1Title : Kernel Probes (Kprobes) 2Authors : Jim Keniston <jkenisto@us.ibm.com> 3 : Prasanna S Panchamukhi <prasanna@in.ibm.com> 4 5CONTENTS 6 71. Concepts: Kprobes, Jprobes, Return Probes 82. Architectures Supported 93. Configuring Kprobes 104. API Reference 115. Kprobes Features and Limitations 126. Probe Overhead 137. TODO 148. Kprobes Example 159. Jprobes Example 1610. Kretprobes Example 17Appendix A: The kprobes debugfs interface 18 191. Concepts: Kprobes, Jprobes, Return Probes 20 21Kprobes enables you to dynamically break into any kernel routine and 22collect debugging and performance information non-disruptively. You 23can trap at almost any kernel code address, specifying a handler 24routine to be invoked when the breakpoint is hit. 25 26There are currently three types of probes: kprobes, jprobes, and 27kretprobes (also called return probes). A kprobe can be inserted 28on virtually any instruction in the kernel. A jprobe is inserted at 29the entry to a kernel function, and provides convenient access to the 30function's arguments. A return probe fires when a specified function 31returns. 32 33In the typical case, Kprobes-based instrumentation is packaged as 34a kernel module. The module's init function installs ("registers") 35one or more probes, and the exit function unregisters them. A 36registration function such as register_kprobe() specifies where 37the probe is to be inserted and what handler is to be called when 38the probe is hit. 39 40There are also register_/unregister_*probes() functions for batch 41registration/unregistration of a group of *probes. These functions 42can speed up unregistration process when you have to unregister 43a lot of probes at once. 44 45The next three subsections explain how the different types of 46probes work. They explain certain things that you'll need to 47know in order to make the best use of Kprobes -- e.g., the 48difference between a pre_handler and a post_handler, and how 49to use the maxactive and nmissed fields of a kretprobe. But 50if you're in a hurry to start using Kprobes, you can skip ahead 51to section 2. 52 531.1 How Does a Kprobe Work? 54 55When a kprobe is registered, Kprobes makes a copy of the probed 56instruction and replaces the first byte(s) of the probed instruction 57with a breakpoint instruction (e.g., int3 on i386 and x86_64). 58 59When a CPU hits the breakpoint instruction, a trap occurs, the CPU's 60registers are saved, and control passes to Kprobes via the 61notifier_call_chain mechanism. Kprobes executes the "pre_handler" 62associated with the kprobe, passing the handler the addresses of the 63kprobe struct and the saved registers. 64 65Next, Kprobes single-steps its copy of the probed instruction. 66(It would be simpler to single-step the actual instruction in place, 67but then Kprobes would have to temporarily remove the breakpoint 68instruction. This would open a small time window when another CPU 69could sail right past the probepoint.) 70 71After the instruction is single-stepped, Kprobes executes the 72"post_handler," if any, that is associated with the kprobe. 73Execution then continues with the instruction following the probepoint. 74 751.2 How Does a Jprobe Work? 76 77A jprobe is implemented using a kprobe that is placed on a function's 78entry point. It employs a simple mirroring principle to allow 79seamless access to the probed function's arguments. The jprobe 80handler routine should have the same signature (arg list and return 81type) as the function being probed, and must always end by calling 82the Kprobes function jprobe_return(). 83 84Here's how it works. When the probe is hit, Kprobes makes a copy of 85the saved registers and a generous portion of the stack (see below). 86Kprobes then points the saved instruction pointer at the jprobe's 87handler routine, and returns from the trap. As a result, control 88passes to the handler, which is presented with the same register and 89stack contents as the probed function. When it is done, the handler 90calls jprobe_return(), which traps again to restore the original stack 91contents and processor state and switch to the probed function. 92 93By convention, the callee owns its arguments, so gcc may produce code 94that unexpectedly modifies that portion of the stack. This is why 95Kprobes saves a copy of the stack and restores it after the jprobe 96handler has run. Up to MAX_STACK_SIZE bytes are copied -- e.g., 9764 bytes on i386. 98 99Note that the probed function's args may be passed on the stack 100or in registers. The jprobe will work in either case, so long as the 101handler's prototype matches that of the probed function. 