요약·해설과 원문, 전문 번역을 서로 분리했습니다. API 이름, symbol, source path는 원문 표기를 사용합니다.
1. 요약·해설
원문의 핵심 논리와 kernel programming 관점의 보충 설명입니다. 아래의 전문 번역과는 별도로 작성했습니다.
2. 영어 원문 전체
번역 기준이 된 Linux v6.18.37 원문입니다. 줄 번호는 이 버전의 파일 좌표입니다.
원문 전체 펼치기
1
.. SPDX-License-Identifier: (GPL-2.0+ OR CC-BY-4.0)
3
======================================================
4
Discovering Linux kernel subsystems used by a workload
5
======================================================
7
:Authors: - Shuah Khan <skhan@linuxfoundation.org>
8
- Shefali Sharma <sshefali021@gmail.com>
9
:maintained-by: Shuah Khan <skhan@linuxfoundation.org>
11
Key Points
12
==========
14
* Understanding system resources necessary to build and run a workload
15
is important.
16
* Linux tracing and strace can be used to discover the system resources
17
in use by a workload. The completeness of the system usage information
18
depends on the completeness of coverage of a workload.
19
* Performance and security of the operating system can be analyzed with
20
the help of tools such as:
21
`perf <https://man7.org/linux/man-pages/man1/perf.1.html>`_,
22
`stress-ng <https://www.mankier.com/1/stress-ng>`_,
23
`paxtest <https://github.com/opntr/paxtest-freebsd>`_.
24
* Once we discover and understand the workload needs, we can focus on them
25
to avoid regressions and use it to evaluate safety considerations.
27
Methodology
28
===========
30
`strace <https://man7.org/linux/man-pages/man1/strace.1.html>`_ is a
31
diagnostic, instructional, and debugging tool and can be used to discover
32
the system resources in use by a workload. Once we discover and understand
33
the workload needs, we can focus on them to avoid regressions and use it
34
to evaluate safety considerations. We use strace tool to trace workloads.
36
This method of tracing using strace tells us the system calls invoked by
37
the workload and doesn't include all the system calls that can be invoked
38
by it. In addition, this tracing method tells us just the code paths within
39
these system calls that are invoked. As an example, if a workload opens a
40
file and reads from it successfully, then the success path is the one that
41
is traced. Any error paths in that system call will not be traced. If there
42
is a workload that provides full coverage of a workload then the method
43
outlined here will trace and find all possible code paths. The completeness
44
of the system usage information depends on the completeness of coverage of a
45
workload.
47
The goal is tracing a workload on a system running a default kernel without
48
requiring custom kernel installs.
50
How do we gather fine-grained system information?
51
=================================================
53
strace tool can be used to trace system calls made by a process and signals
54
it receives. System calls are the fundamental interface between an
55
application and the operating system kernel. They enable a program to
56
request services from the kernel. For instance, the open() system call in
57
Linux is used to provide access to a file in the file system. strace enables
58
us to track all the system calls made by an application. It lists all the
59
system calls made by a process and their resulting output.
61
You can generate profiling data combining strace and perf record tools to
62
record the events and information associated with a process. This provides
63
insight into the process. "perf annotate" tool generates the statistics of
64
each instruction of the program. This document goes over the details of how
65
to gather fine-grained information on a workload's usage of system resources.
67
We used strace to trace the perf, stress-ng, paxtest workloads to illustrate
68
our methodology to discover resources used by a workload. This process can
69
be applied to trace other workloads.
71
Getting the system ready for tracing
72
====================================
74
Before we can get started we will show you how to get your system ready.
75
We assume that you have a Linux distribution running on a physical system
76
or a virtual machine. Most distributions will include strace command. Let’s
77
install other tools that aren’t usually included to build Linux kernel.
78
Please note that the following works on Debian based distributions. You
79
might have to find equivalent packages on other Linux distributions.
81
Install tools to build Linux kernel and tools in kernel repository.
82
scripts/ver_linux is a good way to check if your system already has
83
the necessary tools::
85
sudo apt-get install build-essential flex bison yacc
86
sudo apt install libelf-dev systemtap-sdt-dev libslang2-dev libperl-dev libdw-dev
88
cscope is a good tool to browse kernel sources. Let's install it now::
90
sudo apt-get install cscope
92
Install stress-ng and paxtest::
94
apt-get install stress-ng
95
apt-get install paxtest
97
Workload overview
98
=================
100
As mentioned earlier, we used strace to trace perf bench, stress-ng and
101
paxtest workloads to show how to analyze a workload and identify Linux
102
subsystems used by these workloads. Let's start with an overview of these
103
three workloads to get a better understanding of what they do and how to
104
use them.
106
perf bench (all) workload
107
-------------------------
109
The perf bench command contains multiple multi-threaded microkernel
110
benchmarks for executing different subsystems in the Linux kernel and
111
system calls. This allows us to easily measure the impact of changes,
112
which can help mitigate performance regressions. It also acts as a common
113
benchmarking framework, enabling developers to easily create test cases,
114
integrate transparently, and use performance-rich tooling subsystems.
116
Stress-ng netdev stressor workload
117
----------------------------------
119
stress-ng is used for performing stress testing on the kernel. It allows
120
you to exercise various physical subsystems of the computer, as well as
121
interfaces of the OS kernel, using "stressor-s". They are available for
122
CPU, CPU cache, devices, I/O, interrupts, file system, memory, network,
123
operating system, pipelines, schedulers, and virtual machines. Please refer
124
to the `stress-ng man-page <https://www.mankier.com/1/stress-ng>`_ to
125
find the description of all the available stressor-s. The netdev stressor
126
starts specified number (N) of workers that exercise various netdevice
127
ioctl commands across all the available network devices.
129
paxtest kiddie workload
130
-----------------------
132
paxtest is a program that tests buffer overflows in the kernel. It tests
133
kernel enforcements over memory usage. Generally, execution in some memory
134
segments makes buffer overflows possible. It runs a set of programs that
135
attempt to subvert memory usage. It is used as a regression test suite for
136
PaX, but might be useful to test other memory protection patches for the
137
kernel. We used paxtest kiddie mode which looks for simple vulnerabilities.
