]> git.proxmox.com Git - mirror_ubuntu-bionic-kernel.git/blob - Documentation/trace/ftrace-design.txt
Merge branch 'tip/perf/urgent-3' of git://git.kernel.org/pub/scm/linux/kernel/git...
[mirror_ubuntu-bionic-kernel.git] / Documentation / trace / ftrace-design.txt
1 function tracer guts
2 ====================
3 By Mike Frysinger
4
5 Introduction
6 ------------
7
8 Here we will cover the architecture pieces that the common function tracing
9 code relies on for proper functioning. Things are broken down into increasing
10 complexity so that you can start simple and at least get basic functionality.
11
12 Note that this focuses on architecture implementation details only. If you
13 want more explanation of a feature in terms of common code, review the common
14 ftrace.txt file.
15
16 Ideally, everyone who wishes to retain performance while supporting tracing in
17 their kernel should make it all the way to dynamic ftrace support.
18
19
20 Prerequisites
21 -------------
22
23 Ftrace relies on these features being implemented:
24 STACKTRACE_SUPPORT - implement save_stack_trace()
25 TRACE_IRQFLAGS_SUPPORT - implement include/asm/irqflags.h
26
27
28 HAVE_FUNCTION_TRACER
29 --------------------
30
31 You will need to implement the mcount and the ftrace_stub functions.
32
33 The exact mcount symbol name will depend on your toolchain. Some call it
34 "mcount", "_mcount", or even "__mcount". You can probably figure it out by
35 running something like:
36 $ echo 'main(){}' | gcc -x c -S -o - - -pg | grep mcount
37 call mcount
38 We'll make the assumption below that the symbol is "mcount" just to keep things
39 nice and simple in the examples.
40
41 Keep in mind that the ABI that is in effect inside of the mcount function is
42 *highly* architecture/toolchain specific. We cannot help you in this regard,
43 sorry. Dig up some old documentation and/or find someone more familiar than
44 you to bang ideas off of. Typically, register usage (argument/scratch/etc...)
45 is a major issue at this point, especially in relation to the location of the
46 mcount call (before/after function prologue). You might also want to look at
47 how glibc has implemented the mcount function for your architecture. It might
48 be (semi-)relevant.
49
50 The mcount function should check the function pointer ftrace_trace_function
51 to see if it is set to ftrace_stub. If it is, there is nothing for you to do,
52 so return immediately. If it isn't, then call that function in the same way
53 the mcount function normally calls __mcount_internal -- the first argument is
54 the "frompc" while the second argument is the "selfpc" (adjusted to remove the
55 size of the mcount call that is embedded in the function).
56
57 For example, if the function foo() calls bar(), when the bar() function calls
58 mcount(), the arguments mcount() will pass to the tracer are:
59 "frompc" - the address bar() will use to return to foo()
60 "selfpc" - the address bar() (with mcount() size adjustment)
61
62 Also keep in mind that this mcount function will be called *a lot*, so
63 optimizing for the default case of no tracer will help the smooth running of
64 your system when tracing is disabled. So the start of the mcount function is
65 typically the bare minimum with checking things before returning. That also
66 means the code flow should usually be kept linear (i.e. no branching in the nop
67 case). This is of course an optimization and not a hard requirement.
68
69 Here is some pseudo code that should help (these functions should actually be
70 implemented in assembly):
71
72 void ftrace_stub(void)
73 {
74 return;
75 }
76
77 void mcount(void)
78 {
79 /* save any bare state needed in order to do initial checking */
80
81 extern void (*ftrace_trace_function)(unsigned long, unsigned long);
82 if (ftrace_trace_function != ftrace_stub)
83 goto do_trace;
84
85 /* restore any bare state */
86
87 return;
88
89 do_trace:
90
91 /* save all state needed by the ABI (see paragraph above) */
92
93 unsigned long frompc = ...;
94 unsigned long selfpc = <return address> - MCOUNT_INSN_SIZE;
95 ftrace_trace_function(frompc, selfpc);
96
97 /* restore all state needed by the ABI */
98 }
99
100 Don't forget to export mcount for modules !
101 extern void mcount(void);
102 EXPORT_SYMBOL(mcount);
103
104
105 HAVE_FUNCTION_TRACE_MCOUNT_TEST
106 -------------------------------
107
108 This is an optional optimization for the normal case when tracing is turned off
109 in the system. If you do not enable this Kconfig option, the common ftrace
110 code will take care of doing the checking for you.
