+               _failure, _spec_store, _spec_input, \
+               _final_store, _final_input, _extra_clobber, \
+               _setup, _teardown, _scratch) \
+       __asm__ __volatile__ goto ( \
+               ".pushsection __rseq_table, \"aw\"\n\t" \
+               ".balign 32\n\t" \
+               "3:\n\t" \
+               /* 32-bit only supported on BE */ \
+               ".long 0x0, 1f, 0x0, 2f, 0x0, 4f, 0x0, 0x0\n\t" \
+               ".popsection\n\t" \
+               "1:\n\t" \
+               _setup \
+               RSEQ_INJECT_ASM(1) \
+               "lis %%r17, (3b)@ha\n\t" \
+               "addi %%r17, %%r17, (3b)@l\n\t" \
+               "stw %%r17, 0(%[rseq_cs])\n\t" \
+               RSEQ_INJECT_ASM(2) \
+               "lwz %%r17, %[current_event_counter]\n\t" \
+               "cmpw cr7, %[start_event_counter], %%r17\n\t" \
+               "bne- cr7, 4f\n\t" \
+               RSEQ_INJECT_ASM(3) \
+               _spec_store \
+               _final_store \
+               "2:\n\t" \
+               RSEQ_INJECT_ASM(5) \
+               "li %%r17, 0\n\t" \
+               "stw %%r17, 0(%[rseq_cs])\n\t" \
+               _teardown \
+               "b 5f\n\t" \
+               "4:\n\t" \
+               "li %%r17, 0\n\t" \
+               "std %%r17, 0(%[rseq_cs])\n\t" \
+               _teardown \
+               "b %l[failure]\n\t" \
+               "5:\n\t" \
+               : /* gcc asm goto does not allow outputs */ \
+               : [start_event_counter]"r"((_start_value).event_counter), \
+                 [current_event_counter]"m"((_start_value).rseqp->u.e.event_counter), \
+                 [rseq_cs]"b"(&(_start_value).rseqp->rseq_cs) \
+                 _spec_input \
+                 _final_input \
+                 RSEQ_INJECT_INPUT \
+               : "r17", "memory", "cc" \
+                 _extra_clobber \
+                 RSEQ_INJECT_CLOBBER \
+               : _failure \
+       )
+
+#define RSEQ_FINISH_FINAL_STORE_ASM() \
+               "stw %[to_write_final], 0(%[target_final])\n\t"
+
+#define RSEQ_FINISH_FINAL_STORE_RELEASE_ASM() \
+               "lwsync\n\t" \
+               RSEQ_FINISH_FINAL_STORE_ASM()
+
+#define RSEQ_FINISH_FINAL_STORE_INPUT(_target_final, _to_write_final) \
+               , [to_write_final]"r"(_to_write_final), \
+               [target_final]"b"(_target_final)
+
+#define RSEQ_FINISH_SPECULATIVE_STORE_ASM() \
+               "stw %[to_write_spec], 0(%[target_spec])\n\t" \
+               RSEQ_INJECT_ASM(4)
+
+#define RSEQ_FINISH_SPECULATIVE_STORE_INPUT(_target_spec, _to_write_spec) \
+               , [to_write_spec]"r"(_to_write_spec), \
+               [target_spec]"b"(_target_spec)
+
+/* TODO: implement a faster memcpy. */
+#define RSEQ_FINISH_MEMCPY_STORE_ASM() \
+               "cmpwi %%r19, 0\n\t" \
+               "beq 333f\n\t" \
+               "addi %%r20, %%r20, -1\n\t" \
+               "addi %%r21, %%r21, -1\n\t" \
+               "222:\n\t" \
+               "lbzu %%r18, 1(%%r20)\n\t" \
+               "stbu %%r18, 1(%%r21)\n\t" \
+               "addi %%r19, %%r19, -1\n\t" \
+               "cmpwi %%r19, 0\n\t" \
+               "bne 222b\n\t" \
+               "333:\n\t" \
+               RSEQ_INJECT_ASM(4)
+
+#define RSEQ_FINISH_MEMCPY_STORE_INPUT(_target_memcpy, _to_write_memcpy, _len_memcpy) \
+               , [to_write_memcpy]"r"(_to_write_memcpy), \
+               [target_memcpy]"r"(_target_memcpy), \
+               [len_memcpy]"r"(_len_memcpy)
+
+#define RSEQ_FINISH_MEMCPY_CLOBBER() \
+               , "r18", "r19", "r20", "r21"
+
+#define RSEQ_FINISH_MEMCPY_SCRATCH()
+
+/*
+ * We use extra registers to hold the input registers, and we don't need to
+ * save and restore the input registers.
+ */
+#define RSEQ_FINISH_MEMCPY_SETUP() \
+               "mr %%r19, %[len_memcpy]\n\t" \
+               "mr %%r20, %[to_write_memcpy]\n\t" \
+               "mr %%r21, %[target_memcpy]\n\t" \
+
+#define RSEQ_FINISH_MEMCPY_TEARDOWN()
+
+#endif /* #else #ifdef __PPC64__ */
diff --git a/tools/testing/selftests/rseq/rseq-x86.h b/tools/testing/selftests/rseq/rseq-x86.h
new file mode 100644 (file)
index 0000000..cca5ba2
--- /dev/null
@@ -0,0 +1,306 @@
+/*
+ * rseq-x86.h
+ *
+ * (C) Copyright 2016 - Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#ifdef __x86_64__
+
+#define smp_mb()       __asm__ __volatile__ ("mfence" : : : "memory")
+#define smp_rmb()      barrier()
+#define smp_wmb()      barrier()
+
+#define smp_load_acquire(p)                                            \
+__extension__ ({                                                       \
+       __typeof(*p) ____p1 = READ_ONCE(*p);                            \
+       barrier();                                                      \
+       ____p1;                                                         \
+})
+
+#define smp_acquire__after_ctrl_dep()  smp_rmb()
+
+#define smp_store_release(p, v)                                                \
+do {                                                                   \
+       barrier();                                                      \
+       WRITE_ONCE(*p, v);                                              \
+} while (0)
+
+#define has_fast_acquire_release()     1
+#define has_single_copy_load_64()      1
+
+/*
+ * The __rseq_table section can be used by debuggers to better handle
+ * single-stepping through the restartable critical sections.
