mempool: Introduce rseq_mempool_range_init_numa helper

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Change-Id: Icc0ac582fbd21a71f01212d23b86a9ecd5679465

diff --git a/include/rseq/mempool.h b/include/rseq/mempool.h
index b1ec67586aa74e60e8dc7ff48515aa186c08dac3..3bfff1a2e51c418ffc019968c389100db232b28c 100644 (file)
--- a/include/rseq/mempool.h
+++ b/include/rseq/mempool.h
@@ -459,6 +459,17 @@ int rseq_mempool_attr_set_percpu(struct rseq_mempool_attr *attr,
  */
 int rseq_mempool_attr_set_global(struct rseq_mempool_attr *attr, size_t stride);
 
+/*
+ * rseq_mempool_range_init_numa: NUMA initialization helper for memory range.
+ *
+ * Helper which can be used from mempool_attr @init_func to move a CPU
+ * memory range to the NUMA node associated to its topology.
+ *
+ * Returns 0 on success, -1 with errno set by move_pages(2) on error.
+ * Returns -1, errno=ENOSYS if NUMA support is not present.
+ */
+int rseq_mempool_range_init_numa(void *addr, size_t len, int cpu, int numa_flags);
+
 #ifdef __cplusplus
 }
 #endif

diff --git a/src/rseq-mempool.c b/src/rseq-mempool.c
index afde17a513d90480b2a8bf6bfccdc6c0960196de..a7f57d634e8c1620433c3282a60245f84f06605f 100644 (file)
--- a/src/rseq-mempool.c
+++ b/src/rseq-mempool.c
@@ -153,83 +153,68 @@ void rseq_percpu_zero_item(struct rseq_mempool *pool,
 }
 
 #ifdef HAVE_LIBNUMA
-static
-int rseq_mempool_range_init_numa(struct rseq_mempool *pool, struct rseq_mempool_range *range, int numa_flags)
+int rseq_mempool_range_init_numa(void *addr, size_t len, int cpu, int numa_flags)
 {
        unsigned long nr_pages, page_len;
+       int status[MOVE_PAGES_BATCH_SIZE];
+       int nodes[MOVE_PAGES_BATCH_SIZE];
+       void *pages[MOVE_PAGES_BATCH_SIZE];
        long ret;
-       int cpu;
 
-       if (!numa_flags)
-               return 0;
+       if (!numa_flags) {
+               errno = EINVAL;
+               return -1;
+       }
        page_len = rseq_get_page_len();
-       nr_pages = pool->attr.stride >> rseq_get_count_order_ulong(page_len);
-       for (cpu = 0; cpu < pool->attr.max_nr_cpus; cpu++) {
+       nr_pages = len >> rseq_get_count_order_ulong(page_len);
 
-               int status[MOVE_PAGES_BATCH_SIZE];
-               int nodes[MOVE_PAGES_BATCH_SIZE];
-               void *pages[MOVE_PAGES_BATCH_SIZE];
+       nodes[0] = numa_node_of_cpu(cpu);
+       if (nodes[0] < 0)
+               return -1;
 
-               nodes[0] = numa_node_of_cpu(cpu);
-               if (nodes[0] < 0)
-                       continue;
-               for (size_t k = 1; k < RSEQ_ARRAY_SIZE(nodes); ++k) {
-                       nodes[k] = nodes[0];
-               }
+       for (size_t k = 1; k < RSEQ_ARRAY_SIZE(nodes); ++k) {
+               nodes[k] = nodes[0];
+       }
 
-               for (unsigned long page = 0; page < nr_pages;) {
+       for (unsigned long page = 0; page < nr_pages;) {
 
-                       size_t max_k = RSEQ_ARRAY_SIZE(pages);
-                       size_t left = nr_pages - page;
+               size_t max_k = RSEQ_ARRAY_SIZE(pages);
+               size_t left = nr_pages - page;
 
-                       if (left < max_k) {
-                               max_k = left;
-                       }
+               if (left < max_k) {
+                       max_k = left;
+               }
 
-                       for (size_t k = 0; k < max_k; ++k, ++page) {
-                               pages[k] = __rseq_pool_range_percpu_ptr(range, cpu,
-                                       page * page_len, pool->attr.stride);
-                               status[k] = -EPERM;
-                       }
+               for (size_t k = 0; k < max_k; ++k, ++page) {
+                       pages[k] = addr + (page * page_len);
+                       status[k] = -EPERM;
+               }
 
-                       ret = move_pages(0, max_k, pages, nodes, status, numa_flags);
+               ret = move_pages(0, max_k, pages, nodes, status, numa_flags);
 
-                       if (ret < 0)
-                               return ret;
+               if (ret < 0)
+                       return ret;
 
-                       if (ret > 0) {
-                               fprintf(stderr, "%lu pages were not migrated\n", ret);
-                               for (size_t k = 0; k < max_k; ++k) {
-                                       if (status[k] < 0)
-                                               fprintf(stderr,
-                                                       "Error while moving page %p to numa node %d: %u\n",
-                                                       pages[k], nodes[k], -status[k]);
-                               }
+               if (ret > 0) {
+                       fprintf(stderr, "%lu pages were not migrated\n", ret);
+                       for (size_t k = 0; k < max_k; ++k) {
+                               if (status[k] < 0)
+                                       fprintf(stderr,
+                                               "Error while moving page %p to numa node %d: %u\n",
+                                               pages[k], nodes[k], -status[k]);
                        }
                }
        }
        return 0;
 }
-
-int rseq_mempool_init_numa(struct rseq_mempool *pool, int numa_flags)
-{
-       struct rseq_mempool_range *range;
-       int ret;
-
-       if (!numa_flags)
-               return 0;
-       for (range = pool->ranges; range; range = range->next) {
-               ret = rseq_mempool_range_init_numa(pool, range, numa_flags);
-               if (ret)
-                       return ret;
-       }
-       return 0;
-}
 #else
-int rseq_mempool_init_numa(struct rseq_mempool *pool __attribute__((unused)),
+int rseq_mempool_range_init_numa(void *addr __attribute__((unused)),
+               size_t len __attribute__((unused)),
+               int cpu __attribute__((unused)),
                int numa_flags __attribute__((unused)))
 {
-       return 0;
+       errno = ENOSYS;
+       return -1;
 }
 #endif