Re: [PATCH v1] perf pmu intel: Adjust cpumaks for sub-NUMA clusters on graniterapids

[Arnaldo Carvalho de Melo <acme@kernel.org>]

On Wed, May 14, 2025 at 09:38:18PM -0700, Ian Rogers wrote:
> On graniterapids the cache home agent (CHA) and memory controller
> (IMC) PMUs all have their cpumask set to per-socket information. In
> order for per NUMA node aggregation to work correctly the PMUs cpumask
> needs to be set to CPUs for the relevant sub-NUMA grouping.

I'm blindly applying it as I can't test these changes, and I think this
is bad.

Hopefully we should try to agree on having, say, two Tested-by from
people having access to this.

In the past processing patches for hardware that wasn't available for
testing lead to tons of trouble, but perhaps c'est la vie :-\

Hopefully this doesn't become just a silly rant.

- Arnaldo

> For example, on a 2 socket graniterapids machine with sub NUMA
> clustering of 3, for uncore_cha and uncore_imc PMUs the cpumask is
> "0,120" leading to aggregation only on NUMA nodes 0 and 3:
> ```
> $ perf stat --per-node -e 'UNC_CHA_CLOCKTICKS,UNC_M_CLOCKTICKS' -a sleep 1
> 
>  Performance counter stats for 'system wide':
> 
> N0        1    277,835,681,344      UNC_CHA_CLOCKTICKS
> N0        1     19,242,894,228      UNC_M_CLOCKTICKS
> N3        1    277,803,448,124      UNC_CHA_CLOCKTICKS
> N3        1     19,240,741,498      UNC_M_CLOCKTICKS
> 
>        1.002113847 seconds time elapsed
> ```
> 
> By updating the PMUs cpumasks to "0,120", "40,160" and "80,200" then
> the correctly 6 NUMA node aggregations are achieved:
> ```
> $ perf stat --per-node -e 'UNC_CHA_CLOCKTICKS,UNC_M_CLOCKTICKS' -a sleep 1
> 
>  Performance counter stats for 'system wide':
> 
> N0        1     92,748,667,796      UNC_CHA_CLOCKTICKS
> N0        0      6,424,021,142      UNC_M_CLOCKTICKS
> N1        0     92,753,504,424      UNC_CHA_CLOCKTICKS
> N1        1      6,424,308,338      UNC_M_CLOCKTICKS
> N2        0     92,751,170,084      UNC_CHA_CLOCKTICKS
> N2        0      6,424,227,402      UNC_M_CLOCKTICKS
> N3        1     92,745,944,144      UNC_CHA_CLOCKTICKS
> N3        0      6,423,752,086      UNC_M_CLOCKTICKS
> N4        0     92,725,793,788      UNC_CHA_CLOCKTICKS
> N4        1      6,422,393,266      UNC_M_CLOCKTICKS
> N5        0     92,717,504,388      UNC_CHA_CLOCKTICKS
> N5        0      6,421,842,618      UNC_M_CLOCKTICKS
> 
>        1.003406645 seconds time elapsed
> ```
> 
> In general, having the perf tool adjust cpumasks isn't desirable as
> ideally the PMU driver would be advertising the correct cpumask.
> 
> Signed-off-by: Ian Rogers <irogers@google.com>
> ---
>  tools/perf/arch/x86/util/pmu.c | 259 ++++++++++++++++++++++++++++++++-
>  1 file changed, 254 insertions(+), 5 deletions(-)
> 
> diff --git a/tools/perf/arch/x86/util/pmu.c b/tools/perf/arch/x86/util/pmu.c
> index 8712cbbbc712..38c800c6e9c8 100644
> --- a/tools/perf/arch/x86/util/pmu.c
> +++ b/tools/perf/arch/x86/util/pmu.c
> @@ -8,6 +8,8 @@
