[PATCH V7 0/2] powerpc/nodes: Fix issues with memoryless nodes

powerpc/nodes: Ensure enough nodes avail for operations

powerpc/initnodes: Ensure nodes initialized for hotplug

hotplug/cpu: Fix crash with memoryless nodes

Signed-off-by: Michael Bringmann <redacted>

Michael Bringmann (3):
  powerpc/nodes: Ensure enough nodes avail for operations
  powerpc/initnodes: Ensure nodes initialized for hotplug
  hotplug/cpu: Fix crash with memoryless nodes
---
Changes in V7:
  -- Remove some node initialization/allocation from boot setup
     to later in runtime to try to limit memory needs early on
  -- Add some needed node initialization to runtime code that maps
     CPUs based on VPHN associativity
  -- Add error checks and alternate recovery for compile flag
     CONFIG_MEMORY_HOTPLUG
  -- Add alternate node selection recovery for !CONFIG_MEMORY_HOTPLUG
  -- Add initialization for systems that start using nodes before new
     CPU brought online.
  -- Make function find_cpu_nid() externally visible/usable so that
     it may be used from hotplug-cpu.c

To: linuxppc-dev@lists.ozlabs.org
Cc: Michael Ellerman <mpe@ellerman.id.au>
Cc: Michael Bringmann <redacted>
Cc: John Allen <redacted>
Cc: Nathan Fontenot <redacted>
Cc: Tyrel Datwyler <redacted>
Cc: Thomas Falcon <redacted>
Subject: [PATCH V7 2/3] poserpc/initnodes: Ensure nodes initialized for hotplug

On powerpc systems which allow 'hot-add' of CPU, it may occur that
the new resources are to be inserted into nodes that were not used
for memory resources at bootup.  Many different configurations of
PowerPC resources may need to be supported depending upon the
environment.  Important characteristics of the nodes and operating
environment include:

* Dedicated vs. shared resources.  Shared resources require
  information such as the VPHN hcall for CPU assignment to nodes.
  Associativity decisions made based on dedicated resource rules,
  such as associativity properties in the device tree, may vary
  from decisions made using the values returned by the VPHN hcall.
* memoryless nodes at boot.  Nodes need to be defined as 'possible'
  at boot for operation with other code modules.  Previously, the
  powerpc code would limit the set of possible nodes to those which
  have memory assigned at boot, and were thus online.  Subsequent
  add/remove of CPUs or memory would only work with this subset of
  possible nodes.
* memoryless nodes with CPUs at boot.  Due to the previous restriction
  on nodes, nodes that had CPUs but no memory were being collapsed
  into other nodes that did have memory at boot.  In practice this
  meant that the node assignment presented by the runtime kernel
  differed from the affinity and associativity attributes presented
  by the device tree or VPHN hcalls.  Nodes that might be known to
  the pHyp were not 'possible' in the runtime kernel because they did
  not have memory at boot.

This patch fixes some problems encountered at runtime with
configurations that support memory-less nodes, or that hot-add CPUs
into nodes that are memoryless during system execution after boot.
The problems of interest include,

* Nodes known to powerpc to be memoryless at boot, but to have
  CPUs in them are allowed to be 'possible' and 'online'.  Memory
  allocations for those nodes are taken from another node that does
  have memory until and if memory is hot-added to the node.
* Nodes which have no resources assigned at boot, but which may still
  be referenced subsequently by affinity or associativity attributes,
  are kept in the list of 'possible' nodes for powerpc.  Hot-add of
  memory or CPUs to the system can reference these nodes and bring
  them online instead of redirecting the references to one of the set
  of nodes known to have memory at boot.

Note that this software operates under the context of CPU hotplug.
We are not doing memory hotplug in this code, but rather updating
the kernel's CPU topology (i.e. arch_update_cpu_topology /
numa_update_cpu_topology).  We are initializing a node that may be
used by CPUs or memory before it can be referenced as invalid by a
CPU hotplug operation.  CPU hotplug operations are protected by a
range of APIs including cpu_maps_update_begin/cpu_maps_update_done,
cpus_read/write_lock / cpus_read/write_unlock, device locks, and more.
Memory hotplug operations, including try_online_node, are protected
by mem_hotplug_begin/mem_hotplug_done, device locks, and more.  In
the case of CPUs being hot-added to a previously memoryless node, the
try_online_node operation occurs wholly within the CPU locks with no
overlap.  Using HMC hot-add/hot-remove operations, we have been able
to add and remove CPUs to any possible node without failures.  HMC
operations involve a degree self-serialization, though.

Signed-off-by: Michael Bringmann <redacted>
---
Changes in V6:
  -- Add some needed node initialization to runtime code that maps
     CPUs based on VPHN associativity
  -- Add error checks and alternate recovery for compile flag
     CONFIG_MEMORY_HOTPLUG
  -- Add alternate node selection recovery for !CONFIG_MEMORY_HOTPLUG
  -- Add more information to the patch introductory text
---
 arch/powerpc/mm/numa.c |   51 ++++++++++++++++++++++++++++++++++++++----------
 1 file changed, 40 insertions(+), 11 deletions(-)

diff --git a/arch/powerpc/mm/numa.c b/arch/powerpc/mm/numa.c
index 334a1ff..163f4cc 100644
--- a/arch/powerpc/mm/numa.c
+++ b/arch/powerpc/mm/numa.c

@@ -551,7 +551,7 @@ static int numa_setup_cpu(unsigned long lcpu)
 	nid = of_node_to_nid_single(cpu);
 
 out_present:
-	if (nid < 0 || !node_online(nid))
+	if (nid < 0 || !node_possible(nid))
 		nid = first_online_node;
 
 	map_cpu_to_node(lcpu, nid);

@@ -867,7 +867,7 @@ void __init dump_numa_cpu_topology(void)
 }
 
 /* Initialize NODE_DATA for a node on the local memory */
-static void __init setup_node_data(int nid, u64 start_pfn, u64 end_pfn)
+static void setup_node_data(int nid, u64 start_pfn, u64 end_pfn)
 {
 	u64 spanned_pages = end_pfn - start_pfn;
 	const size_t nd_size = roundup(sizeof(pg_data_t), SMP_CACHE_BYTES);