+ */
+#define RSEQ_FINISH_ASM(_target_final, _to_write_final, _start_value, \
+               _failure, _spec_store, _spec_input, \
+               _final_store, _final_input, _extra_clobber, \
+               _setup, _teardown, _scratch) \
+do { \
+       _scratch \
+       __asm__ __volatile__ goto ( \
+               ".pushsection __rseq_table, \"aw\"\n\t" \
+               ".balign 32\n\t" \
+               "3:\n\t" \
+               ".quad 1f, 2f, 4f, 0x0\n\t" \
+               ".popsection\n\t" \
+               "1:\n\t" \
+               _setup \
+               RSEQ_INJECT_ASM(1) \
+               "movq $3b, %[rseq_cs]\n\t" \
+               RSEQ_INJECT_ASM(2) \
+               "cmpl %[start_event_counter], %[current_event_counter]\n\t" \
+               "jnz 4f\n\t" \
+               RSEQ_INJECT_ASM(3) \
+               _spec_store \
+               _final_store \
+               "2:\n\t" \
+               RSEQ_INJECT_ASM(5) \
+               "movq $0, %[rseq_cs]\n\t" \
+               _teardown \
+               ".pushsection __rseq_failure, \"a\"\n\t" \
+               "4:\n\t" \
+               "movq $0, %[rseq_cs]\n\t" \
+               _teardown \
+               "jmp %l[failure]\n\t" \
+               ".popsection\n\t" \
+               : /* gcc asm goto does not allow outputs */ \
+               : [start_event_counter]"r"((_start_value).event_counter), \
+                 [current_event_counter]"m"((_start_value).rseqp->u.e.event_counter), \
+                 [rseq_cs]"m"((_start_value).rseqp->rseq_cs) \
+                 _spec_input \
+                 _final_input \
+                 RSEQ_INJECT_INPUT \
+               : "memory", "cc" \
+                 _extra_clobber \
+                 RSEQ_INJECT_CLOBBER \
+               : _failure \
+       ); \
+} while (0)
+
+#define RSEQ_FINISH_FINAL_STORE_ASM() \
+               "movq %[to_write_final], %[target_final]\n\t"
+
+/* x86-64 is TSO */
+#define RSEQ_FINISH_FINAL_STORE_RELEASE_ASM() \
+               RSEQ_FINISH_FINAL_STORE_ASM()
+
+#define RSEQ_FINISH_FINAL_STORE_INPUT(_target_final, _to_write_final) \
+               , [to_write_final]"r"(_to_write_final), \
+               [target_final]"m"(*(_target_final))
+
+#define RSEQ_FINISH_SPECULATIVE_STORE_ASM() \
+               "movq %[to_write_spec], %[target_spec]\n\t" \
+               RSEQ_INJECT_ASM(4)
+
+#define RSEQ_FINISH_SPECULATIVE_STORE_INPUT(_target_spec, _to_write_spec) \
+               , [to_write_spec]"r"(_to_write_spec), \
+               [target_spec]"m"(*(_target_spec))
+
+/* TODO: implement a faster memcpy. */
+#define RSEQ_FINISH_MEMCPY_STORE_ASM() \
+               "test %[len_memcpy], %[len_memcpy]\n\t" \
+               "jz 333f\n\t" \
+               "222:\n\t" \
+               "movb (%[to_write_memcpy]), %%al\n\t" \
+               "movb %%al, (%[target_memcpy])\n\t" \
+               "inc %[to_write_memcpy]\n\t" \
+               "inc %[target_memcpy]\n\t" \
+               "dec %[len_memcpy]\n\t" \
+               "jnz 222b\n\t" \
+               "333:\n\t" \
+               RSEQ_INJECT_ASM(4)
+
+#define RSEQ_FINISH_MEMCPY_STORE_INPUT(_target_memcpy, _to_write_memcpy, _len_memcpy) \
+               , [to_write_memcpy]"r"(_to_write_memcpy), \
+               [target_memcpy]"r"(_target_memcpy), \
+               [len_memcpy]"r"(_len_memcpy), \
+               [rseq_scratch0]"m"(rseq_scratch[0]), \
+               [rseq_scratch1]"m"(rseq_scratch[1]), \
+               [rseq_scratch2]"m"(rseq_scratch[2])
+
+#define RSEQ_FINISH_MEMCPY_CLOBBER()   \
+               , "rax"
+
+#define RSEQ_FINISH_MEMCPY_SCRATCH() \
+               uint64_t rseq_scratch[3];
+
+/*
+ * We need to save and restore those input registers so they can be
+ * modified within the assembly.
+ */
+#define RSEQ_FINISH_MEMCPY_SETUP() \
+               "movq %[to_write_memcpy], %[rseq_scratch0]\n\t" \
+               "movq %[target_memcpy], %[rseq_scratch1]\n\t" \
+               "movq %[len_memcpy], %[rseq_scratch2]\n\t"
+
+#define RSEQ_FINISH_MEMCPY_TEARDOWN() \
+               "movq %[rseq_scratch2], %[len_memcpy]\n\t" \
+               "movq %[rseq_scratch1], %[target_memcpy]\n\t" \
+               "movq %[rseq_scratch0], %[to_write_memcpy]\n\t"
+
+#elif __i386__
+
+/*
+ * Support older 32-bit architectures that do not implement fence
+ * instructions.
