From: Kan Liang <[email protected]>
There are two free running counters for client IMC uncore. The custom
event_init() function hardcode their index to 'UNCORE_PMC_IDX_FIXED' and
'UNCORE_PMC_IDX_FIXED + 1'. To support the 'UNCORE_PMC_IDX_FIXED + 1'
case, the generic uncore_perf_event_update is obscurely hacked.
The code quality issue will bring problem when new counter index is
introduced into generic code. For example, free running counter index.
Introduce customized event_read function for client IMC uncore.
The customized function is exactly copied from previous generic
uncore_pmu_event_read.
The 'UNCORE_PMC_IDX_FIXED + 1' case will be isolated for client IMC
uncore only.
Reviewed-by: Thomas Gleixner <[email protected]>
Signed-off-by: Kan Liang <[email protected]>
---
No changes since V6.
arch/x86/events/intel/uncore_snb.c | 33 +++++++++++++++++++++++++++++++--
1 file changed, 31 insertions(+), 2 deletions(-)
diff --git a/arch/x86/events/intel/uncore_snb.c b/arch/x86/events/intel/uncore_snb.c
index aee5e84..df53521 100644
--- a/arch/x86/events/intel/uncore_snb.c
+++ b/arch/x86/events/intel/uncore_snb.c
@@ -450,6 +450,35 @@ static void snb_uncore_imc_event_start(struct perf_event *event, int flags)
uncore_pmu_start_hrtimer(box);
}
+static void snb_uncore_imc_event_read(struct perf_event *event)
+{
+ struct intel_uncore_box *box = uncore_event_to_box(event);
+ u64 prev_count, new_count, delta;
+ int shift;
+
+ /*
+ * There are two free running counters in IMC.
+ * The index for the second one is hardcoded to
+ * UNCORE_PMC_IDX_FIXED + 1.
+ */
+ if (event->hw.idx >= UNCORE_PMC_IDX_FIXED)
+ shift = 64 - uncore_fixed_ctr_bits(box);
+ else
+ shift = 64 - uncore_perf_ctr_bits(box);
+
+ /* the hrtimer might modify the previous event value */
+again:
+ prev_count = local64_read(&event->hw.prev_count);
+ new_count = uncore_read_counter(box, event);
+ if (local64_xchg(&event->hw.prev_count, new_count) != prev_count)
+ goto again;
+
+ delta = (new_count << shift) - (prev_count << shift);
+ delta >>= shift;
+
+ local64_add(delta, &event->count);
+}
+
static void snb_uncore_imc_event_stop(struct perf_event *event, int flags)
{
struct intel_uncore_box *box = uncore_event_to_box(event);
@@ -472,7 +501,7 @@ static void snb_uncore_imc_event_stop(struct perf_event *event, int flags)
* Drain the remaining delta count out of a event
* that we are disabling:
*/
- uncore_perf_event_update(box, event);
+ snb_uncore_imc_event_read(event);
hwc->state |= PERF_HES_UPTODATE;
}
}
@@ -534,7 +563,7 @@ static struct pmu snb_uncore_imc_pmu = {
.del = snb_uncore_imc_event_del,
.start = snb_uncore_imc_event_start,
.stop = snb_uncore_imc_event_stop,
- .read = uncore_pmu_event_read,
+ .read = snb_uncore_imc_event_read,
};
static struct intel_uncore_ops snb_uncore_imc_ops = {
--
2.7.4
From: Kan Liang <[email protected]>
Some uncore has custom pmu. For custom pmu, it does not need to
customize everything. For example, it only needs to customize init()
function for client IMC uncore. Other functions like
add()/del()/start()/stop()/read() can use generic code.
Expose the uncore_pmu_event_add/del/start/stop functions.
Reviewed-by: Thomas Gleixner <[email protected]>
Signed-off-by: Kan Liang <[email protected]>
---
No changes since V6.
arch/x86/events/intel/uncore.c | 8 ++++----
arch/x86/events/intel/uncore.h | 4 ++++
2 files changed, 8 insertions(+), 4 deletions(-)
diff --git a/arch/x86/events/intel/uncore.c b/arch/x86/events/intel/uncore.c
index 0a6f697..15b0737 100644
--- a/arch/x86/events/intel/uncore.c
+++ b/arch/x86/events/intel/uncore.c
@@ -451,7 +451,7 @@ static int uncore_assign_events(struct intel_uncore_box *box, int assign[], int
return ret ? -EINVAL : 0;
}
-static void uncore_pmu_event_start(struct perf_event *event, int flags)
+void uncore_pmu_event_start(struct perf_event *event, int flags)
{
struct intel_uncore_box *box = uncore_event_to_box(event);
int idx = event->hw.idx;
@@ -491,7 +491,7 @@ static void uncore_pmu_event_start(struct perf_event *event, int flags)
}
}
-static void uncore_pmu_event_stop(struct perf_event *event, int flags)
+void uncore_pmu_event_stop(struct perf_event *event, int flags)
{
struct intel_uncore_box *box = uncore_event_to_box(event);
struct hw_perf_event *hwc = &event->hw;
@@ -528,7 +528,7 @@ static void uncore_pmu_event_stop(struct perf_event *event, int flags)
}
}
-static int uncore_pmu_event_add(struct perf_event *event, int flags)
+int uncore_pmu_event_add(struct perf_event *event, int flags)
{
struct intel_uncore_box *box = uncore_event_to_box(event);
struct hw_perf_event *hwc = &event->hw;
@@ -600,7 +600,7 @@ static int uncore_pmu_event_add(struct perf_event *event, int flags)
return 0;
}
-static void uncore_pmu_event_del(struct perf_event *event, int flags)
+void uncore_pmu_event_del(struct perf_event *event, int flags)
{
struct intel_uncore_box *box = uncore_event_to_box(event);
int i;
diff --git a/arch/x86/events/intel/uncore.h b/arch/x86/events/intel/uncore.h
index eb02653..c9e1e0b 100644
--- a/arch/x86/events/intel/uncore.h
+++ b/arch/x86/events/intel/uncore.h
@@ -467,6 +467,10 @@ 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_pmu_start_hrtimer(struct intel_uncore_box *box);
void uncore_pmu_cancel_hrtimer(struct intel_uncore_box *box);
+void uncore_pmu_event_start(struct perf_event *event, int flags);
+void uncore_pmu_event_stop(struct perf_event *event, int flags);
+int uncore_pmu_event_add(struct perf_event *event, int flags);
+void uncore_pmu_event_del(struct perf_event *event, int flags);
void uncore_pmu_event_read(struct perf_event *event);
void uncore_perf_event_update(struct intel_uncore_box *box, struct perf_event *event);
struct event_constraint *
--
2.7.4
From: Kan Liang <[email protected]>
The counters in client IMC uncore are free running counters, not fixed
counters. It should be corrected. The new infrastructure for free
running counter should be applied.
