1 Documentation for support for Intel Processor !! 1 Intel Processor Trace
>> 2 =====================
>> 3
>> 4 Overview
>> 5 ========
>> 6
>> 7 Intel Processor Trace (Intel PT) is an extension of Intel Architecture that
>> 8 collects information about software execution such as control flow, execution
>> 9 modes and timings and formats it into highly compressed binary packets.
>> 10 Technical details are documented in the Intel 64 and IA-32 Architectures
>> 11 Software Developer Manuals, Chapter 36 Intel Processor Trace.
>> 12
>> 13 Intel PT is first supported in Intel Core M and 5th generation Intel Core
>> 14 processors that are based on the Intel micro-architecture code name Broadwell.
>> 15
>> 16 Trace data is collected by 'perf record' and stored within the perf.data file.
>> 17 See below for options to 'perf record'.
>> 18
>> 19 Trace data must be 'decoded' which involves walking the object code and matching
>> 20 the trace data packets. For example a TNT packet only tells whether a
>> 21 conditional branch was taken or not taken, so to make use of that packet the
>> 22 decoder must know precisely which instruction was being executed.
>> 23
>> 24 Decoding is done on-the-fly. The decoder outputs samples in the same format as
>> 25 samples output by perf hardware events, for example as though the "instructions"
>> 26 or "branches" events had been recorded. Presently 3 tools support this:
>> 27 'perf script', 'perf report' and 'perf inject'. See below for more information
>> 28 on using those tools.
>> 29
>> 30 The main distinguishing feature of Intel PT is that the decoder can determine
>> 31 the exact flow of software execution. Intel PT can be used to understand why
>> 32 and how did software get to a certain point, or behave a certain way. The
>> 33 software does not have to be recompiled, so Intel PT works with debug or release
>> 34 builds, however the executed images are needed - which makes use in JIT-compiled
>> 35 environments, or with self-modified code, a challenge. Also symbols need to be
>> 36 provided to make sense of addresses.
>> 37
>> 38 A limitation of Intel PT is that it produces huge amounts of trace data
>> 39 (hundreds of megabytes per second per core) which takes a long time to decode,
>> 40 for example two or three orders of magnitude longer than it took to collect.
>> 41 Another limitation is the performance impact of tracing, something that will
>> 42 vary depending on the use-case and architecture.
>> 43
>> 44
>> 45 Quickstart
>> 46 ==========
>> 47
>> 48 It is important to start small. That is because it is easy to capture vastly
>> 49 more data than can possibly be processed.
>> 50
>> 51 The simplest thing to do with Intel PT is userspace profiling of small programs.
>> 52 Data is captured with 'perf record' e.g. to trace 'ls' userspace-only:
>> 53
>> 54 perf record -e intel_pt//u ls
>> 55
>> 56 And profiled with 'perf report' e.g.
>> 57
>> 58 perf report
>> 59
>> 60 To also trace kernel space presents a problem, namely kernel self-modifying
>> 61 code. A fairly good kernel image is available in /proc/kcore but to get an
>> 62 accurate image a copy of /proc/kcore needs to be made under the same conditions
>> 63 as the data capture. A script perf-with-kcore can do that, but beware that the
>> 64 script makes use of 'sudo' to copy /proc/kcore. If you have perf installed
>> 65 locally from the source tree you can do:
>> 66
>> 67 ~/libexec/perf-core/perf-with-kcore record pt_ls -e intel_pt// -- ls
>> 68
>> 69 which will create a directory named 'pt_ls' and put the perf.data file and
>> 70 copies of /proc/kcore, /proc/kallsyms and /proc/modules into it. Then to use
>> 71 'perf report' becomes:
>> 72
>> 73 ~/libexec/perf-core/perf-with-kcore report pt_ls
>> 74
>> 75 Because samples are synthesized after-the-fact, the sampling period can be
>> 76 selected for reporting. e.g. sample every microsecond
>> 77
>> 78 ~/libexec/perf-core/perf-with-kcore report pt_ls --itrace=i1usge
>> 79
>> 80 See the sections below for more information about the --itrace option.
>> 81
>> 82 Beware the smaller the period, the more samples that are produced, and the
>> 83 longer it takes to process them.
>> 84
>> 85 Also note that the coarseness of Intel PT timing information will start to
>> 86 distort the statistical value of the sampling as the sampling period becomes
>> 87 smaller.
>> 88
>> 89 To represent software control flow, "branches" samples are produced. By default
>> 90 a branch sample is synthesized for every single branch. To get an idea what
>> 91 data is available you can use the 'perf script' tool with no parameters, which
>> 92 will list all the samples.
>> 93
>> 94 perf record -e intel_pt//u ls
>> 95 perf script
>> 96
>> 97 An interesting field that is not printed by default is 'flags' which can be
>> 98 displayed as follows:
>> 99
>> 100 perf script -Fcomm,tid,pid,time,cpu,event,trace,ip,sym,dso,addr,symoff,flags
>> 101
>> 102 The flags are "bcrosyiABEx" which stand for branch, call, return, conditional,
>> 103 system, asynchronous, interrupt, transaction abort, trace begin, trace end, and
>> 104 in transaction, respectively.
