perf intel-pt: Add support for decoding FUP/TIP only
Adrian Hunter
reviewerv2 cover perf intel-pt: Add support for decoding FUP/TIP only
Hi
Here are some fixes and small improvements for Intel PT.
Changes in V2:
For d/e flags, use +/- alphabetic options instead of numbers Update help text
Improve documentation
Adrian Hunter (12):
perf intel-pt: Fix FUP packet state
perf intel-pt: Fix duplicate branch after CBR
perf tools: Improve aux_output not supported error
perf auxtrace: Add missing itrace options to help text
perf auxtrace: Add optional error flags to the itrace 'e' option perf intel-pt: Use itrace error flags to suppress some errors perf auxtrace: Add optional log flags to the itrace 'd' option perf intel-pt: Use itrace debug log flags to suppress some messages perf intel-pt: Time filter logged perf events
perf auxtrace: Add itrace 'q' option for quicker, less detailed decoding perf intel-pt: Add support for decoding FUP/TIP only
perf intel-pt: Add support for decoding PSB+ only
tools/perf/Documentation/itrace.txt | 14 ++ tools/perf/Documentation/perf-intel-pt.txt | 63 +++++- tools/perf/util/auxtrace.c | 50 +++++ tools/perf/util/auxtrace.h | 31 ++- tools/perf/util/evsel.c | 4 + .../perf/util/intel-pt-decoder/intel-pt-decoder.c | 214 +++++++++++++++++++-- .../perf/util/intel-pt-decoder/intel-pt-decoder.h | 1 + tools/perf/util/intel-pt.c | 45 ++++- 8 files changed, 389 insertions(+), 33 deletions(-)
Regards
Adrian
Adrian Hunter
reviewerPATCH v2 1/12 perf intel-pt: Fix FUP packet state
While walking code towards a FUP ip, the packet state is INTEL_PT_STATE_FUP or INTEL_PT_STATE_FUP_NO_TIP. That was mishandled resulting in the state becoming INTEL_PT_STATE_IN_SYNC prematurely. The result was an occasional lost EXSTOP event.
--- .../util/intel-pt-decoder/intel-pt-decoder.c | 21 +++++++------------ 1 file changed, 7 insertions(+), 14 deletions(-)
tools/perf/util/intel-pt-decoder/intel-pt-decoder.c +7 -14
@@ -1164,6 +1164,7 @@ static int intel_pt_walk_fup(struct intel_pt_decoder *decoder)return 0 ;if (err == -EAGAIN ||intel_pt_fup_with_nlip (decoder, &intel_pt_insn, ip, err)) { + decoder->pkt_state =INTEL_PT_STATE_IN_SYNC ;if (intel_pt_fup_event (decoder))return 0 ;return -EAGAIN ;
@@ -1942,17 +1943,13 @@ static int intel_pt_walk_trace(struct intel_pt_decoder *decoder) }if (decoder->set_fup_mwait) no_tip = true; +if (no_tip) + decoder->pkt_state =INTEL_PT_STATE_FUP_NO_TIP ; +else + decoder->pkt_state =INTEL_PT_STATE_FUP ; err =intel_pt_walk_fup (decoder); -if (err != -EAGAIN ) { -if (err) -return err; -if (no_tip) - decoder->pkt_state = -INTEL_PT_STATE_FUP_NO_TIP ; -else - decoder->pkt_state =INTEL_PT_STATE_FUP ; -return 0 ; - } +if (err != -EAGAIN ) +return err;if (no_tip) { no_tip = false;break ;
@@ -2599,15 +2596,11 @@ const struct intel_pt_state *intel_pt_decode(struct intel_pt_decoder *decoder) err =intel_pt_walk_tip (decoder);break ;case INTEL_PT_STATE_FUP : - decoder->pkt_state =INTEL_PT_STATE_IN_SYNC ; err =intel_pt_walk_fup (decoder);if (err == -EAGAIN ) err =intel_pt_walk_fup_tip (decoder); -else if (!err) - decoder->pkt_state =INTEL_PT_STATE_FUP ;break ;case INTEL_PT_STATE_FUP_NO_TIP : - decoder->pkt_state =INTEL_PT_STATE_IN_SYNC ; err =intel_pt_walk_fup (decoder);if (err == -EAGAIN ) err =intel_pt_walk_trace (decoder);
Signature · 1 line
2.17.1
Adrian Hunter
reviewerPATCH v2 2/12 perf intel-pt: Fix duplicate branch after CBR
In reply to Adrian HunterCBR events can result in a duplicate branch event, because the state type defaults to a branch. Fix by clearing the state type.
