metag: perf fixes and OProfile support
cover metag: perf fixes and OProfile support
This patchset fixes some issues in the Meta support for perf events (particularly perf counter interrupts), and then adds OProfile support to the Meta architecture based on perf.
This is aimed at v3.10.
James Hogan (9):
metag: perf: fix core internal / perf channel mux
metag: perf: fix wrap handling in delta calculation
metag: perf: fixes for interrupting perf counters
metag: perf: add missing prev_count updates
metag: perf: fix frequency sampling (dynamic period)
metag: perf: use hard_processor_id() to get thread
metag: perf: don't reset TXTACTCYC
metag: perf: prepare for use by oprofile
metag: OProfile support
arch/metag/Kconfig | 4 ++ arch/metag/Makefile | 2 + arch/metag/kernel/perf/perf_event.c | 74 +++++++++++++++++++++++++------------ arch/metag/oprofile/Makefile | 17 +++++++++ arch/metag/oprofile/backtrace.c | 63 +++++++++++++++++++++++++++++++ arch/metag/oprofile/backtrace.h | 6 +++ arch/metag/oprofile/common.c | 66 +++++++++++++++++++++++++++++++++ 7 files changed, 209 insertions(+), 23 deletions(-) create mode 100644 arch/metag/oprofile/Makefile create mode 100644 arch/metag/oprofile/backtrace.c create mode 100644 arch/metag/oprofile/backtrace.h create mode 100644 arch/metag/oprofile/common.c
Signature · 1 line
1.8.1.2
PATCH 1/9 metag: perf: fix core internal / perf channel mux
The value written to the PERF_ICOREx or PERF_CHANx register to select the performance events for the core internal and perf channel events was (tmp & 0x0f), but tmp was set to (config & 0xf0) so it would always be
0. Correct it to use config instead of tmp.
--- arch/metag/kernel/perf/perf_event.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
arch/metag/kernel/perf/perf_event.c +1 -1
@@ -634,7 +634,7 @@ static void metag_pmu_enable_counter(struct hw_perf_event *event, int idx)break ; } -metag_out32 ((tmp &0x0f ), perf_addr); +metag_out32 ((config &0x0f ), perf_addr);/* * Now we use the high nibble as the performance event to
Signature · 1 line
1.8.1.2
PATCH 2/9 metag: perf: fix wrap handling in delta calculation
In reply to James HoganWhen calculating the delta, mask with MAX_PERIOD (24 bits) to handle wrapping, which particularly happens with periodic sampling since the value is intentionally set so that it will overflow soon.
--- arch/metag/kernel/perf/perf_event.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
arch/metag/kernel/perf/perf_event.c +1 -1
@@ -211,7 +211,7 @@ again:/* * Calculate the delta and add it to the counter. */ - delta = new_raw_count - prev_raw_count; + delta = (new_raw_count - prev_raw_count) &MAX_PERIOD ;local64_add (delta, &event->count); }
Signature · 1 line
1.8.1.2
PATCH 3/9 metag: perf: fixes for interrupting perf counters
In reply to James HoganThe overflow handler needs to read modify write when re-enabling the counter so as not to change the counter value as it may have been changed to ready the next interrupt on overflow. Similarly for interrupting counters metag_pmu_enable_counter needs to leave the counter value unchanged rather than resetting it to zero.
