• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1perf-intel-pt(1)
2================
3
4NAME
5----
6perf-intel-pt - Support for Intel Processor Trace within perf tools
7
8SYNOPSIS
9--------
10[verse]
11'perf record' -e intel_pt//
12
13DESCRIPTION
14-----------
15
16Intel Processor Trace (Intel PT) is an extension of Intel Architecture that
17collects information about software execution such as control flow, execution
18modes and timings and formats it into highly compressed binary packets.
19Technical details are documented in the Intel 64 and IA-32 Architectures
20Software Developer Manuals, Chapter 36 Intel Processor Trace.
21
22Intel PT is first supported in Intel Core M and 5th generation Intel Core
23processors that are based on the Intel micro-architecture code name Broadwell.
24
25Trace data is collected by 'perf record' and stored within the perf.data file.
26See below for options to 'perf record'.
27
28Trace data must be 'decoded' which involves walking the object code and matching
29the trace data packets. For example a TNT packet only tells whether a
30conditional branch was taken or not taken, so to make use of that packet the
31decoder must know precisely which instruction was being executed.
32
33Decoding is done on-the-fly.  The decoder outputs samples in the same format as
34samples output by perf hardware events, for example as though the "instructions"
35or "branches" events had been recorded.  Presently 3 tools support this:
36'perf script', 'perf report' and 'perf inject'.  See below for more information
37on using those tools.
38
39The main distinguishing feature of Intel PT is that the decoder can determine
40the exact flow of software execution.  Intel PT can be used to understand why
41and how did software get to a certain point, or behave a certain way.  The
42software does not have to be recompiled, so Intel PT works with debug or release
43builds, however the executed images are needed - which makes use in JIT-compiled
44environments, or with self-modified code, a challenge.  Also symbols need to be
45provided to make sense of addresses.
46
47A limitation of Intel PT is that it produces huge amounts of trace data
48(hundreds of megabytes per second per core) which takes a long time to decode,
49for example two or three orders of magnitude longer than it took to collect.
50Another limitation is the performance impact of tracing, something that will
51vary depending on the use-case and architecture.
52
53
54Quickstart
55----------
56
57It is important to start small.  That is because it is easy to capture vastly
58more data than can possibly be processed.
59
60The simplest thing to do with Intel PT is userspace profiling of small programs.
61Data is captured with 'perf record' e.g. to trace 'ls' userspace-only:
62
63	perf record -e intel_pt//u ls
64
65And profiled with 'perf report' e.g.
66
67	perf report
68
69To also trace kernel space presents a problem, namely kernel self-modifying
70code.  A fairly good kernel image is available in /proc/kcore but to get an
71accurate image a copy of /proc/kcore needs to be made under the same conditions
72as the data capture. 'perf record' can make a copy of /proc/kcore if the option
73--kcore is used, but access to /proc/kcore is restricted e.g.
74
75	sudo perf record -o pt_ls --kcore -e intel_pt// -- ls
76
77which will create a directory named 'pt_ls' and put the perf.data file (named
78simply 'data') and copies of /proc/kcore, /proc/kallsyms and /proc/modules into
79it.  The other tools understand the directory format, so to use 'perf report'
80becomes:
81
82	sudo perf report -i pt_ls
83
84Because samples are synthesized after-the-fact, the sampling period can be
85selected for reporting. e.g. sample every microsecond
86
87	sudo perf report pt_ls --itrace=i1usge
88
89See the sections below for more information about the --itrace option.
90
91Beware the smaller the period, the more samples that are produced, and the
92longer it takes to process them.
93
94Also note that the coarseness of Intel PT timing information will start to
95distort the statistical value of the sampling as the sampling period becomes
96smaller.
97
98To represent software control flow, "branches" samples are produced.  By default
99a branch sample is synthesized for every single branch.  To get an idea what
100data is available you can use the 'perf script' tool with all itrace sampling
101options, which will list all the samples.
102
103	perf record -e intel_pt//u ls
104	perf script --itrace=ibxwpe
105
106An interesting field that is not printed by default is 'flags' which can be
107displayed as follows:
108
109	perf script --itrace=ibxwpe -F+flags
110
111The flags are "bcrosyiABExghDt" which stand for branch, call, return, conditional,
112system, asynchronous, interrupt, transaction abort, trace begin, trace end,
113in transaction, VM-entry, VM-exit, interrupt disabled, and interrupt disable
114toggle respectively.
115
116perf script also supports higher level ways to dump instruction traces:
117
118	perf script --insn-trace --xed
119
120Dump all instructions. This requires installing the xed tool (see XED below)
121Dumping all instructions in a long trace can be fairly slow. It is usually better
122to start with higher level decoding, like
123
124	perf script --call-trace
125
126or
127
128	perf script --call-ret-trace
129
130and then select a time range of interest. The time range can then be examined
131in detail with
132
133	perf script --time starttime,stoptime --insn-trace --xed
134
135While examining the trace it's also useful to filter on specific CPUs using
136the -C option
137
138	perf script --time starttime,stoptime --insn-trace --xed -C 1
139
140Dump all instructions in time range on CPU 1.
141
142Another interesting field that is not printed by default is 'ipc' which can be
143displayed as follows:
144
145	perf script --itrace=be -F+ipc
146
147There are two ways that instructions-per-cycle (IPC) can be calculated depending
148on the recording.
149
150If the 'cyc' config term (see config terms section below) was used, then IPC is
151calculated using the cycle count from CYC packets, otherwise MTC packets are
152used - refer to the 'mtc' config term.  When MTC is used, however, the values
153are less accurate because the timing is less accurate.
154
155Because Intel PT does not update the cycle count on every branch or instruction,
156the values will often be zero.  When there are values, they will be the number
157of instructions and number of cycles since the last update, and thus represent
158the average IPC since the last IPC for that event type.  Note IPC for "branches"
159events is calculated separately from IPC for "instructions" events.
160
161Also note that the IPC instruction count may or may not include the current
162instruction.  If the cycle count is associated with an asynchronous branch
163(e.g. page fault or interrupt), then the instruction count does not include the
164current instruction, otherwise it does.  That is consistent with whether or not
165that instruction has retired when the cycle count is updated.
166
167Another note, in the case of "branches" events, non-taken branches are not
168presently sampled, so IPC values for them do not appear e.g. a CYC packet with a
169TNT packet that starts with a non-taken branch.  To see every possible IPC
170value, "instructions" events can be used e.g. --itrace=i0ns
171
172While it is possible to create scripts to analyze the data, an alternative
173approach is available to export the data to a sqlite or postgresql database.
174Refer to script export-to-sqlite.py or export-to-postgresql.py for more details,
175and to script exported-sql-viewer.py for an example of using the database.
176
177There is also script intel-pt-events.py which provides an example of how to
178unpack the raw data for power events and PTWRITE. The script also displays
179branches, and supports 2 additional modes selected by option:
180
181 --insn-trace - instruction trace
182 --src-trace - source trace
183
184As mentioned above, it is easy to capture too much data.  One way to limit the
185data captured is to use 'snapshot' mode which is explained further below.
186Refer to 'new snapshot option' and 'Intel PT modes of operation' further below.
187
188Another problem that will be experienced is decoder errors.  They can be caused
189by inability to access the executed image, self-modified or JIT-ed code, or the
190inability to match side-band information (such as context switches and mmaps)
191which results in the decoder not knowing what code was executed.
192
193There is also the problem of perf not being able to copy the data fast enough,
194resulting in data lost because the buffer was full.  See 'Buffer handling' below
195for more details.
196
197
198perf record
199-----------
200
201new event
202~~~~~~~~~
203
204The Intel PT kernel driver creates a new PMU for Intel PT.  PMU events are
205selected by providing the PMU name followed by the "config" separated by slashes.
206An enhancement has been made to allow default "config" e.g. the option
207
208	-e intel_pt//
209
210will use a default config value.  Currently that is the same as
211
212	-e intel_pt/tsc,noretcomp=0/
213
214which is the same as
215
216	-e intel_pt/tsc=1,noretcomp=0/
217
218Note there are now new config terms - see section 'config terms' further below.
219
220The config terms are listed in /sys/devices/intel_pt/format.  They are bit
221fields within the config member of the struct perf_event_attr which is
222passed to the kernel by the perf_event_open system call.  They correspond to bit
223fields in the IA32_RTIT_CTL MSR.  Here is a list of them and their definitions:
224
225	$ grep -H . /sys/bus/event_source/devices/intel_pt/format/*
226	/sys/bus/event_source/devices/intel_pt/format/cyc:config:1
227	/sys/bus/event_source/devices/intel_pt/format/cyc_thresh:config:19-22
228	/sys/bus/event_source/devices/intel_pt/format/mtc:config:9
229	/sys/bus/event_source/devices/intel_pt/format/mtc_period:config:14-17
230	/sys/bus/event_source/devices/intel_pt/format/noretcomp:config:11
231	/sys/bus/event_source/devices/intel_pt/format/psb_period:config:24-27
232	/sys/bus/event_source/devices/intel_pt/format/tsc:config:10
233
234Note that the default config must be overridden for each term i.e.
