Commit | Line | Data |
---|---|---|
d27a4ddd JK |
1 | Title : Kernel Probes (Kprobes) |
2 | Authors : Jim Keniston <jkenisto@us.ibm.com> | |
3 | : Prasanna S Panchamukhi <prasanna@in.ibm.com> | |
4 | ||
5 | CONTENTS | |
6 | ||
7 | 1. Concepts: Kprobes, Jprobes, Return Probes | |
8 | 2. Architectures Supported | |
9 | 3. Configuring Kprobes | |
10 | 4. API Reference | |
11 | 5. Kprobes Features and Limitations | |
12 | 6. Probe Overhead | |
13 | 7. TODO | |
14 | 8. Kprobes Example | |
15 | 9. Jprobes Example | |
16 | 10. Kretprobes Example | |
bf8f6e5b | 17 | Appendix A: The kprobes debugfs interface |
d27a4ddd JK |
18 | |
19 | 1. Concepts: Kprobes, Jprobes, Return Probes | |
20 | ||
21 | Kprobes enables you to dynamically break into any kernel routine and | |
22 | collect debugging and performance information non-disruptively. You | |
23 | can trap at almost any kernel code address, specifying a handler | |
24 | routine to be invoked when the breakpoint is hit. | |
25 | ||
26 | There are currently three types of probes: kprobes, jprobes, and | |
27 | kretprobes (also called return probes). A kprobe can be inserted | |
28 | on virtually any instruction in the kernel. A jprobe is inserted at | |
29 | the entry to a kernel function, and provides convenient access to the | |
30 | function's arguments. A return probe fires when a specified function | |
31 | returns. | |
32 | ||
33 | In the typical case, Kprobes-based instrumentation is packaged as | |
34 | a kernel module. The module's init function installs ("registers") | |
35 | one or more probes, and the exit function unregisters them. A | |
36 | registration function such as register_kprobe() specifies where | |
37 | the probe is to be inserted and what handler is to be called when | |
38 | the probe is hit. | |
39 | ||
40 | The next three subsections explain how the different types of | |
41 | probes work. They explain certain things that you'll need to | |
42 | know in order to make the best use of Kprobes -- e.g., the | |
43 | difference between a pre_handler and a post_handler, and how | |
44 | to use the maxactive and nmissed fields of a kretprobe. But | |
45 | if you're in a hurry to start using Kprobes, you can skip ahead | |
46 | to section 2. | |
47 | ||
48 | 1.1 How Does a Kprobe Work? | |
49 | ||
50 | When a kprobe is registered, Kprobes makes a copy of the probed | |
51 | instruction and replaces the first byte(s) of the probed instruction | |
52 | with a breakpoint instruction (e.g., int3 on i386 and x86_64). | |
53 | ||
54 | When a CPU hits the breakpoint instruction, a trap occurs, the CPU's | |
55 | registers are saved, and control passes to Kprobes via the | |
56 | notifier_call_chain mechanism. Kprobes executes the "pre_handler" | |
57 | associated with the kprobe, passing the handler the addresses of the | |
58 | kprobe struct and the saved registers. | |
59 | ||
60 | Next, Kprobes single-steps its copy of the probed instruction. | |
61 | (It would be simpler to single-step the actual instruction in place, | |
62 | but then Kprobes would have to temporarily remove the breakpoint | |
63 | instruction. This would open a small time window when another CPU | |
64 | could sail right past the probepoint.) | |
65 | ||
66 | After the instruction is single-stepped, Kprobes executes the | |
67 | "post_handler," if any, that is associated with the kprobe. | |
68 | Execution then continues with the instruction following the probepoint. | |
69 | ||
70 | 1.2 How Does a Jprobe Work? | |
71 | ||
72 | A jprobe is implemented using a kprobe that is placed on a function's | |
73 | entry point. It employs a simple mirroring principle to allow | |
74 | seamless access to the probed function's arguments. The jprobe | |
75 | handler routine should have the same signature (arg list and return | |
76 | type) as the function being probed, and must always end by calling | |
77 | the Kprobes function jprobe_return(). | |
78 | ||
79 | Here's how it works. When the probe is hit, Kprobes makes a copy of | |
80 | the saved registers and a generous portion of the stack (see below). | |
