// SPDX-License-Identifier: GPL-2.0
/*
* Performance event support - Processor Activity Instrumentation Extension
* Facility
*
* Copyright IBM Corp. 2022
* Author(s): Thomas Richter <[email protected]>
*/
#define KMSG_COMPONENT "pai_ext"
#define pr_fmt(fmt) KMSG_COMPONENT ": " fmt
#include <linux/kernel.h>
#include <linux/kernel_stat.h>
#include <linux/percpu.h>
#include <linux/notifier.h>
#include <linux/init.h>
#include <linux/export.h>
#include <linux/io.h>
#include <linux/perf_event.h>
#include <asm/ctlreg.h>
#include <asm/pai.h>
#include <asm/debug.h>
#define PAIE1_CB_SZ 0x200 /* Size of PAIE1 control block */
#define PAIE1_CTRBLOCK_SZ 0x400 /* Size of PAIE1 counter blocks */
static debug_info_t *paiext_dbg;
static unsigned int paiext_cnt; /* Extracted with QPACI instruction */
struct pai_userdata {
u16 num;
u64 value;
} __packed;
/* Create the PAI extension 1 control block area.
* The PAI extension control block 1 is pointed to by lowcore
* address 0x1508 for each CPU. This control block is 512 bytes in size
* and requires a 512 byte boundary alignment.
*/
struct paiext_cb { /* PAI extension 1 control block */
u64 header; /* Not used */
u64 reserved1;
u64 acc; /* Addr to analytics counter control block */
u8 reserved2[488];
} __packed;
struct paiext_map {
unsigned long *area; /* Area for CPU to store counters */
struct pai_userdata *save; /* Area to store non-zero counters */
unsigned int active_events; /* # of PAI Extension users */
refcount_t refcnt;
struct perf_event *event; /* Perf event for sampling */
struct paiext_cb *paiext_cb; /* PAI extension control block area */
struct list_head syswide_list; /* List system-wide sampling events */
};
struct paiext_mapptr {
struct paiext_map *mapptr;
};
static struct paiext_root { /* Anchor to per CPU data */
refcount_t refcnt; /* Overall active events */
struct paiext_mapptr __percpu *mapptr;
} paiext_root;
/* Free per CPU data when the last event is removed. */
static void paiext_root_free(void)
{
if (refcount_dec_and_test(&paiext_root.refcnt)) {
free_percpu(paiext_root.mapptr);
paiext_root.mapptr = NULL;
}
debug_sprintf_event(paiext_dbg, 5, "%s root.refcount %d\n", __func__,
refcount_read(&paiext_root.refcnt));
}
/* On initialization of first event also allocate per CPU data dynamically.
* Start with an array of pointers, the array size is the maximum number of
* CPUs possible, which might be larger than the number of CPUs currently
* online.
*/
static int paiext_root_alloc(void)
{
if (!refcount_inc_not_zero(&paiext_root.refcnt)) {
/* The memory is already zeroed. */
paiext_root.mapptr = alloc_percpu(struct paiext_mapptr);
if (!paiext_root.mapptr) {
/* Returning without refcnt adjustment is ok. The
* error code is handled by paiext_alloc() which
* decrements refcnt when an event can not be
* created.
*/
return -ENOMEM;
}
refcount_set(&paiext_root.refcnt, 1);
}
return 0;
}
/* Protects against concurrent increment of sampler and counter member
* increments at the same time and prohibits concurrent execution of
* counting and sampling events.
* Ensures that analytics counter block is deallocated only when the
* sampling and counting on that cpu is zero.
* For details see paiext_alloc().