--- arch/metag/kernel/perf/perf_event.c | 22 +++++++++++++++------- 1 file changed, 15 insertions(+), 7 deletions(-)
arch/metag/kernel/perf/perf_event.c +15 -7
@@ -643,13 +643,15 @@ static void metag_pmu_enable_counter(struct hw_perf_event *event, int idx) config = tmp >>4 ; } -/* -* Enabled counters start from 0. Early cores clear the count on -* write but newer cores don't, so we make sure that the count is -* set to 0. -*/ tmp = ((config &0xf ) <<28 ) | ((1 <<24 ) << cpu_2_hwthread_id[get_cpu ()]); +if (metag_pmu->max_period) +/* +* Cores supporting overflow interrupts may have had the counter +* set to a specific value that needs preserving. +*/ + tmp |=metag_in32 (PERF_COUNT (idx)) &0x00ffffff ; +metag_out32 (tmp,PERF_COUNT (idx)); unlock:raw_spin_unlock_irqrestore (&events->pmu_lock, flags);
@@ -764,10 +766,16 @@ static irqreturn_t metag_pmu_counter_overflow(int irq, void *dev)/* * Enable the counter again once core overflow processing has -* completed. +* completed. Note the counter value may have been modified while it was +* inactive to set it up ready for the next interrupt. */ -if (!perf_event_overflow (event, &sampledata, regs)) +if (!perf_event_overflow (event, &sampledata, regs)) { +__global_lock2 (flags); + counter = (counter &0xff000000 ) | + (metag_in32 (PERF_COUNT (idx)) &0x00ffffff );metag_out32 (counter,PERF_COUNT (idx)); +__global_unlock2 (flags); + }return IRQ_HANDLED ; }
Signature · 1 line
1.8.1.2
PATCH 4/9 metag: perf: add missing prev_count updates
In reply to James HoganThe prev_count needs setting when changing the counter value, otherwise the calculated delta will be wrong, which for frequency sampling (dynamic period sampling) results in sampling at too high a frequency.
For non-interrupting performance counters it should also be cleared when enabling the counter since the write to the PERF_COUNT register will clear the perf counter.
This also includes a minor change to remove the u64 cast from the metag_pmu->write() call as metag_pmu->write() takes a u32 anyway, and in any case GCC is smart enough to optimise away the cast.
--- arch/metag/kernel/perf/perf_event.c | 12 ++++++++++-- 1 file changed, 10 insertions(+), 2 deletions(-)
arch/metag/kernel/perf/perf_event.c +10 -2
@@ -240,8 +240,10 @@ int metag_pmu_event_set_period(struct perf_event *event,if (left > (s64 )metag_pmu->max_period) left = metag_pmu->max_period; -if (metag_pmu->write) - metag_pmu->write (idx, (u64)(-left) &MAX_PERIOD ); +if (metag_pmu->write) { +local64_set (&hwc->prev_count, -(s32 )left); + metag_pmu->write (idx, -left &MAX_PERIOD ); + }perf_event_update_userpage (event);
@@ -651,6 +653,12 @@ static void metag_pmu_enable_counter(struct hw_perf_event *event, int idx)* set to a specific value that needs preserving. */ tmp |=metag_in32 (PERF_COUNT (idx)) &0x00ffffff ; + else +/* +* Older cores reset the counter on write, so prev_count needs +* resetting too so we can calculate a correct delta. +*/ + local64_set(&event->prev_count, 0);metag_out32 (tmp,PERF_COUNT (idx)); unlock:
Signature · 1 line
1.8.1.2
PATCH 5/9 metag: perf: fix frequency sampling (dynamic period)
In reply to James HoganFrequency sampling mode dynamically adjusts the sample period so as to hit a particular frequency of samples. The sample period starts at just 1 and then gets increased if the interrupt rate is too high. This changed sample period needs handling in metag_pmu_event_set_period to update period_left (as the ARM equivalent does). The calculated delta also needs subtracting from period_left in metag_pmu_event_update in order to hit the conditional blocks in metag_pmu_event_set_period which update last_period (which is used in the dynamic sampling period calculation).
--- arch/metag/kernel/perf/perf_event.c | 5 +++++ 1 file changed, 5 insertions(+)
arch/metag/kernel/perf/perf_event.c +5 -0
@@ -214,6 +214,7 @@ again: delta = (new_raw_count - prev_raw_count) &MAX_PERIOD ;local64_add (delta, &event->count); +local64_sub (delta, &hwc->period_left); } int metag_pmu_event_set_period(struct perf_event *event,
@@ -223,6 +224,10 @@ int metag_pmu_event_set_period(struct perf_event *event,s64 period = hwc->sample_period;int ret =0 ; +/* The period may have been changed */ +if (unlikely (period != hwc->last_period)) + left += period - hwc->last_period; +if (unlikely (left <= -period)) { left = period;local64_set (&hwc->period_left, left);
Signature · 1 line
1.8.1.2
PATCH 6/9 metag: perf: use hard_processor_id() to get thread
In reply to James HoganUse hard_processor_id() to get the current thread number rather than get_cpu() and the hardware thread mapping. There was no matching put_cpu(), and in any case this should be slightly more efficient.