235
236	-e intel_pt/noretcomp=0/
237
238is the same as:
239
240	-e intel_pt/tsc=1,noretcomp=0/
241
242So, to disable TSC packets use:
243
244	-e intel_pt/tsc=0/
245
246It is also possible to specify the config value explicitly:
247
248	-e intel_pt/config=0x400/
249
250Note that, as with all events, the event is suffixed with event modifiers:
251
252	u	userspace
253	k	kernel
254	h	hypervisor
255	G	guest
256	H	host
257	p	precise ip
258
259'h', 'G' and 'H' are for virtualization which is not supported by Intel PT.
260'p' is also not relevant to Intel PT.  So only options 'u' and 'k' are
261meaningful for Intel PT.
262
263perf_event_attr is displayed if the -vv option is used e.g.
264
265	------------------------------------------------------------
266	perf_event_attr:
267	type                             6
268	size                             112
269	config                           0x400
270	{ sample_period, sample_freq }   1
271	sample_type                      IP|TID|TIME|CPU|IDENTIFIER
272	read_format                      ID
273	disabled                         1
274	inherit                          1
275	exclude_kernel                   1
276	exclude_hv                       1
277	enable_on_exec                   1
278	sample_id_all                    1
279	------------------------------------------------------------
280	sys_perf_event_open: pid 31104  cpu 0  group_fd -1  flags 0x8
281	sys_perf_event_open: pid 31104  cpu 1  group_fd -1  flags 0x8
282	sys_perf_event_open: pid 31104  cpu 2  group_fd -1  flags 0x8
283	sys_perf_event_open: pid 31104  cpu 3  group_fd -1  flags 0x8
284	------------------------------------------------------------
285
286
287config terms
288~~~~~~~~~~~~
289
290The June 2015 version of Intel 64 and IA-32 Architectures Software Developer
291Manuals, Chapter 36 Intel Processor Trace, defined new Intel PT features.
292Some of the features are reflect in new config terms.  All the config terms are
293described below.
294
295tsc		Always supported.  Produces TSC timestamp packets to provide
296		timing information.  In some cases it is possible to decode
297		without timing information, for example a per-thread context
298		that does not overlap executable memory maps.
299
300		The default config selects tsc (i.e. tsc=1).
301
302noretcomp	Always supported.  Disables "return compression" so a TIP packet
303		is produced when a function returns.  Causes more packets to be
304		produced but might make decoding more reliable.
305
306		The default config does not select noretcomp (i.e. noretcomp=0).
307
308psb_period	Allows the frequency of PSB packets to be specified.
309
310		The PSB packet is a synchronization packet that provides a
311		starting point for decoding or recovery from errors.
312
313		Support for psb_period is indicated by:
314
315			/sys/bus/event_source/devices/intel_pt/caps/psb_cyc
316
317		which contains "1" if the feature is supported and "0"
318		otherwise.
319
320		Valid values are given by:
321
322			/sys/bus/event_source/devices/intel_pt/caps/psb_periods
323
324		which contains a hexadecimal value, the bits of which represent
325		valid values e.g. bit 2 set means value 2 is valid.
326
327		The psb_period value is converted to the approximate number of
328		trace bytes between PSB packets as:
329
330			2 ^ (value + 11)
331
332		e.g. value 3 means 16KiB bytes between PSBs
333
334		If an invalid value is entered, the error message
335		will give a list of valid values e.g.
336
337			$ perf record -e intel_pt/psb_period=15/u uname
338			Invalid psb_period for intel_pt. Valid values are: 0-5
339
340		If MTC packets are selected, the default config selects a value
341		of 3 (i.e. psb_period=3) or the nearest lower value that is
342		supported (0 is always supported).  Otherwise the default is 0.
343
344		If decoding is expected to be reliable and the buffer is large
345		then a large PSB period can be used.
346
347		Because a TSC packet is produced with PSB, the PSB period can
348		also affect the granularity to timing information in the absence
349		of MTC or CYC.
350
351mtc		Produces MTC timing packets.
352
353		MTC packets provide finer grain timestamp information than TSC
354		packets.  MTC packets record time using the hardware crystal
355		clock (CTC) which is related to TSC packets using a TMA packet.
356
357		Support for this feature is indicated by:
358
359			/sys/bus/event_source/devices/intel_pt/caps/mtc
360
361		which contains "1" if the feature is supported and
362		"0" otherwise.
363
364		The frequency of MTC packets can also be specified - see
365		mtc_period below.
366
367mtc_period	Specifies how frequently MTC packets are produced - see mtc
368		above for how to determine if MTC packets are supported.
369
370		Valid values are given by:
371
372			/sys/bus/event_source/devices/intel_pt/caps/mtc_periods
373
374		which contains a hexadecimal value, the bits of which represent
375		valid values e.g. bit 2 set means value 2 is valid.
376
377		The mtc_period value is converted to the MTC frequency as:
378
379			CTC-frequency / (2 ^ value)
380
381		e.g. value 3 means one eighth of CTC-frequency
382
383		Where CTC is the hardware crystal clock, the frequency of which
384		can be related to TSC via values provided in cpuid leaf 0x15.
385
386		If an invalid value is entered, the error message
387		will give a list of valid values e.g.
388
389			$ perf record -e intel_pt/mtc_period=15/u uname
390			Invalid mtc_period for intel_pt. Valid values are: 0,3,6,9
391
392		The default value is 3 or the nearest lower value
393		that is supported (0 is always supported).
394
395cyc		Produces CYC timing packets.
396
397		CYC packets provide even finer grain timestamp information than
398		MTC and TSC packets.  A CYC packet contains the number of CPU
399		cycles since the last CYC packet. Unlike MTC and TSC packets,
400		CYC packets are only sent when another packet is also sent.
401
402		Support for this feature is indicated by:
403
404			/sys/bus/event_source/devices/intel_pt/caps/psb_cyc
405
406		which contains "1" if the feature is supported and
407		"0" otherwise.
408
409		The number of CYC packets produced can be reduced by specifying
410		a threshold - see cyc_thresh below.
411
412cyc_thresh	Specifies how frequently CYC packets are produced - see cyc
413		above for how to determine if CYC packets are supported.
414
415		Valid cyc_thresh values are given by:
416
417			/sys/bus/event_source/devices/intel_pt/caps/cycle_thresholds
418
419		which contains a hexadecimal value, the bits of which represent
420		valid values e.g. bit 2 set means value 2 is valid.
421
422		The cyc_thresh value represents the minimum number of CPU cycles
423		that must have passed before a CYC packet can be sent.  The
424		number of CPU cycles is:
425
426			2 ^ (value - 1)
427
428		e.g. value 4 means 8 CPU cycles must pass before a CYC packet
429		can be sent.  Note a CYC packet is still only sent when another
430		packet is sent, not at, e.g. every 8 CPU cycles.
431
432		If an invalid value is entered, the error message
433		will give a list of valid values e.g.
434
435			$ perf record -e intel_pt/cyc,cyc_thresh=15/u uname
436			Invalid cyc_thresh for intel_pt. Valid values are: 0-12
437
438		CYC packets are not requested by default.
439
440pt		Specifies pass-through which enables the 'branch' config term.
441
442		The default config selects 'pt' if it is available, so a user will
443		never need to specify this term.
444
445branch		Enable branch tracing.  Branch tracing is enabled by default so to
446		disable branch tracing use 'branch=0'.
447
448		The default config selects 'branch' if it is available.
449
450ptw		Enable PTWRITE packets which are produced when a ptwrite instruction
451		is executed.
452
453		Support for this feature is indicated by:
454
455			/sys/bus/event_source/devices/intel_pt/caps/ptwrite
456
457		which contains "1" if the feature is supported and
458		"0" otherwise.
459
460		As an alternative, refer to "Emulated PTWRITE" further below.
461
462fup_on_ptw	Enable a FUP packet to follow the PTWRITE packet.  The FUP packet
463		provides the address of the ptwrite instruction.  In the absence of
464		fup_on_ptw, the decoder will use the address of the previous branch
465		if branch tracing is enabled, otherwise the address will be zero.
466		Note that fup_on_ptw will work even when branch tracing is disabled.
467
468pwr_evt		Enable power events.  The power events provide information about
469		changes to the CPU C-state.