111
112 To support this feature, you only need to check the function_trace_stop
113 variable in the mcount function. If it is non-zero, there is no tracing to be
114 done at all, so you can return.
115
116 This additional pseudo code would simply be:
117 void mcount(void)
118 {
119 /* save any bare state needed in order to do initial checking */
120
121 + if (function_trace_stop)
122 + return;
123
124 extern void (*ftrace_trace_function)(unsigned long, unsigned long);
125 if (ftrace_trace_function != ftrace_stub)
126 ...
127
128
129 HAVE_FUNCTION_GRAPH_TRACER
130 --------------------------
131
132 Deep breath ... time to do some real work. Here you will need to update the
133 mcount function to check ftrace graph function pointers, as well as implement
134 some functions to save (hijack) and restore the return address.
135
136 The mcount function should check the function pointers ftrace_graph_return
137 (compare to ftrace_stub) and ftrace_graph_entry (compare to
138 ftrace_graph_entry_stub). If either of those is not set to the relevant stub
139 function, call the arch-specific function ftrace_graph_caller which in turn
140 calls the arch-specific function prepare_ftrace_return. Neither of these
141 function names is strictly required, but you should use them anyway to stay
142 consistent across the architecture ports -- easier to compare & contrast
143 things.
144
145 The arguments to prepare_ftrace_return are slightly different than what are
146 passed to ftrace_trace_function. The second argument "selfpc" is the same,
147 but the first argument should be a pointer to the "frompc". Typically this is
148 located on the stack. This allows the function to hijack the return address
149 temporarily to have it point to the arch-specific function return_to_handler.
150 That function will simply call the common ftrace_return_to_handler function and
151 that will return the original return address with which you can return to the
152 original call site.
153
154 Here is the updated mcount pseudo code:
155 void mcount(void)
156 {
157 ...
158 if (ftrace_trace_function != ftrace_stub)
159 goto do_trace;
160
161 +#ifdef CONFIG_FUNCTION_GRAPH_TRACER
162 + extern void (*ftrace_graph_return)(...);
163 + extern void (*ftrace_graph_entry)(...);
164 + if (ftrace_graph_return != ftrace_stub ||
165 + ftrace_graph_entry != ftrace_graph_entry_stub)
166 + ftrace_graph_caller();
167 +#endif
168
169 /* restore any bare state */
170 ...
171
172 Here is the pseudo code for the new ftrace_graph_caller assembly function:
173 #ifdef CONFIG_FUNCTION_GRAPH_TRACER
174 void ftrace_graph_caller(void)
175 {
176 /* save all state needed by the ABI */
177
178 unsigned long *frompc = &...;
179 unsigned long selfpc = <return address> - MCOUNT_INSN_SIZE;
180 /* passing frame pointer up is optional -- see below */
181 prepare_ftrace_return(frompc, selfpc, frame_pointer);
182
183 /* restore all state needed by the ABI */
184 }
185 #endif
186
187 For information on how to implement prepare_ftrace_return(), simply look at the
188 x86 version (the frame pointer passing is optional; see the next section for
189 more information). The only architecture-specific piece in it is the setup of
190 the fault recovery table (the asm(...) code). The rest should be the same
191 across architectures.
192
193 Here is the pseudo code for the new return_to_handler assembly function. Note
194 that the ABI that applies here is different from what applies to the mcount
195 code. Since you are returning from a function (after the epilogue), you might
196 be able to skimp on things saved/restored (usually just registers used to pass
197 return values).
