File 03-Resolve-max-count-of-nodes-at-runtime.patch of Package numatop
From 4740ef47cf753420244111df4878cd7128380f17 Mon Sep 17 00:00:00 2001
From: Sandipan Das <sandipan.das@amd.com>
Date: Fri, 27 Jun 2025 15:24:42 +0530
Subject: [PATCH] common: Resolve max count of nodes at runtime
Replace statically defined NNODES_MAX with a variable which is set at
runtime. Use libnuma helpers to determine the OS-defined limits for the
maximum possible number of NUMA nodes in the the system.
Signed-off-by: Sandipan Das <sandipan.das@amd.com>
---
common/include/os/node.h | 2 +-
common/include/types.h | 5 ++--
common/numatop.c | 2 +-
common/os/node.c | 52 +++++++++++++++++++++++++++-------------
common/os/os_perf.c | 4 ++--
common/os/os_util.c | 2 +-
common/win.c | 15 ++++++++----
7 files changed, 54 insertions(+), 28 deletions(-)
diff --git a/common/include/os/node.h b/common/include/os/node.h
index da1ba7c..2c21556 100644
--- a/common/include/os/node.h
+++ b/common/include/os/node.h
@@ -101,7 +101,7 @@ typedef struct _node {
typedef struct _node_group {
pthread_mutex_t mutex;
- node_t nodes[NNODES_MAX];
+ node_t *nodes;
int nnodes;
int cpuid_max;
int intval_ms;
diff --git a/common/include/types.h b/common/include/types.h
index 05b411a..e688225 100644
--- a/common/include/types.h
+++ b/common/include/types.h
@@ -115,13 +115,14 @@ typedef enum {
#define UI_COUNT_NUM 5
-#define NNODES_MAX 64
#define NCPUS_NODE_MAX 256
-#define NCPUS_MAX (NNODES_MAX * NCPUS_NODE_MAX)
+#define NCPUS_MAX (nnodes_max * NCPUS_NODE_MAX)
#define NPROCS_NAX 4096
#define LL_THRESH 128
#define LL_PERIOD 1000
+extern int nnodes_max;
+
typedef struct _count_value {
uint64_t counts[PERF_COUNT_NUM];
} count_value_t;
diff --git a/common/numatop.c b/common/numatop.c
index 122c187..c5c0580 100644
--- a/common/numatop.c
+++ b/common/numatop.c
@@ -236,7 +236,7 @@ main(int argc, char *argv[])
if (node_group_init() != 0) {
stderr_print("The node/cpu number is out of range, \n"
"numatop supports up to %d nodes and %d CPUs\n",
- NNODES_MAX, NCPUS_MAX);
+ nnodes_max, NCPUS_MAX);
goto L_EXIT5;
}
diff --git a/common/os/node.c b/common/os/node.c
index ab0c90d..c2ca7a7 100644
--- a/common/os/node.c
+++ b/common/os/node.c
@@ -36,6 +36,7 @@
#include <pthread.h>
#include <string.h>
#include <assert.h>
+#include <numa.h>
#include "../include/types.h"
#include "../include/util.h"
#include "../include/ui_perf_map.h"
@@ -46,6 +47,8 @@
static node_group_t s_node_group;
int g_ncpus;
+int nnodes_max;
+
static void
node_init(node_t *node, int nid, boolean_t hotadd)
{
@@ -79,13 +82,22 @@ node_group_init(void)
int i;
node_t *node;
+ if (numa_available() < 0)
+ return (-1);
+
+ nnodes_max = numa_num_possible_nodes();
+
(void) memset(&s_node_group, 0, sizeof (node_group_t));
if (pthread_mutex_init(&s_node_group.mutex, NULL) != 0) {
return (-1);
}
+ if ((s_node_group.nodes = zalloc(nnodes_max * sizeof(node_t))) == NULL) {
+ return (-1);
+ }
+
s_node_group.inited = B_TRUE;
- for (i = 0; i < NNODES_MAX; i++) {
+ for (i = 0; i < nnodes_max; i++) {
node = node_get(i);
