• https://github.com/torvalds/linux/blob/v6.5/tools/perf/design.txt

Performance Counters for Linux

Performance counters are special hardware registers available on most modern
CPUs. These registers count the number of certain types of hw events: such
as instructions executed, cachemisses suffered, or branches mis-predicted -
without slowing down the kernel or applications. These registers can also
trigger interrupts when a threshold number of events have passed - and can
thus be used to profile the code that runs on that CPU.

The Linux Performance Counter subsystem provides an abstraction of these
hardware capabilities. It provides per task and per CPU counters, counter
groups, and it provides event capabilities on top of those. It
provides "virtual" 64-bit counters, regardless of the width of the
underlying hardware counters.

Performance counters are accessed via special file descriptors.
There's one file descriptor per virtual counter used.

The special file descriptor is opened via the sys_perf_event_open()
system call:

int sys_perf_event_open(struct perf_event_attr *hw_event_uptr,
pid_t pid, int cpu, int group_fd,
unsigned long flags);

The syscall returns the new fd. The fd can be used via the normal
VFS system calls: read() can be used to read the counter, fcntl()
can be used to set the blocking mode, etc.

Multiple counters can be kept open at a time, and the counters
can be poll()ed.

When creating a new counter fd, 'perf_event_attr' is:

struct perf_event_attr {
        /*
         * The MSB of the config word signifies if the rest contains cpu
         * specific (raw) counter configuration data, if unset, the next
         * 7 bits are an event type and the rest of the bits are the event
         * identifier.
         */
        __u64                   config;

        __u64                   irq_period;
        __u32                   record_type;
        __u32                   read_format;

        __u64                   disabled       :  1, /* off by default        */
                                inherit        :  1, /* children inherit it   */
                                pinned         :  1, /* must always be on PMU */
                                exclusive      :  1, /* only group on PMU     */
                                exclude_user   :  1, /* don't count user      */
                                exclude_kernel :  1, /* ditto kernel          */
                                exclude_hv     :  1, /* ditto hypervisor      */
                                exclude_idle   :  1, /* don't count when idle */
                                mmap           :  1, /* include mmap data     */
                                munmap         :  1, /* include munmap data   */
                                comm           :  1, /* include comm data     */

                                __reserved_1   : 52;

        __u32                   extra_config_len;
        __u32                   wakeup_events;  /* wakeup every n events */

        __u64                   __reserved_2;
        __u64                   __reserved_3;
};

The 'config' field specifies what the counter should count. It
is divided into 3 bit-fields:

raw_type: 1 bit   (most significant bit)	0x8000_0000_0000_0000
type:	  7 bits  (next most significant)	0x7f00_0000_0000_0000
event_id: 56 bits (least significant)		0x00ff_ffff_ffff_ffff

If 'raw_type' is 1, then the counter will count a hardware event
specified by the remaining 63 bits of event_config. The encoding is
machine-specific.

If 'raw_type' is 0, then the 'type' field says what kind of counter
this is, with the following encoding:

enum perf_type_id {
	PERF_TYPE_HARDWARE		= 0,
	PERF_TYPE_SOFTWARE		= 1,
	PERF_TYPE_TRACEPOINT		= 2,
};

A counter of PERF_TYPE_HARDWARE will count the hardware event
specified by 'event_id':

/*
 * Generalized performance counter event types, used by the hw_event.event_id
 * parameter of the sys_perf_event_open() syscall:
 */
enum perf_hw_id {
	/*
	 * Common hardware events, generalized by the kernel:
	 */
	PERF_COUNT_HW_CPU_CYCLES		= 0,
	PERF_COUNT_HW_INSTRUCTIONS		= 1,
	PERF_COUNT_HW_CACHE_REFERENCES		= 2,
	PERF_COUNT_HW_CACHE_MISSES		= 3,
	PERF_COUNT_HW_BRANCH_INSTRUCTIONS	= 4,
	PERF_COUNT_HW_BRANCH_MISSES		= 5,
	PERF_COUNT_HW_BUS_CYCLES		= 6,
	PERF_COUNT_HW_STALLED_CYCLES_FRONTEND	= 7,
	PERF_COUNT_HW_STALLED_CYCLES_BACKEND	= 8,
	PERF_COUNT_HW_REF_CPU_CYCLES		= 9,
};

These are standardized types of events that work relatively uniformly
on all CPUs that implement Performance Counters support under Linux,
although there may be variations (e.g., different CPUs might count
cache references and misses at different levels of the cache hierarchy).
If a CPU is not able to count the selected event, then the system call
will return -EINVAL.