Example: trace 'sleep' and hope for a frequency change
Before:
$ perf record -e intel_pt//u sleep 0.1 [ perf record: Woken up 1 times to write data ] [ perf record: Captured and wrote 0.034 MB perf.data ] $ perf script --itrace=bpe > before.txt
After:
$ perf script --itrace=bpe > after.txt
$ diff -u before.txt after.txt
--- before.txt 2020-07-07 14:42:18.191508098 +0300
+++ after.txt 2020-07-07 14:42:36.587891753 +0300
@@ -29673,7 +29673,6 @@
sleep 93431 [007] 15411.619905: 1 branches:u: 0 [unknown] ([unknown]) => 7f0818abb2e0 clock_nanosleep@@GLIBC_2.17+0x0 (/usr/lib/x86_64-linux-gnu/libc-2.31.so)
sleep 93431 [007] 15411.619905: 1 branches:u: 7f0818abb30c clock_nanosleep@@GLIBC_2.17+0x2c (/usr/lib/x86_64-linux-gnu/libc-2.31.so) => 0 [unknown] ([unknown])
sleep 93431 [007] 15411.720069: cbr: cbr: 15 freq: 1507 MHz ( 56%) 7f0818abb30c clock_nanosleep@@GLIBC_2.17+0x2c (/usr/lib/x86_64-linux-gnu/libc-2.31.so)
- sleep 93431 [007] 15411.720069: 1 branches:u: 7f0818abb30c clock_nanosleep@@GLIBC_2.17+0x2c (/usr/lib/x86_64-linux-gnu/libc-2.31.so) => 0 [unknown] ([unknown])
sleep 93431 [007] 15411.720076: 1 branches:u: 0 [unknown] ([unknown]) => 7f0818abb30e clock_nanosleep@@GLIBC_2.17+0x2e (/usr/lib/x86_64-linux-gnu/libc-2.31.so)
sleep 93431 [007] 15411.720077: 1 branches:u: 7f0818abb323 clock_nanosleep@@GLIBC_2.17+0x43 (/usr/lib/x86_64-linux-gnu/libc-2.31.so) => 7f0818ac0eb7 __nanosleep+0x17 (/usr/lib/x86_64-linux-gnu/libc-2.31.so)
sleep 93431 [007] 15411.720077: 1 branches:u: 7f0818ac0ebf __nanosleep+0x1f (/usr/lib/x86_64-linux-gnu/libc-2.31.so) => 55cb7e4c2827 rpl_nanosleep+0x97 (/usr/bin/sleep)--- tools/perf/util/intel-pt-decoder/intel-pt-decoder.c | 8 ++++++-- 1 file changed, 6 insertions(+), 2 deletions(-)
tools/perf/util/intel-pt-decoder/intel-pt-decoder.c +6 -2
@@ -1977,8 +1977,10 @@ static int intel_pt_walk_trace(struct intel_pt_decoder *decoder)* possibility of another CBR change that gets caught up * in the PSB+. */ -if (decoder->cbr != decoder->cbr_seen) +if (decoder->cbr != decoder->cbr_seen) { + decoder->state.type =0 ;return 0 ; + }break ;case INTEL_PT_PIP :
@@ -2019,8 +2021,10 @@ static int intel_pt_walk_trace(struct intel_pt_decoder *decoder)case INTEL_PT_CBR :intel_pt_calc_cbr (decoder); -if (decoder->cbr != decoder->cbr_seen) +if (decoder->cbr != decoder->cbr_seen) { + decoder->state.type =0 ;return 0 ; + }break ;case INTEL_PT_MODE_EXEC :
Signature · 1 line
2.17.1
Adrian Hunter
reviewerPATCH v2 3/12 perf tools: Improve aux_output not supported error
In reply to Adrian HunterFor example:
Before:
$ perf record -e '{intel_pt/branch=0/,branch-loads/aux-output/ppp}' -- ls -l
Error:
branch-loads: PMU Hardware doesn't support sampling/overflow-interrupts. Try 'perf stat'
After:
$ perf record -e '{intel_pt/branch=0/,branch-loads/aux-output/ppp}' -- ls -l
Error:
branch-loads: PMU Hardware doesn't support 'aux_output' feature--- tools/perf/util/evsel.c | 4 ++++ 1 file changed, 4 insertions(+)
tools/perf/util/evsel.c +4 -0
@@ -2539,6 +2539,10 @@ int evsel__open_strerror(struct evsel *evsel, struct target *target, "No such device - did you specify an out-of-range profile CPU?");break ;case EOPNOTSUPP : +if (evsel->core.attr.aux_output) +return scnprintf (msg, size, +"%s: PMU Hardware doesn't support 'aux_output' feature" , +evsel__name (evsel));if (evsel->core.attr.sample_period !=0 ) return scnprintf(msg, size, "%s: PMU Hardware doesn't support sampling/overflow-interrupts. Try 'perf stat'",
Signature · 1 line
2.17.1
Adrian Hunter
reviewerPATCH v2 4/12 perf auxtrace: Add missing itrace options to help text
In reply to Adrian HunterAdd missing itrace options o, G and L.