--- arch/metag/kernel/perf/perf_event.c | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-)
arch/metag/kernel/perf/perf_event.c +2 -2
@@ -22,9 +22,9 @@#include <linux/slab.h> #include <asm/core_reg.h> -#include <asm/hwthread.h> #include <asm/io.h> #include <asm/irq.h> +#include <asm/processor.h> #include "perf_event.h"
@@ -651,7 +651,7 @@ static void metag_pmu_enable_counter(struct hw_perf_event *event, int idx) } tmp = ((config &0xf ) <<28 ) | - ((1 <<24 ) << cpu_2_hwthread_id[get_cpu ()]); + ((1 <<24 ) <<hard_processor_id ());if (metag_pmu->max_period)/* * Cores supporting overflow interrupts may have had the counter
Signature · 1 line
1.8.1.2
PATCH 7/9 metag: perf: don't reset TXTACTCYC
In reply to James HoganThe thread active cycle counter TXTACTCYC is used in __delay so it shouldn't really be reset to zero by perf. Fix perf to just read the value, and instead of clearing it, record the prev_count value in enable_counter so that the delta calculations know about the previous value.
--- arch/metag/kernel/perf/perf_event.c | 7 ++----- 1 file changed, 2 insertions(+), 5 deletions(-)
arch/metag/kernel/perf/perf_event.c +2 -5
@@ -617,9 +617,7 @@ static void metag_pmu_enable_counter(struct hw_perf_event *event, int idx)WARN_ONCE ((config !=0x100 ),"invalid configuration (%d) for counter (%d)\n" , config, idx); - -/* Reset the cycle count */ -__core_reg_set (TXTACTCYC ,0 ); +local64_set (&event->prev_count,__core_reg_get (TXTACTCYC )); goto unlock; }
@@ -708,9 +706,8 @@ static u64 metag_pmu_read_counter(int idx) {u32 tmp =0 ; -/* The act of reading the cycle counter also clears it */ if (METAG_INST_COUNTER == idx) { -__core_reg_swap (TXTACTCYC , tmp); + tmp =__core_reg_get (TXTACTCYC ); goto out; }
Signature · 1 line
1.8.1.2
PATCH 8/9 metag: perf: prepare for use by oprofile
In reply to James HoganTo allow our perf_events code to work with oprofile the PERF_TYPE_RAW event type attribute is implemented, which allows the internal encoding of events to be used externally (this requires some tweaks so that it handles invalid event types more gracefully), and perf_pmu_name() is adjusted to return metag_pmu->name instead of metag_pmu->pmu.name (which is changed to "meta2").
--- arch/metag/kernel/perf/perf_event.c | 22 ++++++++++++++++------ 1 file changed, 16 insertions(+), 6 deletions(-)
arch/metag/kernel/perf/perf_event.c +16 -6
@@ -40,10 +40,10 @@ static DEFINE_PER_CPU(struct cpu_hw_events, cpu_hw_events);/* PMU admin */ const char *perf_pmu_name (void ) { -if (metag_pmu) -return metag_pmu->pmu.name; +if (!metag_pmu) +return NULL ; -return NULL ; +return metag_pmu->name; }EXPORT_SYMBOL_GPL (perf_pmu_name);
@@ -171,6 +171,7 @@ static int metag_pmu_event_init(struct perf_event *event)switch (event->attr.type) {case PERF_TYPE_HARDWARE :case PERF_TYPE_HW_CACHE : +case PERF_TYPE_RAW : err =_hw_perf_event_init (event);break ;
@@ -556,6 +557,10 @@ static int _hw_perf_event_init(struct perf_event *event)if (err)return err;break ; + +case PERF_TYPE_RAW : + mapping = attr->config; +break ; }/* Return early if the event is unsupported */
@@ -623,7 +628,7 @@ static void metag_pmu_enable_counter(struct hw_perf_event *event, int idx)/* Check for a core internal or performance channel event. */ if (tmp) { -void *perf_addr = (void *)PERF_COUNT (idx); +void *perf_addr;/* * Anything other than a cycle count will write the low-
@@ -637,9 +642,14 @@ static void metag_pmu_enable_counter(struct hw_perf_event *event, int idx)case 0xf0 : perf_addr = (void *)PERF_CHAN (idx);break ; + +default : + perf_addr =NULL ; +break ; } -metag_out32 ((config &0x0f ), perf_addr); +if (perf_addr) +metag_out32 ((config &0x0f ), perf_addr);/* * Now we use the high nibble as the performance event to
@@ -848,7 +858,7 @@ static int __init init_hw_perf_events(void) metag_pmu->max_period =0 ; } - metag_pmu->name ="Meta 2" ; + metag_pmu->name ="meta2" ; metag_pmu->version = version; metag_pmu->pmu = pmu; }
Signature · 1 line
1.8.1.2
PATCH 9/9 metag: OProfile support
In reply to James HoganAdd OProfile support for metag, using the perf backend, and falling back to generic timer based sampling if perf counter interrupt support is disabled.