Introduce free running counter type SNB_PCI_UNCORE_IMC_DATA for data
read and data write counters.
Keep the custom event_init() function compatible with old event
encoding.
Clean up other custom event_* functions.
Reviewed-by: Thomas Gleixner <[email protected]>
Signed-off-by: Kan Liang <[email protected]>
---
No changes since V6.
arch/x86/events/intel/uncore_snb.c | 132 ++++++-------------------------------
1 file changed, 20 insertions(+), 112 deletions(-)
diff --git a/arch/x86/events/intel/uncore_snb.c b/arch/x86/events/intel/uncore_snb.c
index df53521..8527c3e 100644
--- a/arch/x86/events/intel/uncore_snb.c
+++ b/arch/x86/events/intel/uncore_snb.c
@@ -285,6 +285,15 @@ static struct uncore_event_desc snb_uncore_imc_events[] = {
#define SNB_UNCORE_PCI_IMC_DATA_WRITES_BASE 0x5054
#define SNB_UNCORE_PCI_IMC_CTR_BASE SNB_UNCORE_PCI_IMC_DATA_READS_BASE
+enum perf_snb_uncore_imc_freerunning_types {
+ SNB_PCI_UNCORE_IMC_DATA = 0,
+ SNB_PCI_UNCORE_IMC_FREERUNNING_TYPE_MAX,
+};
+
+static struct freerunning_counters snb_uncore_imc_freerunning[] = {
+ [SNB_PCI_UNCORE_IMC_DATA] = { SNB_UNCORE_PCI_IMC_DATA_READS_BASE, 0x4, 0x0, 2, 32 },
+};
+
static struct attribute *snb_uncore_imc_formats_attr[] = {
&format_attr_event.attr,
NULL,
@@ -341,9 +350,8 @@ static u64 snb_uncore_imc_read_counter(struct intel_uncore_box *box, struct perf
}
/*
- * custom event_init() function because we define our own fixed, free
- * running counters, so we do not want to conflict with generic uncore
- * logic. Also simplifies processing
+ * Keep the custom event_init() function compatible with old event
+ * encoding for free running counters.
*/
static int snb_uncore_imc_event_init(struct perf_event *event)
{
@@ -405,11 +413,11 @@ static int snb_uncore_imc_event_init(struct perf_event *event)
switch (cfg) {
case SNB_UNCORE_PCI_IMC_DATA_READS:
base = SNB_UNCORE_PCI_IMC_DATA_READS_BASE;
- idx = UNCORE_PMC_IDX_FIXED;
+ idx = UNCORE_PMC_IDX_FREERUNNING;
break;
case SNB_UNCORE_PCI_IMC_DATA_WRITES:
base = SNB_UNCORE_PCI_IMC_DATA_WRITES_BASE;
- idx = UNCORE_PMC_IDX_FIXED + 1;
+ idx = UNCORE_PMC_IDX_FREERUNNING;
break;
default:
return -EINVAL;
@@ -430,104 +438,6 @@ static int snb_uncore_imc_hw_config(struct intel_uncore_box *box, struct perf_ev
return 0;
}
-static void snb_uncore_imc_event_start(struct perf_event *event, int flags)
-{
- struct intel_uncore_box *box = uncore_event_to_box(event);
- u64 count;
-
- if (WARN_ON_ONCE(!(event->hw.state & PERF_HES_STOPPED)))
- return;
-
- event->hw.state = 0;
- box->n_active++;
-
- list_add_tail(&event->active_entry, &box->active_list);
-
- count = snb_uncore_imc_read_counter(box, event);
- local64_set(&event->hw.prev_count, count);
-
- if (box->n_active == 1)
- uncore_pmu_start_hrtimer(box);
-}
-
-static void snb_uncore_imc_event_read(struct perf_event *event)
-{
- struct intel_uncore_box *box = uncore_event_to_box(event);
- u64 prev_count, new_count, delta;
- int shift;
-
- /*
- * There are two free running counters in IMC.
- * The index for the second one is hardcoded to
- * UNCORE_PMC_IDX_FIXED + 1.
