Root/Documentation/trace/ftrace-design.txt

1        function tracer guts
2        ====================
3        By Mike Frysinger
4
5Introduction
6------------
7
8Here we will cover the architecture pieces that the common function tracing
9code relies on for proper functioning. Things are broken down into increasing
10complexity so that you can start simple and at least get basic functionality.
11
12Note that this focuses on architecture implementation details only. If you
13want more explanation of a feature in terms of common code, review the common
14ftrace.txt file.
15
16
17Prerequisites
18-------------
19
20Ftrace relies on these features being implemented:
21 STACKTRACE_SUPPORT - implement save_stack_trace()
22 TRACE_IRQFLAGS_SUPPORT - implement include/asm/irqflags.h
23
24
25HAVE_FUNCTION_TRACER
26--------------------
27
28You will need to implement the mcount and the ftrace_stub functions.
29
30The exact mcount symbol name will depend on your toolchain. Some call it
31"mcount", "_mcount", or even "__mcount". You can probably figure it out by
32running something like:
33    $ echo 'main(){}' | gcc -x c -S -o - - -pg | grep mcount
34            call mcount
35We'll make the assumption below that the symbol is "mcount" just to keep things
36nice and simple in the examples.
37
38Keep in mind that the ABI that is in effect inside of the mcount function is
39*highly* architecture/toolchain specific. We cannot help you in this regard,
40sorry. Dig up some old documentation and/or find someone more familiar than
41you to bang ideas off of. Typically, register usage (argument/scratch/etc...)
42is a major issue at this point, especially in relation to the location of the
43mcount call (before/after function prologue). You might also want to look at
44how glibc has implemented the mcount function for your architecture. It might
45be (semi-)relevant.
46
47The mcount function should check the function pointer ftrace_trace_function
48to see if it is set to ftrace_stub. If it is, there is nothing for you to do,
49so return immediately. If it isn't, then call that function in the same way
50the mcount function normally calls __mcount_internal -- the first argument is
51the "frompc" while the second argument is the "selfpc" (adjusted to remove the
52size of the mcount call that is embedded in the function).
53
54For example, if the function foo() calls bar(), when the bar() function calls
55mcount(), the arguments mcount() will pass to the tracer are:
56    "frompc" - the address bar() will use to return to foo()
57    "selfpc" - the address bar() (with mcount() size adjustment)
58
59Also keep in mind that this mcount function will be called *a lot*, so
60optimizing for the default case of no tracer will help the smooth running of
61your system when tracing is disabled. So the start of the mcount function is
62typically the bare minimum with checking things before returning. That also
63means the code flow should usually be kept linear (i.e. no branching in the nop
64case). This is of course an optimization and not a hard requirement.
65
66Here is some pseudo code that should help (these functions should actually be
67implemented in assembly):
68
69void ftrace_stub(void)
70{
71    return;
72}
73
74void mcount(void)
75{
76    /* save any bare state needed in order to do initial checking */
77
78    extern void (*ftrace_trace_function)(unsigned long, unsigned long);
79    if (ftrace_trace_function != ftrace_stub)
80        goto do_trace;
81
82    /* restore any bare state */
83
84    return;
85
86do_trace:
87
88    /* save all state needed by the ABI (see paragraph above) */
89
90    unsigned long frompc = ...;
91    unsigned long selfpc = <return address> - MCOUNT_INSN_SIZE;
92    ftrace_trace_function(frompc, selfpc);
93
94    /* restore all state needed by the ABI */
95}
96
97Don't forget to export mcount for modules !
98extern void mcount(void);
99EXPORT_SYMBOL(mcount);
100
101
102HAVE_FUNCTION_TRACE_MCOUNT_TEST
103-------------------------------
104
105This is an optional optimization for the normal case when tracing is turned off
106in the system. If you do not enable this Kconfig option, the common ftrace
107code will take care of doing the checking for you.
108
109To support this feature, you only need to check the function_trace_stop
110variable in the mcount function. If it is non-zero, there is no tracing to be
111done at all, so you can return.
112
113This additional pseudo code would simply be:
114void mcount(void)
115{
116    /* save any bare state needed in order to do initial checking */
117
118+ if (function_trace_stop)
119+ return;
120
121    extern void (*ftrace_trace_function)(unsigned long, unsigned long);
122    if (ftrace_trace_function != ftrace_stub)
123...
124
125
126HAVE_FUNCTION_GRAPH_TRACER
127--------------------------
128
129Deep breath ... time to do some real work. Here you will need to update the
130mcount function to check ftrace graph function pointers, as well as implement
131some functions to save (hijack) and restore the return address.
