Loading...
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 "bcrosyiABExgh" which stand for branch, call, return, conditional,
112system, asynchronous, interrupt, transaction abort, trace begin, trace end,
113in transaction, VM-entry, and VM-exit respectively.
114
115perf script also supports higher level ways to dump instruction traces:
116
117 perf script --insn-trace --xed
118
119Dump all instructions. This requires installing the xed tool (see XED below)
120Dumping all instructions in a long trace can be fairly slow. It is usually better
121to start with higher level decoding, like
122
123 perf script --call-trace
124
125or
126
127 perf script --call-ret-trace
128
129and then select a time range of interest. The time range can then be examined
130in detail with
131
132 perf script --time starttime,stoptime --insn-trace --xed
133
134While examining the trace it's also useful to filter on specific CPUs using
135the -C option
136
137 perf script --time starttime,stoptime --insn-trace --xed -C 1
138
139Dump all instructions in time range on CPU 1.
140
141Another interesting field that is not printed by default is 'ipc' which can be
142displayed as follows:
143
144 perf script --itrace=be -F+ipc
145
146There are two ways that instructions-per-cycle (IPC) can be calculated depending
147on the recording.
148
149If the 'cyc' config term (see config terms section below) was used, then IPC is
150calculated using the cycle count from CYC packets, otherwise MTC packets are
151used - refer to the 'mtc' config term. When MTC is used, however, the values
152are less accurate because the timing is less accurate.
153
154Because Intel PT does not update the cycle count on every branch or instruction,
155the values will often be zero. When there are values, they will be the number
156of instructions and number of cycles since the last update, and thus represent
157the average IPC since the last IPC for that event type. Note IPC for "branches"
158events is calculated separately from IPC for "instructions" events.
159
160Also note that the IPC instruction count may or may not include the current
161instruction. If the cycle count is associated with an asynchronous branch
162(e.g. page fault or interrupt), then the instruction count does not include the
163current instruction, otherwise it does. That is consistent with whether or not
164that instruction has retired when the cycle count is updated.
165
166Another note, in the case of "branches" events, non-taken branches are not
167presently sampled, so IPC values for them do not appear e.g. a CYC packet with a
168TNT packet that starts with a non-taken branch. To see every possible IPC
169value, "instructions" events can be used e.g. --itrace=i0ns
170
171While it is possible to create scripts to analyze the data, an alternative
172approach is available to export the data to a sqlite or postgresql database.
173Refer to script export-to-sqlite.py or export-to-postgresql.py for more details,
174and to script exported-sql-viewer.py for an example of using the database.
175
176There is also script intel-pt-events.py which provides an example of how to
177unpack the raw data for power events and PTWRITE. The script also displays
178branches, and supports 2 additional modes selected by option:
179
180 --insn-trace - instruction trace
181 --src-trace - source trace
182
183As mentioned above, it is easy to capture too much data. One way to limit the
184data captured is to use 'snapshot' mode which is explained further below.
185Refer to 'new snapshot option' and 'Intel PT modes of operation' further below.
186
187Another problem that will be experienced is decoder errors. They can be caused
188by inability to access the executed image, self-modified or JIT-ed code, or the
189inability to match side-band information (such as context switches and mmaps)
190which results in the decoder not knowing what code was executed.
191
192There is also the problem of perf not being able to copy the data fast enough,
193resulting in data lost because the buffer was full. See 'Buffer handling' below
194for more details.
195
196
197perf record
198-----------
199
200new event
201~~~~~~~~~
202
203The Intel PT kernel driver creates a new PMU for Intel PT. PMU events are
204selected by providing the PMU name followed by the "config" separated by slashes.
205An enhancement has been made to allow default "config" e.g. the option
206
207 -e intel_pt//
208
209will use a default config value. Currently that is the same as
210
211 -e intel_pt/tsc,noretcomp=0/
212
213which is the same as
214
215 -e intel_pt/tsc=1,noretcomp=0/
216
217Note there are now new config terms - see section 'config terms' further below.
218
219The config terms are listed in /sys/devices/intel_pt/format. They are bit
220fields within the config member of the struct perf_event_attr which is
221passed to the kernel by the perf_event_open system call. They correspond to bit
222fields in the IA32_RTIT_CTL MSR. Here is a list of them and their definitions:
223
224 $ grep -H . /sys/bus/event_source/devices/intel_pt/format/*
225 /sys/bus/event_source/devices/intel_pt/format/cyc:config:1
226 /sys/bus/event_source/devices/intel_pt/format/cyc_thresh:config:19-22
227 /sys/bus/event_source/devices/intel_pt/format/mtc:config:9
228 /sys/bus/event_source/devices/intel_pt/format/mtc_period:config:14-17
229 /sys/bus/event_source/devices/intel_pt/format/noretcomp:config:11
230 /sys/bus/event_source/devices/intel_pt/format/psb_period:config:24-27
231 /sys/bus/event_source/devices/intel_pt/format/tsc:config:10
232
233Note that the default config must be overridden for each term i.e.
234
235 -e intel_pt/noretcomp=0/
236
237is the same as:
238
239 -e intel_pt/tsc=1,noretcomp=0/
240
241So, to disable TSC packets use:
242
243 -e intel_pt/tsc=0/
244
245It is also possible to specify the config value explicitly:
246
247 -e intel_pt/config=0x400/
248
249Note that, as with all events, the event is suffixed with event modifiers:
250
251 u userspace
252 k kernel
253 h hypervisor
254 G guest
255 H host
256 p precise ip
257
258'h', 'G' and 'H' are for virtualization which is not supported by Intel PT.
259'p' is also not relevant to Intel PT. So only options 'u' and 'k' are
260meaningful for Intel PT.
261
262perf_event_attr is displayed if the -vv option is used e.g.
263
264 ------------------------------------------------------------
265 perf_event_attr:
266 type 6
267 size 112
268 config 0x400
269 { sample_period, sample_freq } 1
270 sample_type IP|TID|TIME|CPU|IDENTIFIER
271 read_format ID
272 disabled 1
273 inherit 1
274 exclude_kernel 1
275 exclude_hv 1
276 enable_on_exec 1
277 sample_id_all 1
278 ------------------------------------------------------------
279 sys_perf_event_open: pid 31104 cpu 0 group_fd -1 flags 0x8
280 sys_perf_event_open: pid 31104 cpu 1 group_fd -1 flags 0x8
281 sys_perf_event_open: pid 31104 cpu 2 group_fd -1 flags 0x8
282 sys_perf_event_open: pid 31104 cpu 3 group_fd -1 flags 0x8
283 ------------------------------------------------------------
284
285
286config terms
287~~~~~~~~~~~~
288
289The June 2015 version of Intel 64 and IA-32 Architectures Software Developer
290Manuals, Chapter 36 Intel Processor Trace, defined new Intel PT features.
291Some of the features are reflect in new config terms. All the config terms are
292described below.
293
294tsc Always supported. Produces TSC timestamp packets to provide
295 timing information. In some cases it is possible to decode
296 without timing information, for example a per-thread context
297 that does not overlap executable memory maps.
298
299 The default config selects tsc (i.e. tsc=1).
300
301noretcomp Always supported. Disables "return compression" so a TIP packet
302 is produced when a function returns. Causes more packets to be
303 produced but might make decoding more reliable.
304
305 The default config does not select noretcomp (i.e. noretcomp=0).
306
307psb_period Allows the frequency of PSB packets to be specified.
308
309 The PSB packet is a synchronization packet that provides a
310 starting point for decoding or recovery from errors.
311
312 Support for psb_period is indicated by:
313
314 /sys/bus/event_source/devices/intel_pt/caps/psb_cyc
315
316 which contains "1" if the feature is supported and "0"
317 otherwise.
318
319 Valid values are given by:
320
321 /sys/bus/event_source/devices/intel_pt/caps/psb_periods
322
323 which contains a hexadecimal value, the bits of which represent
324 valid values e.g. bit 2 set means value 2 is valid.
325
326 The psb_period value is converted to the approximate number of
327 trace bytes between PSB packets as:
328
329 2 ^ (value + 11)
330
331 e.g. value 3 means 16KiB bytes between PSBs
332
333 If an invalid value is entered, the error message
334 will give a list of valid values e.g.
335
336 $ perf record -e intel_pt/psb_period=15/u uname
337 Invalid psb_period for intel_pt. Valid values are: 0-5
338
339 If MTC packets are selected, the default config selects a value
340 of 3 (i.e. psb_period=3) or the nearest lower value that is
341 supported (0 is always supported). Otherwise the default is 0.
342
343 If decoding is expected to be reliable and the buffer is large
344 then a large PSB period can be used.
345
346 Because a TSC packet is produced with PSB, the PSB period can
347 also affect the granularity to timing information in the absence
348 of MTC or CYC.
349
350mtc Produces MTC timing packets.
351
352 MTC packets provide finer grain timestamp information than TSC
353 packets. MTC packets record time using the hardware crystal
354 clock (CTC) which is related to TSC packets using a TMA packet.
355
356 Support for this feature is indicated by:
357
358 /sys/bus/event_source/devices/intel_pt/caps/mtc
359
360 which contains "1" if the feature is supported and
361 "0" otherwise.
362
363 The frequency of MTC packets can also be specified - see
364 mtc_period below.
365
366mtc_period Specifies how frequently MTC packets are produced - see mtc
367 above for how to determine if MTC packets are supported.
368
369 Valid values are given by:
370
371 /sys/bus/event_source/devices/intel_pt/caps/mtc_periods
372
373 which contains a hexadecimal value, the bits of which represent
374 valid values e.g. bit 2 set means value 2 is valid.
375
376 The mtc_period value is converted to the MTC frequency as:
377
378 CTC-frequency / (2 ^ value)
379
380 e.g. value 3 means one eighth of CTC-frequency
381
382 Where CTC is the hardware crystal clock, the frequency of which
383 can be related to TSC via values provided in cpuid leaf 0x15.
384
385 If an invalid value is entered, the error message
386 will give a list of valid values e.g.
387
388 $ perf record -e intel_pt/mtc_period=15/u uname
389 Invalid mtc_period for intel_pt. Valid values are: 0,3,6,9
390
391 The default value is 3 or the nearest lower value
392 that is supported (0 is always supported).
393
394cyc Produces CYC timing packets.
395
396 CYC packets provide even finer grain timestamp information than
397 MTC and TSC packets. A CYC packet contains the number of CPU
398 cycles since the last CYC packet. Unlike MTC and TSC packets,
399 CYC packets are only sent when another packet is also sent.
400
401 Support for this feature is indicated by:
402
403 /sys/bus/event_source/devices/intel_pt/caps/psb_cyc
404
405 which contains "1" if the feature is supported and
406 "0" otherwise.
407
408 The number of CYC packets produced can be reduced by specifying
409 a threshold - see cyc_thresh below.
410
411cyc_thresh Specifies how frequently CYC packets are produced - see cyc
412 above for how to determine if CYC packets are supported.
413
414 Valid cyc_thresh values are given by:
415
416 /sys/bus/event_source/devices/intel_pt/caps/cycle_thresholds
417
418 which contains a hexadecimal value, the bits of which represent
419 valid values e.g. bit 2 set means value 2 is valid.
420
421 The cyc_thresh value represents the minimum number of CPU cycles
422 that must have passed before a CYC packet can be sent. The
423 number of CPU cycles is:
424
425 2 ^ (value - 1)
426
427 e.g. value 4 means 8 CPU cycles must pass before a CYC packet
428 can be sent. Note a CYC packet is still only sent when another
429 packet is sent, not at, e.g. every 8 CPU cycles.
430
431 If an invalid value is entered, the error message
432 will give a list of valid values e.g.
433
434 $ perf record -e intel_pt/cyc,cyc_thresh=15/u uname
435 Invalid cyc_thresh for intel_pt. Valid values are: 0-12
436
437 CYC packets are not requested by default.
438
439pt Specifies pass-through which enables the 'branch' config term.
440
441 The default config selects 'pt' if it is available, so a user will
442 never need to specify this term.
443
444branch Enable branch tracing. Branch tracing is enabled by default so to
445 disable branch tracing use 'branch=0'.
446
447 The default config selects 'branch' if it is available.
448
449ptw Enable PTWRITE packets which are produced when a ptwrite instruction
450 is executed.
451
452 Support for this feature is indicated by:
453
454 /sys/bus/event_source/devices/intel_pt/caps/ptwrite
455
456 which contains "1" if the feature is supported and
457 "0" otherwise.
458
459fup_on_ptw Enable a FUP packet to follow the PTWRITE packet. The FUP packet
460 provides the address of the ptwrite instruction. In the absence of
461 fup_on_ptw, the decoder will use the address of the previous branch
462 if branch tracing is enabled, otherwise the address will be zero.
463 Note that fup_on_ptw will work even when branch tracing is disabled.
464
465pwr_evt Enable power events. The power events provide information about
466 changes to the CPU C-state.
467
468 Support for this feature is indicated by:
469
470 /sys/bus/event_source/devices/intel_pt/caps/power_event_trace
471
472 which contains "1" if the feature is supported and
473 "0" otherwise.
474
475
476AUX area sampling option
477~~~~~~~~~~~~~~~~~~~~~~~~
478
479To select Intel PT "sampling" the AUX area sampling option can be used:
480
481 --aux-sample
482
483Optionally it can be followed by the sample size in bytes e.g.
484
485 --aux-sample=8192
486
487In addition, the Intel PT event to sample must be defined e.g.
488
489 -e intel_pt//u
490
491Samples on other events will be created containing Intel PT data e.g. the
492following will create Intel PT samples on the branch-misses event, note the
493events must be grouped using {}:
494
495 perf record --aux-sample -e '{intel_pt//u,branch-misses:u}'
496
497An alternative to '--aux-sample' is to add the config term 'aux-sample-size' to
498events. In this case, the grouping is implied e.g.
499
500 perf record -e intel_pt//u -e branch-misses/aux-sample-size=8192/u
501
502is the same as:
503
504 perf record -e '{intel_pt//u,branch-misses/aux-sample-size=8192/u}'
505
506but allows for also using an address filter e.g.:
507
508 perf record -e intel_pt//u --filter 'filter * @/bin/ls' -e branch-misses/aux-sample-size=8192/u -- ls
509
510It is important to select a sample size that is big enough to contain at least
511one PSB packet. If not a warning will be displayed:
512
513 Intel PT sample size (%zu) may be too small for PSB period (%zu)
514
515The calculation used for that is: if sample_size <= psb_period + 256 display the
516warning. When sampling is used, psb_period defaults to 0 (2KiB).
517
518The default sample size is 4KiB.
519
520The sample size is passed in aux_sample_size in struct perf_event_attr. The
521sample size is limited by the maximum event size which is 64KiB. It is
522difficult to know how big the event might be without the trace sample attached,
523but the tool validates that the sample size is not greater than 60KiB.
524
525
526new snapshot option
527~~~~~~~~~~~~~~~~~~~
528
529The difference between full trace and snapshot from the kernel's perspective is
530that in full trace we don't overwrite trace data that the user hasn't collected
531yet (and indicated that by advancing aux_tail), whereas in snapshot mode we let
532the trace run and overwrite older data in the buffer so that whenever something
533interesting happens, we can stop it and grab a snapshot of what was going on
534around that interesting moment.
535
536To select snapshot mode a new option has been added:
537
538 -S
539
540Optionally it can be followed by the snapshot size e.g.
541
542 -S0x100000
543
544The default snapshot size is the auxtrace mmap size. If neither auxtrace mmap size
545nor snapshot size is specified, then the default is 4MiB for privileged users
546(or if /proc/sys/kernel/perf_event_paranoid < 0), 128KiB for unprivileged users.
547If an unprivileged user does not specify mmap pages, the mmap pages will be
548reduced as described in the 'new auxtrace mmap size option' section below.
549
550The snapshot size is displayed if the option -vv is used e.g.
551
552 Intel PT snapshot size: %zu
553
554
555new auxtrace mmap size option
556~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
557
558Intel PT buffer size is specified by an addition to the -m option e.g.
559
560 -m,16
561
562selects a buffer size of 16 pages i.e. 64KiB.
563
564Note that the existing functionality of -m is unchanged. The auxtrace mmap size
565is specified by the optional addition of a comma and the value.
566
567The default auxtrace mmap size for Intel PT is 4MiB/page_size for privileged users
568(or if /proc/sys/kernel/perf_event_paranoid < 0), 128KiB for unprivileged users.
569If an unprivileged user does not specify mmap pages, the mmap pages will be
570reduced from the default 512KiB/page_size to 256KiB/page_size, otherwise the
571user is likely to get an error as they exceed their mlock limit (Max locked
572memory as shown in /proc/self/limits). Note that perf does not count the first
573512KiB (actually /proc/sys/kernel/perf_event_mlock_kb minus 1 page) per cpu
574against the mlock limit so an unprivileged user is allowed 512KiB per cpu plus
575their mlock limit (which defaults to 64KiB but is not multiplied by the number
576of cpus).
577
578In full-trace mode, powers of two are allowed for buffer size, with a minimum
579size of 2 pages. In snapshot mode or sampling mode, it is the same but the
580minimum size is 1 page.
581
582The mmap size and auxtrace mmap size are displayed if the -vv option is used e.g.
583
584 mmap length 528384
585 auxtrace mmap length 4198400
586
587
588Intel PT modes of operation
589~~~~~~~~~~~~~~~~~~~~~~~~~~~
590
591Intel PT can be used in 3 modes:
592 full-trace mode
593 sample mode
594 snapshot mode
595
596Full-trace mode traces continuously e.g.
597
598 perf record -e intel_pt//u uname
599
600Sample mode attaches a Intel PT sample to other events e.g.
601
602 perf record --aux-sample -e intel_pt//u -e branch-misses:u
603
604Snapshot mode captures the available data when a signal is sent or "snapshot"
605control command is issued. e.g. using a signal
606
607 perf record -v -e intel_pt//u -S ./loopy 1000000000 &
608 [1] 11435
609 kill -USR2 11435
610 Recording AUX area tracing snapshot
611
612Note that the signal sent is SIGUSR2.
613Note that "Recording AUX area tracing snapshot" is displayed because the -v
614option is used.
615
616The advantage of using "snapshot" control command is that the access is
617controlled by access to a FIFO e.g.
618
619 $ mkfifo perf.control
620 $ mkfifo perf.ack
621 $ cat perf.ack &
622 [1] 15235
623 $ sudo ~/bin/perf record --control fifo:perf.control,perf.ack -S -e intel_pt//u -- sleep 60 &
624 [2] 15243
625 $ ps -e | grep perf
626 15244 pts/1 00:00:00 perf
627 $ kill -USR2 15244
628 bash: kill: (15244) - Operation not permitted
629 $ echo snapshot > perf.control
630 ack
631
632The 3 Intel PT modes of operation cannot be used together.
633
634
635Buffer handling
636~~~~~~~~~~~~~~~
637
638There may be buffer limitations (i.e. single ToPa entry) which means that actual
639buffer sizes are limited to powers of 2 up to 4MiB (MAX_ORDER). In order to
640provide other sizes, and in particular an arbitrarily large size, multiple
641buffers are logically concatenated. However an interrupt must be used to switch
642between buffers. That has two potential problems:
643 a) the interrupt may not be handled in time so that the current buffer
644 becomes full and some trace data is lost.
