There isn't a simple hardware enumeration to indicate to software that
a system is running with Sub-NUMA Cluster enabled.

Compare the number of NUMA nodes with the number of L3 caches to calculate
the number of Sub-NUMA nodes per L3 cache.

When Sub-NUMA cluster mode is enabled in BIOS setup the RMID counters
are distributed equally between the SNC nodes within each socket.

E.g. if there are 400 RMID counters, and the system is configured with
two SNC nodes per socket, then RMID counter 0..199 are used on SNC node
0 on the socket, and RMID counter 200..399 on SNC node 1.

Handle this by initializing a per-cpu RMID offset value. Use this
to calculate the value to write to the RMID field of the IA32_PQR_ASSOC
MSR during context switch, and also to the IA32_QM_EVTSEL MSR when
reading RMID event values.

N.B. this works well for well-behaved NUMA applications that access
memory predominantly from the local memory node. For applications that
access memory across multiple nodes it may be necessary for the user
to read counters for all SNC nodes on a socket and add the values to
get the actual LLC occupancy or memory bandwidth. Perhaps this isn't
all that different from applications that span across multiple sockets
in a legacy system.

Signed-off-by: Tony Luck <tony.luck@intel.com>
---
 arch/x86/include/asm/resctrl.h        |  4 ++-
 arch/x86/kernel/cpu/resctrl/core.c    | 43 +++++++++++++++++++++++++--
 arch/x86/kernel/cpu/resctrl/monitor.c |  2 +-
 3 files changed, 44 insertions(+), 5 deletions(-)

diff --git a/arch/x86/include/asm/resctrl.h b/arch/x86/include/asm/resctrl.h
index 52788f79786f..59b8afd8c53c 100644
--- a/arch/x86/include/asm/resctrl.h
+++ b/arch/x86/include/asm/resctrl.h
@@ -35,6 +35,8 @@ DECLARE_STATIC_KEY_FALSE(rdt_enable_key);
 DECLARE_STATIC_KEY_FALSE(rdt_alloc_enable_key);
 DECLARE_STATIC_KEY_FALSE(rdt_mon_enable_key);
 
+DECLARE_PER_CPU(int, rmid_offset);
+
 /*
  * __resctrl_sched_in() - Writes the task's CLOSid/RMID to IA32_PQR_MSR
  *
@@ -69,7 +71,7 @@ static void __resctrl_sched_in(void)
 	if (static_branch_likely(&rdt_mon_enable_key)) {
 		tmp = READ_ONCE(current->rmid);
 		if (tmp)
-			rmid = tmp;
+			rmid = tmp + this_cpu_read(rmid_offset);
 	}
 
 	if (closid != state->cur_closid || rmid != state->cur_rmid) {
diff --git a/arch/x86/kernel/cpu/resctrl/core.c b/arch/x86/kernel/cpu/resctrl/core.c
index 53b2ab37af2f..0ff739375e3b 100644
--- a/arch/x86/kernel/cpu/resctrl/core.c
+++ b/arch/x86/kernel/cpu/resctrl/core.c
@@ -16,6 +16,7 @@
 
 #define pr_fmt(fmt)	"resctrl: " fmt
 
+#include <linux/cpu.h>
 #include <linux/slab.h>
 #include <linux/err.h>
 #include <linux/cacheinfo.h>
@@ -484,6 +485,13 @@ static int get_domain_id(int cpu, enum resctrl_scope scope)
 	return get_cpu_cacheinfo_id(cpu, scope);
 }
 
+DEFINE_PER_CPU(int, rmid_offset);
+
+static void set_per_cpu_rmid_offset(int cpu, struct rdt_resource *r)
+{
+	this_cpu_write(rmid_offset, (cpu_to_node(cpu) % snc_ways) * r->num_rmid);
+}
+
 /*
  * domain_add_cpu - Add a cpu to a resource's domain list.
  *
@@ -515,6 +523,8 @@ static void domain_add_cpu(int cpu, struct rdt_resource *r)
 		cpumask_set_cpu(cpu, &d->cpu_mask);
 		if (r->cache.arch_has_per_cpu_cfg)
 			rdt_domain_reconfigure_cdp(r);
+		if (r->mon_capable)
+			set_per_cpu_rmid_offset(cpu, r);
 		return;
 	}
 