+ */
+#define smp_mb()       \
+       __asm__ __volatile__ ("lock; addl $0,0(%%esp)" : : : "memory")
+#define smp_rmb()      \
+       __asm__ __volatile__ ("lock; addl $0,0(%%esp)" : : : "memory")
+#define smp_wmb()      \
+       __asm__ __volatile__ ("lock; addl $0,0(%%esp)" : : : "memory")
+
+#define smp_load_acquire(p)                                            \
+__extension__ ({                                                       \
+       __typeof(*p) ____p1 = READ_ONCE(*p);                            \
+       smp_mb();                                                       \
+       ____p1;                                                         \
+})
+
+#define smp_acquire__after_ctrl_dep()  smp_rmb()
+
+#define smp_store_release(p, v)                                                \
+do {                                                                   \
+       smp_mb();                                                       \
+       WRITE_ONCE(*p, v);                                              \
+} while (0)
+
+#define has_fast_acquire_release()     0
+#define has_single_copy_load_64()      0
+
+/*
+ * Use eax as scratch register and take memory operands as input to
+ * lessen register pressure. Especially needed when compiling
+ * do_rseq_memcpy() in O0.
+ */
+#define RSEQ_FINISH_ASM(_target_final, _to_write_final, _start_value, \
+               _failure, _spec_store, _spec_input, \
+               _final_store, _final_input, _extra_clobber, \
+               _setup, _teardown, _scratch) \
+do { \
+       _scratch \
+       __asm__ __volatile__ goto ( \
+               ".pushsection __rseq_table, \"aw\"\n\t" \
+               ".balign 32\n\t" \
+               "3:\n\t" \
+               ".long 1f, 0x0, 2f, 0x0, 4f, 0x0, 0x0, 0x0\n\t" \
+               ".popsection\n\t" \
+               "1:\n\t" \
+               _setup \
+               RSEQ_INJECT_ASM(1) \
+               "movl $3b, %[rseq_cs]\n\t" \
+               RSEQ_INJECT_ASM(2) \
+               "movl %[start_event_counter], %%eax\n\t" \
+               "cmpl %%eax, %[current_event_counter]\n\t" \
+               "jnz 4f\n\t" \
+               RSEQ_INJECT_ASM(3) \
+               _spec_store \
+               _final_store \
+               "2:\n\t" \
+               RSEQ_INJECT_ASM(5) \
+               "movl $0, %[rseq_cs]\n\t" \
+               _teardown \
+               ".pushsection __rseq_failure, \"a\"\n\t" \
+               "4:\n\t" \
+               "movl $0, %[rseq_cs]\n\t" \
+               _teardown \
+               "jmp %l[failure]\n\t" \
+               ".popsection\n\t" \
+               : /* gcc asm goto does not allow outputs */ \
+               : [start_event_counter]"m"((_start_value).event_counter), \
+                 [current_event_counter]"m"((_start_value).rseqp->u.e.event_counter), \
+                 [rseq_cs]"m"((_start_value).rseqp->rseq_cs) \
+                 _spec_input \
+                 _final_input \
+                 RSEQ_INJECT_INPUT \
+               : "memory", "cc", "eax" \
+                 _extra_clobber \
+                 RSEQ_INJECT_CLOBBER \
+               : _failure \
+       ); \
+} while (0)
+
+#define RSEQ_FINISH_FINAL_STORE_ASM() \
+               "movl %[to_write_final], %%eax\n\t" \
+               "movl %%eax, %[target_final]\n\t"
+
+#define RSEQ_FINISH_FINAL_STORE_RELEASE_ASM() \
+               "lock; addl $0,0(%%esp)\n\t" \
+               RSEQ_FINISH_FINAL_STORE_ASM()
+
+#define RSEQ_FINISH_FINAL_STORE_INPUT(_target_final, _to_write_final) \
+               , [to_write_final]"m"(_to_write_final), \
+               [target_final]"m"(*(_target_final))
+
+#define RSEQ_FINISH_SPECULATIVE_STORE_ASM() \
+               "movl %[to_write_spec], %%eax\n\t" \
+               "movl %%eax, %[target_spec]\n\t" \
+               RSEQ_INJECT_ASM(4)
+
+#define RSEQ_FINISH_SPECULATIVE_STORE_INPUT(_target_spec, _to_write_spec) \
+               , [to_write_spec]"m"(_to_write_spec), \
+               [target_spec]"m"(*(_target_spec))
+
+/* TODO: implement a faster memcpy. */
+#define RSEQ_FINISH_MEMCPY_STORE_ASM() \
+               "movl %[len_memcpy], %%eax\n\t" \
+               "test %%eax, %%eax\n\t" \
+               "jz 333f\n\t" \
+               "222:\n\t" \
+               "movb (%[to_write_memcpy]), %%al\n\t" \
+               "movb %%al, (%[target_memcpy])\n\t" \
+               "inc %[to_write_memcpy]\n\t" \
+               "inc %[target_memcpy]\n\t" \
+               "decl %[rseq_scratch2]\n\t" \
+               "jnz 222b\n\t" \
+               "333:\n\t" \
+               RSEQ_INJECT_ASM(4)
+
+#define RSEQ_FINISH_MEMCPY_STORE_INPUT(_target_memcpy, _to_write_memcpy, _len_memcpy) \
+               , [to_write_memcpy]"r"(_to_write_memcpy), \
+               [target_memcpy]"r"(_target_memcpy), \
+               [len_memcpy]"m"(_len_memcpy), \
+               [rseq_scratch0]"m"(rseq_scratch[0]), \
+               [rseq_scratch1]"m"(rseq_scratch[1]), \
+               [rseq_scratch2]"m"(rseq_scratch[2])
+
+#define RSEQ_FINISH_MEMCPY_CLOBBER()
+
+#define RSEQ_FINISH_MEMCPY_SCRATCH() \
+               uint32_t rseq_scratch[3];
+
+/*
+ * We need to save and restore those input registers so they can be
+ * modified within the assembly.
