Introduces numa_maps metrics inside linux_proc pmda and PCP numastat '-p' option support by sourav-sharma796 · Pull Request #2527 · performancecopilot/pcp

sourav-sharma796 · 2026-03-13T08:13:32Z

Implemented option -p in pcp numastat tool same as numactl numastat tool

Add new per-process NUMA maps metrics derived from /proc//numa_maps: proc.numa_maps.huge, proc.numa_maps.heap, proc.numa_maps.stack, proc.numa_maps.private (per-node usage in MB, exported as nodeN: pairs).
Extend pcp numastat/pcp-numastat with -p/--process to display per-process NUMA memory usage in a numastat -p-style layout.
Convert N= counts to MB using system page size (and Hugepagesize where applicable, with fallback).

How to Test

pcp numastat -p (detail view) and pcp numastat -p (summary view)

sourav-sharma796 · 2026-03-13T08:15:06Z

Hi, @natoscott
Please take a look at this new enhancement.

Thanks!

sourav-sharma796 · 2026-03-13T08:21:34Z

Local test with numastat tool

[root@localhost pcp]# pcp numastat -p 2 dbus
Linux  5.15.0-316.196.4.2.el8uek.x86_64  (localhost.localdomain)  03/13/26  x86_64    (2 CPU)
Timestamp            : 03/13/2026 08:18:36
Per-node process memory usage (in MBs)
PID                             Node 0            Total
---------------------  ---------------  ---------------
2 (kthreadd)                      0.00             0.00
1740 (dbus-daemon)                6.70             6.70
5620 (libvirt-dbus)              12.71            12.71
2868087 (dbus-daemon)             5.79             5.79
---------------------  ---------------  ---------------
Total                            25.20            25.20

[root@ssosharm-ol8-1 pcp]# numastat -p 2 dbus

Per-node process memory usage (in MBs)
PID                             Node 0           Total
---------------------  --------------- ---------------
2 (kthreadd)                      0.00            0.00
1740 (dbus-daemon)                6.70            6.70
5620 (libvirt-dbus)              12.70           12.70
2868087 (dbus-daemon)             5.80            5.80
---------------------  --------------- ---------------
Total                            25.20           25.20

Detailed view of a process

[root@localhost pcp]# numastat -p 2868087

Per-node process memory usage (in MBs) for PID 2868087 (dbus-daemon)
                           Node 0           Total
                  --------------- ---------------
Huge                         0.00            0.00
Heap                         0.16            0.16
Stack                        0.02            0.02
Private                      5.61            5.61
----------------  --------------- ---------------
Total                        5.79            5.79
[root@localhost pcp]#
[root@localhost pcp]# pcp numastat -p 2868087
Linux  5.15.0-316.196.4.2.el8uek.x86_64  (localhost.localdomain)  03/13/26  x86_64    (2 CPU)
Timestamp            : 03/13/2026 08:19:17
Per-node process memory usage (in MBs) for PID 2868087 (dbus-daemon)
                           Node 0            Total
                  ---------------  ---------------
Huge                         0.00             0.00
Heap                         0.16             0.16
Stack                        0.02             0.02
Private                      5.61             5.61
----------------  ---------------  ---------------
Total                        5.79             5.79

natoscott

LGTM - few suggested changes in-line.

Implemented option -p in pcp numastat tool same as numactl numastat tool - Add new per-process NUMA maps metrics derived from /proc/<pid>/numa_maps: proc.numa_maps.huge, proc.numa_maps.heap, proc.numa_maps.stack, proc.numa_maps.private (per-node usage in MB, exported as nodeN:<mb> pairs). - Extend pcp numastat/pcp-numastat with -p/--process to display per-process NUMA memory usage in a numastat -p-style layout. - Convert N<node>=<pages> counts to MB using system page size (and Hugepagesize where applicable, with fallback). How to Test - pcp numastat -p <pid> (detail view) and pcp numastat -p <pid1> <pid2> (summary view) Signed-off-by: Sourav Sharma <sourav.ss.sharma@oracle.com>

sourav-sharma796 · 2026-03-16T16:28:24Z

Hi,

@natoscott I closely looked at your above given comments and worked on them.
I researched about your proposed ideas to prevent overhead and slowness.

I have made changes and updated the source.

