[v11,6/8] x86/resctrl: Introduce snc_nodes_per_l3_cache

Message ID 20231109230915.73600-7-tony.luck@intel.com
State New
Headers
Series Add support for Sub-NUMA cluster (SNC) systems |

Commit Message

Luck, Tony Nov. 9, 2023, 11:09 p.m. UTC
  Intel Sub-NUMA Cluster (SNC) is a feature that subdivides the CPU cores
and memory controllers on a socket into two or more groups. These are
presented to the operating system as NUMA nodes.

This may enable some workloads to have slightly lower latency to memory
as the memory controller(s) in an SNC node are electrically closer to the
CPU cores on that SNC node. This cost may be offset by lower bandwidth
since the memory accesses for each core can only be interleaved between
the memory controllers on the same SNC node.

Resctrl monitoring on an Intel system depends upon attaching RMIDs to tasks
to track L3 cache occupancy and memory bandwidth. There is an MSR that
controls how the RMIDs are shared between SNC nodes.

The default mode divides them numerically. E.g. when there are two SNC
nodes on a socket the lower number half of the RMIDs are given to the
first node, the remainder to the second node. This would be difficult
to use with the Linux resctrl interface as specific RMID values assigned
to resctrl groups are not visible to users.

The other mode divides the RMIDs and renumbers the ones on the second
SNC node to start from zero.

Even with this renumbering SNC mode requires several changes in resctrl
behavior for correct operation.

Add a global integer "snc_nodes_per_l3_cache" that will show how many
SNC nodes share each L3 cache. When this is "1", SNC mode is either
not implemented, or not enabled, but all places that need to check
it are updated to take appropriate action when SNC mode is enabled.

Code that needs to take action when SNC is enabled is:
1) The number of logical RMIDs per L3 cache available for use is the
   number of physical RMIDs divided by the number of SNC nodes.
2) Likewise the "mon_scale" value must be divided by the number of SNC
   nodes.
3) The RMID renumbering operates when using the value from the
   IA32_PQR_ASSOC MSR to count accesses by a task. When reading an RMID
   counter, code must adjust from the logical RMID used to the physical
   RMID value for the SNC node that it wishes to read and load the
   adjusted value into the IA32_QM_EVTSEL MSR.
4) The L3 cache is divided between the SNC nodes. So the value
   reported in the resctrl "size" file is divided by the number of SNC
   nodes because the effective amount of cache that can be allocated
   is reduced by that factor.
5) The "-o mba_MBps" mount option must be disabled in SNC mode
   because the monitoring is being done per SNC node, while the
   bandwidth allocation is still done at the L3 cache scope.
   Trying to use this feedback loop might result in contradictory
   changes to the throttling level coming from each of the SNC
   node bandwidth measurements.

Reviewed-by: Peter Newman <peternewman@google.com>
Signed-off-by: Tony Luck <tony.luck@intel.com>
---
 arch/x86/kernel/cpu/resctrl/internal.h |  2 ++
 arch/x86/kernel/cpu/resctrl/core.c     |  6 ++++++
 arch/x86/kernel/cpu/resctrl/monitor.c  | 16 +++++++++++++---
 arch/x86/kernel/cpu/resctrl/rdtgroup.c |  5 +++--
 4 files changed, 24 insertions(+), 5 deletions(-)
  