+ */
+#define RSEQ_FINISH_MEMCPY_SETUP() \
+               "movl %[to_write_memcpy], %[rseq_scratch0]\n\t" \
+               "movl %[target_memcpy], %[rseq_scratch1]\n\t" \
+               "movl %[len_memcpy], %%eax\n\t" \
+               "movl %%eax, %[rseq_scratch2]\n\t"
+
+#define RSEQ_FINISH_MEMCPY_TEARDOWN() \
+               "movl %[rseq_scratch1], %[target_memcpy]\n\t" \
+               "movl %[rseq_scratch0], %[to_write_memcpy]\n\t"
+
+#endif
diff --git a/tools/testing/selftests/rseq/rseq.c b/tools/testing/selftests/rseq/rseq.c
new file mode 100644 (file)
index 0000000..c8193a3
--- /dev/null
@@ -0,0 +1,247 @@
+/*
+ * rseq.c
+ *
+ * Copyright (C) 2016 Mathieu Desnoyers <mathieu.desnoyers@efficios.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; only
+ * version 2.1 of the License.
+ *
+ * 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.
+ */
+
+#define _GNU_SOURCE
+#include <errno.h>
+#include <sched.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <unistd.h>
+#include <syscall.h>
+#include <assert.h>
+#include <signal.h>
+#include <linux/membarrier.h>
+
+#include <rseq.h>
+
+#ifdef __NR_membarrier
+# define membarrier(...)               syscall(__NR_membarrier, __VA_ARGS__)
+#else
+# define membarrier(...)               -ENOSYS
+#endif
+
+struct rseq_thread_state {
+       uint32_t fallback_wait_cnt;
+       uint32_t fallback_cnt;
+       sigset_t sigmask_saved;
+};
+
+__attribute__((weak)) __thread volatile struct rseq __rseq_abi = {
+       .u.e.cpu_id = -1,
+};
+
+static __thread volatile struct rseq_thread_state rseq_thread_state;
+
+int rseq_has_sys_membarrier;
+
+static int sys_rseq(volatile struct rseq *rseq_abi, int flags)
+{
+       return syscall(__NR_rseq, rseq_abi, flags);
+}
+
+int rseq_register_current_thread(void)
+{
+       int rc;
+
+       rc = sys_rseq(&__rseq_abi, 0);
+       if (rc) {
+               fprintf(stderr, "Error: sys_rseq(...) failed(%d): %s\n",
+                       errno, strerror(errno));
+               return -1;
+       }
+       assert(rseq_current_cpu() >= 0);
+       return 0;
+}
+
+int rseq_unregister_current_thread(void)
+{
+       int rc;
+
+       rc = sys_rseq(NULL, 0);
+       if (rc) {
+               fprintf(stderr, "Error: sys_rseq(...) failed(%d): %s\n",
+                       errno, strerror(errno));
+               return -1;
+       }
+       return 0;
+}
+
+int rseq_init_lock(struct rseq_lock *rlock)
+{
+       int ret;
+
+       ret = pthread_mutex_init(&rlock->lock, NULL);
+       if (ret) {
+               errno = ret;
+               return -1;
+       }
+       rlock->state = RSEQ_LOCK_STATE_RESTART;
+       return 0;
+}
+
+int rseq_destroy_lock(struct rseq_lock *rlock)
+{
+       int ret;
+
+       ret = pthread_mutex_destroy(&rlock->lock);
+       if (ret) {
+               errno = ret;
+               return -1;
+       }
+       return 0;
+}
+
+static void signal_off_save(sigset_t *oldset)
+{
+       sigset_t set;
+       int ret;
+
+       sigfillset(&set);
+       ret = pthread_sigmask(SIG_BLOCK, &set, oldset);
+       if (ret)
+               abort();
+}
+
+static void signal_restore(sigset_t oldset)
+{
+       int ret;
+
+       ret = pthread_sigmask(SIG_SETMASK, &oldset, NULL);
+       if (ret)
+               abort();
+}
+
+static void rseq_fallback_lock(struct rseq_lock *rlock)
+{
+       signal_off_save((sigset_t *)&rseq_thread_state.sigmask_saved);
+       pthread_mutex_lock(&rlock->lock);
+       rseq_thread_state.fallback_cnt++;
+       /*
+        * For concurrent threads arriving before we set LOCK:
+        * reading cpu_id after setting the state to LOCK
+        * ensures they restart.
+        */
+       ACCESS_ONCE(rlock->state) = RSEQ_LOCK_STATE_LOCK;
+       /*
+        * For concurrent threads arriving after we set LOCK:
+        * those will grab the lock, so we are protected by
+        * mutual exclusion.
+        */
+}
+
+void rseq_fallback_wait(struct rseq_lock *rlock)
+{
+       signal_off_save((sigset_t *)&rseq_thread_state.sigmask_saved);
+       pthread_mutex_lock(&rlock->lock);
+       rseq_thread_state.fallback_wait_cnt++;
+       pthread_mutex_unlock(&rlock->lock);
+       signal_restore(rseq_thread_state.sigmask_saved);
+}
+
+static void rseq_fallback_unlock(struct rseq_lock *rlock, int cpu_at_start)
+{
+       /*
+        * Concurrent rseq arriving before we set state back to RESTART
+        * grab the lock. Those arriving after we set state back to
+        * RESTART will perform restartable critical sections. The next
+        * owner of the lock will take take of making sure it prevents
+        * concurrent restartable sequences from completing.  We may be
+        * writing from another CPU, so update the state with a store
+        * release semantic to ensure restartable sections will see our
+        * side effect (writing to *p) before they enter their
+        * restartable critical section.
