diff mbox series

[1/5] perf: Add dummy pmu module

Message ID 20241008183501.1354695-2-lucas.demarchi@intel.com (mailing list archive)
State New, archived
Headers show
Series perf: Fix pmu for drivers with bind/unbind | expand

Commit Message

Lucas De Marchi Oct. 8, 2024, 6:34 p.m. UTC
Simple dummy module that mimics what can be done with drivers to
bind/unbind them from the bus, which should trigger resource release.

This is mostly based on how i915 and (pending changes for) xe drivers
are interacting with perf pmu. A few differences due to not having
backing hardware or for simplicity:

	- Instead of using BDF for bind/unbind, use a single number.
	- Unbind is triggered either via debugfs or when removing the
	  module.
	- event::destroy() is always assigned as there should only be a
	  few additional calls

Signed-off-by: Lucas De Marchi <lucas.demarchi@intel.com>
---
 kernel/events/Makefile    |   1 +
 kernel/events/dummy_pmu.c | 426 ++++++++++++++++++++++++++++++++++++++
 2 files changed, 427 insertions(+)
 create mode 100644 kernel/events/dummy_pmu.c

Comments

Jeff Johnson Oct. 15, 2024, 12:26 a.m. UTC | #1
On 10/8/24 11:34, Lucas De Marchi wrote:
...
> +module_init(dummy_init);
> +module_exit(dummy_exit);
> +
> +MODULE_AUTHOR("Lucas De Marchi <lucas.demarchi@intel.com>");
> +MODULE_LICENSE("GPL");

Since commit 1fffe7a34c89 ("script: modpost: emit a warning when the
description is missing"), a module without a MODULE_DESCRIPTION() will
result in a warning when built with make W=1. Recently, multiple
developers have been eradicating these warnings treewide, and very few
(if any) are left, so please don't introduce a new one :)

Please add the missing MODULE_DESCRIPTION()
Lucas De Marchi Oct. 15, 2024, 4:23 a.m. UTC | #2
On Mon, Oct 14, 2024 at 05:26:19PM -0700, Jeff Johnson wrote:
>On 10/8/24 11:34, Lucas De Marchi wrote:
>...
>> +module_init(dummy_init);
>> +module_exit(dummy_exit);
>> +
>> +MODULE_AUTHOR("Lucas De Marchi <lucas.demarchi@intel.com>");
>> +MODULE_LICENSE("GPL");
>
>Since commit 1fffe7a34c89 ("script: modpost: emit a warning when the
>description is missing"), a module without a MODULE_DESCRIPTION() will
>result in a warning when built with make W=1. Recently, multiple
>developers have been eradicating these warnings treewide, and very few
>(if any) are left, so please don't introduce a new one :)
>
>Please add the missing MODULE_DESCRIPTION()

Thanks. This patch is not intended to be applied - it serves only to
demonstrate the issue solved by the other patches. But I will keep that
in mind if I re-submit it.

Lucas De Marchi
Peter Zijlstra Oct. 16, 2024, 8:51 a.m. UTC | #3
On Tue, Oct 08, 2024 at 01:34:57PM -0500, Lucas De Marchi wrote:
> +static int parse_device(const char __user *ubuf, size_t size, u32 *instance)
> +{
> +	char buf[16];
> +	ssize_t len;
> +
> +	if (size > sizeof(buf) - 1)
> +		return -E2BIG;
> +
> +	len = strncpy_from_user(buf, ubuf, sizeof(buf));
> +	if (len < 0 || len >= sizeof(buf) - 1)
> +		return -E2BIG;
> +
> +	if (kstrtou32(buf, 0, instance))
> +		return -EINVAL;
> +
> +	return size;
> +}

I had to change this to:

+static int parse_device(const char __user *ubuf, size_t size, u32 *instance)
+{
+       int ret = kstrtouint_from_user(ubuf, size, 0, instance);
+       if (ret) {
+               printk("suckage: %d\n", ret);
+               return ret;
+       }
+       return size;
+}

because otherwise it didn't want to work for me; I kept getting garbage
at the end and things failing. Specifically, it looked like the string
presented by userspace was not '\0' terminated, ubuf was pointing to
"1...garbage..." and size was 1.
Lucas De Marchi Oct. 18, 2024, 7:30 p.m. UTC | #4
On Wed, Oct 16, 2024 at 10:51:02AM +0200, Peter Zijlstra wrote:
>On Tue, Oct 08, 2024 at 01:34:57PM -0500, Lucas De Marchi wrote:
>> +static int parse_device(const char __user *ubuf, size_t size, u32 *instance)
>> +{
>> +	char buf[16];
>> +	ssize_t len;
>> +
>> +	if (size > sizeof(buf) - 1)
>> +		return -E2BIG;
>> +
>> +	len = strncpy_from_user(buf, ubuf, sizeof(buf));
>> +	if (len < 0 || len >= sizeof(buf) - 1)
>> +		return -E2BIG;
>> +
>> +	if (kstrtou32(buf, 0, instance))
>> +		return -EINVAL;
>> +
>> +	return size;
>> +}
>
>I had to change this to:
>
>+static int parse_device(const char __user *ubuf, size_t size, u32 *instance)
>+{
>+       int ret = kstrtouint_from_user(ubuf, size, 0, instance);
>+       if (ret) {
>+               printk("suckage: %d\n", ret);
>+               return ret;
>+       }
>+       return size;
>+}
>
>because otherwise it didn't want to work for me; I kept getting garbage
>at the end and things failing. Specifically, it looked like the string
>presented by userspace was not '\0' terminated, ubuf was pointing to
>"1...garbage..." and size was 1.

oh... it's the sysfs (that PCI drivers use) that is guaranteed to be
nul-terminated. debugfs is not. And it probably worked for me because of
CONFIG_INIT_STACK_ALL_ZERO=y that comes from my distro.

and this version is also shorter and simpler.

