simple-pt is a simple experimental reference implementation of Intel Processor Trace (PT) on Linux. PT can trace all branches executed by the CPU at the hardware level with moderate overhead. A PT decoder then uses sideband trace data to decode the branch traces.
PT is supported on Intel 5th generation Core (Broadwell), 6th generation Core (Skylake) CPUs, as well as Goldmont based Atom CPUs (Intel Joule, Apollo Lake)
% sptcmd -c tcall taskset -c 0 ./tcall
cpu 0 offset 1027688, 1003 KB, writing to ptout.0
...
Wrote sideband to ptout.sideband
% sptdecode --sideband ptout.sideband --pt ptout.0 | less
TIME DELTA INSNs OPERATION
frequency 32
0 [+0] [+ 1] _dl_aux_init+436
[+ 6] __libc_start_main+455 -> _dl_discover_osversion
...
[+ 13] __libc_start_main+446 -> main
[+ 9] main+22 -> f1
[+ 4] f1+9 -> f2
[+ 2] f1+19 -> f2
[+ 5] main+22 -> f1
[+ 4] f1+9 -> f2
[+ 2] f1+19 -> f2
[+ 5] main+22 -> f1
...
simple-pt consists of a
- kernel driver
- sptcmd to collect data from the kernel driver
- sptdecode to decode PT information
- fastdecode to dump raw PT traces
It uses the libipt PT decoding library
Note that Linux 4.1 and later has an integrated PT implementation as part of Linux perf. gdb 7.10 also supports full debugging on top of PT. Intel VTune also supports PT.
If you want a full production system please use one of these. simple-pt is an experimental implementation.
Simple PT does NOT support:
- It does not support long term tracing of more data than fits in the buffer (no interrupt) (use perf)
- It does not support any sampling (use perf or VTune)
- It requires root rights to collect data (use perf)
- It does not support interactive debugging (use gdb or hardware debuggers)
Simple PT has the following functionality
- set up hardware to processor trace
- supports a ring buffer of branch data, stopped on events
- supports flushing buffer on panic
- does not require patching the kernel (although it cheats a bit using kprobes)
- set up PT filters, such as kernel filter
- start and stop traces at specific kernel addresses
- support tracing multiple processes
- print all function calls in "ftrace" style
- disasseembling all executed instructions (requires udis86 library)
- simple driver that could be ported to older kernel releases or other operating systems
- simple code base that is easily changed.
- modular "unix style" design with simple tools that do only one thing
- can dump branches before panic to kernel log and decode
Build and install libipt. This currently requires a patched version of libipt.
git clone -b simple-pt https://github.com/andikleen/processor-trace
cd processor-trace
cmake .
make
sudo make install
sudo ldconfig
WARNING: Please do not forget the "-b simple-pt" in the above clone line. simple-pt will not build using the master branch!
Install libelf-elf-devel or elfutils-devel or similar depending on your distribution.
Optional install udis86 if you want to see disassembled instructions:
git clone https://udis86.git.sourceforge.net/gitroot/udis86/udis86
cd udis86
./configure
make
Clone simple-pt
git clone https://github.com/andikleen/simple-pt
cd simple-pt
Build the kernel module. May require installing kernel includes from your distribution.
make
Install the kernel module
sudo make modules_install
Build the user tools
make user
If you installed udis86 use
make user UDIS86=1
Check if your system supports PT
./ptfeature
Run a trace
sudo ./sptcmd -c ls ls
sudo ./sptdecode --sideband ptout.sideband --pt ptout.0 | less
sptcmd loads and configures the kernel driver. It runs a program with trace. It always does a global trace. It writes the pt trace data to trace files for each CPU (ptout.N where N is the CPU number). It also writes side band information needed to decode the trace into the ptout.sideband file.
-c sets a command filter, tracing only commands with that name. Otherwise everything global is traced.
sptdecode then decodes the trace for a CPU using the side band information. When it should decode kernel code it needs to run as root to be able to read /proc/kcore. If it's not run as root kernel code will not be shown.
Another way to use simple-pt is to run the workload with PT running in the background and only dump on an event.
Start trace and dump trace on event:
sudo ./sptcmd --enable
<run workload>
<some event of interest happens and triggers:>
sudo ./sptcmd --dump
sudo ./sptdecode --sideband ptout.sideband --pt ptout.0 | less
Another way is to use --stop-address or --stop-range to stop the trace on specific kernel symbols being executed. Note that these options to affect the trace on their current CPU.
Run test suite
sudo ./tester
The kernel driver manages the PT hardware and allocates the trace buffers. It also sets up some custom trace points for the sideband data.