470
471		Support for this feature is indicated by:
472
473			/sys/bus/event_source/devices/intel_pt/caps/power_event_trace
474
475		which contains "1" if the feature is supported and
476		"0" otherwise.
477
478event		Enable Event Trace.  The events provide information about asynchronous
479		events.
480
481		Support for this feature is indicated by:
482
483			/sys/bus/event_source/devices/intel_pt/caps/event_trace
484
485		which contains "1" if the feature is supported and
486		"0" otherwise.
487
488notnt		Disable TNT packets.  Without TNT packets, it is not possible to walk
489		executable code to reconstruct control flow, however FUP, TIP, TIP.PGE
490		and TIP.PGD packets still indicate asynchronous control flow, and (if
491		return compression is disabled - see noretcomp) return statements.
492		The advantage of eliminating TNT packets is reducing the size of the
493		trace and corresponding tracing overhead.
494
495		Support for this feature is indicated by:
496
497			/sys/bus/event_source/devices/intel_pt/caps/tnt_disable
498
499		which contains "1" if the feature is supported and
500		"0" otherwise.
501
502
503AUX area sampling option
504~~~~~~~~~~~~~~~~~~~~~~~~
505
506To select Intel PT "sampling" the AUX area sampling option can be used:
507
508	--aux-sample
509
510Optionally it can be followed by the sample size in bytes e.g.
511
512	--aux-sample=8192
513
514In addition, the Intel PT event to sample must be defined e.g.
515
516	-e intel_pt//u
517
518Samples on other events will be created containing Intel PT data e.g. the
519following will create Intel PT samples on the branch-misses event, note the
520events must be grouped using {}:
521
522	perf record --aux-sample -e '{intel_pt//u,branch-misses:u}'
523
524An alternative to '--aux-sample' is to add the config term 'aux-sample-size' to
525events.  In this case, the grouping is implied e.g.
526
527	perf record -e intel_pt//u -e branch-misses/aux-sample-size=8192/u
528
529is the same as:
530
531	perf record -e '{intel_pt//u,branch-misses/aux-sample-size=8192/u}'
532
533but allows for also using an address filter e.g.:
534
535	perf record -e intel_pt//u --filter 'filter * @/bin/ls' -e branch-misses/aux-sample-size=8192/u -- ls
536
537It is important to select a sample size that is big enough to contain at least
538one PSB packet.  If not a warning will be displayed:
539
540	Intel PT sample size (%zu) may be too small for PSB period (%zu)
541
542The calculation used for that is: if sample_size <= psb_period + 256 display the
543warning.  When sampling is used, psb_period defaults to 0 (2KiB).
544
545The default sample size is 4KiB.
546
547The sample size is passed in aux_sample_size in struct perf_event_attr.  The
548sample size is limited by the maximum event size which is 64KiB.  It is
549difficult to know how big the event might be without the trace sample attached,
550but the tool validates that the sample size is not greater than 60KiB.
551
552
553new snapshot option
554~~~~~~~~~~~~~~~~~~~
555
556The difference between full trace and snapshot from the kernel's perspective is
557that in full trace we don't overwrite trace data that the user hasn't collected
558yet (and indicated that by advancing aux_tail), whereas in snapshot mode we let
559the trace run and overwrite older data in the buffer so that whenever something
560interesting happens, we can stop it and grab a snapshot of what was going on
561around that interesting moment.
562
563To select snapshot mode a new option has been added:
564
565	-S
566
567Optionally it can be followed by the snapshot size e.g.
568
569	-S0x100000
570
571The default snapshot size is the auxtrace mmap size.  If neither auxtrace mmap size
572nor snapshot size is specified, then the default is 4MiB for privileged users
573(or if /proc/sys/kernel/perf_event_paranoid < 0), 128KiB for unprivileged users.
574If an unprivileged user does not specify mmap pages, the mmap pages will be
575reduced as described in the 'new auxtrace mmap size option' section below.
576
577The snapshot size is displayed if the option -vv is used e.g.
578
579	Intel PT snapshot size: %zu
580
581
582new auxtrace mmap size option
583~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
584
585Intel PT buffer size is specified by an addition to the -m option e.g.
586
587	-m,16
588
589selects a buffer size of 16 pages i.e. 64KiB.
590
591Note that the existing functionality of -m is unchanged.  The auxtrace mmap size
592is specified by the optional addition of a comma and the value.
593
594The default auxtrace mmap size for Intel PT is 4MiB/page_size for privileged users
595(or if /proc/sys/kernel/perf_event_paranoid < 0), 128KiB for unprivileged users.
596If an unprivileged user does not specify mmap pages, the mmap pages will be
597reduced from the default 512KiB/page_size to 256KiB/page_size, otherwise the
598user is likely to get an error as they exceed their mlock limit (Max locked
599memory as shown in /proc/self/limits).  Note that perf does not count the first
600512KiB (actually /proc/sys/kernel/perf_event_mlock_kb minus 1 page) per cpu
601against the mlock limit so an unprivileged user is allowed 512KiB per cpu plus
602their mlock limit (which defaults to 64KiB but is not multiplied by the number
603of cpus).
604
605In full-trace mode, powers of two are allowed for buffer size, with a minimum
606size of 2 pages.  In snapshot mode or sampling mode, it is the same but the
607minimum size is 1 page.
608
609The mmap size and auxtrace mmap size are displayed if the -vv option is used e.g.
610
611	mmap length 528384
612	auxtrace mmap length 4198400
613
614
615Intel PT modes of operation
616~~~~~~~~~~~~~~~~~~~~~~~~~~~
617
618Intel PT can be used in 3 modes:
619	full-trace mode
620	sample mode
621	snapshot mode
622
623Full-trace mode traces continuously e.g.
624
625	perf record -e intel_pt//u uname
626
627Sample mode attaches a Intel PT sample to other events e.g.
628
629	perf record --aux-sample -e intel_pt//u -e branch-misses:u
630
631Snapshot mode captures the available data when a signal is sent or "snapshot"
632control command is issued. e.g. using a signal
633
634	perf record -v -e intel_pt//u -S ./loopy 1000000000 &
635	[1] 11435
636	kill -USR2 11435
637	Recording AUX area tracing snapshot
638
639Note that the signal sent is SIGUSR2.
640Note that "Recording AUX area tracing snapshot" is displayed because the -v
641option is used.
642
643The advantage of using "snapshot" control command is that the access is
644controlled by access to a FIFO e.g.
645
646	$ mkfifo perf.control
647	$ mkfifo perf.ack
648	$ cat perf.ack &
649	[1] 15235
650	$ sudo ~/bin/perf record --control fifo:perf.control,perf.ack -S -e intel_pt//u -- sleep 60 &
651	[2] 15243
652	$ ps -e | grep perf
653	15244 pts/1    00:00:00 perf
654	$ kill -USR2 15244
655	bash: kill: (15244) - Operation not permitted
656	$ echo snapshot > perf.control
657	ack
658
659The 3 Intel PT modes of operation cannot be used together.
660
661
662Buffer handling
663~~~~~~~~~~~~~~~
664
665There may be buffer limitations (i.e. single ToPa entry) which means that actual
666buffer sizes are limited to powers of 2 up to 4MiB (MAX_ORDER).  In order to
667provide other sizes, and in particular an arbitrarily large size, multiple
668buffers are logically concatenated.  However an interrupt must be used to switch
669between buffers.  That has two potential problems:
670	a) the interrupt may not be handled in time so that the current buffer
671	becomes full and some trace data is lost.
672	b) the interrupts may slow the system and affect the performance
673	results.
674
675If trace data is lost, the driver sets 'truncated' in the PERF_RECORD_AUX event
676which the tools report as an error.
677
678In full-trace mode, the driver waits for data to be copied out before allowing
679the (logical) buffer to wrap-around.  If data is not copied out quickly enough,
680again 'truncated' is set in the PERF_RECORD_AUX event.  If the driver has to
681wait, the intel_pt event gets disabled.  Because it is difficult to know when
682that happens, perf tools always re-enable the intel_pt event after copying out
683data.
684
685
686Intel PT and build ids
687~~~~~~~~~~~~~~~~~~~~~~
688
689By default "perf record" post-processes the event stream to find all build ids
690for executables for all addresses sampled.  Deliberately, Intel PT is not
691decoded for that purpose (it would take too long).  Instead the build ids for
692all executables encountered (due to mmap, comm or task events) are included
693in the perf.data file.