- */
- if (event->hw.idx >= UNCORE_PMC_IDX_FIXED)
- shift = 64 - uncore_fixed_ctr_bits(box);
- else
- shift = 64 - uncore_perf_ctr_bits(box);
-
- /* the hrtimer might modify the previous event value */
-again:
- prev_count = local64_read(&event->hw.prev_count);
- new_count = uncore_read_counter(box, event);
- if (local64_xchg(&event->hw.prev_count, new_count) != prev_count)
- goto again;
-
- delta = (new_count << shift) - (prev_count << shift);
- delta >>= shift;
-
- local64_add(delta, &event->count);
-}
-
-static void snb_uncore_imc_event_stop(struct perf_event *event, int flags)
-{
- struct intel_uncore_box *box = uncore_event_to_box(event);
- struct hw_perf_event *hwc = &event->hw;
-
- if (!(hwc->state & PERF_HES_STOPPED)) {
- box->n_active--;
-
- WARN_ON_ONCE(hwc->state & PERF_HES_STOPPED);
- hwc->state |= PERF_HES_STOPPED;
-
- list_del(&event->active_entry);
-
- if (box->n_active == 0)
- uncore_pmu_cancel_hrtimer(box);
- }
-
- if ((flags & PERF_EF_UPDATE) && !(hwc->state & PERF_HES_UPTODATE)) {
- /*
- * Drain the remaining delta count out of a event
- * that we are disabling:
- */
- snb_uncore_imc_event_read(event);
- hwc->state |= PERF_HES_UPTODATE;
- }
-}
-
-static int snb_uncore_imc_event_add(struct perf_event *event, int flags)
-{
- struct intel_uncore_box *box = uncore_event_to_box(event);
- struct hw_perf_event *hwc = &event->hw;
-
- if (!box)
- return -ENODEV;
-
- hwc->state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
- if (!(flags & PERF_EF_START))
- hwc->state |= PERF_HES_ARCH;
-
- snb_uncore_imc_event_start(event, 0);
-
- return 0;
-}
-
-static void snb_uncore_imc_event_del(struct perf_event *event, int flags)
-{
- snb_uncore_imc_event_stop(event, PERF_EF_UPDATE);
-}
-
int snb_pci2phy_map_init(int devid)
{
struct pci_dev *dev = NULL;
@@ -559,11 +469,11 @@ int snb_pci2phy_map_init(int devid)
static struct pmu snb_uncore_imc_pmu = {
.task_ctx_nr = perf_invalid_context,
.event_init = snb_uncore_imc_event_init,
- .add = snb_uncore_imc_event_add,
- .del = snb_uncore_imc_event_del,
- .start = snb_uncore_imc_event_start,
- .stop = snb_uncore_imc_event_stop,
- .read = snb_uncore_imc_event_read,
+ .add = uncore_pmu_event_add,
+ .del = uncore_pmu_event_del,
+ .start = uncore_pmu_event_start,
+ .stop = uncore_pmu_event_stop,
+ .read = uncore_pmu_event_read,
};
static struct intel_uncore_ops snb_uncore_imc_ops = {
@@ -581,12 +491,10 @@ static struct intel_uncore_type snb_uncore_imc = {
.name = "imc",
.num_counters = 2,
.num_boxes = 1,
- .fixed_ctr_bits = 32,
- .fixed_ctr = SNB_UNCORE_PCI_IMC_CTR_BASE,
+ .num_freerunning_types = SNB_PCI_UNCORE_IMC_FREERUNNING_TYPE_MAX,
+ .freerunning = snb_uncore_imc_freerunning,
.event_descs = snb_uncore_imc_events,
.format_group = &snb_uncore_imc_format_group,
- .perf_ctr = SNB_UNCORE_PCI_IMC_DATA_READS_BASE,
- .event_mask = SNB_UNCORE_PCI_IMC_EVENT_MASK,
.ops = &snb_uncore_imc_ops,
.pmu = &snb_uncore_imc_pmu,
};
--
2.7.4
From: Kan Liang <[email protected]>
As of Skylake Server, there are a number of free-running counters in
each IIO Box that collect counts for per box IO clocks and per Port
Input/Output x BW/Utilization.
Add a new PMU for these free-running counters. Don't let them share with
the GP counters of IIO box. Otherwise, it will result in some (probably)
unexpected scheduling artifacts.
The free running counter is read-only and always active. Counting will
be suspended only when the IIO Box is powered down.
There are three types of IIO free running counters on Skylake server, IO
CLOCKS counter, BANDWIDTH counters and UTILIZATION counters.
IO CLOCKS counter is to count IO clocks.
BANDWIDTH counters are to count inbound(PCIe->CPU)/outbound(CPU->PCIe)
bandwidth.
UTILIZATION counters are to count input/output utilization.
The bit width of the free running counters is 36-bits.
Signed-off-by: Kan Liang <[email protected]>
---
Changes since V6:
- Add a new PMU for IIO free-running counters.
arch/x86/events/intel/uncore_snbep.c | 82 ++++++++++++++++++++++++++++++++++++
1 file changed, 82 insertions(+)
diff --git a/arch/x86/events/intel/uncore_snbep.c b/arch/x86/events/intel/uncore_snbep.c
index 22ec65b..d6b302e 100644
--- a/arch/x86/events/intel/uncore_snbep.c
+++ b/arch/x86/events/intel/uncore_snbep.c
@@ -3484,6 +3484,87 @@ static struct intel_uncore_type skx_uncore_iio = {
.format_group = &skx_uncore_iio_format_group,
};
+enum perf_uncore_iio_freerunning_type_id {
+ SKX_IIO_MSR_IOCLK = 0,
+ SKX_IIO_MSR_BW = 1,
+ SKX_IIO_MSR_UTIL = 2,
+
+ SKX_IIO_FREERUNNING_TYPE_MAX,
+};
+
+
+static struct freerunning_counters skx_iio_freerunning[] = {
+ [SKX_IIO_MSR_IOCLK] = { 0xa45, 0x1, 0x20, 1, 36 },
+ [SKX_IIO_MSR_BW] = { 0xb00, 0x1, 0x10, 8, 36 },
+ [SKX_IIO_MSR_UTIL] = { 0xb08, 0x1, 0x10, 8, 36 },
+};
+
+static struct uncore_event_desc skx_uncore_iio_freerunning_events[] = {
+ /* Free-Running IO CLOCKS Counter */
+ INTEL_UNCORE_EVENT_DESC(ioclk, "event=0xff,umask=0x10"),
+ /* Free-Running IIO BANDWIDTH Counters */
+ INTEL_UNCORE_EVENT_DESC(bw_in_port0, "event=0xff,umask=0x20"),
+ INTEL_UNCORE_EVENT_DESC(bw_in_port0.scale, "3.814697266e-6"),
+ INTEL_UNCORE_EVENT_DESC(bw_in_port0.unit, "MiB"),
+ INTEL_UNCORE_EVENT_DESC(bw_in_port1, "event=0xff,umask=0x21"),