>> 105
>> 106 While it is possible to create scripts to analyze the data, an alternative
>> 107 approach is available to export the data to a postgresql database. Refer to
>> 108 script export-to-postgresql.py for more details, and to script
>> 109 call-graph-from-postgresql.py for an example of using the database.
>> 110
>> 111 There is also script intel-pt-events.py which provides an example of how to
>> 112 unpack the raw data for power events and PTWRITE.
>> 113
>> 114 As mentioned above, it is easy to capture too much data. One way to limit the
>> 115 data captured is to use 'snapshot' mode which is explained further below.
>> 116 Refer to 'new snapshot option' and 'Intel PT modes of operation' further below.
>> 117
>> 118 Another problem that will be experienced is decoder errors. They can be caused
>> 119 by inability to access the executed image, self-modified or JIT-ed code, or the
>> 120 inability to match side-band information (such as context switches and mmaps)
>> 121 which results in the decoder not knowing what code was executed.
>> 122
>> 123 There is also the problem of perf not being able to copy the data fast enough,
>> 124 resulting in data lost because the buffer was full. See 'Buffer handling' below
>> 125 for more details.
>> 126
>> 127
>> 128 perf record
>> 129 ===========
>> 130
>> 131 new event
>> 132 ---------
>> 133
>> 134 The Intel PT kernel driver creates a new PMU for Intel PT. PMU events are
>> 135 selected by providing the PMU name followed by the "config" separated by slashes.
>> 136 An enhancement has been made to allow default "config" e.g. the option
>> 137
>> 138 -e intel_pt//
>> 139
>> 140 will use a default config value. Currently that is the same as
>> 141
>> 142 -e intel_pt/tsc,noretcomp=0/
>> 143
>> 144 which is the same as
>> 145
>> 146 -e intel_pt/tsc=1,noretcomp=0/
>> 147
>> 148 Note there are now new config terms - see section 'config terms' further below.
>> 149
>> 150 The config terms are listed in /sys/devices/intel_pt/format. They are bit
>> 151 fields within the config member of the struct perf_event_attr which is
>> 152 passed to the kernel by the perf_event_open system call. They correspond to bit
>> 153 fields in the IA32_RTIT_CTL MSR. Here is a list of them and their definitions:
>> 154
>> 155 $ grep -H . /sys/bus/event_source/devices/intel_pt/format/*
>> 156 /sys/bus/event_source/devices/intel_pt/format/cyc:config:1
>> 157 /sys/bus/event_source/devices/intel_pt/format/cyc_thresh:config:19-22
>> 158 /sys/bus/event_source/devices/intel_pt/format/mtc:config:9
>> 159 /sys/bus/event_source/devices/intel_pt/format/mtc_period:config:14-17
>> 160 /sys/bus/event_source/devices/intel_pt/format/noretcomp:config:11
>> 161 /sys/bus/event_source/devices/intel_pt/format/psb_period:config:24-27
>> 162 /sys/bus/event_source/devices/intel_pt/format/tsc:config:10
>> 163
>> 164 Note that the default config must be overridden for each term i.e.
>> 165
>> 166 -e intel_pt/noretcomp=0/
>> 167
>> 168 is the same as:
>> 169
>> 170 -e intel_pt/tsc=1,noretcomp=0/
>> 171
>> 172 So, to disable TSC packets use:
>> 173
>> 174 -e intel_pt/tsc=0/
>> 175
>> 176 It is also possible to specify the config value explicitly:
>> 177
>> 178 -e intel_pt/config=0x400/
>> 179
>> 180 Note that, as with all events, the event is suffixed with event modifiers:
>> 181
>> 182 u userspace
>> 183 k kernel
>> 184 h hypervisor
>> 185 G guest
>> 186 H host
>> 187 p precise ip
>> 188
>> 189 'h', 'G' and 'H' are for virtualization which is not supported by Intel PT.
>> 190 'p' is also not relevant to Intel PT. So only options 'u' and 'k' are
>> 191 meaningful for Intel PT.
>> 192
>> 193 perf_event_attr is displayed if the -vv option is used e.g.
>> 194
>> 195 ------------------------------------------------------------
>> 196 perf_event_attr:
>> 197 type 6
>> 198 size 112
>> 199 config 0x400
>> 200 { sample_period, sample_freq } 1
>> 201 sample_type IP|TID|TIME|CPU|IDENTIFIER
>> 202 read_format ID
>> 203 disabled 1
>> 204 inherit 1
>> 205 exclude_kernel 1
>> 206 exclude_hv 1
>> 207 enable_on_exec 1
>> 208 sample_id_all 1
>> 209 ------------------------------------------------------------
>> 210 sys_perf_event_open: pid 31104 cpu 0 group_fd -1 flags 0x8
>> 211 sys_perf_event_open: pid 31104 cpu 1 group_fd -1 flags 0x8
>> 212 sys_perf_event_open: pid 31104 cpu 2 group_fd -1 flags 0x8
>> 213 sys_perf_event_open: pid 31104 cpu 3 group_fd -1 flags 0x8
>> 214 ------------------------------------------------------------
>> 215
>> 216
>> 217 config terms
>> 218 ------------
>> 219
>> 220 The June 2015 version of Intel 64 and IA-32 Architectures Software Developer
>> 221 Manuals, Chapter 36 Intel Processor Trace, defined new Intel PT features.
