mempool: Receive mempool type (percpu/global) as attribute

Also introduce a smp.c/h helper to get the number of possible cpus,
copied from libside (MIT).

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Change-Id: I4dfc8fa56ae29e5586eb029377abad62c93de6a9

7 files changed, 499 insertions, 66 deletions

diff --git a/include/rseq/mempool.h b/include/rseq/mempool.h
index 428661e..1813a16 100644
--- a/include/rseq/mempool.h
+++ b/include/rseq/mempool.h
@@ -35,7 +35,7 @@ extern "C" {
  * - rseq_percpu_ptr().
  * - rseq_mempool_percpu_free(),
  */
-#define RSEQ_PERCPU_STRIDE	(1U << 16)	/* stride: 64kB */
+#define RSEQ_MEMPOOL_STRIDE	(1U << 16)	/* stride: 64kB */
 
 /*
  * Tag pointers returned by:
@@ -58,21 +58,20 @@ struct rseq_mempool;
 /*
  * rseq_mempool_create: Create a memory pool.
  *
- * Create a per-cpu memory pool for items of size @item_len (rounded to
- * next power of two). The reserved allocation size is @percpu_stride, and
- * the maximum CPU value expected is (@max_nr_cpus - 1). A
- * @percpu_stride of 0 uses the default RSEQ_PERCPU_STRIDE.
+ * Create a memory pool for items of size @item_len (rounded to
+ * next power of two).
  *
  * The @attr pointer used to specify the pool attributes. If NULL, use a
  * default attribute values. The @attr can be destroyed immediately
  * after rseq_mempool_create() returns. The caller keeps ownership
- * of @attr.
+ * of @attr. Default attributes select a per-cpu mempool type.
  *
  * The argument @pool_name can be used to given a name to the pool for
  * debugging purposes. It can be NULL if no name is given.
  *
  * Returns a pointer to the created percpu pool. Return NULL on error,
  * with errno set accordingly:
+ *
  *   EINVAL: Invalid argument.
  *   ENOMEM: Not enough resources (memory or pool indexes) available to
  *           allocate pool.
@@ -84,8 +83,7 @@ struct rseq_mempool;
  * This API is MT-safe.
  */
 struct rseq_mempool *rseq_mempool_create(const char *pool_name,
-		size_t item_len, size_t percpu_stride, int max_nr_cpus,
-		const struct rseq_mempool_attr *attr);
+		size_t item_len, const struct rseq_mempool_attr *attr);
 
 /*
  * rseq_mempool_destroy: Destroy a per-cpu memory pool.
@@ -177,14 +175,14 @@ void *rseq_mempool_zmalloc(struct rseq_mempool *pool)
  *
  * The @stride optional argument to rseq_percpu_free() is a configurable
  * stride, which must match the stride received by pool creation.
- * If the argument is not present, use the default RSEQ_PERCPU_STRIDE.
+ * If the argument is not present, use the default RSEQ_MEMPOOL_STRIDE.
  *
  * This API is MT-safe.
  */
-void librseq_mempool_percpu_free(void __rseq_percpu *ptr, size_t percpu_stride);
+void librseq_mempool_percpu_free(void __rseq_percpu *ptr, size_t stride);
 
 #define rseq_mempool_percpu_free(_ptr, _stride...)		\
-	librseq_mempool_percpu_free(_ptr, RSEQ_PARAM_SELECT_ARG1(_, ##_stride, RSEQ_PERCPU_STRIDE))
+	librseq_mempool_percpu_free(_ptr, RSEQ_PARAM_SELECT_ARG1(_, ##_stride, RSEQ_MEMPOOL_STRIDE))
 
 /*
  * rseq_free: Free memory from a global pool.
@@ -201,14 +199,14 @@ void librseq_mempool_percpu_free(void __rseq_percpu *ptr, size_t percpu_stride);
  *
  * The @stride optional argument to rseq_free() is a configurable
  * stride, which must match the stride received by pool creation. If
- * the argument is not present, use the default RSEQ_PERCPU_STRIDE.
+ * the argument is not present, use the default RSEQ_MEMPOOL_STRIDE.
  * The stride is needed even for a global pool to know the mapping
  * address range.
  *
  * This API is MT-safe.
  */
 #define rseq_mempool_free(_ptr, _stride...)		\
-	librseq_percpu_free((void __rseq_percpu *) _ptr, RSEQ_PARAM_SELECT_ARG1(_, ##_stride, RSEQ_PERCPU_STRIDE))
+	librseq_percpu_free((void __rseq_percpu *) _ptr, RSEQ_PARAM_SELECT_ARG1(_, ##_stride, RSEQ_MEMPOOL_STRIDE))
 