--- tools/perf/util/auxtrace.h | 6 +++++- 1 file changed, 5 insertions(+), 1 deletion(-)
tools/perf/util/auxtrace.h +5 -1
@@ -604,13 +604,15 @@ bool auxtrace__evsel_is_auxtrace(struct perf_session *session, struct evsel *evsel);#define ITRACE_HELP \ -" i: synthesize instructions events\n" \ +" i[period]: synthesize instructions events\n" \ " b: synthesize branches events (branch misses for Arm SPE)\n" \ " c: synthesize branches events (calls only)\n" \ " r: synthesize branches events (returns only)\n" \ " x: synthesize transactions events\n" \ " w: synthesize ptwrite events\n" \ " p: synthesize power events\n" \ +" o: synthesize other events recorded due to the use\n" \ +" of aux-output (refer to perf record)\n" \ " e: synthesize error events\n" \ " d: create a debug log\n" \ " f: synthesize first level cache events\n" \
@@ -618,7 +620,9 @@ bool auxtrace__evsel_is_auxtrace(struct perf_session *session," t: synthesize TLB events\n" \" a: synthesize remote access events\n" \" g[len]: synthesize a call chain (use with i or x)\n" \ +" G[len]: synthesize a call chain on existing event records\n" \" l[len]: synthesize last branch entries (use with i or x)\n" \ +" L[len]: synthesize last branch entries on existing event records\n" \" sNUMBER: skip initial number of events\n" \" PERIOD[ns|us|ms|i|t]: specify period to sample stream\n" \" concatenate multiple options. Default is ibxwpe or cewp\n"
Signature · 1 line
2.17.1
Adrian Hunter
reviewerPATCH v2 5/12 perf auxtrace: Add optional error flags to the itrace 'e' option
In reply to Adrian HunterAllow the 'e' option to be followed by flags which will affect what errors will or will not be reported. Each flag must be preceded by either '+' or '-'. The flags are:
o overflow
l trace data lost
--- tools/perf/Documentation/itrace.txt | 6 ++++ tools/perf/util/auxtrace.c | 44 +++++++++++++++++++++++++++++ tools/perf/util/auxtrace.h | 12 +++++++- 3 files changed, 61 insertions(+), 1 deletion(-)
tools/perf/Documentation/itrace.txt +6 -0
@@ -47,3 +47,9 @@ --itrace=i0nss1000000 skips the first million instructions. + + The 'e' option may be followed by flags which affect what errors will or + will not be reported. Each flag must be preceded by either '+' or '-'. + The flags are: + o overflow + l trace data lost
tools/perf/util/auxtrace.c +44 -0
56 lines not on this page · raw message
tools/perf/util/auxtrace.h +11 -1
@@ -55,6 +55,9 @@ enum itrace_period_type { PERF_ITRACE_PERIOD_NANOSECS, }; +#define AUXTRACE_ERR_FLG_OVERFLOW (1 << ('o' - 'a')) +#define AUXTRACE_ERR_FLG_DATA_LOST (1 << ('l' - 'a')) +/** * struct itrace_synth_opts - AUX area tracing synthesis options. * @set: indicates whether or not options have been set
@@ -91,6 +94,8 @@ enum itrace_period_type {* @cpu_bitmap: CPUs for which to synthesize events, or NULL for all * @ptime_range: time intervals to trace or NULL * @range_num: number of time intervals to trace +* @error_plus_flags: flags to affect what errors are reported +* @error_minus_flags: flags to affect what errors are reported */ struct itrace_synth_opts {bool set;
@@ -124,6 +129,8 @@ struct itrace_synth_opts {unsigned long *cpu_bitmap;struct perf_time_interval *ptime_range;int range_num; +unsigned int error_plus_flags; +unsigned int error_minus_flags; };/**
@@ -613,7 +620,10 @@ bool auxtrace__evsel_is_auxtrace(struct perf_session *session," p: synthesize power events\n" \" o: synthesize other events recorded due to the use\n" \" of aux-output (refer to perf record)\n" \ -" e: synthesize error events\n" \ +" e[flags]: synthesize error events\n" \ +" each flag must be preceded by + or -\n" \ +" error flags are: o (overflow)\n" \ +" l (data lost)\n" \" d: create a debug log\n" \" f: synthesize first level cache events\n" \" m: synthesize last level cache events\n" \
Signature · 1 line
2.17.1
Adrian Hunter
reviewerPATCH v2 6/12 perf intel-pt: Use itrace error flags to suppress some errors
In reply to Adrian HunterThe itrace "e" option may be followed by flags which affect what errors will or will not be reported. Each flag must be preceded by either '+' or '-'. The flags supported by Intel PT are:
-o Suppress overflow errors
-l Suppress trace data lost errors
For example, for errors but not overflow or data lost errors:
--itrace=e-o-l
Suppressing those errors can be useful for testing and debugging because they are not due to decoding.