102 1031.3 Return Probes 104 1051.3.1 How Does a Return Probe Work? 106 107When you call register_kretprobe(), Kprobes establishes a kprobe at 108the entry to the function. When the probed function is called and this 109probe is hit, Kprobes saves a copy of the return address, and replaces 110the return address with the address of a "trampoline." The trampoline 111is an arbitrary piece of code -- typically just a nop instruction. 112At boot time, Kprobes registers a kprobe at the trampoline. 113 114When the probed function executes its return instruction, control 115passes to the trampoline and that probe is hit. Kprobes' trampoline 116handler calls the user-specified return handler associated with the 117kretprobe, then sets the saved instruction pointer to the saved return 118address, and that's where execution resumes upon return from the trap. 119 120While the probed function is executing, its return address is 121stored in an object of type kretprobe_instance. Before calling 122register_kretprobe(), the user sets the maxactive field of the 123kretprobe struct to specify how many instances of the specified 124function can be probed simultaneously. register_kretprobe() 125pre-allocates the indicated number of kretprobe_instance objects. 126 127For example, if the function is non-recursive and is called with a 128spinlock held, maxactive = 1 should be enough. If the function is 129non-recursive and can never relinquish the CPU (e.g., via a semaphore 130or preemption), NR_CPUS should be enough. If maxactive <= 0, it is 131set to a default value. If CONFIG_PREEMPT is enabled, the default 132is max(10, 2*NR_CPUS). Otherwise, the default is NR_CPUS. 133 134It's not a disaster if you set maxactive too low; you'll just miss 135some probes. In the kretprobe struct, the nmissed field is set to 136zero when the return probe is registered, and is incremented every 137time the probed function is entered but there is no kretprobe_instance 138object available for establishing the return probe. 139 1401.3.2 Kretprobe entry-handler 141 142Kretprobes also provides an optional user-specified handler which runs 143on function entry. This handler is specified by setting the entry_handler 144field of the kretprobe struct. Whenever the kprobe placed by kretprobe at the 145function entry is hit, the user-defined entry_handler, if any, is invoked. 146If the entry_handler returns 0 (success) then a corresponding return handler 147is guaranteed to be called upon function return. If the entry_handler 148returns a non-zero error then Kprobes leaves the return address as is, and 149the kretprobe has no further effect for that particular function instance. 150 151Multiple entry and return handler invocations are matched using the unique 152kretprobe_instance object associated with them. Additionally, a user 153may also specify per return-instance private data to be part of each 154kretprobe_instance object. This is especially useful when sharing private 155data between corresponding user entry and return handlers. The size of each 156private data object can be specified at kretprobe registration time by 157setting the data_size field of the kretprobe struct. This data can be 158accessed through the data field of each kretprobe_instance object. 159 160In case probed function is entered but there is no kretprobe_instance 161object available, then in addition to incrementing the nmissed count, 162the user entry_handler invocation is also skipped. 163 1642. Architectures Supported 165 166Kprobes, jprobes, and return probes are implemented on the following 167architectures: 168 169- i386 170- x86_64 (AMD-64, EM64T) 171- ppc64 172- ia64 (Does not support probes on instruction slot1.) 173- sparc64 (Return probes not yet implemented.) 174- arm 175- ppc 176 1773. Configuring Kprobes 178 179When configuring the kernel using make menuconfig/xconfig/oldconfig, 180ensure that CONFIG_KPROBES is set to "y". Under "Instrumentation 181Support", look for "Kprobes". 182 183So that you can load and unload Kprobes-based instrumentation modules, 184make sure "Loadable module support" (CONFIG_MODULES) and "Module 185unloading" (CONFIG_MODULE_UNLOAD) are set to "y". 186 187Also make sure that CONFIG_KALLSYMS and perhaps even CONFIG_KALLSYMS_ALL 188are set to "y", since kallsyms_lookup_name() is used by the in-kernel 189kprobe address resolution code. 190 191If you need to insert a probe in the middle of a function, you may find 192it useful to "Compile the kernel with debug info" (CONFIG_DEBUG_INFO), 193so you can use "objdump -d -l vmlinux" to see the source-to-object 194code mapping. 195 1964. API Reference 197 198The Kprobes API includes a "register" function and an "unregister" 199function for each type of probe. The API also includes "register_*probes" 200and "unregister_*probes" functions for (un)registering arrays of probes. 