645 b) the interrupts may slow the system and affect the performance
646 results.
647
648If trace data is lost, the driver sets 'truncated' in the PERF_RECORD_AUX event
649which the tools report as an error.
650
651In full-trace mode, the driver waits for data to be copied out before allowing
652the (logical) buffer to wrap-around. If data is not copied out quickly enough,
653again 'truncated' is set in the PERF_RECORD_AUX event. If the driver has to
654wait, the intel_pt event gets disabled. Because it is difficult to know when
655that happens, perf tools always re-enable the intel_pt event after copying out
656data.
657
658
659Intel PT and build ids
660~~~~~~~~~~~~~~~~~~~~~~
661
662By default "perf record" post-processes the event stream to find all build ids
663for executables for all addresses sampled. Deliberately, Intel PT is not
664decoded for that purpose (it would take too long). Instead the build ids for
665all executables encountered (due to mmap, comm or task events) are included
666in the perf.data file.
667
668To see buildids included in the perf.data file use the command:
669
670 perf buildid-list
671
672If the perf.data file contains Intel PT data, that is the same as:
673
674 perf buildid-list --with-hits
675
676
677Snapshot mode and event disabling
678~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
679
680In order to make a snapshot, the intel_pt event is disabled using an IOCTL,
681namely PERF_EVENT_IOC_DISABLE. However doing that can also disable the
682collection of side-band information. In order to prevent that, a dummy
683software event has been introduced that permits tracking events (like mmaps) to
684continue to be recorded while intel_pt is disabled. That is important to ensure
685there is complete side-band information to allow the decoding of subsequent
686snapshots.
687
688A test has been created for that. To find the test:
689
690 perf test list
691 ...
692 23: Test using a dummy software event to keep tracking
693
694To run the test:
695
696 perf test 23
697 23: Test using a dummy software event to keep tracking : Ok
698
699
700perf record modes (nothing new here)
701~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
702
703perf record essentially operates in one of three modes:
704 per thread
705 per cpu
706 workload only
707
708"per thread" mode is selected by -t or by --per-thread (with -p or -u or just a
709workload).
710"per cpu" is selected by -C or -a.
711"workload only" mode is selected by not using the other options but providing a
712command to run (i.e. the workload).
713
714In per-thread mode an exact list of threads is traced. There is no inheritance.
715Each thread has its own event buffer.
716
717In per-cpu mode all processes (or processes from the selected cgroup i.e. -G
718option, or processes selected with -p or -u) are traced. Each cpu has its own
719buffer. Inheritance is allowed.
720
721In workload-only mode, the workload is traced but with per-cpu buffers.
722Inheritance is allowed. Note that you can now trace a workload in per-thread
723mode by using the --per-thread option.
724
725
726Privileged vs non-privileged users
727~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
728
729Unless /proc/sys/kernel/perf_event_paranoid is set to -1, unprivileged users
730have memory limits imposed upon them. That affects what buffer sizes they can
731have as outlined above.
732
733The v4.2 kernel introduced support for a context switch metadata event,
734PERF_RECORD_SWITCH, which allows unprivileged users to see when their processes
735are scheduled out and in, just not by whom, which is left for the
736PERF_RECORD_SWITCH_CPU_WIDE, that is only accessible in system wide context,
737which in turn requires CAP_PERFMON or CAP_SYS_ADMIN.
738
739Please see the 45ac1403f564 ("perf: Add PERF_RECORD_SWITCH to indicate context
740switches") commit, that introduces these metadata events for further info.
741
742When working with kernels < v4.2, the following considerations must be taken,
743as the sched:sched_switch tracepoints will be used to receive such information:
744
745Unless /proc/sys/kernel/perf_event_paranoid is set to -1, unprivileged users are
746not permitted to use tracepoints which means there is insufficient side-band
747information to decode Intel PT in per-cpu mode, and potentially workload-only
748mode too if the workload creates new processes.
749
750Note also, that to use tracepoints, read-access to debugfs is required. So if
751debugfs is not mounted or the user does not have read-access, it will again not
752be possible to decode Intel PT in per-cpu mode.
753
754
755sched_switch tracepoint
756~~~~~~~~~~~~~~~~~~~~~~~
757
758The sched_switch tracepoint is used to provide side-band data for Intel PT
759decoding in kernels where the PERF_RECORD_SWITCH metadata event isn't
760available.
761
762The sched_switch events are automatically added. e.g. the second event shown
763below:
764
765 $ perf record -vv -e intel_pt//u uname
766 ------------------------------------------------------------
767 perf_event_attr:
768 type 6
769 size 112
770 config 0x400
771 { sample_period, sample_freq } 1
772 sample_type IP|TID|TIME|CPU|IDENTIFIER
773 read_format ID
774 disabled 1
775 inherit 1
776 exclude_kernel 1
777 exclude_hv 1
778 enable_on_exec 1
779 sample_id_all 1
780 ------------------------------------------------------------
781 sys_perf_event_open: pid 31104 cpu 0 group_fd -1 flags 0x8
782 sys_perf_event_open: pid 31104 cpu 1 group_fd -1 flags 0x8
783 sys_perf_event_open: pid 31104 cpu 2 group_fd -1 flags 0x8
784 sys_perf_event_open: pid 31104 cpu 3 group_fd -1 flags 0x8
785 ------------------------------------------------------------
786 perf_event_attr:
787 type 2
788 size 112
789 config 0x108
790 { sample_period, sample_freq } 1
791 sample_type IP|TID|TIME|CPU|PERIOD|RAW|IDENTIFIER
792 read_format ID
793 inherit 1
794 sample_id_all 1
795 exclude_guest 1
796 ------------------------------------------------------------
797 sys_perf_event_open: pid -1 cpu 0 group_fd -1 flags 0x8
798 sys_perf_event_open: pid -1 cpu 1 group_fd -1 flags 0x8
799 sys_perf_event_open: pid -1 cpu 2 group_fd -1 flags 0x8
800 sys_perf_event_open: pid -1 cpu 3 group_fd -1 flags 0x8
801 ------------------------------------------------------------
802 perf_event_attr:
803 type 1
804 size 112
805 config 0x9
806 { sample_period, sample_freq } 1
807 sample_type IP|TID|TIME|IDENTIFIER
808 read_format ID
809 disabled 1
810 inherit 1
811 exclude_kernel 1
812 exclude_hv 1
813 mmap 1
814 comm 1
815 enable_on_exec 1
816 task 1
817 sample_id_all 1
818 mmap2 1
819 comm_exec 1
820 ------------------------------------------------------------
821 sys_perf_event_open: pid 31104 cpu 0 group_fd -1 flags 0x8
822 sys_perf_event_open: pid 31104 cpu 1 group_fd -1 flags 0x8
823 sys_perf_event_open: pid 31104 cpu 2 group_fd -1 flags 0x8
824 sys_perf_event_open: pid 31104 cpu 3 group_fd -1 flags 0x8
825 mmap size 528384B
826 AUX area mmap length 4194304
827 perf event ring buffer mmapped per cpu
828 Synthesizing auxtrace information
829 Linux
830 [ perf record: Woken up 1 times to write data ]
831 [ perf record: Captured and wrote 0.042 MB perf.data ]
832
833Note, the sched_switch event is only added if the user is permitted to use it
834and only in per-cpu mode.
835
836Note also, the sched_switch event is only added if TSC packets are requested.
837That is because, in the absence of timing information, the sched_switch events
838cannot be matched against the Intel PT trace.
839
840
841perf script
842-----------
843
844By default, perf script will decode trace data found in the perf.data file.
845This can be further controlled by new option --itrace.
846
847
848New --itrace option
849~~~~~~~~~~~~~~~~~~~
850
851Having no option is the same as
852
853 --itrace
854
855which, in turn, is the same as
856
857 --itrace=cepwx
858
859The letters are:
860
861 i synthesize "instructions" events
862 b synthesize "branches" events
863 x synthesize "transactions" events
864 w synthesize "ptwrite" events
865 p synthesize "power" events (incl. PSB events)
866 c synthesize branches events (calls only)
867 r synthesize branches events (returns only)
868 e synthesize tracing error events
869 d create a debug log
870 g synthesize a call chain (use with i or x)
871 G synthesize a call chain on existing event records
872 l synthesize last branch entries (use with i or x)
873 L synthesize last branch entries on existing event records
874 s skip initial number of events
875 q quicker (less detailed) decoding
876 Z prefer to ignore timestamps (so-called "timeless" decoding)
877
878"Instructions" events look like they were recorded by "perf record -e
879instructions".
880
881"Branches" events look like they were recorded by "perf record -e branches". "c"
882and "r" can be combined to get calls and returns.
883
884"Transactions" events correspond to the start or end of transactions. The
885'flags' field can be used in perf script to determine whether the event is a
886tranasaction start, commit or abort.
887
888Note that "instructions", "branches" and "transactions" events depend on code
889flow packets which can be disabled by using the config term "branch=0". Refer
890to the config terms section above.
891
892"ptwrite" events record the payload of the ptwrite instruction and whether
893"fup_on_ptw" was used. "ptwrite" events depend on PTWRITE packets which are
894recorded only if the "ptw" config term was used. Refer to the config terms
895section above. perf script "synth" field displays "ptwrite" information like
896this: "ip: 0 payload: 0x123456789abcdef0" where "ip" is 1 if "fup_on_ptw" was
897used.
898
899"Power" events correspond to power event packets and CBR (core-to-bus ratio)
900packets. While CBR packets are always recorded when tracing is enabled, power
901event packets are recorded only if the "pwr_evt" config term was used. Refer to
902the config terms section above. The power events record information about
903C-state changes, whereas CBR is indicative of CPU frequency. perf script
904"event,synth" fields display information like this:
905 cbr: cbr: 22 freq: 2189 MHz (200%)
906 mwait: hints: 0x60 extensions: 0x1
907 pwre: hw: 0 cstate: 2 sub-cstate: 0
908 exstop: ip: 1
909 pwrx: deepest cstate: 2 last cstate: 2 wake reason: 0x4
910Where:
911 "cbr" includes the frequency and the percentage of maximum non-turbo
912 "mwait" shows mwait hints and extensions
913 "pwre" shows C-state transitions (to a C-state deeper than C0) and
914 whether initiated by hardware
915 "exstop" indicates execution stopped and whether the IP was recorded
916 exactly,
917 "pwrx" indicates return to C0
918For more details refer to the Intel 64 and IA-32 Architectures Software
919Developer Manuals.
920
921PSB events show when a PSB+ occurred and also the byte-offset in the trace.
922Emitting a PSB+ can cause a CPU a slight delay. When doing timing analysis
923of code with Intel PT, it is useful to know if a timing bubble was caused
924by Intel PT or not.
925
926Error events show where the decoder lost the trace. Error events
927are quite important. Users must know if what they are seeing is a complete
928picture or not. The "e" option may be followed by flags which affect what errors
929will or will not be reported. Each flag must be preceded by either '+' or '-'.
930The flags supported by Intel PT are:
931 -o Suppress overflow errors
932 -l Suppress trace data lost errors
933For example, for errors but not overflow or data lost errors:
934
935 --itrace=e-o-l
936
937The "d" option will cause the creation of a file "intel_pt.log" containing all
938decoded packets and instructions. Note that this option slows down the decoder
939and that the resulting file may be very large. The "d" option may be followed
940by flags which affect what debug messages will or will not be logged. Each flag
941must be preceded by either '+' or '-'. The flags support by Intel PT are:
942 -a Suppress logging of perf events
943 +a Log all perf events
944By default, logged perf events are filtered by any specified time ranges, but
945flag +a overrides that.
946
947In addition, the period of the "instructions" event can be specified. e.g.
948
949 --itrace=i10us
950
951sets the period to 10us i.e. one instruction sample is synthesized for each 10
952microseconds of trace. Alternatives to "us" are "ms" (milliseconds),
953"ns" (nanoseconds), "t" (TSC ticks) or "i" (instructions).
954
955"ms", "us" and "ns" are converted to TSC ticks.
956
957The timing information included with Intel PT does not give the time of every
958instruction. Consequently, for the purpose of sampling, the decoder estimates
959the time since the last timing packet based on 1 tick per instruction. The time
960on the sample is *not* adjusted and reflects the last known value of TSC.
961
962For Intel PT, the default period is 100us.
963
964Setting it to a zero period means "as often as possible".
965
966In the case of Intel PT that is the same as a period of 1 and a unit of
967'instructions' (i.e. --itrace=i1i).
968
969Also the call chain size (default 16, max. 1024) for instructions or
970transactions events can be specified. e.g.
971
972 --itrace=ig32
973 --itrace=xg32
974
975Also the number of last branch entries (default 64, max. 1024) for instructions or
976transactions events can be specified. e.g.
977
978 --itrace=il10
979 --itrace=xl10
980
981Note that last branch entries are cleared for each sample, so there is no overlap
982from one sample to the next.
983
984The G and L options are designed in particular for sample mode, and work much
985like g and l but add call chain and branch stack to the other selected events
986instead of synthesized events. For example, to record branch-misses events for
987'ls' and then add a call chain derived from the Intel PT trace:
988
989 perf record --aux-sample -e '{intel_pt//u,branch-misses:u}' -- ls
990 perf report --itrace=Ge
991
992Although in fact G is a default for perf report, so that is the same as just:
993
994 perf report
995
996One caveat with the G and L options is that they work poorly with "Large PEBS".
997Large PEBS means PEBS records will be accumulated by hardware and the written
998into the event buffer in one go. That reduces interrupts, but can give very
999late timestamps. Because the Intel PT trace is synchronized by timestamps,
1000the PEBS events do not match the trace. Currently, Large PEBS is used only in
1001certain circumstances:
1002 - hardware supports it
1003 - PEBS is used
1004 - event period is specified, instead of frequency
1005 - the sample type is limited to the following flags:
1006 PERF_SAMPLE_IP | PERF_SAMPLE_TID | PERF_SAMPLE_ADDR |
1007 PERF_SAMPLE_ID | PERF_SAMPLE_CPU | PERF_SAMPLE_STREAM_ID |
1008 PERF_SAMPLE_DATA_SRC | PERF_SAMPLE_IDENTIFIER |
1009 PERF_SAMPLE_TRANSACTION | PERF_SAMPLE_PHYS_ADDR |
1010 PERF_SAMPLE_REGS_INTR | PERF_SAMPLE_REGS_USER |
1011 PERF_SAMPLE_PERIOD (and sometimes) | PERF_SAMPLE_TIME
1012Because Intel PT sample mode uses a different sample type to the list above,
1013Large PEBS is not used with Intel PT sample mode. To avoid Large PEBS in other
1014cases, avoid specifying the event period i.e. avoid the 'perf record' -c option,
1015--count option, or 'period' config term.
1016
1017To disable trace decoding entirely, use the option --no-itrace.
1018
1019It is also possible to skip events generated (instructions, branches, transactions)
1020at the beginning. This is useful to ignore initialization code.
1021
1022 --itrace=i0nss1000000
1023
1024skips the first million instructions.
1025
1026The q option changes the way the trace is decoded. The decoding is much faster
1027but much less detailed. Specifically, with the q option, the decoder does not
1028decode TNT packets, and does not walk object code, but gets the ip from FUP and
1029TIP packets. The q option can be used with the b and i options but the period
1030is not used. The q option decodes more quickly, but is useful only if the
1031control flow of interest is represented or indicated by FUP, TIP, TIP.PGE, or
1032TIP.PGD packets (refer below). However the q option could be used to find time
1033ranges that could then be decoded fully using the --time option.
1034
1035What will *not* be decoded with the (single) q option:
1036
1037 - direct calls and jmps
1038 - conditional branches
1039 - non-branch instructions
1040
1041What *will* be decoded with the (single) q option:
1042
1043 - asynchronous branches such as interrupts
1044 - indirect branches
1045 - function return target address *if* the noretcomp config term (refer
1046 config terms section) was used
1047 - start of (control-flow) tracing
1048 - end of (control-flow) tracing, if it is not out of context
1049 - power events, ptwrite, transaction start and abort
1050 - instruction pointer associated with PSB packets
1051
1052Note the q option does not specify what events will be synthesized e.g. the p
1053option must be used also to show power events.
1054
1055Repeating the q option (double-q i.e. qq) results in even faster decoding and even
1056less detail. The decoder decodes only extended PSB (PSB+) packets, getting the
1057instruction pointer if there is a FUP packet within PSB+ (i.e. between PSB and
1058PSBEND). Note PSB packets occur regularly in the trace based on the psb_period
1059config term (refer config terms section). There will be a FUP packet if the
1060PSB+ occurs while control flow is being traced.
1061
1062What will *not* be decoded with the qq option:
1063
1064 - everything except instruction pointer associated with PSB packets
1065
1066What *will* be decoded with the qq option:
1067
1068 - instruction pointer associated with PSB packets
1069
1070The Z option is equivalent to having recorded a trace without TSC
1071(i.e. config term tsc=0). It can be useful to avoid timestamp issues when
1072decoding a trace of a virtual machine.
1073
1074
1075dump option
1076~~~~~~~~~~~
1077
1078perf script has an option (-D) to "dump" the events i.e. display the binary
1079data.
1080
1081When -D is used, Intel PT packets are displayed. The packet decoder does not
1082pay attention to PSB packets, but just decodes the bytes - so the packets seen
1083by the actual decoder may not be identical in places where the data is corrupt.
1084One example of that would be when the buffer-switching interrupt has been too
1085slow, and the buffer has been filled completely. In that case, the last packet
1086in the buffer might be truncated and immediately followed by a PSB as the trace
1087continues in the next buffer.
1088
1089To disable the display of Intel PT packets, combine the -D option with
1090--no-itrace.
1091
1092
1093perf report
1094-----------
1095
1096By default, perf report will decode trace data found in the perf.data file.
1097This can be further controlled by new option --itrace exactly the same as
1098perf script, with the exception that the default is --itrace=igxe.
1099
1100
1101perf inject
1102-----------
1103
1104perf inject also accepts the --itrace option in which case tracing data is
1105removed and replaced with the synthesized events. e.g.
1106
1107 perf inject --itrace -i perf.data -o perf.data.new
1108
1109Below is an example of using Intel PT with autofdo. It requires autofdo
1110(https://github.com/google/autofdo) and gcc version 5. The bubble
1111sort example is from the AutoFDO tutorial (https://gcc.gnu.org/wiki/AutoFDO/Tutorial)
1112amended to take the number of elements as a parameter.