@@ -913,10 +913,8 @@ static void __init find_possible_nodes(void)
 		min_common_depth);
 
 	for (i = 0; i < numnodes; i++) {
-		if (!node_possible(i)) {
-			setup_node_data(i, 0, 0);
+		if (!node_possible(i))
 			node_set(i, node_possible_map);
-		}
 	}
 
 out:

@@ -1312,6 +1310,42 @@ static long vphn_get_associativity(unsigned long cpu,
 	return rc;
 }
 
+static inline int find_cpu_nid(int cpu)
+{
+	__be32 associativity[VPHN_ASSOC_BUFSIZE] = {0};
+	int new_nid;
+
+	/* Use associativity from first thread for all siblings */
+	vphn_get_associativity(cpu, associativity);
+	new_nid = associativity_to_nid(associativity);
+	if (new_nid < 0 || !node_possible(new_nid))
+		new_nid = first_online_node;
+
+	if (NODE_DATA(new_nid) == NULL) {
+#ifdef CONFIG_MEMORY_HOTPLUG
+		/*
+		 * Need to ensure that NODE_DATA is initialized
+		 * for a node from available memory (see
+		 * memblock_alloc_try_nid).  If unable to init
+		 * the node, then default to nearest node that
+		 * has memory installed.
+		 */
+		if (try_online_node(new_nid))
+			new_nid = first_online_node;
+#else
+		/*
+		 * Default to using the nearest node that has
+		 * memory installed.  Otherwise, it would be 
+		 * necessary to patch the kernel MM code to deal
+		 * with more memoryless-node error conditions.
+		 */
+		new_nid = first_online_node;
+#endif
+	}
+
+	return new_nid;
+}
+
 /*
  * Update the CPU maps and sysfs entries for a single CPU when its NUMA
  * characteristics change. This function doesn't perform any locking and is

@@ -1379,7 +1413,6 @@ int numa_update_cpu_topology(bool cpus_locked)
 {
 	unsigned int cpu, sibling, changed = 0;
 	struct topology_update_data *updates, *ud;
-	__be32 associativity[VPHN_ASSOC_BUFSIZE] = {0};
 	cpumask_t updated_cpus;
 	struct device *dev;
 	int weight, new_nid, i = 0;

@@ -1417,11 +1450,7 @@ int numa_update_cpu_topology(bool cpus_locked)
 			continue;
 		}
 
-		/* Use associativity from first thread for all siblings */
-		vphn_get_associativity(cpu, associativity);
-		new_nid = associativity_to_nid(associativity);
-		if (new_nid < 0 || !node_online(new_nid))
-			new_nid = first_online_node;
+		new_nid = find_cpu_nid(cpu);
 
 		if (new_nid == numa_cpu_lookup_table[cpu]) {
 			cpumask_andnot(&cpu_associativity_changes_mask,

On powerpc systems which allow 'hot-add' of CPU or memory resources,
it may occur that the new resources are to be inserted into nodes
that were not used for these resources at bootup.  In the kernel,
any node that is used must be defined and initialized.  These empty
nodes may occur when,

* Dedicated vs. shared resources.  Shared resources require
  information such as the VPHN hcall for CPU assignment to nodes.
  Associativity decisions made based on dedicated resource rules,
  such as associativity properties in the device tree, may vary
  from decisions made using the values returned by the VPHN hcall.
* memoryless nodes at boot.  Nodes need to be defined as 'possible'
  at boot for operation with other code modules.  Previously, the
  powerpc code would limit the set of possible nodes to those which
  have memory assigned at boot, and were thus online.  Subsequent
  add/remove of CPUs or memory would only work with this subset of
  possible nodes.
* memoryless nodes with CPUs at boot.  Due to the previous restriction
  on nodes, nodes that had CPUs but no memory were being collapsed
  into other nodes that did have memory at boot.  In practice this
  meant that the node assignment presented by the runtime kernel
  differed from the affinity and associativity attributes presented
  by the device tree or VPHN hcalls.  Nodes that might be known to
  the pHyp were not 'possible' in the runtime kernel because they did
  not have memory at boot.

This patch ensures that sufficient nodes are defined to support
configuration requirements after boot, as well as at boot.  This
patch set fixes a couple of problems.

* Nodes known to powerpc to be memoryless at boot, but to have
  CPUs in them are allowed to be 'possible' and 'online'.  Memory
  allocations for those nodes are taken from another node that does
  have memory until and if memory is hot-added to the node.
* Nodes which have no resources assigned at boot, but which may still
  be referenced subsequently by affinity or associativity attributes,
  are kept in the list of 'possible' nodes for powerpc.  Hot-add of
  memory or CPUs to the system can reference these nodes and bring
  them online instead of redirecting to one of the set of nodes that
  were known to have memory at boot.

This patch extracts the value of the lowest domain level (number of
allocable resources) from the device tree property
"ibm,max-associativity-domains" to use as the maximum number of nodes
to setup as possibly available in the system.  This new setting will
override the instruction,

    nodes_and(node_possible_map, node_possible_map, node_online_map);

presently seen in the function arch/powerpc/mm/numa.c:initmem_init().

If the "ibm,max-associativity-domains" property is not present at boot,
no operation will be performed to define or enable additional nodes, or
enable the above 'nodes_and()'.

Signed-off-by: Michael Bringmann <redacted>
---
Changes in V6:
  -- Remove some node initialization/allocation from boot setup
     to later in runtime to try to limit memory needs early on
  -- Augment descriptive documentation for patch
---
 arch/powerpc/mm/numa.c |   40 +++++++++++++++++++++++++++++++++++++---
 1 file changed, 37 insertions(+), 3 deletions(-)

diff --git a/arch/powerpc/mm/numa.c b/arch/powerpc/mm/numa.c
index eb604b3..334a1ff 100644
--- a/arch/powerpc/mm/numa.c
+++ b/arch/powerpc/mm/numa.c

@@ -892,6 +892,37 @@ static void __init setup_node_data(int nid, u64 start_pfn, u64 end_pfn)
 	NODE_DATA(nid)->node_spanned_pages = spanned_pages;
 }
 