>> 222 Some of the features are reflect in new config terms. All the config terms are
>> 223 described below.
>> 224
>> 225 tsc Always supported. Produces TSC timestamp packets to provide
>> 226 timing information. In some cases it is possible to decode
>> 227 without timing information, for example a per-thread context
>> 228 that does not overlap executable memory maps.
>> 229
>> 230 The default config selects tsc (i.e. tsc=1).
>> 231
>> 232 noretcomp Always supported. Disables "return compression" so a TIP packet
>> 233 is produced when a function returns. Causes more packets to be
>> 234 produced but might make decoding more reliable.
>> 235
>> 236 The default config does not select noretcomp (i.e. noretcomp=0).
>> 237
>> 238 psb_period Allows the frequency of PSB packets to be specified.
>> 239
>> 240 The PSB packet is a synchronization packet that provides a
>> 241 starting point for decoding or recovery from errors.
>> 242
>> 243 Support for psb_period is indicated by:
>> 244
>> 245 /sys/bus/event_source/devices/intel_pt/caps/psb_cyc
>> 246
>> 247 which contains "1" if the feature is supported and "0"
>> 248 otherwise.
>> 249
>> 250 Valid values are given by:
>> 251
>> 252 /sys/bus/event_source/devices/intel_pt/caps/psb_periods
>> 253
>> 254 which contains a hexadecimal value, the bits of which represent
>> 255 valid values e.g. bit 2 set means value 2 is valid.
>> 256
>> 257 The psb_period value is converted to the approximate number of
>> 258 trace bytes between PSB packets as:
>> 259
>> 260 2 ^ (value + 11)
>> 261
>> 262 e.g. value 3 means 16KiB bytes between PSBs
>> 263
>> 264 If an invalid value is entered, the error message
>> 265 will give a list of valid values e.g.
>> 266
>> 267 $ perf record -e intel_pt/psb_period=15/u uname
>> 268 Invalid psb_period for intel_pt. Valid values are: 0-5
>> 269
>> 270 If MTC packets are selected, the default config selects a value
>> 271 of 3 (i.e. psb_period=3) or the nearest lower value that is
>> 272 supported (0 is always supported). Otherwise the default is 0.
>> 273
>> 274 If decoding is expected to be reliable and the buffer is large
>> 275 then a large PSB period can be used.
>> 276
>> 277 Because a TSC packet is produced with PSB, the PSB period can
>> 278 also affect the granularity to timing information in the absence
>> 279 of MTC or CYC.
>> 280
>> 281 mtc Produces MTC timing packets.
>> 282
>> 283 MTC packets provide finer grain timestamp information than TSC
>> 284 packets. MTC packets record time using the hardware crystal
>> 285 clock (CTC) which is related to TSC packets using a TMA packet.
>> 286
>> 287 Support for this feature is indicated by:
>> 288
>> 289 /sys/bus/event_source/devices/intel_pt/caps/mtc
>> 290
>> 291 which contains "1" if the feature is supported and
>> 292 "0" otherwise.
>> 293
>> 294 The frequency of MTC packets can also be specified - see
>> 295 mtc_period below.
>> 296
>> 297 mtc_period Specifies how frequently MTC packets are produced - see mtc
>> 298 above for how to determine if MTC packets are supported.
>> 299
>> 300 Valid values are given by:
>> 301
>> 302 /sys/bus/event_source/devices/intel_pt/caps/mtc_periods
>> 303
>> 304 which contains a hexadecimal value, the bits of which represent
>> 305 valid values e.g. bit 2 set means value 2 is valid.
>> 306
>> 307 The mtc_period value is converted to the MTC frequency as:
>> 308
>> 309 CTC-frequency / (2 ^ value)
>> 310
>> 311 e.g. value 3 means one eighth of CTC-frequency
>> 312
>> 313 Where CTC is the hardware crystal clock, the frequency of which
>> 314 can be related to TSC via values provided in cpuid leaf 0x15.
>> 315
>> 316 If an invalid value is entered, the error message
>> 317 will give a list of valid values e.g.
>> 318
>> 319 $ perf record -e intel_pt/mtc_period=15/u uname
>> 320 Invalid mtc_period for intel_pt. Valid values are: 0,3,6,9
>> 321
>> 322 The default value is 3 or the nearest lower value
>> 323 that is supported (0 is always supported).
>> 324
>> 325 cyc Produces CYC timing packets.
>> 326
>> 327 CYC packets provide even finer grain timestamp information than
>> 328 MTC and TSC packets. A CYC packet contains the number of CPU
>> 329 cycles since the last CYC packet. Unlike MTC and TSC packets,
>> 330 CYC packets are only sent when another packet is also sent.
>> 331
>> 332 Support for this feature is indicated by:
>> 333
>> 334 /sys/bus/event_source/devices/intel_pt/caps/psb_cyc
>> 335
>> 336 which contains "1" if the feature is supported and
>> 337 "0" otherwise.