694
695To see buildids included in the perf.data file use the command:
696
697	perf buildid-list
698
699If the perf.data file contains Intel PT data, that is the same as:
700
701	perf buildid-list --with-hits
702
703
704Snapshot mode and event disabling
705~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
706
707In order to make a snapshot, the intel_pt event is disabled using an IOCTL,
708namely PERF_EVENT_IOC_DISABLE.  However doing that can also disable the
709collection of side-band information.  In order to prevent that,  a dummy
710software event has been introduced that permits tracking events (like mmaps) to
711continue to be recorded while intel_pt is disabled.  That is important to ensure
712there is complete side-band information to allow the decoding of subsequent
713snapshots.
714
715A test has been created for that.  To find the test:
716
717	perf test list
718	...
719	23: Test using a dummy software event to keep tracking
720
721To run the test:
722
723	perf test 23
724	23: Test using a dummy software event to keep tracking     : Ok
725
726
727perf record modes (nothing new here)
728~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
729
730perf record essentially operates in one of three modes:
731	per thread
732	per cpu
733	workload only
734
735"per thread" mode is selected by -t or by --per-thread (with -p or -u or just a
736workload).
737"per cpu" is selected by -C or -a.
738"workload only" mode is selected by not using the other options but providing a
739command to run (i.e. the workload).
740
741In per-thread mode an exact list of threads is traced.  There is no inheritance.
742Each thread has its own event buffer.
743
744In per-cpu mode all processes (or processes from the selected cgroup i.e. -G
745option, or processes selected with -p or -u) are traced.  Each cpu has its own
746buffer. Inheritance is allowed.
747
748In workload-only mode, the workload is traced but with per-cpu buffers.
749Inheritance is allowed.  Note that you can now trace a workload in per-thread
750mode by using the --per-thread option.
751
752
753Privileged vs non-privileged users
754~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
755
756Unless /proc/sys/kernel/perf_event_paranoid is set to -1, unprivileged users
757have memory limits imposed upon them.  That affects what buffer sizes they can
758have as outlined above.
759
760The v4.2 kernel introduced support for a context switch metadata event,
761PERF_RECORD_SWITCH, which allows unprivileged users to see when their processes
762are scheduled out and in, just not by whom, which is left for the
763PERF_RECORD_SWITCH_CPU_WIDE, that is only accessible in system wide context,
764which in turn requires CAP_PERFMON or CAP_SYS_ADMIN.
765
766Please see the 45ac1403f564 ("perf: Add PERF_RECORD_SWITCH to indicate context
767switches") commit, that introduces these metadata events for further info.
768
769When working with kernels < v4.2, the following considerations must be taken,
770as the sched:sched_switch tracepoints will be used to receive such information:
771
772Unless /proc/sys/kernel/perf_event_paranoid is set to -1, unprivileged users are
773not permitted to use tracepoints which means there is insufficient side-band
774information to decode Intel PT in per-cpu mode, and potentially workload-only
775mode too if the workload creates new processes.
776
777Note also, that to use tracepoints, read-access to debugfs is required.  So if
778debugfs is not mounted or the user does not have read-access, it will again not
779be possible to decode Intel PT in per-cpu mode.
780
781
782sched_switch tracepoint
783~~~~~~~~~~~~~~~~~~~~~~~
784
785The sched_switch tracepoint is used to provide side-band data for Intel PT
786decoding in kernels where the PERF_RECORD_SWITCH metadata event isn't
787available.
788
789The sched_switch events are automatically added. e.g. the second event shown
790below:
791
792	$ perf record -vv -e intel_pt//u uname
793	------------------------------------------------------------
794	perf_event_attr:
795	type                             6
796	size                             112
797	config                           0x400
798	{ sample_period, sample_freq }   1
799	sample_type                      IP|TID|TIME|CPU|IDENTIFIER
800	read_format                      ID
801	disabled                         1
802	inherit                          1
803	exclude_kernel                   1
804	exclude_hv                       1
805	enable_on_exec                   1
806	sample_id_all                    1
807	------------------------------------------------------------
808	sys_perf_event_open: pid 31104  cpu 0  group_fd -1  flags 0x8
809	sys_perf_event_open: pid 31104  cpu 1  group_fd -1  flags 0x8
810	sys_perf_event_open: pid 31104  cpu 2  group_fd -1  flags 0x8
811	sys_perf_event_open: pid 31104  cpu 3  group_fd -1  flags 0x8
812	------------------------------------------------------------
813	perf_event_attr:
814	type                             2
815	size                             112
816	config                           0x108
817	{ sample_period, sample_freq }   1
818	sample_type                      IP|TID|TIME|CPU|PERIOD|RAW|IDENTIFIER
819	read_format                      ID
820	inherit                          1
821	sample_id_all                    1
822	exclude_guest                    1
823	------------------------------------------------------------
824	sys_perf_event_open: pid -1  cpu 0  group_fd -1  flags 0x8
825	sys_perf_event_open: pid -1  cpu 1  group_fd -1  flags 0x8
826	sys_perf_event_open: pid -1  cpu 2  group_fd -1  flags 0x8
827	sys_perf_event_open: pid -1  cpu 3  group_fd -1  flags 0x8
828	------------------------------------------------------------
829	perf_event_attr:
830	type                             1
831	size                             112
832	config                           0x9
833	{ sample_period, sample_freq }   1
834	sample_type                      IP|TID|TIME|IDENTIFIER
835	read_format                      ID
836	disabled                         1
837	inherit                          1
838	exclude_kernel                   1
839	exclude_hv                       1
840	mmap                             1
841	comm                             1
842	enable_on_exec                   1
843	task                             1
844	sample_id_all                    1
845	mmap2                            1
846	comm_exec                        1
847	------------------------------------------------------------
848	sys_perf_event_open: pid 31104  cpu 0  group_fd -1  flags 0x8
849	sys_perf_event_open: pid 31104  cpu 1  group_fd -1  flags 0x8
850	sys_perf_event_open: pid 31104  cpu 2  group_fd -1  flags 0x8
851	sys_perf_event_open: pid 31104  cpu 3  group_fd -1  flags 0x8
852	mmap size 528384B
853	AUX area mmap length 4194304
854	perf event ring buffer mmapped per cpu
855	Synthesizing auxtrace information
856	Linux
857	[ perf record: Woken up 1 times to write data ]
858	[ perf record: Captured and wrote 0.042 MB perf.data ]
859
860Note, the sched_switch event is only added if the user is permitted to use it
861and only in per-cpu mode.
862
863Note also, the sched_switch event is only added if TSC packets are requested.
864That is because, in the absence of timing information, the sched_switch events
865cannot be matched against the Intel PT trace.
866
867
868perf script
869-----------
870
871By default, perf script will decode trace data found in the perf.data file.
872This can be further controlled by new option --itrace.
873
874
875New --itrace option
876~~~~~~~~~~~~~~~~~~~
877
878Having no option is the same as
879
880	--itrace
881
882which, in turn, is the same as
883
884	--itrace=cepwx
885
886The letters are:
887
888	i	synthesize "instructions" events
889	b	synthesize "branches" events
890	x	synthesize "transactions" events
891	w	synthesize "ptwrite" events
892	p	synthesize "power" events (incl. PSB events)
893	c	synthesize branches events (calls only)
894	r	synthesize branches events (returns only)
895	o	synthesize PEBS-via-PT events
896	I	synthesize Event Trace events
897	e	synthesize tracing error events
898	d	create a debug log
899	g	synthesize a call chain (use with i or x)
900	G	synthesize a call chain on existing event records
901	l	synthesize last branch entries (use with i or x)
902	L	synthesize last branch entries on existing event records
903	s	skip initial number of events
904	q	quicker (less detailed) decoding
905	Z	prefer to ignore timestamps (so-called "timeless" decoding)
906
907"Instructions" events look like they were recorded by "perf record -e
908instructions".
909
910"Branches" events look like they were recorded by "perf record -e branches". "c"
911and "r" can be combined to get calls and returns.
912
913"Transactions" events correspond to the start or end of transactions. The
914'flags' field can be used in perf script to determine whether the event is a
915transaction start, commit or abort.
916
917Note that "instructions", "branches" and "transactions" events depend on code
918flow packets which can be disabled by using the config term "branch=0".  Refer
919to the config terms section above.
920
921"ptwrite" events record the payload of the ptwrite instruction and whether
922"fup_on_ptw" was used.  "ptwrite" events depend on PTWRITE packets which are
923recorded only if the "ptw" config term was used.  Refer to the config terms
924section above.  perf script "synth" field displays "ptwrite" information like
925this: "ip: 0 payload: 0x123456789abcdef0"  where "ip" is 1 if "fup_on_ptw" was
926used.