139
What is strace and how do we use it?
140
====================================
142
As mentioned earlier, strace which is a useful diagnostic, instructional,
143
and debugging tool and can be used to discover the system resources in use
144
by a workload. It can be used:
146
* To see how a process interacts with the kernel.
147
* To see why a process is failing or hanging.
148
* For reverse engineering a process.
149
* To find the files on which a program depends.
150
* For analyzing the performance of an application.
151
* For troubleshooting various problems related to the operating system.
153
In addition, strace can generate run-time statistics on times, calls, and
154
errors for each system call and report a summary when program exits,
155
suppressing the regular output. This attempts to show system time (CPU time
156
spent running in the kernel) independent of wall clock time. We plan to use
157
these features to get information on workload system usage.
159
strace command supports basic, verbose, and stats modes. strace command when
160
run in verbose mode gives more detailed information about the system calls
161
invoked by a process.
163
Running strace -c generates a report of the percentage of time spent in each
164
system call, the total time in seconds, the microseconds per call, the total
165
number of calls, the count of each system call that has failed with an error
166
and the type of system call made.
168
* Usage: strace <command we want to trace>
169
* Verbose mode usage: strace -v <command>
170
* Gather statistics: strace -c <command>
172
We used the “-c” option to gather fine-grained run-time statistics in use
173
by three workloads we have chose for this analysis.
175
* perf
176
* stress-ng
177
* paxtest
179
What is cscope and how do we use it?
180
====================================
182
Now let’s look at `cscope <https://cscope.sourceforge.net/>`_, a command
183
line tool for browsing C, C++ or Java code-bases. We can use it to find
184
all the references to a symbol, global definitions, functions called by a
185
function, functions calling a function, text strings, regular expression
186
patterns, files including a file.
188
We can use cscope to find which system call belongs to which subsystem.
189
This way we can find the kernel subsystems used by a process when it is
190
executed.
192
Let’s checkout the latest Linux repository and build cscope database::
194
git clone git://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git linux
195
cd linux
196
cscope -R -p10 # builds cscope.out database before starting browse session
197
cscope -d -p10 # starts browse session on cscope.out database
199
Note: Run "cscope -R -p10" to build the database and c"scope -d -p10" to
200
enter into the browsing session. cscope by default cscope.out database.
201
To get out of this mode press ctrl+d. -p option is used to specify the
202
number of file path components to display. -p10 is optimal for browsing
203
kernel sources.
205
What is perf and how do we use it?
206
==================================
208
Perf is an analysis tool based on Linux 2.6+ systems, which abstracts the
209
CPU hardware difference in performance measurement in Linux, and provides
210
a simple command line interface. Perf is based on the perf_events interface
211
exported by the kernel. It is very useful for profiling the system and
212
finding performance bottlenecks in an application.
214
If you haven't already checked out the Linux mainline repository, you can do
215
so and then build kernel and perf tool::
217
git clone git://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git linux
218
cd linux
219
make -j3 all
220
cd tools/perf
221
make
223
Note: The perf command can be built without building the kernel in the
224
repository and can be run on older kernels. However matching the kernel
225
and perf revisions gives more accurate information on the subsystem usage.
227
We used "perf stat" and "perf bench" options. For a detailed information on
228
the perf tool, run "perf -h".
230
perf stat
231
---------
232
The perf stat command generates a report of various hardware and software
233
events. It does so with the help of hardware counter registers found in
234
modern CPUs that keep the count of these activities. "perf stat cal" shows
235
stats for cal command.
237
Perf bench
238
----------
239
The perf bench command contains multiple multi-threaded microkernel
240
benchmarks for executing different subsystems in the Linux kernel and
241
system calls. This allows us to easily measure the impact of changes,
242
which can help mitigate performance regressions. It also acts as a common
243
benchmarking framework, enabling developers to easily create test cases,
244
integrate transparently, and use performance-rich tooling.
246
"perf bench all" command runs the following benchmarks:
248
* sched/messaging
249
* sched/pipe
250
* syscall/basic
251
* mem/memcpy
252
* mem/memset
254
What is stress-ng and how do we use it?
255
=======================================
257
As mentioned earlier, stress-ng is used for performing stress testing on
258
the kernel. It allows you to exercise various physical subsystems of the
259
computer, as well as interfaces of the OS kernel, using stressor-s. They
260
are available for CPU, CPU cache, devices, I/O, interrupts, file system,
261
memory, network, operating system, pipelines, schedulers, and virtual
262
machines.
264
The netdev stressor starts N workers that exercise various netdevice ioctl
265
commands across all the available network devices. The following ioctls are
266
exercised:
268
* SIOCGIFCONF, SIOCGIFINDEX, SIOCGIFNAME, SIOCGIFFLAGS
269
* SIOCGIFADDR, SIOCGIFNETMASK, SIOCGIFMETRIC, SIOCGIFMTU
270
* SIOCGIFHWADDR, SIOCGIFMAP, SIOCGIFTXQLEN
272
The following command runs the stressor::
274
stress-ng --netdev 1 -t 60 --metrics command.
276
We can use the perf record command to record the events and information
277
associated with a process. This command records the profiling data in the
278
perf.data file in the same directory.
280
Using the following commands you can record the events associated with the
281
netdev stressor, view the generated report perf.data and annotate the to
282
view the statistics of each instruction of the program::
284
perf record stress-ng --netdev 1 -t 60 --metrics command.
285
perf report
286
perf annotate
288
What is paxtest and how do we use it?
289
=====================================
291
paxtest is a program that tests buffer overflows in the kernel. It tests
292
kernel enforcements over memory usage. Generally, execution in some memory
293
segments makes buffer overflows possible. It runs a set of programs that
294
attempt to subvert memory usage. It is used as a regression test suite for
295
PaX, and will be useful to test other memory protection patches for the
296
kernel.
298
paxtest provides kiddie and blackhat modes. The paxtest kiddie mode runs
299
in normal mode, whereas the blackhat mode tries to get around the protection
300
of the kernel testing for vulnerabilities. We focus on the kiddie mode here
301
and combine "paxtest kiddie" run with "perf record" to collect CPU stack
302
traces for the paxtest kiddie run to see which function is calling other
303
functions in the performance profile. Then the "dwarf" (DWARF's Call Frame
304
Information) mode can be used to unwind the stack.