More hw_event_types are supported as well, but they are CPU-specific
and accessed as raw events. For example, to count "External bus
cycles while bus lock signal asserted" events on Intel Core CPUs, pass
in a 0x4064 event_id value and set hw_event.raw_type to 1.

A counter of type PERF_TYPE_SOFTWARE will count one of the available
software events, selected by 'event_id':

/*
 * Special "software" counters provided by the kernel, even if the hardware
 * does not support performance counters. These counters measure various
 * physical and sw events of the kernel (and allow the profiling of them as
 * well):
 */
enum perf_sw_ids {
	PERF_COUNT_SW_CPU_CLOCK		= 0,
	PERF_COUNT_SW_TASK_CLOCK	= 1,
	PERF_COUNT_SW_PAGE_FAULTS	= 2,
	PERF_COUNT_SW_CONTEXT_SWITCHES	= 3,
	PERF_COUNT_SW_CPU_MIGRATIONS	= 4,
	PERF_COUNT_SW_PAGE_FAULTS_MIN	= 5,
	PERF_COUNT_SW_PAGE_FAULTS_MAJ	= 6,
	PERF_COUNT_SW_ALIGNMENT_FAULTS	= 7,
	PERF_COUNT_SW_EMULATION_FAULTS	= 8,
};

Counters of the type PERF_TYPE_TRACEPOINT are available when the ftrace event
tracer is available, and event_id values can be obtained from
/debug/tracing/events/*/*/id

Counters come in two flavours: counting counters and sampling
counters. A "counting" counter is one that is used for counting the
number of events that occur, and is characterised by having
irq_period = 0.

A read() on a counter returns the current value of the counter and possible
additional values as specified by 'read_format', each value is a u64 (8 bytes)
in size.

/*
 * Bits that can be set in hw_event.read_format to request that
 * reads on the counter should return the indicated quantities,
 * in increasing order of bit value, after the counter value.
 */
enum perf_event_read_format {
        PERF_FORMAT_TOTAL_TIME_ENABLED  =  1,
        PERF_FORMAT_TOTAL_TIME_RUNNING  =  2,
};

Using these additional values one can establish the overcommit ratio for a
particular counter allowing one to take the round-robin scheduling effect
into account.

A "sampling" counter is one that is set up to generate an interrupt
every N events, where N is given by 'irq_period'. A sampling counter
has irq_period > 0. The record_type controls what data is recorded on each
interrupt:

/*
 * Bits that can be set in hw_event.record_type to request information
 * in the overflow packets.
 */
enum perf_event_record_format {
        PERF_RECORD_IP          = 1U << 0,
        PERF_RECORD_TID         = 1U << 1,
        PERF_RECORD_TIME        = 1U << 2,
        PERF_RECORD_ADDR        = 1U << 3,
        PERF_RECORD_GROUP       = 1U << 4,
        PERF_RECORD_CALLCHAIN   = 1U << 5,
};

Such (and other) events will be recorded in a ring-buffer, which is
available to user-space using mmap() (see below).

The 'disabled' bit specifies whether the counter starts out disabled
or enabled. If it is initially disabled, it can be enabled by ioctl
or prctl (see below).

The 'inherit' bit, if set, specifies that this counter should count
events on descendant tasks as well as the task specified. This only
applies to new descendents, not to any existing descendents at the
time the counter is created (nor to any new descendents of existing
descendents).

The 'pinned' bit, if set, specifies that the counter should always be
on the CPU if at all possible. It only applies to hardware counters
and only to group leaders. If a pinned counter cannot be put onto the
CPU (e.g. because there are not enough hardware counters or because of
a conflict with some other event), then the counter goes into an
'error' state, where reads return end-of-file (i.e. read() returns 0)
until the counter is subsequently enabled or disabled.

