Skip to contentSkip to conversation

perf x86: Exposing IO stack to IO PMON mapping through sysfs

PATCH Superseded [email protected] 4 participants 14 messages Nov 26, 2019 – Dec 6, 2019

Subsystems (2)

v1 of 9 Superseded Latest is v9 →

lore
Share & tools
lore
#0 lore raw

cover perf x86: Exposing IO stack to IO PMON mapping through sysfs

From: Roman Sudarikov <[email protected]>

Intel® Xeon® Scalable processor family (code name Skylake-SP) makes significant changes in the integrated I/O (IIO) architecture. The new solution introduces IIO stacks which are responsible for managing traffic between the PCIe domain and the Mesh domain. Each IIO stack has its own PMON block and can handle either DMI port, x16 PCIe root port, MCP-Link or various built-in accelerators. IIO PMON blocks allow concurrent monitoring of I/O flows up to 4 x4 bifurcation within each IIO stack.

Software is supposed to program required perf counters within each IIO stack and gather performance data. The tricky thing here is that IIO PMON reports data per IIO stack but users have no idea what IIO stacks are - they only know devices which are connected to the platform.

Understanding IIO stack concept to find which IIO stack that particular IO device is connected to, or to identify an IIO PMON block to program for monitoring specific IIO stack assumes a lot of implicit knowledge about given Intel server platform architecture.

This patch set introduces:
An infrastructure for exposing an Uncore unit to Uncore PMON mapping through sysfs-backend A new --iiostat mode in perf stat to provide I/O performance metrics per I/O device

Usage examples:

1. List all devices below IIO stacks
./perf stat --iiostat=show

Sample output w/o libpci:

    S0-RootPort0-uncore_iio_0<00:00.0>
    S1-RootPort0-uncore_iio_0<81:00.0>
    S0-RootPort1-uncore_iio_1<18:00.0>
    S1-RootPort1-uncore_iio_1<86:00.0>
    S1-RootPort1-uncore_iio_1<88:00.0>
    S0-RootPort2-uncore_iio_2<3d:00.0>
    S1-RootPort2-uncore_iio_2<af:00.0>
    S1-RootPort3-uncore_iio_3<da:00.0>

Sample output with libpci:

    S0-RootPort0-uncore_iio_0<00:00.0 Sky Lake-E DMI3 Registers>
    S1-RootPort0-uncore_iio_0<81:00.0 Ethernet Controller X710 for 10GbE SFP+>
    S0-RootPort1-uncore_iio_1<18:00.0 Omni-Path HFI Silicon 100 Series [discrete]>
    S1-RootPort1-uncore_iio_1<86:00.0 Ethernet Controller XL710 for 40GbE QSFP+>
    S1-RootPort1-uncore_iio_1<88:00.0 Ethernet Controller XL710 for 40GbE QSFP+>
    S0-RootPort2-uncore_iio_2<3d:00.0 Ethernet Connection X722 for 10GBASE-T>
    S1-RootPort2-uncore_iio_2<af:00.0 Omni-Path HFI Silicon 100 Series [discrete]>
    S1-RootPort3-uncore_iio_3<da:00.0 NVMe Datacenter SSD [Optane]>

2. Collect metrics for all I/O devices below IIO stack

  ./perf stat --iiostat -- dd if=/dev/zero of=/dev/nvme0n1 bs=1M oflag=direct
    357708+0 records in
    357707+0 records out
    375083606016 bytes (375 GB, 349 GiB) copied, 215.381 s, 1.7 GB/s

Performance counter stats for 'system wide':

device Inbound Read(MB) Inbound Write(MB) Outbound Read(MB) Outbound Write(MB)
00:00.0 0 0 0 0
81:00.0 0 0 0 0
18:00.0 0 0 0 0
86:00.0 0 0 0 0
88:00.0 0 0 0 0
3b:00.0 3 0 0 0
3c:03.0 3 0 0 0
3d:00.0 3 0 0 0
af:00.0 0 0 0 0
da:00.0 358559 44 0 22

215.383783574 seconds time elapsed

3. Collect metrics for comma separted list of I/O devices

  ./perf stat --iiostat=da:00.0 -- dd if=/dev/zero of=/dev/nvme0n1 bs=1M oflag=direct
    381555+0 records in
    381554+0 records out
    400088457216 bytes (400 GB, 373 GiB) copied, 374.044 s, 1.1 GB/s

Performance counter stats for 'system wide':

device Inbound Read(MB) Inbound Write(MB) Outbound Read(MB) Outbound Write(MB)
da:00.0 382462 47 0 23

374.045775505 seconds time elapsed

Roman Sudarikov (6):
perf x86: Infrastructure for exposing an Uncore unit to PMON mapping perf tools: Helper functions to enumerate and probe PCI devices perf stat: Helper functions for list of IIO devices
perf stat: New --iiostat mode to provide I/O performance metrics perf tools: Add feature check for libpci
perf stat: Add PCI device name to --iiostat output

 arch/x86/events/intel/uncore.c                |  61 +-
 arch/x86/events/intel/uncore.h                |  13 +-
 arch/x86/events/intel/uncore_snbep.c          | 144 ++++
 tools/build/Makefile.feature                  |   2 +
 tools/build/feature/Makefile                  |   4 +
 tools/build/feature/test-all.c                |   5 +
 tools/build/feature/test-libpci.c             |  10 +
 tools/perf/Documentation/perf-stat.txt        |  12 +
 tools/perf/Makefile.config                    |  10 +
 tools/perf/arch/x86/util/Build                |   1 +
 tools/perf/arch/x86/util/iiostat.c            | 718 ++++++++++++++++++
 tools/perf/builtin-stat.c                     |  32 +-
 tools/perf/builtin-version.c                  |   1 +
 tools/perf/tests/make                         |   1 +
 tools/perf/util/Build                         |   1 +
 tools/perf/util/evsel.h                       |   1 +
 tools/perf/util/iiostat.h                     |  35 +
 tools/perf/util/pci.c                         |  99 +++
 tools/perf/util/pci.h                         |  27 +
 .../scripting-engines/trace-event-python.c    |   2 +-
 tools/perf/util/stat-display.c                |  53 +-
 tools/perf/util/stat-shadow.c                 |  12 +-
 tools/perf/util/stat.c                        |   3 +-
 tools/perf/util/stat.h                        |   2 +
 24 files changed, 1237 insertions(+), 12 deletions(-)
 create mode 100644 tools/build/feature/test-libpci.c
 create mode 100644 tools/perf/arch/x86/util/iiostat.c
 create mode 100644 tools/perf/util/iiostat.h
 create mode 100644 tools/perf/util/pci.c
 create mode 100644 tools/perf/util/pci.h

base-commit: 219d54332a09e8d8741c1e1982f5eae56099de85

Signature · 1 line
2.19.1
#1 lore raw

PATCH 1/6 perf x86: Infrastructure for exposing an Uncore unit to PMON mapping

From: Roman Sudarikov <[email protected]>

Intel® Xeon® Scalable processor family (code name Skylake-SP) makes significant changes in the integrated I/O (IIO) architecture. The new solution introduces IIO stacks which are responsible for managing traffic between the PCIe domain and the Mesh domain. Each IIO stack has its own PMON block and can handle either DMI port, x16 PCIe root port, MCP-Link or various built-in accelerators. IIO PMON blocks allow concurrent monitoring of I/O flows up to 4 x4 bifurcation within each IIO stack.

Software is supposed to program required perf counters within each IIO stack and gather performance data. The tricky thing here is that IIO PMON reports data per IIO stack but users have no idea what IIO stacks are - they only know devices which are connected to the platform.

Understanding IIO stack concept to find which IIO stack that particular IO device is connected to, or to identify an IIO PMON block to program for monitoring specific IIO stack assumes a lot of implicit knowledge about given Intel server platform architecture.

This patch set introduces:
An infrastructure for exposing an Uncore unit to Uncore PMON mapping through sysfs-backend A new --iiostat mode in perf stat to provide I/O performance metrics per I/O device

Current version supports a server line starting Intel® Xeon® Processor Scalable Family and introduces mapping for IIO Uncore units only.
Other units can be added on demand.

Usage example:
/sys/devices/uncore_<type>_<pmu_idx>/platform_mapping

Each Uncore unit type, by its nature, can be mapped to its own context, for example:
CHA - each uncore_cha_<pmu_idx> is assigned to manage a distinct slice of LLC capacity UPI - each uncore_upi_<pmu_idx> is assigned to manage one link of Intel UPI Subsystem IIO - each uncore_iio_<pmu_idx> is assigned to manage one stack of the IIO module IMC - each uncore_imc_<pmu_idx> is assigned to manage one channel of Memory Controller

Implementation details:
Two callbacks added to struct intel_uncore_type to discover and map Uncore units to PMONs:
int (*get_topology)(struct intel_uncore_type *type)
int (*set_mapping)(struct intel_uncore_type *type)

IIO stack to PMON mapping is exposed through
/sys/devices/uncore_iio_<pmu_idx>/platform_mapping in the following format: domain:bus

Details of IIO Uncore unit mapping to IIO PMON:
Each IIO stack is either a DMI port, x16 PCIe root port, MCP-Link or various built-in accelerators. For Uncore IIO Unit type, the platform_mapping file holds bus numbers of devices, which can be monitored by that IIO PMON block on each die.

For example, on a 4-die Intel Xeon® server platform:

    $ cat /sys/devices/uncore_iio_0/platform_mapping
    0000:00,0000:40,0000:80,0000:c0

Which means:
IIO PMON block 0 on die 0 belongs to IIO stack located on bus 0x00, domain 0x0000 IIO PMON block 0 on die 1 belongs to IIO stack located on bus 0x40, domain 0x0000 IIO PMON block 0 on die 2 belongs to IIO stack located on bus 0x80, domain 0x0000 IIO PMON block 0 on die 3 belongs to IIO stack located on bus 0xc0, domain 0x0000

---
 arch/x86/events/intel/uncore.c       |  61 +++++++++++-
 arch/x86/events/intel/uncore.h       |  13 ++-
 arch/x86/events/intel/uncore_snbep.c | 144 +++++++++++++++++++++++++++
 3 files changed, 214 insertions(+), 4 deletions(-)
arch/x86/events/intel/uncore.c +58 -3
index 86467f85c383..0f779c8fcc05 100644
@@ -18,6 +18,11 @@ struct list_head pci2phy_map_head = LIST_HEAD_INIT(pci2phy_map_head);
 struct pci_extra_dev *uncore_extra_pci_dev;
 static int max_dies;
 
+int get_max_dies(void)
+{
+	return max_dies;
+}
+
 /* mask of cpus that collect uncore events */
 static cpumask_t uncore_cpu_mask;
 
@@ -816,6 +821,16 @@ static ssize_t uncore_get_attr_cpumask(struct device *dev,
 
 static DEVICE_ATTR(cpumask, S_IRUGO, uncore_get_attr_cpumask, NULL);
 
+static ssize_t platform_mapping_show(struct device *dev,
+				struct device_attribute *attr, char *buf)
+{
+	struct intel_uncore_pmu *pmu = dev_get_drvdata(dev);
+
+	return snprintf(buf, PAGE_SIZE - 1, "%s\n", pmu->platform_mapping ?
+		       (char *)pmu->platform_mapping : "0");
+}
+static DEVICE_ATTR_RO(platform_mapping);
+
 static struct attribute *uncore_pmu_attrs[] = {
 	&dev_attr_cpumask.attr,
 	NULL,
@@ -825,6 +840,15 @@ static const struct attribute_group uncore_pmu_attr_group = {
 	.attrs = uncore_pmu_attrs,
 };
 
+static struct attribute *platform_attrs[] = {
+	&dev_attr_platform_mapping.attr,
+	NULL,
+};
+
+static const struct attribute_group uncore_platform_discovery_group = {
+	.attrs = platform_attrs,
+};
+
 static int uncore_pmu_register(struct intel_uncore_pmu *pmu)
 {
 	int ret;
@@ -905,11 +929,27 @@ static void uncore_types_exit(struct intel_uncore_type **types)
 		uncore_type_exit(*types);
 }
 
+static void uncore_type_attrs_compaction(struct intel_uncore_type *type)
+{
+	int i, j;
+
+	for (i = 0, j = 0; i < UNCORE_MAX_NUM_ATTR_GROUP; i++) {
+		if (!type->attr_groups[i])
+			continue;
+		if (i > j) {
+			type->attr_groups[j] = type->attr_groups[i];
+			type->attr_groups[i] = NULL;
+		}
+		j++;
+	}
+}
+
 static int __init uncore_type_init(struct intel_uncore_type *type, bool setid)
 {
 	struct intel_uncore_pmu *pmus;
 	size_t size;
 	int i, j;
+	int ret;
 
 	pmus = kcalloc(type->num_boxes, sizeof(*pmus), GFP_KERNEL);
 	if (!pmus)
@@ -922,8 +962,10 @@ static int __init uncore_type_init(struct intel_uncore_type *type, bool setid)
 		pmus[i].pmu_idx	= i;
 		pmus[i].type	= type;
 		pmus[i].boxes	= kzalloc(size, GFP_KERNEL);
-		if (!pmus[i].boxes)
+		if (!pmus[i].boxes) {
+			ret = -ENOMEM;
 			goto err;
+		}
 	}
 
 	type->pmus = pmus;
@@ -940,8 +982,10 @@ static int __init uncore_type_init(struct intel_uncore_type *type, bool setid)
 
 		attr_group = kzalloc(struct_size(attr_group, attrs, i + 1),
 								GFP_KERNEL);
-		if (!attr_group)
+		if (!attr_group) {
+			ret = -ENOMEM;
 			goto err;
+		}
 
 		attr_group->group.name = "events";
 		attr_group->group.attrs = attr_group->attrs;
@@ -954,6 +998,17 @@ static int __init uncore_type_init(struct intel_uncore_type *type, bool setid)
 
 	type->pmu_group = &uncore_pmu_attr_group;
 
+	/*
+	 * Exposing mapping of Uncore units to corresponding Uncore PMUs
+	 * through /sys/devices/uncore_<type>_<idx>/platform_mapping
+	 */
+	if (type->get_topology && type->set_mapping)
+		if (!type->get_topology(type) && !type->set_mapping(type))
+			type->platform_discovery = &uncore_platform_discovery_group;
+
+	/* For optional attributes, we can safely remove embedded NULL attr_groups elements */
+	uncore_type_attrs_compaction(type);
+
 	return 0;
 
 err:
@@ -961,7 +1016,7 @@ static int __init uncore_type_init(struct intel_uncore_type *type, bool setid)
 		kfree(pmus[i].boxes);
 	kfree(pmus);
 
-	return -ENOMEM;
+	return ret;
 }
 
 static int __init
arch/x86/events/intel/uncore.h +12 -1
index bbfdaa720b45..ce3727b9f7f8 100644
@@ -43,6 +43,8 @@ struct intel_uncore_box;
 struct uncore_event_desc;
 struct freerunning_counters;
 
+#define UNCORE_MAX_NUM_ATTR_GROUP 5
+
 struct intel_uncore_type {
 	const char *name;
 	int num_counters;
@@ -71,13 +73,19 @@ struct intel_uncore_type {
 	struct intel_uncore_ops *ops;
 	struct uncore_event_desc *event_descs;
 	struct freerunning_counters *freerunning;
-	const struct attribute_group *attr_groups[4];
+	const struct attribute_group *attr_groups[UNCORE_MAX_NUM_ATTR_GROUP];
 	struct pmu *pmu; /* for custom pmu ops */
+	void *platform_topology;
+	/* finding Uncore units */
+	int (*get_topology)(struct intel_uncore_type *type);
+	/* mapping Uncore units to PMON ranges */
+	int (*set_mapping)(struct intel_uncore_type *type);
 };
 
 #define pmu_group attr_groups[0]
 #define format_group attr_groups[1]
 #define events_group attr_groups[2]
+#define platform_discovery attr_groups[3]
 
 struct intel_uncore_ops {
 	void (*init_box)(struct intel_uncore_box *);
@@ -99,6 +107,7 @@ struct intel_uncore_pmu {
 	int				pmu_idx;
 	int				func_id;
 	bool				registered;
+	void				*platform_mapping;
 	atomic_t			activeboxes;
 	struct intel_uncore_type	*type;
 	struct intel_uncore_box		**boxes;
@@ -490,6 +499,8 @@ static inline struct intel_uncore_box *uncore_event_to_box(struct perf_event *ev
 	return event->pmu_private;
 }
 
+int get_max_dies(void);
+
 struct intel_uncore_box *uncore_pmu_to_box(struct intel_uncore_pmu *pmu, int cpu);
 u64 uncore_msr_read_counter(struct intel_uncore_box *box, struct perf_event *event);
 void uncore_mmio_exit_box(struct intel_uncore_box *box);
arch/x86/events/intel/uncore_snbep.c +144 -0

168 lines not on this page · raw message

Signature · 1 line
2.19.1
#2 lore raw

PATCH 2/6 perf tools: Helper functions to enumerate and probe PCI devices

From: Roman Sudarikov <[email protected]>

This makes aggregation of performance data per I/O device available in the perf user tools.

