Skip site navigation (1)Skip section navigation (2)
Date:      Thu, 21 Oct 2021 22:31:36 GMT
From:      Alexander Motin <mav@FreeBSD.org>
To:        src-committers@FreeBSD.org, dev-commits-src-all@FreeBSD.org, dev-commits-src-branches@FreeBSD.org
Subject:   git: b5919ea4e6bb - stable/13 - x86: Add NUMA nodes into CPU topology.
Message-ID:  <202110212231.19LMVaFo018836@gitrepo.freebsd.org>

next in thread | raw e-mail | index | archive | help
The branch stable/13 has been updated by mav:

URL: https://cgit.FreeBSD.org/src/commit/?id=b5919ea4e6bb21e22484bc1943665e1f7e3bf888

commit b5919ea4e6bb21e22484bc1943665e1f7e3bf888
Author:     Alexander Motin <mav@FreeBSD.org>
AuthorDate: 2021-09-23 17:41:02 +0000
Commit:     Alexander Motin <mav@FreeBSD.org>
CommitDate: 2021-10-21 22:24:36 +0000

    x86: Add NUMA nodes into CPU topology.
    
    Depending on hardware, NUMA nodes may match last level caches, or
    they may be above them (AMD Zen 2/3) or below (Intel Xeon w/ SNC).
    This information is provided by ACPI instead of CPUID, and it is
    provided for each CPU individually instead of mask widths, but
    this code should be able to properly handle all the above cases.
    
    This change should immediately allow idle stealing in sched_ule(4)
    to prefer load from NUMA-local CPUs to remote ones when the node
    does not match LLC.  Later we may think of how to better handle it
    on sched_pickcpu() side.
    
    MFC after:      1 month
    
    (cherry picked from commit ef50d5fbc39fc39970eab1234222b5ac1d9ba74c)
---
 sys/kern/sched_ule.c |  2 ++
 sys/sys/smp.h        |  1 +
 sys/x86/x86/mp_x86.c | 82 +++++++++++++++++++++++++++++++++++++++++++---------
 3 files changed, 72 insertions(+), 13 deletions(-)

diff --git a/sys/kern/sched_ule.c b/sys/kern/sched_ule.c
index d92436f70db2..98c1f0bca981 100644
--- a/sys/kern/sched_ule.c
+++ b/sys/kern/sched_ule.c
@@ -3098,6 +3098,8 @@ sysctl_kern_sched_topology_spec_internal(struct sbuf *sb, struct cpu_group *cg,
 			sbuf_printf(sb, "<flag name=\"THREAD\">THREAD group</flag>");
 		if ((cg->cg_flags & CG_FLAG_SMT) != 0)
 			sbuf_printf(sb, "<flag name=\"SMT\">SMT group</flag>");
+		if ((cg->cg_flags & CG_FLAG_NODE) != 0)
+			sbuf_printf(sb, "<flag name=\"NODE\">NUMA node</flag>");
 		sbuf_printf(sb, "</flags>\n");
 	}
 
diff --git a/sys/sys/smp.h b/sys/sys/smp.h
index cee1199015a7..1da557212ae2 100644
--- a/sys/sys/smp.h
+++ b/sys/sys/smp.h
@@ -107,6 +107,7 @@ typedef struct cpu_group *cpu_group_t;
 #define	CG_FLAG_HTT	0x01		/* Schedule the alternate core last. */
 #define	CG_FLAG_SMT	0x02		/* New age htt, less crippled. */
 #define	CG_FLAG_THREAD	(CG_FLAG_HTT | CG_FLAG_SMT)	/* Any threading. */
+#define	CG_FLAG_NODE	0x04		/* NUMA node. */
 
 /*
  * Convenience routines for building and traversing topologies.
diff --git a/sys/x86/x86/mp_x86.c b/sys/x86/x86/mp_x86.c
index db40aab28ad5..326b6fdae77d 100644
--- a/sys/x86/x86/mp_x86.c
+++ b/sys/x86/x86/mp_x86.c
@@ -27,6 +27,7 @@
 #include <sys/cdefs.h>
 __FBSDID("$FreeBSD$");
 
+#include "opt_acpi.h"
 #ifdef __i386__
 #include "opt_apic.h"
 #endif
@@ -82,6 +83,11 @@ __FBSDID("$FreeBSD$");
 #include <machine/stack.h>
 #include <x86/ucode.h>
 
+#ifdef DEV_ACPI
+#include <contrib/dev/acpica/include/acpi.h>
+#include <dev/acpica/acpivar.h>
+#endif
+
 static MALLOC_DEFINE(M_CPUS, "cpus", "CPU items");
 
 /* lock region used by kernel profiling */
@@ -501,13 +507,16 @@ topo_probe(void)
 		int type;
 		int subtype;
 		int id_shift;
-	} topo_layers[MAX_CACHE_LEVELS + 4];
+	} topo_layers[MAX_CACHE_LEVELS + 5];
 	struct topo_node *parent;
 	struct topo_node *node;
 	int layer;
 	int nlayers;
 	int node_id;
 	int i;
+#if defined(DEV_ACPI) && MAXMEMDOM > 1
+	int d, domain;
+#endif
 
 	if (cpu_topo_probed)
 		return;
@@ -582,6 +591,31 @@ topo_probe(void)
 	topo_layers[nlayers].id_shift = 0;
 	nlayers++;
 