306
The following command can be used to view resulting report in call-graph
307
format::
309
perf record --call-graph dwarf paxtest kiddie
310
perf report --stdio
312
Tracing workloads
313
=================
315
Now that we understand the workloads, let's start tracing them.
317
Tracing perf bench all workload
318
-------------------------------
320
Run the following command to trace perf bench all workload::
322
strace -c perf bench all
324
**System Calls made by the workload**
326
The below table shows the system calls invoked by the workload, number of
327
times each system call is invoked, and the corresponding Linux subsystem.
329
+-------------------+-----------+-----------------+-------------------------+
330
| System Call | # calls | Linux Subsystem | System Call (API) |
331
+===================+===========+=================+=========================+
332
| getppid | 10000001 | Process Mgmt | sys_getpid() |
333
+-------------------+-----------+-----------------+-------------------------+
334
| clone | 1077 | Process Mgmt. | sys_clone() |
335
+-------------------+-----------+-----------------+-------------------------+
336
| prctl | 23 | Process Mgmt. | sys_prctl() |
337
+-------------------+-----------+-----------------+-------------------------+
338
| prlimit64 | 7 | Process Mgmt. | sys_prlimit64() |
339
+-------------------+-----------+-----------------+-------------------------+
340
| getpid | 10 | Process Mgmt. | sys_getpid() |
341
+-------------------+-----------+-----------------+-------------------------+
342
| uname | 3 | Process Mgmt. | sys_uname() |
343
+-------------------+-----------+-----------------+-------------------------+
344
| sysinfo | 1 | Process Mgmt. | sys_sysinfo() |
345
+-------------------+-----------+-----------------+-------------------------+
346
| getuid | 1 | Process Mgmt. | sys_getuid() |
347
+-------------------+-----------+-----------------+-------------------------+
348
| getgid | 1 | Process Mgmt. | sys_getgid() |
349
+-------------------+-----------+-----------------+-------------------------+
350
| geteuid | 1 | Process Mgmt. | sys_geteuid() |
351
+-------------------+-----------+-----------------+-------------------------+
352
| getegid | 1 | Process Mgmt. | sys_getegid |
353
+-------------------+-----------+-----------------+-------------------------+
354
| close | 49951 | Filesystem | sys_close() |
355
+-------------------+-----------+-----------------+-------------------------+
356
| pipe | 604 | Filesystem | sys_pipe() |
357
+-------------------+-----------+-----------------+-------------------------+
358
| openat | 48560 | Filesystem | sys_opennat() |
359
+-------------------+-----------+-----------------+-------------------------+
360
| fstat | 8338 | Filesystem | sys_fstat() |
361
+-------------------+-----------+-----------------+-------------------------+
362
| stat | 1573 | Filesystem | sys_stat() |
363
+-------------------+-----------+-----------------+-------------------------+
364
| pread64 | 9646 | Filesystem | sys_pread64() |
365
+-------------------+-----------+-----------------+-------------------------+
366
| getdents64 | 1873 | Filesystem | sys_getdents64() |
367
+-------------------+-----------+-----------------+-------------------------+
368
| access | 3 | Filesystem | sys_access() |
369
+-------------------+-----------+-----------------+-------------------------+
370
| lstat | 1880 | Filesystem | sys_lstat() |
371
+-------------------+-----------+-----------------+-------------------------+
372
| lseek | 6 | Filesystem | sys_lseek() |
373
+-------------------+-----------+-----------------+-------------------------+
374
| ioctl | 3 | Filesystem | sys_ioctl() |
375
+-------------------+-----------+-----------------+-------------------------+
376
| dup2 | 1 | Filesystem | sys_dup2() |
377
+-------------------+-----------+-----------------+-------------------------+
378
| execve | 2 | Filesystem | sys_execve() |
379
+-------------------+-----------+-----------------+-------------------------+
380
| fcntl | 8779 | Filesystem | sys_fcntl() |
381
+-------------------+-----------+-----------------+-------------------------+
382
| statfs | 1 | Filesystem | sys_statfs() |
383
+-------------------+-----------+-----------------+-------------------------+
384
| epoll_create | 2 | Filesystem | sys_epoll_create() |
385
+-------------------+-----------+-----------------+-------------------------+
386
| epoll_ctl | 64 | Filesystem | sys_epoll_ctl() |
387
+-------------------+-----------+-----------------+-------------------------+
388
| newfstatat | 8318 | Filesystem | sys_newfstatat() |
389
+-------------------+-----------+-----------------+-------------------------+
390
| eventfd2 | 192 | Filesystem | sys_eventfd2() |
391
+-------------------+-----------+-----------------+-------------------------+
392
| mmap | 243 | Memory Mgmt. | sys_mmap() |
393
+-------------------+-----------+-----------------+-------------------------+
394
| mprotect | 32 | Memory Mgmt. | sys_mprotect() |
395
+-------------------+-----------+-----------------+-------------------------+
396
| brk | 21 | Memory Mgmt. | sys_brk() |
397
+-------------------+-----------+-----------------+-------------------------+
398
| munmap | 128 | Memory Mgmt. | sys_munmap() |
399
+-------------------+-----------+-----------------+-------------------------+
400
| set_mempolicy | 156 | Memory Mgmt. | sys_set_mempolicy() |
401
+-------------------+-----------+-----------------+-------------------------+
402
| set_tid_address | 1 | Process Mgmt. | sys_set_tid_address() |
403
+-------------------+-----------+-----------------+-------------------------+
404
| set_robust_list | 1 | Futex | sys_set_robust_list() |
405
+-------------------+-----------+-----------------+-------------------------+
406
| futex | 341 | Futex | sys_futex() |
407
+-------------------+-----------+-----------------+-------------------------+
408
| sched_getaffinity | 79 | Scheduler | sys_sched_getaffinity() |
409
+-------------------+-----------+-----------------+-------------------------+
410
| sched_setaffinity | 223 | Scheduler | sys_sched_setaffinity() |
411
+-------------------+-----------+-----------------+-------------------------+
412
| socketpair | 202 | Network | sys_socketpair() |
413
+-------------------+-----------+-----------------+-------------------------+
414
| rt_sigprocmask | 21 | Signal | sys_rt_sigprocmask() |
415
+-------------------+-----------+-----------------+-------------------------+
416
| rt_sigaction | 36 | Signal | sys_rt_sigaction() |
417
+-------------------+-----------+-----------------+-------------------------+
418
| rt_sigreturn | 2 | Signal | sys_rt_sigreturn() |
419
+-------------------+-----------+-----------------+-------------------------+
420
| wait4 | 889 | Time | sys_wait4() |
421
+-------------------+-----------+-----------------+-------------------------+
422
| clock_nanosleep | 37 | Time | sys_clock_nanosleep() |
423
+-------------------+-----------+-----------------+-------------------------+
424
| capget | 4 | Capability | sys_capget() |
425
+-------------------+-----------+-----------------+-------------------------+
427
Tracing stress-ng netdev stressor workload
428
------------------------------------------
430
Run the following command to trace stress-ng netdev stressor workload::
432
strace -c stress-ng --netdev 1 -t 60 --metrics
434
**System Calls made by the workload**
436
The below table shows the system calls invoked by the workload, number of
437
times each system call is invoked, and the corresponding Linux subsystem.