+        *
+        * In cases where we observe that we are on the right CPU after the
+        * critical section, program order ensures that following restartable
+        * critical sections will see our stores, so we don't have to use
+        * store-release or membarrier.
+        *
+        * Use sys_membarrier when available to remove the memory barrier
+        * implied by smp_load_acquire().
+        */
+       barrier();
+       if (likely(rseq_current_cpu() == cpu_at_start)) {
+               ACCESS_ONCE(rlock->state) = RSEQ_LOCK_STATE_RESTART;
+       } else {
+               if (!has_fast_acquire_release() && rseq_has_sys_membarrier) {
+                       if (membarrier(MEMBARRIER_CMD_SHARED, 0))
+                               abort();
+                       ACCESS_ONCE(rlock->state) = RSEQ_LOCK_STATE_RESTART;
+               } else {
+                       /*
+                        * Store with release semantic to ensure
+                        * restartable sections will see our side effect
+                        * (writing to *p) before they enter their
+                        * restartable critical section. Matches
+                        * smp_load_acquire() in rseq_start().
+                        */
+                       smp_store_release(&rlock->state,
+                               RSEQ_LOCK_STATE_RESTART);
+               }
+       }
+       pthread_mutex_unlock(&rlock->lock);
+       signal_restore(rseq_thread_state.sigmask_saved);
+}
+
+int rseq_fallback_current_cpu(void)
+{
+       int cpu;
+
+       cpu = sched_getcpu();
+       if (cpu < 0) {
+               perror("sched_getcpu()");
+               abort();
+       }
+       return cpu;
+}
+
+int rseq_fallback_begin(struct rseq_lock *rlock)
+{
+       rseq_fallback_lock(rlock);
+       return rseq_fallback_current_cpu();
+}
+
+void rseq_fallback_end(struct rseq_lock *rlock, int cpu)
+{
+       rseq_fallback_unlock(rlock, cpu);
+}
+
+/* Handle non-initialized rseq for this thread. */
+void rseq_fallback_noinit(struct rseq_state *rseq_state)
+{
+       rseq_state->lock_state = RSEQ_LOCK_STATE_FAIL;
+       rseq_state->cpu_id = 0;
+}
+
+uint32_t rseq_get_fallback_wait_cnt(void)
+{
+       return rseq_thread_state.fallback_wait_cnt;
+}
+
+uint32_t rseq_get_fallback_cnt(void)
+{
+       return rseq_thread_state.fallback_cnt;
+}
+
+void __attribute__((constructor)) rseq_init(void)
+{
+       int ret;
+
+       ret = membarrier(MEMBARRIER_CMD_QUERY, 0);
+       if (ret >= 0 && (ret & MEMBARRIER_CMD_SHARED))
+               rseq_has_sys_membarrier = 1;
+}
diff --git a/tools/testing/selftests/rseq/rseq.h b/tools/testing/selftests/rseq/rseq.h
new file mode 100644 (file)
index 0000000..b0c7434
--- /dev/null
@@ -0,0 +1,477 @@
+/*
+ * rseq.h
+ *
+ * (C) Copyright 2016 - Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#ifndef RSEQ_H
+#define RSEQ_H
+
+#include <stdint.h>
+#include <stdbool.h>
+#include <pthread.h>
+#include <signal.h>
+#include <sched.h>
+#include <errno.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <sched.h>
+#include <linux/rseq.h>
+
+/*
+ * Empty code injection macros, override when testing.
+ * It is important to consider that the ASM injection macros need to be
+ * fully reentrant (e.g. do not modify the stack).
+ */
+#ifndef RSEQ_INJECT_ASM
+#define RSEQ_INJECT_ASM(n)
+#endif
+
+#ifndef RSEQ_INJECT_C
+#define RSEQ_INJECT_C(n)
+#endif
+
+#ifndef RSEQ_INJECT_INPUT
+#define RSEQ_INJECT_INPUT
+#endif
+
+#ifndef RSEQ_INJECT_CLOBBER
+#define RSEQ_INJECT_CLOBBER
+#endif
+
+#ifndef RSEQ_INJECT_FAILED
+#define RSEQ_INJECT_FAILED
+#endif
+
+#ifndef RSEQ_FALLBACK_CNT
+#define RSEQ_FALLBACK_CNT      3
+#endif
+
+uint32_t rseq_get_fallback_wait_cnt(void);
+uint32_t rseq_get_fallback_cnt(void);
+
+extern __thread volatile struct rseq __rseq_abi;
+extern int rseq_has_sys_membarrier;
+
+#define likely(x)              __builtin_expect(!!(x), 1)
+#define unlikely(x)            __builtin_expect(!!(x), 0)
+#define barrier()              __asm__ __volatile__("" : : : "memory")
+
+#define ACCESS_ONCE(x)         (*(__volatile__  __typeof__(x) *)&(x))
+#define WRITE_ONCE(x, v)       __extension__ ({ ACCESS_ONCE(x) = (v); })
+#define READ_ONCE(x)           ACCESS_ONCE(x)
+
+#if defined(__x86_64__) || defined(__i386__)
+#include <rseq-x86.h>
+#elif defined(__ARMEL__)
+#include <rseq-arm.h>
+#elif defined(__PPC__)
+#include <rseq-ppc.h>
+#else
+#error unsupported target
+#endif
+
+enum rseq_lock_state {
+       RSEQ_LOCK_STATE_RESTART = 0,
+       RSEQ_LOCK_STATE_LOCK = 1,
+       RSEQ_LOCK_STATE_FAIL = 2,
+};
+
+struct rseq_lock {
+       pthread_mutex_t lock;
+       int32_t state;          /* enum rseq_lock_state */
+};
+
+/* State returned by rseq_start, passed as argument to rseq_finish. */
+struct rseq_state {
+       volatile struct rseq *rseqp;
+       int32_t cpu_id;         /* cpu_id at start. */
+       uint32_t event_counter; /* event_counter at start. */
+       int32_t lock_state;     /* Lock state at start. */
+};
+
+/*
+ * Register rseq for the current thread. This needs to be called once
+ * by any thread which uses restartable sequences, before they start
+ * using restartable sequences. If initialization is not invoked, or if
+ * it fails, the restartable critical sections will fall-back on locking
+ * (rseq_lock).