+ INTEL_UNCORE_EVENT_DESC(bw_in_port1.scale, "3.814697266e-6"),
+ INTEL_UNCORE_EVENT_DESC(bw_in_port1.unit, "MiB"),
+ INTEL_UNCORE_EVENT_DESC(bw_in_port2, "event=0xff,umask=0x22"),
+ INTEL_UNCORE_EVENT_DESC(bw_in_port2.scale, "3.814697266e-6"),
+ INTEL_UNCORE_EVENT_DESC(bw_in_port2.unit, "MiB"),
+ INTEL_UNCORE_EVENT_DESC(bw_in_port3, "event=0xff,umask=0x23"),
+ INTEL_UNCORE_EVENT_DESC(bw_in_port3.scale, "3.814697266e-6"),
+ INTEL_UNCORE_EVENT_DESC(bw_in_port3.unit, "MiB"),
+ INTEL_UNCORE_EVENT_DESC(bw_out_port0, "event=0xff,umask=0x24"),
+ INTEL_UNCORE_EVENT_DESC(bw_out_port0.scale, "3.814697266e-6"),
+ INTEL_UNCORE_EVENT_DESC(bw_out_port0.unit, "MiB"),
+ INTEL_UNCORE_EVENT_DESC(bw_out_port1, "event=0xff,umask=0x25"),
+ INTEL_UNCORE_EVENT_DESC(bw_out_port1.scale, "3.814697266e-6"),
+ INTEL_UNCORE_EVENT_DESC(bw_out_port1.unit, "MiB"),
+ INTEL_UNCORE_EVENT_DESC(bw_out_port2, "event=0xff,umask=0x26"),
+ INTEL_UNCORE_EVENT_DESC(bw_out_port2.scale, "3.814697266e-6"),
+ INTEL_UNCORE_EVENT_DESC(bw_out_port2.unit, "MiB"),
+ INTEL_UNCORE_EVENT_DESC(bw_out_port3, "event=0xff,umask=0x27"),
+ INTEL_UNCORE_EVENT_DESC(bw_out_port3.scale, "3.814697266e-6"),
+ INTEL_UNCORE_EVENT_DESC(bw_out_port3.unit, "MiB"),
+ /* Free-running IIO UTILIZATION Counters */
+ INTEL_UNCORE_EVENT_DESC(util_in_port0, "event=0xff,umask=0x30"),
+ INTEL_UNCORE_EVENT_DESC(util_out_port0, "event=0xff,umask=0x31"),
+ INTEL_UNCORE_EVENT_DESC(util_in_port1, "event=0xff,umask=0x32"),
+ INTEL_UNCORE_EVENT_DESC(util_out_port1, "event=0xff,umask=0x33"),
+ INTEL_UNCORE_EVENT_DESC(util_in_port2, "event=0xff,umask=0x34"),
+ INTEL_UNCORE_EVENT_DESC(util_out_port2, "event=0xff,umask=0x35"),
+ INTEL_UNCORE_EVENT_DESC(util_in_port3, "event=0xff,umask=0x36"),
+ INTEL_UNCORE_EVENT_DESC(util_out_port3, "event=0xff,umask=0x37"),
+ { /* end: all zeroes */ },
+};
+
+static struct intel_uncore_ops skx_uncore_iio_freerunning_ops = {
+ .read_counter = uncore_msr_read_counter,
+};
+
+static struct attribute *skx_uncore_iio_freerunning_formats_attr[] = {
+ &format_attr_event.attr,
+ &format_attr_umask.attr,
+ NULL,
+};
+
+static const struct attribute_group skx_uncore_iio_freerunning_format_group = {
+ .name = "format",
+ .attrs = skx_uncore_iio_freerunning_formats_attr,
+};
+
+static struct intel_uncore_type skx_uncore_iio_free_running = {
+ .name = "iio_free_running",
+ .num_counters = 17,
+ .num_boxes = 6,
+ .num_freerunning_types = SKX_IIO_FREERUNNING_TYPE_MAX,
+ .freerunning = skx_iio_freerunning,
+ .ops = &skx_uncore_iio_freerunning_ops,
+ .event_descs = skx_uncore_iio_freerunning_events,
+ .format_group = &skx_uncore_iio_freerunning_format_group,
+};
+
static struct attribute *skx_uncore_formats_attr[] = {
&format_attr_event.attr,
&format_attr_umask.attr,
@@ -3557,6 +3638,7 @@ static struct intel_uncore_type *skx_msr_uncores[] = {
&skx_uncore_ubox,
&skx_uncore_chabox,
&skx_uncore_iio,
+ &skx_uncore_iio_free_running,
&skx_uncore_irp,
&skx_uncore_pcu,
NULL,
--
2.7.4
From: Kan Liang <[email protected]>
There are a number of free running counters introduced for uncore, which
provide highly valuable information to a wide array of customers.
However, the generic uncore code doesn't support them yet.
The free running counters will be specially handled based on their
unique attributes
- They are read-only. They cannot be enable/disable.
- The event and the counter are always 1:1 mapped. It doesn't need to
be assigned nor tracked by event_list.
- They are always active. It doesn't need to check the availability.
- They have different bit width.
Also, using inline helpers to replace the check for fixed counter and
free running counter.
Reviewed-by: Thomas Gleixner <[email protected]>
Signed-off-by: Kan Liang <[email protected]>
---
No changes since V6.
arch/x86/events/intel/uncore.c | 68 +++++++++++++++++++++++++++++++++++++++---
1 file changed, 64 insertions(+), 4 deletions(-)
diff --git a/arch/x86/events/intel/uncore.c b/arch/x86/events/intel/uncore.c
index 3b0f93e..0a6f697 100644
--- a/arch/x86/events/intel/uncore.c
+++ b/arch/x86/events/intel/uncore.c
@@ -203,7 +203,7 @@ static void uncore_assign_hw_event(struct intel_uncore_box *box,
hwc->idx = idx;
hwc->last_tag = ++box->tags[idx];
- if (hwc->idx == UNCORE_PMC_IDX_FIXED) {
+ if (uncore_pmc_fixed(hwc->idx)) {
hwc->event_base = uncore_fixed_ctr(box);
hwc->config_base = uncore_fixed_ctl(box);
return;
@@ -218,7 +218,9 @@ void uncore_perf_event_update(struct intel_uncore_box *box, struct perf_event *e
u64 prev_count, new_count, delta;
int shift;
- if (event->hw.idx == UNCORE_PMC_IDX_FIXED)
+ if (uncore_pmc_freerunning(event->hw.idx))
+ shift = 64 - uncore_freerunning_bits(box, event);
+ else if (uncore_pmc_fixed(event->hw.idx))
shift = 64 - uncore_fixed_ctr_bits(box);
else
shift = 64 - uncore_perf_ctr_bits(box);
@@ -454,10 +456,25 @@ static void uncore_pmu_event_start(struct perf_event *event, int flags)
struct intel_uncore_box *box = uncore_event_to_box(event);
int idx = event->hw.idx;
- if (WARN_ON_ONCE(!(event->hw.state & PERF_HES_STOPPED)))
+ if (WARN_ON_ONCE(idx == -1 || idx >= UNCORE_PMC_IDX_MAX))
return;
- if (WARN_ON_ONCE(idx == -1 || idx >= UNCORE_PMC_IDX_MAX))
+ /*
+ * Free running counter is read-only and always active.