198
199 #ifdef CONFIG_FUNCTION_GRAPH_TRACER
200 void return_to_handler(void)
201 {
202 /* save all state needed by the ABI (see paragraph above) */
203
204 void (*original_return_point)(void) = ftrace_return_to_handler();
205
206 /* restore all state needed by the ABI */
207
208 /* this is usually either a return or a jump */
209 original_return_point();
210 }
211 #endif
212
213
214 HAVE_FUNCTION_GRAPH_FP_TEST
215 ---------------------------
216
217 An arch may pass in a unique value (frame pointer) to both the entering and
218 exiting of a function. On exit, the value is compared and if it does not
219 match, then it will panic the kernel. This is largely a sanity check for bad
220 code generation with gcc. If gcc for your port sanely updates the frame
221 pointer under different optimization levels, then ignore this option.
222
223 However, adding support for it isn't terribly difficult. In your assembly code
224 that calls prepare_ftrace_return(), pass the frame pointer as the 3rd argument.
225 Then in the C version of that function, do what the x86 port does and pass it
226 along to ftrace_push_return_trace() instead of a stub value of 0.
227
228 Similarly, when you call ftrace_return_to_handler(), pass it the frame pointer.
229
230
231 HAVE_FTRACE_NMI_ENTER
232 ---------------------
233
234 If you can't trace NMI functions, then skip this option.
235
236 <details to be filled>
237
238
239 HAVE_SYSCALL_TRACEPOINTS
240 ------------------------
241
242 You need very few things to get the syscalls tracing in an arch.
243
244 - Support HAVE_ARCH_TRACEHOOK (see arch/Kconfig).
245 - Have a NR_syscalls variable in <asm/unistd.h> that provides the number
246 of syscalls supported by the arch.
247 - Support the TIF_SYSCALL_TRACEPOINT thread flags.
248 - Put the trace_sys_enter() and trace_sys_exit() tracepoints calls from ptrace
249 in the ptrace syscalls tracing path.
250 - Tag this arch as HAVE_SYSCALL_TRACEPOINTS.
251
252
253 HAVE_FTRACE_MCOUNT_RECORD
254 -------------------------
255
256 See scripts/recordmcount.pl for more info. Just fill in the arch-specific
257 details for how to locate the addresses of mcount call sites via objdump.
258 This option doesn't make much sense without also implementing dynamic ftrace.
259
260
261 HAVE_DYNAMIC_FTRACE
262 -------------------
263
264 You will first need HAVE_FTRACE_MCOUNT_RECORD and HAVE_FUNCTION_TRACER, so
265 scroll your reader back up if you got over eager.
266
267 Once those are out of the way, you will need to implement:
268 - asm/ftrace.h:
269 - MCOUNT_ADDR
270 - ftrace_call_adjust()
271 - struct dyn_arch_ftrace{}
272 - asm code:
273 - mcount() (new stub)
274 - ftrace_caller()
275 - ftrace_call()
276 - ftrace_stub()
277 - C code:
278 - ftrace_dyn_arch_init()
279 - ftrace_make_nop()
280 - ftrace_make_call()
281 - ftrace_update_ftrace_func()
282
283 First you will need to fill out some arch details in your asm/ftrace.h.
284
285 Define MCOUNT_ADDR as the address of your mcount symbol similar to:
286 #define MCOUNT_ADDR ((unsigned long)mcount)
287 Since no one else will have a decl for that function, you will need to:
288 extern void mcount(void);
289
290 You will also need the helper function ftrace_call_adjust(). Most people
291 will be able to stub it out like so:
292 static inline unsigned long ftrace_call_adjust(unsigned long addr)
293 {
294 return addr;
295 }
296 <details to be filled>
297
298 Lastly you will need the custom dyn_arch_ftrace structure. If you need
299 some extra state when runtime patching arbitrary call sites, this is the
300 place. For now though, create an empty struct:
301 struct dyn_arch_ftrace {
302 /* No extra data needed */
303 };
304
305 With the header out of the way, we can fill out the assembly code. While we
306 did already create a mcount() function earlier, dynamic ftrace only wants a
307 stub function. This is because the mcount() will only be used during boot
308 and then all references to it will be patched out never to return. Instead,
309 the guts of the old mcount() will be used to create a new ftrace_caller()
310 function. Because the two are hard to merge, it will most likely be a lot
311 easier to have two separate definitions split up by #ifdefs. Same goes for
312 the ftrace_stub() as that will now be inlined in ftrace_caller().