---
 tools/perf/util/Build |  1 +
 tools/perf/util/pci.c | 53 +++++++++++++++++++++++++++++++++++++++++++
 tools/perf/util/pci.h | 23 +++++++++++++++++++
 3 files changed, 77 insertions(+)
 create mode 100644 tools/perf/util/pci.c
 create mode 100644 tools/perf/util/pci.h
tools/perf/util/Build +1 -0
index 8dcfca1a882f..02b699f8a10a 100644
@@ -23,6 +23,7 @@ perf-y += memswap.o
 perf-y += parse-events.o
 perf-y += perf_regs.o
 perf-y += path.o
+perf-y += pci.o
 perf-y += print_binary.o
 perf-y += rlimit.o
 perf-y += argv_split.o
tools/perf/util/pci.c new file +53 -0

53 lines not on this page · raw message

tools/perf/util/pci.h new file +23 -0
new file mode 100644
index 000000000000..e963b12e10e7
@@ -0,0 +1,23 @@
+/* SPDX-License-Identifier: GPL-2.0*/
+/*
+ *
+ * Copyright (C) 2019, Intel Corporation
+ *
+ * Authors: Roman Sudarikov <[email protected]>
+ *	    Alexander Antonov <[email protected]>
+ */
+#ifndef _PCI_H
+#define _PCI_H
+
+#include <linux/types.h>
+
+struct bdf {
+	u8 busno;
+	u8 devno;
+	u8 funcno;
+};
+
+bool pci_device_probe(struct bdf bdf);
+bool is_pci_device_root_port(struct bdf bdf, u8 *secondary, u8 *subordinate);
+
+#endif /* _PCI_H */
Signature · 1 line
2.19.1
#3 lore raw

PATCH 3/6 perf stat: Helper functions for list of IIO devices

From: Roman Sudarikov <[email protected]>

Helper functions to iterate through and manipulate with list of struct iio_device objects. The following patch will use it.

---
 tools/perf/arch/x86/util/iiostat.c | 178 +++++++++++++++++++++++++++++
 1 file changed, 178 insertions(+)
 create mode 100644 tools/perf/arch/x86/util/iiostat.c
tools/perf/arch/x86/util/iiostat.c new file +178 -0
new file mode 100644
index 000000000000..b93b9b9da418
@@ -0,0 +1,178 @@
+// SPDX-License-Identifier: GPL-2.0
+/*
+ * perf stat --iiostat
+ *
+ * Copyright (C) 2019, Intel Corporation
+ *
+ * Authors: Roman Sudarikov <[email protected]>
+ *	    Alexander Antonov <[email protected]>
+ */
+#include "path.h"
+#include "pci.h"
+#include <api/fs/fs.h>
+#include <linux/kernel.h>
+#include <linux/err.h>
+#include "util/debug.h"
+#include "util/iiostat.h"
+#include "util/counts.h"
+#include <limits.h>
+#include <stdio.h>
+#include <string.h>
+#include <errno.h>
+#include <sys/types.h>
+#include <sys/stat.h>
+#include <fcntl.h>
+#include <dirent.h>
+#include <unistd.h>
+#include <stdlib.h>
+#include <regex.h>
+
+struct dev_info {
+	struct bdf bdf;
+	u8 ch_mask;
+	u8 socket;
+	u8 pmu_idx;
+	u8 root_port_nr;
+};
+
+struct iio_device {
+	struct list_head node;
+	struct dev_info	dev_info;
+	int idx;
+};
+
+struct iio_devs_list {
+	struct list_head devices;
+	int nr_entries;
+};
+
+/**
+ * __iio_devs_for_each_device - iterate thru all the iio devices
+ * @list: list_head instance to iterate
+ * @device: struct iio_device iterator
+ */
+#define __iio_devs_for_each_device(list, device) \
+		list_for_each_entry(device, list, node)
+
+/**
+ * iio_devs_list_for_each_device - iterate thru all the iio devices
+ * @list: iio_devs_list instance to iterate
+ * @device: struct iio_device iterator
+ */
+#define iio_devs_list_for_each_device(list, device) \
+	__iio_devs_for_each_device(&(list->devices), device)
+
+/**
+ * __iio_devs_for_each_device_safe - safely iterate thru all the iio devices
+ * @devices: list_head instance to iterate
+ * @tmp: struct iio_device temp iterator
+ * @device: struct iio_device iterator
+ */
+#define __iio_devs_for_each_device_safe(devices, tmp, device) \
+		list_for_each_entry_safe(device, tmp, devices, node)
+
+/**
+ * iio_devs_list_for_each_device_safe - safely iterate thru all the iio devices
+ * @list: iio_devs_list instance to iterate
+ * @tmp: struct iio_device temp iterator
+ * @device: struct iio_device iterator
+ */
+#define iio_devs_list_for_each_device_safe(list, tmp, device) \
+		__iio_devs_for_each_device_safe(&(list->devices), tmp, device)
+
+#define iio_device_delete_from_list(device) \
+		list_del(&(device->node))
+
+static struct iio_device *iio_device_new(struct dev_info *info)
+{
+	struct iio_device *p =
+		(struct iio_device *)calloc(1, sizeof(struct iio_device));
+
+	if (p) {
+		INIT_LIST_HEAD(&(p->node));
+		p->dev_info = *info;
+		p->idx = -1;
+	}
+	return p;
+}
+
+static void iio_device_delete(struct iio_device *device)
+{
+	if (device) {
+		list_del_init(&(device->node));
+		free(device);
+	}
+}
+
+static void iiostat_device_show(FILE *output,
+			const struct iio_device * const device)
+{
+	if (output && device)
+		fprintf(output, "S%d-RootPort%d-uncore_iio_%d<%02x:%02x.%x>\n",
+			device->dev_info.socket,
+			device->dev_info.root_port_nr, device->dev_info.pmu_idx,
+			device->dev_info.bdf.busno, device->dev_info.bdf.devno,
+			device->dev_info.bdf.funcno);
+}
+
+static struct iio_devs_list *iio_devs_list_new(void)
+{
+	struct iio_devs_list *devs_list =
+		(struct iio_devs_list *)calloc(1, sizeof(struct iio_devs_list));
+
+	if (devs_list)
+		INIT_LIST_HEAD(&(devs_list->devices));
+	return devs_list;
+}
+
+static void iio_devs_list_free(struct iio_devs_list *list)
+{
+	struct iio_device *tmp_device;
+	struct iio_device *device;
+
+	if (list) {
+		iio_devs_list_for_each_device_safe(list, tmp_device, device)
+			iio_device_delete(device);
+		list_del_init(&(list->devices));
+		free(list);
+	}
+}
+
+static bool is_same_iio_device(struct bdf lhd, struct bdf rhd)
+{
+	return (lhd.busno == rhd.busno) && (lhd.devno == rhd.devno) &&
+		(lhd.funcno == rhd.funcno);
+}
+
+static void iio_devs_list_add_device(struct iio_devs_list *list,
+				      struct iio_device * const device)
+{
+	struct iio_device *it;
+
+	if (list && device) {
+		iio_devs_list_for_each_device(list, it)
+			if (is_same_iio_device(it->dev_info.bdf, device->dev_info.bdf))
+				return;
+		device->idx = list->nr_entries++;
+		list_add_tail(&(device->node), &(list->devices));
+	}
+}
+
+static void iio_devs_list_join_list(struct iio_devs_list *dest,
+				     struct iio_devs_list *src)
+{
+	int idx = 0;
+	struct iio_device *it;
+
+	if (dest && src) {
+		if (dest->nr_entries) {
+			it = list_last_entry(&(dest->devices),
+					     struct iio_device, node);
+			idx = it->idx + 1;
+		}
+		iio_devs_list_for_each_device(src, it)
+			it->idx = idx++;
+		list_splice_tail(&(src->devices), &(dest->devices));
+		dest->nr_entries += src->nr_entries;
+	}
+}
Signature · 1 line
2.19.1
#4 lore raw

PATCH 4/6 perf stat: New --iiostat mode to provide I/O performance metrics

From: Roman Sudarikov <[email protected]>

New --iiostat mode in perf stat is intended to provide four I/O performance metrics per each IO device below IIO stacks:

    --Inbound Read(Mb)   - I/O device reads from the host memory, in Mb
    --Inbound Write(Mb)  - I/O device writes to the host memory, in Mb
    --Outbound Read(Mb)  - CPU reads from I/O device, in Mb

--Outbound Write(Mb) - CPU writes to I/O device, in Mb

Each metric requires only one IIO event which increments at every 4B transfer in the corresponding direction. The formula to compute metrics are generic:

    #EventCount * 4B / (1024 * 1024)

This implementation starts from discovering IIO stacks on the platform and all devices below each stack. Next step is to configure a group of four events per each device and tie each event group to its device.

Note: --iiostat introduces new perf data aggregation mode - per I/O device hence -e and -M options are not supported.

Usage examples:

1. List all devices below IIO stacks
./perf stat --iiostat=show

Sample output w/o libpci:

    S0-RootPort0-uncore_iio_0<00:00.0>
    S1-RootPort0-uncore_iio_0<81:00.0>
    S0-RootPort1-uncore_iio_1<18:00.0>
    S1-RootPort1-uncore_iio_1<86:00.0>
    S1-RootPort1-uncore_iio_1<88:00.0>
    S0-RootPort2-uncore_iio_2<3d:00.0>
    S1-RootPort2-uncore_iio_2<af:00.0>
    S1-RootPort3-uncore_iio_3<da:00.0>

Sample output with libpci:

    S0-RootPort0-uncore_iio_0<00:00.0 Sky Lake-E DMI3 Registers>
    S1-RootPort0-uncore_iio_0<81:00.0 Ethernet Controller X710 for 10GbE SFP+>
    S0-RootPort1-uncore_iio_1<18:00.0 Omni-Path HFI Silicon 100 Series [discrete]>
    S1-RootPort1-uncore_iio_1<86:00.0 Ethernet Controller XL710 for 40GbE QSFP+>
    S1-RootPort1-uncore_iio_1<88:00.0 Ethernet Controller XL710 for 40GbE QSFP+>
    S0-RootPort2-uncore_iio_2<3d:00.0 Ethernet Connection X722 for 10GBASE-T>
    S1-RootPort2-uncore_iio_2<af:00.0 Omni-Path HFI Silicon 100 Series [discrete]>
    S1-RootPort3-uncore_iio_3<da:00.0 NVMe Datacenter SSD [Optane]>

2. Collect metrics for all I/O devices below IIO stack

  ./perf stat --iiostat -- dd if=/dev/zero of=/dev/nvme0n1 bs=1M oflag=direct
    357708+0 records in
    357707+0 records out
    375083606016 bytes (375 GB, 349 GiB) copied, 215.381 s, 1.7 GB/s

Performance counter stats for 'system wide':

device Inbound Read(MB) Inbound Write(MB) Outbound Read(MB) Outbound Write(MB)
00:00.0 0 0 0 0
81:00.0 0 0 0 0
18:00.0 0 0 0 0
86:00.0 0 0 0 0
88:00.0 0 0 0 0
3b:00.0 3 0 0 0
3c:03.0 3 0 0 0
3d:00.0 3 0 0 0
af:00.0 0 0 0 0
da:00.0 358559 44 0 22

215.383783574 seconds time elapsed

3. Collect metrics for comma separted list of I/O devices

  ./perf stat --iiostat=da:00.0 -- dd if=/dev/zero of=/dev/nvme0n1 bs=1M oflag=direct
    381555+0 records in
    381554+0 records out
    400088457216 bytes (400 GB, 373 GiB) copied, 374.044 s, 1.1 GB/s

Performance counter stats for 'system wide':

device Inbound Read(MB) Inbound Write(MB) Outbound Read(MB) Outbound Write(MB)
da:00.0 382462 47 0 23

374.045775505 seconds time elapsed

---
 tools/perf/Documentation/perf-stat.txt        |  12 +
 tools/perf/arch/x86/util/Build                |   1 +
 tools/perf/arch/x86/util/iiostat.c            | 533 +++++++++++++++++-
 tools/perf/builtin-stat.c                     |  32 +-
 tools/perf/util/evsel.h                       |   1 +
 tools/perf/util/iiostat.h                     |  35 ++
 .../scripting-engines/trace-event-python.c    |   2 +-
 tools/perf/util/stat-display.c                |  53 +-
 tools/perf/util/stat-shadow.c                 |  12 +-
 tools/perf/util/stat.c                        |   3 +-
 tools/perf/util/stat.h                        |   2 +
 11 files changed, 676 insertions(+), 10 deletions(-)
 create mode 100644 tools/perf/util/iiostat.h
tools/perf/Documentation/perf-stat.txt +12 -0
index 930c51c01201..96a2ec8c6b55 100644
@@ -262,6 +262,18 @@ See perf list output for the possble metrics and metricgroups.
 -A::
 --no-aggr::
 Do not aggregate counts across all monitored CPUs.
+--iiostat::
+This mode is intended to provide four I/O performance metrics per each
+IO device below IIO stacks:
+    --Inbound Read(Mb)   - I/O device reads from the host memory, in Mb
+    --Inbound Write(Mb)  - I/O device writes to the host memory, in Mb
+    --Outbound Read(Mb)  - CPU reads from I/O device, in Mb
+    --Outbound Write(Mb) - CPU writes to I/O device, in Mb
+
+Each metric requiries only one IIO event which increments at every 4B
+transfer in corresponding direction. The formulas to compute metrics
+are generic:
+    #EventCount * 4B / (1024 * 1024)
 
 --topdown::
 Print top down level 1 metrics if supported by the CPU. This allows to
tools/perf/arch/x86/util/Build +1 -0
index 47f9c56e744f..e19566e16e5d 100644
@@ -6,6 +6,7 @@ perf-y += perf_regs.o
 perf-y += group.o
 perf-y += machine.o
 perf-y += event.o
+perf-y += iiostat.o
 
 perf-$(CONFIG_DWARF) += dwarf-regs.o
 perf-$(CONFIG_BPF_PROLOGUE) += dwarf-regs.o
tools/perf/arch/x86/util/iiostat.c +531 -2

563 lines not on this page · raw message

tools/perf/builtin-stat.c +31 -1
index 468fc49420ce..c7516a0182a0 100644
@@ -66,6 +66,7 @@
 #include "util/time-utils.h"
 #include "util/top.h"
 #include "asm/bug.h"
+#include "util/iiostat.h"
 
 #include <linux/time64.h>
 #include <linux/zalloc.h>
@@ -186,6 +187,7 @@ static struct perf_stat_config stat_config = {
 	.metric_only_len	= METRIC_ONLY_LEN,
 	.walltime_nsecs_stats	= &walltime_nsecs_stats,
 	.big_num		= true,
+	.iiostat_run		= false,
 };
 
 static inline void diff_timespec(struct timespec *r, struct timespec *a,
@@ -723,6 +725,13 @@ static int parse_metric_groups(const struct option *opt,
 	return metricgroup__parse_groups(opt, str, &stat_config.metric_events);
 }
 
+__weak int iiostat_parse(const struct option *opt __maybe_unused,
+						const char *str __maybe_unused,
+						int unset __maybe_unused)
+{
+	return 0;
+}
+
 static struct option stat_options[] = {
 	OPT_BOOLEAN('T', "transaction", &transaction_run,
 		    "hardware transaction statistics"),
@@ -803,6 +812,8 @@ static struct option stat_options[] = {
 	OPT_CALLBACK('M', "metrics", &evsel_list, "metric/metric group list",
 		     "monitor specified metrics or metric groups (separated by ,)",
 		     parse_metric_groups),
+	OPT_CALLBACK_OPTARG(0, "iiostat", &evsel_list, &stat_config, "PCIe bandwidth",
+	     "measure PCIe bandwidth per device", iiostat_parse),
 	OPT_END()
 };
 
@@ -908,6 +919,7 @@ static int perf_stat_init_aggr_mode(void)
 		break;
 	case AGGR_GLOBAL:
 	case AGGR_THREAD:
+	case AGGR_DEVICE:
 	case AGGR_UNSET:
 	default:
 		break;
@@ -1072,6 +1084,7 @@ static int perf_stat_init_aggr_mode_file(struct perf_stat *st)
 	case AGGR_NONE:
 	case AGGR_GLOBAL:
 	case AGGR_THREAD:
+	case AGGR_DEVICE:
 	case AGGR_UNSET:
 	default:
 		break;
@@ -1129,6 +1142,12 @@ __weak void arch_topdown_group_warn(void)
 {
 }
 
+__weak int iiostat_print_device_list(struct evlist *evlist __maybe_unused,
+				      struct perf_stat_config *config __maybe_unused)
+{
+	return 0;
+}
+
 /*
  * Add default attributes, if there were no attributes specified or
  * if -d/--detailed, -d -d or -d -d -d is used:
@@ -1358,6 +1377,10 @@ static int add_default_attributes(void)
 		free(str);
 	}
 
+	if (stat_config.iiostat_run &&
+		iiostat_print_device_list(evsel_list, &stat_config) < 0)
+		return -1;
+
 	if (!evsel_list->core.nr_entries) {
 		if (target__has_cpu(&target))
 			default_attrs0[0].config = PERF_COUNT_SW_CPU_CLOCK;
@@ -1680,6 +1703,10 @@ static void setup_system_wide(int forks)
 	}
 }
 