The 'exclusive' bit, if set, specifies that when this counter's group
is on the CPU, it should be the only group using the CPU's counters.
In future, this will allow sophisticated monitoring programs to supply
extra configuration information via 'extra_config_len' to exploit
advanced features of the CPU's Performance Monitor Unit (PMU) that are
not otherwise accessible and that might disrupt other hardware
counters.

The 'exclude_user', 'exclude_kernel' and 'exclude_hv' bits provide a
way to request that counting of events be restricted to times when the
CPU is in user, kernel and/or hypervisor mode.

Furthermore the 'exclude_host' and 'exclude_guest' bits provide a way
to request counting of events restricted to guest and host contexts when
using Linux as the hypervisor.

The 'mmap' and 'munmap' bits allow recording of PROT_EXEC mmap/munmap
operations, these can be used to relate userspace IP addresses to actual
code, even after the mapping (or even the whole process) is gone,
these events are recorded in the ring-buffer (see below).

The 'comm' bit allows tracking of process comm data on process creation.
This too is recorded in the ring-buffer (see below).

The 'pid' parameter to the sys_perf_event_open() system call allows the
counter to be specific to a task:

pid == 0: if the pid parameter is zero, the counter is attached to the
current task.

pid > 0: the counter is attached to a specific task (if the current task
has sufficient privilege to do so)

pid < 0: all tasks are counted (per cpu counters)

The 'cpu' parameter allows a counter to be made specific to a CPU:

cpu >= 0: the counter is restricted to a specific CPU
cpu == -1: the counter counts on all CPUs

(Note: the combination of 'pid == -1' and 'cpu == -1' is not valid.)

A 'pid > 0' and 'cpu == -1' counter is a per task counter that counts
events of that task and 'follows' that task to whatever CPU the task
gets schedule to. Per task counters can be created by any user, for
their own tasks.

A 'pid == -1' and 'cpu == x' counter is a per CPU counter that counts
all events on CPU-x. Per CPU counters need CAP_PERFMON or CAP_SYS_ADMIN
privilege.

The 'flags' parameter is currently unused and must be zero.

The 'group_fd' parameter allows counter "groups" to be set up. A
counter group has one counter which is the group "leader". The leader
is created first, with group_fd = -1 in the sys_perf_event_open call
that creates it. The rest of the group members are created
subsequently, with group_fd giving the fd of the group leader.
(A single counter on its own is created with group_fd = -1 and is
considered to be a group with only 1 member.)

A counter group is scheduled onto the CPU as a unit, that is, it will
only be put onto the CPU if all of the counters in the group can be
put onto the CPU. This means that the values of the member counters
can be meaningfully compared, added, divided (to get ratios), etc.,
with each other, since they have counted events for the same set of
executed instructions.

Like stated, asynchronous events, like counter overflow or PROT_EXEC mmap
tracking are logged into a ring-buffer. This ring-buffer is created and
accessed through mmap().

The mmap size should be 1+2^n pages, where the first page is a meta-data page
(struct perf_event_mmap_page) that contains various bits of information such
as where the ring-buffer head is.

/*
 * Structure of the page that can be mapped via mmap
 */
struct perf_event_mmap_page {
        __u32   version;                /* version number of this structure */
        __u32   compat_version;         /* lowest version this is compat with */

        /*
         * Bits needed to read the hw counters in user-space.
         *
         *   u32 seq;
         *   s64 count;
         *
         *   do {
         *     seq = pc->lock;
         *
         *     barrier()
         *     if (pc->index) {
         *       count = pmc_read(pc->index - 1);
         *       count += pc->offset;
         *     } else
         *       goto regular_read;
         *
         *     barrier();
         *   } while (pc->lock != seq);
         *
         * NOTE: for obvious reason this only works on self-monitoring
         *       processes.
         */
        __u32   lock;                   /* seqlock for synchronization */
        __u32   index;                  /* hardware counter identifier */
        __s64   offset;                 /* add to hardware counter value */

        /*
         * Control data for the mmap() data buffer.
         *
         * User-space reading this value should issue an rmb(), on SMP capable
         * platforms, after reading this value -- see perf_event_wakeup().
         */
        __u32   data_head;              /* head in the data section */
};

NOTE: the hw-counter userspace bits are arch specific and are currently only
implemented on powerpc.