Here's a quick diff for review, as rebase made compare tab huge:

diff --git a/src/pcp/numastat/pcp-numastat.1 b/src/pcp/numastat/pcp-numastat.1
index 4804a5b2c3..2bd9638d47 100644
--- a/src/pcp/numastat/pcp-numastat.1
+++ b/src/pcp/numastat/pcp-numastat.1
@@ -75,7 +75,7 @@ operand is all digits) or regular expressions for command name matching.
 Matching for command names is performed against the full command line
 (as reported by \fBproc.psinfo.psargs\fR).
 When a single process ID operand is provided and it matches exactly one
-process, a detailed table of Huge, Heap, Stack and Private mappings is
+process, a detailed table of Hugepage, Heap, Stack and Private mappings is
 shown (as in \fBnumastat \-p\fR). When multiple process ID operands are
 provided, a per-process summary table is shown (also like
 \fBnumastat \-p\fR) and may be split into multiple node column blocks
diff --git a/src/pcp/numastat/pcp-numastat.py b/src/pcp/numastat/pcp-numastat.py
index 472ea07108..decdf124cc 100755
--- a/src/pcp/numastat/pcp-numastat.py
+++ b/src/pcp/numastat/pcp-numastat.py
@@ -91,14 +91,14 @@ PROCESS_METRICS = [
     "proc.psinfo.pid",
     "proc.psinfo.cmd",
     "proc.psinfo.psargs",
-    "proc.numa_maps.huge",
+    "proc.numa_maps.hugepage",
     "proc.numa_maps.heap",
     "proc.numa_maps.stack",
     "proc.numa_maps.private",
 ]

 PROCESS_NUMA_METRICS = [
-    ("Huge", "proc.numa_maps.huge"),
+    ("Huge", "proc.numa_maps.hugepage"),
     ("Heap", "proc.numa_maps.heap"),
     ("Stack", "proc.numa_maps.stack"),
     ("Private", "proc.numa_maps.private"),
diff --git a/src/pmdas/linux_proc/help b/src/pmdas/linux_proc/help
index fd61456b4c..b1cfdfcfe7 100644
--- a/src/pmdas/linux_proc/help
+++ b/src/pmdas/linux_proc/help
@@ -75,12 +75,18 @@ kernel threads
 @ proc.runq.kernel number of kernel threads
 Instantaneous number of processes with virtual size of zero (kernel threads)

-@ proc.numa_maps.huge per-node hugepage mapped memory in MB (/proc/<pid>/numa_maps)
-Values are reported as comma-separated node/value pairs, e.g.
+@ proc.numa_maps.hugepage per-node hugepage mapped memory in MB (/proc/<pid>/numa_maps)
+Hugepage map values are reported as comma-separated node/value pairs, e.g.
 node0:0.00,node1:4.00, where each value is in megabytes.
 @ proc.numa_maps.heap per-node heap mapped memory in MB (/proc/<pid>/numa_maps)
+Heap map values are reported as comma-separated node/value pairs, e.g.
+node0:0.00,node1:4.00, where each value is in megabytes.
 @ proc.numa_maps.stack per-node stack mapped memory in MB (/proc/<pid>/numa_maps)
+Stack map values are reported as comma-separated node/value pairs, e.g.
+node0:0.00,node1:4.00, where each value is in megabytes.
 @ proc.numa_maps.private per-node private mapped memory in MB (/proc/<pid>/numa_maps)
+Private map values are reported as comma-separated node/value pairs, e.g.
+node0:0.00,node1:4.00, where each value is in megabytes.

 @ proc.control.all.threads process indom includes threads
 If set to one, the process instance domain as reported by pmdaproc
diff --git a/src/pmdas/linux_proc/indom.h b/src/pmdas/linux_proc/indom.h
index abe2596ab2..c5c00a6792 100644
--- a/src/pmdas/linux_proc/indom.h
+++ b/src/pmdas/linux_proc/indom.h
@@ -62,6 +62,7 @@ extern FILE *proc_statsfile(const char *, char *, int);

 /* Generic globals setup during PMDA startup */
 extern size_t _pm_system_pagesize;
+extern size_t _pm_system_hugepagesize;
 extern long _pm_hertz;

 /*
diff --git a/src/pmdas/linux_proc/pmda.c b/src/pmdas/linux_proc/pmda.c
index a531edaca6..b8cc860ebf 100644
--- a/src/pmdas/linux_proc/pmda.c
+++ b/src/pmdas/linux_proc/pmda.c
@@ -66,6 +66,7 @@ static int                    autogroup = -1; /* =1 autogroup enabled */
 static unsigned int            threads;        /* control.all.threads */
 static char *                  cgroups;        /* control.all.cgroups */
 size_t                         _pm_system_pagesize;
+size_t                         _pm_system_hugepagesize;
 long                           _pm_hertz;