+__weak void iiostat_delete_device_list(struct evlist *evlist __maybe_unused)
+{
+}
+
 int cmd_stat(int argc, const char **argv)
 {
 	const char * const stat_usage[] = {
@@ -1844,7 +1871,7 @@ int cmd_stat(int argc, const char **argv)
 	 * --per-thread is aggregated per thread, we dont mix it with cpu mode
 	 */
 	if (((stat_config.aggr_mode != AGGR_GLOBAL &&
-	      stat_config.aggr_mode != AGGR_THREAD) || nr_cgroups) &&
+	      stat_config.aggr_mode != AGGR_THREAD && stat_config.aggr_mode != AGGR_DEVICE) || nr_cgroups) &&
 	    !target__has_cpu(&target)) {
 		fprintf(stderr, "both cgroup and no-aggregation "
 			"modes only available in system-wide mode\n");
@@ -2005,6 +2032,9 @@ int cmd_stat(int argc, const char **argv)
 	perf_stat__exit_aggr_mode();
 	perf_evlist__free_stats(evsel_list);
 out:
+	if (stat_config.iiostat_run)
+		iiostat_delete_device_list(evsel_list);
+
 	zfree(&stat_config.walltime_run);
 
 	if (smi_cost && smi_reset)
tools/perf/util/evsel.h +1 -0
index ddc5ee6f6592..4d1582db3c7b 100644
@@ -100,6 +100,7 @@ struct evsel {
 		perf_evsel__sb_cb_t	*cb;
 		void			*data;
 	} side_band;
+	void			*perf_device;
 };
 
 struct perf_missing_features {
tools/perf/util/iiostat.h new file +35 -0
new file mode 100644
index 000000000000..4c886beec348
@@ -0,0 +1,35 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+/*
+ * perf stat --iiostat
+ *
+ * Copyright (C) 2019, Intel Corporation
+ *
+ * Authors: Roman Sudarikov <[email protected]>
+ *	    Alexander Antonov <[email protected]>
+ */
+
+#ifndef _IIOSTAT_H
+#define _IIOSTAT_H
+
+#include "util/stat.h"
+#include <subcmd/parse-options.h>
+#include "util/parse-events.h"
+#include "util/evlist.h"
+
+struct option;
+struct perf_stat_config;
+struct evlist;
+struct timespec;
+
+int  iiostat_parse(const struct option *opt, const char *str,
+		    int unset __maybe_unused);
+void iiostat_prefix(struct perf_stat_config *config, struct evlist *evlist,
+		     char *prefix, struct timespec *ts);
+void iiostat_print_metric(struct perf_stat_config *config __maybe_unused,
+			   struct evsel *evsel __maybe_unused,
+			   struct perf_stat_output_ctx *out __maybe_unused);
+int iiostat_print_device_list(struct evlist *evlist __maybe_unused,
+			       struct perf_stat_config *config __maybe_unused);
+void iiostat_delete_device_list(struct evlist *evlist);
+
+#endif /* _IIOSTAT_H */
tools/perf/util/scripting-engines/trace-event-python.c +1 -1
index 93c03b39cd9c..65d08daff8d9 100644
@@ -1396,7 +1396,7 @@ static void python_process_stat(struct perf_stat_config *config,
 	struct perf_cpu_map *cpus = counter->core.cpus;
 	int cpu, thread;
 
-	if (config->aggr_mode == AGGR_GLOBAL) {
+	if (config->aggr_mode == AGGR_GLOBAL || config->aggr_mode == AGGR_DEVICE) {
 		process_stat(counter, -1, -1, tstamp,
 			     &counter->counts->aggr);
 		return;
tools/perf/util/stat-display.c +49 -4

137 lines not on this page · raw message

tools/perf/util/stat-shadow.c +11 -1
index 2c41d47f6f83..8c46c172a457 100644
@@ -9,6 +9,8 @@
 #include "expr.h"
 #include "metricgroup.h"
 #include <linux/zalloc.h>
+#include "iiostat.h"
+#include "counts.h"
 
 /*
  * AGGR_GLOBAL: Use CPU 0
@@ -814,6 +816,12 @@ static void generic_metric(struct perf_stat_config *config,
 		zfree(&pctx.ids[i].name);
 }
 
+__weak void iiostat_print_metric(struct perf_stat_config *config __maybe_unused,
+				  struct evsel *evsel __maybe_unused,
+				  struct perf_stat_output_ctx *out __maybe_unused)
+{
+}
+
 void perf_stat__print_shadow_stats(struct perf_stat_config *config,
 				   struct evsel *evsel,
 				   double avg, int cpu,
@@ -829,7 +837,9 @@ void perf_stat__print_shadow_stats(struct perf_stat_config *config,
 	struct metric_event *me;
 	int num = 1;
 
-	if (perf_evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) {
+	if (config->iiostat_run) {
+		iiostat_print_metric(config, evsel, out);
+	} else if (perf_evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) {
 		total = runtime_stat_avg(st, STAT_CYCLES, ctx, cpu);
 
 		if (total) {
tools/perf/util/stat.c +2 -1
index ebdd130557fb..672d33c1cafe 100644
@@ -318,6 +318,7 @@ process_counter_values(struct perf_stat_config *config, struct evsel *evsel,
 		}
 		break;
 	case AGGR_GLOBAL:
+	case AGGR_DEVICE:
 		aggr->val += count->val;
 		aggr->ena += count->ena;
 		aggr->run += count->run;
@@ -377,7 +378,7 @@ int perf_stat_process_counter(struct perf_stat_config *config,
 	if (ret)
 		return ret;
 
-	if (config->aggr_mode != AGGR_GLOBAL)
+	if (config->aggr_mode != AGGR_GLOBAL && config->aggr_mode != AGGR_DEVICE)
 		return 0;
 
 	if (!counter->snapshot)
tools/perf/util/stat.h +2 -0
index edbeb2f63e8d..be65afdaad90 100644
@@ -46,6 +46,7 @@ enum aggr_mode {
 	AGGR_DIE,
 	AGGR_CORE,
 	AGGR_THREAD,
+	AGGR_DEVICE,
 	AGGR_UNSET,
 };
 
@@ -106,6 +107,7 @@ struct perf_stat_config {
 	bool			 big_num;
 	bool			 no_merge;
 	bool			 walltime_run_table;
+	bool			 iiostat_run;
 	FILE			*output;
 	unsigned int		 interval;
 	unsigned int		 timeout;
Signature · 1 line
2.19.1
#5 lore raw

PATCH 5/6 perf tools: Add feature check for libpci

From: Roman Sudarikov <[email protected]>

Add feature check for libpci to show device name in --iiostat mode. libpci support allows device name to b:d:f notion.

---
 tools/build/Makefile.feature      |  2 ++
 tools/build/feature/Makefile      |  4 ++++
 tools/build/feature/test-all.c    |  5 +++++
 tools/build/feature/test-libpci.c | 10 ++++++++++
 tools/perf/Makefile.config        | 10 ++++++++++
 tools/perf/builtin-version.c      |  1 +
 tools/perf/tests/make             |  1 +
 7 files changed, 33 insertions(+)
 create mode 100644 tools/build/feature/test-libpci.c
tools/build/Makefile.feature +2 -0
index 8a19753cc26a..bebdfb99607c 100644
@@ -50,6 +50,7 @@ FEATURE_TESTS_BASIC :=                  \
         libelf-mmap                     \
         libnuma                         \
         numa_num_possible_cpus          \
+        libpci                          \
         libperl                         \
         libpython                       \
         libpython-version               \
@@ -115,6 +116,7 @@ FEATURE_DISPLAY ?=              \
          libelf                 \
          libnuma                \
          numa_num_possible_cpus \
+         libpci                 \
          libperl                \
          libpython              \
          libcrypto              \
tools/build/feature/Makefile +4 -0
index 8499385365c0..f0d5f886602d 100644
@@ -28,6 +28,7 @@ FILES=                                          \
          test-libelf-mmap.bin                   \
          test-libnuma.bin                       \
          test-numa_num_possible_cpus.bin        \
+         test-libpci.bin                        \
          test-libperl.bin                       \
          test-libpython.bin                     \
          test-libpython-version.bin             \
@@ -210,6 +211,9 @@ PERL_EMBED_LIBADD = $(call grep-libs,$(PERL_EMBED_LDOPTS))
 PERL_EMBED_CCOPTS = `perl -MExtUtils::Embed -e ccopts 2>/dev/null`
 FLAGS_PERL_EMBED=$(PERL_EMBED_CCOPTS) $(PERL_EMBED_LDOPTS)
 
+$(OUTPUT)test-libpci.bin:
+	$(BUILD) -lpci
+
 $(OUTPUT)test-libperl.bin:
 	$(BUILD) $(FLAGS_PERL_EMBED)
 
tools/build/feature/test-all.c +5 -0
index 88145e8cde1a..c61d34804a06 100644
@@ -74,6 +74,10 @@
 # include "test-libunwind.c"
 #undef main
 
+#define main main_test_libpci
+# include "test-libpci.c"
+#undef
+
 #define main main_test_libaudit
 # include "test-libaudit.c"
 #undef main
@@ -210,6 +214,7 @@ int main(int argc, char *argv[])
 	main_test_libunwind();
 	main_test_libaudit();
 	main_test_libslang();
+	main_test_libpci();
 	main_test_gtk2(argc, argv);
 	main_test_gtk2_infobar(argc, argv);
 	main_test_libbfd();
tools/build/feature/test-libpci.c new file +10 -0
new file mode 100644
index 000000000000..4bbeb9ffd687
@@ -0,0 +1,10 @@
+// SPDX-License-Identifier: GPL-2.0
+#include "pci/pci.h"
+
+int main(void)
+{
+	struct pci_access *pacc = pci_alloc();
+
+	pci_cleanup(pacc);
+	return 0;
+}
tools/perf/Makefile.config +10 -0
index 46f7fba2306c..1b9d341492c8 100644
@@ -839,6 +839,16 @@ ifndef NO_LIBCAP
   endif
 endif
 
+ifndef NO_PCILIB
+  ifeq ($(feature-libpci), 1)
+    CFLAGS += -DHAVE_LIBPCI_SUPPORT
+    EXTLIBS += -lpci
+  else
+    msg := $(warning No libpci found, show pci devices without names in iiostat mode, please install libpci-dev/pciutils-devel);
+    NO_PCILIB := 1
+  endif
+endif
+
 ifndef NO_BACKTRACE
   ifeq ($(feature-backtrace), 1)
     CFLAGS += -DHAVE_BACKTRACE_SUPPORT
tools/perf/builtin-version.c +1 -0
index 05cf2af9e2c2..ec4e0eb07825 100644
@@ -76,6 +76,7 @@ static void library_status(void)
 	STATUS(HAVE_LIBUNWIND_SUPPORT, libunwind);
 	STATUS(HAVE_DWARF_SUPPORT, libdw-dwarf-unwind);
 	STATUS(HAVE_ZLIB_SUPPORT, zlib);
+	STATUS(HAVE_LIBPCI_SUPPORT, libpci);
 	STATUS(HAVE_LZMA_SUPPORT, lzma);
 	STATUS(HAVE_AUXTRACE_SUPPORT, get_cpuid);
 	STATUS(HAVE_LIBBPF_SUPPORT, bpf);
tools/perf/tests/make +1 -0
index c850d1664c56..0b78cf6e8377 100644
@@ -109,6 +109,7 @@ make_minimal        += NO_LIBNUMA=1 NO_LIBAUDIT=1 NO_LIBBIONIC=1
 make_minimal        += NO_LIBDW_DWARF_UNWIND=1 NO_AUXTRACE=1 NO_LIBBPF=1
 make_minimal        += NO_LIBCRYPTO=1 NO_SDT=1 NO_JVMTI=1 NO_LIBZSTD=1
 make_minimal        += NO_LIBCAP=1
+make_minimal        += NO_LIBPCI=1
 
 # $(run) contains all available tests
 run := make_pure
Signature · 1 line
2.19.1
#6 lore raw

PATCH 6/6 perf stat: Add PCI device name to --iiostat output

From: Roman Sudarikov <[email protected]>

Example:

   $ perf stat --iiostat=show

Sample output w/o libpci:

    S0-RootPort0-uncore_iio_0<00:00.0>
    S1-RootPort2-uncore_iio_2<af:00.0>

Sample output with libpci:

    S0-RootPort0-uncore_iio_0<00:00.0 Sky Lake-E DMI3 Registers>
    S1-RootPort2-uncore_iio_2<af:00.0 Omni-Path HFI Silicon 100 Series [discrete]>
---
 tools/perf/arch/x86/util/iiostat.c | 15 ++++++++--
 tools/perf/util/pci.c              | 46 ++++++++++++++++++++++++++++++
 tools/perf/util/pci.h              |  4 +++
 3 files changed, 63 insertions(+), 2 deletions(-)
tools/perf/arch/x86/util/iiostat.c +13 -2
index 058a01d3a93f..7aad994e4936 100644
@@ -113,6 +113,7 @@ struct dev_info {
 	u8 die;
 	u8 pmu_idx;
 	u8 root_port_nr;
+	char *name;
 };
 
 struct iio_device {
@@ -210,7 +211,12 @@ static struct iio_device *iio_device_new(struct dev_info *info)
 	if (p) {
 		INIT_LIST_HEAD(&(p->node));
 		p->dev_info = *info;
+		p->dev_info.name = strdup(pci_device_name(info->bdf));
 		p->idx = -1;
+		if (!p->dev_info.name) {
+			free(p);
+			p = NULL;
+		}
 	}
 	return p;
 }
@@ -219,6 +225,7 @@ static void iio_device_delete(struct iio_device *device)
 {
 	if (device) {
 		list_del_init(&(device->node));
+		free(device->dev_info.name);
 		free(device);
 	}
 }
@@ -227,11 +234,11 @@ static void iiostat_device_show(FILE *output,
 			const struct iio_device * const device)
 {
 	if (output && device)
-		fprintf(output, "S%d-RootPort%d-uncore_iio_%d<%02x:%02x.%x>\n",
+		fprintf(output, "S%d-RootPort%d-uncore_iio_%d<%02x:%02x.%x %s>\n",
 			device->dev_info.die,
 			device->dev_info.root_port_nr, device->dev_info.pmu_idx,
 			device->dev_info.bdf.busno, device->dev_info.bdf.devno,
-			device->dev_info.bdf.funcno);
+			device->dev_info.bdf.funcno, device->dev_info.name);
 }
 
 static struct iio_devs_list *iio_devs_list_new(void)
@@ -426,9 +433,12 @@ static int iio_devs_scan(struct iio_devs_list **list)
 	if (!rp_nr)
 		return -ENOMEM;
 
+	pci_library_init();
+
 	interim = iio_devs_list_new();
 	if (!interim) {
 		free(rp_nr);
+		pci_library_cleanup();
 		return -ENOMEM;
 	}
 
@@ -457,6 +467,7 @@ static int iio_devs_scan(struct iio_devs_list **list)
 		iio_devs_list_free(interim);
 
 	free(rp_nr);
+	pci_library_cleanup();
 
 	return ret;
 }
tools/perf/util/pci.c +46 -0

58 lines not on this page · raw message

tools/perf/util/pci.h +4 -0
index e963b12e10e7..8d8551360419 100644
@@ -17,6 +17,10 @@ struct bdf {
 	u8 funcno;
 };
 
+void pci_library_init(void);
+void pci_library_cleanup(void);
+
+char *pci_device_name(struct bdf bdf);
 bool pci_device_probe(struct bdf bdf);
 bool is_pci_device_root_port(struct bdf bdf, u8 *secondary, u8 *subordinate);
 
Signature · 1 line
2.19.1
maintainer
#7 lore raw

Re: [PATCH 1/6] perf x86: Infrastructure for exposing an Uncore unit to PMON mapping

In reply to [email protected]
Quoted 62 lines · On Tue, Nov 26, 2019 at 07:36:25PM +0300, [email protected] wrote:

From: Roman Sudarikov <[email protected]>

Intel® Xeon® Scalable processor family (code name Skylake-SP) makes significant changes in the integrated I/O (IIO) architecture. The new solution introduces IIO stacks which are responsible for managing traffic between the PCIe domain and the Mesh domain. Each IIO stack has its own PMON block and can handle either DMI port, x16 PCIe root port, MCP-Link or various built-in accelerators. IIO PMON blocks allow concurrent monitoring of I/O flows up to 4 x4 bifurcation within each IIO stack.

Software is supposed to program required perf counters within each IIO stack and gather performance data. The tricky thing here is that IIO PMON reports data per IIO stack but users have no idea what IIO stacks are - they only know devices which are connected to the platform.

Understanding IIO stack concept to find which IIO stack that particular IO device is connected to, or to identify an IIO PMON block to program for monitoring specific IIO stack assumes a lot of implicit knowledge about given Intel server platform architecture.

This patch set introduces:
An infrastructure for exposing an Uncore unit to Uncore PMON mapping through sysfs-backend A new --iiostat mode in perf stat to provide I/O performance metrics per I/O device

Current version supports a server line starting Intel® Xeon® Processor Scalable Family and introduces mapping for IIO Uncore units only.
Other units can be added on demand.