201Here are terse, mini-man-page specifications for these functions and 202the associated probe handlers that you'll write. See the files in the 203samples/kprobes/ sub-directory for examples. 204 2054.1 register_kprobe 206 207#include <linux/kprobes.h> 208int register_kprobe(struct kprobe *kp); 209 210Sets a breakpoint at the address kp->addr. When the breakpoint is 211hit, Kprobes calls kp->pre_handler. After the probed instruction 212is single-stepped, Kprobe calls kp->post_handler. If a fault 213occurs during execution of kp->pre_handler or kp->post_handler, 214or during single-stepping of the probed instruction, Kprobes calls 215kp->fault_handler. Any or all handlers can be NULL. If kp->flags 216is set KPROBE_FLAG_DISABLED, that kp will be registered but disabled, 217so, it's handlers aren't hit until calling enable_kprobe(kp). 218 219NOTE: 2201. With the introduction of the "symbol_name" field to struct kprobe, 221the probepoint address resolution will now be taken care of by the kernel. 222The following will now work: 223 224 kp.symbol_name = "symbol_name"; 225 226(64-bit powerpc intricacies such as function descriptors are handled 227transparently) 228 2292. Use the "offset" field of struct kprobe if the offset into the symbol 230to install a probepoint is known. This field is used to calculate the 231probepoint. 232 2333. Specify either the kprobe "symbol_name" OR the "addr". If both are 234specified, kprobe registration will fail with -EINVAL. 235 2364. With CISC architectures (such as i386 and x86_64), the kprobes code 237does not validate if the kprobe.addr is at an instruction boundary. 238Use "offset" with caution. 239 240register_kprobe() returns 0 on success, or a negative errno otherwise. 241 242User's pre-handler (kp->pre_handler): 243#include <linux/kprobes.h> 244#include <linux/ptrace.h> 245int pre_handler(struct kprobe *p, struct pt_regs *regs); 246 247Called with p pointing to the kprobe associated with the breakpoint, 248and regs pointing to the struct containing the registers saved when 249the breakpoint was hit. Return 0 here unless you're a Kprobes geek. 250 251User's post-handler (kp->post_handler): 252#include <linux/kprobes.h> 253#include <linux/ptrace.h> 254void post_handler(struct kprobe *p, struct pt_regs *regs, 255 unsigned long flags); 256 257p and regs are as described for the pre_handler. flags always seems 258to be zero. 259 260User's fault-handler (kp->fault_handler): 261#include <linux/kprobes.h> 262#include <linux/ptrace.h> 263int fault_handler(struct kprobe *p, struct pt_regs *regs, int trapnr); 264 265p and regs are as described for the pre_handler. trapnr is the 266architecture-specific trap number associated with the fault (e.g., 267on i386, 13 for a general protection fault or 14 for a page fault). 268Returns 1 if it successfully handled the exception. 269 2704.2 register_jprobe 271 272#include <linux/kprobes.h> 273int register_jprobe(struct jprobe *jp) 274 275Sets a breakpoint at the address jp->kp.addr, which must be the address 276of the first instruction of a function. When the breakpoint is hit, 277Kprobes runs the handler whose address is jp->entry. 278 279The handler should have the same arg list and return type as the probed 280function; and just before it returns, it must call jprobe_return(). 281(The handler never actually returns, since jprobe_return() returns 282control to Kprobes.) If the probed function is declared asmlinkage 283or anything else that affects how args are passed, the handler's 284declaration must match. 285 286register_jprobe() returns 0 on success, or a negative errno otherwise. 287 2884.3 register_kretprobe 289 290#include <linux/kprobes.h> 291int register_kretprobe(struct kretprobe *rp); 292 293Establishes a return probe for the function whose address is 294rp->kp.addr. When that function returns, Kprobes calls rp->handler. 295You must set rp->maxactive appropriately before you call 296register_kretprobe(); see "How Does a Return Probe Work?" for details. 297 298register_kretprobe() returns 0 on success, or a negative errno 299otherwise. 300 301User's return-probe handler (rp->handler): 302#include <linux/kprobes.h> 303#include <linux/ptrace.h> 304int kretprobe_handler(struct kretprobe_instance *ri, struct pt_regs *regs); 305 306regs is as described for kprobe.pre_handler. ri points to the 307kretprobe_instance object, of which the following fields may be 308of interest: 309- ret_addr: the return address 310- rp: points to the corresponding kretprobe object 311- task: points to the corresponding task struct 312- data: points to per return-instance private data; see "Kretprobe 313 entry-handler" for details. 314 315The regs_return_value(regs) macro provides a simple abstraction to 316extract the return value from the appropriate register as defined by 317the architecture's ABI. 318 319The handler's return value is currently ignored. 