+ */
+int rseq_register_current_thread(void);
+
+/*
+ * Unregister rseq for current thread.
+ */
+int rseq_unregister_current_thread(void);
+
+/*
+ * The fallback lock should be initialized before being used by any
+ * thread, and destroyed after all threads are done using it. This lock
+ * should be used by all rseq calls associated with shared data, either
+ * between threads, or between processes in a shared memory.
+ *
+ * There may be many rseq_lock per process, e.g. one per protected data
+ * structure.
+ */
+int rseq_init_lock(struct rseq_lock *rlock);
+int rseq_destroy_lock(struct rseq_lock *rlock);
+
+/*
+ * Restartable sequence fallback prototypes. Fallback on locking when
+ * rseq is not initialized, not available on the system, or during
+ * single-stepping to ensure forward progress.
+ */
+int rseq_fallback_begin(struct rseq_lock *rlock);
+void rseq_fallback_end(struct rseq_lock *rlock, int cpu);
+void rseq_fallback_wait(struct rseq_lock *rlock);
+void rseq_fallback_noinit(struct rseq_state *rseq_state);
+
+/*
+ * Restartable sequence fallback for reading the current CPU number.
+ */
+int rseq_fallback_current_cpu(void);
+
+static inline int32_t rseq_cpu_at_start(struct rseq_state start_value)
+{
+       return start_value.cpu_id;
+}
+
+static inline int32_t rseq_current_cpu_raw(void)
+{
+       return ACCESS_ONCE(__rseq_abi.u.e.cpu_id);
+}
+
+static inline int32_t rseq_current_cpu(void)
+{
+       int32_t cpu;
+
+       cpu = rseq_current_cpu_raw();
+       if (unlikely(cpu < 0))
+               cpu = rseq_fallback_current_cpu();
+       return cpu;
+}
+
+static inline __attribute__((always_inline))
+struct rseq_state rseq_start(struct rseq_lock *rlock)
+{
+       struct rseq_state result;
+
+       result.rseqp = &__rseq_abi;
+       if (has_single_copy_load_64()) {
+               union rseq_cpu_event u;
+
+               u.v = ACCESS_ONCE(result.rseqp->u.v);
+               result.event_counter = u.e.event_counter;
+               result.cpu_id = u.e.cpu_id;
+       } else {
+               result.event_counter =
+                       ACCESS_ONCE(result.rseqp->u.e.event_counter);
+               /* load event_counter before cpu_id. */
+               RSEQ_INJECT_C(6)
+               result.cpu_id = ACCESS_ONCE(result.rseqp->u.e.cpu_id);
+       }
+       /*
+        * Read event counter before lock state and cpu_id. This ensures
+        * that when the state changes from RESTART to LOCK, if we have
+        * some threads that have already seen the RESTART still in
+        * flight, they will necessarily be preempted/signalled before a
+        * thread can see the LOCK state for that same CPU. That
+        * preemption/signalling will cause them to restart, so they
+        * don't interfere with the lock.
+        */
+       RSEQ_INJECT_C(7)
+
+       if (!has_fast_acquire_release() && likely(rseq_has_sys_membarrier)) {
+               result.lock_state = ACCESS_ONCE(rlock->state);
+               barrier();
+       } else {
+               /*
+                * Load lock state with acquire semantic. Matches
+                * smp_store_release() in rseq_fallback_end().
+                */
+               result.lock_state = smp_load_acquire(&rlock->state);
+       }
+       if (unlikely(result.cpu_id < 0))
+               rseq_fallback_noinit(&result);
+       /*
+        * Ensure the compiler does not re-order loads of protected
+        * values before we load the event counter.
+        */
+       barrier();
+       return result;
+}
+
+enum rseq_finish_type {
+       RSEQ_FINISH_SINGLE,
+       RSEQ_FINISH_TWO,
+       RSEQ_FINISH_MEMCPY,
+};
+
+/*
+ * p_spec and to_write_spec are used for a speculative write attempted
+ * near the end of the restartable sequence. A rseq_finish2 may fail
+ * even after this write takes place.
+ *
+ * p_final and to_write_final are used for the final write. If this
+ * write takes place, the rseq_finish2 is guaranteed to succeed.
