Re: [PATCH v5 6/6] powerpc/pseries: Add support for FORM2 associativity

On Thu, Jul 22, 2021 at 01:04:42PM +0530, Aneesh Kumar K.V wrote:

David Gibson [off-list ref] writes:

quoted

On Mon, Jun 28, 2021 at 08:41:17PM +0530, Aneesh Kumar K.V wrote:

quoted

PAPR interface currently supports two different ways of communicating resource
grouping details to the OS. These are referred to as Form 0 and Form 1
associativity grouping. Form 0 is the older format and is now considered
deprecated. This patch adds another resource grouping named FORM2.

Signed-off-by: Daniel Henrique Barboza <redacted>
Signed-off-by: Aneesh Kumar K.V <redacted>
---
 Documentation/powerpc/associativity.rst   | 103 ++++++++++++++
 arch/powerpc/include/asm/firmware.h       |   3 +-
 arch/powerpc/include/asm/prom.h           |   1 +
 arch/powerpc/kernel/prom_init.c           |   3 +-
 arch/powerpc/mm/numa.c                    | 157 ++++++++++++++++++----
 arch/powerpc/platforms/pseries/firmware.c |   1 +
 6 files changed, 242 insertions(+), 26 deletions(-)
 create mode 100644 Documentation/powerpc/associativity.rst

diff --git a/Documentation/powerpc/associativity.rst b/Documentation/powerpc/associativity.rst
new file mode 100644
index 000000000000..31cc7da2c7a6
--- /dev/null
+++ b/Documentation/powerpc/associativity.rst

@@ -0,0 +1,103 @@
+============================
+NUMA resource associativity
+=============================
+
+Associativity represents the groupings of the various platform resources into
+domains of substantially similar mean performance relative to resources outside
+of that domain. Resources subsets of a given domain that exhibit better
+performance relative to each other than relative to other resources subsets
+are represented as being members of a sub-grouping domain. This performance
+characteristic is presented in terms of NUMA node distance within the Linux kernel.
+From the platform view, these groups are also referred to as domains.

Pretty hard to decipher, but that's typical for PAPR.

quoted

+PAPR interface currently supports different ways of communicating these resource
+grouping details to the OS. These are referred to as Form 0, Form 1 and Form2
+associativity grouping. Form 0 is the older format and is now considered deprecated.

Nit: s/older/oldest/ since there are now >2 forms.

updated.

quoted

quoted

+Hypervisor indicates the type/form of associativity used via "ibm,architecture-vec-5 property".
+Bit 0 of byte 5 in the "ibm,architecture-vec-5" property indicates usage of Form 0 or Form 1.
+A value of 1 indicates the usage of Form 1 associativity. For Form 2 associativity
+bit 2 of byte 5 in the "ibm,architecture-vec-5" property is used.
+
+Form 0
+-----
+Form 0 associativity supports only two NUMA distances (LOCAL and REMOTE).
+
+Form 1
+-----
+With Form 1 a combination of ibm,associativity-reference-points, and ibm,associativity
+device tree properties are used to determine the NUMA distance between resource groups/domains.
+
+The “ibm,associativity” property contains a list of one or more numbers (domainID)
+representing the resource’s platform grouping domains.
+
+The “ibm,associativity-reference-points” property contains a list of one or more numbers
+(domainID index) that represents the 1 based ordinal in the associativity lists.
+The list of domainID indexes represents an increasing hierarchy of resource grouping.
+
+ex:
+{ primary domainID index, secondary domainID index, tertiary domainID index.. }
+
+Linux kernel uses the domainID at the primary domainID index as the NUMA node id.
+Linux kernel computes NUMA distance between two domains by recursively comparing
+if they belong to the same higher-level domains. For mismatch at every higher
+level of the resource group, the kernel doubles the NUMA distance between the
+comparing domains.
+
+Form 2
+-------
+Form 2 associativity format adds separate device tree properties representing NUMA node distance
+thereby making the node distance computation flexible. Form 2 also allows flexible primary
+domain numbering. With numa distance computation now detached from the index value in
+"ibm,associativity-reference-points" property, Form 2 allows a large number of primary domain
+ids at the same domainID index representing resource groups of different performance/latency
+characteristics.
+
+Hypervisor indicates the usage of FORM2 associativity using bit 2 of byte 5 in the
+"ibm,architecture-vec-5" property.
+
+"ibm,numa-lookup-index-table" property contains a list of one or more numbers representing
+the domainIDs present in the system. The offset of the domainID in this property is
+used as an index while computing numa distance information via "ibm,numa-distance-table".
+
+prop-encoded-array: The number N of the domainIDs encoded as with encode-int, followed by
+N domainID encoded as with encode-int
+
+For ex:
+"ibm,numa-lookup-index-table" =  {4, 0, 8, 250, 252}. The offset of domainID 8 (2) is used when
+computing the distance of domain 8 from other domains present in the system. For the rest of
+this document, this offset will be referred to as domain distance offset.
+
+"ibm,numa-distance-table" property contains a list of one or more numbers representing the NUMA
+distance between resource groups/domains present in the system.
+
+prop-encoded-array: The number N of the distance values encoded as with encode-int, followed by
+N distance values encoded as with encode-bytes. The max distance value we could encode is 255.
+The number N must be equal to the square of m where m is the number of domainIDs in the
+numa-lookup-index-table.
+
+For ex:
+ibm,numa-lookup-index-table =  {3, 0, 8, 40}
+ibm,numa-distance-table     =  {9, 10, 20, 80, 20, 10, 160, 80, 160, 10}