node_init(node, INVALID_NID, B_FALSE);
}
@@ -102,12 +114,13 @@ node_group_reset(void)
node_t *node;
int i;
- for (i = 0; i < NNODES_MAX; i++) {
+ for (i = 0; i < nnodes_max; i++) {
node = node_get(i);
node_fini(node);
}
s_node_group.nnodes = 0;
+ free(s_node_group.nodes);
}
/*
@@ -161,7 +174,7 @@ cpu_refresh(boolean_t init)
int cpu_arr[NCPUS_NODE_MAX];
node_t *node;
- for (i = 0; i < NNODES_MAX; i++) {
+ for (i = 0; i < nnodes_max; i++) {
node = node_get(i);
if (NODE_VALID(node)) {
if (!os_sysfs_cpu_enum(node->nid, cpu_arr, NCPUS_NODE_MAX, &num)) {
@@ -199,7 +212,7 @@ meminfo_refresh(void)
int i;
node_t *node;
- for (i = 0; i < NNODES_MAX; i++) {
+ for (i = 0; i < nnodes_max; i++) {
node = node_get(i);
if (NODE_VALID(node)) {
if (!os_sysfs_meminfo(node->nid, &node->meminfo)) {
@@ -220,19 +233,23 @@ meminfo_refresh(void)
int
node_group_refresh(boolean_t init)
{
- int node_arr[NNODES_MAX];
- int num, i, j, ret = -1;
+ int *node_arr, num, i, j, ret = -1;
node_t *node;
node_group_lock();
- if (!os_sysfs_node_enum(node_arr, NNODES_MAX, &num)) {
+
+ if ((node_arr = zalloc(nnodes_max * sizeof(int))) == NULL) {
+ goto L_EXIT;
+ }
+
+ if (!os_sysfs_node_enum(node_arr, nnodes_max, &num)) {
goto L_EXIT;
}
- if (num < 0 || num > NNODES_MAX) {
+ if (num < 0 || num > nnodes_max) {
goto L_EXIT;
}
- for (i = 0; i < NNODES_MAX; i++) {
+ for (i = 0; i < nnodes_max; i++) {
node = node_get(i);
if (NODE_VALID(node)) {
if ((j = nid_find(node->nid, node_arr, num)) == -1) {
@@ -244,7 +261,7 @@ node_group_refresh(boolean_t init)
}
}
- for (i = 0; i < NNODES_MAX; i++) {
+ for (i = 0; i < nnodes_max; i++) {
node = node_get(i);
if (!NODE_VALID(node)) {
if ((j = nid_find(i, node_arr, num)) >= 0) {
@@ -272,6 +289,7 @@ node_group_refresh(boolean_t init)
ret = 0;
L_EXIT:
+ free(node_arr);
node_group_unlock();
return (ret);
}
@@ -313,7 +331,7 @@ node_by_cpu(int cpuid)
return (NULL);
}
- for (i = 0; i < NNODES_MAX; i++) {
+ for (i = 0; i < nnodes_max; i++) {
node = node_get(i);
if (!NODE_VALID(node)) {
continue;
@@ -386,7 +404,7 @@ node_cpu_traverse(pfn_perf_cpu_op_t func, void *arg, boolean_t err_ret,
perf_cpu_t *cpu;
int i, j, ret;
- for (i = 0; i < NNODES_MAX; i++) {
+ for (i = 0; i < nnodes_max; i++) {
node = node_get(i);
if (!NODE_VALID(node)) {
continue;
@@ -461,7 +479,7 @@ node_countval_sum(count_value_t *countval_arr, int nid,
return (countval_sum(countval_arr, nid, ui_count_id));
}
- for (i = 0; i < NNODES_MAX; i++) {
+ for (i = 0; i < nnodes_max; i++) {
value += countval_sum(countval_arr, i, ui_count_id);
}
@@ -486,7 +504,7 @@ node_profiling_clear(void)
node_t *node;
int i;
- for (i = 0; i < NNODES_MAX; i++) {
+ for (i = 0; i < nnodes_max; i++) {
node = node_get(i);
(void) memset(&node->countval, 0, sizeof (count_value_t));
}
@@ -498,7 +516,7 @@ node_valid_get(int node_idx)
int i, nvalid = 0;
node_t *node;
- for (i = 0; i < NNODES_MAX; i++) {
+ for (i = 0; i < nnodes_max; i++) {
node = node_get(i);
if (NODE_VALID(node)) {
if (node_idx == nvalid) {
@@ -537,7 +555,7 @@ node_qpi_init(void)
node_group_lock();