>> 338
>> 339 The number of CYC packets produced can be reduced by specifying
>> 340 a threshold - see cyc_thresh below.
>> 341
>> 342 cyc_thresh Specifies how frequently CYC packets are produced - see cyc
>> 343 above for how to determine if CYC packets are supported.
>> 344
>> 345 Valid cyc_thresh values are given by:
>> 346
>> 347 /sys/bus/event_source/devices/intel_pt/caps/cycle_thresholds
>> 348
>> 349 which contains a hexadecimal value, the bits of which represent
>> 350 valid values e.g. bit 2 set means value 2 is valid.
>> 351
>> 352 The cyc_thresh value represents the minimum number of CPU cycles
>> 353 that must have passed before a CYC packet can be sent. The
>> 354 number of CPU cycles is:
>> 355
>> 356 2 ^ (value - 1)
>> 357
>> 358 e.g. value 4 means 8 CPU cycles must pass before a CYC packet
>> 359 can be sent. Note a CYC packet is still only sent when another
>> 360 packet is sent, not at, e.g. every 8 CPU cycles.
>> 361
>> 362 If an invalid value is entered, the error message
>> 363 will give a list of valid values e.g.
>> 364
>> 365 $ perf record -e intel_pt/cyc,cyc_thresh=15/u uname
>> 366 Invalid cyc_thresh for intel_pt. Valid values are: 0-12
>> 367
>> 368 CYC packets are not requested by default.
>> 369
>> 370 pt Specifies pass-through which enables the 'branch' config term.
>> 371
>> 372 The default config selects 'pt' if it is available, so a user will
>> 373 never need to specify this term.
>> 374
>> 375 branch Enable branch tracing. Branch tracing is enabled by default so to
>> 376 disable branch tracing use 'branch=0'.
>> 377
>> 378 The default config selects 'branch' if it is available.
>> 379
>> 380 ptw Enable PTWRITE packets which are produced when a ptwrite instruction
>> 381 is executed.
>> 382
>> 383 Support for this feature is indicated by:
>> 384
>> 385 /sys/bus/event_source/devices/intel_pt/caps/ptwrite
>> 386
>> 387 which contains "1" if the feature is supported and
>> 388 "0" otherwise.
>> 389
>> 390 fup_on_ptw Enable a FUP packet to follow the PTWRITE packet. The FUP packet
>> 391 provides the address of the ptwrite instruction. In the absence of
>> 392 fup_on_ptw, the decoder will use the address of the previous branch
>> 393 if branch tracing is enabled, otherwise the address will be zero.
>> 394 Note that fup_on_ptw will work even when branch tracing is disabled.
>> 395
>> 396 pwr_evt Enable power events. The power events provide information about
>> 397 changes to the CPU C-state.
>> 398
>> 399 Support for this feature is indicated by:
>> 400
>> 401 /sys/bus/event_source/devices/intel_pt/caps/power_event_trace
>> 402
>> 403 which contains "1" if the feature is supported and
>> 404 "0" otherwise.
>> 405
>> 406
>> 407 new snapshot option
>> 408 -------------------
>> 409
>> 410 The difference between full trace and snapshot from the kernel's perspective is
>> 411 that in full trace we don't overwrite trace data that the user hasn't collected
>> 412 yet (and indicated that by advancing aux_tail), whereas in snapshot mode we let
>> 413 the trace run and overwrite older data in the buffer so that whenever something
>> 414 interesting happens, we can stop it and grab a snapshot of what was going on
>> 415 around that interesting moment.
>> 416
>> 417 To select snapshot mode a new option has been added:
>> 418
>> 419 -S
>> 420
>> 421 Optionally it can be followed by the snapshot size e.g.
>> 422
>> 423 -S0x100000
>> 424
>> 425 The default snapshot size is the auxtrace mmap size. If neither auxtrace mmap size
>> 426 nor snapshot size is specified, then the default is 4MiB for privileged users
>> 427 (or if /proc/sys/kernel/perf_event_paranoid < 0), 128KiB for unprivileged users.
>> 428 If an unprivileged user does not specify mmap pages, the mmap pages will be
>> 429 reduced as described in the 'new auxtrace mmap size option' section below.
>> 430
>> 431 The snapshot size is displayed if the option -vv is used e.g.
>> 432
>> 433 Intel PT snapshot size: %zu
>> 434
>> 435
>> 436 new auxtrace mmap size option
>> 437 ---------------------------
>> 438
>> 439 Intel PT buffer size is specified by an addition to the -m option e.g.
>> 440
>> 441 -m,16
>> 442
>> 443 selects a buffer size of 16 pages i.e. 64KiB.
>> 444
>> 445 Note that the existing functionality of -m is unchanged. The auxtrace mmap size
>> 446 is specified by the optional addition of a comma and the value.
>> 447
>> 448 The default auxtrace mmap size for Intel PT is 4MiB/page_size for privileged users
>> 449 (or if /proc/sys/kernel/perf_event_paranoid < 0), 128KiB for unprivileged users.