*/
static DEFINE_MUTEX(paiext_reserve_mutex);
/* Free all memory allocated for event counting/sampling setup */
static void paiext_free(struct paiext_mapptr *mp)
{
kfree(mp->mapptr->area);
kfree(mp->mapptr->paiext_cb);
kvfree(mp->mapptr->save);
kfree(mp->mapptr);
mp->mapptr = NULL;
}
/* Release the PMU if event is the last perf event */
static void paiext_event_destroy_cpu(struct perf_event *event, int cpu)
{
struct paiext_mapptr *mp = per_cpu_ptr(paiext_root.mapptr, cpu);
struct paiext_map *cpump = mp->mapptr;
mutex_lock(&paiext_reserve_mutex);
if (refcount_dec_and_test(&cpump->refcnt)) /* Last reference gone */
paiext_free(mp);
paiext_root_free();
mutex_unlock(&paiext_reserve_mutex);
}
static void paiext_event_destroy(struct perf_event *event)
{
int cpu;
free_page(PAI_SAVE_AREA(event));
if (event->cpu == -1) {
struct cpumask *mask = PAI_CPU_MASK(event);
for_each_cpu(cpu, mask)
paiext_event_destroy_cpu(event, cpu);
kfree(mask);
} else {
paiext_event_destroy_cpu(event, event->cpu);
}
debug_sprintf_event(paiext_dbg, 4, "%s cpu %d\n", __func__,
event->cpu);
}
/* Used to avoid races in checking concurrent access of counting and
* sampling for pai_extension events.
*
* Only one instance of event pai_ext/NNPA_ALL/ for sampling is
* allowed and when this event is running, no counting event is allowed.
* Several counting events are allowed in parallel, but no sampling event
* is allowed while one (or more) counting events are running.
*
* This function is called in process context and it is safe to block.
* When the event initialization functions fails, no other call back will
* be invoked.
*
* Allocate the memory for the event.
*/
static int paiext_alloc_cpu(struct perf_event *event, int cpu)
{
struct paiext_mapptr *mp;
struct paiext_map *cpump;
int rc;
mutex_lock(&paiext_reserve_mutex);
rc = paiext_root_alloc();
if (rc)
goto unlock;
mp = per_cpu_ptr(paiext_root.mapptr, cpu);
cpump = mp->mapptr;
if (!cpump) { /* Paiext_map allocated? */
rc = -ENOMEM;
cpump = kzalloc(sizeof(*cpump), GFP_KERNEL);
if (!cpump)
goto undo;
/* Allocate memory for counter area and counter extraction.
* These are
* - a 512 byte block and requires 512 byte boundary alignment.
* - a 1KB byte block and requires 1KB boundary alignment.
* Only the first counting event has to allocate the area.
*
* Note: This works with commit 59bb47985c1d by default.
* Backporting this to kernels without this commit might
* need adjustment.
*/
mp->mapptr = cpump;
cpump->area = kzalloc(PAIE1_CTRBLOCK_SZ, GFP_KERNEL);
cpump->paiext_cb = kzalloc(PAIE1_CB_SZ, GFP_KERNEL);
cpump->save = kvmalloc_array(paiext_cnt + 1,
sizeof(struct pai_userdata),
GFP_KERNEL);
if (!cpump->save || !cpump->area || !cpump->paiext_cb) {
paiext_free(mp);
goto undo;
}
INIT_LIST_HEAD(&cpump->syswide_list);
refcount_set(&cpump->refcnt, 1);
rc = 0;
} else {
refcount_inc(&cpump->refcnt);
}
undo:
if (rc) {
/* Error in allocation of event, decrement anchor. Since
* the event in not created, its destroy() function is never
* invoked. Adjust the reference counter for the anchor.
*/
paiext_root_free();
}
unlock:
mutex_unlock(&paiext_reserve_mutex);
/* If rc is non-zero, no increment of counter/sampler was done. */
return rc;
}
static int paiext_alloc(struct perf_event *event)
{
struct cpumask *maskptr;
int cpu, rc = -ENOMEM;
maskptr = kzalloc(sizeof(*maskptr), GFP_KERNEL);
if (!maskptr)
goto out;
for_each_online_cpu(cpu) {
rc = paiext_alloc_cpu(event, cpu);
if (rc) {
for_each_cpu(cpu, maskptr)
paiext_event_destroy_cpu(event, cpu);
kfree(maskptr);
goto out;
}
cpumask_set_cpu(cpu, maskptr);
}
/*
* On error all cpumask are freed and all events have been destroyed.
* Save of which CPUs data structures have been allocated for.
* Release them in paicrypt_event_destroy call back function
* for this event.