Comments

Reinette Chatre Nov. 20, 2023, 10:23 p.m. UTC | #1
Hi Tony,

On 11/9/2023 3:09 PM, Tony Luck wrote:
> Intel Sub-NUMA Cluster (SNC) is a feature that subdivides the CPU cores
> and memory controllers on a socket into two or more groups. These are
> presented to the operating system as NUMA nodes.
> 
> This may enable some workloads to have slightly lower latency to memory
> as the memory controller(s) in an SNC node are electrically closer to the
> CPU cores on that SNC node. This cost may be offset by lower bandwidth
> since the memory accesses for each core can only be interleaved between
> the memory controllers on the same SNC node.
> 
> Resctrl monitoring on an Intel system depends upon attaching RMIDs to tasks
> to track L3 cache occupancy and memory bandwidth. There is an MSR that
> controls how the RMIDs are shared between SNC nodes.
> 
> The default mode divides them numerically. E.g. when there are two SNC
> nodes on a socket the lower number half of the RMIDs are given to the
> first node, the remainder to the second node. This would be difficult
> to use with the Linux resctrl interface as specific RMID values assigned
> to resctrl groups are not visible to users.
> 
> The other mode divides the RMIDs and renumbers the ones on the second
> SNC node to start from zero.
> 
> Even with this renumbering SNC mode requires several changes in resctrl
> behavior for correct operation.
> 
> Add a global integer "snc_nodes_per_l3_cache" that will show how many
> SNC nodes share each L3 cache. When this is "1", SNC mode is either
> not implemented, or not enabled, but all places that need to check
> it are updated to take appropriate action when SNC mode is enabled.
> 
> Code that needs to take action when SNC is enabled is:
> 1) The number of logical RMIDs per L3 cache available for use is the
>    number of physical RMIDs divided by the number of SNC nodes.
> 2) Likewise the "mon_scale" value must be divided by the number of SNC
>    nodes.
> 3) The RMID renumbering operates when using the value from the
>    IA32_PQR_ASSOC MSR to count accesses by a task. When reading an RMID
>    counter, code must adjust from the logical RMID used to the physical
>    RMID value for the SNC node that it wishes to read and load the
>    adjusted value into the IA32_QM_EVTSEL MSR.
> 4) The L3 cache is divided between the SNC nodes. So the value
>    reported in the resctrl "size" file is divided by the number of SNC
>    nodes because the effective amount of cache that can be allocated
>    is reduced by that factor.
> 5) The "-o mba_MBps" mount option must be disabled in SNC mode
>    because the monitoring is being done per SNC node, while the
>    bandwidth allocation is still done at the L3 cache scope.
>    Trying to use this feedback loop might result in contradictory
>    changes to the throttling level coming from each of the SNC
>    node bandwidth measurements.
> 

The latter part of this changelog stopped being in imperative mood.
To reduce confusion I slightly reworked it below to address the
parts I noticed. I often get it wrong myself so please check again.

	Add a global integer "snc_nodes_per_l3_cache" that shows how many
	SNC nodes share each L3 cache. When "snc_nodes_per_l3_cache" is "1",
	SNC mode is either not implemented, or not enabled.

	Update all places to take appropriate action when SNC mode is enabled:
	1) The number of logical RMIDs per L3 cache available for use is the
	   number of physical RMIDs divided by the number of SNC nodes.
	2) Likewise the "mon_scale" value must be divided by the number of SNC
	   nodes.
	3) The RMID renumbering operates when using the value from the
	   IA32_PQR_ASSOC MSR to count accesses by a task. When reading an RMID
	   counter, adjust from the logical RMID to the physical
	   RMID value for the SNC node that it wishes to read and load the
	   adjusted value into the IA32_QM_EVTSEL MSR.
	4) Divide the L3 cache between the SNC nodes. Divide the value
	   reported in the resctrl "size" file by the number of SNC
	   nodes because the effective amount of cache that can be allocated
	   is reduced by that factor.
	5) Disable the "-o mba_MBps" mount option in SNC mode
	   because the monitoring is being done per SNC node, while the
	   bandwidth allocation is still done at the L3 cache scope.
	   Trying to use this feedback loop might result in contradictory
	   changes to the throttling level coming from each of the SNC
	   node bandwidth measurements.

> Reviewed-by: Peter Newman <peternewman@google.com>
> Signed-off-by: Tony Luck <tony.luck@intel.com>
> ---

(same comment as previous patch about commit tag ordering)

Reviewed-by: Reinette Chatre <reinette.chatre@intel.com>

Reinette
  

Patch

diff --git a/arch/x86/kernel/cpu/resctrl/internal.h b/arch/x86/kernel/cpu/resctrl/internal.h
index ce3a70657842..e7a75a439c16 100644
--- a/arch/x86/kernel/cpu/resctrl/internal.h
+++ b/arch/x86/kernel/cpu/resctrl/internal.h
@@ -446,6 +446,8 @@  DECLARE_STATIC_KEY_FALSE(rdt_alloc_enable_key);
 
 extern struct dentry *debugfs_resctrl;
 