>> 450 If an unprivileged user does not specify mmap pages, the mmap pages will be
>> 451 reduced from the default 512KiB/page_size to 256KiB/page_size, otherwise the
>> 452 user is likely to get an error as they exceed their mlock limit (Max locked
>> 453 memory as shown in /proc/self/limits). Note that perf does not count the first
>> 454 512KiB (actually /proc/sys/kernel/perf_event_mlock_kb minus 1 page) per cpu
>> 455 against the mlock limit so an unprivileged user is allowed 512KiB per cpu plus
>> 456 their mlock limit (which defaults to 64KiB but is not multiplied by the number
>> 457 of cpus).
>> 458
>> 459 In full-trace mode, powers of two are allowed for buffer size, with a minimum
>> 460 size of 2 pages. In snapshot mode, it is the same but the minimum size is
>> 461 1 page.
>> 462
>> 463 The mmap size and auxtrace mmap size are displayed if the -vv option is used e.g.
>> 464
>> 465 mmap length 528384
>> 466 auxtrace mmap length 4198400
>> 467
>> 468
>> 469 Intel PT modes of operation
>> 470 ---------------------------
>> 471
>> 472 Intel PT can be used in 2 modes:
>> 473 full-trace mode
>> 474 snapshot mode
>> 475
>> 476 Full-trace mode traces continuously e.g.
>> 477
>> 478 perf record -e intel_pt//u uname
>> 479
>> 480 Snapshot mode captures the available data when a signal is sent e.g.
>> 481
>> 482 perf record -v -e intel_pt//u -S ./loopy 1000000000 &
>> 483 [1] 11435
>> 484 kill -USR2 11435
>> 485 Recording AUX area tracing snapshot
>> 486
>> 487 Note that the signal sent is SIGUSR2.
>> 488 Note that "Recording AUX area tracing snapshot" is displayed because the -v
>> 489 option is used.
>> 490
>> 491 The 2 modes cannot be used together.
>> 492
>> 493
>> 494 Buffer handling
>> 495 ---------------
>> 496
>> 497 There may be buffer limitations (i.e. single ToPa entry) which means that actual
>> 498 buffer sizes are limited to powers of 2 up to 4MiB (MAX_ORDER). In order to
>> 499 provide other sizes, and in particular an arbitrarily large size, multiple
>> 500 buffers are logically concatenated. However an interrupt must be used to switch
>> 501 between buffers. That has two potential problems:
>> 502 a) the interrupt may not be handled in time so that the current buffer
>> 503 becomes full and some trace data is lost.
>> 504 b) the interrupts may slow the system and affect the performance
>> 505 results.
>> 506
>> 507 If trace data is lost, the driver sets 'truncated' in the PERF_RECORD_AUX event
>> 508 which the tools report as an error.
>> 509
>> 510 In full-trace mode, the driver waits for data to be copied out before allowing
>> 511 the (logical) buffer to wrap-around. If data is not copied out quickly enough,
>> 512 again 'truncated' is set in the PERF_RECORD_AUX event. If the driver has to
>> 513 wait, the intel_pt event gets disabled. Because it is difficult to know when
>> 514 that happens, perf tools always re-enable the intel_pt event after copying out
>> 515 data.
>> 516
>> 517
>> 518 Intel PT and build ids
>> 519 ----------------------
>> 520
>> 521 By default "perf record" post-processes the event stream to find all build ids
>> 522 for executables for all addresses sampled. Deliberately, Intel PT is not
>> 523 decoded for that purpose (it would take too long). Instead the build ids for
>> 524 all executables encountered (due to mmap, comm or task events) are included
>> 525 in the perf.data file.
>> 526
>> 527 To see buildids included in the perf.data file use the command:
>> 528
>> 529 perf buildid-list
>> 530
>> 531 If the perf.data file contains Intel PT data, that is the same as:
>> 532
>> 533 perf buildid-list --with-hits
>> 534
>> 535
>> 536 Snapshot mode and event disabling
>> 537 ---------------------------------
>> 538
>> 539 In order to make a snapshot, the intel_pt event is disabled using an IOCTL,
>> 540 namely PERF_EVENT_IOC_DISABLE. However doing that can also disable the
>> 541 collection of side-band information. In order to prevent that, a dummy
>> 542 software event has been introduced that permits tracking events (like mmaps) to
>> 543 continue to be recorded while intel_pt is disabled. That is important to ensure
>> 544 there is complete side-band information to allow the decoding of subsequent
>> 545 snapshots.
>> 546
>> 547 A test has been created for that. To find the test:
>> 548
>> 549 perf test list
>> 550 ...
>> 551 23: Test using a dummy software event to keep tracking
>> 552
>> 553 To run the test:
>> 554
>> 555 perf test 23
>> 556 23: Test using a dummy software event to keep tracking : Ok
>> 557
>> 558
>> 559 perf record modes (nothing new here)
>> 560 ------------------------------------
>> 561
>> 562 perf record essentially operates in one of three modes:
>> 563 per thread
>> 564 per cpu
>> 565 workload only
>> 566
>> 567 "per thread" mode is selected by -t or by --per-thread (with -p or -u or just a
>> 568 workload).
