linux/Documentation/kprobes.txt
<<
>>
Prefs
   1Title   : Kernel Probes (Kprobes)
   2Authors : Jim Keniston <jkenisto@us.ibm.com>
   3        : Prasanna S Panchamukhi <prasanna.panchamukhi@gmail.com>
   4        : Masami Hiramatsu <mhiramat@redhat.com>
   5
   6CONTENTS
   7
   81. Concepts: Kprobes, Jprobes, Return Probes
   92. Architectures Supported
  103. Configuring Kprobes
  114. API Reference
  125. Kprobes Features and Limitations
  136. Probe Overhead
  147. TODO
  158. Kprobes Example
  169. Jprobes Example
  1710. Kretprobes Example
  18Appendix A: The kprobes debugfs interface
  19Appendix B: The kprobes sysctl interface
  20
  211. Concepts: Kprobes, Jprobes, Return Probes
  22
  23Kprobes enables you to dynamically break into any kernel routine and
  24collect debugging and performance information non-disruptively. You
  25can trap at almost any kernel code address, specifying a handler
  26routine to be invoked when the breakpoint is hit.
  27
  28There are currently three types of probes: kprobes, jprobes, and
  29kretprobes (also called return probes).  A kprobe can be inserted
  30on virtually any instruction in the kernel.  A jprobe is inserted at
  31the entry to a kernel function, and provides convenient access to the
  32function's arguments.  A return probe fires when a specified function
  33returns.
  34
  35In the typical case, Kprobes-based instrumentation is packaged as
  36a kernel module.  The module's init function installs ("registers")
  37one or more probes, and the exit function unregisters them.  A
  38registration function such as register_kprobe() specifies where
  39the probe is to be inserted and what handler is to be called when
  40the probe is hit.
  41
  42There are also register_/unregister_*probes() functions for batch
  43registration/unregistration of a group of *probes. These functions
  44can speed up unregistration process when you have to unregister
  45a lot of probes at once.
  46
  47The next four subsections explain how the different types of
  48probes work and how jump optimization works.  They explain certain
  49things that you'll need to know in order to make the best use of
  50Kprobes -- e.g., the difference between a pre_handler and
  51a post_handler, and how to use the maxactive and nmissed fields of
  52a kretprobe.  But if you're in a hurry to start using Kprobes, you
  53can skip ahead to section 2.
  54
  551.1 How Does a Kprobe Work?
  56
  57When a kprobe is registered, Kprobes makes a copy of the probed
  58instruction and replaces the first byte(s) of the probed instruction
  59with a breakpoint instruction (e.g., int3 on i386 and x86_64).
  60
  61When a CPU hits the breakpoint instruction, a trap occurs, the CPU's
  62registers are saved, and control passes to Kprobes via the
  63notifier_call_chain mechanism.  Kprobes executes the "pre_handler"
  64associated with the kprobe, passing the handler the addresses of the
  65kprobe struct and the saved registers.
  66
  67Next, Kprobes single-steps its copy of the probed instruction.
  68(It would be simpler to single-step the actual instruction in place,
  69but then Kprobes would have to temporarily remove the breakpoint
  70instruction.  This would open a small time window when another CPU
  71could sail right past the probepoint.)
  72
  73After the instruction is single-stepped, Kprobes executes the
  74"post_handler," if any, that is associated with the kprobe.
  75Execution then continues with the instruction following the probepoint.
  76
  771.2 How Does a Jprobe Work?
  78
  79A jprobe is implemented using a kprobe that is placed on a function's
  80entry point.  It employs a simple mirroring principle to allow
  81seamless access to the probed function's arguments.  The jprobe
  82handler routine should have the same signature (arg list and return
  83type) as the function being probed, and must always end by calling
  84the Kprobes function jprobe_return().
  85
  86Here's how it works.  When the probe is hit, Kprobes makes a copy of
  87the saved registers and a generous portion of the stack (see below).
  88Kprobes then points the saved instruction pointer at the jprobe's
  89handler routine, and returns from the trap.  As a result, control
  90passes to the handler, which is presented with the same register and
  91stack contents as the probed function.  When it is done, the handler
  92calls jprobe_return(), which traps again to restore the original stack
  93contents and processor state and switch to the probed function.