439
+-------------------+-----------+-----------------+-------------------------+
440
| System Call | # calls | Linux Subsystem | System Call (API) |
441
+===================+===========+=================+=========================+
442
| openat | 74 | Filesystem | sys_openat() |
443
+-------------------+-----------+-----------------+-------------------------+
444
| close | 75 | Filesystem | sys_close() |
445
+-------------------+-----------+-----------------+-------------------------+
446
| read | 58 | Filesystem | sys_read() |
447
+-------------------+-----------+-----------------+-------------------------+
448
| fstat | 20 | Filesystem | sys_fstat() |
449
+-------------------+-----------+-----------------+-------------------------+
450
| flock | 10 | Filesystem | sys_flock() |
451
+-------------------+-----------+-----------------+-------------------------+
452
| write | 7 | Filesystem | sys_write() |
453
+-------------------+-----------+-----------------+-------------------------+
454
| getdents64 | 8 | Filesystem | sys_getdents64() |
455
+-------------------+-----------+-----------------+-------------------------+
456
| pread64 | 8 | Filesystem | sys_pread64() |
457
+-------------------+-----------+-----------------+-------------------------+
458
| lseek | 1 | Filesystem | sys_lseek() |
459
+-------------------+-----------+-----------------+-------------------------+
460
| access | 2 | Filesystem | sys_access() |
461
+-------------------+-----------+-----------------+-------------------------+
462
| getcwd | 1 | Filesystem | sys_getcwd() |
463
+-------------------+-----------+-----------------+-------------------------+
464
| execve | 1 | Filesystem | sys_execve() |
465
+-------------------+-----------+-----------------+-------------------------+
466
| mmap | 61 | Memory Mgmt. | sys_mmap() |
467
+-------------------+-----------+-----------------+-------------------------+
468
| munmap | 3 | Memory Mgmt. | sys_munmap() |
469
+-------------------+-----------+-----------------+-------------------------+
470
| mprotect | 20 | Memory Mgmt. | sys_mprotect() |
471
+-------------------+-----------+-----------------+-------------------------+
472
| mlock | 2 | Memory Mgmt. | sys_mlock() |
473
+-------------------+-----------+-----------------+-------------------------+
474
| brk | 3 | Memory Mgmt. | sys_brk() |
475
+-------------------+-----------+-----------------+-------------------------+
476
| rt_sigaction | 21 | Signal | sys_rt_sigaction() |
477
+-------------------+-----------+-----------------+-------------------------+
478
| rt_sigprocmask | 1 | Signal | sys_rt_sigprocmask() |
479
+-------------------+-----------+-----------------+-------------------------+
480
| sigaltstack | 1 | Signal | sys_sigaltstack() |
481
+-------------------+-----------+-----------------+-------------------------+
482
| rt_sigreturn | 1 | Signal | sys_rt_sigreturn() |
483
+-------------------+-----------+-----------------+-------------------------+
484
| getpid | 8 | Process Mgmt. | sys_getpid() |
485
+-------------------+-----------+-----------------+-------------------------+
486
| prlimit64 | 5 | Process Mgmt. | sys_prlimit64() |
487
+-------------------+-----------+-----------------+-------------------------+
488
| arch_prctl | 2 | Process Mgmt. | sys_arch_prctl() |
489
+-------------------+-----------+-----------------+-------------------------+
490
| sysinfo | 2 | Process Mgmt. | sys_sysinfo() |
491
+-------------------+-----------+-----------------+-------------------------+
492
| getuid | 2 | Process Mgmt. | sys_getuid() |
493
+-------------------+-----------+-----------------+-------------------------+
494
| uname | 1 | Process Mgmt. | sys_uname() |
495
+-------------------+-----------+-----------------+-------------------------+
496
| setpgid | 1 | Process Mgmt. | sys_setpgid() |
497
+-------------------+-----------+-----------------+-------------------------+
498
| getrusage | 1 | Process Mgmt. | sys_getrusage() |
499
+-------------------+-----------+-----------------+-------------------------+
500
| geteuid | 1 | Process Mgmt. | sys_geteuid() |
501
+-------------------+-----------+-----------------+-------------------------+
502
| getppid | 1 | Process Mgmt. | sys_getppid() |
503
+-------------------+-----------+-----------------+-------------------------+
504
| sendto | 3 | Network | sys_sendto() |
505
+-------------------+-----------+-----------------+-------------------------+
506
| connect | 1 | Network | sys_connect() |
507
+-------------------+-----------+-----------------+-------------------------+
508
| socket | 1 | Network | sys_socket() |
509
+-------------------+-----------+-----------------+-------------------------+
510
| clone | 1 | Process Mgmt. | sys_clone() |
511
+-------------------+-----------+-----------------+-------------------------+
512
| set_tid_address | 1 | Process Mgmt. | sys_set_tid_address() |
513
+-------------------+-----------+-----------------+-------------------------+
514
| wait4 | 2 | Time | sys_wait4() |
515
+-------------------+-----------+-----------------+-------------------------+
516
| alarm | 1 | Time | sys_alarm() |
517
+-------------------+-----------+-----------------+-------------------------+
518
| set_robust_list | 1 | Futex | sys_set_robust_list() |
519
+-------------------+-----------+-----------------+-------------------------+
521
Tracing paxtest kiddie workload
522
-------------------------------
524
Run the following command to trace paxtest kiddie workload::
526
strace -c paxtest kiddie
528
**System Calls made by the workload**
530
The below table shows the system calls invoked by the workload, number of
531
times each system call is invoked, and the corresponding Linux subsystem.
