[v10,7/8] x86/resctrl: Sub NUMA Cluster detection and enable
Commit Message
There isn't a simple hardware bit that indicates whether a CPU is
running in Sub NUMA Cluster (SNC) mode. Infer the state by comparing
the ratio of NUMA nodes to L3 cache instances.
When SNC mode is detected, reconfigure the RMID counters by updating
the MSR_RMID_SNC_CONFIG MSR on each socket as CPUs are seen.
Clearing bit zero of the MSR divides the RMIDs and renumbers the ones
on the second SNC node to start from zero.
Signed-off-by: Tony Luck <tony.luck@intel.com>
---
Changes since v9
Expand h/w to hardware (commit and code comments)
Remove "earlier commit" reference
s/counnter/counter/
Check for offline CPUs and warn user SNC detection may be broken.
arch/x86/include/asm/msr-index.h | 1 +
arch/x86/kernel/cpu/resctrl/core.c | 100 ++++++++++++++++++++++++++++-
2 files changed, 99 insertions(+), 2 deletions(-)
Comments
Hi Tony,
On 10/31/2023 2:17 PM, Tony Luck wrote:
> There isn't a simple hardware bit that indicates whether a CPU is
> running in Sub NUMA Cluster (SNC) mode. Infer the state by comparing
> the ratio of NUMA nodes to L3 cache instances.
>
> When SNC mode is detected, reconfigure the RMID counters by updating
> the MSR_RMID_SNC_CONFIG MSR on each socket as CPUs are seen.
>
> Clearing bit zero of the MSR divides the RMIDs and renumbers the ones
> on the second SNC node to start from zero.
>
> Signed-off-by: Tony Luck <tony.luck@intel.com>
> ---
> Changes since v9
> Expand h/w to hardware (commit and code comments)
> Remove "earlier commit" reference
> s/counnter/counter/
> Check for offline CPUs and warn user SNC detection may be broken.
>
> arch/x86/include/asm/msr-index.h | 1 +
> arch/x86/kernel/cpu/resctrl/core.c | 100 ++++++++++++++++++++++++++++-
> 2 files changed, 99 insertions(+), 2 deletions(-)
>
> diff --git a/arch/x86/include/asm/msr-index.h b/arch/x86/include/asm/msr-index.h
> index e3fa9cecd599..4285a5ee81fe 100644
> --- a/arch/x86/include/asm/msr-index.h
> +++ b/arch/x86/include/asm/msr-index.h
> @@ -1109,6 +1109,7 @@
> #define MSR_IA32_QM_CTR 0xc8e
> #define MSR_IA32_PQR_ASSOC 0xc8f
> #define MSR_IA32_L3_CBM_BASE 0xc90
> +#define MSR_RMID_SNC_CONFIG 0xca0
> #define MSR_IA32_L2_CBM_BASE 0xd10
> #define MSR_IA32_MBA_THRTL_BASE 0xd50
>
> diff --git a/arch/x86/kernel/cpu/resctrl/core.c b/arch/x86/kernel/cpu/resctrl/core.c
> index 97d2a5a7dd41..034f9797e1fb 100644
> --- a/arch/x86/kernel/cpu/resctrl/core.c
> +++ b/arch/x86/kernel/cpu/resctrl/core.c
> @@ -16,11 +16,14 @@
>
> #define pr_fmt(fmt) "resctrl: " fmt
>
> +#include <linux/cpu.h>
> #include <linux/slab.h>
> #include <linux/err.h>
> #include <linux/cacheinfo.h>
> #include <linux/cpuhotplug.h>
> +#include <linux/mod_devicetable.h>
>
> +#include <asm/cpu_device_id.h>
> #include <asm/intel-family.h>
> #include <asm/resctrl.h>
> #include "internal.h"
> @@ -184,10 +187,10 @@ bool is_mba_sc(struct rdt_resource *r)
>
> /*
> * rdt_get_mb_table() - get a mapping of bandwidth(b/w) percentage values
> - * exposed to user interface and the h/w understandable delay values.
> + * exposed to user interface and the hardware understandable delay values.
> *
> * The non-linear delay values have the granularity of power of two
> - * and also the h/w does not guarantee a curve for configured delay
> + * and also the hardware does not guarantee a curve for configured delay
> * values vs. actual b/w enforced.
> * Hence we need a mapping that is pre calibrated so the user can
> * express the memory b/w as a percentage value.
This seems out of place here. If you want to make such a global change
it can be done as a separate patch. For this work it can just consistently
use "hardware" in own areas changed.
Reinette
@@ -1109,6 +1109,7 @@
#define MSR_IA32_QM_CTR 0xc8e
#define MSR_IA32_PQR_ASSOC 0xc8f
#define MSR_IA32_L3_CBM_BASE 0xc90
+#define MSR_RMID_SNC_CONFIG 0xca0
#define MSR_IA32_L2_CBM_BASE 0xd10
#define MSR_IA32_MBA_THRTL_BASE 0xd50
@@ -16,11 +16,14 @@
#define pr_fmt(fmt) "resctrl: " fmt
+#include <linux/cpu.h>
#include <linux/slab.h>
#include <linux/err.h>
#include <linux/cacheinfo.h>
#include <linux/cpuhotplug.h>
+#include <linux/mod_devicetable.h>
+#include <asm/cpu_device_id.h>
#include <asm/intel-family.h>
#include <asm/resctrl.h>
#include "internal.h"
@@ -184,10 +187,10 @@ bool is_mba_sc(struct rdt_resource *r)
/*
* rdt_get_mb_table() - get a mapping of bandwidth(b/w) percentage values
- * exposed to user interface and the h/w understandable delay values.