  94
  95By convention, the callee owns its arguments, so gcc may produce code
  96that unexpectedly modifies that portion of the stack.  This is why
  97Kprobes saves a copy of the stack and restores it after the jprobe
  98handler has run.  Up to MAX_STACK_SIZE bytes are copied -- e.g.,
  9964 bytes on i386.
 100
 101Note that the probed function's args may be passed on the stack
 102or in registers.  The jprobe will work in either case, so long as the
 103handler's prototype matches that of the probed function.
 104
 1051.3 Return Probes
 106
 1071.3.1 How Does a Return Probe Work?
 108
 109When you call register_kretprobe(), Kprobes establishes a kprobe at
 110the entry to the function.  When the probed function is called and this
 111probe is hit, Kprobes saves a copy of the return address, and replaces
 112the return address with the address of a "trampoline."  The trampoline
 113is an arbitrary piece of code -- typically just a nop instruction.
 114At boot time, Kprobes registers a kprobe at the trampoline.
 115
 116When the probed function executes its return instruction, control
 117passes to the trampoline and that probe is hit.  Kprobes' trampoline
 118handler calls the user-specified return handler associated with the
 119kretprobe, then sets the saved instruction pointer to the saved return
 120address, and that's where execution resumes upon return from the trap.
 121
 122While the probed function is executing, its return address is
 123stored in an object of type kretprobe_instance.  Before calling
 124register_kretprobe(), the user sets the maxactive field of the
 125kretprobe struct to specify how many instances of the specified
 126function can be probed simultaneously.  register_kretprobe()
 127pre-allocates the indicated number of kretprobe_instance objects.
 128
 129For example, if the function is non-recursive and is called with a
 130spinlock held, maxactive = 1 should be enough.  If the function is
 131non-recursive and can never relinquish the CPU (e.g., via a semaphore
 132or preemption), NR_CPUS should be enough.  If maxactive <= 0, it is
 133set to a default value.  If CONFIG_PREEMPT is enabled, the default
 134is max(10, 2*NR_CPUS).  Otherwise, the default is NR_CPUS.
 135
 136It's not a disaster if you set maxactive too low; you'll just miss
 137some probes.  In the kretprobe struct, the nmissed field is set to
 138zero when the return probe is registered, and is incremented every
 139time the probed function is entered but there is no kretprobe_instance
 140object available for establishing the return probe.
 141
 1421.3.2 Kretprobe entry-handler
 143
 144Kretprobes also provides an optional user-specified handler which runs
 145on function entry. This handler is specified by setting the entry_handler
 146field of the kretprobe struct. Whenever the kprobe placed by kretprobe at the
 147function entry is hit, the user-defined entry_handler, if any, is invoked.
 148If the entry_handler returns 0 (success) then a corresponding return handler
 149is guaranteed to be called upon function return. If the entry_handler
 150returns a non-zero error then Kprobes leaves the return address as is, and
 151the kretprobe has no further effect for that particular function instance.
 152
 153Multiple entry and return handler invocations are matched using the unique
 154kretprobe_instance object associated with them. Additionally, a user
 155may also specify per return-instance private data to be part of each
 156kretprobe_instance object. This is especially useful when sharing private
 157data between corresponding user entry and return handlers. The size of each
 158private data object can be specified at kretprobe registration time by
 159setting the data_size field of the kretprobe struct. This data can be
 160accessed through the data field of each kretprobe_instance object.
 161
 162In case probed function is entered but there is no kretprobe_instance
 163object available, then in addition to incrementing the nmissed count,
 164the user entry_handler invocation is also skipped.
 165
 1661.4 How Does Jump Optimization Work?
 167
 168If your kernel is built with CONFIG_OPTPROBES=y (currently this flag
 169is automatically set 'y' on x86/x86-64, non-preemptive kernel) and
 170the "debug.kprobes_optimization" kernel parameter is set to 1 (see
 171sysctl(8)), Kprobes tries to reduce probe-hit overhead by using a jump
 172instruction instead of a breakpoint instruction at each probepoint.