@@ -533,9 +543,12 @@ static void domain_add_cpu(int cpu, struct rdt_resource *r)
 		return;
 	}
 
-	if (r->mon_capable && arch_domain_mbm_alloc(r->num_rmid, hw_dom)) {
-		domain_free(hw_dom);
-		return;
+	if (r->mon_capable) {
+		if (arch_domain_mbm_alloc(r->num_rmid, hw_dom)) {
+			domain_free(hw_dom);
+			return;
+		}
+		set_per_cpu_rmid_offset(cpu, r);
 	}
 
 	list_add_tail(&d->list, add_pos);
@@ -845,11 +858,35 @@ static __init bool get_rdt_resources(void)
 	return (rdt_mon_capable || rdt_alloc_capable);
 }
 
+static __init int find_snc_ways(void)
+{
+	unsigned long *node_caches;
+	int cpu, node, ret;
+
+	node_caches = kcalloc(BITS_TO_LONGS(nr_node_ids), sizeof(*node_caches), GFP_KERNEL);
+	if (!node_caches)
+		return 1;
+
+	cpus_read_lock();
+	for_each_node(node) {
+		cpu = cpumask_first(cpumask_of_node(node));
+		set_bit(get_cpu_cacheinfo_id(cpu, 3), node_caches);
+	}
+	cpus_read_unlock();
+
+	ret = nr_node_ids / bitmap_weight(node_caches, nr_node_ids);
+	kfree(node_caches);
+
+	return ret;
+}
+
 static __init void rdt_init_res_defs_intel(void)
 {
 	struct rdt_hw_resource *hw_res;
 	struct rdt_resource *r;
 
+	snc_ways = find_snc_ways();
+
 	for_each_rdt_resource(r) {
 		hw_res = resctrl_to_arch_res(r);
 
diff --git a/arch/x86/kernel/cpu/resctrl/monitor.c b/arch/x86/kernel/cpu/resctrl/monitor.c
index 3fc63aa68130..bd5ec348d925 100644
--- a/arch/x86/kernel/cpu/resctrl/monitor.c
+++ b/arch/x86/kernel/cpu/resctrl/monitor.c
@@ -160,7 +160,7 @@ static int __rmid_read(u32 rmid, enum resctrl_event_id eventid, u64 *val)
 	 * IA32_QM_CTR.Error (bit 63) and IA32_QM_CTR.Unavailable (bit 62)
 	 * are error bits.
 	 */
-	wrmsr(MSR_IA32_QM_EVTSEL, eventid, rmid);
+	wrmsr(MSR_IA32_QM_EVTSEL, eventid, rmid + this_cpu_read(rmid_offset));
 	rdmsrl(MSR_IA32_QM_CTR, msr_val);
 
