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 all itrace sampling
>> 92 options, which will list all the samples.
>> 93
>> 94 perf record -e intel_pt//u ls
>> 95 perf script --itrace=ibxwpe
>> 96
>> 97 An interesting field that is not printed by default is 'flags' which can be
>> 98 displayed as follows:
>> 99
>> 100 perf script --itrace=ibxwpe -F+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 Another interesting field that is not printed by default is 'ipc' which can be
>> 107 displayed as follows:
>> 108
>> 109 perf script --itrace=be -F+ipc
>> 110
>> 111 There are two ways that instructions-per-cycle (IPC) can be calculated depending
>> 112 on the recording.
>> 113
>> 114 If the 'cyc' config term (see config terms section below) was used, then IPC is
>> 115 calculated using the cycle count from CYC packets, otherwise MTC packets are
>> 116 used - refer to the 'mtc' config term. When MTC is used, however, the values
>> 117 are less accurate because the timing is less accurate.
>> 118
>> 119 Because Intel PT does not update the cycle count on every branch or instruction,
>> 120 the values will often be zero. When there are values, they will be the number
>> 121 of instructions and number of cycles since the last update, and thus represent
>> 122 the average IPC since the last IPC for that event type. Note IPC for "branches"
>> 123 events is calculated separately from IPC for "instructions" events.
>> 124
>> 125 Also note that the IPC instruction count may or may not include the current
>> 126 instruction. If the cycle count is associated with an asynchronous branch
>> 127 (e.g. page fault or interrupt), then the instruction count does not include the
>> 128 current instruction, otherwise it does. That is consistent with whether or not
>> 129 that instruction has retired when the cycle count is updated.
>> 130
>> 131 Another note, in the case of "branches" events, non-taken branches are not
>> 132 presently sampled, so IPC values for them do not appear e.g. a CYC packet with a
>> 133 TNT packet that starts with a non-taken branch. To see every possible IPC
>> 134 value, "instructions" events can be used e.g. --itrace=i0ns
>> 135
>> 136 While it is possible to create scripts to analyze the data, an alternative
>> 137 approach is available to export the data to a sqlite or postgresql database.
>> 138 Refer to script export-to-sqlite.py or export-to-postgresql.py for more details,
>> 139 and to script exported-sql-viewer.py for an example of using the database.
>> 140
>> 141 There is also script intel-pt-events.py which provides an example of how to
>> 142 unpack the raw data for power events and PTWRITE.
>> 143
>> 144 As mentioned above, it is easy to capture too much data. One way to limit the
>> 145 data captured is to use 'snapshot' mode which is explained further below.
>> 146 Refer to 'new snapshot option' and 'Intel PT modes of operation' further below.
>> 147
>> 148 Another problem that will be experienced is decoder errors. They can be caused
>> 149 by inability to access the executed image, self-modified or JIT-ed code, or the
>> 150 inability to match side-band information (such as context switches and mmaps)
>> 151 which results in the decoder not knowing what code was executed.
>> 152
>> 153 There is also the problem of perf not being able to copy the data fast enough,
>> 154 resulting in data lost because the buffer was full. See 'Buffer handling' below
>> 155 for more details.
>> 156
>> 157
>> 158 perf record
>> 159 ===========
>> 160
>> 161 new event
>> 162 ---------
>> 163
>> 164 The Intel PT kernel driver creates a new PMU for Intel PT. PMU events are
>> 165 selected by providing the PMU name followed by the "config" separated by slashes.
>> 166 An enhancement has been made to allow default "config" e.g. the option
>> 167
>> 168 -e intel_pt//
>> 169
>> 170 will use a default config value. Currently that is the same as
>> 171
>> 172 -e intel_pt/tsc,noretcomp=0/
>> 173
>> 174 which is the same as
>> 175
>> 176 -e intel_pt/tsc=1,noretcomp=0/
>> 177
>> 178 Note there are now new config terms - see section 'config terms' further below.
>> 179
>> 180 The config terms are listed in /sys/devices/intel_pt/format. They are bit
>> 181 fields within the config member of the struct perf_event_attr which is
>> 182 passed to the kernel by the perf_event_open system call. They correspond to bit
>> 183 fields in the IA32_RTIT_CTL MSR. Here is a list of them and their definitions:
>> 184
>> 185 $ grep -H . /sys/bus/event_source/devices/intel_pt/format/*
>> 186 /sys/bus/event_source/devices/intel_pt/format/cyc:config:1
>> 187 /sys/bus/event_source/devices/intel_pt/format/cyc_thresh:config:19-22
>> 188 /sys/bus/event_source/devices/intel_pt/format/mtc:config:9
>> 189 /sys/bus/event_source/devices/intel_pt/format/mtc_period:config:14-17
>> 190 /sys/bus/event_source/devices/intel_pt/format/noretcomp:config:11
>> 191 /sys/bus/event_source/devices/intel_pt/format/psb_period:config:24-27
>> 192 /sys/bus/event_source/devices/intel_pt/format/tsc:config:10
>> 193
>> 194 Note that the default config must be overridden for each term i.e.