+ */
+static inline __attribute__((always_inline))
+bool __rseq_finish(struct rseq_lock *rlock,
+               intptr_t *p_spec, intptr_t to_write_spec,
+               void *p_memcpy, void *to_write_memcpy, size_t len_memcpy,
+               intptr_t *p_final, intptr_t to_write_final,
+               struct rseq_state start_value,
+               enum rseq_finish_type type, bool release)
+{
+       RSEQ_INJECT_C(9)
+
+       if (unlikely(start_value.lock_state != RSEQ_LOCK_STATE_RESTART)) {
+               if (start_value.lock_state == RSEQ_LOCK_STATE_LOCK)
+                       rseq_fallback_wait(rlock);
+               return false;
+       }
+       switch (type) {
+       case RSEQ_FINISH_SINGLE:
+               RSEQ_FINISH_ASM(p_final, to_write_final, start_value, failure,
+                       /* no speculative write */, /* no speculative write */,
+                       RSEQ_FINISH_FINAL_STORE_ASM(),
+                       RSEQ_FINISH_FINAL_STORE_INPUT(p_final, to_write_final),
+                       /* no extra clobber */, /* no arg */, /* no arg */,
+                       /* no arg */
+               );
+               break;
+       case RSEQ_FINISH_TWO:
+               if (release) {
+                       RSEQ_FINISH_ASM(p_final, to_write_final, start_value, failure,
+                               RSEQ_FINISH_SPECULATIVE_STORE_ASM(),
+                               RSEQ_FINISH_SPECULATIVE_STORE_INPUT(p_spec, to_write_spec),
+                               RSEQ_FINISH_FINAL_STORE_RELEASE_ASM(),
+                               RSEQ_FINISH_FINAL_STORE_INPUT(p_final, to_write_final),
+                               /* no extra clobber */, /* no arg */, /* no arg */,
+                               /* no arg */
+                       );
+               } else {
+                       RSEQ_FINISH_ASM(p_final, to_write_final, start_value, failure,
+                               RSEQ_FINISH_SPECULATIVE_STORE_ASM(),
+                               RSEQ_FINISH_SPECULATIVE_STORE_INPUT(p_spec, to_write_spec),
+                               RSEQ_FINISH_FINAL_STORE_ASM(),
+                               RSEQ_FINISH_FINAL_STORE_INPUT(p_final, to_write_final),
+                               /* no extra clobber */, /* no arg */, /* no arg */,
+                               /* no arg */
+                       );
+               }
+               break;
+       case RSEQ_FINISH_MEMCPY:
+               if (release) {
+                       RSEQ_FINISH_ASM(p_final, to_write_final, start_value, failure,
+                               RSEQ_FINISH_MEMCPY_STORE_ASM(),
+                               RSEQ_FINISH_MEMCPY_STORE_INPUT(p_memcpy, to_write_memcpy, len_memcpy),
+                               RSEQ_FINISH_FINAL_STORE_RELEASE_ASM(),
+                               RSEQ_FINISH_FINAL_STORE_INPUT(p_final, to_write_final),
+                               RSEQ_FINISH_MEMCPY_CLOBBER(),
+                               RSEQ_FINISH_MEMCPY_SETUP(),
+                               RSEQ_FINISH_MEMCPY_TEARDOWN(),
+                               RSEQ_FINISH_MEMCPY_SCRATCH()
+                       );
+               } else {
+                       RSEQ_FINISH_ASM(p_final, to_write_final, start_value, failure,
+                               RSEQ_FINISH_MEMCPY_STORE_ASM(),
+                               RSEQ_FINISH_MEMCPY_STORE_INPUT(p_memcpy, to_write_memcpy, len_memcpy),
+                               RSEQ_FINISH_FINAL_STORE_ASM(),
+                               RSEQ_FINISH_FINAL_STORE_INPUT(p_final, to_write_final),
+                               RSEQ_FINISH_MEMCPY_CLOBBER(),
+                               RSEQ_FINISH_MEMCPY_SETUP(),
+                               RSEQ_FINISH_MEMCPY_TEARDOWN(),
+                               RSEQ_FINISH_MEMCPY_SCRATCH()
+                       );
+               }
+               break;
+       }
+       return true;
+failure:
+       RSEQ_INJECT_FAILED
+       return false;
+}
+
+static inline __attribute__((always_inline))
+bool rseq_finish(struct rseq_lock *rlock,
+               intptr_t *p, intptr_t to_write,
+               struct rseq_state start_value)
+{
+       return __rseq_finish(rlock, NULL, 0,
+                       NULL, NULL, 0,
+                       p, to_write, start_value,
+                       RSEQ_FINISH_SINGLE, false);
+}
+
+static inline __attribute__((always_inline))
+bool rseq_finish2(struct rseq_lock *rlock,
+               intptr_t *p_spec, intptr_t to_write_spec,
+               intptr_t *p_final, intptr_t to_write_final,
+               struct rseq_state start_value)
+{
+       return __rseq_finish(rlock, p_spec, to_write_spec,
+                       NULL, NULL, 0,
+                       p_final, to_write_final, start_value,
+                       RSEQ_FINISH_TWO, false);
+}
+
+static inline __attribute__((always_inline))
+bool rseq_finish2_release(struct rseq_lock *rlock,
+               intptr_t *p_spec, intptr_t to_write_spec,
+               intptr_t *p_final, intptr_t to_write_final,
+               struct rseq_state start_value)
+{
+       return __rseq_finish(rlock, p_spec, to_write_spec,
+                       NULL, NULL, 0,
+                       p_final, to_write_final, start_value,
+                       RSEQ_FINISH_TWO, true);
+}
+
+static inline __attribute__((always_inline))
+bool rseq_finish_memcpy(struct rseq_lock *rlock,
+               void *p_memcpy, void *to_write_memcpy, size_t len_memcpy,
+               intptr_t *p_final, intptr_t to_write_final,
+               struct rseq_state start_value)
+{
+       return __rseq_finish(rlock, NULL, 0,
+                       p_memcpy, to_write_memcpy, len_memcpy,
+                       p_final, to_write_final, start_value,
+                       RSEQ_FINISH_MEMCPY, false);
+}
+
+static inline __attribute__((always_inline))
+bool rseq_finish_memcpy_release(struct rseq_lock *rlock,
+               void *p_memcpy, void *to_write_memcpy, size_t len_memcpy,
+               intptr_t *p_final, intptr_t to_write_final,
+               struct rseq_state start_value)
+{
+       return __rseq_finish(rlock, NULL, 0,
+                       p_memcpy, to_write_memcpy, len_memcpy,
+                       p_final, to_write_final, start_value,
+                       RSEQ_FINISH_MEMCPY, true);
+}
+
+#define __rseq_store_RSEQ_FINISH_SINGLE(_targetptr_spec, _newval_spec, \
+               _dest_memcpy, _src_memcpy, _len_memcpy,                 \
+               _targetptr_final, _newval_final)                        \
+       do {                                                            \
+               *(_targetptr_final) = (_newval_final);                  \
+       } while (0)
+
+#define __rseq_store_RSEQ_FINISH_TWO(_targetptr_spec, _newval_spec,    \
+               _dest_memcpy, _src_memcpy, _len_memcpy,                 \
+               _targetptr_final, _newval_final)                        \
+       do {                                                            \
+               *(_targetptr_spec) = (_newval_spec);                    \
+               *(_targetptr_final) = (_newval_final);                  \
+       } while (0)
+
+#define __rseq_store_RSEQ_FINISH_MEMCPY(_targetptr_spec,               \
+               _newval_spec, _dest_memcpy, _src_memcpy, _len_memcpy,   \
+               _targetptr_final, _newval_final)                        \
+       do {                                                            \
+               memcpy(_dest_memcpy, _src_memcpy, _len_memcpy);         \
+               *(_targetptr_final) = (_newval_final);                  \
+       } while (0)
+
+/*
+ * Helper macro doing two restartable critical section attempts, and if
+ * they fail, fallback on locking.