- for (i = 0; i < NNODES_MAX; i++) {
+ for (i = 0; i < nnodes_max; i++) {
node = node_get(i);
if (NODE_VALID(node) && (qpi_num > 0)) {
memcpy(node->qpi.qpi_info, qpi_tmp,
@@ -566,7 +584,7 @@ node_imc_init(void)
node_group_lock();
- for (i = 0; i < NNODES_MAX; i++) {
+ for (i = 0; i < nnodes_max; i++) {
node = node_get(i);
if (NODE_VALID(node) && (imc_num > 0)) {
memcpy(node->imc.imc_info, imc_tmp,
diff --git a/common/os/os_perf.c b/common/os/os_perf.c
index d8d634f..9ea93ea 100644
--- a/common/os/os_perf.c
+++ b/common/os/os_perf.c
@@ -1025,7 +1025,7 @@ uncore_stop_all(void)
node_t *node;
int i;
- for (i = 0; i < NNODES_MAX; i++) {
+ for (i = 0; i < nnodes_max; i++) {
node = node_get(i);
if (NODE_VALID(node)) {
if (node->qpi.qpi_num > 0)
@@ -1455,7 +1455,7 @@ int os_uncore_stop(perf_ctl_t *ctl __attribute__((unused)),
pf_uncoreimc_free(node);
}
} else {
- for (i = 0; i < NNODES_MAX; i++) {
+ for (i = 0; i < nnodes_max; i++) {
node = node_get(i);
if (NODE_VALID(node)) {
if (node->qpi.qpi_num > 0)
diff --git a/common/os/os_util.c b/common/os/os_util.c
index 0b862c2..ec3f6e5 100644
--- a/common/os/os_util.c
+++ b/common/os/os_util.c
@@ -863,7 +863,7 @@ static uint64_t cmt_field_value(char *dir, const char *field, int nid)
int i;
if (nid == -1) {
- for (i = 0; i < NNODES_MAX; i++) {
+ for (i = 0; i < nnodes_max; i++) {
if (cmt_task_node_value(dir, i, field,
&tmp) == 0)
val += tmp;
diff --git a/common/win.c b/common/win.c
index 03d096d..dad13d1 100644
--- a/common/win.c
+++ b/common/win.c
@@ -2714,7 +2714,7 @@ accdst_data_show(track_proc_t *proc, dyn_accdst_t *dyn, boolean_t *note_out)
void **addr_arr = NULL;
int *lat_arr = NULL;
int addr_num, i, nnodes, naccess_total = 0;
- map_nodedst_t nodedst_arr[NNODES_MAX];
+ map_nodedst_t *nodedst_arr;
accdst_line_t *lines;
char content[WIN_LINECHAR_MAX], intval_buf[16];
boolean_t ret = B_FALSE;
@@ -2728,14 +2728,19 @@ accdst_data_show(track_proc_t *proc, dyn_accdst_t *dyn, boolean_t *note_out)
return (B_FALSE);
}
+ nodedst_arr = (map_nodedst_t *) malloc(sizeof (map_nodedst_t) * nnodes_max);
+ if (!nodedst_arr) {
+ goto L_EXIT;
+ }
+
if (llrec2addr(proc, lwp, &addr_arr, &lat_arr, &addr_num) != 0) {
goto L_EXIT;
}
- (void) memset(nodedst_arr, 0, sizeof (map_nodedst_t) * NNODES_MAX);
+ (void) memset(nodedst_arr, 0, sizeof (map_nodedst_t) * nnodes_max);
if (addr_num > 0) {
if (map_addr2nodedst(proc->pid, addr_arr, lat_arr, addr_num,
- nodedst_arr, NNODES_MAX, &naccess_total) != 0) {
+ nodedst_arr, nnodes_max, &naccess_total) != 0) {
goto L_EXIT;
}
}
@@ -2784,7 +2789,7 @@ accdst_data_show(track_proc_t *proc, dyn_accdst_t *dyn, boolean_t *note_out)
* Save the per-node data with metrics in scrolling buffer.
*/
for (i = 0; i < nnodes; i++) {
- accdst_data_save(nodedst_arr, NNODES_MAX, naccess_total, i,
+ accdst_data_save(nodedst_arr, nnodes_max, naccess_total, i,
&lines[i]);
}
@@ -2799,6 +2804,8 @@ accdst_data_show(track_proc_t *proc, dyn_accdst_t *dyn, boolean_t *note_out)
reg_refresh_nout(r);
L_EXIT:
+ free(nodedst_arr);
+
if (lwp != NULL) {
lwp_refcount_dec(lwp);
}