This representation doesn't make it clear that the 9 is a u32, but the
rest are u8s.

How do you suggest we specify that? I could do 9:u32 10:u8 etc. But
considering the details are explained in the paragraph above, is that
needed?

Yes, I think it is needed.  The examples are, honestly, a lot easier
to read and follow than the PAPR-ese text, so people are much more
likely to be looking at those than parsing the minutiae of the text.

quoted

quoted

+
+  | 0    8   40
+--|------------
+  |
+0 | 10   20  80
+  |
+8 | 20   10  160
+  |
+40| 80   160  10
+
+A possible "ibm,associativity" property for resources in node 0, 8 and 40
+
+{ 3, 6, 7, 0 }
+{ 3, 6, 9, 8 }
+{ 3, 6, 7, 40}
+
+With "ibm,associativity-reference-points"  { 0x3 }

You haven't actually described how ibm,associativity-reference-points
operates in Form2.

Nothing change w.r.t the definition of associativity-reference-points
w.r.t FORM2. It still will continue to show the increasing hierarchy of
resource groups.

I guess, except that really none of them matter except the primary any
more.

quoted

quoted

+"ibm,lookup-index-table" helps in having a compact representation of distance matrix.
+Since domainID can be sparse, the matrix of distances can also be effectively sparse.
+With "ibm,lookup-index-table" we can achieve a compact representation of
+distance information.

diff --git a/arch/powerpc/include/asm/firmware.h b/arch/powerpc/include/asm/firmware.h
index 60b631161360..97a3bd9ffeb9 100644
--- a/arch/powerpc/include/asm/firmware.h
+++ b/arch/powerpc/include/asm/firmware.h

@@ -53,6 +53,7 @@
 #define FW_FEATURE_ULTRAVISOR	ASM_CONST(0x0000004000000000)
 #define FW_FEATURE_STUFF_TCE	ASM_CONST(0x0000008000000000)
 #define FW_FEATURE_RPT_INVALIDATE ASM_CONST(0x0000010000000000)
+#define FW_FEATURE_FORM2_AFFINITY ASM_CONST(0x0000020000000000)
 
 #ifndef __ASSEMBLY__
 

@@ -73,7 +74,7 @@ enum {
 		FW_FEATURE_HPT_RESIZE | FW_FEATURE_DRMEM_V2 |
 		FW_FEATURE_DRC_INFO | FW_FEATURE_BLOCK_REMOVE |
 		FW_FEATURE_PAPR_SCM | FW_FEATURE_ULTRAVISOR |
-		FW_FEATURE_RPT_INVALIDATE,
+		FW_FEATURE_RPT_INVALIDATE | FW_FEATURE_FORM2_AFFINITY,
 	FW_FEATURE_PSERIES_ALWAYS = 0,
 	FW_FEATURE_POWERNV_POSSIBLE = FW_FEATURE_OPAL | FW_FEATURE_ULTRAVISOR,
 	FW_FEATURE_POWERNV_ALWAYS = 0,

diff --git a/arch/powerpc/include/asm/prom.h b/arch/powerpc/include/asm/prom.h
index df9fec9d232c..5c80152e8f18 100644
--- a/arch/powerpc/include/asm/prom.h
+++ b/arch/powerpc/include/asm/prom.h