--- tools/perf/Documentation/perf-intel-pt.txt | 9 ++++++++- tools/perf/util/intel-pt.c | 9 +++++++++ 2 files changed, 17 insertions(+), 1 deletion(-)
tools/perf/Documentation/perf-intel-pt.txt +8 -1
@@ -871,7 +871,14 @@ Developer Manuals. Error events show where the decoder lost the trace. Error events are quite important. Users must know if what they are seeing is a complete -picture or not. +picture or not. The "e" option may be followed by flags which affect what errors +will or will not be reported. Each flag must be preceded by either '+' or '-'. +The flags supported by Intel PT are: + -o Suppress overflow errors + -l Suppress trace data lost errors +For example, for errors but not overflow or data lost errors: + + --itrace=e-o-l The "d" option will cause the creation of a file "intel_pt.log" containing all decoded packets and instructions. Note that this option slows down the decoder
tools/perf/util/intel-pt.c +9 -0
@@ -1862,6 +1862,15 @@ static int intel_pt_synth_error(struct intel_pt *pt, int code, int cpu,char msg[MAX_AUXTRACE_ERROR_MSG ];int err; +if (pt->synth_opts.error_minus_flags) { +if (code ==INTEL_PT_ERR_OVR && + pt->synth_opts.error_minus_flags &AUXTRACE_ERR_FLG_OVERFLOW ) +return 0 ; +if (code ==INTEL_PT_ERR_LOST && + pt->synth_opts.error_minus_flags &AUXTRACE_ERR_FLG_DATA_LOST ) +return 0 ; + } +intel_pt__strerror (code, msg,MAX_AUXTRACE_ERROR_MSG ); auxtrace_synth_error(&event.auxtrace_error, PERF_AUXTRACE_ERROR_ITRACE,
Signature · 1 line
2.17.1
Adrian Hunter
reviewerPATCH v2 7/12 perf auxtrace: Add optional log flags to the itrace 'd' option
In reply to Adrian HunterAllow the 'd' option to be followed by flags which will affect what debug messages will or will not be reported. Each flag must be preceded by either '+' or '-'. The flags are:
a all perf events
--- tools/perf/Documentation/itrace.txt | 5 +++++ tools/perf/util/auxtrace.c | 3 +++ tools/perf/util/auxtrace.h | 10 +++++++++- 3 files changed, 17 insertions(+), 1 deletion(-)
tools/perf/Documentation/itrace.txt +5 -0
@@ -53,3 +53,8 @@ The flags are: o overflow l trace data lost + + If supported, the 'd' option may be followed by flags which affect what + debug messages will or will not be logged. Each flag must be preceded + by either '+' or '-'. The flags are: + a all perf events
tools/perf/util/auxtrace.c +3 -0
@@ -1483,6 +1483,9 @@ int itrace_parse_synth_opts(const struct option *opt, const char *str,break ;case 'd' : synth_opts->log = true; +if (get_flags (&p, &synth_opts->log_plus_flags, + &synth_opts->log_minus_flags)) + goto out_err;break ;case 'c' : synth_opts->branches = true;
tools/perf/util/auxtrace.h +9 -1
@@ -58,6 +58,8 @@ enum itrace_period_type {#define AUXTRACE_ERR_FLG_OVERFLOW (1 << ('o' - 'a'))#define AUXTRACE_ERR_FLG_DATA_LOST (1 << ('l' - 'a')) +#define AUXTRACE_LOG_FLG_ALL_PERF_EVTS (1 << ('a' - 'a')) +/** * struct itrace_synth_opts - AUX area tracing synthesis options. * @set: indicates whether or not options have been set
@@ -96,6 +98,8 @@ enum itrace_period_type {* @range_num: number of time intervals to trace * @error_plus_flags: flags to affect what errors are reported * @error_minus_flags: flags to affect what errors are reported +* @log_plus_flags: flags to affect what is logged +* @log_minus_flags: flags to affect what is logged */ struct itrace_synth_opts {bool set;
@@ -131,6 +135,8 @@ struct itrace_synth_opts {int range_num;unsigned int error_plus_flags;unsigned int error_minus_flags; +unsigned int log_plus_flags; +unsigned int log_minus_flags; };/**
@@ -624,7 +630,9 @@ bool auxtrace__evsel_is_auxtrace(struct perf_session *session," each flag must be preceded by + or -\n" \" error flags are: o (overflow)\n" \" l (data lost)\n" \ -" d: create a debug log\n" \ +" d[flags]: create a debug log\n" \ +" each flag must be preceded by + or -\n" \ +" log flags are: a (all perf events)\n" \" f: synthesize first level cache events\n" \" m: synthesize last level cache events\n" \" t: synthesize TLB events\n" \
Signature · 1 line
2.17.1
Adrian Hunter
reviewerPATCH v2 8/12 perf intel-pt: Use itrace debug log flags to suppress some messages
In reply to Adrian HunterThe "d" option may be followed by flags which affect what debug messages will or will not be logged. Each flag must be preceded by either '+' or '-'. The flags support by Intel PT are:
-a Suppress logging of perf events
Suppressing perf events is useful for decreasing the size of the log.
--- tools/perf/Documentation/perf-intel-pt.txt | 5 ++++- tools/perf/util/intel-pt.c | 17 ++++++++++------- 2 files changed, 14 insertions(+), 8 deletions(-)
tools/perf/Documentation/perf-intel-pt.txt +4 -1
@@ -882,7 +882,10 @@ For example, for errors but not overflow or data lost errors: The "d" option will cause the creation of a file "intel_pt.log" containing all decoded packets and instructions. Note that this option slows down the decoder -and that the resulting file may be very large. +and that the resulting file may be very large. The "d" option may be followed +by flags which affect what debug messages will or will not be logged. Each flag +must be preceded by either '+' or '-'. The flags support by Intel PT are: + -a Suppress logging of perf events In addition, the period of the "instructions" event can be specified. e.g.