+static void __init find_possible_nodes(void)
+{
+	struct device_node *rtas;
+	u32 numnodes, i;
+
+	if (min_common_depth <= 0)
+		return;
+
+	rtas = of_find_node_by_path("/rtas");
+	if (!rtas)
+		return;
+
+	if (of_property_read_u32_index(rtas,
+				"ibm,max-associativity-domains",
+				min_common_depth, &numnodes))
+		goto out;
+
+	pr_info("numa: Nodes = %d (mcd = %d)\n", numnodes,
+		min_common_depth);
+
+	for (i = 0; i < numnodes; i++) {
+		if (!node_possible(i)) {
+			setup_node_data(i, 0, 0);
+			node_set(i, node_possible_map);
+		}
+	}
+
+out:
+	of_node_put(rtas);
+}
+
 void __init initmem_init(void)
 {
 	int nid, cpu;

@@ -905,12 +936,15 @@ void __init initmem_init(void)
 	memblock_dump_all();
 
 	/*
-	 * Reduce the possible NUMA nodes to the online NUMA nodes,
-	 * since we do not support node hotplug. This ensures that  we
-	 * lower the maximum NUMA node ID to what is actually present.
+	 * Modify the set of possible NUMA nodes to reflect information
+	 * available about the set of online nodes, and the set of nodes
+	 * that we expect to make use of for this platform's affinity
+	 * calculations.
 	 */
 	nodes_and(node_possible_map, node_possible_map, node_online_map);
 
+	find_possible_nodes();
+
 	for_each_online_node(nid) {
 		unsigned long start_pfn, end_pfn;
 

On powerpc systems which allow 'hot-add' of CPU, it may occur that
the new resources are to be inserted into nodes that were not used
for memory resources at bootup.  Many different configurations of
PowerPC resources may need to be supported depending upon the
environment.  Important characteristics of the nodes and operating
environment include:

* Dedicated vs. shared resources.  Shared resources require
  information such as the VPHN hcall for CPU assignment to nodes.
  Associativity decisions made based on dedicated resource rules,
  such as associativity properties in the device tree, may vary
  from decisions made using the values returned by the VPHN hcall.
* memoryless nodes at boot.  Nodes need to be defined as 'possible'
  at boot for operation with other code modules.  Previously, the
  powerpc code would limit the set of possible nodes to those which
  have memory assigned at boot, and were thus online.  Subsequent
  add/remove of CPUs or memory would only work with this subset of
  possible nodes.
* memoryless nodes with CPUs at boot.  Due to the previous restriction
  on nodes, nodes that had CPUs but no memory were being collapsed
  into other nodes that did have memory at boot.  In practice this
  meant that the node assignment presented by the runtime kernel
  differed from the affinity and associativity attributes presented
  by the device tree or VPHN hcalls.  Nodes that might be known to
  the pHyp were not 'possible' in the runtime kernel because they did
  not have memory at boot.

This patch fixes some problems encountered at runtime with
configurations that support memory-less nodes, or that hot-add CPUs
into nodes that are memoryless during system execution after boot.
The problems of interest include,

* Nodes known to powerpc to be memoryless at boot, but to have
  CPUs in them are allowed to be 'possible' and 'online'.  Memory
  allocations for those nodes are taken from another node that does
  have memory until and if memory is hot-added to the node.
* Nodes which have no resources assigned at boot, but which may still
  be referenced subsequently by affinity or associativity attributes,
  are kept in the list of 'possible' nodes for powerpc.  Hot-add of
  memory or CPUs to the system can reference these nodes and bring
  them online instead of redirecting the references to one of the set
  of nodes known to have memory at boot.

Note that this software operates under the context of CPU hotplug.
We are not doing memory hotplug in this code, but rather updating
the kernel's CPU topology (i.e. arch_update_cpu_topology /
numa_update_cpu_topology).  We are initializing a node that may be
used by CPUs or memory before it can be referenced as invalid by a
CPU hotplug operation.  CPU hotplug operations are protected by a
range of APIs including cpu_maps_update_begin/cpu_maps_update_done,
cpus_read/write_lock / cpus_read/write_unlock, device locks, and more.
Memory hotplug operations, including try_online_node, are protected
by mem_hotplug_begin/mem_hotplug_done, device locks, and more.  In
the case of CPUs being hot-added to a previously memoryless node, the
try_online_node operation occurs wholly within the CPU locks with no
overlap.  Using HMC hot-add/hot-remove operations, we have been able
to add and remove CPUs to any possible node without failures.  HMC
operations involve a degree self-serialization, though.

Signed-off-by: Michael Bringmann <redacted>
---
Changes in V6:
  -- Add some needed node initialization to runtime code that maps
     CPUs based on VPHN associativity
  -- Add error checks and alternate recovery for compile flag
     CONFIG_MEMORY_HOTPLUG
  -- Add alternate node selection recovery for !CONFIG_MEMORY_HOTPLUG
  -- Add more information to the patch introductory text
---
 arch/powerpc/mm/numa.c |   51 ++++++++++++++++++++++++++++++++++++++----------
 1 file changed, 40 insertions(+), 11 deletions(-)

diff --git a/arch/powerpc/mm/numa.c b/arch/powerpc/mm/numa.c
index 334a1ff..163f4cc 100644
--- a/arch/powerpc/mm/numa.c
+++ b/arch/powerpc/mm/numa.c

@@ -551,7 +551,7 @@ static int numa_setup_cpu(unsigned long lcpu)
 	nid = of_node_to_nid_single(cpu);
 
 out_present:
-	if (nid < 0 || !node_online(nid))
+	if (nid < 0 || !node_possible(nid))
 		nid = first_online_node;
 
 	map_cpu_to_node(lcpu, nid);

@@ -867,7 +867,7 @@ void __init dump_numa_cpu_topology(void)
 }
 
 /* Initialize NODE_DATA for a node on the local memory */
-static void __init setup_node_data(int nid, u64 start_pfn, u64 end_pfn)
+static void setup_node_data(int nid, u64 start_pfn, u64 end_pfn)
 {
 	u64 spanned_pages = end_pfn - start_pfn;
 	const size_t nd_size = roundup(sizeof(pg_data_t), SMP_CACHE_BYTES);