+ * Use the current counter value as start point.
+ * There is no overflow interrupt for free running counter.
+ * Use hrtimer to periodically poll the counter to avoid overflow.
+ */
+ if (uncore_pmc_freerunning(event->hw.idx)) {
+ list_add_tail(&event->active_entry, &box->active_list);
+ local64_set(&event->hw.prev_count,
+ uncore_read_counter(box, event));
+ if (box->n_active++ == 0)
+ uncore_pmu_start_hrtimer(box);
+ return;
+ }
+
+ if (WARN_ON_ONCE(!(event->hw.state & PERF_HES_STOPPED)))
return;
event->hw.state = 0;
@@ -479,6 +496,15 @@ static void uncore_pmu_event_stop(struct perf_event *event, int flags)
struct intel_uncore_box *box = uncore_event_to_box(event);
struct hw_perf_event *hwc = &event->hw;
+ /* Cannot disable free running counter which is read-only */
+ if (uncore_pmc_freerunning(hwc->idx)) {
+ list_del(&event->active_entry);
+ if (--box->n_active == 0)
+ uncore_pmu_cancel_hrtimer(box);
+ uncore_perf_event_update(box, event);
+ return;
+ }
+
if (__test_and_clear_bit(hwc->idx, box->active_mask)) {
uncore_disable_event(box, event);
box->n_active--;
@@ -512,6 +538,17 @@ static int uncore_pmu_event_add(struct perf_event *event, int flags)
if (!box)
return -ENODEV;
+ /*
+ * The free funning counter is assigned in event_init().
+ * The free running counter event and free running counter
+ * are 1:1 mapped. It doesn't need to be tracked in event_list.
+ */
+ if (uncore_pmc_freerunning(hwc->idx)) {
+ if (flags & PERF_EF_START)
+ uncore_pmu_event_start(event, 0);
+ return 0;
+ }
+
ret = n = uncore_collect_events(box, event, false);
if (ret < 0)
return ret;
@@ -570,6 +607,14 @@ static void uncore_pmu_event_del(struct perf_event *event, int flags)
uncore_pmu_event_stop(event, PERF_EF_UPDATE);
+ /*
+ * The event for free running counter is not tracked by event_list.
+ * It doesn't need to force event->hw.idx = -1 to reassign the counter.
+ * Because the event and the free running counter are 1:1 mapped.
+ */
+ if (uncore_pmc_freerunning(event->hw.idx))
+ return;
+
for (i = 0; i < box->n_events; i++) {
if (event == box->event_list[i]) {
uncore_put_event_constraint(box, event);
@@ -603,6 +648,10 @@ static int uncore_validate_group(struct intel_uncore_pmu *pmu,
struct intel_uncore_box *fake_box;
int ret = -EINVAL, n;
+ /* The free running counter is always active. */
+ if (uncore_pmc_freerunning(event->hw.idx))
+ return 0;
+
fake_box = uncore_alloc_box(pmu->type, NUMA_NO_NODE);
if (!fake_box)
return -ENOMEM;
@@ -690,6 +739,17 @@ static int uncore_pmu_event_init(struct perf_event *event)
/* fixed counters have event field hardcoded to zero */
hwc->config = 0ULL;
+ } else if (is_freerunning_event(event)) {
+ if (!check_valid_freerunning_event(box, event))
+ return -EINVAL;
+ event->hw.idx = UNCORE_PMC_IDX_FREERUNNING;
+ /*
+ * The free running counter event and free running counter
+ * are always 1:1 mapped.
+ * The free running counter is always active.
+ * Assign the free running counter here.
+ */
+ event->hw.event_base = uncore_freerunning_counter(box, event);
} else {
hwc->config = event->attr.config &
(pmu->type->event_mask | ((u64)pmu->type->event_mask_ext << 32));
--
2.7.4
From: Kan Liang <[email protected]>
There is no index which is bigger than UNCORE_PMC_IDX_FIXED. The only
exception is client IMC uncore. It has customized function to deal with
the 'UNCORE_PMC_IDX_FIXED + 1' case. It does not touch the generic code.
For generic code, it is not correct to use >= to check fixed counter.
The code quality issue will bring problem when new counter index is
introduced.
Reviewed-by: Thomas Gleixner <[email protected]>
Signed-off-by: Kan Liang <[email protected]>
---
No changes since V6.
arch/x86/events/intel/uncore.c | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/arch/x86/events/intel/uncore.c b/arch/x86/events/intel/uncore.c
index a7956fc..3b0f93e 100644
--- a/arch/x86/events/intel/uncore.c
+++ b/arch/x86/events/intel/uncore.c
@@ -218,7 +218,7 @@ void uncore_perf_event_update(struct intel_uncore_box *box, struct perf_event *e
u64 prev_count, new_count, delta;
int shift;
- if (event->hw.idx >= UNCORE_PMC_IDX_FIXED)
+ if (event->hw.idx == UNCORE_PMC_IDX_FIXED)
shift = 64 - uncore_fixed_ctr_bits(box);
else
shift = 64 - uncore_perf_ctr_bits(box);
--
2.7.4
From: Kan Liang <[email protected]>
For Nehalem and Westmere, there is only one fixed counter for W-Box.
There is no index which is bigger than UNCORE_PMC_IDX_FIXED.
It is not correct to use >= to check fixed counter.
The code quality issue will bring problem when new counter index is
introduced.
Reviewed-by: Thomas Gleixner <[email protected]>
Signed-off-by: Kan Liang <[email protected]>
---
No changes since V6.
arch/x86/events/intel/uncore_nhmex.c | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/arch/x86/events/intel/uncore_nhmex.c b/arch/x86/events/intel/uncore_nhmex.c
index 93e7a83..173e267 100644
--- a/arch/x86/events/intel/uncore_nhmex.c
+++ b/arch/x86/events/intel/uncore_nhmex.c
@@ -246,7 +246,7 @@ static void nhmex_uncore_msr_enable_event(struct intel_uncore_box *box, struct p
{
struct hw_perf_event *hwc = &event->hw;
- if (hwc->idx >= UNCORE_PMC_IDX_FIXED)
+ if (hwc->idx == UNCORE_PMC_IDX_FIXED)
wrmsrl(hwc->config_base, NHMEX_PMON_CTL_EN_BIT0);
else if (box->pmu->type->event_mask & NHMEX_PMON_CTL_EN_BIT0)
wrmsrl(hwc->config_base, hwc->config | NHMEX_PMON_CTL_EN_BIT22);
--
2.7.4
From: Kan Liang <[email protected]>
There are a number of free running counters introduced for uncore, which
provide highly valuable information to a wide array of customers.
