perf: add support for capturing skid IP
Subsystems (2)
v3 cover perf: add support for capturing skid IP
This patchs adds a new sample record type called PERF_SAMPLE_SKID_IP. The goal is to record the unmodified interrupted instruction pointer (IP) as seen by the kernel and reflected in the machine state.
On some architectures, it is possible to avoid the IP skid using hardware support. For instance, on Intel x86, the use of PEBS helps eliminate the skid on Haswell and later processors. On older Intel processor, software, i.e., the kernel, may succeed in eliminating the skid.
Without this patch, on Haswell processors, if you set:
- attr.precise = 0, then you get the skid IP
- attr.precise = 1, then you get the skid PEBS ip (off-by-1)
- attr.precise = 2, then you get the skidless PEBS ip
The IP is captured when the event has PERF_SAMPLE_IP set in sample_type. However, there are certain measurements where you need to have BOTH the skidless IP and the skid IP. For instance, when studying branches, the skid IP usually points to the target of the branch while the skidless IP points to the branch instruction itself. Today, it is not possible to retrieve both at the same time. This patch makes this possible by specifying PERF_SAMPLE_IP|PERF_SAMPLE_SKID_IP.
As an example, consider the following code snipet:
37.51 42c2ed je 42c2f3
42c2ef add $0x1,%rdx
42c2f3 sub $0x1,%raxWhen using PEBS (precise=2) and sampling on BR_INST_RETIRED.CONDITIONAL, the IP always points to 0x42c2ed. With precise=1, the IP would point to 0x42c2f3. It is interesting to collect both IPs in a single run to determine how often the conditional branch is taken vs. non-taken.
Understanding the skid is also interesting for other precise events.
In V2, we rebased to 10d94ff4d558 (v4.14-rc7).
In V3, code is rebased to 4.14-rc8, LKML comments have been integrated. The new way to specify skid ip is per event:
$ perf record -e cpu/event=0xc5,skid-ip=1/ ....
Stephane Eranian (5):
perf/core: add PERF_RECORD_SAMPLE_SKID_IP record type perf/x86: add PERF_SAMPLE_SKID_IP support for X86 PEBS perf/tools: add support for PERF_SAMPLE_SKID_IP perf/record: add documentation for sampling skid ip perf/script: add support for skid ip
arch/x86/events/intel/ds.c | 7 +++++++ include/linux/perf_event.h | 2 ++ include/uapi/linux/perf_event.h | 4 +++- kernel/events/core.c | 14 ++++++++++++++ tools/include/uapi/linux/perf_event.h | 4 +++- tools/perf/Documentation/perf-record.txt | 8 ++++++++ tools/perf/Documentation/perf-script.txt | 2 +- tools/perf/builtin-script.c | 10 ++++++++-- tools/perf/util/event.h | 1 + tools/perf/util/evsel.c | 11 +++++++++++ tools/perf/util/evsel.h | 2 ++ tools/perf/util/parse-events.c | 7 +++++++ tools/perf/util/parse-events.h | 1 + tools/perf/util/parse-events.l | 1 + tools/perf/util/session.c | 3 +++ 15 files changed, 72 insertions(+), 5 deletions(-)
Signature · 1 line
2.7.4
PATCH v3 1/5 perf/core: add PERF_RECORD_SAMPLE_SKID_IP record type
This patchs adds a new sample record type. The goal
is to record the interrupted instruction pointer (IP)
as seen by the kernel and reflected in the machine state (pt_regs).
On some architectures, it is possible to avoid the IP skid using hardware support. For instance, on Intel x86, the use of PEBS helps eliminate the skid on Haswell and later processors.
Without this patch, on Haswell processors, if you set:
- attr.precise = 0, then you get the skid IP
- attr.precise > 0, then you get the PEBS ip corrected for skid
The IP normally comes when the event has PERF_RECORD_SAMPLE_IP set. However, there are certain measuremewnts where you need to have BOTH the corrected IP and the skid IP. For instance, when studying branches, the skid IP usually points to the target of the branch while the corrected IP point to the branch instruction itself. Today, it is not possible to retrieve both at the same time. This patch makes this possible by specifying PERF_SAMPLE_IP|PERF_SAMPLE_SKID_IP.