927
928"Power" events correspond to power event packets and CBR (core-to-bus ratio)
929packets.  While CBR packets are always recorded when tracing is enabled, power
930event packets are recorded only if the "pwr_evt" config term was used.  Refer to
931the config terms section above.  The power events record information about
932C-state changes, whereas CBR is indicative of CPU frequency.  perf script
933"event,synth" fields display information like this:
934	cbr:  cbr: 22 freq: 2189 MHz (200%)
935	mwait:  hints: 0x60 extensions: 0x1
936	pwre:  hw: 0 cstate: 2 sub-cstate: 0
937	exstop:  ip: 1
938	pwrx:  deepest cstate: 2 last cstate: 2 wake reason: 0x4
939Where:
940	"cbr" includes the frequency and the percentage of maximum non-turbo
941	"mwait" shows mwait hints and extensions
942	"pwre" shows C-state transitions (to a C-state deeper than C0) and
943	whether	initiated by hardware
944	"exstop" indicates execution stopped and whether the IP was recorded
945	exactly,
946	"pwrx" indicates return to C0
947For more details refer to the Intel 64 and IA-32 Architectures Software
948Developer Manuals.
949
950PSB events show when a PSB+ occurred and also the byte-offset in the trace.
951Emitting a PSB+ can cause a CPU a slight delay. When doing timing analysis
952of code with Intel PT, it is useful to know if a timing bubble was caused
953by Intel PT or not.
954
955Error events show where the decoder lost the trace.  Error events
956are quite important.  Users must know if what they are seeing is a complete
957picture or not. The "e" option may be followed by flags which affect what errors
958will or will not be reported.  Each flag must be preceded by either '+' or '-'.
959The flags supported by Intel PT are:
960		-o	Suppress overflow errors
961		-l	Suppress trace data lost errors
962For example, for errors but not overflow or data lost errors:
963
964	--itrace=e-o-l
965
966The "d" option will cause the creation of a file "intel_pt.log" containing all
967decoded packets and instructions.  Note that this option slows down the decoder
968and that the resulting file may be very large.  The "d" option may be followed
969by flags which affect what debug messages will or will not be logged. Each flag
970must be preceded by either '+' or '-'. The flags support by Intel PT are:
971		-a	Suppress logging of perf events
972		+a	Log all perf events
973By default, logged perf events are filtered by any specified time ranges, but
974flag +a overrides that.
975
976In addition, the period of the "instructions" event can be specified. e.g.
977
978	--itrace=i10us
979
980sets the period to 10us i.e. one  instruction sample is synthesized for each 10
981microseconds of trace.  Alternatives to "us" are "ms" (milliseconds),
982"ns" (nanoseconds), "t" (TSC ticks) or "i" (instructions).
983
984"ms", "us" and "ns" are converted to TSC ticks.
985
986The timing information included with Intel PT does not give the time of every
987instruction.  Consequently, for the purpose of sampling, the decoder estimates
988the time since the last timing packet based on 1 tick per instruction.  The time
989on the sample is *not* adjusted and reflects the last known value of TSC.
990
991For Intel PT, the default period is 100us.
992
993Setting it to a zero period means "as often as possible".
994
995In the case of Intel PT that is the same as a period of 1 and a unit of
996'instructions' (i.e. --itrace=i1i).
997
998Also the call chain size (default 16, max. 1024) for instructions or
999transactions events can be specified. e.g.
1000
1001	--itrace=ig32
1002	--itrace=xg32
1003
1004Also the number of last branch entries (default 64, max. 1024) for instructions or
1005transactions events can be specified. e.g.
1006
1007       --itrace=il10
1008       --itrace=xl10
1009
1010Note that last branch entries are cleared for each sample, so there is no overlap
1011from one sample to the next.
1012
1013The G and L options are designed in particular for sample mode, and work much
1014like g and l but add call chain and branch stack to the other selected events
1015instead of synthesized events. For example, to record branch-misses events for
1016'ls' and then add a call chain derived from the Intel PT trace:
1017
1018	perf record --aux-sample -e '{intel_pt//u,branch-misses:u}' -- ls
1019	perf report --itrace=Ge
1020
1021Although in fact G is a default for perf report, so that is the same as just:
1022
1023	perf report
1024
1025One caveat with the G and L options is that they work poorly with "Large PEBS".
1026Large PEBS means PEBS records will be accumulated by hardware and the written
1027into the event buffer in one go.  That reduces interrupts, but can give very
1028late timestamps.  Because the Intel PT trace is synchronized by timestamps,
1029the PEBS events do not match the trace.  Currently, Large PEBS is used only in
1030certain circumstances:
1031	- hardware supports it
1032	- PEBS is used
1033	- event period is specified, instead of frequency
1034	- the sample type is limited to the following flags:
1035		PERF_SAMPLE_IP | PERF_SAMPLE_TID | PERF_SAMPLE_ADDR |
1036		PERF_SAMPLE_ID | PERF_SAMPLE_CPU | PERF_SAMPLE_STREAM_ID |
1037		PERF_SAMPLE_DATA_SRC | PERF_SAMPLE_IDENTIFIER |
1038		PERF_SAMPLE_TRANSACTION | PERF_SAMPLE_PHYS_ADDR |
1039		PERF_SAMPLE_REGS_INTR | PERF_SAMPLE_REGS_USER |
1040		PERF_SAMPLE_PERIOD (and sometimes) | PERF_SAMPLE_TIME
1041Because Intel PT sample mode uses a different sample type to the list above,
1042Large PEBS is not used with Intel PT sample mode. To avoid Large PEBS in other
1043cases, avoid specifying the event period i.e. avoid the 'perf record' -c option,
1044--count option, or 'period' config term.
1045
1046To disable trace decoding entirely, use the option --no-itrace.
1047
1048It is also possible to skip events generated (instructions, branches, transactions)
1049at the beginning. This is useful to ignore initialization code.
1050
1051	--itrace=i0nss1000000
1052
1053skips the first million instructions.
1054
1055The q option changes the way the trace is decoded.  The decoding is much faster
1056but much less detailed.  Specifically, with the q option, the decoder does not
1057decode TNT packets, and does not walk object code, but gets the ip from FUP and
1058TIP packets.  The q option can be used with the b and i options but the period
1059is not used.  The q option decodes more quickly, but is useful only if the
1060control flow of interest is represented or indicated by FUP, TIP, TIP.PGE, or
1061TIP.PGD packets (refer below).  However the q option could be used to find time
1062ranges that could then be decoded fully using the --time option.
1063
1064What will *not* be decoded with the (single) q option:
1065
1066	- direct calls and jmps
1067	- conditional branches
1068	- non-branch instructions
1069
1070What *will* be decoded with the (single) q option:
1071
1072	- asynchronous branches such as interrupts
1073	- indirect branches
1074	- function return target address *if* the noretcomp config term (refer
1075	config terms section) was used
1076	- start of (control-flow) tracing
1077	- end of (control-flow) tracing, if it is not out of context
1078	- power events, ptwrite, transaction start and abort
1079	- instruction pointer associated with PSB packets
1080
1081Note the q option does not specify what events will be synthesized e.g. the p
1082option must be used also to show power events.
1083
1084Repeating the q option (double-q i.e. qq) results in even faster decoding and even
1085less detail.  The decoder decodes only extended PSB (PSB+) packets, getting the
1086instruction pointer if there is a FUP packet within PSB+ (i.e. between PSB and
1087PSBEND).  Note PSB packets occur regularly in the trace based on the psb_period
1088config term (refer config terms section).  There will be a FUP packet if the
1089PSB+ occurs while control flow is being traced.
1090
1091What will *not* be decoded with the qq option:
1092
1093	- everything except instruction pointer associated with PSB packets
1094
1095What *will* be decoded with the qq option:
1096
1097	- instruction pointer associated with PSB packets
1098
1099The Z option is equivalent to having recorded a trace without TSC
1100(i.e. config term tsc=0). It can be useful to avoid timestamp issues when
1101decoding a trace of a virtual machine.
1102
1103
1104dump option
1105~~~~~~~~~~~
1106
1107perf script has an option (-D) to "dump" the events i.e. display the binary
1108data.
1109
1110When -D is used, Intel PT packets are displayed.  The packet decoder does not
1111pay attention to PSB packets, but just decodes the bytes - so the packets seen
1112by the actual decoder may not be identical in places where the data is corrupt.
1113One example of that would be when the buffer-switching interrupt has been too
1114slow, and the buffer has been filled completely.  In that case, the last packet
1115in the buffer might be truncated and immediately followed by a PSB as the trace
1116continues in the next buffer.
1117
1118To disable the display of Intel PT packets, combine the -D option with
1119--no-itrace.
1120
1121
1122perf report
1123-----------
1124
1125By default, perf report will decode trace data found in the perf.data file.
1126This can be further controlled by new option --itrace exactly the same as
1127perf script, with the exception that the default is --itrace=igxe.