For example, Skylake Server has IIO free running counters to collect
Input/Output x BW/Utilization.
There is NO event available on the general purpose counters, that is
exactly the same as the free running counters. The generic uncore code
needs to be enhanced to support the new counters.
In the uncore document, there is no event-code assigned to free running
counters. Some events need to be defined to indicate the free running
counters. The events are encoded as event-code + umask-code.
The event-code for all free running counters is 0xff, which is the same
as the fixed counters.
- It has not been decided what code will be used for common events on
future platforms. 0xff is the only one which will definitely not be
used as any common event-code.
- Cannot re-use current events on the general purpose counters. Because
there is NO event available, that is exactly the same as the free
running counters.
- Even in the existing codes, the fixed counters for core, that have the
same event-code, may count different things. Hence, it should not
surprise the users if the free running counters that share the same
event-code also count different things.
Umask will be used to distinguish the counters.
The umask-code is used to distinguish a fixed counter and a free running
counter, and different types of free running counters.
For fixed counters, the umask-code is 0x0X.
X indicates the index of the fixed counter, which starts from 0.
- Compatible with the old event encoding.
- Currently, there is only one fixed counter. There are still 15
reserved spaces for extension.
For free running counters, the umask-code uses the rest of the space.
It would bare the format of 0xXY.
X stands for the type of free running counters, which starts from 1.
Y stands for the index of free running counters of same type, which
starts from 0.
- The free running counters do different thing. It can be categorized to
several types, according to the MSR location, bit width and
definition. E.g. there are three types of IIO free running counters on
Skylake server to monitor IO CLOCKS, BANDWIDTH and UTILIZATION on
different ports. It makes it easy to locate the free running counter
of a specific type.
- So far, there are at most 8 counters of each type. There are still 8
reserved spaces for extension.
Introduce a new index to indicate the free running counters. Only one
index is enough for all free running counters. Because the free running
countes are always active, and the event and free running counter are
always 1:1 mapped. It does not need extra index to indicate the assigned
counter.
Introduce a new data structure to store free running counters related
information for each type. It includes the number of counters, bit
width, base address, offset between counters and offset between boxes.
Introduce several inline helpers to check index for fixed counter and
free running counter, validate free running counter event, and retrieve
the free running counter information according to box and event.
Reviewed-by: Thomas Gleixner <[email protected]>
Signed-off-by: Kan Liang <[email protected]>
---
No changes since V6.
arch/x86/events/intel/uncore.h | 123 ++++++++++++++++++++++++++++++++++++++++-
1 file changed, 122 insertions(+), 1 deletion(-)
diff --git a/arch/x86/events/intel/uncore.h b/arch/x86/events/intel/uncore.h
index 414dc7e..eb02653 100644
--- a/arch/x86/events/intel/uncore.h
+++ b/arch/x86/events/intel/uncore.h
@@ -12,8 +12,13 @@
#define UNCORE_FIXED_EVENT 0xff
#define UNCORE_PMC_IDX_MAX_GENERIC 8
+#define UNCORE_PMC_IDX_MAX_FIXED 1
+#define UNCORE_PMC_IDX_MAX_FREERUNNING 1
#define UNCORE_PMC_IDX_FIXED UNCORE_PMC_IDX_MAX_GENERIC
-#define UNCORE_PMC_IDX_MAX (UNCORE_PMC_IDX_FIXED + 1)
+#define UNCORE_PMC_IDX_FREERUNNING (UNCORE_PMC_IDX_FIXED + \
+ UNCORE_PMC_IDX_MAX_FIXED)
+#define UNCORE_PMC_IDX_MAX (UNCORE_PMC_IDX_FREERUNNING + \
+ UNCORE_PMC_IDX_MAX_FREERUNNING)
#define UNCORE_PCI_DEV_FULL_DATA(dev, func, type, idx) \
((dev << 24) | (func << 16) | (type << 8) | idx)
@@ -35,6 +40,7 @@ struct intel_uncore_ops;
struct intel_uncore_pmu;
struct intel_uncore_box;
struct uncore_event_desc;
+struct freerunning_counters;
struct intel_uncore_type {
const char *name;
@@ -42,6 +48,7 @@ struct intel_uncore_type {
int num_boxes;
int perf_ctr_bits;
int fixed_ctr_bits;
+ int num_freerunning_types;
unsigned perf_ctr;
unsigned event_ctl;
unsigned event_mask;
@@ -59,6 +66,7 @@ struct intel_uncore_type {
struct intel_uncore_pmu *pmus;
struct intel_uncore_ops *ops;
struct uncore_event_desc *event_descs;
+ struct freerunning_counters *freerunning;
const struct attribute_group *attr_groups[4];
struct pmu *pmu; /* for custom pmu ops */
};
@@ -129,6 +137,14 @@ struct uncore_event_desc {
const char *config;
};
+struct freerunning_counters {
+ unsigned int counter_base;
+ unsigned int counter_offset;
+ unsigned int box_offset;
+ unsigned int num_counters;
+ unsigned int bits;
+};
+
struct pci2phy_map {
struct list_head list;
int segment;
@@ -157,6 +173,16 @@ static ssize_t __uncore_##_var##_show(struct kobject *kobj, \
static struct kobj_attribute format_attr_##_var = \
__ATTR(_name, 0444, __uncore_##_var##_show, NULL)
+static inline bool uncore_pmc_fixed(int idx)
+{
+ return idx == UNCORE_PMC_IDX_FIXED;
+}
+
+static inline bool uncore_pmc_freerunning(int idx)
+{
+ return idx == UNCORE_PMC_IDX_FREERUNNING;
+}
+
static inline unsigned uncore_pci_box_ctl(struct intel_uncore_box *box)
{
return box->pmu->type->box_ctl;
@@ -214,6 +240,60 @@ static inline unsigned uncore_msr_fixed_ctr(struct intel_uncore_box *box)
return box->pmu->type->fixed_ctr + uncore_msr_box_offset(box);
}
+
+/*
+ * In the uncore document, there is no event-code assigned to free running
+ * counters. Some events need to be defined to indicate the free running
+ * counters. The events are encoded as event-code + umask-code.