tools/perf/util/intel-pt.c +10 -7
@@ -249,6 +249,11 @@ static void intel_pt_dump_sample(struct perf_session *session,intel_pt_dump (pt, sample->aux_sample.data, sample->aux_sample.size); } +static bool intel_pt_log_events (struct intel_pt *pt) +{ +return !(pt->synth_opts.log_minus_flags &AUXTRACE_LOG_FLG_ALL_PERF_EVTS ); +} +static int intel_pt_do_fix_overlap (struct intel_pt *pt,struct auxtrace_buffer *a,struct auxtrace_buffer *b) {
@@ -2585,10 +2590,6 @@ static int intel_pt_context_switch(struct intel_pt *pt, union perf_event *event,return -EINVAL ; } -intel_pt_log ("context_switch: cpu %d pid %d tid %d time %" PRIu64" tsc %#" PRIx64"\n" , - cpu, pid, tid, sample->time,perf_time_to_tsc (sample->time, - &pt->tc)); - ret =intel_pt_sync_switch (pt, cpu, tid, sample->time);if (ret <=0 )return ret;
@@ -2745,9 +2746,11 @@ static int intel_pt_process_event(struct perf_session *session,if (!err && event->header.type ==PERF_RECORD_TEXT_POKE ) err =intel_pt_text_poke (pt, event); -intel_pt_log ("event %u: cpu %d time %" PRIu64" tsc %#" PRIx64" " , - event->header.type, sample->cpu, sample->time, timestamp); -intel_pt_log_event (event); +if (intel_pt_enable_logging &&intel_pt_log_events (pt)) { +intel_pt_log ("event %u: cpu %d time %" PRIu64" tsc %#" PRIx64" " , + event->header.type, sample->cpu, sample->time, timestamp); +intel_pt_log_event (event); + }return err; }
Signature · 1 line
2.17.1
Adrian Hunter
reviewerPATCH v2 9/12 perf intel-pt: Time filter logged perf events
In reply to Adrian HunterChange the debug logging (when used with the --time option) to time filter logged perf events, but allow that to be overridden by using "d+a" instead of plain "d".
That can reduce the size of the log file.
--- tools/perf/Documentation/perf-intel-pt.txt | 3 +++ tools/perf/util/intel-pt.c | 19 ++++++++++++++++--- 2 files changed, 19 insertions(+), 3 deletions(-)
tools/perf/Documentation/perf-intel-pt.txt +3 -0
@@ -886,6 +886,9 @@ and that the resulting file may be very large. The "d" option may be followed by flags which affect what debug messages will or will not be logged. Each flag must be preceded by either '+' or '-'. The flags support by Intel PT are: -a Suppress logging of perf events + +a Log all perf events +By default, logged perf events are filtered by any specified time ranges, but +flag +a overrides that. In addition, the period of the "instructions" event can be specified. e.g.
tools/perf/util/intel-pt.c +16 -3
@@ -249,9 +249,22 @@ static void intel_pt_dump_sample(struct perf_session *session,intel_pt_dump (pt, sample->aux_sample.data, sample->aux_sample.size); } -static bool intel_pt_log_events (struct intel_pt *pt) +static bool intel_pt_log_events (struct intel_pt *pt,u64 tm) { -return !(pt->synth_opts.log_minus_flags &AUXTRACE_LOG_FLG_ALL_PERF_EVTS ); +struct perf_time_interval *range = pt->synth_opts.ptime_range; +int n = pt->synth_opts.range_num; + +if (pt->synth_opts.log_plus_flags &AUXTRACE_LOG_FLG_ALL_PERF_EVTS ) +return true; + +if (pt->synth_opts.log_minus_flags &AUXTRACE_LOG_FLG_ALL_PERF_EVTS ) +return false; + +/* perf_time__ranges_skip_sample does not work if time is zero */ +if (!tm) + tm =1 ; + +return !n || !perf_time__ranges_skip_sample (range, n, tm); } static int intel_pt_do_fix_overlap(struct intel_pt *pt, struct auxtrace_buffer *a,
@@ -2746,7 +2759,7 @@ static int intel_pt_process_event(struct perf_session *session,if (!err && event->header.type ==PERF_RECORD_TEXT_POKE ) err =intel_pt_text_poke (pt, event); -if (intel_pt_enable_logging &&intel_pt_log_events (pt)) { +if (intel_pt_enable_logging &&intel_pt_log_events (pt, sample->time)) {intel_pt_log ("event %u: cpu %d time %" PRIu64" tsc %#" PRIx64" " , event->header.type, sample->cpu, sample->time, timestamp);intel_pt_log_event (event);
Signature · 1 line
2.17.1
Adrian Hunter
reviewerPATCH v2 10/12 perf auxtrace: Add itrace 'q' option for quicker, less detailed decoding
In reply to Adrian HunterThe 'q' option is for modes of decoding that are quicker because they skip or omit decoding some aspects of trace data.
If supported, the 'q' option may be repeated to increase the effect.
--- tools/perf/Documentation/itrace.txt | 3 +++ tools/perf/util/auxtrace.c | 3 +++ tools/perf/util/auxtrace.h | 3 +++ 3 files changed, 9 insertions(+)
tools/perf/Documentation/itrace.txt +3 -0
@@ -18,6 +18,7 @@ l synthesize last branch entries (use with i or x) L synthesize last branch entries on existing event records s skip initial number of events + q quicker (less detailed) decoding The default is all events i.e. the same as --itrace=ibxwpe, except for perf script where it is --itrace=ce
@@ -58,3 +59,5 @@ debug messages will or will not be logged. Each flag must be preceded by either '+' or '-'. The flags are: a all perf events + + If supported, the 'q' option may be repeated to increase the effect.