320 3214.4 unregister_*probe 322 323#include <linux/kprobes.h> 324void unregister_kprobe(struct kprobe *kp); 325void unregister_jprobe(struct jprobe *jp); 326void unregister_kretprobe(struct kretprobe *rp); 327 328Removes the specified probe. The unregister function can be called 329at any time after the probe has been registered. 330 331NOTE: 332If the functions find an incorrect probe (ex. an unregistered probe), 333they clear the addr field of the probe. 334 3354.5 register_*probes 336 337#include <linux/kprobes.h> 338int register_kprobes(struct kprobe **kps, int num); 339int register_kretprobes(struct kretprobe **rps, int num); 340int register_jprobes(struct jprobe **jps, int num); 341 342Registers each of the num probes in the specified array. If any 343error occurs during registration, all probes in the array, up to 344the bad probe, are safely unregistered before the register_*probes 345function returns. 346- kps/rps/jps: an array of pointers to *probe data structures 347- num: the number of the array entries. 348 349NOTE: 350You have to allocate(or define) an array of pointers and set all 351of the array entries before using these functions. 352 3534.6 unregister_*probes 354 355#include <linux/kprobes.h> 356void unregister_kprobes(struct kprobe **kps, int num); 357void unregister_kretprobes(struct kretprobe **rps, int num); 358void unregister_jprobes(struct jprobe **jps, int num); 359 360Removes each of the num probes in the specified array at once. 361 362NOTE: 363If the functions find some incorrect probes (ex. unregistered 364probes) in the specified array, they clear the addr field of those 365incorrect probes. However, other probes in the array are 366unregistered correctly. 367 3684.7 disable_*probe 369 370#include <linux/kprobes.h> 371int disable_kprobe(struct kprobe *kp); 372int disable_kretprobe(struct kretprobe *rp); 373int disable_jprobe(struct jprobe *jp); 374 375Temporarily disables the specified *probe. You can enable it again by using 376enable_*probe(). You must specify the probe which has been registered. 377 3784.8 enable_*probe 379 380#include <linux/kprobes.h> 381int enable_kprobe(struct kprobe *kp); 382int enable_kretprobe(struct kretprobe *rp); 383int enable_jprobe(struct jprobe *jp); 384 385Enables *probe which has been disabled by disable_*probe(). You must specify 386the probe which has been registered. 387 3885. Kprobes Features and Limitations 389 390Kprobes allows multiple probes at the same address. Currently, 391however, there cannot be multiple jprobes on the same function at 392the same time. 393 394In general, you can install a probe anywhere in the kernel. 395In particular, you can probe interrupt handlers. Known exceptions 396are discussed in this section. 397 398The register_*probe functions will return -EINVAL if you attempt 399to install a probe in the code that implements Kprobes (mostly 400kernel/kprobes.c and arch/*/kernel/kprobes.c, but also functions such 401as do_page_fault and notifier_call_chain). 402 403If you install a probe in an inline-able function, Kprobes makes 404no attempt to chase down all inline instances of the function and 405install probes there. gcc may inline a function without being asked, 406so keep this in mind if you're not seeing the probe hits you expect. 407 408A probe handler can modify the environment of the probed function 409-- e.g., by modifying kernel data structures, or by modifying the 410contents of the pt_regs struct (which are restored to the registers 411upon return from the breakpoint). So Kprobes can be used, for example, 412to install a bug fix or to inject faults for testing. Kprobes, of 413course, has no way to distinguish the deliberately injected faults 414from the accidental ones. Don't drink and probe. 415 416Kprobes makes no attempt to prevent probe handlers from stepping on 417each other -- e.g., probing printk() and then calling printk() from a 418probe handler. If a probe handler hits a probe, that second probe's 419handlers won't be run in that instance, and the kprobe.nmissed member 420of the second probe will be incremented. 421 422As of Linux v2.6.15-rc1, multiple handlers (or multiple instances of 423the same handler) may run concurrently on different CPUs. 424 425Kprobes does not use mutexes or allocate memory except during 426registration and unregistration. 427 428Probe handlers are run with preemption disabled. Depending on the 429architecture, handlers may also run with interrupts disabled. In any 430case, your handler should not yield the CPU (e.g., by attempting to 431acquire a semaphore). 