 173
 1741.4.1 Init a Kprobe
 175
 176When a probe is registered, before attempting this optimization,
 177Kprobes inserts an ordinary, breakpoint-based kprobe at the specified
 178address. So, even if it's not possible to optimize this particular
 179probepoint, there'll be a probe there.
 180
 1811.4.2 Safety Check
 182
 183Before optimizing a probe, Kprobes performs the following safety checks:
 184
 185- Kprobes verifies that the region that will be replaced by the jump
 186instruction (the "optimized region") lies entirely within one function.
 187(A jump instruction is multiple bytes, and so may overlay multiple
 188instructions.)
 189
 190- Kprobes analyzes the entire function and verifies that there is no
 191jump into the optimized region.  Specifically:
 192  - the function contains no indirect jump;
 193  - the function contains no instruction that causes an exception (since
 194  the fixup code triggered by the exception could jump back into the
 195  optimized region -- Kprobes checks the exception tables to verify this);
 196  and
 197  - there is no near jump to the optimized region (other than to the first
 198  byte).
 199
 200- For each instruction in the optimized region, Kprobes verifies that
 201the instruction can be executed out of line.
 202
 2031.4.3 Preparing Detour Buffer
 204
 205Next, Kprobes prepares a "detour" buffer, which contains the following
 206instruction sequence:
 207- code to push the CPU's registers (emulating a breakpoint trap)
 208- a call to the trampoline code which calls user's probe handlers.
 209- code to restore registers
 210- the instructions from the optimized region
 211- a jump back to the original execution path.
 212
 2131.4.4 Pre-optimization
 214
 215After preparing the detour buffer, Kprobes verifies that none of the
 216following situations exist:
 217- The probe has either a break_handler (i.e., it's a jprobe) or a
 218post_handler.
 219- Other instructions in the optimized region are probed.
 220- The probe is disabled.
 221In any of the above cases, Kprobes won't start optimizing the probe.
 222Since these are temporary situations, Kprobes tries to start
 223optimizing it again if the situation is changed.
 224
 225If the kprobe can be optimized, Kprobes enqueues the kprobe to an
 226optimizing list, and kicks the kprobe-optimizer workqueue to optimize
 227it.  If the to-be-optimized probepoint is hit before being optimized,
 228Kprobes returns control to the original instruction path by setting
 229the CPU's instruction pointer to the copied code in the detour buffer
 230-- thus at least avoiding the single-step.
 231
 2321.4.5 Optimization
 233
 234The Kprobe-optimizer doesn't insert the jump instruction immediately;
 235rather, it calls synchronize_sched() for safety first, because it's
 236possible for a CPU to be interrupted in the middle of executing the
 237optimized region(*).  As you know, synchronize_sched() can ensure
 238that all interruptions that were active when synchronize_sched()
 239was called are done, but only if CONFIG_PREEMPT=n.  So, this version
 240of kprobe optimization supports only kernels with CONFIG_PREEMPT=n.(**)
 241
 242After that, the Kprobe-optimizer calls stop_machine() to replace
 243the optimized region with a jump instruction to the detour buffer,
 244using text_poke_smp().
 245
 2461.4.6 Unoptimization
 247
 248When an optimized kprobe is unregistered, disabled, or blocked by
 249another kprobe, it will be unoptimized.  If this happens before
 250the optimization is complete, the kprobe is just dequeued from the
 251optimized list.  If the optimization has been done, the jump is
 252replaced with the original code (except for an int3 breakpoint in
 253the first byte) by using text_poke_smp().
 254
 255(*)Please imagine that the 2nd instruction is interrupted and then
 256the optimizer replaces the 2nd instruction with the jump *address*
 257while the interrupt handler is running. When the interrupt
 258returns to original address, there is no valid instruction,
 259and it causes an unexpected result.
 260
 261(**)This optimization-safety checking may be replaced with the
 262stop-machine method that ksplice uses for supporting a CONFIG_PREEMPT=y
 263kernel.