*/
PAI_CPU_MASK(event) = maskptr;
rc = 0;
out:
return rc;
}
/* The PAI extension 1 control block supports up to 128 entries. Return
* the index within PAIE1_CB given the event number. Also validate event
* number.
*/
static int paiext_event_valid(struct perf_event *event)
{
u64 cfg = event->attr.config;
if (cfg >= PAI_NNPA_BASE && cfg <= PAI_NNPA_BASE + paiext_cnt) {
/* Offset NNPA in paiext_cb */
event->hw.config_base = offsetof(struct paiext_cb, acc);
return 0;
}
return -EINVAL;
}
/* Might be called on different CPU than the one the event is intended for. */
static int paiext_event_init(struct perf_event *event)
{
struct perf_event_attr *a = &event->attr;
int rc;
/* PMU pai_ext registered as PERF_TYPE_RAW, check event type */
if (a->type != PERF_TYPE_RAW && event->pmu->type != a->type)
return -ENOENT;
/* PAI extension event must be valid and in supported range */
rc = paiext_event_valid(event);
if (rc)
return rc;
/* Allow only event NNPA_ALL for sampling. */
if (a->sample_period && a->config != PAI_NNPA_BASE)
return -EINVAL;
/* Prohibit exclude_user event selection */
if (a->exclude_user)
return -EINVAL;
/* Get a page to store last counter values for sampling */
if (a->sample_period) {
PAI_SAVE_AREA(event) = get_zeroed_page(GFP_KERNEL);
if (!PAI_SAVE_AREA(event))
return -ENOMEM;
}
if (event->cpu >= 0)
rc = paiext_alloc_cpu(event, event->cpu);
else
rc = paiext_alloc(event);
if (rc) {
free_page(PAI_SAVE_AREA(event));
return rc;
}
event->destroy = paiext_event_destroy;
if (a->sample_period) {
a->sample_period = 1;
a->freq = 0;
/* Register for paicrypt_sched_task() to be called */
event->attach_state |= PERF_ATTACH_SCHED_CB;
/* Add raw data which are the memory mapped counters */
a->sample_type |= PERF_SAMPLE_RAW;
/* Turn off inheritance */
a->inherit = 0;
}
return 0;
}
static u64 paiext_getctr(unsigned long *area, int nr)
{
return area[nr];
}
/* Read the counter values. Return value from location in buffer. For event
* NNPA_ALL sum up all events.
*/
static u64 paiext_getdata(struct perf_event *event)
{
struct paiext_mapptr *mp = this_cpu_ptr(paiext_root.mapptr);
struct paiext_map *cpump = mp->mapptr;
u64 sum = 0;
int i;
if (event->attr.config != PAI_NNPA_BASE)
return paiext_getctr(cpump->area,
event->attr.config - PAI_NNPA_BASE);
for (i = 1; i <= paiext_cnt; i++)
sum += paiext_getctr(cpump->area, i);
return sum;
}
static u64 paiext_getall(struct perf_event *event)
{
return paiext_getdata(event);
}
static void paiext_read(struct perf_event *event)
{
u64 prev, new, delta;
prev = local64_read(&event->hw.prev_count);
new = paiext_getall(event);
local64_set(&event->hw.prev_count, new);
delta = new - prev;
local64_add(delta, &event->count);
}
static void paiext_start(struct perf_event *event, int flags)
{
struct paiext_mapptr *mp = this_cpu_ptr(paiext_root.mapptr);
struct paiext_map *cpump = mp->mapptr;
u64 sum;
if (!event->attr.sample_period) { /* Counting */
sum = paiext_getall(event); /* Get current value */
local64_set(&event->hw.prev_count, sum);
} else { /* Sampling */
memcpy((void *)PAI_SAVE_AREA(event), cpump->area,
PAIE1_CTRBLOCK_SZ);