+ *
+ * The event-code for all free running counters is 0xff, which is the same as
+ * the fixed counters.
+ *
+ * The umask-code is used to distinguish a fixed counter and a free running
+ * counter, and different types of free running counters.
+ * - For fixed counters, the umask-code is 0x0X.
+ * X indicates the index of the fixed counter, which starts from 0.
+ * - For free running counters, the umask-code uses the rest of the space.
+ * It would bare the format of 0xXY.
+ * X stands for the type of free running counters, which starts from 1.
+ * Y stands for the index of free running counters of same type, which
+ * starts from 0.
+ *
+ * For example, there are three types of IIO free running counters on Skylake
+ * server, IO CLOCKS counters, BANDWIDTH counters and UTILIZATION counters.
+ * The event-code for all the free running counters is 0xff.
+ * 'ioclk' is the first counter of IO CLOCKS. IO CLOCKS is the first type,
+ * which umask-code starts from 0x10.
+ * So 'ioclk' is encoded as event=0xff,umask=0x10
+ * 'bw_in_port2' is the third counter of BANDWIDTH counters. BANDWIDTH is
+ * the second type, which umask-code starts from 0x20.
+ * So 'bw_in_port2' is encoded as event=0xff,umask=0x22
+ */
+static inline unsigned int uncore_freerunning_idx(u64 config)
+{
+ return ((config >> 8) & 0xf);
+}
+
+#define UNCORE_FREERUNNING_UMASK_START 0x10
+
+static inline unsigned int uncore_freerunning_type(u64 config)
+{
+ return ((((config >> 8) - UNCORE_FREERUNNING_UMASK_START) >> 4) & 0xf);
+}
+
+static inline
+unsigned int uncore_freerunning_counter(struct intel_uncore_box *box,
+ struct perf_event *event)
+{
+ unsigned int type = uncore_freerunning_type(event->attr.config);
+ unsigned int idx = uncore_freerunning_idx(event->attr.config);
+ struct intel_uncore_pmu *pmu = box->pmu;
+
+ return pmu->type->freerunning[type].counter_base +
+ pmu->type->freerunning[type].counter_offset * idx +
+ pmu->type->freerunning[type].box_offset * pmu->pmu_idx;
+}
+
static inline
unsigned uncore_msr_event_ctl(struct intel_uncore_box *box, int idx)
{
@@ -276,11 +356,52 @@ static inline int uncore_fixed_ctr_bits(struct intel_uncore_box *box)
return box->pmu->type->fixed_ctr_bits;
}
+static inline
+unsigned int uncore_freerunning_bits(struct intel_uncore_box *box,
+ struct perf_event *event)
+{
+ unsigned int type = uncore_freerunning_type(event->attr.config);
+
+ return box->pmu->type->freerunning[type].bits;
+}
+
+static inline int uncore_num_freerunning(struct intel_uncore_box *box,
+ struct perf_event *event)
+{
+ unsigned int type = uncore_freerunning_type(event->attr.config);
+
+ return box->pmu->type->freerunning[type].num_counters;
+}
+
+static inline int uncore_num_freerunning_types(struct intel_uncore_box *box,
+ struct perf_event *event)
+{
+ return box->pmu->type->num_freerunning_types;
+}
+
+static inline bool check_valid_freerunning_event(struct intel_uncore_box *box,
+ struct perf_event *event)
+{
+ unsigned int type = uncore_freerunning_type(event->attr.config);
+ unsigned int idx = uncore_freerunning_idx(event->attr.config);
+
+ return (type < uncore_num_freerunning_types(box, event)) &&
+ (idx < uncore_num_freerunning(box, event));
+}
+
static inline int uncore_num_counters(struct intel_uncore_box *box)
{
return box->pmu->type->num_counters;
}
+static inline bool is_freerunning_event(struct perf_event *event)
+{
+ u64 cfg = event->attr.config;
+
+ return ((cfg & UNCORE_FIXED_EVENT) == UNCORE_FIXED_EVENT) &&
+ (((cfg >> 8) & 0xff) >= UNCORE_FREERUNNING_UMASK_START);
+}
+
static inline void uncore_disable_box(struct intel_uncore_box *box)
{
if (box->pmu->type->ops->disable_box)
--
2.7.4
>
> From: Kan Liang <[email protected]>
>
> As of Skylake Server, there are a number of free-running counters in
> each IIO Box that collect counts for per box IO clocks and per Port
> Input/Output x BW/Utilization.
> Add a new PMU for these free-running counters. Don't let them share with
> the GP counters of IIO box. Otherwise, it will result in some (probably)
> unexpected scheduling artifacts.
>
> The free running counter is read-only and always active. Counting will
> be suspended only when the IIO Box is powered down.
>
> There are three types of IIO free running counters on Skylake server, IO
> CLOCKS counter, BANDWIDTH counters and UTILIZATION counters.
> IO CLOCKS counter is to count IO clocks.
> BANDWIDTH counters are to count inbound(PCIe->CPU)/outbound(CPU-
> >PCIe)
> bandwidth.
> UTILIZATION counters are to count input/output utilization.
>
> The bit width of the free running counters is 36-bits.
>
> Signed-off-by: Kan Liang <[email protected]>
> ---
>
> Changes since V6:
> - Add a new PMU for IIO free-running counters.
>
Hi Peter,
Could you please review the patch?
If it's OK for you, could you please merge the patch set?