@@ -149,6 +149,7 @@ extern int of_read_drc_info_cell(struct property **prop,
 #define OV5_XCMO		0x0440	/* Page Coalescing */
 #define OV5_FORM1_AFFINITY	0x0580	/* FORM1 NUMA affinity */
 #define OV5_PRRN		0x0540	/* Platform Resource Reassignment */
+#define OV5_FORM2_AFFINITY	0x0520	/* Form2 NUMA affinity */
 #define OV5_HP_EVT		0x0604	/* Hot Plug Event support */
 #define OV5_RESIZE_HPT		0x0601	/* Hash Page Table resizing */
 #define OV5_PFO_HW_RNG		0x1180	/* PFO Random Number Generator */

diff --git a/arch/powerpc/kernel/prom_init.c b/arch/powerpc/kernel/prom_init.c
index 5d9ea059594f..c483df6c9393 100644
--- a/arch/powerpc/kernel/prom_init.c
+++ b/arch/powerpc/kernel/prom_init.c

@@ -1069,7 +1069,8 @@ static const struct ibm_arch_vec ibm_architecture_vec_template __initconst = {
 #else
 		0,
 #endif
-		.associativity = OV5_FEAT(OV5_FORM1_AFFINITY) | OV5_FEAT(OV5_PRRN),
+		.associativity = OV5_FEAT(OV5_FORM1_AFFINITY) | OV5_FEAT(OV5_PRRN) |
+		OV5_FEAT(OV5_FORM2_AFFINITY),
 		.bin_opts = OV5_FEAT(OV5_RESIZE_HPT) | OV5_FEAT(OV5_HP_EVT),
 		.micro_checkpoint = 0,
 		.reserved0 = 0,

diff --git a/arch/powerpc/mm/numa.c b/arch/powerpc/mm/numa.c
index c6293037a103..c68846fc9550 100644
--- a/arch/powerpc/mm/numa.c
+++ b/arch/powerpc/mm/numa.c

@@ -56,12 +56,17 @@ static int n_mem_addr_cells, n_mem_size_cells;
 
 #define FORM0_AFFINITY 0
 #define FORM1_AFFINITY 1
+#define FORM2_AFFINITY 2
 static int affinity_form;
 
 #define MAX_DISTANCE_REF_POINTS 4
 static int max_associativity_domain_index;
 static const __be32 *distance_ref_points;
 static int distance_lookup_table[MAX_NUMNODES][MAX_DISTANCE_REF_POINTS];
+static int numa_distance_table[MAX_NUMNODES][MAX_NUMNODES] = {
+	[0 ... MAX_NUMNODES - 1] = { [0 ... MAX_NUMNODES - 1] = -1 }
+};
+static int numa_id_index_table[MAX_NUMNODES] = { [0 ... MAX_NUMNODES - 1] = NUMA_NO_NODE };
 
 /*
  * Allocate node_to_cpumask_map based on number of available nodes

@@ -166,6 +171,44 @@ static void unmap_cpu_from_node(unsigned long cpu)
 }
 #endif /* CONFIG_HOTPLUG_CPU || CONFIG_PPC_SPLPAR */
 
+/*
+ * Returns nid in the range [0..nr_node_ids], or -1 if no useful NUMA
+ * info is found.
+ */
+static int associativity_to_nid(const __be32 *associativity)
+{
+	int nid = NUMA_NO_NODE;
+
+	if (!numa_enabled)
+		goto out;
+
+	if (of_read_number(associativity, 1) >= primary_domain_index)
+		nid = of_read_number(&associativity[primary_domain_index], 1);
+
+	/* POWER4 LPAR uses 0xffff as invalid node */
+	if (nid == 0xffff || nid >= nr_node_ids)
+		nid = NUMA_NO_NODE;
+out:
+	return nid;
+}
+
+static int __cpu_form2_relative_distance(__be32 *cpu1_assoc, __be32 *cpu2_assoc)
+{
+	int dist;
+	int node1, node2;
+
+	node1 = associativity_to_nid(cpu1_assoc);
+	node2 = associativity_to_nid(cpu2_assoc);
+
+	dist = numa_distance_table[node1][node2];
+	if (dist <= LOCAL_DISTANCE)
+		return 0;
+	else if (dist <= REMOTE_DISTANCE)
+		return 1;
+	else
+		return 2;

