From 767f30210bee1bb74ab042e237c2a21e77874070 Mon Sep 17 00:00:00 2001 From: Henrik Austad Date: Tue, 4 Mar 2014 09:20:53 +0100 Subject: tile: avoid overflow in ns2cycles In commit 4cecf6d401a ("sched, x86: Avoid unnecessary overflow in sched_clock") and in recent patch "clocksource: avoid unnecessary overflow in cyclecounter_cyc2ns()" https://lkml.org/lkml/2014/3/4/17, the mult-shift approach is replaced by 2 steps to avoid storing a large, intermediate value that could overflow. arch/tile/kernel/time.c has a similar pattern in cycles2ns, and this copies the same pattern in this function CC: John Stultz CC: Mike Galbraith CC: Salman Qazi Signed-off-by: Henrik Austad Signed-off-by: Chris Metcalf --- arch/tile/kernel/time.c | 10 +++++++++- 1 file changed, 9 insertions(+), 1 deletion(-) (limited to 'arch') diff --git a/arch/tile/kernel/time.c b/arch/tile/kernel/time.c index 5d10642..462dcd0 100644 --- a/arch/tile/kernel/time.c +++ b/arch/tile/kernel/time.c @@ -236,7 +236,15 @@ cycles_t ns2cycles(unsigned long nsecs) * clock frequency. */ struct clock_event_device *dev = &__raw_get_cpu_var(tile_timer); - return ((u64)nsecs * dev->mult) >> dev->shift; + + /* + * as in clocksource.h and x86's timer.h, we split the calculation + * into 2 parts to avoid unecessary overflow of the intermediate + * value. This will not lead to any loss of precision. + */ + u64 quot = (u64)nsecs >> dev->shift; + u64 rem = (u64)nsecs & ((1ULL << dev->shift) - 1); + return quot * dev->mult + ((rem * dev->mult) >> dev->shift); } void update_vsyscall_tz(void) -- cgit v1.1 From 2e285458e62fe202cb81b831f41779c7e7e5cbe6 Mon Sep 17 00:00:00 2001 From: Chris Metcalf Date: Thu, 6 Mar 2014 15:31:59 -0500 Subject: tile: don't use __get_cpu_var() with structure-typed arguments This no longer works with the new per-cpu infrastructure. Signed-off-by: Chris Metcalf --- arch/tile/kernel/messaging.c | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) (limited to 'arch') diff --git a/arch/tile/kernel/messaging.c b/arch/tile/kernel/messaging.c index 00331af..7867266 100644 --- a/arch/tile/kernel/messaging.c +++ b/arch/tile/kernel/messaging.c @@ -68,8 +68,8 @@ void hv_message_intr(struct pt_regs *regs, int intnum) #endif while (1) { - rmi = hv_receive_message(__get_cpu_var(msg_state), - (HV_VirtAddr) message, + HV_MsgState *state = this_cpu_ptr(&msg_state); + rmi = hv_receive_message(*state, (HV_VirtAddr) message, sizeof(message)); if (rmi.msglen == 0) break; -- cgit v1.1 From 8e3441ebab48c3537b1a9ae06fb7616a3332bd35 Mon Sep 17 00:00:00 2001 From: Zhigang Lu Date: Mon, 27 Jan 2014 15:11:07 +0800 Subject: tile: Add support for handling PMC hardware The PMC module is used by perf_events, oprofile and watchdogs. Signed-off-by: Zhigang Lu Signed-off-by: Chris Metcalf --- arch/tile/Kconfig | 4 ++ arch/tile/include/asm/pmc.h | 64 +++++++++++++++++++++++ arch/tile/kernel/Makefile | 1 + arch/tile/kernel/intvec_32.S | 13 ++--- arch/tile/kernel/intvec_64.S | 13 ++--- arch/tile/kernel/pmc.c | 121 +++++++++++++++++++++++++++++++++++++++++++ 6 files changed, 204 insertions(+), 12 deletions(-) create mode 100644 arch/tile/include/asm/pmc.h create mode 100644 arch/tile/kernel/pmc.c (limited to 'arch') diff --git a/arch/tile/Kconfig b/arch/tile/Kconfig index b3692ce..3067b15 100644 --- a/arch/tile/Kconfig +++ b/arch/tile/Kconfig @@ -66,6 +66,10 @@ config HUGETLB_SUPER_PAGES config GENERIC_TIME_VSYSCALL def_bool y +# Enable PMC if PERF_EVENTS, OPROFILE, or WATCHPOINTS are enabled. +config USE_PMC + bool + # FIXME: tilegx can implement a more efficient rwsem. config RWSEM_GENERIC_SPINLOCK def_bool y diff --git a/arch/tile/include/asm/pmc.h b/arch/tile/include/asm/pmc.h new file mode 100644 index 0000000..7ae3956 --- /dev/null +++ b/arch/tile/include/asm/pmc.h @@ -0,0 +1,64 @@ +/* + * Copyright 2014 Tilera Corporation. All Rights Reserved. + * + * This program is free software; you can redistribute it and/or + * modify it under the terms of the GNU General Public License + * as published by the Free Software Foundation, version 2. + * + * This program is distributed in the hope that it will be useful, but + * WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or + * NON INFRINGEMENT. See the GNU General Public License for + * more details. + */ + +#ifndef _ASM_TILE_PMC_H +#define _ASM_TILE_PMC_H + +#include + +#define TILE_BASE_COUNTERS 2 + +/* Bitfields below are derived from SPR PERF_COUNT_CTL*/ +#ifndef __tilegx__ +/* PERF_COUNT_CTL on TILEPro */ +#define TILE_CTL_EXCL_USER (1 << 7) /* exclude user level */ +#define TILE_CTL_EXCL_KERNEL (1 << 8) /* exclude kernel level */ +#define TILE_CTL_EXCL_HV (1 << 9) /* exclude hypervisor level */ + +#define TILE_SEL_MASK 0x7f /* 7 bits for event SEL, + COUNT_0_SEL */ +#define TILE_PLM_MASK 0x780 /* 4 bits priv level msks, + COUNT_0_MASK*/ +#define TILE_EVENT_MASK (TILE_SEL_MASK | TILE_PLM_MASK) + +#else /* __tilegx__*/ +/* PERF_COUNT_CTL on TILEGx*/ +#define TILE_CTL_EXCL_USER (1 << 10) /* exclude user level */ +#define TILE_CTL_EXCL_KERNEL (1 << 11) /* exclude kernel level */ +#define TILE_CTL_EXCL_HV (1 << 12) /* exclude hypervisor level */ + +#define TILE_SEL_MASK 0x3f /* 6 bits for event SEL, + COUNT_0_SEL*/ +#define TILE_BOX_MASK 0x1c0 /* 3 bits box msks, + COUNT_0_BOX */ +#define TILE_PLM_MASK 0x3c00 /* 4 bits priv level msks, + COUNT_0_MASK */ +#define TILE_EVENT_MASK (TILE_SEL_MASK | TILE_BOX_MASK | TILE_PLM_MASK) +#endif /* __tilegx__*/ + +/* Takes register and fault number. Returns error to disable the interrupt. */ +typedef int (*perf_irq_t)(struct pt_regs *, int); + +int userspace_perf_handler(struct pt_regs *regs, int fault); + +perf_irq_t reserve_pmc_hardware(perf_irq_t new_perf_irq); +void release_pmc_hardware(void); + +unsigned long pmc_get_overflow(void); +void pmc_ack_overflow(unsigned long status); + +void unmask_pmc_interrupts(void); +void mask_pmc_interrupts(void); + +#endif /* _ASM_TILE_PMC_H */ diff --git a/arch/tile/kernel/Makefile b/arch/tile/kernel/Makefile index 27a2bf3..71d8353 100644 --- a/arch/tile/kernel/Makefile +++ b/arch/tile/kernel/Makefile @@ -25,6 +25,7 @@ obj-$(CONFIG_PCI) += pci_gx.o else obj-$(CONFIG_PCI) += pci.o endif +obj-$(CONFIG_USE_PMC) += pmc.o obj-$(CONFIG_TILE_USB) += usb.o obj-$(CONFIG_TILE_HVGLUE_TRACE) += hvglue_trace.o obj-$(CONFIG_FUNCTION_TRACER) += ftrace.o mcount_64.o diff --git a/arch/tile/kernel/intvec_32.S b/arch/tile/kernel/intvec_32.S index 2cbe6d5..605ffbd 100644 --- a/arch/tile/kernel/intvec_32.S +++ b/arch/tile/kernel/intvec_32.S @@ -313,13 +313,13 @@ intvec_\vecname: movei r3, 0 } .else - .ifc \c_routine, op_handle_perf_interrupt + .ifc \c_routine, handle_perf_interrupt { mfspr r2, PERF_COUNT_STS movei r3, -1 /* not used, but set for consistency */ } .else - .ifc \c_routine, op_handle_aux_perf_interrupt + .ifc \c_routine, handle_perf_interrupt { mfspr r2, AUX_PERF_COUNT_STS movei r3, -1 /* not used, but set for consistency */ @@ -1835,8 +1835,9 @@ int_unalign: /* Include .intrpt array of interrupt vectors */ .section ".intrpt", "ax" -#define op_handle_perf_interrupt bad_intr -#define op_handle_aux_perf_interrupt bad_intr +#ifndef CONFIG_USE_PMC +#define handle_perf_interrupt bad_intr +#endif #ifndef CONFIG_HARDWALL #define do_hardwall_trap bad_intr @@ -1877,7 +1878,7 @@ int_unalign: int_hand INT_IDN_AVAIL, IDN_AVAIL, bad_intr int_hand INT_UDN_AVAIL, UDN_AVAIL, bad_intr int_hand INT_PERF_COUNT, PERF_COUNT, \ - op_handle_perf_interrupt, handle_nmi + handle_perf_interrupt, handle_nmi int_hand INT_INTCTRL_3, INTCTRL_3, bad_intr #if CONFIG_KERNEL_PL == 2 dc_dispatch INT_INTCTRL_2, INTCTRL_2 @@ -1902,7 +1903,7 @@ int_unalign: int_hand INT_SN_CPL, SN_CPL, bad_intr int_hand INT_DOUBLE_FAULT, DOUBLE_FAULT, do_trap int_hand INT_AUX_PERF_COUNT, AUX_PERF_COUNT, \ - op_handle_aux_perf_interrupt, handle_nmi + handle_perf_interrupt, handle_nmi /* Synthetic interrupt delivered only by the simulator */ int_hand INT_BREAKPOINT, BREAKPOINT, do_breakpoint diff --git a/arch/tile/kernel/intvec_64.S b/arch/tile/kernel/intvec_64.S index b8fc497..8f892a5 100644 --- a/arch/tile/kernel/intvec_64.S +++ b/arch/tile/kernel/intvec_64.S @@ -509,10 +509,10 @@ intvec_\vecname: .ifc \c_routine, do_trap mfspr r2, GPV_REASON .else - .ifc \c_routine, op_handle_perf_interrupt + .ifc \c_routine, handle_perf_interrupt mfspr r2, PERF_COUNT_STS .else - .ifc \c_routine, op_handle_aux_perf_interrupt + .ifc \c_routine, handle_perf_interrupt mfspr r2, AUX_PERF_COUNT_STS .endif .endif @@ -1491,8 +1491,9 @@ STD_ENTRY(fill_ra_stack) .global intrpt_start intrpt_start: -#define op_handle_perf_interrupt bad_intr -#define op_handle_aux_perf_interrupt bad_intr +#ifndef CONFIG_USE_PMC +#define handle_perf_interrupt bad_intr +#endif #ifndef CONFIG_HARDWALL #define do_hardwall_trap bad_intr @@ -1540,9 +1541,9 @@ intrpt_start: #endif int_hand INT_IPI_0, IPI_0, bad_intr int_hand INT_PERF_COUNT, PERF_COUNT, \ - op_handle_perf_interrupt, handle_nmi + handle_perf_interrupt, handle_nmi int_hand INT_AUX_PERF_COUNT, AUX_PERF_COUNT, \ - op_handle_perf_interrupt, handle_nmi + handle_perf_interrupt, handle_nmi int_hand INT_INTCTRL_3, INTCTRL_3, bad_intr #if CONFIG_KERNEL_PL == 2 dc_dispatch INT_INTCTRL_2, INTCTRL_2 diff --git a/arch/tile/kernel/pmc.c b/arch/tile/kernel/pmc.c new file mode 100644 index 0000000..db62cc3 --- /dev/null +++ b/arch/tile/kernel/pmc.c @@ -0,0 +1,121 @@ +/* + * Copyright 2014 Tilera Corporation. All Rights Reserved. + * + * This program is free software; you can redistribute it and/or + * modify it under the terms of the GNU General Public License + * as published by the Free Software Foundation, version 2. + * + * This program is distributed in the hope that it will be useful, but + * WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or + * NON INFRINGEMENT. See the GNU General Public License for + * more details. + */ + +#include +#include +#include +#include +#include + +#include +#include + +perf_irq_t perf_irq = NULL; +int handle_perf_interrupt(struct pt_regs *regs, int fault) +{ + int retval; + + if (!perf_irq) + panic("Unexpected PERF_COUNT interrupt %d\n", fault); + + nmi_enter(); + retval = perf_irq(regs, fault); + nmi_exit(); + return retval; +} + +/* Reserve PMC hardware if it is available. */ +perf_irq_t reserve_pmc_hardware(perf_irq_t new_perf_irq) +{ + return cmpxchg(&perf_irq, NULL, new_perf_irq); +} +EXPORT_SYMBOL(reserve_pmc_hardware); + +/* Release PMC hardware. */ +void release_pmc_hardware(void) +{ + perf_irq = NULL; +} +EXPORT_SYMBOL(release_pmc_hardware); + + +/* + * Get current overflow status of each performance counter, + * and auxiliary performance counter. + */ +unsigned long +pmc_get_overflow(void) +{ + unsigned long status; + + /* + * merge base+aux into a single vector + */ + status = __insn_mfspr(SPR_PERF_COUNT_STS); + status |= __insn_mfspr(SPR_AUX_PERF_COUNT_STS) << TILE_BASE_COUNTERS; + return status; +} + +/* + * Clear the status bit for the corresponding counter, if written + * with a one. + */ +void +pmc_ack_overflow(unsigned long status) +{ + /* + * clear overflow status by writing ones + */ + __insn_mtspr(SPR_PERF_COUNT_STS, status); + __insn_mtspr(SPR_AUX_PERF_COUNT_STS, status >> TILE_BASE_COUNTERS); +} + +/* + * The perf count interrupts are masked and unmasked explicitly, + * and only here. The normal irq_enable() does not enable them, + * and irq_disable() does not disable them. That lets these + * routines drive the perf count interrupts orthogonally. + * + * We also mask the perf count interrupts on entry to the perf count + * interrupt handler in assembly code, and by default unmask them + * again (with interrupt critical section protection) just before + * returning from the interrupt. If the perf count handler returns + * a non-zero error code, then we don't re-enable them before returning. + * + * For Pro, we rely on both interrupts being in the same word to update + * them atomically so we never have one enabled and one disabled. + */ + +#if CHIP_HAS_SPLIT_INTR_MASK() +# if INT_PERF_COUNT < 32 || INT_AUX_PERF_COUNT < 32 +# error Fix assumptions about which word PERF_COUNT interrupts are in +# endif +#endif + +static inline unsigned long long pmc_mask(void) +{ + unsigned long long mask = 1ULL << INT_PERF_COUNT; + mask |= 1ULL << INT_AUX_PERF_COUNT; + return mask; +} + +void unmask_pmc_interrupts(void) +{ + interrupt_mask_reset_mask(pmc_mask()); +} + +void mask_pmc_interrupts(void) +{ + interrupt_mask_set_mask(pmc_mask()); +} -- cgit v1.1 From ba67823163c963de7f1f2d87526c9c87f3a3ea0b Mon Sep 17 00:00:00 2001 From: Zhigang Lu Date: Mon, 27 Jan 2014 16:25:28 +0800 Subject: tile: Enable NMIs on return from handle_nmi() without errors NMI interrupts mask ALL interrupts before calling the handler, so we need to unmask NMIs according to the value handle_nmi() returns. If it returns zero, the NMIs should be re-enabled; if it returns a non-zero error, the NMIs should be disabled. Signed-off-by: Zhigang Lu Signed-off-by: Chris Metcalf --- arch/tile/kernel/intvec_32.S | 11 +++++++++++ arch/tile/kernel/intvec_64.S | 11 ++++++++++- 2 files changed, 21 insertions(+), 1 deletion(-) (limited to 'arch') diff --git a/arch/tile/kernel/intvec_32.S b/arch/tile/kernel/intvec_32.S index 605ffbd..cdbda45 100644 --- a/arch/tile/kernel/intvec_32.S +++ b/arch/tile/kernel/intvec_32.S @@ -946,6 +946,13 @@ STD_ENTRY(interrupt_return) bzt r30, .Lrestore_regs 3: + /* We are relying on INT_PERF_COUNT at 33, and AUX_PERF_COUNT at 48 */ + { + moveli r0, lo16(1 << (INT_PERF_COUNT - 32)) + bz r31, .Lrestore_regs + } + auli r0, r0, ha16(1 << (INT_AUX_PERF_COUNT - 32)) + mtspr SPR_INTERRUPT_MASK_RESET_K_1, r0 /* * We now commit to returning from this interrupt, since we will be @@ -1171,6 +1178,10 @@ handle_nmi: PTREGS_PTR(r0, PTREGS_OFFSET_BASE) } FEEDBACK_REENTER(handle_nmi) + { + movei r30, 1 + seq r31, r0, zero + } j interrupt_return STD_ENDPROC(handle_nmi) diff --git a/arch/tile/kernel/intvec_64.S b/arch/tile/kernel/intvec_64.S index 8f892a5..5b67efc 100644 --- a/arch/tile/kernel/intvec_64.S +++ b/arch/tile/kernel/intvec_64.S @@ -971,6 +971,15 @@ STD_ENTRY(interrupt_return) beqzt r30, .Lrestore_regs 3: +#if INT_PERF_COUNT + 1 != INT_AUX_PERF_COUNT +# error Bad interrupt assumption +#endif + { + movei r0, 3 /* two adjacent bits for the PERF_COUNT mask */ + beqz r31, .Lrestore_regs + } + shli r0, r0, INT_PERF_COUNT + mtspr SPR_INTERRUPT_MASK_RESET_K, r0 /* * We now commit to returning from this interrupt, since we will be @@ -1187,7 +1196,7 @@ handle_nmi: FEEDBACK_REENTER(handle_nmi) { movei r30, 1 - move r31, r0 + cmpeq r31, r0, zero } j interrupt_return STD_ENDPROC(handle_nmi) -- cgit v1.1 From 8d61dd7d3e374eb52a174ab04169b04e3d9d729f Mon Sep 17 00:00:00 2001 From: Zhigang Lu Date: Tue, 28 Jan 2014 10:03:50 +0800 Subject: tile/perf: Support perf_events on tilegx and tilepro Add perf support for tile architecture. Signed-off-by: Zhigang Lu Signed-off-by: Chris Metcalf --- arch/tile/Kconfig | 2 + arch/tile/include/asm/perf_event.h | 22 + arch/tile/kernel/Makefile | 1 + arch/tile/kernel/irq.c | 18 + arch/tile/kernel/perf_event.c | 1005 ++++++++++++++++++++++++++++++++++++ 5 files changed, 1048 insertions(+) create mode 100644 arch/tile/include/asm/perf_event.h create mode 100644 arch/tile/kernel/perf_event.c (limited to 'arch') diff --git a/arch/tile/Kconfig b/arch/tile/Kconfig index 3067b15..31c8c62 100644 --- a/arch/tile/Kconfig +++ b/arch/tile/Kconfig @@ -3,6 +3,8 @@ config TILE def_bool y + select HAVE_PERF_EVENTS + select USE_PMC if PERF_EVENTS select HAVE_DMA_ATTRS select HAVE_DMA_API_DEBUG select HAVE_KVM if !TILEGX diff --git a/arch/tile/include/asm/perf_event.h b/arch/tile/include/asm/perf_event.h new file mode 100644 index 0000000..59c5b16 --- /dev/null +++ b/arch/tile/include/asm/perf_event.h @@ -0,0 +1,22 @@ +/* + * Copyright 2014 Tilera Corporation. All Rights Reserved. + * + * This program is free software; you can redistribute it and/or + * modify it under the terms of the GNU General Public License + * as published by the Free Software Foundation, version 2. + * + * This program is distributed in the hope that it will be useful, but + * WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or + * NON INFRINGEMENT. See the GNU General Public License for + * more details. + */ + +#ifndef _ASM_TILE_PERF_EVENT_H +#define _ASM_TILE_PERF_EVENT_H + +#include +DECLARE_PER_CPU(u64, perf_irqs); + +unsigned long handle_syscall_link_address(void); +#endif /* _ASM_TILE_PERF_EVENT_H */ diff --git a/arch/tile/kernel/Makefile b/arch/tile/kernel/Makefile index 71d8353..21f77bf 100644 --- a/arch/tile/kernel/Makefile +++ b/arch/tile/kernel/Makefile @@ -25,6 +25,7 @@ obj-$(CONFIG_PCI) += pci_gx.o else obj-$(CONFIG_PCI) += pci.o endif +obj-$(CONFIG_PERF_EVENTS) += perf_event.o obj-$(CONFIG_USE_PMC) += pmc.o obj-$(CONFIG_TILE_USB) += usb.o obj-$(CONFIG_TILE_HVGLUE_TRACE) += hvglue_trace.o diff --git a/arch/tile/kernel/irq.c b/arch/tile/kernel/irq.c index 0586fdb..906a76b 100644 --- a/arch/tile/kernel/irq.c +++ b/arch/tile/kernel/irq.c @@ -21,6 +21,7 @@ #include #include #include +#include /* Bit-flag stored in irq_desc->chip_data to indicate HW-cleared irqs. */ #define IS_HW_CLEARED 1 @@ -261,6 +262,23 @@ void ack_bad_irq(unsigned int irq) } /* + * /proc/interrupts printing: + */ +int arch_show_interrupts(struct seq_file *p, int prec) +{ +#ifdef CONFIG_PERF_EVENTS + int i; + + seq_printf(p, "%*s: ", prec, "PMI"); + + for_each_online_cpu(i) + seq_printf(p, "%10llu ", per_cpu(perf_irqs, i)); + seq_puts(p, " perf_events\n"); +#endif + return 0; +} + +/* * Generic, controller-independent functions: */ diff --git a/arch/tile/kernel/perf_event.c b/arch/tile/kernel/perf_event.c new file mode 100644 index 0000000..2bf6c9c --- /dev/null +++ b/arch/tile/kernel/perf_event.c @@ -0,0 +1,1005 @@ +/* + * Copyright 2014 Tilera Corporation. All Rights Reserved. + * + * This program is free software; you can redistribute it and/or + * modify it under the terms of the GNU General Public License + * as published by the Free Software Foundation, version 2. + * + * This program is distributed in the hope that it will be useful, but + * WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or + * NON INFRINGEMENT. See the GNU General Public License for + * more details. + * + * + * Perf_events support for Tile processor. + * + * This code is based upon the x86 perf event + * code, which is: + * + * Copyright (C) 2008 Thomas Gleixner + * Copyright (C) 2008-2009 Red Hat, Inc., Ingo Molnar + * Copyright (C) 2009 Jaswinder Singh Rajput + * Copyright (C) 2009 Advanced Micro Devices, Inc., Robert Richter + * Copyright (C) 2008-2009 Red Hat, Inc., Peter Zijlstra + * Copyright (C) 2009 Intel Corporation, + * Copyright (C) 2009 Google, Inc., Stephane Eranian + */ + +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include + +#define TILE_MAX_COUNTERS 4 + +#define PERF_COUNT_0_IDX 0 +#define PERF_COUNT_1_IDX 1 +#define AUX_PERF_COUNT_0_IDX 2 +#define AUX_PERF_COUNT_1_IDX 3 + +struct cpu_hw_events { + int n_events; + struct perf_event *events[TILE_MAX_COUNTERS]; /* counter order */ + struct perf_event *event_list[TILE_MAX_COUNTERS]; /* enabled + order */ + int assign[TILE_MAX_COUNTERS]; + unsigned long active_mask[BITS_TO_LONGS(TILE_MAX_COUNTERS)]; + unsigned long used_mask; +}; + +/* TILE arch specific performance monitor unit */ +struct tile_pmu { + const char *name; + int version; + const int *hw_events; /* generic hw events table */ + /* generic hw cache events table */ + const int (*cache_events)[PERF_COUNT_HW_CACHE_MAX] + [PERF_COUNT_HW_CACHE_OP_MAX] + [PERF_COUNT_HW_CACHE_RESULT_MAX]; + int (*map_hw_event)(u64); /*method used to map + hw events */ + int (*map_cache_event)(u64); /*method used to map + cache events */ + + u64 max_period; /* max sampling period */ + u64 cntval_mask; /* counter width mask */ + int cntval_bits; /* counter width */ + int max_events; /* max generic hw events + in map */ + int num_counters; /* number base + aux counters */ + int num_base_counters; /* number base counters */ +}; + +DEFINE_PER_CPU(u64, perf_irqs); +static DEFINE_PER_CPU(struct cpu_hw_events, cpu_hw_events); + +#define TILE_OP_UNSUPP (-1) + +#ifndef __tilegx__ +/* TILEPro hardware events map */ +static const int tile_hw_event_map[] = { + [PERF_COUNT_HW_CPU_CYCLES] = 0x01, /* ONE */ + [PERF_COUNT_HW_INSTRUCTIONS] = 0x06, /* MP_BUNDLE_RETIRED */ + [PERF_COUNT_HW_CACHE_REFERENCES] = TILE_OP_UNSUPP, + [PERF_COUNT_HW_CACHE_MISSES] = TILE_OP_UNSUPP, + [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = 0x16, /* + MP_CONDITIONAL_BRANCH_ISSUED */ + [PERF_COUNT_HW_BRANCH_MISSES] = 0x14, /* + MP_CONDITIONAL_BRANCH_MISSPREDICT */ + [PERF_COUNT_HW_BUS_CYCLES] = TILE_OP_UNSUPP, +}; +#else +/* TILEGx hardware events map */ +static const int tile_hw_event_map[] = { + [PERF_COUNT_HW_CPU_CYCLES] = 0x181, /* ONE */ + [PERF_COUNT_HW_INSTRUCTIONS] = 0xdb, /* INSTRUCTION_BUNDLE */ + [PERF_COUNT_HW_CACHE_REFERENCES] = TILE_OP_UNSUPP, + [PERF_COUNT_HW_CACHE_MISSES] = TILE_OP_UNSUPP, + [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = 0xd9, /* + COND_BRANCH_PRED_CORRECT */ + [PERF_COUNT_HW_BRANCH_MISSES] = 0xda, /* + COND_BRANCH_PRED_INCORRECT */ + [PERF_COUNT_HW_BUS_CYCLES] = TILE_OP_UNSUPP, +}; +#endif + +#define C(x) PERF_COUNT_HW_CACHE_##x + +/* + * Generalized hw caching related hw_event table, filled + * in on a per model basis. A value of -1 means + * 'not supported', any other value means the + * raw hw_event ID. + */ +#ifndef __tilegx__ +/* TILEPro hardware cache event map */ +static const int tile_cache_event_map[PERF_COUNT_HW_CACHE_MAX] + [PERF_COUNT_HW_CACHE_OP_MAX] + [PERF_COUNT_HW_CACHE_RESULT_MAX] = { +[C(L1D)] = { + [C(OP_READ)] = { + [C(RESULT_ACCESS)] = TILE_OP_UNSUPP, + [C(RESULT_MISS)] = 0x21, /* RD_MISS */ + }, + [C(OP_WRITE)] = { + [C(RESULT_ACCESS)] = TILE_OP_UNSUPP, + [C(RESULT_MISS)] = 0x22, /* WR_MISS */ + }, + [C(OP_PREFETCH)] = { + [C(RESULT_ACCESS)] = TILE_OP_UNSUPP, + [C(RESULT_MISS)] = TILE_OP_UNSUPP, + }, +}, +[C(L1I)] = { + [C(OP_READ)] = { + [C(RESULT_ACCESS)] = 0x12, /* MP_ICACHE_HIT_ISSUED */ + [C(RESULT_MISS)] = TILE_OP_UNSUPP, + }, + [C(OP_WRITE)] = { + [C(RESULT_ACCESS)] = TILE_OP_UNSUPP, + [C(RESULT_MISS)] = TILE_OP_UNSUPP, + }, + [C(OP_PREFETCH)] = { + [C(RESULT_ACCESS)] = TILE_OP_UNSUPP, + [C(RESULT_MISS)] = TILE_OP_UNSUPP, + }, +}, +[C(LL)] = { + [C(OP_READ)] = { + [C(RESULT_ACCESS)] = TILE_OP_UNSUPP, + [C(RESULT_MISS)] = TILE_OP_UNSUPP, + }, + [C(OP_WRITE)] = { + [C(RESULT_ACCESS)] = TILE_OP_UNSUPP, + [C(RESULT_MISS)] = TILE_OP_UNSUPP, + }, + [C(OP_PREFETCH)] = { + [C(RESULT_ACCESS)] = TILE_OP_UNSUPP, + [C(RESULT_MISS)] = TILE_OP_UNSUPP, + }, +}, +[C(DTLB)] = { + [C(OP_READ)] = { + [C(RESULT_ACCESS)] = 0x1d, /* TLB_CNT */ + [C(RESULT_MISS)] = 0x20, /* TLB_EXCEPTION */ + }, + [C(OP_WRITE)] = { + [C(RESULT_ACCESS)] = TILE_OP_UNSUPP, + [C(RESULT_MISS)] = TILE_OP_UNSUPP, + }, + [C(OP_PREFETCH)] = { + [C(RESULT_ACCESS)] = TILE_OP_UNSUPP, + [C(RESULT_MISS)] = TILE_OP_UNSUPP, + }, +}, +[C(ITLB)] = { + [C(OP_READ)] = { + [C(RESULT_ACCESS)] = 0x13, /* MP_ITLB_HIT_ISSUED */ + [C(RESULT_MISS)] = TILE_OP_UNSUPP, + }, + [C(OP_WRITE)] = { + [C(RESULT_ACCESS)] = TILE_OP_UNSUPP, + [C(RESULT_MISS)] = TILE_OP_UNSUPP, + }, + [C(OP_PREFETCH)] = { + [C(RESULT_ACCESS)] = TILE_OP_UNSUPP, + [C(RESULT_MISS)] = TILE_OP_UNSUPP, + }, +}, +[C(BPU)] = { + [C(OP_READ)] = { + [C(RESULT_ACCESS)] = TILE_OP_UNSUPP, + [C(RESULT_MISS)] = TILE_OP_UNSUPP, + }, + [C(OP_WRITE)] = { + [C(RESULT_ACCESS)] = TILE_OP_UNSUPP, + [C(RESULT_MISS)] = TILE_OP_UNSUPP, + }, + [C(OP_PREFETCH)] = { + [C(RESULT_ACCESS)] = TILE_OP_UNSUPP, + [C(RESULT_MISS)] = TILE_OP_UNSUPP, + }, +}, +}; +#else +/* TILEGx hardware events map */ +static const int tile_cache_event_map[PERF_COUNT_HW_CACHE_MAX] + [PERF_COUNT_HW_CACHE_OP_MAX] + [PERF_COUNT_HW_CACHE_RESULT_MAX] = { +[C(L1D)] = { + /* + * Like some other architectures (e.g. ARM), the performance + * counters don't differentiate between read and write + * accesses/misses, so this isn't strictly correct, but it's the + * best we can do. Writes and reads get combined. + */ + [C(OP_READ)] = { + [C(RESULT_ACCESS)] = TILE_OP_UNSUPP, + [C(RESULT_MISS)] = 0x44, /* RD_MISS */ + }, + [C(OP_WRITE)] = { + [C(RESULT_ACCESS)] = TILE_OP_UNSUPP, + [C(RESULT_MISS)] = 0x45, /* WR_MISS */ + }, + [C(OP_PREFETCH)] = { + [C(RESULT_ACCESS)] = TILE_OP_UNSUPP, + [C(RESULT_MISS)] = TILE_OP_UNSUPP, + }, +}, +[C(L1I)] = { + [C(OP_READ)] = { + [C(RESULT_ACCESS)] = TILE_OP_UNSUPP, + [C(RESULT_MISS)] = TILE_OP_UNSUPP, + }, + [C(OP_WRITE)] = { + [C(RESULT_ACCESS)] = TILE_OP_UNSUPP, + [C(RESULT_MISS)] = TILE_OP_UNSUPP, + }, + [C(OP_PREFETCH)] = { + [C(RESULT_ACCESS)] = TILE_OP_UNSUPP, + [C(RESULT_MISS)] = TILE_OP_UNSUPP, + }, +}, +[C(LL)] = { + [C(OP_READ)] = { + [C(RESULT_ACCESS)] = TILE_OP_UNSUPP, + [C(RESULT_MISS)] = TILE_OP_UNSUPP, + }, + [C(OP_WRITE)] = { + [C(RESULT_ACCESS)] = TILE_OP_UNSUPP, + [C(RESULT_MISS)] = TILE_OP_UNSUPP, + }, + [C(OP_PREFETCH)] = { + [C(RESULT_ACCESS)] = TILE_OP_UNSUPP, + [C(RESULT_MISS)] = TILE_OP_UNSUPP, + }, +}, +[C(DTLB)] = { + [C(OP_READ)] = { + [C(RESULT_ACCESS)] = 0x40, /* TLB_CNT */ + [C(RESULT_MISS)] = 0x43, /* TLB_EXCEPTION */ + }, + [C(OP_WRITE)] = { + [C(RESULT_ACCESS)] = 0x40, /* TLB_CNT */ + [C(RESULT_MISS)] = 0x43, /* TLB_EXCEPTION */ + }, + [C(OP_PREFETCH)] = { + [C(RESULT_ACCESS)] = TILE_OP_UNSUPP, + [C(RESULT_MISS)] = TILE_OP_UNSUPP, + }, +}, +[C(ITLB)] = { + [C(OP_READ)] = { + [C(RESULT_ACCESS)] = TILE_OP_UNSUPP, + [C(RESULT_MISS)] = 0xd4, /* ITLB_MISS_INT */ + }, + [C(OP_WRITE)] = { + [C(RESULT_ACCESS)] = TILE_OP_UNSUPP, + [C(RESULT_MISS)] = 0xd4, /* ITLB_MISS_INT */ + }, + [C(OP_PREFETCH)] = { + [C(RESULT_ACCESS)] = TILE_OP_UNSUPP, + [C(RESULT_MISS)] = TILE_OP_UNSUPP, + }, +}, +[C(BPU)] = { + [C(OP_READ)] = { + [C(RESULT_ACCESS)] = TILE_OP_UNSUPP, + [C(RESULT_MISS)] = TILE_OP_UNSUPP, + }, + [C(OP_WRITE)] = { + [C(RESULT_ACCESS)] = TILE_OP_UNSUPP, + [C(RESULT_MISS)] = TILE_OP_UNSUPP, + }, + [C(OP_PREFETCH)] = { + [C(RESULT_ACCESS)] = TILE_OP_UNSUPP, + [C(RESULT_MISS)] = TILE_OP_UNSUPP, + }, +}, +}; +#endif + +static atomic_t tile_active_events; +static DEFINE_MUTEX(perf_intr_reserve_mutex); + +static int tile_map_hw_event(u64 config); +static int tile_map_cache_event(u64 config); + +static int tile_pmu_handle_irq(struct pt_regs *regs, int fault); + +/* + * To avoid new_raw_count getting larger then pre_raw_count + * in tile_perf_event_update(), we limit the value of max_period to 2^31 - 1. + */ +static const struct tile_pmu tilepmu = { +#ifndef __tilegx__ + .name = "tilepro", +#else + .name = "tilegx", +#endif + .max_events = ARRAY_SIZE(tile_hw_event_map), + .map_hw_event = tile_map_hw_event, + .hw_events = tile_hw_event_map, + .map_cache_event = tile_map_cache_event, + .cache_events = &tile_cache_event_map, + .cntval_bits = 32, + .cntval_mask = (1ULL << 32) - 1, + .max_period = (1ULL << 31) - 1, + .num_counters = TILE_MAX_COUNTERS, + .num_base_counters = TILE_BASE_COUNTERS, +}; + +static const struct tile_pmu *tile_pmu __read_mostly; + +/* + * Check whether perf event is enabled. + */ +int tile_perf_enabled(void) +{ + return atomic_read(&tile_active_events) != 0; +} + +/* + * Read Performance Counters. + */ +static inline u64 read_counter(int idx) +{ + u64 val = 0; + + /* __insn_mfspr() only takes an immediate argument */ + switch (idx) { + case PERF_COUNT_0_IDX: + val = __insn_mfspr(SPR_PERF_COUNT_0); + break; + case PERF_COUNT_1_IDX: + val = __insn_mfspr(SPR_PERF_COUNT_1); + break; + case AUX_PERF_COUNT_0_IDX: + val = __insn_mfspr(SPR_AUX_PERF_COUNT_0); + break; + case AUX_PERF_COUNT_1_IDX: + val = __insn_mfspr(SPR_AUX_PERF_COUNT_1); + break; + default: + WARN_ON_ONCE(idx > AUX_PERF_COUNT_1_IDX || + idx < PERF_COUNT_0_IDX); + } + + return val; +} + +/* + * Write Performance Counters. + */ +static inline void write_counter(int idx, u64 value) +{ + /* __insn_mtspr() only takes an immediate argument */ + switch (idx) { + case PERF_COUNT_0_IDX: + __insn_mtspr(SPR_PERF_COUNT_0, value); + break; + case PERF_COUNT_1_IDX: + __insn_mtspr(SPR_PERF_COUNT_1, value); + break; + case AUX_PERF_COUNT_0_IDX: + __insn_mtspr(SPR_AUX_PERF_COUNT_0, value); + break; + case AUX_PERF_COUNT_1_IDX: + __insn_mtspr(SPR_AUX_PERF_COUNT_1, value); + break; + default: + WARN_ON_ONCE(idx > AUX_PERF_COUNT_1_IDX || + idx < PERF_COUNT_0_IDX); + } +} + +/* + * Enable performance event by setting + * Performance Counter Control registers. + */ +static inline void tile_pmu_enable_event(struct perf_event *event) +{ + struct hw_perf_event *hwc = &event->hw; + unsigned long cfg, mask; + int shift, idx = hwc->idx; + + /* + * prevent early activation from tile_pmu_start() in hw_perf_enable + */ + + if (WARN_ON_ONCE(idx == -1)) + return; + + if (idx < tile_pmu->num_base_counters) + cfg = __insn_mfspr(SPR_PERF_COUNT_CTL); + else + cfg = __insn_mfspr(SPR_AUX_PERF_COUNT_CTL); + + switch (idx) { + case PERF_COUNT_0_IDX: + case AUX_PERF_COUNT_0_IDX: + mask = TILE_EVENT_MASK; + shift = 0; + break; + case PERF_COUNT_1_IDX: + case AUX_PERF_COUNT_1_IDX: + mask = TILE_EVENT_MASK << 16; + shift = 16; + break; + default: + WARN_ON_ONCE(idx < PERF_COUNT_0_IDX || + idx > AUX_PERF_COUNT_1_IDX); + return; + } + + /* Clear mask bits to enable the event. */ + cfg &= ~mask; + cfg |= hwc->config << shift; + + if (idx < tile_pmu->num_base_counters) + __insn_mtspr(SPR_PERF_COUNT_CTL, cfg); + else + __insn_mtspr(SPR_AUX_PERF_COUNT_CTL, cfg); +} + +/* + * Disable performance event by clearing + * Performance Counter Control registers. + */ +static inline void tile_pmu_disable_event(struct perf_event *event) +{ + struct hw_perf_event *hwc = &event->hw; + unsigned long cfg, mask; + int idx = hwc->idx; + + if (idx == -1) + return; + + if (idx < tile_pmu->num_base_counters) + cfg = __insn_mfspr(SPR_PERF_COUNT_CTL); + else + cfg = __insn_mfspr(SPR_AUX_PERF_COUNT_CTL); + + switch (idx) { + case PERF_COUNT_0_IDX: + case AUX_PERF_COUNT_0_IDX: + mask = TILE_PLM_MASK; + break; + case PERF_COUNT_1_IDX: + case AUX_PERF_COUNT_1_IDX: + mask = TILE_PLM_MASK << 16; + break; + default: + WARN_ON_ONCE(idx < PERF_COUNT_0_IDX || + idx > AUX_PERF_COUNT_1_IDX); + return; + } + + /* Set mask bits to disable the event. */ + cfg |= mask; + + if (idx < tile_pmu->num_base_counters) + __insn_mtspr(SPR_PERF_COUNT_CTL, cfg); + else + __insn_mtspr(SPR_AUX_PERF_COUNT_CTL, cfg); +} + +/* + * Propagate event elapsed time into the generic event. + * Can only be executed on the CPU where the event is active. + * Returns the delta events processed. + */ +static u64 tile_perf_event_update(struct perf_event *event) +{ + struct hw_perf_event *hwc = &event->hw; + int shift = 64 - tile_pmu->cntval_bits; + u64 prev_raw_count, new_raw_count; + u64 oldval; + int idx = hwc->idx; + u64 delta; + + /* + * Careful: an NMI might modify the previous event value. + * + * Our tactic to handle this is to first atomically read and + * exchange a new raw count - then add that new-prev delta + * count to the generic event atomically: + */ +again: + prev_raw_count = local64_read(&hwc->prev_count); + new_raw_count = read_counter(idx); + + oldval = local64_cmpxchg(&hwc->prev_count, prev_raw_count, + new_raw_count); + if (oldval != prev_raw_count) + goto again; + + /* + * Now we have the new raw value and have updated the prev + * timestamp already. We can now calculate the elapsed delta + * (event-)time and add that to the generic event. + * + * Careful, not all hw sign-extends above the physical width + * of the count. + */ + delta = (new_raw_count << shift) - (prev_raw_count << shift); + delta >>= shift; + + local64_add(delta, &event->count); + local64_sub(delta, &hwc->period_left); + + return new_raw_count; +} + +/* + * Set the next IRQ period, based on the hwc->period_left value. + * To be called with the event disabled in hw: + */ +static int tile_event_set_period(struct perf_event *event) +{ + struct hw_perf_event *hwc = &event->hw; + int idx = hwc->idx; + s64 left = local64_read(&hwc->period_left); + s64 period = hwc->sample_period; + int ret = 0; + + /* + * If we are way outside a reasonable range then just skip forward: + */ + if (unlikely(left <= -period)) { + left = period; + local64_set(&hwc->period_left, left); + hwc->last_period = period; + ret = 1; + } + + if (unlikely(left <= 0)) { + left += period; + local64_set(&hwc->period_left, left); + hwc->last_period = period; + ret = 1; + } + if (left > tile_pmu->max_period) + left = tile_pmu->max_period; + + /* + * The hw event starts counting from this event offset, + * mark it to be able to extra future deltas: + */ + local64_set(&hwc->prev_count, (u64)-left); + + write_counter(idx, (u64)(-left) & tile_pmu->cntval_mask); + + perf_event_update_userpage(event); + + return ret; +} + +/* + * Stop the event but do not release the PMU counter + */ +static void tile_pmu_stop(struct perf_event *event, int flags) +{ + struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events); + struct hw_perf_event *hwc = &event->hw; + int idx = hwc->idx; + + if (__test_and_clear_bit(idx, cpuc->active_mask)) { + tile_pmu_disable_event(event); + cpuc->events[hwc->idx] = NULL; + WARN_ON_ONCE(hwc->state & PERF_HES_STOPPED); + hwc->state |= PERF_HES_STOPPED; + } + + if ((flags & PERF_EF_UPDATE) && !