Intel Processor Trace ===================== Overview ======== Intel Processor Trace (Intel PT) is an extension of Intel Architecture that collects information about software execution such as control flow, execution modes and timings and formats it into highly compressed binary packets. Technical details are documented in the Intel 64 and IA-32 Architectures Software Developer Manuals, Chapter 36 Intel Processor Trace. Intel PT is first supported in Intel Core M and 5th generation Intel Core processors that are based on the Intel micro-architecture code name Broadwell. Trace data is collected by 'perf record' and stored within the perf.data file. See below for options to 'perf record'. Trace data must be 'decoded' which involves walking the object code and matching the trace data packets. For example a TNT packet only tells whether a conditional branch was taken or not taken, so to make use of that packet the decoder must know precisely which instruction was being executed. Decoding is done on-the-fly. The decoder outputs samples in the same format as samples output by perf hardware events, for example as though the "instructions" or "branches" events had been recorded. Presently 3 tools support this: 'perf script', 'perf report' and 'perf inject'. See below for more information on using those tools. The main distinguishing feature of Intel PT is that the decoder can determine the exact flow of software execution. Intel PT can be used to understand why and how did software get to a certain point, or behave a certain way. The software does not have to be recompiled, so Intel PT works with debug or release builds, however the executed images are needed - which makes use in JIT-compiled environments, or with self-modified code, a challenge. Also symbols need to be provided to make sense of addresses. A limitation of Intel PT is that it produces huge amounts of trace data (hundreds of megabytes per second per core) which takes a long time to decode, for example two or three orders of magnitude longer than it took to collect. Another limitation is the performance impact of tracing, something that will vary depending on the use-case and architecture. Quickstart ========== It is important to start small. That is because it is easy to capture vastly more data than can possibly be processed. The simplest thing to do with Intel PT is userspace profiling of small programs. Data is captured with 'perf record' e.g. to trace 'ls' userspace-only: perf record -e intel_pt//u ls And profiled with 'perf report' e.g. perf report To also trace kernel space presents a problem, namely kernel self-modifying code. A fairly good kernel image is available in /proc/kcore but to get an accurate image a copy of /proc/kcore needs to be made under the same conditions as the data capture. A script perf-with-kcore can do that, but beware that the script makes use of 'sudo' to copy /proc/kcore. If you have perf installed locally from the source tree you can do: ~/libexec/perf-core/perf-with-kcore record pt_ls -e intel_pt// -- ls which will create a directory named 'pt_ls' and put the perf.data file and copies of /proc/kcore, /proc/kallsyms and /proc/modules into it. Then to use 'perf report' becomes: ~/libexec/perf-core/perf-with-kcore report pt_ls Because samples are synthesized after-the-fact, the sampling period can be selected for reporting. e.g. sample every microsecond ~/libexec/perf-core/perf-with-kcore report pt_ls --itrace=i1usge See the sections below for more information about the --itrace option. Beware the smaller the period, the more samples that are produced, and the longer it takes to process them. Also note that the coarseness of Intel PT timing information will start to distort the statistical value of the sampling as the sampling period becomes smaller. To represent software control flow, "branches" samples are produced. By default a branch sample is synthesized for every single branch. To get an idea what data is available you can use the 'perf script' tool with no parameters, which will list all the samples. perf record -e intel_pt//u ls perf script An interesting field that is not printed by default is 'flags' which can be displayed as follows: perf script -Fcomm,tid,pid,time,cpu,event,trace,ip,sym,dso,addr,symoff,flags The flags are "bcrosyiABEx" which stand for branch, call, return, conditional, system, asynchronous, interrupt, transaction abort, trace begin, trace end, and in transaction, respectively. While it is possible to create scripts to analyze the data, an alternative approach is available to export the data to a postgresql database. Refer to script export-to-postgresql.py for more details, and to script call-graph-from-postgresql.py for