81 | Kprobes then points the saved instruction pointer at the jprobe's | |
82 | handler routine, and returns from the trap. As a result, control | |
83 | passes to the handler, which is presented with the same register and | |
84 | stack contents as the probed function. When it is done, the handler | |
85 | calls jprobe_return(), which traps again to restore the original stack | |
86 | contents and processor state and switch to the probed function. | |
87 | ||
88 | By convention, the callee owns its arguments, so gcc may produce code | |
89 | that unexpectedly modifies that portion of the stack. This is why | |
90 | Kprobes saves a copy of the stack and restores it after the jprobe | |
91 | handler has run. Up to MAX_STACK_SIZE bytes are copied -- e.g., | |
92 | 64 bytes on i386. | |
93 | ||
94 | Note that the probed function's args may be passed on the stack | |
b5606c2d HH |
95 | or in registers. The jprobe will work in either case, so long as the |
96 | handler's prototype matches that of the probed function. | |
d27a4ddd | 97 | |
f47cd9b5 AS |
98 | 1.3 Return Probes |
99 | ||
100 | 1.3.1 How Does a Return Probe Work? | |
d27a4ddd JK |
101 | |
102 | When you call register_kretprobe(), Kprobes establishes a kprobe at | |
103 | the entry to the function. When the probed function is called and this | |
104 | probe is hit, Kprobes saves a copy of the return address, and replaces | |
105 | the return address with the address of a "trampoline." The trampoline | |
106 | is an arbitrary piece of code -- typically just a nop instruction. | |
107 | At boot time, Kprobes registers a kprobe at the trampoline. | |
108 | ||
109 | When the probed function executes its return instruction, control | |
110 | passes to the trampoline and that probe is hit. Kprobes' trampoline | |
f47cd9b5 AS |
111 | handler calls the user-specified return handler associated with the |
112 | kretprobe, then sets the saved instruction pointer to the saved return | |
113 | address, and that's where execution resumes upon return from the trap. | |
d27a4ddd JK |
114 | |
115 | While the probed function is executing, its return address is | |
116 | stored in an object of type kretprobe_instance. Before calling | |
117 | register_kretprobe(), the user sets the maxactive field of the | |
118 | kretprobe struct to specify how many instances of the specified | |
119 | function can be probed simultaneously. register_kretprobe() | |
120 | pre-allocates the indicated number of kretprobe_instance objects. | |
121 | ||
122 | For example, if the function is non-recursive and is called with a | |
123 | spinlock held, maxactive = 1 should be enough. If the function is | |
124 | non-recursive and can never relinquish the CPU (e.g., via a semaphore | |
125 | or preemption), NR_CPUS should be enough. If maxactive <= 0, it is | |
126 | set to a default value. If CONFIG_PREEMPT is enabled, the default | |
127 | is max(10, 2*NR_CPUS). Otherwise, the default is NR_CPUS. | |
128 | ||
129 | It's not a disaster if you set maxactive too low; you'll just miss | |
130 | some probes. In the kretprobe struct, the nmissed field is set to | |
131 | zero when the return probe is registered, and is incremented every | |
132 | time the probed function is entered but there is no kretprobe_instance | |
133 | object available for establishing the return probe. | |
134 | ||
f47cd9b5 AS |
135 | 1.3.2 Kretprobe entry-handler |
136 | ||
137 | Kretprobes also provides an optional user-specified handler which runs | |
138 | on function entry. This handler is specified by setting the entry_handler | |
139 | field of the kretprobe struct. Whenever the kprobe placed by kretprobe at the | |
140 | function entry is hit, the user-defined entry_handler, if any, is invoked. | |
141 | If the entry_handler returns 0 (success) then a corresponding return handler | |
142 | is guaranteed to be called upon function return. If the entry_handler | |
143 | returns a non-zero error then Kprobes leaves the return address as is, and | |