thanks
Lucas De Marchi
diff mbox series

Patch

diff --git a/kernel/events/Makefile b/kernel/events/Makefile
index 91a62f566743..2993fed2d091 100644
--- a/kernel/events/Makefile
+++ b/kernel/events/Makefile
@@ -4,3 +4,4 @@  obj-y := core.o ring_buffer.o callchain.o
 obj-$(CONFIG_HAVE_HW_BREAKPOINT) += hw_breakpoint.o
 obj-$(CONFIG_HW_BREAKPOINT_KUNIT_TEST) += hw_breakpoint_test.o
 obj-$(CONFIG_UPROBES) += uprobes.o
+obj-m += dummy_pmu.o
diff --git a/kernel/events/dummy_pmu.c b/kernel/events/dummy_pmu.c
new file mode 100644
index 000000000000..cdba3a831e4a
--- /dev/null
+++ b/kernel/events/dummy_pmu.c
@@ -0,0 +1,426 @@ 
+// SPDX-License-Identifier: GPL-2.0
+/*
+ * Copyright © 2024 Intel Corporation
+ */
+
+#define pr_fmt(fmt) "%s: " fmt, KBUILD_MODNAME
+
+#include <linux/debugfs.h>
+#include <linux/init.h>
+#include <linux/list.h>
+#include <linux/module.h>
+#include <linux/mutex.h>
+#include <linux/perf_event.h>
+#include <linux/random.h>
+#include <linux/seq_file.h>
+#include <linux/types.h>
+
+struct dummy_mod {
+	struct dentry *debugfs_root;
+
+	struct list_head device_list;
+	struct mutex mutex;
+};
+
+struct dummy_pmu {
+	struct pmu base;
+	char *name;
+	bool registered;
+};
+
+struct dummy_device {
+	unsigned int instance;
+	struct kref refcount;
+	struct list_head mod_entry;
+	struct dummy_pmu pmu;
+};
+
+static struct dummy_mod dm;
+
+static void device_release(struct kref *ref);
+
+static struct dummy_pmu *event_to_pmu(struct perf_event *event)
+{
+	return container_of(event->pmu, struct dummy_pmu, base);
+}
+
+static struct dummy_device *pmu_to_device(struct dummy_pmu *pmu)
+{
+	return container_of(pmu, struct dummy_device, pmu);
+}
+
+static ssize_t dummy_pmu_events_sysfs_show(struct device *dev,
+					   struct device_attribute *attr,
+					   char *page)
+{
+	struct perf_pmu_events_attr *pmu_attr;
+
+	pmu_attr = container_of(attr, struct perf_pmu_events_attr, attr);
+
+	return sprintf(page, "event=0x%04llx\n", pmu_attr->id);
+}
+
+#define DUMMY_PMU_EVENT_ATTR(name, config)				\
+	PMU_EVENT_ATTR_ID(name, dummy_pmu_events_sysfs_show, config)
+
+PMU_FORMAT_ATTR(event, "config:0-63");
+
+#define DUMMY1_CONFIG 0x01
+#define DUMMY2_CONFIG 0x02
+
+static struct attribute *dummy_pmu_event_attrs[] = {
+	DUMMY_PMU_EVENT_ATTR(test-event-1, DUMMY1_CONFIG),
+	DUMMY_PMU_EVENT_ATTR(test-event-2, DUMMY2_CONFIG),
+	NULL,
+};
+
+static struct attribute *dummy_pmu_format_attrs[] = {
+	&format_attr_event.attr,
+	NULL,
+};
+static const struct attribute_group dummy_pmu_events_attr_group = {
+	.name = "events",
+	.attrs = dummy_pmu_event_attrs,
+};
+static const struct attribute_group dummy_pmu_format_attr_group = {
+	.name = "format",
+	.attrs = dummy_pmu_format_attrs,
+};
+static const struct attribute_group *attr_groups[] = {
+	&dummy_pmu_format_attr_group,
+	&dummy_pmu_events_attr_group,
+	NULL,
+};
+
+static void dummy_pmu_event_destroy(struct perf_event *event)
+{
+	struct dummy_pmu *pmu = event_to_pmu(event);
+	struct dummy_device *d = pmu_to_device(pmu);
+
+	kref_put(&d->refcount, device_release);
+}
+
+static int dummy_pmu_event_init(struct perf_event *event)
+{
+	struct dummy_pmu *pmu = event_to_pmu(event);
+	struct dummy_device *d = pmu_to_device(pmu);
+
+	if (!pmu->registered)
+		return -ENODEV;