Usage example:
/sys/devices/uncore_<type>_<pmu_idx>/platform_mapping

Each Uncore unit type, by its nature, can be mapped to its own context, for example:
CHA - each uncore_cha_<pmu_idx> is assigned to manage a distinct slice of LLC capacity UPI - each uncore_upi_<pmu_idx> is assigned to manage one link of Intel UPI Subsystem IIO - each uncore_iio_<pmu_idx> is assigned to manage one stack of the IIO module IMC - each uncore_imc_<pmu_idx> is assigned to manage one channel of Memory Controller

Implementation details:
Two callbacks added to struct intel_uncore_type to discover and map Uncore units to PMONs:
int (*get_topology)(struct intel_uncore_type *type)
int (*set_mapping)(struct intel_uncore_type *type)

IIO stack to PMON mapping is exposed through
/sys/devices/uncore_iio_<pmu_idx>/platform_mapping in the following format: domain:bus

Details of IIO Uncore unit mapping to IIO PMON:
Each IIO stack is either a DMI port, x16 PCIe root port, MCP-Link or various built-in accelerators. For Uncore IIO Unit type, the platform_mapping file holds bus numbers of devices, which can be monitored by that IIO PMON block on each die.

For example, on a 4-die Intel Xeon® server platform:

    $ cat /sys/devices/uncore_iio_0/platform_mapping
    0000:00,0000:40,0000:80,0000:c0

Which means:
IIO PMON block 0 on die 0 belongs to IIO stack located on bus 0x00, domain 0x0000 IIO PMON block 0 on die 1 belongs to IIO stack located on bus 0x40, domain 0x0000 IIO PMON block 0 on die 2 belongs to IIO stack located on bus 0x80, domain 0x0000 IIO PMON block 0 on die 3 belongs to IIO stack located on bus 0xc0, domain 0x0000

Quoted 3 lines

Kan, can you help these people? There's a ton of process fail with this submission. From SoB chain to CodingStyle to git-sendmail threading.

#8 lore raw

Re: [PATCH 1/6] perf x86: Infrastructure for exposing an Uncore unit to PMON mapping

In reply to [email protected]
Quoted 54 lines · On Tue, Nov 26, 2019 at 8:36 AM <[email protected]> wrote:

From: Roman Sudarikov <[email protected]>

Intel® Xeon® Scalable processor family (code name Skylake-SP) makes significant changes in the integrated I/O (IIO) architecture. The new solution introduces IIO stacks which are responsible for managing traffic between the PCIe domain and the Mesh domain. Each IIO stack has its own PMON block and can handle either DMI port, x16 PCIe root port, MCP-Link or various built-in accelerators. IIO PMON blocks allow concurrent monitoring of I/O flows up to 4 x4 bifurcation within each IIO stack.

Software is supposed to program required perf counters within each IIO stack and gather performance data. The tricky thing here is that IIO PMON reports data per IIO stack but users have no idea what IIO stacks are - they only know devices which are connected to the platform.

Understanding IIO stack concept to find which IIO stack that particular IO device is connected to, or to identify an IIO PMON block to program for monitoring specific IIO stack assumes a lot of implicit knowledge about given Intel server platform architecture.

This patch set introduces:
An infrastructure for exposing an Uncore unit to Uncore PMON mapping through sysfs-backend A new --iiostat mode in perf stat to provide I/O performance metrics per I/O device

Current version supports a server line starting Intel® Xeon® Processor Scalable Family and introduces mapping for IIO Uncore units only.
Other units can be added on demand.

Usage example:
/sys/devices/uncore_<type>_<pmu_idx>/platform_mapping

Each Uncore unit type, by its nature, can be mapped to its own context, for example:
CHA - each uncore_cha_<pmu_idx> is assigned to manage a distinct slice of LLC capacity UPI - each uncore_upi_<pmu_idx> is assigned to manage one link of Intel UPI Subsystem IIO - each uncore_iio_<pmu_idx> is assigned to manage one stack of the IIO module IMC - each uncore_imc_<pmu_idx> is assigned to manage one channel of Memory Controller

Implementation details:
Two callbacks added to struct intel_uncore_type to discover and map Uncore units to PMONs:
int (*get_topology)(struct intel_uncore_type *type)
int (*set_mapping)(struct intel_uncore_type *type)

IIO stack to PMON mapping is exposed through
/sys/devices/uncore_iio_<pmu_idx>/platform_mapping in the following format: domain:bus

Details of IIO Uncore unit mapping to IIO PMON:
Each IIO stack is either a DMI port, x16 PCIe root port, MCP-Link or various built-in accelerators. For Uncore IIO Unit type, the platform_mapping file holds bus numbers of devices, which can be monitored by that IIO PMON block on each die.

For example, on a 4-die Intel Xeon® server platform:

Quoted 9 lines
    $ cat /sys/devices/uncore_iio_0/platform_mapping
    0000:00,0000:40,0000:80,0000:c0

Which means:
IIO PMON block 0 on die 0 belongs to IIO stack located on bus 0x00, domain 0x0000 IIO PMON block 0 on die 1 belongs to IIO stack located on bus 0x40, domain 0x0000 IIO PMON block 0 on die 2 belongs to IIO stack located on bus 0x80, domain 0x0000 IIO PMON block 0 on die 3 belongs to IIO stack located on bus 0xc0, domain 0x0000

You are just looking at one die (package). How does your enumeration help figure out
is the iio_0 is on socket0 of socket1 and then figure out which bus/domain in on which
socket.

And how does that help map actual devices (using the output of lspci) to the IIO?
You need to show how you would do that, which is really what people want, with what you
have in your patch right now.

Quoted 368 lines
---
 arch/x86/events/intel/uncore.c       |  61 +++++++++++-
 arch/x86/events/intel/uncore.h       |  13 ++-
 arch/x86/events/intel/uncore_snbep.c | 144 +++++++++++++++++++++++++++
 3 files changed, 214 insertions(+), 4 deletions(-)
arch/x86/events/intel/uncore.c +58 -3
index 86467f85c383..0f779c8fcc05 100644
@@ -18,6 +18,11 @@ struct list_head pci2phy_map_head = LIST_HEAD_INIT(pci2phy_map_head);
 struct pci_extra_dev *uncore_extra_pci_dev;
 static int max_dies;
 
+int get_max_dies(void)
+{
+       return max_dies;
+}
+
 /* mask of cpus that collect uncore events */
 static cpumask_t uncore_cpu_mask;
 
@@ -816,6 +821,16 @@ static ssize_t uncore_get_attr_cpumask(struct device *dev,
 
 static DEVICE_ATTR(cpumask, S_IRUGO, uncore_get_attr_cpumask, NULL);
 
+static ssize_t platform_mapping_show(struct device *dev,
+                               struct device_attribute *attr, char *buf)
+{
+       struct intel_uncore_pmu *pmu = dev_get_drvdata(dev);
+
+       return snprintf(buf, PAGE_SIZE - 1, "%s\n", pmu->platform_mapping ?
+                      (char *)pmu->platform_mapping : "0");
+}
+static DEVICE_ATTR_RO(platform_mapping);
+
 static struct attribute *uncore_pmu_attrs[] = {
        &dev_attr_cpumask.attr,
        NULL,
@@ -825,6 +840,15 @@ static const struct attribute_group uncore_pmu_attr_group = {
        .attrs = uncore_pmu_attrs,
 };
 
+static struct attribute *platform_attrs[] = {
+       &dev_attr_platform_mapping.attr,
+       NULL,
+};
+
+static const struct attribute_group uncore_platform_discovery_group = {
+       .attrs = platform_attrs,
+};
+
 static int uncore_pmu_register(struct intel_uncore_pmu *pmu)
 {
        int ret;
@@ -905,11 +929,27 @@ static void uncore_types_exit(struct intel_uncore_type **types)
                uncore_type_exit(*types);
 }
 
+static void uncore_type_attrs_compaction(struct intel_uncore_type *type)
+{
+       int i, j;
+
+       for (i = 0, j = 0; i < UNCORE_MAX_NUM_ATTR_GROUP; i++) {
+               if (!type->attr_groups[i])
+                       continue;
+               if (i > j) {
+                       type->attr_groups[j] = type->attr_groups[i];
+                       type->attr_groups[i] = NULL;
+               }
+               j++;
+       }
+}
+
 static int __init uncore_type_init(struct intel_uncore_type *type, bool setid)
 {
        struct intel_uncore_pmu *pmus;
        size_t size;
        int i, j;
+       int ret;
 
        pmus = kcalloc(type->num_boxes, sizeof(*pmus), GFP_KERNEL);
        if (!pmus)
@@ -922,8 +962,10 @@ static int __init uncore_type_init(struct intel_uncore_type *type, bool setid)
                pmus[i].pmu_idx = i;
                pmus[i].type    = type;
                pmus[i].boxes   = kzalloc(size, GFP_KERNEL);
-               if (!pmus[i].boxes)
+               if (!pmus[i].boxes) {
+                       ret = -ENOMEM;
                        goto err;
+               }
        }
 
        type->pmus = pmus;
@@ -940,8 +982,10 @@ static int __init uncore_type_init(struct intel_uncore_type *type, bool setid)
 
                attr_group = kzalloc(struct_size(attr_group, attrs, i + 1),
                                                                GFP_KERNEL);
-               if (!attr_group)
+               if (!attr_group) {
+                       ret = -ENOMEM;
                        goto err;
+               }
 
                attr_group->group.name = "events";
                attr_group->group.attrs = attr_group->attrs;
@@ -954,6 +998,17 @@ static int __init uncore_type_init(struct intel_uncore_type *type, bool setid)
 
        type->pmu_group = &uncore_pmu_attr_group;
 
+       /*
+        * Exposing mapping of Uncore units to corresponding Uncore PMUs
+        * through /sys/devices/uncore_<type>_<idx>/platform_mapping
+        */
+       if (type->get_topology && type->set_mapping)
+               if (!type->get_topology(type) && !type->set_mapping(type))
+                       type->platform_discovery = &uncore_platform_discovery_group;
+
+       /* For optional attributes, we can safely remove embedded NULL attr_groups elements */
+       uncore_type_attrs_compaction(type);
+
        return 0;
 
 err:
@@ -961,7 +1016,7 @@ static int __init uncore_type_init(struct intel_uncore_type *type, bool setid)
                kfree(pmus[i].boxes);
        kfree(pmus);
 
-       return -ENOMEM;
+       return ret;
 }
 
 static int __init
arch/x86/events/intel/uncore.h +12 -1
index bbfdaa720b45..ce3727b9f7f8 100644
@@ -43,6 +43,8 @@ struct intel_uncore_box;
 struct uncore_event_desc;
 struct freerunning_counters;
 
+#define UNCORE_MAX_NUM_ATTR_GROUP 5
+
 struct intel_uncore_type {
        const char *name;
        int num_counters;
@@ -71,13 +73,19 @@ struct intel_uncore_type {
        struct intel_uncore_ops *ops;
        struct uncore_event_desc *event_descs;
        struct freerunning_counters *freerunning;
-       const struct attribute_group *attr_groups[4];
+       const struct attribute_group *attr_groups[UNCORE_MAX_NUM_ATTR_GROUP];
        struct pmu *pmu; /* for custom pmu ops */
+       void *platform_topology;
+       /* finding Uncore units */
+       int (*get_topology)(struct intel_uncore_type *type);
+       /* mapping Uncore units to PMON ranges */
+       int (*set_mapping)(struct intel_uncore_type *type);
 };
 
 #define pmu_group attr_groups[0]
 #define format_group attr_groups[1]
 #define events_group attr_groups[2]
+#define platform_discovery attr_groups[3]
 
 struct intel_uncore_ops {
        void (*init_box)(struct intel_uncore_box *);
@@ -99,6 +107,7 @@ struct intel_uncore_pmu {
        int                             pmu_idx;
        int                             func_id;
        bool                            registered;
+       void                            *platform_mapping;
        atomic_t                        activeboxes;
        struct intel_uncore_type        *type;
        struct intel_uncore_box         **boxes;
@@ -490,6 +499,8 @@ static inline struct intel_uncore_box *uncore_event_to_box(struct perf_event *ev
        return event->pmu_private;
 }
 
+int get_max_dies(void);
+
 struct intel_uncore_box *uncore_pmu_to_box(struct intel_uncore_pmu *pmu, int cpu);
 u64 uncore_msr_read_counter(struct intel_uncore_box *box, struct perf_event *event);
 void uncore_mmio_exit_box(struct intel_uncore_box *box);
arch/x86/events/intel/uncore_snbep.c +144 -1
index b10a5ec79e48..92ce9fbafde1 100644
@@ -273,6 +273,28 @@
 #define SKX_CPUNODEID                  0xc0
 #define SKX_GIDNIDMAP                  0xd4
 
+/*
+ * The CPU_BUS_NUMBER MSR returns the values of the respective CPUBUSNO CSR
+ * that BIOS programmed. MSR has package scope.
+ * |  Bit  |  Default  |  Description
+ * | [63]  |    00h    | VALID - When set, indicates the CPU bus
+ *                       numbers have been initialized. (RO)
+ * |[62:48]|    ---    | Reserved
+ * |[47:40]|    00h    | BUS_NUM_5 — Return the bus number BIOS assigned
+ *                       CPUBUSNO(5). (RO)
+ * |[39:32]|    00h    | BUS_NUM_4 — Return the bus number BIOS assigned
+ *                       CPUBUSNO(4). (RO)
+ * |[31:24]|    00h    | BUS_NUM_3 — Return the bus number BIOS assigned
+ *                       CPUBUSNO(3). (RO)
+ * |[23:16]|    00h    | BUS_NUM_2 — Return the bus number BIOS assigned
+ *                       CPUBUSNO(2). (RO)
+ * |[15:8] |    00h    | BUS_NUM_1 — Return the bus number BIOS assigned
+ *                       CPUBUSNO(1). (RO)
+ * | [7:0] |    00h    | BUS_NUM_0 — Return the bus number BIOS assigned
+ *                       CPUBUSNO(0). (RO)
+ */
+#define SKX_MSR_CPU_BUS_NUMBER         0x300
+
 /* SKX CHA */
 #define SKX_CHA_MSR_PMON_BOX_FILTER_TID                (0x1ffULL << 0)
 #define SKX_CHA_MSR_PMON_BOX_FILTER_LINK       (0xfULL << 9)
@@ -3580,6 +3602,9 @@ static struct intel_uncore_ops skx_uncore_iio_ops = {
        .read_counter           = uncore_msr_read_counter,
 };
 
+static int skx_iio_get_topology(struct intel_uncore_type *type);
+static int skx_iio_set_mapping(struct intel_uncore_type *type);
+
 static struct intel_uncore_type skx_uncore_iio = {
        .name                   = "iio",
        .num_counters           = 4,
@@ -3594,6 +3619,8 @@ static struct intel_uncore_type skx_uncore_iio = {
        .constraints            = skx_uncore_iio_constraints,
        .ops                    = &skx_uncore_iio_ops,
        .format_group           = &skx_uncore_iio_format_group,
+       .get_topology           = skx_iio_get_topology,
+       .set_mapping            = skx_iio_set_mapping,
 };
 
 enum perf_uncore_iio_freerunning_type_id {
@@ -3780,6 +3807,123 @@ static int skx_count_chabox(void)
        return hweight32(val);
 }
 