1113
1114 $ gcc-5 -O3 sort.c -o sort_optimized
1115 $ ./sort_optimized 30000
1116 Bubble sorting array of 30000 elements
1117 2254 ms
1118
1119 $ cat ~/.perfconfig
1120 [intel-pt]
1121 mispred-all = on
1122
1123 $ perf record -e intel_pt//u ./sort 3000
1124 Bubble sorting array of 3000 elements
1125 58 ms
1126 [ perf record: Woken up 2 times to write data ]
1127 [ perf record: Captured and wrote 3.939 MB perf.data ]
1128 $ perf inject -i perf.data -o inj --itrace=i100usle --strip
1129 $ ./create_gcov --binary=./sort --profile=inj --gcov=sort.gcov -gcov_version=1
1130 $ gcc-5 -O3 -fauto-profile=sort.gcov sort.c -o sort_autofdo
1131 $ ./sort_autofdo 30000
1132 Bubble sorting array of 30000 elements
1133 2155 ms
1134
1135Note there is currently no advantage to using Intel PT instead of LBR, but
1136that may change in the future if greater use is made of the data.
1137
1138
1139PEBS via Intel PT
1140-----------------
1141
1142Some hardware has the feature to redirect PEBS records to the Intel PT trace.
1143Recording is selected by using the aux-output config term e.g.
1144
1145 perf record -c 10000 -e '{intel_pt/branch=0/,cycles/aux-output/ppp}' uname
1146
1147Note that currently, software only supports redirecting at most one PEBS event.
1148
1149To display PEBS events from the Intel PT trace, use the itrace 'o' option e.g.
1150
1151 perf script --itrace=oe
1152
1153XED
1154---
1155
1156include::build-xed.txt[]
1157
1158
1159Tracing Virtual Machines
1160------------------------
1161
1162Currently, only kernel tracing is supported and only with either "timeless" decoding
1163(i.e. no TSC timestamps) or VM Time Correlation. VM Time Correlation is an extra step
1164using 'perf inject' and requires unchanging VMX TSC Offset and no VMX TSC Scaling.
1165
1166Other limitations and caveats
1167
1168 VMX controls may suppress packets needed for decoding resulting in decoding errors
1169 VMX controls may block the perf NMI to the host potentially resulting in lost trace data
1170 Guest kernel self-modifying code (e.g. jump labels or JIT-compiled eBPF) will result in decoding errors
1171 Guest thread information is unknown
1172 Guest VCPU is unknown but may be able to be inferred from the host thread
1173 Callchains are not supported
1174
1175Example using "timeless" decoding
1176
1177Start VM
1178
1179 $ sudo virsh start kubuntu20.04
1180 Domain kubuntu20.04 started
1181
1182Mount the guest file system. Note sshfs needs -o direct_io to enable reading of proc files. root access is needed to read /proc/kcore.
1183
1184 $ mkdir vm0
1185 $ sshfs -o direct_io root@vm0:/ vm0
1186
1187Copy the guest /proc/kallsyms, /proc/modules and /proc/kcore
1188
1189 $ perf buildid-cache -v --kcore vm0/proc/kcore
1190 kcore added to build-id cache directory /home/user/.debug/[kernel.kcore]/9600f316a53a0f54278885e8d9710538ec5f6a08/2021021807494306
1191 $ KALLSYMS=/home/user/.debug/[kernel.kcore]/9600f316a53a0f54278885e8d9710538ec5f6a08/2021021807494306/kallsyms
1192
1193Find the VM process
1194
1195 $ ps -eLl | grep 'KVM\|PID'
1196 F S UID PID PPID LWP C PRI NI ADDR SZ WCHAN TTY TIME CMD
1197 3 S 64055 1430 1 1440 1 80 0 - 1921718 - ? 00:02:47 CPU 0/KVM
1198 3 S 64055 1430 1 1441 1 80 0 - 1921718 - ? 00:02:41 CPU 1/KVM
1199 3 S 64055 1430 1 1442 1 80 0 - 1921718 - ? 00:02:38 CPU 2/KVM
1200 3 S 64055 1430 1 1443 2 80 0 - 1921718 - ? 00:03:18 CPU 3/KVM
1201
1202Start an open-ended perf record, tracing the VM process, do something on the VM, and then ctrl-C to stop.
1203TSC is not supported and tsc=0 must be specified. That means mtc is useless, so add mtc=0.
1204However, IPC can still be determined, hence cyc=1 can be added.
1205Only kernel decoding is supported, so 'k' must be specified.
1206Intel PT traces both the host and the guest so --guest and --host need to be specified.
1207Without timestamps, --per-thread must be specified to distinguish threads.
1208
1209 $ sudo perf kvm --guest --host --guestkallsyms $KALLSYMS record --kcore -e intel_pt/tsc=0,mtc=0,cyc=1/k -p 1430 --per-thread
1210 ^C
1211 [ perf record: Woken up 1 times to write data ]
1212 [ perf record: Captured and wrote 5.829 MB ]
1213
1214perf script can be used to provide an instruction trace
1215
1216 $ perf script --guestkallsyms $KALLSYMS --insn-trace --xed -F+ipc | grep -C10 vmresume | head -21
1217 CPU 0/KVM 1440 ffffffff82133cdd __vmx_vcpu_run+0x3d ([kernel.kallsyms]) movq 0x48(%rax), %r9
1218 CPU 0/KVM 1440 ffffffff82133ce1 __vmx_vcpu_run+0x41 ([kernel.kallsyms]) movq 0x50(%rax), %r10
1219 CPU 0/KVM 1440 ffffffff82133ce5 __vmx_vcpu_run+0x45 ([kernel.kallsyms]) movq 0x58(%rax), %r11
1220 CPU 0/KVM 1440 ffffffff82133ce9 __vmx_vcpu_run+0x49 ([kernel.kallsyms]) movq 0x60(%rax), %r12
1221 CPU 0/KVM 1440 ffffffff82133ced __vmx_vcpu_run+0x4d ([kernel.kallsyms]) movq 0x68(%rax), %r13
1222 CPU 0/KVM 1440 ffffffff82133cf1 __vmx_vcpu_run+0x51 ([kernel.kallsyms]) movq 0x70(%rax), %r14
1223 CPU 0/KVM 1440 ffffffff82133cf5 __vmx_vcpu_run+0x55 ([kernel.kallsyms]) movq 0x78(%rax), %r15
1224 CPU 0/KVM 1440 ffffffff82133cf9 __vmx_vcpu_run+0x59 ([kernel.kallsyms]) movq (%rax), %rax
1225 CPU 0/KVM 1440 ffffffff82133cfc __vmx_vcpu_run+0x5c ([kernel.kallsyms]) callq 0xffffffff82133c40
1226 CPU 0/KVM 1440 ffffffff82133c40 vmx_vmenter+0x0 ([kernel.kallsyms]) jz 0xffffffff82133c46
1227 CPU 0/KVM 1440 ffffffff82133c42 vmx_vmenter+0x2 ([kernel.kallsyms]) vmresume IPC: 0.11 (50/445)
1228 :1440 1440 ffffffffbb678b06 native_write_msr+0x6 ([guest.kernel.kallsyms]) nopl %eax, (%rax,%rax,1)
1229 :1440 1440 ffffffffbb678b0b native_write_msr+0xb ([guest.kernel.kallsyms]) retq IPC: 0.04 (2/41)
1230 :1440 1440 ffffffffbb666646 lapic_next_deadline+0x26 ([guest.kernel.kallsyms]) data16 nop
1231 :1440 1440 ffffffffbb666648 lapic_next_deadline+0x28 ([guest.kernel.kallsyms]) xor %eax, %eax
1232 :1440 1440 ffffffffbb66664a lapic_next_deadline+0x2a ([guest.kernel.kallsyms]) popq %rbp
1233 :1440 1440 ffffffffbb66664b lapic_next_deadline+0x2b ([guest.kernel.kallsyms]) retq IPC: 0.16 (4/25)
1234 :1440 1440 ffffffffbb74607f clockevents_program_event+0x8f ([guest.kernel.kallsyms]) test %eax, %eax
1235 :1440 1440 ffffffffbb746081 clockevents_program_event+0x91 ([guest.kernel.kallsyms]) jz 0xffffffffbb74603c IPC: 0.06 (2/30)
1236 :1440 1440 ffffffffbb74603c clockevents_program_event+0x4c ([guest.kernel.kallsyms]) popq %rbx
1237 :1440 1440 ffffffffbb74603d clockevents_program_event+0x4d ([guest.kernel.kallsyms]) popq %r12
1238
1239Example using VM Time Correlation
1240
1241Start VM
1242
1243 $ sudo virsh start kubuntu20.04
1244 Domain kubuntu20.04 started
1245
1246Mount the guest file system. Note sshfs needs -o direct_io to enable reading of proc files. root access is needed to read /proc/kcore.
1247
1248 $ mkdir -p vm0
1249 $ sshfs -o direct_io root@vm0:/ vm0
1250
1251Copy the guest /proc/kallsyms, /proc/modules and /proc/kcore
1252
1253 $ perf buildid-cache -v --kcore vm0/proc/kcore
1254 same kcore found in /home/user/.debug/[kernel.kcore]/cc9c55a98c5e4ec0aeda69302554aabed5cd6491/2021021312450777
1255 $ KALLSYMS=/home/user/.debug/\[kernel.kcore\]/cc9c55a98c5e4ec0aeda69302554aabed5cd6491/2021021312450777/kallsyms
1256
1257Find the VM process
1258
1259 $ ps -eLl | grep 'KVM\|PID'
1260 F S UID PID PPID LWP C PRI NI ADDR SZ WCHAN TTY TIME CMD
1261 3 S 64055 16998 1 17005 13 80 0 - 1818189 - ? 00:00:16 CPU 0/KVM
1262 3 S 64055 16998 1 17006 4 80 0 - 1818189 - ? 00:00:05 CPU 1/KVM
1263 3 S 64055 16998 1 17007 3 80 0 - 1818189 - ? 00:00:04 CPU 2/KVM
1264 3 S 64055 16998 1 17008 4 80 0 - 1818189 - ? 00:00:05 CPU 3/KVM
1265
1266Start an open-ended perf record, tracing the VM process, do something on the VM, and then ctrl-C to stop.
1267IPC can be determined, hence cyc=1 can be added.
1268Only kernel decoding is supported, so 'k' must be specified.
1269Intel PT traces both the host and the guest so --guest and --host need to be specified.
1270
1271 $ sudo perf kvm --guest --host --guestkallsyms $KALLSYMS record --kcore -e intel_pt/cyc=1/k -p 16998
1272 ^C[ perf record: Woken up 1 times to write data ]
1273 [ perf record: Captured and wrote 9.041 MB perf.data.kvm ]
1274
1275Now 'perf inject' can be used to determine the VMX TCS Offset. Note, Intel PT TSC packets are
1276only 7-bytes, so the TSC Offset might differ from the actual value in the 8th byte. That will
1277have no effect i.e. the resulting timestamps will be correct anyway.
1278
1279 $ perf inject -i perf.data.kvm --vm-time-correlation=dry-run
1280 ERROR: Unknown TSC Offset for VMCS 0x1bff6a
1281 VMCS: 0x1bff6a TSC Offset 0xffffe42722c64c41
1282 ERROR: Unknown TSC Offset for VMCS 0x1cbc08
1283 VMCS: 0x1cbc08 TSC Offset 0xffffe42722c64c41
1284 ERROR: Unknown TSC Offset for VMCS 0x1c3ce8
1285 VMCS: 0x1c3ce8 TSC Offset 0xffffe42722c64c41
1286 ERROR: Unknown TSC Offset for VMCS 0x1cbce9
1287 VMCS: 0x1cbce9 TSC Offset 0xffffe42722c64c41
1288
1289Each virtual CPU has a different Virtual Machine Control Structure (VMCS)
1290shown above with the calculated TSC Offset. For an unchanging TSC Offset
1291they should all be the same for the same virtual machine.
1292
1293Now that the TSC Offset is known, it can be provided to 'perf inject'
1294
1295 $ perf inject -i perf.data.kvm --vm-time-correlation="dry-run 0xffffe42722c64c41"
1296
1297Note the options for 'perf inject' --vm-time-correlation are:
1298
1299 [ dry-run ] [ <TSC Offset> [ : <VMCS> [ , <VMCS> ]... ] ]...
1300
1301So it is possible to specify different TSC Offsets for different VMCS.
1302The option "dry-run" will cause the file to be processed but without updating it.
1303Note it is also possible to get a intel_pt.log file by adding option --itrace=d
1304
1305There were no errors so, do it for real
1306
1307 $ perf inject -i perf.data.kvm --vm-time-correlation=0xffffe42722c64c41 --force
1308
1309'perf script' can be used to see if there are any decoder errors
1310
1311 $ perf script -i perf.data.kvm --guestkallsyms $KALLSYMS --itrace=e-o
1312
1313There were none.
1314
1315'perf script' can be used to provide an instruction trace showing timestamps
1316
1317 $ perf script -i perf.data.kvm --guestkallsyms $KALLSYMS --insn-trace --xed -F+ipc | grep -C10 vmresume | head -21
1318 CPU 1/KVM 17006 [001] 11500.262865593: ffffffff82133cdd __vmx_vcpu_run+0x3d ([kernel.kallsyms]) movq 0x48(%rax), %r9
1319 CPU 1/KVM 17006 [001] 11500.262865593: ffffffff82133ce1 __vmx_vcpu_run+0x41 ([kernel.kallsyms]) movq 0x50(%rax), %r10
1320 CPU 1/KVM 17006 [001] 11500.262865593: ffffffff82133ce5 __vmx_vcpu_run+0x45 ([kernel.kallsyms]) movq 0x58(%rax), %r11
1321 CPU 1/KVM 17006 [001] 11500.262865593: ffffffff82133ce9 __vmx_vcpu_run+0x49 ([kernel.kallsyms]) movq 0x60(%rax), %r12
1322 CPU 1/KVM 17006 [001] 11500.262865593: ffffffff82133ced __vmx_vcpu_run+0x4d ([kernel.kallsyms]) movq 0x68(%rax), %r13
1323 CPU 1/KVM 17006 [001] 11500.262865593: ffffffff82133cf1 __vmx_vcpu_run+0x51 ([kernel.kallsyms]) movq 0x70(%rax), %r14
1324 CPU 1/KVM 17006 [001] 11500.262865593: ffffffff82133cf5 __vmx_vcpu_run+0x55 ([kernel.kallsyms]) movq 0x78(%rax), %r15
1325 CPU 1/KVM 17006 [001] 11500.262865593: ffffffff82133cf9 __vmx_vcpu_run+0x59 ([kernel.kallsyms]) movq (%rax), %rax
1326 CPU 1/KVM 17006 [001] 11500.262865593: ffffffff82133cfc __vmx_vcpu_run+0x5c ([kernel.kallsyms]) callq 0xffffffff82133c40
1327 CPU 1/KVM 17006 [001] 11500.262865593: ffffffff82133c40 vmx_vmenter+0x0 ([kernel.kallsyms]) jz 0xffffffff82133c46
1328 CPU 1/KVM 17006 [001] 11500.262866075: ffffffff82133c42 vmx_vmenter+0x2 ([kernel.kallsyms]) vmresume IPC: 0.05 (40/769)
1329 :17006 17006 [001] 11500.262869216: ffffffff82200cb0 asm_sysvec_apic_timer_interrupt+0x0 ([guest.kernel.kallsyms]) clac
1330 :17006 17006 [001] 11500.262869216: ffffffff82200cb3 asm_sysvec_apic_timer_interrupt+0x3 ([guest.kernel.kallsyms]) pushq $0xffffffffffffffff
1331 :17006 17006 [001] 11500.262869216: ffffffff82200cb5 asm_sysvec_apic_timer_interrupt+0x5 ([guest.kernel.kallsyms]) callq 0xffffffff82201160
1332 :17006 17006 [001] 11500.262869216: ffffffff82201160 error_entry+0x0 ([guest.kernel.kallsyms]) cld
1333 :17006 17006 [001] 11500.262869216: ffffffff82201161 error_entry+0x1 ([guest.kernel.kallsyms]) pushq %rsi
1334 :17006 17006 [001] 11500.262869216: ffffffff82201162 error_entry+0x2 ([guest.kernel.kallsyms]) movq 0x8(%rsp), %rsi
1335 :17006 17006 [001] 11500.262869216: ffffffff82201167 error_entry+0x7 ([guest.kernel.kallsyms]) movq %rdi, 0x8(%rsp)
1336 :17006 17006 [001] 11500.262869216: ffffffff8220116c error_entry+0xc ([guest.kernel.kallsyms]) pushq %rdx
1337 :17006 17006 [001] 11500.262869216: ffffffff8220116d error_entry+0xd ([guest.kernel.kallsyms]) pushq %rcx
1338 :17006 17006 [001] 11500.262869216: ffffffff8220116e error_entry+0xe ([guest.kernel.kallsyms]) pushq %rax
1339
1340
1341
1342SEE ALSO
1343--------
1344
1345linkperf:perf-record[1], linkperf:perf-script[1], linkperf:perf-report[1],
1346linkperf:perf-inject[1]
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=iybxwpe
105
106An interesting field that is not printed by default is 'flags' which can be
107displayed as follows:
108
109 perf script --itrace=iybxwpe -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=disasm
119
120or to use the xed disassembler, which requires installing the xed tool
121(see XED below):
122
123 perf script --insn-trace --xed
124
125Dumping all instructions in a long trace can be fairly slow. It is usually better
126to start with higher level decoding, like
127
128 perf script --call-trace
129
130or
131
132 perf script --call-ret-trace
133
134and then select a time range of interest. The time range can then be examined
135in detail with
136
137 perf script --time starttime,stoptime --insn-trace=disasm
138
139While examining the trace it's also useful to filter on specific CPUs using
140the -C option
141
142 perf script --time starttime,stoptime --insn-trace=disasm -C 1
143
144Dump all instructions in time range on CPU 1.
145
146Another interesting field that is not printed by default is 'ipc' which can be
147displayed as follows:
148
149 perf script --itrace=be -F+ipc
150
151There are two ways that instructions-per-cycle (IPC) can be calculated depending
152on the recording.
153
154If the 'cyc' config term (see config terms section below) was used, then IPC
155and cycle events are calculated using the cycle count from CYC packets, otherwise
156MTC packets are used - refer to the 'mtc' config term. When MTC is used, however,
157the values are less accurate because the timing is less accurate.
158
159Because Intel PT does not update the cycle count on every branch or instruction,
160the values will often be zero. When there are values, they will be the number
161of instructions and number of cycles since the last update, and thus represent
162the average IPC cycle count since the last IPC for that event type.
163Note IPC for "branches" events is calculated separately from IPC for "instructions"
164events.
165
166Even with the 'cyc' config term, it is possible to produce IPC information for
167every change of timestamp, but at the expense of accuracy. That is selected by
168specifying the itrace 'A' option. Due to the granularity of timestamps, the
169actual number of cycles increases even though the cycles reported does not.
170The number of instructions is known, but if IPC is reported, cycles can be too
171low and so IPC is too high. Note that inaccuracy decreases as the period of
172sampling increases i.e. if the number of cycles is too low by a small amount,
173that becomes less significant if the number of cycles is large. It may also be
174useful to use the 'A' option in conjunction with dlfilter-show-cycles.so to
175provide higher granularity cycle information.