+
+	if (event->attr.type != event->pmu->type)
+		return -ENOENT;
+
+	/* unsupported modes and filters */
+	if (event->attr.sample_period) /* no sampling */
+		return -EINVAL;
+
+	if (has_branch_stack(event))
+		return -EOPNOTSUPP;
+
+	if (event->cpu < 0)
+		return -EINVAL;
+
+	/* Event keeps a ref to maintain PMU allocated, even if it's unregistered */
+	kref_get(&d->refcount);
+	event->destroy = dummy_pmu_event_destroy;
+
+	return 0;
+}
+
+static void dummy_pmu_event_start(struct perf_event *event, int flags)
+{
+	struct dummy_pmu *pmu = event_to_pmu(event);
+
+	if (!pmu->registered)
+		return;
+
+	event->hw.state = 0;
+}
+
+static void dummy_pmu_event_read(struct perf_event *event)
+{
+	struct dummy_pmu *pmu = event_to_pmu(event);
+	u8 buf;
+
+	if (!pmu->registered) {
+		event->hw.state = PERF_HES_STOPPED;
+		return;
+	}
+
+	get_random_bytes(&buf, 1);
+	buf %= 10;
+
+	switch (event->attr.config & 0xf) {
+	case DUMMY1_CONFIG:
+		break;
+	case DUMMY2_CONFIG:
+		buf *= 2;
+		break;
+	}
+
+	local64_add(buf, &event->count);
+}
+
+static void dummy_pmu_event_stop(struct perf_event *event, int flags)
+{
+	struct dummy_pmu *pmu = event_to_pmu(event);
+
+	if (!pmu->registered)
+		goto out;
+
+	if (flags & PERF_EF_UPDATE)
+		dummy_pmu_event_read(event);
+
+out:
+	event->hw.state = PERF_HES_STOPPED;
+}
+
+static int dummy_pmu_event_add(struct perf_event *event, int flags)
+{
+	struct dummy_pmu *pmu = event_to_pmu(event);
+
+	if (!pmu->registered)
+		return -ENODEV;
+
+	if (flags & PERF_EF_START)
+		dummy_pmu_event_start(event, flags);
+
+	return 0;
+
+}
+
+static void dummy_pmu_event_del(struct perf_event *event, int flags)
+{
+	dummy_pmu_event_stop(event, PERF_EF_UPDATE);
+}
+
+static int device_init(struct dummy_device *d)
+{
+	int ret;
+
+	d->pmu.base = (struct pmu){
+		.attr_groups	= attr_groups,
+		.module		= THIS_MODULE,
+		.task_ctx_nr	= perf_invalid_context,
+		.event_init	= dummy_pmu_event_init,
+		.add		= dummy_pmu_event_add,
+		.del		= dummy_pmu_event_del,
+		.start		= dummy_pmu_event_start,
+		.stop		= dummy_pmu_event_stop,
+		.read		= dummy_pmu_event_read,
+	};
+
+	d->pmu.name = kasprintf(GFP_KERNEL, "dummy_pmu_%u", d->instance);
+	if (!d->pmu.name)
+		return -ENOMEM;
+
+	ret = perf_pmu_register(&d->pmu.base, d->pmu.name, -1);
+	if (ret)
+		return ret;
+
+	d->pmu.registered = true;
+	pr_info("Device registered: %s\n", d->pmu.name);
+
+	return 0;
+}
+
+static void device_exit(struct dummy_device *d)
+{
+	d->pmu.registered = false;
+	perf_pmu_unregister(&d->pmu.base);
+
+	pr_info("Device released: %s\n", d->pmu.name);
+}
+
+static void device_release(struct kref *ref)
+{
+	struct dummy_device *d = container_of(ref, struct dummy_device, refcount);
+
+	kfree(d->pmu.name);
+	kfree(d);
+}
+
+static struct dummy_device *find_device_locked(struct dummy_mod *m, unsigned int instance)
+{
+	struct dummy_device *d;
+
+	list_for_each_entry(d, &m->device_list, mod_entry)
+		if (d->instance == instance)
+			return d;
+
+	return NULL;
+}
+
+static int dummy_add_device(struct dummy_mod *m, unsigned int instance)
+{
+	struct dummy_device *d, *d2;
+	int ret = 0;
+
+	mutex_lock(&m->mutex);
+	d = find_device_locked(m, instance);
+	mutex_unlock(&m->mutex);
+	if (d)
+		return -EINVAL;
+