+static inline u8 skx_iio_topology_byte(void *platform_topology,
+                                       int die, int idx)
+{
+       return *((u8 *)(platform_topology) + die * sizeof(u64) + idx);
+}
+
+static inline bool skx_iio_topology_valid(u64 msr_value)
+{
+       return msr_value & ((u64)1 << 63);
+}
+
+static int skx_msr_cpu_bus_read(int cpu, int die)
+{
+       int ret = rdmsrl_on_cpu(cpu, SKX_MSR_CPU_BUS_NUMBER,
+                               ((u64 *)skx_uncore_iio.platform_topology) + die);
+
+       if (!ret) {
+               if (!skx_iio_topology_valid(*(((u64 *)skx_uncore_iio.platform_topology) + die)))
+                       ret = -1;
+       }
+       return ret;
+}
+
+static int skx_iio_get_topology(struct intel_uncore_type *type)
+{
+       int ret, cpu, die, current_die;
+       struct pci_bus *bus = NULL;
+
+       while ((bus = pci_find_next_bus(bus)) != NULL)
+               if (pci_domain_nr(bus)) {
+                       pr_info("Mapping of I/O stack to PMON ranges is not supported for multi-segment topology\n");
+                       return -1;
+               }
+
+       /* Size of SKX_MSR_CPU_BUS_NUMBER is 8 bytes, the MSR has package scope.*/
+       type->platform_topology =
+               kzalloc(get_max_dies() * sizeof(u64), GFP_KERNEL);
+       if (!type->platform_topology)
+               return -ENOMEM;
+
+       /*
+        * Using cpus_read_lock() to ensure cpu is not going down between
+        * looking at cpu_online_mask.
+        */
+       cpus_read_lock();
+       /* Invalid value to start loop.*/
+       current_die = -1;
+       for_each_online_cpu(cpu) {
+               die = topology_logical_die_id(cpu);
+               if (current_die == die)
+                       continue;
+               ret = skx_msr_cpu_bus_read(cpu, die);
+               if (ret)
+                       break;
+               current_die = die;
+       }
+       cpus_read_unlock();
+
+       if (ret)
+               kfree(type->platform_topology);
+       return ret;
+}
+
+static int skx_iio_set_mapping(struct intel_uncore_type *type)
+{
+       /*
+        * Each IIO stack (PCIe root port) has its own IIO PMON block, so each
+        * platform_mapping holds bus number(s) of PCIe root port(s), which can
+        * be monitored by that IIO PMON block.
+        *
+        * For example, on a 4-die Xeon platform with up to 6 IIO stacks per die
+        * and, therefore, 6 IIO PMON blocks per die, the platform_mapping of IIO
+        * PMON block 0 holds "0000:00,0000:40,0000:80,0000:c0":
+        *
+        * $ cat /sys/devices/uncore_iio_0/platform_mapping
+        * 0000:00,0000:40,0000:80,0000:c0
+        *
+        * Which means:
+        * IIO PMON block 0 on the die 0 belongs to PCIe root port located on bus 0x00, domain 0x0000
+        * IIO PMON block 0 on the die 1 belongs to PCIe root port located on bus 0x40, domain 0x0000
+        * IIO PMON block 0 on the die 2 belongs to PCIe root port located on bus 0x80, domain 0x0000
+        * IIO PMON block 0 on the die 3 belongs to PCIe root port located on bus 0xc0, domain 0x0000
+        */
+
+       int ret = 0;
+       int die, i;
+       char *buf;
+       struct intel_uncore_pmu *pmu;
+       const int template_len = 8;
+
+       for (i = 0; i < type->num_boxes; i++) {
+               pmu = type->pmus + i;
+               /* Root bus 0x00 is valid only for die 0 AND pmu_idx = 0. */
+               if (skx_iio_topology_byte(type->platform_topology, 0, pmu->pmu_idx) || (!pmu->pmu_idx)) {
+                       pmu->platform_mapping =
+                               kzalloc(get_max_dies() * template_len + 1, GFP_KERNEL);
+                       if (pmu->platform_mapping) {
+                               buf = (char *)pmu->platform_mapping;
+                               for (die = 0; die < get_max_dies(); die++)
+                                       buf += snprintf(buf, template_len + 1, "%04x:%02x,", 0,
+                                               skx_iio_topology_byte(type->platform_topology,
+                                                                     die, pmu->pmu_idx));
+
+                               *(--buf) = '\0';
+                       } else {
+                               for (; i >= 0; i--)
+                                       kfree((type->pmus + i)->platform_mapping);
+                               ret = -ENOMEM;
+                               break;
+                       }
+               }
+       }
+
+       kfree(type->platform_topology);
+       return ret;
+}
+
 void skx_uncore_cpu_init(void)
 {
        skx_uncore_chabox.num_boxes = skx_count_chabox();
--

2.19.1

#9 lore raw

Re: [PATCH 1/6] perf x86: Infrastructure for exposing an Uncore unit to PMON mapping

Quoted 11 lines

You are just looking at one die (package). How does your enumeration help figure out
is the iio_0 is on socket0 of socket1 and then figure out which bus/domain in on which
socket.

And how does that help map actual devices (using the output of lspci) to the IIO?
You need to show how you would do that, which is really what people want, with what you
have in your patch right now.

See the rest of the patch series. It implements all of this in the perf tool.

-Andi

#10 lore raw

Re: [PATCH 1/6] perf x86: Infrastructure for exposing an Uncore unit to PMON mapping

In reply to Stephane Eranian
Quoted 69 lines · Stephane Eranian
Quoted 53 lines · On Tue, Nov 26, 2019 at 8:36 AM <[email protected]> wrote:

From: Roman Sudarikov <[email protected]>

Intel® Xeon® Scalable processor family (code name Skylake-SP) makes significant changes in the integrated I/O (IIO) architecture. The new solution introduces IIO stacks which are responsible for managing traffic between the PCIe domain and the Mesh domain. Each IIO stack has its own PMON block and can handle either DMI port, x16 PCIe root port, MCP-Link or various built-in accelerators. IIO PMON blocks allow concurrent monitoring of I/O flows up to 4 x4 bifurcation within each IIO stack.

Software is supposed to program required perf counters within each IIO stack and gather performance data. The tricky thing here is that IIO PMON reports data per IIO stack but users have no idea what IIO stacks are - they only know devices which are connected to the platform.

Understanding IIO stack concept to find which IIO stack that particular IO device is connected to, or to identify an IIO PMON block to program for monitoring specific IIO stack assumes a lot of implicit knowledge about given Intel server platform architecture.

This patch set introduces:
An infrastructure for exposing an Uncore unit to Uncore PMON mapping through sysfs-backend A new --iiostat mode in perf stat to provide I/O performance metrics per I/O device

Current version supports a server line starting Intel® Xeon® Processor Scalable Family and introduces mapping for IIO Uncore units only.
Other units can be added on demand.

Usage example:
/sys/devices/uncore_<type>_<pmu_idx>/platform_mapping

Each Uncore unit type, by its nature, can be mapped to its own context, for example:
CHA - each uncore_cha_<pmu_idx> is assigned to manage a distinct slice of LLC capacity UPI - each uncore_upi_<pmu_idx> is assigned to manage one link of Intel UPI Subsystem IIO - each uncore_iio_<pmu_idx> is assigned to manage one stack of the IIO module IMC - each uncore_imc_<pmu_idx> is assigned to manage one channel of Memory Controller

Implementation details:
Two callbacks added to struct intel_uncore_type to discover and map Uncore units to PMONs:
int (*get_topology)(struct intel_uncore_type *type)
int (*set_mapping)(struct intel_uncore_type *type)

IIO stack to PMON mapping is exposed through
/sys/devices/uncore_iio_<pmu_idx>/platform_mapping in the following format: domain:bus

Details of IIO Uncore unit mapping to IIO PMON:
Each IIO stack is either a DMI port, x16 PCIe root port, MCP-Link or various built-in accelerators. For Uncore IIO Unit type, the platform_mapping file holds bus numbers of devices, which can be monitored by that IIO PMON block on each die.

For example, on a 4-die Intel Xeon® server platform:

Quoted 9 lines
     $ cat /sys/devices/uncore_iio_0/platform_mapping
     0000:00,0000:40,0000:80,0000:c0

Which means:
IIO PMON block 0 on die 0 belongs to IIO stack located on bus 0x00, domain 0x0000 IIO PMON block 0 on die 1 belongs to IIO stack located on bus 0x40, domain 0x0000 IIO PMON block 0 on die 2 belongs to IIO stack located on bus 0x80, domain 0x0000 IIO PMON block 0 on die 3 belongs to IIO stack located on bus 0xc0, domain 0x0000

You are just looking at one die (package). How does your enumeration help figure out
is the iio_0 is on socket0 of socket1 and then figure out which bus/domain in on which
socket.

Quoted 5 lines

And how does that help map actual devices (using the output of lspci) to the IIO?
You need to show how you would do that, which is really what people want, with what you
have in your patch right now.

No. I'm enumerating all IIO PMUs for all sockets on the platform.

Let's take an 4 socket SKX as an example - sysfs exposes 6 instances of IIO PMU and each socket has its own instance of each IIO PMUs, meaning that socket 0 has its own IIO PMU0, socket 1 also has its own IIO PMU0 and so on. Same apply for IIO PMUs 1 through 5. Below is sample output:

$:/sys/devices# cat uncore_iio_0/platform_mapping
0000:00,0000:40,0000:80,0000:c0
$:/sys/devices# cat uncore_iio_1/platform_mapping
0000:16,0000:44,0000:84,0000:c4
$:/sys/devices# cat uncore_iio_2/platform_mapping
0000:24,0000:58,0000:98,0000:d8
$:/sys/devices# cat uncore_iio_3/platform_mapping
0000:32,0000:6c,0000:ac,0000:ec

Technically, the idea is as follows - kernel part of the feature is for locating IIO stacks and creating  IIO PMON to IIO stack mapping. Userspace part of the feature is for locating IO devices connected to each IIO stack on each socket and configure only required IIO counters to provide 4 IO performance metrics - Inbound Read, Inbound Write, Outbound Read, Outbound Write - attributed to each device.

Follow up patches show how users can benefit from the feature; see https://lkml.org/lkml/2019/11/26/451

Below is sample output:

1. show mode

./perf stat --iiostat=show

    S0-RootPort0-uncore_iio_0<00:00.0 Sky Lake-E DMI3 Registers>
    S1-RootPort0-uncore_iio_0<81:00.0 Ethernet Controller X710 for 10GbE SFP+>
    S0-RootPort1-uncore_iio_1<18:00.0 Omni-Path HFI Silicon 100 Series [discrete]>
    S1-RootPort1-uncore_iio_1<86:00.0 Ethernet Controller XL710 for 40GbE QSFP+>
    S1-RootPort1-uncore_iio_1<88:00.0 Ethernet Controller XL710 for 40GbE QSFP+>
    S0-RootPort2-uncore_iio_2<3d:00.0 Ethernet Connection X722 for 10GBASE-T>
    S1-RootPort2-uncore_iio_2<af:00.0 Omni-Path HFI Silicon 100 Series [discrete]>
    S1-RootPort3-uncore_iio_3<da:00.0 NVMe Datacenter SSD [Optane]>

For example, NIC at 81:00.0 is local to S1, connected to its RootPort0 and is covered by IIO PMU0 (on socket 1)

1. collector mode

  ./perf stat --iiostat -- dd if=/dev/zero of=/dev/nvme0n1 bs=1M oflag=direct
    357708+0 records in
    357707+0 records out
    375083606016 bytes (375 GB, 349 GiB) copied, 215.381 s, 1.7 GB/s

Performance counter stats for 'system wide':

device Inbound Read(MB) Inbound Write(MB) Outbound Read(MB) Outbound Write(MB)
00:00.0 0 0 0 0
81:00.0 0 0 0 0
18:00.0 0 0 0 0
86:00.0 0 0 0 0
88:00.0 0 0 0 0
3b:00.0 3 0 0 0
3c:03.0 3 0 0 0
3d:00.0 3 0 0 0
af:00.0 0 0 0 0
da:00.0 358559 44 0 22

215.383783574 seconds time elapsed

Quoted 368 lines
Quoted 368 lines
---
  arch/x86/events/intel/uncore.c       |  61 +++++++++++-
  arch/x86/events/intel/uncore.h       |  13 ++-
  arch/x86/events/intel/uncore_snbep.c | 144 +++++++++++++++++++++++++++
  3 files changed, 214 insertions(+), 4 deletions(-)
arch/x86/events/intel/uncore.c +58 -3
index 86467f85c383..0f779c8fcc05 100644
@@ -18,6 +18,11 @@ struct list_head pci2phy_map_head = LIST_HEAD_INIT(pci2phy_map_head);
  struct pci_extra_dev *uncore_extra_pci_dev;
  static int max_dies;
 
+int get_max_dies(void)
+{
+       return max_dies;
+}
+
  /* mask of cpus that collect uncore events */
  static cpumask_t uncore_cpu_mask;
 
@@ -816,6 +821,16 @@ static ssize_t uncore_get_attr_cpumask(struct device *dev,
 
  static DEVICE_ATTR(cpumask, S_IRUGO, uncore_get_attr_cpumask, NULL);
 
+static ssize_t platform_mapping_show(struct device *dev,
+                               struct device_attribute *attr, char *buf)
+{
+       struct intel_uncore_pmu *pmu = dev_get_drvdata(dev);
+
+       return snprintf(buf, PAGE_SIZE - 1, "%s\n", pmu->platform_mapping ?
+                      (char *)pmu->platform_mapping : "0");
+}
+static DEVICE_ATTR_RO(platform_mapping);
+
  static struct attribute *uncore_pmu_attrs[] = {
         &dev_attr_cpumask.attr,
         NULL,
@@ -825,6 +840,15 @@ static const struct attribute_group uncore_pmu_attr_group = {
         .attrs = uncore_pmu_attrs,
  };
 
+static struct attribute *platform_attrs[] = {
+       &dev_attr_platform_mapping.attr,
+       NULL,
+};
+
+static const struct attribute_group uncore_platform_discovery_group = {
+       .attrs = platform_attrs,
+};
+
  static int uncore_pmu_register(struct intel_uncore_pmu *pmu)
  {
         int ret;
@@ -905,11 +929,27 @@ static void uncore_types_exit(struct intel_uncore_type **types)
                 uncore_type_exit(*types);
  }
 
+static void uncore_type_attrs_compaction(struct intel_uncore_type *type)
+{
+       int i, j;
+
+       for (i = 0, j = 0; i < UNCORE_MAX_NUM_ATTR_GROUP; i++) {
+               if (!type->attr_groups[i])
+                       continue;
+               if (i > j) {
+                       type->attr_groups[j] = type->attr_groups[i];
+                       type->attr_groups[i] = NULL;
+               }
+               j++;
+       }
+}
+
  static int __init uncore_type_init(struct intel_uncore_type *type, bool setid)
  {
         struct intel_uncore_pmu *pmus;
         size_t size;
         int i, j;
+       int ret;
 
         pmus = kcalloc(type->num_boxes, sizeof(*pmus), GFP_KERNEL);
         if (!pmus)
@@ -922,8 +962,10 @@ static int __init uncore_type_init(struct intel_uncore_type *type, bool setid)
                 pmus[i].pmu_idx = i;
                 pmus[i].type    = type;
                 pmus[i].boxes   = kzalloc(size, GFP_KERNEL);
-               if (!pmus[i].boxes)
+               if (!pmus[i].boxes) {
+                       ret = -ENOMEM;
                         goto err;
+               }
         }
 
         type->pmus = pmus;
@@ -940,8 +982,10 @@ static int __init uncore_type_init(struct intel_uncore_type *type, bool setid)
 
                 attr_group = kzalloc(struct_size(attr_group, attrs, i + 1),
                                                                 GFP_KERNEL);
-               if (!attr_group)
+               if (!attr_group) {
+                       ret = -ENOMEM;
                         goto err;
+               }
 
                 attr_group->group.name = "events";
                 attr_group->group.attrs = attr_group->attrs;
@@ -954,6 +998,17 @@ static int __init uncore_type_init(struct intel_uncore_type *type, bool setid)
 
         type->pmu_group = &uncore_pmu_attr_group;
 
+       /*
+        * Exposing mapping of Uncore units to corresponding Uncore PMUs
+        * through /sys/devices/uncore_<type>_<idx>/platform_mapping
+        */
+       if (type->get_topology && type->set_mapping)
+               if (!type->get_topology(type) && !type->set_mapping(type))
+                       type->platform_discovery = &uncore_platform_discovery_group;
+
+       /* For optional attributes, we can safely remove embedded NULL attr_groups elements */
+       uncore_type_attrs_compaction(type);
+
         return 0;
 
  err:
@@ -961,7 +1016,7 @@ static int __init uncore_type_init(struct intel_uncore_type *type, bool setid)
                 kfree(pmus[i].boxes);
         kfree(pmus);
 
-       return -ENOMEM;
+       return ret;
  }
 
  static int __init
arch/x86/events/intel/uncore.h +12 -1
index bbfdaa720b45..ce3727b9f7f8 100644
@@ -43,6 +43,8 @@ struct intel_uncore_box;
  struct uncore_event_desc;
  struct freerunning_counters;
 
+#define UNCORE_MAX_NUM_ATTR_GROUP 5
+
  struct intel_uncore_type {
         const char *name;
         int num_counters;
@@ -71,13 +73,19 @@ struct intel_uncore_type {
         struct intel_uncore_ops *ops;
         struct uncore_event_desc *event_descs;
         struct freerunning_counters *freerunning;
-       const struct attribute_group *attr_groups[4];
+       const struct attribute_group *attr_groups[UNCORE_MAX_NUM_ATTR_GROUP];
         struct pmu *pmu; /* for custom pmu ops */
+       void *platform_topology;
+       /* finding Uncore units */
+       int (*get_topology)(struct intel_uncore_type *type);
+       /* mapping Uncore units to PMON ranges */
+       int (*set_mapping)(struct intel_uncore_type *type);
  };
 
  #define pmu_group attr_groups[0]
  #define format_group attr_groups[1]
  #define events_group attr_groups[2]
+#define platform_discovery attr_groups[3]
 
  struct intel_uncore_ops {
         void (*init_box)(struct intel_uncore_box *);
@@ -99,6 +107,7 @@ struct intel_uncore_pmu {
         int                             pmu_idx;
         int                             func_id;
         bool                            registered;
+       void                            *platform_mapping;
         atomic_t                        activeboxes;
         struct intel_uncore_type        *type;
         struct intel_uncore_box         **boxes;
@@ -490,6 +499,8 @@ static inline struct intel_uncore_box *uncore_event_to_box(struct perf_event *ev
         return event->pmu_private;
  }
 
+int get_max_dies(void);
+
  struct intel_uncore_box *uncore_pmu_to_box(struct intel_uncore_pmu *pmu, int cpu);
  u64 uncore_msr_read_counter(struct intel_uncore_box *box, struct perf_event *event);
  void uncore_mmio_exit_box(struct intel_uncore_box *box);
arch/x86/events/intel/uncore_snbep.c +144 -1
index b10a5ec79e48..92ce9fbafde1 100644
@@ -273,6 +273,28 @@
  #define SKX_CPUNODEID                  0xc0
  #define SKX_GIDNIDMAP                  0xd4
 