--- include/linux/perf_event.h | 2 ++ include/uapi/linux/perf_event.h | 4 +++- kernel/events/core.c | 14 ++++++++++++++ 3 files changed, 19 insertions(+), 1 deletion(-)
include/linux/perf_event.h +2 -0
@@ -917,6 +917,7 @@ struct perf_sample_data {u64 stack_user_size;u64 phys_addr; +u64 skid_ip; }____cacheline_aligned ;/* default value for data source */
@@ -937,6 +938,7 @@ static inline void perf_sample_data_init(struct perf_sample_data *data, data->weight =0 ; data->data_src.val =PERF_MEM_NA ; data->txn =0 ; + data->skid_ip =0 ;/* mark as uinitialized */ } extern void perf_output_sample(struct perf_output_handle *handle,
include/uapi/linux/perf_event.h +3 -1
@@ -141,8 +141,9 @@ enum perf_event_sample_format {PERF_SAMPLE_TRANSACTION =1U <<17 ,PERF_SAMPLE_REGS_INTR =1U <<18 ,PERF_SAMPLE_PHYS_ADDR =1U <<19 , +PERF_SAMPLE_SKID_IP =1U <<20 , - PERF_SAMPLE_MAX = 1U << 20, /* non-ABI */ + PERF_SAMPLE_MAX = 1U << 21, /* non-ABI */ };/*
@@ -817,6 +818,7 @@ enum perf_event_type {* { u64 abi; # enum perf_sample_regs_abi * u64 regs[weight(mask)]; } && PERF_SAMPLE_REGS_INTR * { u64 phys_addr;} && PERF_SAMPLE_PHYS_ADDR +* { u64 skid_ip; } && PERF_SAMPLE_SKID_IP * }; */ PERF_RECORD_SAMPLE = 9,
kernel/events/core.c +14 -0
@@ -1565,6 +1565,9 @@ static void __perf_event_header_size(struct perf_event *event, u64 sample_type)if (sample_type &PERF_SAMPLE_PHYS_ADDR ) size +=sizeof (data->phys_addr); +if (sample_type &PERF_SAMPLE_SKID_IP ) + size +=sizeof (data->skid_ip); + event->header_size = size; }
@@ -5934,6 +5937,9 @@ void perf_output_sample(struct perf_output_handle *handle,if (sample_type &PERF_SAMPLE_PHYS_ADDR )perf_output_put (handle, data->phys_addr); +if (sample_type &PERF_SAMPLE_SKID_IP ) +perf_output_put (handle, data->skid_ip); +if (!event->attr.watermark) {int wakeup_events = event->attr.wakeup_events;
@@ -5999,6 +6005,14 @@ void perf_prepare_sample(struct perf_event_header *header,if (sample_type &PERF_SAMPLE_IP ) data->ip =perf_instruction_pointer (regs); +/* +* if skid_ip has not been set by arch specific code, then +* we initialize it to IP as interrupt-based sampling has +* skid +*/ +if (!data->skid_ip && sample_type &PERF_SAMPLE_SKID_IP ) + data->skid_ip =perf_instruction_pointer (regs); +if (sample_type &PERF_SAMPLE_CALLCHAIN ) {int size =1 ;
Signature · 1 line
2.7.4
PATCH v3 2/5 perf/x86: add PERF_SAMPLE_SKID_IP support for X86 PEBS
In reply to Stephane EranianThis patch adds support for SKID_IP for Intel x86 processors when PEBS mode is enabled.