432 433Since a return probe is implemented by replacing the return 434address with the trampoline's address, stack backtraces and calls 435to __builtin_return_address() will typically yield the trampoline's 436address instead of the real return address for kretprobed functions. 437(As far as we can tell, __builtin_return_address() is used only 438for instrumentation and error reporting.) 439 440If the number of times a function is called does not match the number 441of times it returns, registering a return probe on that function may 442produce undesirable results. In such a case, a line: 443kretprobe BUG!: Processing kretprobe d000000000041aa8 @ c00000000004f48c 444gets printed. With this information, one will be able to correlate the 445exact instance of the kretprobe that caused the problem. We have the 446do_exit() case covered. do_execve() and do_fork() are not an issue. 447We're unaware of other specific cases where this could be a problem. 448 449If, upon entry to or exit from a function, the CPU is running on 450a stack other than that of the current task, registering a return 451probe on that function may produce undesirable results. For this 452reason, Kprobes doesn't support return probes (or kprobes or jprobes) 453on the x86_64 version of __switch_to(); the registration functions 454return -EINVAL. 455 4566. Probe Overhead 457 458On a typical CPU in use in 2005, a kprobe hit takes 0.5 to 1.0 459microseconds to process. Specifically, a benchmark that hits the same 460probepoint repeatedly, firing a simple handler each time, reports 1-2 461million hits per second, depending on the architecture. A jprobe or 462return-probe hit typically takes 50-75% longer than a kprobe hit. 463When you have a return probe set on a function, adding a kprobe at 464the entry to that function adds essentially no overhead. 465 466Here are sample overhead figures (in usec) for different architectures. 467k = kprobe; j = jprobe; r = return probe; kr = kprobe + return probe 468on same function; jr = jprobe + return probe on same function 469 470i386: Intel Pentium M, 1495 MHz, 2957.31 bogomips 471k = 0.57 usec; j = 1.00; r = 0.92; kr = 0.99; jr = 1.40 472 473x86_64: AMD Opteron 246, 1994 MHz, 3971.48 bogomips 474k = 0.49 usec; j = 0.76; r = 0.80; kr = 0.82; jr = 1.07 475 476ppc64: POWER5 (gr), 1656 MHz (SMT disabled, 1 virtual CPU per physical CPU) 477k = 0.77 usec; j = 1.31; r = 1.26; kr = 1.45; jr = 1.99 478 4797. TODO 480 481a. SystemTap (http://sourceware.org/systemtap): Provides a simplified 482programming interface for probe-based instrumentation. Try it out. 483b. Kernel return probes for sparc64. 484c. Support for other architectures. 485d. User-space probes. 486e. Watchpoint probes (which fire on data references). 487 4888. Kprobes Example 489 490See samples/kprobes/kprobe_example.c 491 4929. Jprobes Example 493 494See samples/kprobes/jprobe_example.c 495 49610. Kretprobes Example 497 498See samples/kprobes/kretprobe_example.c 499 500For additional information on Kprobes, refer to the following URLs: 501http://www-106.ibm.com/developerworks/library/l-kprobes.html?ca=dgr-lnxw42Kprobe 502http://www.redhat.com/magazine/005mar05/features/kprobes/ 503http://www-users.cs.umn.edu/~boutcher/kprobes/ 504http://www.linuxsymposium.org/2006/linuxsymposium_procv2.pdf (pages 101-115) 505 506 507Appendix A: The kprobes debugfs interface 508 509With recent kernels (> 2.6.20) the list of registered kprobes is visible 510under the /sys/kernel/debug/kprobes/ directory (assuming debugfs is mounted at //sys/kernel/debug). 511 512/sys/kernel/debug/kprobes/list: Lists all registered probes on the system 513 514c015d71a k vfs_read+0x0 515c011a316 j do_fork+0x0 516c03dedc5 r tcp_v4_rcv+0x0 517 518The first column provides the kernel address where the probe is inserted. 519The second column identifies the type of probe (k - kprobe, r - kretprobe 520and j - jprobe), while the third column specifies the symbol+offset of 521the probe. If the probed function belongs to a module, the module name 522is also specified. Following columns show probe status. If the probe is on 523a virtual address that is no longer valid (module init sections, module 524virtual addresses that correspond to modules that've been unloaded), 525such probes are marked with [GONE]. If the probe is temporarily disabled, 526such probes are marked with [DISABLED]. 527 528/sys/kernel/debug/kprobes/enabled: Turn kprobes ON/OFF forcibly. 529 530Provides a knob to globally and forcibly turn registered kprobes ON or OFF. 531By default, all kprobes are enabled. By echoing "0" to this file, all 532registered probes will be disarmed, till such time a "1" is echoed to this 533file. Note that this knob just disarms and arms all kprobes and doesn't 534change each probe's disabling state. This means that disabled kprobes (marked 535[DISABLED]) will be not enabled if you turn ON all kprobes by this knob. 536