144 | the kretprobe has no further effect for that particular function instance. | |
145 | ||
146 | Multiple entry and return handler invocations are matched using the unique | |
147 | kretprobe_instance object associated with them. Additionally, a user | |
148 | may also specify per return-instance private data to be part of each | |
149 | kretprobe_instance object. This is especially useful when sharing private | |
150 | data between corresponding user entry and return handlers. The size of each | |
151 | private data object can be specified at kretprobe registration time by | |
152 | setting the data_size field of the kretprobe struct. This data can be | |
153 | accessed through the data field of each kretprobe_instance object. | |
154 | ||
155 | In case probed function is entered but there is no kretprobe_instance | |
156 | object available, then in addition to incrementing the nmissed count, | |
157 | the user entry_handler invocation is also skipped. | |
158 | ||
d27a4ddd JK |
159 | 2. Architectures Supported |
160 | ||
161 | Kprobes, jprobes, and return probes are implemented on the following | |
162 | architectures: | |
163 | ||
164 | - i386 | |
8861da31 | 165 | - x86_64 (AMD-64, EM64T) |
d27a4ddd | 166 | - ppc64 |
8861da31 | 167 | - ia64 (Does not support probes on instruction slot1.) |
d27a4ddd | 168 | - sparc64 (Return probes not yet implemented.) |
5de865b4 | 169 | - arm |
d27a4ddd JK |
170 | |
171 | 3. Configuring Kprobes | |
172 | ||
173 | When configuring the kernel using make menuconfig/xconfig/oldconfig, | |
8861da31 JK |
174 | ensure that CONFIG_KPROBES is set to "y". Under "Instrumentation |
175 | Support", look for "Kprobes". | |
176 | ||
177 | So that you can load and unload Kprobes-based instrumentation modules, | |
178 | make sure "Loadable module support" (CONFIG_MODULES) and "Module | |
179 | unloading" (CONFIG_MODULE_UNLOAD) are set to "y". | |
d27a4ddd | 180 | |
09b18203 AM |
181 | Also make sure that CONFIG_KALLSYMS and perhaps even CONFIG_KALLSYMS_ALL |
182 | are set to "y", since kallsyms_lookup_name() is used by the in-kernel | |
183 | kprobe address resolution code. | |
d27a4ddd JK |
184 | |
185 | If you need to insert a probe in the middle of a function, you may find | |
186 | it useful to "Compile the kernel with debug info" (CONFIG_DEBUG_INFO), | |
187 | so you can use "objdump -d -l vmlinux" to see the source-to-object | |
188 | code mapping. | |
189 | ||
190 | 4. API Reference | |
191 | ||
192 | The Kprobes API includes a "register" function and an "unregister" | |
193 | function for each type of probe. Here are terse, mini-man-page | |
194 | specifications for these functions and the associated probe handlers | |
804defea AM |
195 | that you'll write. See the files in the samples/kprobes/ sub-directory |
196 | for examples. | |
d27a4ddd JK |
197 | |
198 | 4.1 register_kprobe | |
199 | ||
200 | #include <linux/kprobes.h> | |
201 | int register_kprobe(struct kprobe *kp); | |
202 | ||
203 | Sets a breakpoint at the address kp->addr. When the breakpoint is | |
204 | hit, Kprobes calls kp->pre_handler. After the probed instruction | |
205 | is single-stepped, Kprobe calls kp->post_handler. If a fault | |
206 | occurs during execution of kp->pre_handler or kp->post_handler, | |
207 | or during single-stepping of the probed instruction, Kprobes calls | |
208 | kp->fault_handler. Any or all handlers can be NULL. | |
209 | ||
09b18203 AM |
210 | NOTE: |
211 | 1. With the introduction of the "symbol_name" field to struct kprobe, | |
212 | the probepoint address resolution will now be taken care of by the kernel. | |
213 | The following will now work: | |
214 | ||
215 | kp.symbol_name = "symbol_name"; | |
216 | ||
217 | (64-bit powerpc intricacies such as function descriptors are handled | |
218 | transparently) | |
219 | ||
220 | 2. Use the "offset" field of struct kprobe if the offset into the symbol | |
221 | to install a probepoint is known. This field is used to calculate the | |
222 | probepoint. | |
223 | ||
224 | 3. Specify either the kprobe "symbol_name" OR the "addr". If both are | |