 /*
  * rseq_percpu_ptr: Offset a per-cpu pointer for a given CPU.
@@ -226,7 +224,7 @@ void librseq_mempool_percpu_free(void __rseq_percpu *ptr, size_t percpu_stride);
  * for the returned pointer, but removes the __rseq_percpu annotation.
  *
  * The macro rseq_percpu_ptr() takes an optional @stride argument. If
- * the argument is not present, use the default RSEQ_PERCPU_STRIDE.
+ * the argument is not present, use the default RSEQ_MEMPOOL_STRIDE.
  * This must match the stride used for pool creation.
  *
  * This API is MT-safe.
@@ -234,7 +232,7 @@ void librseq_mempool_percpu_free(void __rseq_percpu *ptr, size_t percpu_stride);
 #define rseq_percpu_ptr(_ptr, _cpu, _stride...)		\
 	((__typeof__(*(_ptr)) *) ((uintptr_t) (_ptr) +	\
 		((unsigned int) (_cpu) *		\
-			(uintptr_t) RSEQ_PARAM_SELECT_ARG1(_, ##_stride, RSEQ_PERCPU_STRIDE))))
+			(uintptr_t) RSEQ_PARAM_SELECT_ARG1(_, ##_stride, RSEQ_MEMPOOL_STRIDE))))
 
 /*
  * rseq_mempool_set_create: Create a pool set.
@@ -409,6 +407,30 @@ int rseq_mempool_attr_set_mmap(struct rseq_mempool_attr *attr,
  */
 int rseq_mempool_attr_set_robust(struct rseq_mempool_attr *attr);
 
+/*
+ * rseq_mempool_attr_set_percpu: Set pool type as percpu.
+ *
+ * A pool created with this type is a per-cpu memory pool.  The reserved
+ * allocation size is @stride, and the maximum CPU value expected
+ * is (@max_nr_cpus - 1). A @stride of 0 uses the default
+ * RSEQ_MEMPOOL_STRIDE.
+ *
+ * Returns 0 on success, -1 with errno=EINVAL if arguments are invalid.
+ */
+int rseq_mempool_attr_set_percpu(struct rseq_mempool_attr *attr,
+		size_t stride, int max_nr_cpus);
+
+/*
+ * rseq_mempool_attr_set_global: Set pool type as global.
+ *
+ * A pool created with this type is a global memory pool.  The reserved
+ * allocation size is @stride. A @stride of 0 uses the default
+ * RSEQ_MEMPOOL_STRIDE.
+ *
+ * Returns 0 on success, -1 with errno=EINVAL if arguments are invalid.
+ */
+int rseq_mempool_attr_set_global(struct rseq_mempool_attr *attr, size_t stride);
+
 #ifdef __cplusplus
 }
 #endif
diff --git a/src/Makefile.am b/src/Makefile.am
index bb627a1..e313f27 100644
--- a/src/Makefile.am
+++ b/src/Makefile.am
@@ -4,7 +4,7 @@
 lib_LTLIBRARIES = librseq.la
 
 librseq_la_SOURCES = \
-	rseq.c rseq-mempool.c rseq-utils.h
+	rseq.c rseq-mempool.c rseq-utils.h smp.c smp.h
 
 librseq_la_LDFLAGS = -no-undefined -version-info $(RSEQ_LIBRARY_VERSION)
 librseq_la_LIBADD = $(DL_LIBS)
diff --git a/src/rseq-mempool.c b/src/rseq-mempool.c
index 8910dff..cc30d18 100644
--- a/src/rseq-mempool.c
+++ b/src/rseq-mempool.c
@@ -20,6 +20,7 @@
 #endif
 
 #include "rseq-utils.h"
+#include "smp.h"
 
 /*
  * rseq-mempool.c: rseq CPU-Local Storage (CLS) memory allocator.
@@ -63,6 +64,11 @@ struct free_list_node {
 	struct free_list_node *next;
 };
 
+enum mempool_type {
+	MEMPOOL_TYPE_PERCPU = 0,	/* Default */
+	MEMPOOL_TYPE_GLOBAL = 1,
+};
+
 struct rseq_mempool_attr {
 	bool mmap_set;
 	void *(*mmap_func)(void *priv, size_t len);
@@ -70,6 +76,10 @@ struct rseq_mempool_attr {
 	void *mmap_priv;
 
 	bool robust_set;
+
+	enum mempool_type type;
+	size_t stride;
+	int max_nr_cpus;
 };
 
 struct rseq_mempool_range;
@@ -89,9 +99,7 @@ struct rseq_mempool {
 	struct rseq_mempool_range *ranges;
 
 	size_t item_len;
-	size_t percpu_stride;
 	int item_order;
-	int max_nr_cpus;
 