Squashing the full range of distances into just 0, 1 or 2 seems odd.
But then, this whole cpu_distance() thing being distinct from
node_distance() seems odd.

quoted

+}
+
 static int __cpu_form1_relative_distance(__be32 *cpu1_assoc, __be32 *cpu2_assoc)
 {
 	int dist = 0;

@@ -186,8 +229,9 @@ int cpu_relative_distance(__be32 *cpu1_assoc, __be32 *cpu2_assoc)
 {
 	/* We should not get called with FORM0 */
 	VM_WARN_ON(affinity_form == FORM0_AFFINITY);
-
-	return __cpu_form1_relative_distance(cpu1_assoc, cpu2_assoc);
+	if (affinity_form == FORM1_AFFINITY)
+		return __cpu_form1_relative_distance(cpu1_assoc, cpu2_assoc);
+	return __cpu_form2_relative_distance(cpu1_assoc, cpu2_assoc);
 }
 
 /* must hold reference to node during call */

@@ -201,7 +245,9 @@ int __node_distance(int a, int b)
 	int i;
 	int distance = LOCAL_DISTANCE;
 
-	if (affinity_form == FORM0_AFFINITY)
+	if (affinity_form == FORM2_AFFINITY)
+		return numa_distance_table[a][b];
+	else if (affinity_form == FORM0_AFFINITY)
 		return ((a == b) ? LOCAL_DISTANCE : REMOTE_DISTANCE);
 
 	for (i = 0; i < max_associativity_domain_index; i++) {

Hmm.. couldn't we simplify this whole __node_distance function, if we
just update numa_distance_table[][] appropriately for Form0 and Form1
as well?

IIUC what you are suggesting is to look at the possibility of using
numa_distance_table[a][b] even for FORM1_AFFINITY? I can do that as part
of separate patch?

Ok, that's reasonable.

quoted

quoted

@@ -216,27 +262,6 @@ int __node_distance(int a, int b)
 }
 EXPORT_SYMBOL(__node_distance);
 
-/*
- * Returns nid in the range [0..nr_node_ids], or -1 if no useful NUMA
- * info is found.
- */
-static int associativity_to_nid(const __be32 *associativity)
-{
-	int nid = NUMA_NO_NODE;
-
-	if (!numa_enabled)
-		goto out;
-
-	if (of_read_number(associativity, 1) >= primary_domain_index)
-		nid = of_read_number(&associativity[primary_domain_index], 1);
-
-	/* POWER4 LPAR uses 0xffff as invalid node */
-	if (nid == 0xffff || nid >= nr_node_ids)
-		nid = NUMA_NO_NODE;
-out:
-	return nid;
-}
-
 /* Returns the nid associated with the given device tree node,
  * or -1 if not found.
  */

