commit | 218f7482f8ae8de607fbfa8ac3d89de188717e0c | [log] [tgz] |
---|---|---|
author | William Cohen <wcohen@redhat.com> | Thu Dec 06 13:41:01 2018 -0500 |
committer | yonghong-song <ys114321@gmail.com> | Thu Dec 06 10:41:01 2018 -0800 |
tree | 5d6a351743480640b32f85244acbdafc99b3b077 | |
parent | f3fd8e30e1a1e4e61c3d206b10bf4fc11ac26dd9 [diff] |
Wcohen/efficiency (#2063) * Reduce instrumentation overhead with the sys_enter and sys_exit tracepoints The ucalls script initially used kprobes and kretprobes on each of the hundreds of syscalls functions in the system. This approach causes a large number of probes to be set up at the start and removed at the conclusion of the script's execution resulting in slow start up. Like the syscount.py script the ucall syscall instrumentation has been modified to use the sys_enter and sys_exit tracepoints. This only requires the installation and removal of one or two tracepoints to implement and results in much shorter times to start and stop the ucalls script. Another benefit of this change is syscalls on newer kernels will be monitored with the "-S" option. The regular expression used to find the locations for the kprobes and kretprobes for all the possible syscall functions would not would match the syscall function naming convention in newer kernels. * Update ucalls_examples.txt to match current "-S" option output * Add required "import subprocess" and remove unneeded "global syscalls" * Factor out the syscall_name code into a separate python module syscall.py Multiple scripts are going to find the syscall_name() function useful when using the syscall tracepoints. Factoring out this code into a separate python module avoids having to replicate this code in multiple scripts. * Use the syscall_name() function in syscount.py to make it more compact. * Update the default syscall mappings and the way that they were generated The default table was missing some newer syscall mapping. Regenerated the table using the syscallent.h file from Fedora 30 strace-4.25-1.fc30.src.rpm. Also updated the comment with the command actually used to generate the mappings. * Add license information and upsdate the syscalls The default x86_64 syscall dictionary mapping syscalls numbers to names has been updated. The following syscall x86_64 names have been updated: 18: b"pwrite64", 60: b"exit", 166: b"umount2", The following syscall x86_64 have been added: 313: b"finit_module", 314: b"sched_setattr", 315: b"sched_getattr", 316: b"renameat2", 317: b"seccomp", 318: b"getrandom", 319: b"memfd_create", 320: b"kexec_file_load", 321: b"bpf", 322: b"execveat", 323: b"userfaultfd", 324: b"membarrier", 325: b"mlock2", 326: b"copy_file_range", 327: b"preadv2", 328: b"pwritev2", 329: b"pkey_mprotect", 330: b"pkey_alloc", 331: b"pkey_free", 332: b"statx", 333: b"io_pgetevents", 334: b"rseq", * Eliminate stderr output and use of shell features Redirect all stderr output so it isn't seen. Also avoid use of the shell pipe and tail command. Just strip off the first line in the python code instead. * Update lib/ucalls.py smoke test to required linux-4.7 The use of tracepoints in the ucalls.py requires linux-4.7. Changed the test to only run with a suitable kernel. The libs/ucalls.py script is no longer inserting hundreds of kprobes and is much faster as a result, so removed the timeout adjustment and the comment about it being slow.
BCC is a toolkit for creating efficient kernel tracing and manipulation programs, and includes several useful tools and examples. It makes use of extended BPF (Berkeley Packet Filters), formally known as eBPF, a new feature that was first added to Linux 3.15. Much of what BCC uses requires Linux 4.1 and above.
eBPF was described by Ingo Molnár as:
One of the more interesting features in this cycle is the ability to attach eBPF programs (user-defined, sandboxed bytecode executed by the kernel) to kprobes. This allows user-defined instrumentation on a live kernel image that can never crash, hang or interfere with the kernel negatively.
BCC makes BPF programs easier to write, with kernel instrumentation in C (and includes a C wrapper around LLVM), and front-ends in Python and lua. It is suited for many tasks, including performance analysis and network traffic control.
This example traces a disk I/O kernel function, and populates an in-kernel power-of-2 histogram of the I/O size. For efficiency, only the histogram summary is returned to user-level.
# ./bitehist.py Tracing... Hit Ctrl-C to end. ^C kbytes : count distribution 0 -> 1 : 3 | | 2 -> 3 : 0 | | 4 -> 7 : 211 |********** | 8 -> 15 : 0 | | 16 -> 31 : 0 | | 32 -> 63 : 0 | | 64 -> 127 : 1 | | 128 -> 255 : 800 |**************************************|
The above output shows a bimodal distribution, where the largest mode of 800 I/O was between 128 and 255 Kbytes in size.
See the source: bitehist.py. What this traces, what this stores, and how the data is presented, can be entirely customized. This shows only some of many possible capabilities.
See INSTALL.md for installation steps on your platform.
See FAQ.txt for the most common troubleshoot questions.
See docs/reference_guide.md for the reference guide to the bcc and bcc/BPF APIs.
Some of these are single files that contain both C and Python, others have a pair of .c and .py files, and some are directories of files.
Examples:
Tools that help to introspect BPF programs.
BPF guarantees that the programs loaded into the kernel cannot crash, and cannot run forever, but yet BPF is general purpose enough to perform many arbitrary types of computation. Currently, it is possible to write a program in C that will compile into a valid BPF program, yet it is vastly easier to write a C program that will compile into invalid BPF (C is like that). The user won't know until trying to run the program whether it was valid or not.
With a BPF-specific frontend, one should be able to write in a language and receive feedback from the compiler on the validity as it pertains to a BPF backend. This toolkit aims to provide a frontend that can only create valid BPF programs while still harnessing its full flexibility.
Furthermore, current integrations with BPF have a kludgy workflow, sometimes involving compiling directly in a linux kernel source tree. This toolchain aims to minimize the time that a developer spends getting BPF compiled, and instead focus on the applications that can be written and the problems that can be solved with BPF.
The features of this toolkit include:
In the future, more bindings besides python will likely be supported. Feel free to add support for the language of your choice and send a pull request!
At Red Hat Summit 2015, BCC was presented as part of a session on BPF. A multi-host vxlan environment is simulated and a BPF program used to monitor one of the physical interfaces. The BPF program keeps statistics on the inner and outer IP addresses traversing the interface, and the userspace component turns those statistics into a graph showing the traffic distribution at multiple granularities. See the code here.
Already pumped up to commit some code? Here are some resources to join the discussions in the IOVisor community and see what you want to work on.
Looking for more information on BCC and how it's being used? You can find links to other BCC content on the web in LINKS.md.