533
+-------------------+-----------+-----------------+----------------------+
534
| System Call | # calls | Linux Subsystem | System Call (API) |
535
+===================+===========+=================+======================+
536
| read | 3 | Filesystem | sys_read() |
537
+-------------------+-----------+-----------------+----------------------+
538
| write | 11 | Filesystem | sys_write() |
539
+-------------------+-----------+-----------------+----------------------+
540
| close | 41 | Filesystem | sys_close() |
541
+-------------------+-----------+-----------------+----------------------+
542
| stat | 24 | Filesystem | sys_stat() |
543
+-------------------+-----------+-----------------+----------------------+
544
| fstat | 2 | Filesystem | sys_fstat() |
545
+-------------------+-----------+-----------------+----------------------+
546
| pread64 | 6 | Filesystem | sys_pread64() |
547
+-------------------+-----------+-----------------+----------------------+
548
| access | 1 | Filesystem | sys_access() |
549
+-------------------+-----------+-----------------+----------------------+
550
| pipe | 1 | Filesystem | sys_pipe() |
551
+-------------------+-----------+-----------------+----------------------+
552
| dup2 | 24 | Filesystem | sys_dup2() |
553
+-------------------+-----------+-----------------+----------------------+
554
| execve | 1 | Filesystem | sys_execve() |
555
+-------------------+-----------+-----------------+----------------------+
556
| fcntl | 26 | Filesystem | sys_fcntl() |
557
+-------------------+-----------+-----------------+----------------------+
558
| openat | 14 | Filesystem | sys_openat() |
559
+-------------------+-----------+-----------------+----------------------+
560
| rt_sigaction | 7 | Signal | sys_rt_sigaction() |
561
+-------------------+-----------+-----------------+----------------------+
562
| rt_sigreturn | 38 | Signal | sys_rt_sigreturn() |
563
+-------------------+-----------+-----------------+----------------------+
564
| clone | 38 | Process Mgmt. | sys_clone() |
565
+-------------------+-----------+-----------------+----------------------+
566
| wait4 | 44 | Time | sys_wait4() |
567
+-------------------+-----------+-----------------+----------------------+
568
| mmap | 7 | Memory Mgmt. | sys_mmap() |
569
+-------------------+-----------+-----------------+----------------------+
570
| mprotect | 3 | Memory Mgmt. | sys_mprotect() |
571
+-------------------+-----------+-----------------+----------------------+
572
| munmap | 1 | Memory Mgmt. | sys_munmap() |
573
+-------------------+-----------+-----------------+----------------------+
574
| brk | 3 | Memory Mgmt. | sys_brk() |
575
+-------------------+-----------+-----------------+----------------------+
576
| getpid | 1 | Process Mgmt. | sys_getpid() |
577
+-------------------+-----------+-----------------+----------------------+
578
| getuid | 1 | Process Mgmt. | sys_getuid() |
579
+-------------------+-----------+-----------------+----------------------+
580
| getgid | 1 | Process Mgmt. | sys_getgid() |
581
+-------------------+-----------+-----------------+----------------------+
582
| geteuid | 2 | Process Mgmt. | sys_geteuid() |
583
+-------------------+-----------+-----------------+----------------------+
584
| getegid | 1 | Process Mgmt. | sys_getegid() |
585
+-------------------+-----------+-----------------+----------------------+
586
| getppid | 1 | Process Mgmt. | sys_getppid() |
587
+-------------------+-----------+-----------------+----------------------+
588
| arch_prctl | 2 | Process Mgmt. | sys_arch_prctl() |
589
+-------------------+-----------+-----------------+----------------------+
591
Conclusion
592
==========
594
This document is intended to be used as a guide on how to gather fine-grained
595
information on the resources in use by workloads using strace.
597
References
598
==========
600
* `Discovery Linux Kernel Subsystems used by OpenAPS <https://elisa.tech/blog/2022/02/02/discovery-linux-kernel-subsystems-used-by-openaps>`_
601
* `ELISA-White-Papers-Discovering Linux kernel subsystems used by a workload <https://github.com/elisa-tech/ELISA-White-Papers/blob/master/Processes/Discovering_Linux_kernel_subsystems_used_by_a_workload.md>`_
602
* `strace <https://man7.org/linux/man-pages/man1/strace.1.html>`_
603
* `perf <https://man7.org/linux/man-pages/man1/perf.1.html>`_
604
* `paxtest README <https://github.com/opntr/paxtest-freebsd/blob/hardenedbsd/0.9.14-hbsd/README>`_
605
* `stress-ng <https://www.mankier.com/1/stress-ng>`_
606
* `Monitoring and managing system status and performance <https://access.redhat.com/documentation/en-us/red_hat_enterprise_linux/8/html/monitoring_and_managing_system_status_and_performance/index>`_
3. 한국어 전문 번역
영어 원문의 문단 순서와 의미를 유지한 전체 번역입니다. 코드, 함수명, symbol과 URL은 원문 표기를 유지합니다.
Discovering Linux kernel subsystems used by a workload
1-10Shuah Khan과 Shefali Sharma가 작성했으며 Shuah Khan이 관리하는 문서입니다. workload를 추적해 실제로 사용하는 Linux kernel subsystem과 system resource를 식별하는 방법을 설명합니다. 라이선스는 `(GPL-2.0+ OR CC-BY-4.0)`입니다.