an example of using the database. As mentioned above, it is easy to capture too much data. One way to limit the data captured is to use 'snapshot' mode which is explained further below. Refer to 'new snapshot option' and 'Intel PT modes of operation' further below. Another problem that will be experienced is decoder errors. They can be caused by inability to access the executed image, self-modified or JIT-ed code, or the inability to match side-band information (such as context switches and mmaps) which results in the decoder not knowing what code was executed. There is also the problem of perf not being able to copy the data fast enough, resulting in data lost because the buffer was full. See 'Buffer handling' below for more details. perf record =========== new event --------- The Intel PT kernel driver creates a new PMU for Intel PT. PMU events are selected by providing the PMU name followed by the "config" separated by slashes. An enhancement has been made to allow default "config" e.g. the option -e intel_pt// will use a default config value. Currently that is the same as -e intel_pt/tsc,noretcomp=0/ which is the same as -e intel_pt/tsc=1,noretcomp=0/ Note there are now new config terms - see section 'config terms' further below. The config terms are listed in /sys/devices/intel_pt/format. They are bit fields within the config member of the struct perf_event_attr which is passed to the kernel by the perf_event_open system call. They correspond to bit fields in the IA32_RTIT_CTL MSR. Here is a list of them and their definitions: $ grep -H . /sys/bus/event_source/devices/intel_pt/format/* /sys/bus/event_source/devices/intel_pt/format/cyc:config:1 /sys/bus/event_source/devices/intel_pt/format/cyc_thresh:config:19-22 /sys/bus/event_source/devices/intel_pt/format/mtc:config:9 /sys/bus/event_source/devices/intel_pt/format/mtc_period:config:14-17 /sys/bus/event_source/devices/intel_pt/format/noretcomp:config:11 /sys/bus/event_source/devices/intel_pt/format/psb_period:config:24-27 /sys/bus/event_source/devices/intel_pt/format/tsc:config:10 Note that the default config must be overridden for each term i.e. -e intel_pt/noretcomp=0/ is the same as: -e intel_pt/tsc=1,noretcomp=0/ So, to disable TSC packets use: -e intel_pt/tsc=0/ It is also possible to specify the config value explicitly: -e intel_pt/config=0x400/ Note that, as with all events, the event is suffixed with event modifiers: u userspace k kernel h hypervisor G guest H host p precise ip 'h', 'G' and 'H' are for virtualization which is not supported by Intel PT. 'p' is also not relevant to Intel PT. So only options 'u' and 'k' are meaningful for Intel PT. perf_event_attr is displayed if the -vv option is used e.g. ------------------------------------------------------------ perf_event_attr: type 6 size 112 config 0x400 { sample_period, sample_freq } 1 sample_type IP|TID|TIME|CPU|IDENTIFIER read_format ID disabled 1 inherit 1 exclude_kernel 1 exclude_hv 1 enable_on_exec 1 sample_id_all 1 ------------------------------------------------------------ sys_perf_event_open: pid 31104 cpu 0 group_fd -1 flags 0x8 sys_perf_event_open: pid 31104 cpu 1 group_fd -1 flags 0x8 sys_perf_event_open: pid 31104 cpu 2 group_fd -1 flags 0x8 sys_perf_event_open: pid 31104 cpu 3 group_fd -1 flags 0x8 ------------------------------------------------------------ config terms ------------ The June 2015 version of Intel 64 and IA-32 Architectures Software Developer Manuals, Chapter 36 Intel Processor Trace, defined new Intel PT features. Some of the features are reflect in new config terms. All the config terms are described below. tsc Always supported. Produces TSC timestamp packets to provide timing information. In some cases it is possible to decode without timing information, for example a per-thread context that does not overlap executable memory maps. The default config selects tsc (i.e. tsc=1). noretcomp Always supported. Disables "return compression" so a TIP packet is produced when a function returns. Causes more packets to be produced but might make decoding more reliable. The default config does not select noretcomp (i.e. noretcomp=0). psb_period Allows the frequency of PSB packets to be specified. The PSB packet is a synchronization packet that provides a starting point for decoding or recovery from errors. Support for psb_period is indicated by: /sys/bus/event_source/devices/intel_pt/caps/psb_cyc which contains "1" if the feature is supported and "0" otherwise. Valid values are given by: /sys/bus/event_source/devices/intel_pt/caps/psb_periods which contains a hexadecimal value, the bits of which represent valid values e.g. bit 2 set means value 2 is valid. The psb_period value is converted to