tools/perf/util/auxtrace.c +3 -0
@@ -1554,6 +1554,9 @@ int itrace_parse_synth_opts(const struct option *opt, const char *str,case 'a' : synth_opts->remote_access = true;break ; +case 'q' : + synth_opts->quick +=1 ; +break ;case ' ' :case ',' :break ;
tools/perf/util/auxtrace.h +3 -0
@@ -100,6 +100,7 @@ enum itrace_period_type {* @error_minus_flags: flags to affect what errors are reported * @log_plus_flags: flags to affect what is logged * @log_minus_flags: flags to affect what is logged +* @quick: quicker (less detailed) decoding */ struct itrace_synth_opts {bool set;
@@ -137,6 +138,7 @@ struct itrace_synth_opts {unsigned int error_minus_flags;unsigned int log_plus_flags;unsigned int log_minus_flags; +unsigned int quick; };/**
@@ -642,6 +644,7 @@ bool auxtrace__evsel_is_auxtrace(struct perf_session *session," l[len]: synthesize last branch entries (use with i or x)\n" \" L[len]: synthesize last branch entries on existing event records\n" \" sNUMBER: skip initial number of events\n" \ +" q: quicker (less detailed) decoding\n" \" PERIOD[ns|us|ms|i|t]: specify period to sample stream\n" \" concatenate multiple options. Default is ibxwpe or cewp\n"
Signature · 1 line
2.17.1
Re: [PATCH V2 10/12] perf auxtrace: Add itrace 'q' option for quicker, less detailed decoding
Quoted 1 line
+" q: quicker (less detailed) decoding\n" \
Perhaps add '(can be repeated)'
-Andi
Adrian Hunter
reviewerPATCH v2 11/12 perf intel-pt: Add support for decoding FUP/TIP only
In reply to Adrian HunterUse the new itrace 'q' option to add support for a mode of decoding that ignores TNT, does not walk object code, but gets the ip from FUP and TIP packets.
Example:
$ perf record -e intel_pt//u grep -rI pudding drivers [ perf record: Woken up 52 times to write data ] [ perf record: Captured and wrote 57.870 MB perf.data ] $ time perf script --itrace=bi | wc -l 58948289
real 1m23.863s user 1m23.251s sys 0m7.452s $ time perf script --itrace=biq | wc -l 3385694
real 0m4.453s user 0m4.455s sys 0m0.328s
--- tools/perf/Documentation/perf-intel-pt.txt | 31 ++++ .../util/intel-pt-decoder/intel-pt-decoder.c | 167 +++++++++++++++++- .../util/intel-pt-decoder/intel-pt-decoder.h | 1 + tools/perf/util/intel-pt.c | 6 +- 4 files changed, 200 insertions(+), 5 deletions(-)
tools/perf/Documentation/perf-intel-pt.txt +31 -0
@@ -825,6 +825,7 @@ The letters are: l synthesize last branch entries (use with i or x) L synthesize last branch entries on existing event records s skip initial number of events + q quicker (less detailed) decoding "Instructions" events look like they were recorded by "perf record -e instructions".
@@ -969,6 +970,36 @@ at the beginning. This is useful to ignore initialization code. skips the first million instructions. +The q option changes the way the trace is decoded. The decoding is much faster +but much less detailed. Specifically, with the q option, the decoder does not +decode TNT packets, and does not walk object code, but gets the ip from FUP and +TIP packets. The q option can be used with the b and i options but the period +is not used. The q option decodes more quickly, but is useful only if the +control flow of interest is represented or indicated by FUP, TIP, TIP.PGE, or +TIP.PGD packets (refer below). However the q option could be used to find time +ranges that could then be decoded fully using the --time option. + +What will *not* be decoded with the (single) q option: + + - direct calls and jmps + - conditional branches + - non-branch instructions + +What *will* be decoded with the (single) q option: + + - asynchronous branches such as interrupts + - indirect branches + - function return target address *if* the noretcomp config term (refer + config terms section) was used + - start of (control-flow) tracing + - end of (control-flow) tracing, if it is not out of context + - power events, ptwrite, transaction start and abort + - instruction pointer associated with PSB packets + +Note the q option does not specify what events will be synthesized e.g. the p +option must be used also to show power events. + + dump option ~~~~~~~~~~~
tools/perf/util/intel-pt-decoder/intel-pt-decoder.c +163 -4
246 lines not on this page · raw message
tools/perf/util/intel-pt-decoder/intel-pt-decoder.h +1 -0
@@ -250,6 +250,7 @@ struct intel_pt_params {uint32_t tsc_ctc_ratio_n;uint32_t tsc_ctc_ratio_d;enum intel_pt_param_flags flags; +unsigned int quick; };struct intel_pt_decoder ;
tools/perf/util/intel-pt.c +5 -1
@@ -1030,6 +1030,7 @@ static struct intel_pt_queue *intel_pt_alloc_queue(struct intel_pt *pt, params.mtc_period =intel_pt_mtc_period (pt); params.tsc_ctc_ratio_n = pt->tsc_ctc_ratio_n; params.tsc_ctc_ratio_d = pt->tsc_ctc_ratio_d; + params.quick = pt->synth_opts.quick;if (pt->filts.cnt >0 ) params.pgd_ip = intel_pt_pgd_ip;
@@ -1423,7 +1424,10 @@ static int intel_pt_synth_instruction_sample(struct intel_pt_queue *ptq) sample.id = ptq->pt->instructions_id; sample.stream_id = ptq->pt->instructions_id; - sample.period = ptq->state->tot_insn_cnt - ptq->last_insn_cnt; +if (pt->synth_opts.quick) + sample.period =1 ; +else + sample.period = ptq->state->tot_insn_cnt - ptq->last_insn_cnt; sample.cyc_cnt = ptq->ipc_cyc_cnt - ptq->last_in_cyc_cnt;if (sample.cyc_cnt) {
Signature · 1 line
2.17.1
Adrian Hunter
reviewerPATCH v2 12/12 perf intel-pt: Add support for decoding PSB+ only
In reply to Adrian HunterA single q option decodes ip from only FUP/TIP packets. Make it so that repeating the q option (i.e. qq) decodes only PSB+, getting ip if there is a FUP packet within PSB+ (i.e. between PSB and PSBEND).