 	/*
 	 * The free list chains freed items on the CPU 0 address range.
@@ -133,9 +141,9 @@ void rseq_percpu_zero_item(struct rseq_mempool *pool, uintptr_t item_offset)
 {
 	int i;
 
-	for (i = 0; i < pool->max_nr_cpus; i++) {
+	for (i = 0; i < pool->attr.max_nr_cpus; i++) {
 		char *p = __rseq_pool_percpu_ptr(pool, i,
-				item_offset, pool->percpu_stride);
+				item_offset, pool->attr.stride);
 		memset(p, 0, pool->item_len);
 	}
 }
@@ -153,8 +161,8 @@ int rseq_mempool_range_init_numa(struct rseq_mempool *pool, struct rseq_mempool_
 	if (!numa_flags)
 		return 0;
 	page_len = rseq_get_page_len();
-	nr_pages = pool->percpu_stride >> rseq_get_count_order_ulong(page_len);
-	for (cpu = 0; cpu < pool->max_nr_cpus; cpu++) {
+	nr_pages = pool->attr.stride >> rseq_get_count_order_ulong(page_len);
+	for (cpu = 0; cpu < pool->attr.max_nr_cpus; cpu++) {
 
 		int status[MOVE_PAGES_BATCH_SIZE];
 		int nodes[MOVE_PAGES_BATCH_SIZE];
@@ -243,7 +251,7 @@ int create_alloc_bitmap(struct rseq_mempool *pool, struct rseq_mempool_range *ra
 {
 	size_t count;
 
-	count = ((pool->percpu_stride >> pool->item_order) + BIT_PER_ULONG - 1) / BIT_PER_ULONG;
+	count = ((pool->attr.stride >> pool->item_order) + BIT_PER_ULONG - 1) / BIT_PER_ULONG;
 
 	/*
 	 * Not being able to create the validation bitmap is an error
@@ -285,8 +293,8 @@ void check_free_list(const struct rseq_mempool *pool)
 		return;
 
 	for (range = pool->ranges; range; range = range->next) {
-		total_item += pool->percpu_stride >> pool->item_order;
-		total_never_allocated += (pool->percpu_stride - range->next_unused) >> pool->item_order;
+		total_item += pool->attr.stride >> pool->item_order;
+		total_never_allocated += (pool->attr.stride - range->next_unused) >> pool->item_order;
 	}
 	max_list_traversal = total_item - total_never_allocated;
 
@@ -335,7 +343,7 @@ void destroy_alloc_bitmap(struct rseq_mempool *pool, struct rseq_mempool_range *
 	if (!bitmap)
 		return;
 
-	count = ((pool->percpu_stride >> pool->item_order) + BIT_PER_ULONG - 1) / BIT_PER_ULONG;
+	count = ((pool->attr.stride >> pool->item_order) + BIT_PER_ULONG - 1) / BIT_PER_ULONG;
 
 	/* Assert that all items in the pool were freed. */
 	for (size_t k = 0; k < count; ++k)
@@ -357,7 +365,7 @@ int rseq_mempool_range_destroy(struct rseq_mempool *pool,
 	destroy_alloc_bitmap(pool, range);
 	/* range is a header located one page before the aligned mapping. */
 	return pool->attr.munmap_func(pool->attr.mmap_priv, range->header,
-			(pool->percpu_stride * pool->max_nr_cpus) + rseq_get_page_len());
+			(pool->attr.stride * pool->attr.max_nr_cpus) + rseq_get_page_len());
 }
 
 /*
@@ -453,8 +461,8 @@ struct rseq_mempool_range *rseq_mempool_range_create(struct rseq_mempool *pool)
 	page_size = rseq_get_page_len();
 
 	base = aligned_mmap_anonymous(pool, page_size,
-			pool->percpu_stride * pool->max_nr_cpus,
-			pool->percpu_stride,
+			pool->attr.stride * pool->attr.max_nr_cpus,
+			pool->attr.stride,
 			&header, page_size);
 	if (!base)
 		return NULL;
@@ -496,8 +504,7 @@ end:
 }
 
 struct rseq_mempool *rseq_mempool_create(const char *pool_name,
-		size_t item_len, size_t percpu_stride, int max_nr_cpus,
-		const struct rseq_mempool_attr *_attr)
+		size_t item_len, const struct rseq_mempool_attr *_attr)
 {
 	struct rseq_mempool *pool;
 	struct rseq_mempool_attr attr = {};
@@ -515,16 +522,6 @@ struct rseq_mempool *rseq_mempool_create(const char *pool_name,
 	}
 	item_len = 1UL << order;
 