The following 2^n pages are the ring-buffer which contains events of the form:

#define PERF_RECORD_MISC_KERNEL          (1 << 0)
#define PERF_RECORD_MISC_USER            (1 << 1)
#define PERF_RECORD_MISC_OVERFLOW        (1 << 2)

struct perf_event_header {
        __u32   type;
        __u16   misc;
        __u16   size;
};

enum perf_event_type {

        /*
         * The MMAP events record the PROT_EXEC mappings so that we can
         * correlate userspace IPs to code. They have the following structure:
         *
         * struct {
         *      struct perf_event_header        header;
         *
         *      u32                             pid, tid;
         *      u64                             addr;
         *      u64                             len;
         *      u64                             pgoff;
         *      char                            filename[];
         * };
         */
        PERF_RECORD_MMAP                 = 1,
        PERF_RECORD_MUNMAP               = 2,

        /*
         * struct {
         *      struct perf_event_header        header;
         *
         *      u32                             pid, tid;
         *      char                            comm[];
         * };
         */
        PERF_RECORD_COMM                 = 3,

        /*
         * When header.misc & PERF_RECORD_MISC_OVERFLOW the event_type field
         * will be PERF_RECORD_*
         *
         * struct {
         *      struct perf_event_header        header;
         *
         *      { u64                   ip;       } && PERF_RECORD_IP
         *      { u32                   pid, tid; } && PERF_RECORD_TID
         *      { u64                   time;     } && PERF_RECORD_TIME
         *      { u64                   addr;     } && PERF_RECORD_ADDR
         *
         *      { u64                   nr;
         *        { u64 event, val; }   cnt[nr];  } && PERF_RECORD_GROUP
         *
         *      { u16                   nr,
         *                              hv,
         *                              kernel,
         *                              user;
         *        u64                   ips[nr];  } && PERF_RECORD_CALLCHAIN
         * };
         */
};

NOTE: PERF_RECORD_CALLCHAIN is arch specific and currently only implemented
on x86.

Notification of new events is possible through poll()/select()/epoll() and
fcntl() managing signals.

Normally a notification is generated for every page filled, however one can
additionally set perf_event_attr.wakeup_events to generate one every
so many counter overflow events.

Future work will include a splice() interface to the ring-buffer.

Counters can be enabled and disabled in two ways: via ioctl and via
prctl. When a counter is disabled, it doesn't count or generate
events but does continue to exist and maintain its count value.

An individual counter can be enabled with

	ioctl(fd, PERF_EVENT_IOC_ENABLE, 0);

or disabled with

	ioctl(fd, PERF_EVENT_IOC_DISABLE, 0);

For a counter group, pass PERF_IOC_FLAG_GROUP as the third argument.
Enabling or disabling the leader of a group enables or disables the
whole group; that is, while the group leader is disabled, none of the
counters in the group will count. Enabling or disabling a member of a
group other than the leader only affects that counter - disabling an
non-leader stops that counter from counting but doesn't affect any
other counter.

Additionally, non-inherited overflow counters can use

	ioctl(fd, PERF_EVENT_IOC_REFRESH, nr);

to enable a counter for 'nr' events, after which it gets disabled again.

A process can enable or disable all the counter groups that are
attached to it, using prctl:

	prctl(PR_TASK_PERF_EVENTS_ENABLE);

	prctl(PR_TASK_PERF_EVENTS_DISABLE);

This applies to all counters on the current process, whether created
by this process or by another, and doesn't affect any counters that
this process has created on other processes. It only enables or
disables the group leaders, not any other members in the groups.

Arch requirements

If your architecture does not have hardware performance metrics, you can
still use the generic software counters based on hrtimers for sampling.

So to start with, in order to add HAVE_PERF_EVENTS to your Kconfig, you
will need at least this:

• asm/perf_event.h - a basic stub will suffice at first
• support for atomic64 types (and associated helper functions)

If your architecture does have hardware capabilities, you can override the
weak stub hw_perf_event_init() to register hardware counters.

Architectures that have d-cache aliassing issues, such as Sparc and ARM,
should select PERF_USE_VMALLOC in order to avoid these for perf mmap().