Key Points
11-26workload를 빌드하고 실행하는 데 필요한 system resource를 이해하는 것이 중요합니다. Linux tracing과 strace로 사용 resource를 찾을 수 있지만, 정보의 완전성은 workload coverage의 완전성에 달려 있습니다.
`perf`, `stress-ng`, `paxtest` 같은 도구로 운영체제의 performance와 security를 분석할 수 있습니다. 필요를 발견하고 이해한 뒤 그 경로에 집중하면 regression을 피하고 safety 고려사항을 평가할 수 있습니다.
Methodology
27-49`strace`는 진단·교육·debugging 도구이며 workload가 쓰는 system resource를 찾는 데 사용할 수 있습니다. 이 문서의 방법은 strace로 workload를 추적합니다.
한 번의 trace는 workload가 그 실행에서 호출한 system call과 실제로 지난 code path만 보여 줍니다. file open과 read가 성공했다면 success path만 추적되며 error path는 포함되지 않습니다. workload가 모든 경우를 완전히 cover할 때만 가능한 모든 경로를 찾을 수 있습니다.
목표는 custom kernel을 설치하지 않고 기본 kernel을 실행하는 시스템에서 workload를 추적하는 것입니다.
How do we gather fine-grained system information?
50-70strace는 process가 호출한 system call과 받은 signal을 추적합니다. system call은 application과 kernel 사이의 기본 interface이며, 예를 들어 `open()`은 filesystem의 file에 접근하게 합니다. strace는 호출과 그 결과를 모두 나열합니다.
strace와 `perf record`를 결합하면 process 관련 event와 profiling data를 기록할 수 있습니다. `perf annotate`는 program instruction별 통계를 만듭니다. 문서는 perf, stress-ng, paxtest workload를 예시로 사용하지만 같은 절차를 다른 workload에도 적용할 수 있습니다.
Getting the system ready for tracing
71-96물리 시스템이나 virtual machine에 Linux 배포판이 실행 중이라고 가정합니다. 대부분 strace는 포함하지만 kernel과 repository 도구를 빌드할 package는 별도 설치해야 합니다. 아래 명령은 Debian 계열 기준이므로 다른 배포판에서는 대응 package를 찾습니다.
`scripts/ver_linux`는 필요한 build tool이 이미 있는지 확인하기 좋습니다. Linux kernel과 repository 도구의 의존성을 설치합니다.
sudo apt-get install build-essential flex bison yacc
sudo apt install libelf-dev systemtap-sdt-dev libslang2-dev libperl-dev libdw-devkernel source 탐색용 cscope를 설치합니다.
sudo apt-get install cscopestress-ng와 paxtest를 설치합니다.
apt-get install stress-ng
apt-get install paxtestWorkload overview
97-105perf bench, stress-ng, paxtest를 strace로 추적해 workload를 분석하고 사용하는 Linux subsystem을 식별합니다. 먼저 세 workload의 목적과 사용법을 살펴봅니다.
perf bench (all) workload
106-115`perf bench`는 여러 Linux subsystem과 system call을 실행하는 multi-threaded microkernel benchmark 묶음입니다. 변경의 영향을 쉽게 측정해 performance regression을 줄이고, 개발자가 test case를 만들고 통합하며 perf 도구를 활용하는 공통 framework 역할을 합니다.
Stress-ng netdev stressor workload
116-128stress-ng는 CPU, cache, device, I/O, interrupt, filesystem, memory, network, OS, pipeline, scheduler, VM 등 물리 subsystem과 kernel interface를 stressor로 시험합니다. 전체 stressor 설명은 stress-ng man page에 있습니다.
netdev stressor는 N개의 worker를 시작해 사용 가능한 모든 network device에 여러 netdevice ioctl을 실행합니다.
paxtest kiddie workload
129-138paxtest는 kernel의 buffer overflow와 memory 사용 강제를 시험합니다. 실행 가능한 memory segment를 악용하려는 program 집합을 실행하며 PaX regression suite이지만 다른 kernel memory protection patch 시험에도 쓸 수 있습니다. 여기서는 단순 vulnerability를 찾는 kiddie mode를 사용합니다.
What is strace and how do we use it?
139-178strace는 process와 kernel의 상호작용, 실패나 hang 원인, reverse engineering, program 의존 file, application performance, 운영체제 문제를 조사할 수 있습니다.
program 종료 시 system call별 시간, 호출, error 통계를 요약하고 일반 출력을 숨길 수도 있습니다. 이 통계는 wall clock과 별개로 kernel에서 사용한 system CPU time을 보여 주려 합니다.
basic, verbose, stats mode가 있으며 verbose mode는 호출된 system call을 더 자세히 보여 줍니다. `strace -c`는 call별 시간 비율, 총 seconds, call당 microseconds, 총 호출 수, 실패 수, system call 종류를 보고합니다.
- 기본: `strace <command we want to trace>`
- 상세: `strace -v <command>`
- 통계: `strace -c <command>`
문서는 `-c` option으로 perf, stress-ng, paxtest 세 workload의 세밀한 runtime 통계를 수집합니다.
What is cscope and how do we use it?
179-204cscope는 C, C++, Java codebase용 command-line 탐색 도구입니다. symbol 참조와 global definition, 함수의 caller·callee, text·regular expression, file include 관계를 찾을 수 있습니다. 이를 통해 system call이 어느 subsystem에 속하는지 연결합니다.
최신 Linux repository를 받고 cscope database를 만듭니다.
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`cscope -R -p10`은 `cscope.out`을 만들고, `cscope -d -p10`은 기존 database로 browse session을 시작합니다. `Ctrl+D`로 나옵니다. `-p`는 표시할 path component 수이며 kernel source에는 `-p10`이 적절합니다.
What is perf and how do we use it?