-	if (!percpu_stride)
-		percpu_stride = RSEQ_PERCPU_STRIDE;	/* Use default */
-
-	if (max_nr_cpus < 0 || item_len > percpu_stride ||
-			percpu_stride < (size_t) rseq_get_page_len() ||
-			!is_pow2(percpu_stride)) {
-		errno = EINVAL;
-		return NULL;
-	}
-
 	if (_attr)
 		memcpy(&attr, _attr, sizeof(attr));
 	if (!attr.mmap_set) {
@@ -533,14 +530,38 @@ struct rseq_mempool *rseq_mempool_create(const char *pool_name,
 		attr.mmap_priv = NULL;
 	}
 
+	switch (attr.type) {
+	case MEMPOOL_TYPE_PERCPU:
+		if (attr.max_nr_cpus < 0) {
+			errno = EINVAL;
+			return NULL;
+		}
+		if (attr.max_nr_cpus == 0) {
+			/* Auto-detect */
+			attr.max_nr_cpus = get_possible_cpus_array_len();
+			if (attr.max_nr_cpus == 0) {
+				errno = EINVAL;
+				return NULL;
+			}
+		}
+		break;
+	case MEMPOOL_TYPE_GLOBAL:
+		break;
+	}
+	if (!attr.stride)
+		attr.stride = RSEQ_MEMPOOL_STRIDE;	/* Use default */
+	if (item_len > attr.stride || attr.stride < (size_t) rseq_get_page_len() ||
+			!is_pow2(attr.stride)) {
+		errno = EINVAL;
+		return NULL;
+	}
+
 	pool = calloc(1, sizeof(struct rseq_mempool));
 	if (!pool)
 		return NULL;
 
 	memcpy(&pool->attr, &attr, sizeof(attr));
 	pthread_mutex_init(&pool->lock, NULL);
-	pool->percpu_stride = percpu_stride;
-	pool->max_nr_cpus = max_nr_cpus;
 	pool->item_len = item_len;
 	pool->item_order = order;
 
@@ -603,7 +624,7 @@ void __rseq_percpu *__rseq_percpu_malloc(struct rseq_mempool *pool, bool zeroed)
 		addr = (void __rseq_percpu *) (pool->ranges->base + item_offset);
 		goto end;
 	}
-	if (pool->ranges->next_unused + pool->item_len > pool->percpu_stride) {
+	if (pool->ranges->next_unused + pool->item_len > pool->attr.stride) {
 		errno = ENOMEM;
 		addr = NULL;
 		goto end;
@@ -655,13 +676,13 @@ void clear_alloc_slot(struct rseq_mempool *pool, size_t item_offset)
 	bitmap[k] &= ~mask;
 }
 
-void librseq_mempool_percpu_free(void __rseq_percpu *_ptr, size_t percpu_stride)
+void librseq_mempool_percpu_free(void __rseq_percpu *_ptr, size_t stride)
 {
 	uintptr_t ptr = (uintptr_t) _ptr;
-	void *range_base = (void *) (ptr & (~(percpu_stride - 1)));
+	void *range_base = (void *) (ptr & (~(stride - 1)));
 	struct rseq_mempool_range *range = (struct rseq_mempool_range *) (range_base - RANGE_HEADER_OFFSET);
 	struct rseq_mempool *pool = range->pool;
-	uintptr_t item_offset = ptr & (percpu_stride - 1);
+	uintptr_t item_offset = ptr & (stride - 1);
 	struct free_list_node *head, *item;
 
