Debian Codegraph @

doc/SystemTap_Beginners_Guide/en-US/Useful_Scripts-kernelprofiling.xml

raw
1 <?xml version='1.0'?>
2 <!DOCTYPE section PUBLIC "-//OASIS//DTD DocBook XML V4.5//EN" "http://www.oasis-open.org/docbook/xml/4.5/docbookx.dtd" [
3 ]>
4
5
6 <section id="kernelprofsect">
7 <title>Kernel Profiling</title>
8
9
10 <remark>
11 WAR STORY: Kernel Profiling http://sourceware.org/systemtap/wiki/WSKernelProfile?highlight=((WarStories))
12 </remark>
13
14 <remark>
15 http://sourceware.org/systemtap/examples/process/pf2.stp
16 </remark>
17
18
19 <para>In <xref linkend="countcallssect"/>, you can revise the wildcards used in the probe to target <emphasis>all</emphasis> kernel functions. This can be useful if you are interested in indentifying what the kernel is performing over a specific time period.</para>
20
21 <para>However, doing so can cause considerable stress on the machine. In addition, this does not provide any indication of whether a specific function is being called too often during small time increments.</para>
22
23 <para>This section describes how to profile the kernel properly. <xref linkend="kernelprof"/> does this by providing a list of the top ten kernel functions called within a specific time period, and how many times each function was called during that time.</para>
24
25
26 <formalpara id="kernelprof">
27 <title>kernelprof.stp</title>
28 <para>
29 <programlisting>#! /usr/bin/env stap
30
31 global profile, pcount
32 probe timer.profile {
33 pcount &lt;&lt;&lt; 1
34 fn = probefunc ()
35 if (fn != "") profile[fn] &lt;&lt;&lt; 1
36 }
37 probe timer.ms(5000) {
38 printf ("\n--- %d samples recorded:\n", @count(pcount))
39 foreach (f in profile- limit 10) {
40 printf ("%-30s\t%6d\n", f, @count(profile[f]))
41 }
42 delete profile
43 delete pcount
44 }
45 </programlisting>
46 </para>
47 </formalpara>
48
49 <para><xref linkend="kernelprof"/> records and outputs kernel functions called every 5 seconds. You can change this setting by editing <command>probe timer.ms(5000)</command> accordingly. <xref linkend="kernelprofoutput"/> contains an excerpt of the output over a 20-second period:</para>
50
51 <example id="kernelprofoutput">
52 <title><xref linkend="kernelprof"/> Sample Output</title>
53 <screen>[...]
54 --- 10002 samples recorded:
55 sys_recvfrom 1
56 memmove 1
57 __copy_from_user_ll 17
58 __copy_to_user_ll 15
59 mwait_idle 5868
60 link_path_walk 1
61 kfree 1
62 fget_light 1
63 audit_syscall_exit 1
64 __d_lookup 1
65
66 --- 10002 samples recorded:
67 sysfs_read_file 1
68 free_poll_entry 1
69 syscall_exit_work 1
70 profile_hit 1
71 do_page_fault 2
72 _read_lock 3
73 kmap_atomic 4
74 strncpy_from_user 1
75 find_vma_prepare 1
76 __copy_from_user_ll 12
77
78 --- 10000 samples recorded:
79 system_call 4
80 unix_stream_sendmsg 1
81 __d_path 1
82 do_page_fault 2
83 kmap_atomic 1
84 find_vma 1
85 __copy_to_user_ll 16
86 __copy_from_user_ll 10
87 mwait_idle 5759
88 memcpy_fromiovec 1
89
90 --- 10004 samples recorded:
91 syscall_exit 1
92 __d_path 1
93 fput 2
94 do_page_fault 4
95 strncpy_from_user 1
96 may_open 1
97 ide_outb 1
98 ide_outsw 1
99 __copy_to_user_ll 11
100 __copy_from_user_ll 14
101 [...]</screen>
102 </example>
103
104 <para><xref linkend="kernelprof"/> allows you to identify interesting kernel-specific performance facts about the system, such as kernel-intensive workloads and idle times (in this case, <command>mwait_idle</command>. Note that <xref linkend="kernelprof"/> does not count user-space functions, although it counts all kernel functions as part of the sample (even if the script cannot identify from which specific kernel function a call originated.</para>
105
106 </section>
107