(hwc->state & PERF_HES_UPTODATE)) { + /* + * Drain the remaining delta count out of a event + * that we are disabling: + */ + tile_perf_event_update(event); + hwc->state |= PERF_HES_UPTODATE; + } +} + +/* + * Start an event (without re-assigning counter) + */ +static void tile_pmu_start(struct perf_event *event, int flags) +{ + struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events); + int idx = event->hw.idx; + + if (WARN_ON_ONCE(!(event->hw.state & PERF_HES_STOPPED))) + return; + + if (WARN_ON_ONCE(idx == -1)) + return; + + if (flags & PERF_EF_RELOAD) { + WARN_ON_ONCE(!(event->hw.state & PERF_HES_UPTODATE)); + tile_event_set_period(event); + } + + event->hw.state = 0; + + cpuc->events[idx] = event; + __set_bit(idx, cpuc->active_mask); + + unmask_pmc_interrupts(); + + tile_pmu_enable_event(event); + + perf_event_update_userpage(event); +} + +/* + * Add a single event to the PMU. + * + * The event is added to the group of enabled events + * but only if it can be scehduled with existing events. + */ +static int tile_pmu_add(struct perf_event *event, int flags) +{ + struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events); + struct hw_perf_event *hwc; + unsigned long mask; + int b, max_cnt; + + hwc = &event->hw; + + /* + * We are full. + */ + if (cpuc->n_events == tile_pmu->num_counters) + return -ENOSPC; + + cpuc->event_list[cpuc->n_events] = event; + cpuc->n_events++; + + hwc->state = PERF_HES_UPTODATE | PERF_HES_STOPPED; + if (!(flags & PERF_EF_START)) + hwc->state |= PERF_HES_ARCH; + + /* + * Find first empty counter. + */ + max_cnt = tile_pmu->num_counters; + mask = ~cpuc->used_mask; + + /* Find next free counter. */ + b = find_next_bit(&mask, max_cnt, 0); + + /* Should not happen. */ + if (WARN_ON_ONCE(b == max_cnt)) + return -ENOSPC; + + /* + * Assign counter to event. + */ + event->hw.idx = b; + __set_bit(b, &cpuc->used_mask); + + /* + * Start if requested. + */ + if (flags & PERF_EF_START) + tile_pmu_start(event, PERF_EF_RELOAD); + + return 0; +} + +/* + * Delete a single event from the PMU. + * + * The event is deleted from the group of enabled events. + * If it is the last event, disable PMU interrupt. + */ +static void tile_pmu_del(struct perf_event *event, int flags) +{ + struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events); + int i; + + /* + * Remove event from list, compact list if necessary. + */ + for (i = 0; i < cpuc->n_events; i++) { + if (cpuc->event_list[i] == event) { + while (++i < cpuc->n_events) + cpuc->event_list[i-1] = cpuc->event_list[i]; + --cpuc->n_events; + cpuc->events[event->hw.idx] = NULL; + __clear_bit(event->hw.idx, &cpuc->used_mask); + tile_pmu_stop(event, PERF_EF_UPDATE); + break; + } + } + /* + * If there are no events left, then mask PMU interrupt. + */ + if (cpuc->n_events == 0) + mask_pmc_interrupts(); + perf_event_update_userpage(event); +} + +/* + * Propagate event elapsed time into the event. + */ +static inline void tile_pmu_read(struct perf_event *event) +{ + tile_perf_event_update(event); +} + +/* + * Map generic events to Tile PMU. + */ +static int tile_map_hw_event(u64 config) +{ + if (config >= tile_pmu->max_events) + return -EINVAL; + return tile_pmu->hw_events[config]; +} + +/* + * Map generic hardware cache events to Tile PMU. + */ +static int tile_map_cache_event(u64 config) +{ + unsigned int cache_type, cache_op, cache_result; + int code; + + if (!tile_pmu->cache_events) + return -ENOENT; + + cache_type = (config >> 0) & 0xff; + if (cache_type >= PERF_COUNT_HW_CACHE_MAX) + return -EINVAL; + + cache_op = (config >> 8) & 0xff; + if (cache_op >= PERF_COUNT_HW_CACHE_OP_MAX) + return -EINVAL; + + cache_result = (config >> 16) & 0xff; + if (cache_result >= PERF_COUNT_HW_CACHE_RESULT_MAX) + return -EINVAL; + + code = (*tile_pmu->cache_events)[cache_type][cache_op][cache_result]; + if (code == TILE_OP_UNSUPP) + return -EINVAL; + + return code; +} + +static void tile_event_destroy(struct perf_event *event) +{ + if (atomic_dec_return(&tile_active_events) == 0) + release_pmc_hardware(); +} + +static int __tile_event_init(struct perf_event *event) +{ + struct perf_event_attr *attr = &event->attr; + struct hw_perf_event *hwc = &event->hw; + int code; + + switch (attr->type) { + case PERF_TYPE_HARDWARE: + code = tile_pmu->map_hw_event(attr->config); + break; + case PERF_TYPE_HW_CACHE: + code = tile_pmu->map_cache_event(attr->config); + break; + case PERF_TYPE_RAW: + code = attr->config & TILE_EVENT_MASK; + break; + default: + /* Should not happen. */ + return -EOPNOTSUPP; + } + + if (code < 0) + return code; + + hwc->config = code; + hwc->idx = -1; + + if (attr->exclude_user) + hwc->config |= TILE_CTL_EXCL_USER; + + if (attr->exclude_kernel) + hwc->config |= TILE_CTL_EXCL_KERNEL; + + if (attr->exclude_hv) + hwc->config |= TILE_CTL_EXCL_HV; + + if (!hwc->sample_period) { + hwc->sample_period = tile_pmu->max_period; + hwc->last_period = hwc->sample_period; + local64_set(&hwc->period_left, hwc->sample_period); + } + event->destroy = tile_event_destroy; + return 0; +} + +static int tile_event_init(struct perf_event *event) +{ + int err = 0; + perf_irq_t old_irq_handler = NULL; + + if (atomic_inc_return(&tile_active_events) == 1) + old_irq_handler = reserve_pmc_hardware(tile_pmu_handle_irq); + + if (old_irq_handler) { + pr_warn("PMC hardware busy (reserved by oprofile)\n"); + + atomic_dec(&tile_active_events); + return -EBUSY; + } + + switch (event->attr.type) { + case PERF_TYPE_RAW: + case PERF_TYPE_HARDWARE: + case PERF_TYPE_HW_CACHE: + break; + + default: + return -ENOENT; + } + + err = __tile_event_init(event); + if (err) { + if (event->destroy) + event->destroy(event); + } + return err; +} + +static struct pmu tilera_pmu = { + .event_init = tile_event_init, + .add = tile_pmu_add, + .del = tile_pmu_del, + + .start = tile_pmu_start, + .stop = tile_pmu_stop, + + .read = tile_pmu_read, +}; + +/* + * PMU's IRQ handler, PMU has 2 interrupts, they share the same handler. + */ +int tile_pmu_handle_irq(struct pt_regs *regs, int fault) +{ + struct perf_sample_data data; + struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events); + struct perf_event *event; + struct hw_perf_event *hwc; + u64 val; + unsigned long status; + int bit; + + __get_cpu_var(perf_irqs)++; + + if (!atomic_read(&tile_active_events)) + return 0; + + status = pmc_get_overflow(); + pmc_ack_overflow(status); + + for_each_set_bit(bit, &status, tile_pmu->num_counters) { + + event = cpuc->events[bit]; + + if (!event) + continue; + + if (!test_bit(bit, cpuc->active_mask)) + continue; + + hwc = &event->hw; + + val = tile_perf_event_update(event); + if (val & (1ULL << (tile_pmu->cntval_bits - 1))) + continue; + + perf_sample_data_init(&data, 0, event->hw.last_period); + if (!tile_event_set_period(event)) + continue; + + if (perf_event_overflow(event, &data, regs)) + tile_pmu_stop(event, 0); + } + + return 0; +} + +static bool __init supported_pmu(void) +{ + tile_pmu = &tilepmu; + return true; +} + +int __init init_hw_perf_events(void) +{ + supported_pmu(); + perf_pmu_register(&tilera_pmu, "cpu", PERF_TYPE_RAW); + return 0; +} +arch_initcall(init_hw_perf_events); + +/* Callchain handling code. */ + +/* + * Tile specific backtracing code for perf_events. + */ +static inline void perf_callchain(struct perf_callchain_entry *entry, + struct pt_regs *regs) +{ + struct KBacktraceIterator kbt; + unsigned int i; + + /* + * Get the address just after the "jalr" instruction that + * jumps to the handler for a syscall. When we find this + * address in a backtrace, we silently ignore it, which gives + * us a one-step backtrace connection from the sys_xxx() + * function in the kernel to the xxx() function in libc. + * Otherwise, we lose the ability to properly attribute time + * from the libc calls to the kernel implementations, since + * oprofile only considers PCs from backtraces a pair at a time. + */ + unsigned long handle_syscall_pc = handle_syscall_link_address(); + + KBacktraceIterator_init(&kbt, NULL, regs); + kbt.profile = 1; + + /* + * The sample for the pc is already recorded. Now we are adding the + * address of the callsites on the stack. Our iterator starts + * with the frame of the (already sampled) call site. If our + * iterator contained a "return address" field, we could have just + * used it and wouldn't have needed to skip the first + * frame. That's in effect what the arm and x86 versions do. + * Instead we peel off the first iteration to get the equivalent + * behavior. + */ + + if (KBacktraceIterator_end(&kbt)) + return; + KBacktraceIterator_next(&kbt); + + /* + * Set stack depth to 16 for user and kernel space respectively, that + * is, total 32 stack frames. + */ + for (i = 0; i < 16; ++i) { + unsigned long pc; + if (KBacktraceIterator_end(&kbt)) + break; + pc = kbt.it.pc; + if (pc != handle_syscall_pc) + perf_callchain_store(entry, pc); + KBacktraceIterator_next(&kbt); + } +} + +void perf_callchain_user(struct perf_callchain_entry *entry, + struct pt_regs *regs) +{ + perf_callchain(entry, regs); +} + +void perf_callchain_kernel(struct perf_callchain_entry *entry, + struct pt_regs *regs) +{ + perf_callchain(entry, regs); +} -- cgit v1.1 From 603eb2cdee10d1006befa07ff62b6c63129df6c1 Mon Sep 17 00:00:00 2001 From: Kerry Sheh Date: Wed, 2 Apr 2014 10:45:35 +0800 Subject: tile: Fix vDSO compilation issue with allyesconfig make allyesconfig give the following build error on tile: tilegx-linux-gcc: error: arch/tile/kernel/vdso/vgettimeofday32.o: No such file or directory tilegx-linux-objcopy: 'arch/tile/kernel/vdso/vdso32.so.dbg': No such file or directory In case with CONFIG_MODVERSIONS, cmd_cc_o_c generate .tmp_.o from .c only. Fix it by execute rule_cc_o_c instead. Reported-by: Fengguang Wu Signed-off-by: Kerry Sheh Signed-off-by: Chris Metcalf --- arch/tile/kernel/vdso/Makefile | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'arch') diff --git a/arch/tile/kernel/vdso/Makefile b/arch/tile/kernel/vdso/Makefile index e2b7a2f..a025f63 100644 --- a/arch/tile/kernel/vdso/Makefile +++ b/arch/tile/kernel/vdso/Makefile @@ -104,7 +104,7 @@ $(obj-vdso32:%=%): KBUILD_AFLAGS = $(KBUILD_AFLAGS_32) $(obj-vdso32:%=%): KBUILD_CFLAGS = $(KBUILD_CFLAGS_32) $(obj)/vgettimeofday32.o: $(obj)/vgettimeofday.c - $(call if_changed,cc_o_c) + $(call if_changed_rule,cc_o_c) $(obj)/vrt_sigreturn32.o: $(obj)/vrt_sigreturn.S $(call if_changed,as_o_S) -- cgit v1.1 From 5eb0bdf84433eb7b7ad4ba92a80aac57ad4b46ea Mon Sep 17 00:00:00 2001 From: Chris Metcalf Date: Fri, 4 Apr 2014 13:11:40 -0400 Subject: arch/tile: remove unused variable 'devcap' Commit 503275bf37 removed the use of the variable but not the variable itself. Signed-off-by: Chris Metcalf --- arch/tile/kernel/pci.c | 2 -- 1 file changed, 2 deletions(-) (limited to 'arch') diff --git a/arch/tile/kernel/pci.c b/arch/tile/kernel/pci.c index c45593d..1f80a88 100644 --- a/arch/tile/kernel/pci.c +++ b/arch/tile/kernel/pci.c @@ -250,8 +250,6 @@ static void fixup_read_and_payload_sizes(void) /* Scan for the smallest maximum payload size. */ for_each_pci_dev(dev) { - u32 devcap; - if (!pci_is_pcie(dev)) continue; -- cgit v1.1