176
177Also note that the IPC instruction count may or may not include the current
178instruction. If the cycle count is associated with an asynchronous branch
179(e.g. page fault or interrupt), then the instruction count does not include the
180current instruction, otherwise it does. That is consistent with whether or not
181that instruction has retired when the cycle count is updated.
182
183Another note, in the case of "branches" events, non-taken branches are not
184presently sampled, so IPC values for them do not appear e.g. a CYC packet with a
185TNT packet that starts with a non-taken branch. To see every possible IPC
186value, "instructions" events can be used e.g. --itrace=i0ns
187
188While it is possible to create scripts to analyze the data, an alternative
189approach is available to export the data to a sqlite or postgresql database.
190Refer to script export-to-sqlite.py or export-to-postgresql.py for more details,
191and to script exported-sql-viewer.py for an example of using the database.
192
193There is also script intel-pt-events.py which provides an example of how to
194unpack the raw data for power events and PTWRITE. The script also displays
195branches, and supports 2 additional modes selected by option:
196
197 - --insn-trace - instruction trace
198 - --src-trace - source trace
199
200The intel-pt-events.py script also has options:
201
202 - --all-switch-events - display all switch events, not only the last consecutive.
203 - --interleave [<n>] - interleave sample output for the same timestamp so that
204 no more than n samples for a CPU are displayed in a row. 'n' defaults to 4.
205 Note this only affects the order of output, and only when the timestamp is the
206 same.
207
208As mentioned above, it is easy to capture too much data. One way to limit the
209data captured is to use 'snapshot' mode which is explained further below.
210Refer to 'new snapshot option' and 'Intel PT modes of operation' further below.
211
212Another problem that will be experienced is decoder errors. They can be caused
213by inability to access the executed image, self-modified or JIT-ed code, or the
214inability to match side-band information (such as context switches and mmaps)
215which results in the decoder not knowing what code was executed.
216
217There is also the problem of perf not being able to copy the data fast enough,
218resulting in data lost because the buffer was full. See 'Buffer handling' below
219for more details.
220
221
222perf record
223-----------
224
225new event
226~~~~~~~~~
227
228The Intel PT kernel driver creates a new PMU for Intel PT. PMU events are
229selected by providing the PMU name followed by the "config" separated by slashes.
230An enhancement has been made to allow default "config" e.g. the option
231
232 -e intel_pt//
233
234will use a default config value. Currently that is the same as
235
236 -e intel_pt/tsc,noretcomp=0/
237
238which is the same as
239
240 -e intel_pt/tsc=1,noretcomp=0/
241
242Note there are now new config terms - see section 'config terms' further below.
243
244The config terms are listed in /sys/devices/intel_pt/format. They are bit
245fields within the config member of the struct perf_event_attr which is
246passed to the kernel by the perf_event_open system call. They correspond to bit
247fields in the IA32_RTIT_CTL MSR. Here is a list of them and their definitions:
248
249 $ grep -H . /sys/bus/event_source/devices/intel_pt/format/*
250 /sys/bus/event_source/devices/intel_pt/format/cyc:config:1
251 /sys/bus/event_source/devices/intel_pt/format/cyc_thresh:config:19-22
252 /sys/bus/event_source/devices/intel_pt/format/mtc:config:9
253 /sys/bus/event_source/devices/intel_pt/format/mtc_period:config:14-17
254 /sys/bus/event_source/devices/intel_pt/format/noretcomp:config:11
255 /sys/bus/event_source/devices/intel_pt/format/psb_period:config:24-27
256 /sys/bus/event_source/devices/intel_pt/format/tsc:config:10
257
258Note that the default config must be overridden for each term i.e.
259
260 -e intel_pt/noretcomp=0/
261
262is the same as:
263
264 -e intel_pt/tsc=1,noretcomp=0/
265
266So, to disable TSC packets use:
267
268 -e intel_pt/tsc=0/
269
270It is also possible to specify the config value explicitly:
271
272 -e intel_pt/config=0x400/
273
274Note that, as with all events, the event is suffixed with event modifiers:
275
276 u userspace
277 k kernel
278 h hypervisor
279 G guest
280 H host
281 p precise ip
282
283'h', 'G' and 'H' are for virtualization which are not used by Intel PT.
284'p' is also not relevant to Intel PT. So only options 'u' and 'k' are
285meaningful for Intel PT.
286
287perf_event_attr is displayed if the -vv option is used e.g.
288
289 ------------------------------------------------------------
290 perf_event_attr:
291 type 6
292 size 112
293 config 0x400
294 { sample_period, sample_freq } 1
295 sample_type IP|TID|TIME|CPU|IDENTIFIER
296 read_format ID
297 disabled 1
298 inherit 1
299 exclude_kernel 1
300 exclude_hv 1
301 enable_on_exec 1
302 sample_id_all 1
303 ------------------------------------------------------------
304 sys_perf_event_open: pid 31104 cpu 0 group_fd -1 flags 0x8
305 sys_perf_event_open: pid 31104 cpu 1 group_fd -1 flags 0x8
306 sys_perf_event_open: pid 31104 cpu 2 group_fd -1 flags 0x8
307 sys_perf_event_open: pid 31104 cpu 3 group_fd -1 flags 0x8
308 ------------------------------------------------------------
309
310
311config terms
312~~~~~~~~~~~~
313
314The June 2015 version of Intel 64 and IA-32 Architectures Software Developer
315Manuals, Chapter 36 Intel Processor Trace, defined new Intel PT features.
316Some of the features are reflect in new config terms. All the config terms are
317described below.
318
319tsc Always supported. Produces TSC timestamp packets to provide
320 timing information. In some cases it is possible to decode
321 without timing information, for example a per-thread context
322 that does not overlap executable memory maps.
323
324 The default config selects tsc (i.e. tsc=1).
325
326noretcomp Always supported. Disables "return compression" so a TIP packet
327 is produced when a function returns. Causes more packets to be
328 produced but might make decoding more reliable.
329
330 The default config does not select noretcomp (i.e. noretcomp=0).
331
332psb_period Allows the frequency of PSB packets to be specified.
333
334 The PSB packet is a synchronization packet that provides a
335 starting point for decoding or recovery from errors.
336
337 Support for psb_period is indicated by:
338
339 /sys/bus/event_source/devices/intel_pt/caps/psb_cyc
340
341 which contains "1" if the feature is supported and "0"
342 otherwise.
343
344 Valid values are given by:
345
346 /sys/bus/event_source/devices/intel_pt/caps/psb_periods
347
348 which contains a hexadecimal value, the bits of which represent
349 valid values e.g. bit 2 set means value 2 is valid.
350
351 The psb_period value is converted to the approximate number of
352 trace bytes between PSB packets as:
353
354 2 ^ (value + 11)
355
356 e.g. value 3 means 16KiB bytes between PSBs
357
358 If an invalid value is entered, the error message
359 will give a list of valid values e.g.
360
361 $ perf record -e intel_pt/psb_period=15/u uname
362 Invalid psb_period for intel_pt. Valid values are: 0-5
363
364 If MTC packets are selected, the default config selects a value
365 of 3 (i.e. psb_period=3) or the nearest lower value that is
366 supported (0 is always supported). Otherwise the default is 0.
367
368 If decoding is expected to be reliable and the buffer is large
369 then a large PSB period can be used.
370
371 Because a TSC packet is produced with PSB, the PSB period can
372 also affect the granularity to timing information in the absence
373 of MTC or CYC.
374
375mtc Produces MTC timing packets.
376
377 MTC packets provide finer grain timestamp information than TSC
378 packets. MTC packets record time using the hardware crystal
379 clock (CTC) which is related to TSC packets using a TMA packet.
380
381 Support for this feature is indicated by:
382
383 /sys/bus/event_source/devices/intel_pt/caps/mtc
384
385 which contains "1" if the feature is supported and
386 "0" otherwise.
387
388 The frequency of MTC packets can also be specified - see
389 mtc_period below.
390
391mtc_period Specifies how frequently MTC packets are produced - see mtc
392 above for how to determine if MTC packets are supported.
393
394 Valid values are given by:
395
396 /sys/bus/event_source/devices/intel_pt/caps/mtc_periods
397
398 which contains a hexadecimal value, the bits of which represent
399 valid values e.g. bit 2 set means value 2 is valid.
400
401 The mtc_period value is converted to the MTC frequency as:
402
403 CTC-frequency / (2 ^ value)
404
405 e.g. value 3 means one eighth of CTC-frequency
406
407 Where CTC is the hardware crystal clock, the frequency of which
408 can be related to TSC via values provided in cpuid leaf 0x15.
409
410 If an invalid value is entered, the error message
411 will give a list of valid values e.g.
412
413 $ perf record -e intel_pt/mtc_period=15/u uname
414 Invalid mtc_period for intel_pt. Valid values are: 0,3,6,9
415
416 The default value is 3 or the nearest lower value
417 that is supported (0 is always supported).
418
419cyc Produces CYC timing packets.
420
421 CYC packets provide even finer grain timestamp information than
422 MTC and TSC packets. A CYC packet contains the number of CPU
423 cycles since the last CYC packet. Unlike MTC and TSC packets,
424 CYC packets are only sent when another packet is also sent.
425
426 Support for this feature is indicated by:
427
428 /sys/bus/event_source/devices/intel_pt/caps/psb_cyc
429
430 which contains "1" if the feature is supported and
431 "0" otherwise.
432
433 The number of CYC packets produced can be reduced by specifying
434 a threshold - see cyc_thresh below.
435
436cyc_thresh Specifies how frequently CYC packets are produced - see cyc
437 above for how to determine if CYC packets are supported.
438
439 Valid cyc_thresh values are given by:
440
441 /sys/bus/event_source/devices/intel_pt/caps/cycle_thresholds
442
443 which contains a hexadecimal value, the bits of which represent
444 valid values e.g. bit 2 set means value 2 is valid.
445
446 The cyc_thresh value represents the minimum number of CPU cycles
447 that must have passed before a CYC packet can be sent. The
448 number of CPU cycles is:
449
450 2 ^ (value - 1)
451
452 e.g. value 4 means 8 CPU cycles must pass before a CYC packet
453 can be sent. Note a CYC packet is still only sent when another
454 packet is sent, not at, e.g. every 8 CPU cycles.
455
456 If an invalid value is entered, the error message
457 will give a list of valid values e.g.
458
459 $ perf record -e intel_pt/cyc,cyc_thresh=15/u uname
460 Invalid cyc_thresh for intel_pt. Valid values are: 0-12
461
462 CYC packets are not requested by default.
463
464pt Specifies pass-through which enables the 'branch' config term.
465
466 The default config selects 'pt' if it is available, so a user will
467 never need to specify this term.
468
469branch Enable branch tracing. Branch tracing is enabled by default so to
470 disable branch tracing use 'branch=0'.
471
472 The default config selects 'branch' if it is available.
473
474ptw Enable PTWRITE packets which are produced when a ptwrite instruction
475 is executed.
476
477 Support for this feature is indicated by:
478
479 /sys/bus/event_source/devices/intel_pt/caps/ptwrite
480
481 which contains "1" if the feature is supported and
482 "0" otherwise.
483
484 As an alternative, refer to "Emulated PTWRITE" further below.
485
486fup_on_ptw Enable a FUP packet to follow the PTWRITE packet. The FUP packet
487 provides the address of the ptwrite instruction. In the absence of
488 fup_on_ptw, the decoder will use the address of the previous branch
489 if branch tracing is enabled, otherwise the address will be zero.
490 Note that fup_on_ptw will work even when branch tracing is disabled.
491
492pwr_evt Enable power events. The power events provide information about
493 changes to the CPU C-state.
494
495 Support for this feature is indicated by:
496
497 /sys/bus/event_source/devices/intel_pt/caps/power_event_trace
498
499 which contains "1" if the feature is supported and
500 "0" otherwise.
501
502event Enable Event Trace. The events provide information about asynchronous
503 events.
504
505 Support for this feature is indicated by:
506
507 /sys/bus/event_source/devices/intel_pt/caps/event_trace
508
509 which contains "1" if the feature is supported and
510 "0" otherwise.
511
512notnt Disable TNT packets. Without TNT packets, it is not possible to walk
513 executable code to reconstruct control flow, however FUP, TIP, TIP.PGE
514 and TIP.PGD packets still indicate asynchronous control flow, and (if
515 return compression is disabled - see noretcomp) return statements.
516 The advantage of eliminating TNT packets is reducing the size of the
517 trace and corresponding tracing overhead.
518
519 Support for this feature is indicated by:
520
521 /sys/bus/event_source/devices/intel_pt/caps/tnt_disable
522
523 which contains "1" if the feature is supported and
524 "0" otherwise.
525
526
527AUX area sampling option
528~~~~~~~~~~~~~~~~~~~~~~~~
529
530To select Intel PT "sampling" the AUX area sampling option can be used:
531
532 --aux-sample
533
534Optionally it can be followed by the sample size in bytes e.g.
535
536 --aux-sample=8192
537
538In addition, the Intel PT event to sample must be defined e.g.
539
540 -e intel_pt//u
541
542Samples on other events will be created containing Intel PT data e.g. the
543following will create Intel PT samples on the branch-misses event, note the
544events must be grouped using {}:
545
546 perf record --aux-sample -e '{intel_pt//u,branch-misses:u}'
547
548An alternative to '--aux-sample' is to add the config term 'aux-sample-size' to
549events. In this case, the grouping is implied e.g.
550
551 perf record -e intel_pt//u -e branch-misses/aux-sample-size=8192/u
552
553is the same as:
554
555 perf record -e '{intel_pt//u,branch-misses/aux-sample-size=8192/u}'
556
557but allows for also using an address filter e.g.:
558
559 perf record -e intel_pt//u --filter 'filter * @/bin/ls' -e branch-misses/aux-sample-size=8192/u -- ls
560
561It is important to select a sample size that is big enough to contain at least
562one PSB packet. If not a warning will be displayed:
563
564 Intel PT sample size (%zu) may be too small for PSB period (%zu)
565
566The calculation used for that is: if sample_size <= psb_period + 256 display the
567warning. When sampling is used, psb_period defaults to 0 (2KiB).
568
569The default sample size is 4KiB.
570
571The sample size is passed in aux_sample_size in struct perf_event_attr. The
572sample size is limited by the maximum event size which is 64KiB. It is
573difficult to know how big the event might be without the trace sample attached,
574but the tool validates that the sample size is not greater than 60KiB.
575
576
577new snapshot option
578~~~~~~~~~~~~~~~~~~~
579
580The difference between full trace and snapshot from the kernel's perspective is
581that in full trace we don't overwrite trace data that the user hasn't collected
582yet (and indicated that by advancing aux_tail), whereas in snapshot mode we let
583the trace run and overwrite older data in the buffer so that whenever something
584interesting happens, we can stop it and grab a snapshot of what was going on
585around that interesting moment.
586
587To select snapshot mode a new option has been added:
588
589 -S
590
591Optionally it can be followed by the snapshot size e.g.
592
593 -S0x100000
594
595The default snapshot size is the auxtrace mmap size. If neither auxtrace mmap size
596nor snapshot size is specified, then the default is 4MiB for privileged users
597(or if /proc/sys/kernel/perf_event_paranoid < 0), 128KiB for unprivileged users.
598If an unprivileged user does not specify mmap pages, the mmap pages will be
599reduced as described in the 'new auxtrace mmap size option' section below.
600
601The snapshot size is displayed if the option -vv is used e.g.
602
603 Intel PT snapshot size: %zu
604
605
606new auxtrace mmap size option
607~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
608
609Intel PT buffer size is specified by an addition to the -m option e.g.
610
611 -m,16
612
613selects a buffer size of 16 pages i.e. 64KiB.
614
615Note that the existing functionality of -m is unchanged. The auxtrace mmap size
616is specified by the optional addition of a comma and the value.
617
618The default auxtrace mmap size for Intel PT is 4MiB/page_size for privileged users
619(or if /proc/sys/kernel/perf_event_paranoid < 0), 128KiB for unprivileged users.
620If an unprivileged user does not specify mmap pages, the mmap pages will be
621reduced from the default 512KiB/page_size to 256KiB/page_size, otherwise the
622user is likely to get an error as they exceed their mlock limit (Max locked
623memory as shown in /proc/self/limits). Note that perf does not count the first
624512KiB (actually /proc/sys/kernel/perf_event_mlock_kb minus 1 page) per cpu
625against the mlock limit so an unprivileged user is allowed 512KiB per cpu plus
626their mlock limit (which defaults to 64KiB but is not multiplied by the number
627of cpus).
628
629In full-trace mode, powers of two are allowed for buffer size, with a minimum
630size of 2 pages. In snapshot mode or sampling mode, it is the same but the
631minimum size is 1 page.
632
633The mmap size and auxtrace mmap size are displayed if the -vv option is used e.g.
634
635 mmap length 528384
636 auxtrace mmap length 4198400
637
638
639Intel PT modes of operation
640~~~~~~~~~~~~~~~~~~~~~~~~~~~
641
642Intel PT can be used in 3 modes:
643 full-trace mode
644 sample mode
645 snapshot mode
646
647Full-trace mode traces continuously e.g.
648
649 perf record -e intel_pt//u uname
650
651Sample mode attaches a Intel PT sample to other events e.g.
652
653 perf record --aux-sample -e intel_pt//u -e branch-misses:u
654
655Snapshot mode captures the available data when a signal is sent or "snapshot"
656control command is issued. e.g. using a signal
657
658 perf record -v -e intel_pt//u -S ./loopy 1000000000 &
659 [1] 11435
660 kill -USR2 11435
661 Recording AUX area tracing snapshot
662
663Note that the signal sent is SIGUSR2.
664Note that "Recording AUX area tracing snapshot" is displayed because the -v
665option is used.
666
667The advantage of using "snapshot" control command is that the access is
668controlled by access to a FIFO e.g.
669
670 $ mkfifo perf.control
671 $ mkfifo perf.ack
672 $ cat perf.ack &
673 [1] 15235
674 $ sudo ~/bin/perf record --control fifo:perf.control,perf.ack -S -e intel_pt//u -- sleep 60 &
675 [2] 15243
676 $ ps -e | grep perf
677 15244 pts/1 00:00:00 perf
678 $ kill -USR2 15244
679 bash: kill: (15244) - Operation not permitted
680 $ echo snapshot > perf.control
681 ack
682
683The 3 Intel PT modes of operation cannot be used together.
684
685
686Buffer handling
687~~~~~~~~~~~~~~~
688
689There may be buffer limitations (i.e. single ToPa entry) which means that actual
690buffer sizes are limited to powers of 2 up to 4MiB (MAX_PAGE_ORDER). In order to
691provide other sizes, and in particular an arbitrarily large size, multiple
692buffers are logically concatenated. However an interrupt must be used to switch
693between buffers. That has two potential problems:
694 a) the interrupt may not be handled in time so that the current buffer
695 becomes full and some trace data is lost.
696 b) the interrupts may slow the system and affect the performance
697 results.
698
699If trace data is lost, the driver sets 'truncated' in the PERF_RECORD_AUX event
700which the tools report as an error.