>  #include <linux/perf_event.h>
>  #include <linux/zalloc.h>
>  #include <api/fs/fs.h>
> +#include <api/io_dir.h>
> +#include <internal/cpumap.h>
>  #include <errno.h>
>  
>  #include "../../../util/intel-pt.h"
> @@ -16,7 +18,247 @@
>  #include "../../../util/fncache.h"
>  #include "../../../util/pmus.h"
>  #include "mem-events.h"
> +#include "util/debug.h"
>  #include "util/env.h"
> +#include "util/header.h"
> +
> +static bool x86__is_intel_graniterapids(void)
> +{
> +	static bool checked_if_graniterapids;
> +	static bool is_graniterapids;
> +
> +	if (!checked_if_graniterapids) {
> +		const char *graniterapids_cpuid = "GenuineIntel-6-A[DE]";
> +		char *cpuid = get_cpuid_str((struct perf_cpu){0});
> +
> +		is_graniterapids = cpuid && strcmp_cpuid_str(graniterapids_cpuid, cpuid) == 0;
> +		free(cpuid);
> +		checked_if_graniterapids = true;
> +	}
> +	return is_graniterapids;
> +}
> +
> +static struct perf_cpu_map *read_sysfs_cpu_map(const char *sysfs_path)
> +{
> +	struct perf_cpu_map *cpus;
> +	char *buf = NULL;
> +	size_t buf_len;
> +
> +	if (sysfs__read_str(sysfs_path, &buf, &buf_len) < 0)
> +		return NULL;
> +
> +	cpus = perf_cpu_map__new(buf);
> +	free(buf);
> +	return cpus;
> +}
> +
> +static int snc_nodes_per_l3_cache(void)
> +{
> +	static bool checked_snc;
 +	static int snc_nodes;
> +
> +	if (!checked_snc) {
> +		struct perf_cpu_map *node_cpus =
> +			read_sysfs_cpu_map("devices/system/node/node0/cpulist");
> +		struct perf_cpu_map *cache_cpus =
> +			read_sysfs_cpu_map("devices/system/cpu/cpu0/cache/index3/shared_cpu_list");
> +
> +		snc_nodes = perf_cpu_map__nr(cache_cpus) / perf_cpu_map__nr(node_cpus);
> +		perf_cpu_map__put(cache_cpus);
> +		perf_cpu_map__put(node_cpus);
> +		checked_snc = true;
> +	}
> +	return snc_nodes;
> +}
> +
> +static bool starts_with(const char *str, const char *prefix)
> +{
> +	return !strncmp(prefix, str, strlen(prefix));
> +}
> +
> +static int num_chas(void)
> +{
> +	static bool checked_chas;
> +	static int num_chas;
> +
> +	if (!checked_chas) {
> +		int fd = perf_pmu__event_source_devices_fd();
> +		struct io_dir dir;
> +		struct io_dirent64 *dent;
> +
> +		if (fd < 0)
> +			return -1;
> +
> +		io_dir__init(&dir, fd);
> +
> +		while ((dent = io_dir__readdir(&dir)) != NULL) {
> +			/* Note, dent->d_type will be DT_LNK and so isn't a useful filter. */
> +			if (starts_with(dent->d_name, "uncore_cha_"))
> +				num_chas++;
> +		}
> +		close(fd);
> +		checked_chas = true;
> +	}
> +	return num_chas;
> +}
> +
> +#define MAX_SNCS 6
> +
> +static int uncore_cha_snc(struct perf_pmu *pmu)
> +{
> +	// CHA SNC numbers are ordered correspond to the CHAs number.
> +	unsigned int cha_num;
> +	int num_cha, chas_per_node, cha_snc;
> +	int snc_nodes = snc_nodes_per_l3_cache();
> +
> +	if (snc_nodes <= 1)
> +		return 0;
> +
> +	num_cha = num_chas();
> +	if (num_cha <= 0) {
> +		pr_warning("Unexpected: no CHAs found\n");
> +		return 0;
> +	}
> +
> +	/* Compute SNC for PMU. */
> +	if (sscanf(pmu->name, "uncore_cha_%u", &cha_num) != 1) {