+/*
+ * The CPU_BUS_NUMBER MSR returns the values of the respective CPUBUSNO CSR
+ * that BIOS programmed. MSR has package scope.
+ * |  Bit  |  Default  |  Description
+ * | [63]  |    00h    | VALID - When set, indicates the CPU bus
+ *                       numbers have been initialized. (RO)
+ * |[62:48]|    ---    | Reserved
+ * |[47:40]|    00h    | BUS_NUM_5 — Return the bus number BIOS assigned
+ *                       CPUBUSNO(5). (RO)
+ * |[39:32]|    00h    | BUS_NUM_4 — Return the bus number BIOS assigned
+ *                       CPUBUSNO(4). (RO)
+ * |[31:24]|    00h    | BUS_NUM_3 — Return the bus number BIOS assigned
+ *                       CPUBUSNO(3). (RO)
+ * |[23:16]|    00h    | BUS_NUM_2 — Return the bus number BIOS assigned
+ *                       CPUBUSNO(2). (RO)
+ * |[15:8] |    00h    | BUS_NUM_1 — Return the bus number BIOS assigned
+ *                       CPUBUSNO(1). (RO)
+ * | [7:0] |    00h    | BUS_NUM_0 — Return the bus number BIOS assigned
+ *                       CPUBUSNO(0). (RO)
+ */
+#define SKX_MSR_CPU_BUS_NUMBER         0x300
+
  /* SKX CHA */
  #define SKX_CHA_MSR_PMON_BOX_FILTER_TID                (0x1ffULL << 0)
  #define SKX_CHA_MSR_PMON_BOX_FILTER_LINK       (0xfULL << 9)
@@ -3580,6 +3602,9 @@ static struct intel_uncore_ops skx_uncore_iio_ops = {
         .read_counter           = uncore_msr_read_counter,
  };
 
+static int skx_iio_get_topology(struct intel_uncore_type *type);
+static int skx_iio_set_mapping(struct intel_uncore_type *type);
+
  static struct intel_uncore_type skx_uncore_iio = {
         .name                   = "iio",
         .num_counters           = 4,
@@ -3594,6 +3619,8 @@ static struct intel_uncore_type skx_uncore_iio = {
         .constraints            = skx_uncore_iio_constraints,
         .ops                    = &skx_uncore_iio_ops,
         .format_group           = &skx_uncore_iio_format_group,
+       .get_topology           = skx_iio_get_topology,
+       .set_mapping            = skx_iio_set_mapping,
  };
 
  enum perf_uncore_iio_freerunning_type_id {
@@ -3780,6 +3807,123 @@ static int skx_count_chabox(void)
         return hweight32(val);
  }
 
+static inline u8 skx_iio_topology_byte(void *platform_topology,
+                                       int die, int idx)
+{
+       return *((u8 *)(platform_topology) + die * sizeof(u64) + idx);
+}
+
+static inline bool skx_iio_topology_valid(u64 msr_value)
+{
+       return msr_value & ((u64)1 << 63);
+}
+
+static int skx_msr_cpu_bus_read(int cpu, int die)
+{
+       int ret = rdmsrl_on_cpu(cpu, SKX_MSR_CPU_BUS_NUMBER,
+                               ((u64 *)skx_uncore_iio.platform_topology) + die);
+
+       if (!ret) {
+               if (!skx_iio_topology_valid(*(((u64 *)skx_uncore_iio.platform_topology) + die)))
+                       ret = -1;
+       }
+       return ret;
+}
+
+static int skx_iio_get_topology(struct intel_uncore_type *type)
+{
+       int ret, cpu, die, current_die;
+       struct pci_bus *bus = NULL;
+
+       while ((bus = pci_find_next_bus(bus)) != NULL)
+               if (pci_domain_nr(bus)) {
+                       pr_info("Mapping of I/O stack to PMON ranges is not supported for multi-segment topology\n");
+                       return -1;
+               }
+
+       /* Size of SKX_MSR_CPU_BUS_NUMBER is 8 bytes, the MSR has package scope.*/
+       type->platform_topology =
+               kzalloc(get_max_dies() * sizeof(u64), GFP_KERNEL);
+       if (!type->platform_topology)
+               return -ENOMEM;
+
+       /*
+        * Using cpus_read_lock() to ensure cpu is not going down between
+        * looking at cpu_online_mask.
+        */
+       cpus_read_lock();
+       /* Invalid value to start loop.*/
+       current_die = -1;
+       for_each_online_cpu(cpu) {
+               die = topology_logical_die_id(cpu);
+               if (current_die == die)
+                       continue;
+               ret = skx_msr_cpu_bus_read(cpu, die);
+               if (ret)
+                       break;
+               current_die = die;
+       }
+       cpus_read_unlock();
+
+       if (ret)
+               kfree(type->platform_topology);
+       return ret;
+}
+
+static int skx_iio_set_mapping(struct intel_uncore_type *type)
+{
+       /*
+        * Each IIO stack (PCIe root port) has its own IIO PMON block, so each
+        * platform_mapping holds bus number(s) of PCIe root port(s), which can
+        * be monitored by that IIO PMON block.
+        *
+        * For example, on a 4-die Xeon platform with up to 6 IIO stacks per die
+        * and, therefore, 6 IIO PMON blocks per die, the platform_mapping of IIO
+        * PMON block 0 holds "0000:00,0000:40,0000:80,0000:c0":
+        *
+        * $ cat /sys/devices/uncore_iio_0/platform_mapping
+        * 0000:00,0000:40,0000:80,0000:c0
+        *
+        * Which means:
+        * IIO PMON block 0 on the die 0 belongs to PCIe root port located on bus 0x00, domain 0x0000
+        * IIO PMON block 0 on the die 1 belongs to PCIe root port located on bus 0x40, domain 0x0000
+        * IIO PMON block 0 on the die 2 belongs to PCIe root port located on bus 0x80, domain 0x0000
+        * IIO PMON block 0 on the die 3 belongs to PCIe root port located on bus 0xc0, domain 0x0000
+        */
+
+       int ret = 0;
+       int die, i;
+       char *buf;
+       struct intel_uncore_pmu *pmu;
+       const int template_len = 8;
+
+       for (i = 0; i < type->num_boxes; i++) {
+               pmu = type->pmus + i;
+               /* Root bus 0x00 is valid only for die 0 AND pmu_idx = 0. */
+               if (skx_iio_topology_byte(type->platform_topology, 0, pmu->pmu_idx) || (!pmu->pmu_idx)) {
+                       pmu->platform_mapping =
+                               kzalloc(get_max_dies() * template_len + 1, GFP_KERNEL);
+                       if (pmu->platform_mapping) {
+                               buf = (char *)pmu->platform_mapping;
+                               for (die = 0; die < get_max_dies(); die++)
+                                       buf += snprintf(buf, template_len + 1, "%04x:%02x,", 0,
+                                               skx_iio_topology_byte(type->platform_topology,
+                                                                     die, pmu->pmu_idx));
+
+                               *(--buf) = '\0';
+                       } else {
+                               for (; i >= 0; i--)
+                                       kfree((type->pmus + i)->platform_mapping);
+                               ret = -ENOMEM;
+                               break;
+                       }
+               }
+       }
+
+       kfree(type->platform_topology);
+       return ret;
+}
+
  void skx_uncore_cpu_init(void)
  {
         skx_uncore_chabox.num_boxes = skx_count_chabox();
--

2.19.1

#11 lore raw

Re: [PATCH 1/6] perf x86: Infrastructure for exposing an Uncore unit to PMON mapping

Quoted 76 lines · Roman
Quoted 69 lines · Stephane Eranian
Quoted 53 lines · On Tue, Nov 26, 2019 at 8:36 AM <[email protected]> wrote:

From: Roman Sudarikov <[email protected]>

Intel® Xeon® Scalable processor family (code name Skylake-SP) makes significant changes in the integrated I/O (IIO) architecture. The new solution introduces IIO stacks which are responsible for managing traffic between the PCIe domain and the Mesh domain. Each IIO stack has its own PMON block and can handle either DMI port, x16 PCIe root port, MCP-Link or various built-in accelerators. IIO PMON blocks allow concurrent monitoring of I/O flows up to 4 x4 bifurcation within each IIO stack.

Software is supposed to program required perf counters within each IIO stack and gather performance data. The tricky thing here is that IIO PMON reports data per IIO stack but users have no idea what IIO stacks are - they only know devices which are connected to the platform.

Understanding IIO stack concept to find which IIO stack that particular IO device is connected to, or to identify an IIO PMON block to program for monitoring specific IIO stack assumes a lot of implicit knowledge about given Intel server platform architecture.

This patch set introduces:
An infrastructure for exposing an Uncore unit to Uncore PMON mapping through sysfs-backend A new --iiostat mode in perf stat to provide I/O performance metrics per I/O device

Current version supports a server line starting Intel® Xeon® Processor Scalable Family and introduces mapping for IIO Uncore units only.
Other units can be added on demand.

Usage example:
/sys/devices/uncore_<type>_<pmu_idx>/platform_mapping

Each Uncore unit type, by its nature, can be mapped to its own context, for example:
CHA - each uncore_cha_<pmu_idx> is assigned to manage a distinct slice of LLC capacity UPI - each uncore_upi_<pmu_idx> is assigned to manage one link of Intel UPI Subsystem IIO - each uncore_iio_<pmu_idx> is assigned to manage one stack of the IIO module IMC - each uncore_imc_<pmu_idx> is assigned to manage one channel of Memory Controller

Implementation details:
Two callbacks added to struct intel_uncore_type to discover and map Uncore units to PMONs:
int (*get_topology)(struct intel_uncore_type *type)
int (*set_mapping)(struct intel_uncore_type *type)

IIO stack to PMON mapping is exposed through
/sys/devices/uncore_iio_<pmu_idx>/platform_mapping in the following format: domain:bus

Details of IIO Uncore unit mapping to IIO PMON:
Each IIO stack is either a DMI port, x16 PCIe root port, MCP-Link or various built-in accelerators. For Uncore IIO Unit type, the platform_mapping file holds bus numbers of devices, which can be monitored by that IIO PMON block on each die.

For example, on a 4-die Intel Xeon® server platform:

Quoted 9 lines
     $ cat /sys/devices/uncore_iio_0/platform_mapping
     0000:00,0000:40,0000:80,0000:c0

Which means:
IIO PMON block 0 on die 0 belongs to IIO stack located on bus 0x00, domain 0x0000 IIO PMON block 0 on die 1 belongs to IIO stack located on bus 0x40, domain 0x0000 IIO PMON block 0 on die 2 belongs to IIO stack located on bus 0x80, domain 0x0000 IIO PMON block 0 on die 3 belongs to IIO stack located on bus 0xc0, domain 0x0000

You are just looking at one die (package). How does your enumeration help figure out
is the iio_0 is on socket0 of socket1 and then figure out which bus/domain in on which
socket.

Quoted 5 lines

And how does that help map actual devices (using the output of lspci) to the IIO?
You need to show how you would do that, which is really what people want, with what you
have in your patch right now.

Quoted 6 lines

No. I'm enumerating all IIO PMUs for all sockets on the platform.

Let's take an 4 socket SKX as an example - sysfs exposes 6 instances of IIO PMU and each socket has its own instance of each IIO PMUs, meaning that socket 0 has its own IIO PMU0, socket 1 also has its own IIO PMU0 and so on. Same apply for IIO PMUs 1 through 5.

I know that.

Quoted 19 lines

Below is sample output:

$:/sys/devices# cat uncore_iio_0/platform_mapping
0000:00,0000:40,0000:80,0000:c0
$:/sys/devices# cat uncore_iio_1/platform_mapping
0000:16,0000:44,0000:84,0000:c4
$:/sys/devices# cat uncore_iio_2/platform_mapping
0000:24,0000:58,0000:98,0000:d8
$:/sys/devices# cat uncore_iio_3/platform_mapping
0000:32,0000:6c,0000:ac,0000:ec

Technically, the idea is as follows - kernel part of the feature is for locating IIO stacks and creating IIO PMON to IIO stack mapping. Userspace part of the feature is for locating IO devices connected to each IIO stack on each socket and configure only required IIO counters to provide 4 IO performance metrics - Inbound Read, Inbound Write, Outbound Read, Outbound Write - attributed to each device.

Quoted 3 lines

Follow up patches show how users can benefit from the feature; see
https://lkml.org/lkml/2019/11/26/451

I know this is useful. I have done this for internal users a long time ago.

Quoted 26 lines

Below is sample output:

1. show mode

./perf stat --iiostat=show

     S0-RootPort0-uncore_iio_0<00:00.0 Sky Lake-E DMI3 Registers>
     S1-RootPort0-uncore_iio_0<81:00.0 Ethernet Controller X710 for 10GbE SFP+>
     S0-RootPort1-uncore_iio_1<18:00.0 Omni-Path HFI Silicon 100 Series [discrete]>
     S1-RootPort1-uncore_iio_1<86:00.0 Ethernet Controller XL710 for 40GbE QSFP+>
     S1-RootPort1-uncore_iio_1<88:00.0 Ethernet Controller XL710 for 40GbE QSFP+>
     S0-RootPort2-uncore_iio_2<3d:00.0 Ethernet Connection X722 for 10GBASE-T>
     S1-RootPort2-uncore_iio_2<af:00.0 Omni-Path HFI Silicon 100 Series [discrete]>
     S1-RootPort3-uncore_iio_3<da:00.0 NVMe Datacenter SSD [Optane]>

For example, NIC at 81:00.0 is local to S1, connected to its RootPort0 and is covered by IIO PMU0 (on socket 1)

1. collector mode

   ./perf stat --iiostat -- dd if=/dev/zero of=/dev/nvme0n1 bs=1M oflag=direct
     357708+0 records in
     357707+0 records out
     375083606016 bytes (375 GB, 349 GiB) copied, 215.381 s, 1.7 GB/s

Performance counter stats for 'system wide':

Quoted 12 lines

device Inbound Read(MB) Inbound Write(MB) Outbound Read(MB) Outbound Write(MB)
00:00.0 0 0 0 0
81:00.0 0 0 0 0
18:00.0 0 0 0 0
86:00.0 0 0 0 0
88:00.0 0 0 0 0
3b:00.0 3 0 0 0
3c:03.0 3 0 0 0
3d:00.0 3 0 0 0
af:00.0 0 0 0 0
da:00.0 358559 44 0 22

I think this output would be more useful with the socket information. People care about NUMA locality. This output
does not cover that (in a single cmdline). It would also benefit from having the actual Linux device names, e.g., sda, ssda, eth0, ....,

Quoted 371 lines
     215.383783574 seconds time elapsed
Quoted 368 lines
Quoted 368 lines
---
  arch/x86/events/intel/uncore.c       |  61 +++++++++++-
  arch/x86/events/intel/uncore.h       |  13 ++-
  arch/x86/events/intel/uncore_snbep.c | 144 +++++++++++++++++++++++++++
  3 files changed, 214 insertions(+), 4 deletions(-)
arch/x86/events/intel/uncore.c +58 -3
index 86467f85c383..0f779c8fcc05 100644
@@ -18,6 +18,11 @@ struct list_head pci2phy_map_head = LIST_HEAD_INIT(pci2phy_map_head);
  struct pci_extra_dev *uncore_extra_pci_dev;
  static int max_dies;
 
+int get_max_dies(void)
+{
+       return max_dies;
+}
+
  /* mask of cpus that collect uncore events */
  static cpumask_t uncore_cpu_mask;
 
@@ -816,6 +821,16 @@ static ssize_t uncore_get_attr_cpumask(struct device *dev,
 
  static DEVICE_ATTR(cpumask, S_IRUGO, uncore_get_attr_cpumask, NULL);
 
+static ssize_t platform_mapping_show(struct device *dev,
+                               struct device_attribute *attr, char *buf)
+{
+       struct intel_uncore_pmu *pmu = dev_get_drvdata(dev);
+
+       return snprintf(buf, PAGE_SIZE - 1, "%s\n", pmu->platform_mapping ?
+                      (char *)pmu->platform_mapping : "0");
+}
+static DEVICE_ATTR_RO(platform_mapping);
+
  static struct attribute *uncore_pmu_attrs[] = {
         &dev_attr_cpumask.attr,
         NULL,
@@ -825,6 +840,15 @@ static const struct attribute_group uncore_pmu_attr_group = {
         .attrs = uncore_pmu_attrs,
  };
 
+static struct attribute *platform_attrs[] = {
+       &dev_attr_platform_mapping.attr,
+       NULL,
+};
+
+static const struct attribute_group uncore_platform_discovery_group = {
+       .attrs = platform_attrs,
+};
+
  static int uncore_pmu_register(struct intel_uncore_pmu *pmu)
  {
         int ret;
@@ -905,11 +929,27 @@ static void uncore_types_exit(struct intel_uncore_type **types)
                 uncore_type_exit(*types);
  }
 