701
702In full-trace mode, the driver waits for data to be copied out before allowing
703the (logical) buffer to wrap-around. If data is not copied out quickly enough,
704again 'truncated' is set in the PERF_RECORD_AUX event. If the driver has to
705wait, the intel_pt event gets disabled. Because it is difficult to know when
706that happens, perf tools always re-enable the intel_pt event after copying out
707data.
708
709
710Intel PT and build ids
711~~~~~~~~~~~~~~~~~~~~~~
712
713By default "perf record" post-processes the event stream to find all build ids
714for executables for all addresses sampled. Deliberately, Intel PT is not
715decoded for that purpose (it would take too long). Instead the build ids for
716all executables encountered (due to mmap, comm or task events) are included
717in the perf.data file.
718
719To see buildids included in the perf.data file use the command:
720
721 perf buildid-list
722
723If the perf.data file contains Intel PT data, that is the same as:
724
725 perf buildid-list --with-hits
726
727
728Snapshot mode and event disabling
729~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
730
731In order to make a snapshot, the intel_pt event is disabled using an IOCTL,
732namely PERF_EVENT_IOC_DISABLE. However doing that can also disable the
733collection of side-band information. In order to prevent that, a dummy
734software event has been introduced that permits tracking events (like mmaps) to
735continue to be recorded while intel_pt is disabled. That is important to ensure
736there is complete side-band information to allow the decoding of subsequent
737snapshots.
738
739A test has been created for that. To find the test:
740
741 perf test list
742 ...
743 23: Test using a dummy software event to keep tracking
744
745To run the test:
746
747 perf test 23
748 23: Test using a dummy software event to keep tracking : Ok
749
750
751perf record modes (nothing new here)
752~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
753
754perf record essentially operates in one of three modes:
755 per thread
756 per cpu
757 workload only
758
759"per thread" mode is selected by -t or by --per-thread (with -p or -u or just a
760workload).
761"per cpu" is selected by -C or -a.
762"workload only" mode is selected by not using the other options but providing a
763command to run (i.e. the workload).
764
765In per-thread mode an exact list of threads is traced. There is no inheritance.
766Each thread has its own event buffer.
767
768In per-cpu mode all processes (or processes from the selected cgroup i.e. -G
769option, or processes selected with -p or -u) are traced. Each cpu has its own
770buffer. Inheritance is allowed.
771
772In workload-only mode, the workload is traced but with per-cpu buffers.
773Inheritance is allowed. Note that you can now trace a workload in per-thread
774mode by using the --per-thread option.
775
776
777Privileged vs non-privileged users
778~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
779
780Unless /proc/sys/kernel/perf_event_paranoid is set to -1, unprivileged users
781have memory limits imposed upon them. That affects what buffer sizes they can
782have as outlined above.
783
784The v4.2 kernel introduced support for a context switch metadata event,
785PERF_RECORD_SWITCH, which allows unprivileged users to see when their processes
786are scheduled out and in, just not by whom, which is left for the
787PERF_RECORD_SWITCH_CPU_WIDE, that is only accessible in system wide context,
788which in turn requires CAP_PERFMON or CAP_SYS_ADMIN.
789
790Please see the 45ac1403f564 ("perf: Add PERF_RECORD_SWITCH to indicate context
791switches") commit, that introduces these metadata events for further info.
792
793When working with kernels < v4.2, the following considerations must be taken,
794as the sched:sched_switch tracepoints will be used to receive such information:
795
796Unless /proc/sys/kernel/perf_event_paranoid is set to -1, unprivileged users are
797not permitted to use tracepoints which means there is insufficient side-band
798information to decode Intel PT in per-cpu mode, and potentially workload-only
799mode too if the workload creates new processes.
800
801Note also, that to use tracepoints, read-access to debugfs is required. So if
802debugfs is not mounted or the user does not have read-access, it will again not
803be possible to decode Intel PT in per-cpu mode.
804
805
806sched_switch tracepoint
807~~~~~~~~~~~~~~~~~~~~~~~
808
809The sched_switch tracepoint is used to provide side-band data for Intel PT
810decoding in kernels where the PERF_RECORD_SWITCH metadata event isn't
811available.
812
813The sched_switch events are automatically added. e.g. the second event shown
814below:
815
816 $ perf record -vv -e intel_pt//u uname
817 ------------------------------------------------------------
818 perf_event_attr:
819 type 6
820 size 112
821 config 0x400
822 { sample_period, sample_freq } 1
823 sample_type IP|TID|TIME|CPU|IDENTIFIER
824 read_format ID
825 disabled 1
826 inherit 1
827 exclude_kernel 1
828 exclude_hv 1
829 enable_on_exec 1
830 sample_id_all 1
831 ------------------------------------------------------------
832 sys_perf_event_open: pid 31104 cpu 0 group_fd -1 flags 0x8
833 sys_perf_event_open: pid 31104 cpu 1 group_fd -1 flags 0x8
834 sys_perf_event_open: pid 31104 cpu 2 group_fd -1 flags 0x8
835 sys_perf_event_open: pid 31104 cpu 3 group_fd -1 flags 0x8
836 ------------------------------------------------------------
837 perf_event_attr:
838 type 2
839 size 112
840 config 0x108
841 { sample_period, sample_freq } 1
842 sample_type IP|TID|TIME|CPU|PERIOD|RAW|IDENTIFIER
843 read_format ID
844 inherit 1
845 sample_id_all 1
846 exclude_guest 1
847 ------------------------------------------------------------
848 sys_perf_event_open: pid -1 cpu 0 group_fd -1 flags 0x8
849 sys_perf_event_open: pid -1 cpu 1 group_fd -1 flags 0x8
850 sys_perf_event_open: pid -1 cpu 2 group_fd -1 flags 0x8
851 sys_perf_event_open: pid -1 cpu 3 group_fd -1 flags 0x8
852 ------------------------------------------------------------
853 perf_event_attr:
854 type 1
855 size 112
856 config 0x9
857 { sample_period, sample_freq } 1
858 sample_type IP|TID|TIME|IDENTIFIER
859 read_format ID
860 disabled 1
861 inherit 1
862 exclude_kernel 1
863 exclude_hv 1
864 mmap 1
865 comm 1
866 enable_on_exec 1
867 task 1
868 sample_id_all 1
869 mmap2 1
870 comm_exec 1
871 ------------------------------------------------------------
872 sys_perf_event_open: pid 31104 cpu 0 group_fd -1 flags 0x8
873 sys_perf_event_open: pid 31104 cpu 1 group_fd -1 flags 0x8
874 sys_perf_event_open: pid 31104 cpu 2 group_fd -1 flags 0x8
875 sys_perf_event_open: pid 31104 cpu 3 group_fd -1 flags 0x8
876 mmap size 528384B
877 AUX area mmap length 4194304
878 perf event ring buffer mmapped per cpu
879 Synthesizing auxtrace information
880 Linux
881 [ perf record: Woken up 1 times to write data ]
882 [ perf record: Captured and wrote 0.042 MB perf.data ]
883
884Note, the sched_switch event is only added if the user is permitted to use it
885and only in per-cpu mode.
886
887Note also, the sched_switch event is only added if TSC packets are requested.
888That is because, in the absence of timing information, the sched_switch events
889cannot be matched against the Intel PT trace.
890
891
892perf script
893-----------
894
895By default, perf script will decode trace data found in the perf.data file.
896This can be further controlled by new option --itrace.
897
898
899New --itrace option
900~~~~~~~~~~~~~~~~~~~
901
902Having no option is the same as
903
904 --itrace
905
906which, in turn, is the same as
907
908 --itrace=cepwxy
909
910The letters are:
911
912 i synthesize "instructions" events
913 y synthesize "cycles" events
914 b synthesize "branches" events
915 x synthesize "transactions" events
916 w synthesize "ptwrite" events
917 p synthesize "power" events (incl. PSB events)
918 c synthesize branches events (calls only)
919 r synthesize branches events (returns only)
920 o synthesize PEBS-via-PT events
921 I synthesize Event Trace events
922 e synthesize tracing error events
923 d create a debug log
924 g synthesize a call chain (use with i or x)
925 G synthesize a call chain on existing event records
926 l synthesize last branch entries (use with i or x)
927 L synthesize last branch entries on existing event records
928 s skip initial number of events
929 q quicker (less detailed) decoding
930 A approximate IPC
931 Z prefer to ignore timestamps (so-called "timeless" decoding)
932
933"Instructions" events look like they were recorded by "perf record -e
934instructions".
935
936"Cycles" events look like they were recorded by "perf record -e cycles"
937(ie., the default). Note that even with CYC packets enabled and no sampling,
938these are not fully accurate, since CYC packets are not emitted for each
939instruction, only when some other event (like an indirect branch, or a
940TNT packet representing multiple branches) happens causes a packet to
941be emitted. Thus, it is more effective for attributing cycles to functions
942(and possibly basic blocks) than to individual instructions, although it
943is not even perfect for functions (although it becomes better if the noretcomp
944option is active).
945
946"Branches" events look like they were recorded by "perf record -e branches". "c"
947and "r" can be combined to get calls and returns.
948
949"Transactions" events correspond to the start or end of transactions. The
950'flags' field can be used in perf script to determine whether the event is a
951transaction start, commit or abort.
952
953Note that "instructions", "cycles", "branches" and "transactions" events
954depend on code flow packets which can be disabled by using the config term
955"branch=0". Refer to the config terms section above.
956
957"ptwrite" events record the payload of the ptwrite instruction and whether
958"fup_on_ptw" was used. "ptwrite" events depend on PTWRITE packets which are
959recorded only if the "ptw" config term was used. Refer to the config terms
960section above. perf script "synth" field displays "ptwrite" information like
961this: "ip: 0 payload: 0x123456789abcdef0" where "ip" is 1 if "fup_on_ptw" was
962used.
963
964"Power" events correspond to power event packets and CBR (core-to-bus ratio)
965packets. While CBR packets are always recorded when tracing is enabled, power
966event packets are recorded only if the "pwr_evt" config term was used. Refer to
967the config terms section above. The power events record information about
968C-state changes, whereas CBR is indicative of CPU frequency. perf script
969"event,synth" fields display information like this:
970
971 cbr: cbr: 22 freq: 2189 MHz (200%)
972 mwait: hints: 0x60 extensions: 0x1
973 pwre: hw: 0 cstate: 2 sub-cstate: 0
974 exstop: ip: 1
975 pwrx: deepest cstate: 2 last cstate: 2 wake reason: 0x4
976
977Where:
978
979 "cbr" includes the frequency and the percentage of maximum non-turbo
980 "mwait" shows mwait hints and extensions
981 "pwre" shows C-state transitions (to a C-state deeper than C0) and
982 whether initiated by hardware
983 "exstop" indicates execution stopped and whether the IP was recorded
984 exactly,
985 "pwrx" indicates return to C0
986
987For more details refer to the Intel 64 and IA-32 Architectures Software
988Developer Manuals.
989
990PSB events show when a PSB+ occurred and also the byte-offset in the trace.
991Emitting a PSB+ can cause a CPU a slight delay. When doing timing analysis
992of code with Intel PT, it is useful to know if a timing bubble was caused
993by Intel PT or not.
994
995Error events show where the decoder lost the trace. Error events
996are quite important. Users must know if what they are seeing is a complete
997picture or not. The "e" option may be followed by flags which affect what errors
998will or will not be reported. Each flag must be preceded by either '+' or '-'.
999The flags supported by Intel PT are:
1000
1001 -o Suppress overflow errors
1002 -l Suppress trace data lost errors
1003
1004For example, for errors but not overflow or data lost errors:
1005
1006 --itrace=e-o-l
1007
1008The "d" option will cause the creation of a file "intel_pt.log" containing all
1009decoded packets and instructions. Note that this option slows down the decoder
1010and that the resulting file may be very large. The "d" option may be followed
1011by flags which affect what debug messages will or will not be logged. Each flag
1012must be preceded by either '+' or '-'. The flags support by Intel PT are:
1013
1014 -a Suppress logging of perf events
1015 +a Log all perf events
1016 +e Output only on decoding errors (size configurable)
1017 +o Output to stdout instead of "intel_pt.log"
1018
1019By default, logged perf events are filtered by any specified time ranges, but
1020flag +a overrides that. The +e flag can be useful for analyzing errors. By
1021default, the log size in that case is 16384 bytes, but can be altered by
1022linkperf:perf-config[1] e.g. perf config itrace.debug-log-buffer-size=30000
1023
1024In addition, the period of the "instructions" event can be specified. e.g.
1025
1026 --itrace=i10us
1027
1028sets the period to 10us i.e. one instruction sample is synthesized for each 10
1029microseconds of trace. Alternatives to "us" are "ms" (milliseconds),
1030"ns" (nanoseconds), "t" (TSC ticks) or "i" (instructions).
1031
1032"ms", "us" and "ns" are converted to TSC ticks.
1033
1034The timing information included with Intel PT does not give the time of every
1035instruction. Consequently, for the purpose of sampling, the decoder estimates
1036the time since the last timing packet based on 1 tick per instruction. The time
1037on the sample is *not* adjusted and reflects the last known value of TSC.
1038
1039For Intel PT, the default period is 100us.
1040
1041Setting it to a zero period means "as often as possible".
1042
1043In the case of Intel PT that is the same as a period of 1 and a unit of
1044'instructions' (i.e. --itrace=i1i).
1045
1046Also the call chain size (default 16, max. 1024) for instructions or
1047transactions events can be specified. e.g.
1048
1049 --itrace=ig32
1050 --itrace=xg32
1051
1052Also the number of last branch entries (default 64, max. 1024) for instructions or
1053transactions events can be specified. e.g.
1054
1055 --itrace=il10
1056 --itrace=xl10
1057
1058Note that last branch entries are cleared for each sample, so there is no overlap
1059from one sample to the next.
1060
1061The G and L options are designed in particular for sample mode, and work much
1062like g and l but add call chain and branch stack to the other selected events
1063instead of synthesized events. For example, to record branch-misses events for
1064'ls' and then add a call chain derived from the Intel PT trace:
1065
1066 perf record --aux-sample -e '{intel_pt//u,branch-misses:u}' -- ls
1067 perf report --itrace=Ge
1068
1069Although in fact G is a default for perf report, so that is the same as just:
1070
1071 perf report
1072
1073One caveat with the G and L options is that they work poorly with "Large PEBS".
1074Large PEBS means PEBS records will be accumulated by hardware and the written
1075into the event buffer in one go. That reduces interrupts, but can give very
1076late timestamps. Because the Intel PT trace is synchronized by timestamps,
1077the PEBS events do not match the trace. Currently, Large PEBS is used only in
1078certain circumstances:
1079 - hardware supports it
1080 - PEBS is used
1081 - event period is specified, instead of frequency
1082 - the sample type is limited to the following flags:
1083 PERF_SAMPLE_IP | PERF_SAMPLE_TID | PERF_SAMPLE_ADDR |
1084 PERF_SAMPLE_ID | PERF_SAMPLE_CPU | PERF_SAMPLE_STREAM_ID |
1085 PERF_SAMPLE_DATA_SRC | PERF_SAMPLE_IDENTIFIER |
1086 PERF_SAMPLE_TRANSACTION | PERF_SAMPLE_PHYS_ADDR |
1087 PERF_SAMPLE_REGS_INTR | PERF_SAMPLE_REGS_USER |
1088 PERF_SAMPLE_PERIOD (and sometimes) | PERF_SAMPLE_TIME
1089Because Intel PT sample mode uses a different sample type to the list above,
1090Large PEBS is not used with Intel PT sample mode. To avoid Large PEBS in other
1091cases, avoid specifying the event period i.e. avoid the 'perf record' -c option,
1092--count option, or 'period' config term.
1093
1094To disable trace decoding entirely, use the option --no-itrace.
1095
1096It is also possible to skip events generated (instructions, branches, transactions)
1097at the beginning. This is useful to ignore initialization code.
1098
1099 --itrace=i0nss1000000
1100
1101skips the first million instructions.
1102
1103The q option changes the way the trace is decoded. The decoding is much faster
1104but much less detailed. Specifically, with the q option, the decoder does not
1105decode TNT packets, and does not walk object code, but gets the ip from FUP and
1106TIP packets. The q option can be used with the b and i options but the period
1107is not used. The q option decodes more quickly, but is useful only if the
1108control flow of interest is represented or indicated by FUP, TIP, TIP.PGE, or
1109TIP.PGD packets (refer below). However the q option could be used to find time
1110ranges that could then be decoded fully using the --time option.
1111
1112What will *not* be decoded with the (single) q option:
1113
1114 - direct calls and jmps
1115 - conditional branches
1116 - non-branch instructions
1117
1118What *will* be decoded with the (single) q option:
1119
1120 - asynchronous branches such as interrupts
1121 - indirect branches
1122 - function return target address *if* the noretcomp config term (refer
1123 config terms section) was used
1124 - start of (control-flow) tracing
1125 - end of (control-flow) tracing, if it is not out of context
1126 - power events, ptwrite, transaction start and abort
1127 - instruction pointer associated with PSB packets
1128
1129Note the q option does not specify what events will be synthesized e.g. the p
1130option must be used also to show power events.
1131
1132Repeating the q option (double-q i.e. qq) results in even faster decoding and even
1133less detail. The decoder decodes only extended PSB (PSB+) packets, getting the
1134instruction pointer if there is a FUP packet within PSB+ (i.e. between PSB and
1135PSBEND). Note PSB packets occur regularly in the trace based on the psb_period
1136config term (refer config terms section). There will be a FUP packet if the
1137PSB+ occurs while control flow is being traced.
1138
1139What will *not* be decoded with the qq option:
1140
1141 - everything except instruction pointer associated with PSB packets
1142
1143What *will* be decoded with the qq option:
1144
1145 - instruction pointer associated with PSB packets
1146
1147The Z option is equivalent to having recorded a trace without TSC
1148(i.e. config term tsc=0). It can be useful to avoid timestamp issues when
1149decoding a trace of a virtual machine.
1150
1151
1152dlfilter-show-cycles.so
1153~~~~~~~~~~~~~~~~~~~~~~~
1154
1155Cycles can be displayed using dlfilter-show-cycles.so in which case the itrace A
1156option can be useful to provide higher granularity cycle information:
1157
1158 perf script --itrace=A --call-trace --dlfilter dlfilter-show-cycles.so
1159
1160To see a list of dlfilters:
1161
1162 perf script -v --list-dlfilters
1163
1164See also linkperf:perf-dlfilters[1]
1165
1166
1167dump option
1168~~~~~~~~~~~
1169
1170perf script has an option (-D) to "dump" the events i.e. display the binary
1171data.
1172
1173When -D is used, Intel PT packets are displayed. The packet decoder does not
1174pay attention to PSB packets, but just decodes the bytes - so the packets seen
1175by the actual decoder may not be identical in places where the data is corrupt.
1176One example of that would be when the buffer-switching interrupt has been too
1177slow, and the buffer has been filled completely. In that case, the last packet
1178in the buffer might be truncated and immediately followed by a PSB as the trace
1179continues in the next buffer.
1180
1181To disable the display of Intel PT packets, combine the -D option with
1182--no-itrace.