the approximate number of trace bytes between PSB packets as: 2 ^ (value + 11) e.g. value 3 means 16KiB bytes between PSBs If an invalid value is entered, the error message will give a list of valid values e.g. $ perf record -e intel_pt/psb_period=15/u uname Invalid psb_period for intel_pt. Valid values are: 0-5 If MTC packets are selected, the default config selects a value of 3 (i.e. psb_period=3) or the nearest lower value that is supported (0 is always supported). Otherwise the default is 0. If decoding is expected to be reliable and the buffer is large then a large PSB period can be used. Because a TSC packet is produced with PSB, the PSB period can also affect the granularity to timing information in the absence of MTC or CYC. mtc Produces MTC timing packets. MTC packets provide finer grain timestamp information than TSC packets. MTC packets record time using the hardware crystal clock (CTC) which is related to TSC packets using a TMA packet. Support for this feature is indicated by: /sys/bus/event_source/devices/intel_pt/caps/mtc which contains "1" if the feature is supported and "0" otherwise. The frequency of MTC packets can also be specified - see mtc_period below. mtc_period Specifies how frequently MTC packets are produced - see mtc above for how to determine if MTC packets are supported. Valid values are given by: /sys/bus/event_source/devices/intel_pt/caps/mtc_periods which contains a hexadecimal value, the bits of which represent valid values e.g. bit 2 set means value 2 is valid. The mtc_period value is converted to the MTC frequency as: CTC-frequency / (2 ^ value) e.g. value 3 means one eighth of CTC-frequency Where CTC is the hardware crystal clock, the frequency of which can be related to TSC via values provided in cpuid leaf 0x15. If an invalid value is entered, the error message will give a list of valid values e.g. $ perf record -e intel_pt/mtc_period=15/u uname Invalid mtc_period for intel_pt. Valid values are: 0,3,6,9 The default value is 3 or the nearest lower value that is supported (0 is always supported). cyc Produces CYC timing packets. CYC packets provide even finer grain timestamp information than MTC and TSC packets. A CYC packet contains the number of CPU cycles since the last CYC packet. Unlike MTC and TSC packets, CYC packets are only sent when another packet is also sent. Support for this feature is indicated by: /sys/bus/event_source/devices/intel_pt/caps/psb_cyc which contains "1" if the feature is supported and "0" otherwise. The number of CYC packets produced can be reduced by specifying a threshold - see cyc_thresh below. cyc_thresh Specifies how frequently CYC packets are produced - see cyc above for how to determine if CYC packets are supported. Valid cyc_thresh values are given by: /sys/bus/event_source/devices/intel_pt/caps/cycle_thresholds which contains a hexadecimal value, the bits of which represent valid values e.g. bit 2 set means value 2 is valid. The cyc_thresh value represents the minimum number of CPU cycles that must have passed before a CYC packet can be sent. The number of CPU cycles is: 2 ^ (value - 1) e.g. value 4 means 8 CPU cycles must pass before a CYC packet can be sent. Note a CYC packet is still only sent when another packet is sent, not at, e.g. every 8 CPU cycles. If an invalid value is entered, the error message will give a list of valid values e.g. $ perf record -e intel_pt/cyc,cyc_thresh=15/u uname Invalid cyc_thresh for intel_pt. Valid values are: 0-12 CYC packets are not requested by default. new snapshot option ------------------- The difference between full trace and snapshot from the kernel's perspective is that in full trace we don't overwrite trace data that the user hasn't collected yet (and indicated that by advancing aux_tail), whereas in snapshot mode we let the trace run and overwrite older data in the buffer so that whenever something interesting happens, we can stop it and grab a snapshot of what was going on around that interesting moment. To select snapshot mode a new option has been added: -S Optionally it can be followed by the snapshot size e.g. -S0x100000 The default snapshot size is the auxtrace mmap size. If neither auxtrace mmap size nor snapshot size is specified, then the default is 4MiB for privileged users (or if /proc/sys/kernel/perf_event_paranoid < 0), 128KiB for unprivileged users. If an unprivileged user does not specify mmap pages, the mmap pages will be reduced as described in the 'new auxtrace mmap size option' section below. The snapshot size is displayed if the option -vv is used e.g. Intel PT snapshot size: %zu new auxtrace mmap size option --------------------------- Intel PT buffer size is specified by an addition to the -m