 264
 265NOTE for geeks:
 266The jump optimization changes the kprobe's pre_handler behavior.
 267Without optimization, the pre_handler can change the kernel's execution
 268path by changing regs->ip and returning 1.  However, when the probe
 269is optimized, that modification is ignored.  Thus, if you want to
 270tweak the kernel's execution path, you need to suppress optimization,
 271using one of the following techniques:
 272- Specify an empty function for the kprobe's post_handler or break_handler.
 273 or
 274- Execute 'sysctl -w debug.kprobes_optimization=n'
 275
 2762. Architectures Supported
 277
 278Kprobes, jprobes, and return probes are implemented on the following
 279architectures:
 280
 281- i386 (Supports jump optimization)
 282- x86_64 (AMD-64, EM64T) (Supports jump optimization)
 283- ppc64
 284- ia64 (Does not support probes on instruction slot1.)
 285- sparc64 (Return probes not yet implemented.)
 286- arm
 287- ppc
 288- mips
 289
 2903. Configuring Kprobes
 291
 292When configuring the kernel using make menuconfig/xconfig/oldconfig,
 293ensure that CONFIG_KPROBES is set to "y".  Under "Instrumentation
 294Support", look for "Kprobes".
 295
 296So that you can load and unload Kprobes-based instrumentation modules,
 297make sure "Loadable module support" (CONFIG_MODULES) and "Module
 298unloading" (CONFIG_MODULE_UNLOAD) are set to "y".
 299
 300Also make sure that CONFIG_KALLSYMS and perhaps even CONFIG_KALLSYMS_ALL
 301are set to "y", since kallsyms_lookup_name() is used by the in-kernel
 302kprobe address resolution code.
 303
 304If you need to insert a probe in the middle of a function, you may find
 305it useful to "Compile the kernel with debug info" (CONFIG_DEBUG_INFO),
 306so you can use "objdump -d -l vmlinux" to see the source-to-object
 307code mapping.
 308
 3094. API Reference
 310
 311The Kprobes API includes a "register" function and an "unregister"
 312function for each type of probe. The API also includes "register_*probes"
 313and "unregister_*probes" functions for (un)registering arrays of probes.
 314Here are terse, mini-man-page specifications for these functions and
 315the associated probe handlers that you'll write. See the files in the
 316samples/kprobes/ sub-directory for examples.
 317
 3184.1 register_kprobe
 319
 320#include <linux/kprobes.h>
 321int register_kprobe(struct kprobe *kp);
 322
 323Sets a breakpoint at the address kp->addr.  When the breakpoint is
 324hit, Kprobes calls kp->pre_handler.  After the probed instruction
 325is single-stepped, Kprobe calls kp->post_handler.  If a fault
 326occurs during execution of kp->pre_handler or kp->post_handler,
 327or during single-stepping of the probed instruction, Kprobes calls
 328kp->fault_handler.  Any or all handlers can be NULL. If kp->flags
 329is set KPROBE_FLAG_DISABLED, that kp will be registered but disabled,
 330so, its handlers aren't hit until calling enable_kprobe(kp).
 331
 332NOTE:
 3331. With the introduction of the "symbol_name" field to struct kprobe,
 334the probepoint address resolution will now be taken care of by the kernel.
 335The following will now work:
 336
 337        kp.symbol_name = "symbol_name";
 338
 339(64-bit powerpc intricacies such as function descriptors are handled
 340transparently)
 341
 3422. Use the "offset" field of struct kprobe if the offset into the symbol
 343to install a probepoint is known. This field is used to calculate the
 344probepoint.
 345
 3463. Specify either the kprobe "symbol_name" OR the "addr". If both are
 347specified, kprobe registration will fail with -EINVAL.
 348
 3494. With CISC architectures (such as i386 and x86_64), the kprobes code
 350does not validate if the kprobe.addr is at an instruction boundary.
 351Use "offset" with caution.
 352
 353register_kprobe() returns 0 on success, or a negative errno otherwise.