1128
1129
1130perf inject
1131-----------
1132
1133perf inject also accepts the --itrace option in which case tracing data is
1134removed and replaced with the synthesized events. e.g.
1135
1136	perf inject --itrace -i perf.data -o perf.data.new
1137
1138Below is an example of using Intel PT with autofdo.  It requires autofdo
1139(https://github.com/google/autofdo) and gcc version 5.  The bubble
1140sort example is from the AutoFDO tutorial (https://gcc.gnu.org/wiki/AutoFDO/Tutorial)
1141amended to take the number of elements as a parameter.
1142
1143	$ gcc-5 -O3 sort.c -o sort_optimized
1144	$ ./sort_optimized 30000
1145	Bubble sorting array of 30000 elements
1146	2254 ms
1147
1148	$ cat ~/.perfconfig
1149	[intel-pt]
1150		mispred-all = on
1151
1152	$ perf record -e intel_pt//u ./sort 3000
1153	Bubble sorting array of 3000 elements
1154	58 ms
1155	[ perf record: Woken up 2 times to write data ]
1156	[ perf record: Captured and wrote 3.939 MB perf.data ]
1157	$ perf inject -i perf.data -o inj --itrace=i100usle --strip
1158	$ ./create_gcov --binary=./sort --profile=inj --gcov=sort.gcov -gcov_version=1
1159	$ gcc-5 -O3 -fauto-profile=sort.gcov sort.c -o sort_autofdo
1160	$ ./sort_autofdo 30000
1161	Bubble sorting array of 30000 elements
1162	2155 ms
1163
1164Note there is currently no advantage to using Intel PT instead of LBR, but
1165that may change in the future if greater use is made of the data.
1166
1167
1168PEBS via Intel PT
1169-----------------
1170
1171Some hardware has the feature to redirect PEBS records to the Intel PT trace.
1172Recording is selected by using the aux-output config term e.g.
1173
1174	perf record -c 10000 -e '{intel_pt/branch=0/,cycles/aux-output/ppp}' uname
1175
1176Note that currently, software only supports redirecting at most one PEBS event.
1177
1178To display PEBS events from the Intel PT trace, use the itrace 'o' option e.g.
1179
1180	perf script --itrace=oe
1181
1182XED
1183---
1184
1185include::build-xed.txt[]
1186
1187
1188Tracing Virtual Machines
1189------------------------
1190
1191Currently, only kernel tracing is supported and only with either "timeless" decoding
1192(i.e. no TSC timestamps) or VM Time Correlation. VM Time Correlation is an extra step
1193using 'perf inject' and requires unchanging VMX TSC Offset and no VMX TSC Scaling.
1194
1195Other limitations and caveats
1196
1197 VMX controls may suppress packets needed for decoding resulting in decoding errors
1198 VMX controls may block the perf NMI to the host potentially resulting in lost trace data
1199 Guest kernel self-modifying code (e.g. jump labels or JIT-compiled eBPF) will result in decoding errors
1200 Guest thread information is unknown
1201 Guest VCPU is unknown but may be able to be inferred from the host thread
1202 Callchains are not supported
1203
1204Example using "timeless" decoding
1205
1206Start VM
1207
1208 $ sudo virsh start kubuntu20.04
1209 Domain kubuntu20.04 started
1210
1211Mount the guest file system.  Note sshfs needs -o direct_io to enable reading of proc files.  root access is needed to read /proc/kcore.
1212
1213 $ mkdir vm0
1214 $ sshfs -o direct_io root@vm0:/ vm0
1215
1216Copy the guest /proc/kallsyms, /proc/modules and /proc/kcore
1217
1218 $ perf buildid-cache -v --kcore vm0/proc/kcore
1219 kcore added to build-id cache directory /home/user/.debug/[kernel.kcore]/9600f316a53a0f54278885e8d9710538ec5f6a08/2021021807494306
1220 $ KALLSYMS=/home/user/.debug/[kernel.kcore]/9600f316a53a0f54278885e8d9710538ec5f6a08/2021021807494306/kallsyms
1221
1222Find the VM process
1223
1224 $ ps -eLl | grep 'KVM\|PID'
1225 F S   UID     PID    PPID     LWP  C PRI  NI ADDR SZ WCHAN  TTY          TIME CMD
1226 3 S 64055    1430       1    1440  1  80   0 - 1921718 -    ?        00:02:47 CPU 0/KVM
1227 3 S 64055    1430       1    1441  1  80   0 - 1921718 -    ?        00:02:41 CPU 1/KVM
1228 3 S 64055    1430       1    1442  1  80   0 - 1921718 -    ?        00:02:38 CPU 2/KVM
1229 3 S 64055    1430       1    1443  2  80   0 - 1921718 -    ?        00:03:18 CPU 3/KVM
1230
1231Start an open-ended perf record, tracing the VM process, do something on the VM, and then ctrl-C to stop.
1232TSC is not supported and tsc=0 must be specified.  That means mtc is useless, so add mtc=0.
1233However, IPC can still be determined, hence cyc=1 can be added.
1234Only kernel decoding is supported, so 'k' must be specified.
1235Intel PT traces both the host and the guest so --guest and --host need to be specified.
1236Without timestamps, --per-thread must be specified to distinguish threads.
1237
1238 $ sudo perf kvm --guest --host --guestkallsyms $KALLSYMS record --kcore -e intel_pt/tsc=0,mtc=0,cyc=1/k -p 1430 --per-thread
1239 ^C
1240 [ perf record: Woken up 1 times to write data ]
1241 [ perf record: Captured and wrote 5.829 MB ]
1242
1243perf script can be used to provide an instruction trace
1244
1245 $ perf script --guestkallsyms $KALLSYMS --insn-trace --xed -F+ipc | grep -C10 vmresume | head -21
1246       CPU 0/KVM  1440  ffffffff82133cdd __vmx_vcpu_run+0x3d ([kernel.kallsyms])                movq  0x48(%rax), %r9
1247       CPU 0/KVM  1440  ffffffff82133ce1 __vmx_vcpu_run+0x41 ([kernel.kallsyms])                movq  0x50(%rax), %r10
1248       CPU 0/KVM  1440  ffffffff82133ce5 __vmx_vcpu_run+0x45 ([kernel.kallsyms])                movq  0x58(%rax), %r11
1249       CPU 0/KVM  1440  ffffffff82133ce9 __vmx_vcpu_run+0x49 ([kernel.kallsyms])                movq  0x60(%rax), %r12
1250       CPU 0/KVM  1440  ffffffff82133ced __vmx_vcpu_run+0x4d ([kernel.kallsyms])                movq  0x68(%rax), %r13
1251       CPU 0/KVM  1440  ffffffff82133cf1 __vmx_vcpu_run+0x51 ([kernel.kallsyms])                movq  0x70(%rax), %r14
1252       CPU 0/KVM  1440  ffffffff82133cf5 __vmx_vcpu_run+0x55 ([kernel.kallsyms])                movq  0x78(%rax), %r15
1253       CPU 0/KVM  1440  ffffffff82133cf9 __vmx_vcpu_run+0x59 ([kernel.kallsyms])                movq  (%rax), %rax
1254       CPU 0/KVM  1440  ffffffff82133cfc __vmx_vcpu_run+0x5c ([kernel.kallsyms])                callq  0xffffffff82133c40
1255       CPU 0/KVM  1440  ffffffff82133c40 vmx_vmenter+0x0 ([kernel.kallsyms])            jz 0xffffffff82133c46
1256       CPU 0/KVM  1440  ffffffff82133c42 vmx_vmenter+0x2 ([kernel.kallsyms])            vmresume         IPC: 0.11 (50/445)
1257           :1440  1440  ffffffffbb678b06 native_write_msr+0x6 ([guest.kernel.kallsyms])                 nopl  %eax, (%rax,%rax,1)
1258           :1440  1440  ffffffffbb678b0b native_write_msr+0xb ([guest.kernel.kallsyms])                 retq     IPC: 0.04 (2/41)
1259           :1440  1440  ffffffffbb666646 lapic_next_deadline+0x26 ([guest.kernel.kallsyms])             data16 nop
1260           :1440  1440  ffffffffbb666648 lapic_next_deadline+0x28 ([guest.kernel.kallsyms])             xor %eax, %eax
1261           :1440  1440  ffffffffbb66664a lapic_next_deadline+0x2a ([guest.kernel.kallsyms])             popq  %rbp
1262           :1440  1440  ffffffffbb66664b lapic_next_deadline+0x2b ([guest.kernel.kallsyms])             retq     IPC: 0.16 (4/25)
1263           :1440  1440  ffffffffbb74607f clockevents_program_event+0x8f ([guest.kernel.kallsyms])               test %eax, %eax
1264           :1440  1440  ffffffffbb746081 clockevents_program_event+0x91 ([guest.kernel.kallsyms])               jz 0xffffffffbb74603c    IPC: 0.06 (2/30)
1265           :1440  1440  ffffffffbb74603c clockevents_program_event+0x4c ([guest.kernel.kallsyms])               popq  %rbx
1266           :1440  1440  ffffffffbb74603d clockevents_program_event+0x4d ([guest.kernel.kallsyms])               popq  %r12
1267
1268Example using VM Time Correlation
1269
1270Start VM
1271
1272 $ sudo virsh start kubuntu20.04
1273 Domain kubuntu20.04 started
1274
1275Mount the guest file system.  Note sshfs needs -o direct_io to enable reading of proc files.  root access is needed to read /proc/kcore.