1183
1184
1185perf report
1186-----------
1187
1188By default, perf report will decode trace data found in the perf.data file.
1189This can be further controlled by new option --itrace exactly the same as
1190perf script, with the exception that the default is --itrace=igxe.
1191
1192
1193perf inject
1194-----------
1195
1196perf inject also accepts the --itrace option in which case tracing data is
1197removed and replaced with the synthesized events. e.g.
1198
1199 perf inject --itrace -i perf.data -o perf.data.new
1200
1201Below is an example of using Intel PT with autofdo. It requires autofdo
1202(https://github.com/google/autofdo) and gcc version 5. The bubble
1203sort example is from the AutoFDO tutorial (https://gcc.gnu.org/wiki/AutoFDO/Tutorial)
1204amended to take the number of elements as a parameter.
1205
1206 $ gcc-5 -O3 sort.c -o sort_optimized
1207 $ ./sort_optimized 30000
1208 Bubble sorting array of 30000 elements
1209 2254 ms
1210
1211 $ cat ~/.perfconfig
1212 [intel-pt]
1213 mispred-all = on
1214
1215 $ perf record -e intel_pt//u ./sort 3000
1216 Bubble sorting array of 3000 elements
1217 58 ms
1218 [ perf record: Woken up 2 times to write data ]
1219 [ perf record: Captured and wrote 3.939 MB perf.data ]
1220 $ perf inject -i perf.data -o inj --itrace=i100usle --strip
1221 $ ./create_gcov --binary=./sort --profile=inj --gcov=sort.gcov -gcov_version=1
1222 $ gcc-5 -O3 -fauto-profile=sort.gcov sort.c -o sort_autofdo
1223 $ ./sort_autofdo 30000
1224 Bubble sorting array of 30000 elements
1225 2155 ms
1226
1227Note there is currently no advantage to using Intel PT instead of LBR, but
1228that may change in the future if greater use is made of the data.
1229
1230
1231PEBS via Intel PT
1232-----------------
1233
1234Some hardware has the feature to redirect PEBS records to the Intel PT trace.
1235Recording is selected by using the aux-output config term e.g.
1236
1237 perf record -c 10000 -e '{intel_pt/branch=0/,cycles/aux-output/ppp}' uname
1238
1239Originally, software only supported redirecting at most one PEBS event because it
1240was not able to differentiate one event from another. To overcome that, more recent
1241kernels and perf tools add support for the PERF_RECORD_AUX_OUTPUT_HW_ID side-band event.
1242To check for the presence of that event in a PEBS-via-PT trace:
1243
1244 perf script -D --no-itrace | grep PERF_RECORD_AUX_OUTPUT_HW_ID
1245
1246To display PEBS events from the Intel PT trace, use the itrace 'o' option e.g.
1247
1248 perf script --itrace=oe
1249
1250XED
1251---
1252
1253include::build-xed.txt[]
1254
1255
1256Tracing Virtual Machines (kernel only)
1257--------------------------------------
1258
1259Currently, kernel tracing is supported with either "timeless" decoding
1260(i.e. no TSC timestamps) or VM Time Correlation. VM Time Correlation is an extra step
1261using 'perf inject' and requires unchanging VMX TSC Offset and no VMX TSC Scaling.
1262
1263Other limitations and caveats
1264
1265 VMX controls may suppress packets needed for decoding resulting in decoding errors
1266 VMX controls may block the perf NMI to the host potentially resulting in lost trace data
1267 Guest kernel self-modifying code (e.g. jump labels or JIT-compiled eBPF) will result in decoding errors
1268 Guest thread information is unknown
1269 Guest VCPU is unknown but may be able to be inferred from the host thread
1270 Callchains are not supported
1271
1272Example using "timeless" decoding
1273
1274Start VM
1275
1276 $ sudo virsh start kubuntu20.04
1277 Domain kubuntu20.04 started
1278
1279Mount the guest file system. Note sshfs needs -o direct_io to enable reading of proc files. root access is needed to read /proc/kcore.
1280
1281 $ mkdir vm0
1282 $ sshfs -o direct_io root@vm0:/ vm0
1283
1284Copy the guest /proc/kallsyms, /proc/modules and /proc/kcore
1285
1286 $ perf buildid-cache -v --kcore vm0/proc/kcore
1287 kcore added to build-id cache directory /home/user/.debug/[kernel.kcore]/9600f316a53a0f54278885e8d9710538ec5f6a08/2021021807494306
1288 $ KALLSYMS=/home/user/.debug/[kernel.kcore]/9600f316a53a0f54278885e8d9710538ec5f6a08/2021021807494306/kallsyms
1289
1290Find the VM process
1291
1292 $ ps -eLl | grep 'KVM\|PID'
1293 F S UID PID PPID LWP C PRI NI ADDR SZ WCHAN TTY TIME CMD
1294 3 S 64055 1430 1 1440 1 80 0 - 1921718 - ? 00:02:47 CPU 0/KVM
1295 3 S 64055 1430 1 1441 1 80 0 - 1921718 - ? 00:02:41 CPU 1/KVM
1296 3 S 64055 1430 1 1442 1 80 0 - 1921718 - ? 00:02:38 CPU 2/KVM
1297 3 S 64055 1430 1 1443 2 80 0 - 1921718 - ? 00:03:18 CPU 3/KVM
1298
1299Start an open-ended perf record, tracing the VM process, do something on the VM, and then ctrl-C to stop.
1300TSC is not supported and tsc=0 must be specified. That means mtc is useless, so add mtc=0.
1301However, IPC can still be determined, hence cyc=1 can be added.
1302Only kernel decoding is supported, so 'k' must be specified.
1303Intel PT traces both the host and the guest so --guest and --host need to be specified.
1304Without timestamps, --per-thread must be specified to distinguish threads.
1305
1306 $ sudo perf kvm --guest --host --guestkallsyms $KALLSYMS record --kcore -e intel_pt/tsc=0,mtc=0,cyc=1/k -p 1430 --per-thread
1307 ^C
1308 [ perf record: Woken up 1 times to write data ]
1309 [ perf record: Captured and wrote 5.829 MB ]
1310
1311perf script can be used to provide an instruction trace
1312
1313 $ perf script --guestkallsyms $KALLSYMS --insn-trace=disasm -F+ipc | grep -C10 vmresume | head -21
1314 CPU 0/KVM 1440 ffffffff82133cdd __vmx_vcpu_run+0x3d ([kernel.kallsyms]) movq 0x48(%rax), %r9
1315 CPU 0/KVM 1440 ffffffff82133ce1 __vmx_vcpu_run+0x41 ([kernel.kallsyms]) movq 0x50(%rax), %r10
1316 CPU 0/KVM 1440 ffffffff82133ce5 __vmx_vcpu_run+0x45 ([kernel.kallsyms]) movq 0x58(%rax), %r11
1317 CPU 0/KVM 1440 ffffffff82133ce9 __vmx_vcpu_run+0x49 ([kernel.kallsyms]) movq 0x60(%rax), %r12
1318 CPU 0/KVM 1440 ffffffff82133ced __vmx_vcpu_run+0x4d ([kernel.kallsyms]) movq 0x68(%rax), %r13
1319 CPU 0/KVM 1440 ffffffff82133cf1 __vmx_vcpu_run+0x51 ([kernel.kallsyms]) movq 0x70(%rax), %r14
1320 CPU 0/KVM 1440 ffffffff82133cf5 __vmx_vcpu_run+0x55 ([kernel.kallsyms]) movq 0x78(%rax), %r15
1321 CPU 0/KVM 1440 ffffffff82133cf9 __vmx_vcpu_run+0x59 ([kernel.kallsyms]) movq (%rax), %rax
1322 CPU 0/KVM 1440 ffffffff82133cfc __vmx_vcpu_run+0x5c ([kernel.kallsyms]) callq 0xffffffff82133c40
1323 CPU 0/KVM 1440 ffffffff82133c40 vmx_vmenter+0x0 ([kernel.kallsyms]) jz 0xffffffff82133c46
1324 CPU 0/KVM 1440 ffffffff82133c42 vmx_vmenter+0x2 ([kernel.kallsyms]) vmresume IPC: 0.11 (50/445)
1325 :1440 1440 ffffffffbb678b06 native_write_msr+0x6 ([guest.kernel.kallsyms]) nopl %eax, (%rax,%rax,1)
1326 :1440 1440 ffffffffbb678b0b native_write_msr+0xb ([guest.kernel.kallsyms]) retq IPC: 0.04 (2/41)
1327 :1440 1440 ffffffffbb666646 lapic_next_deadline+0x26 ([guest.kernel.kallsyms]) data16 nop
1328 :1440 1440 ffffffffbb666648 lapic_next_deadline+0x28 ([guest.kernel.kallsyms]) xor %eax, %eax
1329 :1440 1440 ffffffffbb66664a lapic_next_deadline+0x2a ([guest.kernel.kallsyms]) popq %rbp
1330 :1440 1440 ffffffffbb66664b lapic_next_deadline+0x2b ([guest.kernel.kallsyms]) retq IPC: 0.16 (4/25)
1331 :1440 1440 ffffffffbb74607f clockevents_program_event+0x8f ([guest.kernel.kallsyms]) test %eax, %eax
1332 :1440 1440 ffffffffbb746081 clockevents_program_event+0x91 ([guest.kernel.kallsyms]) jz 0xffffffffbb74603c IPC: 0.06 (2/30)
1333 :1440 1440 ffffffffbb74603c clockevents_program_event+0x4c ([guest.kernel.kallsyms]) popq %rbx
1334 :1440 1440 ffffffffbb74603d clockevents_program_event+0x4d ([guest.kernel.kallsyms]) popq %r12
1335
1336Example using VM Time Correlation
1337
1338Start VM
1339
1340 $ sudo virsh start kubuntu20.04
1341 Domain kubuntu20.04 started
1342
1343Mount the guest file system. Note sshfs needs -o direct_io to enable reading of proc files. root access is needed to read /proc/kcore.
1344
1345 $ mkdir -p vm0
1346 $ sshfs -o direct_io root@vm0:/ vm0
1347
1348Copy the guest /proc/kallsyms, /proc/modules and /proc/kcore
1349
1350 $ perf buildid-cache -v --kcore vm0/proc/kcore
1351 same kcore found in /home/user/.debug/[kernel.kcore]/cc9c55a98c5e4ec0aeda69302554aabed5cd6491/2021021312450777
1352 $ KALLSYMS=/home/user/.debug/\[kernel.kcore\]/cc9c55a98c5e4ec0aeda69302554aabed5cd6491/2021021312450777/kallsyms
1353
1354Find the VM process
1355
1356 $ ps -eLl | grep 'KVM\|PID'
1357 F S UID PID PPID LWP C PRI NI ADDR SZ WCHAN TTY TIME CMD
1358 3 S 64055 16998 1 17005 13 80 0 - 1818189 - ? 00:00:16 CPU 0/KVM
1359 3 S 64055 16998 1 17006 4 80 0 - 1818189 - ? 00:00:05 CPU 1/KVM
1360 3 S 64055 16998 1 17007 3 80 0 - 1818189 - ? 00:00:04 CPU 2/KVM
1361 3 S 64055 16998 1 17008 4 80 0 - 1818189 - ? 00:00:05 CPU 3/KVM
1362
1363Start an open-ended perf record, tracing the VM process, do something on the VM, and then ctrl-C to stop.
1364IPC can be determined, hence cyc=1 can be added.
1365Only kernel decoding is supported, so 'k' must be specified.
1366Intel PT traces both the host and the guest so --guest and --host need to be specified.
1367
1368 $ sudo perf kvm --guest --host --guestkallsyms $KALLSYMS record --kcore -e intel_pt/cyc=1/k -p 16998
1369 ^C[ perf record: Woken up 1 times to write data ]
1370 [ perf record: Captured and wrote 9.041 MB perf.data.kvm ]
1371
1372Now 'perf inject' can be used to determine the VMX TCS Offset. Note, Intel PT TSC packets are
1373only 7-bytes, so the TSC Offset might differ from the actual value in the 8th byte. That will
1374have no effect i.e. the resulting timestamps will be correct anyway.
1375
1376 $ perf inject -i perf.data.kvm --vm-time-correlation=dry-run
1377 ERROR: Unknown TSC Offset for VMCS 0x1bff6a
1378 VMCS: 0x1bff6a TSC Offset 0xffffe42722c64c41
1379 ERROR: Unknown TSC Offset for VMCS 0x1cbc08
1380 VMCS: 0x1cbc08 TSC Offset 0xffffe42722c64c41
1381 ERROR: Unknown TSC Offset for VMCS 0x1c3ce8
1382 VMCS: 0x1c3ce8 TSC Offset 0xffffe42722c64c41
1383 ERROR: Unknown TSC Offset for VMCS 0x1cbce9
1384 VMCS: 0x1cbce9 TSC Offset 0xffffe42722c64c41
1385
1386Each virtual CPU has a different Virtual Machine Control Structure (VMCS)
1387shown above with the calculated TSC Offset. For an unchanging TSC Offset
1388they should all be the same for the same virtual machine.
1389
1390Now that the TSC Offset is known, it can be provided to 'perf inject'
1391
1392 $ perf inject -i perf.data.kvm --vm-time-correlation="dry-run 0xffffe42722c64c41"
1393
1394Note the options for 'perf inject' --vm-time-correlation are:
1395
1396 [ dry-run ] [ <TSC Offset> [ : <VMCS> [ , <VMCS> ]... ] ]...
1397
1398So it is possible to specify different TSC Offsets for different VMCS.
1399The option "dry-run" will cause the file to be processed but without updating it.
1400Note it is also possible to get a intel_pt.log file by adding option --itrace=d
1401
1402There were no errors so, do it for real
1403
1404 $ perf inject -i perf.data.kvm --vm-time-correlation=0xffffe42722c64c41 --force
1405
1406'perf script' can be used to see if there are any decoder errors
1407
1408 $ perf script -i perf.data.kvm --guestkallsyms $KALLSYMS --itrace=e-o
1409
1410There were none.
1411
1412'perf script' can be used to provide an instruction trace showing timestamps
1413
1414 $ perf script -i perf.data.kvm --guestkallsyms $KALLSYMS --insn-trace=disasm -F+ipc | grep -C10 vmresume | head -21
1415 CPU 1/KVM 17006 [001] 11500.262865593: ffffffff82133cdd __vmx_vcpu_run+0x3d ([kernel.kallsyms]) movq 0x48(%rax), %r9
1416 CPU 1/KVM 17006 [001] 11500.262865593: ffffffff82133ce1 __vmx_vcpu_run+0x41 ([kernel.kallsyms]) movq 0x50(%rax), %r10
1417 CPU 1/KVM 17006 [001] 11500.262865593: ffffffff82133ce5 __vmx_vcpu_run+0x45 ([kernel.kallsyms]) movq 0x58(%rax), %r11
1418 CPU 1/KVM 17006 [001] 11500.262865593: ffffffff82133ce9 __vmx_vcpu_run+0x49 ([kernel.kallsyms]) movq 0x60(%rax), %r12
1419 CPU 1/KVM 17006 [001] 11500.262865593: ffffffff82133ced __vmx_vcpu_run+0x4d ([kernel.kallsyms]) movq 0x68(%rax), %r13
1420 CPU 1/KVM 17006 [001] 11500.262865593: ffffffff82133cf1 __vmx_vcpu_run+0x51 ([kernel.kallsyms]) movq 0x70(%rax), %r14
1421 CPU 1/KVM 17006 [001] 11500.262865593: ffffffff82133cf5 __vmx_vcpu_run+0x55 ([kernel.kallsyms]) movq 0x78(%rax), %r15
1422 CPU 1/KVM 17006 [001] 11500.262865593: ffffffff82133cf9 __vmx_vcpu_run+0x59 ([kernel.kallsyms]) movq (%rax), %rax
1423 CPU 1/KVM 17006 [001] 11500.262865593: ffffffff82133cfc __vmx_vcpu_run+0x5c ([kernel.kallsyms]) callq 0xffffffff82133c40
1424 CPU 1/KVM 17006 [001] 11500.262865593: ffffffff82133c40 vmx_vmenter+0x0 ([kernel.kallsyms]) jz 0xffffffff82133c46
1425 CPU 1/KVM 17006 [001] 11500.262866075: ffffffff82133c42 vmx_vmenter+0x2 ([kernel.kallsyms]) vmresume IPC: 0.05 (40/769)
1426 :17006 17006 [001] 11500.262869216: ffffffff82200cb0 asm_sysvec_apic_timer_interrupt+0x0 ([guest.kernel.kallsyms]) clac
1427 :17006 17006 [001] 11500.262869216: ffffffff82200cb3 asm_sysvec_apic_timer_interrupt+0x3 ([guest.kernel.kallsyms]) pushq $0xffffffffffffffff
1428 :17006 17006 [001] 11500.262869216: ffffffff82200cb5 asm_sysvec_apic_timer_interrupt+0x5 ([guest.kernel.kallsyms]) callq 0xffffffff82201160
1429 :17006 17006 [001] 11500.262869216: ffffffff82201160 error_entry+0x0 ([guest.kernel.kallsyms]) cld
1430 :17006 17006 [001] 11500.262869216: ffffffff82201161 error_entry+0x1 ([guest.kernel.kallsyms]) pushq %rsi
1431 :17006 17006 [001] 11500.262869216: ffffffff82201162 error_entry+0x2 ([guest.kernel.kallsyms]) movq 0x8(%rsp), %rsi
1432 :17006 17006 [001] 11500.262869216: ffffffff82201167 error_entry+0x7 ([guest.kernel.kallsyms]) movq %rdi, 0x8(%rsp)
1433 :17006 17006 [001] 11500.262869216: ffffffff8220116c error_entry+0xc ([guest.kernel.kallsyms]) pushq %rdx
1434 :17006 17006 [001] 11500.262869216: ffffffff8220116d error_entry+0xd ([guest.kernel.kallsyms]) pushq %rcx
1435 :17006 17006 [001] 11500.262869216: ffffffff8220116e error_entry+0xe ([guest.kernel.kallsyms]) pushq %rax
1436
1437
1438Tracing Virtual Machines (including user space)
1439-----------------------------------------------
1440
1441It is possible to use perf record to record sideband events within a virtual machine, so that an Intel PT trace on the host can be decoded.
1442Sideband events from the guest perf.data file can be injected into the host perf.data file using perf inject.
1443
1444Here is an example of the steps needed:
1445
1446On the guest machine:
1447
1448Check that no-kvmclock kernel command line option was used to boot:
1449
1450Note, this is essential to enable time correlation between host and guest machines.
1451
1452 $ cat /proc/cmdline
1453 BOOT_IMAGE=/boot/vmlinuz-5.10.0-16-amd64 root=UUID=cb49c910-e573-47e0-bce7-79e293df8e1d ro no-kvmclock
1454
1455There is no BPF support at present so, if possible, disable JIT compiling:
1456
1457 $ echo 0 | sudo tee /proc/sys/net/core/bpf_jit_enable
1458 0
1459
1460Start perf record to collect sideband events:
1461
1462 $ sudo perf record -o guest-sideband-testing-guest-perf.data --sample-identifier --buildid-all --switch-events --kcore -a -e dummy
1463
1464On the host machine:
1465
1466Start perf record to collect Intel PT trace:
1467
1468Note, the host trace will get very big, very fast, so the steps from starting to stopping the host trace really need to be done so that they happen in the shortest time possible.
