perf/x86: Add ability to sample TSC
Subsystems (2)
Adrian Hunter
reviewerv2 cover perf/x86: Add ability to sample TSC
Hi
TSC is needed to synchronize Intel Processor Trace (Intel PT) with perf event samples. Refer to patch 1 for more details.
There is a description of Intel PT in the Intel Architecture manuals:
http://www.intel.com/content/www/us/en/processors/architectures-software-developer-manuals.html
With the advent of switching perf_clock to CLOCK_MONOTONIC, it will not be possible to convert perf_clock directly to/from TSC. So add the ability to sample TSC instead.
Changes in V2:
perf: Sample additional clock value
Rename "Architecture specific clock" to "Processor trace clock"
Rename PERF_SAMPLE_CLOCK_ARCH -> PERF_SAMPLE_CLOCK_PT etc Expand commit message
perf/x86: Provide TSC for PERF_SAMPLE_CLOCK_PT
Rename PERF_SAMPLE_CLOCK_ARCH -> PERF_SAMPLE_CLOCK_PT etc Expand commit message
Adrian Hunter (2):
perf: Sample additional clock value perf/x86: Provide TSC for PERF_SAMPLE_CLOCK_PT
arch/x86/include/asm/perf_event.h | 6 ++++++ arch/x86/kernel/cpu/perf_event.c | 10 ++++++++++ include/linux/perf_event.h | 3 ++- include/uapi/linux/perf_event.h | 19 +++++++++++++++++-- kernel/events/core.c | 30 ++++++++++++++++++++++++++++++ kernel/events/internal.h | 4 ++++ 6 files changed, 69 insertions(+), 3 deletions(-)
Regards
Adrian
Adrian Hunter
reviewerPATCH v2 1/2 perf: Sample additional clock value
This is needed to allow perf event samples to be synchronized with data from other sources, and in particular, sources like Intel Processor Trace (Intel PT) where the hardware produces a trace with hardware defined timestamps (i.e. TSC).
For example, to decode an Intel PT trace, the decoder must walk the object code. To determine what object code is running, the decoder must track events like sched_switch and MMAP and match them against the trace data using the timestamps.
Note that it is not the accuracy of the time sources that is at issue but instead the ability to correctly order events.
On modern machines, perf_clock is currently directly related to TSC, however that is to change when perf_clock becomes CLOCK_MONOTONIC.
Consequently add PERF_SAMPLE_CLOCK to sample some other clock. The patch allows for 16 possible clock selections with the only initial possibility a processor trace clock which will be TSC on x86.
Although there are only 16 possible clock selections, it is envisioned that POSIX clock ids would be a single selection, with the actual clock id provided in another perf_event_attr member.
--- include/linux/perf_event.h | 3 ++- include/uapi/linux/perf_event.h | 19 +++++++++++++++++-- kernel/events/core.c | 30 ++++++++++++++++++++++++++++++ kernel/events/internal.h | 4 ++++ 4 files changed, 53 insertions(+), 3 deletions(-)
include/linux/perf_event.h +2 -1
@@ -655,7 +655,7 @@ extern void perf_pmu_migrate_context(struct pmu *pmu, int src_cpu, int dst_cpu);extern u64 perf_event_read_value (struct perf_event *event,u64 *enabled,u64 *running); - +u64 perf_sample_clock_pt (void );struct perf_sample_data {/*
@@ -687,6 +687,7 @@ struct perf_sample_data {u32 cpu;u32 reserved; } cpu_entry; +u64 clock;struct perf_callchain_entry *callchain;/*
include/uapi/linux/perf_event.h +17 -2
@@ -139,8 +139,9 @@ enum perf_event_sample_format {PERF_SAMPLE_IDENTIFIER =1U <<16 ,PERF_SAMPLE_TRANSACTION =1U <<17 ,PERF_SAMPLE_REGS_INTR =1U <<18 , +PERF_SAMPLE_CLOCK =1U <<19 , - PERF_SAMPLE_MAX = 1U << 19, /* non-ABI */ + PERF_SAMPLE_MAX = 1U << 20, /* non-ABI */ };/*
@@ -228,6 +229,16 @@ enum { };/* +* Values to determine clock to sample. +*/ +enum perf_sample_clock_type { +/* Processor trace clock (TSC on x86) */ +PERF_SAMPLE_CLOCK_PT =0 , + +PERF_SAMPLE_CLOCK_MAX /* non-ABI */ +}; + +/* * The format of the data returned by read() on a perf event fd, * as specified by attr.read_format: *