>> 195
>> 196 -e intel_pt/noretcomp=0/
>> 197
>> 198 is the same as:
>> 199
>> 200 -e intel_pt/tsc=1,noretcomp=0/
>> 201
>> 202 So, to disable TSC packets use:
>> 203
>> 204 -e intel_pt/tsc=0/
>> 205
>> 206 It is also possible to specify the config value explicitly:
>> 207
>> 208 -e intel_pt/config=0x400/
>> 209
>> 210 Note that, as with all events, the event is suffixed with event modifiers:
>> 211
>> 212 u userspace
>> 213 k kernel
>> 214 h hypervisor
>> 215 G guest
>> 216 H host
>> 217 p precise ip
>> 218
>> 219 'h', 'G' and 'H' are for virtualization which is not supported by Intel PT.
>> 220 'p' is also not relevant to Intel PT. So only options 'u' and 'k' are
>> 221 meaningful for Intel PT.
>> 222
>> 223 perf_event_attr is displayed if the -vv option is used e.g.
>> 224
>> 225 ------------------------------------------------------------
>> 226 perf_event_attr:
>> 227 type 6
>> 228 size 112
>> 229 config 0x400
>> 230 { sample_period, sample_freq } 1
>> 231 sample_type IP|TID|TIME|CPU|IDENTIFIER
>> 232 read_format ID
>> 233 disabled 1
>> 234 inherit 1
>> 235 exclude_kernel 1
>> 236 exclude_hv 1
>> 237 enable_on_exec 1
>> 238 sample_id_all 1
>> 239 ------------------------------------------------------------
>> 240 sys_perf_event_open: pid 31104 cpu 0 group_fd -1 flags 0x8
>> 241 sys_perf_event_open: pid 31104 cpu 1 group_fd -1 flags 0x8
>> 242 sys_perf_event_open: pid 31104 cpu 2 group_fd -1 flags 0x8
>> 243 sys_perf_event_open: pid 31104 cpu 3 group_fd -1 flags 0x8
>> 244 ------------------------------------------------------------
>> 245
>> 246
>> 247 config terms
>> 248 ------------
>> 249
>> 250 The June 2015 version of Intel 64 and IA-32 Architectures Software Developer
>> 251 Manuals, Chapter 36 Intel Processor Trace, defined new Intel PT features.
>> 252 Some of the features are reflect in new config terms. All the config terms are
>> 253 described below.
>> 254
>> 255 tsc Always supported. Produces TSC timestamp packets to provide
>> 256 timing information. In some cases it is possible to decode
>> 257 without timing information, for example a per-thread context
>> 258 that does not overlap executable memory maps.
>> 259
>> 260 The default config selects tsc (i.e. tsc=1).
>> 261
>> 262 noretcomp Always supported. Disables "return compression" so a TIP packet
>> 263 is produced when a function returns. Causes more packets to be
>> 264 produced but might make decoding more reliable.
>> 265
>> 266 The default config does not select noretcomp (i.e. noretcomp=0).
>> 267
>> 268 psb_period Allows the frequency of PSB packets to be specified.
>> 269
>> 270 The PSB packet is a synchronization packet that provides a
>> 271 starting point for decoding or recovery from errors.
>> 272
>> 273 Support for psb_period is indicated by:
>> 274
>> 275 /sys/bus/event_source/devices/intel_pt/caps/psb_cyc
>> 276
>> 277 which contains "1" if the feature is supported and "0"
>> 278 otherwise.
>> 279
>> 280 Valid values are given by:
>> 281
>> 282 /sys/bus/event_source/devices/intel_pt/caps/psb_periods
>> 283
>> 284 which contains a hexadecimal value, the bits of which represent
>> 285 valid values e.g. bit 2 set means value 2 is valid.
>> 286
>> 287 The psb_period value is converted to the approximate number of
>> 288 trace bytes between PSB packets as:
>> 289
>> 290 2 ^ (value + 11)
>> 291
>> 292 e.g. value 3 means 16KiB bytes between PSBs
>> 293
>> 294 If an invalid value is entered, the error message
>> 295 will give a list of valid values e.g.
>> 296
>> 297 $ perf record -e intel_pt/psb_period=15/u uname
>> 298 Invalid psb_period for intel_pt. Valid values are: 0-5
>> 299
>> 300 If MTC packets are selected, the default config selects a value
>> 301 of 3 (i.e. psb_period=3) or the nearest lower value that is
>> 302 supported (0 is always supported). Otherwise the default is 0.
>> 303
>> 304 If decoding is expected to be reliable and the buffer is large
>> 305 then a large PSB period can be used.
>> 306
>> 307 Because a TSC packet is produced with PSB, the PSB period can
>> 308 also affect the granularity to timing information in the absence
>> 309 of MTC or CYC.
>> 310
>> 311 mtc Produces MTC timing packets.
>> 312
>> 313 MTC packets provide finer grain timestamp information than TSC
>> 314 packets. MTC packets record time using the hardware crystal
>> 315 clock (CTC) which is related to TSC packets using a TMA packet.