> +		pr_warning("Unexpected: unable to compute CHA number '%s'\n", pmu->name);
> +		return 0;
> +	}
> +	chas_per_node = num_cha / snc_nodes;
> +	cha_snc = cha_num / chas_per_node;
> +
> +	/* Range check cha_snc. for unexpected out of bounds. */
> +	return cha_snc >= MAX_SNCS ? 0 : cha_snc;
> +}
> +
> +static int uncore_imc_snc(struct perf_pmu *pmu)
> +{
> +	// Compute the IMC SNC using lookup tables.
> +	unsigned int imc_num;
> +	int snc_nodes = snc_nodes_per_l3_cache();
> +	const u8 snc2_map[] = {1, 1, 0, 0, 1, 1, 0, 0};
> +	const u8 snc3_map[] = {1, 1, 0, 0, 2, 2, 1, 1, 0, 0, 2, 2};
> +	const u8 *snc_map;
> +	size_t snc_map_len;
> +
> +	switch (snc_nodes) {
> +	case 2:
> +		snc_map = snc2_map;
> +		snc_map_len = ARRAY_SIZE(snc2_map);
> +		break;
> +	case 3:
> +		snc_map = snc3_map;
> +		snc_map_len = ARRAY_SIZE(snc3_map);
> +		break;
> +	default:
> +		/* Error or no lookup support for SNC with >3 nodes. */
> +		return 0;
> +	}
> +
> +	/* Compute SNC for PMU. */
> +	if (sscanf(pmu->name, "uncore_imc_%u", &imc_num) != 1) {
> +		pr_warning("Unexpected: unable to compute IMC number '%s'\n", pmu->name);
> +		return 0;
> +	}
> +	if (imc_num >= snc_map_len) {
> +		pr_warning("Unexpected IMC %d for SNC%d mapping\n", imc_num, snc_nodes);
> +		return 0;
> +	}
> +	return snc_map[imc_num];
> +}
> +
> +static int uncore_cha_imc_compute_cpu_adjust(int pmu_snc)
> +{
> +	static bool checked_cpu_adjust[MAX_SNCS];
> +	static int cpu_adjust[MAX_SNCS];
> +	struct perf_cpu_map *node_cpus;
> +	char node_path[] = "devices/system/node/node0/cpulist";
> +
> +	/* Was adjust already computed? */
> +	if (checked_cpu_adjust[pmu_snc])
> +		return cpu_adjust[pmu_snc];
> +
> +	/* SNC0 doesn't need an adjust. */
> +	if (pmu_snc == 0) {
> +		cpu_adjust[0] = 0;
> +		checked_cpu_adjust[0] = true;
> +		return 0;
> +	}
> +
> +	/*
> +	 * Use NUMA topology to compute first CPU of the NUMA node, we want to
> +	 * adjust CPU 0 to be this and similarly for other CPUs if there is >1
> +	 * socket.
> +	 */
> +	node_path[24] = pmu_snc; // Shift node0 to be node<pmu_snc>.
> +	node_cpus = read_sysfs_cpu_map(node_path);
> +	cpu_adjust[pmu_snc] = perf_cpu_map__cpu(node_cpus, 0).cpu;
> +	checked_cpu_adjust[pmu_snc] = true;
> +	perf_cpu_map__put(node_cpus);
> +	return cpu_adjust[pmu_snc];
> +}
> +
> +static void gnr_uncore_cha_imc_adjust_cpumask_for_snc(struct perf_pmu *pmu, bool cha)
> +{
> +	// With sub-NUMA clustering (SNC) there is a NUMA node per SNC in the
> +	// topology. For example, a two socket graniterapids machine may be set
> +	// up with 3-way SNC meaning there are 6 NUMA nodes that should be
> +	// displayed with --per-node. The cpumask of the CHA and IMC PMUs
> +	// reflects per-socket information meaning, for example, uncore_cha_60
> +	// on a two socket graniterapids machine with 120 cores per socket will
> +	// have a cpumask of "0,120". This cpumask needs adjusting to "40,160"
> +	// to reflect that uncore_cha_60 is used for the 2nd SNC of each