 /*
@@ -1408,7 +1409,7 @@ static pmdaMetric metrictab[] = {
 * numa_maps cluster
 */

-/* proc.numa_maps.huge */
+/* proc.numa_maps.hugepage */
   { NULL, { PMDA_PMID(CLUSTER_PID_NUMA_MAPS,0), PM_TYPE_STRING, PROC_INDOM,
     PM_SEM_INSTANT, PMDA_PMUNITS(0,0,0,0,0,0)}},
 /* proc.numa_maps.heap */
@@ -3595,9 +3596,9 @@ proc_fetchCallBack(pmdaMetric *mdesc, unsigned int inst, pmAtomValue *atom)
           return 0;

        switch(item) {
-          case 0: /* proc.numa_maps.huge */
-           atom->cp = proc_strings_lookup(entry->numa_maps.huge_id);
-            break;
+          case 0: /* proc.numa_maps.hugepage */
+                   atom->cp = proc_strings_lookup(entry->numa_maps.huge_id);
+                   break;

           case 1: /* proc.numa_maps.heap */
            atom->cp = proc_strings_lookup(entry->numa_maps.heap_id);
@@ -3976,6 +3977,25 @@ proc_gidname_lookup(int gid)
     return "";
 }

+static size_t
+proc_hugepagesize(void)
+{
+    unsigned long      huge_page_size_kb = 0;
+    char               buf[128];
+    FILE               *fs;
+
+    if ((fs = fopen("/proc/meminfo", "r")) == NULL)
+       return 0;
+
+    while (fgets(buf, sizeof(buf), fs)) {
+       if (sscanf(buf, "Hugepagesize: %lu kB", &huge_page_size_kb) == 1)
+           break;
+    }
+    fclose(fs);
+
+    return huge_page_size_kb * 1024;
+}
+
 /*
  * Initialise the agent (both daemon and DSO).
  */
@@ -3997,6 +4017,10 @@ proc_init(pmdaInterface *dp)
        _pm_system_pagesize = atoi(envpath);
     else
        _pm_system_pagesize = getpagesize();
+    if ((envpath = getenv("PROC_HUGEPAGESIZE")) != NULL)
+       _pm_system_hugepagesize = atoi(envpath);
+    else
+       _pm_system_hugepagesize = proc_hugepagesize();
     if ((envpath = getenv("PROC_STATSPATH")) != NULL)
        proc_statspath = envpath;
     if ((envpath = getenv("PROC_THREADS")) != NULL)
diff --git a/src/pmdas/linux_proc/proc_pid.c b/src/pmdas/linux_proc/proc_pid.c
index 36615a9bcf..2efb2ff29e 100644
--- a/src/pmdas/linux_proc/proc_pid.c
+++ b/src/pmdas/linux_proc/proc_pid.c
@@ -2568,9 +2568,6 @@ fetch_proc_pid_fdinfo(int id, proc_pid_t *proc_pid, int *sts)
 #define PROCESS_PRIVATE_INDEX  3
 #define PROCESS_CATEGORY_COUNT 4

-#define SMALL_BUF_SIZE 128
-
-#define KILOBYTE (1024.0)
 #define MEGABYTE (1024.0 * 1024.0)

 static const char *process_mem_tokens[] = {
@@ -2590,29 +2587,6 @@ typedef struct {
     double     values[PROCESS_CATEGORY_COUNT];
 } numa_node_totals_t;

-static double
-get_huge_page_size_in_bytes(void)
-{
-    static double      cached_huge_page_size = -1.0;
-    double     huge_page_size_kb = 0.0;
-    char       buf[SMALL_BUF_SIZE];
-    FILE       *fs;
-
-    if (cached_huge_page_size >= 0.0)
-       return cached_huge_page_size;
-
-    if ((fs = fopen("/proc/meminfo", "r")) == NULL)
-       return 0.0;
-
-    while (fgets(buf, sizeof(buf), fs)) {
-       if (sscanf(buf, "Hugepagesize: %lf kB", &huge_page_size_kb) == 1)
-           break;
-    }
-    fclose(fs);
-    cached_huge_page_size = huge_page_size_kb * KILOBYTE;
-    return cached_huge_page_size;
-}
-
 static int
 append_numa_maps_node(strbuf_t *b, int node_num, double value_mb)
 {
@@ -2718,19 +2692,28 @@ parse_proc_numa_maps(proc_pid_entry_t *ep, size_t buflen, char *buf)
     int                        node_count = 0;
     int                        sts = 0;
     char               *cur = buf;
+    char               *end;