313
314 Before we get confused anymore, let's check out some pseudo code so you can
315 implement your own stuff in assembly:
316
317 void mcount(void)
318 {
319 return;
320 }
321
322 void ftrace_caller(void)
323 {
324 /* implement HAVE_FUNCTION_TRACE_MCOUNT_TEST if you desire */
325
326 /* save all state needed by the ABI (see paragraph above) */
327
328 unsigned long frompc = ...;
329 unsigned long selfpc = <return address> - MCOUNT_INSN_SIZE;
330
331 ftrace_call:
332 ftrace_stub(frompc, selfpc);
333
334 /* restore all state needed by the ABI */
335
336 ftrace_stub:
337 return;
338 }
339
340 This might look a little odd at first, but keep in mind that we will be runtime
341 patching multiple things. First, only functions that we actually want to trace
342 will be patched to call ftrace_caller(). Second, since we only have one tracer
343 active at a time, we will patch the ftrace_caller() function itself to call the
344 specific tracer in question. That is the point of the ftrace_call label.
345
346 With that in mind, let's move on to the C code that will actually be doing the
347 runtime patching. You'll need a little knowledge of your arch's opcodes in
348 order to make it through the next section.
349
350 Every arch has an init callback function. If you need to do something early on
351 to initialize some state, this is the time to do that. Otherwise, this simple
352 function below should be sufficient for most people:
353
354 int __init ftrace_dyn_arch_init(void *data)
355 {
356 /* return value is done indirectly via data */
357 *(unsigned long *)data = 0;
358
359 return 0;
360 }
361
362 There are two functions that are used to do runtime patching of arbitrary
363 functions. The first is used to turn the mcount call site into a nop (which
364 is what helps us retain runtime performance when not tracing). The second is
365 used to turn the mcount call site into a call to an arbitrary location (but
366 typically that is ftracer_caller()). See the general function definition in
367 linux/ftrace.h for the functions:
368 ftrace_make_nop()
369 ftrace_make_call()
370 The rec->ip value is the address of the mcount call site that was collected
371 by the scripts/recordmcount.pl during build time.
372
373 The last function is used to do runtime patching of the active tracer. This
374 will be modifying the assembly code at the location of the ftrace_call symbol
375 inside of the ftrace_caller() function. So you should have sufficient padding
376 at that location to support the new function calls you'll be inserting. Some
377 people will be using a "call" type instruction while others will be using a
378 "branch" type instruction. Specifically, the function is:
379 ftrace_update_ftrace_func()
380
381
382 HAVE_DYNAMIC_FTRACE + HAVE_FUNCTION_GRAPH_TRACER
383 ------------------------------------------------
384
385 The function grapher needs a few tweaks in order to work with dynamic ftrace.
386 Basically, you will need to:
387 - update:
388 - ftrace_caller()
389 - ftrace_graph_call()
390 - ftrace_graph_caller()
391 - implement:
392 - ftrace_enable_ftrace_graph_caller()
393 - ftrace_disable_ftrace_graph_caller()
394
395 <details to be filled>
396 Quick notes:
397 - add a nop stub after the ftrace_call location named ftrace_graph_call;
398 stub needs to be large enough to support a call to ftrace_graph_caller()
399 - update ftrace_graph_caller() to work with being called by the new
400 ftrace_caller() since some semantics may have changed
401 - ftrace_enable_ftrace_graph_caller() will runtime patch the
402 ftrace_graph_call location with a call to ftrace_graph_caller()
403 - ftrace_disable_ftrace_graph_caller() will runtime patch the
404 ftrace_graph_call location with nops