@@ -913,10 +913,8 @@ static void __init find_possible_nodes(void)
 		min_common_depth);
 
 	for (i = 0; i < numnodes; i++) {
-		if (!node_possible(i)) {
-			setup_node_data(i, 0, 0);
+		if (!node_possible(i))
 			node_set(i, node_possible_map);
-		}
 	}
 
 out:

@@ -1312,6 +1310,42 @@ static long vphn_get_associativity(unsigned long cpu,
 	return rc;
 }
 
+static inline int find_cpu_nid(int cpu)
+{
+	__be32 associativity[VPHN_ASSOC_BUFSIZE] = {0};
+	int new_nid;
+
+	/* Use associativity from first thread for all siblings */
+	vphn_get_associativity(cpu, associativity);
+	new_nid = associativity_to_nid(associativity);
+	if (new_nid < 0 || !node_possible(new_nid))
+		new_nid = first_online_node;
+
+	if (NODE_DATA(new_nid) == NULL) {
+#ifdef CONFIG_MEMORY_HOTPLUG
+		/*
+		 * Need to ensure that NODE_DATA is initialized
+		 * for a node from available memory (see
+		 * memblock_alloc_try_nid).  If unable to init
+		 * the node, then default to nearest node that
+		 * has memory installed.
+		 */
+		if (try_online_node(new_nid))
+			new_nid = first_online_node;
+#else
+		/*
+		 * Default to using the nearest node that has
+		 * memory installed.  Otherwise, it would be 
+		 * necessary to patch the kernel MM code to deal
+		 * with more memoryless-node error conditions.
+		 */
+		new_nid = first_online_node;
+#endif
+	}
+
+	return new_nid;
+}
+
 /*
  * Update the CPU maps and sysfs entries for a single CPU when its NUMA
  * characteristics change. This function doesn't perform any locking and is

@@ -1379,7 +1413,6 @@ int numa_update_cpu_topology(bool cpus_locked)
 {
 	unsigned int cpu, sibling, changed = 0;
 	struct topology_update_data *updates, *ud;
-	__be32 associativity[VPHN_ASSOC_BUFSIZE] = {0};
 	cpumask_t updated_cpus;
 	struct device *dev;
 	int weight, new_nid, i = 0;

@@ -1417,11 +1450,7 @@ int numa_update_cpu_topology(bool cpus_locked)
 			continue;
 		}
 
-		/* Use associativity from first thread for all siblings */
-		vphn_get_associativity(cpu, associativity);
-		new_nid = associativity_to_nid(associativity);
-		if (new_nid < 0 || !node_online(new_nid))
-			new_nid = first_online_node;
+		new_nid = find_cpu_nid(cpu);
 
 		if (new_nid == numa_cpu_lookup_table[cpu]) {
 			cpumask_andnot(&cpu_associativity_changes_mask,

On powerpc systems with shared configurations of CPUs and memory and
memoryless nodes at boot, an event ordering problem was observed on
a SLES12 build platforms with the hot-add of CPUs to the memoryless
nodes.

* The most common error occurred when the memory SLAB driver attempted
  to reference the memoryless node to which a CPU was being added
  before the kernel had finished initializing all of the data structures
  for the CPU and exited 'device_online' under DLPAR/hot-add.

  Normally the memoryless node would be initialized through the call
  path device_online ... arch_update_cpu_topology ... find_cpu_nid
  ...  try_online_node.  This patch ensures that the powerpc node will
  be initialized as early as possible, even if it was memoryless and
  CPU-less at the point when we are trying to hot-add a new CPU to it.

Signed-off-by: Michael Bringmann <redacted>
---
Changes in V7:
  -- Make function find_cpu_nid() externally visible/usable so that
     it may be used from hotplug-cpu.c
---
 arch/powerpc/mm/numa.c                       |    3 ++-
 arch/powerpc/platforms/pseries/hotplug-cpu.c |    3 +++
 2 files changed, 5 insertions(+), 1 deletion(-)

diff --git a/arch/powerpc/mm/numa.c b/arch/powerpc/mm/numa.c
index 163f4cc..d6d4f7c 100644
--- a/arch/powerpc/mm/numa.c
+++ b/arch/powerpc/mm/numa.c

@@ -1310,7 +1310,7 @@ static long vphn_get_associativity(unsigned long cpu,
 	return rc;
 }
 
-static inline int find_cpu_nid(int cpu)
+int find_cpu_nid(int cpu)
 {
 	__be32 associativity[VPHN_ASSOC_BUFSIZE] = {0};
 	int new_nid;

@@ -1343,6 +1343,7 @@ static inline int find_cpu_nid(int cpu)
 #endif
 	}
 
+	printk(KERN_INFO "%s:%d cpu %d nid %d\n", __FUNCTION__, __LINE__, cpu, new_nid);
 	return new_nid;
 }
 

diff --git a/arch/powerpc/platforms/pseries/hotplug-cpu.c b/arch/powerpc/platforms/pseries/hotplug-cpu.c
index a7d14aa7..df8c732 100644
--- a/arch/powerpc/platforms/pseries/hotplug-cpu.c
+++ b/arch/powerpc/platforms/pseries/hotplug-cpu.c

@@ -340,6 +340,8 @@ static void pseries_remove_processor(struct device_node *np)
 	cpu_maps_update_done();
 }
 
+extern int find_cpu_nid(int cpu);
+
 static int dlpar_online_cpu(struct device_node *dn)
 {
 	int rc = 0;

@@ -364,6 +366,7 @@ static int dlpar_online_cpu(struct device_node *dn)
 					!= CPU_STATE_OFFLINE);
 			cpu_maps_update_done();
 			timed_topology_update(1);
+			find_cpu_nid(cpu);
 			rc = device_online(get_cpu_device(cpu));
 			if (rc)
 				goto out;

On 11/16/2017 11:24 AM, Michael Bringmann wrote:

numa.c already has a pr_fmt define, no need to pre-pend "numa:" to the
information message.