 	if (msr_val & RMID_VAL_ERROR)
-- 
2.39.1

Hi Tony,


On 1/26/23 12:41, Tony Luck wrote:
> There isn't a simple hardware enumeration to indicate to software that
> a system is running with Sub-NUMA Cluster enabled.
> 
> Compare the number of NUMA nodes with the number of L3 caches to calculate
> the number of Sub-NUMA nodes per L3 cache.
> 
> When Sub-NUMA cluster mode is enabled in BIOS setup the RMID counters
> are distributed equally between the SNC nodes within each socket.
> 
> E.g. if there are 400 RMID counters, and the system is configured with
> two SNC nodes per socket, then RMID counter 0..199 are used on SNC node
> 0 on the socket, and RMID counter 200..399 on SNC node 1.
> 
> Handle this by initializing a per-cpu RMID offset value. Use this
> to calculate the value to write to the RMID field of the IA32_PQR_ASSOC
> MSR during context switch, and also to the IA32_QM_EVTSEL MSR when
> reading RMID event values.
> 
> N.B. this works well for well-behaved NUMA applications that access
> memory predominantly from the local memory node. For applications that
> access memory across multiple nodes it may be necessary for the user
> to read counters for all SNC nodes on a socket and add the values to
> get the actual LLC occupancy or memory bandwidth. Perhaps this isn't
> all that different from applications that span across multiple sockets
> in a legacy system.
> 
> Signed-off-by: Tony Luck <tony.luck@intel.com>
> ---
>  arch/x86/include/asm/resctrl.h        |  4 ++-
>  arch/x86/kernel/cpu/resctrl/core.c    | 43 +++++++++++++++++++++++++--
>  arch/x86/kernel/cpu/resctrl/monitor.c |  2 +-
>  3 files changed, 44 insertions(+), 5 deletions(-)
> 
> diff --git a/arch/x86/include/asm/resctrl.h b/arch/x86/include/asm/resctrl.h
> index 52788f79786f..59b8afd8c53c 100644
> --- a/arch/x86/include/asm/resctrl.h
> +++ b/arch/x86/include/asm/resctrl.h
> @@ -35,6 +35,8 @@ DECLARE_STATIC_KEY_FALSE(rdt_enable_key);
>  DECLARE_STATIC_KEY_FALSE(rdt_alloc_enable_key);
>  DECLARE_STATIC_KEY_FALSE(rdt_mon_enable_key);
>  
> +DECLARE_PER_CPU(int, rmid_offset);
> +
>  /*
>   * __resctrl_sched_in() - Writes the task's CLOSid/RMID to IA32_PQR_MSR
>   *
> @@ -69,7 +71,7 @@ static void __resctrl_sched_in(void)
>  	if (static_branch_likely(&rdt_mon_enable_key)) {
>  		tmp = READ_ONCE(current->rmid);
>  		if (tmp)
> -			rmid = tmp;
> +			rmid = tmp + this_cpu_read(rmid_offset);
>  	}
>  
>  	if (closid != state->cur_closid || rmid != state->cur_rmid) {
> diff --git a/arch/x86/kernel/cpu/resctrl/core.c b/arch/x86/kernel/cpu/resctrl/core.c
> index 53b2ab37af2f..0ff739375e3b 100644
> --- a/arch/x86/kernel/cpu/resctrl/core.c
> +++ b/arch/x86/kernel/cpu/resctrl/core.c
> @@ -16,6 +16,7 @@
>  