>> 316
>> 317 Support for this feature is indicated by:
>> 318
>> 319 /sys/bus/event_source/devices/intel_pt/caps/mtc
>> 320
>> 321 which contains "1" if the feature is supported and
>> 322 "0" otherwise.
>> 323
>> 324 The frequency of MTC packets can also be specified - see
>> 325 mtc_period below.
>> 326
>> 327 mtc_period Specifies how frequently MTC packets are produced - see mtc
>> 328 above for how to determine if MTC packets are supported.
>> 329
>> 330 Valid values are given by:
>> 331
>> 332 /sys/bus/event_source/devices/intel_pt/caps/mtc_periods
>> 333
>> 334 which contains a hexadecimal value, the bits of which represent
>> 335 valid values e.g. bit 2 set means value 2 is valid.
>> 336
>> 337 The mtc_period value is converted to the MTC frequency as:
>> 338
>> 339 CTC-frequency / (2 ^ value)
>> 340
>> 341 e.g. value 3 means one eighth of CTC-frequency
>> 342
>> 343 Where CTC is the hardware crystal clock, the frequency of which
>> 344 can be related to TSC via values provided in cpuid leaf 0x15.
>> 345
>> 346 If an invalid value is entered, the error message
>> 347 will give a list of valid values e.g.
>> 348
>> 349 $ perf record -e intel_pt/mtc_period=15/u uname
>> 350 Invalid mtc_period for intel_pt. Valid values are: 0,3,6,9
>> 351
>> 352 The default value is 3 or the nearest lower value
>> 353 that is supported (0 is always supported).
>> 354
>> 355 cyc Produces CYC timing packets.
>> 356
>> 357 CYC packets provide even finer grain timestamp information than
>> 358 MTC and TSC packets. A CYC packet contains the number of CPU
>> 359 cycles since the last CYC packet. Unlike MTC and TSC packets,
>> 360 CYC packets are only sent when another packet is also sent.
>> 361
>> 362 Support for this feature is indicated by:
>> 363
>> 364 /sys/bus/event_source/devices/intel_pt/caps/psb_cyc
>> 365
>> 366 which contains "1" if the feature is supported and
>> 367 "0" otherwise.
>> 368
>> 369 The number of CYC packets produced can be reduced by specifying
>> 370 a threshold - see cyc_thresh below.
>> 371
>> 372 cyc_thresh Specifies how frequently CYC packets are produced - see cyc
>> 373 above for how to determine if CYC packets are supported.
>> 374
>> 375 Valid cyc_thresh values are given by:
>> 376
>> 377 /sys/bus/event_source/devices/intel_pt/caps/cycle_thresholds
>> 378
>> 379 which contains a hexadecimal value, the bits of which represent
>> 380 valid values e.g. bit 2 set means value 2 is valid.
>> 381
>> 382 The cyc_thresh value represents the minimum number of CPU cycles
>> 383 that must have passed before a CYC packet can be sent. The
>> 384 number of CPU cycles is:
>> 385
>> 386 2 ^ (value - 1)
>> 387
>> 388 e.g. value 4 means 8 CPU cycles must pass before a CYC packet
>> 389 can be sent. Note a CYC packet is still only sent when another
>> 390 packet is sent, not at, e.g. every 8 CPU cycles.
>> 391
>> 392 If an invalid value is entered, the error message
>> 393 will give a list of valid values e.g.
>> 394
>> 395 $ perf record -e intel_pt/cyc,cyc_thresh=15/u uname
>> 396 Invalid cyc_thresh for intel_pt. Valid values are: 0-12
>> 397
>> 398 CYC packets are not requested by default.
>> 399
>> 400 pt Specifies pass-through which enables the 'branch' config term.
>> 401
>> 402 The default config selects 'pt' if it is available, so a user will
>> 403 never need to specify this term.
>> 404
>> 405 branch Enable branch tracing. Branch tracing is enabled by default so to
>> 406 disable branch tracing use 'branch=0'.
>> 407
>> 408 The default config selects 'branch' if it is available.
>> 409
>> 410 ptw Enable PTWRITE packets which are produced when a ptwrite instruction
>> 411 is executed.
>> 412
>> 413 Support for this feature is indicated by:
>> 414
>> 415 /sys/bus/event_source/devices/intel_pt/caps/ptwrite
>> 416
>> 417 which contains "1" if the feature is supported and
>> 418 "0" otherwise.
>> 419
>> 420 fup_on_ptw Enable a FUP packet to follow the PTWRITE packet. The FUP packet
>> 421 provides the address of the ptwrite instruction. In the absence of
>> 422 fup_on_ptw, the decoder will use the address of the previous branch
>> 423 if branch tracing is enabled, otherwise the address will be zero.
>> 424 Note that fup_on_ptw will work even when branch tracing is disabled.
>> 425
>> 426 pwr_evt Enable power events. The power events provide information about
>> 427 changes to the CPU C-state.
>> 428
>> 429 Support for this feature is indicated by:
>> 430
>> 431 /sys/bus/event_source/devices/intel_pt/caps/power_event_trace
>> 432
>> 433 which contains "1" if the feature is supported and
>> 434 "0" otherwise.