-Nathan

On 11/16/2017 11:27 AM, Michael Bringmann wrote:

On powerpc systems which allow 'hot-add' of CPU, it may occur that
the new resources are to be inserted into nodes that were not used
for memory resources at bootup.  Many different configurations of
PowerPC resources may need to be supported depending upon the
environment.  Important characteristics of the nodes and operating
environment include:

* Dedicated vs. shared resources.  Shared resources require

this should be shared CPUs require...since shared CPUs have their
affinity set to node 0 at boot and when hot-added.

  information such as the VPHN hcall for CPU assignment to nodes.
  Associativity decisions made based on dedicated resource rules,
  such as associativity properties in the device tree, may vary
  from decisions made using the values returned by the VPHN hcall.
* memoryless nodes at boot.  Nodes need to be defined as 'possible'
  at boot for operation with other code modules.  Previously, the
  powerpc code would limit the set of possible nodes to those which
  have memory assigned at boot, and were thus online.  Subsequent
  add/remove of CPUs or memory would only work with this subset of
  possible nodes.
* memoryless nodes with CPUs at boot.  Due to the previous restriction
  on nodes, nodes that had CPUs but no memory were being collapsed
  into other nodes that did have memory at boot.  In practice this
  meant that the node assignment presented by the runtime kernel
  differed from the affinity and associativity attributes presented
  by the device tree or VPHN hcalls.  Nodes that might be known to
  the pHyp were not 'possible' in the runtime kernel because they did
  not have memory at boot.

This patch fixes some problems encountered at runtime with
configurations that support memory-less nodes, or that hot-add CPUs
into nodes that are memoryless during system execution after boot.
The problems of interest include,

* Nodes known to powerpc to be memoryless at boot, but to have
  CPUs in them are allowed to be 'possible' and 'online'.  Memory
  allocations for those nodes are taken from another node that does
  have memory until and if memory is hot-added to the node.
* Nodes which have no resources assigned at boot, but which may still
  be referenced subsequently by affinity or associativity attributes,
  are kept in the list of 'possible' nodes for powerpc.  Hot-add of
  memory or CPUs to the system can reference these nodes and bring
  them online instead of redirecting the references to one of the set
  of nodes known to have memory at boot.

Note that this software operates under the context of CPU hotplug.
We are not doing memory hotplug in this code, but rather updating
the kernel's CPU topology (i.e. arch_update_cpu_topology /
numa_update_cpu_topology).  We are initializing a node that may be
used by CPUs or memory before it can be referenced as invalid by a
CPU hotplug operation.  CPU hotplug operations are protected by a
range of APIs including cpu_maps_update_begin/cpu_maps_update_done,
cpus_read/write_lock / cpus_read/write_unlock, device locks, and more.
Memory hotplug operations, including try_online_node, are protected
by mem_hotplug_begin/mem_hotplug_done, device locks, and more.  In
the case of CPUs being hot-added to a previously memoryless node, the
try_online_node operation occurs wholly within the CPU locks with no
overlap.  Using HMC hot-add/hot-remove operations, we have been able
to add and remove CPUs to any possible node without failures.  HMC
operations involve a degree self-serialization, though.

This may be able to be stated as simply saying that cpu hotplug operations
are serialized with the device_hotplug_lock.

Why are you removing the call to setup_node_data() that you added in the previous
patch?

-Nathan

On 11/16/2017 11:28 AM, Michael Bringmann wrote:

This seems like a more likely pr_debug statement.

We don't use the returned node from this call, so I'm not sure why it gets
called. Perhaps its the possible call to try_online_node() that may get called
in find_cpu_nid(), if so perhpas naming the routine something slightly
different would be good, like find_and_online_cpu_nid?

-Nathan

 			rc = device_online(get_cpu_device(cpu));
 			if (rc)
 				goto out;

Nathan Fontenot [off-list ref] writes:

On 11/16/2017 11:24 AM, Michael Bringmann wrote:

...

quoted

diff --git a/arch/powerpc/mm/numa.c b/arch/powerpc/mm/numa.c
index eb604b3..334a1ff 100644
--- a/arch/powerpc/mm/numa.c
+++ b/arch/powerpc/mm/numa.c

@@ -892,6 +892,37 @@ static void __init setup_node_data(int nid, u64 start_pfn, u64 end_pfn)
 	NODE_DATA(nid)->node_spanned_pages = spanned_pages;
 }

+static void __init find_possible_nodes(void)
+{
+	struct device_node *rtas;
+	u32 numnodes, i;
+
+	if (min_common_depth <= 0)
+		return;
+
+	rtas = of_find_node_by_path("/rtas");
+	if (!rtas)
+		return;
+
+	if (of_property_read_u32_index(rtas,
+				"ibm,max-associativity-domains",
+				min_common_depth, &numnodes))
+		goto out;
+
+	pr_info("numa: Nodes = %d (mcd = %d)\n", numnodes,
+		min_common_depth);

numa.c already has a pr_fmt define, no need to pre-pend "numa:" to the
information message.

And in fact no need to print that out here at all, it's covered
elsewhere. So just drop that pr_info() entirely.

cheers

See below.

On 11/20/2017 10:33 AM, Nathan Fontenot wrote:

On 11/16/2017 11:24 AM, Michael Bringmann wrote:

quoted

On powerpc systems which allow 'hot-add' of CPU or memory resources,
it may occur that the new resources are to be inserted into nodes
that were not used for these resources at bootup.  In the kernel,
any node that is used must be defined and initialized.  These empty
nodes may occur when,

* Dedicated vs. shared resources.  Shared resources require
  information such as the VPHN hcall for CPU assignment to nodes.
  Associativity decisions made based on dedicated resource rules,
  such as associativity properties in the device tree, may vary
  from decisions made using the values returned by the VPHN hcall.
* memoryless nodes at boot.  Nodes need to be defined as 'possible'
  at boot for operation with other code modules.  Previously, the
  powerpc code would limit the set of possible nodes to those which
  have memory assigned at boot, and were thus online.  Subsequent
  add/remove of CPUs or memory would only work with this subset of
  possible nodes.
* memoryless nodes with CPUs at boot.  Due to the previous restriction
  on nodes, nodes that had CPUs but no memory were being collapsed
  into other nodes that did have memory at boot.  In practice this
  meant that the node assignment presented by the runtime kernel
  differed from the affinity and associativity attributes presented
  by the device tree or VPHN hcalls.  Nodes that might be known to
  the pHyp were not 'possible' in the runtime kernel because they did
  not have memory at boot.

This patch ensures that sufficient nodes are defined to support
configuration requirements after boot, as well as at boot.  This
patch set fixes a couple of problems.