205-229Perf는 Linux 2.6+의 `perf_events` interface를 기반으로 CPU hardware 차이를 추상화한 command-line 분석 도구입니다. system profiling과 application performance bottleneck 탐색에 유용합니다.
mainline repository가 없다면 kernel과 perf를 함께 빌드합니다.
git clone git://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git linux
cd linux
make -j3 all
cd tools/perf
makeperf command만 따로 빌드해 오래된 kernel에서 실행할 수도 있지만, kernel과 perf revision을 맞추면 subsystem 사용량을 더 정확히 알 수 있습니다. 문서는 `perf stat`과 `perf bench`를 사용하며 자세한 도움말은 `perf -h`입니다.
perf stat
230-236`perf stat`은 현대 CPU의 hardware counter register를 이용해 여러 hardware·software event를 보고합니다. `perf stat cal`은 `cal` 명령의 통계를 보여 줍니다.
Perf bench
237-253`perf bench`는 kernel subsystem과 system call을 실행하는 multi-threaded microkernel benchmark입니다. `perf bench all`은 다음 benchmark를 실행합니다.
- `sched/messaging`
- `sched/pipe`
- `syscall/basic`
- `mem/memcpy`
- `mem/memset`
What is stress-ng and how do we use it?
254-287stress-ng는 여러 hardware subsystem과 OS kernel interface를 stressor로 부하 시험합니다. netdev stressor는 N개 worker가 모든 network device에서 ioctl을 수행합니다.
대상 ioctl은 `SIOCGIFCONF`, `SIOCGIFINDEX`, `SIOCGIFNAME`, `SIOCGIFFLAGS`, `SIOCGIFADDR`, `SIOCGIFNETMASK`, `SIOCGIFMETRIC`, `SIOCGIFMTU`, `SIOCGIFHWADDR`, `SIOCGIFMAP`, `SIOCGIFTXQLEN`입니다.
worker 1개를 60초 실행하고 metric을 수집합니다.
stress-ng --netdev 1 -t 60 --metrics command.`perf record`는 같은 directory의 `perf.data`에 process event와 profiling data를 기록합니다. report를 열고 instruction별 통계를 annotate합니다.
perf record stress-ng --netdev 1 -t 60 --metrics command.
perf report
perf annotateWhat is paxtest and how do we use it?
288-311paxtest는 memory segment 실행을 악용하는 program들로 buffer overflow와 memory protection을 시험합니다. kiddie는 일반 mode이고 blackhat은 kernel protection을 우회하려 시도합니다.
여기서는 `paxtest kiddie`와 `perf record`를 결합해 CPU stack trace와 function call 관계를 수집합니다. `dwarf`, 즉 DWARF Call Frame Information mode로 stack을 unwind하고 call graph를 text report로 봅니다.
perf record --call-graph dwarf paxtest kiddie
perf report --stdioTracing workloads
312-316세 workload의 성격을 이해했으므로 이제 실제 trace와 subsystem 분류를 수행합니다.
Tracing perf bench all workload
317-426perf bench 전체 workload를 system call 통계 mode로 추적합니다.
strace -c perf bench all아래 표는 이 실행에서 호출된 system call, 호출 횟수, 대응 Linux subsystem, kernel API를 원문 값 그대로 보존합니다.
| System Call | # calls | Linux Subsystem | System Call (API) |
|---|---|---|---|
| getppid | 10000001 | Process Mgmt | sys_getpid() |
| clone | 1077 | Process Mgmt. | sys_clone() |
| prctl | 23 | Process Mgmt. | sys_prctl() |
| prlimit64 | 7 | Process Mgmt. | sys_prlimit64() |
| getpid | 10 | Process Mgmt. | sys_getpid() |
| uname | 3 | Process Mgmt. | sys_uname() |
| sysinfo | 1 | Process Mgmt. | sys_sysinfo() |
| getuid | 1 | Process Mgmt. | sys_getuid() |
| getgid | 1 | Process Mgmt. | sys_getgid() |
| geteuid | 1 | Process Mgmt. | sys_geteuid() |
| getegid | 1 | Process Mgmt. | sys_getegid |
| close | 49951 | Filesystem | sys_close() |
| pipe | 604 | Filesystem | sys_pipe() |
| openat | 48560 | Filesystem | sys_opennat() |
| fstat | 8338 | Filesystem | sys_fstat() |
| stat | 1573 | Filesystem | sys_stat() |
| pread64 | 9646 | Filesystem | sys_pread64() |
| getdents64 | 1873 | Filesystem | sys_getdents64() |
| access | 3 | Filesystem | sys_access() |
| lstat | 1880 | Filesystem | sys_lstat() |
| lseek | 6 | Filesystem | sys_lseek() |
| ioctl | 3 | Filesystem | sys_ioctl() |
| dup2 | 1 | Filesystem | sys_dup2() |
| execve | 2 | Filesystem | sys_execve() |
| fcntl | 8779 | Filesystem | sys_fcntl() |
| statfs | 1 | Filesystem | sys_statfs() |
| epoll_create | 2 | Filesystem | sys_epoll_create() |
| epoll_ctl | 64 | Filesystem | sys_epoll_ctl() |
| newfstatat | 8318 | Filesystem | sys_newfstatat() |
| eventfd2 | 192 | Filesystem | sys_eventfd2() |
| mmap | 243 | Memory Mgmt. | sys_mmap() |
| mprotect | 32 | Memory Mgmt. | sys_mprotect() |
| brk | 21 | Memory Mgmt. | sys_brk() |
| munmap | 128 | Memory Mgmt. | sys_munmap() |
| set_mempolicy | 156 | Memory Mgmt. | sys_set_mempolicy() |
| set_tid_address | 1 | Process Mgmt. | sys_set_tid_address() |
| set_robust_list | 1 | Futex | sys_set_robust_list() |
| futex | 341 | Futex | sys_futex() |
| sched_getaffinity | 79 | Scheduler | sys_sched_getaffinity() |
| sched_setaffinity | 223 | Scheduler | sys_sched_setaffinity() |
| socketpair | 202 | Network | sys_socketpair() |
| rt_sigprocmask | 21 | Signal | sys_rt_sigprocmask() |
| rt_sigaction | 36 | Signal | sys_rt_sigaction() |
| rt_sigreturn | 2 | Signal | sys_rt_sigreturn() |
| wait4 | 889 | Time | sys_wait4() |
| clock_nanosleep | 37 | Time | sys_clock_nanosleep() |
| capget | 4 | Capability | sys_capget() |
Tracing stress-ng netdev stressor workload
427-520netdev worker 1개를 60초 실행하며 strace 통계를 수집합니다.