+ * exposed to user interface and the hardware understandable delay values.
*
* The non-linear delay values have the granularity of power of two
- * and also the h/w does not guarantee a curve for configured delay
+ * and also the hardware does not guarantee a curve for configured delay
* values vs. actual b/w enforced.
* Hence we need a mapping that is pre calibrated so the user can
* express the memory b/w as a percentage value.
@@ -738,11 +741,42 @@ static void clear_closid_rmid(int cpu)
wrmsr(MSR_IA32_PQR_ASSOC, 0, 0);
}
+/*
+ * The power-on reset value of MSR_RMID_SNC_CONFIG is 0x1
+ * which indicates that RMIDs are configured in legacy mode.
+ * This mode is incompatible with Linux resctrl semantics
+ * as RMIDs are partitioned between SNC nodes, which requires
+ * a user to know which RMID is allocated to a task.
+ * Clearing bit 0 reconfigures the RMID counters for use
+ * in Sub NUMA Cluster mode. This mode is better for Linux.
+ * The RMID space is divided between all SNC nodes with the
+ * RMIDs renumbered to start from zero in each node when
+ * couning operations from tasks. Code to read the counters
+ * must adjust RMID counter numbers based on SNC node. See
+ * __rmid_read() for code that does this.
+ */
+static void snc_remap_rmids(int cpu)
+{
+ u64 val;
+
+ /* Only need to enable once per package. */
+ if (cpumask_first(topology_core_cpumask(cpu)) != cpu)
+ return;
+
+ rdmsrl(MSR_RMID_SNC_CONFIG, val);
+ val &= ~BIT_ULL(0);
+ wrmsrl(MSR_RMID_SNC_CONFIG, val);
+}
+
static int resctrl_online_cpu(unsigned int cpu)
{
struct rdt_resource *r;
mutex_lock(&rdtgroup_mutex);
+
+ if (snc_nodes_per_l3_cache > 1)
+ snc_remap_rmids(cpu);
+
for_each_capable_rdt_resource(r)
domain_add_cpu(cpu, r);
/* The cpu is set in default rdtgroup after online. */
@@ -997,11 +1031,73 @@ static __init bool get_rdt_resources(void)
return (rdt_mon_capable || rdt_alloc_capable);
}
+/* CPU models that support MSR_RMID_SNC_CONFIG */
+static const struct x86_cpu_id snc_cpu_ids[] __initconst = {
+ X86_MATCH_INTEL_FAM6_MODEL(ICELAKE_X, 0),
+ X86_MATCH_INTEL_FAM6_MODEL(SAPPHIRERAPIDS_X, 0),
+ X86_MATCH_INTEL_FAM6_MODEL(EMERALDRAPIDS_X, 0),
+ X86_MATCH_INTEL_FAM6_MODEL(GRANITERAPIDS_X, 0),
+ {}
+};
+
+/*
+ * There isn't a simple hardware bit that indicates whether a CPU is running
+ * in Sub NUMA Cluster (SNC) mode. Infer the state by comparing the
+ * ratio of NUMA nodes to L3 cache instances.
+ * It is not possible to accurately determine SNC state if the system is
+ * booted with a maxcpus=N parameter. That distorts the ratio of SNC nodes
+ * to L3 caches. It will be OK if system is booted with hyperthreading
+ * disabled (since this doesn't affect the ratio).
+ */
+static __init int snc_get_config(void)
+{
+ unsigned long *node_caches;
+ int mem_only_nodes = 0;
+ int cpu, node, ret;
+ int num_l3_caches;
+
+ if (!x86_match_cpu(snc_cpu_ids))
+ return 1;
+
+ node_caches = bitmap_zalloc(nr_node_ids, GFP_KERNEL);
+ if (!node_caches)
+ return 1;
+
+ cpus_read_lock();
+
+ if (num_online_cpus() != num_present_cpus())
+ pr_warn("Some CPUs offline, SNC detection may be incorrect\n");
+
+ for_each_node(node) {
+ cpu = cpumask_first(cpumask_of_node(node));
+ if (cpu < nr_cpu_ids)
+ set_bit(get_cpu_cacheinfo_id(cpu, 3), node_caches);
+ else
+ mem_only_nodes++;
+ }
+ cpus_read_unlock();
+
+ num_l3_caches = bitmap_weight(node_caches, nr_node_ids);
+ kfree(node_caches);
+
+ if (!num_l3_caches)
+ return 1;
+
+ ret = (nr_node_ids - mem_only_nodes) / num_l3_caches;
+
+ if (ret > 1)
+ rdt_resources_all[RDT_RESOURCE_L3].r_resctrl.mon_scope = RESCTRL_NODE;
+
+ return ret;
+}
+
static __init void rdt_init_res_defs_intel(void)
{
struct rdt_hw_resource *hw_res;
struct rdt_resource *r;
+ snc_nodes_per_l3_cache = snc_get_config();
+
for_each_rdt_resource(r) {
hw_res = resctrl_to_arch_res(r);