>> 435
>> 436
>> 437 AUX area sampling option
>> 438 ------------------------
>> 439
>> 440 To select Intel PT "sampling" the AUX area sampling option can be used:
>> 441
>> 442 --aux-sample
>> 443
>> 444 Optionally it can be followed by the sample size in bytes e.g.
>> 445
>> 446 --aux-sample=8192
>> 447
>> 448 In addition, the Intel PT event to sample must be defined e.g.
>> 449
>> 450 -e intel_pt//u
>> 451
>> 452 Samples on other events will be created containing Intel PT data e.g. the
>> 453 following will create Intel PT samples on the branch-misses event, note the
>> 454 events must be grouped using {}:
>> 455
>> 456 perf record --aux-sample -e '{intel_pt//u,branch-misses:u}'
>> 457
>> 458 An alternative to '--aux-sample' is to add the config term 'aux-sample-size' to
>> 459 events. In this case, the grouping is implied e.g.
>> 460
>> 461 perf record -e intel_pt//u -e branch-misses/aux-sample-size=8192/u
>> 462
>> 463 is the same as:
>> 464
>> 465 perf record -e '{intel_pt//u,branch-misses/aux-sample-size=8192/u}'
>> 466
>> 467 but allows for also using an address filter e.g.:
>> 468
>> 469 perf record -e intel_pt//u --filter 'filter * @/bin/ls' -e branch-misses/aux-sample-size=8192/u -- ls
>> 470
>> 471 It is important to select a sample size that is big enough to contain at least
>> 472 one PSB packet. If not a warning will be displayed:
>> 473
>> 474 Intel PT sample size (%zu) may be too small for PSB period (%zu)
>> 475
>> 476 The calculation used for that is: if sample_size <= psb_period + 256 display the
>> 477 warning. When sampling is used, psb_period defaults to 0 (2KiB).
>> 478
>> 479 The default sample size is 4KiB.
>> 480
>> 481 The sample size is passed in aux_sample_size in struct perf_event_attr. The
>> 482 sample size is limited by the maximum event size which is 64KiB. It is
>> 483 difficult to know how big the event might be without the trace sample attached,
>> 484 but the tool validates that the sample size is not greater than 60KiB.
>> 485
>> 486
>> 487 new snapshot option
>> 488 -------------------
>> 489
>> 490 The difference between full trace and snapshot from the kernel's perspective is
>> 491 that in full trace we don't overwrite trace data that the user hasn't collected
>> 492 yet (and indicated that by advancing aux_tail), whereas in snapshot mode we let
>> 493 the trace run and overwrite older data in the buffer so that whenever something
>> 494 interesting happens, we can stop it and grab a snapshot of what was going on
>> 495 around that interesting moment.
>> 496
>> 497 To select snapshot mode a new option has been added:
>> 498
>> 499 -S
>> 500
>> 501 Optionally it can be followed by the snapshot size e.g.
>> 502
>> 503 -S0x100000
>> 504
>> 505 The default snapshot size is the auxtrace mmap size. If neither auxtrace mmap size
>> 506 nor snapshot size is specified, then the default is 4MiB for privileged users
>> 507 (or if /proc/sys/kernel/perf_event_paranoid < 0), 128KiB for unprivileged users.
>> 508 If an unprivileged user does not specify mmap pages, the mmap pages will be
>> 509 reduced as described in the 'new auxtrace mmap size option' section below.
>> 510
>> 511 The snapshot size is displayed if the option -vv is used e.g.
>> 512
>> 513 Intel PT snapshot size: %zu
>> 514
>> 515
>> 516 new auxtrace mmap size option
>> 517 ---------------------------
>> 518
>> 519 Intel PT buffer size is specified by an addition to the -m option e.g.
>> 520
>> 521 -m,16
>> 522
>> 523 selects a buffer size of 16 pages i.e. 64KiB.
>> 524
>> 525 Note that the existing functionality of -m is unchanged. The auxtrace mmap size
>> 526 is specified by the optional addition of a comma and the value.
>> 527
>> 528 The default auxtrace mmap size for Intel PT is 4MiB/page_size for privileged users
>> 529 (or if /proc/sys/kernel/perf_event_paranoid < 0), 128KiB for unprivileged users.
>> 530 If an unprivileged user does not specify mmap pages, the mmap pages will be
>> 531 reduced from the default 512KiB/page_size to 256KiB/page_size, otherwise the
>> 532 user is likely to get an error as they exceed their mlock limit (Max locked
>> 533 memory as shown in /proc/self/limits). Note that perf does not count the first
>> 534 512KiB (actually /proc/sys/kernel/perf_event_mlock_kb minus 1 page) per cpu
>> 535 against the mlock limit so an unprivileged user is allowed 512KiB per cpu plus
>> 536 their mlock limit (which defaults to 64KiB but is not multiplied by the number
>> 537 of cpus).
>> 538
>> 539 In full-trace mode, powers of two are allowed for buffer size, with a minimum
>> 540 size of 2 pages. In snapshot mode or sampling mode, it is the same but the
>> 541 minimum size is 1 page.