strace -c stress-ng --netdev 1 -t 60 --metrics결과는 filesystem과 memory management 호출이 중심이고 signal, process management, network, time, futex 호출도 포함합니다.
| System Call | # calls | Linux Subsystem | System Call (API) |
|---|---|---|---|
| openat | 74 | Filesystem | sys_openat() |
| close | 75 | Filesystem | sys_close() |
| read | 58 | Filesystem | sys_read() |
| fstat | 20 | Filesystem | sys_fstat() |
| flock | 10 | Filesystem | sys_flock() |
| write | 7 | Filesystem | sys_write() |
| getdents64 | 8 | Filesystem | sys_getdents64() |
| pread64 | 8 | Filesystem | sys_pread64() |
| lseek | 1 | Filesystem | sys_lseek() |
| access | 2 | Filesystem | sys_access() |
| getcwd | 1 | Filesystem | sys_getcwd() |
| execve | 1 | Filesystem | sys_execve() |
| mmap | 61 | Memory Mgmt. | sys_mmap() |
| munmap | 3 | Memory Mgmt. | sys_munmap() |
| mprotect | 20 | Memory Mgmt. | sys_mprotect() |
| mlock | 2 | Memory Mgmt. | sys_mlock() |
| brk | 3 | Memory Mgmt. | sys_brk() |
| rt_sigaction | 21 | Signal | sys_rt_sigaction() |
| rt_sigprocmask | 1 | Signal | sys_rt_sigprocmask() |
| sigaltstack | 1 | Signal | sys_sigaltstack() |
| rt_sigreturn | 1 | Signal | sys_rt_sigreturn() |
| getpid | 8 | Process Mgmt. | sys_getpid() |
| prlimit64 | 5 | Process Mgmt. | sys_prlimit64() |
| arch_prctl | 2 | Process Mgmt. | sys_arch_prctl() |
| sysinfo | 2 | Process Mgmt. | sys_sysinfo() |
| getuid | 2 | Process Mgmt. | sys_getuid() |
| uname | 1 | Process Mgmt. | sys_uname() |
| setpgid | 1 | Process Mgmt. | sys_setpgid() |
| getrusage | 1 | Process Mgmt. | sys_getrusage() |
| geteuid | 1 | Process Mgmt. | sys_geteuid() |
| getppid | 1 | Process Mgmt. | sys_getppid() |
| sendto | 3 | Network | sys_sendto() |
| connect | 1 | Network | sys_connect() |
| socket | 1 | Network | sys_socket() |
| clone | 1 | Process Mgmt. | sys_clone() |
| set_tid_address | 1 | Process Mgmt. | sys_set_tid_address() |
| wait4 | 2 | Time | sys_wait4() |
| alarm | 1 | Time | sys_alarm() |
| set_robust_list | 1 | Futex | sys_set_robust_list() |
Tracing paxtest kiddie workload
521-590paxtest kiddie workload를 strace 통계 mode로 추적합니다.
strace -c paxtest kiddie결과는 filesystem, signal, process management, time, memory management 호출을 보여 줍니다.
| System Call | # calls | Linux Subsystem | System Call (API) |
|---|---|---|---|
| read | 3 | Filesystem | sys_read() |
| write | 11 | Filesystem | sys_write() |
| close | 41 | Filesystem | sys_close() |
| stat | 24 | Filesystem | sys_stat() |
| fstat | 2 | Filesystem | sys_fstat() |
| pread64 | 6 | Filesystem | sys_pread64() |
| access | 1 | Filesystem | sys_access() |
| pipe | 1 | Filesystem | sys_pipe() |
| dup2 | 24 | Filesystem | sys_dup2() |
| execve | 1 | Filesystem | sys_execve() |
| fcntl | 26 | Filesystem | sys_fcntl() |
| openat | 14 | Filesystem | sys_openat() |
| rt_sigaction | 7 | Signal | sys_rt_sigaction() |
| rt_sigreturn | 38 | Signal | sys_rt_sigreturn() |
| clone | 38 | Process Mgmt. | sys_clone() |
| wait4 | 44 | Time | sys_wait4() |
| mmap | 7 | Memory Mgmt. | sys_mmap() |
| mprotect | 3 | Memory Mgmt. | sys_mprotect() |
| munmap | 1 | Memory Mgmt. | sys_munmap() |
| brk | 3 | Memory Mgmt. | sys_brk() |
| getpid | 1 | Process Mgmt. | sys_getpid() |
| getuid | 1 | Process Mgmt. | sys_getuid() |
| getgid | 1 | Process Mgmt. | sys_getgid() |
| geteuid | 2 | Process Mgmt. | sys_geteuid() |
| getegid | 1 | Process Mgmt. | sys_getegid() |
| getppid | 1 | Process Mgmt. | sys_getppid() |
| arch_prctl | 2 | Process Mgmt. | sys_arch_prctl() |
Conclusion
591-596이 문서는 strace를 사용해 workload가 실제로 쓰는 resource에 관한 세밀한 정보를 수집하는 안내서입니다. 결과는 해당 실행의 coverage 범위 안에서 해석해야 합니다.
References
597-606OpenAPS와 ELISA white paper는 이 방법론의 실제 적용 배경을 제공하고, 각 도구의 manual과 README는 command option을 더 자세히 설명합니다. Red Hat 문서는 system 상태와 performance monitoring 전반을 보완합니다.
- Discovery Linux Kernel Subsystems used by OpenAPS
- ELISA White Paper: Discovering Linux kernel subsystems used by a workload
- strace manual
- perf manual
- paxtest README
- stress-ng manual
- Red Hat: Monitoring and managing system status and performance
요약과 해설
workload-tracing.rst:1-606strace는 실제 실행에서 지난 system call path를 보여 주고, cscope는 그 API를 kernel subsystem에 연결하며, perf는 event와 call graph를 보완합니다. 결과의 완전성은 workload가 success·error path를 얼마나 cover했는지에 좌우됩니다.