>  #define pr_fmt(fmt)	"resctrl: " fmt
>  
> +#include <linux/cpu.h>
>  #include <linux/slab.h>
>  #include <linux/err.h>
>  #include <linux/cacheinfo.h>
> @@ -484,6 +485,13 @@ static int get_domain_id(int cpu, enum resctrl_scope scope)
>  	return get_cpu_cacheinfo_id(cpu, scope);
>  }
>  
> +DEFINE_PER_CPU(int, rmid_offset);
> +
> +static void set_per_cpu_rmid_offset(int cpu, struct rdt_resource *r)
> +{
> +	this_cpu_write(rmid_offset, (cpu_to_node(cpu) % snc_ways) * r->num_rmid);
> +}
> +
>  /*
>   * domain_add_cpu - Add a cpu to a resource's domain list.
>   *
> @@ -515,6 +523,8 @@ static void domain_add_cpu(int cpu, struct rdt_resource *r)
>  		cpumask_set_cpu(cpu, &d->cpu_mask);
>  		if (r->cache.arch_has_per_cpu_cfg)
>  			rdt_domain_reconfigure_cdp(r);
> +		if (r->mon_capable)
> +			set_per_cpu_rmid_offset(cpu, r);
>  		return;
>  	}
>  
> @@ -533,9 +543,12 @@ static void domain_add_cpu(int cpu, struct rdt_resource *r)
>  		return;
>  	}
>  
> -	if (r->mon_capable && arch_domain_mbm_alloc(r->num_rmid, hw_dom)) {
> -		domain_free(hw_dom);
> -		return;
> +	if (r->mon_capable) {
> +		if (arch_domain_mbm_alloc(r->num_rmid, hw_dom)) {
> +			domain_free(hw_dom);
> +			return;
> +		}
> +		set_per_cpu_rmid_offset(cpu, r);
>  	}
>  
>  	list_add_tail(&d->list, add_pos);
> @@ -845,11 +858,35 @@ static __init bool get_rdt_resources(void)
>  	return (rdt_mon_capable || rdt_alloc_capable);
>  }
>  
> +static __init int find_snc_ways(void)
> +{
> +	unsigned long *node_caches;
> +	int cpu, node, ret;
> +
> +	node_caches = kcalloc(BITS_TO_LONGS(nr_node_ids), sizeof(*node_caches), GFP_KERNEL);
> +	if (!node_caches)
> +		return 1;
> +
> +	cpus_read_lock();
> +	for_each_node(node) {
> +		cpu = cpumask_first(cpumask_of_node(node));
> +		set_bit(get_cpu_cacheinfo_id(cpu, 3), node_caches);
> +	}
> +	cpus_read_unlock();
> +
> +	ret = nr_node_ids / bitmap_weight(node_caches, nr_node_ids);
> +	kfree(node_caches);
> +
> +	return ret;
> +}
> +
>  static __init void rdt_init_res_defs_intel(void)
>  {
>  	struct rdt_hw_resource *hw_res;
>  	struct rdt_resource *r;
>  
> +	snc_ways = find_snc_ways();
> +
>  	for_each_rdt_resource(r) {
>  		hw_res = resctrl_to_arch_res(r);
>  
> diff --git a/arch/x86/kernel/cpu/resctrl/monitor.c b/arch/x86/kernel/cpu/resctrl/monitor.c
> index 3fc63aa68130..bd5ec348d925 100644
> --- a/arch/x86/kernel/cpu/resctrl/monitor.c
> +++ b/arch/x86/kernel/cpu/resctrl/monitor.c
> @@ -160,7 +160,7 @@ static int __rmid_read(u32 rmid, enum resctrl_event_id eventid, u64 *val)
>  	 * IA32_QM_CTR.Error (bit 63) and IA32_QM_CTR.Unavailable (bit 62)
>  	 * are error bits.
>  	 */
> -	wrmsr(MSR_IA32_QM_EVTSEL, eventid, rmid);
> +	wrmsr(MSR_IA32_QM_EVTSEL, eventid, rmid + this_cpu_read(rmid_offset));

