* [PATCH 0/3] alpha: Implement Hardware Performance Events
@ 2010-04-30 23:55 Michael Cree
2010-04-30 23:55 ` [PATCH 1/3] alpha: Add performance monitor interrupt counter Michael Cree
` (2 more replies)
0 siblings, 3 replies; 10+ messages in thread
From: Michael Cree @ 2010-04-30 23:55 UTC (permalink / raw)
To: linux-kernel
Cc: Michael Cree, Richard Henderson, Ivan Kokshaysky, Matt Turner,
Thomas Gleixner, Peter Zijlstra, Paul Mackerras, Ingo Molnar,
Arnaldo Carvalho de Melo, Tejun Heo, Andrew Morton, linux-alpha
The following patches implement hardware performance events for the Alpha
EV67 and later CPUs. I have had this running on a Compaq XP1000 (EV67,
single CPU) for a few days now. Pretty cool -- discovered that the glibc
exp2() library routine uses on average 985 cycles to execute 777 CPU
instructions whereas Compaq's CPML library version of exp2() uses on average
32 cycles to execute 47 CPU instructions to achieve the same thing!
Michael Cree (3):
alpha: Add performance monitor interrupt counter.
alpha: Add wrperfmon.h header file to aid use of wrperfmon PALcall.
alpha: Implement HW performance events on the EV67 and later CPUs.
arch/alpha/include/asm/hw_irq.h | 1 +
arch/alpha/include/asm/perf_event.h | 8 +-
arch/alpha/include/asm/wrperfmon.h | 93 ++++
arch/alpha/kernel/Makefile | 1 +
arch/alpha/kernel/irq.c | 7 +-
arch/alpha/kernel/irq_alpha.c | 3 +
arch/alpha/kernel/perf_event.c | 836 +++++++++++++++++++++++++++++++++++
7 files changed, 946 insertions(+), 3 deletions(-)
create mode 100644 arch/alpha/include/asm/wrperfmon.h
create mode 100644 arch/alpha/kernel/perf_event.c
^ permalink raw reply [flat|nested] 10+ messages in thread* [PATCH 1/3] alpha: Add performance monitor interrupt counter. 2010-04-30 23:55 [PATCH 0/3] alpha: Implement Hardware Performance Events Michael Cree @ 2010-04-30 23:55 ` Michael Cree 2010-04-30 23:55 ` [PATCH 2/3] alpha: Add wrperfmon.h header file to aid use of wrperfmon PALcall Michael Cree 2010-04-30 23:55 ` [PATCH 3/3] alpha: Implement HW performance events on the EV67 and later CPUs Michael Cree 2 siblings, 0 replies; 10+ messages in thread From: Michael Cree @ 2010-04-30 23:55 UTC (permalink / raw) To: linux-kernel Cc: Michael Cree, Richard Henderson, Ivan Kokshaysky, Matt Turner, Thomas Gleixner, Peter Zijlstra, Tejun Heo, linux-alpha And export the count to user space via /proc/interrupts. Signed-off-by: Michael Cree <mcree@orcon.net.nz> --- arch/alpha/include/asm/hw_irq.h | 1 + arch/alpha/kernel/irq.c | 7 +++++-- 2 files changed, 6 insertions(+), 2 deletions(-) diff --git a/arch/alpha/include/asm/hw_irq.h b/arch/alpha/include/asm/hw_irq.h index a37db0f..5050ac8 100644 --- a/arch/alpha/include/asm/hw_irq.h +++ b/arch/alpha/include/asm/hw_irq.h @@ -3,6 +3,7 @@ extern volatile unsigned long irq_err_count; +DECLARE_PER_CPU(unsigned long, irq_pmi_count); #ifdef CONFIG_ALPHA_GENERIC #define ACTUAL_NR_IRQS alpha_mv.nr_irqs diff --git a/arch/alpha/kernel/irq.c b/arch/alpha/kernel/irq.c index 7f912ba..fe91298 100644 --- a/arch/alpha/kernel/irq.c +++ b/arch/alpha/kernel/irq.c @@ -31,6 +31,7 @@ #include <asm/uaccess.h> volatile unsigned long irq_err_count; +DEFINE_PER_CPU(unsigned long, irq_pmi_count); void ack_bad_irq(unsigned int irq) { @@ -63,9 +64,7 @@ int irq_select_affinity(unsigned int irq) int show_interrupts(struct seq_file *p, void *v) { -#ifdef CONFIG_SMP int j; -#endif int irq = *(loff_t *) v; struct irqaction * action; unsigned long flags; @@ -112,6 +111,10 @@ unlock: seq_printf(p, "%10lu ", cpu_data[j].ipi_count); seq_putc(p, '\n'); #endif + seq_puts(p, "PMI: "); + for_each_online_cpu(j) + seq_printf(p, "%10lu ", per_cpu(irq_pmi_count, j)); + seq_puts(p, " Performance Monitoring\n"); seq_printf(p, "ERR: %10lu\n", irq_err_count); } return 0; -- 1.7.0 ^ permalink raw reply related [flat|nested] 10+ messages in thread
* [PATCH 2/3] alpha: Add wrperfmon.h header file to aid use of wrperfmon PALcall. 2010-04-30 23:55 [PATCH 0/3] alpha: Implement Hardware Performance Events Michael Cree 2010-04-30 23:55 ` [PATCH 1/3] alpha: Add performance monitor interrupt counter Michael Cree @ 2010-04-30 23:55 ` Michael Cree 2010-04-30 23:55 ` [PATCH 3/3] alpha: Implement HW performance events on the EV67 and later CPUs Michael Cree 2 siblings, 0 replies; 10+ messages in thread From: Michael Cree @ 2010-04-30 23:55 UTC (permalink / raw) To: linux-kernel Cc: Michael Cree, Richard Henderson, Ivan Kokshaysky, Matt Turner, linux-alpha Signed-off-by: Michael Cree <mcree@orcon.net.nz> --- arch/alpha/include/asm/wrperfmon.h | 93 ++++++++++++++++++++++++++++++++++++ 1 files changed, 93 insertions(+), 0 deletions(-) create mode 100644 arch/alpha/include/asm/wrperfmon.h diff --git a/arch/alpha/include/asm/wrperfmon.h b/arch/alpha/include/asm/wrperfmon.h new file mode 100644 index 0000000..319bf67 --- /dev/null +++ b/arch/alpha/include/asm/wrperfmon.h @@ -0,0 +1,93 @@ +/* + * Definitions for use with the Alpha wrperfmon PAL call. + */ + +#ifndef __ALPHA_WRPERFMON_H +#define __ALPHA_WRPERFMON_H + +/* Following commands are implemented on all CPUs */ +#define PERFMON_CMD_DISABLE 0 +#define PERFMON_CMD_ENABLE 1 +#define PERFMON_CMD_DESIRED_EVENTS 2 +#define PERFMON_CMD_LOGGING_OPTIONS 3 +/* Following commands on EV5/EV56/PCA56 only */ +#define PERFMON_CMD_INT_FREQ 4 +#define PERFMON_CMD_ENABLE_CLEAR 7 +/* Following commands are on EV5 and better CPUs */ +#define PERFMON_CMD_READ 5 +#define PERFMON_CMD_WRITE 6 +/* Following command are on EV6 and better CPUs */ +#define PERFMON_CMD_ENABLE_WRITE 7 +/* Following command are on EV67 and better CPUs */ +#define PERFMON_CMD_I_STAT 8 +#define PERFMON_CMD_PMPC 9 + + +/* EV5/EV56/PCA56 Counters */ +#define EV5_PCTR_0 (1UL<<0) +#define EV5_PCTR_1 (1UL<<1) +#define EV5_PCTR_2 (1UL<<2) + +#define EV5_PCTR_0_COUNT_SHIFT 48 +#define EV5_PCTR_1_COUNT_SHIFT 32 +#define EV5_PCTR_2_COUNT_SHIFT 16 + +#define EV5_PCTR_0_COUNT_MASK 0xffffUL +#define EV5_PCTR_1_COUNT_MASK 0xffffUL +#define EV5_PCTR_2_COUNT_MASK 0x3fffUL + +/* EV6 Counters */ +#define EV6_PCTR_0 (1UL<<0) +#define EV6_PCTR_1 (1UL<<1) + +#define EV6_PCTR_0_COUNT_SHIFT 28 +#define EV6_PCTR_1_COUNT_SHIFT 6 + +#define EV6_PCTR_0_COUNT_MASK 0xfffffUL +#define EV6_PCTR_1_COUNT_MASK 0xfffffUL + +/* EV67 (and subsequent) counters */ +#define EV67_PCTR_0 (1UL<<0) +#define EV67_PCTR_1 (1UL<<1) + +#define EV67_PCTR_0_COUNT_SHIFT 28 +#define EV67_PCTR_1_COUNT_SHIFT 6 + +#define EV67_PCTR_0_COUNT_MASK 0xfffffUL +#define EV67_PCTR_1_COUNT_MASK 0xfffffUL + + +/* + * The Alpha Architecure Handbook, vers. 4 (1998) appears to have a misprint + * in Table E-23 regarding the bits that set the event PCTR 1 counts. + * Hopefully what we have here is correct. + */ +#define EV6_PCTR_0_EVENT_MASK 0x10UL +#define EV6_PCTR_1_EVENT_MASK 0x0fUL + +/* EV6 Events */ +#define EV6_PCTR_0_CYCLES (0UL << 4) +#define EV6_PCTR_0_INSTRUCTIONS (1UL << 4) + +#define EV6_PCTR_1_CYCLES 0 +#define EV6_PCTR_1_BRANCHES 1 +#define EV6_PCTR_1_BRANCH_MISPREDICTS 2 +#define EV6_PCTR_1_DTB_SINGLE_MISSES 3 +#define EV6_PCTR_1_DTB_DOUBLE_MISSES 4 +#define EV6_PCTR_1_ITB_MISSES 5 +#define EV6_PCTR_1_UNALIGNED_TRAPS 6 +#define EV6_PCTR_1_REPLY_TRAPS 7 + +/* From the Alpha Architecture Reference Manual, 4th edn., 2002 */ +#define EV67_PCTR_MODE_MASK 0x10UL +#define EV67_PCTR_EVENT_MASK 0x0CUL + +#define EV67_PCTR_MODE_PROFILEME (1UL<<4) +#define EV67_PCTR_MODE_AGGREGATE (0UL<<4) + +#define EV67_PCTR_INSTR_CYCLES (0UL<<2) +#define EV67_PCTR_CYCLES_UNDEF (1UL<<2) +#define EV67_PCTR_INSTR_BCACHEMISS (2UL<<2) +#define EV67_PCTR_CYCLES_MBOX (3UL<<2) + +#endif -- 1.7.0 ^ permalink raw reply related [flat|nested] 10+ messages in thread
* [PATCH 3/3] alpha: Implement HW performance events on the EV67 and later CPUs. 2010-04-30 23:55 [PATCH 0/3] alpha: Implement Hardware Performance Events Michael Cree 2010-04-30 23:55 ` [PATCH 1/3] alpha: Add performance monitor interrupt counter Michael Cree 2010-04-30 23:55 ` [PATCH 2/3] alpha: Add wrperfmon.h header file to aid use of wrperfmon PALcall Michael Cree @ 2010-04-30 23:55 ` Michael Cree 2010-05-04 7:40 ` Peter Zijlstra 2 siblings, 1 reply; 10+ messages in thread From: Michael Cree @ 2010-04-30 23:55 UTC (permalink / raw) To: linux-kernel Cc: Michael Cree, Peter Zijlstra, Paul Mackerras, Ingo Molnar, Arnaldo Carvalho de Melo, Richard Henderson, Ivan Kokshaysky, Matt Turner, Andrew Morton, Thomas Gleixner, linux-alpha This implements hardware performance events for the EV67 and later CPUs within the Linux performance events subsystem. Only using the performance monitoring unit in HP/Compaq's so called "Aggregrate mode" is supported. The code has been implemented in a manner that makes extension to other older Alpha CPUs relatively straightforward should some mug wish to indulge themselves. Signed-off-by: Michael Cree <mcree@orcon.net.nz> --- arch/alpha/include/asm/perf_event.h | 8 +- arch/alpha/kernel/Makefile | 1 + arch/alpha/kernel/irq_alpha.c | 3 + arch/alpha/kernel/perf_event.c | 836 +++++++++++++++++++++++++++++++++++ 4 files changed, 847 insertions(+), 1 deletions(-) create mode 100644 arch/alpha/kernel/perf_event.c diff --git a/arch/alpha/include/asm/perf_event.h b/arch/alpha/include/asm/perf_event.h index 3bef852..4157cd3 100644 --- a/arch/alpha/include/asm/perf_event.h +++ b/arch/alpha/include/asm/perf_event.h @@ -2,8 +2,14 @@ #define __ASM_ALPHA_PERF_EVENT_H /* Alpha only supports software events through this interface. */ -static inline void set_perf_event_pending(void) { } +extern void set_perf_event_pending(void); #define PERF_EVENT_INDEX_OFFSET 0 +#ifdef CONFIG_PERF_EVENTS +extern void init_hw_perf_events(void); +#else +static inline void init_hw_perf_events(void) { } +#endif + #endif /* __ASM_ALPHA_PERF_EVENT_H */ diff --git a/arch/alpha/kernel/Makefile b/arch/alpha/kernel/Makefile index 7739a62..e3dc108 100644 --- a/arch/alpha/kernel/Makefile +++ b/arch/alpha/kernel/Makefile @@ -15,6 +15,7 @@ obj-$(CONFIG_SMP) += smp.o obj-$(CONFIG_PCI) += pci.o pci_iommu.o pci-sysfs.o obj-$(CONFIG_SRM_ENV) += srm_env.o obj-$(CONFIG_MODULES) += module.o +obj-$(CONFIG_PERF_EVENTS) += perf_event.o ifdef CONFIG_ALPHA_GENERIC diff --git a/arch/alpha/kernel/irq_alpha.c b/arch/alpha/kernel/irq_alpha.c index cfde865..5f77afb 100644 --- a/arch/alpha/kernel/irq_alpha.c +++ b/arch/alpha/kernel/irq_alpha.c @@ -10,6 +10,7 @@ #include <asm/machvec.h> #include <asm/dma.h> +#include <asm/perf_event.h> #include "proto.h" #include "irq_impl.h" @@ -111,6 +112,8 @@ init_IRQ(void) wrent(entInt, 0); alpha_mv.init_irq(); + + init_hw_perf_events(); } /* diff --git a/arch/alpha/kernel/perf_event.c b/arch/alpha/kernel/perf_event.c new file mode 100644 index 0000000..6bda3e2 --- /dev/null +++ b/arch/alpha/kernel/perf_event.c @@ -0,0 +1,836 @@ +/* + * Hardware performance events for the Alpha. + * + * We implement HW counts on the EV67 and subsequent CPUs only. + * + * (C) 2010 Michael J. Cree + * + * Somewhat based on the Sparc code, and to a lesser extent the PowerPC and + * ARM code, which are copyright by their respective authors. + */ + +#include <linux/perf_event.h> +#include <linux/kprobes.h> +#include <linux/kernel.h> +#include <linux/kdebug.h> +#include <linux/mutex.h> + +#include <asm/hwrpb.h> +#include <asm/atomic.h> +#include <asm/irq.h> +#include <asm/irq_regs.h> +#include <asm/pal.h> +#include <asm/wrperfmon.h> +#include <asm/hw_irq.h> + + +/* The maximum number of PMCs on any Alpha CPU whatsoever. */ +#define MAX_HWEVENTS 3 +#define PMC_NO_INDEX -1 + +/* For tracking PMCs and the hw events they monitor on each CPU. */ +struct cpu_hw_events { + int enabled; + /* Number of events scheduled; also number entries valid in arrays below. */ + int n_events; + /* Number events added since last hw_perf_disable(). */ + int n_added; + /* Events currently scheduled. */ + struct perf_event *event[MAX_HWEVENTS]; + /* Event type of each scheduled event. */ + unsigned long evtype[MAX_HWEVENTS]; + /* Current index of each scheduled event; if not yet determined + * contains PMC_NO_INDEX. + */ + int current_idx[MAX_HWEVENTS]; + /* The active PMCs' config for easy use with wrperfmon(). */ + unsigned long config; + /* The active counters' indices for easy use with wrperfmon(). */ + unsigned long idx_mask; +}; +DEFINE_PER_CPU(struct cpu_hw_events, cpu_hw_events); + + + +/* + * A structure to hold the description of the PMCs available on a particular + * type of Alpha CPU. + */ +struct alpha_pmu_t { + /* Mapping of the perf system hw event types to indigenous event types */ + const int *event_map; + /* The number of entries in the event_map */ + int max_events; + /* The number of PMCs on this Alpha */ + int num_pmcs; + /* + * All PMC counters reside in the IBOX register PCTR. This is the + * LSB of the counter. + */ + int pmc_count_shift[MAX_HWEVENTS]; + /* + * The mask that isolates the PMC bits when the LSB of the counter + * is shifted to bit 0. + */ + unsigned long pmc_count_mask[MAX_HWEVENTS]; + /* The maximum period the PMC can count. */ + unsigned long pmc_max_period[MAX_HWEVENTS]; + /* + * The maximum value that may be written to the counter due to + * hardware restrictions is pmc_max_period - pmc_left. + */ + long pmc_left[3]; + /* Subroutine for allocation of PMCs. Enforces constraints. */ + int (*check_constraints)(struct perf_event **, unsigned long *, int); +}; + +/* + * The Alpha CPU PMU description currently in operation. This is set during + * the boot process to the specific CPU of the machine. + */ +static const struct alpha_pmu_t *alpha_pmu; + + +#define HW_OP_UNSUPPORTED -1 + +/* + * The hardware description of the EV67, EV68, EV69, EV7 and EV79 PMUs + * follow. Since they are identical we refer to them collectively as the + * EV67 henceforth. + */ + +/* + * EV67 PMC event types + * + * There is no one-to-one mapping of the possible hw event types to the + * actual codes that are used to program the PMCs hence we introduce our + * own hw event type identifiers. + */ +enum ev67_pmc_event_type { + EV67_CYCLES = 1, + EV67_INSTRUCTIONS, + EV67_BCACHEMISS, + EV67_MBOXREPLAY, + EV67_LAST_ET +}; +#define EV67_NUM_EVENT_TYPES (EV67_LAST_ET-EV67_CYCLES) + + +/* Mapping of the hw event types to the perf tool interface */ +static const int ev67_perfmon_event_map[] = { + [PERF_COUNT_HW_CPU_CYCLES] = EV67_CYCLES, + [PERF_COUNT_HW_INSTRUCTIONS] = EV67_INSTRUCTIONS, + [PERF_COUNT_HW_CACHE_REFERENCES] = HW_OP_UNSUPPORTED, + [PERF_COUNT_HW_CACHE_MISSES] = EV67_BCACHEMISS, +}; + +struct ev67_mapping_t { + int config; + int idx; +}; + +/* + * The mapping used for one event only - these must be in same order as enum + * ev67_pmc_event_type definition. + */ +static const struct ev67_mapping_t ev67_mapping[] = { + {EV67_PCTR_INSTR_CYCLES, 1}, /* EV67_CYCLES, */ + {EV67_PCTR_INSTR_CYCLES, 0}, /* EV67_INSTRUCTIONS */ + {EV67_PCTR_INSTR_BCACHEMISS, 1}, /* EV67_BCACHEMISS */ + {EV67_PCTR_CYCLES_MBOX, 1} /* EV67_MBOXREPLAY */ +}; + + +/* + * Check that a group of events can be simultaneously scheduled on to the + * EV67 PMU. Also allocate counter indices and config. + */ +static int ev67_check_constraints(struct perf_event **event, + unsigned long *evtype, int n_ev) +{ + int idx0; + unsigned long config; + + idx0 = ev67_mapping[evtype[0]-1].idx; + config = ev67_mapping[evtype[0]-1].config; + if (n_ev == 1) + goto success; + + BUG_ON(n_ev != 2); + + if (evtype[0] == EV67_MBOXREPLAY || evtype[1] == EV67_MBOXREPLAY) { + /* MBOX replay traps must be on PMC 1 */ + idx0 = (evtype[0] == EV67_MBOXREPLAY) ? 1 : 0; + /* Only cycles can accompany MBOX replay traps */ + if (evtype[idx0] == EV67_CYCLES) { + config = EV67_PCTR_CYCLES_MBOX; + goto success; + } + } + + if (evtype[0] == EV67_BCACHEMISS || evtype[1] == EV67_BCACHEMISS) { + /* Bcache misses must be on PMC 1 */ + idx0 = (evtype[0] == EV67_BCACHEMISS) ? 1 : 0; + /* Only instructions can accompany Bcache misses */ + if (evtype[idx0] == EV67_INSTRUCTIONS) { + config = EV67_PCTR_INSTR_BCACHEMISS; + goto success; + } + } + + if (evtype[0] == EV67_INSTRUCTIONS || evtype[1] == EV67_INSTRUCTIONS) { + /* Instructions must be on PMC 0 */ + idx0 = (evtype[0] == EV67_INSTRUCTIONS) ? 0 : 1; + /* By this point only cycles can accompany instructions */ + if (evtype[idx0^1] == EV67_CYCLES) { + config = EV67_PCTR_INSTR_CYCLES; + goto success; + } + } + + /* Otherwise, darn it, there is a conflict. */ + return -1; + +success: + event[0]->hw.idx = idx0; + event[0]->hw.config_base = config; + if (n_ev == 2) { + event[1]->hw.idx = idx0 ^ 1; + event[1]->hw.config_base = config; + } + return 0; +} + + +static const struct alpha_pmu_t ev67_pmu = { + .event_map = ev67_perfmon_event_map, + .max_events = ARRAY_SIZE(ev67_perfmon_event_map), + .num_pmcs = 2, + .pmc_count_shift = {EV67_PCTR_0_COUNT_SHIFT, EV67_PCTR_1_COUNT_SHIFT, 0}, + .pmc_count_mask = {EV67_PCTR_0_COUNT_MASK, EV67_PCTR_1_COUNT_MASK, 0}, + .pmc_max_period = {(1UL<<20) - 1, (1UL<<20) - 1, 0}, + .pmc_left = {16, 4, 0}, + .check_constraints = ev67_check_constraints +}; + + + +/* + * Helper routines to ensure that we read/write only the correct PMC bits + * when calling the wrperfmon PALcall. + */ +static inline void alpha_write_pmc(int idx, unsigned long val) +{ + val &= alpha_pmu->pmc_count_mask[idx]; + val <<= alpha_pmu->pmc_count_shift[idx]; + val |= (1<<idx); + wrperfmon(PERFMON_CMD_WRITE, val); +} + +static inline unsigned long alpha_read_pmc(int idx) +{ + unsigned long val; + + val = wrperfmon(PERFMON_CMD_READ, 0); + val >>= alpha_pmu->pmc_count_shift[idx]; + val &= alpha_pmu->pmc_count_mask[idx]; + return val; +} + +/* Set a new period to sample over */ +static int alpha_perf_event_set_period(struct perf_event *event, + struct hw_perf_event *hwc, int idx) +{ + long left = atomic64_read(&hwc->period_left); + long period = hwc->sample_period; + int ret = 0; + + if (unlikely(left <= -period)) { + left = period; + atomic64_set(&hwc->period_left, left); + hwc->last_period = period; + ret = 1; + } + + if (unlikely(left <= 0)) { + left += period; + atomic64_set(&hwc->period_left, left); + hwc->last_period = period; + ret = 1; + } + + /* + * Hardware restrictions require that the counters must not be + * written with values that are too close to the maximum period. + */ + if (unlikely(left < alpha_pmu->pmc_left[idx])) + left = alpha_pmu->pmc_left[idx]; + + if (left > (long)alpha_pmu->pmc_max_period[idx]) + left = alpha_pmu->pmc_max_period[idx]; + + atomic64_set(&hwc->prev_count, (unsigned long)(-left)); + + alpha_write_pmc(idx, (unsigned long)(-left)); + + perf_event_update_userpage(event); + + return ret; +} + + +/* + * Calculates the count (the 'delta') since the last time the PMC was read. + * + * As the PMCs' full period can