 	pthread_mutex_lock(&pool->lock);
@@ -810,3 +831,29 @@ int rseq_mempool_attr_set_robust(struct rseq_mempool_attr *attr)
 	attr->robust_set = true;
 	return 0;
 }
+
+int rseq_mempool_attr_set_percpu(struct rseq_mempool_attr *attr,
+		size_t stride, int max_nr_cpus)
+{
+	if (!attr) {
+		errno = EINVAL;
+		return -1;
+	}
+	attr->type = MEMPOOL_TYPE_PERCPU;
+	attr->stride = stride;
+	attr->max_nr_cpus = max_nr_cpus;
+	return 0;
+}
+
+int rseq_mempool_attr_set_global(struct rseq_mempool_attr *attr,
+		size_t stride)
+{
+	if (!attr) {
+		errno = EINVAL;
+		return -1;
+	}
+	attr->type = MEMPOOL_TYPE_GLOBAL;
+	attr->stride = stride;
+	attr->max_nr_cpus = 1;
+	return 0;
+}
diff --git a/src/smp.c b/src/smp.c
new file mode 100644
index 0000000..9788e93
--- /dev/null
+++ b/src/smp.c
@@ -0,0 +1,278 @@
+// SPDX-License-Identifier: MIT
+/*
+ * Copyright (C) 2011-2012 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
+ * Copyright (C) 2019 Michael Jeanson <mjeanson@efficios.com>
+ */
+
+#define _LGPL_SOURCE
+#include <assert.h>
+#include <ctype.h>
+#include <dirent.h>
+#include <errno.h>
+#include <fcntl.h>
+#include <limits.h>
+#include <unistd.h>
+#include <pthread.h>
+#include <stdlib.h>
+#include <string.h>
+#include <stdio.h>
+#include <sys/types.h>
+#include <side/macros.h>
+
+#include "smp.h"
+
+#define __max(a,b) ((a)>(b)?(a):(b))
+
+#define SIDE_CPUMASK_SIZE	4096
+
+static int possible_cpus_array_len_cache;
+
+static
+int _get_max_cpuid_from_sysfs(const char *path)
+{
+	long max_cpuid = -1;
+
+	DIR *cpudir;
+	struct dirent *entry;
+
+	assert(path);
+
+	cpudir = opendir(path);
+	if (cpudir == NULL)
+		goto end;
+
+	/*
+	 * Iterate on all directories named "cpu" followed by an integer.
+	 */
+	while ((entry = readdir(cpudir))) {
+		if (entry->d_type == DT_DIR &&
+			strncmp(entry->d_name, "cpu", 3) == 0) {
+
+			char *endptr;
+			long cpu_id;
+
+			cpu_id = strtol(entry->d_name + 3, &endptr, 10);
+			if ((cpu_id < LONG_MAX) && (endptr != entry->d_name + 3)
+					&& (*endptr == '\0')) {
+				if (cpu_id > max_cpuid)
+					max_cpuid = cpu_id;
+			}
+		}
+	}
+
+	if (closedir(cpudir))
+		perror("closedir");
+
+	/*
+	 * If the max CPU id is out of bound, set it to -1 so it results in a
+	 * CPU num of 0.
+	 */
+	if (max_cpuid < 0 || max_cpuid > INT_MAX)
+		max_cpuid = -1;
+
+end:
+	return max_cpuid;
+}
+
+/*
+ * Get the highest CPU id from sysfs.
+ *
+ * Iterate on all the folders in "/sys/devices/system/cpu" that start with
+ * "cpu" followed by an integer, keep the highest CPU id encountered during
+ * this iteration and add 1 to get a number of CPUs.
+ *
+ * Returns the highest CPU id, or -1 on error.
+ */
+static
+int get_max_cpuid_from_sysfs(void)
+{
+	return _get_max_cpuid_from_sysfs("/sys/devices/system/cpu");
+}
+
+/*
+ * As a fallback to parsing the CPU mask in "/sys/devices/system/cpu/possible",
+ * iterate on all the folders in "/sys/devices/system/cpu" that start with
+ * "cpu" followed by an integer, keep the highest CPU id encountered during
+ * this iteration and add 1 to get a number of CPUs.
+ *
+ * Then get the value from sysconf(_SC_NPROCESSORS_CONF) as a fallback and
+ * return the highest one.
+ *
+ * On Linux, using the value from sysconf can be unreliable since the way it
+ * counts CPUs varies between C libraries and even between versions of the same
+ * library. If we used it directly, getcpu() could return a value greater than
+ * this sysconf, in which case the arrays indexed by processor would overflow.
+ *
+ * As another example, the MUSL libc implementation of the _SC_NPROCESSORS_CONF
+ * sysconf does not return the number of configured CPUs in the system but
+ * relies on the cpu affinity mask of the current task.
+ *
+ * Returns 0 or less on error.
+ */
+static
+int get_num_possible_cpus_fallback(void)
+{
+	/*
+	 * Get the sysconf value as a last resort. Keep the highest number.
+	 */
+	return __max(sysconf(_SC_NPROCESSORS_CONF), get_max_cpuid_from_sysfs() + 1);
+}
+
+/*
+ * Get a CPU mask string from sysfs.
+ *
+ * buf: the buffer where the mask will be read.
+ * max_bytes: the maximum number of bytes to write in the buffer.
+ * path: file path to read the mask from.
+ *
+ * Returns the number of bytes read or -1 on error.
+ */
+static
+int get_cpu_mask_from_sysfs(char *buf, size_t max_bytes, const char *path)
+{
+	ssize_t bytes_read = 0;
+	size_t total_bytes_read = 0;
+	int fd = -1, ret = -1;
+
+	assert(path);
+
+	if (buf == NULL)