The oprofile code prepends "metag/" to the perf pmu name to give "metag/meta2" which is more consistent with other oprofile arch names.
The backtrace code makes use of <asm/stacktrace.h> for kernel backtracing, and a simple frame pointer walk for userland backtracing.
--- arch/metag/Kconfig | 4 +++ arch/metag/Makefile | 2 ++ arch/metag/oprofile/Makefile | 17 +++++++++++ arch/metag/oprofile/backtrace.c | 63 +++++++++++++++++++++++++++++++++++++++ arch/metag/oprofile/backtrace.h | 6 ++++ arch/metag/oprofile/common.c | 66 +++++++++++++++++++++++++++++++++++++++++ 6 files changed, 158 insertions(+) create mode 100644 arch/metag/oprofile/Makefile create mode 100644 arch/metag/oprofile/backtrace.c create mode 100644 arch/metag/oprofile/backtrace.h create mode 100644 arch/metag/oprofile/common.c
arch/metag/Kconfig +4 -0
@@ -25,6 +25,7 @@ config METAG select HAVE_MEMBLOCK select HAVE_MEMBLOCK_NODE_MAP select HAVE_MOD_ARCH_SPECIFIC + select HAVE_OPROFILE select HAVE_PERF_EVENTS select HAVE_SYSCALL_TRACEPOINTS select IRQ_DOMAIN
@@ -209,6 +210,9 @@ config METAG_PERFCOUNTER_IRQS When disabled, Performance Counters information will be collected based on Timer Interrupt. +config HW_PERF_EVENTS + def_bool METAG_PERFCOUNTER_IRQS && PERF_EVENTS + config METAG_DA bool "DA support" help
arch/metag/Makefile +2 -0
@@ -49,6 +49,8 @@ core-y += arch/metag/mm/ libs-y += arch/metag/lib/ libs-y += arch/metag/tbx/ +drivers-$(CONFIG_OPROFILE) += arch/metag/oprofile/ + boot := arch/metag/boot boot_targets += uImage
arch/metag/oprofile/Makefile new file +17 -0
@@ -0,0 +1,17 @@ +obj-$(CONFIG_OPROFILE) += oprofile.o + +oprofile-core-y += buffer_sync.o +oprofile-core-y += cpu_buffer.o +oprofile-core-y += event_buffer.o +oprofile-core-y += oprof.o +oprofile-core-y += oprofile_files.o +oprofile-core-y += oprofile_stats.o +oprofile-core-y += oprofilefs.o +oprofile-core-y += timer_int.o +oprofile-core-$(CONFIG_HW_PERF_EVENTS) += oprofile_perf.o + +oprofile-y += backtrace.o +oprofile-y += common.o +oprofile-y += $(addprefix ../../../drivers/oprofile/,$(oprofile-core-y)) + +ccflags-y += -Werror
arch/metag/oprofile/backtrace.c new file +63 -0
63 lines not on this page · raw message
arch/metag/oprofile/backtrace.h new file +6 -0
@@ -0,0 +1,6 @@ +#ifndef _METAG_OPROFILE_BACKTRACE_H +#define _METAG_OPROFILE_BACKTRACE_H + +void metag_backtrace (struct pt_regs *const regs,unsigned int depth); + +#endif
arch/metag/oprofile/common.c new file +66 -0
66 lines not on this page · raw message
Signature · 1 line
1.8.1.2