I am thinking loud here.
When a new monitor group is created, new RMID is assigned. This is done by
alloc_rmid. It does not know about the rmid_offset details. This will
allocate the one of the free RMIDs.

When CPUs are assigned to the group, then per cpu  pqr_state is updated.
At that point, this RMID becomes default_rmid for that cpu.

But CPUs can be assigned from two different Sub-NUMA nodes.

Considering same example you mentioned.

E.g. in 2-way Sub-NUMA cluster with 200 RMID counters there are only
100 available counters to the resctrl code. When running on the first
SNC node RMID values 0..99 are used as before. But when running on the
second node, a task that is assigned resctrl rmid=10 must load 10+100
into IA32_PQR_ASSOC to use RMID counter 110.

#mount -t resctrl resctrl /sys/fs/resctrl/
#cd /sys/fs/resctrl/
#mkdir test  (Lets say RMID 1 is allocated)
#cd test
#echo 1 > cpus_list
#echo 101 > cpus_list

In this case, the following code may run on two different RMIDs even
though it was intended to run on same RMID.

wrmsr(MSR_IA32_QM_EVTSEL, eventid, rmid + this_cpu_read(rmid_offset));

Have you thought of this problem?

Thanks
Babu

On Tue, Feb 28, 2023 at 01:51:32PM -0600, Moger, Babu wrote:
> I am thinking loud here.
> When a new monitor group is created, new RMID is assigned. This is done by
> alloc_rmid. It does not know about the rmid_offset details. This will
> allocate the one of the free RMIDs.
> 
> When CPUs are assigned to the group, then per cpu  pqr_state is updated.
> At that point, this RMID becomes default_rmid for that cpu.
> 
> But CPUs can be assigned from two different Sub-NUMA nodes.
> 
> Considering same example you mentioned.
> 
> E.g. in 2-way Sub-NUMA cluster with 200 RMID counters there are only
> 100 available counters to the resctrl code. When running on the first
> SNC node RMID values 0..99 are used as before. But when running on the
> second node, a task that is assigned resctrl rmid=10 must load 10+100
> into IA32_PQR_ASSOC to use RMID counter 110.
> 
> #mount -t resctrl resctrl /sys/fs/resctrl/
> #cd /sys/fs/resctrl/
> #mkdir test  (Lets say RMID 1 is allocated)
> #cd test
> #echo 1 > cpus_list
> #echo 101 > cpus_list
> 
> In this case, the following code may run on two different RMIDs even
> though it was intended to run on same RMID.
> 
> wrmsr(MSR_IA32_QM_EVTSEL, eventid, rmid + this_cpu_read(rmid_offset));
> 
> Have you thought of this problem?

Now I've thought about this. I don't think it is a problem.

With SNC enabled for two nodes per socket the available RMIDs
are divided between the SNC nodes, but are for some purposes
numbered [0 .. N/2) but in some cases must be viewed as two
separate sets [0 .. N/2) on the first node and [N/2 .. N) on
the second.

In your example RMID 1 is assigned to the group and you have
one CPU from each node in the group. Processes on CPU1 will
load IA32_PQR_ASSOC.RMID = 1, while processes on CPU101 will
set IA32_PQR_ASSOC.RMID = 101. So counts of memory bandwidth
and cache occupancy will be in two different physical RMID
counters.

To read these back the user needs to lookup which $node each CPU
belongs to and then read from the appropriate
mon_data/mon_L3_$node/{llc_occupancy,mbm_local_bytes,mbm_total_bytes}
file.

$ cat mon_data/mon_L3_00/llc_occupancy # reads RMID=1
$ cat mon_data/mon_L3_01/llc_occupancy # reads RMID=101

-Tony

Hi Tony,

On Thu, Jan 26, 2023 at 7:42 PM Tony Luck <tony.luck@intel.com> wrote:
> +static __init int find_snc_ways(void)
> +{
> +       unsigned long *node_caches;
> +       int cpu, node, ret;
> +
> +       node_caches = kcalloc(BITS_TO_LONGS(nr_node_ids), sizeof(*node_caches), GFP_KERNEL);
> +       if (!node_caches)
> +               return 1;
> +
> +       cpus_read_lock();
> +       for_each_node(node) {

Someone tried this patch on a machine with a CPU-less node...

We need to check for this:

+               if (cpumask_empty(cpumask_of_node(node)))
+                       continue;

> +               cpu = cpumask_first(cpumask_of_node(node));
> +               set_bit(get_cpu_cacheinfo_id(cpu, 3), node_caches);
> +       }
> +       cpus_read_unlock();

Thanks!
-Peter

On Mon, Feb 27, 2023 at 02:30:38PM +0100, Peter Newman wrote:
> Hi Tony,
> 
> On Thu, Jan 26, 2023 at 7:42 PM Tony Luck <tony.luck@intel.com> wrote:
> > +static __init int find_snc_ways(void)
> > +{
> > +       unsigned long *node_caches;
> > +       int cpu, node, ret;
> > +
> > +       node_caches = kcalloc(BITS_TO_LONGS(nr_node_ids), sizeof(*node_caches), GFP_KERNEL);
> > +       if (!node_caches)
> > +               return 1;
> > +
> > +       cpus_read_lock();
> > +       for_each_node(node) {
> 
> Someone tried this patch on a machine with a CPU-less node...
> 
> We need to check for this:
> 
> +               if (cpumask_empty(cpumask_of_node(node)))
> +                       continue;
> 
> > +               cpu = cpumask_first(cpumask_of_node(node));
> > +               set_bit(get_cpu_cacheinfo_id(cpu, 3), node_caches);
> > +       }
> > +       cpus_read_unlock();

Peter,

Tell me more about your CPU-less nodes.  Your fix avoids a bad
pointer reference (because cpumask_first() returns cpu >= nr_cpu_ids
for an empty bitmask).

