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.
pt Specifies pass-through which enables the branch config term.
The default config selects 'pt' if it is available, so a user will
never need to specify this term.
branch Enable branch tracing. Branch tracing is enabled by
default so to disable branch tracing use branch=0.
The default config selects 'branch' if it is available.
ptw Enable PTWRITE packets which are produced when a ptwrite
instruction is executed.
Support for this feature is indicated by:
/sys/bus/event_source/devices/intel_pt/caps/ptwrite
which contains "1" if the feature is supported and
"0" otherwise.
fup_on_ptw Enable a FUP packet to follow the PTWRITE packet. The
FUP packet provides the address of the ptwrite instruction. In
the absence of fup_on_ptw, the decoder will use the address of
the previous branch if branch tracing is enabled, otherwise the
address will be zero. Note that fup_on_ptw will work even when
branch tracing is disabled.
pwr_evt Enable power events. The power events provide information
about changes to the CPU C-state.
Support for this feature is indicated by:
/sys/bus/event_source/devices/intel_pt/caps/power_event_trace
which contains "1" if the feature is supported and
"0" otherwise.
AUX area sampling option
To select Intel PT "sampling" the AUX area sampling option can be
used:
--aux-sample
Optionally it can be followed by the sample size in bytes e.g.
--aux-sample=8192
In addition, the Intel PT event to sample must be defined e.g.
-e intel_pt//u
Samples on other events will be created containing Intel PT data
e.g. the following will create Intel PT samples on the
branch-misses event, note the events must be grouped using {}:
perf record --aux-sample -e '{intel_pt//u,branch-misses:u}'
An alternative to --aux-sample is to add the config term
aux-sample-size to events. In this case, the grouping is implied
e.g.
perf record -e intel_pt//u -e branch-misses/aux-sample-size=8192/u
is the same as:
perf record -e '{intel_pt//u,branch-misses/aux-sample-size=8192/u}'
but allows for also using an address filter e.g.:
perf record -e intel_pt//u --filter 'filter * @/bin/ls' -e branch-misses/aux-sample-size=8192/u -- ls
It is important to select a sample size that is big enough to
contain at least one PSB packet. If not a warning will be
displayed:
Intel PT sample size (%zu) may be too small for PSB period (%zu)
The calculation used for that is: if sample_size ⟨ psb_period +
256 display the warning. When sampling is used, psb_period
defaults to 0 (2KiB).
The default sample size is 4KiB.
The sample size is passed in aux_sample_size in struct
perf_event_attr. The sample size is limited by the maximum event
size which is 64KiB. It is difficult to know how big the event
might be without the trace sample attached, but the tool
validates that the sample size is not greater than 60KiB.
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 or sampling
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 3 modes: full-trace mode sample mode
snapshot mode
Full-trace mode traces continuously e.g.
perf record -e intel_pt//u uname
Sample mode attaches a Intel PT sample to other events e.g.
perf record --aux-sample -e intel_pt//u -e branch-misses:u
Snapshot mode captures the available data when a signal is sent
or "snapshot" control command is issued. e.g. using a signal
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 advantage of using "snapshot" control command is that the
access is controlled by access to a FIFO e.g.
$ mkfifo perf.control
$ mkfifo perf.ack
$ cat perf.ack &
[1] 15235
$ sudo ~/bin/perf record --control fifo:perf.control,perf.ack -S -e intel_pt//u -- sleep 60 &
[2] 15243
$ ps -e | grep perf
15244 pts/1 00:00:00 perf
$ kill -USR2 15244
bash: kill: (15244) - Operation not permitted
$ echo snapshot > perf.control
ack
The 3 Intel PT modes of operation 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_PERFMON or 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.