easily be exceeded within the perf system + * sampling period we cannot use any high order bits as a guard bit in the + * PMCs to detect overflow as is done by other architectures. The code here + * calculates the delta on the basis that there is no overflow when ovf is + * zero. The value passed via ovf by the interrupt handler corrects for + * overflow. + * + * This can be racey on rare occasions -- a call to this routine can occur + * with an overflowed counter just before the PMI service routine is called. + * The check for delta negative hopefully always rectifies this situation. + */ +static unsigned long alpha_perf_event_update(struct perf_event *event, + struct hw_perf_event *hwc, int idx, long ovf) +{ + long prev_raw_count, new_raw_count; + long delta; + +again: + prev_raw_count = atomic64_read(&hwc->prev_count); + new_raw_count = alpha_read_pmc(idx); + + if (atomic64_cmpxchg(&hwc->prev_count, prev_raw_count, + new_raw_count) != prev_raw_count) + goto again; + + delta = (new_raw_count - (prev_raw_count & alpha_pmu->pmc_count_mask[idx])) + ovf; + + /* It is possible on very rare occasions that the PMC has overflowed + * but the interrupt is yet to come. Detect and fix this situation. + */ + if (unlikely(delta < 0)) { + delta += alpha_pmu->pmc_max_period[idx] + 1; + } + + atomic64_add(delta, &event->count); + atomic64_sub(delta, &hwc->period_left); + + return new_raw_count; +} + + +/* + * Collect all HW events into the array event[]. + */ +static int collect_events(struct perf_event *group, int max_count, + struct perf_event *event[], unsigned long *evtype, + int *current_idx) +{ + struct perf_event *pe; + int n = 0; + + if (!is_software_event(group)) { + if (n >= max_count) + return -1; + event[n] = group; + evtype[n] = group->hw.event_base; + current_idx[n++] = PMC_NO_INDEX; + } + list_for_each_entry(pe, &group->sibling_list, group_entry) { + if (!is_software_event(pe) && pe->state != PERF_EVENT_STATE_OFF) { + if (n >= max_count) + return -1; + event[n] = pe; + evtype[n] = pe->hw.event_base; + current_idx[n++] = PMC_NO_INDEX; + } + } + return n; +} + + + +/* + * Check that a group of events can be simultaneously scheduled on to the PMU. + */ +static int alpha_check_constraints(struct perf_event **events, + unsigned long *evtypes, int n_ev) +{ + + /* No HW events is possible from hw_perf_group_sched_in(). */ + if (n_ev == 0) + return 0; + + if (n_ev > alpha_pmu->num_pmcs) + return -1; + + return alpha_pmu->check_constraints(events, evtypes, n_ev); +} + + +/* + * If new events have been scheduled then update cpuc with the new + * configuration. This may involve shifting cycle counts from one PMC to + * another. + */ +static void maybe_change_configuration(struct cpu_hw_events *cpuc) +{ + int j; + + if (cpuc->n_added == 0) + return; + + /* Find counters that are moving to another PMC and update */ + for (j = 0; j < cpuc->n_events; j++) { + struct perf_event *pe = cpuc->event[j]; + + if (cpuc->current_idx[j] != PMC_NO_INDEX && + cpuc->current_idx[j] != pe->hw.idx) { + alpha_perf_event_update(pe, &pe->hw, cpuc->current_idx[j], 0); + cpuc->current_idx[j] = PMC_NO_INDEX; + } + } + + /* Assign to counters all unassigned events. */ + cpuc->idx_mask = 0; + for (j = 0; j < cpuc->n_events; j++) { + struct perf_event *pe = cpuc->event[j]; + struct hw_perf_event *hwc = &pe->hw; + int idx = hwc->idx; + + if (cpuc->current_idx[j] != PMC_NO_INDEX) { + cpuc->idx_mask |= (1<<cpuc->current_idx[j]); + continue; + } + + alpha_perf_event_set_period(pe, hwc, idx); + cpuc->current_idx[j] = idx; + cpuc->idx_mask |= (1<<cpuc->current_idx[j]); + } + cpuc->config = cpuc->event[0]->hw.config_base; +} + + + +/* Schedule perf HW event on to PMU. + * - this function is called from outside this module via the pmu struct + * returned from perf event initialisation. + */ +static int alpha_pmu_enable(struct perf_event *event) +{ + struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events); + int n0; + int ret; + unsigned long flags; + + /* + * The Sparc code has the IRQ disable first followed by the perf + * disable, however this can lead to an overflowed counter with the + * PMI disabled on rare occasions. The alpha_perf_event_update() + * routine should detect this situation by noting a negative delta, + * nevertheless we disable the PMCs first to enable a potential + * final PMI to occur before we disable interrupts. + */ + perf_disable(); + local_irq_save(flags); + + /* Default to error to be returned */ + ret = -EAGAIN; + + /* Insert event on to PMU and if successful modify ret to valid return */ + n0 = cpuc->n_events; + if (n0 < alpha_pmu->num_pmcs) { + cpuc->event[n0] = event; + cpuc->evtype[n0] = event->hw.event_base; + cpuc->current_idx[n0] = PMC_NO_INDEX; + + if (!alpha_check_constraints(cpuc->event, cpuc->evtype, n0+1)) { + cpuc->n_events++; + cpuc->n_added++; + ret = 0; + } + } + + local_irq_restore(flags); + perf_enable(); + + return ret; +} + + + +/* Disable performance monitoring unit + * - this function is called from outside this module via the pmu struct + * returned from perf event initialisation. + */ +static void alpha_pmu_disable(struct perf_event *event) +{ + struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events); + struct hw_perf_event *hwc = &event->hw; + unsigned long flags; + int j; + + perf_disable(); + local_irq_save(flags); + + for (j = 0; j < cpuc->n_events; j++) { + if (event == cpuc->event[j]) { + int idx = cpuc->current_idx[j]; + + /* Shift remaining entries down into the existing + * slot. + */ + while (++j < cpuc->n_events) { + cpuc->event[j - 1] = cpuc->event[j]; + cpuc->evtype[j - 1] = cpuc->evtype[j]; + cpuc->current_idx[j - 1] = + cpuc->current_idx[j]; + } + + /* Absorb the final count and turn off the event. */ + alpha_perf_event_update(event, hwc, idx, 0); + perf_event_update_userpage(event); + + cpuc->idx_mask &= ~(1UL<<idx); + cpuc->n_events--; + break; + } + } + + local_irq_restore(flags); + perf_enable(); +} + + +static void alpha_pmu_read(struct perf_event *event) +{ + struct hw_perf_event *hwc = &event->hw; + + alpha_perf_event_update(event, hwc, hwc->idx, 0); +} + + +static void alpha_pmu_unthrottle(struct perf_event *event) +{ + struct hw_perf_event *hwc = &event->hw; + + wrperfmon(PERFMON_CMD_ENABLE, (1<<hwc->idx)); +} + + +/* + * Check that CPU performance counters are supported. + * - currently support EV67 and later CPUs. + * - actually some later revisions of the EV6 have the same PMC model as the + * EV67 but we don't do suffiently deep CPU detection to detect them. + * Bad luck to the very few people who might have one, I guess. + */ +static int supported_cpu(void) +{ + struct percpu_struct *cpu; + unsigned long cputype; + + /* Get cpu type from HW */ + cpu = (struct percpu_struct *)((char *)hwrpb + hwrpb->processor_offset); + cputype = cpu->type & 0xffffffff; + /* Include all of EV67, EV68, EV7, EV79 and EV69 as supported. */ + return (cputype >= EV67_CPU) && (cputype <= EV69_CPU); +} + + + +static void hw_perf_event_destroy(struct perf_event *event) +{ + /* Nothing to be done! */ + return; +} + + + +static int __hw_perf_event_init(struct perf_event *event) +{ + struct perf_event_attr *attr = &event->attr; + struct hw_perf_event *hwc = &event->hw; + struct perf_event *evts[MAX_HWEVENTS]; + unsigned long evtypes[MAX_HWEVENTS]; + int idx_rubbish_bin[MAX_HWEVENTS]; + int ev; + int n; + + /* We only support a limited range of HARDWARE event types with one + * only programmable via a RAW event type. + */ + if (attr->type == PERF_TYPE_HARDWARE) { + if (attr->config >= alpha_pmu->max_events) + return -EINVAL; + ev = alpha_pmu->event_map[attr->config]; + } else if (attr->type == PERF_TYPE_HW_CACHE) { + return -EOPNOTSUPP; + } else if (attr->type == PERF_TYPE_RAW) { + ev = attr->config & 0xff; + } else { + return -EOPNOTSUPP; + } + + if (ev < 0) { + return ev; + } + + /* The EV67 does not support mode exclusion */ + if (attr->exclude_kernel || attr->exclude_user + || attr->exclude_hv || attr->exclude_idle) { + return -EPERM; + } + + /* + * We place the event type in event_base here and leave calculation + * of the codes to programme the PMU for alpha_pmu_enable() because + * it is only then we will know what HW events are actually + * scheduled on to the PMU. At that point the code to programme the + * PMU is put into config_base and the PMC to use is placed into + * idx. We initialise idx (below) to PMC_NO_INDEX to indicate that + * it is yet to be determined. + */ + hwc->event_base = ev; + + /* Collect events in a group together suitable for calling + * alpha_check_constraints() to verify that the group as a whole can + * be scheduled on to the PMU. + */ + n = 0; + if (event->group_leader != event) { + n = collect_events(event->group_leader, + alpha_pmu->num_pmcs - 1, + evts, evtypes, idx_rubbish_bin); + if (n < 0) + return -EINVAL; + } + evtypes[n] = hwc->event_base; + evts[n] = event; + + if (alpha_check_constraints(evts, evtypes, n + 1)) + return -EINVAL; + + /* Indicate that PMU config and idx are yet to be determined. */ + hwc->config_base = 0; + hwc->idx = PMC_NO_INDEX; + + event->destroy = hw_perf_event_destroy; + + /* + * Most architectures reserve the PMU for their use at this point. + * As there is no existing mechanism to arbitrate usage and there + * appears to be no other user of