+static void uncore_type_attrs_compaction(struct intel_uncore_type *type)
+{
+       int i, j;
+
+       for (i = 0, j = 0; i < UNCORE_MAX_NUM_ATTR_GROUP; i++) {
+               if (!type->attr_groups[i])
+                       continue;
+               if (i > j) {
+                       type->attr_groups[j] = type->attr_groups[i];
+                       type->attr_groups[i] = NULL;
+               }
+               j++;
+       }
+}
+
  static int __init uncore_type_init(struct intel_uncore_type *type, bool setid)
  {
         struct intel_uncore_pmu *pmus;
         size_t size;
         int i, j;
+       int ret;
 
         pmus = kcalloc(type->num_boxes, sizeof(*pmus), GFP_KERNEL);
         if (!pmus)
@@ -922,8 +962,10 @@ static int __init uncore_type_init(struct intel_uncore_type *type, bool setid)
                 pmus[i].pmu_idx = i;
                 pmus[i].type    = type;
                 pmus[i].boxes   = kzalloc(size, GFP_KERNEL);
-               if (!pmus[i].boxes)
+               if (!pmus[i].boxes) {
+                       ret = -ENOMEM;
                         goto err;
+               }
         }
 
         type->pmus = pmus;
@@ -940,8 +982,10 @@ static int __init uncore_type_init(struct intel_uncore_type *type, bool setid)
 
                 attr_group = kzalloc(struct_size(attr_group, attrs, i + 1),
                                                                 GFP_KERNEL);
-               if (!attr_group)
+               if (!attr_group) {
+                       ret = -ENOMEM;
                         goto err;
+               }
 
                 attr_group->group.name = "events";
                 attr_group->group.attrs = attr_group->attrs;
@@ -954,6 +998,17 @@ static int __init uncore_type_init(struct intel_uncore_type *type, bool setid)
 
         type->pmu_group = &uncore_pmu_attr_group;
 
+       /*
+        * Exposing mapping of Uncore units to corresponding Uncore PMUs
+        * through /sys/devices/uncore_<type>_<idx>/platform_mapping
+        */
+       if (type->get_topology && type->set_mapping)
+               if (!type->get_topology(type) && !type->set_mapping(type))
+                       type->platform_discovery = &uncore_platform_discovery_group;
+
+       /* For optional attributes, we can safely remove embedded NULL attr_groups elements */
+       uncore_type_attrs_compaction(type);
+
         return 0;
 
  err:
@@ -961,7 +1016,7 @@ static int __init uncore_type_init(struct intel_uncore_type *type, bool setid)
                 kfree(pmus[i].boxes);
         kfree(pmus);
 
-       return -ENOMEM;
+       return ret;
  }
 
  static int __init
arch/x86/events/intel/uncore.h +12 -1
index bbfdaa720b45..ce3727b9f7f8 100644
@@ -43,6 +43,8 @@ struct intel_uncore_box;
  struct uncore_event_desc;
  struct freerunning_counters;
 
+#define UNCORE_MAX_NUM_ATTR_GROUP 5
+
  struct intel_uncore_type {
         const char *name;
         int num_counters;
@@ -71,13 +73,19 @@ struct intel_uncore_type {
         struct intel_uncore_ops *ops;
         struct uncore_event_desc *event_descs;
         struct freerunning_counters *freerunning;
-       const struct attribute_group *attr_groups[4];
+       const struct attribute_group *attr_groups[UNCORE_MAX_NUM_ATTR_GROUP];
         struct pmu *pmu; /* for custom pmu ops */
+       void *platform_topology;
+       /* finding Uncore units */
+       int (*get_topology)(struct intel_uncore_type *type);
+       /* mapping Uncore units to PMON ranges */
+       int (*set_mapping)(struct intel_uncore_type *type);
  };
 
  #define pmu_group attr_groups[0]
  #define format_group attr_groups[1]
  #define events_group attr_groups[2]
+#define platform_discovery attr_groups[3]
 
  struct intel_uncore_ops {
         void (*init_box)(struct intel_uncore_box *);
@@ -99,6 +107,7 @@ struct intel_uncore_pmu {
         int                             pmu_idx;
         int                             func_id;
         bool                            registered;
+       void                            *platform_mapping;
         atomic_t                        activeboxes;
         struct intel_uncore_type        *type;
         struct intel_uncore_box         **boxes;
@@ -490,6 +499,8 @@ static inline struct intel_uncore_box *uncore_event_to_box(struct perf_event *ev
         return event->pmu_private;
  }
 
+int get_max_dies(void);
+
  struct intel_uncore_box *uncore_pmu_to_box(struct intel_uncore_pmu *pmu, int cpu);
  u64 uncore_msr_read_counter(struct intel_uncore_box *box, struct perf_event *event);
  void uncore_mmio_exit_box(struct intel_uncore_box *box);
arch/x86/events/intel/uncore_snbep.c +144 -1
index b10a5ec79e48..92ce9fbafde1 100644
@@ -273,6 +273,28 @@
  #define SKX_CPUNODEID                  0xc0
  #define SKX_GIDNIDMAP                  0xd4
 
+/*
+ * The CPU_BUS_NUMBER MSR returns the values of the respective CPUBUSNO CSR
+ * that BIOS programmed. MSR has package scope.
+ * |  Bit  |  Default  |  Description
+ * | [63]  |    00h    | VALID - When set, indicates the CPU bus
+ *                       numbers have been initialized. (RO)
+ * |[62:48]|    ---    | Reserved
+ * |[47:40]|    00h    | BUS_NUM_5 — Return the bus number BIOS assigned
+ *                       CPUBUSNO(5). (RO)
+ * |[39:32]|    00h    | BUS_NUM_4 — Return the bus number BIOS assigned
+ *                       CPUBUSNO(4). (RO)
+ * |[31:24]|    00h    | BUS_NUM_3 — Return the bus number BIOS assigned
+ *                       CPUBUSNO(3). (RO)
+ * |[23:16]|    00h    | BUS_NUM_2 — Return the bus number BIOS assigned
+ *                       CPUBUSNO(2). (RO)
+ * |[15:8] |    00h    | BUS_NUM_1 — Return the bus number BIOS assigned
+ *                       CPUBUSNO(1). (RO)
+ * | [7:0] |    00h    | BUS_NUM_0 — Return the bus number BIOS assigned
+ *                       CPUBUSNO(0). (RO)
+ */
+#define SKX_MSR_CPU_BUS_NUMBER         0x300
+
  /* SKX CHA */
  #define SKX_CHA_MSR_PMON_BOX_FILTER_TID                (0x1ffULL << 0)
  #define SKX_CHA_MSR_PMON_BOX_FILTER_LINK       (0xfULL << 9)
@@ -3580,6 +3602,9 @@ static struct intel_uncore_ops skx_uncore_iio_ops = {
         .read_counter           = uncore_msr_read_counter,
  };
 
+static int skx_iio_get_topology(struct intel_uncore_type *type);
+static int skx_iio_set_mapping(struct intel_uncore_type *type);
+
  static struct intel_uncore_type skx_uncore_iio = {
         .name                   = "iio",
         .num_counters           = 4,
@@ -3594,6 +3619,8 @@ static struct intel_uncore_type skx_uncore_iio = {
         .constraints            = skx_uncore_iio_constraints,
         .ops                    = &skx_uncore_iio_ops,
         .format_group           = &skx_uncore_iio_format_group,
+       .get_topology           = skx_iio_get_topology,
+       .set_mapping            = skx_iio_set_mapping,
  };
 
  enum perf_uncore_iio_freerunning_type_id {
@@ -3780,6 +3807,123 @@ static int skx_count_chabox(void)
         return hweight32(val);
  }
 
+static inline u8 skx_iio_topology_byte(void *platform_topology,
+                                       int die, int idx)
+{
+       return *((u8 *)(platform_topology) + die * sizeof(u64) + idx);
+}
+
+static inline bool skx_iio_topology_valid(u64 msr_value)
+{
+       return msr_value & ((u64)1 << 63);
+}
+
+static int skx_msr_cpu_bus_read(int cpu, int die)
+{
+       int ret = rdmsrl_on_cpu(cpu, SKX_MSR_CPU_BUS_NUMBER,
+                               ((u64 *)skx_uncore_iio.platform_topology) + die);
+
+       if (!ret) {
+               if (!skx_iio_topology_valid(*(((u64 *)skx_uncore_iio.platform_topology) + die)))
+                       ret = -1;
+       }
+       return ret;
+}
+
+static int skx_iio_get_topology(struct intel_uncore_type *type)
+{
+       int ret, cpu, die, current_die;
+       struct pci_bus *bus = NULL;
+
+       while ((bus = pci_find_next_bus(bus)) != NULL)
+               if (pci_domain_nr(bus)) {
+                       pr_info("Mapping of I/O stack to PMON ranges is not supported for multi-segment topology\n");
+                       return -1;
+               }
+
+       /* Size of SKX_MSR_CPU_BUS_NUMBER is 8 bytes, the MSR has package scope.*/
+       type->platform_topology =
+               kzalloc(get_max_dies() * sizeof(u64), GFP_KERNEL);
+       if (!type->platform_topology)
+               return -ENOMEM;
+
+       /*
+        * Using cpus_read_lock() to ensure cpu is not going down between
+        * looking at cpu_online_mask.
+        */
+       cpus_read_lock();
+       /* Invalid value to start loop.*/
+       current_die = -1;
+       for_each_online_cpu(cpu) {
+               die = topology_logical_die_id(cpu);
+               if (current_die == die)
+                       continue;
+               ret = skx_msr_cpu_bus_read(cpu, die);
+               if (ret)
+                       break;
+               current_die = die;
+       }
+       cpus_read_unlock();
+
+       if (ret)
+               kfree(type->platform_topology);
+       return ret;
+}
+
+static int skx_iio_set_mapping(struct intel_uncore_type *type)
+{
+       /*
+        * Each IIO stack (PCIe root port) has its own IIO PMON block, so each
+        * platform_mapping holds bus number(s) of PCIe root port(s), which can
+        * be monitored by that IIO PMON block.
+        *
+        * For example, on a 4-die Xeon platform with up to 6 IIO stacks per die
+        * and, therefore, 6 IIO PMON blocks per die, the platform_mapping of IIO
+        * PMON block 0 holds "0000:00,0000:40,0000:80,0000:c0":
+        *
+        * $ cat /sys/devices/uncore_iio_0/platform_mapping
+        * 0000:00,0000:40,0000:80,0000:c0
+        *
+        * Which means:
+        * IIO PMON block 0 on the die 0 belongs to PCIe root port located on bus 0x00, domain 0x0000
+        * IIO PMON block 0 on the die 1 belongs to PCIe root port located on bus 0x40, domain 0x0000
+        * IIO PMON block 0 on the die 2 belongs to PCIe root port located on bus 0x80, domain 0x0000
+        * IIO PMON block 0 on the die 3 belongs to PCIe root port located on bus 0xc0, domain 0x0000
+        */
+
+       int ret = 0;
+       int die, i;
+       char *buf;
+       struct intel_uncore_pmu *pmu;
+       const int template_len = 8;
+
+       for (i = 0; i < type->num_boxes; i++) {
+               pmu = type->pmus + i;
+               /* Root bus 0x00 is valid only for die 0 AND pmu_idx = 0. */
+               if (skx_iio_topology_byte(type->platform_topology, 0, pmu->pmu_idx) || (!pmu->pmu_idx)) {
+                       pmu->platform_mapping =
+                               kzalloc(get_max_dies() * template_len + 1, GFP_KERNEL);
+                       if (pmu->platform_mapping) {
+                               buf = (char *)pmu->platform_mapping;
+                               for (die = 0; die < get_max_dies(); die++)
+                                       buf += snprintf(buf, template_len + 1, "%04x:%02x,", 0,
+                                               skx_iio_topology_byte(type->platform_topology,
+                                                                     die, pmu->pmu_idx));
+
+                               *(--buf) = '\0';
+                       } else {
+                               for (; i >= 0; i--)
+                                       kfree((type->pmus + i)->platform_mapping);
+                               ret = -ENOMEM;
+                               break;
+                       }
+               }
+       }
+
+       kfree(type->platform_topology);
+       return ret;
+}
+
  void skx_uncore_cpu_init(void)
  {
         skx_uncore_chabox.num_boxes = skx_count_chabox();
--

2.19.1

#12 lore raw

Re: [PATCH 1/6] perf x86: Infrastructure for exposing an Uncore unit to PMON mapping

Quoted 4 lines · Stephane Eranian

does not cover that (in a single cmdline).

It would also benefit from
having the actual Linux device names, e.g., sda, ssda, eth0, ....,

Some example code to do that mapping in the other direction is here

https://github.com/numactl/numactl/blob/master/affinity.c

-Andi

#13 lore raw

Re: [PATCH 1/6] perf x86: Infrastructure for exposing an Uncore unit to PMON mapping

In reply to Stephane Eranian
Quoted 147 lines · Stephane Eranian
Quoted 75 lines · Roman
Quoted 69 lines · Stephane Eranian
Quoted 53 lines · On Tue, Nov 26, 2019 at 8:36 AM <[email protected]> wrote:

From: Roman Sudarikov <[email protected]>

Intel® Xeon® Scalable processor family (code name Skylake-SP) makes significant changes in the integrated I/O (IIO) architecture. The new solution introduces IIO stacks which are responsible for managing traffic between the PCIe domain and the Mesh domain. Each IIO stack has its own PMON block and can handle either DMI port, x16 PCIe root port, MCP-Link or various built-in accelerators. IIO PMON blocks allow concurrent monitoring of I/O flows up to 4 x4 bifurcation within each IIO stack.

Software is supposed to program required perf counters within each IIO stack and gather performance data. The tricky thing here is that IIO PMON reports data per IIO stack but users have no idea what IIO stacks are - they only know devices which are connected to the platform.

Understanding IIO stack concept to find which IIO stack that particular IO device is connected to, or to identify an IIO PMON block to program for monitoring specific IIO stack assumes a lot of implicit knowledge about given Intel server platform architecture.

This patch set introduces:
An infrastructure for exposing an Uncore unit to Uncore PMON mapping through sysfs-backend A new --iiostat mode in perf stat to provide I/O performance metrics per I/O device

Current version supports a server line starting Intel® Xeon® Processor Scalable Family and introduces mapping for IIO Uncore units only.
Other units can be added on demand.

Usage example:
/sys/devices/uncore_<type>_<pmu_idx>/platform_mapping

Each Uncore unit type, by its nature, can be mapped to its own context, for example:
CHA - each uncore_cha_<pmu_idx> is assigned to manage a distinct slice of LLC capacity UPI - each uncore_upi_<pmu_idx> is assigned to manage one link of Intel UPI Subsystem IIO - each uncore_iio_<pmu_idx> is assigned to manage one stack of the IIO module IMC - each uncore_imc_<pmu_idx> is assigned to manage one channel of Memory Controller

Implementation details:
Two callbacks added to struct intel_uncore_type to discover and map Uncore units to PMONs:
int (*get_topology)(struct intel_uncore_type *type)
int (*set_mapping)(struct intel_uncore_type *type)

IIO stack to PMON mapping is exposed through
/sys/devices/uncore_iio_<pmu_idx>/platform_mapping in the following format: domain:bus

Details of IIO Uncore unit mapping to IIO PMON:
Each IIO stack is either a DMI port, x16 PCIe root port, MCP-Link or various built-in accelerators. For Uncore IIO Unit type, the platform_mapping file holds bus numbers of devices, which can be monitored by that IIO PMON block on each die.

For example, on a 4-die Intel Xeon® server platform:

Quoted 9 lines
      $ cat /sys/devices/uncore_iio_0/platform_mapping
      0000:00,0000:40,0000:80,0000:c0

Which means:
IIO PMON block 0 on die 0 belongs to IIO stack located on bus 0x00, domain 0x0000 IIO PMON block 0 on die 1 belongs to IIO stack located on bus 0x40, domain 0x0000 IIO PMON block 0 on die 2 belongs to IIO stack located on bus 0x80, domain 0x0000 IIO PMON block 0 on die 3 belongs to IIO stack located on bus 0xc0, domain 0x0000

You are just looking at one die (package). How does your enumeration help figure out
is the iio_0 is on socket0 of socket1 and then figure out which bus/domain in on which
socket.

Quoted 5 lines

And how does that help map actual devices (using the output of lspci) to the IIO?
You need to show how you would do that, which is really what people want, with what you
have in your patch right now.

Quoted 6 lines

No. I'm enumerating all IIO PMUs for all sockets on the platform.

Let's take an 4 socket SKX as an example - sysfs exposes 6 instances of IIO PMU and each socket has its own instance of each IIO PMUs, meaning that socket 0 has its own IIO PMU0, socket 1 also has its own IIO PMU0 and so on. Same apply for IIO PMUs 1 through 5.

I know that.

Quoted 19 lines

Below is sample output:

$:/sys/devices# cat uncore_iio_0/platform_mapping
0000:00,0000:40,0000:80,0000:c0
$:/sys/devices# cat uncore_iio_1/platform_mapping
0000:16,0000:44,0000:84,0000:c4
$:/sys/devices# cat uncore_iio_2/platform_mapping
0000:24,0000:58,0000:98,0000:d8
$:/sys/devices# cat uncore_iio_3/platform_mapping
0000:32,0000:6c,0000:ac,0000:ec

Technically, the idea is as follows - kernel part of the feature is for locating IIO stacks and creating IIO PMON to IIO stack mapping. Userspace part of the feature is for locating IO devices connected to each IIO stack on each socket and configure only required IIO counters to provide 4 IO performance metrics - Inbound Read, Inbound Write, Outbound Read, Outbound Write - attributed to each device.

Quoted 3 lines

Follow up patches show how users can benefit from the feature; see
https://lkml.org/lkml/2019/11/26/451

I know this is useful. I have done this for internal users a long time ago.