1469
1470 $ sudo perf record -o guest-sideband-testing-host-perf.data -m,64M --kcore -a -e intel_pt/cyc/
1471
1472On the guest machine:
1473
1474Run a small test case, just 'uname' in this example:
1475
1476 $ uname
1477 Linux
1478
1479On the host machine:
1480
1481Stop the Intel PT trace:
1482
1483 ^C
1484 [ perf record: Woken up 1 times to write data ]
1485 [ perf record: Captured and wrote 76.122 MB guest-sideband-testing-host-perf.data ]
1486
1487On the guest machine:
1488
1489Stop the Intel PT trace:
1490
1491 ^C
1492 [ perf record: Woken up 1 times to write data ]
1493 [ perf record: Captured and wrote 1.247 MB guest-sideband-testing-guest-perf.data ]
1494
1495And then copy guest-sideband-testing-guest-perf.data to the host (not shown here).
1496
1497On the host machine:
1498
1499With the 2 perf.data recordings, and with their ownership changed to the user.
1500
1501Identify the TSC Offset:
1502
1503 $ perf inject -i guest-sideband-testing-host-perf.data --vm-time-correlation=dry-run
1504 VMCS: 0x103fc6 TSC Offset 0xfffffa6ae070cb20
1505 VMCS: 0x103ff2 TSC Offset 0xfffffa6ae070cb20
1506 VMCS: 0x10fdaa TSC Offset 0xfffffa6ae070cb20
1507 VMCS: 0x24d57c TSC Offset 0xfffffa6ae070cb20
1508
1509Correct Intel PT TSC timestamps for the guest machine:
1510
1511 $ perf inject -i guest-sideband-testing-host-perf.data --vm-time-correlation=0xfffffa6ae070cb20 --force
1512
1513Identify the guest machine PID:
1514
1515 $ perf script -i guest-sideband-testing-host-perf.data --no-itrace --show-task-events | grep KVM
1516 CPU 0/KVM 0 [000] 0.000000: PERF_RECORD_COMM: CPU 0/KVM:13376/13381
1517 CPU 1/KVM 0 [000] 0.000000: PERF_RECORD_COMM: CPU 1/KVM:13376/13382
1518 CPU 2/KVM 0 [000] 0.000000: PERF_RECORD_COMM: CPU 2/KVM:13376/13383
1519 CPU 3/KVM 0 [000] 0.000000: PERF_RECORD_COMM: CPU 3/KVM:13376/13384
1520
1521Note, the QEMU option -name debug-threads=on is needed so that thread names
1522can be used to determine which thread is running which VCPU as above. libvirt seems to use this by default.
1523
1524Create a guestmount, assuming the guest machine is 'vm_to_test':
1525
1526 $ mkdir -p ~/guestmount/13376
1527 $ sshfs -o direct_io vm_to_test:/ ~/guestmount/13376
1528
1529Inject the guest perf.data file into the host perf.data file:
1530
1531Note, due to the guestmount option, guest object files and debug files will be copied into the build ID cache from the guest machine, with the notable exception of VDSO.
1532If needed, VDSO can be copied manually in a fashion similar to that used by the perf-archive script.
1533
1534 $ perf inject -i guest-sideband-testing-host-perf.data -o inj --guestmount ~/guestmount --guest-data=guest-sideband-testing-guest-perf.data,13376,0xfffffa6ae070cb20
1535
1536Show an excerpt from the result. In this case the CPU and time range have been to chosen to show interaction between guest and host when 'uname' is starting to run on the guest machine:
1537
1538Notes:
1539
1540 - the CPU displayed, [002] in this case, is always the host CPU
1541 - events happening in the virtual machine start with VM:13376 VCPU:003, which shows the hypervisor PID 13376 and the VCPU number
1542 - only calls and errors are displayed i.e. --itrace=ce
1543 - branches entering and exiting the virtual machine are split, and show as 2 branches to/from "0 [unknown] ([unknown])"
1544
1545 $ perf script -i inj --itrace=ce -F+machine_pid,+vcpu,+addr,+pid,+tid,-period --ns --time 7919.408803365,7919.408804631 -C 2
1546 CPU 3/KVM 13376/13384 [002] 7919.408803365: branches: ffffffffc0f8ebe0 vmx_vcpu_enter_exit+0xc0 ([kernel.kallsyms]) => ffffffffc0f8edc0 __vmx_vcpu_run+0x0 ([kernel.kallsyms])
1547 CPU 3/KVM 13376/13384 [002] 7919.408803365: branches: ffffffffc0f8edd5 __vmx_vcpu_run+0x15 ([kernel.kallsyms]) => ffffffffc0f8eca0 vmx_update_host_rsp+0x0 ([kernel.kallsyms])
1548 CPU 3/KVM 13376/13384 [002] 7919.408803365: branches: ffffffffc0f8ee1b __vmx_vcpu_run+0x5b ([kernel.kallsyms]) => ffffffffc0f8ed60 vmx_vmenter+0x0 ([kernel.kallsyms])
1549 CPU 3/KVM 13376/13384 [002] 7919.408803461: branches: ffffffffc0f8ed62 vmx_vmenter+0x2 ([kernel.kallsyms]) => 0 [unknown] ([unknown])
1550 VM:13376 VCPU:003 uname 3404/3404 [002] 7919.408803461: branches: 0 [unknown] ([unknown]) => 7f851c9b5a5c init_cacheinfo+0x3ac (/usr/lib/x86_64-linux-gnu/libc-2.31.so)
1551 VM:13376 VCPU:003 uname 3404/3404 [002] 7919.408803567: branches: 7f851c9b5a5a init_cacheinfo+0x3aa (/usr/lib/x86_64-linux-gnu/libc-2.31.so) => 0 [unknown] ([unknown])
1552 CPU 3/KVM 13376/13384 [002] 7919.408803567: branches: 0 [unknown] ([unknown]) => ffffffffc0f8ed80 vmx_vmexit+0x0 ([kernel.kallsyms])
1553 CPU 3/KVM 13376/13384 [002] 7919.408803596: branches: ffffffffc0f6619a vmx_vcpu_run+0x26a ([kernel.kallsyms]) => ffffffffb2255c60 x86_virt_spec_ctrl+0x0 ([kernel.kallsyms])
1554 CPU 3/KVM 13376/13384 [002] 7919.408803801: branches: ffffffffc0f66445 vmx_vcpu_run+0x515 ([kernel.kallsyms]) => ffffffffb2290b30 native_write_msr+0x0 ([kernel.kallsyms])
1555 CPU 3/KVM 13376/13384 [002] 7919.408803850: branches: ffffffffc0f661f8 vmx_vcpu_run+0x2c8 ([kernel.kallsyms]) => ffffffffc1092300 kvm_load_host_xsave_state+0x0 ([kernel.kallsyms])
1556 CPU 3/KVM 13376/13384 [002] 7919.408803850: branches: ffffffffc1092327 kvm_load_host_xsave_state+0x27 ([kernel.kallsyms]) => ffffffffc1092220 kvm_load_host_xsave_state.part.0+0x0 ([kernel.kallsyms])
1557 CPU 3/KVM 13376/13384 [002] 7919.408803862: branches: ffffffffc0f662cf vmx_vcpu_run+0x39f ([kernel.kallsyms]) => ffffffffc0f63f90 vmx_recover_nmi_blocking+0x0 ([kernel.kallsyms])
1558 CPU 3/KVM 13376/13384 [002] 7919.408803862: branches: ffffffffc0f662e9 vmx_vcpu_run+0x3b9 ([kernel.kallsyms]) => ffffffffc0f619a0 __vmx_complete_interrupts+0x0 ([kernel.kallsyms])
1559 CPU 3/KVM 13376/13384 [002] 7919.408803872: branches: ffffffffc109cfb2 vcpu_enter_guest+0x752 ([kernel.kallsyms]) => ffffffffc0f5f570 vmx_handle_exit_irqoff+0x0 ([kernel.kallsyms])
1560 CPU 3/KVM 13376/13384 [002] 7919.408803881: branches: ffffffffc109d028 vcpu_enter_guest+0x7c8 ([kernel.kallsyms]) => ffffffffb234f900 __srcu_read_lock+0x0 ([kernel.kallsyms])
1561 CPU 3/KVM 13376/13384 [002] 7919.408803897: branches: ffffffffc109d06f vcpu_enter_guest+0x80f ([kernel.kallsyms]) => ffffffffc0f72e30 vmx_handle_exit+0x0 ([kernel.kallsyms])
1562 CPU 3/KVM 13376/13384 [002] 7919.408803897: branches: ffffffffc0f72e3d vmx_handle_exit+0xd ([kernel.kallsyms]) => ffffffffc0f727c0 __vmx_handle_exit+0x0 ([kernel.kallsyms])
1563 CPU 3/KVM 13376/13384 [002] 7919.408803897: branches: ffffffffc0f72b15 __vmx_handle_exit+0x355 ([kernel.kallsyms]) => ffffffffc0f60ae0 vmx_flush_pml_buffer+0x0 ([kernel.kallsyms])
1564 CPU 3/KVM 13376/13384 [002] 7919.408803903: branches: ffffffffc0f72994 __vmx_handle_exit+0x1d4 ([kernel.kallsyms]) => ffffffffc10b7090 kvm_emulate_cpuid+0x0 ([kernel.kallsyms])
1565 CPU 3/KVM 13376/13384 [002] 7919.408803903: branches: ffffffffc10b70f1 kvm_emulate_cpuid+0x61 ([kernel.kallsyms]) => ffffffffc10b6e10 kvm_cpuid+0x0 ([kernel.kallsyms])
1566 CPU 3/KVM 13376/13384 [002] 7919.408803941: branches: ffffffffc10b7125 kvm_emulate_cpuid+0x95 ([kernel.kallsyms]) => ffffffffc1093110 kvm_skip_emulated_instruction+0x0 ([kernel.kallsyms])
1567 CPU 3/KVM 13376/13384 [002] 7919.408803941: branches: ffffffffc109311f kvm_skip_emulated_instruction+0xf ([kernel.kallsyms]) => ffffffffc0f5e180 vmx_get_rflags+0x0 ([kernel.kallsyms])
1568 CPU 3/KVM 13376/13384 [002] 7919.408803951: branches: ffffffffc109312a kvm_skip_emulated_instruction+0x1a ([kernel.kallsyms]) => ffffffffc0f5fd30 vmx_skip_emulated_instruction+0x0 ([kernel.kallsyms])
1569 CPU 3/KVM 13376/13384 [002] 7919.408803951: branches: ffffffffc0f5fd79 vmx_skip_emulated_instruction+0x49 ([kernel.kallsyms]) => ffffffffc0f5fb50 skip_emulated_instruction+0x0 ([kernel.kallsyms])
1570 CPU 3/KVM 13376/13384 [002] 7919.408803956: branches: ffffffffc0f5fc68 skip_emulated_instruction+0x118 ([kernel.kallsyms]) => ffffffffc0f6a940 vmx_cache_reg+0x0 ([kernel.kallsyms])
1571 CPU 3/KVM 13376/13384 [002] 7919.408803964: branches: ffffffffc0f5fc11 skip_emulated_instruction+0xc1 ([kernel.kallsyms]) => ffffffffc0f5f9e0 vmx_set_interrupt_shadow+0x0 ([kernel.kallsyms])
1572 CPU 3/KVM 13376/13384 [002] 7919.408803980: branches: ffffffffc109f8b1 vcpu_run+0x71 ([kernel.kallsyms]) => ffffffffc10ad2f0 kvm_cpu_has_pending_timer+0x0 ([kernel.kallsyms])
1573 CPU 3/KVM 13376/13384 [002] 7919.408803980: branches: ffffffffc10ad2fb kvm_cpu_has_pending_timer+0xb ([kernel.kallsyms]) => ffffffffc10b0490 apic_has_pending_timer+0x0 ([kernel.kallsyms])
1574 CPU 3/KVM 13376/13384 [002] 7919.408803991: branches: ffffffffc109f899 vcpu_run+0x59 ([kernel.kallsyms]) => ffffffffc109c860 vcpu_enter_guest+0x0 ([kernel.kallsyms])
1575 CPU 3/KVM 13376/13384 [002] 7919.408803993: branches: ffffffffc109cd4c vcpu_enter_guest+0x4ec ([kernel.kallsyms]) => ffffffffc0f69140 vmx_prepare_switch_to_guest+0x0 ([kernel.kallsyms])
1576 CPU 3/KVM 13376/13384 [002] 7919.408803996: branches: ffffffffc109cd7d vcpu_enter_guest+0x51d ([kernel.kallsyms]) => ffffffffb234f930 __srcu_read_unlock+0x0 ([kernel.kallsyms])
1577 CPU 3/KVM 13376/13384 [002] 7919.408803996: branches: ffffffffc109cd9c vcpu_enter_guest+0x53c ([kernel.kallsyms]) => ffffffffc0f609b0 vmx_sync_pir_to_irr+0x0 ([kernel.kallsyms])
1578 CPU 3/KVM 13376/13384 [002] 7919.408803996: branches: ffffffffc0f60a6d vmx_sync_pir_to_irr+0xbd ([kernel.kallsyms]) => ffffffffc10adc20 kvm_lapic_find_highest_irr+0x0 ([kernel.kallsyms])
1579 CPU 3/KVM 13376/13384 [002] 7919.408804010: branches: ffffffffc0f60abd vmx_sync_pir_to_irr+0x10d ([kernel.kallsyms]) => ffffffffc0f60820 vmx_set_rvi+0x0 ([kernel.kallsyms])
1580 CPU 3/KVM 13376/13384 [002] 7919.408804019: branches: ffffffffc109ceca vcpu_enter_guest+0x66a ([kernel.kallsyms]) => ffffffffb2249840 fpregs_assert_state_consistent+0x0 ([kernel.kallsyms])
1581 CPU 3/KVM 13376/13384 [002] 7919.408804021: branches: ffffffffc109cf10 vcpu_enter_guest+0x6b0 ([kernel.kallsyms]) => ffffffffc0f65f30 vmx_vcpu_run+0x0 ([kernel.kallsyms])
1582 CPU 3/KVM 13376/13384 [002] 7919.408804024: branches: ffffffffc0f6603b vmx_vcpu_run+0x10b ([kernel.kallsyms]) => ffffffffb229bed0 __get_current_cr3_fast+0x0 ([kernel.kallsyms])
1583 CPU 3/KVM 13376/13384 [002] 7919.408804024: branches: ffffffffc0f66055 vmx_vcpu_run+0x125 ([kernel.kallsyms]) => ffffffffb2253050 cr4_read_shadow+0x0 ([kernel.kallsyms])
1584 CPU 3/KVM 13376/13384 [002] 7919.408804030: branches: ffffffffc0f6608d vmx_vcpu_run+0x15d ([kernel.kallsyms]) => ffffffffc10921e0 kvm_load_guest_xsave_state+0x0 ([kernel.kallsyms])
1585 CPU 3/KVM 13376/13384 [002] 7919.408804030: branches: ffffffffc1092207 kvm_load_guest_xsave_state+0x27 ([kernel.kallsyms]) => ffffffffc1092110 kvm_load_guest_xsave_state.part.0+0x0 ([kernel.kallsyms])
1586 CPU 3/KVM 13376/13384 [002] 7919.408804032: branches: ffffffffc0f660c6 vmx_vcpu_run+0x196 ([kernel.kallsyms]) => ffffffffb22061a0 perf_guest_get_msrs+0x0 ([kernel.kallsyms])
1587 CPU 3/KVM 13376/13384 [002] 7919.408804032: branches: ffffffffb22061a9 perf_guest_get_msrs+0x9 ([kernel.kallsyms]) => ffffffffb220cda0 intel_guest_get_msrs+0x0 ([kernel.kallsyms])
1588 CPU 3/KVM 13376/13384 [002] 7919.408804039: branches: ffffffffc0f66109 vmx_vcpu_run+0x1d9 ([kernel.kallsyms]) => ffffffffc0f652c0 clear_atomic_switch_msr+0x0 ([kernel.kallsyms])
1589 CPU 3/KVM 13376/13384 [002] 7919.408804040: branches: ffffffffc0f66119 vmx_vcpu_run+0x1e9 ([kernel.kallsyms]) => ffffffffc0f73f60 intel_pmu_lbr_is_enabled+0x0 ([kernel.kallsyms])
1590 CPU 3/KVM 13376/13384 [002] 7919.408804042: branches: ffffffffc0f73f81 intel_pmu_lbr_is_enabled+0x21 ([kernel.kallsyms]) => ffffffffc10b68e0 kvm_find_cpuid_entry+0x0 ([kernel.kallsyms])
1591 CPU 3/KVM 13376/13384 [002] 7919.408804045: branches: ffffffffc0f66454 vmx_vcpu_run+0x524 ([kernel.kallsyms]) => ffffffffc0f61ff0 vmx_update_hv_timer+0x0 ([kernel.kallsyms])
1592 CPU 3/KVM 13376/13384 [002] 7919.408804057: branches: ffffffffc0f66142 vmx_vcpu_run+0x212 ([kernel.kallsyms]) => ffffffffc10af100 kvm_wait_lapic_expire+0x0 ([kernel.kallsyms])
1593 CPU 3/KVM 13376/13384 [002] 7919.408804057: branches: ffffffffc0f66156 vmx_vcpu_run+0x226 ([kernel.kallsyms]) => ffffffffb2255c60 x86_virt_spec_ctrl+0x0 ([kernel.kallsyms])
1594 CPU 3/KVM 13376/13384 [002] 7919.408804057: branches: ffffffffc0f66161 vmx_vcpu_run+0x231 ([kernel.kallsyms]) => ffffffffc0f8eb20 vmx_vcpu_enter_exit+0x0 ([kernel.kallsyms])
1595 CPU 3/KVM 13376/13384 [002] 7919.408804057: branches: ffffffffc0f8eb44 vmx_vcpu_enter_exit+0x24 ([kernel.kallsyms]) => ffffffffb2353e10 rcu_note_context_switch+0x0 ([kernel.kallsyms])
1596 CPU 3/KVM 13376/13384 [002] 7919.408804057: branches: ffffffffb2353e1c rcu_note_context_switch+0xc ([kernel.kallsyms]) => ffffffffb2353db0 rcu_qs+0x0 ([kernel.kallsyms])
1597 CPU 3/KVM 13376/13384 [002] 7919.408804066: branches: ffffffffc0f8ebe0 vmx_vcpu_enter_exit+0xc0 ([kernel.kallsyms]) => ffffffffc0f8edc0 __vmx_vcpu_run+0x0 ([kernel.kallsyms])
1598 CPU 3/KVM 13376/13384 [002] 7919.408804066: branches: ffffffffc0f8edd5 __vmx_vcpu_run+0x15 ([kernel.kallsyms]) => ffffffffc0f8eca0 vmx_update_host_rsp+0x0 ([kernel.kallsyms])
1599 CPU 3/KVM 13376/13384 [002] 7919.408804066: branches: ffffffffc0f8ee1b __vmx_vcpu_run+0x5b ([kernel.kallsyms]) => ffffffffc0f8ed60 vmx_vmenter+0x0 ([kernel.kallsyms])
1600 CPU 3/KVM 13376/13384 [002] 7919.408804162: branches: ffffffffc0f8ed62 vmx_vmenter+0x2 ([kernel.kallsyms]) => 0 [unknown] ([unknown])
1601 VM:13376 VCPU:003 uname 3404/3404 [002] 7919.408804162: branches: 0 [unknown] ([unknown]) => 7f851c9b5a5c init_cacheinfo+0x3ac (/usr/lib/x86_64-linux-gnu/libc-2.31.so)
1602 VM:13376 VCPU:003 uname 3404/3404 [002] 7919.408804273: branches: 7f851cb7c0e4 _dl_init+0x74 (/usr/lib/x86_64-linux-gnu/ld-2.31.so) => 7f851cb7bf50 call_init.part.0+0x0 (/usr/lib/x86_64-linux-gnu/ld-2.31.so)
1603 VM:13376 VCPU:003 uname 3404/3404 [002] 7919.408804526: branches: 55e0c00136f0 _start+0x0 (/usr/bin/uname) => ffffffff83200ac0 asm_exc_page_fault+0x0 ([kernel.kallsyms])
1604 VM:13376 VCPU:003 uname 3404/3404 [002] 7919.408804526: branches: ffffffff83200ac3 asm_exc_page_fault+0x3 ([kernel.kallsyms]) => ffffffff83201290 error_entry+0x0 ([kernel.kallsyms])
1605 VM:13376 VCPU:003 uname 3404/3404 [002] 7919.408804534: branches: ffffffff832012fa error_entry+0x6a ([kernel.kallsyms]) => ffffffff830b59a0 sync_regs+0x0 ([kernel.kallsyms])
1606 VM:13376 VCPU:003 uname 3404/3404 [002] 7919.408804631: branches: ffffffff83200ad9 asm_exc_page_fault+0x19 ([kernel.kallsyms]) => ffffffff830b8210 exc_page_fault+0x0 ([kernel.kallsyms])
1607 VM:13376 VCPU:003 uname 3404/3404 [002] 7919.408804631: branches: ffffffff830b82a4 exc_page_fault+0x94 ([kernel.kallsyms]) => ffffffff830b80e0 __kvm_handle_async_pf+0x0 ([kernel.kallsyms])
1608 VM:13376 VCPU:003 uname 3404/3404 [002] 7919.408804631: branches: ffffffff830b80ed __kvm_handle_async_pf+0xd ([kernel.kallsyms]) => ffffffff830b80c0 kvm_read_and_reset_apf_flags+0x0 ([kernel.kallsyms])
1609
1610
1611Tracing Virtual Machines - Guest Code
1612-------------------------------------
1613
1614A common case for KVM test programs is that the test program acts as the
1615hypervisor, creating, running and destroying the virtual machine, and
1616providing the guest object code from its own object code. In this case,
1617the VM is not running an OS, but only the functions loaded into it by the
1618hypervisor test program, and conveniently, loaded at the same virtual
1619addresses. To support that, option "--guest-code" has been added to perf script
1620and perf kvm report.