* Nodes known to powerpc to be memoryless at boot, but to have
  CPUs in them are allowed to be 'possible' and 'online'.  Memory
  allocations for those nodes are taken from another node that does
  have memory until and if memory is hot-added to the node.
* Nodes which have no resources assigned at boot, but which may still
  be referenced subsequently by affinity or associativity attributes,
  are kept in the list of 'possible' nodes for powerpc.  Hot-add of
  memory or CPUs to the system can reference these nodes and bring
  them online instead of redirecting to one of the set of nodes that
  were known to have memory at boot.

This patch extracts the value of the lowest domain level (number of
allocable resources) from the device tree property
"ibm,max-associativity-domains" to use as the maximum number of nodes
to setup as possibly available in the system.  This new setting will
override the instruction,

    nodes_and(node_possible_map, node_possible_map, node_online_map);

presently seen in the function arch/powerpc/mm/numa.c:initmem_init().

If the "ibm,max-associativity-domains" property is not present at boot,
no operation will be performed to define or enable additional nodes, or
enable the above 'nodes_and()'.

Signed-off-by: Michael Bringmann <redacted>
---
Changes in V6:
  -- Remove some node initialization/allocation from boot setup
     to later in runtime to try to limit memory needs early on
  -- Augment descriptive documentation for patch
---
 arch/powerpc/mm/numa.c |   40 +++++++++++++++++++++++++++++++++++++---
 1 file changed, 37 insertions(+), 3 deletions(-)

diff --git a/arch/powerpc/mm/numa.c b/arch/powerpc/mm/numa.c
index eb604b3..334a1ff 100644
--- a/arch/powerpc/mm/numa.c
+++ b/arch/powerpc/mm/numa.c

@@ -892,6 +892,37 @@ static void __init setup_node_data(int nid, u64 start_pfn, u64 end_pfn)
 	NODE_DATA(nid)->node_spanned_pages = spanned_pages;
 }

+static void __init find_possible_nodes(void)
+{
+	struct device_node *rtas;
+	u32 numnodes, i;
+
+	if (min_common_depth <= 0)
+		return;
+
+	rtas = of_find_node_by_path("/rtas");
+	if (!rtas)
+		return;
+
+	if (of_property_read_u32_index(rtas,
+				"ibm,max-associativity-domains",
+				min_common_depth, &numnodes))
+		goto out;
+
+	pr_info("numa: Nodes = %d (mcd = %d)\n", numnodes,
+		min_common_depth);

numa.c already has a pr_fmt define, no need to pre-pend "numa:" to the
information message.

-Nathan

Okay.

quoted

+
+	for (i = 0; i < numnodes; i++) {
+		if (!node_possible(i)) {
+			setup_node_data(i, 0, 0);
+			node_set(i, node_possible_map);
+		}
+	}
+
+out:
+	of_node_put(rtas);
+}
+
 void __init initmem_init(void)
 {
 	int nid, cpu;

@@ -905,12 +936,15 @@ void __init initmem_init(void)
 	memblock_dump_all();

 	/*
-	 * Reduce the possible NUMA nodes to the online NUMA nodes,
-	 * since we do not support node hotplug. This ensures that  we
-	 * lower the maximum NUMA node ID to what is actually present.
+	 * Modify the set of possible NUMA nodes to reflect information
+	 * available about the set of online nodes, and the set of nodes
+	 * that we expect to make use of for this platform's affinity
+	 * calculations.
 	 */
 	nodes_and(node_possible_map, node_possible_map, node_online_map);

+	find_possible_nodes();
+
 	for_each_online_node(nid) {
 		unsigned long start_pfn, end_pfn;

-- 
Michael W. Bringmann
Linux Technology Center
IBM Corporation
Tie-Line  363-5196
External: (512) 286-5196
Cell:       (512) 466-0650
mwb@linux.vnet.ibm.com

See below.

On 11/22/2017 05:17 AM, Michael Ellerman wrote:

Nathan Fontenot [off-list ref] writes:

quoted

On 11/16/2017 11:24 AM, Michael Bringmann wrote:

...

quoted

quoted

diff --git a/arch/powerpc/mm/numa.c b/arch/powerpc/mm/numa.c
index eb604b3..334a1ff 100644
--- a/arch/powerpc/mm/numa.c
+++ b/arch/powerpc/mm/numa.c

@@ -892,6 +892,37 @@ static void __init setup_node_data(int nid, u64 start_pfn, u64 end_pfn)
 	NODE_DATA(nid)->node_spanned_pages = spanned_pages;
 }

+static void __init find_possible_nodes(void)
+{
+	struct device_node *rtas;
+	u32 numnodes, i;
+
+	if (min_common_depth <= 0)
+		return;
+
+	rtas = of_find_node_by_path("/rtas");
+	if (!rtas)
+		return;
+
+	if (of_property_read_u32_index(rtas,
+				"ibm,max-associativity-domains",
+				min_common_depth, &numnodes))
+		goto out;
+
+	pr_info("numa: Nodes = %d (mcd = %d)\n", numnodes,
+		min_common_depth);

numa.c already has a pr_fmt define, no need to pre-pend "numa:" to the
information message.

And in fact no need to print that out here at all, it's covered
elsewhere. So just drop that pr_info() entirely.

cheers

Okay.  pr_info() removed.

-- 
Michael W. Bringmann
Linux Technology Center
IBM Corporation
Tie-Line  363-5196
External: (512) 286-5196
Cell:       (512) 466-0650
mwb@linux.vnet.ibm.com

See below.

On 11/20/2017 10:45 AM, Nathan Fontenot wrote:

On 11/16/2017 11:27 AM, Michael Bringmann wrote:

quoted

On powerpc systems which allow 'hot-add' of CPU, it may occur that
the new resources are to be inserted into nodes that were not used
for memory resources at bootup.  Many different configurations of
PowerPC resources may need to be supported depending upon the
environment.  Important characteristics of the nodes and operating
environment include:

* Dedicated vs. shared resources.  Shared resources require

this should be shared CPUs require...since shared CPUs have their
affinity set to node 0 at boot and when hot-added.