option e.g. -m,16 selects a buffer size of 16 pages i.e. 64KiB. Note that the existing functionality of -m is unchanged. The auxtrace mmap size is specified by the optional addition of a comma and the value. The default auxtrace mmap size for Intel PT is 4MiB/page_size for privileged users (or if /proc/sys/kernel/perf_event_paranoid < 0), 128KiB for unprivileged users. If an unprivileged user does not specify mmap pages, the mmap pages will be reduced from the default 512KiB/page_size to 256KiB/page_size, otherwise the user is likely to get an error as they exceed their mlock limit (Max locked memory as shown in /proc/self/limits). Note that perf does not count the first 512KiB (actually /proc/sys/kernel/perf_event_mlock_kb minus 1 page) per cpu against the mlock limit so an unprivileged user is allowed 512KiB per cpu plus their mlock limit (which defaults to 64KiB but is not multiplied by the number of cpus). In full-trace mode, powers of two are allowed for buffer size, with a minimum size of 2 pages. In snapshot mode, it is the same but the minimum size is 1 page. The mmap size and auxtrace mmap size are displayed if the -vv option is used e.g. mmap length 528384 auxtrace mmap length 4198400 Intel PT modes of operation --------------------------- Intel PT can be used in 2 modes: full-trace mode snapshot mode Full-trace mode traces continuously e.g. perf record -e intel_pt//u uname Snapshot mode captures the available data when a signal is sent e.g. perf record -v -e intel_pt//u -S ./loopy 1000000000 & [1] 11435 kill -USR2 11435 Recording AUX area tracing snapshot Note that the signal sent is SIGUSR2. Note that "Recording AUX area tracing snapshot" is displayed because the -v option is used. The 2 modes cannot be used together. Buffer handling --------------- There may be buffer limitations (i.e. single ToPa entry) which means that actual buffer sizes are limited to powers of 2 up to 4MiB (MAX_ORDER). In order to provide other sizes, and in particular an arbitrarily large size, multiple buffers are logically concatenated. However an interrupt must be used to switch between buffers. That has two potential problems: a) the interrupt may not be handled in time so that the current buffer becomes full and some trace data is lost. b) the interrupts may slow the system and affect the performance results. If trace data is lost, the driver sets 'truncated' in the PERF_RECORD_AUX event which the tools report as an error. In full-trace mode, the driver waits for data to be copied out before allowing the (logical) buffer to wrap-around. If data is not copied out quickly enough, again 'truncated' is set in the PERF_RECORD_AUX event. If the driver has to wait, the intel_pt event gets disabled. Because it is difficult to know when that happens, perf tools always re-enable the intel_pt event after copying out data. Intel PT and build ids ---------------------- By default "perf record" post-processes the event stream to find all build ids for executables for all addresses sampled. Deliberately, Intel PT is not decoded for that purpose (it would take too long). Instead the build ids for all executables encountered (due to mmap, comm or task events) are included in the perf.data file. To see buildids included in the perf.data file use the command: perf buildid-list If the perf.data file contains Intel PT data, that is the same as: perf buildid-list --with-hits Snapshot mode and event disabling --------------------------------- In order to make a snapshot, the intel_pt event is disabled using an IOCTL, namely PERF_EVENT_IOC_DISABLE. However doing that can also disable the collection of side-band information. In order to prevent that, a dummy software event has been introduced that permits tracking events (like mmaps) to continue to be recorded while intel_pt is disabled. That is important to ensure there is complete side-band information to allow the decoding of subsequent snapshots. A test has been created for that. To find the test: perf test list ... 23: Test using a dummy software event to keep tracking To run the test: perf test 23 23: Test using a dummy software event to keep tracking : Ok perf record modes (nothing new here) ------------------------------------ perf record essentially operates in one of three modes: per thread per cpu workload only "per thread" mode is selected by -t or by --per-thread (with -p or -u or just a workload). "per cpu" is selected by -C or -a. "workload only" mode is selected by not using the other options but providing a command to run (i.e. the workload). In per-thread mode an exact list of threads is traced. There is no inheritance. Each thread has its own event buffer. In per-cpu mode all processes (or processes from