-    (void)buflen;
+    if (buf == NULL || buflen == 0)
+       return 0;
+    /*
+     * Ensure the proc buffer is NUL-terminated so the string routines below
+     * cannot read past the end.  read_proc_entry() allocates len+1 bytes and
+     * uses buflen (len) for the bytes-read value passed here.
+     */
+    buf[buflen] = '\0';
+    end = buf + buflen;

-    page_size_bytes = (double)sysconf(_SC_PAGESIZE);
+    page_size_bytes = (double)_pm_system_pagesize;
     if (page_size_bytes <= 0.0)
        page_size_bytes = 4096.0;

-    huge_page_size_bytes = get_huge_page_size_in_bytes();
+    huge_page_size_bytes = (double)_pm_system_hugepagesize;
     if (huge_page_size_bytes <= 0.0)
        huge_page_size_bytes = page_size_bytes;

-    while (cur && *cur) {
-       char    *nl = strchr(cur, '\n');
+    while (cur < end && *cur) {
+       char    *nl = memchr(cur, '\n', (size_t)(end - cur));
        char    *tok, *saveptr = NULL;
        int     category;

diff --git a/src/pmdas/linux_proc/root_proc b/src/pmdas/linux_proc/root_proc
index 79c9765be6..d80915e9a9 100644
--- a/src/pmdas/linux_proc/root_proc
+++ b/src/pmdas/linux_proc/root_proc
@@ -459,7 +459,7 @@ proc {
 }

 proc.numa_maps {
-    huge               PROC:81:0
+    hugepage           PROC:81:0
     heap               PROC:81:1
     stack              PROC:81:2
     private            PROC:81:3

Please take a look.

Thanks for your time reviewing this code.

natoscott · 2026-03-20T00:21:49Z

@sourav-sharma796 I notice the values have a trailing comma - could you fix that?

proc.numa_maps.private
    inst [785462 or "785462 -bash"] value "node0:3.69,"
    inst [429983 or "429983 /var/lib/pcp/pmdas/proc/pmdaproc"] value "node0:5.69,"

There's also a few failing QA tests as a result of this change, I'll take a quick look at those.

sourav-sharma796 · 2026-03-20T03:33:14Z

@sourav-sharma796 I notice the values have a trailing comma - could you fix that?
proc.numa_maps.private
    inst [785462 or "785462 -bash"] value "node0:3.69,"
    inst [429983 or "429983 /var/lib/pcp/pmdas/proc/pmdaproc"] value "node0:5.69,"
There's also a few failing QA tests as a result of this change, I'll take a quick look at those.

Yes, I can fix this. I'll open up a PR asap.
Thanks!

sourav-sharma796 · 2026-03-20T06:51:12Z

@sourav-sharma796 I notice the values have a trailing comma - could you fix that?
proc.numa_maps.private
    inst [785462 or "785462 -bash"] value "node0:3.69,"
    inst [429983 or "429983 /var/lib/pcp/pmdas/proc/pmdaproc"] value "node0:5.69,"
There's also a few failing QA tests as a result of this change, I'll take a quick look at those.
Yes, I can fix this. I'll open up a PR asap. Thanks!

See: #2544

natoscott · 2026-03-20T07:17:48Z

Another thing I forgot about - the overnight Coverity scan came back with this...

*** CID 502014:         Insecure data handling  (INTEGER_OVERFLOW)
/src/pmdas/linux_proc/pmda.c: 3996             in proc_hugepagesize()
3990         while (fgets(buf, sizeof(buf), fs)) {
3991     	if (sscanf(buf, "Hugepagesize: %lu kB", &huge_page_size_kb) == 1)
3992     	    break;
3993         }
3994         fclose(fs);
3995     
>>>     CID 502014:         Insecure data handling  (INTEGER_OVERFLOW)
>>>     "huge_page_size_kb * 1024UL", which might have overflowed, is returned from the function.
3996         return huge_page_size_kb * 1024;
3997     }
3998     
3999     /*
4000      * Initialise the agent (both daemon and DSO).
4001      */

It's benign since the kernel wont give us garbage (hopefully ;) - but we could easily guard against overflow here (eg just check the huge page size is less than something that'd overflow) - can you take a look?