+	d = kcalloc(1, sizeof(*d), GFP_KERNEL);
+	if (!d)
+		return -ENOMEM;
+
+	kref_init(&d->refcount);
+	d->instance = instance;
+
+	ret = device_init(d);
+	if (ret < 0)
+		goto fail_put;
+
+	mutex_lock(&m->mutex);
+	d2 = find_device_locked(m, instance);
+	if (d2) {
+		mutex_unlock(&m->mutex);
+		ret = -EINVAL;
+		goto fail_exit;
+	}
+	list_add(&d->mod_entry, &m->device_list);
+	mutex_unlock(&m->mutex);
+
+	return 0;
+
+fail_exit:
+	device_exit(d);
+fail_put:
+	kref_put(&d->refcount, device_release);
+	return ret;
+}
+
+static int dummy_del_device(struct dummy_mod *m, unsigned int instance)
+{
+	struct dummy_device *d, *found = NULL;
+
+	mutex_lock(&m->mutex);
+	list_for_each_entry(d, &m->device_list, mod_entry) {
+		if (d->instance == instance) {
+			list_del(&d->mod_entry);
+			found = d;
+			break;
+		}
+	}
+	mutex_unlock(&m->mutex);
+
+	if (!found)
+		return -EINVAL;
+
+	device_exit(found);
+	kref_put(&found->refcount, device_release);
+
+	return 0;
+}
+
+static int parse_device(const char __user *ubuf, size_t size, u32 *instance)
+{
+	char buf[16];
+	ssize_t len;
+
+	if (size > sizeof(buf) - 1)
+		return -E2BIG;
+
+	len = strncpy_from_user(buf, ubuf, sizeof(buf));
+	if (len < 0 || len >= sizeof(buf) - 1)
+		return -E2BIG;
+
+	if (kstrtou32(buf, 0, instance))
+		return -EINVAL;
+
+	return size;
+}
+
+static int bind_show(struct seq_file *s, void *unused)
+{
+	struct dummy_mod *m = s->private;
+	struct dummy_device *d;
+
+	mutex_lock(&m->mutex);
+	list_for_each_entry(d, &m->device_list, mod_entry)
+		seq_printf(s, "%u\n", d->instance);
+	mutex_unlock(&m->mutex);
+
+	return 0;
+}
+
+static ssize_t bind_write(struct file *f, const char __user *ubuf,
+			  size_t size, loff_t *pos)
+{
+	struct dummy_mod *m = file_inode(f)->i_private;
+	u32 instance;
+	ssize_t ret;
+
+	ret = parse_device(ubuf, size, &instance);
+	if (ret < 0)
+		return ret;
+
+	ret = dummy_add_device(m, instance);
+	if (ret < 0)
+		return ret;
+
+	return size;
+}
+DEFINE_SHOW_STORE_ATTRIBUTE(bind);
+
+static int unbind_show(struct seq_file *s, void *unused)
+{
+	return -EPERM;
+}
+
+static ssize_t unbind_write(struct file *f, const char __user *ubuf,
+			    size_t size, loff_t *pos)
+{
+	struct dummy_mod *m = file_inode(f)->i_private;
+	unsigned int instance;
+	ssize_t ret;
+
+	ret = parse_device(ubuf, size, &instance);
+	if (ret < 0)
+		return ret;
+
+	ret = dummy_del_device(m, instance);
+	if (ret < 0)
+		return ret;
+
+	return size;
+}
+DEFINE_SHOW_STORE_ATTRIBUTE(unbind);
+
+static int __init dummy_init(void)
+{
+	struct dentry *dir;
+
+	dir = debugfs_create_dir(KBUILD_MODNAME, NULL);
+	debugfs_create_file("bind", 0600, dir, &dm, &bind_fops);
+	debugfs_create_file("unbind", 0200, dir, &dm, &unbind_fops);
+
+	dm.debugfs_root = dir;
+	INIT_LIST_HEAD(&dm.device_list);
+	mutex_init(&dm.mutex);
+
+	return 0;
+}
+
+static void dummy_exit(void)
+{
+	struct dummy_device *d, *tmp;
+
+	debugfs_remove_recursive(dm.debugfs_root);
+
+	mutex_lock(&dm.mutex);
+	list_for_each_entry_safe(d, tmp, &dm.device_list, mod_entry) {
+		device_exit(d);
+		kref_put(&d->refcount, device_release);
+	}
+	mutex_unlock(&dm.mutex);
+}
+
+module_init(dummy_init);
+module_exit(dummy_exit);
+
+MODULE_AUTHOR("Lucas De Marchi <lucas.demarchi@intel.com>");
+MODULE_LICENSE("GPL");