 354
 355User's pre-handler (kp->pre_handler):
 356#include <linux/kprobes.h>
 357#include <linux/ptrace.h>
 358int pre_handler(struct kprobe *p, struct pt_regs *regs);
 359
 360Called with p pointing to the kprobe associated with the breakpoint,
 361and regs pointing to the struct containing the registers saved when
 362the breakpoint was hit.  Return 0 here unless you're a Kprobes geek.
 363
 364User's post-handler (kp->post_handler):
 365#include <linux/kprobes.h>
 366#include <linux/ptrace.h>
 367void post_handler(struct kprobe *p, struct pt_regs *regs,
 368        unsigned long flags);
 369
 370p and regs are as described for the pre_handler.  flags always seems
 371to be zero.
 372
 373User's fault-handler (kp->fault_handler):
 374#include <linux/kprobes.h>
 375#include <linux/ptrace.h>
 376int fault_handler(struct kprobe *p, struct pt_regs *regs, int trapnr);
 377
 378p and regs are as described for the pre_handler.  trapnr is the
 379architecture-specific trap number associated with the fault (e.g.,
 380on i386, 13 for a general protection fault or 14 for a page fault).
 381Returns 1 if it successfully handled the exception.
 382
 3834.2 register_jprobe
 384
 385#include <linux/kprobes.h>
 386int register_jprobe(struct jprobe *jp)
 387
 388Sets a breakpoint at the address jp->kp.addr, which must be the address
 389of the first instruction of a function.  When the breakpoint is hit,
 390Kprobes runs the handler whose address is jp->entry.
 391
 392The handler should have the same arg list and return type as the probed
 393function; and just before it returns, it must call jprobe_return().
 394(The handler never actually returns, since jprobe_return() returns
 395control to Kprobes.)  If the probed function is declared asmlinkage
 396or anything else that affects how args are passed, the handler's
 397declaration must match.
 398
 399register_jprobe() returns 0 on success, or a negative errno otherwise.
 400
 4014.3 register_kretprobe
 402
 403#include <linux/kprobes.h>
 404int register_kretprobe(struct kretprobe *rp);
 405
 406Establishes a return probe for the function whose address is
 407rp->kp.addr.  When that function returns, Kprobes calls rp->handler.
 408You must set rp->maxactive appropriately before you call
 409register_kretprobe(); see "How Does a Return Probe Work?" for details.
 410
 411register_kretprobe() returns 0 on success, or a negative errno
 412otherwise.
 413
 414User's return-probe handler (rp->handler):
 415#include <linux/kprobes.h>
 416#include <linux/ptrace.h>
 417int kretprobe_handler(struct kretprobe_instance *ri, struct pt_regs *regs);
 418
 419regs is as described for kprobe.pre_handler.  ri points to the
 420kretprobe_instance object, of which the following fields may be
 421of interest:
 422- ret_addr: the return address
 423- rp: points to the corresponding kretprobe object
 424- task: points to the corresponding task struct
 425- data: points to per return-instance private data; see "Kretprobe
 426        entry-handler" for details.
 427
 428The regs_return_value(regs) macro provides a simple abstraction to
 429extract the return value from the appropriate register as defined by
 430the architecture's ABI.
 431
 432The handler's return value is currently ignored.
 433
 4344.4 unregister_*probe
 435
 436#include <linux/kprobes.h>
 437void unregister_kprobe(struct kprobe *kp);
 438void unregister_jprobe(struct jprobe *jp);
 439void unregister_kretprobe(struct kretprobe *rp);
 440
 441Removes the specified probe.  The unregister function can be called
 442at any time after the probe has been registered.
 443
 444NOTE:
 445If the functions find an incorrect probe (ex. an unregistered probe),
 446they clear the addr field of the probe.
 447
 4484.5 register_*probes
 449
 450#include <linux/kprobes.h>
 451int register_kprobes(struct kprobe **kps, int num);
 452int register_kretprobes(struct kretprobe **rps, int num);
 453int register_jprobes(struct jprobe **jps, int num);
 454
 455Registers each of the num probes in the specified array.  If any
 456error occurs during registration, all probes in the array, up to
 457the bad probe, are safely unregistered before the register_*probes
 458function returns.