>> 569 "per cpu" is selected by -C or -a.
>> 570 "workload only" mode is selected by not using the other options but providing a
>> 571 command to run (i.e. the workload).
>> 572
>> 573 In per-thread mode an exact list of threads is traced. There is no inheritance.
>> 574 Each thread has its own event buffer.
>> 575
>> 576 In per-cpu mode all processes (or processes from the selected cgroup i.e. -G
>> 577 option, or processes selected with -p or -u) are traced. Each cpu has its own
>> 578 buffer. Inheritance is allowed.
>> 579
>> 580 In workload-only mode, the workload is traced but with per-cpu buffers.
>> 581 Inheritance is allowed. Note that you can now trace a workload in per-thread
>> 582 mode by using the --per-thread option.
>> 583
>> 584
>> 585 Privileged vs non-privileged users
>> 586 ----------------------------------
>> 587
>> 588 Unless /proc/sys/kernel/perf_event_paranoid is set to -1, unprivileged users
>> 589 have memory limits imposed upon them. That affects what buffer sizes they can
>> 590 have as outlined above.
>> 591
>> 592 The v4.2 kernel introduced support for a context switch metadata event,
>> 593 PERF_RECORD_SWITCH, which allows unprivileged users to see when their processes
>> 594 are scheduled out and in, just not by whom, which is left for the
>> 595 PERF_RECORD_SWITCH_CPU_WIDE, that is only accessible in system wide context,
>> 596 which in turn requires CAP_SYS_ADMIN.
>> 597
>> 598 Please see the 45ac1403f564 ("perf: Add PERF_RECORD_SWITCH to indicate context
>> 599 switches") commit, that introduces these metadata events for further info.
>> 600
>> 601 When working with kernels < v4.2, the following considerations must be taken,
>> 602 as the sched:sched_switch tracepoints will be used to receive such information:
>> 603
>> 604 Unless /proc/sys/kernel/perf_event_paranoid is set to -1, unprivileged users are
>> 605 not permitted to use tracepoints which means there is insufficient side-band
>> 606 information to decode Intel PT in per-cpu mode, and potentially workload-only
>> 607 mode too if the workload creates new processes.
>> 608
>> 609 Note also, that to use tracepoints, read-access to debugfs is required. So if
>> 610 debugfs is not mounted or the user does not have read-access, it will again not
>> 611 be possible to decode Intel PT in per-cpu mode.
>> 612
>> 613
>> 614 sched_switch tracepoint
>> 615 -----------------------
>> 616
>> 617 The sched_switch tracepoint is used to provide side-band data for Intel PT
>> 618 decoding in kernels where the PERF_RECORD_SWITCH metadata event isn't
>> 619 available.
>> 620
>> 621 The sched_switch events are automatically added. e.g. the second event shown
>> 622 below:
>> 623
>> 624 $ perf record -vv -e intel_pt//u uname
>> 625 ------------------------------------------------------------
>> 626 perf_event_attr:
>> 627 type 6
>> 628 size 112
>> 629 config 0x400
>> 630 { sample_period, sample_freq } 1
>> 631 sample_type IP|TID|TIME|CPU|IDENTIFIER
>> 632 read_format ID
>> 633 disabled 1
>> 634 inherit 1
>> 635 exclude_kernel 1
>> 636 exclude_hv 1
>> 637 enable_on_exec 1
>> 638 sample_id_all 1
>> 639 ------------------------------------------------------------
>> 640 sys_perf_event_open: pid 31104 cpu 0 group_fd -1 flags 0x8
>> 641 sys_perf_event_open: pid 31104 cpu 1 group_fd -1 flags 0x8
>> 642 sys_perf_event_open: pid 31104 cpu 2 group_fd -1 flags 0x8
>> 643 sys_perf_event_open: pid 31104 cpu 3 group_fd -1 flags 0x8
>> 644 ------------------------------------------------------------
>> 645 perf_event_attr:
>> 646 type 2
>> 647 size 112
>> 648 config 0x108
>> 649 { sample_period, sample_freq } 1
>> 650 sample_type IP|TID|TIME|CPU|PERIOD|RAW|IDENTIFIER
>> 651 read_format ID
>> 652 inherit 1
>> 653 sample_id_all 1
>> 654 exclude_guest 1
>> 655 ------------------------------------------------------------
>> 656 sys_perf_event_open: pid -1 cpu 0 group_fd -1 flags 0x8
>> 657 sys_perf_event_open: pid -1 cpu 1 group_fd -1 flags 0x8
>> 658 sys_perf_event_open: pid -1 cpu 2 group_fd -1 flags 0x8
>> 659 sys_perf_event_open: pid -1 cpu 3 group_fd -1 flags 0x8
>> 660 ------------------------------------------------------------
>> 661 perf_event_attr:
>> 662 type 1
>> 663 size 112
>> 664 config 0x9
>> 665 { sample_period, sample_freq } 1
>> 666 sample_type IP|TID|TIME|IDENTIFIER
>> 667 read_format ID
>> 668 disabled 1
>> 669 inherit 1
>> 670 exclude_kernel 1
>> 671 exclude_hv 1
>> 672 mmap 1
>> 673 comm 1
>> 674 enable_on_exec 1
>> 675 task 1
>> 676 sample_id_all 1
>> 677 mmap2 1
>> 678 comm_exec 1
>> 679 ------------------------------------------------------------
>> 680 sys_perf_event_open: pid 31104 cpu 0 group_fd -1 flags 0x8
>> 681 sys_perf_event_open: pid 31104 cpu 1 group_fd -1 flags 0x8
>> 682 sys_perf_event_open: pid 31104 cpu 2 group_fd -1 flags 0x8
>> 683 sys_perf_event_open: pid 31104 cpu 3 group_fd -1 flags 0x8
>> 684 mmap size 528384B
>> 685 AUX area mmap length 4194304
>> 686 perf event ring buffer mmapped per cpu
>> 687 Synthesizing auxtrace information
>> 688 Linux
>> 689 [ perf record: Woken up 1 times to write data ]
>> 690 [ perf record: Captured and wrote 0.042 MB perf.data ]
>> 691
>> 692 Note, the sched_switch event is only added if the user is permitted to use it
>> 693 and only in per-cpu mode.