225 | specified, kprobe registration will fail with -EINVAL. | |
226 | ||
227 | 4. With CISC architectures (such as i386 and x86_64), the kprobes code | |
228 | does not validate if the kprobe.addr is at an instruction boundary. | |
229 | Use "offset" with caution. | |
230 | ||
d27a4ddd JK |
231 | register_kprobe() returns 0 on success, or a negative errno otherwise. |
232 | ||
233 | User's pre-handler (kp->pre_handler): | |
234 | #include <linux/kprobes.h> | |
235 | #include <linux/ptrace.h> | |
236 | int pre_handler(struct kprobe *p, struct pt_regs *regs); | |
237 | ||
238 | Called with p pointing to the kprobe associated with the breakpoint, | |
239 | and regs pointing to the struct containing the registers saved when | |
240 | the breakpoint was hit. Return 0 here unless you're a Kprobes geek. | |
241 | ||
242 | User's post-handler (kp->post_handler): | |
243 | #include <linux/kprobes.h> | |
244 | #include <linux/ptrace.h> | |
245 | void post_handler(struct kprobe *p, struct pt_regs *regs, | |
246 | unsigned long flags); | |
247 | ||
248 | p and regs are as described for the pre_handler. flags always seems | |
249 | to be zero. | |
250 | ||
251 | User's fault-handler (kp->fault_handler): | |
252 | #include <linux/kprobes.h> | |
253 | #include <linux/ptrace.h> | |
254 | int fault_handler(struct kprobe *p, struct pt_regs *regs, int trapnr); | |
255 | ||
256 | p and regs are as described for the pre_handler. trapnr is the | |
257 | architecture-specific trap number associated with the fault (e.g., | |
258 | on i386, 13 for a general protection fault or 14 for a page fault). | |
259 | Returns 1 if it successfully handled the exception. | |
260 | ||
261 | 4.2 register_jprobe | |
262 | ||
263 | #include <linux/kprobes.h> | |
264 | int register_jprobe(struct jprobe *jp) | |
265 | ||
266 | Sets a breakpoint at the address jp->kp.addr, which must be the address | |
267 | of the first instruction of a function. When the breakpoint is hit, | |
268 | Kprobes runs the handler whose address is jp->entry. | |
269 | ||
270 | The handler should have the same arg list and return type as the probed | |
271 | function; and just before it returns, it must call jprobe_return(). | |
272 | (The handler never actually returns, since jprobe_return() returns | |
b5606c2d HH |
273 | control to Kprobes.) If the probed function is declared asmlinkage |
274 | or anything else that affects how args are passed, the handler's | |
275 | declaration must match. | |
d27a4ddd JK |
276 | |
277 | register_jprobe() returns 0 on success, or a negative errno otherwise. | |
278 | ||
279 | 4.3 register_kretprobe | |
280 | ||
281 | #include <linux/kprobes.h> | |
282 | int register_kretprobe(struct kretprobe *rp); | |
283 | ||
284 | Establishes a return probe for the function whose address is | |
285 | rp->kp.addr. When that function returns, Kprobes calls rp->handler. | |
286 | You must set rp->maxactive appropriately before you call | |
287 | register_kretprobe(); see "How Does a Return Probe Work?" for details. | |
288 | ||
289 | register_kretprobe() returns 0 on success, or a negative errno | |
290 | otherwise. | |
291 | ||
292 | User's return-probe handler (rp->handler): | |
293 | #include <linux/kprobes.h> | |
294 | #include <linux/ptrace.h> | |
295 | int kretprobe_handler(struct kretprobe_instance *ri, struct pt_regs *regs); | |
296 | ||
297 | regs is as described for kprobe.pre_handler. ri points to the | |
298 | kretprobe_instance object, of which the following fields may be | |
299 | of interest: | |
300 | - ret_addr: the return address | |
301 | - rp: points to the corresponding kretprobe object | |
302 | - task: points to the corresponding task struct | |
f47cd9b5 AS |
303 | - data: points to per return-instance private data; see "Kretprobe |
304 | entry-handler" for details. | |
09b18203 AM |
305 | |
306 | The regs_return_value(regs) macro provides a simple abstraction to | |
307 | extract the return value from the appropriate register as defined by | |
308 | the architecture's ABI. | |
309 | ||
d27a4ddd JK |