+		goto end;
+
+	fd = open(path, O_RDONLY);
+	if (fd < 0)
+		goto end;
+
+	do {
+		bytes_read = read(fd, buf + total_bytes_read,
+				max_bytes - total_bytes_read);
+
+		if (bytes_read < 0) {
+			if (errno == EINTR) {
+				continue;	/* retry operation */
+			} else {
+				goto end;
+			}
+		}
+
+		total_bytes_read += bytes_read;
+		assert(total_bytes_read <= max_bytes);
+	} while (max_bytes > total_bytes_read && bytes_read > 0);
+
+	/*
+	 * Make sure the mask read is a null terminated string.
+	 */
+	if (total_bytes_read < max_bytes)
+		buf[total_bytes_read] = '\0';
+	else
+		buf[max_bytes - 1] = '\0';
+
+	if (total_bytes_read > INT_MAX)
+		goto end;
+	ret = (int) total_bytes_read;
+end:
+	if (fd >= 0 && close(fd) < 0)
+		perror("close");
+	return ret;
+}
+
+/*
+ * Get the CPU possible mask string from sysfs.
+ *
+ * buf: the buffer where the mask will be read.
+ * max_bytes: the maximum number of bytes to write in the buffer.
+ *
+ * Returns the number of bytes read or -1 on error.
+ */
+static
+int get_possible_cpu_mask_from_sysfs(char *buf, size_t max_bytes)
+{
+	return get_cpu_mask_from_sysfs(buf, max_bytes,
+			"/sys/devices/system/cpu/possible");
+}
+
+/*
+ * Get the highest CPU id from a CPU mask.
+ *
+ * pmask: the mask to parse.
+ * len: the len of the mask excluding '\0'.
+ *
+ * Returns the highest CPU id from the mask or -1 on error.
+ */
+static
+int get_max_cpuid_from_mask(const char *pmask, size_t len)
+{
+	ssize_t i;
+	unsigned long cpu_index;
+	char *endptr;
+
+	/* We need at least one char to read */
+	if (len < 1)
+		goto error;
+
+	/* Start from the end to read the last CPU index. */
+	for (i = len - 1; i > 0; i--) {
+		/* Break when we hit the first separator. */
+		if ((pmask[i] == ',') || (pmask[i] == '-')) {
+			i++;
+			break;
+		}
+	}
+
+	cpu_index = strtoul(&pmask[i], &endptr, 10);
+
+	if ((&pmask[i] != endptr) && (cpu_index < INT_MAX))
+		return (int) cpu_index;
+
+error:
+	return -1;
+}
+
+static void update_possible_cpus_array_len_cache(void)
+{
+	char buf[SIDE_CPUMASK_SIZE];
+	int ret;
+
+	/* Get the possible cpu mask from sysfs, fallback to sysconf. */
+	ret = get_possible_cpu_mask_from_sysfs((char *) &buf, SIDE_CPUMASK_SIZE);
+	if (ret <= 0)
+		goto fallback;
+
+	/* Parse the possible cpu mask, on failure fallback to sysconf. */
+	ret = get_max_cpuid_from_mask((char *) &buf, ret);
+	if (ret >= 0) {
+		/* Add 1 to convert from max cpuid to an array len. */
+		ret++;
+		goto end;
+	}
+
+fallback:
+	/* Fallback to sysconf. */
+	ret = get_num_possible_cpus_fallback();
+
+end:
+	/* If all methods failed, don't store the value. */
+	if (ret < 1)
+		return;
+
+	possible_cpus_array_len_cache = ret;
+}
+
+/*
+ * Returns the length of an array that could contain a per-CPU element for each
+ * possible CPU id for the lifetime of the process.
+ *
+ * We currently assume CPU ids are contiguous up the maximum CPU id.
+ *
+ * If the cache is not yet initialized, get the value from
+ * "/sys/devices/system/cpu/possible" or fallback to sysconf and cache it.
+ *
+ * If all methods fail, don't populate the cache and return 0.
+ */
+int get_possible_cpus_array_len(void)
+{
+	if (side_unlikely(!possible_cpus_array_len_cache))
+		update_possible_cpus_array_len_cache();
+
+	return possible_cpus_array_len_cache;
+}
diff --git a/src/smp.h b/src/smp.h
new file mode 100644
index 0000000..6f9240f
--- /dev/null
+++ b/src/smp.h
@@ -0,0 +1,12 @@
+// SPDX-License-Identifier: MIT
+/*
+ * Copyright (C) 2011-2012 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
+ * Copyright (C) 2019 Michael Jeanson <mjeanson@efficios.com>
+ */
+
+#ifndef _RSEQ_SMP_H
+#define _RSEQ_SMP_H
+
+int get_possible_cpus_array_len(void) __attribute__((visibility("hidden")));
+
+#endif /* _RSEQ_SMP_H */
diff --git a/tests/mempool_test.c b/tests/mempool_test.c
index 8911fde..6e19246 100644
--- a/tests/mempool_test.c
+++ b/tests/mempool_test.c
@@ -44,10 +44,10 @@ static void test_mempool_fill(size_t stride)
 	ok(attr, "Create pool attribute");
 	ret = rseq_mempool_attr_set_robust(attr);
 	ok(ret == 0, "Setting mempool robust attribute");
-
+	ret = rseq_mempool_attr_set_percpu(attr, stride, CPU_SETSIZE);
+	ok(ret == 0, "Setting mempool percpu type");
 	mempool = rseq_mempool_create("test_data",
-			sizeof(struct test_data),
-			stride, CPU_SETSIZE, attr);
+			sizeof(struct test_data), attr);
 	ok(mempool, "Create mempool of size %zu", stride);
 	rseq_mempool_attr_destroy(attr);
 