>> 694
>> 695 Note also, the sched_switch event is only added if TSC packets are requested.
>> 696 That is because, in the absence of timing information, the sched_switch events
>> 697 cannot be matched against the Intel PT trace.
>> 698
>> 699
>> 700 perf script
>> 701 ===========
>> 702
>> 703 By default, perf script will decode trace data found in the perf.data file.
>> 704 This can be further controlled by new option --itrace.
>> 705
>> 706
>> 707 New --itrace option
>> 708 -------------------
>> 709
>> 710 Having no option is the same as
>> 711
>> 712 --itrace
>> 713
>> 714 which, in turn, is the same as
>> 715
>> 716 --itrace=ibxwpe
>> 717
>> 718 The letters are:
>> 719
>> 720 i synthesize "instructions" events
>> 721 b synthesize "branches" events
>> 722 x synthesize "transactions" events
>> 723 w synthesize "ptwrite" events
>> 724 p synthesize "power" events
>> 725 c synthesize branches events (calls only)
>> 726 r synthesize branches events (returns only)
>> 727 e synthesize tracing error events
>> 728 d create a debug log
>> 729 g synthesize a call chain (use with i or x)
>> 730 l synthesize last branch entries (use with i or x)
>> 731 s skip initial number of events
>> 732
>> 733 "Instructions" events look like they were recorded by "perf record -e
>> 734 instructions".
>> 735
>> 736 "Branches" events look like they were recorded by "perf record -e branches". "c"
>> 737 and "r" can be combined to get calls and returns.
>> 738
>> 739 "Transactions" events correspond to the start or end of transactions. The
>> 740 'flags' field can be used in perf script to determine whether the event is a
>> 741 tranasaction start, commit or abort.
>> 742
>> 743 Note that "instructions", "branches" and "transactions" events depend on code
>> 744 flow packets which can be disabled by using the config term "branch=0". Refer
>> 745 to the config terms section above.
>> 746
>> 747 "ptwrite" events record the payload of the ptwrite instruction and whether
>> 748 "fup_on_ptw" was used. "ptwrite" events depend on PTWRITE packets which are
>> 749 recorded only if the "ptw" config term was used. Refer to the config terms
>> 750 section above. perf script "synth" field displays "ptwrite" information like
>> 751 this: "ip: 0 payload: 0x123456789abcdef0" where "ip" is 1 if "fup_on_ptw" was
>> 752 used.
>> 753
>> 754 "Power" events correspond to power event packets and CBR (core-to-bus ratio)
>> 755 packets. While CBR packets are always recorded when tracing is enabled, power
>> 756 event packets are recorded only if the "pwr_evt" config term was used. Refer to
>> 757 the config terms section above. The power events record information about
>> 758 C-state changes, whereas CBR is indicative of CPU frequency. perf script
>> 759 "event,synth" fields display information like this:
>> 760 cbr: cbr: 22 freq: 2189 MHz (200%)
>> 761 mwait: hints: 0x60 extensions: 0x1
>> 762 pwre: hw: 0 cstate: 2 sub-cstate: 0
>> 763 exstop: ip: 1
>> 764 pwrx: deepest cstate: 2 last cstate: 2 wake reason: 0x4
>> 765 Where:
>> 766 "cbr" includes the frequency and the percentage of maximum non-turbo
>> 767 "mwait" shows mwait hints and extensions
>> 768 "pwre" shows C-state transitions (to a C-state deeper than C0) and
>> 769 whether initiated by hardware
>> 770 "exstop" indicates execution stopped and whether the IP was recorded
>> 771 exactly,
>> 772 "pwrx" indicates return to C0
>> 773 For more details refer to the Intel 64 and IA-32 Architectures Software
>> 774 Developer Manuals.
>> 775
>> 776 Error events show where the decoder lost the trace. Error events
>> 777 are quite important. Users must know if what they are seeing is a complete
>> 778 picture or not.