Patch description updated to include this modification.

quoted

  information such as the VPHN hcall for CPU assignment to nodes.
  Associativity decisions made based on dedicated resource rules,
  such as associativity properties in the device tree, may vary
  from decisions made using the values returned by the VPHN hcall.
* memoryless nodes at boot.  Nodes need to be defined as 'possible'
  at boot for operation with other code modules.  Previously, the
  powerpc code would limit the set of possible nodes to those which
  have memory assigned at boot, and were thus online.  Subsequent
  add/remove of CPUs or memory would only work with this subset of
  possible nodes.
* memoryless nodes with CPUs at boot.  Due to the previous restriction
  on nodes, nodes that had CPUs but no memory were being collapsed
  into other nodes that did have memory at boot.  In practice this
  meant that the node assignment presented by the runtime kernel
  differed from the affinity and associativity attributes presented
  by the device tree or VPHN hcalls.  Nodes that might be known to
  the pHyp were not 'possible' in the runtime kernel because they did
  not have memory at boot.

This patch fixes some problems encountered at runtime with
configurations that support memory-less nodes, or that hot-add CPUs
into nodes that are memoryless during system execution after boot.
The problems of interest include,

* Nodes known to powerpc to be memoryless at boot, but to have
  CPUs in them are allowed to be 'possible' and 'online'.  Memory
  allocations for those nodes are taken from another node that does
  have memory until and if memory is hot-added to the node.
* Nodes which have no resources assigned at boot, but which may still
  be referenced subsequently by affinity or associativity attributes,
  are kept in the list of 'possible' nodes for powerpc.  Hot-add of
  memory or CPUs to the system can reference these nodes and bring
  them online instead of redirecting the references to one of the set
  of nodes known to have memory at boot.

Note that this software operates under the context of CPU hotplug.
We are not doing memory hotplug in this code, but rather updating
the kernel's CPU topology (i.e. arch_update_cpu_topology /
numa_update_cpu_topology).  We are initializing a node that may be
used by CPUs or memory before it can be referenced as invalid by a
CPU hotplug operation.  CPU hotplug operations are protected by a
range of APIs including cpu_maps_update_begin/cpu_maps_update_done,
cpus_read/write_lock / cpus_read/write_unlock, device locks, and more.
Memory hotplug operations, including try_online_node, are protected
by mem_hotplug_begin/mem_hotplug_done, device locks, and more.  In
the case of CPUs being hot-added to a previously memoryless node, the
try_online_node operation occurs wholly within the CPU locks with no
overlap.  Using HMC hot-add/hot-remove operations, we have been able
to add and remove CPUs to any possible node without failures.  HMC
operations involve a degree self-serialization, though.

This may be able to be stated as simply saying that cpu hotplug operations
are serialized with the device_hotplug_lock.

quoted

Signed-off-by: Michael Bringmann <redacted>
---
Changes in V6:
  -- Add some needed node initialization to runtime code that maps
     CPUs based on VPHN associativity
  -- Add error checks and alternate recovery for compile flag
     CONFIG_MEMORY_HOTPLUG
  -- Add alternate node selection recovery for !CONFIG_MEMORY_HOTPLUG
  -- Add more information to the patch introductory text
---
 arch/powerpc/mm/numa.c |   51 ++++++++++++++++++++++++++++++++++++++----------
 1 file changed, 40 insertions(+), 11 deletions(-)

diff --git a/arch/powerpc/mm/numa.c b/arch/powerpc/mm/numa.c
index 334a1ff..163f4cc 100644
--- a/arch/powerpc/mm/numa.c
+++ b/arch/powerpc/mm/numa.c

@@ -551,7 +551,7 @@ static int numa_setup_cpu(unsigned long lcpu)
 	nid = of_node_to_nid_single(cpu);

 out_present:
-	if (nid < 0 || !node_online(nid))
+	if (nid < 0 || !node_possible(nid))
 		nid = first_online_node;

 	map_cpu_to_node(lcpu, nid);

@@ -867,7 +867,7 @@ void __init dump_numa_cpu_topology(void)
 }

 /* Initialize NODE_DATA for a node on the local memory */
-static void __init setup_node_data(int nid, u64 start_pfn, u64 end_pfn)
+static void setup_node_data(int nid, u64 start_pfn, u64 end_pfn)
 {
 	u64 spanned_pages = end_pfn - start_pfn;
 	const size_t nd_size = roundup(sizeof(pg_data_t), SMP_CACHE_BYTES);

@@ -913,10 +913,8 @@ static void __init find_possible_nodes(void)
 		min_common_depth);

 	for (i = 0; i < numnodes; i++) {
-		if (!node_possible(i)) {
-			setup_node_data(i, 0, 0);

Why are you removing the call to setup_node_data() that you added in the previous
patch?

-Nathan

Two reasons:

1) Trying to defer the amount of memory allocated to possibly unused nodes,
   at boot, especially since there may now often be 256 nodes to try per
   "ibm,max-associativity-domains".
2) The function 'try_online_node' called from 'find_cpu_nid' provides
   the same functionality, plus a whole lot more.  it seemed unnecessary
   to have both calls.

quoted

+		if (!node_possible(i))
 			node_set(i, node_possible_map);
-		}
 	}

 out:

@@ -1312,6 +1310,42 @@ static long vphn_get_associativity(unsigned long cpu,
 	return rc;
 }

+static inline int find_cpu_nid(int cpu)
+{
+	__be32 associativity[VPHN_ASSOC_BUFSIZE] = {0};
+	int new_nid;
+
+	/* Use associativity from first thread for all siblings */
+	vphn_get_associativity(cpu, associativity);
+	new_nid = associativity_to_nid(associativity);
+	if (new_nid < 0 || !node_possible(new_nid))
+		new_nid = first_online_node;
+
+	if (NODE_DATA(new_nid) == NULL) {
+#ifdef CONFIG_MEMORY_HOTPLUG
+		/*
+		 * Need to ensure that NODE_DATA is initialized
+		 * for a node from available memory (see
+		 * memblock_alloc_try_nid).  If unable to init
+		 * the node, then default to nearest node that
+		 * has memory installed.
+		 */
+		if (try_online_node(new_nid))
+			new_nid = first_online_node;
+#else
+		/*
+		 * Default to using the nearest node that has
+		 * memory installed.  Otherwise, it would be 
+		 * necessary to patch the kernel MM code to deal
+		 * with more memoryless-node error conditions.
+		 */
+		new_nid = first_online_node;
+#endif
+	}
+
+	return new_nid;
+}
+
 /*
  * Update the CPU maps and sysfs entries for a single CPU when its NUMA
  * characteristics change. This function doesn't perform any locking and is