Example:
$ perf record -e intel_pt//u grep -rI pudding drivers [ perf record: Woken up 52 times to write data ] [ perf record: Captured and wrote 57.870 MB perf.data ] $ time perf script --itrace=bi | wc -l 58948289
real 1m23.863s user 1m23.251s sys 0m7.452s $ time perf script --itrace=biq | wc -l 3385694
real 0m4.453s user 0m4.455s sys 0m0.328s $ time perf script --itrace=biqq | wc -l 1883
real 0m0.047s user 0m0.043s sys 0m0.009s
--- tools/perf/Documentation/perf-intel-pt.txt | 15 +++++++++++++++ .../util/intel-pt-decoder/intel-pt-decoder.c | 18 ++++++++++++++++++ 2 files changed, 33 insertions(+)
tools/perf/Documentation/perf-intel-pt.txt +15 -0
@@ -999,6 +999,21 @@ What *will* be decoded with the (single) q option: Note the q option does not specify what events will be synthesized e.g. the p option must be used also to show power events. +Repeating the q option (double-q i.e. qq) results in even faster decoding and even +less detail. The decoder decodes only extended PSB (PSB+) packets, getting the +instruction pointer if there is a FUP packet within PSB+ (i.e. between PSB and +PSBEND). Note PSB packets occur regularly in the trace based on the psb_period +config term (refer config terms section). There will be a FUP packet if the +PSB+ occurs while control flow is being traced. + +What will *not* be decoded with the qq option: + + - everything except instruction pointer associated with PSB packets + +What *will* be decoded with the qq option: + + - instruction pointer associated with PSB packets + dump option ~~~~~~~~~~~
tools/perf/util/intel-pt-decoder/intel-pt-decoder.c +18 -0
@@ -113,6 +113,7 @@ struct intel_pt_decoder {bool in_psb;bool hop;bool hop_psb_fup; +bool leap;enum intel_pt_param_flags flags;uint64_t pos;uint64_t last_ip;
@@ -240,6 +241,7 @@ struct intel_pt_decoder *intel_pt_decoder_new(struct intel_pt_params *params) decoder->return_compression = params->return_compression; decoder->branch_enable = params->branch_enable; decoder->hop = params->quick >=1 ; + decoder->leap = params->quick >=2 ; decoder->flags = params->flags;
@@ -1903,9 +1905,18 @@ static int intel_pt_resample(struct intel_pt_decoder *decoder)#define HOP_RETURN 2#define HOP_AGAIN 3 +static int intel_pt_scan_for_psb (struct intel_pt_decoder *decoder); +/* Hop mode: Ignore TNT, do not walk code, but get ip from FUPs and TIPs */ static int intel_pt_hop_trace (struct intel_pt_decoder *decoder,bool *no_tip,int *err) { +/* Leap from PSB to PSB, getting ip from FUP within PSB+ */ +if (decoder->leap && !decoder->in_psb && decoder->packet.type !=INTEL_PT_PSB ) { + *err =intel_pt_scan_for_psb (decoder); +if (*err) +return HOP_RETURN ; + } +switch (decoder->packet.type) {case INTEL_PT_TNT :return HOP_IGNORE ;
@@ -2681,6 +2692,7 @@ static int intel_pt_sync(struct intel_pt_decoder *decoder) decoder->ip =0 ;intel_pt_clear_stack (&decoder->stack); +leap: err =intel_pt_scan_for_psb (decoder);if (err)return err;
@@ -2702,6 +2714,12 @@ static int intel_pt_sync(struct intel_pt_decoder *decoder) decoder->pkt_state =INTEL_PT_STATE_RESAMPLE ; else decoder->pkt_state = INTEL_PT_STATE_IN_SYNC; + }else if (decoder->leap) { +/* +* In leap mode, only PSB+ is decoded, so keeping leaping to the +* next PSB until there is an ip. +*/ + goto leap; }else {return intel_pt_sync_ip (decoder); }
Signature · 1 line
2.17.1
Re: [PATCH V2 12/12] perf intel-pt: Add support for decoding PSB+ only
Quoted 39 lines · Adrian Hunter
A single q option decodes ip from only FUP/TIP packets. Make it so that repeating the q option (i.e. qq) decodes only PSB+, getting ip if there is a FUP packet within PSB+ (i.e. between PSB and PSBEND).