 459- kps/rps/jps: an array of pointers to *probe data structures
 460- num: the number of the array entries.
 461
 462NOTE:
 463You have to allocate(or define) an array of pointers and set all
 464of the array entries before using these functions.
 465
 4664.6 unregister_*probes
 467
 468#include <linux/kprobes.h>
 469void unregister_kprobes(struct kprobe **kps, int num);
 470void unregister_kretprobes(struct kretprobe **rps, int num);
 471void unregister_jprobes(struct jprobe **jps, int num);
 472
 473Removes each of the num probes in the specified array at once.
 474
 475NOTE:
 476If the functions find some incorrect probes (ex. unregistered
 477probes) in the specified array, they clear the addr field of those
 478incorrect probes. However, other probes in the array are
 479unregistered correctly.
 480
 4814.7 disable_*probe
 482
 483#include <linux/kprobes.h>
 484int disable_kprobe(struct kprobe *kp);
 485int disable_kretprobe(struct kretprobe *rp);
 486int disable_jprobe(struct jprobe *jp);
 487
 488Temporarily disables the specified *probe. You can enable it again by using
 489enable_*probe(). You must specify the probe which has been registered.
 490
 4914.8 enable_*probe
 492
 493#include <linux/kprobes.h>
 494int enable_kprobe(struct kprobe *kp);
 495int enable_kretprobe(struct kretprobe *rp);
 496int enable_jprobe(struct jprobe *jp);
 497
 498Enables *probe which has been disabled by disable_*probe(). You must specify
 499the probe which has been registered.
 500
 5015. Kprobes Features and Limitations
 502
 503Kprobes allows multiple probes at the same address.  Currently,
 504however, there cannot be multiple jprobes on the same function at
 505the same time.  Also, a probepoint for which there is a jprobe or
 506a post_handler cannot be optimized.  So if you install a jprobe,
 507or a kprobe with a post_handler, at an optimized probepoint, the
 508probepoint will be unoptimized automatically.
 509
 510In general, you can install a probe anywhere in the kernel.
 511In particular, you can probe interrupt handlers.  Known exceptions
 512are discussed in this section.
 513
 514The register_*probe functions will return -EINVAL if you attempt
 515to install a probe in the code that implements Kprobes (mostly
 516kernel/kprobes.c and arch/*/kernel/kprobes.c, but also functions such
 517as do_page_fault and notifier_call_chain).
 518
 519If you install a probe in an inline-able function, Kprobes makes
 520no attempt to chase down all inline instances of the function and
 521install probes there.  gcc may inline a function without being asked,
 522so keep this in mind if you're not seeing the probe hits you expect.
 523
 524A probe handler can modify the environment of the probed function
 525-- e.g., by modifying kernel data structures, or by modifying the
 526contents of the pt_regs struct (which are restored to the registers
 527upon return from the breakpoint).  So Kprobes can be used, for example,
 528to install a bug fix or to inject faults for testing.  Kprobes, of
 529course, has no way to distinguish the deliberately injected faults
 530from the accidental ones.  Don't drink and probe.
 531
 532Kprobes makes no attempt to prevent probe handlers from stepping on
 533each other -- e.g., probing printk() and then calling printk() from a
 534probe handler.  If a probe handler hits a probe, that second probe's
 535handlers won't be run in that instance, and the kprobe.nmissed member
 536of the second probe will be incremented.
 537
 538As of Linux v2.6.15-rc1, multiple handlers (or multiple instances of
 539the same handler) may run concurrently on different CPUs.
 540
 541Kprobes does not use mutexes or allocate memory except during
 542registration and unregistration.
 543
 544Probe handlers are run with preemption disabled.  Depending on the
 545architecture and optimization state, handlers may also run with
 546interrupts disabled (e.g., kretprobe handlers and optimized kprobe
 547handlers run without interrupt disabled on x86/x86-64).  In any case,
 548your handler should not yield the CPU (e.g., by attempting to acquire
 549a semaphore).