>> 542
>> 543 The mmap size and auxtrace mmap size are displayed if the -vv option is used e.g.
>> 544
>> 545 mmap length 528384
>> 546 auxtrace mmap length 4198400
>> 547
>> 548
>> 549 Intel PT modes of operation
>> 550 ---------------------------
>> 551
>> 552 Intel PT can be used in 2 modes:
>> 553 full-trace mode
>> 554 sample mode
>> 555 snapshot mode
>> 556
>> 557 Full-trace mode traces continuously e.g.
>> 558
>> 559 perf record -e intel_pt//u uname
>> 560
>> 561 Sample mode attaches a Intel PT sample to other events e.g.
>> 562
>> 563 perf record --aux-sample -e intel_pt//u -e branch-misses:u
>> 564
>> 565 Snapshot mode captures the available data when a signal is sent e.g.
>> 566
>> 567 perf record -v -e intel_pt//u -S ./loopy 1000000000 &
>> 568 [1] 11435
>> 569 kill -USR2 11435
>> 570 Recording AUX area tracing snapshot
>> 571
>> 572 Note that the signal sent is SIGUSR2.
>> 573 Note that "Recording AUX area tracing snapshot" is displayed because the -v
>> 574 option is used.
>> 575
>> 576 The 2 modes cannot be used together.
>> 577
>> 578
>> 579 Buffer handling
>> 580 ---------------
>> 581
>> 582 There may be buffer limitations (i.e. single ToPa entry) which means that actual
>> 583 buffer sizes are limited to powers of 2 up to 4MiB (MAX_ORDER). In order to
>> 584 provide other sizes, and in particular an arbitrarily large size, multiple
>> 585 buffers are logically concatenated. However an interrupt must be used to switch
>> 586 between buffers. That has two potential problems:
>> 587 a) the interrupt may not be handled in time so that the current buffer
>> 588 becomes full and some trace data is lost.
>> 589 b) the interrupts may slow the system and affect the performance
>> 590 results.
>> 591
>> 592 If trace data is lost, the driver sets 'truncated' in the PERF_RECORD_AUX event
>> 593 which the tools report as an error.
>> 594
>> 595 In full-trace mode, the driver waits for data to be copied out before allowing
>> 596 the (logical) buffer to wrap-around. If data is not copied out quickly enough,
>> 597 again 'truncated' is set in the PERF_RECORD_AUX event. If the driver has to
>> 598 wait, the intel_pt event gets disabled. Because it is difficult to know when
>> 599 that happens, perf tools always re-enable the intel_pt event after copying out
>> 600 data.
>> 601
>> 602
>> 603 Intel PT and build ids
>> 604 ----------------------
>> 605
>> 606 By default "perf record" post-processes the event stream to find all build ids
>> 607 for executables for all addresses sampled. Deliberately, Intel PT is not
>> 608 decoded for that purpose (it would take too long). Instead the build ids for
>> 609 all executables encountered (due to mmap, comm or task events) are included
>> 610 in the perf.data file.
>> 611
>> 612 To see buildids included in the perf.data file use the command:
>> 613
>> 614 perf buildid-list
>> 615
>> 616 If the perf.data file contains Intel PT data, that is the same as:
>> 617
>> 618 perf buildid-list --with-hits
>> 619
>> 620
>> 621 Snapshot mode and event disabling
>> 622 ---------------------------------
>> 623
>> 624 In order to make a snapshot, the intel_pt event is disabled using an IOCTL,
>> 625 namely PERF_EVENT_IOC_DISABLE. However doing that can also disable the
>> 626 collection of side-band information. In order to prevent that, a dummy
>> 627 software event has been introduced that permits tracking events (like mmaps) to
>> 628 continue to be recorded while intel_pt is disabled. That is important to ensure
>> 629 there is complete side-band information to allow the decoding of subsequent
>> 630 snapshots.
>> 631
>> 632 A test has been created for that. To find the test:
>> 633
>> 634 perf test list
>> 635 ...
>> 636 23: Test using a dummy software event to keep tracking
>> 637
>> 638 To run the test:
>> 639
>> 640 perf test 23
>> 641 23: Test using a dummy software event to keep tracking : Ok
>> 642
>> 643
>> 644 perf record modes (nothing new here)
>> 645 ------------------------------------
>> 646
>> 647 perf record essentially operates in one of three modes:
>> 648 per thread
>> 649 per cpu
>> 650 workload only
>> 651
>> 652 "per thread" mode is selected by -t or by --per-thread (with -p or -u or just a
>> 653 workload).
>> 654 "per cpu" is selected by -C or -a.
>> 655 "workload only" mode is selected by not using the other options but providing a
>> 656 command to run (i.e. the workload).
>> 657
>> 658 In per-thread mode an exact list of threads is traced. There is no inheritance.
>> 659 Each thread has its own event buffer.
>> 660
>> 661 In per-cpu mode all processes (or processes from the selected cgroup i.e. -G
>> 662 option, or processes selected with -p or -u) are traced. Each cpu has its own
>> 663 buffer. Inheritance is allowed.
>> 664
>> 665 In workload-only mode, the workload is traced but with per-cpu buffers.