Example:
$ perf record -e intel_pt//u grep -rI pudding drivers [ perf record: Woken up 52 times to write data ] [ perf record: Captured and wrote 57.870 MB perf.data ] $ time perf script --itrace=bi | wc -l 58948289real 1m23.863s user 1m23.251s sys 0m7.452s $ time perf script --itrace=biq | wc -l 3385694real 0m4.453s user 0m4.455s sys 0m0.328s $ time perf script --itrace=biqq | wc -l 1883real 0m0.047s user 0m0.043s sys 0m0.009sSigned-off-by Adrian Hunter <[email protected]>--- tools/perf/Documentation/perf-intel-pt.txt | 15 +++++++++++++++ .../util/intel-pt-decoder/intel-pt-decoder.c | 18 ++++++++++++++++++ 2 files changed, 33 insertions(+)tools/perf/Documentation/perf-intel-pt.txt +0
-0
tools/perf/Documentation/perf-intel-pt.txt What *will* be decoded with the (single) q option: · line 1007
Quoted 9 lines
tools/perf/Documentation/perf-intel-pt.txt +6
-0@@ -999,6 +999,21 @@ What *will* be decoded with the (single) q option: Note the q option does not specify what events will be synthesized e.g. the p option must be used also to show power events. +Repeating the q option (double-q i.e. qq) results in even faster decoding and even +less detail. The decoder decodes only extended PSB (PSB+) packets, getting the +instruction pointer if there is a FUP packet within PSB+ (i.e. between PSB and +PSBEND). Note PSB packets occur regularly in the trace based on the psb_period +config term (refer config terms section). There will be a FUP packet if the +PSB+ occurs while control flow is being traced.
Some estimate would be good how frequent that is.
If we assume one bit per instruction then a 2K period it's roughly 16k instructions, with the 16K period roughly 128K instructions.
Could be added in a followon patch.
But looks overall the patches look good to me now.
(for the whole series)
Re: [PATCH V2 12/12] perf intel-pt: Add support for decoding PSB+ only
Em Mon, Jul 20, 2020 at 03:25:02PM -0700, Andi Kleen escreveu:
Quoted 4 lines
But looks overall the patches look good to me now.
(for the whole series)
Reviewed-by Andi Kleen <[email protected]>
Thanks, applied.
- Arnaldo
Adrian Hunter
reviewerQuoted 3 lines · Adrian Hunter
Hi
Here are some fixes and small improvements for Intel PT.
Andi added his "Reviewed-by" with 2 comments to tweak the documentation.
The patches still apply, so do you want me to send a V3?
Quoted 35 lines
Changes in V2:
For d/e flags, use +/- alphabetic options instead of numbers Update help text
Improve documentationAdrian Hunter (12):
perf intel-pt: Fix FUP packet state
perf intel-pt: Fix duplicate branch after CBR
perf tools: Improveaux_outputnot supported error
perf auxtrace: Add missing itrace options to help text
perf auxtrace: Add optional error flags to the itrace 'e' option perf intel-pt: Use itrace error flags to suppress some errors perf auxtrace: Add optional log flags to the itrace 'd' option perf intel-pt: Use itrace debug log flags to suppress some messages perf intel-pt: Time filter logged perf events
perf auxtrace: Add itrace 'q' option for quicker, less detailed decoding perf intel-pt: Add support for decoding FUP/TIP only
perf intel-pt: Add support for decoding PSB+ onlytools/perf/Documentation/itrace.txt | 14 ++ tools/perf/Documentation/perf-intel-pt.txt | 63 +++++- tools/perf/util/auxtrace.c | 50 +++++ tools/perf/util/auxtrace.h | 31 ++- tools/perf/util/evsel.c | 4 + .../perf/util/intel-pt-decoder/intel-pt-decoder.c | 214 +++++++++++++++++++-- .../perf/util/intel-pt-decoder/intel-pt-decoder.h | 1 + tools/perf/util/intel-pt.c | 45 ++++- 8 files changed, 389 insertions(+), 33 deletions(-)Regards
Adrian
Arnaldo Carvalho de Melo
maintainerEm Tue, Aug 04, 2020 at 04:34:36PM +0300, Adrian Hunter escreveu:
Quoted 9 lines · Adrian Hunter
Quoted 3 lines
Hi
Here are some fixes and small improvements for Intel PT.
Andi added his "Reviewed-by" with 2 comments to tweak the documentation.
The patches still apply, so do you want me to send a V3?
Thanks for the reminder, will apply and test build,
- Arnaldo
Quoted 36 lines
Quoted 35 lines
Changes in V2:
For d/e flags, use +/- alphabetic options instead of numbers Update help text
Improve documentationAdrian Hunter (12):
perf intel-pt: Fix FUP packet state
perf intel-pt: Fix duplicate branch after CBR
perf tools: Improveaux_outputnot supported error
perf auxtrace: Add missing itrace options to help text
perf auxtrace: Add optional error flags to the itrace 'e' option perf intel-pt: Use itrace error flags to suppress some errors perf auxtrace: Add optional log flags to the itrace 'd' option perf intel-pt: Use itrace debug log flags to suppress some messages perf intel-pt: Time filter logged perf events
perf auxtrace: Add itrace 'q' option for quicker, less detailed decoding perf intel-pt: Add support for decoding FUP/TIP only
perf intel-pt: Add support for decoding PSB+ onlytools/perf/Documentation/itrace.txt | 14 ++ tools/perf/Documentation/perf-intel-pt.txt | 63 +++++- tools/perf/util/auxtrace.c | 50 +++++ tools/perf/util/auxtrace.h | 31 ++- tools/perf/util/evsel.c | 4 + .../perf/util/intel-pt-decoder/intel-pt-decoder.c | 214 +++++++++++++++++++-- .../perf/util/intel-pt-decoder/intel-pt-decoder.h | 1 + tools/perf/util/intel-pt.c | 45 ++++- 8 files changed, 389 insertions(+), 33 deletions(-)Regards
Adrian
Signature · 1 line
- Arnaldo