/* Enable context switch callback for system-wide sampling */
if (!(event->attach_state & PERF_ATTACH_TASK)) {
list_add_tail(PAI_SWLIST(event), &cpump->syswide_list);
perf_sched_cb_inc(event->pmu);
} else {
cpump->event = event;
}
}
}
static int paiext_add(struct perf_event *event, int flags)
{
struct paiext_mapptr *mp = this_cpu_ptr(paiext_root.mapptr);
struct paiext_map *cpump = mp->mapptr;
struct paiext_cb *pcb = cpump->paiext_cb;
if (++cpump->active_events == 1) {
get_lowcore()->aicd = virt_to_phys(cpump->paiext_cb);
pcb->acc = virt_to_phys(cpump->area) | 0x1;
/* Enable CPU instruction lookup for PAIE1 control block */
local_ctl_set_bit(0, CR0_PAI_EXTENSION_BIT);
}
if (flags & PERF_EF_START)
paiext_start(event, PERF_EF_RELOAD);
event->hw.state = 0;
return 0;
}
static void paiext_have_sample(struct perf_event *, struct paiext_map *);
static void paiext_stop(struct perf_event *event, int flags)
{
struct paiext_mapptr *mp = this_cpu_ptr(paiext_root.mapptr);
struct paiext_map *cpump = mp->mapptr;
if (!event->attr.sample_period) { /* Counting */
paiext_read(event);
} else { /* Sampling */
if (!(event->attach_state & PERF_ATTACH_TASK)) {
list_del(PAI_SWLIST(event));
perf_sched_cb_dec(event->pmu);
} else {
paiext_have_sample(event, cpump);
cpump->event = NULL;
}
}
event->hw.state = PERF_HES_STOPPED;
}
static void paiext_del(struct perf_event *event, int flags)
{
struct paiext_mapptr *mp = this_cpu_ptr(paiext_root.mapptr);
struct paiext_map *cpump = mp->mapptr;
struct paiext_cb *pcb = cpump->paiext_cb;
paiext_stop(event, PERF_EF_UPDATE);
if (--cpump->active_events == 0) {
/* Disable CPU instruction lookup for PAIE1 control block */
local_ctl_clear_bit(0, CR0_PAI_EXTENSION_BIT);
pcb->acc = 0;
get_lowcore()->aicd = 0;
}
}
/* Create raw data and save it in buffer. Returns number of bytes copied.
* Saves only positive counter entries of the form
* 2 bytes: Number of counter
* 8 bytes: Value of counter
*/
static size_t paiext_copy(struct pai_userdata *userdata, unsigned long *area,
unsigned long *area_old)
{
int i, outidx = 0;
for (i = 1; i <= paiext_cnt; i++) {
u64 val = paiext_getctr(area, i);
u64 val_old = paiext_getctr(area_old, i);
if (val >= val_old)
val -= val_old;
else
val = (~0ULL - val_old) + val + 1;
if (val) {
userdata[outidx].num = i;
userdata[outidx].value = val;
outidx++;
}
}
return outidx * sizeof(*userdata);
}
/* Write sample when one or more counters values are nonzero.
*
* Note: The function paiext_sched_task() and paiext_push_sample() are not
* invoked after function paiext_del() has been called because of function
* perf_sched_cb_dec().
* The function paiext_sched_task() and paiext_push_sample() are only
* called when sampling is active. Function perf_sched_cb_inc()
* has been invoked to install function paiext_sched_task() as call back
* to run at context switch time (see paiext_add()).
*
* This causes function perf_event_context_sched_out() and
* perf_event_context_sched_in() to check whether the PMU has installed an
* sched_task() callback. That callback is not active after paiext_del()
* returns and has deleted the event on that CPU.