the Alpha PMU we just assume + * that we can just use it, hence a NO-OP here. + * + * Maybe an alpha_reserve_pmu() routine should be implemented but is + * anything else ever going to use it? + */ + + if (!hwc->sample_period) { + hwc->sample_period = alpha_pmu->pmc_max_period[0]; + hwc->last_period = hwc->sample_period; + atomic64_set(&hwc->period_left, hwc->sample_period); + } + + return 0; +} + +static const struct pmu pmu = { + .enable = alpha_pmu_enable, + .disable = alpha_pmu_disable, + .read = alpha_pmu_read, + .unthrottle = alpha_pmu_unthrottle, +}; + + +/* + * Main entry point to initialise a HW performance event. + */ +const struct pmu *hw_perf_event_init(struct perf_event *event) +{ + int err; + + if (!alpha_pmu) + return ERR_PTR(-ENODEV); + + /* Do the real initialisation work. */ + err = __hw_perf_event_init(event); + + if (err) + return ERR_PTR(err); + + return &pmu; +} + + + +/* + * Main entry point - enable HW performance counters. + */ +void hw_perf_enable(void) +{ + struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events); + + if (cpuc->enabled) + return; + + cpuc->enabled = 1; + barrier(); + + if (cpuc->n_events > 0) { + /* Update cpuc with information from any new scheduled events. */ + maybe_change_configuration(cpuc); + + /* Start counting the desired events. */ + wrperfmon(PERFMON_CMD_LOGGING_OPTIONS, EV67_PCTR_MODE_AGGREGATE); + wrperfmon(PERFMON_CMD_DESIRED_EVENTS, cpuc->config); + wrperfmon(PERFMON_CMD_ENABLE, cpuc->idx_mask); + } +} + + +/* + * Main entry point - disable HW performance counters. + */ + +void hw_perf_disable(void) +{ + struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events); + + if (!cpuc->enabled) + return; + + cpuc->enabled = 0; + cpuc->n_added = 0; + + wrperfmon(PERFMON_CMD_DISABLE, cpuc->idx_mask); +} + + +/* + * Main entry point - don't know when this is called but it + * obviously dumps debug info. + */ +void perf_event_print_debug(void) +{ + unsigned long flags; + unsigned long pcr; + int pcr0, pcr1; + int cpu; + + if (!supported_cpu()) + return; + + local_irq_save(flags); + + cpu = smp_processor_id(); + + pcr = wrperfmon(PERFMON_CMD_READ, 0); + pcr0 = (pcr >> alpha_pmu->pmc_count_shift[0]) & alpha_pmu->pmc_count_mask[0]; + pcr1 = (pcr >> alpha_pmu->pmc_count_shift[1]) & alpha_pmu->pmc_count_mask[1]; + + pr_info("CPU#%d: PCTR0[%06x] PCTR1[%06x]\n", cpu, pcr0, pcr1); + + local_irq_restore(flags); +} + + +/* + * Performance Monitoring Interrupt Service Routine called when a PMC + * overflows. The PMC that overflowed is passed in la_ptr. + */ +static void alpha_perf_event_irq_handler(unsigned long la_ptr, + struct pt_regs *regs) +{ + struct cpu_hw_events *cpuc; + struct perf_sample_data data; + struct perf_event *event; + struct hw_perf_event *hwc; + int idx, j; + + __get_cpu_var(irq_pmi_count)++; + cpuc = &__get_cpu_var(cpu_hw_events); + + /* la_ptr is the counter that overflowed. */ + + if (la_ptr >= perf_max_events) { + /* This should never occur! */ + irq_err_count++; + pr_warning("PMI: silly index %ld\n", la_ptr); + return; + } + + idx = la_ptr; + + perf_sample_data_init(&data, 0); + for (j = 0; j < cpuc->n_events; j++) { + if (cpuc->current_idx[j] == idx) + break; + } + + if (j == cpuc->n_events) { + /* This can occur if the event is disabled right on a PMC overflow. */ + return; + } + + event = cpuc->event[j]; + + if (!event) { + /* This should never occur! */ + irq_err_count++; + pr_warning("PMI: No event at index %d!\n", idx); + return; + } + + hwc = &event->hw; + alpha_perf_event_update(event, hwc, idx, alpha_pmu->pmc_max_period[idx]+1); + data.period = event->hw.last_period; + + if (alpha_perf_event_set_period(event, hwc, idx)) { + if (perf_event_overflow(event, 0, &data, regs)) { + /* Interrupts coming too quickly; "throttle" the + * counter, i.e., disable it for a little while. + */ + wrperfmon(PERFMON_CMD_DISABLE, 1<<idx); + } + } + + return; +} + + +/* I wonder what this is for ??? */ +void set_perf_event_pending(void) +{ +} + + + +/* + * Init call to initialise performance events at kernel startup. + */ +void __init init_hw_perf_events(void) +{ + pr_info("Performance events: "); + + if (!supported_cpu()) { + pr_cont("No support for your CPU.\n"); + return; + } + + pr_cont("Supported CPU type!\n"); + + /* Override performance counter IRQ vector */ + + perf_irq = alpha_perf_event_irq_handler; + + /* And set up PMU specification */ + alpha_pmu = &ev67_pmu; + perf_max_events = alpha_pmu->num_pmcs; +} + -- 1.7.0 ^ permalink raw reply related [flat|nested] 10+ messages in thread
* Re: [PATCH 3/3] alpha: Implement HW performance events on the EV67 and later CPUs. 2010-04-30 23:55 ` [PATCH 3/3] alpha: Implement HW performance events on the EV67 and later CPUs Michael Cree @ 2010-05-04 7:40 ` Peter Zijlstra 2010-05-24 21:58 ` Michael Cree ` (2 more replies) 0 siblings, 3 replies; 10+ messages in thread From: Peter Zijlstra @ 2010-05-04 7:40 UTC (permalink / raw) To: Michael Cree Cc: linux-kernel, Paul Mackerras, Ingo Molnar, Arnaldo Carvalho de Melo, Richard Henderson, Ivan Kokshaysky, Matt Turner, Andrew Morton, Thomas Gleixner, linux-alpha On Sat, 2010-05-01 at 11:55 +1200, Michael Cree wrote: > +/* I wonder what this is for ??? */ > +void set_perf_event_pending(void) > +{ > +} If the PMU Interrupt comes in as an NMI we cannot take locks and such from its handler. So what we do is queue that work to be ran later. The kernel has a fallback to run this stuff from the timer hardirq, but if the PMI is IRQ context (like ARM) you can simply call perf_event_do_pending() at its tail. Alternatively you can self-IPI like x86 does and run perf_event_do_pending() from there. ^ permalink raw reply [flat|nested] 10+ messages in thread
* Re: [PATCH 3/3] alpha: Implement HW performance events on the EV67 and later CPUs. 2010-05-04 7:40 ` Peter Zijlstra @ 2010-05-24 21:58 ` Michael Cree 2010-07-27 10:29 ` Michael Cree 2010-07-27 10:29 ` [PATCH 3/3 V2] " Michael Cree 2 siblings, 0 replies; 10+ messages in thread From: Michael Cree @ 2010-05-24 21:58 UTC (permalink / raw) To: Peter Zijlstra Cc: linux-kernel, Paul Mackerras, Ingo Molnar, Arnaldo Carvalho de Melo, Richard Henderson, Ivan Kokshaysky, Matt Turner, Andrew Morton, Thomas Gleixner, linux-alpha On 4/05/2010, at 7:40 PM, Peter Zijlstra wrote: > On Sat, 2010-05-01 at 11:55 +1200, Michael Cree wrote: >> +/* I wonder what this is for ??? */ >> +void set_perf_event_pending(void) >> +{ >> +} > > If the PMU Interrupt comes in as an NMI we cannot take locks and such > from its handler. So what we do is queue that work to be ran later. > The > kernel has a fallback to run this stuff from the timer hardirq, but if > the PMI is IRQ context (like ARM) you can simply call > perf_event_do_pending() at its tail. My understanding from the Alpha HW Ref. Man. and a brief look at the kernel code is that the PMI comes in through the normal IRQ entry point on the Alpha. It is a high priority interrupt and only a machine check (i.e. detected hardware failure) can interrupt the PMU interrupt handler. I don't know whether that is an issue for taking locks or not and need advice on that. I originally based the Alpha implementation on the Sparc implementation and would've copied the set_perf_event_pending() definition from the Sparc code, but I now see that newer Sparc code does not have that definition nor does it call perf_event_do_pending() in the interrupt handler. So I need advice on whether to: 1) Just remove the set_perf_event_pending() definition so we are like the Sparc code, 2) Add perf_event_do_pending() to the tail of the IRQ handler so we are like the ARM code, or 3) Implement the following suggestion: > Alternatively you can self-IPI like x86 does and run > perf_event_do_pending() from there. Another question is what are the observeable symptoms if an incorrect option above is implemented? Regards Michael. ^ permalink raw reply [flat|nested] 10+ messages in thread