@@ -167,14 +167,19 @@ static void run_robust_tests(void)
 {
 	struct rseq_mempool_attr *attr;
 	struct rseq_mempool *pool;
+	int ret;
 
 	attr = rseq_mempool_attr_create();
+	ok(attr, "Create mempool attributes");
+
+	ret = rseq_mempool_attr_set_robust(attr);
+	ok(ret == 0, "Setting mempool robust attribute");
 
-	rseq_mempool_attr_set_robust(attr);
+	ret = rseq_mempool_attr_set_percpu(attr, RSEQ_MEMPOOL_STRIDE, 1);
+	ok(ret == 0, "Setting mempool percpu type");
 
 	pool = rseq_mempool_create("mempool-robust",
-				sizeof(void*), RSEQ_PERCPU_STRIDE, 1,
-				attr);
+				sizeof(void*), attr);
 
 	rseq_mempool_attr_destroy(attr);
 
diff --git a/tests/param_test.c b/tests/param_test.c
index ac4f12c..62614c3 100644
--- a/tests/param_test.c
+++ b/tests/param_test.c
@@ -496,14 +496,25 @@ static void test_percpu_spinlock(void)
 	struct spinlock_test_data __rseq_percpu *data;
 	struct spinlock_thread_test_data thread_data[num_threads];
 	struct rseq_mempool *mempool;
+	struct rseq_mempool_attr *attr;
 
+	attr = rseq_mempool_attr_create();
+	if (!attr) {
+		perror("rseq_mempool_attr_create");
+		abort();
+	}
+	ret = rseq_mempool_attr_set_percpu(attr, RSEQ_MEMPOOL_STRIDE, CPU_SETSIZE);
+	if (ret) {
+		perror("rseq_mempool_attr_set_percpu");
+		abort();
+	}
 	mempool = rseq_mempool_create("spinlock_test_data",
-			sizeof(struct spinlock_test_data),
-			0, CPU_SETSIZE, NULL);
+			sizeof(struct spinlock_test_data), attr);
 	if (!mempool) {
 		perror("rseq_mempool_create");
 		abort();
 	}
+	rseq_mempool_attr_destroy(attr);
 	data = (struct spinlock_test_data __rseq_percpu *)rseq_mempool_percpu_zmalloc(mempool);
 	if (!data) {
 		perror("rseq_mempool_percpu_zmalloc");
@@ -592,14 +603,25 @@ static void test_percpu_inc(void)
 	struct inc_test_data __rseq_percpu *data;
 	struct inc_thread_test_data thread_data[num_threads];
 	struct rseq_mempool *mempool;
+	struct rseq_mempool_attr *attr;
 
+	attr = rseq_mempool_attr_create();
+	if (!attr) {
+		perror("rseq_mempool_attr_create");
+		abort();
+	}
+	ret = rseq_mempool_attr_set_percpu(attr, RSEQ_MEMPOOL_STRIDE, CPU_SETSIZE);
+	if (ret) {
+		perror("rseq_mempool_attr_set_percpu");
+		abort();
+	}
 	mempool = rseq_mempool_create("inc_test_data",
-			sizeof(struct inc_test_data),
-			0, CPU_SETSIZE, NULL);
+			sizeof(struct inc_test_data), attr);
 	if (!mempool) {
 		perror("rseq_mempool_create");
 		abort();
 	}
+	rseq_mempool_attr_destroy(attr);
 	data = (struct inc_test_data __rseq_percpu *)rseq_mempool_percpu_zmalloc(mempool);
 	if (!data) {
 		perror("rseq_mempool_percpu_zmalloc");
@@ -766,13 +788,25 @@ static void test_percpu_list(void)
 	pthread_t test_threads[num_threads];
 	cpu_set_t allowed_cpus;
 	struct rseq_mempool *mempool;
+	struct rseq_mempool_attr *attr;
 