+extern unsigned int snc_nodes_per_l3_cache;
+
 enum resctrl_res_level {
 	RDT_RESOURCE_L3,
 	RDT_RESOURCE_L2,
diff --git a/arch/x86/kernel/cpu/resctrl/core.c b/arch/x86/kernel/cpu/resctrl/core.c
index 0f2b5ee429b0..f10b68b45342 100644
--- a/arch/x86/kernel/cpu/resctrl/core.c
+++ b/arch/x86/kernel/cpu/resctrl/core.c
@@ -48,6 +48,12 @@  int max_name_width, max_data_width;
  */
 bool rdt_alloc_capable;
 
+/*
+ * Number of SNC nodes that share each L3 cache.  Default is 1 for
+ * systems that do not support SNC, or have SNC disabled.
+ */
+unsigned int snc_nodes_per_l3_cache = 1;
+
 static void
 mba_wrmsr_intel(struct rdt_ctrl_domain *d, struct msr_param *m,
 		struct rdt_resource *r);
diff --git a/arch/x86/kernel/cpu/resctrl/monitor.c b/arch/x86/kernel/cpu/resctrl/monitor.c
index 4e145f5620b0..30b7c3b9b517 100644
--- a/arch/x86/kernel/cpu/resctrl/monitor.c
+++ b/arch/x86/kernel/cpu/resctrl/monitor.c
@@ -148,8 +148,18 @@  static inline struct rmid_entry *__rmid_entry(u32 rmid)
 
 static int __rmid_read(u32 rmid, enum resctrl_event_id eventid, u64 *val)
 {
+	struct rdt_resource *r = &rdt_resources_all[RDT_RESOURCE_L3].r_resctrl;
+	int cpu = smp_processor_id();
+	int rmid_offset = 0;
 	u64 msr_val;
 
+	/*
+	 * When SNC mode is on, need to compute the offset to read the
+	 * physical RMID counter for the node to which this CPU belongs.
+	 */
+	if (snc_nodes_per_l3_cache > 1)
+		rmid_offset = (cpu_to_node(cpu) % snc_nodes_per_l3_cache) * r->num_rmid;
+
 	/*
 	 * As per the SDM, when IA32_QM_EVTSEL.EvtID (bits 7:0) is configured
 	 * with a valid event code for supported resource type and the bits
@@ -158,7 +168,7 @@  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 + rmid_offset);
 	rdmsrl(MSR_IA32_QM_CTR, msr_val);
 
 	if (msr_val & RMID_VAL_ERROR)
@@ -783,8 +793,8 @@  int __init rdt_get_mon_l3_config(struct rdt_resource *r)
 	int ret;
 
 	resctrl_rmid_realloc_limit = boot_cpu_data.x86_cache_size * 1024;
-	hw_res->mon_scale = boot_cpu_data.x86_cache_occ_scale;
-	r->num_rmid = boot_cpu_data.x86_cache_max_rmid + 1;
+	hw_res->mon_scale = boot_cpu_data.x86_cache_occ_scale / snc_nodes_per_l3_cache;
+	r->num_rmid = (boot_cpu_data.x86_cache_max_rmid + 1) / snc_nodes_per_l3_cache;
 	hw_res->mbm_width = MBM_CNTR_WIDTH_BASE;
 
 	if (mbm_offset > 0 && mbm_offset <= MBM_CNTR_WIDTH_OFFSET_MAX)
diff --git a/arch/x86/kernel/cpu/resctrl/rdtgroup.c b/arch/x86/kernel/cpu/resctrl/rdtgroup.c
index 21bbd832f3f2..79d57dade568 100644
--- a/arch/x86/kernel/cpu/resctrl/rdtgroup.c
+++ b/arch/x86/kernel/cpu/resctrl/rdtgroup.c
@@ -1425,7 +1425,7 @@  unsigned int rdtgroup_cbm_to_size(struct rdt_resource *r,
 		}
 	}
 
-	return size;
+	return size / snc_nodes_per_l3_cache;
 }
 
 /*
@@ -2298,7 +2298,8 @@  static bool supports_mba_mbps(void)
 	struct rdt_resource *r = &rdt_resources_all[RDT_RESOURCE_MBA].r_resctrl;
 
 	return (is_mbm_local_enabled() &&
-		r->alloc_capable && is_mba_linear());
+		r->alloc_capable && is_mba_linear() &&
+		snc_nodes_per_l3_cache == 1);
 }
 
 /*