 550
 551Since a return probe is implemented by replacing the return
 552address with the trampoline's address, stack backtraces and calls
 553to __builtin_return_address() will typically yield the trampoline's
 554address instead of the real return address for kretprobed functions.
 555(As far as we can tell, __builtin_return_address() is used only
 556for instrumentation and error reporting.)
 557
 558If the number of times a function is called does not match the number
 559of times it returns, registering a return probe on that function may
 560produce undesirable results. In such a case, a line:
 561kretprobe BUG!: Processing kretprobe d000000000041aa8 @ c00000000004f48c
 562gets printed. With this information, one will be able to correlate the
 563exact instance of the kretprobe that caused the problem. We have the
 564do_exit() case covered. do_execve() and do_fork() are not an issue.
 565We're unaware of other specific cases where this could be a problem.
 566
 567If, upon entry to or exit from a function, the CPU is running on
 568a stack other than that of the current task, registering a return
 569probe on that function may produce undesirable results.  For this
 570reason, Kprobes doesn't support return probes (or kprobes or jprobes)
 571on the x86_64 version of __switch_to(); the registration functions
 572return -EINVAL.
 573
 574On x86/x86-64, since the Jump Optimization of Kprobes modifies
 575instructions widely, there are some limitations to optimization. To
 576explain it, we introduce some terminology. Imagine a 3-instruction
 577sequence consisting of a two 2-byte instructions and one 3-byte
 578instruction.
 579
 580        IA
 581         |
 582[-2][-1][0][1][2][3][4][5][6][7]
 583        [ins1][ins2][  ins3 ]
 584        [<-     DCR       ->]
 585           [<- JTPR ->]
 586
 587ins1: 1st Instruction
 588ins2: 2nd Instruction
 589ins3: 3rd Instruction
 590IA:  Insertion Address
 591JTPR: Jump Target Prohibition Region
 592DCR: Detoured Code Region
 593
 594The instructions in DCR are copied to the out-of-line buffer
 595of the kprobe, because the bytes in DCR are replaced by
 596a 5-byte jump instruction. So there are several limitations.
 597
 598a) The instructions in DCR must be relocatable.
 599b) The instructions in DCR must not include a call instruction.
 600c) JTPR must not be targeted by any jump or call instruction.
 601d) DCR must not straddle the border between functions.
 602
 603Anyway, these limitations are checked by the in-kernel instruction
 604decoder, so you don't need to worry about that.
 605
 6066. Probe Overhead
 607
 608On a typical CPU in use in 2005, a kprobe hit takes 0.5 to 1.0
 609microseconds to process.  Specifically, a benchmark that hits the same
 610probepoint repeatedly, firing a simple handler each time, reports 1-2
 611million hits per second, depending on the architecture.  A jprobe or
 612return-probe hit typically takes 50-75% longer than a kprobe hit.
 613When you have a return probe set on a function, adding a kprobe at
 614the entry to that function adds essentially no overhead.
 615
 616Here are sample overhead figures (in usec) for different architectures.
 617k = kprobe; j = jprobe; r = return probe; kr = kprobe + return probe
 618on same function; jr = jprobe + return probe on same function
 619
 620i386: Intel Pentium M, 1495 MHz, 2957.31 bogomips
 621k = 0.57 usec; j = 1.00; r = 0.92; kr = 0.99; jr = 1.40
 622
 623x86_64: AMD Opteron 246, 1994 MHz, 3971.48 bogomips
 624k = 0.49 usec; j = 0.76; r = 0.80; kr = 0.82; jr = 1.07
 625
 626ppc64: POWER5 (gr), 1656 MHz (SMT disabled, 1 virtual CPU per physical CPU)
 627k = 0.77 usec; j = 1.31; r = 1.26; kr = 1.45; jr = 1.99
 628
 6296.1 Optimized Probe Overhead
 630
 631Typically, an optimized kprobe hit takes 0.07 to 0.1 microseconds to
 632process. Here are sample overhead figures (in usec) for x86 architectures.
 633k = unoptimized kprobe, b = boosted (single-step skipped), o = optimized kprobe,
 634r = unoptimized kretprobe, rb = boosted kretprobe, ro = optimized kretprobe.