* [PATCH 3/3] alpha: Implement HW performance events on the EV67 and later CPUs. 2010-05-04 7:40 ` Peter Zijlstra 2010-05-24 21:58 ` Michael Cree @ 2010-07-27 10:29 ` Michael Cree 2010-07-27 10:29 ` [PATCH 3/3 V2] " Michael Cree 2 siblings, 0 replies; 10+ messages in thread From: Michael Cree @ 2010-07-27 10:29 UTC (permalink / raw) To: Peter Zijlstra Cc: Michael Cree, linux-kernel, Paul Mackerras, Ingo Molnar, Arnaldo Carvalho de Melo, Richard Henderson, Ivan Kokshaysky, Matt Turner, Andrew Morton, Thomas Gleixner, linux-alpha On 04/05/10 19:40, Peter Zijlstra wrote: > On Sat, 2010-05-01 at 11:55 +1200, Michael Cree wrote: >> +/* I wonder what this is for ??? */ >> +void set_perf_event_pending(void) >> +{ >> +} > > If the PMU Interrupt comes in as an NMI we cannot take locks and such > from its handler. So what we do is queue that work to be ran later. The > kernel has a fallback to run this stuff from the timer hardirq, but if > the PMI is IRQ context (like ARM) you can simply call > perf_event_do_pending() at its tail. > > Alternatively you can self-IPI like x86 does and run > perf_event_do_pending() from there. OK, I couldn't see how to self-IPI on the Apha (there is no obvious way in the Alpha arch code, and the Alpha HW Ref. Man. only describes such a feature being supported by the PALcode for OpenVMS which we are obviously not running) and the PMI on Alpha is at such a high priority there is no chance that the lock can be freed while in the PMI handler, so I have put the call to perf_event_do_pending() in the timer interrupt, but unlike the PowerPC code I don't know any tricks to force the timer interrupt to happen early when there is pending work, so there is a potential maximum delay of 1ms from pending work being notified to calling perf_event_do_pending(). Is this delay acceptable? I also fix a nasty, but heretofore hidden, bug that could completely lock up a machine with continuous PMIs due to a throttled PMC being accidently re-enabled by the other PMC. It only showed up when I set the NMI flag true in the call to perf_event_overflow() and had two events counting simultaneously with a very short sample period. I also see the x86 and Sparc code in 2.6.35 have reimplemented the call to hw_perf_group_sched_in() in terms of new functions start_txn(), cancel_txn() and commit_txn(). Is this change necessary to get the Alpha code accepted for 2.6.36? If so it would be helpful if I could have a brief description of what each function does. The updated patch follows this message. Cheers Michael. Michael Cree (1): alpha: Implement HW performance events on the EV67 and later CPUs. arch/alpha/include/asm/perf_event.h | 8 +- arch/alpha/kernel/Makefile | 1 + arch/alpha/kernel/irq_alpha.c | 3 + arch/alpha/kernel/perf_event.c | 842 +++++++++++++++++++++++++++++++++++ arch/alpha/kernel/time.c | 26 ++ 5 files changed, 879 insertions(+), 1 deletions(-) create mode 100644 arch/alpha/kernel/perf_event.c ^ permalink raw reply [flat|nested] 10+ messages in thread
* [PATCH 3/3 V2] alpha: Implement HW performance events on the EV67 and later CPUs. 2010-05-04 7:40 ` Peter Zijlstra 2010-05-24 21:58 ` Michael Cree 2010-07-27 10:29 ` Michael Cree @ 2010-07-27 10:29 ` Michael Cree 2010-07-28 19:53 ` Andrew Morton 2 siblings, 1 reply; 10+ messages in thread From: Michael Cree @ 2010-07-27 10:29 UTC (permalink / raw) To: Peter Zijlstra Cc: Michael Cree, linux-kernel, Paul Mackerras, Ingo Molnar, Arnaldo Carvalho de Melo, Richard Henderson, Ivan Kokshaysky, Matt Turner, Andrew Morton, Thomas Gleixner, linux-alpha This implements hardware performance events for the EV67 and later CPUs within the Linux performance events subsystem. Only using the performance monitoring unit in HP/Compaq's so called "Aggregrate mode" is supported. The code has been implemented in a manner that makes extension to other older Alpha CPUs relatively straightforward should some mug wish to indulge his or herself. Signed-off-by: Michael Cree <mcree@orcon.net.nz> --- arch/alpha/include/asm/perf_event.h | 8 +- arch/alpha/kernel/Makefile | 1 + arch/alpha/kernel/irq_alpha.c | 3 + arch/alpha/kernel/perf_event.c | 842 +++++++++++++++++++++++++++++++++++ arch/alpha/kernel/time.c | 26 ++ 5 files changed, 879 insertions(+), 1 deletions(-) create mode 100644 arch/alpha/kernel/perf_event.c diff --git a/arch/alpha/include/asm/perf_event.h b/arch/alpha/include/asm/perf_event.h index 3bef852..4157cd3 100644 --- a/arch/alpha/include/asm/perf_event.h +++ b/arch/alpha/include/asm/perf_event.h @@ -2,8 +2,14 @@ #define __ASM_ALPHA_PERF_EVENT_H /* Alpha only supports software events through this interface. */ -static inline void set_perf_event_pending(void) { } +extern void set_perf_event_pending(void); #define PERF_EVENT_INDEX_OFFSET 0 +#ifdef CONFIG_PERF_EVENTS +extern void init_hw_perf_events(void); +#else +static inline void init_hw_perf_events(void) { } +#endif + #endif /* __ASM_ALPHA_PERF_EVENT_H */ diff --git a/arch/alpha/kernel/Makefile b/arch/alpha/kernel/Makefile index 5a62fb4..1ee9b5b 100644 --- a/arch/alpha/kernel/Makefile +++ b/arch/alpha/kernel/Makefile @@ -15,6 +15,7 @@ obj-$(CONFIG_SMP) += smp.o obj-$(CONFIG_PCI) += pci.o pci_iommu.o pci-sysfs.o obj-$(CONFIG_SRM_ENV) += srm_env.o obj-$(CONFIG_MODULES) += module.o +obj-$(CONFIG_PERF_EVENTS) += perf_event.o ifdef CONFIG_ALPHA_GENERIC diff --git a/arch/alpha/kernel/irq_alpha.c b/arch/alpha/kernel/irq_alpha.c index cfde865..5f77afb 100644 --- a/arch/alpha/kernel/irq_alpha.c +++ b/arch/alpha/kernel/irq_alpha.c @@ -10,6 +10,7 @@ #include <asm/machvec.h> #include <asm/dma.h> +#include <asm/perf_event.h> #include "proto.h" #include "irq_impl.h" @@ -111,6 +112,8 @@ init_IRQ(void) wrent(entInt, 0); alpha_mv.init_irq(); + + init_hw_perf_events(); } /* diff --git a/arch/alpha/kernel/perf_event.c b/arch/alpha/kernel/perf_event.c new file mode 100644 index 0000000..51c39fa --- /dev/null +++ b/arch/alpha/kernel/perf_event.c @@ -0,0 +1,842 @@ +/* + * Hardware performance events for the Alpha. + * + * We implement HW counts on the EV67 and subsequent CPUs only. + * + * (C) 2010 Michael J. Cree + * + * Somewhat based on the Sparc code, and to a lesser extent the PowerPC and + * ARM code, which are copyright by their respective authors. + */ + +#include <linux/perf_event.h> +#include <linux/kprobes.h> +#include <linux/kernel.h> +#include <linux/kdebug.h> +#include <linux/mutex.h> + +#include <asm/hwrpb.h> +#include <asm/atomic.h> +#include <asm/irq.h> +#include <asm/irq_regs.h> +#include <asm/pal.h> +#include <asm/wrperfmon.h> +#include <asm/hw_irq.h> + + +/* The maximum number of PMCs on any Alpha CPU whatsoever. */ +#define MAX_HWEVENTS 3 +#define PMC_NO_INDEX -1 + +/* For tracking PMCs and the hw events they monitor on each CPU. */ +struct cpu_hw_events { + int enabled; + /* Number of events scheduled; also number entries valid in arrays below. */ + int n_events; + /* Number events added since last hw_perf_disable(). */ + int n_added; + /* Events currently scheduled. */ + struct perf_event *event[MAX_HWEVENTS]; + /* Event type of each scheduled event. */ + unsigned long evtype[MAX_HWEVENTS]; + /* Current index of each scheduled event; if not yet determined + * contains PMC_NO_INDEX. + */ + int current_idx[MAX_HWEVENTS]; + /* The active PMCs' config for easy use with wrperfmon(). */ + unsigned long config; + /* The active counters' indices for easy use with wrperfmon(). */ + unsigned long idx_mask; +}; +DEFINE_PER_CPU(struct cpu_hw_events, cpu_hw_events); + + + +/* + * A structure to hold the description of the PMCs available on a particular + * type of Alpha CPU. + */ +struct alpha_pmu_t { + /* Mapping of the perf system hw event types to indigenous event types */ + const int *event_map; + /* The number of entries in the event_map */ + int max_events; + /* The number of PMCs on this Alpha */ + int num_pmcs; + /* + * All PMC counters reside in the IBOX register PCTR. This is the + * LSB of the counter. + */ + int pmc_count_shift[MAX_HWEVENTS]; + /* + * The mask that isolates the PMC bits when the LSB of the counter + * is shifted to bit 0. + */ + unsigned long pmc_count_mask[MAX_HWEVENTS]; + /* The maximum period the PMC can count. */ + unsigned long pmc_max_period[MAX_HWEVENTS]; + /* + * The maximum value that may be written to the counter due to + * hardware restrictions is pmc_max_period - pmc_left. + */ + long pmc_left[3]; + /* Subroutine for allocation of PMCs. Enforces constraints. */ + int (*check_constraints)(struct perf_event **, unsigned long *, int); +}; + +/* + * The Alpha CPU PMU description currently in operation. This is set during + * the boot process to the specific CPU of the machine. + */ +static const struct alpha_pmu_t *alpha_pmu; + + +#define HW_OP_UNSUPPORTED -1 + +/* + * The hardware description of the EV67, EV68, EV69, EV7 and EV79 PMUs + * follow. Since they are identical we refer to them collectively as the + * EV67 henceforth. + */ + +/* + * EV67 PMC event types + * + * There is no one-to-one mapping of the possible hw event types to the + * actual codes that are used to program the PMCs hence we introduce our + * own hw event type identifiers. + */ +enum ev67_pmc_event_type { + EV67_CYCLES = 1, + EV67_INSTRUCTIONS, + EV67_BCACHEMISS, + EV67_MBOXREPLAY, + EV67_LAST_ET +}; +#define EV67_NUM_EVENT_TYPES (EV67_LAST_ET-EV67_CYCLES) + + +/* Mapping of the hw event types to the perf tool interface */ +static const int ev67_perfmon_event_map[] = { + [PERF_COUNT_HW_CPU_CYCLES] = EV67_CYCLES, + [PERF_COUNT_HW_INSTRUCTIONS] = EV67_INSTRUCTIONS, + [PERF_COUNT_HW_CACHE_REFERENCES] = HW_OP_UNSUPPORTED, + [PERF_COUNT_HW_CACHE_MISSES] = EV67_BCACHEMISS, +}; + +struct ev67_mapping_t { + int config; + int idx; +}; + +/* + * The mapping used for one event only - these must be in same order as enum + * ev67_pmc_event_type definition. + */ +static const struct ev67_mapping_t ev67_mapping[] = { + {EV67_PCTR_INSTR_CYCLES, 1}, /* EV67_CYCLES, */ + {EV67_PCTR_INSTR_CYCLES, 0}, /* EV67_INSTRUCTIONS */ + {EV67_PCTR_INSTR_BCACHEMISS, 1}, /* EV67_BCACHEMISS */ + {EV67_PCTR_CYCLES_MBOX, 1} /* EV67_MBOXREPLAY */ +}; + + +/* + * Check that a group of events can be simultaneously scheduled on to the + * EV67 PMU. Also allocate counter indices and config. + */ +static int ev67_check_constraints(struct perf_event **event, + unsigned long *evtype, int n_ev) +{ + int idx0; + unsigned long config; + + idx0 = ev67_mapping[evtype[0]-1].idx; + config = ev67_mapping[evtype[0]-1].config; + if (n_ev == 1) + goto success; + + BUG_ON(n_ev != 2); + + if (evtype[0] == EV67_MBOXREPLAY || evtype[1] == EV67_MBOXREPLAY) { + /* MBOX replay traps must be on PMC 1 */ + idx0 = (evtype[0] == EV67_MBOXREPLAY) ? 