On Tue, Jan 31, 2023 at 03:11:05PM -0700, Shuah Khan wrote: > Add a new section to the admin-guide with information of interest to > application developers and system integrators doing analysis of the > Linux kernel for safety critical applications. > > This section will contain documents supporting analysis of kernel > interactions with applications, and key kernel subsystems expectations. > > Add a new workload-tracing document to this new section. > > Signed-off-by: Shefali Sharma <sshefali021@xxxxxxxxx> > Signed-off-by: Shuah Khan <skhan@xxxxxxxxxxxxxxxxxxx> > --- > Changes since v2: Addressed review comments on v2 I think you haven't addressed my comments there [1], so I have to write the improv: ---- >8 ---- diff --git a/Documentation/admin-guide/workload-tracing.rst b/Documentation/admin-guide/workload-tracing.rst index 5fad64b4ebd66f..ac60ff9dec8f0e 100644 --- a/Documentation/admin-guide/workload-tracing.rst +++ b/Documentation/admin-guide/workload-tracing.rst @@ -27,10 +27,10 @@ Methodology =========== `strace <https://man7.org/linux/man-pages/man1/strace.1.html>`_ is a -diagnostic, instructional, and debugging tool and can be used to discover -the system resources in use by a workload. Once we discover and understand -the workload needs, we can focus on them to avoid regressions and use it -to evaluate safety considerations. We use strace tool to trace workloads. +diagnostic, instructional, and debugging tool and can be used to +discover the system resources in use by a workload by tracing it. Once +we discover and understand the workload needs, we can focus on them to +avoid regressions and use it to evaluate safety considerations. This method of tracing using strace tells us the system calls invoked by the workload and doesn't include all the system calls that can be invoked @@ -43,7 +43,7 @@ outlined here will trace and find all possible code paths. The completeness of the system usage information depends on the completeness of coverage of a workload. -The goal is tracing a workload on a system running a default kernel without +The goal is to trace workloads on a system running a default kernel without requiring custom kernel installs. How do we gather fine-grained system information? @@ -63,9 +63,9 @@ insight into the process. "perf annotate" tool generates the statistics of each instruction of the program. This document goes over the details of how to gather fine-grained information on a workload's usage of system resources. -We used strace to trace the perf, stress-ng, paxtest workloads to illustrate -our methodology to discover resources used by a workload. This process can -be applied to trace other workloads. +In this document, we use strace to trace the perf, stress-ng, paxtest +workloads to illustrate our methodology to discover resources used by a +workload. This process can be applied to trace other workloads. Getting the system ready for tracing ==================================== @@ -73,34 +73,38 @@ Getting the system ready for tracing Before we can get started we will show you how to get your system ready. We assume that you have a Linux distribution running on a physical system or a virtual machine. Most distributions will include strace command. Let’s -install other tools that aren’t usually included to build Linux kernel. +install other tools that aren’t usually pre-installed to build Linux kernel. Please note that the following works on Debian based distributions. You might have to find equivalent packages on other Linux distributions. Install tools to build Linux kernel and tools in kernel repository. scripts/ver_linux is a good way to check if your system already has -the necessary tools: :: +the necessary tools:: sudo apt-get build-essentials flex bison yacc sudo apt install libelf-dev systemtap-sdt-dev libaudit-dev libslang2-dev libperl-dev libdw-dev -cscope is a good tool to browse kernel sources. Let's install it now: :: +cscope is a good tool to browse kernel sources. Let's install it now:: sudo apt-get install cscope -Install stress-ng and paxtest: :: +Install stress-ng and paxtest:: apt-get install stress-ng apt-get install paxtest +You will also need to clone Linus's mainline tree, which can be grabbed +by:: + + git clone git://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git linux + Workload overview ================= -As mentioned earlier, we used strace to trace perf bench, stress-ng and -paxtest workloads to show how to analyze a workload and identify Linux -subsystems used by these workloads. Let's start with an overview of these -three workloads to get a better understanding of what they do and how to -use them. +As mentioned earlier, the workloads to be analyzed here are perf, +stress-ng, and paxtest. Let's start with an overview of these three +workloads to get a better understanding of what they do and how to use +them. perf bench (all) workload ------------------------- @@ -108,32 +112,34 @@ perf bench (all) workload The perf bench command contains multiple multi-threaded microkernel benchmarks for executing different subsystems in the Linux kernel and system calls. This allows us to easily measure the impact of changes, -which can help mitigate performance regressions. It also acts as a common -benchmarking framework, enabling developers to easily create test cases, -integrate transparently, and use performance-rich tooling subsystems. +which can help to mitigate performance regressions. It also acts as a +common benchmarking framework, enabling developers to easily create test +cases, integrate transparently, and use performance-rich tooling +subsystems. Stress-ng netdev stressor workload ---------------------------------- stress-ng is used for performing stress testing on the kernel. It allows you to exercise various physical subsystems of the computer, as well as -interfaces of the OS kernel, using "stressor-s". They are available for +interfaces of the OS kernel, using stressors. They are available for CPU, CPU cache, devices, I/O, interrupts, file system, memory, network, -operating system, pipelines, schedulers, and