Thanks,
Kan
> arch/x86/events/intel/uncore_snbep.c | 82
> ++++++++++++++++++++++++++++++++++++
> 1 file changed, 82 insertions(+)
>
> diff --git a/arch/x86/events/intel/uncore_snbep.c
> b/arch/x86/events/intel/uncore_snbep.c
> index 22ec65b..d6b302e 100644
> --- a/arch/x86/events/intel/uncore_snbep.c
> +++ b/arch/x86/events/intel/uncore_snbep.c
> @@ -3484,6 +3484,87 @@ static struct intel_uncore_type skx_uncore_iio = {
> .format_group = &skx_uncore_iio_format_group,
> };
>
> +enum perf_uncore_iio_freerunning_type_id {
> + SKX_IIO_MSR_IOCLK = 0,
> + SKX_IIO_MSR_BW = 1,
> + SKX_IIO_MSR_UTIL = 2,
> +
> + SKX_IIO_FREERUNNING_TYPE_MAX,
> +};
> +
> +
> +static struct freerunning_counters skx_iio_freerunning[] = {
> + [SKX_IIO_MSR_IOCLK] = { 0xa45, 0x1, 0x20, 1, 36 },
> + [SKX_IIO_MSR_BW] = { 0xb00, 0x1, 0x10, 8, 36 },
> + [SKX_IIO_MSR_UTIL] = { 0xb08, 0x1, 0x10, 8, 36 },
> +};
> +
> +static struct uncore_event_desc skx_uncore_iio_freerunning_events[] = {
> + /* Free-Running IO CLOCKS Counter */
> + INTEL_UNCORE_EVENT_DESC(ioclk,
> "event=0xff,umask=0x10"),
> + /* Free-Running IIO BANDWIDTH Counters */
> + INTEL_UNCORE_EVENT_DESC(bw_in_port0,
> "event=0xff,umask=0x20"),
> + INTEL_UNCORE_EVENT_DESC(bw_in_port0.scale,
> "3.814697266e-6"),
> + INTEL_UNCORE_EVENT_DESC(bw_in_port0.unit, "MiB"),
> + INTEL_UNCORE_EVENT_DESC(bw_in_port1,
> "event=0xff,umask=0x21"),
> + INTEL_UNCORE_EVENT_DESC(bw_in_port1.scale,
> "3.814697266e-6"),
> + INTEL_UNCORE_EVENT_DESC(bw_in_port1.unit, "MiB"),
> + INTEL_UNCORE_EVENT_DESC(bw_in_port2,
> "event=0xff,umask=0x22"),
> + INTEL_UNCORE_EVENT_DESC(bw_in_port2.scale,
> "3.814697266e-6"),
> + INTEL_UNCORE_EVENT_DESC(bw_in_port2.unit, "MiB"),
> + INTEL_UNCORE_EVENT_DESC(bw_in_port3,
> "event=0xff,umask=0x23"),
> + INTEL_UNCORE_EVENT_DESC(bw_in_port3.scale,
> "3.814697266e-6"),
> + INTEL_UNCORE_EVENT_DESC(bw_in_port3.unit, "MiB"),
> + INTEL_UNCORE_EVENT_DESC(bw_out_port0,
> "event=0xff,umask=0x24"),
> + INTEL_UNCORE_EVENT_DESC(bw_out_port0.scale,
> "3.814697266e-6"),
> + INTEL_UNCORE_EVENT_DESC(bw_out_port0.unit, "MiB"),
> + INTEL_UNCORE_EVENT_DESC(bw_out_port1,
> "event=0xff,umask=0x25"),
> + INTEL_UNCORE_EVENT_DESC(bw_out_port1.scale,
> "3.814697266e-6"),
> + INTEL_UNCORE_EVENT_DESC(bw_out_port1.unit, "MiB"),
> + INTEL_UNCORE_EVENT_DESC(bw_out_port2,
> "event=0xff,umask=0x26"),
> + INTEL_UNCORE_EVENT_DESC(bw_out_port2.scale,
> "3.814697266e-6"),
> + INTEL_UNCORE_EVENT_DESC(bw_out_port2.unit, "MiB"),
> + INTEL_UNCORE_EVENT_DESC(bw_out_port3,
> "event=0xff,umask=0x27"),
> + INTEL_UNCORE_EVENT_DESC(bw_out_port3.scale,
> "3.814697266e-6"),
> + INTEL_UNCORE_EVENT_DESC(bw_out_port3.unit, "MiB"),
> + /* Free-running IIO UTILIZATION Counters */
> + INTEL_UNCORE_EVENT_DESC(util_in_port0,
> "event=0xff,umask=0x30"),
> + INTEL_UNCORE_EVENT_DESC(util_out_port0,
> "event=0xff,umask=0x31"),
> + INTEL_UNCORE_EVENT_DESC(util_in_port1,
> "event=0xff,umask=0x32"),
> + INTEL_UNCORE_EVENT_DESC(util_out_port1,
> "event=0xff,umask=0x33"),
> + INTEL_UNCORE_EVENT_DESC(util_in_port2,
> "event=0xff,umask=0x34"),
> + INTEL_UNCORE_EVENT_DESC(util_out_port2,
> "event=0xff,umask=0x35"),
> + INTEL_UNCORE_EVENT_DESC(util_in_port3,
> "event=0xff,umask=0x36"),
> + INTEL_UNCORE_EVENT_DESC(util_out_port3,
> "event=0xff,umask=0x37"),
> + { /* end: all zeroes */ },
> +};
> +
> +static struct intel_uncore_ops skx_uncore_iio_freerunning_ops = {
> + .read_counter = uncore_msr_read_counter,
> +};
> +
> +static struct attribute *skx_uncore_iio_freerunning_formats_attr[] = {
> + &format_attr_event.attr,
> + &format_attr_umask.attr,
> + NULL,
> +};
> +
> +static const struct attribute_group
> skx_uncore_iio_freerunning_format_group = {
> + .name = "format",
> + .attrs = skx_uncore_iio_freerunning_formats_attr,
> +};
> +
> +static struct intel_uncore_type skx_uncore_iio_free_running = {
> + .name = "iio_free_running",
> + .num_counters = 17,
> + .num_boxes = 6,
> + .num_freerunning_types = SKX_IIO_FREERUNNING_TYPE_MAX,
> + .freerunning = skx_iio_freerunning,
> + .ops = &skx_uncore_iio_freerunning_ops,
> + .event_descs = skx_uncore_iio_freerunning_events,
> + .format_group =
> &skx_uncore_iio_freerunning_format_group,
> +};
> +
> static struct attribute *skx_uncore_formats_attr[] = {
> &format_attr_event.attr,
> &format_attr_umask.attr,
> @@ -3557,6 +3638,7 @@ static struct intel_uncore_type *skx_msr_uncores[]
> = {
> &skx_uncore_ubox,
> &skx_uncore_chabox,
> &skx_uncore_iio,
> + &skx_uncore_iio_free_running,
> &skx_uncore_irp,
> &skx_uncore_pcu,
> NULL,
> --
> 2.7.4