1 : 0; + /* Only cycles can accompany MBOX replay traps */ + if (evtype[idx0] == EV67_CYCLES) { + config = EV67_PCTR_CYCLES_MBOX; + goto success; + } + } + + if (evtype[0] == EV67_BCACHEMISS || evtype[1] == EV67_BCACHEMISS) { + /* Bcache misses must be on PMC 1 */ + idx0 = (evtype[0] == EV67_BCACHEMISS) ? 1 : 0; + /* Only instructions can accompany Bcache misses */ + if (evtype[idx0] == EV67_INSTRUCTIONS) { + config = EV67_PCTR_INSTR_BCACHEMISS; + goto success; + } + } + + if (evtype[0] == EV67_INSTRUCTIONS || evtype[1] == EV67_INSTRUCTIONS) { + /* Instructions must be on PMC 0 */ + idx0 = (evtype[0] == EV67_INSTRUCTIONS) ? 0 : 1; + /* By this point only cycles can accompany instructions */ + if (evtype[idx0^1] == EV67_CYCLES) { + config = EV67_PCTR_INSTR_CYCLES; + goto success; + } + } + + /* Otherwise, darn it, there is a conflict. */ + return -1; + +success: + event[0]->hw.idx = idx0; + event[0]->hw.config_base = config; + if (n_ev == 2) { + event[1]->hw.idx = idx0 ^ 1; + event[1]->hw.config_base = config; + } + return 0; +} + + +static const struct alpha_pmu_t ev67_pmu = { + .event_map = ev67_perfmon_event_map, + .max_events = ARRAY_SIZE(ev67_perfmon_event_map), + .num_pmcs = 2, + .pmc_count_shift = {EV67_PCTR_0_COUNT_SHIFT, EV67_PCTR_1_COUNT_SHIFT, 0}, + .pmc_count_mask = {EV67_PCTR_0_COUNT_MASK, EV67_PCTR_1_COUNT_MASK, 0}, + .pmc_max_period = {(1UL<<20) - 1, (1UL<<20) - 1, 0}, + .pmc_left = {16, 4, 0}, + .check_constraints = ev67_check_constraints +}; + + + +/* + * Helper routines to ensure that we read/write only the correct PMC bits + * when calling the wrperfmon PALcall. + */ +static inline void alpha_write_pmc(int idx, unsigned long val) +{ + val &= alpha_pmu->pmc_count_mask[idx]; + val <<= alpha_pmu->pmc_count_shift[idx]; + val |= (1<<idx); + wrperfmon(PERFMON_CMD_WRITE, val); +} + +static inline unsigned long alpha_read_pmc(int idx) +{ + unsigned long val; + + val = wrperfmon(PERFMON_CMD_READ, 0); + val >>= alpha_pmu->pmc_count_shift[idx]; + val &= alpha_pmu->pmc_count_mask[idx]; + return val; +} + +/* Set a new period to sample over */ +static int alpha_perf_event_set_period(struct perf_event *event, + struct hw_perf_event *hwc, int idx) +{ + long left = atomic64_read(&hwc->period_left); + long period = hwc->sample_period; + int ret = 0; + + if (unlikely(left <= -period)) { + left = period; + atomic64_set(&hwc->period_left, left); + hwc->last_period = period; + ret = 1; + } + + if (unlikely(left <= 0)) { + left += period; + atomic64_set(&hwc->period_left, left); + hwc->last_period = period; + ret = 1; + } + + /* + * Hardware restrictions require that the counters must not be + * written with values that are too close to the maximum period. + */ + if (unlikely(left < alpha_pmu->pmc_left[idx])) + left = alpha_pmu->pmc_left[idx]; + + if (left > (long)alpha_pmu->pmc_max_period[idx]) + left = alpha_pmu->pmc_max_period[idx]; + + atomic64_set(&hwc->prev_count, (unsigned long)(-left)); + + alpha_write_pmc(idx, (unsigned long)(-left)); + + perf_event_update_userpage(event); + + return ret; +} + + +/* + * Calculates the count (the 'delta') since the last time the PMC was read. + * + * As the PMCs' full period can easily be exceeded within the perf system + * sampling period we cannot use any high order bits as a guard bit in the + * PMCs to detect overflow as is done by other architectures. The code here + * calculates the delta on the basis that there is no overflow when ovf is + * zero. The value passed via ovf by the interrupt handler corrects for + * overflow. + * + * This can be racey on rare occasions -- a call to this routine can occur + * with an overflowed counter just before the PMI service routine is called. + * The check for delta negative hopefully always rectifies this situation. + */ +static unsigned long alpha_perf_event_update(struct perf_event *event, + struct hw_perf_event *hwc, int idx, long ovf) +{ + long prev_raw_count, new_raw_count; + long delta; + +again: + prev_raw_count = atomic64_read(&hwc->prev_count); + new_raw_count = alpha_read_pmc(idx); + + if (atomic64_cmpxchg(&hwc->prev_count, prev_raw_count, + new_raw_count) != prev_raw_count) + goto again; + + delta = (new_raw_count - (prev_raw_count & alpha_pmu->pmc_count_mask[idx])) + ovf; + + /* It is possible on very rare occasions that the PMC has overflowed + * but the interrupt is yet to come. Detect and fix this situation. + */ + if (unlikely(delta < 0)) { + delta += alpha_pmu->pmc_max_period[idx] + 1; + } + + atomic64_add(delta, &event->count); + atomic64_sub(delta, &hwc->period_left); + + return new_raw_count; +} + + +/* + * Collect all HW events into the array event[]. + */ +static int collect_events(struct perf_event *group, int max_count, + struct perf_event *event[], unsigned long *evtype, + int *current_idx) +{ + struct perf_event *pe; + int n = 0; + + if (!is_software_event(group)) { + if (n >= max_count) + return -1; + event[n] = group; + evtype[n] = group->hw.event_base; + current_idx[n++] = PMC_NO_INDEX; + } + list_for_each_entry(pe, &group->sibling_list, group_entry) { + if (!is_software_event(pe) && pe->state != PERF_EVENT_STATE_OFF) { + if (n >= max_count) + return -1; + event[n] = pe; + evtype[n] = pe->hw.event_base; + current_idx[n++] = PMC_NO_INDEX; + } + } + return n; +} + + + +/* + * Check that a group of events can be simultaneously scheduled on to the PMU. + */ +static int alpha_check_constraints(struct perf_event **events, + unsigned long *evtypes, int n_ev) +{ + + /* No HW events is possible from hw_perf_group_sched_in(). */ + if (n_ev == 0) + return 0; + + if (n_ev > alpha_pmu->num_pmcs) + return -1; + + return alpha_pmu->check_constraints(events, evtypes, n_ev); +} + + +/* + * If new events have been scheduled then update cpuc with the new + * configuration. This may involve shifting cycle counts from one PMC to + * another. + */ +static void maybe_change_configuration(struct cpu_hw_events *cpuc) +{ + int j; + + if (cpuc->n_added == 0) + return; + + /* Find counters that are moving to another PMC and update */ + for (j = 0; j < cpuc->n_events; j++) { + struct perf_event *pe = cpuc->event[j]; + + if (cpuc->current_idx[j] != PMC_NO_INDEX && + cpuc->current_idx[j] != pe->hw.idx) { + alpha_perf_event_update(pe, &pe->hw, cpuc->current_idx[j], 0); + cpuc->current_idx[j] = PMC_NO_INDEX; + } + } + + /* Assign to counters all unassigned events. */ + cpuc->idx_mask = 0; + for (j = 0; j < cpuc->n_events; j++) { + struct perf_event *pe = cpuc->event[j]; + struct hw_perf_event *hwc = &pe->hw; + int idx = hwc->idx; + + if (cpuc->current_idx[j] != PMC_NO_INDEX) { + cpuc->idx_mask |= (1<<cpuc->current_idx[j]); + continue; + } + + alpha_perf_event_set_period(pe, hwc, idx); + cpuc->current_idx[j] = idx; + cpuc->idx_mask |= (1<<cpuc->current_idx[j]); + } + cpuc->config = cpuc->event[0]->hw.config_base; +} + + + +/* Schedule perf HW event on to PMU. + * - this function is called from outside this module via the pmu struct + * returned from perf event initialisation. + */ +static int alpha_pmu_enable(struct perf_event *event) +{ + struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events); + int n0; + int ret; + unsigned long flags; + + /* + * The Sparc code has the IRQ disable first followed by the perf + * disable, however this can lead to an overflowed counter with the + * PMI disabled on rare occasions. The alpha_perf_event_update() + * routine should detect this situation by noting a negative delta, + * nevertheless we disable the PMCs first to enable a potential + * final PMI to occur before we disable interrupts. + */ + perf_disable(); + local_irq_save(flags); + + /* Default to error to be returned */ + ret = -EAGAIN; + + /* Insert event on to PMU and if successful modify ret to valid return */ + n0 = cpuc->n_events; + if (n0 < alpha_pmu->num_pmcs) { + cpuc->event[n0] = event; + cpuc->evtype[n0] = event->hw.event_base; + cpuc->current_idx[n0] = PMC_NO_INDEX; + + if (!alpha_check_constraints(cpuc->event, cpuc->evtype, n0+1)) { + cpuc->n_events++; + cpuc->n_added++; + ret = 0; + } + } + + local_irq_restore(flags); + perf_enable(); + + return ret; +} + + + +/* Disable performance monitoring unit + * - this function is called from outside this module via the pmu struct + * returned from perf event initialisation. + */ +static void alpha_pmu_disable(struct perf_event *event) +{ + struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events); + struct hw_perf_event *hwc = &event->hw; + unsigned long flags; + int j; + + perf_disable(); + local_irq_save(flags); + + for (j = 0; j < cpuc->n_events; j++) { + if (event == cpuc->event[j]) { + int idx = cpuc->current_idx[j]; + + /* Shift remaining entries down into the existing + * slot. + */ + while (++j < cpuc->n_events) { + cpuc->event[j - 1] = cpuc->event[j]; + cpuc->evtype[j - 1] = cpuc->evtype[j]; + cpuc->current_idx[j - 1] = + cpuc->current_idx[j]; + } + + /* Absorb the final count and turn off the event. */ + alpha_perf_event_update(event, hwc, idx, 0); + perf_event_update_userpage(event); + + cpuc->idx_mask &= ~(1UL<<idx); + cpuc->n_events--; + break; + } + } + + local_irq_restore(flags); + perf_enable(); +} + + +static void alpha_pmu_read(struct perf_event *event) +{ + struct hw_perf_event *hwc = &event->hw; + + alpha_perf_event_update(event, hwc, hwc->idx, 0); +} + + +static void alpha_pmu_unthrottle(struct perf_event *event) +{ + struct hw_perf_event *hwc = &event->hw; + struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events); + + cpuc->idx_mask |= 1UL<<hwc->idx; + wrperfmon(PERFMON_CMD_ENABLE, (1UL<<hwc->idx)); +} + + +/* + * Check that CPU performance counters are supported. + * - currently support EV67 and later CPUs. + * - actually some later revisions of the EV6 have the same PMC model as the + * EV67 but we don't do suffiently deep CPU detection to detect them. + * Bad luck to the very few people who might have one, I guess. + */ +static int supported_cpu(void) +{ + struct percpu_struct *cpu; + unsigned long cputype; + + /* Get cpu type from HW */ + cpu = (struct percpu_struct *)((char *)hwrpb + hwrpb->processor_offset); + cputype = cpu->type & 0xffffffff; + /* Include all of EV67, EV68, EV7, EV79 and EV69 as supported. */ + return (cputype >= EV67_CPU) && (cputype <= EV69_CPU); +} + + + +static void hw_perf_event_destroy(struct perf_event *event) +{ + /* Nothing to be done! */ + return; +} + + + +static int __hw_perf_event_init(struct perf_event *event) +{ + struct perf_event_attr *attr = &event->attr; + struct hw_perf_event *hwc = &event->hw; + struct perf_event *evts[MAX_HWEVENTS]; + unsigned long evtypes[MAX_HWEVENTS]; + int idx_rubbish_bin[MAX_HWEVENTS]; + int ev; + int n; + + /* We only support a limited range of HARDWARE event types with one + * only programmable via a RAW event type. + */ + if (attr->type == PERF_TYPE_HARDWARE) { + if (attr->config >= alpha_pmu->max_events) + return -EINVAL; + ev = alpha_pmu->event_map[attr->config]; + } else if (attr->type == PERF_TYPE_HW_CACHE) { + return -EOPNOTSUPP; + } else if (attr->type == PERF_TYPE_RAW) { + ev = attr->config & 0xff; + } else { + return -EOPNOTSUPP; + } + + if (ev < 0) { + return ev; + } + + /* The EV67 does not support mode exclusion */ + if (attr->exclude_kernel || attr->exclude_user + || attr->exclude_hv || attr->exclude_idle) { + return -EPERM; + } + + /* + * We place the event type in event_base here and leave calculation + * of the codes to programme the PMU for alpha_pmu_enable() because + * it is only then we will know what HW events are actually + * scheduled on to the PMU. At that point the code to programme the + * PMU is put into config_base and the PMC to use is placed into + * idx. We initialise idx (below) to PMC_NO_INDEX to indicate that + * it is yet to be determined. + */ + hwc->event_base = ev; + + /* Collect events in a group together suitable for calling + * alpha_check_constraints() to verify that the group as a whole can + * be scheduled on to the PMU. + */ + n = 0; + if (event->group_leader != event) { + n = collect_events(event->group_leader, + alpha_pmu->num_pmcs - 1, + evts, evtypes, idx_rubbish_bin); + if (n < 0) + return -EINVAL; + } + evtypes[n] = hwc->event_base; + evts[n] = event; + + if (alpha_check_constraints(evts, evtypes, n + 1)) + return -EINVAL; + + /* Indicate that PMU config and idx are yet to be determined. */ + hwc->config_base = 0; + hwc->idx = PMC_NO_INDEX; + + event->destroy = hw_perf_event_destroy; + + /* + * Most architectures reserve the PMU for their use at this point. + * As there is no existing mechanism to arbitrate usage and there + * appears to be no other user of the Alpha PMU we just assume + * that we can just use it, hence a NO-OP here. + * + * Maybe an alpha_reserve_pmu() routine should be implemented but is + * anything else ever going to use it? + */ + + if (!hwc->sample_period) { + hwc->sample_period = alpha_pmu->pmc_max_period[0]; + hwc->last_period = hwc->sample_period; + atomic64_set(&hwc->period_left, hwc->sample_period); + } + + return 0; +} + +static const struct pmu pmu = { + .enable = alpha_pmu_enable, + .disable = alpha_pmu_disable, + .read = alpha_pmu_read, + .unthrottle = alpha_pmu_unthrottle, +}; + + +/* + * Main entry point to initialise a HW performance event. + */ +const struct pmu *hw_perf_event_init(struct perf_event *event) +{ + int err; + + if (!alpha_pmu) + return ERR_PTR(-ENODEV); + + /* Do the real initialisation work. */ + err = __hw_perf_event_init(event); + + if (err) + return ERR_PTR(err); + + return &pmu; +} + + + +/* + * Main entry point - enable HW performance counters. + */ +void hw_perf_enable(void) +{ + struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events); + + if (cpuc->enabled) + return; + + cpuc->enabled = 1; + barrier(); + + if (cpuc->n_events > 0) { + /* Update cpuc with information from any new scheduled events. */ + maybe_change_configuration(cpuc); + + /* Start counting the desired events. */ + wrperfmon(PERFMON_CMD_LOGGING_OPTIONS, EV67_PCTR_MODE_AGGREGATE); + wrperfmon(PERFMON_CMD_DESIRED_EVENTS, cpuc->config); + wrperfmon(PERFMON_CMD_ENABLE, cpuc->idx_mask); + } +} + + +/* + * Main entry point - disable HW performance counters. + */ + +void hw_perf_disable(void) +{ + struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events); + + if (!cpuc->enabled) + return; + + cpuc->enabled = 0; + cpuc->n_added = 0; + + wrperfmon(PERFMON_CMD_DISABLE, cpuc->idx_mask); +} + + +/* + * Main entry point - don't know when this is called but it + * obviously dumps debug info. + */ +void perf_event_print_debug(void) +{ + unsigned long flags; + unsigned long pcr; + int pcr0, pcr1; + int cpu; + + if (!supported_cpu()) + return; + + local_irq_save(flags); + + cpu = smp_processor_id(); + + pcr = wrperfmon(PERFMON_CMD_READ, 0); + pcr0 = (pcr >> alpha_pmu->pmc_count_shift[0]) & alpha_pmu->pmc_count_mask[0]; + pcr1 = (pcr >> alpha_pmu->pmc_count_shift[1]) & alpha_pmu->pmc_count_mask[1]; + + pr_info("CPU#%d: PCTR0[%06x] PCTR1[%06x]\n", cpu, pcr0, pcr1); + + local_irq_restore(flags); +} + + +/* + * Performance Monitoring Interrupt Service Routine called when a PMC + * overflows. The PMC that overflowed is passed in la_ptr. + */ +static void alpha_perf_event_irq_handler(unsigned long la_ptr, + struct pt_regs *regs) +{ + struct cpu_hw_events *cpuc; + struct perf_sample_data data; + struct perf_event *event; + struct hw_perf_event *hwc; + int idx, j; + + __get_cpu_var(irq_pmi_count)++; + cpuc = &__get_cpu_var(cpu_hw_events); + + /* Completely counting through the PMC's period to trigger a new PMC + * overflow interrupt while in this interrupt routine is utterly + * disastrous! The EV6 and EV67 counters are sufficiently large to + * prevent this but to be really sure disable the PMCs. + */ + wrperfmon(PERFMON_CMD_DISABLE, cpuc->idx_mask); + + /* la_ptr is the counter that overflowed. */ + if (unlikely(la_ptr >= perf_max_events)) { + /* This should never occur! */ + irq_err_count++; + pr_warning("PMI: silly index %ld\n", la_ptr); + wrperfmon(PERFMON_CMD_ENABLE, cpuc->idx_mask); + return; + } + + idx = la_ptr; + + perf_sample_data_init(&data, 0); + for (j = 0; j < cpuc->n_events; j++) { + if (cpuc->current_idx[j] == idx) + break; + } + + if (unlikely(j == cpuc->n_events)) { + /* This can occur if the event is disabled right on a PMC overflow. */ + wrperfmon(PERFMON_CMD_ENABLE, cpuc->idx_mask); + return; + } + + event = cpuc->event[j]; + + if (unlikely(!event)) { + /* This should never occur! */ + irq_err_count++; + pr_warning("PMI: No event at index %d!\n", idx); + wrperfmon(PERFMON_CMD_ENABLE, cpuc->idx_mask); + return; + } + + hwc = &event->hw; + alpha_perf_event_update(event, hwc, idx, alpha_pmu->pmc_max_period[idx]+1); + data.period = event->hw.last_period; + + if (alpha_perf_event_set_period(event, hwc, idx)) { + if (perf_event_overflow(event, 1, &data, regs)) { + /* Interrupts coming too quickly; "throttle" the + * counter, i.e., disable it for a little while. + */ + cpuc->idx_mask &= ~(1UL<<idx); + } + } + wrperfmon(PERFMON_CMD_ENABLE, cpuc->idx_mask); + + return; +} + + + +/* + * Init call to initialise performance events at kernel startup. + */ +void __init init_hw_perf_events(void) +{ + pr_info("Performance events: "); + + if (!supported_cpu()) { + pr_cont("No support for your CPU.\n"); + return; + } + + pr_cont("Supported CPU type!\n"); + + /* Override performance counter IRQ vector */ + + perf_irq = alpha_perf_event_irq_handler; + + /* And set up PMU specification */ + alpha_pmu = &ev67_pmu; + perf_max_events = alpha_pmu->num_pmcs; +} + diff --git a/arch/alpha/kernel/time.c b/arch/alpha/kernel/time.c index 1efbed8..eacceb2 100644 --- a/arch/alpha/kernel/time.c +++ b/arch/alpha/kernel/time.c @@ -41,6 +41,7 @@ #include <linux/init.h> #include <linux/bcd.h> #include <linux/profile.h> +#include <linux/perf_event.h> #include <asm/uaccess.h> #include <asm/io.h> @@ -82,6 +83,26 @@ static struct { unsigned long est_cycle_freq; +#ifdef CONFIG_PERF_EVENTS + +DEFINE_PER_CPU(u8, perf_event_pending); + +#define set_perf_event_pending_flag() __get_cpu_var(perf_event_pending) = 1 +#define test_perf_event_pending() __get_cpu_var(perf_event_pending) +#define clear_perf_event_pending() __get_cpu_var(perf_event_pending) = 0 + +void set_perf_event_pending(void) +{ + set_perf_event_pending_flag(); +} + +#else /* CONFIG_PERF_EVENTS */ + +#define test_perf_event_pending() 0 +#define clear_perf_event_pending() + +#endif /* CONFIG_PERF_EVENTS */ + static inline __u32 rpcc(void) { @@ -175,6 +196,11 @@ irqreturn_t timer_interrupt(int irq, void *dev) update_process_times(user_mode(get_irq_regs())); #endif + if (test_perf_event_pending()) { + clear_perf_event_pending(); + perf_event_do_pending(); + } + return IRQ_HANDLED; } -- 1.7.1 ^ permalink raw reply related [flat|nested] 10+ messages in thread