@@ -328,7 +339,9 @@ struct perf_event_attr { exclude_callchain_user : 1, /* exclude user callchains */ mmap2 : 1, /* include mmap with inode data */ comm_exec : 1, /* flag comm events that are due to an exec */ - __reserved_1 : 39; + /* clock: see enum perf_sample_clock_type */ + clock : 4, /* which clock */ + __reserved_1 : 35;union {__u32 wakeup_events;/* wakeup every n events */
@@ -601,6 +614,7 @@ enum perf_event_type {* { u64 id; } && PERF_SAMPLE_ID * { u64 stream_id;} && PERF_SAMPLE_STREAM_ID * { u32 cpu, res; } && PERF_SAMPLE_CPU +* { u64 clock; } && PERF_SAMPLE_CLOCK * { u64 id; } && PERF_SAMPLE_IDENTIFIER * } && perf_event_attr::sample_id_all *
@@ -746,6 +760,7 @@ enum perf_event_type {* { u64 transaction; } && PERF_SAMPLE_TRANSACTION * { u64 abi; # enum perf_sample_regs_abi * u64 regs[weight(mask)]; } && PERF_SAMPLE_REGS_INTR +* { u64 clock; } && PERF_SAMPLE_CLOCK * }; */ PERF_RECORD_SAMPLE = 9,
kernel/events/core.c +30 -0
@@ -1323,6 +1323,9 @@ static void perf_event__id_header_size(struct perf_event *event)if (sample_type &PERF_SAMPLE_CPU ) size +=sizeof (data->cpu_entry); +if (sample_type &PERF_SAMPLE_CLOCK ) + size +=sizeof (data->clock); + event->id_header_size = size; }
@@ -4915,6 +4918,11 @@ perf_output_sample_ustack(struct perf_output_handle *handle, u64 dump_size, } } +u64 __weak perf_sample_clock_pt (void ) +{ +return 0 ; +} +static void __perf_event_header__init_id (struct perf_event_header *header,struct perf_sample_data *data,struct perf_event *event)
@@ -4943,6 +4951,16 @@ static void __perf_event_header__init_id(struct perf_event_header *header, data->cpu_entry.cpu =raw_smp_processor_id (); data->cpu_entry.reserved =0 ; } + +if (sample_type &PERF_SAMPLE_CLOCK ) { +switch (event->attr.clock) { +case PERF_SAMPLE_CLOCK_PT : + data->clock =perf_sample_clock_pt (); +break ; +default : + data->clock =0 ; + } + } } void perf_event_header__init_id(struct perf_event_header *header,
@@ -4973,6 +4991,9 @@ static void __perf_event__output_id_sample(struct perf_output_handle *handle,if (sample_type &PERF_SAMPLE_CPU )perf_output_put (handle, data->cpu_entry); +if (sample_type &PERF_SAMPLE_CLOCK ) +perf_output_put (handle, data->clock); +if (sample_type &PERF_SAMPLE_IDENTIFIER )perf_output_put (handle, data->id); }
@@ -5218,6 +5239,9 @@ void perf_output_sample(struct perf_output_handle *handle, } } +if (sample_type &PERF_SAMPLE_CLOCK ) +perf_output_put (handle, data->clock); +if (!event->attr.watermark) {int wakeup_events = event->attr.wakeup_events;
@@ -7632,6 +7656,12 @@ static int perf_copy_attr(struct perf_event_attr __user *uattr,if (attr->sample_type &PERF_SAMPLE_REGS_INTR ) ret =perf_reg_validate (attr->sample_regs_intr); + +if ((attr->sample_type &PERF_SAMPLE_CLOCK ) && + (attr->clock >=PERF_SAMPLE_CLOCK_MAX || + (!HAVE_PERF_SAMPLE_CLOCK_PT && + attr->clock ==PERF_SAMPLE_CLOCK_PT ))) +return -EINVAL ; out:return ret;
kernel/events/internal.h +4 -0
@@ -228,4 +228,8 @@ static inline bool arch_perf_have_user_stack_dump(void)#define perf_user_stack_pointer (regs) 0#endif /* CONFIG_HAVE_PERF_USER_STACK_DUMP */ +#ifndef HAVE_PERF_SAMPLE_CLOCK_PT +#define HAVE_PERF_SAMPLE_CLOCK_PT 0 +#endif +#endif /* _KERNEL_EVENTS_INTERNAL_H */
Signature · 1 line
1.9.1
Adrian Hunter
reviewerPATCH v2 2/2 perf/x86: Provide TSC for PERF_SAMPLE_CLOCK_PT
In reply to Adrian HunterProvide TSC for PERF_SAMPLE_CLOCK_PT. This is needed to match perf events against hardware traces like Intel Processor Trace (Intel PT) which is the purpose for which PERF_SAMPLE_CLOCK_PT was invented.