The simple-pt kernel driver is configured using module parameters. Many can be changed at runtime through /sys/module/simple_pt/parameters. A few need a driver reload
Use modinfo simple-pt.ko
to show all allowed parameters. For most parameters sptcmd has options to set them up. That is the recommended interface.
sptcmd configures the driver, starts the trace and runs the trace command. The driver sets up a ring buffer and runs the the processor trace for each CPU until stopped. Then it calls sptdump to write the buffer for each CPU to a ptout.N file (N is the number of the CPU)
For the side band information ftrace with some custom trace points defined by the driver is used. sptsideband converts the ftrace output into the .sideband files used by the decoder.
sptdecode then reads the PT data, the sideband data, the executables, the kernel code through /proc/kcore, and uses the libipt decoder to reconstruct the trace.
To change the PT buffer size the driver needs to be loaded manually. The PT buffer size can be changed with the pt_buffer_order parameter.
rmmod simple_pt # if it was loaded
modprobe simple_pt pt_buffer_order=10
The size is specified in 2^n 4K pages. The default is 9 (2MB). The maximum limit is the kernel's MAX_ORDER limit, typically 8MB. The allocation may also fail if the kernel memory is too fragmented. In this case quitting a large process may help.
When ptfeature shows the "multiple toPA entries" feature it is possible to allocate multiple PT buffers with the pt_num_buffers parameter. All the buffers are logically concatenated. The default is one buffer. The maximum is 511 buffers.
simple-pt can be used to print a number of branches before a panic.
insmod simple-pt.ko start=1 print_panic_psbs=4
<panic system>
<collect log from serial console>
The number after print_panic_psbs specifies the length of the logged trace (expressed in number of PT sync points)
The PT information is logged in base64 format to the kernel log. It can be recovered with the base64log.py utility
base64log.py < log > ptlog
sptdecode --elf vmlinux --pt ptlog
This method currently does not support modules or ring 3 code, or multiple PT buffers.
-
To limit the program to one CPU use sptcmd taskset -c CPU ..
-
To demangle C++ symbols pipe output through c++filt
-
To start/stop around specific user code bracket it with dummy syscalls that you can then put a kernel trigger on. The test suite uses personality(21212212) and prctl(12341234). This will be improved in the future.
-
perf or the BIOS may be already using the PT hardware. If you know it's safe you can take over the PT hardware with --force -d.
-
When configuring the driver manually you need to manually reset any parameters you do not want anymore. sptcmd takes care of that automatically.
-
Some Debian kernels are built without CONFIG_KALLSYMS_ALL. When you see an "Cannot find task_lock" error message load the simple_pt module like this
insmod simple_pt.ko tasklist_lock_ptr=0x$(grep tasklist_lock /boot/System.map-$(uname -r) | awk ' {print $1}')
-
Various older Linux kernels have problems with ftrace in kernel modules. simple-pt relies on ftrace output for its sideband. "tester" has a special test. If there are problems likely the workarounds in "compat.h" (e.g. the ifdefs) need to be adjusted. Upgrading to a newer kernel fixes the problem too.
-
When kernel tracing is disabled (-K) multiple processes cannot be distinguished by the decoder.
-
Enabling/Disabling tracing causes the kernel to modify itself, which can cause the PT decoder to lose synchronization. sptcmd disables trace points. Workaround is to keep trace points running after the trace ends with -k, or disable kernel tracing. This can sometimes affect the test suite. If this happens try "tester -k"
-
sptcmd does not continuously save side band data, so events at the beginning of a trace may not be saved. For complex workloads it may be needed to increase the trace buffers in /sys/kernel/debug/tracing/buffer_size_kb.
-
The decoder does not (currently) support reusing the same address region in a process for different code (for example after dlclose/dlopen)
-
Tracing JITed code is not supported.
-
On Skylake the trace time occasionally jumps backwards after frequency changes.
-
Decoder loses synchronization in some cases where it shouldn't.
-
Binaries with spaces in the name are not supported (due to limitations in sptsideband.py)
There is some Linux specific code in the driver, but the basic PT hardware configuration should be straight forward to adapt to other environments. The minimum support needed is memory allocation, a mechanism to call a callback on all CPUs (IPIs), and a mechanism to establish a shared buffer with the decoding tool (implemented using mmap on a character device). When suspend-to-ram is supported it's also useful to have a callback after resume to reinitialize the hardware.
The kernel driver is configured using global variables with Linux's moduleparams mechanism. This can be replaced with simple hard coded variables.
The driver supports Linux "kprobes" and "kallsyms" to set custom triggers. That code is all optional and can be removed. Such optional code is generally marked as optional.
The user tools should be portable to POSIX C99 based systems. The code to access the kernel image will need to be adapted. Porting to non DWARF/ELF based systems will need more work.
For bugs please file a github issue.
Andi Kleen