+ */
+#define __do_rseq(_type, _lock, _rseq_state, _cpu, _result,            \
+               _targetptr_spec, _newval_spec,                          \
+               _dest_memcpy, _src_memcpy, _len_memcpy,                 \
+               _targetptr_final, _newval_final, _code, _release)       \
+       do {                                                            \
+               _rseq_state = rseq_start(_lock);                        \
+               _cpu = rseq_cpu_at_start(_rseq_state);                  \
+               _result = true;                                         \
+               _code                                                   \
+               if (unlikely(!_result))                                 \
+                       break;                                          \
+               if (likely(__rseq_finish(_lock,                         \
+                               _targetptr_spec, _newval_spec,          \
+                               _dest_memcpy, _src_memcpy, _len_memcpy, \
+                               _targetptr_final, _newval_final,        \
+                               _rseq_state, _type, _release)))         \
+                       break;                                          \
+               _rseq_state = rseq_start(_lock);                        \
+               _cpu = rseq_cpu_at_start(_rseq_state);                  \
+               _result = true;                                         \
+               _code                                                   \
+               if (unlikely(!_result))                                 \
+                       break;                                          \
+               if (likely(__rseq_finish(_lock,                         \
+                               _targetptr_spec, _newval_spec,          \
+                               _dest_memcpy, _src_memcpy, _len_memcpy, \
+                               _targetptr_final, _newval_final,        \
+                               _rseq_state, _type, _release)))         \
+                       break;                                          \
+               _cpu = rseq_fallback_begin(_lock);                      \
+               _result = true;                                         \
+               _code                                                   \
+               if (likely(_result))                                    \
+                       __rseq_store_##_type(_targetptr_spec,           \
+                                _newval_spec, _dest_memcpy,            \
+                               _src_memcpy, _len_memcpy,               \
+                               _targetptr_final, _newval_final);       \
+               rseq_fallback_end(_lock, _cpu);                         \
+       } while (0)
+
+#define do_rseq(_lock, _rseq_state, _cpu, _result, _targetptr, _newval,        \
+               _code)                                                  \
+       __do_rseq(RSEQ_FINISH_SINGLE, _lock, _rseq_state, _cpu, _result,\
+               NULL, 0, NULL, NULL, 0, _targetptr, _newval, _code, false)
+
+#define do_rseq2(_lock, _rseq_state, _cpu, _result,                    \
+               _targetptr_spec, _newval_spec,                          \
+               _targetptr_final, _newval_final, _code)                 \
+       __do_rseq(RSEQ_FINISH_TWO, _lock, _rseq_state, _cpu, _result,   \
+               _targetptr_spec, _newval_spec,                          \
+               NULL, NULL, 0,                                          \
+               _targetptr_final, _newval_final, _code, false)
+
+#define do_rseq2_release(_lock, _rseq_state, _cpu, _result,            \
+               _targetptr_spec, _newval_spec,                          \
+               _targetptr_final, _newval_final, _code)                 \
+       __do_rseq(RSEQ_FINISH_TWO, _lock, _rseq_state, _cpu, _result,   \
+               _targetptr_spec, _newval_spec,                          \
+               NULL, NULL, 0,                                          \
+               _targetptr_final, _newval_final, _code, true)
+
+#define do_rseq_memcpy(_lock, _rseq_state, _cpu, _result,              \
+               _dest_memcpy, _src_memcpy, _len_memcpy,                 \
+               _targetptr_final, _newval_final, _code)                 \
+       __do_rseq(RSEQ_FINISH_MEMCPY, _lock, _rseq_state, _cpu, _result,\
+               NULL, 0,                                                \
+               _dest_memcpy, _src_memcpy, _len_memcpy,                 \
+               _targetptr_final, _newval_final, _code, false)
+
+#define do_rseq_memcpy_release(_lock, _rseq_state, _cpu, _result,      \
+               _dest_memcpy, _src_memcpy, _len_memcpy,                 \
+               _targetptr_final, _newval_final, _code)                 \
+       __do_rseq(RSEQ_FINISH_MEMCPY, _lock, _rseq_state, _cpu, _result,\
+               NULL, 0,                                                \
+               _dest_memcpy, _src_memcpy, _len_memcpy,                 \
+               _targetptr_final, _newval_final, _code, true)
+
+#endif  /* RSEQ_H_ */
This page took 0.067226 seconds and 5 git commands to generate.