> +	// socket. Without the adjustment events on uncore_cha_60 will appear in
> +	// node 0 and node 3 (in our example 2 socket 3-way set up), but with
> +	// the adjustment they will appear in node 1 and node 4. The number of
> +	// CHAs is typically larger than the number of cores. The CHA numbers
> +	// are assumed to split evenly and inorder wrt core numbers. There are
> +	// fewer memory IMC PMUs than cores and mapping is handled using lookup
> +	// tables.
> +	static struct perf_cpu_map *cha_adjusted[MAX_SNCS];
> +	static struct perf_cpu_map *imc_adjusted[MAX_SNCS];
> +	struct perf_cpu_map **adjusted = cha ? cha_adjusted : imc_adjusted;
> +	int idx, pmu_snc, cpu_adjust;
> +	struct perf_cpu cpu;
> +	bool alloc;
> +
> +	// Cpus from the kernel holds first CPU of each socket. e.g. 0,120
> +	assert(perf_cpu_map__cpu(pmu->cpus, 0).cpu == 0);
> +
> +	pmu_snc = cha ? uncore_cha_snc(pmu) : uncore_imc_snc(pmu);
> +	if (pmu_snc == 0) {
> +		// No adjustment necessary for the first SNC.
> +		return;
> +	}
> +
> +	alloc = adjusted[pmu_snc] == NULL;
> +	if (alloc) {
> +		// Hold onto the perf_cpu_map globally to avoid recomputation.
> +		cpu_adjust = uncore_cha_imc_compute_cpu_adjust(pmu_snc);
> +		adjusted[pmu_snc] = perf_cpu_map__empty_new(perf_cpu_map__nr(pmu->cpus));
> +		if (!adjusted[pmu_snc])
> +			return;
> +	}
> +
> +	perf_cpu_map__for_each_cpu(cpu, idx, pmu->cpus) {
> +		// Compute the new cpu map values or if not allocating, assert
> +		// that they match expectations. asserts will be removed to
> +		// avoid overhead in NDEBUG builds.
> +		if (alloc) {
> +			adjusted[pmu_snc]->map[idx].cpu = cpu.cpu + cpu_adjust;
> +		} else if (idx == 0) {
> +			cpu_adjust = perf_cpu_map__cpu(adjusted[pmu_snc], idx).cpu - cpu.cpu;
> +			assert(uncore_cha_imc_compute_cpu_adjust(pmu_snc) == cpu_adjust);
> +		} else {
> +			assert(perf_cpu_map__cpu(adjusted[pmu_snc], idx).cpu ==
> +			       cpu.cpu + cpu_adjust);
> +		}
> +	}
> +
> +	perf_cpu_map__put(pmu->cpus);
> +	pmu->cpus = perf_cpu_map__get(adjusted[pmu_snc]);
> +}
>  
>  void perf_pmu__arch_init(struct perf_pmu *pmu)
>  {
> @@ -49,10 +291,17 @@ void perf_pmu__arch_init(struct perf_pmu *pmu)
>  
>  		perf_mem_events__loads_ldlat = 0;
>  		pmu->mem_events = perf_mem_events_amd_ldlat;
> -	} else if (pmu->is_core) {
> -		if (perf_pmu__have_event(pmu, "mem-loads-aux"))
> -			pmu->mem_events = perf_mem_events_intel_aux;
> -		else
> -			pmu->mem_events = perf_mem_events_intel;
> +	} else {
> +		if (pmu->is_core) {
> +			if (perf_pmu__have_event(pmu, "mem-loads-aux"))
> +				pmu->mem_events = perf_mem_events_intel_aux;
> +			else
> +				pmu->mem_events = perf_mem_events_intel;
> +		} else if (x86__is_intel_graniterapids()) {
> +			if (starts_with(pmu->name, "uncore_cha_"))
> +				gnr_uncore_cha_imc_adjust_cpumask_for_snc(pmu, /*cha=*/true);
> +			else if (starts_with(pmu->name, "uncore_imc_"))
> +				gnr_uncore_cha_imc_adjust_cpumask_for_snc(pmu, /*cha=*/false);
> +		}
>  	}
>  }
> -- 
> 2.49.0.1045.g170613ef41-goog
>