310 | The handler's return value is currently ignored. |
311 | ||
312 | 4.4 unregister_*probe | |
313 | ||
314 | #include <linux/kprobes.h> | |
315 | void unregister_kprobe(struct kprobe *kp); | |
316 | void unregister_jprobe(struct jprobe *jp); | |
317 | void unregister_kretprobe(struct kretprobe *rp); | |
318 | ||
319 | Removes the specified probe. The unregister function can be called | |
320 | at any time after the probe has been registered. | |
321 | ||
322 | 5. Kprobes Features and Limitations | |
323 | ||
8861da31 JK |
324 | Kprobes allows multiple probes at the same address. Currently, |
325 | however, there cannot be multiple jprobes on the same function at | |
326 | the same time. | |
d27a4ddd JK |
327 | |
328 | In general, you can install a probe anywhere in the kernel. | |
329 | In particular, you can probe interrupt handlers. Known exceptions | |
330 | are discussed in this section. | |
331 | ||
8861da31 JK |
332 | The register_*probe functions will return -EINVAL if you attempt |
333 | to install a probe in the code that implements Kprobes (mostly | |
334 | kernel/kprobes.c and arch/*/kernel/kprobes.c, but also functions such | |
335 | as do_page_fault and notifier_call_chain). | |
d27a4ddd JK |
336 | |
337 | If you install a probe in an inline-able function, Kprobes makes | |
338 | no attempt to chase down all inline instances of the function and | |
339 | install probes there. gcc may inline a function without being asked, | |
340 | so keep this in mind if you're not seeing the probe hits you expect. | |
341 | ||
342 | A probe handler can modify the environment of the probed function | |
343 | -- e.g., by modifying kernel data structures, or by modifying the | |
344 | contents of the pt_regs struct (which are restored to the registers | |
345 | upon return from the breakpoint). So Kprobes can be used, for example, | |
346 | to install a bug fix or to inject faults for testing. Kprobes, of | |
347 | course, has no way to distinguish the deliberately injected faults | |
348 | from the accidental ones. Don't drink and probe. | |
349 | ||
350 | Kprobes makes no attempt to prevent probe handlers from stepping on | |
351 | each other -- e.g., probing printk() and then calling printk() from a | |
8861da31 JK |
352 | probe handler. If a probe handler hits a probe, that second probe's |
353 | handlers won't be run in that instance, and the kprobe.nmissed member | |
354 | of the second probe will be incremented. | |
355 | ||
356 | As of Linux v2.6.15-rc1, multiple handlers (or multiple instances of | |
357 | the same handler) may run concurrently on different CPUs. | |
358 | ||
359 | Kprobes does not use mutexes or allocate memory except during | |
d27a4ddd JK |
360 | registration and unregistration. |
361 | ||
362 | Probe handlers are run with preemption disabled. Depending on the | |
363 | architecture, handlers may also run with interrupts disabled. In any | |
364 | case, your handler should not yield the CPU (e.g., by attempting to | |
365 | acquire a semaphore). | |
366 | ||
367 | Since a return probe is implemented by replacing the return | |
368 | address with the trampoline's address, stack backtraces and calls | |
369 | to __builtin_return_address() will typically yield the trampoline's | |
370 | address instead of the real return address for kretprobed functions. | |
371 | (As far as we can tell, __builtin_return_address() is used only | |
372 | for instrumentation and error reporting.) | |
373 | ||
8861da31 JK |
374 | If the number of times a function is called does not match the number |
375 | of times it returns, registering a return probe on that function may | |
bf8f6e5b AM |
376 | produce undesirable results. In such a case, a line: |
377 | kretprobe BUG!: Processing kretprobe d000000000041aa8 @ c00000000004f48c | |
378 | gets printed. With this information, one will be able to correlate the | |
379 | exact instance of the kretprobe that caused the problem. We have the | |
380 | do_exit() case covered. do_execve() and do_fork() are not an issue. | |