+#if defined(DEV_ACPI) && MAXMEMDOM > 1
+	if (vm_ndomains > 1) {
+		for (layer = 0; layer < nlayers; ++layer) {
+			for (i = 0; i <= max_apic_id; ++i) {
+				if ((i & ((1 << topo_layers[layer].id_shift) - 1)) == 0)
+					domain = -1;
+				if (!cpu_info[i].cpu_present)
+					continue;
+				d = acpi_pxm_get_cpu_locality(i);
+				if (domain >= 0 && domain != d)
+					break;
+				domain = d;
+			}
+			if (i > max_apic_id)
+				break;
+		}
+		KASSERT(layer < nlayers, ("NUMA domain smaller than PU"));
+		memmove(&topo_layers[layer+1], &topo_layers[layer],
+		    sizeof(*topo_layers) * (nlayers - layer));
+		topo_layers[layer].type = TOPO_TYPE_NODE;
+		topo_layers[layer].subtype = CG_SHARE_NONE;
+		nlayers++;
+	}
+#endif
+
 	topo_init_root(&topo_root);
 	for (i = 0; i <= max_apic_id; ++i) {
 		if (!cpu_info[i].cpu_present)
@@ -589,7 +623,12 @@ topo_probe(void)
 
 		parent = &topo_root;
 		for (layer = 0; layer < nlayers; ++layer) {
-			node_id = i >> topo_layers[layer].id_shift;
+#if defined(DEV_ACPI) && MAXMEMDOM > 1
+			if (topo_layers[layer].type == TOPO_TYPE_NODE) {
+				node_id = acpi_pxm_get_cpu_locality(i);
+			} else
+#endif
+				node_id = i >> topo_layers[layer].id_shift;
 			parent = topo_add_node_by_hwid(parent, node_id,
 			    topo_layers[layer].type,
 			    topo_layers[layer].subtype);
@@ -598,7 +637,12 @@ topo_probe(void)
 
 	parent = &topo_root;
 	for (layer = 0; layer < nlayers; ++layer) {
-		node_id = boot_cpu_id >> topo_layers[layer].id_shift;
+#if defined(DEV_ACPI) && MAXMEMDOM > 1
+		if (topo_layers[layer].type == TOPO_TYPE_NODE)
+			node_id = acpi_pxm_get_cpu_locality(boot_cpu_id);
+		else
+#endif
+			node_id = boot_cpu_id >> topo_layers[layer].id_shift;
 		node = topo_find_node_by_hwid(parent, node_id,
 		    topo_layers[layer].type,
 		    topo_layers[layer].subtype);
@@ -773,14 +817,18 @@ x86topo_add_sched_group(struct topo_node *root, struct cpu_group *cg_root)
 	int i;
 
 	KASSERT(root->type == TOPO_TYPE_SYSTEM || root->type == TOPO_TYPE_CACHE ||
-	    root->type == TOPO_TYPE_GROUP,
+	    root->type == TOPO_TYPE_NODE || root->type == TOPO_TYPE_GROUP,
 	    ("x86topo_add_sched_group: bad type: %u", root->type));
 	CPU_COPY(&root->cpuset, &cg_root->cg_mask);
 	cg_root->cg_count = root->cpu_count;
-	if (root->type == TOPO_TYPE_SYSTEM)
+	if (root->type == TOPO_TYPE_CACHE)
+		cg_root->cg_level = root->subtype;
+	else
 		cg_root->cg_level = CG_SHARE_NONE;
+	if (root->type == TOPO_TYPE_NODE)
+		cg_root->cg_flags = CG_FLAG_NODE;
 	else
-		cg_root->cg_level = root->subtype;
+		cg_root->cg_flags = 0;
 
 	/*
 	 * Check how many core nodes we have under the given root node.
@@ -801,7 +849,7 @@ x86topo_add_sched_group(struct topo_node *root, struct cpu_group *cg_root)
 
 	if (cg_root->cg_level != CG_SHARE_NONE &&
 	    root->cpu_count > 1 && ncores < 2)
-		cg_root->cg_flags = CG_FLAG_SMT;
+		cg_root->cg_flags |= CG_FLAG_SMT;
 
 	/*
 	 * Find out how many cache nodes we have under the given root node.
@@ -813,10 +861,18 @@ x86topo_add_sched_group(struct topo_node *root, struct cpu_group *cg_root)
 	nchildren = 0;
 	node = root;
 	while (node != NULL) {
-		if ((node->type != TOPO_TYPE_GROUP &&
-		    node->type != TOPO_TYPE_CACHE) ||
-		    (root->type != TOPO_TYPE_SYSTEM &&
-		    CPU_CMP(&node->cpuset, &root->cpuset) == 0)) {
+		if (CPU_CMP(&node->cpuset, &root->cpuset) == 0) {
+			if (node->type == TOPO_TYPE_CACHE &&
+			    cg_root->cg_level < node->subtype)
+				cg_root->cg_level = node->subtype;
+			if (node->type == TOPO_TYPE_NODE)
+				cg_root->cg_flags |= CG_FLAG_NODE;
+			node = topo_next_node(root, node);
+			continue;
+		}
+		if (node->type != TOPO_TYPE_GROUP &&
+		    node->type != TOPO_TYPE_NODE &&
+		    node->type != TOPO_TYPE_CACHE) {
 			node = topo_next_node(root, node);
 			continue;
 		}
@@ -841,9 +897,9 @@ x86topo_add_sched_group(struct topo_node *root, struct cpu_group *cg_root)
 	i = 0;
 	while (node != NULL) {
 		if ((node->type != TOPO_TYPE_GROUP &&
+		    node->type != TOPO_TYPE_NODE &&
 		    node->type != TOPO_TYPE_CACHE) ||
-		    (root->type != TOPO_TYPE_SYSTEM &&
-		    CPU_CMP(&node->cpuset, &root->cpuset) == 0)) {
+		    CPU_CMP(&node->cpuset, &root->cpuset) == 0) {
 			node = topo_next_node(root, node);
 			continue;
 		}



Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?202110212231.19LMVaFo018836>