--- arch/x86/events/intel/ds.c | 7 +++++++ 1 file changed, 7 insertions(+)
arch/x86/events/intel/ds.c +7 -0
@@ -1190,6 +1190,13 @@ static void setup_pebs_sample_data(struct perf_event *event, x86_pmu.intel_cap.pebs_format >= 1) data->addr = pebs->dla; +/* +* unmodified, skid IP which is guaranteed to be the next +* dynamic instruction +*/ +if (sample_type &PERF_SAMPLE_SKID_IP ) + data->skid_ip = pebs->ip; +if (x86_pmu.intel_cap.pebs_format >=2 ) {/* Only set the TSX weight when no memory weight. */ if ((sample_type &PERF_SAMPLE_WEIGHT ) && !fll)
Signature · 1 line
2.7.4
PATCH v3 3/5 perf/tools: add support for PERF_SAMPLE_SKID_IP
In reply to Stephane EranianThis patch adds the support code to handle the PERF_SAMPLE_SKID_IP record type. This is done as an event term and as such can be enabled per event: cpu/event=xxx,skid-ip=1/. This is a boolean term which is false by default.
--- tools/include/uapi/linux/perf_event.h | 4 +++- tools/perf/util/event.h | 1 + tools/perf/util/evsel.c | 11 +++++++++++ tools/perf/util/evsel.h | 2 ++ tools/perf/util/parse-events.c | 7 +++++++ tools/perf/util/parse-events.h | 1 + tools/perf/util/parse-events.l | 1 + tools/perf/util/session.c | 3 +++ 8 files changed, 29 insertions(+), 1 deletion(-)
tools/include/uapi/linux/perf_event.h +3 -1
@@ -141,8 +141,9 @@ enum perf_event_sample_format {PERF_SAMPLE_TRANSACTION =1U <<17 ,PERF_SAMPLE_REGS_INTR =1U <<18 ,PERF_SAMPLE_PHYS_ADDR =1U <<19 , +PERF_SAMPLE_SKID_IP =1U <<20 , - PERF_SAMPLE_MAX = 1U << 20, /* non-ABI */ + PERF_SAMPLE_MAX = 1U << 21, /* non-ABI */ };/*
@@ -817,6 +818,7 @@ enum perf_event_type {* { u64 abi; # enum perf_sample_regs_abi * u64 regs[weight(mask)]; } && PERF_SAMPLE_REGS_INTR * { u64 phys_addr;} && PERF_SAMPLE_PHYS_ADDR +* { u64 skid_ip; } && PERF_SAMPLE_SKID_IP * }; */ PERF_RECORD_SAMPLE = 9,
tools/perf/util/event.h +1 -0
@@ -202,6 +202,7 @@ struct perf_sample {u32 raw_size;u64 data_src;u64 phys_addr; +u64 skid_ip;u32 flags;u16 insn_len;u8 cpumode;
tools/perf/util/evsel.c +11 -0
@@ -775,6 +775,10 @@ static void apply_config_terms(struct perf_evsel *evsel,case PERF_EVSEL__CONFIG_TERM_OVERWRITE : attr->write_backward = term->val.overwrite ?1 :0 ;break ; +case PERF_EVSEL__CONFIG_TERM_SKID_IP : +if (term->val.skid_ip) +perf_evsel__set_sample_bit (evsel,SKID_IP ); +break ;default :break ; }
@@ -1478,6 +1482,7 @@ static void __p_sample_type(char *buf, size_t size, u64 value)bit_name (BRANCH_STACK ),bit_name (REGS_USER ),bit_name (STACK_USER ),bit_name (IDENTIFIER ),bit_name (REGS_INTR ),bit_name (DATA_SRC ),bit_name (WEIGHT ),bit_name (PHYS_ADDR ), +bit_name (SKID_IP ), { .name = NULL, } };#undef bit_name
@@ -2225,6 +2230,12 @@ int perf_evsel__parse_sample(struct perf_evsel *evsel, union perf_event *event, array++; } + data->skid_ip =0 ; +if (type &PERF_SAMPLE_SKID_IP ) { + data->skid_ip = *array; + array++; + } +return 0 ; }