>> 666 Inheritance is allowed. Note that you can now trace a workload in per-thread
>> 667 mode by using the --per-thread option.
>> 668
>> 669
>> 670 Privileged vs non-privileged users
>> 671 ----------------------------------
>> 672
>> 673 Unless /proc/sys/kernel/perf_event_paranoid is set to -1, unprivileged users
>> 674 have memory limits imposed upon them. That affects what buffer sizes they can
>> 675 have as outlined above.
>> 676
>> 677 The v4.2 kernel introduced support for a context switch metadata event,
>> 678 PERF_RECORD_SWITCH, which allows unprivileged users to see when their processes
>> 679 are scheduled out and in, just not by whom, which is left for the
>> 680 PERF_RECORD_SWITCH_CPU_WIDE, that is only accessible in system wide context,
>> 681 which in turn requires CAP_SYS_ADMIN.
>> 682
>> 683 Please see the 45ac1403f564 ("perf: Add PERF_RECORD_SWITCH to indicate context
>> 684 switches") commit, that introduces these metadata events for further info.
>> 685
>> 686 When working with kernels < v4.2, the following considerations must be taken,
>> 687 as the sched:sched_switch tracepoints will be used to receive such information:
>> 688
>> 689 Unless /proc/sys/kernel/perf_event_paranoid is set to -1, unprivileged users are
>> 690 not permitted to use tracepoints which means there is insufficient side-band
>> 691 information to decode Intel PT in per-cpu mode, and potentially workload-only
>> 692 mode too if the workload creates new processes.
>> 693
>> 694 Note also, that to use tracepoints, read-access to debugfs is required. So if
>> 695 debugfs is not mounted or the user does not have read-access, it will again not
>> 696 be possible to decode Intel PT in per-cpu mode.
>> 697
>> 698
>> 699 sched_switch tracepoint
>> 700 -----------------------
>> 701
>> 702 The sched_switch tracepoint is used to provide side-band data for Intel PT
>> 703 decoding in kernels where the PERF_RECORD_SWITCH metadata event isn't
>> 704 available.
>> 705
>> 706 The sched_switch events are automatically added. e.g. the second event shown
>> 707 below:
>> 708
>> 709 $ perf record -vv -e intel_pt//u uname
>> 710 ------------------------------------------------------------
>> 711 perf_event_attr:
>> 712 type 6
>> 713 size 112
>> 714 config 0x400
>> 715 { sample_period, sample_freq } 1
>> 716 sample_type IP|TID|TIME|CPU|IDENTIFIER
>> 717 read_format ID
>> 718 disabled 1
>> 719 inherit 1
>> 720 exclude_kernel 1
>> 721 exclude_hv 1
>> 722 enable_on_exec 1
>> 723 sample_id_all 1
>> 724 ------------------------------------------------------------
>> 725 sys_perf_event_open: pid 31104 cpu 0 group_fd -1 flags 0x8
>> 726 sys_perf_event_open: pid 31104 cpu 1 group_fd -1 flags 0x8
>> 727 sys_perf_event_open: pid 31104 cpu 2 group_fd -1 flags 0x8
>> 728 sys_perf_event_open: pid 31104 cpu 3 group_fd -1 flags 0x8
>> 729 ------------------------------------------------------------
>> 730 perf_event_attr:
>> 731 type 2
>> 732 size 112
>> 733 config 0x108
>> 734 { sample_period, sample_freq } 1
>> 735 sample_type IP|TID|TIME|CPU|PERIOD|RAW|IDENTIFIER
>> 736 read_format ID
>> 737 inherit 1
>> 738 sample_id_all 1
>> 739 exclude_guest 1
>> 740 ------------------------------------------------------------
>> 741 sys_perf_event_open: pid -1 cpu 0 group_fd -1 flags 0x8
>> 742 sys_perf_event_open: pid -1 cpu 1 group_fd -1 flags 0x8
>> 743 sys_perf_event_open: pid -1 cpu 2 group_fd -1 flags 0x8
>> 744 sys_perf_event_open: pid -1 cpu 3 group_fd -1 flags 0x8
>> 745 ------------------------------------------------------------
>> 746 perf_event_attr:
>> 747 type 1
>> 748 size 112
>> 749 config 0x9
>> 750 { sample_period, sample_freq } 1
>> 751 sample_type IP|TID|TIME|IDENTIFIER
>> 752 read_format ID
>> 753 disabled 1
>> 754 inherit 1
>> 755 exclude_kernel 1
>> 756 exclude_hv 1
>> 757 mmap 1
>> 758 comm 1
>> 759 enable_on_exec 1
>> 760 task 1
>> 761 sample_id_all 1
>> 762 mmap2 1
>> 763 comm_exec 1
>> 764 ------------------------------------------------------------
>> 765 sys_perf_event_open: pid 31104 cpu 0 group_fd -1 flags 0x8
>> 766 sys_perf_event_open: pid 31104 cpu 1 group_fd -1 flags 0x8
>> 767 sys_perf_event_open: pid 31104 cpu 2 group_fd -1 flags 0x8
>> 768 sys_perf_event_open: pid 31104 cpu 3 group_fd -1 flags 0x8
>> 769 mmap size 528384B
>> 770 AUX area mmap length 4194304
>> 771 perf event ring buffer mmapped per cpu
>> 772 Synthesizing auxtrace information
>> 773 Linux
>> 774 [ perf record: Woken up 1 times to write data ]
>> 775 [ perf record: Captured and wrote 0.042 MB perf.data ]
>> 776
>> 777 Note, the sched_switch event is only added if the user is permitted to use it
>> 778 and only in per-cpu mode.