1276
1277 $ mkdir -p vm0
1278 $ sshfs -o direct_io root@vm0:/ vm0
1279
1280Copy the guest /proc/kallsyms, /proc/modules and /proc/kcore
1281
1282 $ perf buildid-cache -v --kcore vm0/proc/kcore
1283 same kcore found in /home/user/.debug/[kernel.kcore]/cc9c55a98c5e4ec0aeda69302554aabed5cd6491/2021021312450777
1284 $ KALLSYMS=/home/user/.debug/\[kernel.kcore\]/cc9c55a98c5e4ec0aeda69302554aabed5cd6491/2021021312450777/kallsyms
1285
1286Find the VM process
1287
1288 $ ps -eLl | grep 'KVM\|PID'
1289 F S   UID     PID    PPID     LWP  C PRI  NI ADDR SZ WCHAN  TTY          TIME CMD
1290 3 S 64055   16998       1   17005 13  80   0 - 1818189 -    ?        00:00:16 CPU 0/KVM
1291 3 S 64055   16998       1   17006  4  80   0 - 1818189 -    ?        00:00:05 CPU 1/KVM
1292 3 S 64055   16998       1   17007  3  80   0 - 1818189 -    ?        00:00:04 CPU 2/KVM
1293 3 S 64055   16998       1   17008  4  80   0 - 1818189 -    ?        00:00:05 CPU 3/KVM
1294
1295Start an open-ended perf record, tracing the VM process, do something on the VM, and then ctrl-C to stop.
1296IPC can be determined, hence cyc=1 can be added.
1297Only kernel decoding is supported, so 'k' must be specified.
1298Intel PT traces both the host and the guest so --guest and --host need to be specified.
1299
1300 $ sudo perf kvm --guest --host --guestkallsyms $KALLSYMS record --kcore -e intel_pt/cyc=1/k -p 16998
1301 ^C[ perf record: Woken up 1 times to write data ]
1302 [ perf record: Captured and wrote 9.041 MB perf.data.kvm ]
1303
1304Now 'perf inject' can be used to determine the VMX TCS Offset. Note, Intel PT TSC packets are
1305only 7-bytes, so the TSC Offset might differ from the actual value in the 8th byte. That will
1306have no effect i.e. the resulting timestamps will be correct anyway.
1307
1308 $ perf inject -i perf.data.kvm --vm-time-correlation=dry-run
1309 ERROR: Unknown TSC Offset for VMCS 0x1bff6a
1310 VMCS: 0x1bff6a  TSC Offset 0xffffe42722c64c41
1311 ERROR: Unknown TSC Offset for VMCS 0x1cbc08
1312 VMCS: 0x1cbc08  TSC Offset 0xffffe42722c64c41
1313 ERROR: Unknown TSC Offset for VMCS 0x1c3ce8
1314 VMCS: 0x1c3ce8  TSC Offset 0xffffe42722c64c41
1315 ERROR: Unknown TSC Offset for VMCS 0x1cbce9
1316 VMCS: 0x1cbce9  TSC Offset 0xffffe42722c64c41
1317
1318Each virtual CPU has a different Virtual Machine Control Structure (VMCS)
1319shown above with the calculated TSC Offset. For an unchanging TSC Offset
1320they should all be the same for the same virtual machine.
1321
1322Now that the TSC Offset is known, it can be provided to 'perf inject'
1323
1324 $ perf inject -i perf.data.kvm --vm-time-correlation="dry-run 0xffffe42722c64c41"
1325
1326Note the options for 'perf inject' --vm-time-correlation are:
1327
1328 [ dry-run ] [ <TSC Offset> [ : <VMCS> [ , <VMCS> ]... ]  ]...
1329
1330So it is possible to specify different TSC Offsets for different VMCS.
1331The option "dry-run" will cause the file to be processed but without updating it.
1332Note it is also possible to get a intel_pt.log file by adding option --itrace=d
1333
1334There were no errors so, do it for real
1335
1336 $ perf inject -i perf.data.kvm --vm-time-correlation=0xffffe42722c64c41 --force
1337
1338'perf script' can be used to see if there are any decoder errors
1339
1340 $ perf script -i perf.data.kvm --guestkallsyms $KALLSYMS --itrace=e-o
1341
1342There were none.
1343
1344'perf script' can be used to provide an instruction trace showing timestamps
1345
1346 $ perf script -i perf.data.kvm --guestkallsyms $KALLSYMS --insn-trace --xed -F+ipc | grep -C10 vmresume | head -21
1347       CPU 1/KVM 17006 [001] 11500.262865593:  ffffffff82133cdd __vmx_vcpu_run+0x3d ([kernel.kallsyms])                 movq  0x48(%rax), %r9
1348       CPU 1/KVM 17006 [001] 11500.262865593:  ffffffff82133ce1 __vmx_vcpu_run+0x41 ([kernel.kallsyms])                 movq  0x50(%rax), %r10
1349       CPU 1/KVM 17006 [001] 11500.262865593:  ffffffff82133ce5 __vmx_vcpu_run+0x45 ([kernel.kallsyms])                 movq  0x58(%rax), %r11
1350       CPU 1/KVM 17006 [001] 11500.262865593:  ffffffff82133ce9 __vmx_vcpu_run+0x49 ([kernel.kallsyms])                 movq  0x60(%rax), %r12
1351       CPU 1/KVM 17006 [001] 11500.262865593:  ffffffff82133ced __vmx_vcpu_run+0x4d ([kernel.kallsyms])                 movq  0x68(%rax), %r13
1352       CPU 1/KVM 17006 [001] 11500.262865593:  ffffffff82133cf1 __vmx_vcpu_run+0x51 ([kernel.kallsyms])                 movq  0x70(%rax), %r14
1353       CPU 1/KVM 17006 [001] 11500.262865593:  ffffffff82133cf5 __vmx_vcpu_run+0x55 ([kernel.kallsyms])                 movq  0x78(%rax), %r15
1354       CPU 1/KVM 17006 [001] 11500.262865593:  ffffffff82133cf9 __vmx_vcpu_run+0x59 ([kernel.kallsyms])                 movq  (%rax), %rax
1355       CPU 1/KVM 17006 [001] 11500.262865593:  ffffffff82133cfc __vmx_vcpu_run+0x5c ([kernel.kallsyms])                 callq  0xffffffff82133c40
1356       CPU 1/KVM 17006 [001] 11500.262865593:  ffffffff82133c40 vmx_vmenter+0x0 ([kernel.kallsyms])             jz 0xffffffff82133c46
1357       CPU 1/KVM 17006 [001] 11500.262866075:  ffffffff82133c42 vmx_vmenter+0x2 ([kernel.kallsyms])             vmresume         IPC: 0.05 (40/769)
1358          :17006 17006 [001] 11500.262869216:  ffffffff82200cb0 asm_sysvec_apic_timer_interrupt+0x0 ([guest.kernel.kallsyms])           clac
1359          :17006 17006 [001] 11500.262869216:  ffffffff82200cb3 asm_sysvec_apic_timer_interrupt+0x3 ([guest.kernel.kallsyms])           pushq  $0xffffffffffffffff
1360          :17006 17006 [001] 11500.262869216:  ffffffff82200cb5 asm_sysvec_apic_timer_interrupt+0x5 ([guest.kernel.kallsyms])           callq  0xffffffff82201160
1361          :17006 17006 [001] 11500.262869216:  ffffffff82201160 error_entry+0x0 ([guest.kernel.kallsyms])               cld
1362          :17006 17006 [001] 11500.262869216:  ffffffff82201161 error_entry+0x1 ([guest.kernel.kallsyms])               pushq  %rsi
1363          :17006 17006 [001] 11500.262869216:  ffffffff82201162 error_entry+0x2 ([guest.kernel.kallsyms])               movq  0x8(%rsp), %rsi
1364          :17006 17006 [001] 11500.262869216:  ffffffff82201167 error_entry+0x7 ([guest.kernel.kallsyms])               movq  %rdi, 0x8(%rsp)
1365          :17006 17006 [001] 11500.262869216:  ffffffff8220116c error_entry+0xc ([guest.kernel.kallsyms])               pushq  %rdx
1366          :17006 17006 [001] 11500.262869216:  ffffffff8220116d error_entry+0xd ([guest.kernel.kallsyms])               pushq  %rcx
1367          :17006 17006 [001] 11500.262869216:  ffffffff8220116e error_entry+0xe ([guest.kernel.kallsyms])               pushq  %rax
1368
1369
1370Event Trace
1371-----------
1372
1373Event Trace records information about asynchronous events, for example interrupts,
1374faults, VM exits and entries.  The information is recorded in CFE and EVD packets,
1375and also the Interrupt Flag is recorded on the MODE.Exec packet.  The CFE packet
1376contains a type field to identify one of the following:
1377
1378	 1	INTR		interrupt, fault, exception, NMI
1379	 2	IRET		interrupt return
1380	 3	SMI		system management interrupt
1381	 4	RSM		resume from system management mode
1382	 5	SIPI		startup interprocessor interrupt
1383	 6	INIT		INIT signal
1384	 7	VMENTRY		VM-Entry
1385	 8	VMEXIT		VM-Entry
1386	 9	VMEXIT_INTR	VM-Exit due to interrupt
1387	10	SHUTDOWN	Shutdown
1388
1389For more details, refer to the Intel 64 and IA-32 Architectures Software
1390Developer Manuals (version 076 or later).