132
133The mcount function should check the function pointers ftrace_graph_return
134(compare to ftrace_stub) and ftrace_graph_entry (compare to
135ftrace_graph_entry_stub). If either of those is not set to the relevant stub
136function, call the arch-specific function ftrace_graph_caller which in turn
137calls the arch-specific function prepare_ftrace_return. Neither of these
138function names is strictly required, but you should use them anyway to stay
139consistent across the architecture ports -- easier to compare & contrast
140things.
141
142The arguments to prepare_ftrace_return are slightly different than what are
143passed to ftrace_trace_function. The second argument "selfpc" is the same,
144but the first argument should be a pointer to the "frompc". Typically this is
145located on the stack. This allows the function to hijack the return address
146temporarily to have it point to the arch-specific function return_to_handler.
147That function will simply call the common ftrace_return_to_handler function and
148that will return the original return address with which you can return to the
149original call site.
150
151Here is the updated mcount pseudo code:
152void mcount(void)
153{
154...
155    if (ftrace_trace_function != ftrace_stub)
156        goto do_trace;
157
158+#ifdef CONFIG_FUNCTION_GRAPH_TRACER
159+ extern void (*ftrace_graph_return)(...);
160+ extern void (*ftrace_graph_entry)(...);
161+ if (ftrace_graph_return != ftrace_stub ||
162+ ftrace_graph_entry != ftrace_graph_entry_stub)
163+ ftrace_graph_caller();
164+#endif
165
166    /* restore any bare state */
167...
168
169Here is the pseudo code for the new ftrace_graph_caller assembly function:
170#ifdef CONFIG_FUNCTION_GRAPH_TRACER
171void ftrace_graph_caller(void)
172{
173    /* save all state needed by the ABI */
174
175    unsigned long *frompc = &...;
176    unsigned long selfpc = <return address> - MCOUNT_INSN_SIZE;
177    /* passing frame pointer up is optional -- see below */
178    prepare_ftrace_return(frompc, selfpc, frame_pointer);
179
180    /* restore all state needed by the ABI */
181}
182#endif
183
184For information on how to implement prepare_ftrace_return(), simply look at the
185x86 version (the frame pointer passing is optional; see the next section for
186more information). The only architecture-specific piece in it is the setup of
187the fault recovery table (the asm(...) code). The rest should be the same
188across architectures.
189
190Here is the pseudo code for the new return_to_handler assembly function. Note
191that the ABI that applies here is different from what applies to the mcount
192code. Since you are returning from a function (after the epilogue), you might
193be able to skimp on things saved/restored (usually just registers used to pass
194return values).
195
196#ifdef CONFIG_FUNCTION_GRAPH_TRACER
197void return_to_handler(void)
198{
199    /* save all state needed by the ABI (see paragraph above) */
200
201    void (*original_return_point)(void) = ftrace_return_to_handler();
202
203    /* restore all state needed by the ABI */
204
205    /* this is usually either a return or a jump */
206    original_return_point();
207}
208#endif
209
210
211HAVE_FUNCTION_GRAPH_FP_TEST
212---------------------------
213
214An arch may pass in a unique value (frame pointer) to both the entering and
215exiting of a function. On exit, the value is compared and if it does not
216match, then it will panic the kernel. This is largely a sanity check for bad
217code generation with gcc. If gcc for your port sanely updates the frame
218pointer under different opitmization levels, then ignore this option.
219
220However, adding support for it isn't terribly difficult. In your assembly code
221that calls prepare_ftrace_return(), pass the frame pointer as the 3rd argument.
222Then in the C version of that function, do what the x86 port does and pass it
223along to ftrace_push_return_trace() instead of a stub value of 0.
224
225Similarly, when you call ftrace_return_to_handler(), pass it the frame pointer.
226
227
228HAVE_FTRACE_NMI_ENTER
229---------------------
230
231If you can't trace NMI functions, then skip this option.
232
233<details to be filled>
234
235
236HAVE_SYSCALL_TRACEPOINTS
237---------------------
238
239You need very few things to get the syscalls tracing in an arch.
240
241- Support HAVE_ARCH_TRACEHOOK (see arch/Kconfig).
242- Have a NR_syscalls variable in <asm/unistd.h> that provides the number
243  of syscalls supported by the arch.
244- Support the TIF_SYSCALL_TRACEPOINT thread flags.
245- Put the trace_sys_enter() and trace_sys_exit() tracepoints calls from ptrace
246  in the ptrace syscalls tracing path.
247- Tag this arch as HAVE_SYSCALL_TRACEPOINTS.
248
249
250HAVE_FTRACE_MCOUNT_RECORD
251-------------------------
252
253See scripts/recordmcount.pl for more info.
254
255<details to be filled>
256
257
258HAVE_DYNAMIC_FTRACE
259---------------------
260
261<details to be filled>
262

Archive Download this file



interactive