the selected cgroup i.e. -G option, or processes selected with -p or -u) are traced. Each cpu has its own buffer. Inheritance is allowed. In workload-only mode, the workload is traced but with per-cpu buffers. Inheritance is allowed. Note that you can now trace a workload in per-thread mode by using the --per-thread option. Privileged vs non-privileged users ---------------------------------- Unless /proc/sys/kernel/perf_event_paranoid is set to -1, unprivileged users have memory limits imposed upon them. That affects what buffer sizes they can have as outlined above. The v4.2 kernel introduced support for a context switch metadata event, PERF_RECORD_SWITCH, which allows unprivileged users to see when their processes are scheduled out and in, just not by whom, which is left for the PERF_RECORD_SWITCH_CPU_WIDE, that is only accessible in system wide context, which in turn requires CAP_SYS_ADMIN. Please see the 45ac1403f564 ("perf: Add PERF_RECORD_SWITCH to indicate context switches") commit, that introduces these metadata events for further info. When working with kernels < v4.2, the following considerations must be taken, as the sched:sched_switch tracepoints will be used to receive such information: Unless /proc/sys/kernel/perf_event_paranoid is set to -1, unprivileged users are not permitted to use tracepoints which means there is insufficient side-band information to decode Intel PT in per-cpu mode, and potentially workload-only mode too if the workload creates new processes. Note also, that to use tracepoints, read-access to debugfs is required. So if debugfs is not mounted or the user does not have read-access, it will again not be possible to decode Intel PT in per-cpu mode. sched_switch tracepoint ----------------------- The sched_switch tracepoint is used to provide side-band data for Intel PT decoding in kernels where the PERF_RECORD_SWITCH metadata event isn't available. The sched_switch events are automatically added. e.g. the second event shown below: $ perf record -vv -e intel_pt//u uname ------------------------------------------------------------ perf_event_attr: type 6 size 112 config 0x400 { sample_period, sample_freq } 1 sample_type IP|TID|TIME|CPU|IDENTIFIER read_format ID disabled 1 inherit 1 exclude_kernel 1 exclude_hv 1 enable_on_exec 1 sample_id_all 1 ------------------------------------------------------------ sys_perf_event_open: pid 31104 cpu 0 group_fd -1 flags 0x8 sys_perf_event_open: pid 31104 cpu 1 group_fd -1 flags 0x8 sys_perf_event_open: pid 31104 cpu 2 group_fd -1 flags 0x8 sys_perf_event_open: pid 31104 cpu 3 group_fd -1 flags 0x8 ------------------------------------------------------------ perf_event_attr: type 2 size 112 config 0x108 { sample_period, sample_freq } 1 sample_type IP|TID|TIME|CPU|PERIOD|RAW|IDENTIFIER read_format ID inherit 1 sample_id_all 1 exclude_guest 1 ------------------------------------------------------------ sys_perf_event_open: pid -1 cpu 0 group_fd -1 flags 0x8 sys_perf_event_open: pid -1 cpu 1 group_fd -1 flags 0x8 sys_perf_event_open: pid -1 cpu 2 group_fd -1 flags 0x8 sys_perf_event_open: pid -1 cpu 3 group_fd -1 flags 0x8 ------------------------------------------------------------ perf_event_attr: type 1 size 112 config 0x9 { sample_period, sample_freq } 1 sample_type IP|TID|TIME|IDENTIFIER read_format ID disabled 1 inherit 1 exclude_kernel 1 exclude_hv 1 mmap 1 comm 1 enable_on_exec 1 task 1 sample_id_all 1 mmap2 1 comm_exec 1 ------------------------------------------------------------ sys_perf_event_open: pid 31104 cpu 0 group_fd -1 flags 0x8 sys_perf_event_open: pid 31104 cpu 1 group_fd -1 flags 0x8 sys_perf_event_open: pid 31104 cpu 2 group_fd -1 flags 0x8 sys_perf_event_open: pid 31104 cpu 3 group_fd -1 flags 0x8 mmap size 528384B AUX area mmap length 4194304 perf event ring buffer mmapped per cpu Synthesizing auxtrace information Linux [ perf record: Woken up 1 times to write data ] [ perf record: Captured and wrote 0.042 MB perf.data ] Note, the sched_switch event is only added if the user is permitted to use it and only in per-cpu mode. Note also, the sched_switch event is only added if TSC packets are requested. That is because, in the absence of timing information, the sched_switch events cannot be matched against the Intel PT trace. perf script =========== By default, perf script will decode trace data found in the perf.data file. This can be further controlled by new option --itrace. New --itrace option ------------------- Having no option is the same as --itrace which, in turn, is the same as --itrace=ibxe The letters are: i synthesize "instructions" events b synthesize "branches" events x synthesize "transactions" events c synthesize branches events (calls only) r synthesize branches events (returns only) e synthesize tracing error events d create a debug log g synthesize a call chain (use with i or x) l synthesize last branch entries (use with i or x) s skip initial number of events "Instructions" events look like they were recorded by "perf record -e instructions". "Branches" events look like they were recorded by "perf record -e branches". "c" and "r" can be combined to get calls and returns. "Transactions" events correspond to the start or end of transactions. The 'flags' field can be used in perf script to determine whether the event is a tranasaction start, commit or abort. Error events are new. They show where the decoder lost the trace. Error events are quite important. Users must know if what they are seeing is a complete picture or not. The "d" option will cause the creation of a file "intel_pt.log" containing all decoded packets and instructions. Note that this option slows down the decoder and that the resulting file may be very large. In addition, the period of the "instructions" event can be specified. e.g. --itrace=i10us sets the period to 10us i.e. one instruction sample is synthesized for each 10 microseconds of trace. Alternatives to "us" are "ms" (milliseconds), "ns" (nanoseconds), "t" (TSC ticks) or "i" (instructions). "ms", "us" and "ns" are converted to TSC ticks. The timing information included with Intel PT does not give the time of every instruction. Consequently, for the purpose of sampling, the decoder estimates the time since the last timing packet based on 1 tick per instruction. The time on the sample is *not* adjusted and reflects the last known value of TSC. For Intel PT, the default period is 100us. Setting it to a zero period means "as often as possible". In the case of Intel PT that is the same as a period of 1 and a unit of 'instructions' (i.e. --itrace=i1i). Also the call chain size (default 16, max. 1024) for instructions or transactions events can be specified. e.g. --itrace=ig32 --itrace=xg32 Also the number of last branch entries (default 64, max. 1024) for instructions or transactions events can be specified. e.g. --itrace=il10 --itrace=xl10 Note that last branch entries are cleared for each sample, so there is no overlap from one sample to the next. To disable trace decoding entirely, use the option --no-itrace. It is also possible to skip events generated (instructions, branches, transactions) at the beginning. This is useful to ignore initialization code. --itrace=i0nss1000000 skips the first million instructions. dump option ----------- perf script has an option (-D) to "dump" the events i.e. display the binary data. When -D is used, Intel PT packets are displayed. The packet decoder does not pay attention to PSB packets, but just decodes the bytes - so the packets seen by the actual decoder may not be identical in places where the data is corrupt. One example of that would be when the buffer-switching interrupt has been too slow, and the buffer has been filled completely. In that case, the last packet in the buffer might be truncated and immediately followed by a PSB as the trace continues in the next buffer. To disable the display of Intel PT packets, combine the -D option with --no-itrace. perf report =========== By default, perf report will decode trace data found in the perf.data file. This can be further controlled by new option --itrace exactly the same as perf script, with the exception that the default is --itrace=igxe. perf inject =========== perf inject also accepts the --itrace option in which case tracing data is removed and replaced with the synthesized events. e.g. perf inject --itrace -i perf.data -o perf.data.new Below is an example of using Intel PT with autofdo. It requires autofdo (https://github.com/google/autofdo) and gcc version 5. The bubble sort example is from the AutoFDO tutorial (https://gcc.gnu.org/wiki/AutoFDO/Tutorial) amended to take the number of elements as a parameter. $ gcc-5 -O3 sort.c -o sort_optimized $ ./sort_optimized 30000 Bubble sorting array of 30000 elements 2254 ms $ cat ~/.perfconfig [intel-pt] mispred-all $ perf record -e intel_pt//u ./sort 3000 Bubble sorting array of 3000 elements 58 ms [ perf record: Woken up 2 times to write data ] [ perf record: Captured and wrote 3.939 MB perf.data ] $ perf inject -i perf.data -o inj --itrace=i100usle --strip $ ./create_gcov --binary=./sort --profile=inj --gcov=sort.gcov -gcov_version=1 $ gcc-5 -O3 -fauto-profile=sort.gcov sort.c -o sort_autofdo $ ./sort_autofdo 30000 Bubble sorting array of 30000 elements 2155 ms Note there is currently no advantage to using Intel PT instead of LBR, but that may change in the future if greater use is made of the data.