 635
 636i386: Intel(R) Xeon(R) E5410, 2.33GHz, 4656.90 bogomips
 637k = 0.80 usec; b = 0.33; o = 0.05; r = 1.10; rb = 0.61; ro = 0.33
 638
 639x86-64: Intel(R) Xeon(R) E5410, 2.33GHz, 4656.90 bogomips
 640k = 0.99 usec; b = 0.43; o = 0.06; r = 1.24; rb = 0.68; ro = 0.30
 641
 6427. TODO
 643
 644a. SystemTap (http://sourceware.org/systemtap): Provides a simplified
 645programming interface for probe-based instrumentation.  Try it out.
 646b. Kernel return probes for sparc64.
 647c. Support for other architectures.
 648d. User-space probes.
 649e. Watchpoint probes (which fire on data references).
 650
 6518. Kprobes Example
 652
 653See samples/kprobes/kprobe_example.c
 654
 6559. Jprobes Example
 656
 657See samples/kprobes/jprobe_example.c
 658
 65910. Kretprobes Example
 660
 661See samples/kprobes/kretprobe_example.c
 662
 663For additional information on Kprobes, refer to the following URLs:
 664http://www-106.ibm.com/developerworks/library/l-kprobes.html?ca=dgr-lnxw42Kprobe
 665http://www.redhat.com/magazine/005mar05/features/kprobes/
 666http://www-users.cs.umn.edu/~boutcher/kprobes/
 667http://www.linuxsymposium.org/2006/linuxsymposium_procv2.pdf (pages 101-115)
 668
 669
 670Appendix A: The kprobes debugfs interface
 671
 672With recent kernels (> 2.6.20) the list of registered kprobes is visible
 673under the /sys/kernel/debug/kprobes/ directory (assuming debugfs is mounted at //sys/kernel/debug).
 674
 675/sys/kernel/debug/kprobes/list: Lists all registered probes on the system
 676
 677c015d71a  k  vfs_read+0x0
 678c011a316  j  do_fork+0x0
 679c03dedc5  r  tcp_v4_rcv+0x0
 680
 681The first column provides the kernel address where the probe is inserted.
 682The second column identifies the type of probe (k - kprobe, r - kretprobe
 683and j - jprobe), while the third column specifies the symbol+offset of
 684the probe. If the probed function belongs to a module, the module name
 685is also specified. Following columns show probe status. If the probe is on
 686a virtual address that is no longer valid (module init sections, module
 687virtual addresses that correspond to modules that've been unloaded),
 688such probes are marked with [GONE]. If the probe is temporarily disabled,
 689such probes are marked with [DISABLED]. If the probe is optimized, it is
 690marked with [OPTIMIZED].
 691
 692/sys/kernel/debug/kprobes/enabled: Turn kprobes ON/OFF forcibly.
 693
 694Provides a knob to globally and forcibly turn registered kprobes ON or OFF.
 695By default, all kprobes are enabled. By echoing "0" to this file, all
 696registered probes will be disarmed, till such time a "1" is echoed to this
 697file. Note that this knob just disarms and arms all kprobes and doesn't
 698change each probe's disabling state. This means that disabled kprobes (marked
 699[DISABLED]) will be not enabled if you turn ON all kprobes by this knob.
 700
 701
 702Appendix B: The kprobes sysctl interface
 703
 704/proc/sys/debug/kprobes-optimization: Turn kprobes optimization ON/OFF.
 705
 706When CONFIG_OPTPROBES=y, this sysctl interface appears and it provides
 707a knob to globally and forcibly turn jump optimization (see section
 7081.4) ON or OFF. By default, jump optimization is allowed (ON).
 709If you echo "0" to this file or set "debug.kprobes_optimization" to
 7100 via sysctl, all optimized probes will be unoptimized, and any new
 711probes registered after that will not be optimized.  Note that this
 712knob *changes* the optimized state. This means that optimized probes
 713(marked [OPTIMIZED]) will be unoptimized ([OPTIMIZED] tag will be
 714removed). If the knob is turned on, they will be optimized again.
 715
 716