sourav-sharma796 · 2026-03-20T07:20:46Z

Another thing I forgot about - the overnight Coverity scan came back with this...
*** CID 502014:         Insecure data handling  (INTEGER_OVERFLOW)
/src/pmdas/linux_proc/pmda.c: 3996             in proc_hugepagesize()
3990         while (fgets(buf, sizeof(buf), fs)) {
3991     	if (sscanf(buf, "Hugepagesize: %lu kB", &huge_page_size_kb) == 1)
3992     	    break;
3993         }
3994         fclose(fs);
3995     
>>>     CID 502014:         Insecure data handling  (INTEGER_OVERFLOW)
>>>     "huge_page_size_kb * 1024UL", which might have overflowed, is returned from the function.
3996         return huge_page_size_kb * 1024;
3997     }
3998     
3999     /*
4000      * Initialise the agent (both daemon and DSO).
4001      */
It's benign since the kernel wont give us garbage (hopefully ;) - but we could easily guard against overflow here (eg just check the huge page size is less than something that'd overflow) - can you take a look?

Yes, I'm looking into it.

I appreciate your efforts.
Thanks!

sourav-sharma796 · 2026-03-20T07:40:47Z

Another thing I forgot about - the overnight Coverity scan came back with this...
*** CID 502014:         Insecure data handling  (INTEGER_OVERFLOW)
/src/pmdas/linux_proc/pmda.c: 3996             in proc_hugepagesize()
3990         while (fgets(buf, sizeof(buf), fs)) {
3991     	if (sscanf(buf, "Hugepagesize: %lu kB", &huge_page_size_kb) == 1)
3992     	    break;
3993         }
3994         fclose(fs);
3995     
>>>     CID 502014:         Insecure data handling  (INTEGER_OVERFLOW)
>>>     "huge_page_size_kb * 1024UL", which might have overflowed, is returned from the function.
3996         return huge_page_size_kb * 1024;
3997     }
3998     
3999     /*
4000      * Initialise the agent (both daemon and DSO).
4001      */
It's benign since the kernel wont give us garbage (hopefully ;) - but we could easily guard against overflow here (eg just check the huge page size is less than something that'd overflow) - can you take a look?

Hi @natoscott , following change should fix this:

--- a/src/pmdas/linux_proc/pmda.c
+++ b/src/pmdas/linux_proc/pmda.c
@@ -3993,7 +3993,16 @@ proc_hugepagesize(void)
         }
         fclose(fs);
     
-        return huge_page_size_kb * 1024;
+        /* Guard against overflow (CID 502014) */
+        if (huge_page_size_kb > ULONG_MAX / 1024UL) {
+            /* Unexpected value from /proc - fail safely */
+            return 0;
+        }
+
+        return huge_page_size_kb * 1024UL;
 }

What do you think on this?

natoscott · 2026-03-20T23:00:36Z

@sourav-sharma796 LGTM

sourav-sharma796 · 2026-03-23T11:31:44Z

@sourav-sharma796 LGTM

Hi @natoscott ,
I've opened a PR #2548 for this.
Please check.

Thanks!

sourav-sharma796 changed the title ~~Introduces numa_maps metrics inside linux_proc pmda~~ Introduces numa_maps metrics inside linux_proc pmda and PCP numastat '-p' option support Mar 13, 2026

sourav-sharma796 force-pushed the add_numa_metrics branch from 0f100d5 to 6614d5e Compare March 13, 2026 08:31

natoscott reviewed Mar 15, 2026

View reviewed changes

sourav-sharma796 force-pushed the add_numa_metrics branch from 6614d5e to 0ea70c7 Compare March 16, 2026 16:09

sourav-sharma796 requested a review from natoscott March 17, 2026 04:23

natoscott merged commit 5c9d2da into performancecopilot:main Mar 18, 2026
23 checks passed

Uh oh!

Conversation

sourav-sharma796 commented Mar 13, 2026

Uh oh!

sourav-sharma796 commented Mar 13, 2026

Uh oh!

sourav-sharma796 commented Mar 13, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

natoscott left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

sourav-sharma796 commented Mar 16, 2026

Uh oh!

Uh oh!

natoscott commented Mar 20, 2026

Uh oh!

sourav-sharma796 commented Mar 20, 2026

Uh oh!

sourav-sharma796 commented Mar 20, 2026

Uh oh!

natoscott commented Mar 20, 2026

Uh oh!

sourav-sharma796 commented Mar 20, 2026

Uh oh!

sourav-sharma796 commented Mar 20, 2026

Uh oh!

natoscott commented Mar 20, 2026

Uh oh!

sourav-sharma796 commented Mar 23, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

sourav-sharma796 commented Mar 13, 2026 •

edited

Loading

sourav-sharma796 commented Mar 23, 2026 •

edited

Loading