>> 779
>> 780 Note also, the sched_switch event is only added if TSC packets are requested.
>> 781 That is because, in the absence of timing information, the sched_switch events
>> 782 cannot be matched against the Intel PT trace.
>> 783
>> 784
>> 785 perf script
>> 786 ===========
>> 787
>> 788 By default, perf script will decode trace data found in the perf.data file.
>> 789 This can be further controlled by new option --itrace.
>> 790
>> 791
>> 792 New --itrace option
>> 793 -------------------
>> 794
>> 795 Having no option is the same as
>> 796
>> 797 --itrace
>> 798
>> 799 which, in turn, is the same as
>> 800
>> 801 --itrace=cepwx
>> 802
>> 803 The letters are:
>> 804
>> 805 i synthesize "instructions" events
>> 806 b synthesize "branches" events
>> 807 x synthesize "transactions" events
>> 808 w synthesize "ptwrite" events
>> 809 p synthesize "power" events
>> 810 c synthesize branches events (calls only)
>> 811 r synthesize branches events (returns only)
>> 812 e synthesize tracing error events
>> 813 d create a debug log
>> 814 g synthesize a call chain (use with i or x)
>> 815 l synthesize last branch entries (use with i or x)
>> 816 s skip initial number of events
>> 817
>> 818 "Instructions" events look like they were recorded by "perf record -e
>> 819 instructions".
>> 820
>> 821 "Branches" events look like they were recorded by "perf record -e branches". "c"
>> 822 and "r" can be combined to get calls and returns.
>> 823
>> 824 "Transactions" events correspond to the start or end of transactions. The
>> 825 'flags' field can be used in perf script to determine whether the event is a
>> 826 tranasaction start, commit or abort.
>> 827
>> 828 Note that "instructions", "branches" and "transactions" events depend on code
>> 829 flow packets which can be disabled by using the config term "branch=0". Refer
>> 830 to the config terms section above.
>> 831
>> 832 "ptwrite" events record the payload of the ptwrite instruction and whether
>> 833 "fup_on_ptw" was used. "ptwrite" events depend on PTWRITE packets which are
>> 834 recorded only if the "ptw" config term was used. Refer to the config terms
>> 835 section above. perf script "synth" field displays "ptwrite" information like
>> 836 this: "ip: 0 payload: 0x123456789abcdef0" where "ip" is 1 if "fup_on_ptw" was
>> 837 used.
>> 838
>> 839 "Power" events correspond to power event packets and CBR (core-to-bus ratio)
>> 840 packets. While CBR packets are always recorded when tracing is enabled, power
>> 841 event packets are recorded only if the "pwr_evt" config term was used. Refer to
>> 842 the config terms section above. The power events record information about
>> 843 C-state changes, whereas CBR is indicative of CPU frequency. perf script
>> 844 "event,synth" fields display information like this:
>> 845 cbr: cbr: 22 freq: 2189 MHz (200%)
>> 846 mwait: hints: 0x60 extensions: 0x1
>> 847 pwre: hw: 0 cstate: 2 sub-cstate: 0
>> 848 exstop: ip: 1
>> 849 pwrx: deepest cstate: 2 last cstate: 2 wake reason: 0x4
>> 850 Where:
>> 851 "cbr" includes the frequency and the percentage of maximum non-turbo
>> 852 "mwait" shows mwait hints and extensions
>> 853 "pwre" shows C-state transitions (to a C-state deeper than C0) and
>> 854 whether initiated by hardware
>> 855 "exstop" indicates execution stopped and whether the IP was recorded
>> 856 exactly,
>> 857 "pwrx" indicates return to C0
>> 858 For more details refer to the Intel 64 and IA-32 Architectures Software
>> 859 Developer Manuals.
>> 860
>> 861 Error events show where the decoder lost the trace. Error events
>> 862 are quite important. Users must know if what they are seeing is a complete
>> 863 picture or not.
>> 864
>> 865 The "d" option will cause the creation of a file "intel_pt.log" containing all
>> 866 decoded packets and instructions. Note that this option slows down the decoder
>> 867 and that the resulting file may be very large.
>> 868
>> 869 In addition, the period of the "instructions" event can be specified. e.g.
>> 870
>> 871 --itrace=i10us
>> 872
>> 873 sets the period to 10us i.e. one instruction sample is synthesized for each 10
>> 874 microseconds of trace. Alternatives to "us" are "ms" (milliseconds),
>> 875 "ns" (nanoseconds), "t" (TSC ticks) or "i" (instructions).
>> 876
>> 877 "ms", "us" and "ns" are converted to TSC ticks.