virtual machines. Please refer -to the `stress-ng man-page <https://www.mankier.com/1/stress-ng>`_ to -find the description of all the available stressor-s. The netdev stressor -starts specified number (N) of workers that exercise various netdevice -ioctl commands across all the available network devices. +operating system, pipelines, schedulers, and virtual machines. Please +refer to the `stress-ng man-page <https://www.mankier.com/1/stress-ng>`_ +for the details of all available stressors. The netdev stressor starts +specified number (N) of workers that exercise various netdevice ioctl +commands across all the available network devices. paxtest kiddie workload ----------------------- paxtest is a program that tests buffer overflows in the kernel. It tests -kernel enforcements over memory usage. Generally, execution in some memory -segments makes buffer overflows possible. It runs a set of programs that -attempt to subvert memory usage. It is used as a regression test suite for -PaX, but might be useful to test other memory protection patches for the -kernel. We used paxtest kiddie mode which looks for simple vulnerabilities. +kernel enforcements over memory usage. Generally, execution in some +memory segments makes buffer overflows possible. It runs a set of +programs that attempt to subvert memory usage. It is originally intended +as a regression test suite for PaX, but can also useful to test other +memory protection patches for the kernel. Here, we use paxtest kiddie +mode which looks for simple vulnerabilities. What is strace and how do we use it? ==================================== @@ -155,51 +161,40 @@ suppressing the regular output. This attempts to show system time (CPU time spent running in the kernel) independent of wall clock time. We plan to use these features to get information on workload system usage. -strace command supports basic, verbose, and stats modes. strace command when -run in verbose mode gives more detailed information about the system calls -invoked by a process. +strace command supports basic ("strace <process>", verbose ("strace -v +<process>"), and stats ("strace -v") modes. In verbose mode, strace +gives more detailed information about syscalls invoked by a process. -Running strace -c generates a report of the percentage of time spent in each -system call, the total time in seconds, the microseconds per call, the total -number of calls, the count of each system call that has failed with an error -and the type of system call made. - - * Usage: strace <command we want to trace> - * Verbose mode usage: strace -v <command> - * Gather statistics: strace -c <command> - -We used the “-c” option to gather fine-grained run-time statistics in use -by three workloads we have chose for this analysis. - - * perf - * stress-ng - * paxtest +In stats mode, strace generates fine-grained run-time statistics report +which consisted of: percentage of time spent in each system call; the +total time in seconds; the microseconds per call; the total number of +calls, the count of each system call that has failed with an error and +the type of system call made. What is cscope and how do we use it? ==================================== Now let’s look at `cscope <https://cscope.sourceforge.net/>`_, a command -line tool for browsing C, C++ or Java code-bases. We can use it to find +line tool for browsing C, C++ or Java code-bases. You can use it to find all the references to a symbol, global definitions, functions called by a function, functions calling a function, text strings, regular expression patterns, files including a file. -We can use cscope to find which system call belongs to which subsystem. -This way we can find the kernel subsystems used by a process when it is -executed. +In context of this document, you can use cscope to find which system +call belongs to which subsystem. This way you can find the kernel +subsystems used by a process when it is executed. -Let’s checkout the latest Linux repository and build cscope database: :: +To begin using cscope, cd to the kernel sources directory and build the +database:: - git clone git://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git linux cd linux cscope -R -p10 # builds cscope.out database before starting browse session cscope -d -p10 # starts browse session on cscope.out database -Note: Run "cscope -R -p10" to build the database and c"scope -d -p10" to -enter into the browsing session. cscope by default cscope.out database. -To get out of this mode press ctrl+d. -p option is used to specify the -number of file path components to display. -p10 is optimal for browsing -kernel sources. +Here, "cscope -R -p10" builds the database and "cscope -d -p10" to +browses the resulting database, which is by default in cscope.out. To +quit browsing, type ctrl+d. -p option is used to specify the directory +depth. The 10-level depth is sufficient for browsing kernel sources. What is perf and how do we use it? ================================== @@ -210,21 +205,20 @@ a simple command line interface. Perf is based on the perf_events interface exported by the kernel. It is very useful for profiling the system and finding performance bottlenecks in an application. -If you haven't already checked out the Linux mainline repository, you can do -so and then build kernel and perf tool: :: +Change to the kernel sources directory and build both kernel and perf tool:: - git clone git://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git linux cd linux make -j3 all cd tools/perf make -Note: The perf command can be built without building the kernel in the -repository and can be run on older kernels. However matching the kernel -and perf revisions gives more accurate information on the subsystem usage. +.. note:: + The perf command can be built without building the kernel in the + repository and can be run on any kernels. However matching the kernel + and perf revisions gives more accurate information on the subsystem usage. -We used "perf stat" and "perf bench" options. For a detailed information on -the perf tool, run "perf -h". +Below, we will describe "perf stat" and "perf bench" options. For +detailed help on perf tool, see "perf -h". perf stat --------- @@ -268,17 +262,17 @@ exercised: * SIOCGIFADDR, SIOCGIFNETMASK, SIOCGIFMETRIC, SIOCGIFMTU * SIOCGIFHWADDR, SIOCGIFMAP, SIOCGIFTXQLEN -The following command runs the stressor: :: +To run the netdev stressor:: stress-ng --netdev 1 -t 60 --metrics command. -We can use the perf record command to record the events and information -associated with a process. This command records the profiling data in the -perf.data file in the same directory. +Then you can use "perf record" command to record the events and +information associated with a process. This command records the +profiling data in the perf.data file in the same directory. -Using the following commands you can record the events associated with the -netdev stressor, view the generated report perf.data and annotate the to -view the statistics of each instruction of the program: :: +For example, to record stress-ng stressor above, view the generated +report and annotate it to gather statistics of each instruction of the +program:: perf record stress-ng --netdev 1 -t 60 --metrics command. perf report @@ -288,22 +282,21 @@ What is paxtest and how do we use it? ===================================== paxtest is a program that tests buffer overflows in the kernel. It tests -kernel enforcements over memory usage. Generally, execution in some memory -segments makes buffer overflows possible. It runs a set of programs that -attempt to subvert memory usage. It is used as a regression test suite for -PaX, and will be useful to test other memory protection patches for the -kernel. +kernel enforcements over memory usage. Generally, execution in some +memory segments makes buffer overflows possible. It runs a set of +programs that attempt to subvert memory usage. It is originally intend +as a regression test suite for PaX, but it can also be useful to test +other memory protection patches for the kernel. -paxtest provides kiddie and blackhat modes. The paxtest kiddie mode runs -in normal mode, whereas the blackhat mode tries to get around the protection +paxtest provides kiddie and blackhat modes. The former runs +in normal mode, whereas the latter tries to get around the protection of the kernel testing for vulnerabilities. We focus on the kiddie mode here -and combine "paxtest kiddie" run with "perf record" to collect CPU stack -traces for the paxtest kiddie run to see which function is calling other -functions in the performance profile. Then the "dwarf" (DWARF's Call Frame -Information) mode can be used to unwind the stack. +and combine it with perf to collect CPU stack +traces for the paxtest run to see which function is calling other +functions in the performance profile. Then stack unwinding can be done +by specifying "--call-graph dwarf" option to perf. -The following command can be used to view resulting report in call-graph -format: :: +Thus, the combined commands are:: perf record --call-graph dwarf paxtest kiddie perf report --stdio @@ -316,14 +309,17 @@ Now that we understand the workloads, let's start tracing them. Tracing perf bench all workload ------------------------------- -Run the following command to trace perf bench all workload: :: +To trace all workloads under perf benchmark:: - strace -c perf bench all + strace -c perf bench all -**System Calls made by the workload** +The below table is the list of invoked syscalls, with number of +times each is invoked, and the corresponding Linux subsystem. -The below table shows the system calls invoked by the workload, number of -times each system call is invoked, and the corresponding Linux subsystem. +.. note:: + + The syscall tables below are generated from example workloads. The actual + figures may differ depending on workload being traced. +-------------------+-----------+-----------------+-------------------------+ | System Call | # calls | Linux Subsystem | System Call (API) | @@ -426,14 +422,11 @@ times each system call is invoked, and the corresponding Linux subsystem. Tracing stress-ng netdev stressor workload ------------------------------------------ -Run the following command to trace stress-ng netdev stressor workload: :: +To trace netdev stress-ng workload:: strace -c stress-ng --netdev 1 -t 60 --metrics -**System Calls made by the workload** - -The below table shows the system calls invoked by the workload, number of -times each system call is invoked, and the corresponding Linux subsystem. +The corresponding syscall table is: +-------------------+-----------+-----------------+-------------------------+ | System Call | # calls | Linux Subsystem | System Call (API) | @@ -520,14 +513,11 @@ times each system call is invoked, and the corresponding Linux subsystem. Tracing paxtest kiddie workload ------------------------------- -Run the following command to trace paxtest kiddie workload: :: +To trace paxtest (kiddie mode) workload:: - strace -c paxtest kiddie + strace -c paxtest kiddie -**System Calls made by the workload** - -The below table shows the system calls invoked by the workload, number of -times each system call is invoked, and the corresponding Linux subsystem. +The corresponding syscall table is: +-------------------+-----------+-----------------+----------------------+ | System Call | # calls | Linux Subsystem | System Call (API) | @@ -590,8 +580,10 @@ times each system call is invoked, and the corresponding Linux subsystem. Conclusion ========== -This document is intended to be used as a guide on how to gather fine-grained -information on the resources in use by workloads using strace. +This document is intended to be used as a guide on how to gather +fine-grained information on the resources in use by workloads using +strace. You may want to consult references below in case you want to run +strace for your workload needs. References ========== Thanks. [1]: https://lore.kernel.org/linux-doc/Y9STCwt2FnYf4%2FX4@xxxxxxxxx/ -- An old man doll... just what I always wanted! - Clara
Attachment:
signature.asc
Description: PGP signature