Quoted 26 lines

Below is sample output:

1. show mode

./perf stat --iiostat=show

      S0-RootPort0-uncore_iio_0<00:00.0 Sky Lake-E DMI3 Registers>
      S1-RootPort0-uncore_iio_0<81:00.0 Ethernet Controller X710 for 10GbE SFP+>
      S0-RootPort1-uncore_iio_1<18:00.0 Omni-Path HFI Silicon 100 Series [discrete]>
      S1-RootPort1-uncore_iio_1<86:00.0 Ethernet Controller XL710 for 40GbE QSFP+>
      S1-RootPort1-uncore_iio_1<88:00.0 Ethernet Controller XL710 for 40GbE QSFP+>
      S0-RootPort2-uncore_iio_2<3d:00.0 Ethernet Connection X722 for 10GBASE-T>
      S1-RootPort2-uncore_iio_2<af:00.0 Omni-Path HFI Silicon 100 Series [discrete]>
      S1-RootPort3-uncore_iio_3<da:00.0 NVMe Datacenter SSD [Optane]>

For example, NIC at 81:00.0 is local to S1, connected to its RootPort0 and is covered by IIO PMU0 (on socket 1)

1. collector mode

./perf stat --iiostat -- dd if=/dev/zero of=/dev/nvme0n1 bs=1M oflag=direct
357708+0 records in
357707+0 records out
375083606016 bytes (375 GB, 349 GiB) copied, 215.381 s, 1.7 GB/s

Performance counter stats for 'system wide':

Quoted 12 lines

device Inbound Read(MB) Inbound Write(MB) Outbound Read(MB) Outbound Write(MB)
00:00.0 0 0 0 0
81:00.0 0 0 0 0
18:00.0 0 0 0 0
86:00.0 0 0 0 0
88:00.0 0 0 0 0
3b:00.0 3 0 0 0
3c:03.0 3 0 0 0
3d:00.0 3 0 0 0
af:00.0 0 0 0 0
da:00.0 358559 44 0 22

Quoted 4 lines

I think this output would be more useful with the socket information. People care about NUMA locality. This output
does not cover that (in a single cmdline). It would also benefit from having the actual Linux device names, e.g., sda, ssda, eth0, ....,

Hi Stephane,

I still think we should keep b:d.f notion as a part of the output and, sure, we can add socket and device name information, so it will look like this:

Before:
   Performance counter stats for 'system wide':

  deviceInbound Read(MB) Inbound Write(MB)
 da:00.0

After:
   Performance counter stats for 'system wide':

           device             Inbound Read(MB)    Inbound Write(MB) S1<nvme0>da:00.0

Are you OK with that approach?

BTW, to address it we need code changes at userspace part only. Can we proceed with kernel patch review and once finalized, I'll send userspace part and we will figure out right output format?

Thanks,
Roman

Quoted 370 lines
Quoted 370 lines
      215.383783574 seconds time elapsed
Quoted 367 lines
Quoted 367 lines
---
   arch/x86/events/intel/uncore.c       |  61 +++++++++++-
   arch/x86/events/intel/uncore.h       |  13 ++-
   arch/x86/events/intel/uncore_snbep.c | 144 +++++++++++++++++++++++++++
   3 files changed, 214 insertions(+), 4 deletions(-)
arch/x86/events/intel/uncore.c +58 -3
index 86467f85c383..0f779c8fcc05 100644
@@ -18,6 +18,11 @@ struct list_head pci2phy_map_head = LIST_HEAD_INIT(pci2phy_map_head);
   struct pci_extra_dev *uncore_extra_pci_dev;
   static int max_dies;
 
+int get_max_dies(void)
+{
+       return max_dies;
+}
+
   /* mask of cpus that collect uncore events */
   static cpumask_t uncore_cpu_mask;
 
@@ -816,6 +821,16 @@ static ssize_t uncore_get_attr_cpumask(struct device *dev,
 
   static DEVICE_ATTR(cpumask, S_IRUGO, uncore_get_attr_cpumask, NULL);
 
+static ssize_t platform_mapping_show(struct device *dev,
+                               struct device_attribute *attr, char *buf)
+{
+       struct intel_uncore_pmu *pmu = dev_get_drvdata(dev);
+
+       return snprintf(buf, PAGE_SIZE - 1, "%s\n", pmu->platform_mapping ?
+                      (char *)pmu->platform_mapping : "0");
+}
+static DEVICE_ATTR_RO(platform_mapping);
+
   static struct attribute *uncore_pmu_attrs[] = {
          &dev_attr_cpumask.attr,
          NULL,
@@ -825,6 +840,15 @@ static const struct attribute_group uncore_pmu_attr_group = {
          .attrs = uncore_pmu_attrs,
   };
 
+static struct attribute *platform_attrs[] = {
+       &dev_attr_platform_mapping.attr,
+       NULL,
+};
+
+static const struct attribute_group uncore_platform_discovery_group = {
+       .attrs = platform_attrs,
+};
+
   static int uncore_pmu_register(struct intel_uncore_pmu *pmu)
   {
          int ret;
@@ -905,11 +929,27 @@ static void uncore_types_exit(struct intel_uncore_type **types)
                  uncore_type_exit(*types);
   }
 
+static void uncore_type_attrs_compaction(struct intel_uncore_type *type)
+{
+       int i, j;
+
+       for (i = 0, j = 0; i < UNCORE_MAX_NUM_ATTR_GROUP; i++) {
+               if (!type->attr_groups[i])
+                       continue;
+               if (i > j) {
+                       type->attr_groups[j] = type->attr_groups[i];
+                       type->attr_groups[i] = NULL;
+               }
+               j++;
+       }
+}
+
   static int __init uncore_type_init(struct intel_uncore_type *type, bool setid)
   {
          struct intel_uncore_pmu *pmus;
          size_t size;
          int i, j;
+       int ret;
 
          pmus = kcalloc(type->num_boxes, sizeof(*pmus), GFP_KERNEL);
          if (!pmus)
@@ -922,8 +962,10 @@ static int __init uncore_type_init(struct intel_uncore_type *type, bool setid)
                  pmus[i].pmu_idx = i;
                  pmus[i].type    = type;
                  pmus[i].boxes   = kzalloc(size, GFP_KERNEL);
-               if (!pmus[i].boxes)
+               if (!pmus[i].boxes) {
+                       ret = -ENOMEM;
                          goto err;
+               }
          }
 
          type->pmus = pmus;
@@ -940,8 +982,10 @@ static int __init uncore_type_init(struct intel_uncore_type *type, bool setid)
 
                  attr_group = kzalloc(struct_size(attr_group, attrs, i + 1),
                                                                  GFP_KERNEL);
-               if (!attr_group)
+               if (!attr_group) {
+                       ret = -ENOMEM;
                          goto err;
+               }
 
                  attr_group->group.name = "events";
                  attr_group->group.attrs = attr_group->attrs;
@@ -954,6 +998,17 @@ static int __init uncore_type_init(struct intel_uncore_type *type, bool setid)
 
          type->pmu_group = &uncore_pmu_attr_group;
 
+       /*
+        * Exposing mapping of Uncore units to corresponding Uncore PMUs
+        * through /sys/devices/uncore_<type>_<idx>/platform_mapping
+        */
+       if (type->get_topology && type->set_mapping)
+               if (!type->get_topology(type) && !type->set_mapping(type))
+                       type->platform_discovery = &uncore_platform_discovery_group;
+
+       /* For optional attributes, we can safely remove embedded NULL attr_groups elements */
+       uncore_type_attrs_compaction(type);
+
          return 0;
 
   err:
@@ -961,7 +1016,7 @@ static int __init uncore_type_init(struct intel_uncore_type *type, bool setid)
                  kfree(pmus[i].boxes);
          kfree(pmus);
 
-       return -ENOMEM;
+       return ret;
   }
 
   static int __init
arch/x86/events/intel/uncore.h +12 -1
index bbfdaa720b45..ce3727b9f7f8 100644
@@ -43,6 +43,8 @@ struct intel_uncore_box;
   struct uncore_event_desc;
   struct freerunning_counters;
 
+#define UNCORE_MAX_NUM_ATTR_GROUP 5
+
   struct intel_uncore_type {
          const char *name;
          int num_counters;
@@ -71,13 +73,19 @@ struct intel_uncore_type {
          struct intel_uncore_ops *ops;
          struct uncore_event_desc *event_descs;
          struct freerunning_counters *freerunning;
-       const struct attribute_group *attr_groups[4];
+       const struct attribute_group *attr_groups[UNCORE_MAX_NUM_ATTR_GROUP];
          struct pmu *pmu; /* for custom pmu ops */
+       void *platform_topology;
+       /* finding Uncore units */
+       int (*get_topology)(struct intel_uncore_type *type);
+       /* mapping Uncore units to PMON ranges */
+       int (*set_mapping)(struct intel_uncore_type *type);
   };
 
   #define pmu_group attr_groups[0]
   #define format_group attr_groups[1]
   #define events_group attr_groups[2]
+#define platform_discovery attr_groups[3]
 
   struct intel_uncore_ops {
          void (*init_box)(struct intel_uncore_box *);
@@ -99,6 +107,7 @@ struct intel_uncore_pmu {
          int                             pmu_idx;
          int                             func_id;
          bool                            registered;
+       void                            *platform_mapping;
          atomic_t                        activeboxes;
          struct intel_uncore_type        *type;
          struct intel_uncore_box         **boxes;
@@ -490,6 +499,8 @@ static inline struct intel_uncore_box *uncore_event_to_box(struct perf_event *ev
          return event->pmu_private;
   }
 
+int get_max_dies(void);
+
   struct intel_uncore_box *uncore_pmu_to_box(struct intel_uncore_pmu *pmu, int cpu);
   u64 uncore_msr_read_counter(struct intel_uncore_box *box, struct perf_event *event);
   void uncore_mmio_exit_box(struct intel_uncore_box *box);
arch/x86/events/intel/uncore_snbep.c +144 -1
index b10a5ec79e48..92ce9fbafde1 100644
@@ -273,6 +273,28 @@
   #define SKX_CPUNODEID                  0xc0
   #define SKX_GIDNIDMAP                  0xd4
 
+/*
+ * The CPU_BUS_NUMBER MSR returns the values of the respective CPUBUSNO CSR
+ * that BIOS programmed. MSR has package scope.
+ * |  Bit  |  Default  |  Description
+ * | [63]  |    00h    | VALID - When set, indicates the CPU bus
+ *                       numbers have been initialized. (RO)
+ * |[62:48]|    ---    | Reserved
+ * |[47:40]|    00h    | BUS_NUM_5 — Return the bus number BIOS assigned
+ *                       CPUBUSNO(5). (RO)
+ * |[39:32]|    00h    | BUS_NUM_4 — Return the bus number BIOS assigned
+ *                       CPUBUSNO(4). (RO)
+ * |[31:24]|    00h    | BUS_NUM_3 — Return the bus number BIOS assigned
+ *                       CPUBUSNO(3). (RO)
+ * |[23:16]|    00h    | BUS_NUM_2 — Return the bus number BIOS assigned
+ *                       CPUBUSNO(2). (RO)
+ * |[15:8] |    00h    | BUS_NUM_1 — Return the bus number BIOS assigned
+ *                       CPUBUSNO(1). (RO)
+ * | [7:0] |    00h    | BUS_NUM_0 — Return the bus number BIOS assigned
+ *                       CPUBUSNO(0). (RO)
+ */
+#define SKX_MSR_CPU_BUS_NUMBER         0x300
+
   /* SKX CHA */
   #define SKX_CHA_MSR_PMON_BOX_FILTER_TID                (0x1ffULL << 0)
   #define SKX_CHA_MSR_PMON_BOX_FILTER_LINK       (0xfULL << 9)
@@ -3580,6 +3602,9 @@ static struct intel_uncore_ops skx_uncore_iio_ops = {
          .read_counter           = uncore_msr_read_counter,
   };
 
+static int skx_iio_get_topology(struct intel_uncore_type *type);
+static int skx_iio_set_mapping(struct intel_uncore_type *type);
+
   static struct intel_uncore_type skx_uncore_iio = {
          .name                   = "iio",
          .num_counters           = 4,
@@ -3594,6 +3619,8 @@ static struct intel_uncore_type skx_uncore_iio = {
          .constraints            = skx_uncore_iio_constraints,
          .ops                    = &skx_uncore_iio_ops,
          .format_group           = &skx_uncore_iio_format_group,
+       .get_topology           = skx_iio_get_topology,
+       .set_mapping            = skx_iio_set_mapping,
   };
 
   enum perf_uncore_iio_freerunning_type_id {
@@ -3780,6 +3807,123 @@ static int skx_count_chabox(void)
          return hweight32(val);
   }
 
+static inline u8 skx_iio_topology_byte(void *platform_topology,
+                                       int die, int idx)
+{
+       return *((u8 *)(platform_topology) + die * sizeof(u64) + idx);
+}
+
+static inline bool skx_iio_topology_valid(u64 msr_value)
+{
+       return msr_value & ((u64)1 << 63);
+}
+
+static int skx_msr_cpu_bus_read(int cpu, int die)
+{
+       int ret = rdmsrl_on_cpu(cpu, SKX_MSR_CPU_BUS_NUMBER,
+                               ((u64 *)skx_uncore_iio.platform_topology) + die);
+
+       if (!ret) {
+               if (!skx_iio_topology_valid(*(((u64 *)skx_uncore_iio.platform_topology) + die)))
+                       ret = -1;
+       }
+       return ret;
+}
+
+static int skx_iio_get_topology(struct intel_uncore_type *type)
+{
+       int ret, cpu, die, current_die;
+       struct pci_bus *bus = NULL;
+
+       while ((bus = pci_find_next_bus(bus)) != NULL)
+               if (pci_domain_nr(bus)) {
+                       pr_info("Mapping of I/O stack to PMON ranges is not supported for multi-segment topology\n");
+                       return -1;
+               }
+
+       /* Size of SKX_MSR_CPU_BUS_NUMBER is 8 bytes, the MSR has package scope.*/
+       type->platform_topology =
+               kzalloc(get_max_dies() * sizeof(u64), GFP_KERNEL);
+       if (!type->platform_topology)
+               return -ENOMEM;
+
+       /*
+        * Using cpus_read_lock() to ensure cpu is not going down between
+        * looking at cpu_online_mask.
+        */
+       cpus_read_lock();
+       /* Invalid value to start loop.*/
+       current_die = -1;
+       for_each_online_cpu(cpu) {
+               die = topology_logical_die_id(cpu);
+               if (current_die == die)
+                       continue;
+               ret = skx_msr_cpu_bus_read(cpu, die);
+               if (ret)
+                       break;
+               current_die = die;
+       }
+       cpus_read_unlock();
+
+       if (ret)
+               kfree(type->platform_topology);
+       return ret;
+}
+
+static int skx_iio_set_mapping(struct intel_uncore_type *type)
+{
+       /*
+        * Each IIO stack (PCIe root port) has its own IIO PMON block, so each
+        * platform_mapping holds bus number(s) of PCIe root port(s), which can
+        * be monitored by that IIO PMON block.
+        *
+        * For example, on a 4-die Xeon platform with up to 6 IIO stacks per die
+        * and, therefore, 6 IIO PMON blocks per die, the platform_mapping of IIO
+        * PMON block 0 holds "0000:00,0000:40,0000:80,0000:c0":
+        *
+        * $ cat /sys/devices/uncore_iio_0/platform_mapping
+        * 0000:00,0000:40,0000:80,0000:c0
+        *
+        * Which means:
+        * IIO PMON block 0 on the die 0 belongs to PCIe root port located on bus 0x00, domain 0x0000
+        * IIO PMON block 0 on the die 1 belongs to PCIe root port located on bus 0x40, domain 0x0000
+        * IIO PMON block 0 on the die 2 belongs to PCIe root port located on bus 0x80, domain 0x0000
+        * IIO PMON block 0 on the die 3 belongs to PCIe root port located on bus 0xc0, domain 0x0000
+        */
+
+       int ret = 0;
+       int die, i;
+       char *buf;
+       struct intel_uncore_pmu *pmu;
+       const int template_len = 8;
+
+       for (i = 0; i < type->num_boxes; i++) {
+               pmu = type->pmus + i;
+               /* Root bus 0x00 is valid only for die 0 AND pmu_idx = 0. */
+               if (skx_iio_topology_byte(type->platform_topology, 0, pmu->pmu_idx) || (!pmu->pmu_idx)) {
+                       pmu->platform_mapping =
+                               kzalloc(get_max_dies() * template_len + 1, GFP_KERNEL);
+                       if (pmu->platform_mapping) {
+                               buf = (char *)pmu->platform_mapping;
+                               for (die = 0; die < get_max_dies(); die++)
+                                       buf += snprintf(buf, template_len + 1, "%04x:%02x,", 0,
+                                               skx_iio_topology_byte(type->platform_topology,
+                                                                     die, pmu->pmu_idx));
+
+                               *(--buf) = '\0';
+                       } else {
+                               for (; i >= 0; i--)
+                                       kfree((type->pmus + i)->platform_mapping);
+                               ret = -ENOMEM;
+                               break;
+                       }
+               }
+       }
+
+       kfree(type->platform_topology);
+       return ret;
+}
+
   void skx_uncore_cpu_init(void)
   {
          skx_uncore_chabox.num_boxes = skx_count_chabox();
--

2.19.1