*/
static int paiext_push_sample(size_t rawsize, struct paiext_map *cpump,
struct perf_event *event)
{
struct perf_sample_data data;
struct perf_raw_record raw;
struct pt_regs regs;
int overflow;
/* Setup perf sample */
memset(®s, 0, sizeof(regs));
memset(&raw, 0, sizeof(raw));
memset(&data, 0, sizeof(data));
perf_sample_data_init(&data, 0, event->hw.last_period);
if (event->attr.sample_type & PERF_SAMPLE_TID) {
data.tid_entry.pid = task_tgid_nr(current);
data.tid_entry.tid = task_pid_nr(current);
}
if (event->attr.sample_type & PERF_SAMPLE_TIME)
data.time = event->clock();
if (event->attr.sample_type & (PERF_SAMPLE_ID | PERF_SAMPLE_IDENTIFIER))
data.id = event->id;
if (event->attr.sample_type & PERF_SAMPLE_CPU)
data.cpu_entry.cpu = smp_processor_id();
if (event->attr.sample_type & PERF_SAMPLE_RAW) {
raw.frag.size = rawsize;
raw.frag.data = cpump->save;
perf_sample_save_raw_data(&data, &raw);
}
overflow = perf_event_overflow(event, &data, ®s);
perf_event_update_userpage(event);
/* Save NNPA lowcore area after read in event */
memcpy((void *)PAI_SAVE_AREA(event), cpump->area,
PAIE1_CTRBLOCK_SZ);
return overflow;
}
/* Check if there is data to be saved on schedule out of a task. */
static void paiext_have_sample(struct perf_event *event,
struct paiext_map *cpump)
{
size_t rawsize;
if (!event)
return;
rawsize = paiext_copy(cpump->save, cpump->area,
(unsigned long *)PAI_SAVE_AREA(event));
if (rawsize) /* Incremented counters */
paiext_push_sample(rawsize, cpump, event);
}
/* Check if there is data to be saved on schedule out of a task. */
static void paiext_have_samples(void)
{
struct paiext_mapptr *mp = this_cpu_ptr(paiext_root.mapptr);
struct paiext_map *cpump = mp->mapptr;
struct perf_event *event;
list_for_each_entry(event, &cpump->syswide_list, hw.tp_list)
paiext_have_sample(event, cpump);
}
/* Called on schedule-in and schedule-out. No access to event structure,
* but for sampling only event NNPA_ALL is allowed.
*/
static void paiext_sched_task(struct perf_event_pmu_context *pmu_ctx, bool sched_in)
{
/* We started with a clean page on event installation. So read out
* results on schedule_out and if page was dirty, save old values.
*/
if (!sched_in)
paiext_have_samples();
}
/* Attribute definitions for pai extension1 interface. As with other CPU
* Measurement Facilities, there is one attribute per mapped counter.
* The number of mapped counters may vary per machine generation. Use
* the QUERY PROCESSOR ACTIVITY COUNTER INFORMATION (QPACI) instruction
* to determine the number of mapped counters. The instructions returns
* a positive number, which is the highest number of supported counters.
* All counters less than this number are also supported, there are no
* holes. A returned number of zero means no support for mapped counters.
*
* The identification of the counter is a unique number. The chosen range
* is 0x1800 + offset in mapped kernel page.
* All CPU Measurement Facility counters identifiers must be unique and
* the numbers from 0 to 496 are already used for the CPU Measurement
* Counter facility. Number 0x1000 to 0x103e are used for PAI cryptography
* counters.
* Numbers 0xb0000, 0xbc000 and 0xbd000 are already
* used for the CPU Measurement Sampling facility.
*/
PMU_FORMAT_ATTR(event, "config:0-63");
static struct attribute *paiext_format_attr[] = {
&format_attr_event.attr,
NULL,
};
static struct attribute_group paiext_events_group = {
.name = "events",
.attrs = NULL, /* Filled in attr_event_init() */
};
static struct attribute_group paiext_format_group = {