tools/perf/util/evsel.h +2 -0
@@ -49,6 +49,7 @@ enum {PERF_EVSEL__CONFIG_TERM_OVERWRITE ,PERF_EVSEL__CONFIG_TERM_DRV_CFG ,PERF_EVSEL__CONFIG_TERM_BRANCH , +PERF_EVSEL__CONFIG_TERM_SKID_IP , PERF_EVSEL__CONFIG_TERM_MAX, };
@@ -66,6 +67,7 @@ struct perf_evsel_config_term {bool inherit;bool overwrite;char *branch; +bool skid_ip; } val; };
tools/perf/util/parse-events.c +7 -0
@@ -918,6 +918,7 @@ static const char *config_term_names[__PARSE_EVENTS__TERM_TYPE_NR] = { [PARSE_EVENTS__TERM_TYPE_OVERWRITE] = "overwrite", [PARSE_EVENTS__TERM_TYPE_NOOVERWRITE] = "no-overwrite", [PARSE_EVENTS__TERM_TYPE_DRV_CFG] = "driver-config", + [PARSE_EVENTS__TERM_TYPE_SKID_IP] = "skid-ip", };static bool config_term_shrinked;
@@ -1026,6 +1027,9 @@ do { \case PARSE_EVENTS__TERM_TYPE_MAX_STACK :CHECK_TYPE_VAL (NUM );break ; +case PARSE_EVENTS__TERM_TYPE_SKID_IP : +CHECK_TYPE_VAL (NUM ); +break ;default : err->str =strdup ("unknown term" ); err->idx = term->err_term;
@@ -1159,6 +1163,9 @@ do { \case PARSE_EVENTS__TERM_TYPE_DRV_CFG :ADD_CONFIG_TERM (DRV_CFG , drv_cfg, term->val.str);break ; +case PARSE_EVENTS__TERM_TYPE_SKID_IP : +ADD_CONFIG_TERM (SKID_IP , skid_ip, term->val.num ?1 :0 ); +break ;default :break ; }
tools/perf/util/parse-events.h +1 -0
@@ -74,6 +74,7 @@ enum {PARSE_EVENTS__TERM_TYPE_NOOVERWRITE ,PARSE_EVENTS__TERM_TYPE_OVERWRITE ,PARSE_EVENTS__TERM_TYPE_DRV_CFG , +PARSE_EVENTS__TERM_TYPE_SKID_IP , __PARSE_EVENTS__TERM_TYPE_NR, };
tools/perf/util/parse-events.l +1 -0
@@ -257,6 +257,7 @@ inherit { return term(yyscanner, PARSE_EVENTS__TERM_TYPE_INHERIT); } no-inherit { return term(yyscanner, PARSE_EVENTS__TERM_TYPE_NOINHERIT); } overwrite { return term(yyscanner, PARSE_EVENTS__TERM_TYPE_OVERWRITE); } no-overwrite { return term(yyscanner, PARSE_EVENTS__TERM_TYPE_NOOVERWRITE); } +skid-ip { return term(yyscanner, PARSE_EVENTS__TERM_TYPE_SKID_IP); } , { return ','; } "/" { BEGIN(INITIAL); return '/'; } {name_minus} { return str(yyscanner, PE_NAME); }
tools/perf/util/session.c +3 -0
@@ -1131,6 +1131,9 @@ static void dump_sample(struct perf_evsel *evsel, union perf_event *event,if (sample_type &PERF_SAMPLE_READ )sample_read__printf (sample, evsel->attr.read_format); + +if (sample_type &PERF_SAMPLE_SKID_IP ) +printf ("... skid_ip: %" PRIu64"\n" , sample->skid_ip); }static void dump_read (struct perf_evsel *evsel,union perf_event *event)
Signature · 1 line
2.7.4
PATCH v3 4/5 perf/record: add documentation for sampling skid ip
In reply to Stephane EranianThis patch adds documentation to describe how to use the skid ip support with perf record. The sample type can be provided per event as follows: pmu_instance/...,skid-ip=1/
For instance on Intel X86:
$ perf record -e cpu/event=0xc5,skid-ip=1/pp
does record the precise address of retired branches and their target.