1391
1392The capability to do Event Trace is indicated by the
1393/sys/bus/event_source/devices/intel_pt/caps/event_trace file.
1394
1395Event trace is selected for recording using the "event" config term. e.g.
1396
1397	perf record -e intel_pt/event/u uname
1398
1399Event trace events are output using the --itrace I option. e.g.
1400
1401	perf script --itrace=Ie
1402
1403perf script displays events containing CFE type, vector and event data,
1404in the form:
1405
1406	  evt:   hw int            (t)  cfe: INTR IP: 1 vector: 3 PFA: 0x8877665544332211
1407
1408The IP flag indicates if the event binds to an IP, which includes any case where
1409flow control packet generation is enabled, as well as when CFE packet IP bit is
1410set.
1411
1412perf script displays events containing changes to the Interrupt Flag in the form:
1413
1414	iflag:   t                      IFLAG: 1->0 via branch
1415
1416where "via branch" indicates a branch (interrupt or return from interrupt) and
1417"non branch" indicates an instruction such as CFI, STI or POPF).
1418
1419In addition, the current state of the interrupt flag is indicated by the presence
1420or absence of the "D" (interrupt disabled) perf script flag.  If the interrupt
1421flag is changed, then the "t" flag is also included i.e.
1422
1423		no flag, interrupts enabled IF=1
1424	t	interrupts become disabled IF=1 -> IF=0
1425	D	interrupts are disabled IF=0
1426	Dt	interrupts become enabled  IF=0 -> IF=1
1427
1428The intel-pt-events.py script illustrates how to access Event Trace information
1429using a Python script.
1430
1431
1432TNT Disable
1433-----------
1434
1435TNT packets are disabled using the "notnt" config term. e.g.
1436
1437	perf record -e intel_pt/notnt/u uname
1438
1439In that case the --itrace q option is forced because walking executable code
1440to reconstruct the control flow is not possible.
1441
1442
1443Emulated PTWRITE
1444----------------
1445
1446Later perf tools support a method to emulate the ptwrite instruction, which
1447can be useful if hardware does not support the ptwrite instruction.
1448
1449Instead of using the ptwrite instruction, a function is used which produces
1450a trace that encodes the payload data into TNT packets.  Here is an example
1451of the function:
1452
1453 #include <stdint.h>
1454
1455 void perf_emulate_ptwrite(uint64_t x)
1456 __attribute__((externally_visible, noipa, no_instrument_function, naked));
1457
1458 #define PERF_EMULATE_PTWRITE_8_BITS \
1459                 "1: shl %rax\n"     \
1460                 "   jc 1f\n"        \
1461                 "1: shl %rax\n"     \
1462                 "   jc 1f\n"        \
1463                 "1: shl %rax\n"     \
1464                 "   jc 1f\n"        \
1465                 "1: shl %rax\n"     \
1466                 "   jc 1f\n"        \
1467                 "1: shl %rax\n"     \
1468                 "   jc 1f\n"        \
1469                 "1: shl %rax\n"     \
1470                 "   jc 1f\n"        \
1471                 "1: shl %rax\n"     \
1472                 "   jc 1f\n"        \
1473                 "1: shl %rax\n"     \
1474                 "   jc 1f\n"
1475
1476 /* Undefined instruction */
1477 #define PERF_EMULATE_PTWRITE_UD2        ".byte 0x0f, 0x0b\n"
1478
1479 #define PERF_EMULATE_PTWRITE_MAGIC        PERF_EMULATE_PTWRITE_UD2 ".ascii \"perf,ptwrite  \"\n"
1480
1481 void perf_emulate_ptwrite(uint64_t x __attribute__ ((__unused__)))
1482 {
1483          /* Assumes SysV ABI : x passed in rdi */
1484         __asm__ volatile (
1485                 "jmp 1f\n"
1486                 PERF_EMULATE_PTWRITE_MAGIC
1487                 "1: mov %rdi, %rax\n"
1488                 PERF_EMULATE_PTWRITE_8_BITS
1489                 PERF_EMULATE_PTWRITE_8_BITS
1490                 PERF_EMULATE_PTWRITE_8_BITS
1491                 PERF_EMULATE_PTWRITE_8_BITS
1492                 PERF_EMULATE_PTWRITE_8_BITS
1493                 PERF_EMULATE_PTWRITE_8_BITS
1494                 PERF_EMULATE_PTWRITE_8_BITS
1495                 PERF_EMULATE_PTWRITE_8_BITS
1496                 "1: ret\n"
1497         );
1498 }
1499
1500For example, a test program with the function above:
1501
1502 #include <stdio.h>
1503 #include <stdint.h>
1504 #include <stdlib.h>
1505
1506 #include "perf_emulate_ptwrite.h"
1507
1508 int main(int argc, char *argv[])
1509 {
1510         uint64_t x = 0;
1511
1512         if (argc > 1)
1513                 x = strtoull(argv[1], NULL, 0);
1514         perf_emulate_ptwrite(x);
1515         return 0;
1516 }
1517
1518Can be compiled and traced:
1519
1520 $ gcc -Wall -Wextra -O3 -g -o eg_ptw eg_ptw.c
1521 $ perf record -e intel_pt//u ./eg_ptw 0x1234567890abcdef
1522 [ perf record: Woken up 1 times to write data ]
1523 [ perf record: Captured and wrote 0.017 MB perf.data ]
1524 $ perf script --itrace=ew
1525           eg_ptw 19875 [007]  8061.235912:     ptwrite:  IP: 0 payload: 0x1234567890abcdef      55701249a196 perf_emulate_ptwrite+0x16 (/home/user/eg_ptw)
1526 $
1527
1528
1529Pipe mode
1530---------
1531Pipe mode is a problem for Intel PT and possibly other auxtrace users.
1532It's not recommended to use a pipe as data output with Intel PT because
1533of the following reason.
1534
1535Essentially the auxtrace buffers do not behave like the regular perf
1536event buffers.  That is because the head and tail are updated by
1537software, but in the auxtrace case the data is written by hardware.
1538So the head and tail do not get updated as data is written.
1539
1540In the Intel PT case, the head and tail are updated only when the trace
1541is disabled by software, for example:
1542    - full-trace, system wide : when buffer passes watermark
1543    - full-trace, not system-wide : when buffer passes watermark or
1544                                    context switches
1545    - snapshot mode : as above but also when a snapshot is made
1546    - sample mode : as above but also when a sample is made
1547
1548That means finished-round ordering doesn't work.  An auxtrace buffer
1549can turn up that has data that extends back in time, possibly to the
1550very beginning of tracing.
1551
1552For a perf.data file, that problem is solved by going through the trace
1553and queuing up the auxtrace buffers in advance.
1554
1555For pipe mode, the order of events and timestamps can presumably
1556be messed up.
1557
1558
1559EXAMPLE
1560-------
1561
1562Examples can be found on perf wiki page "Perf tools support for Intel® Processor Trace":
1563
1564https://perf.wiki.kernel.org/index.php/Perf_tools_support_for_Intel%C2%AE_Processor_Trace
1565
1566
1567SEE ALSO
1568--------
1569
1570linkperf:perf-record[1], linkperf:perf-script[1], linkperf:perf-report[1],
1571linkperf:perf-inject[1]
1572