>> 779
>> 780 The "d" option will cause the creation of a file "intel_pt.log" containing all
>> 781 decoded packets and instructions. Note that this option slows down the decoder
>> 782 and that the resulting file may be very large.
>> 783
>> 784 In addition, the period of the "instructions" event can be specified. e.g.
>> 785
>> 786 --itrace=i10us
>> 787
>> 788 sets the period to 10us i.e. one instruction sample is synthesized for each 10
>> 789 microseconds of trace. Alternatives to "us" are "ms" (milliseconds),
>> 790 "ns" (nanoseconds), "t" (TSC ticks) or "i" (instructions).
>> 791
>> 792 "ms", "us" and "ns" are converted to TSC ticks.
>> 793
>> 794 The timing information included with Intel PT does not give the time of every
>> 795 instruction. Consequently, for the purpose of sampling, the decoder estimates
>> 796 the time since the last timing packet based on 1 tick per instruction. The time
>> 797 on the sample is *not* adjusted and reflects the last known value of TSC.
>> 798
>> 799 For Intel PT, the default period is 100us.
>> 800
>> 801 Setting it to a zero period means "as often as possible".
>> 802
>> 803 In the case of Intel PT that is the same as a period of 1 and a unit of
>> 804 'instructions' (i.e. --itrace=i1i).
>> 805
>> 806 Also the call chain size (default 16, max. 1024) for instructions or
>> 807 transactions events can be specified. e.g.
>> 808
>> 809 --itrace=ig32
>> 810 --itrace=xg32
>> 811
>> 812 Also the number of last branch entries (default 64, max. 1024) for instructions or
>> 813 transactions events can be specified. e.g.
>> 814
>> 815 --itrace=il10
>> 816 --itrace=xl10
>> 817
>> 818 Note that last branch entries are cleared for each sample, so there is no overlap
>> 819 from one sample to the next.
>> 820
>> 821 To disable trace decoding entirely, use the option --no-itrace.
>> 822
>> 823 It is also possible to skip events generated (instructions, branches, transactions)
>> 824 at the beginning. This is useful to ignore initialization code.
>> 825
>> 826 --itrace=i0nss1000000
>> 827
>> 828 skips the first million instructions.
>> 829
>> 830 dump option
>> 831 -----------
>> 832
>> 833 perf script has an option (-D) to "dump" the events i.e. display the binary
>> 834 data.
>> 835
>> 836 When -D is used, Intel PT packets are displayed. The packet decoder does not
>> 837 pay attention to PSB packets, but just decodes the bytes - so the packets seen
>> 838 by the actual decoder may not be identical in places where the data is corrupt.
>> 839 One example of that would be when the buffer-switching interrupt has been too
>> 840 slow, and the buffer has been filled completely. In that case, the last packet
>> 841 in the buffer might be truncated and immediately followed by a PSB as the trace
>> 842 continues in the next buffer.
>> 843
>> 844 To disable the display of Intel PT packets, combine the -D option with
>> 845 --no-itrace.
>> 846
>> 847
>> 848 perf report
>> 849 ===========
>> 850
>> 851 By default, perf report will decode trace data found in the perf.data file.
>> 852 This can be further controlled by new option --itrace exactly the same as
>> 853 perf script, with the exception that the default is --itrace=igxe.
>> 854
>> 855
>> 856 perf inject
>> 857 ===========
>> 858
>> 859 perf inject also accepts the --itrace option in which case tracing data is
>> 860 removed and replaced with the synthesized events. e.g.
>> 861
>> 862 perf inject --itrace -i perf.data -o perf.data.new
>> 863
>> 864 Below is an example of using Intel PT with autofdo. It requires autofdo
>> 865 (https://github.com/google/autofdo) and gcc version 5. The bubble
>> 866 sort example is from the AutoFDO tutorial (https://gcc.gnu.org/wiki/AutoFDO/Tutorial)
>> 867 amended to take the number of elements as a parameter.
>> 868
>> 869 $ gcc-5 -O3 sort.c -o sort_optimized
>> 870 $ ./sort_optimized 30000
>> 871 Bubble sorting array of 30000 elements
>> 872 2254 ms
>> 873
>> 874 $ cat ~/.perfconfig
>> 875 [intel-pt]
>> 876 mispred-all
>> 877
>> 878 $ perf record -e intel_pt//u ./sort 3000
>> 879 Bubble sorting array of 3000 elements
>> 880 58 ms
>> 881 [ perf record: Woken up 2 times to write data ]
>> 882 [ perf record: Captured and wrote 3.939 MB perf.data ]
>> 883 $ perf inject -i perf.data -o inj --itrace=i100usle --strip
>> 884 $ ./create_gcov --binary=./sort --profile=inj --gcov=sort.gcov -gcov_version=1
>> 885 $ gcc-5 -O3 -fauto-profile=sort.gcov sort.c -o sort_autofdo
>> 886 $ ./sort_autofdo 30000
>> 887 Bubble sorting array of 30000 elements
>> 888 2155 ms
>> 889
>> 890 Note there is currently no advantage to using Intel PT instead of LBR, but
>> 891 that may change in the future if greater use is made of the data.
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.