.name = "format",
.attrs = paiext_format_attr,
};
static const struct attribute_group *paiext_attr_groups[] = {
&paiext_events_group,
&paiext_format_group,
NULL,
};
/* Performance monitoring unit for mapped counters */
static struct pmu paiext = {
.task_ctx_nr = perf_hw_context,
.event_init = paiext_event_init,
.add = paiext_add,
.del = paiext_del,
.start = paiext_start,
.stop = paiext_stop,
.read = paiext_read,
.sched_task = paiext_sched_task,
.attr_groups = paiext_attr_groups,
};
/* List of symbolic PAI extension 1 NNPA counter names. */
static const char * const paiext_ctrnames[] = {
[0] = "NNPA_ALL",
[1] = "NNPA_ADD",
[2] = "NNPA_SUB",
[3] = "NNPA_MUL",
[4] = "NNPA_DIV",
[5] = "NNPA_MIN",
[6] = "NNPA_MAX",
[7] = "NNPA_LOG",
[8] = "NNPA_EXP",
[9] = "NNPA_IBM_RESERVED_9",
[10] = "NNPA_RELU",
[11] = "NNPA_TANH",
[12] = "NNPA_SIGMOID",
[13] = "NNPA_SOFTMAX",
[14] = "NNPA_BATCHNORM",
[15] = "NNPA_MAXPOOL2D",
[16] = "NNPA_AVGPOOL2D",
[17] = "NNPA_LSTMACT",
[18] = "NNPA_GRUACT",
[19] = "NNPA_CONVOLUTION",
[20] = "NNPA_MATMUL_OP",
[21] = "NNPA_MATMUL_OP_BCAST23",
[22] = "NNPA_SMALLBATCH",
[23] = "NNPA_LARGEDIM",
[24] = "NNPA_SMALLTENSOR",
[25] = "NNPA_1MFRAME",
[26] = "NNPA_2GFRAME",
[27] = "NNPA_ACCESSEXCEPT",
};
static void __init attr_event_free(struct attribute **attrs, int num)
{
struct perf_pmu_events_attr *pa;
struct device_attribute *dap;
int i;
for (i = 0; i < num; i++) {
dap = container_of(attrs[i], struct device_attribute, attr);
pa = container_of(dap, struct perf_pmu_events_attr, attr);
kfree(pa);
}
kfree(attrs);
}
static int __init attr_event_init_one(struct attribute **attrs, int num)
{
struct perf_pmu_events_attr *pa;
/* Index larger than array_size, no counter name available */
if (num >= ARRAY_SIZE(paiext_ctrnames)) {
attrs[num] = NULL;
return 0;
}
pa = kzalloc(sizeof(*pa), GFP_KERNEL);
if (!pa)
return -ENOMEM;
sysfs_attr_init(&pa->attr.attr);
pa->id = PAI_NNPA_BASE + num;
pa->attr.attr.name = paiext_ctrnames[num];
pa->attr.attr.mode = 0444;
pa->attr.show = cpumf_events_sysfs_show;
pa->attr.store = NULL;
attrs[num] = &pa->attr.attr;
return 0;
}
/* Create PMU sysfs event attributes on the fly. */
static int __init attr_event_init(void)
{
struct attribute **attrs;
int ret, i;
attrs = kmalloc_array(paiext_cnt + 2, sizeof(*attrs), GFP_KERNEL);
if (!attrs)
return -ENOMEM;
for (i = 0; i <= paiext_cnt; i++) {
ret = attr_event_init_one(attrs, i);
if (ret) {
attr_event_free(attrs, i);
return ret;
}
}
attrs[i] = NULL;
paiext_events_group.attrs = attrs;
return 0;
}
static int __init paiext_init(void)
{
struct qpaci_info_block ib;
int rc = -ENOMEM;
if (!test_facility(197))
return 0;
qpaci(&ib);
paiext_cnt = ib.num_nnpa;
if (paiext_cnt >= PAI_NNPA_MAXCTR)
paiext_cnt = PAI_NNPA_MAXCTR;
if (!paiext_cnt)
return 0;
rc = attr_event_init();
if (rc) {
pr_err("Creation of PMU " KMSG_COMPONENT " /sysfs failed\n");
return rc;
}
/* Setup s390dbf facility */
paiext_dbg = debug_register(KMSG_COMPONENT, 2, 256, 128);
if (!paiext_dbg) {
pr_err("Registration of s390dbf " KMSG_COMPONENT " failed\n");
rc = -ENOMEM;
goto out_init;
}
debug_register_view(paiext_dbg, &debug_sprintf_view);
rc = perf_pmu_register(&paiext, KMSG_COMPONENT, -1);
if (rc) {
pr_err("Registration of " KMSG_COMPONENT " PMU failed with "
"rc=%i\n", rc);
goto out_pmu;
}
return 0;
out_pmu:
debug_unregister_view(paiext_dbg, &debug_sprintf_view);
debug_unregister(paiext_dbg);
out_init:
attr_event_free(paiext_events_group.attrs,
ARRAY_SIZE(paiext_ctrnames) + 1);
return rc;
}
device_initcall(paiext_init);
Codebase Browser
linux