-	mempool = rseq_mempool_create("percpu_list", sizeof(struct percpu_list),
-			0, CPU_SETSIZE, NULL);
+	attr = rseq_mempool_attr_create();
+	if (!attr) {
+		perror("rseq_mempool_attr_create");
+		abort();
+	}
+	ret = rseq_mempool_attr_set_percpu(attr, RSEQ_MEMPOOL_STRIDE, CPU_SETSIZE);
+	if (ret) {
+		perror("rseq_mempool_attr_set_percpu");
+		abort();
+	}
+	mempool = rseq_mempool_create("percpu_list",
+			sizeof(struct percpu_list), attr);
 	if (!mempool) {
 		perror("rseq_mempool_create");
 		abort();
 	}
+	rseq_mempool_attr_destroy(attr);
 	list = (struct percpu_list __rseq_percpu *)rseq_mempool_percpu_zmalloc(mempool);
 	if (!list) {
 		perror("rseq_mempool_percpu_zmalloc");
@@ -977,13 +1011,25 @@ static void test_percpu_buffer(void)
 	pthread_t test_threads[num_threads];
 	cpu_set_t allowed_cpus;
 	struct rseq_mempool *mempool;
+	struct rseq_mempool_attr *attr;
 
-	mempool = rseq_mempool_create("percpu_buffer", sizeof(struct percpu_buffer),
-			0, CPU_SETSIZE, NULL);
+	attr = rseq_mempool_attr_create();
+	if (!attr) {
+		perror("rseq_mempool_attr_create");
+		abort();
+	}
+	ret = rseq_mempool_attr_set_percpu(attr, RSEQ_MEMPOOL_STRIDE, CPU_SETSIZE);
+	if (ret) {
+		perror("rseq_mempool_attr_set_percpu");
+		abort();
+	}
+	mempool = rseq_mempool_create("percpu_buffer",
+			sizeof(struct percpu_buffer), attr);
 	if (!mempool) {
 		perror("rseq_mempool_create");
 		abort();
 	}
+	rseq_mempool_attr_destroy(attr);
 	buffer = (struct percpu_buffer __rseq_percpu *)rseq_mempool_percpu_zmalloc(mempool);
 	if (!buffer) {
 		perror("rseq_mempool_percpu_zmalloc");
@@ -1217,14 +1263,25 @@ static void test_percpu_memcpy_buffer(void)
 	pthread_t test_threads[num_threads];
 	cpu_set_t allowed_cpus;
 	struct rseq_mempool *mempool;
+	struct rseq_mempool_attr *attr;
 
+	attr = rseq_mempool_attr_create();
+	if (!attr) {
+		perror("rseq_mempool_attr_create");
+		abort();
+	}
+	ret = rseq_mempool_attr_set_percpu(attr, RSEQ_MEMPOOL_STRIDE, CPU_SETSIZE);
+	if (ret) {
+		perror("rseq_mempool_attr_set_percpu");
+		abort();
+	}
 	mempool = rseq_mempool_create("percpu_memcpy_buffer",
-			sizeof(struct percpu_memcpy_buffer),
-			0, CPU_SETSIZE, NULL);
+			sizeof(struct percpu_memcpy_buffer), attr);
 	if (!mempool) {
 		perror("rseq_mempool_create");
 		abort();
 	}
+	rseq_mempool_attr_destroy(attr);
 	buffer = (struct percpu_memcpy_buffer __rseq_percpu *)rseq_mempool_percpu_zmalloc(mempool);
 	if (!buffer) {
 		perror("rseq_mempool_percpu_zmalloc");
@@ -1390,7 +1447,7 @@ void *test_membarrier_worker_thread(void *arg)
 
 			ret = rseq_load_add_load_load_add_store__ptr(RSEQ_MO_RELAXED, RSEQ_PERCPU,
 				(intptr_t *) &args->percpu_list_ptr,
-				(RSEQ_PERCPU_STRIDE * cpu) + offsetof(struct percpu_list, head),
+				(RSEQ_MEMPOOL_STRIDE * cpu) + offsetof(struct percpu_list, head),
 				1, cpu);
 		} while (rseq_unlikely(ret));
 	}
@@ -1463,13 +1520,25 @@ void *test_membarrier_manager_thread(void *arg)
 	struct rseq_mempool *mempool;
 	int ret;
 	long long total_count = 0;
+	struct rseq_mempool_attr *attr;
 
-	mempool = rseq_mempool_create("percpu_list", sizeof(struct percpu_list),
-			0, CPU_SETSIZE, NULL);
+	attr = rseq_mempool_attr_create();
+	if (!attr) {
+		perror("rseq_mempool_attr_create");
+		abort();
+	}
+	ret = rseq_mempool_attr_set_percpu(attr, RSEQ_MEMPOOL_STRIDE, CPU_SETSIZE);
+	if (ret) {
+		perror("rseq_mempool_attr_set_percpu");
+		abort();
+	}
+	mempool = rseq_mempool_create("percpu_list",
+			sizeof(struct percpu_list), attr);
 	if (!mempool) {
 		perror("rseq_mempool_create");
 		abort();
 	}
+	rseq_mempool_attr_destroy(attr);
 	args->mempool = mempool;
 
 	if (rseq_register_current_thread()) {