>> 878
>> 879 The timing information included with Intel PT does not give the time of every
>> 880 instruction. Consequently, for the purpose of sampling, the decoder estimates
>> 881 the time since the last timing packet based on 1 tick per instruction. The time
>> 882 on the sample is *not* adjusted and reflects the last known value of TSC.
>> 883
>> 884 For Intel PT, the default period is 100us.
>> 885
>> 886 Setting it to a zero period means "as often as possible".
>> 887
>> 888 In the case of Intel PT that is the same as a period of 1 and a unit of
>> 889 'instructions' (i.e. --itrace=i1i).
>> 890
>> 891 Also the call chain size (default 16, max. 1024) for instructions or
>> 892 transactions events can be specified. e.g.
>> 893
>> 894 --itrace=ig32
>> 895 --itrace=xg32
>> 896
>> 897 Also the number of last branch entries (default 64, max. 1024) for instructions or
>> 898 transactions events can be specified. e.g.
>> 899
>> 900 --itrace=il10
>> 901 --itrace=xl10
>> 902
>> 903 Note that last branch entries are cleared for each sample, so there is no overlap
>> 904 from one sample to the next.
>> 905
>> 906 To disable trace decoding entirely, use the option --no-itrace.
>> 907
>> 908 It is also possible to skip events generated (instructions, branches, transactions)
>> 909 at the beginning. This is useful to ignore initialization code.
>> 910
>> 911 --itrace=i0nss1000000
>> 912
>> 913 skips the first million instructions.
>> 914
>> 915 dump option
>> 916 -----------
>> 917
>> 918 perf script has an option (-D) to "dump" the events i.e. display the binary
>> 919 data.
>> 920
>> 921 When -D is used, Intel PT packets are displayed. The packet decoder does not
>> 922 pay attention to PSB packets, but just decodes the bytes - so the packets seen
>> 923 by the actual decoder may not be identical in places where the data is corrupt.
>> 924 One example of that would be when the buffer-switching interrupt has been too
>> 925 slow, and the buffer has been filled completely. In that case, the last packet
>> 926 in the buffer might be truncated and immediately followed by a PSB as the trace
>> 927 continues in the next buffer.
>> 928
>> 929 To disable the display of Intel PT packets, combine the -D option with
>> 930 --no-itrace.
>> 931
>> 932
>> 933 perf report
>> 934 ===========
>> 935
>> 936 By default, perf report will decode trace data found in the perf.data file.
>> 937 This can be further controlled by new option --itrace exactly the same as
>> 938 perf script, with the exception that the default is --itrace=igxe.
>> 939
>> 940
>> 941 perf inject
>> 942 ===========
>> 943
>> 944 perf inject also accepts the --itrace option in which case tracing data is
>> 945 removed and replaced with the synthesized events. e.g.
>> 946
>> 947 perf inject --itrace -i perf.data -o perf.data.new
>> 948
>> 949 Below is an example of using Intel PT with autofdo. It requires autofdo
>> 950 (https://github.com/google/autofdo) and gcc version 5. The bubble
>> 951 sort example is from the AutoFDO tutorial (https://gcc.gnu.org/wiki/AutoFDO/Tutorial)
>> 952 amended to take the number of elements as a parameter.
>> 953
>> 954 $ gcc-5 -O3 sort.c -o sort_optimized
>> 955 $ ./sort_optimized 30000
>> 956 Bubble sorting array of 30000 elements
>> 957 2254 ms
>> 958
>> 959 $ cat ~/.perfconfig
>> 960 [intel-pt]
>> 961 mispred-all = on
>> 962
>> 963 $ perf record -e intel_pt//u ./sort 3000
>> 964 Bubble sorting array of 3000 elements
>> 965 58 ms
>> 966 [ perf record: Woken up 2 times to write data ]
>> 967 [ perf record: Captured and wrote 3.939 MB perf.data ]
>> 968 $ perf inject -i perf.data -o inj --itrace=i100usle --strip
>> 969 $ ./create_gcov --binary=./sort --profile=inj --gcov=sort.gcov -gcov_version=1
>> 970 $ gcc-5 -O3 -fauto-profile=sort.gcov sort.c -o sort_autofdo
>> 971 $ ./sort_autofdo 30000
>> 972 Bubble sorting array of 30000 elements
>> 973 2155 ms
>> 974
>> 975 Note there is currently no advantage to using Intel PT instead of LBR, but
>> 976 that may change in the future if greater use is made of the data.
>> 977
>> 978
>> 979 PEBS via Intel PT
>> 980 =================
>> 981
>> 982 Some hardware has the feature to redirect PEBS records to the Intel PT trace.
>> 983 Recording is selected by using the aux-output config term e.g.
>> 984
>> 985 perf record -c 10000 -e '{intel_pt/branch=0/,cycles/aux-output/ppp}' uname
>> 986
>> 987 Note that currently, software only supports redirecting at most one PEBS event.
>> 988
>> 989 To display PEBS events from the Intel PT trace, use the itrace 'o' option e.g.
>> 990
>> 991 perf script --itrace=oe
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.