--- arch/x86/include/asm/perf_event.h | 6 ++++++ arch/x86/kernel/cpu/perf_event.c | 10 ++++++++++ 2 files changed, 16 insertions(+)
arch/x86/include/asm/perf_event.h +6 -0
@@ -261,6 +261,12 @@ struct perf_guest_switch_msr {extern struct perf_guest_switch_msr *perf_guest_get_msrs (int *nr);extern void perf_get_x86_pmu_capability (struct x86_pmu_capability *cap);extern void perf_check_microcode (void ); + +#ifdef CONFIG_X86_TSC +#define HAVE_PERF_SAMPLE_CLOCK_PT 1 +u64 perf_sample_clock_pt (void ); +#endif +#else static inline struct perf_guest_switch_msr *perf_guest_get_msrs (int *nr) {
arch/x86/kernel/cpu/perf_event.c +10 -0
@@ -2256,3 +2256,13 @@ void perf_get_x86_pmu_capability(struct x86_pmu_capability *cap) cap->events_mask_len = x86_pmu.events_mask_len; }EXPORT_SYMBOL_GPL (perf_get_x86_pmu_capability); + +#ifdef CONFIG_X86_TSC +u64 perf_sample_clock_pt (void ) +{ +u64 tsc; + +rdtscll (tsc); +return tsc; +} +#endif
Signature · 1 line
1.9.1
Re: [PATCH V2 1/2] perf: Sample additional clock value
In reply to Adrian Hunterinclude/linux/perf_event.h extern void perf_pmu_migrate_context(struct pmu *pmu, · line 658
Quoted 10 lines · Adrian Hunter
include/linux/perf_event.h +1
-1@@ -655,7 +655,7 @@ extern void perf_pmu_migrate_context(struct pmu *pmu, int src_cpu, int dst_cpu);extern u64 perf_event_read_value (struct perf_event *event,u64 *enabled,u64 *running); - +u64 perf_sample_clock_pt (void );
Core functions should not be arch specific. PT == x86.
Quoted 18 lines
include/linux/perf_event.h +7
-0@@ -4915,6 +4918,11 @@ perf_output_sample_ustack(struct perf_output_handle *handle, u64 dump_size, } } +u64 __weak perf_sample_clock_pt (void ) +{ +return 0 ; +} +static void __perf_event_header__init_id (struct perf_event_header *header,struct perf_sample_data *data,struct perf_event *event)@@ -4943,6 +4951,16 @@ static void __perf_event_header__init_id(struct perf_event_header *header, data->cpu_entry.cpu =raw_smp_processor_id (); data->cpu_entry.reserved =0 ; } + +if (sample_type &PERF_SAMPLE_CLOCK ) {
include/linux/perf_event.h static void __perf_event_header__init_id(struct perf_event_header *header, · line 4965
Quoted 10 lines
include/linux/perf_event.h +8
-0+switch (event->attr.clock) { +case PERF_SAMPLE_CLOCK_PT : + data->clock =perf_sample_clock_pt (); +break ; +default : + data->clock =0 ; + } + } }
Ditto here. This should be transparent for arch specific clocks.
David
Adrian Hunter
reviewerRe: [PATCH V2 1/2] perf: Sample additional clock value
Quoted 13 lines · David Ahern
include/linux/perf_event.h extern void perf_pmu_migrate_context(struct pmu *pmu, · line 658
Quoted 10 lines · Adrian Hunter
include/linux/perf_event.h +1
-1@@ -655,7 +655,7 @@ extern void perf_pmu_migrate_context(struct pmu *pmu, int src_cpu, int dst_cpu);extern u64 perf_event_read_value (struct perf_event *event,u64 *enabled,u64 *running); - +u64 perf_sample_clock_pt (void );Core functions should not be arch specific. PT == x86.
Actually it was one of the ARM guys that asked for it to be called "Processor Trace". It was "arch" in V1.
http://marc.info/?l=linux-kernel&m=142436891806015
But it has been superseded completely by patches from Peter, so it is not going further anyway.