@@ -305,12 +330,84 @@ static void initialize_form1_numa_distance(struct device_node *node)
  */
 void update_numa_distance(struct device_node *node)
 {
+	int nid;
+
 	if (affinity_form == FORM0_AFFINITY)
 		return;
 	else if (affinity_form == FORM1_AFFINITY) {
 		initialize_form1_numa_distance(node);
 		return;
 	}
+
+	/* FORM2 affinity  */
+	nid = of_node_to_nid_single(node);
+	if (nid == NUMA_NO_NODE)
+		return;
+
+	/*
+	 * With FORM2 we expect NUMA distance of all possible NUMA
+	 * nodes to be provided during boot.
+	 */
+	WARN(numa_distance_table[nid][nid] == -1,
+	     "NUMA distance details for node %d not provided\n", nid);
+}
+
+/*
+ * ibm,numa-lookup-index-table= {N, domainid1, domainid2, ..... domainidN}
+ * ibm,numa-distance-table = { N, 1, 2, 4, 5, 1, 6, .... N elements}
+ */
+static void initialize_form2_numa_distance_lookup_table(struct device_node *root)
+{
+	int i, j;
+	const __u8 *numa_dist_table;
+	const __be32 *numa_lookup_index;
+	int numa_dist_table_length;
+	int max_numa_index, distance_index;
+
+	numa_lookup_index = of_get_property(root, "ibm,numa-lookup-index-table", NULL);
+	max_numa_index = of_read_number(&numa_lookup_index[0], 1);
+
+	/* first element of the array is the size and is encode-int */
+	numa_dist_table = of_get_property(root, "ibm,numa-distance-table", NULL);
+	numa_dist_table_length = of_read_number((const __be32 *)&numa_dist_table[0], 1);
+	/* Skip the size which is encoded int */
+	numa_dist_table += sizeof(__be32);
+
+	pr_debug("numa_dist_table_len = %d, numa_dist_indexes_len = %d\n",
+		 numa_dist_table_length, max_numa_index);
+
+	for (i = 0; i < max_numa_index; i++)
+		/* +1 skip the max_numa_index in the property */
+		numa_id_index_table[i] = of_read_number(&numa_lookup_index[i + 1], 1);
+
+
+	if (numa_dist_table_length != max_numa_index * max_numa_index) {
+
+		WARN(1, "Wrong NUMA distance information\n");
+		/* consider everybody else just remote. */
+		for (i = 0;  i < max_numa_index; i++) {
+			for (j = 0; j < max_numa_index; j++) {
+				int nodeA = numa_id_index_table[i];
+				int nodeB = numa_id_index_table[j];
+
+				if (nodeA == nodeB)
+					numa_distance_table[nodeA][nodeB] = LOCAL_DISTANCE;
+				else
+					numa_distance_table[nodeA][nodeB] = REMOTE_DISTANCE;
+			}
+		}
+	}
+
+	distance_index = 0;
+	for (i = 0;  i < max_numa_index; i++) {
+		for (j = 0; j < max_numa_index; j++) {
+			int nodeA = numa_id_index_table[i];
+			int nodeB = numa_id_index_table[j];
+
+			numa_distance_table[nodeA][nodeB] = numa_dist_table[distance_index++];
+			pr_debug("dist[%d][%d]=%d ", nodeA, nodeB, numa_distance_table[nodeA][nodeB]);
+		}
+	}
 }
 
 static int __init find_primary_domain_index(void)

@@ -323,6 +420,9 @@ static int __init find_primary_domain_index(void)
 	 */
 	if (firmware_has_feature(FW_FEATURE_OPAL)) {
 		affinity_form = FORM1_AFFINITY;
+	} else if (firmware_has_feature(FW_FEATURE_FORM2_AFFINITY)) {
+		dbg("Using form 2 affinity\n");
+		affinity_form = FORM2_AFFINITY;
 	} else if (firmware_has_feature(FW_FEATURE_FORM1_AFFINITY)) {
 		dbg("Using form 1 affinity\n");
 		affinity_form = FORM1_AFFINITY;

@@ -367,8 +467,17 @@ static int __init find_primary_domain_index(void)
 
 		index = of_read_number(&distance_ref_points[1], 1);
 	} else {
+		/*
+		 * Both FORM1 and FORM2 affinity find the primary domain details
+		 * at the same offset.
+		 */
 		index = of_read_number(distance_ref_points, 1);
 	}
+	/*
+	 * If it is FORM2 also initialize the distance table here.
+	 */
+	if (affinity_form == FORM2_AFFINITY)
+		initialize_form2_numa_distance_lookup_table(root);

Ew.  Calling a function called "find_primary_domain_index" to also
initialize the main distance table is needlessly counterintuitive.
Move this call to parse_numa_properties().

The reason I ended up doing it here is because 'root' is already fetched
here. But I agree it is confusing. I will move fetching of root inside
initialize_form2_numa_distance_lookup_table() and move the function
outside primary_index lookup.

Ok.  This is not a hot path anyway, so looking up root twice isn't
really a big deal anyway.

-- 
David Gibson			| I'll have my music baroque, and my code
david AT gibson.dropbear.id.au	| minimalist, thank you.  NOT _the_ _other_
				| _way_ _around_!
http://www.ozlabs.org/~dgibson