But now I'm worried about whether I have the right values in the
formula:

	nr_node_ids / bitmap_weight(node_caches, nr_node_ids);

This fix avoids counting the L3 from a non-existent CPU, but still
counts the node in the numerator.

Is your CPU-less node a full (non-SNC) node? Like this:

          Socket 0                 Socket 1
    +--------------------+  +--------------------+
    |          .         |  |          .         |
    |  SNC 0.0 . SNC 0.1 |  |  zero    . zero    |
    |          .         |  |  CPUs    . CPUs    |
    |          .         |  |          .         |
    |          .         |  |          .         |
    +--------------------+  +--------------------+
    |      L3 Cache      |  |      L3 Cache      |
    +--------------------+  +--------------------+

I could fix this case by counting how many CPU-less
nodes I find, and reducing the numerator (the denominator
didn't count the L3 cache from socket 1 because there
are no CPUs there)

	(nr_node_ids - n_empty_nodes) / bitmap_weight(node_caches, nr_node_ids);

	=> 2 / 1

But that won't work if your CPU-less node is an SNC node
and the other SNC node in the same socket does have some
CPUs:

          Socket 0                 Socket 1
    +--------------------+  +--------------------+
    |          .         |  |          .         |
    |  SNC 0.0 . SNC 0.1 |  |  zero    . SNC 1.1 |
    |          .         |  |  CPUs    .         |
    |          .         |  |          .         |
    |          .         |  |          .         |
    +--------------------+  +--------------------+
    |      L3 Cache      |  |      L3 Cache      |
    +--------------------+  +--------------------+

This would get 3 / 2 ... i.e. I should still count the
empty node because its cache was counted by its SNC
buddy.

-Tony

Hi Tony,

On Fri, Mar 10, 2023 at 6:30 PM Tony Luck <tony.luck@intel.com> wrote:
> Tell me more about your CPU-less nodes. Your fix avoids a bad
> pointer reference (because cpumask_first() returns cpu >= nr_cpu_ids
> for an empty bitmask).
>
> But now I'm worried about whether I have the right values in the
> formula:
>
> nr_node_ids / bitmap_weight(node_caches, nr_node_ids);
>
> This fix avoids counting the L3 from a non-existent CPU, but still
> counts the node in the numerator.
>
> Is your CPU-less node a full (non-SNC) node? Like this:
>
> Socket 0 Socket 1
> +--------------------+ +--------------------+
> | . | | . |
> | SNC 0.0 . SNC 0.1 | | zero . zero |
> | . | | CPUs . CPUs |
> | . | | . |
> | . | | . |
> +--------------------+ +--------------------+
> | L3 Cache | | L3 Cache |
> +--------------------+ +--------------------+

In the case I saw, the nodes were AEP DIMMs, so all-memory nodes.

Browsing sysfs, they are listed in has_memory, but not
has_normal_memory or has_cpu.

I imagine CXL.mem would have similar issues.

-Peter

> In the case I saw, the nodes were AEP DIMMs, so all-memory nodes.

Peter,

Thanks. This helps a lot.

Ok. I will add code to count the number of memory only nodes and subtract
that from the numerator of "nodes / L3-caches".

I'll ignore the weird case of a memory-only SNC node when other SNC
nodes on the same socket do have CPUs until such time as someone
convinces me that there is a real-world reason to enable SNC and then
disable the CPUs in one node. It would seem much better to keep SNC
turned off so that the remaining CPUs on the socket get access to all
of the L3.

-Tony