381 | We're unaware of other specific cases where this could be a problem. | |
8861da31 JK |
382 | |
383 | If, upon entry to or exit from a function, the CPU is running on | |
384 | a stack other than that of the current task, registering a return | |
385 | probe on that function may produce undesirable results. For this | |
386 | reason, Kprobes doesn't support return probes (or kprobes or jprobes) | |
387 | on the x86_64 version of __switch_to(); the registration functions | |
388 | return -EINVAL. | |
d27a4ddd JK |
389 | |
390 | 6. Probe Overhead | |
391 | ||
392 | On a typical CPU in use in 2005, a kprobe hit takes 0.5 to 1.0 | |
393 | microseconds to process. Specifically, a benchmark that hits the same | |
394 | probepoint repeatedly, firing a simple handler each time, reports 1-2 | |
395 | million hits per second, depending on the architecture. A jprobe or | |
396 | return-probe hit typically takes 50-75% longer than a kprobe hit. | |
397 | When you have a return probe set on a function, adding a kprobe at | |
398 | the entry to that function adds essentially no overhead. | |
399 | ||
400 | Here are sample overhead figures (in usec) for different architectures. | |
401 | k = kprobe; j = jprobe; r = return probe; kr = kprobe + return probe | |
402 | on same function; jr = jprobe + return probe on same function | |
403 | ||
404 | i386: Intel Pentium M, 1495 MHz, 2957.31 bogomips | |
405 | k = 0.57 usec; j = 1.00; r = 0.92; kr = 0.99; jr = 1.40 | |
406 | ||
407 | x86_64: AMD Opteron 246, 1994 MHz, 3971.48 bogomips | |
408 | k = 0.49 usec; j = 0.76; r = 0.80; kr = 0.82; jr = 1.07 | |
409 | ||
410 | ppc64: POWER5 (gr), 1656 MHz (SMT disabled, 1 virtual CPU per physical CPU) | |
411 | k = 0.77 usec; j = 1.31; r = 1.26; kr = 1.45; jr = 1.99 | |
412 | ||
413 | 7. TODO | |
414 | ||
8861da31 JK |
415 | a. SystemTap (http://sourceware.org/systemtap): Provides a simplified |
416 | programming interface for probe-based instrumentation. Try it out. | |
417 | b. Kernel return probes for sparc64. | |
418 | c. Support for other architectures. | |
419 | d. User-space probes. | |
420 | e. Watchpoint probes (which fire on data references). | |
d27a4ddd JK |
421 | |
422 | 8. Kprobes Example | |
423 | ||
804defea | 424 | See samples/kprobes/kprobe_example.c |
d27a4ddd JK |
425 | |
426 | 9. Jprobes Example | |
427 | ||
804defea | 428 | See samples/kprobes/jprobe_example.c |
d27a4ddd JK |
429 | |
430 | 10. Kretprobes Example | |
431 | ||
804defea | 432 | See samples/kprobes/kretprobe_example.c |
d27a4ddd JK |
433 | |
434 | For additional information on Kprobes, refer to the following URLs: | |
435 | http://www-106.ibm.com/developerworks/library/l-kprobes.html?ca=dgr-lnxw42Kprobe | |
436 | http://www.redhat.com/magazine/005mar05/features/kprobes/ | |
09b18203 AM |
437 | http://www-users.cs.umn.edu/~boutcher/kprobes/ |
438 | http://www.linuxsymposium.org/2006/linuxsymposium_procv2.pdf (pages 101-115) | |
bf8f6e5b AM |
439 | |
440 | ||
441 | Appendix A: The kprobes debugfs interface | |
442 | ||
443 | With recent kernels (> 2.6.20) the list of registered kprobes is visible | |
444 | under the /debug/kprobes/ directory (assuming debugfs is mounted at /debug). | |
445 | ||
446 | /debug/kprobes/list: Lists all registered probes on the system | |
447 | ||
448 | c015d71a k vfs_read+0x0 | |
449 | c011a316 j do_fork+0x0 | |
450 | c03dedc5 r tcp_v4_rcv+0x0 | |
451 | ||
452 | The first column provides the kernel address where the probe is inserted. | |
453 | The second column identifies the type of probe (k - kprobe, r - kretprobe | |
454 | and j - jprobe), while the third column specifies the symbol+offset of | |
455 | the probe. If the probed function belongs to a module, the module name | |
456 | is also specified. | |
457 | ||
458 | /debug/kprobes/enabled: Turn kprobes ON/OFF | |
459 | ||
460 | Provides a knob to globally turn registered kprobes ON or OFF. By default, | |
461 | all kprobes are enabled. By echoing "0" to this file, all registered probes | |
462 | will be disarmed, till such time a "1" is echoed to this file. |