--- tools/perf/Documentation/perf-record.txt | 8 ++++++++ 1 file changed, 8 insertions(+)
tools/perf/Documentation/perf-record.txt +8 -0
@@ -57,6 +57,14 @@ OPTIONS FP mode, "dwarf" for DWARF mode, "lbr" for LBR mode and "no" for disable callgraph. - 'stack-size': user stack size for dwarf mode + - 'skid_ip' : boolean, captures the unmodified interrupt instruction pointer + (IP) in each sample. Usually with event-based sampling, the IP + has skid and rarely point to the instruction which caused the + event to overflow. On some architectures, the hardware can eliminate + the skid and perf_events returns it as the IP with precise sampling is + enabled. But for certain measurements, it may be useful to have both + the correct and skid ip. This option enable capturing the skid ip in + additional to the corrected ip. Default is: false See the linkperf:perf-list[1] man page for more parameters.
Signature · 1 line
2.7.4
Re: [PATCH v3 4/5] perf/record: add documentation for sampling skid ip
Quoted 6 lines
tools/perf/Documentation/perf-record.txt +0
-0@@ -57,6 +57,14 @@ OPTIONS FP mode, "dwarf" for DWARF mode, "lbr" for LBR mode and
tools/perf/Documentation/perf-record.txt OPTIONS · line 67
Quoted 10 lines
tools/perf/Documentation/perf-record.txt +8
-0"no" for disable callgraph. - 'stack-size': user stack size for dwarf mode + - 'skid_ip' : boolean, captures the unmodified interrupt instruction pointer + (IP) in each sample. Usually with event-based sampling, the IP + has skid and rarely point to the instruction which caused the + event to overflow. On some architectures, the hardware can eliminate + the skid and perf_events returns it as the IP with precise sampling is + enabled. But for certain measurements, it may be useful to have both + the correct and skid ip. This option enable capturing the skid ip in + additional to the corrected ip. Default is: false
Actually in the scanner you use 'skid-ip' (with minus). Documentation should at least match the scanner.
_ is usually safer than -
-Andi
Re: [PATCH v3 4/5] perf/record: add documentation for sampling skid ip
Quoted 11 lines · Andi Kleen
Quoted 11 lines
tools/perf/Documentation/perf-record.txt +3
-0@@ -57,6 +57,14 @@ OPTIONS FP mode, "dwarf" for DWARF mode, "lbr" for LBR mode and "no" for disable callgraph. - 'stack-size': user stack size for dwarf mode + - 'skid_ip' : boolean, captures the unmodified interrupt instruction pointer + (IP) in each sample. Usually with event-based sampling, the IP + has skid and rarely point to the instruction which caused the
Quoted 9 lines
tools/perf/Documentation/perf-record.txt OPTIONS · line 67
Quoted 5 lines
tools/perf/Documentation/perf-record.txt +5
-0+ event to overflow. On some architectures, the hardware can eliminate + the skid and perf_events returns it as the IP with precise sampling is + enabled. But for certain measurements, it may be useful to have both + the correct and skid ip. This option enable capturing the skid ip in + additional to the corrected ip. Default is: falseActually in the scanner you use 'skid-ip' (with minus). Documentation should at least match the scanner.
I have fixed the documentation. I used - because many other terms and options use - instead of _.
Quoted 2 lines
_ is usually safer than -
I tend to agree with you.
Quoted 1 line
-Andi
PATCH v3 5/5 perf/script: add support for skid ip
In reply to Stephane EranianThis patch adds a skid_ip field to perf script
to dump the raw value of the PERF_SAMPLE_SKID_IP
field in each sample.
$ perf script -F +ip,+skid_ip ......
The field is not enabled by default.