@@ -1379,7 +1413,6 @@ int numa_update_cpu_topology(bool cpus_locked)
 {
 	unsigned int cpu, sibling, changed = 0;
 	struct topology_update_data *updates, *ud;
-	__be32 associativity[VPHN_ASSOC_BUFSIZE] = {0};
 	cpumask_t updated_cpus;
 	struct device *dev;
 	int weight, new_nid, i = 0;

@@ -1417,11 +1450,7 @@ int numa_update_cpu_topology(bool cpus_locked)
 			continue;
 		}

-		/* Use associativity from first thread for all siblings */
-		vphn_get_associativity(cpu, associativity);
-		new_nid = associativity_to_nid(associativity);
-		if (new_nid < 0 || !node_online(new_nid))
-			new_nid = first_online_node;
+		new_nid = find_cpu_nid(cpu);

 		if (new_nid == numa_cpu_lookup_table[cpu]) {
 			cpumask_andnot(&cpu_associativity_changes_mask,

-- 
Michael W. Bringmann
Linux Technology Center
IBM Corporation
Tie-Line  363-5196
External: (512) 286-5196
Cell:       (512) 466-0650
mwb@linux.vnet.ibm.com

See below.

On 11/20/2017 10:50 AM, Nathan Fontenot wrote:

On 11/16/2017 11:28 AM, Michael Bringmann wrote:

quoted

On powerpc systems with shared configurations of CPUs and memory and
memoryless nodes at boot, an event ordering problem was observed on
a SLES12 build platforms with the hot-add of CPUs to the memoryless
nodes.

* The most common error occurred when the memory SLAB driver attempted
  to reference the memoryless node to which a CPU was being added
  before the kernel had finished initializing all of the data structures
  for the CPU and exited 'device_online' under DLPAR/hot-add.

  Normally the memoryless node would be initialized through the call
  path device_online ... arch_update_cpu_topology ... find_cpu_nid
  ...  try_online_node.  This patch ensures that the powerpc node will
  be initialized as early as possible, even if it was memoryless and
  CPU-less at the point when we are trying to hot-add a new CPU to it.

Signed-off-by: Michael Bringmann <redacted>
---
Changes in V7:
  -- Make function find_cpu_nid() externally visible/usable so that
     it may be used from hotplug-cpu.c
---
 arch/powerpc/mm/numa.c                       |    3 ++-
 arch/powerpc/platforms/pseries/hotplug-cpu.c |    3 +++
 2 files changed, 5 insertions(+), 1 deletion(-)

diff --git a/arch/powerpc/mm/numa.c b/arch/powerpc/mm/numa.c
index 163f4cc..d6d4f7c 100644
--- a/arch/powerpc/mm/numa.c
+++ b/arch/powerpc/mm/numa.c

@@ -1310,7 +1310,7 @@ static long vphn_get_associativity(unsigned long cpu,
 	return rc;
 }

-static inline int find_cpu_nid(int cpu)
+int find_cpu_nid(int cpu)
 {
 	__be32 associativity[VPHN_ASSOC_BUFSIZE] = {0};
 	int new_nid;

@@ -1343,6 +1343,7 @@ static inline int find_cpu_nid(int cpu)
 #endif
 	}

+	printk(KERN_INFO "%s:%d cpu %d nid %d\n", __FUNCTION__, __LINE__, cpu, new_nid);

This seems like a more likely pr_debug statement.

Yes.  Changed.

quoted

 	return new_nid;
 }

diff --git a/arch/powerpc/platforms/pseries/hotplug-cpu.c b/arch/powerpc/platforms/pseries/hotplug-cpu.c
index a7d14aa7..df8c732 100644
--- a/arch/powerpc/platforms/pseries/hotplug-cpu.c
+++ b/arch/powerpc/platforms/pseries/hotplug-cpu.c

@@ -340,6 +340,8 @@ static void pseries_remove_processor(struct device_node *np)
 	cpu_maps_update_done();
 }

+extern int find_cpu_nid(int cpu);
+
 static int dlpar_online_cpu(struct device_node *dn)
 {
 	int rc = 0;

@@ -364,6 +366,7 @@ static int dlpar_online_cpu(struct device_node *dn)
 					!= CPU_STATE_OFFLINE);
 			cpu_maps_update_done();
 			timed_topology_update(1);
+			find_cpu_nid(cpu);

We don't use the returned node from this call, so I'm not sure why it gets
called. Perhaps its the possible call to try_online_node() that may get called
in find_cpu_nid(), if so perhpas naming the routine something slightly
different would be good, like find_and_online_cpu_nid?

The function find_cpu_nid() gets called here for its operations to ensure that
the node is online and initialized during the 'device_online()' operation for
the new CPU that immediately follows.  As I tried to explain in the patch
description, I ran into a case during the SLES12 SP3 backport test where failure
to initialize/online the node early on led to a crash in the memory SLAB driver.
I don't know why that driver was being called so early, but ensuring that the
node was setup as early as possible provided the simplest and smallest patch.

As all of the functionality had been placed in the function 'find_cpu_nid' in
'numa.c' during a previous patch, I reused it here.  Though I did not need
the actual node Id during the call dlpar_online_cpu().

Yes, I can provide a separate calling interface named 'find_and_online_cpu_nid'
for the call from 'dlpar_online_cpu'.  They would both do the same operations.

Or I could just use the new name for the current implementation of 'find_cpu_nid',
instead, and migrate that name to the previous patches.

If I don't hear any further remarks on this point, I will implement/migrate
the function name 'find_and_online_cpu_nid' in the next version of this patch.

-Nathan

Michael

quoted

 			rc = device_online(get_cpu_device(cpu));
 			if (rc)
 				goto out;

-- 
Michael W. Bringmann
Linux Technology Center
IBM Corporation
Tie-Line  363-5196
External: (512) 286-5196
Cell:       (512) 466-0650
mwb@linux.vnet.ibm.com