1621
1622Here is an example tracing a test program from the kernel's KVM selftests:
1623
1624 # perf record --kcore -e intel_pt/cyc/ -- tools/testing/selftests/kselftest_install/kvm/tsc_msrs_test
1625 [ perf record: Woken up 1 times to write data ]
1626 [ perf record: Captured and wrote 0.280 MB perf.data ]
1627 # perf script --guest-code --itrace=bep --ns -F-period,+addr,+flags
1628 [SNIP]
1629 tsc_msrs_test 18436 [007] 10897.962087733: branches: call ffffffffc13b2ff5 __vmx_vcpu_run+0x15 (vmlinux) => ffffffffc13b2f50 vmx_update_host_rsp+0x0 (vmlinux)
1630 tsc_msrs_test 18436 [007] 10897.962087733: branches: return ffffffffc13b2f5d vmx_update_host_rsp+0xd (vmlinux) => ffffffffc13b2ffa __vmx_vcpu_run+0x1a (vmlinux)
1631 tsc_msrs_test 18436 [007] 10897.962087733: branches: call ffffffffc13b303b __vmx_vcpu_run+0x5b (vmlinux) => ffffffffc13b2f80 vmx_vmenter+0x0 (vmlinux)
1632 tsc_msrs_test 18436 [007] 10897.962087836: branches: vmentry ffffffffc13b2f82 vmx_vmenter+0x2 (vmlinux) => 0 [unknown] ([unknown])
1633 [guest/18436] 18436 [007] 10897.962087836: branches: vmentry 0 [unknown] ([unknown]) => 402c81 guest_code+0x131 (/home/user/git/work/tools/testing/selftests/kselftest_install/kvm/tsc_msrs_test)
1634 [guest/18436] 18436 [007] 10897.962087836: branches: call 402c81 guest_code+0x131 (/home/user/git/work/tools/testing/selftests/kselftest_install/kvm/tsc_msrs_test) => 40dba0 ucall+0x0 (/home/user/git/work/tools/testing/selftests/kselftest_install/kvm/tsc_msrs_test)
1635 [guest/18436] 18436 [007] 10897.962088248: branches: vmexit 40dba0 ucall+0x0 (/home/user/git/work/tools/testing/selftests/kselftest_install/kvm/tsc_msrs_test) => 0 [unknown] ([unknown])
1636 tsc_msrs_test 18436 [007] 10897.962088248: branches: vmexit 0 [unknown] ([unknown]) => ffffffffc13b2fa0 vmx_vmexit+0x0 (vmlinux)
1637 tsc_msrs_test 18436 [007] 10897.962088248: branches: jmp ffffffffc13b2fa0 vmx_vmexit+0x0 (vmlinux) => ffffffffc13b2fd2 vmx_vmexit+0x32 (vmlinux)
1638 tsc_msrs_test 18436 [007] 10897.962088256: branches: return ffffffffc13b2fd2 vmx_vmexit+0x32 (vmlinux) => ffffffffc13b3040 __vmx_vcpu_run+0x60 (vmlinux)
1639 tsc_msrs_test 18436 [007] 10897.962088270: branches: return ffffffffc13b30b6 __vmx_vcpu_run+0xd6 (vmlinux) => ffffffffc13b2f2e vmx_vcpu_enter_exit+0x4e (vmlinux)
1640 [SNIP]
1641 tsc_msrs_test 18436 [007] 10897.962089321: branches: call ffffffffc13b2ff5 __vmx_vcpu_run+0x15 (vmlinux) => ffffffffc13b2f50 vmx_update_host_rsp+0x0 (vmlinux)
1642 tsc_msrs_test 18436 [007] 10897.962089321: branches: return ffffffffc13b2f5d vmx_update_host_rsp+0xd (vmlinux) => ffffffffc13b2ffa __vmx_vcpu_run+0x1a (vmlinux)
1643 tsc_msrs_test 18436 [007] 10897.962089321: branches: call ffffffffc13b303b __vmx_vcpu_run+0x5b (vmlinux) => ffffffffc13b2f80 vmx_vmenter+0x0 (vmlinux)
1644 tsc_msrs_test 18436 [007] 10897.962089424: branches: vmentry ffffffffc13b2f82 vmx_vmenter+0x2 (vmlinux) => 0 [unknown] ([unknown])
1645 [guest/18436] 18436 [007] 10897.962089424: branches: vmentry 0 [unknown] ([unknown]) => 40dba0 ucall+0x0 (/home/user/git/work/tools/testing/selftests/kselftest_install/kvm/tsc_msrs_test)
1646 [guest/18436] 18436 [007] 10897.962089701: branches: jmp 40dc1b ucall+0x7b (/home/user/git/work/tools/testing/selftests/kselftest_install/kvm/tsc_msrs_test) => 40dc39 ucall+0x99 (/home/user/git/work/tools/testing/selftests/kselftest_install/kvm/tsc_msrs_test)
1647 [guest/18436] 18436 [007] 10897.962089701: branches: jcc 40dc3c ucall+0x9c (/home/user/git/work/tools/testing/selftests/kselftest_install/kvm/tsc_msrs_test) => 40dc20 ucall+0x80 (/home/user/git/work/tools/testing/selftests/kselftest_install/kvm/tsc_msrs_test)
1648 [guest/18436] 18436 [007] 10897.962089701: branches: jcc 40dc3c ucall+0x9c (/home/user/git/work/tools/testing/selftests/kselftest_install/kvm/tsc_msrs_test) => 40dc20 ucall+0x80 (/home/user/git/work/tools/testing/selftests/kselftest_install/kvm/tsc_msrs_test)
1649 [guest/18436] 18436 [007] 10897.962089701: branches: jcc 40dc37 ucall+0x97 (/home/user/git/work/tools/testing/selftests/kselftest_install/kvm/tsc_msrs_test) => 40dc50 ucall+0xb0 (/home/user/git/work/tools/testing/selftests/kselftest_install/kvm/tsc_msrs_test)
1650 [guest/18436] 18436 [007] 10897.962089878: branches: vmexit 40dc55 ucall+0xb5 (/home/user/git/work/tools/testing/selftests/kselftest_install/kvm/tsc_msrs_test) => 0 [unknown] ([unknown])
1651 tsc_msrs_test 18436 [007] 10897.962089878: branches: vmexit 0 [unknown] ([unknown]) => ffffffffc13b2fa0 vmx_vmexit+0x0 (vmlinux)
1652 tsc_msrs_test 18436 [007] 10897.962089878: branches: jmp ffffffffc13b2fa0 vmx_vmexit+0x0 (vmlinux) => ffffffffc13b2fd2 vmx_vmexit+0x32 (vmlinux)
1653 tsc_msrs_test 18436 [007] 10897.962089887: branches: return ffffffffc13b2fd2 vmx_vmexit+0x32 (vmlinux) => ffffffffc13b3040 __vmx_vcpu_run+0x60 (vmlinux)
1654 tsc_msrs_test 18436 [007] 10897.962089901: branches: return ffffffffc13b30b6 __vmx_vcpu_run+0xd6 (vmlinux) => ffffffffc13b2f2e vmx_vcpu_enter_exit+0x4e (vmlinux)
1655 [SNIP]
1656
1657 # perf kvm --guest-code --guest --host report -i perf.data --stdio | head -20
1658
1659 # To display the perf.data header info, please use --header/--header-only options.
1660 #
1661 #
1662 # Total Lost Samples: 0
1663 #
1664 # Samples: 12 of event 'instructions'
1665 # Event count (approx.): 2274583
1666 #
1667 # Children Self Command Shared Object Symbol
1668 # ........ ........ ............. .................... ...........................................
1669 #
1670 54.70% 0.00% tsc_msrs_test [kernel.vmlinux] [k] entry_SYSCALL_64_after_hwframe
1671 |
1672 ---entry_SYSCALL_64_after_hwframe
1673 do_syscall_64
1674 |
1675 |--29.44%--syscall_exit_to_user_mode
1676 | exit_to_user_mode_prepare
1677 | task_work_run
1678 | __fput
1679
1680
1681Event Trace
1682-----------
1683
1684Event Trace records information about asynchronous events, for example interrupts,
1685faults, VM exits and entries. The information is recorded in CFE and EVD packets,
1686and also the Interrupt Flag is recorded on the MODE.Exec packet. The CFE packet
1687contains a type field to identify one of the following:
1688
1689 1 INTR interrupt, fault, exception, NMI
1690 2 IRET interrupt return
1691 3 SMI system management interrupt
1692 4 RSM resume from system management mode
1693 5 SIPI startup interprocessor interrupt
1694 6 INIT INIT signal
1695 7 VMENTRY VM-Entry
1696 8 VMEXIT VM-Entry
1697 9 VMEXIT_INTR VM-Exit due to interrupt
1698 10 SHUTDOWN Shutdown
1699
1700For more details, refer to the Intel 64 and IA-32 Architectures Software
1701Developer Manuals (version 076 or later).
1702
1703The capability to do Event Trace is indicated by the
1704/sys/bus/event_source/devices/intel_pt/caps/event_trace file.
1705
1706Event trace is selected for recording using the "event" config term. e.g.
1707
1708 perf record -e intel_pt/event/u uname
1709
1710Event trace events are output using the --itrace I option. e.g.
1711
1712 perf script --itrace=Ie
1713
1714perf script displays events containing CFE type, vector and event data,
1715in the form:
1716
1717 evt: hw int (t) cfe: INTR IP: 1 vector: 3 PFA: 0x8877665544332211
1718
1719The IP flag indicates if the event binds to an IP, which includes any case where
1720flow control packet generation is enabled, as well as when CFE packet IP bit is
1721set.
1722
1723perf script displays events containing changes to the Interrupt Flag in the form:
1724
1725 iflag: t IFLAG: 1->0 via branch
1726
1727where "via branch" indicates a branch (interrupt or return from interrupt) and
1728"non branch" indicates an instruction such as CFI, STI or POPF).
1729
1730In addition, the current state of the interrupt flag is indicated by the presence
1731or absence of the "D" (interrupt disabled) perf script flag. If the interrupt
1732flag is changed, then the "t" flag is also included i.e.
1733
1734 no flag, interrupts enabled IF=1
1735 t interrupts become disabled IF=1 -> IF=0
1736 D interrupts are disabled IF=0
1737 Dt interrupts become enabled IF=0 -> IF=1
1738
1739The intel-pt-events.py script illustrates how to access Event Trace information
1740using a Python script.
1741
1742
1743TNT Disable
1744-----------
1745
1746TNT packets are disabled using the "notnt" config term. e.g.
1747
1748 perf record -e intel_pt/notnt/u uname
1749
1750In that case the --itrace q option is forced because walking executable code
1751to reconstruct the control flow is not possible.
1752
1753
1754Emulated PTWRITE
1755----------------
1756
1757Later perf tools support a method to emulate the ptwrite instruction, which
1758can be useful if hardware does not support the ptwrite instruction.
1759
1760Instead of using the ptwrite instruction, a function is used which produces
1761a trace that encodes the payload data into TNT packets. Here is an example
1762of the function:
1763
1764 #include <stdint.h>
1765
1766 void perf_emulate_ptwrite(uint64_t x)
1767 __attribute__((externally_visible, noipa, no_instrument_function, naked));
1768
1769 #define PERF_EMULATE_PTWRITE_8_BITS \
1770 "1: shl %rax\n" \
1771 " jc 1f\n" \
1772 "1: shl %rax\n" \
1773 " jc 1f\n" \
1774 "1: shl %rax\n" \
1775 " jc 1f\n" \
1776 "1: shl %rax\n" \
1777 " jc 1f\n" \
1778 "1: shl %rax\n" \
1779 " jc 1f\n" \
1780 "1: shl %rax\n" \
1781 " jc 1f\n" \
1782 "1: shl %rax\n" \
1783 " jc 1f\n" \
1784 "1: shl %rax\n" \
1785 " jc 1f\n"
1786
1787 /* Undefined instruction */
1788 #define PERF_EMULATE_PTWRITE_UD2 ".byte 0x0f, 0x0b\n"
1789
1790 #define PERF_EMULATE_PTWRITE_MAGIC PERF_EMULATE_PTWRITE_UD2 ".ascii \"perf,ptwrite \"\n"
1791
1792 void perf_emulate_ptwrite(uint64_t x __attribute__ ((__unused__)))
1793 {
1794 /* Assumes SysV ABI : x passed in rdi */
1795 __asm__ volatile (
1796 "jmp 1f\n"
1797 PERF_EMULATE_PTWRITE_MAGIC
1798 "1: mov %rdi, %rax\n"
1799 PERF_EMULATE_PTWRITE_8_BITS
1800 PERF_EMULATE_PTWRITE_8_BITS
1801 PERF_EMULATE_PTWRITE_8_BITS
1802 PERF_EMULATE_PTWRITE_8_BITS
1803 PERF_EMULATE_PTWRITE_8_BITS
1804 PERF_EMULATE_PTWRITE_8_BITS
1805 PERF_EMULATE_PTWRITE_8_BITS
1806 PERF_EMULATE_PTWRITE_8_BITS
1807 "1: ret\n"
1808 );
1809 }
1810
1811For example, a test program with the function above:
1812
1813 #include <stdio.h>
1814 #include <stdint.h>
1815 #include <stdlib.h>
1816
1817 #include "perf_emulate_ptwrite.h"
1818
1819 int main(int argc, char *argv[])
1820 {
1821 uint64_t x = 0;
1822
1823 if (argc > 1)
1824 x = strtoull(argv[1], NULL, 0);
1825 perf_emulate_ptwrite(x);
1826 return 0;
1827 }
1828
1829Can be compiled and traced:
1830
1831 $ gcc -Wall -Wextra -O3 -g -o eg_ptw eg_ptw.c
1832 $ perf record -e intel_pt//u ./eg_ptw 0x1234567890abcdef
1833 [ perf record: Woken up 1 times to write data ]
1834 [ perf record: Captured and wrote 0.017 MB perf.data ]
1835 $ perf script --itrace=ew
1836 eg_ptw 19875 [007] 8061.235912: ptwrite: IP: 0 payload: 0x1234567890abcdef 55701249a196 perf_emulate_ptwrite+0x16 (/home/user/eg_ptw)
1837 $
1838
1839
1840Pipe mode
1841---------
1842Pipe mode is a problem for Intel PT and possibly other auxtrace users.
1843It's not recommended to use a pipe as data output with Intel PT because
1844of the following reason.
1845
1846Essentially the auxtrace buffers do not behave like the regular perf
1847event buffers. That is because the head and tail are updated by
1848software, but in the auxtrace case the data is written by hardware.
1849So the head and tail do not get updated as data is written.
1850
1851In the Intel PT case, the head and tail are updated only when the trace
1852is disabled by software, for example:
1853 - full-trace, system wide : when buffer passes watermark
1854 - full-trace, not system-wide : when buffer passes watermark or
1855 context switches
1856 - snapshot mode : as above but also when a snapshot is made
1857 - sample mode : as above but also when a sample is made
1858
1859That means finished-round ordering doesn't work. An auxtrace buffer
1860can turn up that has data that extends back in time, possibly to the
1861very beginning of tracing.
1862
1863For a perf.data file, that problem is solved by going through the trace
1864and queuing up the auxtrace buffers in advance.
1865
1866For pipe mode, the order of events and timestamps can presumably
1867be messed up.
1868
1869
1870EXAMPLE
1871-------
1872
1873Examples can be found on perf wiki page "Perf tools support for Intel® Processor Trace":
1874
1875https://perf.wiki.kernel.org/index.php/Perf_tools_support_for_Intel%C2%AE_Processor_Trace
1876
1877
1878SEE ALSO
1879--------
1880
1881linkperf:perf-record[1], linkperf:perf-script[1], linkperf:perf-report[1],
1882linkperf:perf-inject[1]