--- tools/perf/Documentation/perf-script.txt | 2 +- tools/perf/builtin-script.c | 10 ++++++++-- tools/perf/util/session.c | 2 +- 3 files changed, 10 insertions(+), 4 deletions(-)
tools/perf/Documentation/perf-script.txt +1 -1
@@ -117,7 +117,7 @@ OPTIONS Comma separated list of fields to print. Options are: comm, tid, pid, time, cpu, event, trace, ip, sym, dso, addr, symoff, srcline, period, iregs, uregs, brstack, brstacksym, flags, bpf-output, brstackinsn, - brstackoff, callindent, insn, insnlen, synth, phys_addr. + brstackoff, callindent, insn, insnlen, synth, phys_addr, skid_ip. Field list can be prepended with the type, trace, sw or hw, to indicate to which event type the field list applies. e.g., -F sw:comm,tid,time,ip,sym and -F trace:time,cpu,trace
tools/perf/builtin-script.c +8 -2
@@ -90,6 +90,7 @@ enum perf_output_field {PERF_OUTPUT_SYNTH =1U <<25 ,PERF_OUTPUT_PHYS_ADDR =1U <<26 ,PERF_OUTPUT_UREGS =1U <<27 , + PERF_OUTPUT_SKID_IP = 1U << 28, }; struct output_option {
@@ -124,6 +125,7 @@ struct output_option { {.str = "brstackoff", .field = PERF_OUTPUT_BRSTACKOFF}, {.str = "synth", .field = PERF_OUTPUT_SYNTH}, {.str = "phys_addr", .field = PERF_OUTPUT_PHYS_ADDR}, + {.str = "skid_ip", .field = PERF_OUTPUT_SKID_IP}, }; enum {
@@ -1563,7 +1565,11 @@ static void process_event(struct perf_script *script,if (PRINT_FIELD (PHYS_ADDR ))fprintf (fp,"%16" PRIx64, sample->phys_addr); -fprintf (fp,"\n" ); + +if (PRINT_FIELD (SKID_IP )) +printf (" %" PRIx64" " , sample->skid_ip); + +printf ("\n" ); }static struct scripting_ops *scripting_ops;
@@ -2915,7 +2921,7 @@ int cmd_script(int argc, const char **argv) "Valid types: hw,sw,trace,raw,synth. " "Fields: comm,tid,pid,time,cpu,event,trace,ip,sym,dso," "addr,symoff,period,iregs,uregs,brstack,brstacksym,flags," - "bpf-output,callindent,insn,insnlen,brstackinsn,synth,phys_addr", + "bpf-output,callindent,insn,insnlen,brstackinsn,synth,phys_addr,skid_ip", parse_output_fields),OPT_BOOLEAN ('a' ,"all-cpus" , &system_wide, "system-wide collection from all CPUs"),
tools/perf/util/session.c +1 -1
@@ -1133,7 +1133,7 @@ static void dump_sample(struct perf_evsel *evsel, union perf_event *event,sample_read__printf (sample, evsel->attr.read_format);if (sample_type &PERF_SAMPLE_SKID_IP ) -printf ("... skid_ip: %" PRIu64"\n" , sample->skid_ip); +printf ("... skid_ip: 0x%" PRIx64"\n" , sample->skid_ip); }static void dump_read (struct perf_evsel *evsel,union perf_event *event)
Signature · 1 line
2.7.4
Re: [PATCH v3 5/5] perf/script: add support for skid ip
On Tue, Nov 07, 2017 at 11:56:37PM -0800, Stephane Eranian wrote:
SNIP
tools/perf/util/session.c static void dump_sample(struct perf_evsel *evsel, union perf_event *event, · line 1136
Quoted 10 lines
tools/perf/util/session.c +1
-1@@ -1133,7 +1133,7 @@ static void dump_sample(struct perf_evsel *evsel, union perf_event *event,sample_read__printf (sample, evsel->attr.read_format);if (sample_type &PERF_SAMPLE_SKID_IP ) -printf ("... skid_ip: %" PRIu64"\n" , sample->skid_ip); +printf ("... skid_ip: 0x%" PRIx64"\n" , sample->skid_ip);
should be part of the patch 3/5
jirka