* Re: [PATCH 3/3 V2] alpha: Implement HW performance events on the EV67 and later CPUs. 2010-07-27 10:29 ` [PATCH 3/3 V2] " Michael Cree @ 2010-07-28 19:53 ` Andrew Morton 2010-07-28 20:52 ` Michael Cree 0 siblings, 1 reply; 10+ messages in thread From: Andrew Morton @ 2010-07-28 19:53 UTC (permalink / raw) To: Michael Cree Cc: Peter Zijlstra, linux-kernel, Paul Mackerras, Ingo Molnar, Arnaldo Carvalho de Melo, Richard Henderson, Ivan Kokshaysky, Matt Turner, Thomas Gleixner, linux-alpha On Tue, 27 Jul 2010 22:29:39 +1200 Michael Cree <mcree@orcon.net.nz> wrote: > This implements hardware performance events for the EV67 and later > CPUs within the Linux performance events subsystem. Only using the > performance monitoring unit in HP/Compaq's so called "Aggregrate > mode" is supported. > > The code has been implemented in a manner that makes extension to > other older Alpha CPUs relatively straightforward should some mug > wish to indulge his or herself. Below is the incremental diff. I have a note here that Peter had issues with the earlier version of this patch. But I see no info here regarding what those issues were, nor whether or how they were addressed. arch/alpha/kernel/perf_event.c | 32 ++++++++++++++++++------------- arch/alpha/kernel/time.c | 26 +++++++++++++++++++++++++ 2 files changed, 45 insertions(+), 13 deletions(-) diff -puN arch/alpha/kernel/perf_event.c~a arch/alpha/kernel/perf_event.c --- a/arch/alpha/kernel/perf_event.c~a +++ a/arch/alpha/kernel/perf_event.c @@ -517,8 +517,10 @@ static void alpha_pmu_read(struct perf_e static void alpha_pmu_unthrottle(struct perf_event *event) { struct hw_perf_event *hwc = &event->hw; + struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events); - wrperfmon(PERFMON_CMD_ENABLE, (1<<hwc->idx)); + cpuc->idx_mask |= 1UL<<hwc->idx; + wrperfmon(PERFMON_CMD_ENABLE, (1UL<<hwc->idx)); } @@ -756,12 +758,19 @@ static void alpha_perf_event_irq_handler __get_cpu_var(irq_pmi_count)++; cpuc = &__get_cpu_var(cpu_hw_events); - /* la_ptr is the counter that overflowed. */ + /* Completely counting through the PMC's period to trigger a new PMC + * overflow interrupt while in this interrupt routine is utterly + * disastrous! The EV6 and EV67 counters are sufficiently large to + * prevent this but to be really sure disable the PMCs. + */ + wrperfmon(PERFMON_CMD_DISABLE, cpuc->idx_mask); - if (la_ptr >= perf_max_events) { + /* la_ptr is the counter that overflowed. */ + if (unlikely(la_ptr >= perf_max_events)) { /* This should never occur! */ irq_err_count++; pr_warning("PMI: silly index %ld\n", la_ptr); + wrperfmon(PERFMON_CMD_ENABLE, cpuc->idx_mask); return; } @@ -773,17 +782,19 @@ static void alpha_perf_event_irq_handler break; } - if (j == cpuc->n_events) { + if (unlikely(j == cpuc->n_events)) { /* This can occur if the event is disabled right on a PMC overflow. */ + wrperfmon(PERFMON_CMD_ENABLE, cpuc->idx_mask); return; } event = cpuc->event[j]; - if (!event) { + if (unlikely(!event)) { /* This should never occur! */ irq_err_count++; pr_warning("PMI: No event at index %d!\n", idx); + wrperfmon(PERFMON_CMD_ENABLE, cpuc->idx_mask); return; } @@ -792,24 +803,19 @@ static void alpha_perf_event_irq_handler data.period = event->hw.last_period; if (alpha_perf_event_set_period(event, hwc, idx)) { - if (perf_event_overflow(event, 0, &data, regs)) { + if (perf_event_overflow(event, 1, &data, regs)) { /* Interrupts coming too quickly; "throttle" the * counter, i.e., disable it for a little while. */ - wrperfmon(PERFMON_CMD_DISABLE, 1<<idx); + cpuc->idx_mask &= ~(1UL<<idx); } } + wrperfmon(PERFMON_CMD_ENABLE, cpuc->idx_mask); return; } -/* I wonder what this is for ??? */ -void set_perf_event_pending(void) -{ -} - - /* * Init call to initialise performance events at kernel startup. diff -puN arch/alpha/kernel/time.c~a arch/alpha/kernel/time.c --- a/arch/alpha/kernel/time.c~a +++ a/arch/alpha/kernel/time.c @@ -41,6 +41,7 @@ #include <linux/init.h> #include <linux/bcd.h> #include <linux/profile.h> +#include <linux/perf_event.h> #include <asm/uaccess.h> #include <asm/io.h> @@ -82,6 +83,26 @@ static struct { unsigned long est_cycle_freq; +#ifdef CONFIG_PERF_EVENTS + +DEFINE_PER_CPU(u8, perf_event_pending); + +#define set_perf_event_pending_flag() __get_cpu_var(perf_event_pending) = 1 +#define test_perf_event_pending() __get_cpu_var(perf_event_pending) +#define clear_perf_event_pending() __get_cpu_var(perf_event_pending) = 0 + +void set_perf_event_pending(void) +{ + set_perf_event_pending_flag(); +} + +#else /* CONFIG_PERF_EVENTS */ + +#define test_perf_event_pending() 0 +#define clear_perf_event_pending() + +#endif /* CONFIG_PERF_EVENTS */ + static inline __u32 rpcc(void) { @@ -175,6 +196,11 @@ irqreturn_t timer_interrupt(int irq, voi update_process_times(user_mode(get_irq_regs())); #endif + if (test_perf_event_pending()) { + clear_perf_event_pending(); + perf_event_do_pending(); + } + return IRQ_HANDLED; } _ ^ permalink raw reply [flat|nested] 10+ messages in thread
* Re: [PATCH 3/3 V2] alpha: Implement HW performance events on the EV67 and later CPUs. 2010-07-28 19:53 ` Andrew Morton @ 2010-07-28 20:52 ` Michael Cree 0 siblings, 0 replies; 10+ messages in thread From: Michael Cree @ 2010-07-28 20:52 UTC (permalink / raw) To: Andrew Morton Cc: Peter Zijlstra, linux-kernel, Paul Mackerras, Ingo Molnar, Arnaldo Carvalho de Melo, Richard Henderson, Ivan Kokshaysky, Matt Turner, Thomas Gleixner, linux-alpha On 29/07/2010, at 7:53 AM, Andrew Morton wrote: > On Tue, 27 Jul 2010 22:29:39 +1200 > Michael Cree <mcree@orcon.net.nz> wrote: > >> This implements hardware performance events for the EV67 and later >> CPUs within the Linux performance events subsystem. Only using the >> performance monitoring unit in HP/Compaq's so called "Aggregrate >> mode" is supported. >> >> The code has been implemented in a manner that makes extension to >> other older Alpha CPUs relatively straightforward should some mug >> wish to indulge his or herself. > > Below is the incremental diff. > > I have a note here that Peter had issues with the earlier version of > this patch. But I see no info here regarding what those issues were, > nor whether or how they were addressed. Yeah, I addressed those in the cover letter, which I include here FYI: On 04/05/10 19:40, Peter Zijlstra wrote: > On Sat, 2010-05-01 at 11:55 +1200, Michael Cree wrote: >> +/* I wonder what this is for ??? */ >> +void set_perf_event_pending(void) >> +{ >> +} > > If the PMU Interrupt comes in as an NMI we cannot take locks and such > from its handler. So what we do is queue that work to be ran later. > The > kernel has a fallback to run this stuff from the timer hardirq, but if > the PMI is IRQ context (like ARM) you can simply call > perf_event_do_pending() at its tail. > > Alternatively you can self-IPI like x86 does and run > perf_event_do_pending() from there. OK, I couldn't see how to self-IPI on the Apha (there is no obvious way in the Alpha arch code, and the Alpha HW Ref. Man. only describes such a feature being supported by the PALcode for OpenVMS which we are obviously not running) and the PMI on Alpha is at such a high priority there is no chance that the lock can be freed while in the PMI handler, so I have put the call to perf_event_do_pending() in the timer interrupt, but unlike the PowerPC code I don't know any tricks to force the timer interrupt to happen early when there is pending work, so there is a potential maximum delay of 1ms from pending work being notified to calling perf_event_do_pending(). Is this delay acceptable? I also fix a nasty, but heretofore hidden, bug that could completely lock up a machine with continuous PMIs due to a throttled PMC being accidently re-enabled by the other PMC. It only showed up when I set the NMI flag true in the call to perf_event_overflow() and had two events counting simultaneously with a very short sample period. I also see the x86 and Sparc code in 2.6.35 have reimplemented the call to hw_perf_group_sched_in() in terms of new functions start_txn(), cancel_txn() and commit_txn(). Is this change necessary to get the Alpha code accepted for 2.6.36? If so it would be helpful if I could have a brief description of what each function does. Cheers Michael. ^ permalink raw reply [flat|nested] 10+ messages in thread
end of thread, other threads:[~2010-07-28 20:52 UTC | newest] Thread overview: 10+ messages (download: mbox.gz follow: Atom feed -- links below jump to the message on this page -- 2010-04-30 23:55 [PATCH 0/3] alpha: Implement Hardware Performance Events Michael Cree 2010-04-30 23:55 ` [PATCH 1/3] alpha: Add performance monitor interrupt counter Michael Cree 2010-04-30 23:55 ` [PATCH 2/3] alpha: Add wrperfmon.h header file to aid use of wrperfmon PALcall Michael Cree 2010-04-30 23:55 ` [PATCH 3/3] alpha: Implement HW performance events on the EV67 and later CPUs Michael Cree 2010-05-04 7:40 ` Peter Zijlstra 2010-05-24 21:58 ` Michael Cree 2010-07-27 10:29 ` Michael Cree 2010-07-27 10:29 ` [PATCH 3/3 V2] " Michael Cree 2010-07-28 19:53 ` Andrew Morton 2010-07-28 20:52 ` Michael Cree
This is a public inbox, see mirroring instructions for how to clone and mirror all data and code used for this inbox